SENSOR

ARRAY SIGNAL
PROCESSING

This Technical tutorial can be found on https://www.jotrin.com

 SENSOR
ARRAY SIGNAL
PROCESSING

Prabhakar S. Naidu

CRC Press
Boca Raton London New York Washington, D.C.
1195/Disclaimer Page 1 Monday, June 5, 2000 3:20 PM

Library of Congress Cataloging-in-Publication Data

Naidu, Prabhakar S.
Sensor array signal processing / Prabhakar S. Naidu.
p. cm.
Includes bibliographical references and index.
ISBN 0-8493-1195-0 (alk. paper)
1. Singal processing–Digital techniques. 2. Multisensor data fusion.
I. Title.
TK5102.9.N35 2000
621.382'2—dc21 00-030409
CIP

This book contains information obtained from authentic and highly regarded sources. Reprinted material
is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable
efforts have been made to publish reliable data and information, but the author and the publisher cannot
assume responsibility for the validity of all materials or for the consequences of their use.

Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic
or mechanical, including photocopying, microfilming, and recording, or by any information storage or
retrieval system, without prior permission in writing from the publisher.

The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for
creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC
for such copying.

Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation, without intent to infringe.

© 2001 by CRC Press LLC

No claim to original U.S. Government works

International Standard Book Number 0-8493-1195-0
Library of Congress Card Number 00-030409
Printed in the United States of America 1 2 3 4 5 6 7 8 9 0
Printed on acid-free paper
 Prologue

An array of sensors is often used in many diverse fields of science and

engineering, particularly where the goal is to study propagating wavefields.
Some examples are astronomy (radio astronomy), medical diagnosis, radar,
communication, sonar, nonrestrictive testing, seismology, and seismic
exploration (see [1] for different applications of the array signal processing).
The main goal of array signal processing is to deduce the following
information through an analysis of wavefields:

• (a) Source localization as in radar, sonar, astronomy, and seismology, etc.

• (b) Source waveform estimation as in communication, etc.
• (c) Source characterization as in seismology
• (d) Imaging of the scattering medium as in medical diagnosis, seismic
exploration, etc.

The tools of array signal processing remain the same, cutting across the
boundaries of different disciplines. For example, the basic tool of
beamformation is used in many areas mentioned above. The present book aims
at unraveling the underlying basic principles of array signal processing without
a reference to any particular application. However, an attempt is made to
include as many tools as possible from different disciplines in an order which
reflects the underlying principle.
In the real world, different types of wavefields are used in different
applications, for example, acoustic waves in sonar, mechanical waves in
seismic exploration, electromagnetic waves in radar and radio astronomy.
Fortunately, all wavefields can be characterized under identical mathematical
framework. This common mathematical framework is briefly summarized in
chapter 1. Here we have described the basic equations underlying different
wavefields and the structure of array signals and the background noise when the
noise sources follow some simple geometrical distribution. The topics covered
are wavefield in open space, bounded space including multipath propagation and
layered medium. Also covered is the weak scattering phenomenon which is the
basis for tomographic imaging. In chapter 2 we study different types of sensor
configurations. The emphasis is however on commonly used uniform linear
array (ULA), uniform circular array (UCA). Many practical sensor array
systems can be studied in terms of the basic ULA and UCA systems
(cylindrical array in radar and sonar, cross array in astronomy and seismology).
Like sensors, the sources can also be configured in the form of an array. The

© 2001 by CRC Press LLC

 source array is useful in synthesizing a desired wavefront and/or waveform. In
chapter 3 we examine the issues connected with the design of 2D digital filters
for wavefield analysis. Since the propagating wavefields possess some
interesting spectral characteristics in frequency wavenumber domain, for
example, the spectrum of a propagating wavefront is always on a radial line, it
is natural to take into account these features in the design of digital filters for
separation of interfering wavefields. Specifically, we cover in detail the design
of a fan filter and quadrant filter. Also, the classical Wiener filter as an
optimum least squares filter is covered in this chapter.
The theme in chapters 4 and 5 is localization of a source. In chapter 4
we describe the classical methods based on the frequency wavenumber spectrum
of the observed array output. We start with the Blackman Tukey type frequency
wavenumber spectrum and then go on to modern nonlinear high resolution
spectrum analysis methods such as Capon’s maximum likelihood spectrum
which is also known as minimum variance distortionless response (MVDR)
beamformer and maximum entropy spectrum. Localization essentially involves
estimation of parameters pertaining to the source position, for example,
azimuth and elevation angles, range, speed if the source is moving, etc. In the
last two decades a host of new methods of source localization have been
invented. We elaborate these new approaches in chapter 5. These include
subspace based methods, use of man-made signals such as in communication
and finally multipath environment. Quite often localization must be done in the
real time and it may be necessary to track a moving source. Adaptive techniques
are best suited for such tasks. A brief discussion on adaptive approach is
included. In chapter 6 we look into methods for source waveform separation and
estimation. The direction of arrival (DOA) is assumed to be known or has been
estimated. We shall describe a Wiener filter which minimizes the mean square
error in the estimation of the desired signal coming from a known direction and
a Capon filter which, while minimizing the power, ensures that the desired
signal is not distorted. We also talk about the estimation of direction of arrival
in a multipath environment encountered in wireless communication.
The next two chapters are devoted to array processing for imaging
purposes. Firstly, in chapter 7 we look at different types of tomographic
imaging systems: nondiffracting, diffracting and reflection tomography. The
received wavefield is inverted under the assumption of weak scattering to map
any one or more physical properties of the medium, for example, sound speed
variations in a medium. For objects of regular shape, scattering points play an
important role in geometrical diffraction theory. Estimation of these scattering
points for the determination of shape is also discussed. In chapter 8 we study
the method of wavefield extrapolation for imaging, extensively used in seismic

© 2001 by CRC Press LLC

exploration. The raw seismic traces are stacked in order to produce an output
trace from a hypothetical sensor kept close to the source (with zero- offset). A
suite of such stacked traces may be modeled as a wavefield recorded in an
imaginary experiment wherein small charges are placed on the reflector and
exploded at the same time. The zero-offset wavefield is used for imaging of
reflectors. The imaging process may be looked upon as a downward
continuation of the wavefield or inverse source problem or propagation
backward in time, i.e., depropagation to the reflector. All three view points are
very briefly described.
The book is based on a course entitled “Digital Array Processing”
offered to the graduate students who had already taken a course on digital signal
processing (DSP) and a course on modern spectrum analysis (MSA). It has
been my conviction that a student should be exposed to all basic concepts
cutting across the different disciplines without being burdened with the
questions of practical applications which are usually dealt with in specialty
courses. The most satisfying experience is that there is a common thread that
connects seemingly different tools used in different disciplines. An example is
beamformation, a commonly used tool in radar/sonar, which has a close
similarity with stacking used in seismic exploration. I have tried to bring out
in this exposition the common thread that exists in the analysis of wavefields
used in a wide variety of application areas. The proposed book has a
significantly different flavor, both in coverage and depth in comparison with the
ones on the market [1-5]. The first book, edited by Haykin, is a collection of
chapters, each devoted to an application. It rapidly surveys the state of art in
respective application areas but does not go deep enough and describe the basic
mathematical theory required for the understanding of array processing. The
second book by Ziomek is entirely devoted to array signal processing in
underwater acoustics. It covers in great depth the topic of beamformation by
linear and planar arrays but confines to linear methods. Modern array processing
tools do not find a place in this book. The third book by Pillai [3] has a very
narrow scope as it deals with in great detail only the subspace based methods.
The fourth book by Bouvet and Bienvenu (Eds) is again a collection of papers
largely devoted to modern subspace techniques. It is not suitable as a text.
Finally, the present book has some similarities with a book by Johnson and
Dudgeon [3] but differs in one important respect, namely, it does not cover the
application of arrays to imaging though a brief mention of tomography is
made. Also, the present book covers newer material which was not available at
the time of the publication of the book by Johnson and Dudgeon. During the
last two decades there has been intense research activity in the area of array
signal processing. There have been at least two review papers summarizing the

© 2001 by CRC Press LLC

 new results obtained during this period. The present book is not a research
monograph but it is an advanced level text which focuses on the important
developments which, the author believes, should be taught to give a broad
“picture” of array signal processing.
I have adopted the following plan of teaching. As the entire book
cannot be covered in one semester (about 35 hours) I preferred to cover it in two
parts in alternate semesters. In the first part, I covered chapter 1 (exclude §1.6),
chapter 2, chapters 4, 5 and 6. In the second part, I covered chapter 1, chapter 2
(exclude §2.3), chapter 3 (exclude §3.5), chapters 7 and 8. Exercises are given at
the end of each chapter. (The solution guide may be obtained from the
publisher).

1. S.Haykin (Ed), Array Signal Processing, Prentice Hall, Englewood Cliffs,

NJ, 1985.
2. L. J. Ziomek, Underwater Acoustics, A Linear Systems Theory, Academic
Press, Orlando, 1985.
3. S. U. Pillai, Array Signal Processing, Springer-Verlag, New York, 1989.
4. M. Bouvet and G. Bienvenu, High Resolution Methods in Underwater
Acoustics, Springer-Verlag, Berlin, 1991.
5. D. H. Johnson and D. E. Dudgeon, Array Signal Processing, Prentice Hall,
Englewood Cliffs, NJ, 1993.
6. H. Krim and M. Viberg, Two decades of array signl processing, IEEE Signal
Proc. Mag., July 1996, pp. 67-94.
7. T. Chen (Ed) Highlights of statistical signal and array processing, IEEE
Signal Proc. Mag., pp. 21-64, Sept. 1998.

Prabhakar S. Naidu February, 2000

Prof, Dept of ECE,
Indian Institute of Science,
Bangalore 560012, India.

© 2001 by CRC Press LLC

 Sensor Array Signal Processing

Contents

Chapter One

An Overview of Wavefields

1.1 Types of wavefields and the governing equations

1.2 Wavefield in open space
1.3 Wavefield in bounded space
1.4 Stochastic wavefield
1.5 Multipath propagation
1.6 Propagation through random medium
1.7 Exercises

Chapter Two

Sensor Array Systems

2.1 Uniform linear array (ULA)

2.2 Planar array
2.3 Broadband sensor array
2.4 Source and sensor arrays
2.5 Exercises

Chapter Three

Frequency Wavenumber Processing

3.1 Digital filters in the ω-k domain

3.2 Mapping of 1D into 2D filters
3.3 Multichannel Wiener filters
3.4 Wiener filters for ULA and UCA
3.5 Predictive noise cancellation
3.6 Exercises

Chapter Four

Source Localization: Frequency Wavenumber Spectrum

4.1 Frequency wavenumber spectrum

4.2 Beamformation
4.3 Capon's ω-k spectrum
4.4 Maximum entropy ω-k spectrum
4.5 Exercises
© 2001 by CRC Press LLC
 Chapter Five

Source Localization: Subspace Methods

5.1 Subspace methods (Narrowband)

5.2 Subspace methods (Broadband)
5.3 Coded signals
5.4 Array calibration
5.5 Source in bounded space
5.6 Exercises

Chapter Six

Source Estimation

6.1 Wiener filters

6.2 Minimum variance (Capon method)
6.3 Adaptive beamformation
6.4 Beamformation with coded signals
6.5 Multipath channel
6.6 Exercises

Chapter Seven

Tomographic Imaging

7.1 Nondiffracting radiation

7.2 Diffracting radiation
7.3 Broadband illumination
7.4 Reflection tomography
7.5 Object shape estimation
7.6 Exercises

Chapter Eight

Imaging by Wavefield Extrapolation

8.1 Migration
8.2 Exploding reflector model
8.3 Extrapolation in ω-k plane
8.4 Focused beam
8.5 Estimation of wave speed
8.6 Exercises

© 2001 by CRC Press LLC

 Acknowledgment
The thought of formalizing the lecture notes into a text occurred to me when I
was visiting the Rurh Universitaet, Bochum, Germany in 1996 as a Humboldt
Fellow. Much of the ground work was done during this period. I am grateful to
AvH Foundation who supported my stay. Prof Dr. J. F Boehme was my host. I
am grateful to him for the hospitality extended to me. Many of my students,
who credited the course on Array Signal Processing have contributed by way of
working out the exercises cited in the text. I am particularly grateful to the
following: S. Jena, S. S. Arun, P. Sexena, P. D. Pradeep, G. Viswanath, K.
Ganesh Kumar, Joby Joseph, V. Krishnagiri, N. B. Barkar. My graduate
students, Ms. A. Vasuki and Ms. A Buvaneswari, have significantly contributed
to chapter 7. Dr. K. V. S. Hari read the manuscript at an early stage and made
many constructive suggestions. I wish to thank the CRC Press Inc., in
particular, Ms. Nora Konopka and Ms Maggie Mogck for their promptness and
patience.

Finally I owe a deep gratitude to my family; my wife, Madhumati and

sons Srikanth, Sridhar and Srinath for their forbearance. I must specially thank
my son, Srinath who carefully scrutinized parts of the manuscript.

© 2001 by CRC Press LLC

 Dedication

This work is dedicated to the memory of the great visionary , J. R. D Tata

who shaped the Indian Institute of Science for many decades.

© 2001 by CRC Press LLC

 Chapter One

An Overview of Wavefields

A sensor array is used to measure wavefields and extract information about the
sources and the medium through which the wavefield propagates. It is therefore
imperative that some background in different types of wavefields and the basic
equations governing the wavefield must be acquired for complete understanding
of the principles of Array Signal Processing (ASP). In an idealistic
environment of open space, homogeneous medium and high frequency (where
ray approach is valid), a thorough understanding of the wave phenomenon may
not be necessary (those who are lucky enough to work in such an idealistic
environment may skip this chapter). But in a bounded inhomogeneous medium
and at low frequencies where diffraction phenomenon is dominating, the physics
of the waves plays a significant role in ASP algorithms. In this chapter our
aim is essentially to provide the basics of the physics of the waves which will
enable us to understand the complexities of the ASP problems in a more
realistic situation. The subject of wave physics is vast and naturally no attempt
is made to cover all its complexities.

§1.1 Types of Wavefields and Governing Equations:

The most commonly encountered wavefields are: (i) Acoustic waves including
sound waves, (ii) Mechanical waves in solids including vibrations and (iii)
Electromagnetic waves including light. The wavefields may be classified into
two types, namely, scalar and vector waves. In the scalar wavefield we have a
scalar physical quantity that propagates through the space, for example,
hydrostatic pressure is the physical quantity in acoustic scalar wavefields. In a
vector wavefield, the physical quantity involved is a vector, for example, the
displacement vector in mechanical waves, electric and magnetic vectors in
electromagnetic waves. A vector has three components all of which travel
independently in a homogeneous medium without any exchange of energy. But
at an interface separating two different media the components do interact. For
example, at an interface separating two solids a pressure wave will produce a
shear wave and vice versa. In a homogeneous medium without any reflecting
boundaries there is no energy transfer among components. Each component of a
vector field then behaves as if it is a scalar field, like an acoustic pressure field.
1.1.1 Acoustic Field: Acoustic field is a pressure (hydrostatic) field. The energy
is transmitted by means of propagation of compression and rarefaction waves.
The governing equation in a homogeneous medium is given by

ρ d 2φ
∇2φ = (1.1a)
γφ 0 dt 2

© 2001 by CRC Press LLC

 where φ 0 is ambient pressure, γ is ratio of specific heats at constant pressure
and volume and ρ is density. The wave propagation speed is given by

γφ 0 κ
c= = (1.1b)
ρ ρ

where κ is compressibility modulus and the wave propagates radially away

from the source. In an inhomogeneous medium the wave equation is given by

1 d 2φ 1
= ρ(r)∇ ⋅ ( ∇φ)
2
c (r) dt 2
ρ(r)
(1.2a)
1
= ρ(r)∇( ) ⋅ ∇φ + ∇ 2 φ
ρ(r)

After rearranging the terms in (1.2a) we obtain

1 d 2 φ ∇ρ(r)
∇2φ − = ⋅ ∇φ (1.2b)
c 2 (r) dt 2 ρ(r)

where r stands for position vector. The acoustic impedance is equal to the
product of density and propagation speed ρc and the admittance is given by the
inverse of the impedance or it is also defined in terms of the fluid speed and the
pressure,

fluid speed ∇φ
Acoustic admittance = =
pressure jωφ

Note that the acoustic impedance in air is 42 but in water it is 1.53 x105.
1.1.2 Mechanical Waves in Solids: The physical quantity which propagates is
the displacement vector, that is, particle displacement with respect to its
stationary position. Let d stand for the displacement vector. The wave equation
in a homogeneous medium is given by [1, p142]

∂2d
ρ = (2µ + λ)grad divd − µcurlcurld
∂t 2

where µ is shear constant and λ is Young’s modulus. In terms of these two

basic lame constants we define other more familiar parameters:

© 2001 by CRC Press LLC

 (2µ + λ)
Pressure wave speed: α=
ρ
µ
Shear wave speed: β=
ρ
λ
Poisson ratio: σ=
2(µ + λ)
2
Bulk modules κ = ( µ + λ)
3
The above parameters are observable from experimental data.
A displacement vector can be expressed as a sum of gradient of a scalar function
φ and curl of a vector function ψ (Helmholtz theorem)

d = ∇φ + ∇ × ψ (1.3a)

φ and ψ satisfy two different wave equations:

1 ∂2φ
∇2φ =
α 2 ∂t 2
(1.3b)
1 ∂2 ψ
∇×∇×ψ = − 2 2 ∇.ψ = 0
β ∂t

where ∇ × is a curl operator on a vector. The operator is defined as follows:

e x e y ez 
 
∂ ∂ ∂ 
∇ × F = det 
 ∂x ∂y ∂z 
 
 f x f y f z 
∂f z ∂f y ∂f ∂f ∂f y ∂f x
=( − )e x + ( x − z )e y + ( − )e z
∂y ∂z ∂z ∂x ∂x ∂y

where e x , e y and e z are unit vectors in the direction of x, y, and z,

respectively and f x , f y and f z components of vector F. The scalar potential
gives rise to longitudinal waves or pressure waves (p-waves) and the vector

© 2001 by CRC Press LLC

 potential gives rise to transverse waves or shear waves (s-waves). The p-waves
travel with speed α and s-waves travel with speed β . The components of
displacement vector can be expressed in terms of φ and ψ . From (1.3a) we
obtain

d = (d x , d y , dz )
∂φ ∂ψ z ∂ψ y
dx = + −
∂x ∂y ∂z
∂φ ∂ψ x ∂ψ z
dy = + − (1.4)
∂y ∂z ∂x
∂φ ∂ψ y ∂ψ x
dz = + −
∂z ∂x ∂y

where ψ = (ψ x , ψ y , ψ z ) . In solids we must speak of stress and strain

tensors. An element of solid is not only compressed but also twisted while an
element of fluid is only capable of being compressed but not twisted. We have
to use tensors for characterizing the phenomenon of twisting. We shall define
the stress and strain tensors and relate them through Hooke’s law. A stress
tensor is a matrix of nine components

sxx syx szx 

 
s = sxy syy szy  (1.5a)
s s s 
 xz yz zz 

The components of the stress tensor represent stresses on different faces of a

cube (see fig. 1.1). A strain tensor is given by

exx eyx ezx 

 
ε = exy eyy ezy  (1.5b)
e e e 
 xz yz zz 

The first subscript refers to the plane perpendicular to the axis denoted by the
subscript and the second subscript denotes the direction in which the vector is
pointing. For example, sxx is a stress in a plane perpendicular to the x-axis
(i.e., y-z plane) and pointing along the x-axis.

© 2001 by CRC Press LLC

 z

sxz
sxy
sxx

Figure 1.1: An element of volume (cuboid) and stresses are shown on a face
perpendicular to the x-axis.

The stress components on different faces of a cuboid are shown in fig. 1.1. The
torque on the cuboid should not cause any rotation. For this, we must have
sxy = syx and similarly all other nondiagonal elements in the stress matrix.
Thus, s must be a symmetric matrix. The components of a strain matrix are
related to the displacement vector

∂d x ∂d ∂d
ε xx = , ε yy = y , ε zz = z
∂x ∂y ∂z
∂d ∂d
ε xy = ε yx = x + y ;
∂y ∂x
∂d y ∂dz
ε yz = ε zy = + ; (1.6)
∂z ∂y
∂dz ∂d x
ε zx = ε xz = + .
∂x ∂z

Finally, the stress and strain components are related through Hooke’s Law:

sxx = ρα 2 ε xx + ρ(α 2 − 2β 2 )(ε yy + ε zz )

syy = ρα 2 ε yy + ρ(α 2 − 2β 2 )(ε xx + ε zz )
szz = ρα 2 ε zz + ρ(α 2 − 2β 2 )(ε xx + ε yy ) (1.7)

© 2001 by CRC Press LLC

 sxy = syx = ρβ2 ε xy
syz = szy = ρβ2 ε yz
szx = sxz = ρβ2 ε zx

Using equations (1.4), (1.6) and (1.7) we can express all nine stress
components in terms of the scalar and vector potential functions, φ and ψ .
For example, it is possible to show that sxx is given by

∂2φ ∂2φ ∂2φ 2 ∂ ψz

2
∂2 ψ y
sxx = ρα + ρ(α − 2β )( 2 + 2 ) + 2ρβ (
2 2 2
− )
∂x 2 ∂y ∂z ∂x∂y ∂x∂z

A general solution of (1.3b) may be given by

∞
1
∫ ∫ Φ(u,v,ω)e
− u 2 + v 2 − k α2 z
φ(x, y, z,ω) = 2 e j (ux + vy) dudv (1.8a)
4π −∞

1 ∞ − u 2 +v 2 −kβ2 z j(ux +vy)

ψ(x, y, z, ω ) = 2 ∫ ∫ Ψ(u,v, ω )e e dudv
4 π −∞
(1.8b)
ω, ω
where kα = kβ = . Φ(u,v,ω) and Ψ(u,v,ω) are respectively the
α β
Fourier transforms of the displacement potentials φ and ψ evaluated on the
surface z=0. Furthermore, ψ must satisfy zero divergence condition (1.3b).
This will place additional constraints on Ψ(u,v,ω) , namely,

juΨ x (u,v,ω) + jvΨ y (u,v,ω) − u 2 + v 2 − kβ2 Ψ z (u,v,ω) = 0 (1.8c)

Recall that the pressure waves (p-waves) travel at speed α and the shear waves
travel at speed β , where α is generally greater than β . The displacement
vector is in the direction of the gradient of the scalar potential but it is in the
direction of curl of the vector potential (1.4), that is, normal to the vector
potential. Thus there is a fundamental difference in the nature of propagation of
the shear and the pressure waves. The shear waves are polarized; the
displacement vector is always perpendicular to the direction of wave
propagation. The displacement vector executes a motion depending upon the
phase difference between the components of the displacement vector; a line

© 2001 by CRC Press LLC

when the phase difference is zero, a circular path when the phase difference is
900, or a random path when the phase difference is randomly varying. These
factors play an important role in the design of sensor array systems and
processing of vector potential signals.
1.1.3 Electromagnetic Fields: In electromagnetic fields there are two vectors,
namely, electric vector E and magnetic vector H, each with three components,
thus a six component vector field. The basic equations governing the
electromagnetic (EM) fields are the Maxwell’s equations (in mks units),

∂B 
∇×E= − 
∂t  Faraday’s law (1.9a)
∇⋅D = ρ 
∂D 
∇×H= J+ 
∂t  Ampere’s law (1.9b)
∇⋅B = 0 
where
D = εE ε: dielectric constant
B = µH µ: magnetic susceptibility (1.9c)
J = σE σ: conductivity

We shall introduce two potential functions φ , a scalar potential, and ψ , a

vector potential, in terms of which the components of the EM field are
expressed. The electric and magnetic fields are defined in terms of φ and ψ as
follows:

H=∇×ψ
∂ψ (1.10a)
Ε = −∇φ −
∂t

Th vector potential is further required to satisfy (Lorentz gauge)

∂φ
∇ ⋅ ψ + εµ =0 (1.10b)
∂t
Using (1.10) in (1.9) we obtain two decoupled equations,

∂2φ 1
∇ φ − εµ 2 = − ρ
2
(1.11a)
∂t ε

© 2001 by CRC Press LLC

 ∂ 2ψ
∇ × ∇ × ψ − εµ = −µJ (1.11b)
∂t 2
When there is no free charge, that is, ρ =0, the scalar potential may be canceled
by a suitable choice, using the principle of gauge transformation [1, p207].
Then, equations in (1.10a) reduce to

H=∇×ψ
∂ψ (1.12)
Ε=−
∂t
and the vector potential satisfies

∂ 2ψ
∇ × ∇ × ψ − εµ = −µJ
∂t 2 (1.13)
∇⋅ψ = 0

Both electric and magnetic fields travel with the same speed unlike p and s
waves in solids. The electric and magnetic vectors lie in a plane ⊥ to the
direction of propagation. The tip of a field vector will execute a smooth curve
known as a polarization ellipse. For example, in a vertically polarized electric
field the tip of the electric vector lies on a vertical line while the magnetic field
lies on a horizontal line. The minor and major axes of the polarization ellipse
a
are a and b respectively. Define an angle χ = tan −1 ( ) . The major axis of the
b
polarization ellipse is inclined at an angle ζ (see fig. 1.2). Consider a plane
wave with its wave vector in y-z plane making an angle θ with z axis. The
electric vector will lie in a plane ⊥ to direction of propagation. This plane is
known as the x-θ plane. The electric vector can be split into two components,

Ex and Eθ, lying in x-θ plane.

Ex = E cos γ horizontal component

Eθ = E sin γ e jκ vertical component

where γ and κ are related to χ and ζ [2],

π
cos(2γ ) = cos(2χ)cos(2ζ) 0≤γ ≤
2

© 2001 by CRC Press LLC

 z Wave vector
in y-z plane

-E x

θ Eθ

(a)
x

Eθ

b
a ζ
Ex

Polarization
ellipse

(b)

Figure 1.2: (a) EM wave propagation vector is in y-z plane. Electric and magnetic
vectors lie in a plane ⊥ to the propagation vector and it contains the x-axis. (b)
The tip of the electric vector executes a smooth curve, for example, an ellipse. The
electric vector can be split into two components, Ex and Eθ.

tan(κ) = tan(2χ)csc(2ζ) −π≤κ≤π

© 2001 by CRC Press LLC

 A plane EM wave is characterized by four parameters, namely, E (amplitude), θ
(direction), χ and ζ (polarization parameters).
When there is no free charge current (free moving charge, i.e., ρ = 0)
the vector potential for a plane wave EM field is given by

ψ = a p A exp( j(k ⋅ r − ωt))

ω
where k= a k and A is a complex constant. a p and a k are mutually ⊥
c
unit vectors. The electric and magnetic field vectors are given by

E = jω a p Aexp( j(k ⋅ r − ω t))

(1.14)
H = jω (a k × a p )Aexp( j(k ⋅ r − ω t)) = a k × E

1.1.4 Properties of Wavefields: We shall briefly state some properties of the

wavefields: (a) Radiation condition: A solution to a wave equation must satisfy
the radiation condition or Sommerfeld condition given by [26, p. 499]

∂φ 1
r→∞

→ O( )
∂r r
∂φ 1
+ jk φ r→∞
→ o( )
∂r r
(b) Time Reversal: All wave equations without time varying coefficients or
time derivatives of odd order share an interesting property of time reversal. If
φ(x, y, z,t) is a solution of the wave equation it is easy to verify that
φ(x, y, z,t0 − t) is also a solution of the same wave equation for any constant
t 0 [3]. This fact has been used for reverse time propagation of seismic
wavefields. The seismic wavefield observed on a surface is time reversed and fed
into a loudspeaker broadcasting the recorded wavefield.
1.1.5 Sensing of Wavefields: A device is required to convert a physical
wavefield, say, pressure wave, into an electrical signal which is then sampled
and digitized. The resulting numbers are stored in computer memory for further
processing. Such a device is known as a sensor. While the mechanism of
conversion of a physical wavefield into an electrical signal is not important
from the point of array signal processing, the speed of conversion and the
dynamic range are very relevant; accordingly the sampling rate and the number
of bits per sample are fixed. The speed of conversion controls the bandwidth of
a sensor, faster is the conversion larger is its bandwidth. As we have two types
of wavefields, scalar wavefield and vector wavefield, there are also two types of

© 2001 by CRC Press LLC

sensors. A scaler sensor is used to sense a scaler wavefield such as pressure or
any one component of the EM field. The most common example of the scaler
sensor is a microphone or a hydrophone. A vector sensor will measure all
components of the vector wavefield, three components of mechanical waves in
solid or six components of EM field. Three component seisometers are
sometimes used in seismic exploration. Six component EM sensors are likely
to be available very soon off the shelf [4]. Modern sensor arrays consist of
several tens or hundreds of sensors. One major problem is the lack of
uniformity in sensor response. Ideally, it is assumed that all sensors are
omnidirectional with unit response. But in practice, the response of a sensor
may depend upon the direction of incident wavefront and it may be complex.
This behaviour in itself is not so disturbing as the variation of this behaviour
from sensor to sensor. It is often required to carefully estimate the response of
each sensor in a large array. This process is known as array calibration.
Commonly used sensors measure just one component of the wavefield. This
seems to be adequate as all components travel with the same speed (except p
and s waves in solids) yielding the same delay information on which many
array processing algorithms are based. Additional information such as the
polarization and the particle motion has also been lately used in array
processing. Some interesting developments in sensor technology, which will
have considerable bearing on the future of sensor array processing, are taking
place. For example, the vector sensors capable of measuring all six components
of the EM field and four components of the mechanical field (pressure and three
particle velocity components) have become commercially available [4].
In addition to the above wavefield sensors, we have chemical sensors
capable of detecting a very small quantity of chemicals in the vapour state. The
distribution of vapour is governed by a diffusion equation in place of a wave
equation. Extremely sensitive detectors of magnetic fields based on the principle
of superconducting quantum interference have also appeared and have been used
in magnetoencephalography.

§1.2 Wavefield in Open Space:

In a homegenous space without any reflecting boundaries the wave equation is
easily solved. The propagation of wavefields may be described within the
framework of filter theory. The propagating and nonpropagating (transient)
waves are well demarcated in the frequency wavenumber domain. Continuation
wavefield from one plane to another plane is easily achieved through a filtering
operation. We shall introduce some of these basic concepts here in this section.
Another important constraint imposed by practical considerations is that the
wavefield is measured on a set of discrete points with discrete sensors. Thus,
the spatial sampling is always implied; in contrast a temporal sampling of the
analog sensor output is required only when digital processing is desired.
1.2.1 Fourier Representation of Wave Field: The generic form of a wave
equation in a homogeneous medium is given by

© 2001 by CRC Press LLC

 1 ∂2 f
∇2 f = (1.15)
c 2 ∂t 2

where f (r,t) stands for any one of the wave types, for example, pressure or
one of the components of a vector field. We introduce the Fourier integral
representation of a wavefield,

∞
1
f (r,t) = 3 ∫ ∫ ∫ F(u,v,ω)H(u,v, z)e j (ωt −ux − vy) dudvdω (1.16)
8π − ∞

in wave equation (1.15). We observe that H(u,v, z) must satisfy an ordinary

differential equation given by

d 2 H(u,v, z) ω2
= (u 2
+ v 2
− )H(u,v, z) (1.17)
dz 2 c2
whose solution is

H(u,v, z) = exp(± (u 2 + v 2 − k 2 z) (1.18)

ω
where k= is known as a wavenumber. When (u 2 + v 2 > k , we choose
c
(-) sign for z>0 and (+) sign for z<0 so that the wavefield does not diverge. In
both cases the field will rapidly decay as z → ∞ . These are known as
evanescent waves. When (u
2
+ v 2 < k we get propagating waves whose
integral representation reduces to

f (x, y,t) =
∞
1 (1.19)
8π 3 ∫ −∫∞∫
k 2 −u 2 − v 2 z ) j ( ωt −ux − vy)
F(u,v, ω )e ± j ( e dudvdω

± j( k 2 −u 2 −v 2 z)
where the sign in e is selected depending on whether the
waves are diverging or converging. The convention is (-) sign for diverging
waves and (+) sign for converging waves (see fig. 1.3). Note that in a bounded
space both diverging and converging waves can coexist and hence it would be
necessary to use both signs in describing wavefields in a bounded space.

© 2001 by CRC Press LLC

 (-) sign (+) sign
Diverging wavefronts Converging wavefronts

Figure 1.3: Sign convention in diverging and converging waves.

Equation (1.19) suggests an interesting possibility, that is, if a

wavefield is observed on an x-y plane, it is possible to extrapolate it into the
space above or below the plane of observation. Further, (1.19) may be looked
upon as a sum of an infinitely large number of plane waves of the type
k 2 −u 2 −v 2 z)
e j(ωt −ux −vy− with complex amplitude F(u,v,ω) . The
direction of propagation of individual wave is prescribed by the spatial
frequencies u and v,

u = k sin θ cosϕ
v = k sin θsin ϕ

where ϕ and θ are respectively azimuth and elevation angles of a plane wave.
The elevation angle is an angle between the z-axis and the wave vector and
azimuth angle is an angle between the x-axis and the projection of the wave
vector on the x-y plane. The representation given in (1.19) is also known as a
plane wave decomposition of a wavefield. In (u, v, ω) space a single frequency
plane wave is represented by a point and a wideband wave by a line passing
through the center of the coordinate system (see fig. 1.4). The slope of the line
is inversely proportional to the direction cosines (defined later on page 17) of
the vector ⊥ to the plane wavefront and directly proportional to the speed of
propagation.

© 2001 by CRC Press LLC

 ω

c
b

a
0

Figure 1.4: Representation of a plane wave in (u,v, ω) space. Point c represents a

single frequency (narrowband) plane wave and line ab represents a wideband plane
wave.

1.2.2 Domain of Propagation: We have noted earlier that propagation of waves

is possible only when (u 2 + v 2 < k . For a fixed ω, the spatial frequencies u
ω
and v must lie within a circle of radius k (= ), that is, a disc defined by
c

ω
(u 2 + v 2 ) ≤ ( )2 (1.20)
c

Equation (1.20) represents, as a function of ω, a conical surface in (u,v, ω)

space. It is a vertical cone with an apex at the center of the coordinates and the
angle of the cone is inversely proportional to the speed of propagation. For a
real signal, plane wave representation in (u,v, ω) space (see fig.1.4) extends
below the (u,v) plane. The domain of wave propagation is obtained by
reflecting the conical surface below (u,v) plane. This results into an hourglass
figure shown in fig. 1.5.
1.2.3 Apparent Propagation Speed: Apparent propagation speed refers to the
speed with which a wave appears to travel across an array of sensors placed on a
line. For example, consider an array of sensors on the x-axis and a plane
wavefront incident at angle θ and ϕ as shown in fig. 1.6. Travel time from p to

© 2001 by CRC Press LLC

 po cos(θ)
o is dt = . Hence, the speed of propagation along the vertical axis
c
is

po c
= cz = (1.21a)
dt cos(θ)

To compute travel time from q to o, do the following construction:

i) draw a wavefront qs2q1 ⊥ to incident ray
ii) draw a vertical plane through os2. Since qq1 is in the horizontal
π
plane xoy and or is in the vertical plane, ∠orq = .
2
π
iii) Since rs2 is in the wavefront plane, ∠rs2 o =
2
From (i), (ii), and (iii) we obtain the following result:

or = oq cosϕ
os2 = or sin θ
= oqsin θ cosϕ

os2 oq
The travel time from s 2 to o is dt = = sin θ cosϕ . Hence the
c c
apparent speed along the x-axis is given by

oq c
cx = = (1.21b)
dt sin θ cosϕ

Similarly we can show that

c
cy = (1.21c)
sin θsin ϕ

Equations (1.21b) and (1.21c) can be rewritten in terms of spatial frequencies

ω ω
cx = =
k sin θ cosϕ u
ω ω
cy = =
k sin θsin ϕ v

© 2001 by CRC Press LLC

 ω

Figure 1.5: Domain of propagation in (u,v, ω) space (an hourglass figure).

q
1 r ϕ
o q x
s2
s
1

Figure 1.6: Geometrical derivation of apparent speeds.

© 2001 by CRC Press LLC

 ω ω
cz = = (1.22)
k cosθ k − u2 − v2
2

From (1.22) it follows that

1 1 1 1
2
+ 2+ 2 = 2 (1.23)
c x c y cz c

and that the apparent speeds are always greater than or equal to the wave speed.
Closely related to the apparent speed are the so-called ray parameters,

sin θ cosϕ sin θsin ϕ cosθ

px = , py = , pz = (1.24)
c c c
The significance of ray parameters is that as a ray propagates through media of
different wave speeds, the angles of incidence and emergence will change such
that the ray parameters remain fixed, equal to that at the start of the ray. From
(1.22) it is easy to show that the ray parameters are related to the apparent
speeds,

1 1 1
px = , py = , pz = (1.25a)
cx cy cz

The direction cosines of a ray are defined as

α = sin θ cosϕ, β = sin θsin ϕ, γ = cosθ (1.25b)

From (1.24) and (1.25) the ray parameters can be expressed in terms of direction
cosines

α = px c, β = py c, γ = pz c

1.2.4 Continuation of Wavefield: Consider a thin layer of sources on the z=0

plane. Let f 0 (x, y,t) be the wavefield observed close to the z=0 plane. The
wavefield on a horizontal plane z=z1 can be expressed in terms of that on z=0
plane. For this we make use of (1.19),

f (x, y, z1 ,t) =

© 2001 by CRC Press LLC

 ∞
1
3 ∫ ∫∫ 0
F (u,v,ω)e ± j ( k 2 −u 2 − v 2 z1 ) j (ωt −ux − vy)
e dudvdω (1.26)
8π − ∞

where F0 (u,v, ω ) is the Fourier transform of the wavefield observed on the

z=0 surface. Thus, it is possible to extrapolate a wavefield observed on one
plane to another plane provided the intervening space is source free. Note that
the Fourier transforms of the wavefield observed on two parallel surfaces differ
by a phase factor only. In the case of horizontal surfaces, the phase factor is
± j ( k 2 −u 2 + v 2 ∆z )
simply given by e where ∆z is vertical separation. Each
component of the plane wave decomposition is subjected to a phase shift whose
magnitude depends upon the spatial frequencies, u and v, or on the direction of
the wave vector. For example, for ∆z = λ the phase shift applied to different
plane waves is shown in fig. 1.7 as a function of u (keeping v=0). It may be
recalled that the Fourier transform of the wavefield for (u + v ) > k is
2 2 2

rapidly vanishing and hence the phase shift is set to zero in this range.
1.2.5 Point Source: A point source is often used as a source of illumination,
but it will only generate spherical waves which produce a much more complex
(mathematically speaking) response from a target than what a plane wave does.
Fortunately, a point source wavefield may be written as a sum of infinitely
many plane waves; a plane wave decomposition was described in subsection
(1.2.1). Additionally, a response of a target may be obtained as a sum of plane
wave responses. To obtain the Fourier integral representation of a point source
wavefield we go back to (1.19) where we shall assume that the Fourier
transform of the wavefield has radial symmetry,

f (x, y, z,t)
∞
1
= 3 ∫ ∫ ∫ F0 (u,v, ω )e ± j ( k 2 −u 2 − v 2 z ) j ( ωt −ux − vy)
e dudvdω (1.27a)
8π − ∞
∞ 2π
1 1
2 ∫ 0
= F (s, ω )e jωt dω ∫ se ± j ( k 2 −s2 z)
ds ∫ e − j (sr cos( θ − ϕ )) dθ
4π 0
2π 0

where F0 (u,v,ω) is the Fourier transform of f (x, y, z = 0,t) . We can

further simplify (1.27a) as

∞ ∞
1
∫ F (s,ω)e dω ∫ sJ0 (sr)e ± j ( k 2 −s2 z)
f (r, z,t) = 2 0
jωt
ds (1.27b)
4π −∞ 0

© 2001 by CRC Press LLC

 3.2

2.4

1.6

0.8

0.0
-3.2 -1.6 0.0 1.6 3.2
u

Figure 1.7: Phase shift as a function of u (keeping v=0) for vertical separation equal
to wavelength (assumed to one).

π
where s = u 2 + v 2 = k sin γ where 0 ≤ γ ≤ . In (1.27b) we replace s by
2
k sin γ and rewrite (1.27b) as
π
∞ 2 z
1 jω(t ± cos γ )
f (r, z,t) = 2 ∫ dω ∫ F (k sin γ,ω)e k sin(2γ )J0 (kr sin γ )dγ
c 2
0
8π −∞ 0
(1.28)

Equation (1.28) is a plane wave decomposition (PWD) of a point source

wavefield. In an r-z plane, for a fixed ω and γ the integrand in (1.28) represents
a plane wave component, traveling with an angle of incidence γ as shown in
z
fig. 1.8 and ∆t = cos γ is propagation time from surface to a depth z. The
c
inverse of (1.28) is given by

© 2001 by CRC Press LLC

 point source
r

Figure 1.8: In a homogeneous medium the wavefield due to a point source may be
expressed as a sum of plane waves traveling with an angle of incidence γ , where
π
0≤γ ≤ .
2

F0 (s,ω)e ± j( k 2 −s2 z)
= ∫ F(r, z,ω)rJ0 (sr)dr (1.29)
0

where F(r, z,ω) is a Fourier transform (temporal) of f (r, z,t) . On the z=0
surface (1.29) reduces to

F0 (k sin γ,ω) = ∫ F(r, z = 0,ω)rJ0 (kr sin γ )dr (1.30)

0

where we have substituted s = k sin γ . Equation (1.30) in the time domain

may be expressed as

 

∞
2 
f (t,sin γ ) = ∫  f (r, z = 0,t) * rdr (1.31)
0
r 
( sin γ )2 − t 2
 c 

© 2001 by CRC Press LLC

Equation (1.31) enables us to convert the wavefield due to a point source into
that due to a plane wave source. A linear array of sensors is used in a radial
direction with respect to point source. The output of each sensor is first filtered
with a filter having response

2
h(t) =
r
( sin γ )2 − t 2
c

and then summed over all sensors. The impulse response function may be
expressed as an integral involving a delta function [5],

π
2
h(t) = = ∫ δ(t − t0 cosβ)dβ
r
( sin γ )2 − t 2 −π
c

r
where t0 = ( sin γ ) . Using the above integral representation of the impulse
c
response function in (1.31) we obtain

 t 0
∞
2 
f (t,sin γ ) = ∫  ∫ f (r, z = 0,t − t ′ ) 2 d t ′ rdr
0  0
−t t 0 − ( t ′ ) 2

∞ t 0 π 
= ∫  ∫ f (r, z = 0,t − t ′ )dt ′ ∫ δ(t ′ − t0 cosβ)dβ rdr (1.32)
0 −t 0 −π 

∞ π

= ∫  ∫ f (r, z = 0,t − t0 cosβ)dβ rdr
0 − π 

For a discrete array the integral is replaced by a sum,

π ∞

f (n∆t0 ,sin γ ) = ∫ ∑ n∆r f (n∆r, z = 0,t − n∆t0 cosβ)dβ

2
(1.33)
− πn = 0

∆r
where ∆t0 = sin γ . The inner sum in (1.33) is the sum-after-delay
c
operation, commonly known as slant stacking in seismic exploration. Later in
chapter 2 we shall show how this operation is related to the radon transform.

© 2001 by CRC Press LLC

 1.2.6 Spatial Sampling and Aliasing: A wavefield is measured on a set of
discrete points with discrete sensors. Thus, spatial sampling is always implied;
in contrast a temporal sampling of an analog sensor output is used only when
digital processing is desired. Often the temporal sampling is preceded by some
kind of low pass filtering, which determines the maximum frequency and the
maximum sampling rate according to the well known sampling theorem. It is
thus possible to avoid the temporal aliasing but it is a different story with
spatial aliasing which is intimately related to propagation speed. For simplicity
of analysis let us assume only spatial sampling. Consider a broadband plane
wave incident on an infinitely long array of sensors,

x ∞
f (x,t) = f (t − ) ∑ δ(x − i∆x)
cx i = − ∞
∞ j (t − )ω
x ∞
1
=
2π −∫∞
F(ω)e cx
dω ∑ δ(x − i∆x)
i=−∞

Note that cx stands for wave speed in the direction of x-axis or apparent speed.
Taking the 2D Fourier transform of f(x,t)

+∞

F(u,ω) = F(u,ω) = ∫ ∫ f (x,t)e j (ωt −ux ) dxdt

−∞
∞ ∞ +∞ x
j (t − ) ω ′
∫ F(ω ′)dω ′ ∑ ∫ ∫ e
− j (ωt −ux )
= e c
δ(x − i∆x)dxdt
−∞ i=−∞ − ∞

∞ ∞ ∞ ω′
j (u − )x
= ∫ F(ω ′)δ(ω ′ − ω)dω ′ ∑ ∫ e δ(x − i∆x)dx
cx

−∞ i=−∞ −∞

∞ ∞ ω′
j (u − )i∆x (1.34)
= ∫ F(ω ′)δ(ω ′ − ω)dω ′ ∑ e
cx

−∞ i=−∞
∞ ∞
ω ′ 2π
= ∫ F(ω ′)δ(ω ′ − ω)dω ′ ∑ δ(u −
−∞ k =−∞
−
cx ∆x
k)

The spatial Fourier transform of the array output is sketched in fig. 1.9. The
signal spectrum is concentrated on sloping lines in the u-w plane as seen in fig.
1.9.
Depending upon the signal bandwidth and the apparent speed, an alias
will show up within the principal band. In fig. (1.9) there is no aliasing when

© 2001 by CRC Press LLC

 ω k=-1

aliased
k=0

k=1

aliased

4π
−∆ - 2∆πx 0 2π u 4π
x ∆x ∆x

Principal
band

Figure 1.9: The Fourier transform of spatially sampled wavefield (plane wave) lies
on a set of sloping lines (with slope= 1/ cx ) corresponding to different values of k
in (1.34).

the signal bandwidth is as shown by the dark square, which in this case
λ min π
corresponds to a sensor spacing ∆x = or ∆x = cx . Note that,
2 ω max
π
since the sampling interval ∆t = , ∆x = ∆t cx . There is no aliasing
ω max
whenever the sensor spacing and the time sampling interval are such that
∆x
≤ cx . For the vertical angle of incidence, since cx = ∞ , there is no
∆t
aliasing effect for any ∆x .
1.2.7 Dispersion: A medium is said to be dispersive when a wavefront travels
at a speed which is a function of frequency. Consider a wavefront at a fixed

© 2001 by CRC Press LLC

 temporal frequency, ω 0 . Let us track a particular point, say, the crest of a
sinusoidal waveform. The speed with which a point on the waveform travels is
λ0
said to be phase speed, which is equal to where λ 0 is the wavelength and
T0
T0 is the period of a single frequency waveform. Thus, phase speed, by
definition, is given by

λ0 ω0
c ph = = =c
T0 k0

and it is equal to the propagation speed.

Now consider a group of waves whose temporal spectrum is centered at
frequency ω 0 and it is spread over a frequency interval ∆ω . The wavefield
may be expressed as

∆ω
ω0 +
2
1
f (r,t) =
2π ∫ F(ω)exp( j(ωt − k ⋅ r)dω
∆ω
(1.35a)
ω0 −
2

ω
where k = (u,v, ( )2 − u 2 − v 2 ) is the wave vector. We assume that the
c
propagation speed is a function of frequency, and hence the wave vector is also
a function of frequency. Using a Taylor’s series expansion of the wave vector

dk(ω)
k(ω) = k(ω 0 ) + (ω − ω 0 )+...
dω ω =ω 0

in (1.35a) we obtain

∆ω
2
1 dk(ω)
f (r,t) = F(ω 0 )e j (ω 0 t − k(ω 0 )⋅r)
2π ∫
∆ω
exp[ j(t −
dω ω =ω 0
⋅ r)ω̃]dω̃
−
2

∆ω dk(ω) ∆ω
= F(ω 0 )e j (ω 0 t − k(ω 0 )⋅r) sin c[(t − ⋅ r) ]
2π dω ω =ω 0 2
(1.35b)

© 2001 by CRC Press LLC

 1

0.5
Field

0
0

-0.5

-1

1 1

0.5 0.5
Field

0 0
Field

-0.5
-0.5

-1
-1

0 50 100 150 200 250 300

Time
Figure 1.10: A narrowband signal with a center frequency at 0.23 Hz and a bandwidth
equal to 0.02 Hz. The envelope travels at the group speed which in this is equal to
the phase speed (assumed to be one). The signal in the lower panel arrives 32 time
units later.

where ω̃ = (ω − ω 0 ) . Observe that the second term modulates the carrier

wave (first term). The second term is a propagating waveform which travels at a
speed of cgp which satisfies an equation

dk(ω)
cgp = 1 (1.36a)
dω ω =ω 0
or
1
cgp = (1.36b)
dk(ω)
dω ω =ω 0

The group speed differs from the phase speed only when the medium is
dispersive. In a nondispervive medium both speeds are equal. The modulating
waveform (second term in (1.35b) travels at the same speed as the carrier wave
(first term). An example of propagation of a narrowband signal is shown in fig.

© 2001 by CRC Press LLC

 1.10 where we have assumed that c ph = cgp =1. The arrival of the signal is
denoted by the arrival of the crest of the modulating wave which travels with a
speed equal to cgp . Hence the energy is transmitted at a speed equal to the
group speed.

§1.3 Wavefield in Bounded Space:

Wavefield in a space bounded by a plane reflecting boundaries is of interest in
many practical problems. The examples of bounded space are (i) acoustic field
in a room, (ii) acoustic field in a shallow water channel and (iii) mechanical
waves in a layered medium, for example, in shallow earth, etc. The wavefield in
a bounded space is normally studied under two different approaches, namely,
rays and wavefronts, and the exact solution of the wave equation leading to
modes. While the former is more versatile and easy to use, it is less accurate.
The second approach is mathematically accurate but difficult to use except in
simple geometries.
1.3.1 Ray Propagation: A point source emits a spherical wavefront, a surface of
equal phase. Lines ⊥ to the wavefront are rays along which the wave energy is
transmitted. A spherical wavefront at a large distance from a point source may
be approximated by a plane

∞
1
∫
j[ωt −u 0 ( x − x 0 )− v 0 ( y − y 0 )− k 2 − s 02 ( z − z 0 )]
f (r,t) = F(ω)e dω
2π − ∞

where (x0 , y0 , z0 ) is the position of a point source and r =

((x − x0 ),(y − y0 ),(z − z0 )) is the position vector of a point on a plane
wavefront. Note that the direction cosines are related to the spatial frequencies,
u0 = kα 0 and v0 = kβ 0 where α 0 ,β 0 , 1 − α 20 − β 20 are direction cosines
of a ray. If a sensor is placed at point (x1 , y1 , z1 ) , the wavefield received at
(x1 , y1 , z1 ) is given by
∞
1
2π −∫∞
f (r1 ,t) = F(ω)e j[ωt −φ1 ]dω

where φ1 = u0 (x1 − x0 ) + v0 (y1 − y0 ) + k 2 − s02 (z1 − z0 ) is a constant

phase. We can also write φ1 = ωt1 where

© 2001 by CRC Press LLC

 incident Normal reflected

θi θr

Fluid Medium #1
Fluid Medium #2

θt
refracted

Figure 1.11: Laws of reflection and refraction: (1) Incident wave, reflected wave,
refracted wave and the normal to the interface at the point of incidence lie in the
sin θ i sin θ t
same plane. (2). θi=θr (3) = = p (ray parameter).
c1 c2

(x1 − x0 )2 + (y1 − y0 )2 + (z1 − z0 )2

t1 =
c
(x − x0 ) (y1 − y0 ) (z1 − z0 )
= 1 + +
cx cy cz

where cx ,cy ,cz are apparent speeds in x, y, z directions, respectively. When a

ray encounters an interface separating two contrasting media it splits itself into
two rays, a reflected and a refracted ray. The laws of reflection and refraction are
summarized in fig. 1.11. The reflection and refraction coefficients are related to
the impedance of the media on both sides of the interface. For example, the
reflection and transmission coefficients at the interface separating two fluid
media are given by

ρ2 c
cosθi − ( 2 )2 − sin 2 θi
ρ c1
r̂ = 1 (1.37a)
ρ2 c
cosθi + ( 2 )2 − sin 2 θi
ρ1 c1

© 2001 by CRC Press LLC

 incident(p) reflected (s) reflected (p)

θ' r
θ i θr

Elastic Medium #1
Elastic Medium #2

θt

θ't refracted(p)

refracted(s)

Figure 1.12: Reflection and refraction of a p-wave at an interface. Four rays as

shown are generated. All rays including the incident ray and the normal to the
interface lie in a plane. The angle of incidence and the angles of reflection and
sin θ i sin θ ′r sin θ t sin θ ′t
refraction are related. θ i = θ r and = = II = II = const.
c pI csI cp cs

2 cosθi
tˆ = (1.37b)
ρ2 c
cosθi + ( 2 )2 − sin 2 θi
ρ1 c1

Reflection and refraction at an interface separating two elastic media are more
complex. An incident longitudinal wave will give rise to two reflected and two
refracted rays. The first ray is a p-wave and the second ray is a s-wave. The laws
of reflection and refraction for a p-wave incident at the interface are summarized
in fig. 1.12. The coefficients of reflection and refraction, ( r̂ p , r̂s , tˆp , tˆs ) , are
obtained by solving the following system of four linear equations [6, p3-101]:

r̂ p cosθi − r̂s sin θ′r + tˆp cosθt − tˆs sin θ′t = cosθi
(1.38a)
− r̂ p sin θi − r̂s cos θ′r + tˆp sin θt + tˆs cos θ′t = sin θi

© 2001 by CRC Press LLC

 (λ1 + 2µ1 − 2µ1 sin 2 θ r ) µ
− r̂ p I
+ r̂s I1 sin 2θ′r
cp cs
(λ 2 + 2µ 2 − 2µ 2 sin 2 θt ) ˆ µ 2 (λ1 − 2µ1 − 2µ1 sin 2 θ r )
+ tˆp − t s II sin 2 θ ′
t =
c pII cs c pI
µ1 µ µ µ µ
r̂ p I
sin 2θ r + r̂s I1 cos2θ′r + tˆp II2 sin 2θt − tˆs II2 cos2θ′t = I1 sin 2θ r
cp cs cp cs cp
(1.38b)

where (λ1 ,µ1 ) are the elastic constants and (c pI ,csI ) are respectively the p-
wave and the s-wave speed in the upper medium. Other parameters refer to the
lower medium. For vertical incidence, that is, θi = 0 from (1.38), it follows
that r̂s = tˆs = 0 . This is a consequence of the fact that there is no s-wave
generation for vertical incidence.
Polarization in the plane of incidence: Finally, an EM wave at an interface
between two different media undergoes reflection and refraction. However, there
is only one reflected and one transmitted wave which travel at the same speed.
The process of reflection and refraction is strongly influenced by the
polarization angle. We shall consider one simple case of polarization in the
plane of incidence (fig. 1.13). The coefficient of reflection and transmission is
given by (Fresnel’s equations)

χ−ζ ˆ 2
r̂ E = , tE =
χ+ζ χ+ζ (1.39)

cosθt µc
where χ= and ζ = 1 1 . Note, when χ = ζ , r̂ E = 0 and tˆE = 1 ;
cosθi µ 2 c2
that is, there is no reflected energy. This phenomenon takes place at a specific
angle of incidence known as Brewster angle given by
1 − ζ2
sin 2 θ B = (1.40)
c
( 2 )2 − ζ 2
c1

1.3.2 Propagation in Channel - Ray Theory: Consider a source inside a fluid

channel which is bounded from above by a free surface and from below by
another fluid (this is known as the Pekeris model). The reflection coefficient at
the

© 2001 by CRC Press LLC

 Ei
Bi Br
Er
θ i θi µ1 c1
medium 1
medium 2
µ 2 c2
θt Et
Bt

Figure 1.13: The electric vector is in the plane of incidence, that is, in the plane
containing the incident ray, normal to interface and reflected and refracted rays.

free surface is (-1) and that at the bottom is r̂b (see (1.37)). Because of multiple
reflections at the free surface and at the bottom (see fig. 1.14) many waves will
reach a sensor array at different times. It is convenient to model all these waves
as emanating from a series of images whose position can be determined by
following simple geometrical optics rules. An infinite set of images is formed
between two parallel reflecting surfaces. We index these images with two
integers (i,k); the first integer represents a group and the second integer
represents an image within a group. There are four images in each group. For
example, in the i=0 group the images are: s01 and s02 which are equal but
opposite in strength, that is, a surface dipole, and s03 and s04 , which are
caused by the reflection of the surface dipole onto the bottom. The surface
dipole and its image are separated by a distance of 2H where H is the depth of
the fluid channel. The next group of images, that is, for i=1, is obtained by
sliding the surface dipole two depth units above and the image dipole two depth
units below the bottom. The vertical distance with respect to the top sensor in
the array to different images in the ith group is given by

Hi1 = 2iH + zr − zs
Hi2 = 2iH + zr + zs
(1.41)
Hi3 = 2(i + 1)H − zr − zs
Hi 4 = 2(i + 1)H − zr + zs i = 0,1,2,...

© 2001 by CRC Press LLC

 s o2

Air

zs

Source zr
s o1

H sensor

Fluid medium 1
Fluid medium 2

so 3

so
4

Figure 1.14: A shallow water channel with a sound source (•) at depth zs and a
sensor at depth zr. Rays starting from the source are reflected at the surface and
bottom and finally reach the sensor after one or more reflections. Some of the
images (•) are also shown.

An image whose second index is 1 or 2 lies above the bottom and an image
whose second image is 3 or 4 lies below the bottom. The strength of an image
depends upon the number of bounces the ray has undergone before reaching the

© 2001 by CRC Press LLC

 sensor and also, on account of geometrical spreading and absorption, on the
actual distance traveled which is given by

li 0 = d 2 + Hi20 li1 = d 2 + Hi12

(1.42)
li2 = d 2 + Hi22 li3 = d 2 + Hi32
where d is the horizontal distance to the sensor from the source. The strength of
the images is given by

e −βli0 e −βli1
i +1 i
α i 0 = (−1) r̂
i i
b α i1 = (−1) r̂ b
li 0 li1
(1.43)
e −βli2 e −βli3
α i2 = (−1)i r̂bi+1 α i3 = (−1)i +1 r̂bi+1
li2 li3

where β is the attenuation coefficient in the top liquid layer. The signal
reaching the sensor (fig 1.14) may be expressed as

∞
1 ∞ 3
p(t) = ∑
2π i = 0
∑ ∫α
m=0 −∞
im P(ω)e j (t − τ im )ω dω (1.44)

where p(t) is the pressure field received by a sensor and P(ω) is the Fourier
transform of the waveform emitted by the source. τ im is propagation delay
lim
τ im = . The pressure field given by (1.44) is a complex field. The
c
magnitude and phase are found to vary rapidly with the depth of the sensor. For
example, for a channel with the following parameters:

H=100 m (meters)
d=2000 m
zs= 50 m
ρ1 = 1, c1 = 1500
ρ2 = 2,c2 = 1600

the magnitude and phase variations as a function of the depth of a sensor are
shown in fig. 1.15. Eighty images were taken into account in computing the
pressure field.
As the range increases, while the position of image sources remain
unchanged, the strength of the images may increase on account of the
phenomenon of total internal reflection. Recalling equation (1.37) it may be

© 2001 by CRC Press LLC

 c1
noticed that whenever sin(θi ) = ( c1 < c2 ), rb = 1. This angle of
c2
incidence is known as the critical angle, θ c . For θi > θ c , rb becomes purely
imaginary. The wave energy travels along the interface and as it travels some
energy is re-radiated back into the channel at the critical angle. It is possible to
account for long distance transmission of acoustic energy using the above ray
model even at frequencies as low as 200Hz [7]. By introducing certain
corrections (i.e., beam displacement) the ray model has been improved in the
low frequency region (60 - 140 Hz)[8]. The ray model has been used for source
localization in shallow water [9].
1.3.3 Propagation in a Channel - Normal Mode Theory: The pressure field due
to a point source in a fluid channel, as depicted in fig. 1.14, is given by

j ∞ ∞
p(r, z,t) = ∫ ∑ φ n (z)φ n (zs )H01 (kn r)P(ω )e jωt dω (1.45)
2 −∞ n=0

where φ n (z) is an eigenfunction and kn is its corresponding eigenvalue

obtained by solving a differential equation with homogeneous boundary
conditions

d 2φ
2
+ γ 2n φ = 0
dz
φ(0) = 0 (1.46)
dφ
= 0 (rigid bottom)
dz z = H
ω2
where kn + γ n = 2 . kn is a horizontal wavenumber and γ n is a vertical
2 2

c (z)
wavenumber. The source is assumed to be at a depth zs . Note that both
wavenumbers are functions of frequency and also of depth when the sound speed
is a function of depth. Consider a special case of isospeed channel and large r.
Firstly, the Hankel function can be approximated as
π
2 j ( kn r + 4 )
H (kn r) ≈
1
e ; secondly both wavenumbers are independent of
kn πr
0

depth and therefore the solution of (1.46) is given by

© 2001 by CRC Press LLC

 Figure 1.15: (a) Magnitude and (b) phase of the computed pressure field in a shallow
water channel. Eighty images were taken into consideration for evaluating the net

© 2001 by CRC Press LLC

field at a desired point. Note the highly complex structure of the field caused by
interference.

e jγ n z − e − jγ n z
φ n (z) = sin(γ n z) =
2j
1
(n − )π
where γ n = 2 . For this special case the pressure field in the fluid
H
channel with the rigid bottom is given by

p(r, z,t) =
π
j
4 ∞ ∞
e
2 ∫∑
1
2kn πr
{ }
sin(γ n zs ) e j ( γ n z + k n r +ωt ) − e j ( − γ n z + k n r +ωt ) P(ω)dω
−∞ n=0
(1.47)

From (1.47) it is possible to infer that each mode is a sum of two plane
γ
wavefronts traveling in the vertical plane at angle ±θ n , where tan θ n = n
kn
with respect to the horizontal plane. In three dimensions the wavefront is a
π
conical wavefront (see fig.1.16); the angle of the cone is − θ n . Note that the
2
direction of propagation of the wavefronts is solely dependent on the channel
characteristics.
The comparison of ray and normal mode approaches to propagation in
a shallow water channel is instructive. The ray approach is essentially a high
frequency approximation to the exact solution of the wave equation. The
normal mode approach on the other hand is a series approximation to the exact
solution of the wave equation. The conical wavefronts obtained by
decomposing each mode may be looked upon as a result of constructive
interference of spherical wavefronts from the source and its images [8]. At low
frequency the accuracy of the ray approach is enhanced if one were to use the
concept of beam displacement. At a point of reflection the beam is found to be
laterally displaced, a phenomenon first observed in optics [10]. Both numerical
and experimental evidence in support of the fact that beam displacement does
help to increase the accuracy of the ray approach in relation to the exact
solution has been widely reported [11, 12].
1.3.4 Propagation Through Layered Medium: In seismic exploration a
horizontally layered medium is often used as a model. Vertically
incidentlongitudinal waves (p-waves) are preferred as there is no loss of energy
through conversion into s-waves at each interface. Also, as s-waves arrive
slightly later,

© 2001 by CRC Press LLC

 Air

Wavefronts
Shallow
water Source

Rigid Bottom

Figure 1.16: A mode can be decomposed into two conical wavefronts. In the vertical
γn
plane the wavefronts travel at the angle ±θ n , where tan θ n = .
kn

ρ1 c1
ρ2c2

ρn cn

Figure 1.17: A stack of uniform layers. All layers are of the same thickness but with
different impedances.

they would interfere with late arriving p-wave signals. A layered medium is
modeled as uniform horizontal layers stacked one above the other (see fig.
1.17). A vertically propagating plane wavefront is repeatedly reflected and
transmitted at each interface, thus producing a complex reverberation pattern
which may be conveniently described within the frame work of filter theory
[13].

© 2001 by CRC Press LLC

 incident
wavefront reflected refracted

refracted incident reflected

wavefront
(a) (b)

Figure 1.18: Reflected and transmitted waves at an interface. (a) Incident from above
and (b) incident from below.

A plane wave of unit amplitude is vertically incident on an interface. It

is split into two waves, the first one traveling upwards (reflected wavefront) and
the second one traveling downwards (refracted wavefront). Similarly, a plane
wave incident from below is split into two plane waves (see fig. 1.18). The
amplitude of the reflected wave is equal to − r̂ (reflection coefficient) and that
of the refracted wave is equal to tˆ (transmission coefficient). Now consider a
layer sandwiched between two semi-infinite layers. The thickness of the layer is
measured in units of return travel time ∆t , that is, one unit corresponds to a
c
physical thickness ∆h where ∆h = ∆t . Fig. 1.19 shows repeated
2
reflections and refractions at two faces of the layer along with their respective
amplitudes. r0 and t0 are respectively reflection and refraction coefficients at
the top face of the layer; similarly r1 and t1 are those at the bottom face of the
layer.
Let f refl (t) be the reflected waveform, which consists of a sum of all
successively reflected wave components,

∞
1 r0 + r1e − jω jωt
2π −∫∞
f refl (t) = F (ω) e (1.48a)
1 + r0 r1e − jω
0

and similarly the transmitted waveform is given by

© 2001 by CRC Press LLC

 Reflection
^t 0 ^r ^t 0 ^t 0 ^r1 ^r ^r ^t 0
^r 1 0 1
0

Layer 0

^t 0 ^t 0 ^r 1 - ^t 0 ^r 1 r 0
^ ^
- ^t 0 ^r 1 r 0 ^r 1
travel time
unit return

Layer 1

Layer 2

^t 0 ^t ^ ^
1 - ^t 0 r 1 r 0 ^t 1
Transmission

Figure 1.19: Repeated reflections at the two faces of a layer produce a sequence of
reflected and transmitted waves. Some amount of energy is also trapped inside the
layer. All waves travel vertically though for the sake of clarity the waves are shown
as inclined.

ω
∞ −j
1 t0 t1e 2 jωt
2π −∫∞
f trans (t) = F (ω) e (1.48b)
1 + r0 r1e − jω
0

Let z = e jω and define the reflection and transmission response functions as

r0 + r1z −1
1
t0 t1 −
R(z) = , T(z) = z 2 (1.49)
1 + r0 r1z −1 1 + r0 r1z −1

We can now express the z-transforms of the reflected and the transmitted
waveforms as

© 2001 by CRC Press LLC

 Frefl (z) = F0 (z)R(z)
Ftrans (z) = F0 (z)T(z)

It is interesting to note that T(z) has the form of a first order AR process
filter and R(z) has the form of a ARMA process filter of order (1,1) [14].
Since r0 r1 ≤ 1, the only pole of the transfer function (1.49) lies within the
unit circle, making R(z) and T(z) stable. The signal flow diagrams of
transmission and reflection filters are shown in fig. 1.20.

The above characterization of a single layer model has been extended to a

multilayer model. The basic structure of the transmission and reflection filters,
however, remains unchanged. Further exposition of this approach may be found
in [13].

§1.4 Stochastic Wavefield:

A signal emitted by a source, even if it is a simple sinusoid, may be modeled
as a stochastic process because the signal may be controlled by a set of
unknown random parameters, for example, a random phase in a sinusoid. The
noise in the array output may consist of thermal noise, the wavefield emitted
by numerous sources, either natural or man-made, and transmitted signal which
has been scattered by numerous random scatterers or reflectors. In all these
models there is an element of stochastic nature which imparts a stochastic
character to the wavefield. In this and in the next section we shall study some
of these stochastic models of the wavefield from the point of array signal
processing. First, we consider the wavefield emitted by a large number of
random sources which are distributed in two or three dimensional space, both in
open space and bounded space. Basically we shall study the correlation and the
spectral characteristics of the output of an elementary array of two sensors.
Next, in §1.5, we shall study scattering or reflections produced by many
random scatterers which act as obstacles in radio communication or micro-
variations of speed and density in underwater detection communication.
1.4.1 Frequency-Wavenumber Spectrum: A stochastic wavefield observed over
an infinite line or plane is best characterized in terms of a frequency
wavenumber spectrum. For a stochastic wavefield, in place of Fourier
representation (1.19), we have spectral representation (on z=0 plane)

∞
1
3 ∫ ∫∫
f (x, y,t) = dF(u,v,ω)e j (ωt −ux − vy)
8π − ∞

which is analogous to the spectral representation of a 1D stationary stochastic

process [15]. dF(u,v,ω) is a differential of the generalized Fourier transform

© 2001 by CRC Press LLC

 Figure 1.20: The response of a layer to an incident wave is described within the
framework of filter theory. (a) reflection filter and (b) transmission filter.

© 2001 by CRC Press LLC

of the stochastic wavefield having the property that
E{dF(u,v,ω) } ∝ S f (u,v,ω) which we shall use to relate the covariance
2

function to the frequency wavenumber spectrum.

∞
1
3 ∫ ∫∫ f
C f (∆x, ∆y, τ) = S (u,v,ω)e j (ωτ −u∆x − v∆y) dudvdω (1.50)
8π − ∞

where C f (∆x, ∆y, τ) = E{f (x, y,t) f (x + ∆x, y + ∆y,t + τ)} is the
covariance function and S f (u,v,ω) is the frequency wavenumber spectrum. In
the spectrum analysis of time series (1.50) is known as the Wiener-Khinchin
relation of great significance [14]. S f (u,v,ω) represents power received at a
given temporal frequency ω and spatial frequencies u and v. Since u and v are
related to the direction cosines, S f (u,v,ω) may be looked upon at a fixed
temporal frequency as a function of the direction cosines.
1.4.2 Open Space: Consider noise sources in the far field region, distributed
over a sphere or a circle (in two dimensions). Each point source emits a
stationary stochastic waveform uncorrelated with all other sources. Let f i (t)
be the stochastic waveform emitted by the ith source at angular distance
ϕ 0 + θi where θi is a random variable uniformly distributed over an interval
±θ 0 (see fig. 1.21). The signal received at the upper sensor is given
d
by f i (t − sin(ϕ 0 + θi )) and that at the lower sensor is
2c
d
f i (t + sin(ϕ 0 + θi )) where d is the sensor separation. The total signal
2c
obtained by summing over all sources is given by

d
g1 (t) = ∑ f i (t − sin(ϕ 0 + θi ))
i 2c
(1.51a)
d
g2 (t) = ∑ f i (t + sin(ϕ 0 + θi ))
i 2c

We shall now replace the random function in (1.51a) with its spectral
representation [14] and rewrite it as

d
1 ∞ j (ω(t − sin(ϕ 0 +θ i ))
g1 (t) = ∑
2π ∫− ∞ i
dFi (ω)e 2c

© 2001 by CRC Press LLC

 y

θi
2θ 0

ϕ0

Figure 1.21: Point sources are uniformly distributed over an arc of a large circle in
the x, y plane. An elementary array of two sensors is on the y-axis.

d
1 ∞ j (ω(t + sin(ϕ 0 +θ i )))
g2 (t) = ∑
2π ∫− ∞ i
dFi (ω)e 2c
(1.51b)

We will now compute the cross-covariance function between the two outputs,

E{g1 (t)g2 (t + τ)}

E{dFi∗ (ω)dFk (ω ′ )} ×
1 ∞ ∞  − j (ω(t − 2cd sin(ϕ 0 +θ i )))  (1.52a)
=
4π 2 ∫− ∞ ∫− ∞ ∑i ∑i E e dωdω ′
 d 
×e j ( ω ′ (t + τ + 2c sin(ϕ 0 +θ k ))) 
 

Noting the properties of the generalized Fourier transforms of stationary

processes [14] (1.52a) may be simplified to yield the following result:

© 2001 by CRC Press LLC

 1 ∞  j (ω( τ + dc sin(ϕ 0 +θ i ) 
c12 (τ) = ∑ Sii (ω)E e
2π ∫− ∞ i
dω (1.52b)
 

where we have assumed that the sources are uncorrelated with an identical
spectrum. Since θi is a uniformly distributed random variable the expected
value in (1.52b) may be replaced by an integral

θ0 d
1 ∞ 1 j (ω( τ + sin(ϕ 0 +θ)))
c12 (τ) = c12 (τ) = ∫
2π − ∞
S0 (ω)dω
2θ 0 ∫e
−θ 0
c
dθ (1.53a)

The integral over θ can be evaluated in the form of a series [16],

θ0 d
j (ω( τ + sin(ϕ 0 +θ)))
∫e
−θ 0
c
dθ

  ωd sin(nθ 0 )cos(nϕ 0 ) 
 cos(ωτ) ∑ δ n Jn ( c ) n
+ 
   n = 0,2,4... 

  ωd sin(nθ 0 )sin(nϕ 0 ) 
  sin(ωτ) ∑ 2J n (
c
)
n 
  n =1,3,5...

=2
  ωd sin(nθ 0 )cos(nϕ 0 )   (1.53b)
 sin(ωτ) ∑ δ n Jn ( ) − 
  n = 0,2,4... c n 
 j 
ωd sin(nθ 0 )sin(nϕ 0 )  
 cos(ωτ) ∑ 2Jn ( )
 
  n =1,3,5... c n  

where δ 0 = 1 and δ n = 2 for all n. In the limiting case of circularly

ωd
distributed noise sources, θ 0 = π , (1.53b) reduces to 2πJ0 ( ).
c
∞
1 ωd
c12 (τ) =
π0∫ S0 (ω)J0 ( )cos(ωτ)dω
c
(1.53c)

The normalized spatial covariance function at zero lag as a function of sensor

separation is shown in fig. 1.22. Notice that the sensor outputs become
increasingly uncorrelated as the angular extent of the noise sources increases. In
the limiting case of circularly distributed noise sources the correlation becomes

© 2001 by CRC Press LLC

 Figure 1.22: Spatial covariance function (at zero lag) as a function of sensor
d
spacing, . The covariance function is shown for three different angular widths of
λ
the distant noise sources. (1) 1800 , (2) 22.50 , (3) 5.60 .

negligible even for a separation of the order of one wavelength. When the
sources are uniformly distributed on a sphere of large radius, the spatial
sin(kd)
covariance function is given by c12 (0) = [17].
kd
1.4.3 Channel: In this model we shall consider a situation where the noise
sources are uniformly distributed on one of the faces of a channel. This
situation is close to an ocean channel where all noise sources are on or close to
the ocean surface (fig. 1.23) [18].

© 2001 by CRC Press LLC

 Noise Sources

Array of
Sensors

channel

Figure 1.23: The noise sources on the surface of a ocean channel. Correlation
structure of wavefield in the vertical direction is investigated.

Let ψ(r,ϕ,t) , a homogeneous random function, represent the noise

sources on the top face of the channel and Sψ (s,α,ω) be its spectrum where s
is radial frequency and α is azimuth angle. The acoustic power radiated per unit
area at a fixed temporal frequency is given by

∞ 2π
1
Sψ (ω) = 2
4π ∫ ∫ sS
0 0
ψ (s,α,ω)dαds

For wind generated noise on the ocean surface the frequency wavenumber
spectrum is approximately modeled as

2p s 2 m −1
Sψ (s,α,ω) = 2 Γ(m)(1 − 2 ) S0 (ω) s<k
k k
=0 s≥k

where p and m are constants ( 2 ≤ m ≤ 3 ) [17, 19]. Note the spectral

representation of noise sources, at a fixed point on the plane,

∞
1
ψ(r,ϕ,t) = ∫
2π − ∞
dΨ(r,ϕ,ω)e jωt (1.54a)

© 2001 by CRC Press LLC

 and over the entire plane

+∞
1
3 ∫∫∫
ψ(r,ϕ,t) = dΨ(u,v,ω)e j (ωt −ux − vy) (1.54b)
(2π) − ∞

The pressure field due to a small element of noise sources at a sensor at a depth
z1 is given by (1.45),

∞
1 ω
p1 (t) = ∫ dΨ(r,ϕ,ω)e jωt ∑ φ m (z1 )φ*m (zs )H01 (km r)rdrdϕ
2π − ∞ 2H m
(1.55)

where dΨ(r,ϕ,ω)rdrdϕ is the strength of the noise sources. The total

pressure is obtained by integrating (1.55) over the entire plane. The spectral
representation (1.54b) is used in carrying out the integration.
∞
φ m (z1 )φ*m (zs )
1 ω
p1 (t) = ∫ 2H e ∑ +∞ ∞
jωt
(1.56)
(2π)2 ∫ ∫ dΨ(u,v,ω)∫ H0 (kmr)J0 (sr)rdr
1
−∞ m
−∞ 0

Finally, we note that [17]

∞
2 1
∫ H (k r)J0 (sr)rdr =
1

π km2 − s 2
0 m
0

and obtain

∞ +∞
1 ω jωt φ (z )φ* (z )
p1 (t) = 3 ∫ e ∫ ∫ dΨ(u,v,ω)∑ m 21 m2 s (1.57a)
2π − ∞ 2H −∞ m km − s

The pressure field at another sensor placed at a depth z2 is given by

∞ +∞
1 ω jωt φ (z )φ* (z )
p2 (t) = 3 ∫ e ∫ ∫ dΨ(u,v,ω)∑ m 22 m2 s (1.57b)
2π − ∞ 2H −∞ m km − s

The cross-spectrum between these two sensor outputs may be evaluated as

follows

© 2001 by CRC Press LLC

 S12 (ω)dω = E  dP1 (ω) dP2* (ω)
1 1 1
2π  2π 2π 
φ m (z1 )φ*m (zs )φ*n (z2 )φ n (zs ) (1.58)
ω 1 2
= 3 2 dω ∑ ∑ ∞ sSψ (s,ω)
2π H 2π m n ∫ 2 ds
0 (k − s )(kn − s )
22 * 2
m

∞ ∞
1 1
where p1 (t) = ∫
2π − ∞
dP1 (ω)e jωt and p2 (t) = ∫
2π − ∞
dP2 (ω)e jωt .

Simplifying (1.58) we obtain

φ m (z1 )φ*m (zs )φ*n (z2 )φ n (zs )

ω 2
S12 (ω) = ∑∑∞
2π 3 H 2 m n ∫
sSψ (s,ω)
ds
(1.59)
*2
0 (km2 − s 2 )(k − s 2 )
n

For spatially white noise sources, Sψ (s,ω) = S0 (ω) , and the integral in
(1.59) reduces to

∞
sSψ (s,ω) 4 k S (ω)
∫ (k
0
2
m
2 *2
− s )(k − s )
n
2
ds =
π 2
ln( m* ) 2 0 *2
kn (km − kn )

For details on the derivation and numerical results the reader is urged to see [20]
and [21] where a more general model is dealt with. As an illustration of the
general variation of the spectrum and the coherence, the numerical results for a
channel where H= 4λ are shown in fig. 1.24. The noise sources are located on
an annular ring with inner radius = 100λ and outer radius =1000λ. The noise is
presumed to have been generated by wind (m=1). For coherence calculation one
sensor is kept fixed at 2λ and the other sensor is moved along the depth axis.
We considered two types of bottoms, rigid and soft bottoms (ρ 1 =1gms/cc,
c1=1500 m/sec, ρ2=2.0 gms/cc, c2=1600 m/sec). The results are shown in fig.
Sz (ω)
1.24. The y-axis in fig. 1.24a represents a ratio, where Sz (ω) is the
Sψ (ω)
spectrum of the pressure field at z. It is interesting to note that in the hard
bottom channel, because of the trapping of energy, the spectrum is always
greater than that in the soft bottom channel and it increases with depth. The
coherence as a function of the sensor separation (fig.1.24b) is highly
oscillatory, roughly following the interference pattern of vertically travelling
modes.

© 2001 by CRC Press LLC

 0.02
Spectrum

1
0.01

0.00
0 1 2 3 4

Depth/lambda
(a)
Coherence Magnitude Phase
0 0.5 1.0 -180 -90 0 90 180
-2

-1

(2)
Separation in wavelengths

(t)

(t)

(2)

(b)
Figure 1.24: Spectrum and Coherence as a function of depth. (a) Spectrum. Empty
circles represent soft bottom and the filled circles, hard bottom. (b) Coherence as a
function of sensor separation. (1) Hard bottom. (2) Soft bottom. The magnitude is
shown in the left panel and phase in the right panel.

© 2001 by CRC Press LLC

 The stochastic wavefield in a bounded medium is extremely complex
largely due to the interference of multiply reflected waves. What we have
described above is a simple example involving just two reflectors. Imagine the
complexity of wavefield in a room with six reflectors! Often, in such acomplex
situation, we model the wavefield as a diffused wavefield which we have briefly
discussed in (1.4.2).

§1.5 Multipath Propagation:

When wave energy travels from point A to point B along more than one path,
the propagation is said to be multipath propagation. Such a propagation regime
is the result of local micro inhomogeneities or point scatterers. Since the local
micro inhomogeneities or point scatterers are stochastic in nature it is
appropriate to characterize the resulting multipath propagation under stochastic
frame work. It often happens that the local inhomogeneities as well as the
point scatterers are time varying, consequently multipath propagation is also
time varying, sometimes very rapidly as in mobile wireless communication.
The phenomenon of bending of a ray as it passes through a blob of
inhomogeneity is illustrated in fig. 1.25. Scattering from a point scatterer
produces a similar effect but is more pronounced as the wave energy is scattered
in all directions.
The dominant effect of the multipath propagation is a large collection
of rays, known as microrays [22] impinging on an array from different
directions. In a weakly scattering medium a micro-ray follows a path very close
to that of an unperturbed ray. Hence all micro-rays would have gone through
similar macro variations but with slight path variations. A sensor will then
receive a train of coherent signals arriving at different time intervals, possibly
Doppler shifted or scaled versions of the transmitted signal, when the scatterers
are in motion. The amplitude variation among the microrays is likely to be less
and of lower significance. Multipath propagation can cause severe loss of signal
fidelity particularly in those problems where the wave propagation is through
natural channels, such as in underwater, solid earth, and urban environment, for
wireless communication. We shall consider two types of multipath
environments. First, we shall consider a random distribution of blobs of
inhomogeneity as encountered in underwater channels and in subsurface
channels. Next we shall consider a collection of scattering points around a
transmitter typically encountered in wireless propagation. There are other
possibilities but only the above two environments are well understood. The
inhomogeneities in the medium are either stationary or very slowly moving;
hence the Doppler effect can be ignored in the first instance.
1.5.1 Random Filter Model: The conceptual picture of many microrays closely
following an unperturbed ray and reaching a sensor is mathematically described
through the random filter model. The output of a sensor may be expressed as
follows:

© 2001 by CRC Press LLC

 1
f n (t) = ∑ ak g0 (t − tn − τ k − δτ k (t))
Nrays k
(1.60)

where
g0(t): waveform transmitted by a source.
t : arrival time of unperturbed ray at nth sensor.
n
τ k : relative arrival time of kth micro-ray.
δτ k (t) : relative delay due to time varying scattering effect.
ak : coefficient of attenuation for kth ray.
Nrays : Number of rays.

In the frequency domain (1.60) may be expressed as

1
∞
 1  jωt
f n (t) =
2π −∫∞
dG0 (ω)  ∑
 Nrays k
ak e − jω(t n + τ k +δτ k (t ))
e

(1.61)
∞
1
= ∫
2π − ∞
dG0 (ω)H(ω,t)e jωt

where
1
H(ω,t) = ∑ ak e − jω(tn + τ k +δτ k (t ))
Nrays k

is the time-varying random filter representation of the multipath propagation

effect. Here both τ k and δτ k (t) are modeled as random variables; hence the
filter is known as a random time-varying filter. Sea experimental results as
reported in [23] indicate that the delays may be approximated as a uniformly
distributed random variable and that they are uncorrelated when the sensor
separation is of the order of a few hundred feet. Using the above model of
fluctuations we can compute the mean and the variance of the time-varying
filter transfer function. The mean is given by

© 2001 by CRC Press LLC

 A B

Tr Rc

prism

(a)

A B

Tr Rc

(b)

Figure 1.25: Phenomenon of a ray bending as it passes through a blob of

inhomogeneity is shown above. For comparison a ray is shown to bend as it passes
through a prism. Scattering by a point scatterer produces a similar effect.

 1 
E{Hn (ω,t)} = E  ∑ ak e − jω(t n + τ k +δτ k (t )) 
 Nrays k 

© 2001 by CRC Press LLC

 =
1 − jωt n
Nrays
e ∑ { } {
E{ak } E e − jωτ k E e − jωδτ k (t ) }
k (1.62)
− jωt n
= An e Φ 0 (ω)Φ1 (ω,t)

where An =
1
∑ E{ak },
Nrays k
{
Φ 0 (ω) = E e − jωτ k } and Φ1 (ω,t) =

{ }
E e − jωδτ k (t ) . For a uniformly distributed random variable in the interval
∆t
± ,
2
∆t
sin(ω )
Φ 0 (ω) = 2
∆t
ω
2
1.5.2 Point Scatterers: In radio communication, the electromagnetic waves
travelling along straight line ray paths may encounter obstacles which would
reflect or scatter the incident wavefield. Since frequency used is very high (800-
1000 MHz, λ ≈ 1meter ) most of the obstacles are likely to be much larger
than the wavelength. There will be significant reflections and corner
diffractions. The air medium is however assumed to be homogeneous and free
from any scattering. Furthermore, a transceiver used in a modern wireless
communication system is likely to be in motion causing a significant Doppler
shift. One consequence of the reflection or the scattering of waves is the
possibility of more than one ray path connecting a transmitter and a receiver.
Such multipath propagation in wireless communication is quite common. As a
result of this, the different components of the wavefield reach a receiver at
slightly different time instants, at different angles and with different Doppler
shifts, but coherently. Thus, multipaths are characterized by the following
attributes: (i) delay diversity (0-10 micro seconds), (ii) angular diversity (5-6
degrees) and (iii) Doppler shift (0 ± 50 Hz).
Delay diversity: The signals arrive at a receiver (usually a single sensor) at
different time intervals, the delay being due to different path lengths. A long
delay implies a weak signal due to multiple reflections, attenuation in the air
and also by geometrical spreading. The quantity of great interest is the power
received at a sensor as a function of delay, known as power delay profile. A
typical power delay profile is sketched in fig. 1.26. Referring to this figure we
define an excess delay spread τ e as the delay within which 90 percent of the
total power reaches the sensor. There are two other parameters commonly used
to characterize a power delay profile,

© 2001 by CRC Press LLC

 1
Power

.1

τe

.01
0 1 2
Delay (in microsecs)

Figure 1.26: A sketch of the power delay profile. Excess delay parameter τe is
shown.

Mean delay (τ) =

∑p τ k k

∑p k

rms delay spread =

∑p τ 2
k k
− (τ)2 1.63)
∑p k

where pk is the power in the kth path arriving with a delay τ k . A cumulative
m m
plot of power, a plot of ∑pk =1
k versus ∑τ
k =1
k , is useful in deciphering different

groups of multipaths.

© 2001 by CRC Press LLC

 Angular diversity: The angle of arrival of different multipaths is a random
variable. We define mean angle and rms angle spread as follows:

Mean angle (ϕ) =

∑p ϕ k k

∑p k

rms angle spread =

∑p ϕ k
2
k
− (ϕ)2 (1.64)
∑p k

Doppler Shift: It is common that a transceiver is moving, often at high speed,

while the base station antenna is stationary. Consider a speeding car carrying a
transceiver past a stationary scatterer (see fig 1.27). The component of car
velocity in the direction of scatterer is v cos(ϕ) where v is the speed of the car
moving along the road.
v
The doppler shift will be given by ∆ω = ω c cos(ϕ) where ω c is
c
the carrier frequency. Plot of the Doppler shift as a function angle is shown in
fig. 1.27 (inset). As a numerical example, let v=120 km/hour, i.e., 100/3
meters/sec, c=3x108 meters/sec, and ω c =18π 10 the Doppler shift is equal
8

to ∆ω max = 200π .
It is interesting to compute a power delay profile for a simple model of
a street lined with buildings which act as obstacles. For simplicity we have
assumed regularly spaced building (spacing=15m and gap=5m) on either sides of
a street, twenty meters wide (see fig. 1.28a). A stationary source in the middle
of the street emits a spike waveform. A suite of rays starting from the source
and ending at the base station antenna, which is 200 meters away from the
street, were traced taking into effect all possible reflections at different
buildings. A sample of five rays thus traced are shown in fig. 1.28a. Perfect
reflection was assumed (reflection coefficient =1) but geometrical spreading was
taken into account. The power delay profile was computed by noting the power
received at computed delay. The computed power delay profile is shown in fig.
1.28b. Other parameters are mean delay=0.042 microsec, rms delay=0.149
microsec and excess delay spread=0.05 microsec.
1.5.3 Local Scattering: A signal emitted by a transceiver is often scattered by
point scatterers in the immediate neighborhood of the transceiver. What reaches
a distant array is a collection of plane wavefronts, differing in phase, angle of
arrival and amplitude but all wavefronts remain correlated. Let (θ 0 + δθ k ) be
the direction-of-arrival (DOA) of the scattered wavefront from the kth scatterer
and θ 0 is the nominal DOA of the direct wavefront from the source (see fig.
1.29). We

© 2001 by CRC Press LLC

 ϕ
Y
Mobile

v
c ω0
Doppler shift

Base station
0
-2 -1 0 1 2
angle

Figure 1.27: A moving car and stationary scatterer will produce a Doppler shift at
the base station antenna. Inset shows the Doppler shift as a function of angle.

assume that there are L scatterers in the vicinity of the source. The array output
may be expressed as

 L −1 
f(t) = ∑ α k e − jω c δt k a(θ 0 + δθ k ) f 0 (t) (1.65)
k = 0 

where a(θ 0 + δθ k ) is the array response vector defined as

a(θ 0 + δθ k ) =
ωcd ωc 2d ω c ( M −1)d
−j sin(θ 0 +δθ k ) −j sin(θ 0 +δθ k ) −j sin(θ 0 +δθ k )
c c c
[1,e ,e ,...,e ]

for a equispaced (spacing = d) linear array (see chapter 2 for more details on the
array response), α k is the coefficient of scattering (complex) and δ tk is the
delay (with respect to direct wavefront) of a wavefront from the kth scatterer.
Note that k=0 refers to the direct wavefront, for which α 0 = 1 and δ t0 = 0 . It

© 2001 by CRC Press LLC

 Figure 1.28: A simple model of a street lined with buildings which act as obstacles.
(a) A sample of five rays is shown as they propagate through the street. (b)
Computed power delay profile.

© 2001 by CRC Press LLC

is assumed that the source emits a narrowband signal, that is, bandwidth
2π
<< . We shall express a(θ 0 + δθ k ) in Taylor's series expansion,
δ tk

∂a(θ) (δθ k )2 ∂ 2 a(θ)

a(θ 0 + δθ k ) = a(θ 0 ) + δθ k + +...
∂θ θ=θ 0 2! ∂θ2 θ=θ 0
≈ a(θ 0 ) + δθ k a ′(θ 0 )
(1.66)

Using (1.66) with the first two terms only in (1.65) we obtain the following
approximate result [24]

f(t) ≈ [φ 0 a(θ 0 ) + φ1a ′(θ 0 )] f 0 (t) (1.67a)

and for uniform linear array

2 πd
 φ1 2πd j sin θ 0 
f(t) ≈ φ 0 1, (1 + j cosθ 0 )e λ , + ... f 0 (t) (1.67b)
 φ0 λ 
L −1 L −1
where φ 0 = ∑ α k e − jω c δt k and φ1 = ∑ α k δθ k e − jω c δt k . The covariance
k =0 k =0
matrix of the array output is of interest in the DOA estimation. Using the first
order approximation in (1.67a), we obtain

c f ≈ Lσ θ2 σ α2 [a(θ 0 )a H (θ 0 ) + σ θ2 a ′(θ 0 )a ′ H (θ 0 )] (1.68)

where we have assumed that α k and δθ k are independent random variables

whose variances are σ and σ θ , respectively. Note c f is a sum of two rank
2 2
α
one matrices; hence its maximum rank will be two.
For uniform linear array we can derive a more specific result. For
small δθ , sin δθ ≈ δθ and cosδθ ≈ 1 we have a(θ 0 + δθ) m = [ ]
2 πd
j m(sin θ 0 +δθ cos θ 0 )
λ
e . The coefficient of scattering is assumed to be
uncorrelated. The covariance matrix is given by

© 2001 by CRC Press LLC

 Linear Array
YY Y Y Y Y Y

θ0

2∆

×
×
Source
× ×
× Scatterers

Figure 1.29: Local scattering model. A signal source (transceiver) is surrounded by

scatterers. The signal reaching array consists of a suite of plane waves arriving with
different DOAs and delays.

{c } f mn
= Lσ 2f 0 σ α2 E{a(θ 0 + δθ)a H (θ 0 + δθ)}mn
2 πd
j ( m − n) sin θ 0  j 2 πd ( m − n)δθ cos θ 0  (1.69a)
= Lσ 2f 0 σ α2 e λ
E e λ 
 

where σ 2f 0 is the variance of the source signal and σ α2 is the variance of the
coefficient of scattering. Assuming δθ is uniformly distributed over a range
±∆ the expected value in (1.69a) may be shown to be

d
 j 2 λπd ( m − n)δθ cos θ 0  sin 2π λ ∆(m − n)cosθ 0
E e = d
(1.69b)
  2π ∆(m − n)cosθ 0
λ

Using (1.69b) in (1.69a) we obtain

c f ≈ Lσ 2f 0 σ α2 D(θ 0 )QD H (θ 0 ) (1.70)

© 2001 by CRC Press LLC

where
d
sin 2π ∆(m − n)cosθ 0
{Q}mn = λ
d
2π ∆(m − n)cosθ 0
λ
and
 j 2 λπd sin θ 0 j 4 λπd sin θ 0 j 6πd sin θ 0 
D = diag 1, e ,e ,e λ ,...  .
 

Note that Q is a symmetric toeplitz matrix with real eigenvalues. Let

M
Q = ∑ λ i vi viH be the eigen decomposition of Q matrix. It is known that
i=1
λ1 ≈ λ 2 ≈ λ 3 ≈ ... λ r ≈ 1 where r is the rank of Q and the remaining
eigenvalues are insignificant. The rank is approximately given by

r ≈  2∆ cosθ 0  where [x] stands for the largest integer greater than x.
Md
 λ 
The eigenvectors corresponding to the significant eigenvalues are known as
discrete prolate spheroidal sequences (DSSP) [25]. We have computed the rank
of the matrix Q as shown in table 1.1

§1.6 Propagation through Random Medium:

Imaging of a medium through which a wave is propagating is one of the
important applications of the sensor arrays, in particular, in medical
tomography and in seismic exploration where reflected or scattered field is used
for imaging. When the inhomogeneous medium is made of simple layers as in
subsurface imaging, the reflected field is all that is required for imaging
purposes. This is by and large true in seismic imaging which we shall cover in
chapter 8. But, when the medium is highly inhomogeneous, the scattering
dominates. Currently, only weak scattering has been extensively used for
imaging. In this section we would like to review the weak scattering and point
out how the scattered field enables imaging of an inhomogeneous medium.
Later in chapter 7 we shall describe the use of sensor arrays for imaging.
1.6.1 Acoustic Field: The acoustic pressure field in an inhomogeneous medium
satisfies the following wave equation,

1 ∂2 f
∇ f−
2
=0 (1.71)
c(x, y, z)2 ∂t 2

© 2001 by CRC Press LLC

 ∆cosθ0 rank
0.03 rad 2
0.1 7
0.15 10

Table 1.1: Rank of Q matrix for different values of ∆cosθ 0 and M=64 and
λ=d/2.

where c is the acoustic wave speed which is a function of the space coordinates.
We shall assume that the density remains constant. (Equation (1.2b) deals with
both speed and density variations.) Let c = c0 + δc(x, y, z) where c0 is the
mean wave speed in the medium and δc(x, y, z) is the fluctuation around the
mean value. A medium is said to be weakly inhomogeneous if
δc(x, y, z) << c0 . The wave equation for inhomogeneous medium (1.71)
reduces to

1 ∂2 f 2δc ∂ 2 f
∇ f − 2 2 =− 3
2
(1.72)
c0 ∂t c0 ∂t 2

where the term on the right hand side of (1.72) represents the contribution due
δc
to speed fluctuations in the medium. Let us represent = εδc̃ where δc̃ is
c0
a normalized function with unit root mean square (rms) magnitude and ε (<<1)
is a constant. Eq(1.72) may be expressed as

1 ∂2 f 2ε ∂ 2 f
∇2 f − = − δc̃ 2 (1.73)
c02 ∂t 2 c02 ∂t

We shall now try to find a series solution of (1.73). Let the series solution be
given by

f = f 0 + εf 1 + ε 2 f 2 +...+ε P f P (1.74)

On substituting (1.74) in (1.73) and equating the coefficients of the powers of

ε to zero we obtain the following system of partial differential equations,

1 ∂2 f 0
∇2 f 0 − =0
c02 ∂t 2

© 2001 by CRC Press LLC

 1 ∂2 f 1 2 ∂2 f 0
∇2 f 1 − = − δc̃
c02 ∂t 2 c02 ∂t 2
... (1.75)

1 ∂ fp ∂ 2 f p−1
2
2
∇2 f p − = − δc̃
c02 ∂t 2 c02 ∂t 2

We like to solve these equations in an unbounded medium. f i (r) , for i=1, 2,

..., must satisfy the radiation condition. Note that the Green’s function is the
same for all equations in (1.75); only the driving function on the right hand
side differs. The solution of ith equation may be expressed as

2 ∂ 2 f i −1
+∞ − c̃( x ′ , y ′ , z ′ )
1 c02 ∂t 2 jk⋅(r − r ′ )
4π ∫− ∫∞ ∫
f i (r,t) = e dx ′dy ′dz ′ (1.76)
r − r′

Consider the case of plane wave illumination,

f 0 (r, t) = A0 e j (ωt −u0 x − v0 y − w 0 z ) , where u0 , v0 , and w0 are related to the
direction cosines of the traveling plane wave (see p. 9). Using the above
illumination function in (1.76) we obtain the following expression for the first
order term, also known as Born approximation,

+∞
1 2k02δc̃e j (k 0 r − r ′ )
4π ∫− ∫∞ ∫
f 1 (r, t) = f 0 (r' , t)dx ′dy ′dz ′ (1.77)
r − r′

We shall now use the Sommerfeld formula [26]

e j ( k0 r − r ′ ) e j ( k0 ρ + z )
2 2

=
r − r′ ρ2 + z 2
(1.78)
∞ 2π
1 λdλ
∫ ∫
− λ2
− k 02 z−z′
= e j (λ( x − x ′ ) cos θ+ λ( y − y ′ ) sin θe dθ
2π 0 0 λ2 − k02

to simplify (1.77)

© 2001 by CRC Press LLC

 f 1 (r, t) =

λdλdθ  δc̃( x ′, y ′, z ′ ) f 0 (r' , t) 

∞ 2π
k02
4π 2 ∫∫
−∞ 0
∫
λ2 − k02  Γ e j (λ( x − x ′ ) cos θ+ λ( y − y ′ ) sin θ)e − λ2 − k 02 z−z′

dx ′dy ′dz ′ 
(1.79)

where Γ represents the space occupied by the scattering medium (see fig. 1.30).
Define u ′ = λ cos θ and v ′ = λ sin θ . It follows that du ′dv ′ = λ dλ dθ .
Equation (1.79) can be expressed as

f 1 (x, y,ω) =
+∞
− jk02
∫ ∫ ∫ δc̃( x ′, y′, z ′)e
− j (( u ′ −u 0 ) x ′ +( v ′ − v 0 ) y ′ +( k 02 − u ′ 2 − v ′ 2 − w 0 ) z ′
dx ′dy ′dz ′
4π 2 −∞Γ

− j ( k 02 −( u ′ 2 + v ′ 2 ) − w 0 )l
e
e j ( u ′x + v ′y) du′dv ′
k − (u ′ + v ′ )
2
0
2 2

(1.80)

Notice that the inner integral over Γ represents a three dimensional Fourier
transform of δ c̃( x ′, y ′, z ′ ) . Hence, (1.80) can be written in the frequency
domain as follows:

f 1 (x, y, ω)
∆c̃(u′ − u0 , v ′ − v0 , k02 − u′ 2 − v ′ 2 − w0 )
2 +∞
− jk (1.81)
=
4π
0
2
−∞
∫∫e − j ( k 02 −( u ′ 2 + v ′ 2 ) − w 0 )l
e j ( u ′x + v ′y) du′dv ′
k − (u ′ + v ′ )
2
0
2 2

j k 02 −( u ′ 2 + v ′ 2 )l
Note that the factor e rapidly decays for (u′ 2 + v ′ 2 ) > k02 and
l > 0 . Such waves correspond to evanescent waves, which will be significant
only in the immediate neighborhood of the scattering object. The presence of
density fluctuation merely introduces an extra term in (1.81). We will not go
into the details but cite a reference where the density fluctuation is accounted for
[27]. For a two dimensional object the scattered field has been obtained by Kak
[28].

© 2001 by CRC Press LLC

 y
Observation Plane
x

z
l

Illumination

Figure 1.30: Γ represents the space occupied by the scattering medium. δ c=0 outside
Γ. The scattered field is evaluated on observation plane l units above Γ .

+∞ j ( k 02 − u ′ 2 − v 0 )l
jk 2 e
f 1 (x, ω) = 0 ∫ ∆c̃(u′ − u , k − u′ − v0 ) e ju ′x du′
2 2
0 0
4π −∞ k − u′
2
0
2

(1.82)

1.6.2 Far Field Approximation: When the scattering object is finite and the
array is placed quite far from the object we shall apply in (1.77a) the far field
e jk⋅(r − r ′ ) e jk (r − r ′ ⋅r̂)
approximation, namely, ≈ where r̂ is the unit vector in
r − r′ r
the direction of r . The error due to this approximation, specially in the
numerator, is illustrated in fig. 1.31. To assess the quantitative effect, consider
the binomial expansion of

r′ r′
( )2 − 2 cos θ
r − r′ = r 2 + r ′ 2 − 2rr ′ cos θ ≈ r(1 + r r +...)
2
r ′2
= (r + − r ′ cos θ +...) ≈ (r − r ′ cos θ )
2r

© 2001 by CRC Press LLC

 Sensor
s

y r − r′
r

r' o'
x
o

Figure 1.31: In far field approximation r − r′ is approximated by o’s.

r′2
The error is of the order of in the first term of binomial expansion. This
2r
πr ′ 2
error will introduce a phase error, . For the phase error to be small we
λr
must have r ′ << λr or the largest dimension of the object must be much
2

smaller than λr . Using this as far field approximation in (1.77a) we obtain

f 1 (r, t) =

e − j (ωt − k 0 r ) (1.83)
∫ 2k δc̃( x ′, y′, z ′)e
− jk 0 ( r ′ ⋅ r̂) − j (u 0 x ′ + v 0 y ′ + w 0 z ′ )
2
0 e dx ′dy ′dz ′
4πr
Γ

In (1.83) we note that

© 2001 by CRC Press LLC

 e − jk 0 ( r ′ ⋅r̂) = e − jk 0 (ax ′ +βy ′ + γz ′ )
= e − j (ux ′ + vy ′ + wz ′ )
where u = k0 α, v = k0β, w = k0 γ and (α,β, γ ) are the direction cosines of
unit vector r̂ . Using this simplification in (1.83) we obtain

f 1 (r, t) =
e − j (ωt − k 0 r )
4πr ∫Γ
2k02δc̃( x ′, y ′, z ′ )e − j[(u −u0 ) x ′ +(v − v0 ) y ′ +(w − w 0 ) z ′ ]dx ′dy ′dz ′

e − j (ωt − k 0 r ) 2
= 2k0 ∆c̃(u − u0 ,v − v0 , w − w0 ) (1.84)
4πr

where ∆c̃(.) is the Fourier transform of δc̃ . The result derived in (1.84) has
some significance. In the far field region the first order scattered field has the
form of a point scatterer (term outside the square brackets), that is, spherical
waves. The term inside the square brackets depends on the Fourier transform of
the speed fluctuations, evaluated at spatial frequencies determined by the
direction of illumination and the direction of sensor. We shall later in chapter 6
exploit this result for reconstruction of speed fluctuations.
1.6.3 Multisource Illumination: The basic fact used in tomography is that
when an object is illuminated from different directions the scattered field
contains useful information for three dimensional reconstruction. This property
of the wavefield is elaborated in this simple example. Consider a source and
sensor array on opposite sides of an object to be imaged (see fig. 1.32). Let the
a
mth source be fired and the scattered field be sensed by the sensor array. Let r m
be a vector to the mth source and r be a vector to the nth sensor. r ′ is vector
b
n
to a scattering element. The scattered field due to the scattering element at nth
sensor is given by

jk r b − r ′ jk r ′ −r a
1 2k02δc̃e 0 n e 0 m
∆f 1 (m, n) = dx ′dy ′dz ′ (1.85)
(4π)2 r′ − r bn r ′ − r am

Note that the source and sensor arrays are equispaced linear arrays in the y=0
a b
plane. Therefore, the tips of the vectors r m and r n will lie at
[md,0, −La , m = 0, ±1, ±2,⋅⋅⋅] and [nd, 0, Lb , n = 0, ±1, ±2,⋅⋅⋅],
respectively. Using Sommerfeld formula (p. 61) to express

© 2001 by CRC Press LLC

 a
r - r' b
m rn - r '
b
a rn
r
m m n

r′

object
La Lb

Sensor array
Source array

x
y
z

Figure 1.32: An array of sources is used to illuminate an object and the scattered
wavefield is received by another array of sensors. The object Fourier transform can
be directly related to the 2D Fourier transform of the received wavefield.

jk r a − r ′ +∞
e − j k 0 −u − v ( z ′ + La ) j[(md − x ′ )u − y ′v ]
2 2 2
e 0 m −j
2 π ∫− ∫∞ k02 − u 2 − v 2
= e dudv (1.86a)
r am − r ′
z ′ < La

jk r b − r ′ +∞
e − j k 0 −u − v (− z ′ + Lb ) j[(nd − x ′ )u − y ′v ]
2 2 2
e 0 n −j
2 π ∫− ∫∞
= e dudv (1.86b)
r′ − r bn k02 − u 2 − v 2
z ′ < Lb

Next we compute the 2D discrete Fourier transform of ∆f 1 (m, n) . From

(1.85) it may be seen that the discrete Fourier transform of ∆f 1 (m, n) is equal
to the product of discrete Fourier transforms of Green’s functions (i.e., left hand
side of (1.86)). These are given as follows:

© 2001 by CRC Press LLC

  e jk 0 r m − r ′ 
a

DFT  a 
 r m − r′ m
(1.87a)
+∞ − j k 02 −u 2 − v 2 z ′ + La
−j e
= ∫ ∫
2π − ∞ k0 − u − v
2 2 2
δ(ud − u1 )e − j[ x ′u + y ′v]dudv

 e jk 0 r n − r ′ 
a

DFT  
 r′ − r n
a
n
(1.87b)
+∞ − j k 02 −u 2 − v 2 − z ′ + La
−j e
= ∫ ∫
2π − ∞ k02 − u 2 − v 2
δ(ud − u2 )e − j[ x ′u + y ′v]dudv

The subscripts m and n on the left hand side refer to discrete Fourier transforms
with respect to index m and index n, respectively. Using (1.87) in (1.85) we
obtain the 2D Fourier transform of the response of a scattering element

∆F1 (u1 , u2 ) =
 − j k 02 −( u1 ) 2 − v ′ 2 ( z ′ + La ) − j k 02 −( u2 ) 2 − v 2 ( − z ′ + Lb ) 
e d
e d

+∞  
−k02 u u (1.88)
2 ∫ ∫
 k0
2
− ( 1 2
) − v ′ 2
k0
2
− ( 2 2
) − v 2
dx ′dy ′dz ′
2π − ∞  d d 
 − j[( u1 + u2 ) x ′ +( v ′ + v) y ′ ] 
δc̃e d d dv ′dv 

Equation (1.88) is now summed over all scattering elements covering the entire
object. We obtain

F1 (u1 , u2 ) =
+∞
−k02 u1 u2  u u 
2π 2 ∫ ∫ ∆c̃( d +
−∞
d
, v + v ′,  k02 − ( 1 )2 − v 2 − k02 − ( 2 )2 − v ′ 2  )
 d d 

© 2001 by CRC Press LLC

 u2 2 u1 2 2
− j k 02 −( ) − v ′ 2 La − j k 02 −( ) − v Lb
d d
e e
× dv ′dv (1.89)
u u
k − ( 1 )2 − v ′ 2
2
0 k − ( 2 )2 − v 2
2
0
d d

As a special case we consider an object having a slow variation in the y

direction. Then, δC(u,v, w) ≈ δC(u, w)δ(v) and (1.89) reduces to

F1 (u1 , u2 ) =
∞
k02 u1 u2  2 u1 2 u 
π ∫ ∆c̃( d +
−∞
,  k0 − ( ) − v 2 − k02 − ( 2 )2 − v 2  )
d  d d 
u1 2 2 u2 2 2
− j k 02 −( ) − v La − j k 02 −( ) − v Lb
d d
e e
× dv (1. 90)
u u
k − ( 1 )2 − v 2
2
0 k − ( 2 )2 − v 2
2
0
d d

If we were to use a line source in place of a point source (1.90) reduces to, by
letting v=0,

F1 (u1 , u2 ) =
u1 u2 u u
2k02 ∆c̃( + , k02 − ( 1 )2 − k02 − ( 2 )2 ) ×
d d d d (1.91)
u u
− j k 02 −( 1 ) 2 La − j k 02 −( 2 ) 2 Lb
d d
e e
u u2 2
k02 − ( 1 )2 k02 − ( )
d d

The above result is identical to that given in [29].

1.6.4 Scattering of EM Field: The electric properties of a medium are dielectric
permitivity (ε), magnetic permeability (µ) and conductivity (σ). The external
electromagnetic wavefield will induce electric current in inhomogeneous
medium. The induced current will in turn create electromagnetic wavefields
outside the medium. The induced current density at a point is given by

J1 (x, y, z) = (κ12 (x, y, z) − κ 20 )E(x, y, z) (1.92)

© 2001 by CRC Press LLC

where κ12 (x, y, z) = ω 2 µ 0 ε1 (x, y, z) − jωµ 0 σ1 (x, y, z) , κ 02 = ω 2 µ 0 ε 0
and E(x, y, z) is an electric field which induces electric the current [27, 30]. It
is assumed that the space outside the inhomogeneous medium is air, and hence
σ 0 = 0 . The electric field at a point outside the inhomogeneous medium is
given by

exp( jk0 r − r ′ )
E1 (x, y, z, ω ) = jω ∫ J1 ( x ′, y ′, z ′, ω ) dx ′dy ′dz ′ (1.93a)
Γ
4 π r − r′

and the magnetic field is given by

exp( jk0 r − r ′ )
H1 (x, y, z, ω ) = ∫ ∇ × J1 ( x ′, y ′, z ′, ω ) dx ′dy ′dz ′ (1.93b)
Γ
4 π r − r′

Under Born approximation E(x, y, z) ≈ E 0 e jk0 ( αx + β y+ γz) where (α,β, γ )

are direction cosines of the wave vector and E 0 is the incident electric field.
Note that E1 is in the same direction as E 0 but H1 is ⊥ to the incident
vector. Let the incident electric field be in the z direction, then the scattered
magnetic field will in the (x,y) plane. The x and y components of the magnetic
field are given by

H1x (x, y, z,ω) = (1 + jk0 )E0 ×

exp( jk0 r − r ′ ) (1.94a)
∫ (κ (x' , y' , z' ) − κ 20 )(y − y ′ ) dx ′dy ′dz ′
2
1 2
Γ
4π r − r ′

and

H1y (x, y, z,ω) =

exp( jk0 r − r ′ )
(1 + jk0 )E0 ∫ (κ12 (x' , y' , z' ) − κ 20 )(x − x ′ ) 2 dx ′dy ′dz ′
Γ
4π r − r ′
(1.94b)

Analogous to (1.81) we can write, in the frequency domain, an expression for

the scattered electric and magnetic fields as

© 2001 by CRC Press LLC

 H1y (x, y, z,ω) = (1 + jk0 )E0 ×
exp( jk0 r − r ′ ) (1.95a)
∫ (κ1 (x' , y' , z' ) − κ 0 )(x − x ′) dx ′dy ′dz ′
2 2
2
Γ
4π r − r ′
and

H1x (x, y, l,ω)

Κ1 (u′ − u0 , v ′ − v0 , k02 − u′ 2 − v ′ 2 − w0 )
E  +∞  (1.95b)
= 02 ∫ ∫  e j ( k 02 −( u ′ 2 + v ′ 2 ) − w 0 )l  du′dv ′
4π − ∞  ( jv ′ )e j ( u ′x + v ′y) 
 k0 − (u′ + v ′ )
2 2 2


H1y (x, y, l,ω)

Κ1 (u′ − u0 , v ′ − v0 , k02 − u′ 2 − v ′ 2 − w0 )
E  +∞  (1.95c)
= 02 ∫ ∫  e j ( k 02 −( u ′ 2 + v ′ 2 ) − w 0 )l  du′dv ′
4π − ∞  ( ju′ )e j ( u ′x + v ′y) 
 k0 − (u′ + v ′ )
2 2 2


where Κ1 stands for the 3D Fourier transform of

( ω µ 0 ε1 (x, y, z) − jωµ 0 σ1 (x, y, z) − ω µ 0 ε 0 ). From equations (1.81) and
2 2

(1.95) it may be inferred that a strong similarity between the scattered acoustic
field and the scattered EM field exists.

§1.7 Exercises:

1. Show that if second order approximation is used in (1.66) an additional term

L 2 4
will appear in (1.68), namely, σ α νθ a ′′(θ 0 )a ′′ H (θ 0 ) , where
4
νθ4 = E{δθ 4 } . The pdf of δθ must be symmetric.
2. In equation (1.53c) let the noise be a bandlimited process with a spectrum
δ δ
limited to ω0 − ≤ ω ≤ ω 0 + . Show that
2 2
d
c12 (τ) = c11 (τ)J0 (ω 0 )
c

© 2001 by CRC Press LLC

 d
under the condition that δ << 1. What is the implication of this
c
condition?[31].
3. Consider a diffused field in 3D space. One possible model is that point
sources are assumed to be uniformly distributed over a large sphere of radius R.
Each point source emits a stochastic waveform which is uncorrelated with the
waveforms emitted by all other point sources. Show that the field at a sensor
d
placed on the z-axis at a distance units from the origin of the coordinates is
2
given by

ω
−j R ∞ d
c jω(t − cos θ)
1 e
f 1 (t) ≈
2π R ∫ dF(ω,θ,ϕ)e
−∞
2c

where θ,ϕ are respectively elevation and azimuth of the point source. Place
d
another sensor, also on the z-axis, at a distance - . Show that the cross
2
correlation between the outputs is given by

∞
1 ωd
c12 (τ) ≈ ∫
2π − ∞
S f (ω)sin c( )e jωτ dω
c

4. Derive the reflection and transmission response of a single layer (shown in

fig 1.19) when it is illuminated from below. Derive the reflection response of
two layers illuminated from top.
5. In local scattering model show that the covariance matrix of the array output
becomes diagonal, that is, the wavefields become spatially uncorrelated when
the angular width subtended by the cloud of scatterers at the array is equal to
λ
. Consequently, it is not possible to estimate the DOA of the
d cosθ 0
transceivers.

References

1. P. M. Morse and H. Feshbach, Methods of Theoretical Physics, vol. I,

McGraw-Hill, New York, 1953.
2. J. Li and R. T. Compton, Angle and polarization estimation using ESPRIT
with polarization sensitive array, IEEE Trans., AP-39, pp. 1376-1383, 1991.
3. S. A. Levin, Geophysics, vol. 49, No.5, pp. 581-583, 1984.

© 2001 by CRC Press LLC

 4. A. Nehorai, Advanced-sensor signal processing in Highlights of Statistical
Signal and Array Processing, IEEE Signal Proc. Mag., pp.43-45, 1998
5. S. Treitel, P. R. Gutowski, and D. Wagner, Plane wave decomposition of
seismograms, Geophy., vol. 47, pp. 1375-1401, 1982.
6. P. M. Morse, Vibrations of elastic bodies; Wave propagation in elastic
solids, in Handbook of Physics, Eds. E. U. Condon and H. Odishaw, McGraw -
Hill Book Co., New York, 1958.
7. K. V. Mackenzie, Long range shallow water transmission, J. Acoust. Soc.
Am, vol. 33, pp. 1505-1514, 1961.
8. C. T. Tindle and G. E. J. Bold, Improved ray calculations in shallow water,
J. Acoust. Soc. Am, vol. 70, pp. 813-819, 1981.
9. P. S. Naidu and P. G. Krishna Mohan, Signal subspace approach in
localization of sound source in shallow water, Signal Proc., vol. 24, pp. 31-42,
1991.
10. Von F. Goos and H. Hanchen, Ein neuer und fundamentaler Versuch zur
Totalreflexion, Annales der Physik, vol. 1, 1947, pp. 333-346.
11. C. T. Tindle, Ray calculations with beam displacement, J. Acoust. Soc.
Am, vol. 73, pp. 1581-1586, 1983.
12. E. K. Westwood and C. T. Tindle, Shallow water time series simulation
using ray theory, J. Acoust. Soc. Am, vol. 81, pp.1752-1761, 1987.
13. E. A. Robinson, Iterative leastsquares procedure for ARMA spectral
estimation, in Nonlinear Methods of Spectral Analysis, Ed. S. Haykin,
Springer Verlag, pp. 127-154, 1979.
14. P. S. Naidu, Modern spectrum analysis of time series, CRC Press, Boca
Raton, Fl., USA, 1996.
15. A. H. Yaglom, Introduction to theory of stationary random functions,
Prentice-Hall , Englewood Cliffs, 1962.
16. M. Subbarayudu, Performance of the eigenvector (EV) method in the
presence of coloured noise, Ph. D. Thesis, Indian Institite of Science,
Bangalore, 1985.
17. B. F. Cron and C. H. Sherman, Spatial-correlation functions for various
noise models, J. Acoust. Soc. of Am, vol. 34, pp. 1732-1736, 1962.
18. W. A. Kuperman and F. Ingenito, Spatial correlation of surface generated
noise field in a stratified ocean, J. Acoust. Soc. Am, vol. 67, pp. 1988-1996,
1980.
19. R. M. Hampson, The theoretical response of vertical and horizontal line
arrays to wind generated noise in shallow water, J. Acoust Soc. of Am., vol.
78, pp. 1702-1712, 1985.
20. P. S. Naidu and P. G. Krishna Mohan, A study of the spectrum of an
acoustic field in shallow water due to noise sources at the surface, J. Acoust
Soc. of Am, vol. 85, pp. 716-725, 1989.
21. P. G. Krishna Mohan, Source localization by signal subspace approach and
ambient noise modelling in shallow water, Ph. D. Thesis, Indian Institute of
Science, Bangalore, 1988.

© 2001 by CRC Press LLC

22. S. M. Flatte, Wave propagation through random media: Contribution from
ocean acoustics, Proc. of the IEEE, vol. 71, pp.1267-1294, 1983.
23. H. W. Broek, Temporal and spatial fluctuations in single-path underwater
acoustic wave fronts, I-III, J. Acoust. Soc. of Am, vol. 72, pp. 1527-1543,
1982.
24. D. Asztely, B. Ottersten, and A. L. Swindlehurst, Generalized array
manifold model for wireless communication channels with local scattering, IEE
Proc. Radar Sonar, Navigation, vol. 145, pp.51-57, 1998.
25. D. Slepian, H. O. Pollack, and H. J. Landau, Prolate spheroidal wave
functions, Fourier analysis and uncertainty, Bell Systems Tech. J., vol. 40, pp.
43-84, 1961.
26. A. N. Tychonoff and A. A. Samarski, Differentialgleichungen der
Mathematischen Physik, VEB Verlag der Wissenschaften, Berlin, 1959.
27. S. J. Norton, Generation of separate density and compressibility images in
tissue, Utrasonic Imaging, vol. 5, pp 240-252 (1983).
28. A. C. Kak, Tomographic Imaging with diffracting and nondiffracting
sources, in Array Processing, Ed. S. Haykin, Springer-Verlag, pp. 351-433,
1985.
29. Ru-Shan Wu and M. N. Toksoz, Diffraction tomography and multisource
holography applied to seismic imaging, Geophy., vol. 52, pp. 11-25, 1987.
30. C. Pichot, L. Jofre, G. Peronnet, and J. C. Bolomey, Active microwave
imaging of inhomogeneous bodies, IEEE Trans., AP-32, pp. 416-424, 1985.
31 N. T. Gaarder, The design of point detector arrays, I, IEEE Trans., IT-13,
pp. 42-53, 1967.

© 2001 by CRC Press LLC

 Chapter Two

Sensor Array Systems

An array of sensors, distributed over a horizontal plane surface, is used to

receive a propagating wavefield with the following objectives:

1) To localize a source.
2) To receive a message from a distant source.
3) To image a medium through which the wavefield is
propagating.

In this chapter we shall study the basic structure of a sensor array system and in
the sequel learn how the above objectives are achieved. The most commonly
used array geometries are uniform linear array (ULA) and uniform circular array
(UCA). A uniform planar array (UPA) where sensors are placed on an
equispaced rectangular grid is more common in large military phased array
systems. A wavefront which propagates across the array of sensors is picked up
by all sensors. Thus, we have not one but many outputs which constitute an
array signal. In the simplest case, all components of the array signals are
simply delayed replicas of a basic signal waveform. In the worst case, individual
sensor outputs are strongly corrupted with noise and other interference, leaving
a very little resemblance among them. Array processing now involves
combining all sensor outputs in some optimal manner so that the coherent
signal emitted by the source is received and all other inputs are maximally
discarded. The aperture of an array, that is, the spatial extent of the sensor
distribution, is a limiting factor on resolution. However, the aperture can be
synthetically increased by moving a source or sensor. The synthetic aperture
concepts are extensively used in mapping radars and sonars. In this chapter we
concentrate on sensor array systems which will form the basic material for the
subsequent chapters.

§2.1 Uniform Linear Array (ULA):

2.1.1 Array Response: Consider a plane wavefront, having a temporal
waveform f(t) incident on a uniform linear array (ULA) of sensors (see fig. 2.1)
at an angle θ . In signal processing literature the angle of incidence is also
known as direction of arrival (DOA). Note that the DOA is always measured
with respect to the normal to array aperture, while another related quantity
azimuth, which was introduced in chapter 1, is measured with respect to the x-
axis, independent of array orientation. In this work θ stands for DOA and ϕ
stands for azimuth. We shall assume that a source emits a stationary stochastic
signal f(t). Let f m (t) , m=0, 1, 2,..., M-1 be the outputs of the sensors. The
signal arrives at successive sensors with an incremental delay. The output of the

© 2001 by CRC Press LLC

 Broadside

f(t)
θ

wave vector

ULA

7 2 1 0

Reference sensor

Figure 2.1: Uniform linear array of sensors. Note the convention of sensor
indexing. The left most sensor is the reference sensor with respect to which all time
delays are measured.

first sensor is f 0 (t) = f (t) , the output of the second sensor is f 1 (t) = f (t − ∆t)
and so on. Thus, the output of the mth sensor is f m (t) = f (t − m∆t). Some
times it is convenient to represent the sensor output in the frequency domain

∞ md
1 jω(t − sin θ)
f m (t) = ∫
2π − ∞
dF(ω)e c
(2.1)

where we have used the spectral representation of a stationary stochastic process

[1]. The simplest form of array signal processing is to sum all sensor outputs
without any delay.

∞
1 M −1 M −1 − jω md
1 sin θ
jωt 1
g(t) = ∑ m 2π ∫
M m=0
f (t) = dF(ω)e
M
∑= 0 e c

−∞ m
∞
1
= ∫
2π − ∞
dF(ω)H(ωτ)e jωt (2.2)

© 2001 by CRC Press LLC

 d
where H(ωτ) is the array response function, τ = sin θ, and d is sensor
c
spacing. The array response function for a ULA is given by

M
M −1 jω md sin θ sin( ωτ) j M −1 ωτ
1
H(ωτ) = ∑e
M m=0
c
= 2
ωτ
e 2 (2.3a)
M sin
2

When the sensor output is weighted with complex coefficients, am ,

m=0,1,...,M-1, the array response becomes

1 M −1
md
sin θ
∑
jω
H(ωτ) = am e c
(2.3b)
M m=0

A few samples of the frequency response function (magnitude only) are shown
in fig. 2.2 for different values of M, that is, array size. The response function is
periodic with a period 2 π . The maximum occurs at ωτ = 2nπ . The peak at
n=0 is known as the main lobe and other peaks at n = ±1, ±2,... are known as
grating lobes. Since the magnitude of the array response is plotted, the period
becomes π as seen in fig. 2.2. The grating lobes can be avoided if we restrict
the range of ωτ to ±π , that is, at a fixed frequency the direction of arrival
d 1 π
must satisfy the relation sin θ ≤ . For θ in the interval ± this
λ 2 2
d 1
requirement is satisfied if ≤ . If the range of θ is reduced it is possible to
λ 2
π π
increase the sensor spacing, for example, for − ≤θ+ the sensor spacing
4 4
d 1
need satisfies the constraint ≤ . The phase of the frequency response is a
λ 2
linear function of ωτ . This useful property of a ULA is lost when the sensors
are nonuniformly spaced (see p. 94).
The array response is a function of the product of frequency ω and
d
delay τ or, more explicitly, ω sin θ . The implication of this dependence is
λ
that two wavefronts whose waveform is a simple sinusoid but with different
frequencies (ω1 , ω 2 )arriving at different angles (θ1 , θ 2 ) will produce identical
array response if ω1 sin θ1 = ω 2 sin θ 2 . We shall discuss later such ambiguity
issues when we look into the broadband beamformation. The response function
has a main lobe which is surrounded by many sidelobes of decreasing magnitude
just as we find in spectral windows. The first zero is at

© 2001 by CRC Press LLC

 1

M=8
0.8

0.6

0.4

0.2

0
1

0.8 M=64
Response (magnitude)

0.6

0.4

0.2

0.8 M=128

0.6

0.4

0.2

0
-6 -4 -2 0 2 4 6
Radians
Figure 2.2: Array response function (magnitude) for different values of M. Notice
that the main lobe becomes sharper as the array size is increased.

© 2001 by CRC Press LLC

 λ
θ zero = sin −1 (2.4)
Md
which, for large M, becomes inversely proportional to the array length
expressed in terms of wavelength. The first sidelobe is 13.5 dB below the main
lobe. It is well known that both width of the main lobe and magnitude of the
sidelobes can be controlled by using a suitable weight function as in spectrum
analysis [1].
2.1.2 Array Steering: We have seen that the array response is maximum when
the direction of arrival (DOA) is on broad side ( θ =0). The maximum, however,
can be changed to any direction through a simple act of introducing a time delay
to each sensor output before summation. This is known as array steering. Let
an incremental delay of τ per channel be introduced. The sum output of the
array is now given by

1 M −1
g(t) = ∑ f m (t + mτ)
M m=0
∞ M −1 j ( τ − sin θ )ωm
d
1 jωt 1
= ∫ ∑
0
dF(ω)e e c (2.5)
2π − ∞ M m=0
∞
1 d
= ∫
2π − ∞
dF(ω)H((τ − sin θ 0 )ω)e jωt
c

d
where we have assumed that the DOA is θ 0 . Let τ = sin θ. Then the array
c
response is maximum whenever θ = θ 0 . We say that the array is steered in the
direction θ 0 , that is, in the direction of arrival of the incident wavefront. The
array response is now a function of DOA. This is demonstrated in fig. 2.3. It is
interesting to note that the width of the main lobe increases with increasing
DOA. To further understand this broadening effect we shall study the array
d
response function around its maximum, that is, at τ = sin θ 0 . The first zero
c
will occur at

M d
sin(θ 0 − ∆θ) = π
d
ω sin θ 0 − ω (2.6a)
2  c c 

Upon simplifying (2.6a) we get an equation

© 2001 by CRC Press LLC

 λ
sin θ 0 − sin(θ 0 − ∆θ) = (2.6b)
Md
whose solution is given by

 λ 
∆θ = θ 0 − sin −1 sin θ 0 − (2.6c)
 Md 

The dependence of ∆θ on the DOA for different array sizes is illustrated in fig.
2.4. The broadening of the main lobe is due to reduction in the array aperture
for a wavefront which is incident away from the broadside. The response is
maximum whenever

d
ω (sin θ 0 − sin θ) = 2πn
c
or
d
(sin θ 0 − sin θ) = n
λ
d 1 d 1
For ≤ , the acceptable solution is θ = θ 0 for which n=0. For > there
λ 2 λ 2
d
is more than one solution, one for each grating lobe. For example, let = 1; a
λ
solution of θ = sin −1 (sin θ 0 − n) exists only for n=0 and ±1.
d k
Now, let τ = . The array response function can be written as a
c M
2πd
discrete Fourier transform of a complex sinusoid, exp(− j sin θ 0 m) ,
λ

2 πd 2 πkm d
1 M −1 − j sin θ 0 m
∑e
j
H(k) = λ
e M λ
(2.7)
M m=0

Now H(k) is the kth discrete Fourier transform coefficient which should
k
correspond to the array response at a steering angle, sin −1 ( ) . The array
M
response is thus computed only at a set of discrete angles. Since M is finite,
π
usually a few tens, the angular range of ± is coarsely sampled. To overcome
2
this limitation it may be necessary to pad zeros to the complex sinusoid before
computing the discrete Fourier transform. However, it must be remembered that

© 2001 by CRC Press LLC

 1
0
0.8 30

0.6

0.4

0.2

0
1
0
60
0.8
Response (magnitude)

0.6

0.4

0.2

1
0
72
0.8

0.6

0.4

0.2

0
-1.5 -1 -0.5 0 0.5 1 1.5
DOA in radians

Figure 2.3: The effect of angle of arrival of a wavefront on the array response. The
mainlobe broadens and the sidelobes become asymmetric.

© 2001 by CRC Press LLC

 1.2

0.8
Width in radians

16

0.6

0.4
128

0.2
1024

0
0 0.5 1 1.5 2
DOA in radians

Figure 2.4 Width of the main lobe as a function of DOA for three different sizes of
the sensor array (M=16, 128,1024). The sensor spacing is assumed to be λ/2.

this step does not enhance the resolution but only improves the sampling of the
otherwise windowed Fourier transform. Use of discrete Fourier transform for
beamformation was first suggested in [2].
2.1.3 Broadband Source: Often a remote source such as broadband radar, engine
noise, or earthquake, etc. emits a broadband stochastic waveform. The simplest
approach to DOA estimation in such a situation is to compute the spectrum of
the sum of the sensor outputs. From (2.2) we have

∞
1 M −1 1
g(t) = ∑
M m=0
f m (t) = ∫
2π − ∞
dF(ω)H(ωτ)e jωt (2.8)

which may be considered as a spectral representation of the sum of the sensor

2
outputs. Hence its spectrum is given by Sg (ω) = S f (ω) H(ωτ) [1] where
S f (ω) is the spectrum of the waveform emitted by the source. We can
approximate S f (ω) by the spectrum of the output of any one sensor. Thus, we
obtain

© 2001 by CRC Press LLC

 Sg (ω)
H(ωτ) =
2
(2.9)
S f (ω)

2
Now consider a plot of H(ωτ) as a function of ω . There is always one peak
at ω =0 and a stream of peaks caused by the incident wavefront [3], at positions
given by the equation below

c
ω peak = 2πn (2.10a)
d sin θ
where n=0, ± 1, ± 2,... We introduce a quantity called minimum array
c
sampling frequency ω min = 2π . An array of sensors may be considered as a
d
waveform sampler which samples the waveform as it propagates across the
d
array. The sampling interval is sin θ and the maximum interval or minimum
c
π
sampling frequency occurs when θ = . In terms of the minimum array
2
sampling frequency the peak may be written as

ω min n
ω peak = (2.10b)
sin θ 0

Evidently, ω peak must be in the range ω min n ≤ ω peak ≤ ∞ . For the sake of
illustration, let the signal spectrum be of infinite width. Now, a plot of
2 ω
H(ωτ) will show an infinite set of peaks spaced at an interval min . For
sin θ
2
example, for θ = 450 an idealised plot of H(ωτ) is shown in fig. 2.5a. A
numerical example is shown in fig. 2.5b where we have assumed a 16 sensor
ULA with spacing d=15 meters. A broadband signal with bandwidth = ( ±200
Hz) is incident at DOA angle equal to 45o. The average spacing of peaks is
42.0 Hz against the theoretical value of 42.43 Hz.
Angular Spectrum: Recall that the frequency wavenumber spectrum of a plane
wave is a line passing through the origin with a slope inversely proportional to
the direction cosines of the wave vector, in particular, on pages 15-17, chapter 1
we have

c c
ω= u, ω = v (2.11)
α β

© 2001 by CRC Press LLC

 (a)
2
H(ω ν)

0 ω min
ω

400

350 (b)

300

250
Sg (ω)
200
S f (ω)
150

100

50

0
0 0.1 0.2 0.3 0.4 0.5
frequency(Hz) x512

Figure 2.5: (a) A plot of H(ων)

2
for θ 0 = 41.810 . ω min is the minimum array
sampling frequency. The position of the first peak (excluding the one at ω = 0 ) or
the interval between the peaks can be used for estimation of DOA. (b) The ratio of
the spectrum of the sum of all sensor outputs divided by the spectrum of the first
2
sensor output. There are four peaks of H(ων) within the available bandwidth. A
16 sensor ULA (d=15 m) is illuminated by a broadband signal ( ±200 Hz) incident at
45o DOA. Wave speed is 1500m/s.

© 2001 by CRC Press LLC

 x 10 4
2.5

1.5
Power

0.5

0
10 20 30 40 50 60 70 80
Angle in deg

Figure 2.6: Angular spectrum obtained by averaging along the radial lines. Two
broadband sources (solid line: 80-120 Hz) and two narrowband sources (dashed line:
98-102 Hz). 16 sensor array, 128 samples with sampling interval 0.005 sec.
snr=0dB.

where α and β are direction cosines. Conversely, the spectrum on a line

joining the origin with a point in the frequency wavenumber space corresponds
to the power of a plane wave propagating with the direction cosines as in
(2.11). Consider the frequency wavenumber spectrum of the output of a ULA
and integrate the power along a series of radial lines. The integrated power thus
obtained is plotted as a function of DOA θ . Such a plot will reveal the
presence of plane waves incident on the ULA. The background incoherent noise
will tend to average out giving constant or slowly varying power. As an
example, we consider a ULA of 16 sensors along x-axis and two incoherent
broadband (80 to 120 Hz) sources at DOAs 40o and 50o. The angular spectrum
obtained by averaging over different radial lines clearly resolves two peaks but
fails to resolve the narrowband signals (see fig. 2.6). Based on the above
property of the frequency wavenumber spectrum a method of estimating the
direction of arrival by projecting the spectrum along a radial line onto ω = π
line has been proposed in [4]. It was mentioned earlier in chapter 1 that the
frequency wavenumber spectrum may be considered as a directional spectrum.

© 2001 by CRC Press LLC

 Indeed, it is possible to estimate the direction of arrivals by integrating over the
temporal frequency band

S̃ f (u,v) = ∫ S (u,v,ω)dω
f
freq band
(2.12)

Such a method was in fact proposed in [3] where the integrated wavenumber
spectrum S̃ f (u, v) was called zero-delay wavenumber spectrum.
Slant Stacking: Closely related to the idea of array steering (or beamforming) is
slant stacking used extensively in seismic exploration. Stacking is also related
to the radon transform [5]. Consider a wavefield, f(t,x), where we shall replace t
1 u
by τ + px x where px = = . The stacking operation is defined as
cx ω

f̂ (τ, px ) = ∫ f (τ + p x, x)dx
−∞
x (2.13)

Let u = px ω for a fixed ω

+∞ ∞
1
∫ ∫ ωF(ω, pω)e dωdp ∫ e x
j ( p ωx − pωx )
f̂ (τ, px ) = jωτ
dx
4π 2 −∞ −∞
+∞
1
= ∫
2π − ∞
F(ω, px ω)e jωτ dω (2.14a)

Taking the inverse Fourier transform we obtain from (2.14a)

F̂(ω, px ) = F(ω, px ω) (2.14b)

Thus, 1D Fourier transform of the stacked output is equal to 2D Fourier

transform of the wavefield. Beam steering assumes a plane wave model but
stacking does not require this assumption. As shown in (2.14) the stacked
output is directly proportional to the spatial Fourier transform of the wavefield,
which is equal to array response function (2.7). When the incident wavefield is
nonplanar the right thing to do is plane wave decomposition which is achieved
through stacking (see (1.33)). Such a situation arises in seismic exploration
where a low frequency (wavelengths on the order of a few hundred meters are
common) source is used to energize rock strata. The wavefield observed on the
surface is a function of two spatial coordinates and time. We shall first derive a
result similar to that in (2.14) but for a two-dimensional wavefield.

© 2001 by CRC Press LLC

The stacked output of a planar array is defined as

+∞

f̂ (τ, px , py ) = ∫ ∫ f (τ + px x + py y, x, y)dxdy (2.15)

−∞

In (2.15) we shall replace the integrand by its Fourier representation,

+∞
1 j (ωt − p x ωx − p y ωy)
3 ∫∫
f (t, x, y) = ω 2 F(ω, px ω, pyω)e dωdpx dpy
8π − ∞

where we have used the relations u = px ω and v = py ω and obtain

f̂ ( τ , px , py ) =
+∞ +∞
1 j ( ωτ +( p x′ − p x ) ωx +( p y′ − p y ) ωy)
∫∫ ∫∫ω F(ω , px ω , py ω )e dωdpx′ dpy′
2
dxdy 3
8π
−∞ −∞
∞
1
∫ F(ω , p′ ω , p′ ω )e
jωτ
= x y dω
2π
−∞

Hence,
F̂(ω, px , py ) = F(ω, px ω, pyω)
(2.16a)

When there is axial symmetry, as it happens in horizontally stratified rocks,

(2.16a) takes a different form [6]

F̂(ω, p) = ∫ rdrJ0 (ωpr)F(ω,r) (2.16b)

0

2.1.4 Matrix Formulation: When the incident signal is a narrowband signal the
output of an array, in particular a ULA, may be conveniently represented in a
matrix format which reveals some interesting properties. This is also true of a
broadband signal but the processing has to be in the frequency domain.
Representation of Narrowband Signals: A narrowband signal f nb (t) may be
represented as

f nb (t) = s0 (t)cos(ω c t + ϕ 0 (t)) (2.17a)

© 2001 by CRC Press LLC

 where s0 (t) is a slowly varying waveform, often called envelope and ϕ 0 (t) is
also a slowly varying phase. cos(ω c t) is a rapidly varying sinusoid, often
known as a carrier and ω c is known as carrier frequency. Many active array
systems radiate narrowband signals, for example, a phased array radar. Equation
(2.17a) may be expressed as

f nb (t) = f i (t)cos(ω c t) − f q (t)sin(ω c t)

where

f i (t) = s0 (t)cos(ϕ 0 (t))

f q (t) = s0 (t)sin(ϕ 0 (t))

f i (t) is known as an inphase component and f q (t) is a quadrature component.

The inphase and quadrature components are uncorrelated. They have the same
spectral density function. The inphase and quadrature can be uniquely recovered
from a narrowband signal by a process known as mixing which involves
multiplication with 2 cos(ω c t) and −2 sin(ω c t) and low pass filtering [7]. A
complex analytical signal is defined as f c (t) = f i (t) + jf q (t) . Consider a
narrowband signal delayed by one quarter period. Assuming that both inphase
and quadrature components are slowly varying signals we get the following
approximate result:

τ0 τ τ
f nb ((t − )) = f i (t)cos(ω c (t − 0 )) − f q (t)sin(ω c (t − 0 ))
4 4 4
≈ f i (t)sin(ω c t) + f q (t)cos(ω c t)
= f nbHilb (t)

We define a complex analytical signal as

τ0
f nb (t) + jf nb (t − ) = f i (t)e jω c t + jf q (t)e jω c t
4 (2.17b)
= f c (t)e jω c t

The representation given by (2.17b) is useful in narrowband beamformation.

The process described in (2.17b) is often referred to as quadrature filtering,
which is illustrated in fig. 2.7. Note that the input to quadrature filter is real but
the output is complex.

© 2001 by CRC Press LLC

 +
real complex
input π/2 × output
j

Figure 2.7. Quadrature filter structure. Since the phase change due to propagation
appears in the complex sinusoid at the output it is easy to introduce phase
adjustments for beamformation.

Consider the mth sensor of a ULA. The complex output of the

quadrature filter is

d d τ
f m (t) = f nb (t − m ) + jf nb (t − m − 0 )
cx cx 4
(2.17c)
d
jω c t − jmω c
= f c (t)e cx

The term representing propagation delay now appears in the complex sinusoid.
Naturally, in order to form a beam, it is convenient to introduce phase
adjustments. Let w0 , w1 ,..., wM −1 be a set of complex weight coefficients for
beamformation. The beam output will be given by

M−1 d

∑w e
jω ct+ jmω c
cx
output= f c (t) m .
m=0

Through complex weight coefficients it is possible to adjust both amplitude and

phase so that the resulting response is closest to any desired response.
Matrix: A snapshot is a vector representing the outputs of all sensors taken at
the same time instant t. Let f (t) = col { f 0 (t), f 1 (t),..., f M−1 (t)} be a snapshot,
where f 0 (t), f 1 (t),..., f M−1 (t) stand for the sensor outputs at time t. When the
incident signal is narrowband, the signal varies slowly with time (assume that
the carrier has been removed). In the noise-free case a single time shot is
adequate as it contains all available information. A snapshot vector for
narrowband signal may be expressed using (2.17c) as

© 2001 by CRC Press LLC

  − jω c c x 
d d
cx 
− j ( M −1)ω c
f(t) = f c (t)col 1,e ,...,e 
  (2.17d)

= f c (t)φ(θ 0 )

c
where, it may be recalled that, the apparent speed c x = . Further let the
sin θ 0
sensor response matrix be α(θ 0 ) = diag{α 0 (θ 0 ), α1 (θ 0 ),...α M−1 (θ 0 )}, in which
each element represents the response of a sensor as a function of the angle of
incidence of the wavefront. φ(θ 0 ) represents the propagation effect of the
medium on a wavefront propagating across the array. φ(θ 0 ) and α(θ 0 ) together
form a direction vector a(θ 0 ) = α(θ 0 )φ(θ 0 ) representing the response of an
array to a wavefront incident at angle θ 0 (DOA). Finally, the array output may
be expressed as follows:

f(t) = f c (t)α(θ 0 )φ(θ 0 )

(2.17e)
= f c (t)a(θ 0 )

When there are P narrowband sources radiating simultaneously the array output
may be expressed as a linear combination of P terms of the type shown in
(2.17e)

 f c0 (t) 
 
 f c1 (t) 
. 
α(θ 0 )φ(θ 0 ),α(θ1 )φ(θ1 ),...,α(θ P −1 )φ(θ P −1 )  
f(t) =   .  + η(t)
 (M × P)  
.
 
 f c P−1 (t)
 
(P × 1) 
(2.18a)

where η(t) is the noise vector assumed to be uncorrelated with the signal
terms. Equationq (2.18a) may be written in a more compact form where P
columns

f(t) = As + η(t) (2.18b)

© 2001 by CRC Press LLC

of A matrix are P direction vectors pointing to P sources. The matrix
representation of the array output model as in (2.18b) plays a very crucial role
in the development of high resolution methods for DOA estimation.
The array steering can also be represented in terms of a matrix operation. To
steer an array to a desired direction, θ , we form an inner product of the steering
vector and the array snapshot

a H (θ)f(t) = f c (t)a H (θ)a(θ 0 ) (2.19a)

The output power is given by

{
a H (θ)C f a(θ) = σ 2s0 E a H (θ)a(θ 0 )
2
}
2 (2.19b)
d
= σ M H(ω (sin θ − sin θ 0 ))
2
s0
2

{ }
where C f = E f (t)f H (t) is the spatial covariance matrix (SCM). Whenever
θ = θ 0 , that is, when the steering angle is equal to the DOA, the left hand side
of (2.19b) equals σ 2s0 M 2 giving the power of the source.
The M dimensional steering vector will span an M-dimensional space
known as an array manifold. The tip of the steering vector traces a closed curve
in the array manifold or a closed surface when the steering vector is a function
of two variables, for example, azimuth and elevation. Consider the case of
identical sensors, that is,

α 0 (θ 0 ) = α1 (θ 0 ) =...= α M −1 (θ 0 ) = α(θ 0 )

In this case the direction vector is given by

d 2d (M-1)d
- jω sin θ 0 - jω sin θ 0 - jω sin θ 0
a(θ 0 ) = α(θ 0 )col{1,e c
,e c
...e c
}

In the event of sensors being omnidirectional, that is, α(θ 0 ) = constant, the
array manifold becomes a closed curve on a sphere (in M-dimensional space).
For uniqueness the array manifold must not intersect; otherwise, at the point of
intersection, the steering vector will point to two different directions, θ1 and
d
θ 2 , such that a(θ1 ) = a(θ 2 ). Such a possibility exists only when > 0.5. To
λ
show this, consider the steering vector for omnidirectional sensors. Let θ1 and
θ 2 be two such directions for which a(θ1 ) = a(θ 2 ), that is, for all m

© 2001 by CRC Press LLC

 d d
j2π msin θ j2π msin θ
e λ 1
=e λ 2

This is possible when

d
[sin θ1 − sin θ 2 ] = 1
λ
or
λ
sin θ1 = + sin θ 2 (2.20)
d
λ λ
A solution of (2.20) exists only when < 2 ; for example, when = 1. 2 the
d d
following pairs of directions are the possible solutions: (36.87o , -36.87o ),
(23.58o,-53.13o), and (11.54o,-90o).
The steering vector satisfies the following properties:

(a) a(θ) = a(π − θ)

∗
(b) a (θ) = a(−θ)
π λ
(c) a(θ) is periodic with a period ± only if d = .
2 2

Property (a) implies a wavefront coming from the north and another
symmetrically opposite from the south (a and b in fig 2.8) cannot be
distinguished (north-south ambiguity). Property (b) implies a wavefront coming
from the east and another symmetrically opposite from the west (a and c in fig.
2.8) can be distinguished only if the signal is complex (east-west ambiguity).
To show this recall (2.17a) and compare the outputs of a ULA for a real input
signal incident at angle θ and −θ. Let f θ (t) be output of a ULA for an
incident angle, θ , and f −θ (t) be the output for an incident angle, −θ. For a real
signal f θ (t) = f ∗−θ (t) but for a complex signal f θ (t) ≠ f ∗−θ (t). Property (c)
π
implies that there is no grating lobe in the range ± when the sensor spacing
2
λ
is d ≤ .
2
The steering vector is closely related to the array response function. To
show this we define a unit vector, 1=col{1, 1, 1,...,1}, and consider a dot
product

© 2001 by CRC Press LLC

 N

c a

π −θ

−θ θ
x
E

b
.

Figure 2.8: A ULA cannot distinguish wavefronts a and b (north-south ambiguity).

However, it can distinguish wavefronts a and c if the signal is complex (east-west
ambiguity).

M −1 jωm d sin(θ)
a(θ) 1 =
H
∑e
m=0
c
(2.21a)

which follows from (2.3). We have assumed for the sake of simplicity that all
sensors are identical and omnidirectional. In real array, the individual sensor
response is likely to be directional and varying. Then, the array response is
given by

M −1 d

∑ α m (θ)e
jωm sin(θ)
a(θ) H 1 = c
(2.21b)
m=0

An array is steered to a desired direction by introducing delays to each sensor

output. The response of such a steered array is given by (from (2.5))

© 2001 by CRC Press LLC

 1 M −1 j ( τ − c sin θ 0 )ωm
d
d
H(τ − ω sin θ 0 ) =
c
∑e
M m=0
(2.22)
1
= a H (θ)a(θ 0 )
M
where θ is the desired direction to which the array is steered and θ 0 is the
direction of arrival of a wavefront. The response of the steered array is expressed
as an inner product of the steering vector and direction vector as shown in
(2.22).
2.1.5 Nonuniform Linear Arrays : There are reasons for having to consider non
uniform linear arrays. These are: (a) Redundant sensors are removed and
employed to increase the array aperture, (b) Certain sensors in a long ULA may
fail as a result of factors beyond our control, and (c) Array spacing is
intentionally made nonuniform in order to derive certain benefits, for example,
there is no aliasing effect if the periodicity of a ULA is destroyed. We shall
show how by means of nonuniform array the above objectives may be achieved.
Redundant sensors: Let us first consider the case of redundant sensors. We shall
rewrite (2.22) as

d 1
H((τ − sin ϕ 0 )ω) = b H φ(r) (2.23)
c M

d
where b = col{α 0 (θ 0 )α ∗0 (θ), α1 (θ 0 )α1∗ (θ),...α P−1 (θ 0 )α ∗P−1 (θ)} and r=ω
c
× (sin θ − sin θ 0 ) . The power output of the array is simply proportional to
the square of the transfer function

1 H
power ∝ H = b φ(r)φ H (r)b
2
2
(2.24)
M

Let us expand matrix φ(r)φ H (r)

1 e jr e j 2r e j 3r ... e j ( p−1)r 
 − jr 
e 1 e jr e j 2r ... e j ( p−2)r 
H = φ(r)φ H (r) = e − j 2r e − jr 1 e jr ... e j ( p−3)r  (2.25)
 
 ... 
 − j ( p−1)r − j ( p−2)r 
e e ... 1 

© 2001 by CRC Press LLC

 3 λ Array

Sensor
No sensor

Figure 2.9: A four sensor array spread over a 3 λ aperture array will produce all
entries of the matrix on (2.25).

Array length in Sensor locations for

λ minimum redundancy
units of
2
6 0, 2, 5,6

15 0, 1, 2, 3, 7, 11, 15
0, 1, 3, 6, 10, 14, 15
0, 1, 4, 8, 13, 14, 15
0, 2, 4, 5, 8, 14, 15

31 0, 1, 2, 3, 4, 5, 12, 18, 25, 31

0, 1, 2, 3, 6, 11, 18, 25, 27, 31
0, 1, 2, 3, 8, 14, 18, 22, 27, 31
0, 1, 2, 3, 12, 18, 22, 26, 29, 31

Table 2.1: Sensor locations for minimum redundancy. For array length of six there is
only one arrangement but for array lengths 15 and 31 there are 77 and 888
arrangements, respectively.

It may be observed that H is a toeplitz matrix, that is, along any diagonal the
entries are repeated even though they refer to different sensors in the array. For
example, consider the second upper diagonal where the entries refer to a pair of
sensors whose indices are m, and m-2, where m=2,3,...,M-2; explicitly, the
pairs of sensors involved in creating the terms on this diagonal are (2,0), (3,1),
(4,2), etc. Thus, insofar as the second diagonal is concerned there are several
redundant pairs of sensors. This redundancy can be removed by selectively
removing sensors [8]. For example, consider the seven sensors uniform array
shown in fig. 2.9. All entries in the H matrix (2.25) can be obtained from just
four sensors shown by filled circles; for example, the first diagonal may be
obtained from sensors at position 5 and 6, the second diagonal from sensors at

© 2001 by CRC Press LLC

 position 0 and 2, the third diagonal from sensors 2 and 5, the fourth diagonal
from sensors 2 and 6, the fifth diagonal from sensors 0 and 5 and finally the last
diagonal from sensors 0 and 6. Thus we are able to generate a complete 7x7
matrix just from a four sensor nonuniformly spaced array by removing all
redundant sensors. In general, it is possible to arrive at a distribution of the
minimum number of sensors required to fill all entries in the matrix in (2.25)
[9]. For a given array length the minimum number of sensors required to
generate the H matrix is found through an exhaustive search. As the array size
increases there exists more than one distribution of a zero redundancy array, that
is, an array having sensors just enough to generate all diagonals of the H
λ
matrix. For example, for an array length of 15 (16 sensors spaced at ) the
2
minimum redundant array has seven sensors and there are 77 different
λ
distributions. For array length of 31 (32 sensors spaced at ) the minimum
2
redundant array has ten sensors and there are 888 different distributions. A few
sample distributions are shown in table 2.1
The theoretical response of a minimum redundancy array is identical with that
of a normal array (nonredundant array) but in the presence of background noise
the response of a redundant array is considerably inferior. To demonstrate this
we have considered an array of length 15 but having only seven sensors in the
redundant array. Broadband array transfer was computed with background noise
variance equal to one. First, the H matrix was computed. To simulate a
uniform linear array, a complex noise of unit variance was added to each
element of the matrix. But to simulate a minimum redundancy array, a complex
noise of unit variance was added to each diagonal (same noise element to all
elements in a given diagonal). The transfer functions are shown in fig. 2.10.
The transfer functions show the main lobe and the first grating lobe at correct
frequency in both cases; however, the noisy sidelobes are about ten times
greater in the case of minimum redundancy array. The upper and lower bounds
on the snr (signal to noise ratio) gain were theoretically derived in [10].
Missing Sensors: In a long chain of sensors such as in a towed array there may
be a few sensors which have either failed or are malfunctioning. Such sensors
are often skipped, leading to loss of fidelity in the array steering response. Here
we shall analyse the effect of a missing sensor on the array performance. Let p
be the probability that a sensor will be malfunctioning and it is independent of
all other sensors. Let xl , l = 0,1, 2,... N − 1be locations of live sensors. Note
that xl is a multiple of sensor spacing. Let ∆xl = xl+1 − xl be the spacing
between the l + 1 st and l th sensors. It may be expressed as k ⋅ d , where d is a
nominal sensor spacing and k -1 is a Bernoulli random variable so that

prob{∆xl = k ⋅ d} = (1 − p) p k −1 (2.26)

© 2001 by CRC Press LLC

 10

(a)

1
Transfer function (magnitude)

.1

.01

.001
0 32 64 96 128 160 192 224 256
Frequency in Hz
10

(b)

1
Transfer function (magnitude)

.1

.01

.001
0 32 64 96 128 160 192 224 256
Frequency in Hz

Figure 2.10: (a) Transfer function of a 16 sensor ULA and (b) Transfer function of a 7
sensor minimum redundancy array of same length as ULA in (a). Sensor spacing=15
meters, wave speed=1500 m/s and θ = 450 .

© 2001 by CRC Press LLC

 The output of the nth sensor may be expressed as

n
ω
+∞ j (ωt − sin θ 0 ∑ ∆x l )
1
∫
c
f n (t) = dF(ω)e l=0

2π − ∞
(2.27)
+∞ n ω
1 −j sin θ 0 ∆x l
= ∫ dF(ω)e j (ωt ) ∏ e c
2π − ∞ l=0

where θ 0 is the direction of arrival of a wavefront. We further assume that

x 0 = 0 . We consider a delay and sum type of beam steering. Each sensor output
ω
is delayed by an amount equal to τ n = x n sin θ where θ is the angle to
c
which the array is steered

1 N −1
g(t) = ∑ f n (t − τ n )
N n=0
+∞ ω
(2.28)
N −1 n − j (sin θ 0 −sin θ)∆x l
1 j (ωt ) 1
= ∫
2π − ∞
dF(ω)e ∑ ∏
N n=0 l=0
e c

We shall now evaluate the expected value of the delay and sum processor given
by (2.28)

+∞
1 j (ωt ) 1
N −1 n
 − j ωc (sin θ 0 −sin θ)∆x l 
E{g(t)} =
2π −∫∞
dF(ω)e ∑ ∏ E e
N n=0 l=0 
 (2.29)


We need to evaluate the expectation of the expression inside curly brackets in

(2.29). Using the distribution of the Bernoulli random variable (2.26) in (2.29)
we obtain

 − j ωc (sin θ 0 −sin θ)∆x l  ∞ ω

− j (sin θ 0 −sin θ)k d
 = ∑ (1 − p) p e c
k −1
E e
  k =1
ω ω
−j d (sin θ 0 −sin θ) ∞ −j (sin θ 0 −sin θ)( k −1) d
= (1 − p)e c
∑p
k =1
k −1
e c

ω
−j d (sin θ 0 −sin θ)
(1 − p)e c
= ω (2.30)
−j d (sin θ 0 −sin θ)
1 − pe c

© 2001 by CRC Press LLC

Using the above result in (2.29) we obtain

ω n
+∞ 
N −1
− j d (sin θ 0 −sin θ) 
 (1 − p)e
c
1 1 
E{g(t)} = ∫
2π − ∞
dF(ω)e j (ωt ) ∑
N n=0  − j d (sin θ 0 −sin θ) 
ω

 1 − pe c 
+∞
j (ωt ) 1 1 − Q
N
1
2π −∫∞
= dF(ω)e (2.31)
N 1− Q

where Q stands for the quantity inside the square brackets of (2.31). Note that
ω
−j d(sin θ 0 −sin θ)
for p=0, Q = e c and (2.31) reduces to a known expression for the
response function of a ULA. The response of an array with missing sensors,
given by

1 1 − QN
H(ωτ) =
N 1− Q

d
where τ = (sin θ 0 − sin θ) , is shown in fig. 2.11 for two different values of
c
probability of malfunctioning. It is assumed that the total number of live
sensors is the same in both cases, namely, 16. Notice that the magnitude
response has fewer sidelobes but the phase characteristics appear to be grossly
different from those of a ULA, which is a linear function of ωτ (see (2.3)). A
nonlinear phase response results in a distortion of the received waveform
particularly when it is broadband.
Random Array: A third type of nonuniform array is one where the sensors are
spaced at random intervals, in particular, an exponential distribution for which a
closed form solution can be derived. Let x n , n = 0,1,... N − 1 be the locations of
the sensors; all of which, however, lie on a straight line. Let us assume an
exponential distribution for the sensor spacing

pdf (∆xn ) = βe −β∆x ∆x ≥ 0

(2.32)
=0 ∆x < 0

where pdf stands for probability density function and β is a parameter in the
exponential distribution. The output of the nth sensor may be written as in
(2.27). The delay and sum type of processing would result in the following
array output (from 2.29)

© 2001 by CRC Press LLC

 1.0

0.8 (a)

p=0.2

0.6
Magnitude

0.4

p=0.4

0.2

0.0
0.0

-0.5
p=0.2

-1.0 p=0.4 (b)

Phase

-1.5

-2.0

-2.5
0.0 0.8 1.6 2.4 3.2
ωτ

Figure 2.11: (a) Magnitude response and (b) Phase response of an array with
malfunctioning sensors. The total number of live sensors is assumed to be the same
in both cases, namely, 16.

© 2001 by CRC Press LLC

 E{g(t)} =
+∞ N −1 n ∞ ω
1 − j (sin θ 0 −sin θ)∆x l
j (ωt ) 1
∫
2π − ∞
dF(ω)e ∑ ∏ ∫
N n=0 l=0 0
pdf (∆xl )e c d∆xl (2.33a)

Using the exponential distribution function (2.32) in (2.33a) we obtain

N
 
 β 
1−  
ω
+∞  j (sin ϕ 0 − sin ϕ) + β 
E{g(t)} =
1 1  c 
2π −∫∞
dF(ω)e j (ωt )
N 1− β
ω
j (sin ϕ 0 − sin ϕ) + β
c
(2.33b)

The array response may be expressed in terms of the product of the wavelength
and parameter β,

N
 βλ 
1− 
1  j2π(sin θ 0 − sin θ) + βλ 
H(ν,βλ) = (2.33c)
N 1− βλ
j2π(sin θ 0 − sin θ) + βλ

where ν = 2π(sin θ 0 − sin θ). We have plotted the array response function for
different values of βλ in fig. 2.12. While the magnitude response is free from
sidelobes, the phase response is highly nonlinear in the range, ν=0.0 to 1.6,
where the magnitude response is significant. This behavior was also noticed in
the case of an array with missing sensors.
2.1.6 Flexible Array: We have so far considered a sensor array which is rigidly
fixed to the ground or to a platform. We now consider an array where the
sensors are held in position by means of a flexible rope which allows a sensor
to move over a circular arc of fixed radius. The sensor spacing, however,
remains unchanged. An important consequence of this freedom allowed to the
array is to alter the shape of the array when it is being towed or it is under the
influence of ocean currents. A commonly used array model is that M sensors are
separated by straight line segments of fixed length d [11] (see fig 2.13a). Let
(x m , ym , zm ) represent coordinates of the mth sensor with respect to m+1st
sensor.

© 2001 by CRC Press LLC

 1.0

0.8 (a)

βλ=2
0.6
Magnitude

βλ=1
0.4

0.2

0.0
0.0 1.6 3.2 4.8 6.4
v

0.0
βλ=1

-0.5 (b)

βλ=2
-1.0
Phase

-1.5

-2.0
0.0 1.6 3.2 4.8 6.4
v

Figure 2.12: Frequency response of a random linear array. (a) magnitude response.
(b) phase response. 16 sensors spaced at random intervals having an exponential
distribution.

Since the distance between the sensors remains fixed (x m , ym , zm ) must satsify a

© 2001 by CRC Press LLC

 z

β2 d
y
α2

x
(a)

0.5
0.4

0.3
y-axis

0.2

0.1

0.0
0.0 2.0 4.0 6.0
x-axis

(b)
Figure 2.13: (a) Model of a flexible array where the adjacent sensors are held at a
fixed distance but the azimuth and elevation of the line segment joining the sensors
are random variables. (b) A bow shaped 16 sensor array in the x-y plane. The
distance between the adjacent sensors is 0.5 λ . All dimensions are in units of
wavelength .

relation x m2 + ym2 + zm2 = d , alternatively,

xm = d sinβ m sin α m
ym = d sinβ m cosα m (2.34a)
zm = d cosβ m

where (α m ,β m ) are azimuth and elevation of a line segment joining mth sensor
with m+1st sensor (see fig 2.13a). It is assumed that x 0 = y0 = z0 = 0 .

© 2001 by CRC Press LLC

 Let a plane wavefront be incident on the array shown in fig. 2.13. The
output of mth sensor is given by

 m ω 
ω c t − ∑ xi c sin θ 0 sin ϕ 0
 i=0 c 
f m (t) = s0 (t) cos  + ηm (t) (2.34b)
 − ∑ yi ω c sin θ 0 cos ϕ 0 − ∑ zi ω c sin θ 0 
m m

 i=0 c i=0 c 

where (ϕ 0 , θ 0 ) are respectively azimuth and elevation of the incident wavefront.

Using (2.34a) in (2.34b) we get

  m∑ sin β i cos α i sin θ 0 sin ϕ 0 

 i = 0m 
f m (t )= s 0 (t ) cos ω c t −ω c  + ∑ sin β i sin α i sin θ 0 cos ϕ 0  + η m (t )
d
c
  i =m0 
  + ∑ cos β i sin θ 0
 i=0 

Transforming into a complex analytical signal (2.17d) the array output may be
expressed in a matrix form

 − jω c dc [γ 1 sin θ 0 sin ϕ 0 + ε 1 sin θ 0 cos ϕ 0 + ξ1 sin θ 0 ] 

1,e ,...,
f(t) = f c (t)col  d  + η(t)
e − jω c c [ γ M −1 sin θ 0 sin ϕ 0 + ε M −1 sin θ 0 cos ϕ 0 + ξ M −1 sin θ 0 ] 
 
m m m
where γ m = ∑ sin β i cos α i , ε m = ∑ sin β i sin α i and ξ m = ∑ cosβ i . These
i=0 i=0 i=0
parameters may be expressed in a recursive form

γ m = γ m −1 + sinβ m cosα m
ε m = ε m −1 + sinβ m sin α m
ξ m = ξ m −1 + cosβ m

When the array is perfectly linear, γ 0 = γ 1 =, ..., γ M−1 = sin β 0 cos α 0 ,

ε 0 = ε1 =, ..., ε M−1 = sin β 0 sin α 0 and ξ 0 = 1 − γ 20 − ε 20 .
We like to demonstrate the effect of array deformation on its response
function. Consider a 16 sensor ULA which is deformed into a bow shaped curve
in an x-y plane as shown in fig. 2.12b. The maximum displacement along the
y-axis is half wavelength. The array responses are shown in fig. 2.14.

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

0.4

0.2

0
0 20 40 60 80 100
Angle in deg.

Figure 2.14: Response of a bow shaped array (solid curve) shown in fig. 2.13b and
the dashed curve is for an undeformed array.

Notice the extremely broad main lobe which is fortunately located at the right
position. If we further increase the deformation, for example, the bow height is
increased to 2.3λ the array response is found to be totally distorted. Even the
main lobe is found to be wrongly placed.

§2.2 Planar Array:

A planar array has its sensors distributed over a plane. When a wavefront has
two unknown parameters, azimuth and elevation, we need a planar array for
estimation of a pair of parameters. Since a plane has two dimensions, there are
many possible array geometries; some of these are illustrated in fig. 2.15. The
natural extension of a ULA in two dimensions is a square or rectangular array
where the sensors are placed on a square or rectangular grid. Other geometries
are essentially sparse versions of the square or the rectangular array. We shall
first study the rectangular array and then look into the sparse arrays, in
particular, the circular array which has found many applications.
2.2.1 Uniform Planar Array (UPA): Sensors are placed on a rectangular grid
where the nodes are spaced d1 along x-axis and d2 along y-axis (when d1=d2=d
we get a square grid). Let a plane wavefront be incident at azimuth angle ϕ and
elevation θ (see fig. 2.16). A plane wavefront, given by

© 2001 by CRC Press LLC

 Uniform Square Array Cross Array

Circular Array
Triangular Array

Figure 2.15: Planar array geometries are shown above. The square or rectangular
array is a natural extension of ULA and other geometries are sparse versions of
square or rectangular array.

+∞
1
f (t, x, y) = ∫
2π − ∞
F(ω)e j (ωt −ux − vy) dω (2.35)

ω ω
where u = sin θ cos ϕ and v = sin θ sin ϕ , is incident on a UPA. The
c c
output of the (m,n)th sensor is

© 2001 by CRC Press LLC

 z

θ
y

ϕ
x

d2
d1

Figure 2.16: A plane wavefront incident on a UPA at azimuth ϕ and elevation θ .

+∞
1
f m1 m2 (t) = ∫
2π − ∞
F(ω)e j (ωt −um1 d1 − vm2 d 2 ) dω (2.36)

All sensor outputs are summed in phase, yielding

M1 M2

∑ ∑a
1
g(t) = m1 m 2 f m1 m2 (t)
M1 M2 m1 =1m 2 =1
+∞ M1 M 2

∑ ∑a
1 1
∫ F(ω )e
jωt − j (um1 d1 + vm 2 d 2 )
= m1 m 2 e dω
2π M1 M2 m1 =1m 2 =1
−∞
(2.37)
+∞
1
∫ F(ω )e
jωt
= H(ud1 , vd2 )dω
2π
−∞

where

M1 M2
1
H(ud1 ,vd2 ) =
M1 M2
∑ ∑a
m1 =1m 2 =1
m1 m 2 e − j (um1 d1 + vm2 d 2 ) (2.38a)

which, for constant weighting coefficients, becomes

© 2001 by CRC Press LLC

 M1ud1 M vd
) sin( 2 2 ) − j [( M1 −1)u +( M 2 −1)v]
sin(
H(ud1 ,vd2 ) = 2 2 e 2
(2.38b)
ud1 vd2
M1 sin M2 sin
2 2
The frequency response function of a UPA given by (2.38b) can be written as a
product of two frequency response functions of two ULAs, one in the x
direction and the other in the y direction. This is also true in the weighted sum
case provided the weighting coefficients can be expressed as a product of two
coefficient sets, that is, amn = α mβ n where (α m1 , m1 = 0,1,... M1 − 1) and
(β m2 , m2 = 0,1,... M2 − 1). This is the motivation behind the use of a cross
array (fig. 2.15) in place of a UPA.
A UPA may be steered both in azimuth and elevation by means of
appropriate delays introduced before summation. The delay to be introduced in
the (m,n)th sensor is

m1d1 md
τ m1 ,m2 = sin θ cosϕ + 2 2 sin θsin ϕ
c c
where ϕ and θ are respectively azimuth and elevation to which the array is
required to be steered. In place of (2.37) we have

M1 M2
g(t) = ∑∑f
m1 =1m 2 =1
m1 m 2 (t + τ m1 ,m2 )

+∞ M M
1 1 2

= ∫
2π − ∞
F(ω)e jωt
∑ ∑
m1 =1m 2 =1
e − j[(u0 −u)m1 d1 +(v0 − v)m2 d 2 ]dω (2.39)

+∞
1
2π −∫∞
= F(ω)e jωt H((u0 − u)d1 ,(v0 − v)d2 )dω

where

H((u0 − u)d1 ,(v0 − v)d2 )

M1 (u0 − u)d1 M (v − v)d2
sin( ) sin( 2 0 ) − j [( M1 −1)(u0 −u)+( M 2 −1)(v0 − v)]
= 2 2 e 2
(u0 − u)d1 (v0 − v)d2
M1 sin M2 sin
2 2
where

© 2001 by CRC Press LLC

 ω
u0 = sin θ 0 cos ϕ 0 ,
c
ω
v0 = sin θ 0 sin ϕ 0
c

and (ϕ 0 , θ 0 ) are respectively azimuth and elevation angles.

Let the steering angles ϕ and θ be chosen such that
d1 k d l
sin θ cos ϕ = and 2 sin θ sin ϕ = where k and l are integers,
λ M1 λ M2
0 ≤ k ≤ M1 and 0 ≤ l ≤ M2 . Equation (2.39) may be expressed as a 2D spatial
discrete Fourier transform (DFT) of the wavefield incident on the array.

+∞ M M 2 πkm1 2 πlm 2
j( + )
1 1 2

g(t) =
2π −∫∞
F(ω)e jωt
∑∑
m1 =1m 2 =1
e − j (u 0 m1 d1 + v 0 m 2 d 2 )
e M1 M2
dω
(2.40a)
+∞
1
2π −∫∞
= F(ω)e jωt DFT{e − j (u0 m1 d1 + v0 m2 d 2 )}dω

and

H((u0 − u)d1 ,(v0 − v)d2 ) = DFT{e − j (u0 m1 d1 + v0 m2 d 2 )} (2.40b)

This result is an extension of (2.7) for a ULA to a UPA. The spatial frequencies
u and v, which appear in (2.40b), are linearly related to the frequency numbers,
k and l, in DFT

2πk 2πl
u= , v=
M1d1 M1d1

Given the frequency numbers, we can get the azimuth and elevation angles. For
λ l
example, assuming d1 = d2 = and M1 = M2 = M , ϕ = sin −1 and
2 k + l2
2

2 k2 + l2 M M
θ = sin −1 . Note that 0 ≤ k, l ≤ and 0 ≤ k 2 + l 2 ≤ . Thus, the
M 2 2
acceptable domain for k and l is as shown in fig. 2.17. For those values of k
and l lying outside this domain we get nonpropagating waves (see chapter 1,
page 11).
Random Array: An array with its sensors at random locations, modelled as an
independent, identically distributed random variable, is known as a totally
random array [12]. The root mean square (rms) error in the estimation of the

© 2001 by CRC Press LLC

 l

M/2

Propagating
domain

k
M/2-1 0 M/2

M/2-1

Figure 2.17: Acceptable domain for frequency numbers, k and l, is shown above. For
those values of k and l lying outside this domain we get non propagating waves.

azimuth angle (measured with respect to the x-axis) is given in [12] and
reproduced here without derivation

rms error = E{(ϕ̂ 0 − ϕ 0 )2 }

λ0 1 1
≈
2π 2snrM σ

where σ is the standard deviation of the random variable representing the

location of the sensors. The above approximate result is valid for large M.
Interestingly, the rms error decreases if the random array is spread out over a
large area. The practical interest in a random array arises in the context of
sonobuoy arrays used in submarine detection.
2.2.2 Uniform Circular Array (UCA): The sensors may be placed on a plane in
a polar grid. For a fixed radial distance we have a circle on which the sensors are
placed, forming a circular array. Consider a circular array of radius a with M
sensors, symmetrically placed on the circumference (see fig. 2.18). Let a plane
wavefront be incident on the array at angles ϕ and θ . The output of the mth
sensor is given by

© 2001 by CRC Press LLC

 z

ϕ x
a

Figure 2.18: Sensors are uniformly spaced on the circumference of a circle of radius
a. A plane wave is incident at an azimuth angle ϕ and an elevation angle θ .

∞ ωa 2 πm 2 πm
1 j[ωt − (cos ϕ sin θ cos +sin ϕ sin θ sin )]
f m (t) = ∫
2π − ∞
F(ω)e c M M
dω
(2.41)
∞ ωa 2 πm
1 j[ωt − (sin θ cos( −ϕ))]
= ∫
2π − ∞
F(ω)e c M
dω

Note that time is measured with respect to the time of arrival of the wavefront
at the center of the array. First, we evaluate the frequency response function.
The sum of all outputs in the frequency domain is given by

∞ M −1 − j[ ωa 2 πm
1 (sin θ cos( −ϕ))]
jωt 1
g(t) =
2π −∫∞
F(ω)e ∑e
M m=0
c M
dω
(2.42)
∞
1
= ∫
2π − ∞
F(ω)H(ω,ϕ,θ)e jωt dω

where the frequency response function H(ω, ϕ, θ) is given by

© 2001 by CRC Press LLC

 ωa 2 πm
1 M −1 − j[ (sin θ cos( −ϕ))]
H(ω,ϕ,θ) = ∑e
M m=0
c M
(2.43a)

For large M (for example, M> 48 when a=6λ and M>32 when a=4λ) the
summation in (2.43a) may be replaced by an integral and the result is

ωa
H(ω , ϕ , θ ) ≈ J0 ( sin θ ) (2.43b)
c
We shall call such a UCA a fully populated array. The most interesting
property of a circular array is that the frequency response function is independent
of ϕ . The property arises from (2.43b). Taking the distance to the first zero as
the effective half width of the main lobe, the angular width will be equal to
c
∆θ = sin −1 (2. 45 ). The height of the first (largest) sidelobe is 0.4025 at
ωa
c
θ = sin −1 (0.8 ).
ωa
A circular array may be steered to any desired direction just like a ULA
or a UPA. A delay τ m is introduced at each sensor output before summation,
where

a 2πm 2πm
τ m = [ (cosϕ sin θ cos + sin ϕ sin θsin )]
c M M
and ϕ and θ respectively are the desired azimuth and elevation angles. The
delayed outputs of all sensors are summed, for the time being without any
weighting.

1 M −1
g(t) = ∑ f m (t + τ m )
M m=0
∞ M −1 − j[ ωa 2 πm 2 πm
1 (sin θ 0 cos( −ϕ 0 )−sin θ cos( −ϕ))]
jωt 1
=
2π −∫∞
F(ω)e ∑e
M m=0
c M M
dω

(2.44)

where ϕ 0 and θ 0 respectively are the unknown azimuth and elevation angles of
the incident wavefront. Let

ωa 2 πm 2 πm
ωa 1 M −1 − j[ (sin θ 0 cos( −ϕ 0 )−sin θ cos( −ϕ))]
H( ,θ 0 ,ϕ 0 ,θ,ϕ) =
c
∑e
M m=0
c M M

© 2001 by CRC Press LLC

The output power of the array, steered to any chosen direction ϕ and θ , is
given by

ωa
2
output power= F(ω)H( ,θ 0 ,ϕ 0 ,θ,ϕ)
c

Earlier we had noted that steering of an array is equivalent to the spatial Fourier
transform of the array output. This result holds in a slightly different form for a
circular array. We will demonstrate how the spatial Fourier transform can be
used for estimation of azimuth [13] and elevation angles. Consider the spatial
discrete Fourier transform of the circular array output.

2 πm
1 M −1 −j
∑
k
gk (t) = f m (t)e M
M m=0
∞ M −1 − j[ ωa (sin θ cos( 2 πm − ϕ ))]
1 jωt 1
=
2 π −∫∞
F(ω )e ∑e c
M m=0
M
dω
(2.45)
∞
1
= ∫
2π −∞
F(ω )H(ω , ϕ , θ )e jωt dω

Taking the temporal Fourier transform of (2.45) we obtain an important result,

π
ωa jk
Gk (ω ) ≈ F(ω )Jk ( sin θ 0 )e 2 e − jkϕ 0 (2.46)
c
ωa
which is valid for k < kmax ≈ [15] and for sensor spacing approximately
c
λ
equal to [14]. Consider the following quantity:
2

ωa
Jk +1 ( sin θ 0 )
Gk +1 (ω) c
= je − jϕ 0 (2.47)
Gk (ω) ωa
Jk ( sin θ 0 )
c

Referring to the recurrence relation of Bessel functions [15],

2k
Jk +1 (x) = Jk (x) − Jk −1 (x) ,
x

© 2001 by CRC Press LLC

 we can write

Jk +1 (x) 2k Jk −1 (x)
= −
Jk (x) x Jk (x)

which we use in (2.47) and derive a basic result for the estimation of ϕ 0 and
θ0

Gk +1 (ω) 2k G (ω)
− je jϕ 0 = − je − jϕ 0 k −1 (2.48)
Gk (ω) ωa Gk (ω)
sin θ 0
c
Equation (2.48) may be solved for ϕ 0 and θ 0 As an example, we consider a 16
sensor circular array of 3λ radius and a source in a far field emitting a
bandlimited random signal. The center frequency is 100Hz and the bandwidth is
10Hz. The azimuth and the elevation angle of the source are respectively 10o
(0.1745 rad) and 45o (0.7854 rad). The sampling rate was 500 samples/sec. The
estimates were averaged over all frequency bins lying within the bandwidth. The
results are shown in fig. 2.19. Notice that the standard deviation of the
estimates decreases considerably when a reference sensor is used at the center.
The decrease is more pronounced at a very low snr, e. g., at 0 dB the decrease is
by a factor of three or more. An analysis of errors has shown that the standard
deviation is dominated by a few outliers which are caused by random noise in
the array output. Unless these outliers are eliminated the mean and the standard
deviation of the estimate gets severely affected. To overcome this problem
median in place of mean may be considered. It was observed through computer
simulation that the median is a better estimate of the azimuth than the mean.
When there is more than one source (say, P sources), equation (2.46)
takes a form

π
p−1
ωa
Gk (ω) = ∑ Fi (ω)Jk (
jk
sin θi )e 2 e − jkϕ i + ηk (ω) (2.49a)
i=0 c

We shall rewrite (2.49a) in a matrix form. For this purpose we define the
following vectors amd matrices:

G r (ω) = col{G0 (ω),G1 (ω),...Gr −1 (ω)}

ωa ωa ωa
Di (ω ) = diag  J0 ( sin θ i ), J1 ( sin θ i ), ... Jr −1 ( sin θ i ),
 c c c 

© 2001 by CRC Press LLC

 0.14

0.12

0.10
Standard deviation

0.08

0.06

0.04

0.02

0.00
512 1024 1536 2048
Data length

Figure 2.19: Standard deviation of azimuth and elevation estimates as a function of

data length. Thick line: with a sensor at the center, thin line: without a sensor at the
center, filled circle: azimuth and empty circle: elevation. snr=10 dB.

 − j (ϕ i − π2 ) − j 2(ϕ i − π2 ) π
− j (r −1)(ϕ i − ) 
Zi = col 1,e ,e ,...e 2

 
ηr (ω) = col{η0 (ω), η1 (ω),...ηr −1 (ω)}

where r is an integer (P ≤ r ≤ M) .
P −1
G r (ω) = ∑ Di (ω)Zi Fi (ω) + ηr (ω) (2.49b)
i=0

Let us assume that all P sources emit stationary uncorrelated stochastic

processes. Since the array output will also be a stationary stochastic process,
the spectral matrix is given by
1
S (ω )dω = E
2π g {
1
2π
1
dG r (ω ) dG rH (ω )
2π } where we
have used the generalized Fourier transform in place of the ordinary Fourier
transform. From (2.49b) it follows that

© 2001 by CRC Press LLC

  (D 0 Z0 ) 
H
 S0
 S  H

 1  (D1 Z 1 ) 
 . 
 . 
[
Sg = D 0 Z0 ,D1Z1 ,...D p−1Z p−1 ] 
.  .
   + σ 2ηI


 . .  
  
 Sp−1  (D p−1Z p−1 ) H 
 
(2.50)
(rxr) (rxp) (pxp) (pxr)

Exploiting the above signal structure a subspace approach has been developed in
[14, 16] for the estimation of azimuth and elevation.
Steering Vector: For circular array we define a steering vector as

a(ϕ , θ )
 − j[ ωa (sin θ cos( ϕ ))] − j[ ωca (sin θ cos( 2Mπ − ϕ ))] − j[
ωa
(sin θ cos(
2 π ( M −1)
− ϕ ))] 
= col e c ,e ,...e c M

 
(2.51)

Each term in the steering vector can be expanded in a series form [15]

ωa 2 πm ∞ 2 πm
− j[ (sin θ cos( −ϕ))] ωa −ϕ)
∑
jk (
e c M
= (− j)k Jk ( sin θ)e M

k =−∞ c

Define a matrix

m = 0,1,... M − 1
2 πkm
j
{W}km = e M
 
k = 0, ±1,...∞ 
and a vector

π
ωa − j ( +ϕ)k
{V(ϕ,θ)}k = Jk ( sin θ)e 2 , k = 0, ±1,...∞
c
In terms of matrix W and vector V we can express (2.51) as

a(ϕ , θ ) = W V(ϕ , θ )
M ×1 M × ∞ ∞ ×1

© 2001 by CRC Press LLC

Observe that matrix W is independent of the azimuth and elevation angles,
which are confined only to the vector V(ϕ, θ) . The size of V(ϕ, θ) depends on
the argument of the Bessel function which may be approximated, for large
order, as [15]

ωa
e sin θ
ωa 1 c
Jk ( sin θ) ≈ ( )k
c 2πk 2k

ωa
Hence, the size of vector V(ϕ, θ) will be of the order of e sin θ . For
c
π
example, for a = 8λ and θ = the size of the vector is about 140. The
2
steering vector for a circular array possesses some interesting properties,
namely,

(a) a(ϕ,θ) ≠ a(−ϕ,θ) ,

∗
(b) a (ϕ,θ) = a(π − ϕ,θ) and
(c) a(ϕ,θ) periodic in ϕ with period 2π, and independent of sensor
spacing.

Property (a) implies a wavefront coming from the north can be distinguished
from one coming from the south (north-south ambiguity). Property (b) implies
that a complex signal coming from the east can be distinguished from the one
coming from the west (east-west ambiguity; see fig. 2.20 for an illustration)
and property (c) implies that for any sensor spacing there is no grating lobe in
the range of ±π . A circular array differs from a linear array in respect of
properties (a- c).
Boundary Array: An array of sensors for localization of an object in the near
field region may take a form of a boundary array where the sensors are placed all
around the object as its approximate location is known before hand. A circular
array enclosing a source is an example of boundary array. With three coplanar
sensors and accurate time delay measurements it is possible to localize a point
target in the plane of array. There is a vast literature on time delay estimation
but, on account of space limitation, the topic of time delay estimation will not
be covered in this book. It is, on the other hand, possible for a boundary array
to determine the time delays from the phase measurements. The basic idea is to
consider a pair of sensors and estimate the phase difference at a fixed frequency
ω . The source will fall on one of the phase trajectories drawn with a phase
difference ω∆τ = 2πs + φ where ∆τ is the unknown time delay, s is an integer
and φ is the actual observed phase difference. To estimate the time delay ∆τ ,

© 2001 by CRC Press LLC

 N

c a

π −ϕ
ϕ
x E
−ϕ

Figure 2.20: Circular array does not suffer from north-south ambiguity, that is,
wavefronts a and b can be distinguished. There is no east-west ambiguity for
complex signals, that is, wavefronts a and c can be distinguished.

ω∆τ 
we need to know the integer constant s in the range ± s0 where s0 = Int 
 2π 
and Int[ x ] stands for the largest integer less than x. For a given φ , ∆τ will
assume a set of 2s0 + 1 values and for each value of s there corresponds a locus
of points called the phase trajectory [17]. For every pair of adjacent sensors we
can draw a suite of trajectories as shown in fig. 2.21a. The unknown source
must lie on any one of the trajectories. Next, consider another sensor pair and
draw another suite of trajectories. Any one of the points of intersection is a
possible location of the unknown source (see fig. 2.21b). Since there are M
pairs of sensors there will be M suites of trajectories. The true position of the
unknown source is then given by the intersection of all M trajectories, one
from each suite. At true source location all phase estimates obtained from
different pairs of sensors must match with the theoretically evaluated phases.
Let φ̂1k , k = 0,1,..., M − 1 be the estimated phases from M pairs of adjacent
sensors and φ1k , k = 0,1,..., M − 1 be the theoretically computed phases. Define
an error vector

© 2001 by CRC Press LLC

 (a)

(b)

F igure 2.21: (a) For every pair of adjacent sensors we draw a suite of equiphase
trajectories. (b) Intersection of two sets of equiphase trajectories. The unknown
source must lie at one of the intersections. For every adjacent pair of sensors the
search is carried out within the dotted quadrilateral.

{ 0 0 1 1 M −1
ε = col (e jφ1 − e jφ̂1 ),(e jφ1 − e jφ̂1 ),...,(e jφ1
M −1
− e jφ̂1 ) } (2.52)

© 2001 by CRC Press LLC

 Figure 2.22 A plot of inverse error power. The source is at range 50 meters and
azimuth -60o . A circular array of radius 100 meters and having eight equispaced
sensors is assumed.

H
and error power= ε ε . The error power will be zero at the true location of the
source. This property may be utilized to spot the true location from all
available intersections of any two suites of trajectories. Evaluate the error power
at each and every intersection. That intersection which yields zero (or
minimum) error power is the most likely location of the unknown source.
Finally, to reconstruct the true phase we need to know the integers which may
be obtained from the order of the trajectory for every pair of sensors passing
through the source location. However, in any practical problem, this step may
not be required as the interest is usually in localization and not in the true phase
retrieval. The results of a numerical experiment are shown in fig. 2.22. The
question of minimum size (number of sensors) of the array, required for unique
localization, has not been answered. However, numerical experiments suggest a
minimum array size of five sensors placed the circumference of a large circle. It
is not necessary that the array be perfectly circular. Any closed curve will do,
provided the phase trajectories are drawn for each pair.

2.2.3 Distributed Dipoles: A planar array with sensors randomly distributed in a

plane constitutes a random planar array. Here we shall consider a random planar
array of dipole sensors. A dipole sensor consists of a pair of identical sensors
displaced by ∆ (a vector). In fig. 2.23 an example of a dipole planar array is
shown. Let r i , i = 0,1,... M − 1 represent locations of M dipole sensors (the
midpoint of the sensor pair). It may be emphasized that the spatial distribution
of the dipoles can be quite arbitrary. Define the following data vector in
frequency domain:

© 2001 by CRC Press LLC

 ∆

Figure 2.23: Randomly distributed planar dipole sensors. Each dipole consists of
two identical sensors displaced by ∆ (vector).

F1 
F̃ =   (2.53)
F 2 
where
F1 = col{F0+ (ω), F1+ (ω)... FM+ −1 (ω)}
F 2 = col{F0− (ω), F1− (ω)... FM− −1 (ω)}

where superscript + refers to the upper sensor and - refers to the lower sensor.
Let P plane wavefronts from P sources be incident on the array.

[ ]
F1 = a 0 ,a1 ,...a p−1 F(ω) + dη1 (ω) (2.54)
where
 − j ∆ ⋅δ i − j ω (r1 + ∆ )⋅δ i ω ∆
− j (r M −1 + )⋅δ i 
a i = col e 2 ,e c 2 ,...,e c 2

 
∆⋅δ i
 − j ωc r1 ⋅δ i ω
− j r M −1 ⋅δ i  − j
= col 1,e ,...,e c
e 2
 

© 2001 by CRC Press LLC

 where δ i = col {sin θ i , cos θ i } and F(ω ) = [F0 (ω ), F1 (ω )... FP−1 (ω )]T is a
vector whose components are Fourier transforms of the waveforms emitted by P
sources. Equation (2.54) may be expressed as follows

F1 = AΓ + F(ω) + dη1 (ω) (2.55a)

where

∆⋅δ p−1
 − j ∆⋅δ 0 − j ∆⋅δ1 −j 
Γ + = diag e 2
,e 2
,...,e 2 
 
and

1 1 ... 1 
 − j ω r ⋅δ − j ω r ⋅δ ω
− j r 1 ⋅δ p−1

e c 1 0 e c 1 1 e c 
 
. 
A= 
.
 
. 
 ω ω ω 
e − j c r M −1 ⋅δ 0 e − j c r M −1 ⋅δ1 ... e − j c r M −1 ⋅δ p−1 
 
Similarly,

F 2 = AΓ − F(ω) + dη2 (ω) (2.55b)

where
∆⋅δ p−1
 j ∆⋅δ 0 j ∆⋅δ1 j 
Γ − = diag e 2
,e 2
,...,e 2 
 

We shall now compute the spectral matrix of dipole array output. For this the
incident signal will be treated as a stationary stochastic process. Then, in place
of the ordinary Fourier transform we need to invoke the generalized Fourier
transform of a stationary stochastic process (see p.40).

© 2001 by CRC Press LLC

  1 dF1  1 dF1  
H
1
2π
S f̃ dω = E { 1
2π
dF̃ 1
2π
dF̃ H
} = E 2π   2π   
 dF 2  dF 2  
(2.56)
AS f A H AS f Γ H A H  I 0  σ η1 
2
1 
= 2π   dω +   2  dω
1
2π 
0 I  σ η 2 
H
AΓS f A AS f A H 

where
{
Γ = Γ − Γ H+ = diag e jω∆⋅δ 0 ,e jω∆⋅δ1 ,...,e
jω∆⋅δ p−1
}
{
S f = diag S f 0 , S f 1 ,..., S f p−1 }
Equation (2.56) may be expressed as

I 0  σ η1 
H 2
A  A 
S f̃ =  S f   +   2  (2.57)
AΓ  AΓ  0 I  σ η2 

The DOA information is buried in Γ while the A matrix contains dipole

location information, which is known. We shall show later in chapter 5 how Γ
can be estimated using the subspace approach.
Electric Dipole Array: Now consider a ULA of electric dipoles. There are two
possible arrangements: a dipole axis along an x-axis or along an y-axis as
illustrated in fig. 2.24. Note that an electric dipole will measure the electric
field component along its axis. Let a plane EM wave with its wave vector in
the y-z plane be incident on both types of arrays (see figs. 2.24 a&b) of electric
dipoles.
Thus, we are simultaneously measuring the x and y components of the electric
field. The array outputs are given by

(a) for dipoles oriented in the x direction

T
 j ωdc sin θ j 2 ωdc sin θ j ( M −1)
ωd
sin θ 
j (ωt +ϕ)
f x = − cos γ 1,e ,e ,...,e c
 Ee
 
= − cos γ aEe j (ωt +ϕ)

(b) for dipoles oriented in the y direction

© 2001 by CRC Press LLC

 z

(a)
z

x
(b)

Figure 2.24: A ULA of electric dipoles oriented in the x direction (a) and in the y
direction (b).

T
 j ωd sin θ j 2 ωdc sin θ j ( M −1)
ωd
sin θ 
j (ωt +ϕ)
f y = sin γ cosθe 1,e cjκ
,e ,...,e c
 Ee
 
= sin γ cosθe jκ a Ee j (ωt +ϕ)

© 2001 by CRC Press LLC

where θ is DOA and γ and κ are defined in terms of the polarization
parameters (see p. 8). When there are P wavefronts incident on a ULA of
dipoles, the array output may be expressed as follows:

f x = AΓ1S0 e jωt + η1 (t) (2.58)

where
Γ1 = −diag{cos γ 0 ,cos γ 1 ,...,cos γ P −1}
A = {a 0 ,a1 ,...,a P −1}
{
S0 = col E0 e jϕ 0 , E1e jϕ1 ,..., EP −1e
jϕ p−1
}
and η1 (t) = col {η0′ (t), η1′ (t),..., η M−1
′ (t)} is the background noise vector. A
similar expression for an array of dipoles oriented along the y-axis is given by

f y = AΓ 2S0 e jωt + η2 (t) (2.59)

where

{
Γ 2 = diag sin γ 0 cosθ 0 e jκ 0 ,sin γ 1 cosθ1e jκ 1 ,...,sin γ P −1 cosθ P −1e jκ P−1 }
η2 (t) = col{η′′0 (t), η1′′(t),..., η′′M −1 (t)}

§2.3 Broadband Sensor Array:

Very often an array of sensors is required to receive broadband signals, which
include both natural signals (e.g., seismic and sonar signals) or man-made
signals (e.g., communication signals). Not only must the sensors be broadband
but also the special processing technique must be devised to exploit the
broadband property of the signal. The bandwidth of the array response function
depends upon the time taken by the wavefront to sweep across the array
aperture, hence on the DOA for ULA but only on the diameter of UCA. A
broadband signal may be considered as an aggregate of many narrowband signals
covering the entire bandwidth. Since every narrowband signal is capable of
determining the source parameters, we have considerable redundant information
which may be used to fight against the noise and the model uncertainties.
Alternatively, a broadband signal may also be represented by a collection of
independent temporal samples (sampled at Nyquist rate) or snapshots, each
carrying information about the source parameters. In this section we shall
introduce the concept of array bandwidth, and frequency and time snapshots.
2.3.1 Bandwidth of an Array: The bandwidth of an array is the bandwidth of its
transfer function which for a ULA is given by (2.3a)

© 2001 by CRC Press LLC

 M
sin( ωτ) j M −1 ωτ
1 M −1 jω
md
sin θ
H(ωτ) = ∑e
M m=0
c
= 2
ωτ
e 2
M sin
2
d
A plot of H(ω sin θ ) with θ held fixed but ω varied will be same as in fig.
c
2.2 except that the x-axis variation is now linear in ω , in place of a noninear
variation in θ . If we define the bandwidth as one half of the distance between
the first nulls, we obtain

2π c 2π 2π
∆ω = = = (2.60)
M d sin θ Md τ
cx

where τ is the time required to sweep across the array aperture. Note that the
bandwidth is infinite when the wavefront is incident on the broadside, that is,
2 πc
the array is steered in the direction of source; but it is equal to when the
Md
wavefront is incident from the endfire. For a circular array, however, the
wavefront has to sweep a constant aperture equal to the diameter of the circle,
independent of the azimuth angle. This essential difference between a ULA and a
UCA is illustrated in fig. 2.25. The bandwidth of a long ULA or a large
diameter UCA is very small and hence much of the energy of a broadband
source will be lost, unless the array is steered in the direction of the desired
source. Further, as pointed in §2.1, when there is more than one source
radiating at different frequencies there is a possibility for ambiguity. Let P
narrowband sources with DOAs θ i , i = 0,1,..., P − 1 and center frequencies ω i ,
i = 0,1,..., P − 1 be incident on a ULA. Further, we assume that the center
frequency and DOA pair satisfies the following relation:

d
ωi sin θi = τ 0 (cons tan t), i = 0,1,..., P − 1 (2.61)
c
where τ 0 is the delay per sensor introduced by all sources. Now, through a
process of sum and delay, the array is steered simultaneously, to all sources; in
other words, the array will “see” all P sources at the same time. When the
sources are broadband with overlapping spectrum we can always find a set of
frequencies which satisfies (2.61). As a result, when the array is steered to one
of the sources the output may be contaminated with the power derived from
other sources. Such an interference is unacceptable particularly when waveform

© 2001 by CRC Press LLC

 Aperture=(M-1)d

(a)

2a

(b)

Figure 2.25: The essential difference between a ULA and UCA is that the effective
aperture (dashed line) is azimuth dependent for a ULA but independent of the azimuth
for a UCA. (a) Linear array, the effective aperture= (M − 1)d sin θ . (b) Circular
array, the effective aperture = diameter of the circle.

© 2001 by CRC Press LLC

 estimation is the goal. Some of these problems can be overcome by introducing
temporal processing over and above the delay and sum type of spatial
processing.
A snapshot may be expressed in the frequency domain as

∞
1
f(t) = ∫
2π − ∞
a(ω,θ 0 )dF0 (ω)e jωt (2.62)

where

 − jω d sinc θ 0 − jω( M −1)

d sin θ 0

a(ω,θ 0 ) = col 1,e ,...,e c

 
On account of the properties of the steering vector listed on page 92, we can
conclude the following about the broadband array output:

(a) f θ 0 (t) = f π −θ 0 (t) . Follows from property (a) on page 92.

(b) f θ 0 (t) ≠ f −θ 0 (t) . From property (b) and (2.61) we can write

∞
1
f −θ 0 (t) = ∫
2π − ∞
a∗ (ω,θ 0 )dF0 (ω)e jωt ≠ f θ 0 (t)

except in the unlikely event of dF0 (ω ) being real. Thus, the east-west
ambiguity shown for the narrowband complex signal does not apply to the
broadband signals.
2.3.2 Broadband Signals: In the case of thebroadband signal, since a snapshot
may vary rapidly, it is necessary that many snapshots must be collected at
different time instants in the past; for example, f (t), f (t − ∆t), f (t − 2∆t),...,
f (t − ( N − 1)∆t) are N past or delayed snapshots (see fig. 2.26). An alternate
approach, in the case of broadband signal, is to go over to the frequency domain
(temporal frequency). The output of each sensor, consisting of N samples, is
subjected to Fourier analysis (DFT). A collection of the Fourier coefficients,
one from each sensor at a fixed frequency, constitutes a frequency snapshot. The
array signal processing of broadband signals using the frequency snapshots
closely follows the time domain approach for narrowband signals (after
removing the carrier), widely used in radar signal processing. In place of a
covariance matrix we use a spectral matrix which indeed is a spatial covariance
matrix (SCM). In this book we shall emphasize the frequency domain approach
as the time domain approach for wideband signals turns out to be conceptually a
bit more involved. We begin with time domain approach. First, let us introduce

© 2001 by CRC Press LLC

some new quantities required in the processing. All delayed snapshots are
stacked one above the other to form one large vector of size M ⋅ N ,

f MN (t) = col{f T (t),f T (t − ∆t),f T (t − 2∆t),...,f T (t − (N − 1)∆t)}

(2.63)

We define a covariance matrix known as spatio-temporal covariance matrix,

STCM,

CSTCM = E{f ML (t)f HML (t)} (2.64)

As an example, consider a two sensor array (M=2) and two delayed snapshots
(N=2). The STCM is given by

C STCM =

C f (0) C f (τ 0 ) C f (∆t) C f (∆t + τ 0 )

 
C f (−τ 0 ) C f (0) C f (∆t − τ 0 ) C f (∆t) 
 
C f (∆t) C f (∆t − τ 0 ) C f (0) C f (τ 0 ) 
 
C f (∆t + τ 0 ) C f (∆t) C f (−τ 0 ) C f (0) 
 

d sin θ 0
where τ 0 = is a propagation delay per sensor. We express the stacked
c
vector by using (2.62) in (2.63),

∞
1
f stacked (t) = ∫
2π − ∞
ζ(ω) ⊗ a(ω,θ 0 )dF0 (ω)e jωt (2.65)

where

ζ(ω) = col{1,e − jω∆t ,...,e − jω( L −1)∆t }

and ⊗ stands for the Kronecker product. Define a direction vector h(ω , θ 0 ) =
ζ(ω ) ⊗ a(ω , θ 0 ) and rewrite (2.65) as

© 2001 by CRC Press LLC

 f 0 (t) f 0 (t- ∆ t) f 0 (t- 2 ∆ t) f 0 (t- 3 ∆ t)
∆t ∆t ∆t

f 1 (t) f 1 (t- ∆ t) f 1 (t- 2 ∆ t) f 1 (t- 3 ∆ t)

∆t ∆t ∆t
Sensor array

f 7 (t) f 7 (t- ∆ t) f 7 (t- 2 ∆ t) f 7 (t- 3 ∆ t)

∆t ∆t ∆t

Figure 2.26: The output of an eight sensor ULA is sampled at different time instants
in the past. All outputs taken at the same time instant are grouped into a vector
called a temporal snapshot. We have as many temporal snapshots as the number of
samples or taps taken from each sensor output.

∞
1
f stacked (t) = ∫
2π − ∞
h(ω,θ 0 )dF0 (ω)e jωt (2.66)

which may be considered as an extended output of the array (stacked vector) due
to a source at azimuth angle θ 0 . Using (2.66) we can express STCM as

CSTCM = E{f stacked (t)f (t)} =

1
2π −∫∞
H
stacked h(ω,θ 0 )h H (ω,θ 0 )S f 0 (ω)dω

(2.67)

© 2001 by CRC Press LLC

Let dF stacked (ω ) be the generalized Fourier transform of the extended vector
output, f stacked (t) ,

dF stacked (ω) = h(ω,θ 0 )dF0 (ω) (2.68)

Premultiply on both sides of (2.68) with the steering vector in some desired
direction and evaluate the expected value of the magnitude square. Dividing by
(M ⋅ N ) 2 we get the output power

1
h H (ω, θ)S f (ω)h(ω, θ) =
(M ⋅ N )2 stacked
2 2 (2.69)
h H (ω, θ)h(ω, θ 0 ) a H (ω, θ)a(ω, θ 0
S0 (ω) = S0 (ω)
M⋅N M

where S f stacked (ω) is the spectral matrix of an extended array signal.

Interestingly, in spite of delayed snapshots used in deriving (2.69), the output
power remains the same as in (2.19b). However, the STCM is different from
the spatial covariance matrix (SCM). For example, reconsider the frequency-
direction ambiguity due to the fact that if there is a set of DOAs and frequencies
such that

d
ωi sin θi = τ 0 (cons tan t), i = 0,1,..., P − 1,
c

the steering vector a(ω i ,θi ) remains unchanged. This, however, is not true
for a ULA with a tapped delay line, as the steering vectors, h(ω i , θ i ) =
ζ(ω i ) ⊗ a(ω i , θ i ), i = 0,1,..., P − 1, will be different because of ζ(ω i ). It may
be noted that for a circular array there is no frequency-direction ambiguity of the
type described above even when no tapped delay line is used [18].

§2.4 Source and Sensor Arrays:

In nonoptical imaging systems such as in radar, sonar, seismic, and biomedical
imaging systems, we often use an array of sources and an array of sensors. The
main idea is to illuminate an element to be imaged from different directions and
collect as much of the scattered radiation as possible for the purpose of
imaging. To achieve this goal in a straightforward manner would require an
impractically large array. It is known that a large aperture array can be
synthesized from a small array of transmitters and receivers (transceivers). This
leads to the concept of coarray used in radar imaging. In tomographic imaging
the element to be imaged has to be illuminated from all directions. We,

© 2001 by CRC Press LLC

 therefore, need a boundary array of transceivers or a single source and a line of
sensors going round the object as in some biomedical imaging systems. In
seismic exploration a source array is often used to provide the additional degree
of freedom required to combat the high noise and the interference level, and also
to estimate the wave speed in different rock layers.
2.4.1 Coarray: An imaging system is basically concerned with faithful mapping
of a point target in the object space into a point in the image space. The point
spread function (PSF) of the imaging system describes the mapping operation.
Ideally, it is a delta function, but in practice a point may be mapped into a
small area, the size of which will depend upon the aperture of the array used in
the imaging system. The coarray is required to synthesize an arbitrary PSF
using a finite array of sensors and sources. A source array (transmit array) is
weighted with a complex function t(x) and the sensor array (receive array) is
also weighted with another complex weight function r(y). The response
function of a source array in the direction of the scatterer (in far field) is given
by
+∞

T(u,v) = ∫ ∫ t(x)δ(y)e j (ux + vy) dxdy (2.70a)

−∞

and the response function of the sensor array in the direction of the scatterer is
given by
+∞

R(u,v) = ∫ ∫ r(y)δ(x)e j (ux + vy) dxdy (2.70b)

−∞

2π
where u = k sin θ cos ϕ , v = k sin θ sin ϕ and k = is the wavenumber. Let
λ
ρ (u, v) be the reflection coefficient as a function of the azimuth angle ϕ and
the elevation angle θ . The total response of the source/sensor array is given by

O(u,v) = T(u,v)ρ(u,v)R(u,v)

This is the image of a point scatterer in the frequency domain. If we assume

that the scatterer is omnidirectional, that is, ρ (u, v) = ρ 0 (constant), the PSF of
the array will be given by the inverse Fourier transform of T(u, v)R(u, v)

PSF = ∑ ∑ t(x)δ(y)r(q − y)δ( p − x)

x y

= t( p)r(q)

Thus, we note that the combined PSF of a source array along the x-axis and a
sensor array along the y-axis is equal to the product of two weighting functions;

© 2001 by CRC Press LLC

for example, when the source and sensor arrays are uniformly weighted, the
PSF of the combined system is a rectangular function [19]. This is illustrated
in fig. 2.27 for an L-shaped array. For an arbitrary distribution of sources and
sensors the PSF is given by a convolution of the source and sensor weighting
functions,

+∞

PSF = ∫ ∫ t(x, y)r( p − x,q − y)dxdy (2.71)

−∞

The PSF at point (p,q) is obtained by summing the product of two weight
functions over two lines p=x+x´ and q=y+y´ where (x,y) refer to a point in the
source array and (x´,y´) refer to a point in the sensor array. Such a coarray is
also known as a sum coarray. Since we have a discrete set of point sources and
sensors, the discrete version of equation (2.71) is given by R T T where
T = col {t0 , t1 ,...t M−1 } and R = col {r 0 , r1 ,... r M−1 } .
An arbitrary PSF may be synthesized through eigen-decomposition.
Let P be the desired PSF given in the form of a matrix which we shall assume
to have a hermitian symmetry. We can express its eigen-decomposition as
M−1
P= ∑λ u
l=0
l
H
l u l where λ l is lth eigenvalue (real) and u l is the corresponding
eigenvector. We let T = R = u l , that is, both source and sensor arrays are
weighed by the same eigenvector of the desired PSF matrix, and thus an image
is formed. This step is repeated over all significant eigenvectors. All such
images are linearly combined after weighting each with the corresponding
eigenvalue [20].
Let us consider an example of circular array of eight transceivers
uniformly distributed on the circumference. A transceiver consists of a source
and a sensor physically placed at the same place. The transceivers are located at

(a, 0), ( a , a ), (0, a), (− a , a ), (0, −a), 

 2 2 2 2 
 
(− a , − a ), (−a, 0), ( a , − a ) 
 2 2 2 2 

where a stands for radius of the circle. The sum coarray consists of 64 locations
whose coordinates may be found by summing the 1st column and the 1st row
entries from table 2.2. The actual coordinates, thus obtained, are also shown in
N2
the table. There are in all + 1 distinct nodes. The coarray nodes lie on a set
2
of concentric circles of radii (0, 0.76a, 1.141a, 1.85a, 2a) as shown in fig. 2.28.

© 2001 by CRC Press LLC

 (a)
Sensor array

Source Array

(b)

PSF

Figure 2.27: (a) Source and sensor array. (b) Point spread function (PSF) on a grid,
known as coarray.

The synthesized aperture at (m,n)th node of the circular coarray is given by

2π 2π
t (a cos(m ), asin(m ))
M M

© 2001 by CRC Press LLC

Column entry m
(a,0) (
a
, (0, a) (−
a
, (0, −a) −(
a
, (−a,0) (
a
,
2 2 2 2
a a a a
) ) ) − )
2 2 2 2
(a,0) (2a,0) a+
a
, (a,a) a−
a
, (a,-a) a−
a
, (0,0) a+
a
,
2 2 2 2
a a a a
− −
2 2 2 2
a a a a (0,0) a
( , a+ , 2a, , (0, 2a) , − a, ( 2a, 0)
2 2 2 2 2
a a 2a a a a
) a+ −a
2 2 2 2 2
(0,a) (a,a) a (0,2a) a (0,0) a (-a,a) a
, − , − , ,
2 2 2 2
a a a a
a+ a+ a− a−
2 2 2 2

(−
a
, a−
a
, (0, 2a) −
a
, − 2a, −
a
, − 2a, 0 −a−
a
, (0,0)
2 2 2 2 2
a a a 2a a a
) a+ −a +
2 2 2 2 2
(0,− a) (a,-a) a (0,0) a (0.2a) a (-a,-a) a
, − , − , ,
2 2 2 2
a a a a
2
−a −a + −a − −a −
2 2 2
−(
a
, a−
a
, (0,0) −
a
, − 2a, 0 −
a
, − 2a, −a−
a
, (0,
2 2 2 2 2
a a a a − 2a a − 2a)
) − a− −a − − ,
2 2 2 2 2
( − a,0) (0,0) a
− a, (-a,a) −a−
a
, (-a,-a) −a−
a
, (-2a,0) −a+
a
,
2 2 2 2
a a a a
− , −
2 2 2 2

(
a
, a+
a
, ( 2a, 0)
a
, (0,0) a
, 0, a
− a, 2a,
2 2 2 2 2
a a a a − 2a a
− 2a
− ) − a− −a − −
2 2 2 2 2

Row entry

Table 2.2: The sum coarray consists of 64 locations whose coordinates are obtained
by summing the 1st column (left shaded column) and the respective entries from the
1st row (top shaded row). The coordinates are shown in clear cells.

© 2001 by CRC Press LLC

 2π 2π
r (a cos(n ), asin(n ))
M M
The PSF at (p,q) where p = x m + x n and q = ym + yn may be obtained from the
discrete equivalent of (2.70). Note that the coordinates of the mth point are
(x m , ym ) and those of the nth point are (x n , yn ).

PSF( p, q)
2π 2π 2π 2π
= t(a cos(m ), a sin(m ))r(a cos(m ), a sin(m ))
M M M M
m=n
t(a cos(i 2 π ), a sin(i 2 π ))r(a cos((i + M ) 2 π ), 
 M M 2 M 
 
−1 a sin((i + M ) 2 π )) +
M

2
 
= ∑ 
i = 0  r(a cos(i
2π
2 M
2π M 2π 
), a sin(i ))t(a cos((i + ) ), 
 M M 2 M 
 M 2π 
 a sin((i + ) )) 
 2 M 
M
m−n =
2
2π 2π 2π 2π
= t(a cos(m ), a sin(m ))r(a cos(n ), a sin(n ))
M M M M
(2.72)
2π 2π 2π 2π
+r(a cos(m ), a sin(m ))t(a cos(n ), a sin(n ))
M M M M
m≠n
M
m−n ≠
2

As an illustration, consider uniform transmitter strength and receiver

sensitivity; let t = t0 and r = r 0 . Clearly, since the array has a circular
symmertry, PSF will also be radially symmetric. It is straightforward to show
that

PSF(0,0) = 8t0 r0
(2.73)
PSF( p 2 + q 2 ) = 2t0 r0

The PSF has a tall peak at the center, four times the background level.

© 2001 by CRC Press LLC

 (a)

(b)

Figure 2.28 (a) A circular array (radius=1.5cm) of eight transceivers. (b) Coarray:
Nodes are shown by empty circles. There are 33 nodes. The diameter of the outermost
circle is 3.0cm. The PSF is defined at the nodes. For uniform array, with constant
source strength and sensor sensitivity, the PSF at the center is equal to 8 t0 r 0 and
elsewhere it is equal to 2 t0 r 0 .

In general, with M sensors PSF(0, 0) = Mt0 r 0 but PSF( p 2 + q 2 ) = 2t0 r 0 .

Thus, for large M the PSF tends to a delta function, the ideal condition for a
perfect imaging system. A reverse problem of synthesizing a circular array of
transceivers given the desired PSF is proposed in [19].

© 2001 by CRC Press LLC

 2.4.2 Passive Imaging: In passive imaging the object to be imaged is itself a
source of energy, for example, a distant star or an earthquake deep inside the
earth. It is of some interest to estimate the distribution of the energy as a
function of the azimuth and elevation or the spatial frequencies u and v
2π 2π
(u= sinθ cosϕ and v= sinθ sinϕ). The imaging system consists of two
λ λ
receiving arrays each with its own weighting functions. Let r1 (x, y) and
r2 (x, y) be the weighting functions. Let Ρ(u, v) be the source distribution. The
outputs of the two arrays, in frequency domain, can be written as

O1 (u,v) = Ρ(u,v)R1 (u,v)

(2.74)
O2 (u,v) = Ρ(u,v)R2 (u,v)

We form a cross-correlation of the two outputs. Using (2.74) we obtain

I(u,v) = O1 (u,v)O2* (u,v)

(2.75a)
= Ρ(u,v) R1 (u,v)R2* (u,v)
2

Or

+ ∞ 
I(u,v) = Ρ(u,v) ∫ ∫  ∫ ∫ r1 (x, y)r2 (x + p, y + q)dxdy e j (up+ vq) dpdq
2

p q − ∞ 
(2.75b)

The quantity inside the square brackets in (2.75b) represents the cross
correlation of two weighting functions.

+∞

Cr1r 2 ( p,q) = ∫ ∫ r1 (x, y)r2 (x + p, y + q)dxdy (2.76)

−∞

Let x ′ = x + p and y ′ = y + q . Then the cross-correlation may be looked upon

as an integral of the product of weight functions at r1 (x, y)and r2 ( x ′, y ′ ) for
fixed p and q; alternatively, for a fixed difference, p = x ′ − x and q = y ′ − y.
Such a coarray is also known as a difference array. For example, for a L-shaped
receiver array (sensors on both arms) the difference array is on a square grid in
the fourth quadrant (see fig. 2.29). Recall that the sum coarray is also on a
square grid but in the 1st quadrant (see fig. 2.27).
2.4.3 Synthetic Aperture: A long aperture is the primary requirement for
achieving high spatial resolution; however there is a limit on the size of a
physical array that one can deploy. This is particularly true in imaging systems

© 2001 by CRC Press LLC

where it is uneconomical to use a large physical array required to cover a large
target area with high resolution. The main idea in synthetic aperture is that by
moving a transceiver, preferably at a constant speed and in a straight line, an
array of large aperture can be synthesized through subsequent processing. A
simple illustration of this principle is shown in fig. 2.30. A transceiver is
moving at a constant speed cs in a straight line parallel to the ground. Both
transmit and receive beams are wide angle, as shown in the figure. At position
p1 the beam just begins to illuminate a scattering particle on the ground. The
particle remains under illumination until the transceiver reaches position p2.
The energy scattered by the particle is received over an interval p1p2; hence the
effective array aperture is equal to p1p2. This is the synthesized aperture. Let us
now look into how to process the echoes received over p1 p 2 . These echoes
reach the receiver with different delays depending upon the position of the
transceiver. Secondly, since the transceiver is moving, there will be a Doppler
shift proportional to the component of transceiver speed in the direction of the
radius vector joining the transceiver and the scattering particle. From fig. 2.30b
it can be shown that p1 p2 = 2l0 tan θ 0 . Let us assume that the transceiver is
a directional sensor with effective aperture L; hence the beam width is 2 θ 0
where sin θ 0 = λ/L. The beamwidth is measured between the first two nulls of
the array response function (see p. 77-82). Therefore, the synthesized aperture
may be expressed as

λ
p1 p2 = 2l0 (2.77)
L − λ2
2

where l0 is the height of the moving transceiver above ground. The smallest
size of an object that can be seen on the ground with the help of an array with
aperture given by (2.77) will be approximately equal to L2 − λ 2 . The
underlying assumption is that p1 p2 >> λ which would allow us to simplify
the Rayleigh resolution criterion (see p. 222) and then the result follows. The
requirement that p1 p2 >> λ is easily met by selecting a small antenna as a
transmitter. The two way (one way in a passive system) travel time from the
scatterer to the transceiver is given by

2l 2 l0 + (x − x' ) 2 l02 + (cs t)2

2 2
= = (2.78)
c c c

© 2001 by CRC Press LLC

 q

Figure 2.29 Difference coarray obtained from L-shaped receive array shown in fig.
2.24(a).

where we have expressed the horizontal distance as a product of time and

transceiver speed, (x − x' ) = cs t . Let the transmitted waveform be a sinusoid
2l
(real), cos(ωt + ϕ ). The received waveform which arrives after a delay of is
c

2 l02 + (cs t)2

f (t) = r0 cos(ω(t − ) + ϕ) (2.79)
c
Since the phase of the received waveform is time dependent, the instantaneous
frequency [1] will be different from the frequency of the transmitted waveform.
This difference is the Doppler shift. The instantaneous frequency is given by

2ωcs cs t
ω(t) = ω −
c l + (cs t)2
2
0

© 2001 by CRC Press LLC

 p2 p1

(a)

Beam Beam

Ground
p2 p1
x

l
2θ 0 (b)

l0 θ

Beam Beam
x'

Ground

Figure 2.30: (a) A simple illustration of the principle of synthetic aperture. The
energy scattered by a particle on the ground is received by a transceiver at p. (b) The
two way travel time from the scatterer at x´ to the transceiver at x.

Hence, the Doppler shift is equal to

2ω cs cs t 2ω cs
∆ω = − =− sin θ (2.80)
c l02 + (cs t)2 c

In the case of passive synthetic aperture sonar, the Doppler shift can be used for
estimation of the direction of arrival of an unknown source [21].
At a fixed place and time the signal received consists of a sum of the
scattered wavefields from a patch on the ground which is coherently illuminated
(see fig. 2.31). The size of the patch is equal to the synthesized aperture. The
wavefield from different scatterers reaches the sensor with delays as given by
(2.78). The receiver output is given by a convolutional integral [22],

© 2001 by CRC Press LLC

 ∞
2 l02 + (x − x ′ )2
f (x,t) = ∫
−∞
r0 (x' )w(x − x ′ )cos(ω(t −
c
) + ϕ)dx ′

(2.81)

L
where w(x − x ′ ) =1 for x − x ′ ≤ and L=p1 p 2 , the length of the synthetic
2
aperture. For x − x' << l0 we can approximate (2.81) as

p2
(x − x' )2
f (x,t) = ∫ r0 (x' )w(x − x ′ )cos(ω(t − τ 0 − ) + ϕ)dx ′
p1
cl0
(2.82a)

which simplifies to

∞
(x − x' )2 ϕ
f (x,t) = cos(ω(t − τ 0 )) ∫ r0 (x' )w(x − x ′ )cos(ω( ) + )dx ′
−∞
cl0 2
∞
(x − x' )2 ϕ
+ sin(ω(t − τ 0 )) ∫ r0 (x' )w(x − x ′ )sin(ω( ) + )dx ′
−∞
cl0 2
(2.82b)

We can recover r 0 (x) from f (x, t) , using the Fourier transform method, that
is,

 

−1 
r̃0 (x) = FT 
FT f̂ (x) { } 

 (2.83)
 FT w(x)cos(ω( x )) 
2

  cl0  

where

f̂ (x) = ∫ f (x,t)cos(ω(t − τ
one period
0 ))dt

© 2001 by CRC Press LLC

and FT stands for the Fourier transform. For a large aperture, we get
r̃ 0 (x) L→∞
→ r 0 (x) , that is, an exact reconstruction. It is important to
remember that the increased aperture and therefore the increased resolution is the
p2 p1

Figure 2.31: At a fixed place and time the signal received consists of a sum of
scattered wavefields from a patch of scatterers which is coherently illuminated. The
width of the patch is equal to the synthesized aperture.

result of geometry of the data collection as shown in fig. 2.30. Indeed, the
transceiver need not move at all during pulse transmission and reception.

§2.5 Exercises:
1. The angle of arrival of a broadband signal at ULA may be estimated from the
spectrum of the sum of the output of all sensors (without delays) and the
spectrum of the output of anyone sensor. Find the bandwidth required to
estimate the angle of arrival equal to 45o given that c=1500 m/s and d=15m.
What is the role of the sensor spacing?
2. It is desired to measure the speed of wave propagation in a medium. A ULA
with sensor spacing d meters is employed for this purpose. A broadband signal
from a known direction is sweeping across the array. How do you go about
estimating the speed of propagation?
3. Consider a circular transmit and receive array as shown in fig. 2.32. (a) Show
the sum coarray along with weight coefficients assuming that the physical array
has unit coefficient. (b) Do the same for the difference coarray.
4. Compare the outputs of crossed electric dipole arrays given by (2.58 & 2.59)
with that of displaced identical sensors pairs (also known as dipoles) given by
equations (2.55 a & b). Obtain an expression for the spectral matrix, similar to
the one given in (2.57), of the output of the crossed dipole arrays.

© 2001 by CRC Press LLC

 Sensor Source

Figure 2.32: A circular sensor and source array. Compute sum and difference coarray

a b

Figure 2.33: Triangular (equilateral) array and a source.

5. Consider an equilateral triangular array inscribed in a circle of radius=64 λ . A

point source is located at r=32 λ and ϕ =45 o (see fig. 2.33). Draw all phase
trajectories passing through the point where the source is located.
6. A ULA is placed on a 100m deep sea bed and a broadband acoustic source is
placed 1000m away from the array and 50m below the sea surface. Sketch the
frequency wavenumber spectrum of the array signal. [Hint: Consider a direct and
two reflected wavefronts.]
7. A broadband wavefront is incident on a ULA. Show that there is no aliasing
λ
for any angle of incidence if the sensor spacing is ≤ min 2 .

© 2001 by CRC Press LLC

 References

1. P. S. Naidu, Modern spectrum analysis of time series, CRC Press, Boca

Raton, Fl, 1996.
2. J. R. Williams, Fast beam-forming algorithm, J. Acoust. Soc. of Am., vol.
44, pp. 1454-1455, 1968.
3. S. H. Nawab, F. U. Dowla, R. T. Lacoss, Direction determination for wide
band signals, IEEE Trans, ASSP-33, pp. 1114-1122, 1985.
4. M. Allam and A. Moghaddamjoo, Two dimensional DFT projection for
wideband direction-of-arrival estimation, IEEE Trans. on Signal Proc., vol. 43,
pp. 1728-1732, 1995
5. R. H. Tatham, Multidimensional filtering of seismic data, IEEE Proc., vol.
72, pp. 1357-1369, 1984.
6. H. Brysk and D. W. McCowan, Geophysics, vol. 51, pp. 1370-1386, 1986.
7. C. W. Therrien, Discrete random signals and statistical signal processing,
Prentice Hall, Englewood Cliffs, NJ , 1992.
8 A. T. Moffet, Minimum redundancy linear arrays, IEEE Trans. Antenna
Propagation, vol. AP-16, pp. 172-175, 1968.
9 Lee and S. U.Pillai, An algorithm for optimal placement of sensor elements,
ICASSP-88, pp. 2674-2677, 1988.
10N. M. J. Earwicker, Signal to noise ratio gain of sparse array processors, J.
Acoust. Soc. of Am., vol. 88, pp. 1129-1134, 1980.
11. B. G. Quinn, R. F. Barrett, P. J. Kootsookos and S. J. Searle, The
estimation of the shape of an array using a hidden markov model, IEEE Jour.
Ocean Engineering, vol. 18, pp. 557-564, 1993.
12. M. Hinich, Estimating signal and noise using a random array, J. Acoust.
Soc. of Am., vol. 71, pp. 97-99, 1982.
13. M. P. Moody, Resolution of coherent sources incident on a circular antenna
array, Proc. IEEE, vol.68, pp. 276-277, 1980.
14. C. P. Mathews and M. D. Zoltowaski, Eigen structure technique for 2D
angle estimation with uniform circular arrays, IEEE Trans. SP-42, pp. 2395-
2407, 1994.
15. M. Abramowitz and I. A. Stegun (Editors), Handbook of Mathematical
Functions, National Bureau of Standards, Applied Mathematics Series 55,
1964.
16. A. H. Tewfik and H. Wong, On the application of uniform linear array
bearing estimation techniques to uniform circular arrays, IEEE Trans. SP-40,
pp. 1008-1011, 1992.
17. A. G. Peirsol, Time delay estimation using phase data, IEEE Trans.,
ASSP-29, pp.471-477, 1981.
18. C. U. Padmini and P. S. Naidu, Circular array and estimation of direction
of arrival of a broadband source, Signal Processing, vol. 37, pp. 243-254,
1994.

© 2001 by CRC Press LLC

 19. R. J. Kozick, Linear imaging with source array on convex polygonal
boundary, IEEE Trans on System, Man and Cybernetics, vol.21, pp. 1155- ,
1991.
20. R. J. Kozick and S. A. Kassam, Coarray synthesis with circular and
elliptical boundary arrays, IEEE, Trans on Image Processing, vol. 1, pp. 391-
404, 1992.
21. R. Williams and Bernard Harris, Passive acoustic sythetic aperture
processing techniques, IEEE Jour. of Ocean Eng., vol 17, pp. 8-14, 1992.
22. D. C. Munson, Jr., An introduction to strip-mapping sythetic aperture
radar, IEEE, Proc of ICASSP, pp. 2245-2248, 1987.

© 2001 by CRC Press LLC

 Chapter Three

Frequency Wavenumber Processing

In signal processing, extraction of a signal buried in noise has been a primary

goal of lasting interest. A digital filter is often employed to modify the
spectrum of a signal in some prescribed manner. Notably the proposed filter is
designed to possess unit response in the spectral region where the desired signal
is residing and low response where the undesired signal and noise are residing.
This strategy will work only when the spectrum of the desired signal does not
overlap (or only partially overlapping) with the spectrum of the undesired signal
and noise. This is also true in wavefield processing. Spectrum shaping is
necessary whenever the aim is to enhance certain types of wavefields and
suppress other types of unwanted wavefields, often termed as noise or
interference. Such a selective enhancement is possible on the basis of spectral
differences in the frequency wavenumber (ω-k) domain. For example, it is
possible, using a digital filter, to enhance the wavefield traveling at a speed
different from that of the noise or the interference as their spectra lie on different
radial lines in the (ω-k) plane. Other situations where spectral shaping is
required are (a) Prediction of wavefields, (b) Interpolation of wavefields, and (c)
Extrapolation of wavefield into space where the field could be measured. In this
chapter we shall consider the design of pass filters, specially useful in wavefield
processing such as a fan and a quadrant filter. When signal and noise are
overlapping in the frequency wavenumber domain simple pass filters are
inadequate. Optimum filters such as Wiener filters are required. We cover this
topic in some depth in view of its importance. Next, we introduce the concept
of noise cancellation through prediction. Later, in chapter 6 we shall evaluate
the effectiveness of some of the techniques described here.

§3.1 Digital Filters in the ω -k Domain:

In wavefield processing digital filters are often used to remove the interference
corrupting the signal of interest (SOI) which arrives from one or more known
directions while the interference arrives from different directions, which are often
unknown. The purpose of a digital filter is to accept the wave energy arriving
from a given direction or an angular sector and to reject all wave energy
(interference) lying outside the assumed passband which consists of a vertical
cone symmetric about the temporal frequency axis. Such a filter is known as a
fan filter in seismic exploration (fig. 3.1) or as a quadrant filter (fig. 3.6) when
the axis is tilted. In this section we shall study in detail the fan and the quadrant
filters. The issues connected with the design and implementation will be
discussed. Finally, we shall also examine the role of sampling and the
associated question of aliasing.

© 2001 by CRC Press LLC

 3.1.1 Two Dimensional Filters: Two dimensional filters are extensively used in
image processing [1], and geophysical map processing [2]. However, because of
the fundamental differences between the wavefield and the image signal, the 2D
filters used in the respective applications are of different types. For example, in
wavefield processing the filter must satisfy the causality condition in time but
there is no such requirement in image processing. The causality requirement is
waived whenever delay in the output can be tolerated or the data is prerecorded as
in seismic exploration or nondestructive testing. Another important difference is
that the spectrum of a propagating wavefield is necessarily confined to a conical
domain in the frequency wavenumber space (see chapter 1). Only locally
generated disturbances (evanescent waves) and instrument generated noise are not
subject to the above restriction. Hence, in wavefield applications a typical
lowpass filter has a shape of a hand-held fan. Such a filter is known as a fan
filter, widely used in seismic data processing. The design and implementation of
the fan filter is naturally of great interest in wavefield processing. As for other
types of filters, such as lowpass filters with circular or elliptical passbands, a
ring type bandpass filter, Laplacian filter for edge detection, deblurring filter,
etc., which are widely used in image processing, they are outside the scope of
this book.
3.1.2 Fan Filters: The goal is to extract a broadband signal traveling with a
c
horizontal apparent speed between cHor and ∞ (note cHor = ). The
sin θ
desired fan filter will have the following transfer function:

 ω < ω max

H(s,ω) = 1  ω
s < c (s = u 2 + v 2 ) (3.1)
 Hor

=0 otherwise

which is illustrated in fig. 3.1. The filter will pass all propagating waves whose
c
angle of elevation lies between 0o and ± sin −1 ( ) (the elevation angle is
cHor
measured with respect to the vertical axis). Fail and Grau [3] were the first to
introduce the fan filter in 1963. Independently, Embree [4] came up with a
similar filter which he called a “Pie Slice Filter”. Since the filter has a circular
symmetry in a spatial frequency plane we need a series of concentric uniform
circular arrays (in effect, a planar array) to measure the direction of arrival in the
horizontal plane and horizontal apparent speed, when the wave speed is known.
For the purpose of filter design it is enough if we consider any one radial line.

© 2001 by CRC Press LLC

 ω

Slope= c hor

0 u

Figure 3.1: Fan filter. The passband is a cone with an axis along the ω-axis and the
apex is at the center. All waves whose horizontal apparent speed lies inside the cone
are allowed and others are attenuated.

The filter may be rotated about the vertical axis to yield a circularly symmetric
filter. For simplicity we consider a single radial line of sensors coinciding with
the x-axis and assume that all waves are propagating along the x-axis with
different horizontal apparent speeds.
In order to compute the fan filter coefficients we take the inverse
Fourier transform of the transfer function given in (3.1),

+∞
1
∫ ∫ H(ω,u)e
j (ωn∆t +um∆x )
hm,n = 2 dωdu
4π −∞

© 2001 by CRC Press LLC

 ω
+ω max c Hor
1
∫ ∫e
j (ωn∆t +um∆x )
= dωdu
4π 2 −ω max ω
−
c Hor
(3.2)
ω
+ω max sin(m∆x )
1 cHor jωn∆t
=
2π 2 ∫
−ω max
m∆x
e dω

where ∆t and ∆x are temporal and spatial sampling intervals, respectively.

ω
Since sin(m∆x ) is a symmetric function, the last integral in (3.2) can be
cHor
simplified as

hm,n =
+ω max
1  m∆x m∆x 
2 ∫ sin( c − n∆t)ω + sin( + n∆t)ω  dω (3.3)
2π m∆x 0  Hor cHor 

We are free to choose the spatial and temporal sampling intervals but within the
constraints imposed by the sampling theorem. These are discussed in §3.2. Let
∆x
the temporal sampling interval ∆t be so chosen that ∆t = . In practice
cHor
∆x is held fixed; therefore, to alter cHor it is necessary to resample the whole
signal with a different sampling interval, consistent with the above choice.
Equation (3.3) is easily evaluated to yield

1 1 − cos(m − n)π 1 − cos(m + n)π 

hm,n = +
2π m∆x∆t   (3.4)
2
(m − n) (m + n) 

Equation (3.4) takes particularly a simple form if we shift the origin to the half-
way point between two sensors

1 1
hm,n = (3.5)
π ∆x∆t m − n 2
2 2

1 1 1
where m = ± , ±1 , ±2 ,... (see fig. 3.2). Note that since m ≠ n , hm,n
2 2 2

© 2001 by CRC Press LLC

 t

∆x
1 1 1 1
-1 - m= 1
2 2 2 2

Figure 3.2: Centered array (ULA) with origin lying between two sensors. This is also
the physical center of the array.

will always be finite. The frequency response (magnitude) of a fan filter given
by (3.5) is shown in fig. 3.3
3.1.3 Fast Algorithm: In order to implement the filter the following
convolution sum will have to be evaluated

f̂ (n∆t) = ∑ ∑ hm, n ′ f (m∆x,(n − n′ )∆t) (3.6)

n′ m

The computational load for evaluating the convolutional sum can be greatly
reduced by exploiting the filter structure [5]. We define a new index
µ = (1 + 2 m ′ ) where m ′ = 0, ±1, ±2,... . We note that the new index takes
values, µ = ±1, ± 3, ± 5, ... , for different values of m ′ . Equation (3.5) may
be reduced to

 
1  1 
hm,n = 2 2 
π µ
 − n2 
 4 

which may be further split into two components

© 2001 by CRC Press LLC

 3 0.3
0.5

0.7
2

0.3 0.1 0.05

0
ω

-1

-2

-3
-3 -2 -1 0 1 2 3
u
Figure 3.3: Frequency response (magnitude) of a fan filter (64 sensors and 64 time
samples).

 
1  1 1 
hm,n = 2 +
π µ  µ − n µ + n 
2 2 
1
[
= rµ ,n + qµ ,n
µ
]
The components rµ ,n and qµ ,n possess many symmetry properties as listed
below:

µ
• rµ ,n and qµ ,n are antisymmetric in time index (n) about a point n = and
2
µ
n=− , respectively.
2

© 2001 by CRC Press LLC

• rµ ,n and qµ ,n are of the same shape as r1,n except for the time shift.
rµ ,n = r µ −1 and qµ ,n = −r µ +1 .
1,n − 1,n +
2 2
• qµ ,n = −rµ ,n + µ .

Using the last property we obtain hm,n =

1
µ
[ ]
rµ ,n − rµ ,n + µ and using the
second property we obtain

1  
hm,n =  r µ −1 − r µ +1  (3.7)
π µ  1,n − 2
2
1,n +
2 

Substituting (3.7) in (3.6) we obtain

1  
f̂ (n∆t) = ∑ ∑ 2 
r µ −1 − r µ +1  f (µ∆x,(n − n′ )∆t)
n′ µ π µ  1, n ′ − 2 1, n ′ +
2 

 f (µ∆x, (n − n′ − µ − 1 )∆t)  (3.8)

1 1 l
2 
= 2 ∑ r1, n ′ ∑  
π n′ µ = −l µ  µ +1
− f (µ∆x, (n − n′ + )∆t)
 2 

where l=M-1 and M (an even number) stands for the number of sensors. In (3.8)
there is only one convolution to be carried out in place of M/2 convolutions in
(3.6). The inner summation in (3.8) stands for spatial filtering and the outer
summation stands for temporal filtering. Let the output of the spatial filter be
f 1 (n∆t) , where

 f (µ∆x, (n − n′ − µ − 1 )∆t) 
1
l
2 
f 1 (n∆t) = ∑  
µ = −l µ  µ −1
− f (µ∆x, (n − n′ + )∆t)
 2 

Equation (3.8) takes the form

f̂ (n∆t) = ∑ r1, n ′ f 1 ((n − n′ )∆t) (3.9)

n′

© 2001 by CRC Press LLC

 Taking z transform on both sides of (3.9) we shall obtain

F̂(z) = R1 (z)F1 (z)

where F̂(z) , F1 (z) , R1 (z) are z transforms of f̂ (n), f (n),r1,n ,

respectively. In particular,

N1
1
R1 (z) = ∑ 1
zn
n = − N1 +1 ( − n)
2
N1 −1 N1 −1
1 1
= −z ∑ zn + ∑ z−n (3.10)
1 1
n=0 ( + n) n=0 ( + n)
2 2
= −zRN1 (z) + RN1 (z −1 )

where
N1 −1
1
RN1 (z) = ∑ 1
zn
n=0 ( + n)
2

and the number of time samples are assumed to be 2N1 . Note that RN1 (z)
−1
acts on future time and RN1 (z ) acts on past time. We assume that the data are
pre-recorded; hence the question of causality is irrelevant. It is shown in [5] that
as N1 → ∞ , RN1 (z) may be approximated by a stable pole-zero filter, in
particular,

2(1 − 0.65465z)
RN1 (z) →
1 − 0.98612z + 0.13091z 2 (3.11)
N1 → ∞

A comparison of the filter response N1=64 and N1 → ∞ is shown in fig. 3.4.

A block diagram for fast realization of the fan is shown in fig. 3.5. The
realization consists of two parts, the temporal processor (upper part) and the
spatial processor (lower part). The temporal part has two pole-zero filters, and
one acting on past time and the other on future time. A pole-zero filter can be
implemented recursively; for example, the left arm of the temporal processor
may be implemented via a difference equation given below

© 2001 by CRC Press LLC

 7

5
Magnitude

1
0 0.5 1 1.5 2 2.5 3
radians

Figure 3.4: A comparison of the filter response N1 =64 (dashed line) and N1 →∞
(solid line).

f̂ l (n∆t) − 0.98612 f̂ l ((n + 1)∆t) + 0.13091 f̂ l ((n + 2)∆t)

= 2 f 1 ((n + 1)∆t) − 1.3093 f 1 ((n + 2)∆t)
and the right arm via

f̂ r (n∆t) − 0.98612 f̂ r ((n − 1)∆t) + 0.13091 f̂ r ((n − 2)∆t)

= 2 f 1 (n∆t) − 1.3093 f 1 ((n − 1)∆t)

The output of the temporal processor is finally given by

f̂ (n∆t) = f̂ l (n∆t) + f̂ r (n∆t) .
3.1.4 Quadrant Filter: The fan filter is a half-plane filter; hence it is insensitive
to the direction of propagation as long as the horizontal apparent speed lies
within specified limits. A wave traveling with apparent speed cHor cannot be
distinguished from a wave traveling with apparent speed of −cHor . For such a
discriminatory property the desired filter must possess a passband in only one
quadrant, either the 1st or 2nd quadrant. Note that for the filter to be real the
passband must be reflected diagonally into the opposite quadrant. An example of
such a passband is shown in fig. 3.6. We will call such a filter a quadrant filter.

© 2001 by CRC Press LLC

 2/ π 2

z +
Temporal
Processor

+
1/3 1 -1 -1/3

Spatial
Processor

-
++ -
++ -
++ -
++
-2 +1 -1 0 0 -1 +1 -2
z z z z z z z z

-3 -1 µ= 1 3

Figure 3.5: A block diagram of fan filter implementation. The filter structure has two
parts, namely, spatial and temporal parts, which are independent of each other.

© 2001 by CRC Press LLC

 ω ω=bu
ω= ωmax
ωmax

ω =au

- ωmax

Figure 3.6: Quadrant filter in ω-u plane.

The passband is bounded with two radial lines with slopes a and b respectively.
Further the passband is terminated by two horizontal lines ω = ±ω max where
ω max is cut-off temporal frequency. When b > 1 and a < 1 the passband will
be terminated by two cut-off lines as shown in fig. 3.6. Draw a diagonal line
and divide the passband into two triangles. Straight forward integration yields
m ∆x
the filter impulse response function for m ≠ 0, ( + n) ≠ 0 and
a ∆t
m ∆x
( + n) ≠ 0 [6].
b ∆t

ω max  ωa 
1  
hm,n =
4π 2 ∫  ∫ 2 cos(m∆xu + n∆tω)du  dω
0
 ωb 

© 2001 by CRC Press LLC

  (1 − cos( m ∆x + n)∆tω ) 
 a ∆t
max 
 m ∆x 
 +n 
1  a ∆t 
= 2 (3.12a)
2π m∆x∆t  m ∆x 
 (1 − cos( b ∆t + n)∆tω max ) 
− m ∆x 
 +n 
 b ∆t 

The special cases are given by

(b − a)ω 2max
h0,0 = m=n=0
4π 2 ab

(b − a) nω max ∆t sin(nω max ∆t)

hm,0 = m = 0, n ≠ 0
2π 2 n 2 ∆t 2 ab  + cos(nω max ∆t) − 1 

m ∆x
(1 − cos( + n)ω max ∆t)
−1 b ∆t m ∆x
hm,n = ( + n) = 0, n ≠ 0,
2π 2 m ∆x a ∆t
n∆x∆t( + n)
b ∆t
m ∆x
( + n) ≠ 0
b ∆t
m ∆x
(1 − cos( + n)ω max ∆t)
1 a ∆t m ∆x
hm,n = ( + n) = 0, n ≠ 0,
2π 2 m ∆x b ∆t
n∆x∆t( + n)
a ∆t
m ∆x
( + n) ≠ 0
a ∆t
(3.12b)

∆x
We now set ∆x =1 and ∆t =1, hence =1. Using these settings (3.12a)
∆t
1
reduces to equation (6) in [6]. The highest temporal frequency is f hi = Hz (or
2
ω hi = π in angular frequency) and the corresponding lowest wavelength,

© 2001 by CRC Press LLC

 3

2 0.9
0.1

1
Temporal Frequency

0.5
0
0.5

-1
0.1

-2
0.9

-3
-3 -2 -1 0 1 2 3
Spatial frequency

Figure 3.7: Frequency response of a quadrant filter (a=1, b=2). 16 sensors and 64
time samples are assumed. The maximum temporal frequency is 0.8π. The contour
values are as shown in the figure.

λ lowest = 2 . The sensor spacing will be one. In real terms, consider a signal
with f hi = 1500Hz propagating underwater where the wave speed is 1500
m/s. Since the wavelength is 1 meter, the sensor spacing is ∆x = 0.5 meter.
1
The sampling interval is ∆t = ms. Let us redefine units of time and distance.
3
1
Let one unit of time be ms and one unit of distance be 0.5 meter. Here
3
onwards we shall measure the time and the distance in terms of these newly
defined units. The response function of a quadrant filter for parameters listed
below is shown in fig. 3.7 Parameters: 64 sensors, 64 time samples,
ω max =0.8π, a=1 and b=4. In (3.12a) if we let a=-b=1, cHor = 1 , ∆t = 1, and

© 2001 by CRC Press LLC

 No. of Sensors rms error maximum error
4 0.1183 0.9959
8 0.0775 0.9958
16 0.0418 0.9912
32 0.0240 0.9743
64 0.0153 0.9303
Table 3.1: The number of temporal samples is fixed at 64. While the rms error
decreases with increasing number of sensors the maximum error shows only
marginal decrease.

ω max =π the resulting filter becomes a fan filter with its axis turned 90o and the
filter coefficients are

1  (1 − cos(m + n)π) (1 − cos(m − n)π) 

hm,n = +
2mπ 2  m+n m−n 

which indeed are the same as in (3.4).

3.1.5 Weighted Least Squares Filter: The filter size controls the root mean
square (rms) difference between the ideal infinite length filter and finite length
filter. In general, the rms error decreases with increasing filter size but only
slowly as revealed in table 3.1 where we considered a discrete fan filter whose
ideal response is one in the region bounded by two rays with slopes 2 and 8 and
zero outside. The maximum temporal frequency is set at 78% of the Nyquist
frequency. The impulse response function of the quadrant filter is computed by
taking the inverse discrete Fourier transform of the ideal quadrant filter embedded
in a large matrix (256x256) of zeros. The impulse response function is then
truncated to 64 temporal samples and a variable number of spatial samples: 4,
8, 16, 32, and 64. Both root mean square (rms) error and maximum difference
between the ideal filter (frequency) response and truncated filter (frequency)
response, which is now interpolated to the same grid as the ideal impulse
response, are shown in table 3.1.
The response in the reject region, where the ideal response is zero, also
decreases with increasing filter size. From a practical point of view, for a given
filter size, the response in the pass region must be as close to unity as possible
while the response in the reject region or at least in some identified parts of the
reject region must be close to zero. This goal can be achieved by using a
weighted least squares approach [7]. Let W(ω,u) be a weighting function which
takes a large value in the reject region, where it is desired to have a response
close to zero. Elsewhere, W(ω,u)=1. The filter coefficients are obtained by
minimizing the following quantity:

© 2001 by CRC Press LLC

 π π 2
∆x ∆t
1
4π 2 ∫ ∫ W(ω ,u) H(ω ,u) − ∑ ∑ h
π π m n
mn e − j ( ωn∆t −um∆x ) dω du = min
− −
∆x ∆t
(3.13)

Minimization with respect to the filter coefficients leads to the following

normal equations:

∑∑h
m′ n′
r
mn m − m ′ ,n − n ′ = gmn (3.14a)

where

π π
∆x ∆t
1
∫ ∫ W(ω ,u)e
+ j ( ωn∆t −um∆x )
rm,n = dω du
4π 2 π π
− −
∆x ∆t

and

π π
∆x ∆t
1
∫ ∫ W(ω ,u)H(ω ,u)e
+ j ( ωn∆t −um∆x )
gm,n = dω du
4π 2 π π
− −
∆x ∆t

Note that, when W(ω,u)=1 over the entire plane, (3.14a) reduces to the expected
result, namely,

π π
∆x ∆t
1
∫ ∫ H(ω ,u)e
+ j ( ωn∆t −um∆x )
hm,n = dω du
4π 2 π π
− −
∆x ∆t

In order to understand the role of the weighting function, consider (3.14a) with
finite limits on the summation signs

M −1 N −1

∑ ∑h
m′ =0 n′ =0
r
mn m − m ′ ,n − n ′ = gmn (3.14b)

and take a finite Fourier transform on both sides of (3.14b). We obtain

© 2001 by CRC Press LLC

 [ H ] finite [W ] finite = [WH ] finite (3.15)

M −1 N −1
where [ F ] finite = ∑ ∑ f mne − j ( ωm +un) stands for the finite discrete Fourier
m=0 n=0
2π 2π
transform (DFT) for ω= k and u = l . Note that [ F ] finite → F as
M N
M → ∞ and N → ∞ . From (3.15) we obtain

[WH ] finite
[ H ] finite = (3.16)
[W ] finite

Although WH is a bandpass filter (because H is a bandpass filter) [WH ] finite

will not be zero outside the passband. The presence of W in the denominator
will however greatly reduce (if W>>1) the out of band magnitude of
[WH ] finite . Hence, [ H ] finite on the left hand side of (3.16) will have low
magnitude outside the passband. In fig. 3.8 we illustrate this phenomenon. A
quadrant filter with unit response in a region bounded by two radial lines with
slopes 3 and 8 and with maximum temporal frequency 0.78 times the Nyquist
frequency is considered. We selected a weighting function whose value is six in
a fan-shaped region bounded by two radial lines with slopes 0.5 and 2 and with
the maximum temporal frequency 0.78 times the Nyquist frequency and equal
to one elsewhere. The impulse response functions of WH and W are now
limited to 64 x16 (64 temporal samples and 16 spatial samples). The finite
impulse functions are next used to compute the weighted least squares filter as
given in (3.16). The plain (unweighted) finite filter response function along
with the region where the weight function is six is shown in fig. 3.8(a) and the
weighted least squares filter response function is shown in fig. 3.8(b). Notice
how a side lobe falling within the region of the high weight function has been
attenuated.
3.1.6 Aliasing Effect in Fan Filter and Quadrant Filter: In array signal
processing there is a basic limitation imposed by discrete sensors.
Consequently, the wavefield is sampled in the spatial dimension but not in the
temporal dimension. It is possible to avoid temporal aliasing through a lowpass
filtering and by sampling according to the sampling theorem, but it is a
different story with the spatial aliasing which is intimately related to the
propagation speed. This phenomenon was briefly discussed in chapter 1 (see p.
22). Here we shall examine the effect of spatial sampling on digital filters for
wavefields, in particular, a quadrant filter whose pass region is defined by two
radial lines with slopes cLo and cHi as shown in fig. 3.9. We further assume
that sensor outputs have been prefiltered to confine the spectra to ±ω max .

© 2001 by CRC Press LLC

 (a)

(b)

Figure 3.8: The role of a weighting function is shown here. (a) unweighted case. (b)
weighted case (equal to six inside a fan shaped region bounded by two radial lines).
Notice how a side lobe falling within the region of high weight function has been
attenuated.

© 2001 by CRC Press LLC

 The sampled version of the quadrant filter will be periodic both in spatial and
temporal frequencies but as the sensor outputs have been prefiltered replication
in the temporal frequency would not cause any aliasing. However, aliasing can
take place due to the replication in spatial frequency. This is shown in fig. 3.9
where we have drawn three replications, that is, three rectangles including the
principal rectangle. Notice the intrusion of a radial line with a slope cLo from
the neighboring rectangles into the principal rectangle. Such an intrusion is the
cause of aliasing. Clearly, the aliasing can be prevented if

ω max ∆x
cLo ≥ = (3.17)
B0 ∆t

A similar requirement was shown (in chapter 1, p. 23) to be necessary to avoid

aliasing of a broadband signal traveling with a horizontal speed cHor .

§3.2 Mapping of 1D into 2D Filters:

As in a one dimensional filter, a sharp transition from passband to rejectband in
a two dimensional filter (in frequency wavenumber space) results in large
ripples in passband and rejectband. To overcome this drawback, a sharp
transition is replaced by a smooth transition, but this will degrade the quality of
the filter. Thus, in the design of a pass filter there is a contradictory
requirement, calling for an optimum design strategy of minimum transition
width while maintaining the ripple height below a prescribed limit. This
optimization problem has been largely solved in one dimension [8] but in two
or higher dimensions the procedure of optimization becomes highly
cumbersome and computationally intensive. It has been found that optimally
designed 1D filters may be mapped into 2D filters having similar transition and
ripple height properties as those of 1D filters [9]. The 2D filters of interest in
the wavefield analysis are fan filter and quadrant filter. The fan filter may be
obtained by transforming an optimally designed 1D filter.
3.2.1 McClellan’s Transformation: Consider a real linear phase FIR (finite
impulse response) filter h0 , h±1 , h±2 ... h± N where for linear phase we must
have hk = h− k . The transfer function of the filter is given by

N
H(ω ′ ) = h0 + 2 ∑ hk cos(ω ′k) (3.18)
k =1
which may be expressed as a polynomial in cos(ω ′ ) ,

© 2001 by CRC Press LLC

 ω

b' a' ωmax

c Hi
c' cLo

- B0 B0 2B
0 u

- ω maxa b

Figure 3.9: The aliasing is caused by intrusion of a radial line with slope cLo from
the neighboring rectangles into the principal rectangle. Triangles abc and a´b ´c´ are
the aliased pass regions. Aliasing can be prevented if we were to sample an analog
quadrant filter according to (3.17).

N
H(ω ′ ) = ∑ bk cos k (ω ′ ) (3.19)
k =0

where the coefficients bk , k=0, 1, ... N are expressed in terms of the FIR filter
coefficients. A point on the frequency axis is mapped onto a closed contour in
the frequency wavenumber space using a linear transformation
P Q
cos(ω ′ ) = ∑ ∑ t pq cos( pω)cos(qu) (3.20a)
p= 0 q = 0

© 2001 by CRC Press LLC

 3

-1

-2

-3
-3 -2 -1 0 1 2 3

Figure 3.10: Mapping produced by (3.21). The inner contours corresponding to low
frequency are close to circular shape but the outer contours corresponding to high
frequency are only approximately circular.

or
ω′ P Q
ω u
sin 2 ( ) = ∑ ∑ t pq ′ sin 2 p ( )sin 2q ( ) (3.20b)
2 p= 0 q = 0 2 2

where ′ are yet to be determined coefficients. The shape of the

t pq and t pq
′ . For example, we obtain an
contour depends upon the coefficients t pq or t pq
approximate circular contour (see fig. 3.10) for

1
t00 = −
2
1
t01 = t10 = t11 =
2

© 2001 by CRC Press LLC

The mapping relation (3.20a) takes a form

1
cos(ω ′ ) = {−1 + cos(u) + cos(ω) + cos(u)cos(ω)} (3.21)
2

Contours corresponding to a set of fixed values of cos(ω ′ ) are shown in fig.

3.10. Circular approximation is better in the low frequency range. Using
(3.20a) in (3.19) we obtain a 2D filter transfer function,

k
P Q
N 
H(ω,u) = ∑ bk  ∑ ∑ t pq cos( pω)cos(qu) (3.22)
k =0  p= 0 q = 0 

The mapping function (3.20a) must satisfy the following conditions:

P Q
a) When ω ′ =0, ω = u = 0 . This requires ∑∑t
p= 0 q = 0
pq = 1. The

condition ensures that a lowpass filter remains a lowpass filter even after the
transformation, that is, H(0,0)=H(0).
P Q
b) When ω ′ =π, ω = u = π . This requires ∑∑t
p= 0 q = 0
pq (−1) p+ q = −1.

The condition ensures that a highpass filter remains a highpass filter after
transformation, that is, H(π,π)=H(π).
P
c) When ω =0, H(0,u) = H(u) . This requires ∑t
p= 0
p1 = 1 and
P

∑t
p= 0
pq = 0 for q ≠ 1 .
Q
d) When u =0, H(ω,0) = H(ω) . This requires ∑t
q=0
1q = 1 and
Q

∑t
q=0
pq = 0 for p ≠ 1 .

For circularly symmetric mapping, conditions (b) and (c) must hold good. For
P=Q=1, a general solution for the transformation coefficients satisfying the
1 1
above conditions is given by t00 = − , t01 = t10 = t11 = . By relaxing (b),
2 2
the solution is given by

© 2001 by CRC Press LLC

 t00 = −a, t01 = t10 = a, t11 = 1 − a (3.23)

having one free constant which may be optimally chosen for the best fit.
The next important question is to determine the coefficients bk ,
k=0,1,2,..., N in (3.19) given the coefficients of a digital filter. Let
h0 ,h1 ,h2 ,...hN be the given 1D filter coefficients.
N
H(ω ′ ) = ∑ h̃k cos(kω ′ )
k =0
(3.24)
N
= ∑ h̃k T k (cos(ω ′ ))
k =0

where h̃0 = h0 , h̃k = 2hk , k = 1,..., N , and T n (cos(ω ′ )) is a Chebyshev

polynomial which may be expressed as a polynomial in its argument,

k
T k (cos(ω ′ )) = ∑c
m=0
k
m cos m (ω ′ ) (3.25)

Using (3.25) in (3.24) and comparing with (3.19) we obtain the following
system of equations:

N N N
b0 = ∑ c0n h̃n , b1 = ∑ c1n h̃n , ... bk = ∑ ckn h̃n , ... bN = cNN h̃N
n=0 n =1 n=k

n
The coefficients ck are listed in [10, p. 795]. Finally, the 2D filter transfer
function for the case P=Q=1 can be written as

{t00 + t01 cos(u) +

k
N

H(ω,u) = ∑ bk   (3.26a)
k =0 t10 cos(ω) + t11 cos(u)cos(ω)}

Let F(ω,u) be the input and G(ω,u) be the output of a 2D filter,

t00 + t01 cos(u) + t10 cos(ω)

N k

G(ω,u) = ∑ bk   F(ω,u)
k =0  +t11 cos(u)cos(ω)  (3.26b)
N
= ∑ bk [ H0 (ω,u)] F(ω,u)
k

k =0

© 2001 by CRC Press LLC

which may be written in a recursive fashion. Let

F0 (ω ,u) = F(ω ,u)

F1 (ω ,u) = [ H0 (ω ,u)]F0 (ω ,u)
F2 (ω ,u) = [ H0 (ω ,u)]F1 (ω ,u)
... (3.26c)
...
...
FN (ω ,u) = [ H0 (ω ,u)]FN −1 (ω ,u)

whereH0 (ω ,u) = t00 + t01 cos(u) + t10 cos(ω ) + t11 cos(ω )cos(u) . Note
that F0 (ω,u), F1 (ω,u), ..., FN (ω,u) do not depend upon the filter
coefficients but only on mapping coefficients and the input. Using (3.26c), the
filter output may be written as

N
G(ω,u) = ∑ bk Fk (ω,u) (3.27)
k =0

In filter implementation it is possible to generate F0 (ω, u), F1 (ω, u), ...,

F N (ω, u) in a recursive manner and then combine them after weighting each
with bk . The filter structure is illustrated in fig. 3.11.
3.2.2 Fan Filter: To get fan-shaped contours the required mapping coefficients
are

1
t00 = t11 = 0 and t01 = −t10 = (3.28a)
2
and the mapping function is given by [11]

1
cos(ω ′ ) = [cos(ω) − cos(u)] (3.28b)
2
The contours generated by (3.28b) are shown in fig. 3.12. The zero valued
π
diagonal contour corresponds to . When ω ′ = 0 , ω = 0 and
ω′ = ±
2
u = ±π ; and when ω ′ = ±π , ω = ±π and u = 0 . The positive contours

© 2001 by CRC Press LLC

 Figure 3.11: A recursive implementation of a 2D filter obtained by transforming a
1D filter. H0 (ω ,u) solely depends upon the mapping coefficients while the
coefficients and bk 's depend on 1D filter coefficients.

π π
correspond to 0 ≤ ω′ ≤ and the negative contours to < ω′ ≤ π .
2 2
π
Ideally, a lowpass filter having a unit response in the range ± and zero outside
2

© 2001 by CRC Press LLC

 250
+
.8
200

150 .0 -.8

- -

100

50

+
0
0 50 100 150 200 250

Figure 3.12: Contours generated by a mapping function given by (3.28a). The two
diagonals make up the zero contour which divides the display area into four
π
triangles. The positive contours correspond to 0 ≤ ω′ ≤ and the negative
2
π
contours to < ω′ ≤ π .
2

when mapped into two dimensions using the mapping function (3.28b) will
result in a fan filter with a unit response in the top and bottom triangles and a
zero response in the left and right triangles. The value of the 2D filter transfer
function may be found by first computing the frequency corresponding to the
−1
index of each contour, that is, ω ′ = cos (contour index) and then evaluating
the 1D filter response at the desired frequency.
It is interesting to note that the mapping coefficients given by (3.28a)
do not satisfy any of the conditions listed on page 167. This is expected as the

© 2001 by CRC Press LLC

 mapping function does not satisfy circular symmetry nor is it necessary to map
a lowpass filter into a lowpass filter or a highpass filter into a highpass filter.
A fan filter in 3D can also be obtained by transformation of a 1D filter. The
transformation function is given by [12]

1 1 1 1
cos(ω ′ ) = cos(ω) − cos(u) − cos(v) + (3.29)
2 2 2 2
§3.3 Multichannel Wiener Filters:
In Wiener filtering the goal is to make the filter output as close as possible, in
the least square sense, to a desired signal. The sensor output and the desired
signal are assumed to be stationary stochastic signals which are characterized
through their covariance functions. The Wiener filters are known after Norbert
Wiener who did pioneering work on the prediction of a trajectory of a moving
object from its past observations [13]. A recursive algorithm for the solution of
a discrete version of the Wiener-Hopf equation was developed by Levinson [14]
and by Durbin [15] in the context of time series model fitting. Multichannel
extension was made by Wiggins and Robinson [16]. In this section we briefly
describe the Wiener filter as applied to array signals where we like to extract a
signal traveling in some known direction and to optimally suppress all other
propagating waves and noise. Here, a straightforward solution of the Wiener-
Hopf equation requires invertion of a large block toeplitz covariance matrix,
leading to a dramatic increase in the computational load over the single time
series version. Therefore, it is worthwhile to spend some effort to understand
the principles of the Levinson-Durbin recursive algorithm for the multichannel
Wiener filter [16].
3.3.1 Planar Array: We consider a planar array of sensors, not necessarily
uniformly distributed (see fig. 3.13). Let f p (t) be the output of the pth sensor
located at (x p , y p ) . We shall model the output as a sum of two random
processes, namely, a desired signal, ξ p (t) , and unwanted noise, η p (t) ,

f p (t) = ξ p (t) + η p (t), p = 0,1,... P − 1 (3.30)

A multichannel filter hp (t), p = 0,1,... P − 1 is sought such that the

output, as given by (3.31), is closest to the signal at one of the sensors, for
example, at p=0. The filter output, given by,

P −1 ∞
f̂ (t) = ∑ ∫ hp (τ) f p (t − τ)dτ (3.31)
p= 0 0

© 2001 by CRC Press LLC

 x P-1y P-1

x 1y 1

x 2y 2

x 0y 0

Figure 3.13: A distribution of sensors in a planar array. All sensors are identical but
for their position.

must be closest to ξ 0 (t) in the sense

{
E f̂ (t) − ξ 0 (t)
2
} = min
that is, the mean square error (mse) is minimum. This requires minimization of
an expression for mse,

P −1 ∞
mse = cξ 0 (0) − 2 ∑ ∫ hp (τ)cξ 0 f p (τ)dτ
p= 0 0
(3.32)
P −1 P −1 ∞ ∞
+∑ ∑ ∫ ∫ h (τ)h (τ′)[c
p= 0 q = 0 0 0
p q ξ pξq ]
(τ − τ ′ ) + cη p η q (τ − τ ′ ) dτdτ ′

We shall minimize (3.32) with respect to hp (τ) . So we need to differentiate

(3.32) with respect to hp (τ) which occurs inside an integral. To see how such

© 2001 by CRC Press LLC

 a differentiation may be carried out let us express the integral as a limiting sum.
For example,
∞ ∞

∫ hp (τ)cξ 0 ξ p (t − τ)dτ → ∆τ∑ hp (n∆τ)cξ 0 ξ p (t − n∆τ)

0 n=0 (3.33)
∆τ → 0

Now, differentiate the sum on the right hand side in (3.33) with respect to
hp (n∆τ) for fixed n. The result is ∆τcξ ξ (t − n∆τ) . A similar approach is
0 p

used to differentiate an expression involving a double integral. The derivative is

first set to zero and then a limiting operation ( ∆τ → 0 ) is carried out. We
finally obtain a set of equations known as normal equations:

P −1 ∞

∑ ∫ h (τ′)c
q=0 0
q f p fq (τ − τ ′ )dτ ′ = cξ 0 f p (τ), p = 0,1,... P − 1 (3.34a)

The minimum mean square error (mmse) may be derived by using (3.34a) in
(3.32). We obtain

P −1 ∞
mse min = cξ 0 (0) − ∑ ∫ hp (τ)cξ 0 f p (τ)dτ (3.34b)
p= 0 0

Let us express (3.34) in discrete form. In order to do this we introduce the

following vectors and matrix:

c f p f q (0) c f p f q (∆τ) ... c f p f q ((N − 1)∆τ) 

 
c f p f q (−∆τ) c f p f q (0) ... c f p f q ((N − 1)∆τ)
 
 ... 
c f p fq =
 ... 
 
 ... 
c f f ((1 − N)∆τ) c f f ((2 − N)∆τ), ... c f f (0) 
 p q p q p q 

[ ]
T
h p = hp (0), hp (∆τ), hp (2∆τ), ... hp ((N − 1)∆τ)
and

© 2001 by CRC Press LLC

 [ ]
T
c ξ 0 f p = cξ 0 f p (0), cξ 0 f p (∆τ), cξ 0 f p (2∆τ), ... cξ 0 f p ((N − 1)∆τ)

where ∆τ is the sampling interval and [ ]T

stands for matrix or vector
transpose. Equation (3.34) may now be expressed using the above vectors and
matrices

P −1

∑c
q=0
f p fq hq = cξ 0 f p , p = 0,1,2,... P − 1 (3.35a)

P −1
mse min = cξ 0 (0) − ∑ c Tξ 0 f p h p (3.35b)
p= 0

An alternate representation of (3.34) is through block matrices defined as

h = [h(0), h(∆τ), ... h((N − 1)∆τ) ]

T

where

h(n∆τ) = [h0 (n∆τ), h1 (n∆τ), ... hP −1 (n∆τ)]

T

C0 = [c 0 (0), c 0 (∆τ), ... c 0 ((N − 1)∆τ)]

T

where

[ ]
T
c 0 (n∆τ) = cξ 0 f 0 (n∆τ), cξ 0 f 1 (n∆τ), cξ 0 f 2 (n∆τ), ... cξ 0 f P−1 (n∆τ)

c(0), c(∆τ), ... c((N − 1)∆τ) 

c(−∆τ), c(0), ... c((N − 2)∆τ)
 
C= ... 
 
 ... 
c((1 − N)∆τ), ... c(0) 
(PNxPN)

where each element is a block matrix of the type shown below

© 2001 by CRC Press LLC

 c f 0 f 0 (n∆τ) c f 0 f 1 (n∆τ) ... c f 0 f P−1 (n∆τ) 
 
c f 1 f 0 (n∆τ) c f 1 f 1 (n∆τ) ... c f 1 f P−1 (n∆τ) 
 ... 
c(n∆τ) =  
 ... 
 ... 
 
c f P−1 f 0 (n∆τ) c f P−1 f 0 (n∆τ), ... c f P−1 f P−1 (n∆τ)
(PxP)

As an example consider a three sensor array with two time samples (N=2). The
above quantities, C , h , and C0 , become

c f 0 f 0 (0) c f 0 f 1 (0) c f 0 f 2 (0) c f 0 f 0 (1) c f 0 f 1 (1) c f 0 f 2 (1)

 
c f 1 f 0 (0) c f 1 f 1 (0) c f 1 f 2 (0) c f 1 f 0 (1) c f 1 f 1 (1) c f 1 f 2 (1) 
c (0) c f 2 f 1 (0) c f 2 f 2 (0) c f 2 f 0 (1) c f 2 f 1 (1) c f 2 f 2 (1) 
C= 
f2 f0

c f 0 f 0 (−1) c f 0 f 1 (−1) c f 0 f 2 (−1) c f 0 f 0 (0) c f 0 f 1 (0) c f 0 f 2 (0) 

 
c f 1 f 0 (−1) c f 1 f 1 (−1) c f 1 f 2 (−1) c f 1 f 0 (0) c f 1 f 1 (0) c f 1 f 2 (0) 
 
c f 2 f 0 (−1) c f 2 f 1 (−1) c f 2 f 2 (−1) c f 2 f 0 (0) c f 2 f 1 (0) c f 2 f 2 (0) 

h = [h0 (0),h1 (0),h2 (0),h0 (1),h1 (1),h2 (1)]

T

and

[ ]
T
C0 = cξ 0 f 0 (0),cξ 0 f 1 (0),cξ 0 f 2 (0),cξ 0 f 0 (1),cξ 0 f 1 (1),cξ 0 f 2 (1)

Equation (3.35) may be expressed in a compact form as

C h = C0 (3.36a)

mse min = cξ 0 (0) − CT0 h (3.36b)

3.3.2 Frequency Domain: Sometimes it is advantageous to express the normal

equations (3.34) in the frequency domain. Taking the Fourier transform on both
sides of (3.34) we obtain the following result:

© 2001 by CRC Press LLC

P −1

∑ H (ω)S
q =1
q f p fq (ω) = Sξ 0 f p (ω), p = 0,1,... P − 1 (3.37a)

where Hq (ω) is the transfer function of the qth filter given by

∞
Hq (ω) = ∫ hq (τ)e jωτ dτ
0

S f p f q (ω) is the cross-spectrum between f q (t) and f p (t) and similarly

Sξ 0 f p (ω) is the cross-spectrum between ξ 0 (t) and f p (t) . Similarly,
minimum mean square error (mmse) in the frequency domain may be obtained
from (3.34b)

∞
 P −1 
mse min = ∫− ∞  ξ 0
S (ω) − ∑ H p (ω)Sξ 0 f p (−ω)dω (3.37b)
 p= 0 

To write (3.37) in a matrix form define the following vectors and matrix:

 S0,0 (ω) S0,1 (ω) ... S0, P −1 (ω) 

 S (ω) S (ω) ... S (ω) 
 1,0 1,1 1, P −1 
 ... 
S f (ω) =  
 ... 
 ... 
 
 SP −1, 0 (ω) SP −1, 1 (ω) ... SP −1, P −1 (ω)

H(ω) = [ H0 (ω), H1 (ω) ... HP −1 (ω)]

T

and

S0 (ω) = [ S00 (ω), S01 (ω) ... S0 P −1 (ω)]

T

The normal equations in the frequency domain (3.37) may be expressed in a

compact form

© 2001 by CRC Press LLC

 S f (ω)H(ω) = S0 (ω) (3.38a)

for all ω in the range ±∞ . Formally, the solution of (3.38) may be expressed
as

H(ω) = S−1
f (ω)S 0 (ω) (3.38b)

Now consider a plane wave sweeping across an array of sensors. Let the
background noise be spatially white. The spectral matrix for this model is given
by

S(ω) = A(ω)Sη (ω) (3.39)

where
1 + T(ω) T(ω)e j (u0 x1,0 + v0 y1,0 ) ... T(ω)e j (u0 x P−1, 0 + v0 y P−1, 0 ) 
 − j (u x + v y ) j (u x +v y )

T(ω)e 0 1,0 0 1,0 1 + T(ω) ... T(ω)e 0 P−1, 1 0 P−1, 1 
 ... 
A(ω) =  
 ... 
 
 ... 
T(ω)e − j (u0 x P−1, 0 + v0 y P−1, 0 ) T(ω)e − j (u0 x P−1, 1 + v0 y P−1, 1 ) ... 1 + T(ω) 
 

S0 (ω)
T(ω) = , x p q = x p − xq , y p q = y p − yq , S0 (ω) is the signal
Sη (ω)
spectrum, and Sη (ω ) is the noise spectrum common to all sensors. Similarly,
the vector on the right hand side of (3.38) may be expressed as

[
S 0 (ω) = 1, e
− j (u 0 x1,0 + v 0 y1,0 )
,e
− j (u 0 x 2,0 + v 0 y 2,0 )
, ..., e
− j (u 0 x P−1, 0 + v 0 y P−1, 0 )
]S (ω)
0

= B(ω)S0 (ω)
(3.40)

Using (3.39) and (3.40) in (3.38) we obtain special normal equations

S0 (ω)
A(ω)H(ω) = B(ω)
Sη (ω) (3.41a)
= B(ω)T(ω)

© 2001 by CRC Press LLC

or
H(ω) = A −1 (ω)B(ω)T(ω) (3.41b)

We note that A(ω) has an useful structure in that it can be expressed as

A(ω) = I + T(ω)a(ω)a H (ω) (3.42a)

where

[ ]
T
a(ω) = e j (u0 x 0 + v0 y0 ) , e j (u0 x1 + v0 y1 ) , e j (u0 x 2 + v0 y2 ) , ... e j (u0 x P−1 + v0 y P−1 )
(3.42b)

3.3.3 Constrained Minimization: Let the filtered signal output of an array of

sensors be given by

P −1 ∞
f̂ (t) = ∑ ∫ hp (τ)ξ p (t − τ)dτ (3.43)
p= 0 0

Assume a signal model where a plane wave is sweeping across the array
maintaining its waveform unchanged. The outputs of any two sensors differ
only in propagation delays. The signal at the pth sensor may be given by
ξ p (t) = ξ 0 (t − τ p ) where τ p is the propagation delay at the p t h
sensor ( τ0 = 0 when delays are measured with respect to the 0th sensor). Since
the geometry of the array, speed of propagation and direction of arrival are
known or can be estimated independently the propagation delays are presumed to
be known. The output of each sensor is advanced to make it in phase with the
output of the reference sensor. Equation (3.43) reduces to

P −1 ∞
f̂ (t) = ∑ ∫ hp (τ)ξ 0 (t − τ)dτ
p= 0 0
(3.44)
∞
 P −1 
= ∫  ∑ hp (τ) ξ 0 (t − τ)dτ
0  p= 0 

The filters are to be chosen to satisfy the condition that f̂ (t) = ξ 0 (t) . From
(3.44) it is clear that this constraint can be satisfied if

P −1

∑ h (t) = δ(t)
p= 0
p (3.45a)

© 2001 by CRC Press LLC

 or in the frequency domain

P −1

∑ H (ω) = 1
p= 0
p (3.45b)

Thus for distortion free extraction of waveforms the filters must satisfy (3.45)
[17]. While an individual filter transfer function is allowed to be of any form,
the sum must be a constant.
Another type of constraint arises when it is required to minimize the
background noise power. The noise in the output of an array processor is given
by

P −1 ∞
η̂(t) = ∑ ∫ hp (τ)η p (t − τ)dτ
p= 0 0

and the noise power is given by

∞∞
2
η̂
p q
{ }
σ = ∑ ∑ ∫ ∫ hp (τ)hq ( τ ′ )E η p (t − τ)ηq (t − τ ′ ) dτdτ ′
0 0

∞∞

= ∑ ∑ ∫ ∫ hp (τ)hq ( τ ′ )c pq ( τ ′ − τ )dτdτ ′
p q 0 0
(3.46)
∞
1
= ∑∑ ∫ H p (ω)Hq* (ω)Sη pq (ω)dω
p q 2π 0

For spatially and temporally uncorrelated noise (3.46) reduces to

P −1 ∞
1
σ =∑
2
∫ dω σ 2η
2
η̂
H p (ω)
p= 0 2π 0

The noise power in the array output shall be minimum whenever the filter
transfer functions satisfy the condition

P −1 ∞
1
∑
2
∫
p= 0 2π 0
H p (ω) dω = min (3.47)

© 2001 by CRC Press LLC

The trivial solution, namely, H p (ω) = 0 for all p, is not acceptable as it will
set the output signal power also to zero. The constraint (3.45) or (3.47) is
usually imposed along with other constraints. Sometimes the output noise
power is set to a given fraction (noise reduction factor, NRF) of the input noise
power.

P −1 ∞
1
∑
2
∫
p= 0 2π 0
H p (ω) dω = NRF < 1 (3.48)

§3.4 Wiener Filters for ULA and UCA:

We shall now turn to some specific array geometries; in particular, we consider
uniform linear array (ULA) and uniform circular array (UCA).
3.4.1 Uniform Linear Array (ULA): As in the last section, the signal model
assumed here is a plane wavefront sweeping across a linear array of sensors
spaced at an interval d on x-axis. The noise is assumed to be spatially white. In
(3.41a) the matrix A(ω) and the column B(ω) take the following form:

1 + T(ω) T(ω)e ju0 d ... T(ω)e j (u0 ( P −1)d 

 − ju 0 d 
T(ω)e 1 + T(ω) ... T(ω)e j (u0 ( p−2)d 
 ... 
AULA (ω) =  
 ... 
 ... 
 
T(ω)e − ju0 ( P −1)d T(ω)e − ju0 ( P −2)d ... 1 + T(ω) 
and
BULA = col[1, e − ju0 d , e − ju0 2 d , ... e − ju0 ( P −1)d ]

AULA (ω) can be expressed as (I + T(ω)a(ω)a H (ω)) (see (3.42)) where the
vector a(ω) for a ULA is given by

a(ω) = col[1, e − ju0 d , e − ju0 2 d , ... e − ju0 ( P −1)d ]

Remember that in a ULA the right most sensor is conventionally taken as the
reference sensor. Using the Woodbury’s identity [18] we obtain its inverse in a
closed form,

© 2001 by CRC Press LLC

 −1 1
AULA (ω) = ×
1 + PT(ω)
1 + (P − 1)T(ω) − T(ω)e ju0 d ... − T(ω)e ju0 ( P −1)d 
 − ju 0 d 
−T(ω)e 1 + (P − 1)T(ω) ... − T(ω)e ju0 ( p−2)d 
 ... 
 
 ... 
 ... 
 
−T(ω)e − ju0 ( P −1)d − T(ω)e − ju0 ( P −2)d ... 1 + (P − 1)T(ω)
(3.49)

Using (3.49) in (3.41b) we obtain transfer functions of Wiener filters

T(ω)
H0 (ω) =
1 + PT(ω)
T(ω)
H1 (ω) = e − ju0 d
1 + PT(ω)
T(ω)
H2 (ω) = e − j 2u0 d
1 + PT(ω)
.
(3.50)
.
.
T(ω)
HP −1 (ω) = e − ju0 ( P −1)d
1 + PT(ω)

The frequency wavenumber response of the Wiener filter for T=4 and sixteen
sensor array (ULA) is shown in fig. 3.14. Notice that whenever PT(ω) >>1,
S0 (ω)
that is, either T(ω) = >>1 or P>>1 or both or the spectra of the
Sη (ω)
signal and noise are nonoverlapping, the Wiener filter reduces to a simple delay
filter

1 − jk u0 d
Hk (ω) = e (3.51)
P

© 2001 by CRC Press LLC

3.4.2 Uniform Circular Array: The sensors are uniformly spaced on the
circumference of a circular array of radius R units (see chapter 2 for more on
circular arrays). For circular array the matrix A(ω) takes the form,

AUCA (ω) = I + T(ω)a(ω)a H (ω) (3.52a)

where

 − ju R 2π 2π 
− j[u 0 R cos + v 0 R sin ]
e 0 e P P ... 
a(ω) = col   (3.52b)
2π 2π
 − j[u0 R cos( P −1) P + v 0 R sin( P −1) P ] 
e 
and
BUCA (ω) = a(ω) (3.53)

The reference point is at the center of the circle; however, there is no sensor
physically present there. The Wiener filter is designed to predict the waveform
as seen by an hypothetical sensor kept at the center of the circle. To solve for
the Wiener filter (3.41b) we need to invert the A(ω) matrix. We shall once
again use Woodbury’s identity and obtain

−1
AUCA (ω) = I −
T(ω)
1 + PT(ω)
[
a(ω)a H (ω) ] (3.54)

Using (3.53) and (3.54) in (3.41b) we get

T(ω)
H(ω) = a(ω) (3.55)
1 + PT(ω)

The Wiener filter for a circular array is similar to that for a linear array
except for the difference arising out of the definition of a(ω) . The frequency
wavenumber response of 16 sensor circular array of radius 10λ is shown in fig.
3.15. Although the mainlobe width is narrow, the sidelobe level is quite high.
This is clearly brought out in a cross-sectional plot passing through the
maximum (see fig. 3.16). When we increase the number of sensors to 64 the
sidelobe level is brought down considerably, but the main lobe width remains
practically unchanged. It may be emphasized that the sensors need not be spaced
at ≤ 0.5λ as in a ULA [19]. What is gained by increasing the number of
sensors (keeping the radius fixed) is the reduction of the sidelobe level. In
contrast, in case of a ULA, by increasing the number of sensors the array
aperture is increased which in turn sharpens the mainlobe but does not reduce
the sidelobe level.

© 2001 by CRC Press LLC

 3

0.0
2
0.0

1 0.9
Temporal frequency

0.0
0.0
.
0 0.5
0.5
0.0

-1
0.9 0.0

0.0
-2
0.9

-3
-3 -2 -1 0 1 2
Spatial frequency

Figure 3.14: Frequency wavenumber response of the Wiener filter for ULA. The
direction of arrival (DOA) is assumed to be known. In the above example it is equal
to 30o .

3.4.3 Robustification: The Wiener filters given by (3.51) and (3.55) require a
knowledge of a(ω) for which we need to know the apparent speed of the
wavefront sweeping across the array. Prior to waveform estimation it is a
common practice to estimate the direction of arrival (DOA) of a wavefront.
(This topic is covered in some detail in chapter 5). The DOA estimation is not
without error. Hence, it would be nice if the Wiener filters are made robust so
that the degradation in its performance is minimum. We shall confine to a
linear array (ULA). An error in DOA estimation will introduce an error in the
wavenumber. For a ULA, the erroneous wavenumber may be expressed as
(u0 + ε) where ε is auniformly distributed random variable. Naturally, there
AULA (ω) and BULA (ω) . We shall use the
will be an error in each element of
stochastically averaged AULA (ω) and BULA (ω) matrices in (3.41b). This
approach was suggested in [20] in the context of optimal velocity filters in

© 2001 by CRC Press LLC

 3

1
v

-1

-2

-3
-2 0 2
u

Figure 3.15: Frequency wavenumber response of Wiener filter for UCA. The DOAs
are known (azimuth=elevation=45o and ω=π /2). 16 sensors and constant snr=4 are
assumed.

seismic exploration. Observe that in each element of the matrices there is an

extra term which does not permit the matrix to be written as an outer product of
two vectors as in (3.52a).

AULA (ω ) =

(3.56a)
ju 0 ( P −1)d
1+T (ω ) ju 0 d
T (ω ) sin c(ε 0 d )e ... T (ω ) sin c(ε 0 ( P −1)d )e 
 T (ω ) sin c(ε 0 d )e − ju0 d 1+T (ω ) ... T (ω ) sin c(ε 0 ( P −2)d )e 0
ju ( p−2)d 
 ... 
 ...
...

 − ju 0 ( P −1)d − ju 0 ( P −2)d 
 T (ω ) sin c(ε 0 ( P −1)d )e T (ω ) sin c(ε 0 ( P −2)d )e ... 1+T (ω ) 

and

1, sin c(ε 0 d)e − ju0 d , sin c(2ε 0 d)e − ju0 2 d , ..., 
BULA = col   (3.56b)
sin c((P − 1)ε 0 d)e − ju0 ( P −1)d 

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

16
0.4

0.2

0 64
-4 -2 0 2 4
Spatial frequency

Figure 3.16: A cross-section of the Wiener filter response taken through the peak.
Number of sensors are 16 and 64. The sidelobe level relative to the peak has been
reduced when the number of sensors is increased from 16 to 64. Radius of circular
aperture is 10 units.

The effectiveness of the proposed approach is demonstrated through frequency

wavenumber response of the Wiener filters before and after robustification. We
compute the frequency wavenumber response given that the DOA is known to
an accuracy of ±2.0 . The plot is shown in fig. 3.17a for a ULA with 16
o

sensors and constant snr (=4). Notice the splitting of the peak particularly in the
higher temporal frequency range. Next, we compute the frequency wavenumber
response of a UCA designed to tolerate an error of 2o . The frequency
wavenumber response is shown in fig. 3.17b. The main lobe shape remains
practically unchanged but there is an increase in the sidelobe level. Cross-
sectional plots passing through the maximum of the response function of the
UCA with and without DOA error are shown in fig. 3.18. While the shape of
the main lobe remains practically unchanged the side lobe level seems to have
slightly increased. This is the price one has to pay for the lack of exact
knowledge of the DOA.
3.4.4 Levinson-Durbin Algorithm: In (3.36), in order to solve for h, we have
to invert a large covariance matrix (e.g., with the array size, M=24, and filter
length, N=16, the size of the covariance matrix will be 384x384). The

© 2001 by CRC Press LLC

 3

0.2
Temporal frequency

2
1
0.4
0.8
1

0
-2 -1 0 1 2
Spatial frequency
(a)

ω 0

-1

-2

-3
-2 0 2
u
(b)

Figure 3.17: Frequency wavenumber response of Wiener filter when the error in DOA
estimate is ±2.0 o . (a) ULA and (b) UCA. 16 sensors and 64 time samples are
assumed.

© 2001 by CRC Press LLC

 2.5

1.5
Response

0.5

0
-4 -2 0 2 4
Spatial frequency

Figure 3.18: A cross-section through the maximum (v0=0.7854) for UCA. Solid line
shows response when there is no error and the dashed line shows one with DOA error
( ±2.0 ). 16 sensors and 64 time samples are assumed.
o

computational load for inverting such a large matrix will be very high. We
shall outline a recursive method applicable to a ULA. For any other geometry
of an array the covariance matrix becomes a block symmetric matrix, but only
for a ULA does the covariance matrix become toeplitz. This important property
enables us to devise a recursive algorithm, known as Levinson-Durbin
algorithm, which requires inversion of a matrix of size M × M in place of a
matrix of size MN × MN. Briefly, the algorithm is as follows [16]: Let h N
[ ]
( h N = col h(0), h(∆τ), ... h((N − 1)∆τ) ) be the solution of the Nt h
order normal equations, (3.36). Let us now increase the size of the covariance
matrix by padding one row and one column of covariance matrices as below,

C N c(N)
  h N  C0 
  = 
.
. (3.57)
 .
 0   γ N 
c(N) ... c(0) 

© 2001 by CRC Press LLC

where
γN = [c(N) ... c(1)] h N
(M × 1) (M × MN) (MN × 1)

Note that the square matrix on the left hand side of (3.57) is a block covariance
matrix of size M(N + 1) × M(N + 1) ; therefore, (3.57) is similar to (3.36)
but of order N+1. Indeed, if we subtract (3.57) from (3.36) of the same order,
we shall obtain

 N 
h  0 
C N +1 h N +1 −    =  
 (3.58)
0   c (N) − γ 
    0 N

Define a set of auxiliary coefficient matrices

b N = col{b NN ,b NN −1 ,...b N1 , I}
( N +1) M × M

where b NN −i , i = 0, ... , N − 1 are M × M matrices yet to be defined, but

they satisfy the following recurrence relation:

h N +1 (i) = b NN −i h N +1 (N) + h N (i), i = 0,1,..., N − 1 (3.59)

b NN h N +1 (N) 
b  0 
NN −1h N +1 (N)
C N +1   =  
 (3.60a)
M  β N h N +1 (N)

 
h N +1 (N) 

where β N h N +1 (N) = c 0 (N) − γ N . Eliminating h N +1 (N) from both sides

of (3.60a) we obtain

 0 
 
C N +1b N =   (3.60b)


β 
N

We shall once again increase the order of C N +1 in (3.60b) by padding one more
row and one more column. We get

© 2001 by CRC Press LLC

 β′N 
0   
C N + 2   = 0  (3.60c)
b N   
β N 

where

β′N =[c(1), c(2),...c(N + 1),] b N

(M × M(N + 1)) M(N + 1) × M

We shall now introduce another set of auxiliary coefficients,

{ }
a N = col I, a N 1 , a N 2 , ..., a N N as a solution of the following system of
equations:

α N 
C N +1a N =   (3.61a)
0 
where

[
α N = c(0)I + c(1)a N 1 + ... + c(N)a NN ]
M × M

Let us now increase the size of C N +1 in (3.61a) by padding one more row and
one more column of covariance matrices. We obtain

α N 
a N   
C N + 2   = 0  (3.61b)
0   
α ′N 
where
[
α ′N = c(N + 1)I + c(N)a N 1 + ... + c(1)a NN ]
M × M

We linearly combine (3.60c) and (3.61b) such that the resulting equation is the
(N+2)th order equivalent of (3.60b). Let the linear combination be given by

© 2001 by CRC Press LLC

 β′N  α N  
 0  a N        
C N + 2    +   δ N  =  0  + 0  δ N  (3.62a)
b N  0   β  α ′  
 N   N  

where δ N is a M × M matrix of constants which we shall select in such a

manner that the right hand side of (3.62a) resembles the right hand side of
(3.60b). This may be achieved by requiring

β′N + α N δ N = 0
or
δ N = −α −1
N β ′N (3.62b)

The resulting equation is equivalent to (3.60) but of order N+2. Then, we have

0  a N 
b N +1 =   +  δ N (3.63a)
b N   0 

β N +1 = β N + α ′N δ N (3.63b)

Further we take a linear combination of (3.60c) and (3.61b) such that the
resulting equation resembles (3.61a) but it is of order N+2. This may be
achieved by requiring

∆ N β N + α ′N = 0
or
∆ N = −β −1
N α ′N (3.64a)

where ∆ N is also a M × M matrix of constants for linear combination. Now

we have

a N  0 
a N +1 =   +   ∆ N (3.64b)
 0  b N 

α N +1 = α N + β′N ∆ N (3.64c)

Eqs. (3.63a) and (3.64b) form a set of recursive relations to compute a N +1 and
b N +1 given a N and b N . Similarly, equations (3.63b) and (3.64c) form another

© 2001 by CRC Press LLC

 set of recursive relations to compute α N +1 and β N +1 given α N and β N . The
initial conditions are:

1) b0,0 = a 0,0 = I , 2) α 0 = β 0 = c(0) and 3) h0 = 0

Finally, to compute the filter coefficients we need h N +1 (N) which we get

from

h N +1 (N) = [β N ] (c 0 (N) − γ N )
−1

It must be emphasized that, since the Levinson-Durbin algorithm exploits the

toeplitz property of the covariance matrix, the array signal must be stationary
and sufficiently large data must be available for the estimation of statistically
averaged covariance matrix. Such a filter will naturally be optimum to an
ensemble of time series having the same second order structure. In practice, this
is often difficult to realize; consequently, the toeplitz character is often lost and
we cannot use the Levinson-Durbin algorithm altogether. Later in chapter 6 we
shall describe a deterministic least squares approach which uses the actual
measured signal in place of covariance function.

§3.5 Predictive Noise Cancellation:

Noise suppression is based on the principle of differential spectral properties of
signal and noise. A filter is often used to maximally attenuate the noise power
but at the same time to minimally attenuate the signal power. There is an
alternate approach to noise attenuation known as noise cancellation. This
involves prediction of the noise that is actually corrupting the signal. For
prediction we need a sample of noise which is not contaminated with the signal
but which is correlated with the noise present in the actual observed signal plus
noise sample. In principle, it is possible to devise a Wiener filter for the
prediction of the unknown noise from the sample of noise which is correlated
with the unknown. A signal free noise sample can be obtained as follows:
(a) Place an extra sensor within the correlation distance of the noise but away
from the signal source. This is possible when the signal source is in the near
field region but the noise sources are in the far field region, for example, when a
speaker is close to a microphone and the noise sources are far away from the
microphone.
(b) Both signal and noise are in the far field region but reach the microphone
array from different directions. An array of sensors may be simultaneously used
to receive the signal and the noise coming from different directions. However,
since the array response is finite in a direction other than the direction to which
it is tuned, some amount of noise will leak into the array output. The noise,
which has leaked into the array output, will be strongly correlated with the

© 2001 by CRC Press LLC

incident noise and hence it may be predicted using the noise output when the
array is steered in the direction of the noise.
In the above approach the noise is canceled electronically. It is also
possible to achieve the noise cancellation acoustically. This would, however,
require use of many coherently generated noise sources whose combined effect is
to produce a noise wavefront with a phase equal but opposite to that of the
noise present in the observed waveform.
3.5.1 Signal Source in Near Field: Consider a speaker close to a microphone
( M1 ) and another microphone ( M2 ) away from the speaker but well within the
correlation distance of the noise (see fig. 3.19). It is assumed that all noise
sources are in the far field region. Let f 1 (t) be the output of M1 , f 2 (t) be
the output ofM2 , ξ 0 (t) be the signal emitted by the speaker and η1 (t) be
the noise in M1 . The signal and noise in M2 are related to those in M1
through impulse response functions, h1 (t) and h2 (t) .

f 1 (t) = ξ 0 (t) + η1 (t)

∞ ∞
(3.65)
f 2 (t) = ∫ ξ 0 (t− t ′ )h1 (t ′ )dt ′ + ∫ η1 (t− t ′ )h2 (t ′ )dt ′
0 0

The output of M2 is passed through a prediction filter, hpred (t) , which is

found by minimizing a quantity,

 ∞ 2

E  f 1 (t) − ∫ f 2 (t ′ − t)hpred (t ′ )dt ′  = min
 0 

or in the frequency domain by minimizing the following

∞  S (ω) + S (ω) H 2 
1 pred (ω) −
∫  1  dω
2
2π  S (ω)H ∗ (ω) − S ∗ (ω)H 
− ∞  12 pred 12 pred (ω) 

We obtain

S12 (ω) S0 (ω)H1∗ (ω) + Sη1 (ω)H2∗ (ω)

H pred (ω) = = 2 2 (3.66a)
S2 (ω) S0 (ω) H1 (ω) + Sη1 (ω) H2 (ω)

and

© 2001 by CRC Press LLC

 Speaker
M f1(t)
⊃1 + -

h1
M2
⊃ + h
h2 f2(t)

Figure 3.19: Source is in near field and noise is in far field. Microphone M2
(reference microphone) receives very little of the signal.

[ ]
∞
1 2
Error min = ∫
2π − ∞
S1 (ω) − S2 (ω) H pred (ω) dω (3.66b)

Some Special Cases:

(a) h1 (t) = 0 . The signal from the speaker does not reach M2 .
h2 (t) = δ(t − τ 0 ) . The noise reaching the microphone M2 is simply a
delayed version of the noise reaching microphone M1 . For this case, the
prediction filter is simply a delay filter, H pred (ω) = e 0 and the minimum
jωτ

error is equal to the signal power,

∞
1
2π −∫∞
Error min = S0 (ω)dω

In this special case complete noise cancellation takes place.

(b) h1 (t) = h2 (t) ≠ 0 . The output of the reference microphone is a filtered
version of the output of M1 . For this case, the prediction filter is given by

1
H pred (ω) =
H1 (ω)

© 2001 by CRC Press LLC

and the minimum error is equal to zero. Apparently, both signal and noise are
canceled and the output power is zero. For noise cancellation to take place we
must have h2 (t) > h1 (t) .
Define gain G as a ratio of snr at the output to snr at the input. The
output power is given by

∞
1
∫ [S (ω) − S (ω) H
2
Output power = 1 2 pred (ω) ]dω
2π
−∞

∞  S0 (ω) + Sη (ω) − (S0 (ω) H1 (ω) 2 + 

1 
∫
1
=  2  dω
2π 2
−∞  Sη (ω) H2 (ω) ) H pred (ω)
 1


∞

∫ {S (ω)(1 − H (ω) }
1 2 2
= 0 1 H pred (ω) ) dω Signal power
2π
−∞
∞

∫ {S }
1 2 2
+ η1 (ω)(1 − H2 (ω) H pred (ω) ) dω Noise power
2π
−∞

The snr at the output as a function of frequency is given by

=
{S (ω)(1 − H (ω) H
0 1
2
pred (ω) )
2
}
{S (ω)(1 − H (ω) H (ω) )}
SNRoutput
2 2
η1 2 pred

from which the gain as defined here turns out to be

2 2
[ SNR ]output 1 − H1 (ω) H pred (ω)
G= [ SNR ]input
= 2 2 (3.67)
1 − H2 (ω) H pred (ω)

2 2
For G>1 we must have H1 (ω) < H2 (ω) .
3.5.2 Source in Far Field: Both signal and noise sources are in the far field
region but the DOAs of their wavefronts are different. Let τ 0 and τ1 be the
incremental delays produced by the signal wavefront and noise wavefront
respectively. The array can be steered to receive the signal or noise at the same
time (see fig. 3.20)[21]. Let f 1 (t) be the output of an array when it is steered
to the signal wavefront and f 2 (t) be the output of an array when steered to the
noise wavefront. Since the array response function has finite side lobes, some

© 2001 by CRC Press LLC

 amount of wave energy will leak through the sidelobes. Hence, we model the
array output as

∞
1
2π −∫∞
f 1 (t) = ξ 0 (t) + Ν1 (ω)H(ω(τ 0 − τ1 ))e jωt dω (3.68a)

and
∞
1
2π −∫∞
f 2 (t) = η1 (t) + Ξ0 (ω)H(ω(τ1 − τ 0 ))e jωt dω (3.68b)

where Ξ0 (ω) is the Fourier transform of the signal and Ν1 (ω) is that of the
noise. By comparing (3.65) with (3.68) it is possible to write

H1 (ω) = H (ω(τ1 − τ 0 )) (3.69a)

and

1
H2 (ω) = (3.69b)
H (ω(τ 0 − τ1 ))

Using (3.69) in (3.66a) we obtain a filter to predict the noise in f 1 (t) ,

Sη1 (ω)
Sξ 0 (ω)H ∗ (ω(τ1 − τ 0 )) + ∗
H (ω(τ 0 − τ1 ))
H pred (ω) =
2 Sη1 (ω)
Sξ 0 (ω) H(ω(τ1 − τ 0 )) + 2
H(ω(τ 0 − τ1 ))
1 + SNRinput
= 2 H(ω(τ 0 − τ1 ))
1 + SNRinput H(ω(τ1 − τ 0 ))

Let us now consider a few special cases:

2
(a) When SNRinput >>1 and SNRinput H(ω(τ1 − τ 0 )) >>1

1
H pred (ω) ≈
H(ω(τ1 − τ 0 ))

If this filter is used on f 2 (t) (see eq.(3.68b)) for predicting the noise in f 1 (t) ,
the signal component will be restored causing the cancellation of the signal.

© 2001 by CRC Press LLC

 f1 (t) f2 (t)

1
1
1
1

4τ
3τ
2τ
τ

3τ0
2τ0

4τ0
0
τ

Noise
Signal

Figure 3.20: An array of sensors can be steered simultaneously in the direction of the
signal and in the direction of noise. When the array is steered in the direction of
signal the output f 1 (t) is mostly signal and when it is steered in the direction of
noise the output f 2 (t) is mostly noise.

2
(b) SNRinput H(ω(τ1 − τ 0 )) <<1

H pred (ω) ≈ (1 + SNRinput )H(ω(τ 0 − τ1 ))

© 2001 by CRC Press LLC

 If this filter is used on f 2 (t) (see eq.(3.68b)) for predicting the noise
component in f 1 (t) the noise component will be largely canceled without
canceling the signal. As an illustration, we consider two pure sinusoidal signals
(of same frequency) arriving with different DOAs (0o and 5.7o) at a ULA of 16
sensors spaced at λ/2 spacing. The second sinusoid arrives 50 time units later
with an amplitude of 0.8. Figure 3.21(a) shows a sum of the two tones as
received by the first sensor. The array is steered in the direction of the first
sinusoid and at the same time in the direction of the second sinusoid. The array
outputs are described by (3.68), which is now considerably simplified for pure
sinusoidal inputs.

f 1 (t) = s1 (t) + s2 (t)H(ω 0 (τ 0 − τ1 ))

(3.70a)
f 2 (t) = s2 (t) + s1 (t)H(ω 0 (τ 0 − τ1 ))

where s1 (t) and s2 (t) are the first and the second sinusoid, respectively and
ω 0 is the frequency of the sinusoids. Solving (3.70) we obtain, for
2
H(ω 0 (τ 0 − τ1 )) < 1,

f 1 (t) − f 2 (t)H(ω 0 (τ 0 − τ1 ))
s1 (t) =
1 − H 2 (ω 0 (τ 0 − τ1 ))
(3.70b)
f 2 (t) − f 1 (t)H(ω 0 (τ 0 − τ1 ))
s2 (t) =
1 − H 2 (ω 0 (τ 0 − τ1 ))

The results are shown in figs. 3.21 (b & c).

We explore another possibility, where, instead of steering the array in
the direction of interference, we steer a null in the direction of a signal with
finite response in the direction of interference [22]. Equation (3.68b) may be
written as

∞
1
f 2 (t) = ∫
H
H null (ω)Ν1 (ω)e jωt dω (3.71a)
2π − ∞

where

 a (ω)a 0H (ω) 
H null (ω) = I − 0  a1 (ω)
 M 

© 2001 by CRC Press LLC

 2
(a)

1.5

0.5
Amplitude

-0.5

-1

-1.5

-2
0 50 100 150 200 250 300
1
(b)

0.5
Amplitude

-0.5

-1
0 50 100 150 200 250 300
1
(c)

0.5
Amplitude

-0.5

-1
0 50 100 150 200 250 300
Time

Figure 3.21: (a) sum of two sinusoids, (b) first sinusoid after subtraction and (c)
second sinusoid after subtraction.

© 2001 by CRC Press LLC

 is a filter with a null in the direction of the signal. Noting that
Ν1 (ω) = a1 (ω)Ν1 (ω) , (3.71a) may be written as

1
∞  a1H (ω)a 0 (ω) 
2

2π −∫∞ 
f 2 (t) = M 1 −  Ν1 (ω)e jωt dω (3.71b)
M2 


Define a filter

a 0H (ω)a1 (ω)
w= M (3.71c)
 a1 (ω)a 0 (ω) 
H 2

M 1 − 
 M2 

If we now pass f 2 (t) through above filter (3.71c) the output will be exactly
the same as the noise term in (3.68a); therefore it may be removed by simple
subtraction.
3.5.3 Adaptive Filter: We have derived the multichannel Wiener filter in §3.3.
A single channel version may be derived along the same lines (see [22] for
derivation). Here we shall state the final result. The Wiener filter which predicts
f 1 (t) from f 2 (t) is given by h = C −1 f 2 C f 1 f 2 where C f 2 is the covariance
matrix of f 2 (t) and C f 2 f 1 is the cross-covariance matrix between f 2 (t) and
f 1 (t) . For real time estimation of the filter and also to account for temporal
variations in the covariance functions it is appropriate to devise an adaptive
approach which in the limiting case reduces to the Wiener solution. Let
h = [h0 , h1 , h2 ... hN −1 ] be the prediction filter vector and
T

f 2 = [ f 2 (t), f 2 (t − ∆t), f 2 (t − 2∆t), ... f 2 (t − (N − 1)∆t)]

T

T
be the data vector. The filter output is given by h f 2 , which is required to be
as close as possible to f 1 (t) . This filter is known as a transversal filter acting
on the delayed outputs (see fig. 3.22). For this we need to minimize the mean
square error

{
E ε(t)
2
} = E{ f (t) − h f
1
T
2
2
}

© 2001 by CRC Press LLC

 Input
-1 -1 ... -1
z z z

h0 h1 hN-1

+
Output

Figure 3.22: Structure of transversal filter. The filter coefficients are made adaptable
to the changing input.

with respect to the prediction filter coefficients. To minimize the error power
we need to compute a gradient of {
E ε(t)
2
} with respect to
h0 , h1 , h2 ... hN −1 and go down the path of the steepest descent until a
minimum (possibly a local minimum) is encountered. The difficulty, however,
is in estimating {
E ε(t)
2
}, which requires averaging over a finite interval
(ideally infinite) of time. Instead, in the least mean squares (LMS) algorithm, it
is proposed to use ε(t) in place of E ε(t)
2
{ 2
}. The gradient of ε(t)
2
is
now easily computed,

 ∂ε(t) ∂ε(t) ∂ε(t) 

∇ ε(t) = 2ε(t)
2
, , ... ,
 ∂h0 ∂h1 ∂hN −1  (3.72)
= −2ε(t)f 2 (t)

In the steepest descent search method the current filter vector is adjusted by an
amount proportional to negative of the gradient of the error function, ∇ ε(t)
2

[24]. The idea of adaptation is illustrated in fig. 3.23. The current filter vector is
updated by an amount proportional to the product of prediction error and current
input,

hi +1 = hi + 2µε(t)f 2 (t) (3.73)

where µ is a gain constant which regulates the speed of adaptation.

© 2001 by CRC Press LLC

 It is interesting to note that the filter vector converges to the Wiener
−1
filter, that is, as i → ∞ hi → C f 2 C f 1 f 2 . To show this consider the expected
value of hi +1

E{hi +1} = E{hi } + E{2µε(t)f 2 (t)}

= E{hi } + 2µE{f 2 (t)( f 1 (t) − f T2 (t)hi )} (3.74)

= E{hi } + 2µCf 2 f 1 − 2µCf 2 E{hi }

where we have assumed that hi and f 2 (t) are independent. Let h̃ represent
filter coefficients obtained by solving the normal equation (3.36), that is,
h̃ = Cf−12 Cf 2 f 1 . Equation (3.74) reduces to

{ } { }
E hi +1 − h̃ = E hi − h̃ + 2µh̃Cf 2 − 2µCf 2 E{hi } (3.75a)

E{∆hi +1} = E{∆hi } − 2µCf 2 E{∆hi } (3.75b)

E{∆hi +1} = (I − 2µCf 2 )E{∆hi } (3.75c)

Let us use the eigendecomposition of the covariance matrix, Cf 2 = VΛV H ,

in (3.75c) and obtain

E{V H ∆hi +1} = (I − 2µΛ)E{V H ∆hi } (3.76)

The solution of the above difference equation (3.76) is given by

E{V H ∆hi } = (I − 2µΛ)i χ 0 where χ 0 is the initial condition. As i → ∞
E{V H ∆hi } → 0 provided (I − 2µΛ)i → 0 for i → ∞ . This is possible
iff(I − 2µλ l ) < 1 for all i. This can be easily achieved if we were to select
1
µ such that 0 < µ < where λ max stands for the maximum eigenvalue
λ max
of the covariance matrix. Note that { }
λ max ≤ tr Cf 2 = the sum of the
eigenvalues of the covariance matrix. Hence, E{∆hi } → 0 as i → ∞ . From
this result it follows that E{hi } → h̃ as i → ∞ .

© 2001 by CRC Press LLC

 f1(t)
+ ε( t )

f2(t)
hpred

Figure 3.23: The idea of adaptive prediction filter is illustrated in the above figure.
The prediction error modifies the filter so as to reduce the prediction error.

§3.6 Exercises:
1. A tapered fan filter is defined as [3],

1 u u
H(u, ω ) = rect( ) • rect( )
ω ω 2ω

where • stands for the convolution in u. Sketch the filter frequency response
function. Compute the impulse response function.
2. In the weighted least squares filter design the minimum mean square error,
which is given by

WH − [WH ] finite
+ π 2
1
ε = ∫∫
2
dudv ,
4π 2
min
− π
W

becomes independent of the weight function as the filter size increases, ideally
at infinity.
3. The output of a ULA with its sensors spaced at one meter apart is sampled at
the rate of 5kHz (Nyquist rate). A quadrant filter is desired with upper and lower
cut off speeds 7 km/sec and 3 km/sec, respectively. Sketch the pass regions
including the aliased part, if any.
4. The following are the low pass 1D filter coefficients:

h(0)=0.52
h(1)=0.3133176
h(2)=-0.01808986
h(3)=-0.09138802
h(4)=0.01223454
h(5)=0.04000004

© 2001 by CRC Press LLC

 h(6)=-0.001945309
h(7)=-0.014112893

Compute the coefficients bk , k = 0,1,..., N appearing in (3.26b) on page 168

and then, using these coefficients, evaluate a circular and a fan shaped 2D filters.
5. A UCA is split into two UCAs. The first UCA has all even sensors and the
second UCA has all odd sensors. Show that the response of the first UCA may
be obtained by rotating the response of the second UCA through an angle equal
to angular separation between the sensors. Using this property give a heuristic
explanation on the behaviour of the sidelobe as a function of the number of
sensors (see p. 180).

References

1. H. C. Andrews and B. R. Hunt, Digital Image Restoration, Prentice-Hall,

Englewood Cliffs, NJ, 1977.
2. P. S. Naidu and M. P. Mathew, Geophysical Potential Field Analysis- A
Digital Signal Processing Approach, Advances in Geophysical Exploration vol.
5, Elsevier Science Publishers, Amsterdam, 1998.
3. J. P. Fail and G. Grau, Les filtere en eventail, Geophy. Prosp., vol. 11, pp.
131-163, 1963.
4. P. Embree, J. P. Burg, and M. M. Backus, Wide-band velocity filtering- The
pie slice process, Geophysics, vol. 28, pp. 948-974, 1963.
5. S. Treitel, J. L. Shanks and C. W. Frasier, Some aspects of fan filtering,
Geophysics, vol. 32, pp. 789-800, 1967.
6. Soo-Chang Pei and Sy-Been Jaw, Two dimensional general fan-type FIR
digital filter design, Signal Processing (Eurosip), vol. 37, pp. 265-274, 1994
7. R. A. Wiggins, ω-k filter design, Geophysical Prospecting, vol. 14, pp.
427-440, 1966.
8. J. G. Proakis and D. G. Manolakis, Digital Signal Processing (chapter
eight), Second Edition, Prentice-Hall of India Private Ltd, New Delhi, 1995.
9. J. H. McClellan, The design of two dimensional digital filter by
transformation, Proc. of 7th Annual Princeton Conf on Information Sciences
and Systems, pp. 247-251, 1973.
10. M. Abramowitz and I. A. Stegun (Editors), Handbook of Mathematical
Functions, National Bureau of Standards, Applied Mathematics Series 55,
1964.
11. J. H. McClellan, T. W. Parks, Equiripple approximation of fan filter,
Geophysics, vol. 37, pp. 573-583, 1972.
12. D. E. Dudgeon, Fundamentals of digital array processing, Proc. IEEE, vol.
65, pp. 898-905, 1975.
13. N Wiener, Extrapolation, interpolation and smoothing of stationary time
series, The MIT Press, Cambridge, MA, 1949.

© 2001 by CRC Press LLC

14. N. Levinson, The Wiener RMS (root mean square) error criterion in filter
design and prediction, J. of Mathematics and Physics, vol. 25, pp. 261-278,
1947 (see also Appendix B in [13]).
15. J. Durbin, The fitting of time series models, Revue de l’Instutute
International de Statistique, vol. 28, pp. 233-243, 1960.
16. R. A. Wiggins and E. A. Robinson, Recursive solution to the
multichannel filtering problem, Journal of Geophysical Research, vol. 70, pp.
1885-1891, 1965.
17. P. E. Green, E. J. Kelly, Jr., and M. J. Levin, A comparison of seismic
array processing methods, Geophy. J. Roy Soc, vol. 11, pp. 67-84, 1966.
18. S. M. Kay, Modern Spectrum Analysis, Prentice-Hall, Englewood Cliffs,
NJ, p. 24, 1989.
19. C. Usha Padmini and Prabhakar S. Naidu, Circular array and estimation of
direction of arrival of a broadband source, Signal Processing, vol. 37, pp. 243-
254, 1994.
20. R. C. Sengbush and M. R. Foster, Design and application of optimal
velocity filters in seismic exploration, IEEE Trans G21, pp. 648-654, 1972.
21. A. Cantoni and L. C. Godara, Performance of a postbeamformer interference
canceller in the presence of broadband directional signals, J. Acoust. Soc. Am.,
vol. 76, pp. 128-138, 1984.
22. L. C. Godara, A robust adaptive array processor, IEEE Trans., vol. CAS-
34, pp.721-730, 1987.
23. C. W. Therrien, Discrete random signals and statistical signal processing,
Prentice-Hall, Englewood Cliffs, NJ, 1992. (1D Wiener filter, p. 422).
24. B. Widrow and J. D. Stearms, Adaptive Signal Processing, Prentice-Hall,
Englewood Cliffs, NJ, p. 47, 1985.

© 2001 by CRC Press LLC

 Chapter Four

Source Localization:
Frequency Wavenumber Spectrum

In this chapter we consider the most important problem in sensor array signal
processing which is estimation of the coordinates of a source emitting a signal
(passive localization) or a point target illuminated by an external signal (active
localization). A point in three dimensional space is defined by three parameters,
namely, range (r), azimuth ( ϕ ) and elevation (θ) . The range is often measured
by means of return time of travel in active systems and by means of time delays
measured at a number of sensors in passive systems. The azimuth and elevation
angles are obtained from the measurements of direction of arrival (DOA) by an
array of sensors. A horizontal array of sensors is required for azimuth
measurement and a vertical array for elevation measurement. The basic quantity
used for estimation of location parameters is the frequency wavenumber (ω , k)
spectrum (see chapter 2). A source is assumed to be present where there is a
concentration of power. We shall describe three different methods; namely,
beamformation, Capon spectrum and maximum entropy spectrum. The last two
methods fall under the nonlinear category while the first method belongs to the
linear category. The important difference between the linear and nonlinear
methods lies in their response to an input which consists of a sum of two or
more uncorrelated signals. The output of a linear method will be a sum of the
spectra of input signals but the output of a nonlinear method may contain an
additional cross term. In spite of this drawback the nonlinear methods have
become quite popular [1].

§4.1 Frequency Wavenumber Spectrum:

A wavefield produced by sources in the far field region may be expressed as a
sum of plane waves with random phase (see plane wave decomposition in
chapter 1, p. 13). The quantity of interest is power (or energy when transient
waves are involved) as a function of azimuth and elevation. This is the
frequency wavenumber (ω , k) spectrum which we had introduced in chapter 2.
As the array is of limited size, the task of estimating the frequency wavenumber
spectrum becomes too ambitious. What can be estimated with a reasonable
certainty is the spectral matrix by treating the array output as a multichannel
time series. The spectral matrix is indeed related to the frequency wavenumber
spectrum. Fortunately, for the purpose of source localization it is enough if we
can accurately estimate the spectral matrix.
4.1.1 Spectral Representation of the Wavefield: As in chapter 2 we shall model
the wavefield as a stochastic process; hence the basic tool will be the spectral
representation of the wavefield,

© 2001 by CRC Press LLC

 +π
1
3 ∫ ∫∫
f (x, y, z,t) = dF(ω ,u,v)e j ( ωt −ux − vy − k 2 −s2 z)

(2 π ) − π

ω
where k= and s = u + v . Using the stochastic properties of the
2 2

c
spectral representation of a wavefield the frequency wavenumber (ω , k)
spectrum may be given by

1
S f (ω,u,v)dωdudv
(2π)3

 1 1 
= E 3
dF(ω,u,v) 3
dF * (ω,u,v) (4.1)
 (2π) (2π) 

Note that S f (ω ,u,v) ≥ 0 and it satisfies symmetry relations

S f (ω , +u, −v) = S f (ω , −u, +v)

S f (ω ,u,v) = S f (ω , −u, −v)

Further, the (ω , k) spectrum, for a propagating wavefield, must satisfy a

condition, S f (ω ,u,v) = 0 for u 2 + v 2 > k . The spectrum is also defined
as the Fourier transform of the spatio-temporal covariance function,

∞
1
3 ∫ ∫∫ f
S f (ω,u,v) = c (x, y, τ)e j (ωτ −ux − vy ) dxdydτ (4.2)
(2π) − ∞
where
c f (x, y, τ) = E{ f (x, y,t) f * (x + x, y + y,t + τ)}

is the spatio-temporal covariance function of the wavefield on the z=0 surface.

Consider an example of a wideband plane wavefront. Let cx and cy be the
apparent wave speed in the x and y directions respectively. The spectrum of the
wavefront is given by

S f (ω,u,v) = S0 (ω)δ(ω − cx u)δ(ω − cy v) (4.3)

© 2001 by CRC Press LLC

where S0 (ω) is the spectrum of the temporal waveform. The spectrum of a
plane wavefront given in (4.3) is a line in the (ω , k) space passing through
the origin with direction cosines α = sin θ cosϕ , β = sin θsin ϕ , and
c c
γ = cosθ (see fig. 4.1). Note that cx = and cy = . A stochastic
α β
wavefield may be modeled as a sum of uncorrelated plane waves; therefore its
spectrum will consist of a set of radial line segments.
4.1.2 Aliasing: In the spatial domain a wavefield is always measured at a set of
discrete points (using point detectors) while in the temporal domain the
wavefield is measured as a continuous function of time or over dense sample
points, as required. Sampling in the spatial domain is dictated by the cost of
deploying a large array of sensors but in the temporal domain the sampling rate
may be as high as required, only at a marginally higher cost. Hence, the
phenomenon of aliasing in the spatial domain becomes important. Consider a
ULA of infinite length with interelement spacing equal to d. The highest spatial
frequency beyond which there is a reflection of power resulting in aliasing is
π
. This phenomenon is illustrated in fig. 3.9 with reference to a digital filter.
d
Let us assume that the signal has been prefiltered to limit the temporal
spectrum to a band, ± ω max . The effect of spatial sampling on the spectrum of
a broadband plane wave is illustrated in fig. 4.2. The (ω , k) spectrum lies on a
line abcd. The segments ab and cd lie outside the principal domain but reappear
as c'd' and a'b' as shown. Since, in practice we have a finite array, the line
spectrum will be broadened. To show this consider the Fourier transform of the
output of a finite ULA.

M −1 + ∞ −j
2π

∑∫
mk
− jωt
F(ω, k) = f (t, md)e dt e M

m=0 −∞

M −1 jm( ω d − 2 π k ) + ∞
= ∑e
m=0
cx M
∫f 0 (t)e − jωt dt (4.4)
−∞

ω 2π
= F0 (ω)H( d− k)
cx M

ω 2π
where H( d− k) is the response function of a ULA of length M. Note
cx M
that

© 2001 by CRC Press LLC

 ω

ω2
v
ω1

Figure 4.1: Frequency-wavenumber spectrum of a broadband plane wavefront. It lies

on a line passing through the center and has direction cosines α , β and γ . The
temporal frequency bandwidth is from ω1 to ω 2 .

ω 2π ω
H( d− k) → δ( d − u)
cx M cx
as M → ∞

2π
where k → u as M, k → ∞ . To avoid the aliasing error in the spectrum
M
the temporal sampling interval and sensor spacing must satisfy the relation
shown in (1.34b) which, when expressed in terms of ω max and d, reduces to

d c
ω max ≤ (4.5a)
π sin θ

For fixed ω max , d and c, in order to avoid aliasing error, the angle of incidence
will have to satisfy the following inequality,

π c
θ ≤ sin −1 ( ) (4.5b)
ω max d

© 2001 by CRC Press LLC

 ω

d' ωmax d

c'
c

2
π π π 2π u
d - d d d

b b'

a - ωmaxa'
Figure 4.2: Aliasing error due to spatial sampling of a broadband plane wave.

λ π
From (4.5b) it may be seen that for d= and ∆t = there is no
2 ω max
aliasing for any angle of incidence. Aliasing error will occur whenever the
above requirements are not satisfied. As an example, consider a stochastic plane
wave, incident on a ULA with 16 sensors at an angle of 45o. The sensors are
15 meters apart. The bandwidth of the waveform is ±100Hz and it is sampled
with a sampling interval of 0.005 sec. The aliasing error is present in the top
left and bottom right corners (see fig. 4.3).
4.1.3 Spectral Matrix: The output of an array of sensors may be treated as a
collection of time series or vector time series. A spectral matrix whose
elements

© 2001 by CRC Press LLC

 100

80

60
40
Frequency (Hz)

20
0

-20
-40
-60
-80

-0.1 0 0.1 0.2

Spatial frequency (u)

Figure 4.3: Aliasing error due to spatial undersampling. A stochastic plane wave is
incident on a ULA at 45°. The ULA consists of 16 sensors spaced 15 meters apart.
The wave speed is 1500m/s. The bandwidth of the waveform is ±100Hz and it is
sampled with a sampling interval of 0.005sec.

are the spectra and cross-spectra of a pair of outputs provides a complete

characterization, particularly when the outputs are Gaussian. We would like to
relate a spectral matrix to the (ω , k) spectrum. Note that the output of the mth
sensor, f m (t) , has the following spectral representation:

∞
1
∫ ∫ dF(ω,u)e
j (ωt −umd )
f m (t) = f (t, x = md) = 2 (4.6)
4π −∞

Using (4.6), the cross-covariance function between two sensor outputs is given
by
cmn (τ) = E{ f m (t) f n (t + τ)}
∞
1 (4.7)
∫ ∫ s (ω,u)e
jωτ − j ( m − n)ud
= f e dωdu
4π 2 −∞
Further, the spectral representation of a cross-covariance function [2] is

© 2001 by CRC Press LLC

 ∞
1
cmn (τ) = ∫
2π − ∞
smn (ω)e jωτ dω (4.8)

Comparing (4.7) and (4.8) we obtain the following relation between elements
of the spectral matrix and (ω , k) spectrum:

∞
1
2π −∫∞
Smn (ω) = S f (ω,u)e jd ( m − n)u du
(4.9a)
∞
1
2π −∫∞
Smm (ω) = S f (ω,u)du

The reverse relation, that is, (ω , k) spectrum in terms the elements of spectral
matrix, is

∞ ∞
S f (ω,u) = ∑ ∑S
m=−∞ n=−∞
mn (ω)e − jd ( m − n)u (4.9b)

The spectral matrix has Hermitian symmetry. Additionally, for a ULA, it has
toeplitz symmetry. Consider an example of a stochastic plane wave incident on
a ULA. The output of the mth sensor is given by

∞ d
1 jω(t − m sin θ)

2π −∫∞
f m (t) = dF(ω)e c

and the cross-covariance function between two outputs is given by

∞ d
1 jω( τ −( m − n) sin θ)

2π −∫∞
cmn (τ) = S f (ω)e c
dω (4.10)

From (4.10) the cross-spectrum between the outputs of the mth and nth sensors
is given by

d
− jω( m − n) sin θ
Smn (ω) = S f (ω)e c
(4.11a)

which may also be expressed as

© 2001 by CRC Press LLC

 ∞
1
Smn (ω) = ∫
2π − ∞
S f (ω)δ(ω − u)e − jd ( m − n)u du (4.11b)

ω
where u= sin θ . Comparing (4.11b) with (4.9) we obtain
c
S f (ω,u) = S f (ω)δ(ω − u) .
The spectral matrix for this model has a very useful representation, that
is, as an outer product of two vectors

S f (ω) = S f (ω)a(ω)a H (ω) (4.12a)

where
 − jω dc sin θ − j ( M −1)ω sin θ 
d
a(ω) = col 1,e ,...,e c

 

is the direction vector of the incident plane wave. Eq. (4.12a) may be easily
generalized for P uncorrelated sources,

P −1
S f (ω) = ∑ S f i (ω)a i (ω)a i H (ω)
i=0

which we shall express in matrix form. Define the following matrices:

{ }
S0 (ω) = diag S f 0 (ω), S f 1 (ω),..., S f P−1 (ω)
A(ω) = [a 0 (ω),a1 (ω),...,a P −1 (ω)]

The spectral matrix for a case of P uncorrelated waves and uncorrelated white
background noise is given by

S f (ω) = A(ω)S0 (ω)A H (ω) + σ 2ηI (4.12b)

The columns of matrix A(ω) possess an interesting structure (for a ULA

only), namely, each column can be expressed as powers of a constant,
col{1,µ1 ,µ 2 ,...,µ M −1} , where µ is a constant. Let µ m and µ n be the
constants corresponding to mth and nth columns. If µ m ≠ µ n for all m and n,
m ≠ n the matrix. A(ω) will have full column rank, that is, equal to P.
Such a matrix is also known as a Vandermonde matrix [3]. Even when the

© 2001 by CRC Press LLC

plane waves are correlated, as in multipath propagation, the representation of a
spectral matrix given by (4.12b) still hold good, with the only difference that
S0 (ω) is no more a diagonal matrix. The nondiagonal terms will represent
cross-spectra between the outputs of sources. An important consequence of the
nondiagonal character of S0 (ω) is the loss of the toeplitz character (see (4.9))
of the spectral matrix. Interestingly, the toeplitz character of the spectral matrix
is lost whenever the sensors of the ULA are disturbed. In a more general
situation where we have a large number of random plane waves incident on a
ULA the spectral matrix is a spatial covariance matrix at a fixed frequency.
Propagation Matrix: The matrix A(ω) is partitioned as follows:

A1 ( P × P) 
A(ω) =  
A 2 ( M − P × P) 

Since A(ω) has full column rank, there is a unique linear operator known as
propagation matrix, Γ (P × M − P) , such that Γ A1 = A 2 which may also
H

be written as

Γ 
AH   = AHQ = 0 (4.13a)
−I 

It follows that Q spans the null space of A . Now let us use the partitioned
A(ω) matrix in (4.12b)
P M − P

G1 H1  A1S0 A1 A1S0 A 2  P

H H

S f (ω) =  =  + σ 2ηI

 2 2  A 2S0 A1 A 2S0 A 2 
H M−P
G H H

Note that the spectral matrix has also been partitioned so that

G1 = A1S0 A1H + σ 2ηI P , G 2 = A 2S0 A1H

(4.13b)
H1 = A1S0 A 2H , H 2 = A 2S0 A 2H + σ 2ηI M − P

where I P stands for a unit matrix of size PxP. It may be shown from (4.13b)
that
G 2 = Γ H A1S0 A1H = Γ H (G1 − σ 2ηI P )
and hence,

© 2001 by CRC Press LLC

 Γ H = G 2 (G1 − σ 2ηI P )−1 (4.13c)

Thus, the propagation matrix may be derived from the partitioned spectral
matrix. The background noise variance is assumed to be known (see Exercises,
4.4). Q may be used to find the direction of arrival in place of eigenvectors
corresponding to noise eigenvalues as in the MUSIC algorithm to be described
later in chapter 5 [4, 5].
4.1.4 Eigenstructure: The spectral matrix possesses interesting eigenstructure.
S f (ω) is a hermitian symmetric toeplitz (only for a ULA) matrix; hence its
e(ω) be
eigenvalues are real. Further, they are positive as shown below: Let
H
some arbitrary vector and consider a quadratic form e (ω)S f (ω)e(ω) . It
follows from (4.13a) for any e(ω) ,

P −1
e H (ω)S f (ω)e(ω) = ∑ S f i (ω)e H (ω)a i (ω)a i H (ω)e(ω)
i=0
P −1
= ∑ S f i (ω) e H (ω)a i (ω) ≥ 0
2

i=0

Therefore, a spectral matrix is always positive definite or positive semidefinite

(when P<M and there is no noise), that is, all its eigenvalues are positive or
zero. Let v m , m = 0,1,..., M − 1 be the eigenvectors of S f (ω) and the
corresponding eigenvalues be λ m , m = 0,1,..., M − 1. From (4.13b) it
follows that

v mH S f (ω)v m = v mH A(ω)S0 (ω)A H (ω)v m + σ 2ηv mH Iv m

λm αm σ 2η
hence,
λ m = α m + σ 2η (4.14a)

where α m is an eigenvalue of the noise-free spectral matrix. Note that

A(ω)S0 (ω)A H (ω) is a rank P matrix, as all P columns of the Vandermonde
matrix A(ω) are independent, and S0 (ω) is a PxP diagonal matrix. Hence,
the remaining M-P eigenvalues must be equal to zero, that is,
α P ,α P +1 ,...,α M −1 = 0 . It therefore follows that

v mH A(ω)S0 (ω)A H (ω)v m = 0 m = P, P + 1,..., M − 1 (4.14b)

© 2001 by CRC Press LLC

Since S0 (ω) is positive definite, by assumption, for (4.14b) to be valid we
must have

v mH A = 0, for m = P, P + 1,..., M − 1 (4.14c)

that is, v m , m = P, P + 1,..., M − 1 are ⊥ to direction vectors of all

incident wavefronts. Equations (4.14a) and (4.14c) are of great importance as it
will be evident in the sequel (chapter 5). Some of these properties of a spectral
matrix were first noted by Nakhamkin et al. [25] in connection with the
separation of seismic wavefronts.
4.1.5 Frequency Wavenumber Spectrum: When a large number of random plane
waves are incident on an array, the (ω , k) spectrum may be computed from the
spectral matrix using (4.9b). Writing (4.9b) in a matrix form for a finite array
we obtain
S(ω,θ) = a H S f (ω)a (4.15a)

Since the above spectrum is analogous to the BT (Blackman-Tuckey) spectrum

in time series literature [2], we shall call it a BT frequency wavenumber
spectrum.
The (ω , k) spectrum turns into an angular spectrum when integrated
over the bandwidth of the incident signal for a fixed angle of propagation

ω max
1 ω
S(θ) =
2ω max ∫
−ω max
S f (ω,
c
sin θ)dω (4.15b)

In the (ω , k) domain the integration is carried out over a radial line sloping at
angle θ (see fig. 4.4). A peak in the angular spectrum is an indication of wave
energy arriving from a direction where the peak is found.
The angular spectrum defined in (4.15b) is also the beam power integrated over
the frequency bandwidth as a function of the look angle.
4.1.6 Parametric Spectrum: The frequency wavenumber spectrum defined in
(4.15a) assumes a plane wave model which is more appropriate in open space
(see chapter 1 for wavefield representation in open and bounded space). In
bounded space the wavefronts are far from planar. Such nonplanar wavefronts
may be represented in terms of source location parameters measured with
reference to the bounded space geometry. For example, in s hallow water the

© 2001 by CRC Press LLC

 ω
θ0

ωmax

0 u

Figure 4.4: In the frequency wavenumber plane the spectrum is averaged over a series
of radial lines. The spectrum of a plane wave which is incident at angle θ 0 is shown
by the bold line.

source location parameters are range, depth (measured from the surface) and
azimuth. In place of frequency wavenumber spectrum, where the parameters of
interest are frequency and wavenumbers which depend on azimuth and elevation,
we introduce a similarly defined quantity called parametric spectrum

S(ω ,θ) = a H (θ)S f (ω )a(θ) (4.16)

where θ now stands for generalized location parameters and a(θ) is the
wavefield which the array would sense if the source were to be located at θ .
S(ω ,θ) is computed over a range of values of θ spanning the entire
parameter space. The actual source position is indicated by the position of the
maximum in S(ω ,θ) . Evidently, a(θ) must be computed for the assumed
geometry and the boundary conditions there on. Since the central idea is to
match the computed field with the observed field, the processing is also known
as matched field processing. It was first introduced by Bucker [6] in 1976 and
since then a lot of research effort has been devoted toward its development as a
tool in underwater detection. An account of this effort is summarized in a
monograph [7]. The chief obstacle in the use of matched field processing lies in
the requirement of an exact knowledge of the propagation environment for the
purpose of computing the wavefield.

© 2001 by CRC Press LLC

§4.2 Beamformation:
In this section we shall deal with the estimation of the (ω , k) spectrum,
starting with the beamformation both in time and frequency domains including
fast Fourier transform (fft) based method. Next, we describe nonlinear methods:
Capon’s ML method and maximum entropy method. The nonlinear methods
provide higher resolution when the signal-to-noise ratio (snr) is reasonably
high. The importance of spectrum estimation arises on account of the fact that a
signal wavefield approaching an array of sensors in a particular direction will
produce a strong peak in the frequency wavenumber spectrum. Given the peak
position we can estimate the center frequency of radiation and the direction of
approach, that is, the directions of arrival (azimuth and elevation).
4.2.1 Beamformation: A beam in a desired direction is formed by introducing
delays before summation. The required delay per sensor in a ULA is equal to
d
τ= sin θ and in a UCA the delay for the mth sensor is equal to
c
a 2 πm
τ m = sin θ cos( − ϕ ) . In analog beamformation, introduction of
c M
continuously varying delays is achieved through analog delay lines, but in
digital beamformation the delays can be achieved only as integral steps of
λ
sampling time units. Consider a ULA with sensor spacing d equal to and
2
d
time sampling interval, ∆t , equal to . As noted in the previous section there
c
will be no aliasing, spatial or temporal for this choice of parameters. However,
we can form just one beam, namely, for θ = 0 (excluding endfire beam).
Clearly, to form more beams we need more samples between two Nyquist
samples. Assume that we have sampled at q times the Nyquist rate, that is, we
have q equispaced samples between two Nyquist samples which will enable us
i
to form beams at angles, θ i , i=0,1,..., q-1, where θ i = sin −1 ( ) . For
q
example, let q=8, the beam angles are: 0o , 7.18o , 14.48o , 22.04o , 30.0o ,
38.68 o , 48.59o , 61.04o. Evidently, only a fixed number of beams can be
formed for a given oversampling rate. It is not possible to form a beam in any
arbitrary direction. The situation with a UCA is far more difficult as for no
direction of arrival can a uniformly sampled sensor output be used for
a
beamformation. For example, consider a UCA of 16 sensors and = 8 time
c
samples. The delays to be introduced in the sensor outputs, in units of the
temporal sampling interval, for θ = 90 and ϕ = 0 are: 8.00, 7.39, 5.66,
0

3.06, 0.0, -3.06, -5.65, -7.39, -8.00, -7.39, -5.66, -3.06, 0.0, 3.06, 5.66, 7.39
(rounded to second decimal place). All these delays are with respect to a

© 2001 by CRC Press LLC

 hypothetical sensor at the center of the circle. Notice that the delays are not in
integral steps of the sampling interval. This leaves us with the only alternative
of nonuniform sampling through interpolation of uniformly sampled sensor
output. To minimize the computational load a simple linear interpolation has
been suggested [8].
Narrowband: For narrowband signals the delays applied to the sensor outputs
before summation may be expressed in terms of phase rotation. A narrowband
signal output of the mth sensor may be represented as

∆ω
ω0 +
2 d
1 jω(t − m sin θ 0 )
f m (t) ≈
2π ∫ ∆ω
Fnb (ω)e c dω
ω0 −
2
 ω (t − m d sin θ ) 
∆ω  0 0 
j c
2 d 
1  +δω(t − m sin θ 0 
)
=
2π ∫ F(ω
∆ω
0 + δω)e  c  dω (4.17a)
−
2
d
− jω 0 m sin θ 0 )
=e c f nb (t)

where the subscript nb stands for narrowband. The approximation is valid when
d
the bandwidth satisfies the condition, ∆ω (m sin θ 0 ) << 2 π for all m,
c
which implies that the time taken for a wave to sweep across the array must be
much less than the inverse of the bandwidth, expressed in Hertz. In vector
notation (4.17a) may be expressed as

f(t) = a 0 f nb (t) (4.17b)

 − jω 0 dc sin θ 0 − jω 0 ( M −1) sin θ 0 

d
where a 0 = 1,e ,...,e c
 is the direction vector on
 
the incident wavefront. The delays applied to sensor outputs may be expressed
in terms of a vector dot product. Define a vector,
 − jω 0 dc sin θ − jω 0 ( M −1) sin θ 
d
a = 1,e ,...,e c
 , known as the steering vector, which
 
d
will rotate the phase of each sensor output by an amount equal to ω 0 m sin θ
c
for the mth sensor. Thus, a narrowband beam is formed in the direction θ as

© 2001 by CRC Press LLC

 a H f(t) = a H a 0 f nb (t)

or in terms of beam power, that is, the (ω , k) spectrum is given by

{
S(ω 0 ,θ) = E a H a 0 f nb (t)
2
}= a aH
0
2
σ 2f (4.17c)

Window: The sensor outputs are often weighted before summation, the purpose
being to reduce the sidelobes of the response function just as in spectrum
estimation where a window is used to reduce the sidelobes and thereby reduce
the power leakage. As this topic is extensively covered under spectrum
estimation, for example see [2,9], we shall not pursue any further. Instead, we
like to explain the use of a weight vector to reduce the background noise
variance or to increase the snr. Let us select a weight vector, w , such that the
signal amplitude is preserved but the noise power is minimized.

w H a 0 = 1 and w H c η w = min (4.118a)

where c η is the noise covariance function. The solution to the constrained

minimization problem in (4.18a) results in

c −1
η a0
w= H −1
(4.18b)
a0 cη a0
It may be observed that for spatially white noise c η = σ 2ηI and therefore
a0
w= . In other words, the weights are simply phase shifts or delays as in
M
beamformation. The variance of the noise in the output is equal to
σ 2η̂ = σ 2η /M.
A weight vector may be chosen to maximize the snr. The output
power of an array with a weight vector w , when there is no noise, is given by
w H c s w and when there is no signal, the output power is w H c η w . We select
that weight vector which will maximize the output power ratio

w H csw
= max (with respect to w )
w H c ηw

© 2001 by CRC Press LLC

 The solution is simply the generalized eigenvector corresponding to the largest
eigenvalue of the pencil matrix [c ,c ][10]. There is a large body of
s η

knowledge on how to obtain an optimum weight vector which meets different

types of constraints. A brief review of the relevant literature is given in [10].
Rayleigh Resolution: When two wavefronts are simultaneously incident on an
array, naturally we would like to measure their directions of arrival. For this to
be possible the spectrum given by (4.17c) must show two distinct peaks. Let
f nb1 (t) and f nb2 (t) be two uncorrelated signals incident at angles θ1 and θ 2 ,
with the center frequencies being the same for both signals. The beam power is
given by

2 2
s(ω 0 , θ ) = a H a1 σ 2f 1 + a H a 2 σ 2f 2

2
In order that each signal gives rise to a distinct peak, a H a1 σ 2f 1 when plotted
2
as a function of θ should not overlap with a H a 2 σ 2f 2 . A condition for
nonoverlap is necessarily arbitrary as the array response to an incident wavefront
is strictly not limited to a fixed angular range The Rayleigh resolution criterion
states that two wavefronts are resolved when the peak of the array response due
to the first wavefront falls on the first zero of the array response due to the
λ
second wavefront. The first zero is located at an angle, sin −1 , away from
Md
the direction of arrival (broad side). Thus, two wavefronts are resolved,
according to the Rayleigh resolution criterion when their directions of arrival
λ
differ by sin −1 . An example of resolution is shown in fig. 4.5. For a UCA
Md
we can derive a simple expression when it is fully populated. In this case, its
response function is a Bessal function of 0th order (see eq. 2.43b). The first zero
of the Bessal function of 0th order is at 2.45. Two wavefronts are said to be
resolved according to the Rayleigh resolution criterion when the angular
separation is greater than sin −1 (
1. 225λ
). Let us compare the resolution
πa
properties of a ULA and a UCA having equal aperture, for example, a 16 sensor
ULA with 7.5λ aperture and the corresponding UCA with a radius equal to
3.75λ but fully populated with more than 32 sensors. The relative performance

© 2001 by CRC Press LLC

 1.0

0.8

0.6
Spectrum

0.4

0.2

0.0
-90 -70 -50 -30 -10 10 30 50 70 90
Angle in deg

Figure 4.5: Two uncorrelated wavefronts with DOA’s, 0o and 7.18o, are incident on a
16-sensor ULA. The waves are clearly resolved. The DOAs were chosen to satisfy the
Rayleigh resolution criterion.

No of Sensors Rayleigh Resolution Rayleigh Resolution

Angle in deg.(ULA) Angle in deg.(UCA)
4 30 30.61
8 14.48 12.67
16 7.18 5.84
32 3.58 2.83
64 1.79 1.39

Table 4.1: The Rayleigh resolution angle as a function of the number of sensors
λ
(ULA with sensor spacing).
2

is shown in table 4.1. The performance of the UCA is marginally better than
that of the ULA. Beamformation in the frequency domain requires the 2D
Fourier transform. For a fixed temporal frequency, the magnitude of the spatial
Fourier transform coefficients is related to the power of a wave coming from a

© 2001 by CRC Press LLC

 direction which may be computed from the spatial frequency number,
k
θ = sin −1 ( ) (see chapter 2, page 80) where k is the spatial frequency number.
M
Here too, only a finite number of fixed beams are formed. This number is equal
to the number of sensors. However, the discrete Fourier transform allows
interpolation between fixed beams through a simple means of padding zeros or
placing dummy sensors giving no output. Consider an example of a wavefront
incident at an angle of 21.06o. on a 16 sensor array. The output is subjected to
the temporal Fourier transform. The spatial Fourier transform is performed
before and after padding zeros. In fig. 4.6a the spatial Fourier transform before
padding is shown. The peak appears at frequency number 6 corresponding to an
angle of 22.02o (sin-1(6/16)). Next, the sequence is padded with 48 zeros before
Fourier transformation. The result is shown in fig. 4.6b where the peak appears
at frequency number 23 corresponding to an angle 21.06o (sin-1(23/64)) which
is the correct figure. Note that the correct peak position lies between frequency
numbers 5 and 6 (closer to 6). By padding zeros we are able to interpolate
between the frequency numbers 5 and 6 and are thus able to capture the peak at
its correct position. Further, the peak is better defined but the peak width
remains unchanged. It may be emphasized that by introducing dummy sensors
(zeros) we cannot achieve higher resolution.
Sources of Error: In practical beamformation we encounter several sources of
phase errors such as those caused by sensor position errors, variable propagation
conditions, sensor and associated electronics phase errors, quantization error in
the phase shifter, etc. The array response is highly prone to such phase errors.
Nominally, the array response may be expressed as

d d
H(ω) = a H (ω sin θ)a(ω sin θ 0 )
c c

where θ 0 is DOA of the incident wave and θ is the steering angle. We shall
model two types of phase errors, namely, those caused by position errors and
phase errors caused by all other sources lumped into one. The corrupted
direction vector has the following form:

 − jφ 0 − j (ω d +c∆d1 sin θ 0 + φ1 ) − j (ω( M −1)

d + ∆d M −1
sin θ 0 + φ M −1 ) 
ã = col e ,e ,...,e c
 (4.19)
 

where ∆d1 is the position error of the ith sensor and φ1 is the phase error. We
have assumed that the first sensor is a reference sensor and hence there is no

© 2001 by CRC Press LLC

 15

(a)

10
Magnitude

0
0 2 4 6 8 10 12 14
20

(b)
15
Magnitude

10

0
0 8 16 24 32 40 48 56
Frequency Numbers

Figure 4.6: The role of padding zeros or introducing dummy sensors is to correctly
position the peak (a) before padding zeros and (b) after padding zeros. The beam
width remains unchanged.

position error. We shall illustrate the effect of position and phase errors on the
array response function. We assume that the ULA has 16 sensors which are
λ
equispaced but with some position error. Let d= and ∆d be a uniformly
2

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

0.4

0.2

0
-100 -50 0 50 100
Angle in deg

Figure 4.7: Response of a ULA with position errors which are uniformly distributed
λ
in the range ± (solid curve). Compare this with the response of the ULA without
4
any position errors (dashed curve).

λ
distributed random variable in the range ± . The resulting response is shown
16
in fig. 4.7. The array response due to phase errors, caused by other factors, is
shown in fig. 4.8. The phase errors seem to cause less harm compared to the
position errors. The sensor position and phase errors largely affect the sidelobe
structure of the response function while the main lobe position and the width
remain unchanged.
4.2.2 Broadband Beamformation: Beamformation with a broadband signal can
also be written in a form similar to that for a narrowband signal. We must first
Fourier transform (temporal) the broadband signal output from each sensor and
treat each Fourier coefficient as a Fourier transform of a narrowband signal
whose bandwidth is approximately equal to the inverse of the time duration of
the signal. The frequency wavenumber spectrum in this case is given by

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

0.4

0.2

0
-100 -50 0 50 100
Angle in deg

Figure 4.8: Response of a ULA with phase errors which are uniformly distributed in
π
the range ±
4

{
S(ω,θ) = E a H a 0 F(ω)
2
}= a aH
0
2
S f (ω) (4.20)

By integrating over the temporal frequency as shown in (4.15b) we get an

estimate of the total power arriving from the direction θ (over an angular
interval determined by the array). We had previously called this quantity an
angular spectrum. If the objective is to estimate power received from a given
direction the angular spectrum meets the requirement. On the other hand, when
the aim is to estimate the waveform arriving from a given direction it is not
obviously enough if we estimate the power.
Delayed Snapshots: In chapter 2 we have introduced the concept of delayed
snapshots to represent a broadband wavefield. We shall make use of that
representation in devising a 2D spatio-temporal filter for beamformation. A
desired frequency wavenumber response is shown in fig. 3.6 where we have a
passband lying between two radial lines with prescribed slopes and a
horizontal

© 2001 by CRC Press LLC

 line representing the maximum temporal frequency. Let
wmn , m = 0,1,..., M − 1; and n = 0,1,..., N − 1 , where M stands for the
number of sensors and N for the number of delayed snapshots be the required
finite 2D filter whose response is as close to the desired frequency wavenumber
response as possible. A block diagram showing the filter structure is given in
fig. 4.9.
We shall express the frequency wavenumber response of the filter in a
matrix form. For this purpose we define the following quantities:

w0,0 , w1,0 ,..., wM −1,0 ;w0,1 , w1,1 ,..., wM −1,1 ;...;

w = col   (4.21a)
w0, N −1 , w1, N −1 ,..., wM −1, N −1 

1,e jud ,...,e ju( M −1)d ;e jω∆t ,e j (ud + ω∆t ) ,...,e j (ud ( M −1)+ ω∆t ) ;...;
A = col  jω ( N −1)∆t j (ud + ω ( N −1)∆t ) 
e ,e ,...,e j (ud ( M −1)+ ω ( N −1)∆t ) 
(4.21b)
It is easy to show that the response function can be expressed as an inner
product of two vectors defined in (4.21)

M −1 N −1
H(u,ω) = ∑∑w
m=0 n=0
mn e − j (umd +ωn∆t )
(4.22)
=A w H

The energy output of the filter (4.22) is given by

 1 ω2ω b H 
Output energy = w  2 ∫ ∫ AA dudω  w
H
(4.23a)
 4π ω 1 ω a 

which we like to maximize with respect to w in relation to the total energy,

 1 ω max π d H 
Total energy = w  2 ∫ ∫ AA dudω  w
H
(4.23b)
 4π ω min − π d 

where a and b are slopes of the radial lines defining the passband (see fig. 3.6),
ω 2 and ω1 are respectively the upper and the lower cut-off frequency for the
beam, and ω max and ω min refer to the maximum and the minimum frequency
present in the signal, respectively. The problem may be expressed as a problem

© 2001 by CRC Press LLC

 f0(t) f 0(t- ∆) t f 0(t-2 ∆) t f (t-3 )
∆t ∆t ∆t 0

f1(t) f1(t - ∆)t f1(t-2 ∆ ) t f (t-3 ∆ ) t

∆t ∆t ∆t 1
x w 10 x x w 12 x w
w 11 13
+ + + +
Sensor array

f7(t) f7(t - ∆ ) t f7(t-2 ∆ ) t f 7 (t-3 ∆ ) t

∆t 1 ∆t ∆t
x w 70 x w 71 x w 72 x w 73
+ + + +

Figure 4.9: The structure of a 2D filter for broadband beamformation. M=8 and N=4.

w H Γ 1w
in maximization of a ratio, = λ , which is achieved by solving the
w HΓ2w
following generalized eigendecomposition problem:

Γ 1w = λ Γ 2 w (4.24)

ω2ω b ω max π d
1 1
where Γ1 = ∫ω ω∫aAA dudω and Γ 2 = 4 π 2 ∫ ∫ AA dudω . The
H H

4π 2 1 ω min − π d
solution is given by the eigenvector corresponding to the largest eigenvalue and
the maximum relative power is equal to the largest eigenvalue.
To evaluate Γ1 and Γ 2 , we must first simplify the elements of
AA H ,

© 2001 by CRC Press LLC

 [AA ]
H
pq
= e j ( p1ud + p2 ω∆t )e − j (q1ud + q2 ω∆t )
j ( ( p1 − q1 )ud +( p 2 − q 2 )ω∆t )
=e

p1 = p − Int   M , p2 = Int   and Int[x] stands for the largest

p p
where
M M
integer less than or equal to x. q1 and q2 are similarly defined. The elements of
Γ1 may be evaluated as in (3.12a) and those of Γ 2 are given below:

 ω max sin c(( p − q )ω ∆t) 

1  π 2 2 max 
[Γ 2 ]p,q = sin c(( p1 − q1 )π )   (4.25)
d − ω min sin c(( p − q )ω ∆t)
 π 2 2 min


The principle of maximizing the power in the passband was first suggested in
[11] in the context of optimum window for spectrum estimation of time series.
Later, this principle with additional constraints on the magnitude and derivative
has been applied in beamformation [12, 13]. In fig. 4.10 we show a numerical
example of spatio-temporal filter for broadband beamformation. The desired
response is unity in the region bounded by two radial lines and the upper and
the lower frequency cutoff lines. The actual response of the maximum energy
filter is contoured in the same figure. It is observed that the maximum side lobe
level is about 4dB less than that in the simple quadrant filter shown in fig. 3.7.
§4.3 Capon’s ω-k Spectrum:
We consider a stochastic plane wave incident on a ULA. Let us represent the
array output in a matrix form,

∞
1
f(t) = ∫
2π − ∞
dF(ω)e jωt (4.26a)

where
d
dF(ω) = dF 0 (ω)a(ω sin θ) + dη(ω) (4.26b)
c

where dF 0 (ω) is the generalized Fourier transform of the stochastic

d
waveform, a(ω sin θ) is the direction vector and dη(ω) is the background
c

© 2001 by CRC Press LLC

noise. From (4.26) and using the properties of the generalized Fourier transform
we can derive an expression for the spectral matrix
π
3

2
0.5
0.9

1
Temporal frequency

0.1

0.1
-1 0.5
0.9

-2

-3
−π
-3 -2 -1 0 1 2 3 π
Spatial frequency

Figure 4.10: Response of filter for broadband beamformation. The slopes of the
radial lines are a=1 (45o ) and b=2 (63.4o ). ω1 = 0 and ω 2 = 0.6π . The maximum
energy in the passband is 94% ( λ max = 0.94). 16 sensors and 16 delayed samples.
∆x = 1 and ∆t = 1.

d d
S f (ω) = S0 (ω)a(ω sin θ)a H (ω sin θ) + S η (ω)
c c

where both S f (ω) and S η (ω) are MxM matrices but S0 (ω) is a scalar. We
like to find a weight vector w acting on the array output such that it
minimizes the power output of the array and is transparent to all waves
propagating through a narrow cone with a cone angle ∆θ and its axis pointing
in the direction of θ 0 . Thus, the beamwidth is made intentionally wider to
allow for possible variation in the direction of arrival. This model is useful

© 2001 by CRC Press LLC

 when the direction of the incoming wavefront is likely to be different from the
assumed or given direction. Translated into a mathematical statement we obtain

w H S f (ω)w = min
(4.27a)
w H Γw = 1
where
∆θ
θ0 +
2
1 d d
Γ=
∆θ ∫ ∆θ
a(ω sin θ)a H (ω sin θ)dθ
c c
(4.27b)
θ0 −
2

and S f (ω) is the array signal spectrum under the assumption that the source
bearing lies in the range θ 0 ± ∆θ .
The constrained minimization problem specified in (4.27) is solved by
the Lagrange method,

w H S f (ω)w + λ(1 − w H Γw) = min (4.28)

Differentiating (4.28) with respect to w and setting the derivative to zero we

obtain

S f w = λΓw (4.29a)
or
λ−1w = S−1
f Γw (4.29b)

From (4.29b) it is clear that w is an eigenvector of S−1

f Γ and the

corresponding eigenvalue is λ−1 . Note that from (4.29) w H S f w = λ , that is,

equal to the output power of the array weighted by vector w . In order that the
−1
array output power is minimum we must select the largest eigenvalue of S f Γ
and the corresponding eigenvector as the weight vector, w .
Special Case: Let ∆θ =0, that is, the beamwidth is zero. Therefore,

d d
Γ = a(ω sin θ 0 )a H (ω sin θ 0 )
c c
Equation (4.29b) now becomes

© 2001 by CRC Press LLC

 d d
λ−1w = S−1
f a(ω sin θ 0 )a H (ω sin θ 0 )w (4.30)
c c

d
By premultiplying both sides by a H (ω sin θ 0 ) we find that
c

d d
λ−1 = a H (ω sin θ 0 )S−1
f a(ω sin θ 0 ) (4.31a)
c c
It turns out that
d
S−1
f a(ω sin θ 0 )
w= c (4.31b)
d d
a H (ω sin θ 0 )S−1
f a(ω sin θ 0 )
c c

satisfies (4.30). We can express the array output power, which we shall call as
Capon spectrum,

1
sCap (ω,θ 0 ) = (4.32)
d d
a (ω sin θ 0 )S−1
H
f a(ω sin θ 0 )
c c

Capon [14], who first suggested the above measure of spectrum, however,
called it maximum likelihood spectrum. It is also known as the minimum
variance distortionless response (MVDR) beamformer or a linearly constrained
minimum variance (LCMV) beamformer [15]. Since θ is related to the spatial
d
frequency, u = ω sin θ , sCap (ω,θ) is indeed a (ω , k) spectrum as a
c
function of θ or u.
4.3.1 Resolution: The Capon spectrum has a better resolution compared to the
BT ω-k spectrum. We shall demonstrate this by considering two uncorrelated
wavefronts in the presence of white noise. The spectral matrix is given by

S f (ω) = s0 a 0 (ω,θ 0 )a 0H (ω,θ 0 ) + s1a1 (ω,θ1 )a1H (ω,θ1 ) + σ 2ηI (4.33)

where θ 0 and θ1 are directions of arrival and s0 and s1 are powers of two
plane wavefronts and σ η is noise variance. The inverse of the spectral matrix
2

in (4.33) may be computed following the procedure described in [16].

© 2001 by CRC Press LLC

 −1 V −1a1 (ω,θ1 )a1 H (ω,θ1 )V − H
−1
S (ω) = V − s1 (4.34a)
1 + s1a1 H (ω,θ1 )V −1a1 (ω,θ1 )
f

where

 s0
a 0 (ω,θ 0 )a 0 H (ω,θ 0 ) 
1  σ 2 
V −1 = 2 I − η  (4.34b)
ση  s0
1+ 2 M 
 ση 

Using (4.34a) in (4.32) we obtain the Capon spectrum for the two source
model,

σ 2η
sCap (ω, θ) = 2 (4.35)
s0 H H
2 a a 0 a 0 a1
s1 H σ η
a a1 −
s0 H 2 σ 2η s
1 + 02 M
a a0
σ 2η ση
M− −
s
1 + 02 M  s0 H 2 
ση  2 a 0 a1 
s1 σ η
1+ 2 M − 
ση  1 +
s0
M 
 σ 2η 
 
where for the sake of compactness we have dropped the arguments of the vectors
a , a 0 , and a1 . When the steering vector points to one of the sources, for
example, when a = a0

σ 2η σ 2η
sCap (ω, θ 0 ) = 2 ≈ s0 + (4.36a)
s0 H M
2 Ma 0 a1
s1 H σ η
a 0 a1 −
σ 2η s
1 + 02 M
M ση
−
s
1 + 02 M  s0 H 2 
ση  2 a 0 a1 
s1 σ η
1+ 2 M − 
ση  1 +
s0
M 
 σ 2η 
 
and when a = a1

© 2001 by CRC Press LLC

  s0 H 2 
a 0 a1
s1  σ 2η 
1+ 2 M − 
ση  s
1 + 02 M 
 ση  σ 2η
sCap (ω,θ1 ) = σ η
2
≈ s1 + (4.36b)
s0 H 2 M
a a
σ 2η
1 0

M−
s
1 + 02 M
ση

The approximation shown in (4.36) is valid for a1H a 0 << M . From the
above it is clear that when the wavefronts are well resolved the peak amplitude
approximately equals the power of the source. The noise power is reduced by a
factor equal to the number of sensors.
We will examine the resolution power of the Capon spectrum.
Consider again two equal power wavefronts incident at angles θ 0 and θ1 . The
peaks corresponding to two wavefronts are resolved when a valley is formed in
between them. Let sCap (ω , θ̃ ) be the spectrum at θ̃ midway between θ 0 and
θ1 . Define the ratio ρ as

sCap (ω,θ 0 )
ρ=
sCap (ω, θ̃)
  a 0H a1
2
ã H a1 
2

  1 − α − (1 + α) 2
 s1 M  M M  ã H
a 
1 + 2  −α 1

 ση   ã a1a1 a 0 a 0 ã  
H H H M  (4.37a)
  +2α Re    
   M3  
= H 2
a a
1− α 0 1
M

sM
Assuming 1 >> 1 (4.37a) simplifies to
σ2
η

© 2001 by CRC Press LLC

   2 2
 ã H a a H a a H ã  
 s1M  aHa ã H a
1 + 2 1 −
0 1 −2 1 + 2 Re  1 1 0 0  
 σ  M M  M 3  
η   
≈ ,
H 2
a a
1− 0 1
M
(4.37b)

s1 M sM
where α = ≈ 1 for 1 2 >> 1. A valley is formed iff ρ > 1. Let
1 + s1 M ση
a 0 = a1 = ã , that is, when two wavefronts merge into a single wavefront we
notice that ρ = 1 for all snr, which means that these two wavefronts can never
2 2
aHa ã H a1
be resolved. Next, let 0 1 = ≈ 0 which means that the wavefronts
M M
sM
are well separated. Then, ρ = 1 + 1 2 >1 except when s1 =0. The wavefronts
ση
can then always be resolved. All the above conclusions follow from common
∆θ
sense. We now consider two wavefronts with DOAs ± respectively and
2
sM
compute ρ for different 1 2 and ∆θ . A plot of ∆θ for which ρ is just
ση
greater than one as a function of array snr is shown in fig. 4.11.
4.3.2 Robust Beamformation: The sensitivity of beamformation to errors in the
sensor position and other phase errors has been demonstrated in figs. 4.7 and
4.8. These drawbacks may be reduced through an appropriate choice of
weighting coefficients. In this section we shall show how such coefficients can
be obtained [17] following a constraint used in deriving Capon’s filter and the
{ }
associated spectrum. Let w = col w0 , w1 ,...wM −1 be a coefficient vector.
The array response may be expressed as

d
H(ω) = w H a(ω sin θ) (4.38)
c
We shall model two types of phase errors, namely, those caused by position
errors and those caused by all other sources of errors lumped into a single phase

© 2001 by CRC Press LLC

 10

6
Angle in deg.

0
.01 .1 1 10 100
Array SNR

Figure 4.11: Resolution properties of the Capon spectrum as a function of array snr.
A 16 sensor ULA is assumed. The angles of incidence are 30o and 30o + angular
separation as shown on the y-axis. Simple beamformation (BT ω-k Spectrum) yields
a resolution of 7.18o shown in the figure by a thick line for comparison. Resolution
gain by the Capon Spectrum is possible only for high array snr.

error. The steering vector given by (4.19). We have assumed that the first
sensor is a reference sensor and hence there is no position error. Let H0 be the
desired response, for example, equal to 1 and H̃ be corrupted response,

H̃ = w H ã (4.39)

The weighting coefficients are selected so as to minimize the mean square

difference between H0 and H̃ , defined as
2
mse = ∫ ... ∫ Ω(ζ 0 ,ζ1 ,...,ζ M −1 ) H0 − w H ã dζ 0 dζ1 ...dζ M −1 (4.40a)

© 2001 by CRC Press LLC

 where Ω(ζ 0 ,ζ1 ,...,ζ M −1 ) is the probability density function of the random
variables appearing in (4.40a). We can rewrite (4.40a) in a compact form

2
mse = w H Qw − (H0 P H w + H0H w H P) + H0 (4.40b)
where

P = ∫ ... ∫ Ω(ζ 0 ,ζ1 ,...,ζ M −1 )ãd ζ 0 d ζ1 ...d ζ M −1

Q = ∫ ... ∫ Ω(ζ 0 ,ζ1 ,...,ζ M −1 )ãã H d ζ 0 d ζ1 ...d ζ M −1

The mean square error is minimum for w 0 which is a solution of the

following equation:

Qw 0 = P (4.41)

We shall rewrite (4.40b) in terms of w0

mse = ( w 0 − w ) Q( w 0 − w ) + H0 − w 0H Qw 0
H 2
(4.42)

An increased robustness in beamformation is sought by requiring that the

output power of the beamformer be minimum [17],

H
Output power= w S f w =min (4.43a)

subject to a quadratic constraint on the weight vector, namely,

( w 0 − w ) H Q( w 0 − w ) ≤ ε 2 (4.43b)

2
where ε 2 = mse − H0 + w 0H Qw 0 is a prescribed number which represents
an excess error over the minimum that can be achieved by satisfying (4.41).
Note that S f is a spectral matrix of the array output. Using the standard
primal-dual method we can solve the constrained optimization problem [18].
The solution is given by

w = w 0 − (S f + λ Q)−1 S f w 0 (4.44)

© 2001 by CRC Press LLC

where λ is a Lagrange multiplier and it is given as a root of the following
rational function:

w 0H S f (S f + λ Q)−1 Q(S f + λ Q)−1 S f w 0 = ε 2 (4.45)

A simple approach to solve for λ is to plot λ vs ε 2 for the given S f , P and

Q and pick a value of λ for a prescribed ε . It is demonstrated in [17] that for
2

a circular array consisting of two concentric rings with random interring spacing
the array gain remains practically unaltered if the actual spacing is well within
the bounds used in the design of the weight coefficients. However, the sidelobe
characteristics of the response function of the weight coefficients are not
known.
4.3.3 High Resolution Capon Spectrum: The resolution capability of Capon’s
frequency wavenumber spectrum may be improved by noise cancellation by
subtracting an estimated white noise power from the diagonal elements of the
spectral matrix [19] and stretching the eigenvalue spread of the spectral matrix
[20]. A predetermined quantity is subtracted from the diagonal elements of the
spectral matrix, thus increasing the ratio between the maximum and minimum
eigenvalues. This process is called stretching the eigenvalue spread. The spectral
matrix must, however, remain positive definite. Consider a model of P plane
waves and white background noise. The spectral matrix, given in (4.12b), is
subjected to stretching of the eigenvalue spread by subtracting a fixed number
σ 20 from the diagonal elements,

S̃ f (ω) = S f (ω) − σ 20 I
(4.46)
= vΓ 0 v H

where
{ }
Γ 0 = diag α m + σ 2η − σ 20 , m = 1,2,..., M − 1
= diag{γ m , m = 1,2,..., M − 1}

We now introduce an improved Capon spectrum by using the spectral matrix

whose eigenvalue spread has been stretched in (4.32) to obtain

© 2001 by CRC Press LLC

 1
s̃Cap (ω,θ) =
d d
a H (ω sin θ)S̃−1
f a(ω sin θ)
c c
(4.47)
1 1
= H −1 H = M −1
a vΓ 0 v a 1
∑
2
a H vm
m=0 γ m

By selecting σ 02 as close to σ η2 as possible we can make γ m ≈ 0 , m=P,P+1,

..., M-1 and γ m ≈ α m , m = 0,1,..., P − 1. Next we note that from (4.14c)
a H v m = 0 for m=P, P+1, ...,M-1 and when a l is equal to one of the direction
vectors of the incident wavefronts. Conversely, when a does not belong to that
set of direction vectors, a v m ≠ 0 . Let us rewrite (4.47) as
H

1
s̃Cap (ω , θ ) = P −1 (4.48)
1 H 2 M −1 1 H 2
∑ a v m + m∑= P γ a v m
m=0 γ m m

In the denominator of (4.48) the second term dominates whenever a does not
belong to the set of direction vectors of the incident wavefronts and vice versa.
Hence,

s̃Cap (ω , θ ) = sCap (ω , θ ), θ = θ i i = 0,1,..., P − 1

and
s̃Cap (ω,θ) ≈ 0 , θ ≠ θi

2
As an example, consider a single source, that is, P=1, then a H v 0 = M and
α 0 = Mp0 . On account of (4.14c) a H v m = 0 (but γ m ≠ 0 ) for m=1, 2,...,
M-1. Therefore, s̃Cap (ω,θ 0 ) = p0 (power of the incident wavefront). For all
other values of θ s̃Cap (ω,θ) ≈ 0 . To demonstrate the resolution power of the
high resolution Capon spectrum we consider a 16 sensor ULA and two
wavefronts incident at angles 30o and 35o and 0dB snr. The resolution of the
Capon spectrum as shown in fig. 4.12 has dramatically improved when the
eigenvalue spread is increased from 30 to 1000. The above result is for an error-
free spectral matrix, that is, with infinite data. To study the effect of finite data

© 2001 by CRC Press LLC

 1.8
1.6

1.4 a

1.2
Spectrum

1
0.8

0.6

0.4
0.2
0
10 20 30 40 50

1.2

1.0
1

0.8
Spectrum

0.6

0.4

0.2

0
10 20 30 40 50
DOA in deg

Figure 4.12: (a) Capon spectrum. (b) High resolution Capon spectrum. Eigenvalue
spread is 1000. 16 sensor ULA with sensor spacing of λ/2 and two uncorrelated
wavefronts incident at 30o and 33o were assumed. The high resolution Capon
spectrum yields correct amplitude and DOA information.

a numerical experiment was carried out [20] on an eight sensor ULA with two
wavefronts incident at angles 45o and 53o in presence of white background

© 2001 by CRC Press LLC

 noise (-5dB). The spectral matrix was computed by averaging over ten
independent segments. The Capon spectrum is barely able to resolve the peaks
but the high resolution Capon spectrum shows two clear peaks (see fig. 2 in
[20]). The minimum resolvable angular separation between two uncorrelated
wavefronts as a function of snr is shown in fig. 4.13. The spectral matrix was
computed using 10 data segments. The eigenvalue spread was set at 2000 (note
that the Rayleigh resolution is equal to 16.6o). Further, a study of the bias, the
mean square error (mse) and the probability of resolution was also carried out in
the above experiment. The results are summarized in table 4.2. The total bias is
a sum of the bias (magnitude) in both peaks. Similarly, the total standard
deviation is a sum of the standard deviations of both peaks. The probability of
resolution was computed by noting the number of times the peaks were clearly
resolved in a hundred trials. Finally, by increasing the eigenvalue spread the
spectral peaks become sharper but soon instability sets in while inverting the
modified spectral matrix.
§4.4 Maximum Entropy ω-k Spectrum:
The maximum entropy (ME) spectrum is another example of the nonlinear
spectrum estimation method, originally developed for time series. The basic
idea is to find a frequency wavenumber spectrum which is consistent with the
observed spectral matrix of the array signal but is maximally noncommittal on
the wavefield which has not been observed simply because the array happened to
be finite. The requirement of being maximally noncommittal is translated into
maximization of entropy [2]. As in time series, there is an implied prediction of
the wavefield outside the array aperture. Many of the properties of the ME
spectrum of time series naturally hold good in the wavefield analysis. We shall
elaborate in this section some of these aspects of the ME spectrum.
4.4.1 Maximum Entropy: We shall start with an assumption that the spectral
matrix of the array signal is available. An element of the spectral matrix and the
frequency wavenumber spectrum are related through (4.9a), reproduced here for
convenience

∞
1
Smn (ω) = ∫
2π − ∞
S f (ω,u)e − jd ( m − n)u du (4.9a)

The entropy gain is given by

∞
1
4π −∫∞
∆H = log S f (ω,u)du (4.49)

© 2001 by CRC Press LLC

 15

13

11
Angle in deg.

5
-15 -10 -5 0
SNR in dB

Figure 4.13: Resolution capability of high resolution Capon spectrum as a function

of snr. Thick line: High resolution Capon spectrum and Thin line: Capon spectrum.

Method Total bias Total std. Probability of

dev. resolution
Capon spectrum 1.420 0.860 58.3%
High resolution Capon 0.690 0.800 92.2%
spectrum
Table 4.2: Comparison of performance of Capon spectrum and its high resolution
version. Parameters: 8 sensor ULA, two wavefronts with angular separation = 6o
incident on broadside, snr=0 dB, number of segments used =50, number of
trials=500 and eigenvalue spread= 2000.

It is proposed to find S f (ω,u) which satisfies (4.9a) and maximizes the

entropy gain (4.49). We shall keep the temporal frequency ω fixed throughout.
The above statement of estimation using the principle of maximum entropy is
exactly similar to that in the estimation of time series [2]. The approach to
maximization is essentially the same as in the time series case. We shall only
briefly outline the approach leaving out all the details which may be found in

© 2001 by CRC Press LLC

 [2]. Maximization of entropy is achieved when the spectrum can be expressed in
a parametric form,

1
S f (ω,u) = p= p 0 (4.50)
∑λ e
p= − p 0
p
jup

where λ p , p=0, ± 1, ± 2, ± p0 are Lagrange coefficients. Let us further

express the denominator in (4.50) as

1
S f (ω,u) =
H(u)H H (u)
1 (4.51)
= p0 p0

∑h e ∑h e
m=0
m
jum

m=0
m
− jum

The phase of H(u) is yet to be specified. We will choose H(u) as a

minimum phase function whose all poles and zeros are to the left of imaginary
axis in the complex frequency plane. The coefficients hm , m = 0,1,..., p0 may
be obtained by solving the following equation,

1
SfH = δ (4.52a)
h0

where S f is a spectral matrix, H = col h0 ,h1 ,h2 ,...,hp0 { } and

δ = col{1,0,0,...,0} . The solution of (4.52a) is given by

1 −1
H= Sf δ (4.52b)
h0

Notice that S−1 −1

f δ refers to the first column of S f and h0 is the first element of
that column. Using (4.52b) in (4.51) we can express the maximum entropy
spectrum as

© 2001 by CRC Press LLC

 h02
SME (ω,u) = 2 (4.52c)
a H S−1
f δ

where { }
a = col 1,e ju ,e j 2u ,...,e jp0 u is the steering vector.
We shall look at an alternate approach which will lead to an equation
identical to (4.52b). At a fixed temporal frequency the output of an array (ULA)
may be expressed as a sum of complex sinusoids,

∞ p d
1 0 jω(t − m sin θ p )
f m (t) = ∫ ∑
2π − ∞ p= 0
dGp (ω)e c
(4.53a)

Let

P ωd
− jm sin θ p
dFm (ω) = ∑ dGp (ω)e c
, m = 0,1,..., M − 1 (4.53b)
p= 0

where we have assumed that P plane wavefronts are incident on a ULA. The
sources are assumed to radiate stationary stochastic but uncorrelated signals.
Clearly, (4.53b) is a sum of P random spatial complex sinusoids. A sum of
P random sinusoids (real) are known as a deterministic random process and it
can be predicted without error from 2 P past samples [2]. In the case of P
complex sinusoids we will require P past samples for prediction. Error-free
prediction is not possible when there is background noise. The prediction
equation is given by

P
dFm (ω) + ∑ hp dFm − p (ω) = ηm (4.54a)
p=1

where ηm is the prediction error. We express (4.54a) in a matrix form,

H H dF m (ω) = ηm (4.54b)

where {
dF m (ω) = col dF m , dF m −1 ,..., dF m − p0 } and h =1. The prediction
0

error is given by

H H S f (ω)H = σ 2η (4.55)

© 2001 by CRC Press LLC

 which we wish to minimize, subject to the constraint that h0 =1 or H H δ = 1 .
This leads to the following equation:

S f (ω)H = σ 2ηδ
or
H = σ 2ηS−1
f (ω)δ (4.56)

which is identical to (4.52b), except for a scale factor. Using (4.56) in (4.55)
1
the minimum prediction error is equal to σ 2η = −1
. The
min δ S (ω)δ
H
f
prediction filter vector corresponding to minimum error is given by

S−1 (ω)δ
H = H f −1 (4.57)
δ S f (ω)δ

The maximum entropy spectrum defined in (4.51) may be expressed as follows:

(δ )
2
H
S−1
f (ω)δ
SME (ω,u) = 2 (4.58)
a H S−1
f (ω)δ

Note that the numerator of (4.58) is equal to the first element of the first
−1
column of S f . The maximum entropy spectrum given by (4.52c) and the
spectrum obtained by minimizing the prediction error, (4.58), are identical.
Thus, the maximum entropy principle leads to a simple interpretation in the
form of linear prediction.
4.4.2 Resolution: We shall now examine some of the properties of the
maximum entropy spectrum and compare them with those of the Capon
spectrum. As earlier we shall consider two wavefronts incident on a ULA. The
spectral matrix is given by (4.33). Using the inversion formula (4.34) in
(4.52c) we obtain an expression for the maximum entropy spectrum for two
wavefronts in presence of white noise.

h02 σ 4η
sME (ω,θ) = (4.59)
den
where

© 2001 by CRC Press LLC

 2
 a a 0 a 0H a1  
s0 H s0 H 
a a
s1  H ση  σ2 1 0 
2
η
 a a1 −  1 − 
s0 H
a a0 σ 2
η  1 +
s0
M   1 +
s0
M 
σ 2η  σ 2η   σ 2η 
den= 1 − −
s
1 + 02 M  s0 H 2 
a 0 a1
ση s1  σ 2η 
1+ 2 M − 
ση  s
1 + 02 M 
 ση 

where all symbols are as in (4.35). An example of maximum entropy of two

uncorrelated wavefronts incident at 30o and 35o on a 16 sensor ULA is shown
in fig 4.14a and the corresponding Capon spectrum is shown in fig. 4.14b. In
both cases we have assumed white background noise and snr=1.0.
Let us now evaluate (4.59) at θ = θ 0 and θ = θ1 under the
2
assumption that a 0H a1 ≈ 0 . We obtain

sME (ω,θ 0 ) = h02 (σ 2η + s0 M)2 (4.60a)

sME (ω,θ1 ) = h02 (σ 2η + s1 M)2 (4.60b)

The height of the spectral peak grows with the array signal-to-noise ratio,
increasing to infinity as sM → ∞ . This is demonstrated in fig. 4.15 for the
two wavefront model, whose spectrum is plotted in fig. 4.14.
From figs. 4.14a and 4.14b it may be conjectured that the maximum
entropy spectrum has a better resolution property than that of the Capon
spectrum. The depth of the valley for the maximum entropy spectrum is much
larger than the one for the Capon spectrum. We have carried out a series
computation to find out the minimum snr required to resolve two equal
amplitude wavefronts separated by a specified angular distance. The criterion for
resolution was formation of a nascent valley between the spectral peaks. While
this is not a quantitative criterion it serves the purpose of comparison. The
results are shown in fig. 4.16 which may be compared with fig. 4.11 for the
Capon spectrum. Clearly, the maximum entropy spectrum has a better
resolution capability, but its peak amplitude does not bear any simple relation
to the actual spectrum value. In the time series context it was shown in [21]
that the Capon spectrum and maximum entropy spectrum of different orders are
related,

© 2001 by CRC Press LLC

 100

80

60
Spectrum

40

20

0
0 20 40 60
DOA in deg
1.2

0.8
Spectrum

0.6

0.4

0.2

0
0 20 40 60
DOA in deg

Figure 4.14: (a) Maximum entropy spectrum and (b) Capon spectrum. Two unit
amplitude plane wavefronts are incident at angles 30o and 35o on a 16 sensor ULA.
The amplitude of the peaks of Capon spectrum is close to the actual amplitude but the
amplitude of ME spectrum is much higher. According to (4.53), valid for large
separation, the computed amplitude is equal to 200.

© 2001 by CRC Press LLC

 10 6

10 5

10 4
Peak spectrum

10 3

10 2

10 1

10 0
0 20 40 60 80 100
Array SNR

Figure 4.15: Peak spectrum as a function of array snr, that is, Msnr. Two equal
amplitude uncorrelated wavefronts incident at 30o and 35o on a 16 sensor ULA.

1 1 M −1 1
= ∑ m (ω)
SCap (ω) M m = 0 SME
(4.61)

where M stands for the size of the covariance matrix. In array processing, M
stands for array size.
4.4.3 Finite Data Effects: So far we have tacitly assumed that the spectral
matrix is known and the incident wavefield and noise confirm with the assumed
model; for example, the wavefronts are planar and uncorrelated and the noise is
white. In practice, however, the spectral matrix needs to be computed from the
available data. Since the spectral matrix is a statistically defined quantity
involving the operation of expectation, there is bound to be some error in its
estimation when only finite length data is available. The effect of the errors in
the spectral matrix on wavenumber spectrum has been a subject of investigation
by many researchers [22, 23]. Here we shall briefly outline the important
results. The mean and variance of the BT frequency wavenumber spectrum
(linear), the Capon frequency wavenumber spectrum (nonlinear), are tabulated in

© 2001 by CRC Press LLC

 8

6
Angle in deg.

0
.01 .1 1 10 100
Array SNR

Figure 4.16: Resolution property of ME spectrum. Two equal amplitude uncorrelated

wavefronts are incident on a 16 sensor ULA. The directions of arrival are 30o and
30 o + the angular separation, as shown on the y-axis.

Method Mean Variance

BT spectrum S(ω,θ) S 2 (ω,θ)
N
Capon spectrum N − M +1 N − M +1 2
SCap (ω,θ) SCap (ω,θ)
N N2
Table 4.3 Mean and variance of frequency wavenumber spectrum estimated from
finite data. N stands for number of independent snapshots and M for array size.

table 4.3. Unfortunately, we do not have simple expressions for mean and
variance of ME spectrum estimates; only experimental results are available
[22]. The ME spectrum is more variable and hence it needs much larger data to

© 2001 by CRC Press LLC

stabilize. It is reported in [22] that to get a stable ME spectrum the number of
2
snapshots must be far greater than M and M SNR . the effect of finite data
will result in (a) a loss in resolution, that is, closely spaced wavefronts cannot
be resolved; (b) a shift in the position of peaks, that is, an erroneous estimate
of the direction of arrival.
4.4.4 Iterative Inversion of Spectral Matrix: The most important computational
step in Capon and ME methods is the computation of an inverse of the spectral
matrix (covariance matrix for narrowband signal). The spectral matrix is
computed as an average of the outer product of the frequency snapshots. The
output of an array is divided into equal duration and overlapping segments and
each one is Fourier transformed. A frequency snapshot of an array, analogous to
the time snapshot, is an output of an array at a fixed (temporal) frequency. Let
Fi (ω) be the ith frequency snapshot obtained by Fourier transforming the ith
segment of the array output. The spectral matrix is estimated as

1 N
Ŝ f (ω) = ∑
N 1
Fi (ω)FiH (ω)

When a new time snapshot arrives a new segment is formed with the newly
received snapshot and the past snapshots to form a required length segment and
then a new frequency snapshot is formed. The spectral matrix is updated by
incorporating the outer product of the newly formed frequency snapshot,

1 N 
Ŝ Nf +1 (ω) =  ∑
N +1 1
Fi (ω)FiH (ω) + F N +1 (ω)F HN +1 (ω)
 (4.62a)
N 1
= Ŝ Nf (ω) + F N +1 (ω)F HN +1 (ω)
N +1 N +1
The recursion may be commenced with an initial value
Ŝ (ω) = F1 (ω)F (ω) . We can obtain a recursive expression for the (ω , k)
1
f
H
1
spectrum by using (4.62a) in (4.16)

 N S N (ω,θ) + 
N +1 
N +1
S (ω,θ) =  2
 (4.62b)
 1 a H (ω d sin θ)F (ω) 
 N + 1 c
N +1 

Using the matrix inversion formula given in (4.34) we can get a recursive
relation between the inverse of spectral matrix Ŝ Nf (ω) and Ŝ Nf +1 (ω)

© 2001 by CRC Press LLC

 [ ]
−1
Ŝ Nf +1 (ω) =


[ ] [ ]  (4.62c)
1 N −1 −H
Ŝ f (ω) F N +1 (ω)F HN +1 (ω) Ŝ Nf (ω)
N +1 N
[ ] 
−1
 Ŝ f (ω) − N 
[ ]
1 −1
N  1 + F HN +1 (ω) Ŝ Nf (ω) F N +1 (ω) 
 N 

The above recursion can be commenced only after Ŝ Nf (ω) becomes full rank.
This will require a minimum of M frequency snapshots. There is yet another
recursive approach to spectral matrix inversion. It is based on diagonalization of
the spectral matrix through a transformation,

Γ = QŜ Nf (ω)Q H (4.63a)

or equivalently,

Ŝ Nf (ω) = QΓQ H (4.63b)

where Γ = diag{γ 1 , γ 1 ,... γ M } and Q is the upper triangular matrix of

prediction coefficients which are computed through a recursive algorithm [24].
From (4.63b) we can express the inverse of the spectral matrix as

[Ŝ (ω)]
−1
N
f = QΓ −1Q H (4.64)

§4.5 Exercises:
1. The spatial undersampling has resulted into aliasing of the frequency
wavenumber spectrum as illustrated in figs. 4.2 and 4.3. Now consider temporal
undersampling. Sketch the (ω , k) spectrum of a stochastic plane wave which
has been undersampled temporally.
2. Apply the principle of Rayleigh resolution to wideband signals. In chapter 2
we have shown how the DOA of an incident wideband signal can be estimated
from the position of the spectral peaks of the transfer functions. Show that for
resolution the wavefronts must be separated by an angle greater than ∆θ ,
where

tan θ
∆θ =
M

© 2001 by CRC Press LLC

3. Show that the power output of a beamformer given by (4.16) is always
greater than the Capon spectrum. Use Schwarze inequality (see, p.19, [23]).
4. The noise variance is often required to be known but in practice this is not
likely to be true. It has to be estimated from the observed data. One possible
approach is to use the partitioning of the spectral matrix as on page 215. Show
that

tr(H 2 π)
σ 2η =
tr(π)

where π = Ι M-P − G 2 G 2# and G 2# is a pseudoinverse of G 2 .

5. A plane wavefront is incident on a ULA from a known direction θ 0 in
presence of coloured background noise. Consider the following noise reduction
strategy. The array is steered to the known direction along with a weight vector.
The steering vector is w1 = Γw 0 where w 0 is an unknown weight vector and

 − jω d sin θ 0 − j ( M −1)ω sin θ 0 

d

Γ = diag 1, e c ,..., e c

 

It is proposed to findw 0 which shall minimize the noise variance and at the
same time preserve the signal power. Show that w 0 is equal to the eigenvector
corresponding to the least eigenvalue of Γ S ηΓ where S η is the spectral
H

matrix of the noise.

6. It was shown on page 222 that the weight vector which maximizes the snr is
the eigenvector corresponding to the largest eigenvalue of a pencil matrix
[c ,c ]. Let a single wavefront be incident on the array. The covariance
s η

matrix, when there is no noise, is c s = s0 a 0 a 0H . Show that w = αc −1

η a0
where α is a normalizing constant and the maximum snr is equal to
s0 a 0H c −1
η a0 .

7. Let B be a MxQ (Q<M) matrix satisfying a property B B = I . We define a

H

reduced array output G i (ω) = B F i (ω) where F i (ω) is a frequency snapshot

H

(see 4.62a). Show that the eigenstructure of the spectral matrix of the reduced
output is identical to that of the spectral matrix of the normal array output, in
particular, B a ⊥ v η . This property forms the basis for the beamspace
H

subspace method where B acts as a spatial filter to restrict the incident energy
to a preselected angular sector [26, 27].
8. Let the columns of B be the eigenvectors corresponding to the significant
eigenvalues of Q in (1.70). Assume that noise sources are distributed over an

© 2001 by CRC Press LLC

 angular sector ±∆ ( θ 0 = 0 ). Show that the reduced noise (as defined in exercise
7 above) becomes white.

References

1. S. Haykin (Ed), Nonlinear methods of spectrum analysis, Springer-Verlag,

Berlin, 1979.
2. P. S. Naidu, Modern Spectrum Analysis of Time Series, CRC Press, Boca
Raton, FL. 1996.
3. G. H. Golub and C. F. Van Loan, Matrix Computations, The Johns
Hopkins Univ. Press, Baltimore, 1983.
4. S. Marcos, A. Marsal, and M. Benidie, The propagation method for source
bearing estimation, Signal Proc., vol.42, pp. 121-138, 1995.
5. J. Munier and G. Y. Deliste, Spatial analysis using new properties of the
cross spectral matrix, IEEE, Trans. Signal Proc., vol.37, pp. 1991.
6. H. P. Bucker, Use of calculated sound fields and matched field detection to
locate sound sources in shallow water, J. Acoust. Soc. of Am., vol. 59, pp.
368-373, 1976.
7. A. Tolstoy, Matched field processing, WorldScientific, Singapore, 1995.
8. I. D. Rathjen, G. Boedecker, and M. Siegel, Omnidirectional beam forming
for linear antennas by means of interpolated signals, IEEE Jour of Ocean Eng.
vol. OE-10, pp. 360-368, 1985.
9. F. J. Harris, On the use of windows for harmonic analysis with the discrete
Fourier transform, Proc IEEE, vol. 66, pp. 51-83, 1978.
10. R. Monzingo and T. Miller, Introduction to Adaptive Arrays, Wiley and
Sons, New York, 1980.
11. A. Eberhard, An optimal discrete window for calculation of power spectra,
IEEE Trans., AU-21, pp. 37-43, 1973.
12. K. Yao, Maximum energy windows with constrained spectral values,
Signal Processing, pp. 157-168, 1986.
13. D. Korompis, K. Yao, and F. Lorenzelli, Broadband maximum energy array
with user imposed spatial and frequency constraints, IEEE, ICASSP, pp. IV-
529-532, 1994.
14. J. Capon, High resolution frequency-wavenumber spectrum analysis, Proc.
IEEE, 57, pp. 1408-1418, 1969.
15. B. D. Van Veen and K. M. Buckley, Beamforming: A versatile approach to
spatial filtering, IEEE ASSP Mag., pp.4-24, April 1988.
16. H. Cox, Resolving power and sensitivity to mismatch of optimum array
processors, J. Acoust. Soc. of Am., vol. 54, pp. 771-785, 1973.
17. M. H. Er and A Cantoni, A unified approach to the design of robust
narrowband antenna processor, IEEE Trans. on Antenna and Propagation, vol.
38, pp. 17-23, 1990.

© 2001 by CRC Press LLC

18. D. G. Luenberger, Optimization by vector space methods, Wiley, New
York, 1969.
19. J. Munier and G. R. Deliste, Spatial analysis in passive listening using
adaptive techniques, Proc. IEEE, vol. 75, pp. 1458-1471, 1987.
20. P. S. Naidu and V. V. Krishna, Improved maximum likelihood (IML)
spectrum estimation: Performance analysis, Jour. I. E. T. E, vol. 34, pp. 383-
390, 1988.
21. J. P. Burg, The relationship between maximum entropy spectra and
maximum likelihood spectra, Geophysics, vol. 37, pp. 375-376, 1972.
22. S. R. De Graaf, and D. H. Johnson, Capability of array processing
algorithms to estimate source bearings, IEEE Trans ASSP-33, pp. 1368-1379,
1985.
23. S. U. Pillai, Array Signal Processing, Springer-Verlag, New York, 1989.
24. T. S. Durrani and K. C. Sharman, Eigenfilter approaches to adaptive array
processing, IEE Proc., vol. 130, Pts F & H, pp. 22-28, 1983.
25. S. A. Nakhamkin, L. G. Tyurikov, and A. V. Malik, Construction models
and decomposition algorithm of seismic fields from the characteristic numbers
and vectors of spectral matrices, Izvestia, Earth Physics, No. 12, pp. 785-791,
1975.
26. G. Bienvenu and L Kopp, Decreasing high resolution method sensitivity by
conventional beamformer preprocessing, IEEE, ICASSP-84, pp.33.2.1-33.2.4,
1984.
27. P. Forster and G. Vezzosi, Application of spheroidal sequences to array
processing, IEEE, ICASSP-87, pp. 2268-2271, 1987.

© 2001 by CRC Press LLC

 Chapter Five

Source Localization: Subspace Methods

The location parameters are estimated directly without having to search for
peaks as in frequency-wavenumber spectrum (the approach described in chapter
4). In open space the direction of arrival (DOA), that is, azimuth or elevation or
both, is estimated using the subspace properties of the spatial covariance matrix
or spectral matrix. MUSIC is a well known algorithm where we define a
positive quantity which becomes infinity whenever the assumed parameter(s) is
equal to the true parameter. We shall call this quantity as a spectrum even
though it does not possess the units of power as in the true spectrum. The
MUSIC algorithm in its original form does involve scanning and searching,
often very fine scanning lest we may miss the peak. Later extensions of the
MUSIC, like root MUSIC, ESPRIT, etc. have overcome this limitation of the
original MUSIC algorithm. When a source is located in a bounded space, such
as a duct, the wavefront reaching an array of sensors is necessarily nonplanar
due to multipath propagation in the bounded space. In this case all three
position parameters can be measured by means of an array of sensors. But the
complexity of the problem of localization is such that a good prior knowledge
of the channel becomes mandatory for successful localization. In active
systems, since one has control over the source, it is possible to design
waveforms which possess the property that is best suited for localization; for
example, a binary phase shift key (BPSK) signal with its narrow
autocorrelation function is best suited for time delay estimation. Source
tracking of a moving source is another important extension of the source
localization problem.

§5.1 Subspace Methods (Narrowband):

Earlier we showed an interesting property of eigenvalues and eigenvectors of a
spectral matrix of a wavefield which consists of uncorrelated plane waves in the
presence of white noise. Specifically, equations (4.14a) and (4.14b) form the
basis for the signal subspace method for DOA estimation. For convenience,
those two equations are reproduced here

λ m = α m + σ 2η (4.14a)

v mH A = 0, for m = P, P + 1,..., M − 1 (4.14b)

Equation (4.14b) implies that the space spanned by the columns of A , that is,
the direction vectors of incident wavefronts, is orthogonal to the space spanned
by the eigenvectors, v m , m = P, P + 1,..., M − 1, often known as noise

© 2001 by CRC Press LLC

 subspace,v η . The space spanned by the columns of A is known as signal
subspace, v s . Consider the space spanned by a steering vector,
 − jω d sin θ − j 2ω dc sin θ − j ( M −1)ω sin θ 
d
a(ω,θ) = col 1,e c ,e ,...,e c
 , as the steering
 
π
angle is varied over a range ± . The intersection of the array manifold with
2
the signal subspace yields the direction vectors of the signals.
5.1.1 MUSIC: On account of (4.14b) a steering vector pointing in the direction
of one of the incident wavefronts will be orthogonal to the noise subspace,

v mH a(ω,θ) = 0, m = P, P + 1,..., M − 1 (5.1)

where θ = θ 0 ,θ1 ,...,θ P −1 . For narrowband signals the matrix A is given by

A = a(ω 0 sin θ 0 ),a(ω1 sin θ1 ),...,a(ω P −1 sin θ P −1 )

d d d
(5.2)
 c c c 

where we take a wavefront with a center frequency ω p to be incident at an

angle θ p . Further, we assume that the center frequencies of the narrowband
signals are known. We define a quadratic function involving a steering vector
and noise subspace,

1
SMusic (ω,θ) = (5.3)
a (ω,θ)v ηv ηH a(ω,θ)
H

SMusic (ω,θ) , also known as the eigenvector spectrum, will show sharp peaks
whenever θ = θ 0 ,θ1 ,...,θ P −1 . The subscript Music stands for M ultiple
Signal Classification. This acronym was coined by Schmidt [1] who discovered
the subspace algorithm. At about the same time but independently Bienvenu
and Kopp [2] proposed a similar algorithm. Pisarenko [3] had previously
published a subspace based algorithm in the context of harmonic analysis of
time series. Note that, although we refer to SMusic (ω,θ) as spectrum, it does
not have the units of power; hence it is not a true spectrum. Let us express the
steering vector in terms of the product of frequency and time delay
d
τ= sin θ ,
c
a(ωτ) = col{1,e jωτ ,e j 2ωτ ,...,e j ( M −1)ωτ }

© 2001 by CRC Press LLC

The peaks of SMusic (ω,θ) will now be at ωτ = ω 0 τ 0 ,ω1τ1 ,...,ω P −1τ P −1
and given the frequencies we can estimate the delays but there is one difficulty.
It is possible that two wavefronts with different center frequencies may arrive at
d d
such angles that ωp sin θ p = ω p' sin θ p' , in which case the two
c c
direction vectors will become identical, causing a loss of the full column rank
property of the matrix A . When the frequencies are unknown, evidently, the
delays or DOA's cannot be uniquely estimated.
We now turn to the signal subspace spanned by the eigenvectors
corresponding to eigenvalues λ m = α m + σ η , m = 0,1,..., P − 1 . We
2

shall show that the signal subspace is the same as the space spanned by the
columns of the matrix A . Since A is a full rank matrix its polar
decomposition gives

A = TG (5.4)

where G is a full rank PxP matrix and T is a MxP matrix satisfying the
following property,

THT = I (PxP unit matrix)

In (4.14a) we shall replace A by its polar decomposition (5.4)

v mH AC0 A H v m = α m (5.5a)

or in matrix form

[ ]
v sH T GC0G H T H v s = diag{α m , m = 0,1,..., P − 1} (5.5b)

Let T = v s , which is consistent with the assumed properties of T . Equation

(5.5b) now reduces to

GC0G H = diag{α m , m = 0,1,..., P − 1} (5.6a)

and
A = v sG (5.6b)

We can estimate G from (5.6b) by premultiplying both sides by v sH and

obtain

© 2001 by CRC Press LLC

 G = v sH A (5.6c)

Thus, the space spanned by A and v sH are identical. Further, by eliminating

G from (5.6b) and (5.6c) we obtain an interesting result

A = v s v sH A (5.6d)

Let us define a complementary orthogonal space (I − v s v sH ) which will also

be orthogonal to A ; therefore, A (I − v s v s )A = 0 . An equivalent
H H

definition of Music spectrum may be given using (I − v s v s ) ,

H

1
SMusic (ω,θ) = (5.7)
a (ω,θ)(I − v s v sH )a(ω,θ)
H

Signal Eigenvalues and Source Power: The signal eigenvalues and source power
are related, although the relationship is rather involved except for the single
source case. Let us first consider a single source case. The covariance matrix is
given by C0 = a 0 s0 a 0 which we use in (5.5a) and obtain
H

α0
s0 = H 2 (5.8)
v a0
0

Since v 0 and a 0 span the same space, v 0 = g0 a 0 where g0 is a constant

1
which may be obtained by requiring that v 0 v 0 = 1 . We obtain g0 =
H
.
M
α
Equation (5.8) reduces to s0 = 0 . Next we consider the two-source case. The
M
covariance function is given by

s0 0 
C f = [a 0 ,a1 ] [ a 0 ,a1 ]
H

0 s1 

and using (5.5a) we obtain

v sH AC0 A H v s = diag{α 0 ,α1} (5.9a)

© 2001 by CRC Press LLC

 azimuths power power
#1 source #2 source
30o, 40o 1.00 1.00
30o, 35o 1.00 1.00
30o, 33o 1.00 1.00
30o, 31o 1.00 1.00
30 o, 30.5o 1.00 1.00
Table 5.1 Estimation of power given the azimuths of two equal power wavefronts
incident on a 16 sensor ULA. Power estimates are error-free even when the
wavefronts are closely spaced.

[ ] [
diag{α 0 ,α1} A H v s ]
−1 −1
C0 = v sH A (5.9b)

Equation (5.9b) has been used to estimate the powers of two equal power
(snr=1.0) wavefronts incident on a 16 sensor ULA and the results are tabulated
in Table 5.1. The estimation is error-free right down to a half-a-degree
separation; however, this good performance deteriorates in the presence of
model and estimation errors.
Aliasing: We pointed out earlier (p.211) that the sensor spacing in a ULA must
be less than half the wavelength so that there is no aliasing in the frequency-
wavenumber spectrum. The same requirement exists in all high resolution
methods, namely the Capon spectrum, Maximum entropy spectrum and Music
spectrum. The basic reason for aliasing lies in the fact that the direction vector
is periodic. Consider anyone column of the matrix A ,

 − jω p dc sin θ p − j 2ω p dc sin θ p − j ( M −1)ω p sin θ p 

d
a p (ωτ) = col 1,e ,e ,...,e c

 

λp λp
Now let d= − δ p where 0 ≤ δ p ≤ , then a p (ωτ )
sin θ p 2sin θ p

becomes

d  − j 2 π λd sin θ p − j 4 π λd sin θ p − j ( M −1)2 π

d
λp
sin θ p 

a p (ω p sin θ p ) = col 1,e p
,e p
,...,e 
c  

© 2001 by CRC Press LLC

  j 2 π δ p sin θ p j 4 π δ p sin θ p j ( M −1)2 π
δp
sin θ p 

λp λp λp
= col 1,e ,e ,...,e 
 
−δ
= a p (ω p p sin θ p )
c

d δ
We can find an angle θ̂ p such that sin θ̂ p = − p sin θ p and hence
λp λp

d d
a p (ω p sin θ p ) = a p (ω p sin θ̂ p )
c c

An aliased spectral peak would appear at θ̂ p which is related to θ p ,

δ 
θ̂ p = sin −1  p sin θ p  (5.10)
 d 

d
As an example, consider an array with sensor spacing, = 1 and a wavefront
λp
incident at an angle θ p =60°. For this choice of array and wave parameters
δ p = 0.1547 and from (5.10) we get the angle where the aliased peak will be
located, θ̂ p = −7.6993°. The wave number spectrum computed by all four
methods is shown in fig. 5.1.
Aliasing is on account of periodicity of a direction vector which in
turn is caused by periodicity present in a ULA. Thus, to avoid aliasing, it
would be necessary to break this periodicity; for example, we may space the
sensors nonuniformly. In a circular array, though sensors are uniformly spaced
(e.g. UCA), the time delays are nonuniform; therefore a UCA will yield an
alias-free spectrum [4]. This is demonstrated in fig. 5.2 where we consider a
wavefront which is incident at 60o (with respect to x-axis) on a circular array
consisting of 16 sensors uniformly spread over a circle of radius 8λ . The
Capon spectrum is shown for this case. The aliasing phenomenon is not
encountered in random arrays where the sensors are spaced at random intervals.
But as shown in chapter 2 the random array possesses a highly nonlinear phase
response.

© 2001 by CRC Press LLC

 10
(a)
8

6
Aliased Actual
4

2
0
1.5
(b)
1

0.5
Spectrum

0
6000

5000 (c)

4000

3000

2000

1000

0
6
x 10
5

4 (d)

0
-100 -50 0 50 100
Angle in deg

Figure 5.1 The aliasing effect due to undersampling of the wavefield (d=λ). All four
methods of spectrum estimation have been used. (a) Bartlett spectrum, (b) Capon
spectrum, (c) Maximum entropy spectrum and (d) Music spectrum. While the actual
peak is at 60o the aliased peak appears at -7.69o.

© 2001 by CRC Press LLC

 5.1.2 Correlated Sources: We have so far assumed that the source matrix S0 is
a full rank matrix. When sources are fully uncorrelated S0 is a diagonal matrix
with the nonzero elements representing the power of the sources. The source
matrix is naturally full rank. Let us now consider a situation where the sources
are partially or fully correlated. We model the source matrix as

S0 = s ρ s H (5.11)
where

s = diag { s0 , s1 ,..., sP −1 }
s0 , s1 ,..., sP −1 represent the power of P sources and ρ is the coherence matrix
whose (m,n)th element represents the normalized coherence between the mth
and nth sources. The signal eigenvalues of the spectral matrix for P=2 are given
by [5]

M
λ0 = (s0 + s1 ) + M s0 s1 Re{ρ12 }ψ +
2

{[ }
1
M
]
(s0 + s1 ) + 2 s0 s1 Re{ρ12 }ψ − 4s0 s1 (1 − ψ )(1 − ρ12 )
2 2 2 2
+ σ 2η
2
M
λ1 = (s0 + s1 ) + M s0 s1 Re{ρ12 }ψ −
2

{[ }
1
M
]
(s0 + s1 ) + 2 s0 s1 Re{ρ12 }ψ − 4s0 s1 (1 − ψ )(1 − ρ12 )
2 2 2 2
+ σ 2η
2
(5.12a)

where
d
M(sin θ 0 − sin θ1 ))
sin(π
ψ(M) = λ
d
M sin(π (sin θ 0 − sin θ1 ))
λ

The sum of the signal eigenvalues, that is,

(λ 0 − σ 2η ) + (λ1 − σ 2η ) =
(5.12b)
M(s0 + s1 ) + 2M s0 s1 Re{ρ12 }ψ(M)

© 2001 by CRC Press LLC

 1
Spectrum

0.5

0
-100 -50 0 50 100
Angle in deg.

Figure 5.2: No aliasing effect is seen with a circular array. Sixteen sensor UCA with
radius= 8λ (sensor spacing=3.12l) is used. A plane wavefront is incident on the
array at 60o .

represents coherent addition of power. For uncorrelated sources ρ12 = 0 (5.12a)

reduces to

1
M
λ 0 = (s0 + s1 ) +
2
M
2
{ 2
(s0 + s1 )2 − 4s0 s1 (1 − ψ (M) ) } 2
+ σ η2
(5.13)
1
M
λ 1 = (s0 + s1 ) −
2
M
2
{ 2
(s0 + s1 )2 − 4s0 s1 (1 − ψ (M) ) } 2
+σ 2
η

Also, note that when the sources are in the same direction λ 0 = M(s0 + s1 )
and λ1 = 0 .
The source spectral matrix may be modified by spatial smoothing of
the array outputs. This is achieved by averaging the spectral matrices of
subarray outputs over all possible subarrays (see fig. 5.3). The ith subarray
(size µ ) signal vector at a fixed temporal frequency is given by

{ }
Fi = col Fi (ω ), Fi +1 (ω ),..., Fi + µ −1 (ω ) 0 ≤ i ≤ M − µ +1
= I i, µ F

© 2001 by CRC Press LLC

 ...

4th subarray

Figure 5.3: Overlapping subarrays are formed as shown above. Each subarray has
four sensors. It shares three sensors with its immediate neighbours.

where I i, µ is a diagonal matrix,

0,...0, 1,...,1, 0,...,0 

I i,µ = diag  
 0 to i −1, i to i +µ −1, i +µ to M 
F = col{F0 (ω ),..., FM −1 (ω )}

The spectral matrix of the i t h subarray is now given by

Si, µ = E{Fi F H
i }=I i, µ
H
S I . We shall now average all subarray spectral
i, µ
matrices

M −µ +1 M −µ +1
1 1
S= ∑
M − µ + 1 i=0
S i,µ = ∑
M − µ + 1 i=0
H
I i,µ S I i,µ
(5.14a)
M −µ +1
1
= ∑
M − µ + 1 i=0
H
I i,µ A(ω)S0 (ω)A H I i,µ + σ 2ηI

where we have used the spectral matrix of the signal model plane waves in the
presence of white noise (4.12b). We can show that

I i,µ A(ω) = I i,µ [a 0 ,a1 ,...,a P −1 ]

 − j 2 πi dc sin θ 0 − j 2 πi sin θ P−1 
d
(5.14b)
= [â 0 , â1 ,..., â P −1 ]diag e ,...,e c

 

© 2001 by CRC Press LLC

where

 − jω d sin θ i − j 2ω dc sin θ i − j (µ −1)ω sin θ i 

d
â i (ωτ) = col 1,e c ,e ,...,e c

 

We shall use (5.14b) in (5.14a) and obtain

M −µ +1
 1 
S = Â  ∑ φ i S0 (ω) φ Hi  Â H + σ 2ηI (5.15a)
 M − µ + 1 i=0 

where

 = [â 0 , â1 ,..., â P −1 ]

 − j 2 πi dc sin θ 0 − j 2 πi dc sin θ1 − j 2 πi sin θ P−1 
d
φ i = diag e ,e ,...,e c

 

The quantity inside the square brackets in (5.15a) may be computed by actual
multiplication followed by summation,

M −µ +1 d
1 − j 2 π i ( sin θ m −sin θ n )
[ ]mn = sm sn ρmn ∑
M − µ + 1 i=0
e c

(5.15b)
d
− jπ ( M −µ ) ( sin θ m −sin θ n )
= sm sn ρmn ψ(M − µ + 1)e c

From (5.15b) it follows that the coherence after spatial smoothing may be
written as

d
− jπ ( M −µ ) ( sin θ m −sin θ n )
ρmn = ρmn ψ(M − µ + 1)e c
(5.16)

The magnitude of the off-diagonal terms in the coherence matrix, ρ (see

(5.11)), is reduced by a factor, ψ(M − µ + 1) , which is indeed small (<<1) for
large (M − µ + 1) . This contributes to the increase of the rank of the smoothed
coherence matrix. It is shown in [6] that ρ becomes full rank when
M − µ +1≥ P.
Let us now examine the effect of the spatial smoothing on the
eigenvalues for the two source case. From

© 2001 by CRC Press LLC

 µ
λ0 = (s0 + s1 ) + µ s0 s1 Re{ρ̃01}ψ(µ)ψ(M − µ + 1) +
2
1

[
µ  (s0 + s1 ) + 2 s0 s1 Re{ρ̃01}ψ(µ)ψ(M − µ + 1) 
  ]
2 2

 + ση
2

2 −4s s (1 − ψ(µ) )(1 − ρ ψ(M − µ + 1) )
2 2

 01 01

µ
λ1 = (s0 + s1 ) + µ s0 s1 Re{ρ̃01}ψ(µ)ψ(M − µ + 1) −
2
1

[
µ  (s0 + s1 ) + 2 s0 s1 Re{ρ̃01}ψ(µ)ψ(M − µ + 1)
 ] 2
2

 + ση
2

2 −4s s (1 − ψ(µ) 2 )(1 − ρ ψ(M − µ + 1) 2 ) 
 01 01
(5.17)

d
jπ ( M − µ ) ( sin θ 0 −sin θ 1 )
where ρ̃01 = ρ01e
λ
. The sum of the signal eigenvalues is
given by

(λ 0 − σ 2η ) + (λ1 − σ 2η ) =
M(s0 + s1 ) + 2M s0 s1 Re{ρ̃01}ψ(µ)ψ(M − µ + 1)

If we select µ such that

d
π (M − µ + 1) (sin θ 0 − sin θ1 ) ≈ π
λ

then ψ(M − µ + 1) ≈ 0 , λ 0 = Ms0 and λ1 = Ms1 and the desired subarray size
would be

2
µ ≈ (M + 1) −
(sin θ 0 − sin θ1 )

1
Let (sin θ 0 − sin θ1 ) ≈ . Then the desired subarray size is
µ

1
µ ≈ (M + 1) (5.18)
3

© 2001 by CRC Press LLC

 2.4

2.2

2.0
2nd Eigenvalue

1.8

1.6

1.4

1.2

1.0
4 12 20 28 36 44 52
µ

Figure 5.4: The 2nd eigenvalue relative to noise variance is shown as a function of
the subarray size (µ).

λ1
This result is very close to that given in [7]. We have evaluated as a
σ 2η
function of the subarray size for two perfectly correlated equal power waves
1
incident at angles θ 0 and θ1 such that sin θ 0 − sin θ1 = on a ULA with
µ
64 sensors and array snr =10. The results are shown in fig. 5.4.

5.1.3 Direct Estimation: In MUSIC it is required to scan the entire angular

π
range of ± . Since the spectral peaks are extremely fine the scanning must be
2
done at a very fine interval; otherwise there is a high risk of missing the peak
altogether. This is a serious drawback of the subspace methods based on
scanning. Alternatively, there are direct methods which enable us to estimate
the DOA directly. We shall describe three such methods, namely, Pisarenko’s
method, minimum norm, and root Music and show how they are interrelated.

© 2001 by CRC Press LLC

 Pisarenko’s Method: Let the array size be M = P + 1 where P stands for the
number of uncorrelated sources. The noise subspace is now spanned by a single
eigenvector, v M −1 . Hence, the direction vector, a , will be orthogonal to
v M −1 ,

v HM −1a = 0 (5.19)

The direction vector of the pth source may be expressed as

 − j 2 π λd sin θ p − j 2 π( M −1) sin θ p 

d
a p = col 1,e ,...,e λ

 
{ }
= col 1, z p , z 2p ,..., z pM −1 , p = 0,1,..., P − 1

d
− j 2 π sin θ p
where zp = e λ
. Consider a polynomial

[ ]
T
v HM −1 1, z, z 2 ,..., z M −1 =0 (5.20)

whose roots are indeed z0 , z1 ,..., z P −1 . In the complex plane all roots lie on a
unit circle. The angular coordinate of a root is related to the DOA. For
example, let the pth root be located at ϕ p . Then ϕ p = πsin θ p . This method
was first suggested by Pisarenko [3], many years before MUSIC was invented !
Root Music: The Pisarenko’s method has been extended taking into account the
noise space spanned by two or more eigenvectors [8]. The extended version is
often known as root Music. Define a polynomial,

M −1
S(z) = ∑ v [1, z, z ,..., z ]
m= P
H
m
2 M −1 T
(5.21)

1
The roots of S(z) or D(z) = S(z)S( ) lying on the unit circle will
z
correspond to the DOAs of the sources and the remaining M-P roots will fall
inside the unit circle (and also at inverse complex conjugate positions outside
the circle).
In the minimum norm method a vector a is found which is a solution
of a S f a = min under the constraint that a a = 1 . It turns out that the
H H

solution is an eigenvector of S f corresponding to the smallest eigenvalue.

© 2001 by CRC Press LLC

 Radial Angular Estimation
coordinate coordinate error variance=.01
(in radians)
Radial Angular
33.3162 0.9139 2.4854 -1.8515
2.2755 2.5400 1.5593 -0.7223
1.5793 2.9561 1.3047 -3.0944
0.8500 -1.7578 0.7244 -2.2490
0.8809 -0.9386 0.9336 1.6835
1.0000 0.5455 1.0050 0.5647
1.0000 0.7600 1.0057 0.7864
Table 5.2: Direct estimation of DOA by computing signal zeros of a polynomial
given by (5.20) or (5.21). The results shown in columns one and two are for error
free spectral matrix and those in columns three and four are for a spectral matrix with
random errors (zero mean and 0.01 variance).

Thus, the minimum norm method belongs to the same class of direct methods
of DOA estimation initiated by Pisarenko [3]. An example of DOA obtained by
computing zeros on the unit circle under the ideal condition of no errors in the
spectral matrix is shown in table 5.2. Two equal power signal wavefronts are
assumed to be incident on eight sensor ULA at angles 10o and 14o . The DOA
estimates from the error-free spectral matrix are exact (columns one and two in
table 5.2), but in the presence of even a small error in the estimate of the
spectral matrix, a significant error in DOA estimation may be encountered (see
columns three and four in the table).
The zeros of the polynomial defined in any one of the above methods
are important from the point of DOA estimation. In particular, the zeros which
fall on the unit circle or close to it represent the DOAs of the incident
wavefronts. These zeros are often called signal zeros and the remaining zeros,
located deep inside the unit circle, are called the noise zeros. In MUSIC, a peak
in the spectrum represents the DOA of the wavefront. But the height of the
peak is greatly influenced by the position of the signal zero. The peak is
infinite when the zero is right on the unit circle but it rapidly diminishes when
the zero moves away from the unit circle. The peak may be completely lost,
particularly in the presence of another signal zero in the neighborhood but
closer to the unit circle. Hence, it is possible that, while the spectral peaks
remain unresolved, the signal zeros are well separated. The shift of the signal
zero may be caused by errors in the estimation of the spectral matrix from finite
data. Let a signal zero at zi be displaced to ẑi . The displacement, both radial
and angular, is given by

© 2001 by CRC Press LLC

 ∆zi ≈ (zi − ẑi ) ≈ δ r e j (u + δu)

where we have assumed that δ u << 1 . Perturbation analysis reported in [9] for
time series shows that the mean square error in δ r and δ u are related

{ }
E δ ri
2 
= 2N 
cos( θ i 
)

2N
≤ 2
E{ δθ }
(5.22)
2  ωd  π
 
 c 
i

where N stand for the number of time samples (or snapshots). From (5.22) it
follows that E δri { } >> E{ δθ } , particularly when the wavefronts are
2
i
2

incident close to broadside. On account of the above finding the magnitude of a

spectral peak in MUSIC is likely to be more adversely affected by the errors due
to finite data.
Subspace Rotation: The direct estimation of DOA discussed in the previous
section requires a ULA. This itself is a restriction in many practical situations.
A new principle of DOA estimation, known as subspace rotation, exploits a
property of a dipole sensor array, where all dipoles are held in the same
direction, which is the subspace spanned by the direction vectors pertaining to
the upper sensors in the dipole array and those pertaining to lower sensors are
related through a rotation. The principle is explained in chapter 2 (see page
120). The basic starting equation is (2.57), which we reproduce here for
convenience,

I 0  σ η1 
H 2
A  A 
S f̃ =  S f   +   2  (2.57)
AΓ  AΓ  0 I  σ η2 

where S f̃ is the spectral matrix of the dipole array sensors,

S f 1 f 1 S f 1 f 2  AS f A AS f Γ H A H 
H

S f̃ =  = 
S f 2 f 1 S f 2 f 2  AΓS f A
H
AS f A H 

Note that S f 1 f 1 is the spectral matrix of the upper sensor outputs and S f 2 f 2 is
the spectral matrix of the lower sensor outputs. S f 1 f 2 is the cross-spectral
matrix between the upper sensor and lower sensor outputs. From the

© 2001 by CRC Press LLC

eigenvalues of S f 1 f 1 we can estimate the noise variance σ 2η1 (and σ 2η2 from
S f 2 f 2 ) and subtract it from the spectral matrix. Define S0f 1 f 1 = S f 1 f 1 − σ 2η1 .
Consider the generalized eigenvector (GEV) problem of the matrix pencil
{S 0
f1 f1 }
, S f1 f 2 ,

S0f 1 f 1 v = γS f 1 f 2 v (5.23)

where γ is a generalized eigenvalue and v is a corresponding eigenvector.

Consider the following:

[ ] [
v H S0f 1 f 1 − γS f 1 f 2 v = v H AS f A H − γAS f Γ H A H v ]
[ ]
= v H AS f I − γΓ H A H v
(5.24)
=v H
AS [ I − γΓ ]A
f
H H
v
= v H Qv = 0

Since S f is full rank as the sources are assumed to be uncorrelated and A is

assumed to be full rank, then, for the right hand side of (5.24) to be true, we
ω
−j ∆.δ i
must have γ = e c . The rank of Q will be reduced by one and a vector v
can be found to satisfy (5.24). Thus the generalized eigenvalues of the matrix
ω

{ }
−j ∆.δ i
pencil S0f 1 f 1 , S f 1 f 2 are γ i = e c , i = 0,1,..., P − 1 from which we can
estimate θi ,i = 0,1,..., P − 1 . This method of DOA estimation is known as
ESPRIT (Estimation of signal parameters via rotation invariance technique)
[10]. An example of DOA obtained via ESPRIT is in table 5.3. The
eigenvalues of the pencil matrix (5.23) under the ideal condition of no errors in
the spectral matrix as well as with errors in the spectral matrix are shown in
table 5.3. Two equal power signal wavefronts are assumed to be incident on
eight sensor ULA at angles 10o and 14o. The eigenvalues estimated from error
free spectral matrix are exact (columns one and two in table 5.3), but in the
presence of a small error in the estimation of the spectral matrix, a significant
error in DOA estimation may be encountered (see columns three and four in the
table).
The generalized eigenvector v i , corresponding to eigenvalue γ i ,
possesses an interesting property. Let us write (5.24) in expanded form

© 2001 by CRC Press LLC

 Radial Angular Estimation
coordinate coordinate error variance=.01
(in radians)
Radial Angular
4.6642 1.8540 3.4640 -2.1554
2.0198 0.7276 2.5202 0.4402
0.8182 -2.3272 1.7050 -0.4834
0.2108 -2.6779 0.5738 -1.6783
0.5224 -0.3753 0.4586 1.6290
0.5193 1.1199 0.6753 0.9060
1.0000 0.5455 0.9969 0.5533
1.0000 0.7600 1.0087 0.7720
Table 5.3: Generalized eigenvalues of pencil matrix (5.23).

[viH a0 ,..., viH ai − 1,viH ai + 1,..., viH a P − 1]Ŝ f

v H a 
 i 0 
. 
 
 ωd  . 
 j sin θ 0   
c .
1 − γ i e ... 0  
  v H a 
   i i − 1,  = 0
  H 
  vi ai + 1 
ωd
 j sin θ P − 1  . 
0 c   
 ... 1 − γ ie  . 
 
. 
 H 
 v i a P − 1 

where Ŝ f represents the source matrix where the ith column and the ith row are
ω0
−j ∆.δ k
deleted. Since γ i ≠ e c , i ≠ k , the diagonal matrix will be full rank and
also Ŝ f is full rank by assumption, we must have v iH a k, = 0 for all k ≠ i .
{ 0
Thus, the generalized signal eigenvector of the pencil matrix S f 1 f 1 , S f 1 f 2 is }
orthogonal to all direction vectors except the ith direction vector. We shall
exploit this property for signal separation and estimation later in chapter 6.

© 2001 by CRC Press LLC

5.1.4 Diffused Source: A point source, on account of local scattering or
fluctuating medium or reflections from an uneven surface, may appear as a
diffused source, that is, the main source surrounded by many secondary sources.
A sensor array will receive a large number of rays, all of which would have
actually started from the same point (main source) but have traveled different
paths due to scattering or refraction or reflection. Such a model was considered
in chapter 1 (p. 49). Here we shall explore the possibilities of localizing such a
diffused source, in particular, a diffused source due to local scattering (see fig.
1.29). The model of the diffused source signal is given by (1.65)

f(t) = ãs(t) + η(t) (5.25)

L −1
where ã = ∑ α l e − jω c δt l a(θ + δθl ) where α l , δtl and δθl are for the lt h
l=0
ray, complex amplitude, time delay and direction of arrival with respect to a
direct ray from the main source, respectively. Assume that there are L rays
reaching the sensor array. For the direct ray we shall assume that α 0 = 1,
δt0 = 0 and δθ 0 = 0 . The covariance matrix of the array (ULA) output is
given by (see eq. 1.70)

c f ≈ σ 2s Lσ α2 D(θ)QD H (θ) + σ 2ηI (5.26)

where we have assumed that the scatterers are uniformly distributed in the
angular range ±∆ . The Q matrix for the assumed uniform distribution is
given by

d
sin 2π (m − n)∆ cosθ
{Q}mn = λ
d
2π (m − n)∆ cosθ
λ
and
 j 2 λπd sin θ j
2 π( M −1)d
sin θ 
D = diag 1, e ,..., e λ

 

Q shows some interesting properties [11]. There

The eigendecomposition of
 d 
are r, where r = 2 M∆ cosθ , significant eigenvalues close to unity ([x]
 λ 
stands for the largest integer greater than x). The remaining (M-r) eigenvalues

© 2001 by CRC Press LLC

 are insignificant. Let v η be a matrix whose columns are M-r eigenvectors
corresponding to M-r insignificant eigenvalues (cf. the noise subspace in the
Music algorithm). Now it follows that

[ ]
v ηH c f − σ 2ηI v η = 0 (a matrix M-r x M-r zeros) (5.27)

Using (5.26) in (5.27) we obtain

r −1
σ 2s Lσ α2 ∑ λ m v ηH D(θ)e m e mH D H (θ)v η = 0 (5.28)
m=0

r −1
where we have used eigendecomposition of Q≈ ∑λ
m=0
e e . Here λ m
H
m m m

( ≈ 1 ) are significant eigenvalues of Q . Since σ 2f 0 , and λ m >0 for all m<r,

the following must hold good for all m<r:

2
v ηH D(θ)e m = 0 (5.29)

for all m<r. The azimuth of the center of the cluster can be estimated using the
orthogonality property demonstrated in (5.29). The eigenvectors e m , discrete
prolate spheroidal sequence (DPSS), are obtained by eigendecomposition of the
Q matrix. But in Q there is an unknown parameter pertaining to the width of
the cluster, namely, ∆ cos θ which has to be estimated. We have in chapter 1
(page 59) indicated that the rank of the covariance matrix (noise free) is closely
related to this unknown parameter.
5.1.5 Adaptive Subspace: Adaptive methods for the estimation of the
eigenvector corresponding to the minimum (or maximum) eigenvalue have
been described by many researchers[12-14]. The method is based on inverse
power iteration [15]

v̂ 0 = [1,0,...,0]
T

ṽ k +1 = S−1
f (ω)v̂
k
(5.30)

k +1 ṽ k +1
v̂ =
ṽ k +1 ṽ k +1
T

k
where v̂ is an eigenvector at kth iteration and the last equation is meant for
normalization. To estimate the eigenvector corresponding to the largest

© 2001 by CRC Press LLC

eigenvalue the algorithm shown in (5.30) may be used with the difference that
the spectral matrix, instead of its inverse, is used.
An attempt to extend the inverse power iteration method to the
estimation of the entire subspace (signal or noise subspace) has also been
reported [13]. All the above methods require computation of the gradient of a
cost function which is minimized. The gradient computation is slow;
furthermore, in the presence of noise, it is unstable. Additionally, the rate of
convergence of a gradient method is known to be very slow [14]. The alternate
approach is to use the well-known power method for estimation of an
eigenvector corresponding to the minimum (or maximum) eigenvalue [15]. The
power method is easy to understand and easy to implement on a computer. We
shall briefly describe the power method for estimation of the entire signal or
noise subspace. Let v̂ m , m = 1,2,... P be the eigenvectors based on the data
N

till time N , that is, the eigenvectors of the spectral matrix, Ŝ Nf (ω) . When a
new sample arrives the eigenvectors are updated using the following recursive
equation followed by Gram-Schmidt orthonormalization (GSO),

Ŝ Nf +1v̂ mN = g mN +1
(5.31)
v̂ mN +1 = GSO{g mN +1}, m = 1,2,... P

The choice of initial eigenvectors to start the recursion is important for rapid
convergence. A suggested choice in [14] is to orthonormalize P initial
frequency snapshots and use them as the initial eigenvectors. The starting value
of the spectral matrix is chosen as

Ŝ Pf = (I + F 0 F 0H )

The Gram-Schmidt orthonormalization process involves finding p orthonormal

vectors which are linearly equivalent to data vectors [50]. Let F1 , F 2 ,..., F P be
a set of frequency snapshots and I1 , I 2 ,..., I P be the set of linearly equivalent
orthonormal vectors which satisfy the following conditions

I k ⊥F r for r < k

2
I k = 1 for all k (5.32)

The Gram-Schmidt method determines the orthonormal vectors by solving a

system of linear equations

© 2001 by CRC Press LLC

 I1 = γ 11F1
I 2 = γ 12 F1 + γ 22 F 2

I 3 = γ 12 F1 + γ 22 F 2 + γ 32 F3
... (5.33)
I P = γ 1P F1 + γ 2P F 2 +...+γ PP F P

§5.2 Subspace Methods (Broadband):

Since the location parameters are independent of frequency, we ought to get
identical location estimates at different frequencies. But, on account of errors in
the estimation of the spectral matrix, the location estimates are likely to be
different at different frequencies. The errors in the estimation of a spectral
matrix are likely to be uncorrelated random variables. By suitably combining
the spectral matrix estimates obtained at different frequencies, it is hoped that
the error in the final location estimate can be reduced. This is the motivation
for broadband processing. In the process of combining all spectral matrices it is
important that the basic structure of the spectral matrices should not be
destroyed. For example, a spectral matrix at all signal frequencies is of rank one
when there is only one source or rank two when there are two uncorrelated
sources, and so on. Consider a ULA and a single source. The spectral matrix at
frequency ω is given by S(ω ) = a(ω , θ )S0 (ω )a (ω , θ ) + σ η I . Notice
H 2

that the frequency dependence of direction vectors is explicitly shown. Let us

consider a generalized linear transformation of the form, T(ω )S(ω )T (ω ) ,
H

where T(ω ) is a transformation matrix, yet to be chosen. The transformation

matrix must map the direction vector into another direction vector at a chosen
frequency and the background noise continues to remain white. The transformed
spectral matrix is given by

T(ω )S(ω )T H (ω ) =
(5.34)
T(ω )a(ω , θ )S0 (ω )a H (ω , θ )T H (ω ) + σ η2 T(ω )T H (ω )

where the transformation matrix T(ω ) must possess the following properties:

T(ω )a(ω , θ ) = a(ω 0 , θ )

(5.35)
T(ω )T H (ω ) = I

© 2001 by CRC Press LLC

where ω 0 is a selected frequency. Using (5.35) in (5.34) and averaging over a
band of frequencies we obtain

S(ω 0 ) = ∑ T(ω )S(ω )T (ω )

i∈bw
i i
H
i

= a(ω ,θ) ∑ S (ω )a (ω ,θ) + σ I

0 0 i
H
0
2
η (5.36)
i∈bw

where bw stands for the signal bandwidth. In (5.36) we observe that the signal
power spread over a band of frequencies has been focused at one frequency,
namely, ω 0 . This idea of focusing of energy has been actively pursued in [16,
17] for DOA estimation.
In the direction vector there is a parameter, namely, sensor spacing
ωd
which may be changed according to the frequency such that remains
c
constant. This will require the sensor spacing at frequency ω i should be equal
ω 0 d0
to di = , i ∈bw . In practice the required change in the physical
ωi
separation is difficult to achieve, but resampling of the wavefield through
interpolation is possible. This approach to focusing was suggested in [18]. In
chapter 2 we introduced the spatio temporal covariance matrix (STCM) which
contains all spatio temporal information present in a wavefield. It has been
extensively used for source localization [19]. We shall probe into the signal and
noise subspace structure of STCM of ULA as well as UCA and show how the
structure can be exploited for source localization.
5.2.1 Wideband Focusing: Consider P uncorrelated point sources in a far field
and a ULA for DOA estimation. As in (4.12b) the spectral matrix of an array
signal may be expressed as S f (ω) = A(ω, θ)S 0 (ω)A H (ω, θ) where the
columns of A(ω, θ) matrix are the direction vectors. We seek a transformation
matrix T(ω) which will map the direction vectors at frequency ω into
direction vectors at the preselected frequency ω 0 as in (5.35). There is no
unique solution but a least squares solution is possible,

[ ]
−1
T(ω) ≈ A(ω 0 ,θ) A H (ω,θ)A(ω,θ) A H (ω,θ) (5.37)

The transformation matrix given by (5.37) depends upon the unknown azimuth
information. However, it is claimed in [20] that approximate estimates obtained
through beamformation or any other simple approach are adequate. As an
example, consider a single source case. For this, A(ω,θ) =

© 2001 by CRC Press LLC

  − jω dc sin θ − jω( M −1) sin θ 
d
col 1,e ,...,e c
 and the transformation matrix given by
 
1
(5.37) is equal to T(ω) ≈ A(ω 0 ,θ)A H (ω,θ) . To show that this is not a
N
unique answer, consider the following transformation matrix,

 − j (ω 0 −ω ) dc sin θ − j (ω 0 −ω )( M −1) sin θ 

d
T(ω) = diag 1,e ,...,e c
 (5.38a)
 

Clearly, using (5.38a) we can also achieve the desired transformation. In (5.38a)
let ω 0 =0; the transformation matrix becomes

 jω dc sin θ jω( M −1) sin θ 

d
T(ω) = diag 1,e ,...,e c
 (5.38b)
 

which we shall use as a filter on the array output. The filtered output is given
by

∞
1
f 1 (t) = ∫
2π − ∞
T(ω)F(ω)e jωt dω
(5.39)
= col  f (t), f (t − sin θ),..., f (t − (M − 1)sin θ)
d d
 c c 

which indeed is the same as the progressively delayed array output in

beamformation processing. Let us next compute the covariance function of the
filtered output,

c f 1 = E{f 1 (t)f 1H (t)}

 d  
 f (t − m c sin θ), m = 0,1,..., M − 1 
= E H
 f (t − m d sin θ), m = 0,1,..., M − 1 
 c  
d
= c f ((m − n) (sin θ 0 − sin θ)) = c f
c
[ ] mn

© 2001 by CRC Press LLC

wherec f is the covariance matrix of the signal emitted by the source. c f 1 is
known as the steered covariance matrix [21]. When θ = θ 0 all elements of c f 1
are equal to a constant equal to the total signal power, that is,
∞
1
2π −∫∞
S f (ω)dω . Thus, in beamformation we seem to focus all signal power

to zero frequency. When there is more than one source we should compute a
series of steered covariance matrices over a range of azimuth angles. Whenever a
steering angle matches with one of the directions of arrival the steered
covariance matrix will display a strong ‘dc’ term, equal to the power of the
source.
The main drawback of wideband focusing is the requirement that the
DOAs of incident wavefronts must be known, at least approximately, but the
resulting estimate is likely to have large bias and variance [20]. To overcome
this drawback an alternate approach has been suggested in [22]. Let A and B
be two square matrices and consider a two sided transformation, TBT H , which
is closest to A . It is shown in [22] that this can be achieved if T = v A v B
H

where v A is a matrix whose columns are eigenvectors of A and v B is

similarly the eigenvector matrix of B. Now let A and B be the spectral
matrices at two frequencies, A = S f (ω 0 ) and B = S f (ω i ). To transform
S f (ω i ) into S f (ω 0 ) the desired matrix is

T(ω 0 ,ω i ) = v s (ω 0 )v sH (ω i ) (5.40)

where v s is a matrix whose columns are signal eigenvectors. It is easy to show

that T(ω 0 , ω i )T H (ω 0 , ω i ) = I . Applying the transformation on a spectral
matrix at frequency ω i we get

T(ω 0 ,ω i )S f (ω i )T H (ω 0 ,ω i )
= v s (ω 0 )v sH (ω i )S f (ω i )v s (ω i )v sH (ω 0 )
= v s (ω 0 )λ(ω i )v sH (ω 0 )

where

λ(ω i ) = diag{λ 0 (ω i ),λ1 (ω i ),...,λ P −1 (ω i ),0,...,0} ,

are the eigenvalues of the noise free spectral matrix, S f (ω i ) . Next, we average
all transformed matrices and show that

© 2001 by CRC Press LLC

 1 N
S f (ω 0 ) = ∑ T(ω 0 ,ω i )S f (ω i )T H (ω 0 ,ω i )
N i =1
1 N
= v s (ω 0 ) ∑ λ(ω i )v sH (ω 0 )
N i =1
(5.41)

= v s (ω 0 )λv sH (ω 0 )

S f (ω 0 ) is the focused spectral matrix whose eigenvectors are the same as

those of S f (ω 0 ) (before focusing) but its eigenvalues are equal to the averaged
eigenvalues of all spectral matrices taken over the frequency band. We may now
use the focused spectral matrix for DOA estimation.
5.2.2 Spatial Resampling: The basic idea is to resample the wavefield sensed
by a fixed ULA so as to create a virtual ULA with a different sensor spacing
depending upon the frequency. The spectral matrices computed at different
frequencies are then simply averaged. Let us take a specific case of two
wideband uncorrelated sources in the far field. The spectral matrix is given by

a(ω d sin θ )S (ω)a H (ω d sin θ ) + 

 c
0 0
c
0 
S f (ω) =   + σ ηI
2
d d
a(ω sin θ )S (ω)a H (ω sin θ ) 
 c
1 1
c
1


where S0 (ω ) and S1 (ω ) are spectra of the first and the second source
respectively. We compute the spectral matrix at two different frequencies, ω 0
and ω1 , with different sensor spacings, d0 and d1 , where

ω0
d1 = d0
ω1
and form a sum.

S f (ω 0 ) + S f (ω1 ) =
a(ω d0 sin θ )S (ω )a H (ω d0 sin θ ) 
 0
c
0 0 0 0
c
0 
 + σ ηI
2

d d
+a(ω 0 0 sin θ1 )S1 (ω 0 )a H (ω 0 0 sin θ1 )
 c0 c 

© 2001 by CRC Press LLC

 a(ω d1 sin θ )S (ω )a H (ω d1 sin θ ) 
 1
c
0 0 1 1
c
0 

+  + σ ηI
2
d d
+a(ω1 1 sin θ1 )S1 (ω1 )a H (ω1 1 sin θ1 )
 c c 
d d
= a(ω 0 0 sin θ 0 )[ S0 (ω 0 ) + S0 (ω i )]a H (ω 0 0 sin θ 0 )
c c
d0 d
+ a(ω 0 sin θ1 )[ S1 (ω 0 ) + S1 (ω i )]a H (ω 0 0 sin θ1 ) + 2σ 2ηI
c c

S f (ω 0 ) + S f (ω1 ) =
a(ω d0 sin θ )S (ω )a H (ω d0 sin θ ) 
 0
c
0 0 0 0
c
0 
 + σ ηI
2

d d
+a(ω 0 0 sin θ1 )S1 (ω 0 )a (ω 0 0 sin θ1 )
H
 c0 c 
a(ω d1 sin θ )S (ω )a H (ω d1 sin θ ) 
 1
c
0 0 1 1
c
0 

+  + σ ηI
2
(5.42)
+a(ω1 d1 sin θ1 )S1 (ω1 )a H (ω1 d1 sin θ1 )
 c c 
d d
= a(ω 0 0 sin θ 0 )[ S0 (ω 0 ) + S0 (ω i )]a H (ω 0 0 sin θ 0 )
c c
d d
+a(ω 0 0 sin θ1 )[ S1 (ω 0 ) + S1 (ω i )]a H (ω 0 0 sin θ1 ) + 2σ 2ηI
c c
Let us assume that N virtual arrays, each with different sensor spacing, have
been created by a resampling process. We can generalize (5.42),

1 N −1 d0 1 N −1 d
∑
N i=0
S f (ω i ) = a(ω 0
c
sin θ 0 ) ∑
N i=0
S0 (ω i ) a H (ω 0 0 sin θ 0 )
c
d0 1 N −1 d
+ a(ω 0 sin θ1 ) ∑ S1 (ω i ) a H (ω 0 0 sin θ1 ) + σ 2ηI
c N i=0 c
(5.43)

The above procedure for focusing may be extended to P sources; some

of them may be correlated. While the idea of resampling is conceptually simple

© 2001 by CRC Press LLC

 its implementation would require us to perform interpolation of the wavefield
in between physical sensors. Fortunately, this is possible because in a
homogeneous medium the wavefield is limited to a spatial frequency range,
ω2
u2 + v2 ≤ (see chapter 1, p. 14). In two dimensions, the spatial frequency
c2
ω ω
range is − ≤ u ≤ ; hence the wavefield is spatially bandlimited. This fact
c c
has been exploited in [21] for resampling.
The interpolation filter is a lowpass filter with its passband given by
ω ω λ
− ≤ u ≤ . Consider a ULA with sensor spacing equal to d0 ( d0 = ).
c c 2
c
The maximum temporal frequency will be ω max = π . The low pass filter
d0
is given by

(x − md0 )
sin ω max
h(m) = c , m = 0, ±1,..., ±∞ (5.44)
(x − md0 )
ω max
c

where x denotes the point where interpolation is desired, for example,

d0ω max
x = m ′dk where dk = .
ωk
5.2.3 Spatio Temporal Covariance Matrix (STCM): In chapter 2 we introduced
the concept of extended direction vector (chapter 2, p. 129) and STCM using
the extended direction vector and the source spectrum (2.67) which we reproduce
here for convenience,

CSTCM =
∞

E{f stacked (t)f (t)} =

1 (2.67)
∫
H
stacked h(ω,ϕ 0 )h H (ω,ϕ 0 )S f 0 (ω)dω
2π − ∞

In this subsection we shall show how the STCM can be used for the direction
of arrival estimation of a broadband source. For this it is necessary that the
eigenstructure, in particular, the rank of STCM, will have to be ascertained.
This has been done for a ULA in [23] and for a UCA in [24, 4]. We shall
assume that the source spectrum is a smooth function and that it may be
approximated by a piecewise constant function

© 2001 by CRC Press LLC

 l=L
S f 0 (ω) ≈ ∑S
l=−L
0l rect(ω − ∆ωl) (5.45)

where

1 1 1
rect(ω − ∆ωl) = for ∆ω(l − ) ≤ ω ≤ ∆ω(l + )
∆ω 2 2
=0 otherwise

and S0l (S0,l = S0,−l ) is the average spectrum in the lth frequency bin, ∆ωl .
Using (5.45) in (2.67) we obtain

1
∆ω(l − )
l=L 2
1 1
CSTCM = ∑ S0l
2π l = − L ∆ω ∫ h(ω,ϕ 0 )h H (ω,ϕ 0 )dω (5.46)
1
− ∆ω(l − )
2

where we have divided the frequency interval ±π into 2L+1 nonoverlapping

2π
frequency bins (= ) . Using the definition of the extended direction vector
∆ω
given in (2.67) in the integral in (5.46) we obtain the following result:

1
∆ω(l − )
2
1
2π∆ω ∫ h(ω,ϕ
1
0 )h H (ω,ϕ 0 )dω = g l Qg lH
− ∆ω(l − )
2

where

gl =
e − jωτ 0 ,e − jωτ1 ,...,e − jωτ M −1 ;e − jω( ∆t + τ 0 ) ,e − jω( ∆t + τ1 ) ,...,
 
diag e − jω( ∆t + τ M −1 ) ;...;e − jω(( N −1)∆t + τ 0 ) ,e − jω(( N −1)∆t + τ1 ) ,..., 
 − jω(( N −1)∆t + τ M −1 ) 
e  ω = ∆ωl

and
∆ω 
[Q]α,α' = sin c ((n − n' )∆t + τ m − τ m' )
 2 

© 2001 by CRC Press LLC

 where α = m + n × M , α' = m' +n' ×M , n,n' = 0,1,..., N − 1 and
m, m' = 0,1,..., M − 1. Note that the time delays (τ 0 , τ1 ,..., τ M −1 ) depend
upon the array geometry and, of course, on the azimuth; for example, for a
d d
(0, sin ϕ,...,(M − 1) sin ϕ) and for a
linear array, the time delays are
c c
2π
circular array the time delays are (δ cosϕ,δ cos( − ϕ),...,
M
2π a
δ cos( (M − 1) − ϕ)) where δ = and a is the radius of the circular
M c
array. Equation (5.46) may be expressed as

l=L
CSTCM = ∑S
l=−L
0l g l Qg lH (5.47)

The rank of CSTCM is determined by the rank of matrix Q . It follows from the
results in [11, 24] that 99.99% energy is contained in the first
(( N − 1)∆t + 2a ) ∆ω + 1 eigenvalues of Q where [x] denotes the next
 c 2π 
integer greater than x. A comparison of the theoretical and numerically
determined rank of CSTCM is given in fig. 5.5 for a circular array. For a linear
array the corresponding number is given by
(( N − 1)∆t + − ϕ ∆ω
+ 1 [19]. Note that the dimension of the
(M 1)d cos
 )
c 2π 
signal subspace is approximately equal to the time bandwidth product.
When the size of the observation space, that is, dimensions of
2a ∆ω 
CSTCM , is larger than the rank of CSTCM , NM > (( N − 1)∆t + ) + 1 ,
 c 2π 
there exists a null subspace, a subspace of the observation space, of dimension
equal to

2a ∆ω 
Dim{v null } = NM − ((N − 1)∆t + ) + 1
 c 2π 

where v null eigenvectors correspond to the insignificant eigenvalues. Note that

the null subspace thus defined will not be an ideal null space with zero power;
some residual power (<0.01%) will be present.

© 2001 by CRC Press LLC

 70

60

50
Rank

Exp

40

Theory
30

20
0 10 20 30 40 50 60 70
Radius / λ min

Figure 5.5: The effective rank of signal only STCM as a function of the radius of a
circular array. The drop in the rank at radius 30λ min is due to a transition from
smooth spectrum to line spectrum. We have considered a sixteen sensor equispaced
circular array with other parameters N=4, B=0.8 (normalized bandwidth) and one
source.(From [4]. c 1994, With permission from Elsevier Science)

Now, consider the following quadratic,

∞
1
2π −∫∞
v iH CSTCM v i = v iH h(ω,ϕ 0 )h H (ω,ϕ 0 )v i S f 0 (ω)dω
(5.48)
∞
1 2
= ∫
2π − ∞
v iH h(ω,ϕ 0 ) S f 0 (ω)dω ≈ 0

Assume that S f 0 (ω) > 0 over some frequency band. For (5.48) to hold good
we must have in that frequency band

2
v iH h(ω,ϕ 0 ) ≈ 0 i ∈null space (5.49)

Equation (5.49) is the basic result used for estimation of ϕ 0 . For the purpose
of locating the null we can define a parametric spectrum as in narrowband
MUSIC.

© 2001 by CRC Press LLC

 5.2.4 Number of Sources: We have so far assumed that the number of sources
is known a priori or can be estimated from the knowledge of the significant and
the repeated eigenvalues of the spectral matrix. The need for this information
arises in the estimation of signal and noise subspaces. In time series analysis
the equivalent problem is estimation of the number of sinusoids or estimation
of the order of a time series model. A lot of effort has gone into this problem
in time series analysis [25]. In this section we shall describe a test known as
the sphericity test and its modern versions for estimating the number of
repeated eigenvalues. Let P plane wavefronts be incident on an M sensor array.
A covariance matrix of size Μ × Μ (Μ>P) has been computed using a finite
number of N snapshots. Let λ̂1 > λ̂ 2 > λ̂ 3 >...> λ̂ M be the eigenvalues of
the estimated spectral matrix. We begin with the hypothesis that there are p
wavefronts; then a subset of smaller eigenvalues, λ̂ p+1 > λ̂ p+ 2 >...> λ̂ M , is
equal (or repeated). The null hypothis, that the smallest eigenvalues has a
multiplicity of M-p, is tested starting at p=0 till the test fails. Mauchley [26]
suggested a test known as the sphericity test to verify the equality of all
eigenvalues of an estimated covariance matrix. Define a likelihood ratio statistic

M− p N
 1 M  
 ∑ i  
λ̂
  M − p i = p+1 
Γ(λ̂ p+1 , λ̂ p+ 2 ,..., λ̂ M ) = ln  M  (5.50a)
 ∏ i λ̂ 
 i = p+1 
 

where N stands for number of snapshots. The log likelihood ratio is then
compared with a threshold γ. Whenever

Γ(λ̂ p+1 , λ̂ p+ 2 ,..., λ̂ M ) > γ

the test is said to have failed. Modified forms of the sphericity tests have been
suggested in References [27, 28].
The choice of γ is subjective. Alternate approaches which do not
require a subjective judgment have been proposed [29]. The number of
wavefronts is given by that value of p in the range 0 to M − 1 where, either

Γ(λ̂ p+1 , λ̂ p+ 2 ,..., λ̂ M ) + p(2M − p) (5.50b)

or

© 2001 by CRC Press LLC

 p
Γ(λ̂ p+1 , λ̂ p+ 2 ,..., λ̂ M ) + (2M − p)ln N (5.50c)
2
is minimum. The first form is known as Akaike information criterion (AIC)
and the second form is known as minimum description length (MDL). The
expression consists of two parts; the first part is the log likelihood ratio which
decreases monotonically and the second part is a penalty term which increases
monotonically as p increases from 0 to M − 1 . The penalty term is equal to
free adjusted parameters. The log likelihood ratio plot around p=P undergoes a
rapid change of slope, which is responsible for the occurrence of a minimum at
p=P. This is demonstrated in fig. 5.6.
The performance of AIC and MDL criteria has been studied in the context
of detection of plane waves incident on a linear array in [30]. They define a
probability of error as

prob.of error = prob( pmin < P HP ) + prob( pmin > P HP )

where pmin stands for the position of the minimum of (5.50) and HP denotes
the hypothesis that the true number of signal sources is P. The probability of
error for the MDL criterion monotonically goes to zero for a large number of
snapshots or high snr. In contrast, the probability of error for the AIC criterion
tends to a small finite value. However, for a small number of snapshots or low
snr the performance of AIC is better than that of MDL.

§5.3 Coded Signals:

In communication a message bearing signal is specifically designed to carry
maximum information with minimum degradation. In analog communication,
amplitude modulated and frequency modulated signals are often used while in
digital communication each pulse representing a bit is modulated with a sine
wave (frequency shift keying) or with a pseudorandom sequence (spread
spectrum). From the signal processing point, as these signals belong to a class
of cyclostationary processes whose covariance function is periodic, but the
background noise is simple ordinary stochastic process, these differences have
been exploited in array processing for the direction arrival estimation.
Sometimes, it is necessary to localize an active source for improved
communication as in dense wireless telephone user environment. Here we
would like to emphasize how a specially coded signal leads us to newer
approaches to localization. We shall demonstrate with three different types of
coded signals, namely, (i) multitone signal where the tones are spaced at
prescribed frequency intervals but with random phases, (ii) binary phase shift
keying (BPSK) signal and (iii) cyclostationary signals.
5.3.1 Multitone Signal: Consider a circular boundary array (see chapter 2) of M
equispaced, omnidirectional, wideband sensors (fig. 5.7) with a target which is

© 2001 by CRC Press LLC

 1000

(a)
800

600
Likelihood Ratio

400

200

0
0 2 4 6 8 10 12 14
p

1000

(b)
800
Measure

600

400

200
0 2 4 6 8 10 12 14
p

Figure 5.6: (a) Log likelihood ratio as a function of p, number of assumed

wavefronts, (b) AIC (filled circles) and MDL (empty circles) as a function of p. Note
the minimum at p=2. 20 frequency snapshots. 16 sensor ULA with λ/2 spacing. Two
uncorrelated wavefronts are incident at angles 24o and 32o. snr=10 db per source.

© 2001 by CRC Press LLC

surrounded by the array. A source is assumed at a point with polar coordinates
(r,ϕ) and it radiates a signal which is a sum of harmonically related random
sinusoids,

Q−1
f (t) = ∑ α k e − j (ω 0 + k∆ω )t (5.51)
k =0

where α k is the complex amplitude of kth sinusoid. The output of the ith
sensor is given by f i (t) = f (t + ∆τ i ) + ηi (t) , where ∆τ i is time delay at
the ith sensor with respect to a fictitious sensor at the center of the array. Each
sensor output is tapped at N time instants and arranged in a vector form,

f i = col{f i (0), f i (1),... f i (N − 1)}

The output data vector of the ith sensor may be written in a compact form
shown below:

f i = a(ri )HA i ε + ηi 5.52)

where
H = [h 0 ,h1 ,...hQ−1 ]
h k = col{1,e jω k ,...e j ( N −1)ω k }
ω k = ω 0 + k∆ ω
jω Q−1 ∆τ i
A i = diag{e jω 0 ∆τ i ,e jω 1 ∆τ i ,...e }
ε = col[α 0 ,α1 ,...α Q−1 ]

Next, we shall stack up all data vectors into one large vector F of dimension
MNx1. It may be expressed as

F = TDE + η (5.53)

where

T = diag{H,H,...H} ,(NMxMQ),
D = diag{a(r0 )A 0 , a(r1 )A1 ,...a(r M −1 )A M −1} , (MQxMQ) and
T
E = [εT, ε T, ε T, ... ] , (MQx1)

The covariance matrix (STCM) of F is given by

© 2001 by CRC Press LLC

 S
r 1
Source 1

ri r r
0

S ϕ
S0
i
(x i , y i )

Figure 5.7: A circular array of sensors and a source inside the circle. The range (r) and
azimuth (ϕ) of the source and all time delays are measured with respect to the center
of the array as shown.

c f = TDΓ 0 D H T H + σ 2ηI

{ }
where Γ 0 = E EE H = 1 ⊗ Γ where 1 is a square matrix of size MxM whose
elements are all equal to 1, Γ = diag{γ 0 , γ 1 ,... γ Q−1} and γ 0 , γ 1 ,...γ Q−1 are
powers of the random sinusoids. Symbol ⊗ stands for Kronecker product. We
will assume hereafter that the noise variance σ 2η is known or has been
estimated from the array output when there is no signal transmission or by
averaging the noise eigenvalues of the covariance matrix and that it has been
subtracted from the covariance matrix.
Let us consider the structure of the mth column of the STCM. By
straightforward multiplication it can be shown that the mth column of the
covariance matrix is given by

[ ]
T T T T
c m = c 0m ,c1m ,...,c Q−1
m (5.54)

© 2001 by CRC Press LLC

where c im = a(ri )a(r0 )HA i ΓA oH H H u m and u m is a Mx1 vector consisting of
H
all zeros except at mth location where there is one. Note A iΓA 0 is a
diagonal matrix given by

jω Q−1 µ i
A i ΓA 0H = diag{γ 0 e jω 0 µ i , γ 1e jω 1µ i ,...γ Q−1e } (5.55)

where µ i = ∆τ i − ∆τ 0 . Further, we note that

H −jmω − jmω −jmω

H u m = col{e 0
,e 1
, ... e Q− 1
} (5.56)

Using (5.55) and (5.56) in (5.54) we obtain

Cim = a(ri )a(r0 ) ×

{ }
jω Q−1 (µ i − m)
(5.57)
Hcol γ 0 e jω 0 (µ i − m) , γ 1e jω 1 (µ i − m) ,...γ Q−1e

Multiply on both sides of (5.57) by H # = (H H H) −1 H H , the pseudo-inverse of

H, and obtain

{
H # Cim = a(ri )a(r0 )col γ 0 e jω 0 (µ i − m) , γ 1e jω 1 (µ i − m) ,...γ Q−1e
jω Q−1 (µ i − m)
}
We now define a matrix D whose columns are H # Cim , i=0, 1, 2, ... M-1,

D = {H # C0m ,H # C1m ,...H # CmM −1}

(5.58)
{
= a 2 (r0 )diag γ 0 e − jω 0 m , γ 1e − jω 1 m ,...γ Q−1e
− jω Q−1 m
}
where
1, b1e jω 0 µ 1 , ...bM −1e jω 0 µ M −1 
 jω µ jω µ

1, b1e 1 1 , ...bM −1e 1 M −1 
G= ... 
 
 ... 
 jω Q−1 µ 1 jω Q−1 µ M −1 
1, b1e ,...bM −1e 

© 2001 by CRC Press LLC

 Source #1 Source #2
Range Azimuth Range Azimuth
24.99 50.01 30.00 50.02
(25.00) (50.00) (30.00) (50.00)
89.99 90.00 70.01 -90.00
(90.00) (90.00) (70.00) (-90.00)
9.99 29.99 50.02 150.02
(10.00) (30.00) (50.00) (150.00)
50.01 -60.00 49.99 -60.00
(50.1) (-60.00) (50.00) (-60.00)
Table 5.4: Localization of two sources by a circular (boundary) array of 8 sensors.
snr=10 dB. The STCM was averaged over one hundred independent estimates. The
wave speed was assumed as 1500 meters/sec. The numbers inside brackets represent
the true values (range in meters and azimuth in degrees).

a(ri )
and bi = , i=0,1,...,M-1. Since the location information is present in
a(r0 )
µ 0 ,µ1 ,...µ M −1 , our aim naturally is to estimate these from the columns of
G. Each column may be considered as a complex sinusoid whose frequency can
be estimated, provided ∆ωµ i ≤ π . For a boundary array, specifically a
circular array, this limitation can be overcome. In chapter 2 we have given an
algorithm to estimate time delays or the location of the source from the phase
estimates.
As a numerical example we have considered two sources each emitting
eight random tones. The frequencies emitted by the first source are (0, 200,
400, 600, 800, 1000, 1200, 1400 Hz) and those by the second are (1600, 1800,
2000, 2200, 2400, 2600, 2800, 3000 Hz). An eight sensor UCA of radius 100
meters surrounds both sources. Sixteen delayed snapshots (taps) were used
making the size of STCM as 128x128. The results are displayed in table 5.4.
5.3.2 BPSK Signal: Binary phase shift keying signal consists of square pulses
of fixed width with amplitude given by a Bernoulli random variable, an
outcome of a coin tossing experiment where head equals +1 and tail equals -1.
A typical BPSK sequence is shown in fig. 5.8. The analytic representation of
BPSK signal is given by
L −1
c0 (t) = ∑ c0,n hT c (t − nTc ) (5.59a)
n=0

where c0,n , n=0,1,...,L-1 are Bernoulli random variables and hT c (t) is a

© 2001 by CRC Press LLC

 1
Amplitude

-1

0 127 255
Time units

Figure 5.8: BPSK waveform. A train of rectangular pulses with amplitude alternating
between +1 and -1.

rectangular function of width Tc . One of the most useful properties of a BPSK

sequence is its narrow autocorrelation function with sidelobes whose variance is
less than the inverse of the sequence length. We can easily generate many
uncorrelated BPSK sequences. BPSK sequences have beenused for time delay
measurement [31], digital communication [32], and mobile communication
[33] where it is of interest to exploit the spatial diversity for the purpose of
increasing the capacity. Here we shall briefly show how the azimuth of a source
emitting a known BPSK signal may be estimated.
The signal model considered here is as in [34]. There are Q uncorrelated
sources emitting data bits which are encoded with a user specific random
sequence, a BPSK sequence, for example.

Q
xm (t) = ∑ pk am (θ k )∑ bk,l ck (t − lT s − τ k ) + ηm (t)
k =0 l (5.59b)
m = 0,1,2,... M − 1
where
am (θ k ) : response of mth sensor to a signal coming from kth user

© 2001 by CRC Press LLC

 pk : signal amplitude of kth user
bk,l : data stream from kth user
ck (t) : random code of kth user
Ts: bit duration
τ k : delay of a signal from kth user
η (t) : Noise at mth sensor
m
Q : number of users

There are two different approaches to DOA estimation with BPSK coded signal.
In the first approach the usual covariance matrix is computed and in the second
approach, due to the fact that the code used by the user of interest is known, the
received signal is first cross-correlated with that code. The output of the cross-
correlator is then used to compute the covariance matrix. Both approaches yield
similar results; though in the second approach the interference from the users of
no interest is reduced by a factor proportional to the code length.
Precorrelation Covariance Matrix: The outputs of the sensor array, after
removing the carrier, are arranged in a vector form,

Q
f(t) = ∑ pk a(θ k )∑ bk,l ck (t − lT s − τ k ) + η(t) (5.60a)
k =0 l

where f(t) , a(θ k ) , and η(t) are Mx1 vectors. The precorrelation covariance
matrix is given by

c f = E{f(t)f H (t)} (5.60b)

Using (5.60a) in (5.60b) we obtain

 2 Q Q

 η ∑ ∑ pk pk ′ a(θ k )a(θ k ′ )
σ I + 
 k =0 k′=0 
cf =   (5.60c)
 
∑
l
∑l bk,l ck (t − lTs − τ k )bk ′,l ′ck ′ (t − l ′Ts − τ k ′ )
′ 

We shall assume that the data bits coming from different users are independent
and codes assigned to different users are uncorrelated. As a result, (5.60c)
reduces to

© 2001 by CRC Press LLC

 Q 2  
c f = ∑ pk a(θ k )a(θ k )∑ bk,l
2 2
ck (t − lT s − τ k ) + σ 2ηI
k = 0  l 

But the quantity inside the inner curly brackets is always equal to 1; hence we
obtain

Q 2 
c f = ∑ pk a(θ k )a(θ k ) + σ 2ηI
k = 0 

which may be expressed in a standard form (4.12b) as

[ ]
c f = a(θ 0 ),a(θ1 ),...,a(θQ−1 ) diag p02 , p12 ,..., pQ−1
2
{ }
[a(θ ),a(θ ),...,a(θ )] + σ I
H 2
0 1 Q−1 η (5.60d)

= A diag{ p , p ,..., p }A + σ I
2
0
2
1
2
Q−1
H 2
η

Postcorrelation Covariance Matrix: The outputs of the antenna array, after

removing the carrier, are correlated with the desired user code assuming
synchronization has been achieved. The postcorrelation array signal vector
corresponding to the lth bit is given by

(l +1)T s
1
g 0 (l) =
Ts ∫ f(t)c (t − lT )dt
lT s
0 s (5.61)

where f(t) stands for array signal vector without carrier. Using the signal
model given by (5.60a), eq. (5.61) may be written as

g 0 (l) =
(l +1)T s ∞
1
p0 a(θ 0 ) ∫ [ ∑ b0 j c0 (t − jT s )]c0 (t − lT s )dt
Ts lT s j =−∞

(l +1)T s ∞
1 Q−1
+ ∑ pk a(θk )
T s k =1 ∫ [∑ bkj ck (t − jT s + τ k )]c0 (t − lT s )dt
−∞
(5.62)
lT s
(l +1)T s
1
+
Ts ∫ η(t)c (t − lT )dt
lT s
0 s

© 2001 by CRC Press LLC

 where we have set τ 0 = 0 . In (5.62) the first term, equal to b0 k , is the signal
term. The second term represents the interference. It is evaluated as follows:
The integral is split into two parts, lT s to lT s + τ k , and lT s + τ k to (l + 1)T s .
After a change of variable it reduces to

τk τk
1 Q−1  
∑
T s k =1
pk a(θ k ) bkl −1 ∫ ck (t − T s − τ k )c0 (t)dt + bkl ∫ ck (t − τ k )c0 (t)dt 
 
0 0

= b1l + bl2 (5.63)

Finally, the noise term reduces to

T
1 s
T s ∫0
ηl = η(t + lT s )c0 (t)dt

The postcorrelation covariance matrix of the array signal is given by

Cg0 g0 = E{g 0 (l)g 0H (l)}

(5.64a)
H H
= p02 a(θ 0 )a H (θ 0 ) + E{b1l b1l } + E{bl2 bl2 } + E{ηl ηlH }
H
All cross terms vanish. First, we shall evaluate E{b1l b1l }.

1 Q 2
E{b b } = 2 ∑ pk a(θ k )a(θ k ) H ρ1
1 1H
l l (5.64b)
T s k =1

where

{ τk
ρ1 = E [ ∫ ck (t + T s − τ k )c0 (t)dt]2
0 }
and the sources are assumed to be uncorrelated. Using the representation of
BPSK signal (5.59a) we obtain

ρ1 =
 τ k L −1 L −1  
2

E  ∫ ∑ ∑ c0, m0 ck, mk hT c (t − m0 Tc )hT c (t + T s − mk Tc − τ k )dt  

 m0 mk  
0

© 2001 by CRC Press LLC

which, on account of the fact that c0, m0 and ck, mk are uncorrelated random
variables, reduces to

 τ k hT c (t − m0 Tc )
2
1 T s L −1 L −1 
ρ1 = ∫ ∑ ∑  ∫0  dτ k (5.65a)
Ts  hT c (t + T s − mk Tc − τ k )dt 
m0 mk 
0

where τ k , the arrival time of the signal from the kth user, is assumed to be
uniformly distributed over 0 to T s , the symbol duration. The integral over τ k
may be expressed as a sum over L integrals, one for each chip. When hT c is a
rectangular function the integral with respect to τ k in (5.65a) can be evaluated
in a closed form. After some algebraic manipulations we obtain

L2 3
ρ1 = Tc (5.65b)
3T s

Substituting (5.65b) in (5.64b) we obtain

1 Q 2
∑
H
E{b1l b1l } = pk a(θ k )a(θ k ) H (5.66a)
3L k =1

2 2H
Evaluation of E{bl bl } proceeds on the same lines as above. In fact the
result is identical, that is,

1 Q 2
∑
H H
E{bl2 bl2 } = pk a(θ k )a(θ k ) H = E{b1l b1l } (5.66b)
3L k =1

Finally, we shall evaluate the noise term in (5.62).

1  Ts Ts

E{ηl ηlH } = ∫ ∫ η(t + lT )η (t + lT )c (t )c (t )dt dt 
H
E 1 s 2 s 0 1 0 2 1 2
T s2  0 0

L −1 E {η(t + lT )η (t + lT )} H

∑∫ ∫
1 Tc Tc 1 c 2 c
=
T s2 m1 = 0
0 0 hT c (t1 − m1T c )hT c (t2 − m1T c )dt1dt2
σ 2η
= I (5.67)
L

© 2001 by CRC Press LLC

 where we assumed that the noise is spatially uncorrelated but temporally
correlated over the chip width T c . Using equations (5.65b), (5.66) and (5.67) in
(5.64) the postcorrelation covariance matrix may be expressed as

2 Q 2 σ 2

c g0 g0 = p a(θ 0 )a (θ 0 ) +
2
0
H
∑
3L k =1
pk a(θ k )a H (θ k ) + η I
L
(5.68)

which may be expressed in a standard form (4.12b)

[ ]
c g0 g0 = a(θ 0 ),a(θ1 ),...,a(θQ−1 ) diag  p02 ,
 3L
2 2
p1 ,...,
2 2 
3L
pQ−1 

σ 2η
[a(θ ),a(θ ),...,a(θ )]
H
0 1 Q−1 + I
L
2 2  H ση
2
 2 2
= Adiag  p0 , 2
p1 ,..., pQ−1 A + I (5.69)
 3L 3L  L

The directions of arrival (DOA) may be estimated following the subspace

approach described earlier in §5.1 and §5.2.
5.3.3 Cyclostationary Signals: A signal defined in (5.58) and other similar
communication signals possess an important property, namely, its temporal
covariance function is periodic. This is a result of transmission of symbols at a
fixed rate and also due to the use of a periodic carrier. Indeed the periodicity in
the covariance function of the signal, also known as cyclic frequency α , is
equal to lf b where l is an integer and f b is baud rate, that is, the number of
symbols transmitted per second. The baud rate is a unique property associated
with each source and it is not affected by propagation. Since the baud rate of a
signal of interest (SOI) is known a priori and it is different from that of the
other interfering signals, it is possible to distinguish the SOI from the
interference and the system noise whose covariance function is known to be
aperiodic. In this section we shall describe how a subspace method, i.e.,
MUSIC, may be devised by exploiting the property of cyclostationarity. The
background information on cyclostationary process will not be covered here as
such material is already available in a book [35] and in a popular exposition
[36]. However, we shall introduce enough material that is essential for the
understanding of its application to DOA estimation.

Let f (t) =  f 0 (t), f 0 (t − sin θ 0 ),..., f 0 (t − (M − 1) sin θ 0 ) 

d d
 c c 

© 2001 by CRC Press LLC

be the output of a M sensor array on which a cyclostationary signal f 0 (t) is
incident with an angle of incidence θ 0 . We define frequency shifted versions of
f(t)
f + (t) = f(t)e jπαt
(5.70)
f − (t) = f(t)e − jπαt

and cyclic covariance matrix

τ τ
E f − (t + )f +H (t − )
2
1
c αf (τ) = ∑
T t=− T  2 2 
2
T

τ τ
E f(t + )f H (t − )e − j 2 παt
2
1
= ∑
T t=− T  2 2 
2

T→∞

Let us show how to evaluate the (k,l)th element of the matrix c αf (τ) .

[c (τ)]
α
f kl
T
1  ∗
2 d τ d τ − j 2 παt 
= ∑ E  f 0 (t + k sin θ 0 − ) f 0 (t + l sin θ 0 + )e 
T t=− 
T c 2 c 2 
2

T→∞
d
d jπα(k +l ) sin θ 0
α
= c (τ + (k − l) sin θ 0 )e
f0
c
c
(5.71a)

When a source is narrowband with center frequency ω c so that following

jω c τ
approximation holds good, f 0 (t + τ) ≈ f 0 (t)e (see (4.15b)), the (k,l)th
α
element of the matrix c f (τ) may be approximated as
d d

[ ]
jω c ( k −l ) sin θ 0 jπα(k +l ) sin θ 0
c αf (τ) ≈ c αf 0 (τ)e c
e c
, which may be further
kl

© 2001 by CRC Press LLC

 d

[c (τ)]
jω c (k −l ) sin θ 0
α α
approximated as f ≈ c (τ)e
f0
c
for ω c >> 2πα . Using
kl
the narrowband approximation and the assumption ω c >> 2πα we can
express the cyclic covariance matrix as

c αf (τ) ≈ a 0 a 0H c αf 0 (τ) (5.71b)

The above relation is quite similar to (4.12b), which was the starting point in
the subspace algorithm, for example, in MUSIC. Naturally, based on (5.71b), a
subspace algorithm known as Cyclic MUSIC has been proposed in [37].
Although, in deriving (5.71b) we have assumed a single source, it holds good
even in the presence of multiple signals with different cyclic frequencies and
any type of stationary noise.
Let us consider the diagonal terms of the cyclic covariance matrix.
From (5.71a) the diagonal terms, k=l, are given by

[ ]
j 2 παl sin θ 0
c αf (τ) = c αf 0 (τ)e c
(5.72a)
k =l

which we shall express in a matrix form. Let

{[
c̃ αf (τ) = col c αf (τ) ]k =l = 0
[
, c αf (τ) ] k =l =1
[
,..., c αf (τ) ]
k =l = M −1 }
 j 2 πα dc sin θ 0 j 2 πα( M −1) sin θ 0 
d
a(α,θ 0 ) = col 1,e ,...,e c

 

and using these vectors, (5.72a) may be written as

c̃ αf (τ) = c αf 0 (τ)a(α,θ 0 ) (5.72b)

and for P uncorrelated sources, but with the same cyclic frequency, we obtain

c αf 0 (τ) 
 
c̃ αf (τ) = {a(α,θ 0 ),a(α,θ1 ),...,a(α,θ P −1 )}...  (5.72c)
c α (τ)
 f P−1 

Note that in deriving (5.72) we have not used the narrowband approximation
which was earlier used in deriving (5.71b). The computer simulation results
reported in [38] showed improved results obtained by using (5.72c) over (5.71b)

© 2001 by CRC Press LLC

except when the bandwidth is only a tiny fraction (<0.05) of the carrier
frequency. In (5.72b) or (5.72c) the noise term is absent as it is not a
cyclostationary process. This is a very important feature of any system using
the cyclostationary signals. Whether the noise is correlated or white its
cyclostationary autocorrelation function at any cyclic frequency vanishes.
Another important feature of great signficance is that the condition of two
sources being mutually cyclically uncorrelated is weaker than the condition of
mutually uncorrelated. Two cyclostationary signals f(t) and g(t) are said to be
uncorrelated if their cross-cyclic covariance function defined as

τ τ
c αfg (τ) = ∑ E  f ∗ (t − )g(t + )e − j 2 παt
1 2
T t=− T  2 2  (5.73)
2

T→∞

vanishes for all τ and α is equal to the cyclic frequency either of f(t) or g(t).
To demonstrate this property let us consider f (t) = s(t)cos(ω1t + θ1 ) and
g(t) = s(t)cos(ω 2 t + θ 2 ) and substituting in (5.73) we obtain

c αfg (τ)=
T cos((ω + ω )t − (ω − ω ) τ + θ + θ ) 
1 2  1 2 1 2
2
1 2  − j 2 παt
cs (τ) ∑  e
2T t = − T  + cos((ω1 − ω 2 )t − (ω1 + ω 2 ) τ + θ1 − θ 2 )
2  2 
T→∞

which becomes zero except when 2πα = ±(ω1 + ω 2 ) or = ±(ω1 − ω 2 ) .

Thus, two sources transmitting the same message but with different carriers
become cyclically uncorrelated. Generalizing the above result, P sources
transmitting even the same message but with different carrier frequencies may
become cyclically uncorrelated. Further, we notice that even when the carrier
frequencies are the same, the sources can become cyclically uncorrelated unless
we choose 2πα = ±2ω1 or 0. As a result, it is possible through an
appropriate choice of α to selectively cancel all signals of no interest. As a
numerical example, we computed the cyclic cross-correlation function between
two sources radiating the same stochastic signal with different carrier
frequencies. In fig. 5.9a the cyclic cross-correlation function is shown as a
function of the cyclic frequency. The carrier frequencies are 0.15 Hz and 0.20
Hz. (The Nyquist frequency is 0.5 Hz and 1024 time samples were used.)

© 2001 by CRC Press LLC

 0.25

0.20 (a)
Cyclic cross-correlation

0.15

0.10

0.05

0.00
0.0 0.1 0.2 0.3 0.4
Frequency in Hz

0.25
(b)

0.20
Cyclic cross-correlation

0.15

0.10

0.05

0.00
0.150 0.151 0.152
Frequency in Hz
Figure 5.9: Cyclic cross-correlation function at zero lag as a function of (a) cyclic
frequency and (b) carrier frequency in Hz.

© 2001 by CRC Press LLC

Note the peaks at the sum and difference frequencies. In fig. 5.9a the cyclic
cross-correlation function is shown as a function of the carrier frequency of the
second source.

§5.4 Array Calibration:

Throughout we have implicitly assumed that all sensors are ideal, that is, they
are omnidirectional point sensors with constant response. Further, the shape of
the array is fully known. If we are dealing with a ULA we assume that all
sensors are on a perfectly straight line and sensors are equispaced. Any
deviations from the assumptions can cause considerable loss of performance,
particularly in the use of subspace methods [39]. One way out of this
limitation is to actually measure the properties of the array, including its shape,
and use this information in the design of the processor. Often this requires a
source whose location is known. The properties of the array can be measured, a
process known as array calibration. We shall describe two types of calibrations.
In the first instance we shall give a method of computing the amplitude and the
phase variations. Next we shall describe a method for the shape estimation,
which is vital in the use of flexible arrays.
5.4.1 Amplitude and Phase Variation of a Sensor:
From (2.17e) the direction vector may be expressed as a product of two
components

a(θ 0 ) = α(θ 0 )φ(θ 0 )

where

α(θ 0 ) = diag{α 0 (θ 0 ),α1 (θ 0 ),...α M −1 (θ 0 )}

d 2d (M-1)d
- jω sin θ 0 - jω sin θ 0 - jω sin θ 0
φ(θ 0 ) = col{1,e c
,e c
...e c
}

The purpose of array calibration is to estimate α(θ 0 ) from actual array

observations. We preclude the possibility of direct in-situ measurement of
sensor sensitivity (hardware calibration). Recall the relation we had derived
H
between the direction vectors and eigenvectors of v s v s (5.6d). Consider a
single source whose direction of arrival is known. For single source (5.6d) may
be expressed as

α(θ 0 )φ(θ 0 ) = v s v sH α(θ 0 )φ(θ 0 ) (5.74)

Clearly α(θ 0 )φ(θ 0 ) is an eigenvector of v s v s and the corresponding

H

eigenvalue is 1. Since the direction of arrival of the source is known we can

compute φ(θ 0 ) and demodulate the eigenvector. We can thus estimate α(θ 0 )

© 2001 by CRC Press LLC

 but for a complex constant. Assume that α(θ 0 ) is independent of θ 0 . We
shall express (5.74) in a different form. Note that

d (M-1)d
- jω sin θ 0 - jω sin θ 0
α φ(θ 0 ) = diag{α 0 ,α1e c
,...α M −1e c
}
d (M-1)d
- jω sin θ 0 - jω sin θ 0
= diag{1,e c
,...e c
}col{α 0 , ...α M −1}
= φ d (θ 0 )α c

Using the above result in (5.74) we obtain

α c = φ dH (θ 0 )v s v sH φ d (θ 0 )α c (5.75a)

When there are P sources with known DOAs, θ p , p = 0,1,..., P − 1 (5.75a)

may be expressed as α c = Qα c where

P −1
Q = ∑ φ dH (θ p )v s v sH φ d (θ p ) (5.75b)
p= 0

and α c is the eigenvector of Q corresponding to its largest eigenvalue equal to

P [40].
5.4.2 Shape Estimation: In any practical towed array system used in sonar or
seismic exploration the position of sensors is continuously monitored.
Additionally, it is also possible to estimate the shape from the array output.
Here from the point of array signal processing our natural interest is in the
second approach. Consider a ULA which has been deformed either on account of
towing or ocean currents. Assume that we have a known source radiating a
narrowband signal and the background noise is spatially uncorrelated. We have
shown in (5.4) the eigenvector of the spatial covariance matrix corresponding to
the largest eigenvalue is proportional to the direction vector. In fact, the
1
relation is given by v0 = a 0 . For deformed array the direction vector
M
may be obtained from (2.34c)

 − jω c dc [γ 1 sin θ 0 sin ϕ 0 + ε 1 sin θ 0 cos ϕ 0 + ξ1 sin θ 0 ] 

1,e ,...,
a(θ 0 ,ϕ 0 ) = col  d  (5.76)
e − jω c c [ γ M −1 sin θ 0 sin ϕ 0 + ε M −1 sin θ 0 cos ϕ 0 + ξ M −1 sin θ 0 ] 
 

© 2001 by CRC Press LLC

Therefore, we can relate the eigenvector to the direction vector

 − jω c dc [ γ 1 sin θ 0 sin ϕ 0 + ε 1 sin θ 0 cos ϕ 0 + ξ1 sin θ 0 ] 

1 1,e ,...,
v0 = col  d  (5.77a)
M e − jω c c [ γ M −1 sin θ 0 sin ϕ 0 + ε M −1 sin θ 0 cos ϕ 0 + ξ M −1 sin θ 0 ] 
 

Note that since the source is known, its azimuth and elevation are known a
priori. The unknowns are γ m ,ε m , m = 0,1,..., M − 1 . (Note
ξ m , m = 0,1,..., M − 1 are dependent on γ m ,ε m .) To solve for a pair of
unknowns we would need one more source, say, with different azimuth and
elevation, (θ1 ,ϕ1 ) . Once again the largest eigenvector may be related to the
direction vector of the second source.

 − jω c dc [ γ 1 sin θ1 sin ϕ1 + ε 1 sin θ1 cos ϕ1 + ξ1 sin θ1 ] 

1 1,e ,...,
v1 = col  d  (5.77b)
M e − jω c c [ γ M −1 sin θ1 sin ϕ1 + ε M −1 sin θ1 cos ϕ1 + ξ M −1 sin θ1 ] 
 

We shall, for the sake of simplicity, assume that the deformed array is in the x-
y plane and also place the calibrating source in the x-y plane. Then, a single
source is enough for estimation of γ m and ε m which now take a form

m
γ m = ∑ cosα i , m = 0,1,..., M − 1
i=0
(5.78)
m
ε m = ∑ sin α i , m = 0,1,..., M − 1
i=0

Using (5.78) in (5.77b) we obtain the following basic result

λc
2π
[
∠{v 0 }m − ∠{v 0 }m −1 + λ c n = ] (5.79)
d cosα m cosϕ 0 + d sin α m sin ϕ 0 = ∆xm cosϕ 0 + ∆ym sin ϕ 0

where n is an unknown integer and ∆xm and ∆ym are x,y coordinates of mth
sensor relative to m-1st sensor. The ambiguity, arising out of the unknown
integer, may be resolved through geometrical considerations. Note that the mth
sensor must be on a circle of radius d centered at

© 2001 by CRC Press LLC

 p1

n=2
d
m-1st p2

n=1

n=0

Figure 5.10 The ambiguity in (5.77) is resolved by requiring that the line it
represents must intersect the circle of radius d drawn at m-1st sensor. There are two
intersection points. The sensor may be at any one of the two intersections.

m-1 st sensor. Further the position of the sensor must satisfy (5.79), which,
incidentally, is an equation of a straight line [41]. For illustration let the line
corresponding to n=1 intersect the circle at two points p1 and p2 (see fig. 5.10).
The sensor can either be at p1 or p2. This ambiguity is resolved by choosing a
point which results in minimum array distortion [42].

§5.5 Source in Bounded Space:

When a source is in a bounded space a sensor array will receive signals reflected
from the boundaries enclosing the space, for example, radar returns from a low
flying object [43], an acoustic source in shallow water [44], a speaker in a
room [45]. In all these cases the reflections are strongly correlated with the
direct signal and come from a set of predictable directions. The complexity of
the multipath structure increases with the increasing number of reflecting
surfaces, as illustrated in fig. 5.11. We shall in this section briefly consider two
simple cases involving one reflecting surface (a low flying object) and two
reflecting surfaces (an acoustic source in shallow water) shown in fig. 5.11a &
b.

© 2001 by CRC Press LLC

 Source Array

(a)

Array
Source

(b)

Array

Source

(c)

Figure 5.11: Three examples of multipath structure with increasing complexity.

Assuming the boundaries are well defined we can predict the multipaths given the
source and array locations.

5.5.1 Single Reflecting Surface: Assume that the source emits a stationary
stochastic signal. A vertical sensor array is used to receive the signal radiated by
the source. The array output in frequency domain may be written in terms of
the radiated signal as follows:

© 2001 by CRC Press LLC

 dF(ω) = a 0 dF0 (ω) + w1 e − jωτ1 a1dF0 (ω) + dΝ(ω)
(5.80)
[ ]
= [a 0 , a1 ] 1, w1e − jωτ1 dF0 (ω) + dΝ(ω)
T

where dF0 (ω ) is the differential of the generalized Fourier transform of

stochastic signal emitted by the source, dΝ(ω) is the differential of the
generalized Fourier transform of the background noise presumed to be spatially
and temporally white and a 0 , and a1 are the direction vectors of direct and
reflected signals, respectively,

 − jω d sin θ 0 − jω ( M −1) sin θ 0 

d
a 0 = col 1,e c ,...,e c

 
 − jω d sin θ1 − jω ( M −1) sin θ 1 
d
a1 = col 1,e c ,...,e c

 

τ1 is the delay of the reflected signal relative to the direct arrival and w1 stands
for reflection coefficient. The spectral matrix of the array output is easily
derived from (5.80)

[
S f (ω) = [a 0 , a1 ] 1, w1e − jωτ1 ] [1, w e ][a , a ]
T − jωτ 1 H
1 0 1 S0 (ω) + σ 2ηI
(5.81)

[
ã = [a 0 , a1 ] 1, w1e − jωτ1 ]
T
Define = a 0 + w1e − jωτ1 a1 and rewrite (5.81) as

S f (ω) = ãS0 (ω)ã H + σ 2ηI (5.82)

Note that (5.82) is of the same form as (4.12b). Naturally, it is possible to

derive a subspace algorithm to estimate the parameters, θ 0 and θ1 ;
alternatively, range and height of the source above the reflecting surface, in
terms of which θ 0 and θ1 , can be expressed. The Music spectrum is now a
function of two parameters, range and height, instead of frequency as in the
conventional Music spectrum. For this reason it may be worthwhile to call it a
parametric spectrum. A numerical example of parametric spectrum for a source
situated above a reflecting surface and an eight sensor vertical ULA is given in
fig. 5.12. The peak of the parametric spectrum is correctly located but this good
result has been achieved because we have used the exact value of the reflection
coefficients. Even a small error, on the order of 1%, appear to completely alter

© 2001 by CRC Press LLC

 x 1015

0
150
1200
100 1100
He 1000
igh 50 900 ge
t Ran

Figure 5.12: Parametric spectrum for a source situated above a reflecting surface
(100 λ ) and 1000 λ away from an eight element vertical ULA. The array center is at
1 6 λ above the reflecting surface.

the picture; in particular, the range estimation becomes very difficult.

5.5.2 Two Reflecting Surfaces: To keep the analysis reasonably simple we
shall confine ourselves to two parallel reflecting surfaces and a vertical ULA. A
broadband signal is assumed, as simple frequency averaging seems to yield good
results (see fig. 5.13). Consider a uniform channel of depth H meters where a
vertical array of equispaced (spacing d0<λ/2) sensors is located at depth zR
(depth is measured from the surface of the channel to the top sensor of the
array) and a source of radiation at a depth zs and is horizontally separated by a
distance R0 from the array. We have noted in chapter 1 (page 26) that depending
upon the range and the channel characteristics the array would be able to see a
certain number of significant images, say P. The waveform received from P
images at mth sensor (m=0 is the top sensor) is given by

P −1
αp
f m (t) = ∑ f 0 (t − τ pm ) + ηm (t) (5.83)
p= 0 Rp

where f 0 (t) is the signal emitted by the source, presumed to be a stationary

stochastic process, τ pm is time delay from pth image to mth sensor, Rp is the

© 2001 by CRC Press LLC

 (a)

(b)

Figure 5.13: (a) Parametric spectrum at 1600 Hz is shown as a function of range and
depth. A broadband acoustic source is assumed at 2000m (range) and 25m (depth) in
a channel of depth 100m. Twenty frequency snapshots were used to compute the
spectral matrix. Next, the parametric spectra computed at 50 frequencies, equispaced
in the band 1000Hz to 2000Hz, were averaged. The averaged spectrum is shown in
(b). (From [47]. c 1999, With permission from Elsevier Science)

distance to the pth image and α p is the reflection coefficient for the ray arriving
from the pth image. ηm (t) is noise received by the mth sensor. It is easy to
show that

d
τ pm = τ p + (m − 1) sin θ p (5.84)
c

© 2001 by CRC Press LLC

where τ p is travel time from pth image to the top sensor and θ p is azimuth
angle (with respect to the horizontal plane) of the wave vector from pth image
and c is sound speed in sea water. This angle can be computed from the
geometry of the image structure as shown in fig.1.14. Using the spectral
representation of the stationary stochastic process (5.83) may be written in the
frequency domain as

P −1 d
j ( m −1)2 π sin θ p
dFm (ω) = ∑ w p dF0 (ω)e λ
+ dΝ m (ω) (5.85)
p= 0

α p jωτ p
where wp = e [44]. Let us express (5.85) in a compact matrix
Rp
notation by defining

1 ... 1 
 jω d sin θ d
jω sin θ P−1

e c 0 ... e c 
A= 
 ... 
 jω d ( M −1) sin θ d
jω ( M −1) sin θ P−1

e c 0
c 
 ... e 

and w = col[ w0 , w1 ,..., wP −1 ]. Equation (5.85) may be written in a matrix

form,

dF(ω) = AwdF0 (ω) + dΝ(ω) (5.86)

The spectral matrix is obtained from (5.86) by squaring and averaging

S f (ω) = Aww H A H S f 0 (ω) + σ 2ηI (5.87)

S f 0 (ω) (scalar) is the spectrum of the source radiation, and σ η is the

2
where

variance of the background noise. Define a vector à = Aw and rewrite (5.87)

as

S f (ω) = ÃS f 0 (ω)Ã H + σ 2ηI (5.88)

© 2001 by CRC Press LLC

 Note that the structure of spectral matrix in (5.88) is the same as in (4.12b).
This enables us to develop a subspace algorithm to estimate unknown
parameters, range and depth of the source (also azimuth, if a horizontal array is
used) [44].
Source localization in a bounded space is prone to errors in the
assumed model as well as the estimation errors due to finite available data. The
geometry of the channel needs` to be specified with an accuracy of a fraction of
wavelength and the wavespeed must be known accurately. In practice these
requirements are often difficult to satisfy. However, it is possible, at the cost of
increased array length, to achieve good results when the channel is only
partially known [46]. Good results have also been obtained with limited data
but using a broadband signal as demonstrated in fig. 5.13 [47].

§5.6 Exercises:
1. The line represented by (5.79) intersects the circle at two points p1 and p2
(see fig. 5.10). Let two adjacent sensors be on the x-axis. Show that one of the
points will be at the intersection of the x-axis and circle. Where will be the
second point?
2. In chapter 4 (4.13a) it was shown that A H Q = 0 where Q was defined in
terms of partitions of the spectral matrix. Is Q itself the noise subspace?
Remember that in obtaining Q no eigendecomposition was required.
3. Show that, taking into account the variation in the sensitivity of the
sensors, equation (5.72b) takes the form

c̃ αf (τ) = c αf (τ)α 2 (θ 0 )a(α,θ 0 )

where
2 2 2
α 2 (θ 0 ) = diag{α 0 (θ 0 ) , α1 (θ 0 ) ,..., α M −1 (θ 0 ) }.

Interestingly, phase variations themselves do not affect the relation given in

(5.71b).
4. How would you cancel the interference coming from the users of no interest
in equations (5.60d) and (5.69)? Can you also cancel the term due to the noise?
5. In subsection 5.1.2 we have shown how to restore the rank of a rank
deficient spectral matrix by smoothing the spectral matrices of subarrays. Show
that this can also be achieved by smoothing of the signal subspace eigenvectors
of the spectral matrix of the full array. This approach is suggested in [48] but
on a covariance matrix.
6. Let J be an exchange matrix which collects all odd and even elements of a
vector; for example, a 4x4 exchange matrix is

© 2001 by CRC Press LLC

  1,0,0,0
0,0,1,0
J = 0,1,0,0

0,0,0,1


Consider a ULA with M sensors and P uncorrelated wavefronts that are incident
on the array. Let v s be a matrix whose columns are the signal eigenvectors of
A 
the array spectral matrix. Show that Jv s =  G −1 where matrices A
AΓ 
M
( × P) and G (P × P) are as defined in (5.6c) and the Γ (P × P) is as
2
defined in (2.57). This result provides an alternate approach to the ESPRIT
algorithm described in §5.1. It can be used to extend the concept of subspace
rotation to multiple subarrays as described in [49].

References

1. R. O. Schmidt, Multiple emitter location and signal parameter estimation,

in Proc. of RADC Spectral Estimation Workshop, Griffiths AFB, New York,
pp. 243-258, Reprinted in IEEE Trans. AP-34, pp. 276-280, 1986.
2. G. Bienvenu and L. Kopp, Principe de la goniometri passive adaptive, Proc.
Colloque, CRESTI, Nice (France), pp. 106/1-10, 1979.
3. V. F. Pisarenko, The retrieval of harmonics from covariance function,
Geophy. J. R. Astron. Soc., vol. 33, pp. 347-366, 1973.
4. C. U. Padmini and P. S. Naidu, Circular array and estimation of direction of
arrival of a broadband source, Signal Proc., vol. 37, pp. 243-254, 1994.
5. A. B. Gershman and V. T. Ermolaev, Electronics Letters, vol. 28, pp. 1114-
1115, 1992.
6. S. U. Pillai, Array Signal Processing, Springer-Verlag, New York, 1989.
7. A. B. Gershman and V. T. Ermolaev, Optimal subarray size for spatial
smoothing, IEEE SP Letters, vol.2, pp. 28-30, 1995.
8. A. J. Barabell, Improving the resolution performance of eigenstructure based
direction finding algorithm, IEEE Proc ICASSP, pp. 336-339, 1983.
9. B. D. Rao, Performance analysis of root Music, IEEE Trans., ASSP-37, pp.
1939-1049, 1989.
10. R. Roy, A. Paulraj, and T. Kailath, Direction of arrival estimation by
subspace rotation method - ESPRIT, IEEE. Proc ICASSP, Tokyo, pp. 2495-
2498, 1986.
11. D. Slepian and H. D. Pollack, and H. J. Landau, Prolate spheroidal wave
function, Fourier analysis and uncertainty, Bell Systems Tech Jour., vol. 40,
pp. 43-84, 1961.

© 2001 by CRC Press LLC

 12. U. Reddy, B. Egart, and T. Kailath, Least squares type algorithm for
adaptive implementation of Pisarenko’s harmonic retrieval method, IEEE
Trans. ASSP-30, pp. 339-405, 1982.
13. P. Yang and M. Kaveh, Adaptive eigensubspace algorithm for direction or
frequency estimation and tracking, IEEE Trans, ASSP-36, pp. 241-251, 1988.
14. J. Karhunen, Adaptive algorithms for estimation of eigenvectors of
correlation type matrices, IEEE, Proc. ICASSP-84, pp. 14.6.1-4, 1984.
15. G. Stewart, Introduction to Matrix Computations, Academic Press, New
York, 1973.
16. H. Wang and M. Kaveh, Coherent signal subspace processing for the
detection and estimation of multiple wideband sources, IEEE, ASSP-30, pp.
823-831, 1985.
17. H. Wang and M Kaveh, Focussing matrices for coherent signal subspace
processing, IEEE ASSP-36, pp. 1272-1281, 1988.
18. J. Krolick and D. N. Swinger, Focussed wideband array processing via
spatial sampling, IEEE, ASSP-38, pp. 356-360, 1990.
19. K. M. Buckly and L. Griffiths, Broadband signal subspace spatial spectrum
(BASS-ALE) estimation, IEEE ASSP-36, pp. 953-964, 1988.
20. D. N. Swinger and J. Krolik, Source location bias in the coherently focused
high resolution broadband beamformation, IEEE ASSP-37, pp. 143-144, 1989.
21. J. Krolik and D. Swingler, Multiple broadband source location using
steered covariance matrices, IEEE Trans. ASSP-37, pp. 1481-1494, 1989.
22. S. Valae and P. Kabal, Wideband array processing using a two sided
correlation transformation, IEEE Trans SP-43, pp. 160-172, 1995.
23. K. M. Buckley and L. Griffiths, Eigenstructure based broadband source
localization, IEEE, ICASSP-86, pp. 1867-1872, 1986.
24. C. Usha Padmini, Wideband source localization, ME dissertation, ECE
Dept, IISc, Bangalore, India, 1991.
25. P. S. Naidu, Modern Spectrum Analysis of Time Series, CRC Press, Boca
Raton,FL.,1996.26. J. W. Mauchley, Significance test for sphericity of a
normal n-variate distribution, Ann. Math. Stat., vol.11. pp. 204-209, 1940.
27. W. S. Ligget, Jr., Passive sonar: fitting models to multiple time series, in
Signal Processing: Proc. NATO ASI Signal Processing, J. W. R. Griffiths and
P.L. Stocklin, Eds., Academic, New York, pp. 327-345, 1973.
28. D. B. Williams and D. H. Johnson, Using the sphericity test for source
detection with narrow band passive array, IEEE Trans., ASSP-38, pp. 2008-
2014, 1990.
29. M. Wax and T. Kailath, Detection of signals by information theoretic
criteria, IEEE Trans., ASSP-33, pp. 387-392, 1985.
30. Qi-Tu Zhang, Kon Max Wong, P. C. Yip, and J. P. Reilly, Statistical
analysis of the performance of information theoretic criteria in the detection of
the number of signals in array processing, IEEE Trans., ASSP-39, pp. 1557-
1567,1989.

© 2001 by CRC Press LLC

31. J. P. Henrioux and G. Jourdain, Use of large bandwidth-duration binary
phase shift keying signals for bearing measurements in active sonar
classification, JASA, vol. 97, pp. 1737-1746, 1995.
32. R. L. Pickholz, D. L. Schilling, and L. B. Milstein, Theory of spread
spectrum communications-A Tutorial, IEEE Trans COM-30, pp. 855-884,
1982.
33. R. L. Pickholz, L. B. Milstein, and D. L. Schilling, Spread spectrum for
mobile communications, IEEE Trans.Vehicular Technology, vol. 40, pp. 313-
322, 1991.
34. A. F. Naguib, A. Paulraj, and T. Kailath, Capacity improvement with base
station antenna array in cellular CDMA, IEEE Trans.,vol. 43, pp. 691-698,
1994.
35. W. A. Gardner, Introduction to Random Processes with Applicarion to
Signals and Systems, New York, MacMillan, 1985.
36. W. A Gardner, Exploitation of spectral redundancy in cyclostationarity
signals, IEEE., Signal Proc. Mag., pp. 14-36, April 1991.
37. S. V. Schell, R. A. Calabretta, W. A. Gardner and B. G. Agee, Cyclic
MUSIC algorithm for signal selective direction finding, Proc. ICASSP89, vol.
4, pp. 2278-2281, 1989.
38. G. Xu and T. Kailath, Direction of arrival estimation via exploitation of
cyclostationarity- A combination of temporal and spatial processing, IEEE
Trans., SP-40, pp. 1775-1786, 1992.
39.R. T. Compton, Pointing accuracy and dynamic range in steered beam
adaptive array, IEEE Trans., AES-16, pp. 280-287, 1980.
40. V. C. Soon, L. Tong, Y. F. Huang, and R. Liu, A subspace method for
estimating sensor gains and phases, IEEE Trans., SP-42, pp. 973-976, 1994.
41. J. J. Smith, Y. H. Leung, and A. Cantoni, The partitioned eigenvector
method for towed array shape estimation, IEEE Trans SP-44, pp. 2273-2283,
1996.
42. D. A. Gray, W. O. Wolfe and J. L. Riley, An eigenvector method for
estimating the positions of the elements of an array of receivers, Proc. ASSPA
89, Signal Processing theories, implementation, and applications, R. F.
Barrett (Ed.), Adelaide Australia, pp. 391-393, 1989.
43. S. B. Kesler, J. Kesler and G. Levita, Resolution of coherent target in the
near field of an array antenna, IEEE, Proc. ICASSP-86, pp. 1937-1940, 1986.
44. P. S. Naidu, A new method of acoustic source localization in shallow sea.
Proc. Institute of Acoustics, vol. 18, part 5, pp. 39-48, 1996.
45. H. F. Silverman, Some analysis of microphone arrays for speech data
acquisition, IEEE Trans., ASSP-35, pp. 1699-1712, 1987.
46. P. S. Naidu and T. Ganesan, Source localization in a partially known
shallow water channel, Journ Acoust Soc of Am, vol. 98, pp. 2554-2559,
1995.
47. P. S. Naidu and H. Uday Shankar, Broadband source localization in shallow
water, Signal Processing, vol.72, pp. 107-116, 1999.

© 2001 by CRC Press LLC

 48. S. S. Reddi and A.B. Gershman, An alternative approach to coherent source
location proble, IEEE Trans. SP-59, pp. 221-233, 1997.
49. A. L. Swindlehurst, B. Ottersten, R. Roy, and T. Kailath, Multiple
invariance ESPRIT, IEEE Trans. SP-40, pp. 867-881, 1992.
50. A. Papoulis, Probability, Random Variables and Stochastic Processes,
Third edition, McGraw-Hill Book Co., New York, 1991.

© 2001 by CRC Press LLC

 Chapter Six

Source Estimation

The temporal signals radiated by sources are separated on the principle of

nonoverlapping or partially overlapping temporal spectral characteristics of the
signals. A perfect separation is possible only when the signal spectra are
nonoverlapping. The spatio-temporal signals possess an additional degree of
variability, namely, the spatial spectrum. The differences in the spatial spectra
can be used, in addition to the differences in the temporal spectra, for the
purpose of signal separation. The signals coming from widely different
directions will have nonoverlapping spatial spectra and therefore they can be
separated using an array of sensors. However, when the signal sources are quite
close, perfect separation is not possible. There will be some cross talk. We
shall evaluate the Wiener filter, which was derived in chapter 3, from the point
of cross talk power in relation to the total signal power. Suppression of
unwanted signal or interference is achieved by placing a null or a collection of
nulls in the spatial frequency band occupied by the interference. The
effectiveness of nulls is enhanced when additional constraints are placed on the
filter; for example, the filter response is unity in the direction of useful signal.
This leads to the well known Capon’s filter which is also known as the
minimum variance filter. It is found that the Capon’s filter is quite effective
when the signal and the interference sources are highly directional. The filter
will automatically place a null wherever there is a strong interference. Finally,
when the direction of interference is known a priori, it is possible to devise a
filter which will place a sharp null at the spatial frequency corresponding to the
direction of arrival of the interference. The null can be steered to any desired
position, depending upon how the interference is changing its direction. Thus,
null steering can be effectively used to suppress slowly varying interference.

§6.1 Wiener Filters:

Previously, in chapter 3 we have derived Wiener filters for extraction of a
wavefront incident at M sensor ULA and UCA. It may be recalled that the
Wiener filter minimizes the mean square error between the filter output and the
desired signal. The signal and noise are assumed to be stationary stochastic
processes. The basic equation (3.38) is reproduced here for convenience,

S f (ω)H(ω) = S0 (ω) (3.38a)

where S f (ω) is the spectral matrix of the array output,

© 2001 by CRC Press LLC

  S0,0 (ω) S0,1 (ω) ... S0, M −1 (ω) 
 S (ω) S (ω) ... S 
 1,0 1,1 1, M −1 (ω) 
 ... 
S f (ω) =  
 ... 
 ... 
 
 SM −1, 0 (ω) SM −1, 1 (ω) ... SM −1, M −1 (ω)

and S0 (ω) is the cross-spectral vector between the desired output and the array
output. We had derived a specific result for a single wavefront with white
background noise. It was shown that the filter function is given by

H W (ω,θ 0 ) = Q a 0 (ω,θ 0 ) (6.1a)

S0 (ω)
σ 2η
where Q=
S (ω)
1+ M 0 2
ση

a 0 (ω , θ 0 ) is a steering vector and S0 (ω) is the spectrum of the incident

signal.
6.1.1 Filter Output: We shall now use this filter on the Fourier transformed
array output, F̂(ω)

F̂(ω ) = H WH (ω , θ 0 )F(ω )
= Qa 0H (ω , θ 0 )[a 0 (ω , θ 0 )Ξ0 (ω ) + Ν(ω )] (6.1b)

= Ξ0 (ω ) + Qa 0H (ω , θ 0 )Ν(ω )

Note that the signal component in the output remains undisturbed and the noise
1
variance is reduced by a factor of . The filter response to an incident plane
M
wave coming from a different direction with a direction vector a(ω , θ ) is
given by Qa 0 (ω , θ 0 )a(ω , θ ) . The response of the Wiener filter will be
H

compared later with that of the Capon filter in fig. 6.4.

Array Gain: Array gain is defined as a ratio of the output signal-to-noise ratio
(snr) to the input signal-to-noise ratio. For the purpose of illustration let us

© 2001 by CRC Press LLC

consider a single source in presence of white noise. The spectral matrix of the
array output (4.12b) which we shall write in a slightly different form showing
the signal power explicitly

S f (ω) = σ 2s a 0 (ω)S̃0 (ω)a 0 H (ω) + σ 2ηI

where S̃0 (ω) is a normalized spectrum. The input signal-to-noise ratio is

σ 2s
evidently equal to 2 . We assume that the Wiener filter (6.1a) is used to
ση
extract the signal whose DOA is known a priori. The output power is given by

H WH (ω,θ 0 )S f (ω)H W (ω,θ 0 ) =

2 2 2
σ 2s Q2 a 0 (ω) S̃0 (ω) a 0 (ω) + σ 2ηQ2 a 0 (ω)

σ 2s 2
It is easy to show the output snr as a 0 (ω) . By definition, the array gain
ση2

2
is equal to a 0 (ω) which, for ideal sensor array with omnidirectional and unit
response sensors, is equal to M (the number of sensors).
6.1.2 Two Source Case: We shall now consider two wavefronts incident on an
array (ULA) in presence of white background noise. The sources are assumed to
be uncorrelated between themselves as well as with the background noise. The
aim is to extract a first source signal while we suppress the signal from the
second source as well as the background noise. Corresponding to this model the
spectral matrix and the cross-spectral vector in the Wiener filter equation take
the following form,

a 0 (ω,θ 0 )a 0H (ω,θ 0 )S0 (ω) 

S(ω) =  2 
+a1 (ω,θ1 )a1 (ω,θ1 )S1 (ω) + σ ηI 
H
(6.2)

S0 (ω) = a 0 (ω,θ 0 )S0 (ω)

where

ω ω ω
−j d sin θ 0 −j 2 d sin θ 0 −j ( M −1)d sin θ 0
a 0 (ω , θ 0 ) = col[1, e c
,e c
, ... e c
]

and

© 2001 by CRC Press LLC

 ω ω ω
−j d sin θ 1 −j 2 d sin θ 1 −j ( M −1)d sin θ 1
a1 (ω , θ1 ) = col[1, e c
,e c
, ... e c
]

where θ 0 and θ1 are DOA angles of the first and the second wavefront,
respectively. The filter response for extraction of the first source signal is given
by

−1
a 0 (ω,θ 0 )a 0H (ω,θ 0 ) + 
 
H W (ω,θ 0 ) =  S1 (ω) σ 2η  a 0 (ω,θ 0 ) (6.3)
a1 (ω,θ1 )a1 (ω,θ1 ) S (ω) + S (ω) I 
H

 0 0 

The inverse of the quantity inside the square brackets in (6.3) has been derived
in chapter 4 (p. 234). Using that result we obtain

V −1 − S1 (ω) × 
 S0 (ω) 
 
H W (ω,θ 0 ) =  V −1a1 (ω,θ1 )a1H (ω,θ1 ))V −1  a 0 (ω,θ 0 ) (6.4a)
 S (ω) H 
1 + 1 a1 (ω,θ1 )V −1a1 (ω) 
 S0 (ω) 

where

S0 (ω )
V −1 =
ση 2 [
I − Qa 0 (ω , θ 0 )a 0H (ω , θ 0 ) ] (6.4b)

where
S 0 (ω )
σ η2
Q=
S (ω )
1+ M 0 2
ση

Though the filter is tuned to receive the first wavefront some amount of energy
from the second wavefront will leak into the filter output. This is known as
cross talk. Ideally, the cross talk should be zero. Let a1 (ω)Ξ1 (ω) be the
Fourier transform of the signal emitted by the second source. The output of the
filter tuned to the first source is given by

© 2001 by CRC Press LLC

 H WH (ω,θ 0 )a 0 (ω,θ 0 )Ξ0 (ω) + 
output =  H  (6.5a)
H W (ω,θ 0 )a1 (ω,θ1 )Ξ1 (ω) + H WH (ω,θ 0 )Ν(ω)

Further, to evaluate different terms in (6.5a) we need the following results:

S0 (ω ) a 0H (ω , θ 0 )a1 (ω , θ1 )
a 0H (ω , θ 0 )V −1a1 (ω , θ1 ) = (6.5b)
σ η2 S (ω )
1+ M 0 2
ση
and

a1H (ω,θ1 )V −1a1 (ω,θ1 ) =

S0 (ω)
ση 2 [
M − Q a1H (ω,θ1 )a 0 (ω,θ 0 )
2
] (6.5c)

Using (6.5b) and (6.5c) in (6.5a) we obtain

H WH (ω,θ 0 )a 0 (ω,θ 0 )Ξ0 (ω) =

2
S (ω) Q2 a1H (ω,θ1 )a 0 (ω,θ 0 ) Ξ0 (ω)
MQΞ0 (ω) − 1 (6.6a)
S
S0 (ω) 1 + 1 (ω)
σ 2η [
M − Q a1 (ω,θ1 )a 0 (ω,θ 0 )
H 2
]
H WH (ω,θ 0 )a1 (ω,θ1 )Ξ1 (ω)
(6.6b)
a 0H (ω,θ 0 )a1 (ω,θ1 )Ξ1 (ω)
=Q
S (ω)
[
1 + 1 2 M − Q a1H (ω,θ1 )a 0 (ω,θ 0 )
ση
2
]
We shall assume that the array signal-to-noise ratio (asnr) is much greater than
MS0 (ω)
one, that is, >>1. Using this approximation in (6.6) the array
σ 2η
output power may be approximated as

Output power ≈ S0 (ω) +

© 2001 by CRC Press LLC

  a1H (ω,θ1 )a 0 (ω,θ 0 )
4

 
2 S (ω) 
 S1 (ω)  ×
0
M4 
2
 S0 (ω)   S (ω)  a (ω,θ1 )a 0 (ω,θ 0 )  
H 2

 1 + 1 2 M 1 − 1  
  ση  M2   
   
 H
a1 (ω,θ1 )a 0 (ω,θ 0 )
2 
 S (ω) 
−2 M2
1

  H 2
 
 1 + S1 (ω) M 1 − a1 (ω,θ1 )a 0 (ω,θ 0 )  
 σ 2η  M 2
 
 
 H
a 0 (ω,θ 0 )a1 (ω,θ1 )
2 
 S (ω) 
+ M2
1

   H 2

2
 (6.7)
 1 + S1 (ω) M 1 − a1 (ω,θ1 )a 0 (ω,θ 0 )   
  ση 2
 M 2
  
  
+H H (ω,θ )H (ω,θ )σ 2 (ω) 
 W 0 W 0 η 

The first term is the desired signal power. The remaining terms represent the
interference. Of these three, the magnitude of the first is much lower than that
of the second and third terms. The magnitude of the first is proportional to
4
a1H (ω,θ1 )a 0 (ω,θ 0 )
while the magnitude of the second and third terms is
M4
2
a1H (ω,θ1 )a 0 (ω,θ 0 )
proportional to . Hence, we drop the first term from
M2
the interference expression. The cross talk, defined as a ratio of the power leaked
from the second source and the actual power in the second source, is given by

2
a 0H (ω,θ 0 )a1 (ω,θ1 )
cross talk ≈ ×
M2

© 2001 by CRC Press LLC

   S (ω) a H
(ω,θ )a (ω,θ )
2
 2

  1 2 M(1 − 0 0 1 1
) 
  σ η M2 

1 − 2 (6.8)
  S (ω) a H
(ω,θ )a (ω,θ )
2
 
 1 + 1 2 M(1 − 0 0 1 1
) 
  ση M2  

2
a1H (ω,θ1 )a 0 (ω,θ 0 )
Note that represents the square of cosine of the angle
M2
between two direction vectors a 0 (ω,θ 0 ) and a1 (ω,θ1 ) . When
2
a1H (ω,θ1 )a 0 (ω,θ 0 )
a 0 (ω,θ 0 ) = a1 (ω,θ1 ) , =1 the cross talk =1, as
M2
2
a 0H (ω,θ 0 )a1 (ω,θ1 )
expected. But when a 0 (ω,θ 0 )⊥a1 (ω,θ1 )) , =0 and
M2
cross talk is zero. Aside from these two extreme situations the cross talk may
MS1 (ω)
be reduced to zero if >>1, that is, the array signal-to-noise ratio
σ 2η
(asnr) for the second source is very large. The cross-talk is shown in fig. 6.1.
6.1.3 Linear Least Squares Estimate (LLSE): As previously stated the signal
model is

ξ 0 (t) 
 
 ξ1 (t) 
f(t) = [a 0 ,a1 ,...a P −1 ] + η(t)
  (6.9)
 
ξ P −1 (t)
= Aξ(t) + η(t), t = 0,1,..., N

where a 0 ,a1 ,...a P −1 are steering vectors of P sources. It is assumed that A is

known and ξ(t) is unknown but deterministic. The least squares estimate is
obtained by minimizing with respect to ξ(t)

2
f(t) − Aξ(t) = min (6.10)

© 2001 by CRC Press LLC

 1

0.9

0.8

0.7
Cross talk

0.6

0.5

0.4

0.3

0.2
0 1 2 3 4 5 6
Angle in degrees

Figure 6.1: Wiener Filter: Cross talk as a function of angular distance between two
sources. The first source is on broadside (DOA=0o ). Array signal-to-noise ratio
(asnr)=10 and eight sensor ULA is assumed.

Differentiating (6.10) with respect to ξ(t) and setting the derivative to zero we
obtain

[] A f(t)
−1
ξ̂(t) = A H A H

(6.11)
= ξ(t) + [A A] A
−1
H H
η(t)

We observe that the signal term is extracted without any distortion but the
[A A]
−1
noise term, given by
H
A H η(t) , behaves differently; for example, the
noise becomes correlated. The output noise covariance matrix is given by

[ ]
−1
Cη̂ = A H A σ η2 (6.12)

H
When A A is singular or close to being singular, that is, with a large
eigenvalue spread, the noise in the output may get amplified. Consider a case of

© 2001 by CRC Press LLC

 800

80

8
Amplitude

.8

.08

.008
0 1 2 3 4 5
Angle in degrees

Figure 6.2: The noise amplification factor as a function of DOA of source #2. Source
#1 is held fixed at 0o and source #2 is moved. Array size M=8.

two sources with direction vectors a 0 and a1 and corresponding to directions

of arrival θ 0 and θ1 , respectively. It is easy to show that

 a 0H a1 
2

[ H
]
det A A = M 1 −
M 
2
(6.13)


Using (6.13) in (6.12) we obtain the variance of the noise in all array outputs

1
σ η̂2 = σ η2 (6.14)
 a 0H a1 
2

M 1 −
 M 

We have plotted in fig. 6.2 the noise amlpification factor, the factor
multiplying σ η , as a function of the angular separation between two sources.
2

© 2001 by CRC Press LLC

 From the plot we note that the noise is amplified only when the angular
separation is a fraction of a degree.
In the stochastic model, the signal waveform is (stationary) stochastic
and the LLSE turns out to be the same as the Wiener filter, which is easily
derived as (Exercise #6.1),

H W (ω) = S−1
f (ω)AS 0 (ω)

where A = [a 0 ,a1 ,...,a P −1 ] . In the Wiener filter we require spectral matrix

and cross-spectral vectors; both of these are obtained by statistical averaging.
Such a filter will naturally be applicable to an ensemble of time series having
the same second order statistical properties. In the deterministic approach the
LLSE filter is adapted to a particular data set. It is claimed in [1] that the
stochastic LLSE results in a lower mean square error than the deterministic
LLSE.
Circular Array: The result on cross talk (6.8) is of a general nature, valid for
any array geometry. The direction vector needs to be appropriately defined. We
now consider a circular array (UCA) for which the direction vector is given by
(2.51) where we let θ = 90 so that the source is in the same plane as the
o

circular array,

 − j ωa cos(ϕ) − j ωac cos( 2Mπ −ϕ) −j

ωa
cos(
2 π( M −1)
−ϕ) 
a(ϕ) = col e c ,e ,...,e c M
 (6.15)
 

The interest is to find out how a circular array fares in comparison with a linear
array with respect to cross talk capability. For a given number of sensors (say,
(M − 1)λ
M) the maximum aperture of an ULA is fixed at but the aperture of
2
a circular array can be very large, at least in principle. Since the array aperture
is the main factor deciding the cross talk it is expected that the circular array
ought to perform better in terms of lower cross talk. Using the same number of
sensors, cross talk may be reduced when arranged over a large circle. This is
shown in fig. 6.3 for an eight sensor UCA. The cross talk has been reduced
considerably when the array radius is increased from four to thirty-two
wavelengths. A linear array of eight sensors will have aperture of 3.5λ. An
UCA having an aperture of 3.5λ was found to show much higher cross talk
than that of the ULA shown in fig. 6.1. Thus, performance of an UCA is
significantly better only when the radius is increased considerably.
6.1.4 Effects of Errors in DOA : Waveform estimation requires a knowledge of
the direction of arrival of wavefronts. But the DOA estimates are subject to

© 2001 by CRC Press LLC

 1

0.8
4λ

0.6
Cross talk

0.4 32 λ

0.2

0
0 1 2 3 4 5 6
Angle in degrees

Figure 6.3: Cross talk as a function of angular separation of two sources. Eight
sensors are uniformly spaced over a circle of radius 4λ (solid curve) and 32λ (dashed
curve).

errors largely on account of finite data length used in their estimate. In this
section we investigate the effects of errors in DOA estimates on interference
from signals of no interest. A comprehensive analysis of the effects of model
errors including the errors in DOA may be found in [2]. Let θ̂ p = θ p + ∆θ p ,
p = 0,1,..., P − 1 be the estimated DOAs, where θ p is correct DOA and
∆θ p is estimation error. We shall rewrite (6.11) showing explicitly the
dependence on the estimated DOAs

[ ]
−1
ξ̂(t) = A H (θ̂)A(θ̂) A H (θ̂)f(t) (6.16)

[ ]
Let us express A(θ̂) = B c where B is a matrix whose columns are the
direction vectors of all interfering sources (signals of no interest) and c is a
vector representing the direction vector of the signal of interest (soi).

[A ]
−1
H
(θ̂)A(θ̂) may be simplified [2] as

© 2001 by CRC Press LLC

 [A ] = [[B c] [B c]]
−1 H −1
H
(θ̂)A(θ̂)

−1
B H B BH c
= H 
c B cHc 
−1 −1
 H B H cc H B   H B H cc H B  B H c 
 B B − 2  − B B − 2  c2
=  c   c  
 H  (6.17)
−[c c − c Pc] c B1 [c c − c Pc]
H −1 H H H −1


1 (B H B)−1 c H (I − P)c + B1cc H B1H − B1c 

= H  
c (I − P)c −c H B1H 1 

−1
where P = B(B B) B
H H
and B1 = (B H B)−1 B H . Using (6.17) in (6.16)
and simplifying we obtain

[A ]
−1
H
(θ̂)A(θ̂) A H (θ̂)

1 (B H B)−1 c H (I − P)c + B1cc H B1H − B1c  B H 

= H     (6.18)
c (I − P)c −c H B1H 1  c H 

1 B1[c H (I − P)cI − cc H (I − P)]

= H  
c (I − P)c c H (I − P) 

Using (6.18) in (6.16) an estimate of the signal of interest (soi) is obtained

c H (I − P)
ξ̂ soi (t) = f(t) = d H (θ̂)f(t) (6.19)
c (I − P)c
H

We now expand d H (θ̂) in a Taylor’s series and retain the first derivative term
only, d(θ̂) ≈ d(θ) + d1 (θ)∆θ , where d1 (θ) is the first derivative of
d(θ) with respect to θ . The estimated signal of interest reduces to

© 2001 by CRC Press LLC

 ξ̂ soi (t) = d H (θ)Aξ(t) + ∆θ H d1H (θ)Aξ(t) + d(θ̂)η(t)
(6.20)
= ξ soi (t) + ∆θ H d1H (θ)Aξ(t) + d(θ̂)η(t)

The interference term in (6.20) is ∆θ

H
d1H (θ)Aξ(t) which we shall rewrite
showing the contribution of each DOA

P −1

[
∆ξ soi (t) = ∑ β H ξ(t) p ∆θ p ] (6.21)
p= 0

where [ ]
∆ξ soi (t) stands for the error in ξ̂ soi (t) , β = A H d1 (θ) and [.] p
stands for the pth element of a column vector.
The errors in the calibration of an array can cause serious errors in the
waveform estimation [3]. It is, therefore, imperative to carefully calibrate the
array response, particularly when subspace methods are used for localization or
waveform estimation. One simple solution to this problem of high sensitivity
to calibration errors is to reduce the degrees of freedom by using a low rank
approximation of the spectral matrix [4, 5]. In this approximation a spectral
matrix is simply approximated by its signal subspace, that is,

S f (ω) = v s α s v sH + v ηα ηv ηH
(6.22)
≈ v s α s v sH

where α s = diag{α 0 ,α1 ,...,α P −1} and α η = diag{α P ,α P +1 ,...,α M −1} .

§6.2 Minimum Variance (Capon Method):

In §4.3 we have devised a filter (in the frequency domain) which minimizes the
output power while maintaining unit response within an angular sector
(extended source) whose axis is directed in the desired direction. This leads to
the following equation,

1
H = (S−1
f Γ)H (6.23a)
λ
where
∆θ
θ+
2
1
Γ= ∫ a(ω,φ)a
H
(ω,φ)dφ (6.23b)
∆θ ∆θ
θ−
2

© 2001 by CRC Press LLC

 a(ω,φ) is the steering vector and λ = H H S f H is output power. To further
minimize the output power we must select a filter vector as the eigenvector of
(S−1
f Γ) corresponding to the largest eigenvalue; in which case the output

power will be equal to the inverse of the largest eigenvalue of (S−1

f Γ) . A
closed form solution of (6.23) can be obtained for ∆θ = 0 (point source).
Equation (6.23) simplifies to

H
H
= (S−1 H
f a(ω,θ)a (ω,θ)H (6.24)
H SfH

Multiply both sides of (6.24) by a H (ω,θ) and simplify the resulting

expression,

1
HHS f H =
a (ω,θ)S−1
H
f a(ω,θ) (6.25)

By direct substitution in (6.24) it may be verified that the solution is given by

S−1a(ω,θ)
H cap = H f −1 (6.26)
a (ω,θ)S f a(ω,θ)
1
and the output power is equal to H The optimum filter
a (ω,θ)S−1
f a(ω,θ)
given in (6.26) is also known as the Applebaum filter [6].
6.2.1 Extended Source: For ULA the Γ matrix takes the following form

∆θ
sin(k )
ωd
[Γ ]m,n = ∑ ε k Jk ((m − n) )cos(k θ) ∆θ + 2
k = 0,2,4... c k
2
∆θ
sin(k )
ωd
j ∑ 2Jk ((m − n) )sin(k θ) 2
k =1,3,5... c ∆θ
k
2
where ε 0 = 1 and ε n = 2 for all k. To evaluate the integral in (6.23b) we
have used the result derived in (1.53b). For UCA the Γ matrix takes the
following form:

© 2001 by CRC Press LLC

 Γ = W Θ WH
( M × M ) ( M ×α ) (α ×α ) (α × M ) α→∞

where

m = 0,1,... M − 1
2 πkm
j
{W}km = e M
 
k = 0, ±1,...∞ 
π
ωa ωa − j ( + ϕ 0 )( k −l ) ∆θ
[Θ]k,l = Jk ( )Jl ( )e 2 sin c((k − l) )
c c 2

valid for all values of ∆θ . We have used a series expansion of the steering
vector for a circular array (2.51).
6.2.2 Single Point Source Case: Let there be a single source emitting a plane
wave with a direction vector a 0 (ω,θ 0 ) in presence of white noise. The
spectral matrix is given by S f = a 0 (ω,θ 0 )a 0H (ω,θ 0 )Sξ (ω) + σ 2ηI , and its
inverse is given by

S−1
f =
1
ση
2 [
I − Qa 0 (ω , θ 0 )a 0H (ω , θ 0 ) ] (6.27)

where
Sξ (ω)
σ 2η
Q=
S (ω)
1+ M ξ 2
ση

S ξ (ω )
For large array signal-to-noise ratio, >> 1 , we may approximate
σ η2
1
Q≈ . Using (6.27) in (6.26) we obtain
M

 

H cap
1
= 
0 [
 a(ω , θ ) − Q a H (ω , θ )a(ω , θ ) a (ω , θ ) 
0 0 0
 ] (6.28)
M 
2
a (ω , θ )a 0 (ωθ 0 )
H

 1− Q 
 M 

© 2001 by CRC Press LLC

 Let us obtain the filter output for the signal model of a single source in the
presence of white noise

F̂(ω ) = H cap
H
[a 0 (ω , θ 0 )Ξ0 (ω ) + Ν(ω )] (6.29a)

The signal term turns out to be

H
H cap a 0 (ω,θ 0 )Ξ0 (ω)
 
 
H
a (ω,θ)a 0 (ω,θ 0 )  1 − QM 
= 2 Ξ 0 (ω) (6.29b)
M  a (ω,θ)a 0 (ωθ 0 ) 
H

1 − QM 
 M2 
= Ξ0 (ω) when a H (ω,θ) = a 0 (ω,θ 0 )

and the noise term

H
H cap Ν(ω)
 
1
= 
[0 0 ] 0 0 Ν(ω)
 a(ω,θ) − Q a H (ω,θ )a(ω,θ) a (ω,θ ) 
(6.29c)
M 
2
a H (ω,θ)a 0 (ωθ 0 )
 1 − QM 
 M 2

a 0H (ω,θ 0 )
= Ν(ω) when a H (ω,θ) = a 0 (ω,θ 0 )
M
The variance of the noise in the filter output (6.29c), when
σ N2
a H (ω , θ ) = a 0 (ω , θ 0 ) , turns out to be . The response of the filter was
M
computed from (6.29a) for different steering directions. The DOA of the
incident wavefront was assumed to be at 0o. An ULA with eight sensors was
considered. The response of the filter depends upon the signal-to-noise ratio
(snr). For snr=10 (or asnr=80) the response becomes extremely sharp as
demonstrated in fig. 6.4. For comparison we have also plotted a Wiener filter
response (dashed curve) where we have assumed the signal-to-noise ratio equal
to ten while other parameters remain the same as for the Capon filter.

© 2001 by CRC Press LLC

 1

Wiener filter
0.8

0.6
Response

Capon filter snr=1

0.4

snr=10
0.2

0
0 0.5 1 1.5 2 2.5
Angle in deg.
Figure 6.4: Response of Capon filter as a function of the angle of the steering
vector. An ULA of 8 elements was assumed. Equivalent Wiener filter response is also
shown by the dashed curve.

The response of the Capon filter is much superior to that of an

equivalent Wiener filter. The shape of the Capon filter is strongly dependent on
signal-to-noise ratio; on the other hand the dependence of the Wiener filter on
the signal-to-noise ratio is merely to the extent of changing the scale factor.
Even at a very low signal-to-noise ratio the Capon filter seems to outperform
the Wiener filter.
6.2.3 Two Source Case: As in the case of the Wiener filter, we shall now
consider the Capon filter specifically for two uncorrelated wavefronts in the
presence of white noise. The spectral matrix is given by (6.2) and its inverse is
computed in the same manner as shown there. Using the expression for inverse
of the spectral matrix in (6.26) we obtain an expression for the Capon filter,
num
H Cap (ω, θ 0 ) = where
den

 
 −1 S (ω) −1 H
V a1 (ω,θ1 )a1 (ω,θ1 )V −1 
num = V − 1  a 0 (ω,θ 0 )
 S0 (ω) 1 + S1 (ω) a H (ω,θ )V −1a (ω,θ ) 
1 1 1 1
 S0 (ω) 
(6.30a)

© 2001 by CRC Press LLC

 den = a 0H (ω,θ 0 ) ⋅ num and V is defined in (6.4b). The denominator may
be further reduced to

S1 (ω)  a1H (ω,θ1 )a 0 (ω,θ 0 ) 

2

M(1 + 2 1 − )
ση  M2 
den=
S0 (ω)
(1 − QM) 
σ 2η S1 (ω)  a1H (ω,θ1 )a 0 (ω,θ 0 ) 
2

(1 + 2 1 − QM )
ση  M2 

(6.30b)

The filter output may be shown to be

Ξ0 (ω) 
output =  H  (6.31)
+H Cap (ω,θ 0 )a1 (ω,θ1 )Ξ1 (ω) + H Cap (ω,θ 0 )Ν(ω)
H

Notice that the signal term is equal to the actual signal. But this was not so in
the case of the Wiener filter where we had to assume a large array snr in order to
arrive at this result.
The contribution of the second source is represented by the second
term in (6.31). The cross talk is then given by

Cross talk = a1H (ω,θ1 )H Cap (ω,θ 0 )H Cap

H
(ω,θ 0 )a1 (ω,θ1 )

2
a 0H (ω,θ1 )a1 (ω,θ 0 )
M
= 2 (6.32)
 S (ω)  a1H (ω,θ1 )a 0 (ω,θ 0 )  
2

1 + 2 1 − 
1

 σ  M 2
 
η
 
The cross talk is plotted in fig. 6.5 for an ULA and in fig. 6.6 for an UCA.
Compare these two figures with figures 6.1 and 6.3 where we have plotted the
cross talk for the Wiener filter. Clearly, the Capon filter performs better for
both types of array geometries. An expression for snr which agrees with (6.32)
is derived in [7].
Finally, we shall evaluate the noise term, that is, the leftover noise in
the array output,

© 2001 by CRC Press LLC

 1

0.8

0.6
Cross talk

0.4

0.2

0
0 1 2 3 4 5 6
Angle in deg
Figure 6.5: Capon Filter: Cross talk as a function of angular distance between two
sources. The first source is on broadside (DOA=0o ). Array signal-to-noise ratio
(asnr)=10. Continuous curve: 8 sensor ULA and Dashed curve: 16 sensor ULA.

H
H Cap (ω,θ 0 )Ν(ω)
 
 S (ω) V −1
a (ω,θ )a H
(ω,θ )V −1 
a 0H (ω,θ 0 ) V −1 − 1 1 1 1 1
Ν
 S0 (ω) 1 + S1 (ω) a H (ω,θ )V −1a (ω,θ ) 
1 1 1 1
 S0 (ω) 
= (6.33)
den
a 0H (ω,θ 0 )V −1Ν − ψa1H (ω,θ1 )V −1Ν
=
den
where the denominator term, den, is given in (6.30b) and

S1 (ω) H
a 0 (ω,θ 0 )V −1a1 (ω,θ1 )
S0 (ω)
ψ=
S (ω) H
1+ 1 a1 (ω,θ1 )V −1a1 (ω,θ1 )
S0 (ω)

© 2001 by CRC Press LLC

 1

0.8

4λ

0.6
Cross talk

0.4

32 λ
0.2

0
0 1 2 3 4 5 6
Angle in deg

Figure 6.6: Capon Filter: Cross talk as a function of angular separation of two
sources. Eight sensors are uniformly spaced over a circle of radius 4λ (solid curve)
and 32λ (dashed curve).

The variance of the leftover noise may be computed from (6.33). We obtain
after simplification

H
Var H Cap {
(ω,θ 0 )Ν(ω) }
σ 2η  a1H (ω,θ1 )a 0 (ω,θ 0 ) 
= 1 − ψ
M  M 

 S1 (ω) a1H (ω,θ1 )a 0 (ω,θ 0 )
2
` (6.34a)
2 
1 + M(1 − ) 
σ ση 2
M
= η 
M a (ω,θ1 )a 0 (ω,θ 0 ) 
H 2
S (ω)
1 + 1 2 M(1 − QM 1 )
 ση M 

When the array snr is large, then QM ≈ 1 and (6.34) reduces to a simple form

© 2001 by CRC Press LLC

 σ 2η
{ H
Var H Cap (ω,θ 0 )Ν(ω) ≈ } M
(6.34b)

§6.3 Adaptive Beamformation:

In adaptive beamformation the array processor is so designed that it receives a
signal coming from a desired direction and it automatically suppresses signals
(that is, interference) coming from all other directions. This is achieved by
means of a filter which adapts itself to the incoming signal and interference.
Suppression of interference is achieved through predictive cancellation; an
example of this approach was described in chapter 2 in connection with noise
cancellation. The weight coefficients are upgraded through an algorithm, such
as LMS or one of its kind. There are excellent texts on the topic of adaptive
signal processing which includes adaptive beamformation [8,9]. We do not
intend to cover this topic in any detail. The aim here is to briefly discuss a few
selected topics, namely, null steering, adaptive interference canceller and
adaptive Capon filter.
6.3.1 Null Steering: A basic step in beamformation is the weighted and delayed
sum of the array outputs

M −1
f̂ (t) = ∑a
m=0
m f m (t − m τ) (6.35a)

where τ is delay per sensor and am is a real weight coefficient. In temporal

frequency domain (6.35a) may be expressed as

P −1
F̂(ω) = ∑ (a
m=0
m e − jmτω )F(md,ω)
π
(6.35b)
1
= ∫
2π − π
A(u)F * (u,ω)du

where

P −1
A(u) = ∑ ã
m=0
m e − jmu

ãm = am e − jmτω

© 2001 by CRC Press LLC

 jy
z-plane

θ0
x

Figure 6.7: Position of a null and two poles in its immediate neighborhood but
slightly inside the unit circle.

P −1
F(u,ω) = ∑ F(md,ω)e
m=0
− jmu

Through a process of selecting the complex weight coefficients we can assign

the desired property to the response function of the weight function; for
example, A(u) may be required to have a deep null in some desired direction
and unit response in all directions. Let us express A(u) as a z-transform of
P −1
ãm ,, A(z) = ∑ ã
m=0
m z −1 where z = e ju . We can now design a notch filter in
the z-plane. A simple approach is to place a zero on the unit circle and two
jπ sin θ 0
poles, one on either side of the zero. Let the zero be at z0 = e and the
∆θ ∆θ
jπ sin(θ 0 + ) jπ sin(θ 0 − )
poles at z+ = re and z− = re2 2
(see fig. 6.7) where
0. 9 ≤ r ≤ 0. 99. We have assumed that d = 0. 5λ .

(z − z0 )
A(z) = (6.36)
(z − z+ )(z − z− )

The response of the notch filter given by (6.36) is shown in fig. 6.8. See
exercise 3 for another type of null steering filter independent of frequency.
6.3.2 Beam Steered Adaptive Array: In Capon filter (6.26) let us replace the
inverse of the spectral matrix by its eigenvalue eigenvector representation,

© 2001 by CRC Press LLC

 1.2

1.0

(a)
0.8
Magnitude

0.6

0.4

0.2

0.0
-90 -60 -30 0 30 60 90
Angle (deg.)

(b)
1
Phase (radians)

-1

-2

-3
-90 -60 -30 0 30 60 90
Angle (deg.)

Figure 6.8: Wavenumber response of the notch filter (a) amplitude response and (b)
phase response.

© 2001 by CRC Press LLC

 S−1
f a(ω,θ)
H cap =
a (ω,θ)S−1
H
f a(ω,θ)
(6.37a)
=
{v λ v + v λ v }a(ω,θ)
s
−1 H
s s η
−1 H
η η

aH (ω,θ){v λ v + v λ v }a(ω,θ)
s
−1 H
s s η
−1 H
η η

We assume that the look direction vector a(ω,θ) lies in the signal subspace;
hence a(ω,θ)⊥v η . Hence (6.37a) simplifies to

H cap =
{v λ v }a(ω,θ)
s
−1 H
s s
(6.37b)
aH (ω,θ){v λ v }a(ω,θ)
s
−1 H
s s

The Capon filter has an interesting property, that is, when the look direction
coincides with one of the signal directions, for example, a(ω,θ) = a(ω,θ 0 )
it is approximately orthogonal to all direction vectors,
a(ω,θ m ), m = 1,2,..., P − 1. To show this, consider

a H (ω,θ 0 ){v s λ s−1v sH }a(ω,θ m )

H
H cap a(ω,θ m ) = , m≠0 (6.38a)
a H (ω,θ 0 ){v s λ s−1v sH }a(ω,θ 0 )

In (6.38a) the inverse of the signal eigenvalues may be approximated, for large
σ 2η
snr or large array size, by λ−1m ≈ α −1
m − . Therefore, (6.38a) may be
α 2m
expressed as

 a H (ω,θ 0 ){v s α s−1v sH }a(ω,θ m ) 

 H 
 a (ω,θ 0 ){v s λ s v s }a(ω,θ 0 )
−1 H

 
H cap a(ω,θ m ) ≈ 
H
P −1
ση2
, m ≠ 0 (6.38b)
 a (ω,θ 0 ) ∑ 2 v m v m a(ω,θ m ) 
H H

− m=0 αm 
 a (ω,θ 0 ){v s λ s−1v sH }a(ω,θ 0 ) 
H

We shall show that the first term inside the brackets is indeed zero. Let us
consider the signal term alone from (4.12b) and compute the pseudoinverse
(denoted by #) on both sides of the equation

© 2001 by CRC Press LLC

 H
S #f = A # S0−1A # (6.39)

On premultiplying by A H and postmultiplying by A both sides of (6.39) we

obtain

H
A H S #f A = A H A # S0−1A # A = S0−1 (6.40)

When the sources are uncorrelated, S0 as well as S0−1 will be diagonal

H #
matrices. It, therefore, follows that all cross product terms in A S f A must be
equal to zero. The numerator of the first term is one such cross product term.
The second term in (6.38b) will be small when the asnr (array snr) is high
Thus, we obtain the following approximate result

H
H cap a(ω,θ m ) ≈ 0 m = 1,2,..., P − 1 (6.41)

A modified Capon filter is defined as

Ss# (ω)a(ω,θ)
H̃ cap = (6.42)
a H (ω,θ)Ss# (ω)a(ω,θ)
#
where Ss (ω) is the pseudoinverse of the signal-only spectral matrix. It is
shown in [10] that the modified Capon filter is robust against look direction
errors. It was shown in chapter 5 (page 274) that the generalized eigenvector
corresponding to a given direction of arrival (DOA) is orthogonal to all other
direction vectors.
6.3.3 Adaptive Capon filter: In chapter 3 we have derived two different filters to
extract a wavefront coming from a specified direction. It may be recalled that
the Wiener filter minimizes the mean square error between the filter output and
the desired signal and the Capon filter minimizes the output power while
maintaining unit response in the desired direction. The Wiener filter is given by
H W (ω) = S−1 f (ω)S 0 (ω) and the Capon filter is given by
S−1
f (ω)a(ω,θ)
H Cap (ω) = where a(ω,θ) is the direction vector
a (ω,θ)S−1
H
f (ω)a(ω,θ)
in the desired direction. Note that S0 (ω) = a(ω,θ)S0 (ω) where S0 (ω) is
power from the desired direction. In the Wiener filter this power is assumed to
be a constant but in the Capon filter the power is estimated

© 2001 by CRC Press LLC

 1
(= H
) and it is a function of the desired direction. This
−1
a (ω,θ)S (ω)a(ω,θ) f
subtle difference is probably responsible for the improved performance of the
Capon filter.
To implement the Wiener or Capon filter we need the inverse of the
−1
spectral matrix S f (ω) . In practical terms with the arrival of new data in the
form of new time or frequency snapshot we should be able to improve upon the
available estimate of the spectral matrix and its inverse. In chapter 4 we have
shown how to recursively estimate these quantities. We shall rewrite equations
(4.62a and 4.62c) in a more general fashion

Ŝ Nf +1 (ω) = µŜ Nf (ω) + (1 − µ)F N +1 (ω)F HN +1 (ω) (6.43a)

[Ŝ (ω)]
−1
N +1
f =

[Ŝ (ω)] (1 − µ)zz H

1 −1 (6.43b)
N
−
[ ]
−1
µ µ  (1 − µ)F HN +1 (ω) Ŝ Nf (ω) F N +1 (ω) + µ
f

[ ]
−1
where z = Ŝ Nf (ω) F N +1 (ω) and µ ∈(0,1) , which is a free parameter to
be chosen depending upon how fast the estimated spectral matrix changes from
snapshot to snapshot. For stationary process where the change is small, µ ≈ 1.
Using (6.43a) and (6.43b) we can recursively compute the Wiener and Capon
filters. We will do this for the Capon filter. Multiply both sides of (6.43b) by
a(ω,θ)
. Noting the definition of the Capon filter and
[ ]
−1
a H (ω,θ) Ŝ Nf (ω) a(ω,θ)
assuming that the power from the desired direction does not change much from
snapshot to snapshot we obtain

N +1 1 N a(ω,θ)
H Cap = H Cap − βzz H (6.43c)
[ ]
−1
µ H
a (ω,θ) Ŝ Nf (ω) a(ω,θ)

where

(1 − µ)
β=
[ ]
−1
µ  (1 − µ)F HN +1 (ω) Ŝ Nf (ω) F N +1 (ω) + µ

© 2001 by CRC Press LLC

But, we can show that

a(ω,θ)
zH = F HN +1 (ω)H Cap
N

[ ]
−1
H N
a (ω,θ) Ŝ (ω)f a(ω,θ)

and reduce (6.43c) to

N +1 1 N
H Cap = H Cap − βzF HN +1 (ω)H Cap
N
(6.43d)
µ

It is shown in [11] that (6.43d) yields a stable estimate of the Capon filter.

§6.4 Beamformation with Coded Signals

In modern communication systems coded signals are used for transmitting
information, which consists of a bit stream of ones and zeros. The bits are
suitably coded into identifiable waveforms; for example ‘1’ may be coded into a
sinusoid of frequency f1 and ‘0’ is coded into another sinusoid of frequency f0
( f 1 ≠ f 0 ) as in FSK (frequency shift keying) modulation. The bits thus coded
after mixing with a carrier are sequentially transmitted. Since the physical
channel is a shielded cable (including optical cable) there is much less cross
channel interference. But, in the radio communication scenario, both transmitter
and receiver are in the open space. Naturally a sensor will receive signals from
more than one source. It is therefore of great interest to minimize this co-
channel interference, a problem unique to radio communication. To overcome
the problem of co-channel interference modern cellular radio communication has
been devised. A user needs to communicate to the nearest base station which in
turn is connected to a central exchange. Thus, it is possible to communicate
with a distant user without having to radiate a lot of power causing a drain on
the battery (in case of a mobile transmitter) and creating interference to other
users. Even this system seems to fail when many users in the same cell are
trying to reach the base station. To further mitigate the problem of co-channel
interference it is proposed to use a highly directional antenna at the base
station. This would enable the base station to separate the users having different
bearings and reduce the co-channel interference among them. In the urban
environment there is the additional problem of multipath propagation due to
scattering. Since the multipaths are likely to arrive from different directions, the
use of a directional array, it is hoped, will help to alleviate the problem of
fading, loss of bits, etc. However, as this is still a research problem we shall
not discuss this aspect.
The essential step in beamformation is estimation of the direction
vector of each source. Given the direction vector a beam may be formed in that

© 2001 by CRC Press LLC

 direction using one of the beamforming methods described in §6.1 and §6.2.
When there are a large number of sources (users), ordinarily we need to have a
large sensor array (more sensors than the number of users) for the estimation of
direction vectors. The advantage of using coded signals is that this limitation
no longer exists. We can have more number of users than the number of
sensors.
6.4.1 Co-channel Interference: We have derived in §5.2 a postcorrelation
covariance matrix (see (5.68)) and expressed the same in a form (5.67) suitable
for application of the subspace algorithm. The direction vectors to all users
may be estimated provided the array has more sensors than the number of users.
When this is not satisfied, the presence of users in excess of the number of
sensors will only introduce interference, known as co-channel interference. We
shall show how by employing extra information, which is available but not
used, we can overcome the problem of co-channel interference.
At the base station we will assume an array of sensors (EM dipoles).
The preferred array shape is a circular array (UCA) with uniform response in all
directions as the users are likely to be all around the base station. A linear array
(ULA) may also be used particularly when most users are on the broadside
where the array has the best possible response.
The postcorrelation covariance matrix is reproduced here for convenience.

Q
σ 2η
∑
2
C z 0 z 0 = p02 a(θ 0 )a(θ 0 ) H + pk2 a(θ k )a(θ k ) H + I (5.68)
3L k =1
L

The first term on the right hand side is of interest as we would like to estimate
the direction vector of the user of interest. The second term represents the co-
channel interference from all other sources. Notice that this term will be small
for large L.
6.4.2 Estimation of All Direction Vectors: We like to estimate the direction
vectors of all users in the same cell. For this we shall compute the
postcorrelation covariance matrices for all users. Thus, we will have Q
equations of the type given by (5.66)

Q
σ 2η
∑
2
C g k g k = pk2 a(θ k )a(θ k ) H + pi2 a(θ i )a(θ i ) H + I
3L i=0
L (6.44)
k ≠i

k = 0,1,..., Q − 1

We have in (6.44) Q matrix equations and Q matrix unknowns a(θ k )a(θ k ) H ,

k=0, ... Q-1. We shall express (6.44) in a matrix form

© 2001 by CRC Press LLC

 1 2
L ...
2 
L
Cz 0 z 0    3 3   p02 a(θ 0 )a(θ 0 ) H 
 I 
  2 2  I 
Cz1 z1   L 1 ... L   p1 a(θ1 )a(θ1 )
2 H

3   
.  3  ...  σ 2 . 
  = . ... .  + η  

.  . .    T s . 
...
... 
.    ... . 
  . ... .    
  pQ−1a(θQ−1 )a(θQ−1 ) 
Cz Q−1 z Q−1   H
  2 2
2
I 
 L L ... 1 
3 3 
Q ×1 Q×Q Q ×1 Q ×1
(6.45a)
In a compact form
σ 2η
C = £Θ + II (6.45b)
Ts

Multiplying by £-1 on both sides of (6.45b) we can express it as

σ 2η -1
Θ = £-1 C + £ II (6.46)
Ts
The error term in (6.46) may be expressed as a product of a diagonal matrix and
a column unit matrix. The elements of the diagonal matrix are equal to row
-1
sums of £ . Thus, the noise covariance matrix in the estimated direction
matrix remains diagonal. The variance of the noise may be estimated from the
eigenvalues of the direction matrix. The power, transmitted by each user, that
2
is, pi , can also be estimated from the largest eigenvalue of the direction
matrix.
6.4.3 Simulation Results: The estimated direction vector of a user is compared
with the known direction vector. A dot product between the two vectors is
computed as a measure of similarity,

â H (θl )a(θl )
εl =
â H (θl ) a(θl )

Note that 0 ≤ ε l ≤ 1 , the lower limit represents the worst estimate and the
upper limit represents the best estimate. We have computed the mean and the

© 2001 by CRC Press LLC

 snr Mean Variance
-20 dB 0.6029 0.1059
-10 dB 0.9746 3.8480e-04
-5 dB 0.9945 8.8470e-06
0 dB 0.9979 8.6691e-07
No noise 0.9983 1.0934e-06
Table 6.1: Error in the estimation of direction vector for different snrs. 10 sensor
ULA, randomly distributed 10 users, 100 snapshots, code length = 63 chips.

Number of Mean Variance

users
10 0.9983 1.0934e-06
20 0.9949 6.7374e-06
40 0.9930 1.5846e-05
50 0.9909 2.3899e-05
60 0.9878 3.6955e-05
Table 6.2: Error in the estimation of direction vector with increasing number of
users. 10 sensor ULA, randomly distributed users, 100 snapshots, no noise, and
code length = 63 chips.

variance of ε l as a measure of quality of estimate. The results are shown in

table 6.1. The postcorrelation matrix approach for the estimation of direction
vectors as described here is not limited by the requirement that the number of
sensors must be greater than the number of users as in the approach described in
[12] using both pre- and postcorrelation matrices. In fact to verify this claim,
above simulation was repeated with no noise for a different number of users.
The results are shown in table 6.2. There is, however, a slight decrease in the
quality of estimate.
6.4.4 Beamforming with Cyclostationary Signals: We consider P sources
emitting cyclostationary signals with different but known cyclic frequencies.
We like to find a set of weight coefficients which forms a beam in the direction
of a source having a specified cyclic frequency. The array output f(t) is
governed by the signal model given in (2.18). The noise is stationary but not
necessarily white. We use the frequency shifted version of f(t) defined in
(5.70). Let w + and w − be the beamforming weight coefficient vectors for
f + (t) and f − (t) , respectively. The cross-correlation of the outputs is given
by

© 2001 by CRC Press LLC

 ρ = E{w − H f − (t)f + H (t)w + }
(6.47)
= w − H c αf w +

where c αf is the cyclic covariance matrix defined in (5.71a) for zero lag. The
filter coefficients are chosen to maximize the cross-correlation (6.47) or its
magnitude square. Further, we require that w + and w − are unit norm vectors.

2
w − H c αf (τ)w + = max, w + H w + = 1, w − H w − = 1 (6.48a)

The solution of (6.48a) is given by left and right singular vectors corresponding
α
to the largest singular value of c f (τ) . Also w − and w + are, respectively, the
eigenvectors corresponding to the largest eigenvalues of c αf (τ)c αf (τ) H and
c αf (τ) H c αf (τ) [5].The cyclic covariance matrix appearing in (6.48) is in
practice replaced by the cyclic autocorrelation function defined in terms of time
average,

1 2
ĉ αf = ∑ f − (t)f +H (t) (6.48b)
T −T
2

When the carrier frequencies of different sources are sufficiently apart and the
signal duration T is large the cyclic autocorrelation matrix given by (6.48)
approaches the cyclic covariance matrix of a single source which is a rank one
matrix as shown in (5.71b) whose left and right singular vectors corresponding
1
to the largest singular value are equal to a 0 (for ULA). Thus, we have an
M
interesting result
1
w+ = w− = a0 (6.48c)
M

§6.5 Multipath Channel:

In many real life situations a signal (also interference) may reach an array of
sensors via more than one path. The resulting signal will be an overlap of
many copies of the actual signal emitted by the source. The different copies of
the signal may differ in arrival time and amplitude but remain highly correlated
unless they have traveled along widely different paths through a random

© 2001 by CRC Press LLC

 1

0.8

0.6

0.4
Amplitude

0.2

-0.2
-0.4
0 20 40 60 80
Time

Figure 6.9: A source emits a waveform (sinc function) shown by solid curve. The
signal reaches a sensor via four paths as in equation (6.49). The sensor output is
shown by dashed curve.

medium. For the time being we shall not consider such a situation. We shall
assume that all copies of the signal remain correlated. Even in this simplified
model of multipath propagation the waveform received by a sensor may have
no resemblance with the signal emitted by the source. To appreciate the kind of
deterioration a signal may suffer let us consider a source emitting a signal
sin(x)
described by sinc function (= ) and a multipath channel having four
x
paths. The relative amplitudes and delays are assumed to be random numbers as
given below:

 f 0 (t) − 0.0562 f 0 (t − 7.0119) 

f (t) =   (6.49)
+0.5135 f 0 (t + 9.1032) + 0.3967 f 0 (t + 2.6245)
where the signal emitted by the source is a sinc function,
f 0 (t) = sin c(0. 2π(t − 32)) .The sum waveform is shown in fig. 6.9.
Generally, the broadband signals are more distorted than the narrowband ones.
In chapter 4, in the context of source localization in a bounded
medium, we have modeled the array signal as

F(ω) = AwF0 (ω) + Ν(ω) (6.50)

© 2001 by CRC Press LLC

where [ ]
A = a 0 ,a1 ,...a q −1 and a 0 ,a1 ,...a q −1 are direction vectors pointing
to q paths ( a 0 is assumed to point to the direct path), and w stands for
complex weight vector, [ ]
w = w0 , w1 ,..., wq −1 where w0 , w1 ,..., wq −1 are
complex weighting coefficients applied to the temporal Fourier coefficients of
the array signal. They represent attenuation and phase change due to
propagation delays. If the channel is well characterized both A and w can be
estimated as a part of the source localization step, which, we will assume, has
been carried out prior to waveform estimation. Thus, as a first step, assuming
that Aw is known let us explore how well f 0 (t) can be recovered in the
presence of background noise and interference from other sources.
6.5.1 Known Channel: A few examples of simple channels, whose
characteristics are known or can be modeled reasonably well, are sketched in fig.
6.10. The channel shown in fig. 6.10a is a good approximation to surface
reflected radar signal, surface reflected seismic signal, etc. The reflection
coefficient of the surface is the unknown quantity but the direction of arrival of
the direct wavefront as well as the reflected wavefront can be easily computed
given the source location. For example, let the source be located at (l,hs )
where l is range and hs is depth from the surface. The direction of the arrival
of the direct wavefront is given by

hs − ha hs + ha
tan θ 0 = and tan θ1 = − (6.51)
l l

and ha is the depth of the array (midpoint) from the surface. In deriving (6.51)
we have used the method of images as outlined in chapter 1. For this channel
the direction vectors and weight vector, respectively, are A = a 0 ,a1 where [ ]
 − j 2 π d sin(θ 0 ) − j 2λπ 2 d sin(θ 0 ) −j
2π
( M −1)d sin(θ 0 ) 
a 0 = col 1,e λ ,e ,...,e λ 
 
 − j 2 π d sin(θ1 ) − j 2λπ 2 d sin(θ1 ) −j
2π
( M −1)d sin(θ1 ) 
a1 = col 1,e λ ,e ,...,e λ 
 

w = col{1,r}

where r is the coefficient of reflection which we shall assume for simplicity

independent of the angle of incidence. We can now express Aw vector in terms

© 2001 by CRC Press LLC

 Source
Array

(a)

Source

Array

(b)

Figure 6.10: Two types of simple channels for which the Aw vector may be
estimated from a knowledge of source location and channel characteristics.

of the direction vectors of the direct wavefront from the source and the one from
the image of the source.

Aw = [a 0 + ra1 ] (6.52)

First, let us try the linear least squares estimate (LLSE) of F0 (ω) . We shall
transform (6.11) into frequency domain, leading to

[ ]
−1
F̂0 (ω ) = w H A H Aw w H A H F(ω )
(6.53a)
[ ]
−1
= F0 (ω ) + w H A H Aw w H A H Ν(ω )

The quantity inside the square brackets in (6.53a) may be evaluated using
(6.52). It is given by

© 2001 by CRC Press LLC

 w H A H Aw = a 0H a 0 + r H a1H a 0 + ra 0H a1 + r a1H a1
2

 r H a1H a 0 + ra 0H a1  (6.53b)
= 2M 1 + 
 2M 

The variance of the noise power in the filtered array output is given by

σ η2
σ η̂2 =
[w H A H Aw]
σ η2 (6.54)
=
 r H a1H a 0 + ra 0H a1 
M 1 + r +
2

 M 

Compare (6.54) with (6.14), which was derived for two uncorrelated sources. In
the present case both sources are correlated (the second source is an image of the
primary source). The variance of the noise is reduced by a factor four, when r=1
and a1 = a 0 . The multipath propagation has indeed helped to improve signal
estimation.
Next, we shall try the Capon filter to estimate the waveform in the
presence of interference, another source at known location, and the usual
background white noise. The position of the sources and the receiving array are
shown in fig. 6.11. We shall assume that the sources are uncorrelated. The
directions of arrival of the direct and the reflected wavefronts are given by

hs − ha hs + ha
tan θ 00 = , tan θ10 = −
l l

hs + ∆h − ha h + ∆h + ha
tan θ 01 = , tan θ11 = s
l l
For simplicity we assume that the coefficient of reflection r is the same for
both sources. The direction vectors are given by

A 0 w = [a 0 + ra1 ]
A1w = [a 01 + ra11 ]

© 2001 by CRC Press LLC

 Array

s1

s0

Figure 6.11: Two sources at the same range but at different heights. A vertical array
of sensors is assumed.

The array output is modeled as

F(ω ) = A 0 wF0 (ω ) + A1wF1 (ω ) + Ν(ω ) (6.55)

To compute the Capon filter given by (6.26) we need the spectral matrix of the
array output. Since the sources are assumed uncorrelated and the background
noise is white, the spectral matrix may expressed as

S(ω ) = A 0 ww H A 0H S0 (ω ) + A1ww H A1H S1 (ω ) + σ η2 I (6.56)

The Capon filter is specified by (6.26). To estimate the waveform emitted by

the zeroth source the required filter is given by

S−1 Aw
H cap = H H f 0 −1 (6.57)
w A 0 (ω , θ )S f A 0 w

Applying the filter given in (6.57) to the array output we obtain

w H A 0H S−1
f A1w w H A 0H S−1Ν(ω )
H F(ω ) = F0 (ω ) + H H −1
H
cap F1 (ω ) + H H f−1 (6.58)
w A0 S f A0 w w A0 S f A0 w

© 2001 by CRC Press LLC

 1

0.8

0.6
Cross talk

0.4

0.2

0
50 55 60 65 70
Separation in meters

Figure 6.12: Cross talk as a function of separation between two sources. The range
is the same for both sources. Solid curve: range=1000m and dashed curve=5000m.
The first source is 50 meters above the surface.

While the signal from the zeroth source has been fully extracted there are two
terms in (6.58) interfering with the desired signal. Here we shall consider the
term representing the interference from the source #1. The second term, that is,
the noise term, will not be considered as it follows the same approach used
previously in connection with the single source case. The interference due to
second source will be measured in terms of cross talk as defined previously. In
the present case the cross talk is given by

(w H A1H S−1 H H −1
f A 0 w)(w A 0 S f A1w)
cross talk = 2 (6.59)
w H A 0H S−1
f A0 w

The cross talk as a function of source separation for two uncorrelated sources is
plotted in fig. 6.12. A 16 sensor vertical ULA was assumed. Notice that for a
short range the second source has little influence on the first source. But this
influence grows rapidly as the range increases.
6.5.2 Partially Known Channel: In many real life problems the channel
characteristics are never fully known as it is impossible to measure the micro
level variations causing path length variations on the order of a fraction of
wavelength. Such variations are known to affect the performance of source
localization algorithms, particularly those belonging to a high resolution class
[13]. On the other hand we may have fairly good knowledge about the general

© 2001 by CRC Press LLC

 features of a channel but not enough to characterize at micro level. One such
example is a shallow water sea where the sea bottom variability is high. While
the general geometry of the shallow water channel, broad undulations of the sea
bottom, water temperature variations, etc. are available from the actual
measurements, the factors which affect the performance of waveform estimation
such as the details of sea bottom, particularly sediment distribution, and sea
surface undulation are unknown. In chapter 5 we have shown how source
localization can be carried out with a full knowledge of a shallow water
channel. Here we shall describe an approach capable of estimating the source
location as well as signal waveform emitted by the source given only partial
knowledge.
Consider a single source in a shallow water channel and a vertical array
of sensors. We have shown in chapter 5 that the eigenvector corresponding to
the largest eigenvalue is related to Aw ,

Aw
Es = , Aw ≠ 0 (6.60)
Aw

In (6.60) the weight vector w is dependent on the channel parameters and the
columns of A matrix on the source position. Notice that the weight vector w
in (6.60) occurs linearly while the source parameters in A occur nonlinearly.
A least squares method of estimating the nonlinear parameters by first
eliminating the linear parameters followed by minimizing the norm of the error
vector was first suggested by Guttman et al.[14] and applied to a signal
processing problem by Tuft and Kumaresan [15]. We shall exploit here this
approach. In (6.60) we assume that the source position is approximately known
and write the equation in terms of the unknown w vector,

w
Es = A = Aw̃ (6.61)
Aw
#
Let A be the pseudoinverse of A. The least squares estimate of w̃ will be
given by

˜ˆ = A # E s
w

Substitute back into (6.60) and obtain an estimate of E s . The mean square
error is given by

© 2001 by CRC Press LLC

 error 2 = (I − AA # )E s 2
= P A⊥ E s 2

⊥
where P A is the orthogonal projection complement of matrix A . The mean
square error is now minimized with respect to the source location parameters.
We define a parametric spectrum as,

1
S(R 0 , Zs) = H ⊥
E s P A ES (6.62)

When the source location is exactly known, P ⊥A = (I − AA # ) = E ηE #η , the

parametric spectrum defined in (6.62) turns out to be same as in (4).
In practical implementation the parametric spectrum is computed over
a dense grid in the range-depth space. At each grid point the projection matrix
⊥
PA is computed and (6.62) is evaluated. Thus, in the proposed method, the
subspace spanned by the columns of the A matrix (range space) is steered until
⊥ H
PA coincides with Eη Eη , that is, when the parametric spectrum becomes
very large (ideally infinite). After obtaining the source location parameters we
use them to estimate the weight vector,

˜ˆ = A max
w #
Es (6.63)

# #
where A max is A evaluated where the parametric spectrum is maximum.
˜ˆ is the least mean square estimate of w̃ . An example of parametric
Thus, w
spectrum is shown in fig. 6.13. A low power source (-10dB) is assumed at
range 4600m and depth 50m in a Pekeris channel of depth 200 meters. A
vertical ULA is placed at a depth of 70m. For the same channel the reflection
coefficients were computed (6.63) from the eigenvector corresponding to the
largest eigenvalue. The reflection coefficients are normalized with respect to
Aw , which may be obtained from the fact that the weighting coefficient
corresponding to the direct path is by definition equal to one; hence its actual
observed value must be equal to Aw . In table 6.3 the estimated and the actual
reflection coefficients for the first eight images out of twenty multipaths used
in computation are listed for two different array lengths.
In computer simulation we have found that, for good results, the
number of sensors has got to be many more than the lower limit given in [16].
The least mean square error in the estimated reflection coefficients for different
number of sensors is shown in fig. 6.14. Here the number of significant images

© 2001 by CRC Press LLC

 10
1000
De
pth
ge
Ran
180 8000

Figure 6.13: The parametric spectrum for a single source located at range 4600 m and
depth 50 m. The source power is -10dB (relative to the background noise). Under
water channel is 200m deep Pekeris channel (soft bottom, speed: 1600m/s and
relative density: 2.0). A vertical ULA consisting of 40 sensors is placed at a depth
of 70m from the surface. Range scan is from 1000 m to 8000 m in steps of 200 m.
Depth scan is from 10 m to 180 m in steps of 20 m.

Image True reflection Estimated reflection coefficients

# coefficients
M=40 M=60
Real Imag Real Imag Real Imag
1 1.0 0.0 1.0 0.0 1.0 0.0
2 -.861 -.506 -.847 -.718 -.870 -.490
3 .467 .882 .612 .829 .416 .886
4 .995 -.064 1.075 -.260 .998 -.091
5 .981 .167 1.395 -.053 .925 .205
6 .905 .407 1.040 .251 .858 .459
7 .626 -.768 .705 -.783 .641 -.752
8 .107 .981 -.013 .964 .107 .986
Table 6.3: A comparison of estimated reflection coefficients (first eight
coefficients) with true reflection coefficients, computed for a channel described in
fig. 6.13. The results for M=40 and 60.

© 2001 by CRC Press LLC

 -50

-40

-30
mse in dB

-20

-10

0
10 20 30 40 50 60 70
No of sensors
Figure 6.14: The role of array size on the mean square error in the estimated
reflection coefficients is shown above. Twenty multipaths were assumed.

is twenty (P=20) and hence, according to the lower limit (> 2P+ 2)
the minimum number of sensors ought to be more than 42. We observe that an
array of sixty sensors appears to be optimum.
In addition to unknown reflection coefficients we have background
noise which is likely to be both spatially and temporally correlated. But, since
it is uncorrelated with the signal, it occurs in a linear combination; as such it
may be estimated using the approach used for the estimation of the reflection
coefficients. Indeed, such an approach has been used by Boehme [17] for the
estimation of the noise spectrum first using an approximate knowledge of the
channel. In the next step, the previously estimated noise spectrum is used in
the expression for the likelihood ratio which is then maximized with respect to
the unknown channel parameters and the source position. The maximization of
the likelihood ratio is, however, highly computation intensive [18].

§6.6 Exercises:
1. Consider P uncorrelated wavefronts incident on an M sensor ULA. Let the
background noise be uncorrelated with all signals. Show that the Wiener filter
in the frequency domain is given by

H W (ω) = S−1
f (ω)AS 0 (ω)

© 2001 by CRC Press LLC

 where A is a matrix whose columns are direction vectors of the wavefronts
which are incident on the array.

2. Show that the Wiener filter (also Capon filter) can be written as

H W = v s α s−1v sH a 0 S0

where a 0 is the direction vector of the wavefront from the first source and S0
is signal power, (Hint: Use (3.38b) and the property given in (4.14b).) Such a
filter is robust against calibration error (see p.342).

3. In §6.1 and §6.2 we have seen that in estimating a waveform in the presence
of interference there is always some cross talk, that is, leakage of power from
the interfering signal to the desired signal. It is possible to devise a set of
filters, Hm (ω), m = 0,1,..., M − 1 , which will null the interference without
M −1
distorting, that is, ∑H
m=0
m (ω) = 1, the desired signal (but with no noise).
The array has been steered to receive the desired signal. The DOA of the desired
signal and that of the interference are known. Show that

M −1
M − e jωτ m ∑ e − jωτ i
Hm (ω) = i=0
2
M −1
M −
2
∑e
i=0
− jωτ i

where τ i is the time delay of the interference at ith sensor. The interference is
nulled except when ωτ i is equal to an integer multiple of 2π [4].

References
1. B. Otterstein, R. Roy, and T. Kailath, Signal waveform estimation in sensor
array processing, Asilomar Conf, pp. 787-791, 1989.
2. B. Friedlander and A.J. Weiss, Effects of model errors on waveform
estimation using the MUSIC algorithm, IEEE Trans., SP-42, pp. 147-155,
1994.
3. R. T. Compton, Pointing accuracy and dynamic range in steered beam
adaptive array, IEEE Trans., AES-16, pp. 280-287, 1980.
4. M.T. Hanna and M. Simaan, Absolutely optimum array filters for sensor
array, IEEE Trans., ASSP-33, pp. 1380-1386, 1985.
5. Q. Wu and K. Wong, Blind adaptive beamforming for cyclostationary
signals, IEEE Trans., SP-44, pp. 2757-2767, 1996.

© 2001 by CRC Press LLC

6. S.P. Applebaum, Adaptive array, IEEE Trans., AP-24, pp. 585-598, 1976.
7. L. C. Godara, Error analysis of the optimal antennal array processors, IEEE
Trans, AES-22, pp. 395-409, 1986.
8. B. Widrow and S. D. Stearms, Adaptive Signal Processing, Prentice-Hall,
Inc, Englewood Cliffs, NJ, 1985.
9. J. R. Treichler, C. R. Johnson, Jr., and M. G. Larimore, Theory and design
of adaptive filters, Wiley Interscience, New York, 1987.
10. J. W. Kim and C. K. Un, A robust adaptive array based on signal subspace
approach, IEEE Trans., SP-41, pp. 3166-3171, 1993.
11. R. Schreiber, Implementation of adaptive array algorithms, IEEE Trans.,
ASSP-34, pp. 1038-1045, 1986.
12. A. F. Naguib, A. Paulraj, and T. Kailath, Capacity improvement with
base-station antenna array in cellular CDMA, IEEE Trans. on Vehicular
Technology, vol. 43, pp. 691-698, 1994.
13. J. R. Daugherty and J. F. Lynch, Surface wave, internal wave, and source
motion effects on matched field processing in a shallow water waveguide, J.
Acoust. Soc. Am., vol. 87, pp. 2503-2526, 1990.
14. I. Guttman, V. Peereyra, and H. D. Scholnik, Least squares estimation for
a class of nonlinear models, Technometrics, vol. 15, pp. 209-218,1973.
15.D. W. Tuft and R. Kumaresan, Improved spectral resolution II, Proc.
ICASSP, Denver, CO, 1980, pp. 392-397.
16. P. S. Naidu and T. Ganesan, Source localization in a partially known
shallow water, J. Acoust. Soc. of Am., vol 98, Pt. 1, pp. 2554-2559, 1995.
17. J. F. Boehme, Array Processing in Advances in Spectrum Analysis and
Array Processing, S. Haykin (Ed), pp. 1-63, Prentice-Hall, Englewood Cliffs,
NJ, 1991.
18. C. F. Mecklenbraeuker, A. B. Gershman, and J. F. Boehme, ML
estimation of environmental parameters in shallow ocean using unknown
broadband sources, Proc. IEEE ICNNSP-95, Nanjing, China, 1995.

© 2001 by CRC Press LLC

 Chapter Seven

Tomographic Imaging

As a wavefield propagates through a medium it is subjected to time delays and

loss of power. The wavefield is reflected from interfaces separating media of
different impedances, and is scattered by inhomogeneities present in the
medium. By observing these effects it is possible to study the characteristics of
the medium through which the field has propagated. Seismic exploration, on
which depends the future discoveries of the petroleum deposits, exploits these
effects of propagation to produce a detailed image of the subsurface geologic
structure which may be conducive to accumulation of the hydrocarbon deposits.
Likewise the ultrasonic imaging used in medical diagnoses and in
nondestructive testing also exploits the propagation effects of the wavefield. In
this chapter we shall study these effects of propagation for the purpose of
constructing a three dimensional image of the medium. Tomography refers to
cross-sectional imaging of objects from either transmitted, reflected or diffracted
wavefields. Accordingly, there are three different types of tomographic imaging
methods. One or more effects of propagation such as accumulated attenuation,
travel time, wavefield produced by diffraction or scattering are observed in all
directions (360o for 3D imaging). The observations such as travel time delays
or accumulated attenuation are inverted by solving a system of linear equations.
Where the medium is a diffracting type, that is, the size of inhomogeneities is
comparable to the wave length of the illuminating wavefield, the preferred
approach is Fourier inversion. The subject of tomographic imaging is covered
in the next four sections. In the last section we investigate how to estimate the
shape of an object from its scattered field.

§7.1 Nondiffracting Radiation:

When the wavelength of illuminating radiation (e.g., x-rays, ultrasound) is
much smaller than the dimensions of the inhomogeneities in the propagating
medium the concept of ray propagation becomes useful. The rays may travel in
straight line or along a curved line depending upon the average wave speed: a
straight line when the average speed is constant or curved path when the average
speed is spatially variable as in a layered medium. The local speed variation is
assumed to have a negligible effect on the ray paths. The propagation has two
effects on the wave, namely, wave attenuation and delay, both of which are of
great significance from the point of tomographic imaging. Typically in x-ray
tomography, wave attenuation is used and in ultrasound tomography and also in
seismic tomography total time delay is used.
7.1.1 Absorption: Consider for example an object cross section represented by a
function f(x,y). A straight line ray intersects the object and suffers a certain
amount of attenuation depending upon the length of the ray path lying inside

© 2001 by CRC Press LLC

363
 source
y

ds
t
x
B
Object

sensor

Figure 7:1. A ray passing through an absorbing object suffers an attenuation

proportional to the integral over the path lying within the object.

the object (see fig. 7.1). Let N in be the number of photons incident on the
object at point A and Nd be the number of photons coming out at point B
within the time interval of measurement. N in and N d are related as below
B
N d = N in exp[− ∫ f ( x, y)ds ] (7.1a)
A

Nd
Define attenuation as negative of log , which turns out to be equal to the
Nin
integral of f (x, y) along path AB,
B
Nd
Nin ∫A
p = − log e = f (x, y)ds (7.1b)

where p is also known as projection. The equation of line AB may be expressed

as

x cos(θ) + ysin(θ) = t

© 2001 by CRC Press LLC

 y

Parallel
beam
A
ds

t θ
x
B

(a)
Scanning
sensor
y
Fan
beam

α
t θ x

Scanning
(b) sensor

Figure 7.2: (a) A parallel projection is obtained by illuminating an object with a

parallel beam. (b) A fan projection is obtained by illuminating an object with a fan
beam generated by a point source at a radial distance d and angular distance (90+α)o.

© 2001 by CRC Press LLC

 where θ is slope of the line and t stands for perpendicular distance to the line
from the origin (see fig. 7.1). Equation (7.1b) may be expressed as follows:

B
N
pθ (t) = − log e d = ∫ f (x, y)ds
Nin A
(7.2)
∞ ∞

= ∫ ∫ f (x, y)δ(x cos(θ) + ysin(θ) − t)dxdy

−∞ −∞

pθ (t) is known as Radon transform (see page 21). For a fixed θ and variable t,
we obtain p θ (t) , a continuous function of t, known as parallel projection
which may be generated by illuminating an object with a parallel beam and
scanning the output with a receiver (see fig. 7.2a)
Taking Fourier transform on both sides of (7.2) we obtain

Pθ (ω)
∞

= ∫ p (t)exp(− jωt)dt
−∞
θ

∞ ∞ ∞

= ∫∫
−∞ −∞
f (x, y)dxdy ∫ δ(x cos(θ) + ysin(θ) − t)exp(− jωt)dt
−∞
∞ ∞

= ∫ ∫ f (x, y)exp( jω(x cos(θ) + ysin(θ))dxdy

−∞ −∞
(7.3)

= F(ω cos(θ),ω sin(θ))

where F(ω cos(θ),ω sin(θ)) is the 2D Fourier transform of the object

function f(x,y) evaluated on u = ω cos(θ) and v = ω sin(θ) . The Fourier
transform of a projection of an object function taken at an angle θ is equal to
the slice of 2D Fourier transform of the object function evaluated on a radial
line at an angle θ as shown in fig. 7.3. This is known as Fourier slice
theorem [1, p. 372]. By changing θ continuously the 2D Fourier transform of
the object function is evaluated over a series of radial lines as shown in fig.
7.3.
A fan beam is more appropriate, as a point source at finite distance
emits a spherical beam which, when used to illuminate a finite target, may be
considered as a conical beam in three dimensions or a fan beam in two
dimensions. Both slope and the perpendicular distance of a ray depend upon the
ray angle measured with respect to radius vector of the source. An exact

© 2001 by CRC Press LLC

reconstruction algorithm for a fan beam has been worked out after some tedious
geometrical simplifications [1], but a simpler approach, where a series of fan
beams covering 360o may be regrouped into a series of parallel beams, is of
greater interest from the array signal processing point of view. This approach is
outlined below.
7.1.2 Filtered Backprojection Algorithm: The reconstruction algorithm consists
of inverse Fourier transformation of a radially sampled object Fourier transform
which in turn is obtained from parallel projection data. Let us express the 2D
inverse Fourier transform in polar coordinates.

+∞
1
∫ ∫ F(u,v)e
j (ux + vy)
f (x, y) = dudv
4π 2 −∞
(7.4)
∞ 2π
1
= 2 ∫ sds ∫ F(s,θ)e js( x cos θ+ y sin θ) dθ
4π 0 0

The inner integral may be expressed as a sum of two integrals,

2π js(x cosθ + ysin θ)

∫ F(s,θ)e dθ
0
π js(x cosθ + ysin θ) 2π js(x cosθ + ysin θ)
= ∫ F(s,θ)e dθ + ∫ F(s,θ)e dθ
0 π
π js(x cosθ + ysin θ)
= ∫ F(s,θ)e dθ
0
π js(x cos(π + θ) + ysin(π + θ))
+ ∫ F(s, π + θ)e dθ
0
Using the mapping

F(s, π + θ) = F(−s,θ)

equation (7.4) may be written as

© 2001 by CRC Press LLC

 v

u
F(u,v)

Figure 7.3: The Fourier transform of the object function is scanned along a series of
radial lines, one for each projection.

π ∞
1 1 
f (x, y) = ∫  ∫
2π 0  2π − ∞
F(s,θ) s e js( x cos θ+ y sin θ) ds dθ

(7.5)
π ∞
1 1 
= ∫  ∫
2π 0  2π − ∞
F(s,θ) s e jst ds dθ


Note that, for a fixed θ , F(s, θ ) is equal to Pθ (ω) given in (7.3). Thus, the
quantity inside the curly brackets in (7.5) may be obtained from the parallel
projection by simply filtering it with a filter having a transfer function,
H(s) = s .
π
1
2π ∫0
f (x, y) = f̃ (x cosθ + ysin θ)dθ (7.6)

where

1 ∞ 
∫ θ
jω( x cos θ+ y sin θ)
f̃ (x cosθ + ysin θ) =  P (ω) ω e dω  (7.7)
 2π − ∞ 

© 2001 by CRC Press LLC

is often known as a filtered projection. In (7.7) the projection data is projected
back onto a section t = x cos θ + y sin θ for a fixed θ and hence the
reconstruction procedure is called the
back projection algorithm. The process of reconstruction consists of filtering
each projection with a filter whose transfer function is H(ω) = ω and then
backprojecting according to (7.6).
7.1.3 Algebraic Reconstruction: The back projection algorithm requires a ULA
capable of going round the target object. Such an idealistic experimental setup
cannot be achieved atleast in one important area of application, namely,
exploration for earth resources. The sensor array tends to be nonuniform and
generally distributed over a large area. Furthermore, since experimental
observations are necessarily confined to the earth’s surface or a few deep
borewells, it is practically impossible to go round the target object to get a 4 π
solid angle coverage. The back projection algorithm cannot be applied in most
real situations except perhaps in seismic exploration for oil where a near ideal
experimental setup can be realized. It is, therefore, necessary to devise an
alternate approach, albeit less accurate and of lower resolution. The basic idea in
this alternate approach is to divide the target object into as many homogeneous
cells as possible. The cell size is small enough to allow the assumption of no
variation in physical parameters (e.g., wave speed) over the size of a cell and
large enough to allow the validity of ray propagation. As a ray propagates
through the target object, it passes through many cells lying in its path. The
ray as it reaches a sensor carries the cumulative effect of all cells, for example,
sum of all delays introduced by all those cells lying in its path. It is further
assumed that a ray does not suffer any refraction (or reflection) on the boundary
of a cell and hence it travels in a straight line joining source and sensor. For
this assumption to hold good the change in the acoustic impedance must be
small (on the order of 10%). It is, however, possible to relax this assumption
but only at the cost of increased computation as the ray paths have to be traced
numerically. The reconstruction process becomes iterative where, starting from
an initial estimate of the speed variation, successive corrections are introduced
consistent with the available time delay information. The ray paths will have to
be traced afresh at the beginning of each iteration. Each cell would introduce a
delay proportional to the path length inside the cell and the unknown wave
speed in the cell. Let d represent the length of each side of a cell and cm be the
wave speed in mth cell. The maximum delay introduced by mth cell will be
2δ
τm = . The contribution of a cell toward the total delay observed at a
cm
sensor would depend upon the sensor position. It is possible that some cells do
not contribute at all. Indeed in any typical sensor array distribution there are
many cells through which a ray joining the source to sensor will not pass at
all; hence there can be no contribution from such cells. This may be easily seen
in fig. 7.4 where out of 56 cells a ray passes through less than 10 cells. Let

© 2001 by CRC Press LLC

 Source

δx δ

δ mn

to n th sensor

Figure 7.4: A target is divided into many square cells (or cubes in 3D). The wave
speed is assumed to be constant in a cell. The path delay introduced by mth cell is
δ mn
given by where δ mn is path length in mth cell of a ray going to nth sensor.
cm

wm,n represent a weighting coefficient which when used along with the
maximum delay gives a delay contributed by mth cell to nth sensor. In terms of
path length of a ray in the mth cell the weighting coefficient is given by
δ mn
wm,n = (see fig 7.4). Thus, the total delay observed at the nth sensor is
2δ
given by
M −1
tn = ∑w
m=0
τ
m,n m n=0, 1,...,N-1 (7.8a)

where we have assumed that there are M cells and N sensors. In matrix form (1)
may be expressed as

t = wτ (7.8b)

© 2001 by CRC Press LLC

where t = col{t0 , t1 , ... ,t N }, τ = col{τ 0 , τ1 , ... , τ M } and w is a N × M
matrix of weight coefficients. When N ≥ M a unique solution of (7.8b) may
be given by

τ = (w T w)−1 w T t (7.9)

T
provided w w is nonsingular. The question of the rank of the weight matrix
w has no quantitative answer but we can give some qualitative guidelines:

i) The ray path lengths in different cells must be quite different so that
there is correlation among weight coefficients. In fig.7.4, this is more likely
to happen with a fan beam than with a parallel beam.
ii) The weight matrix is fully determined by the sensor array geometry.
For example, if sensors are too closely spaced all rays will travel through the
same group of cells and each cell will contribute roughly the same delay. The
weight matrix will then tend to be more singular.
iii) The sensor and source arrays must be so designed that the rays pass
through different cells in different directions. More on this possibility will be
discussed in the next subsection on borehole tomography.

It may be noted that since the weight matrix is generally a large sparse matrix,
efficient techniques have been designed for fast and economical (in terms of
memory requirements) inversion of the weight matrix. This is, however,
beyond the scope of this book. The interested reader may like to review an article
by S. Ivansson [2].
Borehole Tomography: The use of a source array often improves the rank
condition of the weight matrix. Consider a P source array arranged in some
unspecified form. Equation (7.8b) may be used to express the output as

t p = w p τ, p = 0,1, ... P − 1 (7.10)

where t p is the array output due to pth source and w p is the weight matrix
corresponding to the position of pth source. Next, we stack up all array vectors
into a single vector. Note that τ is independent of the source position.
Equation (7.10) reduces to

t̃ = w̃τ (7.11)
where t̃ = {t 0 ,t1 ,...,t P −1 } and w̃ = {w 0 , w1 ,..., w P −1 } . The solution of
(7.11) may be expressed as

© 2001 by CRC Press LLC

 borehole#1 borehole#2

Source Sensor
array (a) array

Sensor array borehole

Source
array (b)

Figure 7.5: Source and sensor arrays in borehole tomography. (a) A target lies
between two boreholes. In borehole# 1 sources are arranged as a ULA of sources and
in borehole# 2 sensors are arranged as ULA of sensors. (b) In another arrangement
the sensor array is on the surface and the source array is in the borehole.

© 2001 by CRC Press LLC

 source sensor
array array

Figure 7.6: A target consisting of 16 cells (unit cells) lies between two boreholes
separated by 16 units. An eight sensor array is located in the right hole and the
source array (up to five sources) is in the left hole. The center of sensor array and
source array is at the average depth of the target.

τ = ( w̃ T w̃)−1 w̃ T t̃ (7.12)

We shall now consider a specific example of source and sensor arrays used in
borehole tomography in geophysical exploration [3]. A typical arrangement in
borehole tomography is shown in fig. 7.5. The sources are fired sequentially
and the transmitted signals are recorded for later processing.
Source/Sensor Array Design: For successful reconstruction of wave speed
T
variations the primary requirement is that w̃ w̃ in (7.12) must be invertible.
Since w̃ is entirely determined by the source and sensor array geometry, it is
possible to come up with a proper design for the source and sensor arrays which
would make the rank of w̃ equal to the number of cells. A simple numerical
T
example is worked out to show how the rank of w̃ w̃ depends on the number
of sources, source spacing, and sensor spacing. The interesting outcome of this
exercise is the fact that the number of sensors need not be greater than the
number of cells. One may achieve by using multiple sources what could be
achieved by using more sensors.
The source and sensor array geometry along with the target location are
shown in fig. 7.6. The sensor spacing (d units) and source spacing (s units) are
the variable parameters and the rest of the geometry remains fixed. From each

© 2001 by CRC Press LLC

 Number of
sources d=1 d=1.5
s=0.5 =1.0 =1.5 s=1.0 s=1.5
1 6 7
3 13 14 15 16 16
(1350) (1850)
5 15 15 16 16 16
(2200) (2200) (835)
T
Table 7.1: Rank of w̃ w̃ as a function of source and sensor spacing. The
bracketed quantity represents the eigenvalue spread.

source eight rays (straight lines) were drawn toward eight sensors. The line
intercept in each cell was found and the weight coefficient was computed as
described on page 370 (also see fig. 7.4). The weight matrix w̃ is first
T
computed and then the rank of w̃ w̃ , whose inverse is used in (7.12). The
T
results are shown in table 7.1. To achieve the full rank property for w̃ w̃ we
must have three to five sources and the sensor spacing should be around 1.5
units. Note that we have considered only eight sensors which is half the
number of cells. We have compensated for this deficiency by using three to five
sources. When the source array is close to the target the angular width of the
illuminating beam becomes large, which in turn requires a large aperture sensor
array to capture the wide illuminating beam. However, indefinite increase of the
sensor spacing will not help. There exists a range of sensor separation over
T
which not only w̃ w̃ is full rank but it is also stable as shown in fig 7.7
whereas for sensor separation between 1.5 and 3.0 units the eigenvalue spread
T
of w̃ w̃ is low and the matrix becomes singular outside the range 1.0 to 4.0.
The above findings are specific to the geometry of source and sensor arrays and
the target; nevertheless similar behavior is expected in other situations.

§7.2 Diffracting Radiation:

Diffraction becomes important whenever the inhomogeneities in an object are
comparable in size to the wavelength of the wavefield used for imaging. In
tomographic imaging, an object is illuminated from many different directions,
either sequentially or simultaneously, and the image is reconstructed from the
scattered wave field collected by an array of sensors, usually a linear array. Early
workers who attacked the problem of deriving the inversion algorithm for
tomography with diffracting wavefields were Iwata and Nagata [4] and Mueller

© 2001 by CRC Press LLC

 10 4

10 3

10 2
0 1 2 3 4 5
Sensor separation

Figure 7.7: Eigenvalue spread as a function of sensor separation. Five sources are
spaced at interval 1.5 units. The target used is same as in fig. 7.6.

et al. [5] who based their work on Wolf's work [6] on the inverse scattering
problem assuming the first order Born approximation. A good review of
diffraction tomography may be found in [1, 7].
Linear Array: An object is illuminated from various directions with a diffracting
source of radiation such as acoustic waves whose wavelength is comparable to
the scale of inhomogeneities. The incident wave energy is scattered in all
directions by the diffraction process within the object. A long linear array of
sensors facing the incident wave field is used to record the forward scatter (see
fig. 7.8). In §1.6 we derived an expression for the scattered field in the x-y
plane due to a plane wave traveling in z direction and illuminating a three
dimensional object (spherical). A similar result for a two-dimensional object
(cylindrical) was also given. For the sake of simplicity we shall talk about
tomographic imaging of a two-dimensional object. Consider an arrangement
wherein a cylindrical object is illuminated with a plane wave traveling at right
angle to the axis and a linear array of sensors located on the opposite side as
shown in fig. 7.8. The Fourier transform of the scattered field, which is
measured at a set of discrete points by the sensor array, may be obtained from
(1.82) where set u0=0 and v0=k0,

© 2001 by CRC Press LLC

 Array

Object

Plane wave

Figure 7.8: A basic experimental setup for tomographic imaging. A linear array,
ideally of infinite aperture, is used to receive the forward scatter from the object.

j k 2 −u 2 l
jk 2 e 0
P0 (u) = 0 ∆c̃(u, k02 − u 2 − k0 )
2 k0 − u
2 2
(7.13a)

u ≤ k0

where ∆c̃(u,v) is a 2D Fourier transform of δ c(x,y). As u varies from - k0 to

+k 0 P0 (u) traces a cross section of ∆c̃(u,v) along a semicircular arc as
shown in fig. 7.9. The circle is centered at (0, − k0 ) and the radius is equal to
k0 . The circular arc is described by an equation v = k02 − u 2 − k0 , u ≤ k0 .
The entire ∆c̃(u,v) may be sampled over a series of arcs either by rotating the
object but keeping the direction of illumination and the position of array fixed
or vice versa. We consider the first case. When the object is rotated so is its
Fourier transform through the same angle. Hence, for an object, which is
rotated through an angle ϕ , the scattered field, analogous to (7.13a), is given
by the following equation:

© 2001 by CRC Press LLC

 v

k0 2k0

Figure 7.9: As u varies from - k0 to +k0 P0 (u) traces a cross section of ∆c̃(u, v)
along a semicircular arc (thick curve).

j k 2 −u 2 l
jk02 e 0  u cosϕ + ( k02 − u 2 − k0 )sin ϕ, 
Pϕ (u) = ∆c̃  
2 k02 − u 2  −usin ϕ + ( k 2 − u 2 − k )cosϕ 
 0 0  (7.13b)
π π
u ≤ k0 and − ≤ ϕ ≤
2 2
The center of the circular arcs will all lie on a circle of radius k0 centered at
(−k0 sin ϕ , − k0 cos ϕ ) (see figure 7.10).
7.2.1 Filtered Backpropagation Algorithm: This is an adaptation of the filtered
backprojection algorithm developed for nondiffracting radiation to diffracting
radiation. The essential difference is that the sampling paths are now arcs of a
circle instead of radial lines in nondiffracting radiation. We start by expressing
the Fourier integral in polar coordinates (see (7.5-7.7))

π ∞
1 1 
2π ∫0  2π −∫∞
js( x cos ϕ + y sin ϕ)
f (x, y) =  F(s,ϕ) s e ds dϕ


© 2001 by CRC Press LLC

 v

2k0

Figure 7.10: The entire 2D Fourier transform, ∆c̃(u, v) , is sampled along a series of
semicircular arcs by rotating the object keeping the transmitter and array fixed.
Because u must lie within - k0 to +k0 the radius of the disc spanned by the
semicircular sampling arcs is equal to 2k0 .

π ∞
1 1 
2π ∫0  2π −∫∞
=  F(s,ϕ) s e jst
ds dϕ

π
1
2π ∫0
= f̃ (x cosϕ + ysin ϕ)dϕ

where

1 ∞ 
f̃ (x cosϕ + ysin ϕ) =  ∫ F(s,ϕ) s e js( x cos ϕ + y sin ϕ) ds 
 2π − ∞ 

We shall use the above representation of 2D Fourier transform in polar

coordinates. Note that in place of f(x,y) we have δc̃ and in place of F(s,ϕ) we
have ∆c̃ . Equation (7.13b) relates the Fourier transform of the scattered field to
the object Fourier transform, that is,

∆c̃(u cosϕ + u′ sin ϕ, − usin ϕ + u′ cosϕ)

© 2001 by CRC Press LLC

 2 −j k 02 −u 2 l
= 2
k02 − u 2 e Pϕ (u)
jk0
where u ≤ k0 and u′ = ( k02 − u 2 − k0 ) . Using the above relation we
obtain the reconstruction

π
1
2π ∫0
δc̃(x, y) = ˜ cosϕ + ysin ϕ)dϕ
δc̃(x (7.14a)

where

 k0
2 − j k 2 −u 2 l
k02 − u 2 e 0 
1 2 
δc̃(x
˜ cosϕ + ysin ϕ) =  ∫ jk0  (7.14b)
 2π − k 0 ×P (u) u e ju( x cos ϕ + y sin ϕ) du 
 ϕ 

According to (7.14b) the scattered field measured by the linear array is filtered
with a filter whose transfer function is given by

− j k 02 −u 2 l
H(u) = k02 − u 2 e u (7.14c)

− j k 2 −u 2 l
The term of interest in the filter transfer function is e 0
which
represents backward propagation of the wavefield from the plane of observation
to the target (see §1.2.4). For this reason the reconstruction algorithm described
above is called the filtered backpropagation algorithm. Except for this difference
the algorithm is quite similar to the filtered backprojection algorithm.
7.2.2 Multisource Illumination: There are many situations where it is not
possible to turn an object around for multiple illuminations nor is it possible
to turn around the source-array configuration, keeping the object fixed as the
space around the object may not be accessible as is the case in geophysical
imaging, nondestructive testing, remote monitoring, etc. In such a situation it
is recommended to employ an array of sources, often arranged as a linear
equispaced array. A typical example is borehole tomography which we have
already considered in the previous section in the context of algebraic
reconstruction. In this section we shall reconsider the same in the context of the
backpropagation algorithm [8]. But, first, let us look at a simpler system, a
circular array of transceivers of interest in medical imaging (see fig. 7.11).
Circular Array: The back scatter is lost in a linear array tomographic system;
naturally, some potentially useful information is also lost, in the sense that
only a part of the object spectrum lying within a disc of radius equal to
2k0

© 2001 by CRC Press LLC

 y
(a)

Forward
Scatter
χ0
χ
x

Backward
Scatter

Transmitter/ receiver
v

(b)
2k0 Direction of
illumination

χ0 0
45
u

Figure 7.11: (a) An experimental setup of transceivers for tomographic imaging. (b)
∆C(u, v) is now sampled over a circumference of a circle of radius k0 as shown. The
angle of incidence of the plane wave illumination is 45o.

is utilized. An alternate tomographic system consisting of a circular array of

sensors (transceivers) encircling the object is proposed [9]. In this configuration
both forward scatter and backward scatter are captured. This results in the
doubling of the area of the spectrum coverage, a disc of radius 2k0.

© 2001 by CRC Press LLC

 Arbitrarily shaped measurement boundaries were suggested in [10, 11]
whose authors have shown that on a straight line boundary or on a circular
boundary it is enough to measure either the diffracted field or its normal
derivative. For a completely arbitrary boundary, we need both types of
measurements [11]. A circular transducer array was used to illuminate an object
with a pulse (broadband) from different directions [12, 13], and the back scatter
alone, measured as a function of time, was employed for the purpose of object
reconstruction. A circular array for ultrasound holographic imaging was used by
Qin et al. [14] but they have approximated a circular array by a series of linear
arrays and then applied the backpropagation algorithm. A circular array of
transceivers was suggested for microwave diffraction tomography [15] where a
near field diffraction phenomenon was used. The object Fourier transform was
related to the scattered field through a two dimensional convolution relation.
The scattered field measured with a large circular array surrounding the object
(see fig. 7.11a) is proportional to the Fourier transform of the object profile
taken on the circumference of a circle of radius equal to the wave number and
centered at (-k0cos χ 0 , -k0sin χ 0 ) . This result is called here a Fourier
diffraction theorem (FDT) for a circular array [9].

Ps (R, χ, χ 0 ) =
π
k02 j ( k 0 R+ 4 ) 2 (7.15)
e ∆C(k0 (cos χ − cos χ 0 ), k0 (sin χ − sin χ 0 )
4 πk0 R

The left hand side is simply the observed scattered field on a large circle. The
right hand side is a Fourier transform of the object function which is evaluated
on a circle of radius k0 and centered at kx = −k0 cos χ 0 and ky = −k0 sin χ 0
(see fig. 7.11).
By changing the angle of incidence of the wavefront, χ 0 , it is
possible to cover the Fourier plane with a series of circles spanning a disc of
radius equal to 2k0 (see fig. 7.12). The increased coverage has been possible
because we captured the back scatter as well. Note the important difference is
that the scattered field measured with a circular array, being in the far field
region, directly yields the object Fourier transform. On the contrary, with a
linear array we need to Fourier transform the observed field. This important
result is first verified against the measured scattered field.
Verification of Fourier Diffraction Theorem: First we shall verify the FDT
through an example where an exact scattered field as well as its object Fourier
transform are known. Consider a liquid cylinder in water and assume that its
λ
refractive index is slightly above that of the water, δn ≤ where δn is
4a

© 2001 by CRC Press LLC

 v

2 k0

Figure 7.12: By changing the angle of incidence of the wavefront, χ 0 , it is possible

to cover the Fourier plane with a series of circles spanning a disc of radius equal to
2k0 .

change in the refractive index, a is radius of the cylinder and λ is wavelength of

illuminating wave (Born approximation). The scattered field due to a liquid
cylinder, when it is illuminated by a plane wave, was experimentally studied by
[16] and theoretically by [17]. We have computed the scattered field using the
derivation given in [17]. Next, we evaluate the object profile Fourier transform

~ J ( k a)
O( k 0 (s i n χ − s i n χ 0),k 0 (cos χ − cos χ 0 )) = 2 πδ n a 1
k
where

2 2
k= (s i n χ − s i n χ 0 ) + (cos χ − cos χ 0) k0

as a function of χ for a fixed direction of illumination; in the present case,

χ 0 = 0 o . The scattered field measured by a circular array is now compared with
the Fourier transform of the object profile evaluated on a circle of radius k0

© 2001 by CRC Press LLC

 −k −k
0 0
centered at ( 2
, 2
) (see fig. 7.13).
7.2.3 Bilinear Interpolation Algorithm: As the Fourier transform of sound
speed fluctuations is obtained from the scattered field either from a linear array
or circular array, in principle it is possible, by inverse Fourier transformation,
to estimate the sound speed fluctuations. In practice, however, this is not a
trivial step. As the object Fourier transform is measured over a series of arcs it
will be necessary to interpolate to the nearest square grid point and thus create a
discrete Fourier transform matrix for inversion. Alternatively, the back
propagation method originally developed for nondiffracting radiation may also
be used. We shall use the interpolation approach.
Consider a circular array of transceivers. The angle of illumination χ 0
is varied over 360o by switching on a sensor to transmit mode one at a time
and keeping the remaining sensors in the receive mode. The received field can
be expressed as a function of two parameters, namely, the angle of illumination
and the angular coordinate of each sensor, that is, (χ, χ 0 ) . We must then map
every point in (χ, χ 0 ) space onto (kx, ky) space. The reverse mapping, i.e.,
from (kx, ky) space onto (χ, χ 0 ) space, is, however, more convenient to use.
The to-and-fro mapping functions are as follows:

kx k
cos χ − cos χ 0 = , sin χ − sin χ 0 = y (7.16)
k0 k0

Solving the above equations for χ 0 , we get the following inverse mapping
functions:

 2   2 
 −ky − kx −1  ky − kx −1
−1  p  −1  p 
χ 0 = tan   χ = tan   (7.17)
 −kx + ky 2 2
−1  kx + ky −1
 p   p 

kx2 + ky2
where p = . Equations (7.16) and (7.17) together give a set of
2k02
transformation equations that can be used to map from the k-plane into the χ -
plane. Every point in the k-plane is mapped onto the χ -plane. The values
of (χ, χ 0 ) thus obtained may not correspond to any of those points where the

© 2001 by CRC Press LLC

 0.04

0.03
Field in arbitrary units

0.02

0.01

0.00

-0.01
0 90 180 270 360
Angle in degrees

F igure 7.13: The scattered field measured by a circular array is now compared with
the Fourier transform of a uniform cylindrical object evaluated on a circle centered at
−k0 −k0
( , ) as shown in fig. 7.11. The mean square error is 1.759x10-5.
2 2

scattered field is observed; then we must take recourse to some form of

interpolation. For example, bilinear interpolation is given by

O(χ, χ 0 ) = ∑ ∑ O(χ , χ )h (χ − χ )h (χ
i j
i j 1 i 2 0 − χj)

χ χ
where h1 (χ) = 1 −
, for χ ≤ ∆χ otherwise =0 and h2 (χ 0 ) = 1 − 0 ,
∆χ ∆χ 0
for χ 0 ≤ ∆χ 0 otherewise =0. Here ∆χ and ∆χ 0 are the sampling intervals.
Once the values of the Fourier transform are obtained over a rectangular grid in
(kx, ky) space, the inverse two dimensional Fourier transform can be computed
to obtain the object profile. The above algorithm is essentially an adaptation of
the frequency domain interpolation algorithm, which is known to be very fast
[1].
7.2.4 Imaging with a Circular Array: Since a circular array captures the entire
diffracted energy, that is, both forward and backward scattered energy, a greater
part of the object spectrum is utilized, indeed twice that of forward-scatter-only
(linear array) setup. Consequently, we expect a better resolution of small
inhomogeneities. To demonstrate this, we have c arried out the following

© 2001 by CRC Press LLC

 0.012
(a)
0.008

0.004
Relative refractive index

0.000

0.012
(b)

0.008

0.004

0.000
-16 -8 0 8 16
Distance in λ

F igure 7.14: An example of improved performance of the circular array over linear
array. A cylinder of radius 2λ with a small inhomogeneity of radius 0.25λ embedded
in it (see inset in (b)) is used as a test target. (a) Reconstruction using a linear array
and (b) using a circular array.(From [18] with permission.)

numerical experiment [18]. A small inhomogeneity of radius 0.25λ is

embedded in a larger cylinder of radius 2λ having a refractive index contrasts
with respect to the surrounding medium of 0.01 and 0.005, respectively (fig.
7.14b). The target is insonified with a narrowband plane wave radiation of
wavelength 1.0λ. A circular array of 64 transceivers is assumed. For
comparison we have also

© 2001 by CRC Press LLC

 .015
a b
.010

.005

.000
Relative refractive index

-.005

.015
c d
.010

.005

.000

-.005
-4 -2 0 2 4 -4 -2 0 2 4

Distance in λ

Figure 7.15: A comparison of performance of a linear array of finite size with that of
a circular array. The number of illuminations in all cases was 64. (From [18] with
permission.)

considered a linear array of the same length and one transmitter located on the
broad side, and the scattered field was calculated using the object Fourier
transform over semicircular arcs. For a circular array, however, the scattered
field was computed using the exact solution given in [16]. The reconstruction
(a central slice) of the target is shown in fig. 7.14. The reconstruction obtained
using a linear array is shown in fig. (7.14a) and that obtained using a circular
array is shown in fig. (7.14b). Clearly the circular array outperforms the
equivalent linear array as the small inhomogeneity is more accurately located.
Next, we would like to emphasize the role of the array size on object (fig. 7.15)

© 2001 by CRC Press LLC

 ky ky
(a) Direction of (b)
illumination

χ 0
0 45
kx

Circle of radius k 0

Figure 7.16: (a) For narrowband the scattered field is proportional to the object
Fourier transform evaluated on a circle. (b) For finite band, the scattered field is
proportional to the object Fourier transform evaluated inside a crescent shaped
region. (From [18] with permission.)

reconstruction. When using a linear array, it is necessary that the array output
be Fourier transformed before it is used for reconstruction. Consequently, the
errors in the Fourier transformation due to finite size of the array will degrade
the reconstruction. This effect is demonstrated in fig. 7.15. The first three
figures (7.15a, b, c) were obtained using a linear array of three different sizes,
namely, 64, 128, and 512 receivers spaced at 2λ and 100 λ away from the
object, and a cylinder of radius 1λ with a refractive index contrast of 0.01. The
scattered field was computed using the exact solution given in [16]. The
reconstruction shown in fig. (7.15d) was obtained using a 64 element circular
array (radius=100λ). The reconstruction obtained with the circular array is
superior to that obtained with a linear array of a much larger size (512
receivers). Notice that the side lobes have practically disappeared .

§7.3 Broadband Illumination: The object of interest is illuminated with a

broadband signal which results into a better coverage of the spatial spectrum of
the object with fewer illuminations. We shall in this section examine the
effectiveness of broadband illumination.
7.3.1 Spectrum Coverage: If the object is illuminated from one direction with a
set of different frequencies covering a spectral band, the scattered field will
correspond to a set of circular arcs covering a crescent shaped region in the
object Fourier plane, as shown in fig. 7.16b. The radii of the inner and outer
circles forming the crescent are related to the lower and upper cut-off
frequencies. Let f l and f u be the lower and upper cut-off frequencies

© 2001 by CRC Press LLC

 2πf l 2πf u
respectively; the radii of the circles are and , respectively. By
c0 c0
suitably selecting the lower or upper cut-off frequencies it is possible to
emphasize the low or high frequency spatial spectrum of the object. With a
single illumination it is thus possible to cover a fraction r of the disc of
maximum size of radius 2 f u , where

f u2 − f l2
r= (7.18)
4 f u2

Note that r ≤ 0.25 , where the maximum is achieved when f l = 0 . To cover

the entire disc we need four or more illuminations, but certainly far fewer than
the number of illuminations for narrowband illumination. With more than two
illuminations certain regions are likely to be covered more than once; thus the
fraction of covered area will be less than that predicted by (7.18). For example,
for four illuminations the area covered is shown in fig. 7.17 where we have
assumed that f l = 1.5 kHz and f u = 3. 0 kHz . The fraction of the Fourier plane
covered is 75%. It is further possible to increase this fraction, in particular, by
decreasing the lower cut-off frequency. It is straightforward to analytically
compute the area covered by one or more crescents, although it becomes
difficult with increasing number of illuminations. In table 7.2 we list the
fraction of area covered as a function of lower cut-off but keep the upper cut-off
frequency fixed at 3 kHz. Note that one hundred percent coverage is not possible
even when f l = 0 .
Further, to demonstrate the effect of the lower cut-off, we have carried
out computer reconstruction using four illuminations and 64 sensors (circular
array). The results are shown in fig. 7.18. There is an overall improvement in
the reconstruction with decreasing lower cut-off frequency. Evidently, increasing
the number of illuminations will also increase the coverage of the Fourier
plane. For example, in table 7.3 we show how the coverage increases as we
increase the number of illuminations. Here the lower and upper cut-off
frequencies are held fixed at 250 and 3000 Hz respectively. Four to eight
illuminations seem to be ideal as not much is gained by going beyond eight
illuminations, which gives almost 95% coverage.
7.3.2 Signal Processing Issues: The major issue in tomographic imaging from
the point of signal processing relates to the fact that the object Fourier
transform is sampled on a nonstandard grid, such as polar rastor in
nondiffracting tomography or circular arcs in diffraction tomography. All
existing methods of reconstruction require interpolation to convert from
nonstandard grid to standard square grid. Another issue relates to the fact that in
any practical implementation of the tomographic imaging scheme only a

© 2001 by CRC Press LLC

Figure 7.17 : Broadband illumination provides a better coverage of the object
Fourier transform. Just with four illuminations it is possible to get a 75% coverage.
The lower and upper cut-off frequencies are 1.5 and 3.0 kHz, respectively.(From [18]
with permission.)

Lower cutoff Area covered

frequency(Hz) %
2000 34.35
1500 75.00
750 80.72
500 81.37
250 81.71
0 82.00
Table 7.2: The fraction of the disc area covered with a broadband signal whose lower
cutoff is varied and upper cutoff is held fixed (3.0kHz). We have assumed four
illuminations.(Source [18] with permission.)

limited number of views, often covering a finite angular interval, are likely to
be available, leaving large gaps in the Fourier plane. As shown earlier,
broadband illumination can help to reduce the gaps. Signal processing tools
have also been suggested for extrapolation of the observed Fourier transform
into the missing gaps. A priori information about the object, such as a limit
on the support of the object function either in space or frequency domain, does

© 2001 by CRC Press LLC

 0.020

a
0.010

0.000

-0.010

0.010
Relative refractive index

b
0.006

0.002

-0.002

0.015
c
0.010

0.005

0.000

-0.005
-16 -8 0 8 16
Distance in λ

Figure 7.18: The effect of the lower cut-off frequency on the reconstruction. (a)
1500Hz, (b) 500Hz and (c) 250Hz. The upper cut-off frequency is 3000 Hz. (From
[18] with permission.)

not help to uniquely reconstruct the object function from a limited number of
samples on an algebraic contour in the Fourier plane [19]. Extrapolation
outside the frequency domain, where the observations are available, has been
attempted using the principle of maximum entropy [21, 22] which is familiar
to the signal processing community as it is extensively used to extrapolate the
covariance function for spectrum estimation [20].

© 2001 by CRC Press LLC

 Number of Area covered
illuminations %
1 24.83
2 49.66
4 81.71
8 94.9
12 97.2
Table 7.3: Fraction of the disc area covered with a broadband signal (250-3000Hz)
for a different number of illuminations. (Source [18] with permission.)

7.3.3 Cross Borehole Tomography: In chapter 1 we have derived a relationship

between the scattered wavefield from a target which has been illuminated by a
source array positioned in one borehole and receiver array in another borehole
(pp. 65-68). Consider a two dimensional target for which the relationship is
given by (1.91) which we reproduce here for convenience,

u u u u
F1 (u1 , u2 ) = 2k02 ∆c̃( 1 + 2 , k02 − ( 1 )2 − k02 − ( 2 )2 )
d d d d
u u
− j k02 − ( 1 )2 La − j k02 − ( 2 )2 Lb (1.91)
e d e d
×
u u
k02 − ( 1 )2 k02 − ( 2 )2
d d

λ0 2u 2u
Let d = , u1′ = 1 , and u2′ = 2 . For π ≤ u1 ,u2 ≤ −π it turns out that
2 λ0 λ0
k0 ≤ u1′, u2′ ≤ −k0 . A point in the Fourier plane, (u1′, u2′ ) , would correspond
to a point (u,v) in the Fourier plane of ∆c̃ where

u = u1′ + u2′

{ }
(7.19)
v=± k02 − (u1′)2 − k02 − (u2′ )2

Eliminating u2′ from two equations in (7.19) we obtain

[v ± ] + (u − u′)
2
k02 − (u1′)2 1
2
= k02 (7.20)

For a given value of u1′ equation (7.20) describes a circle with radius k0 and

centered at (u1′, ± k02 − (u1′)2 ) , for example; for u1′ = 0 the two circles are

© 2001 by CRC Press LLC

 v
A

A'

- k0 o k0 u

B'

B (a)

A'

B'

B
(b)

Figure 7.19: (a) The object Fourier transform lying inside two circular disks is
scanned along a series of semicircular arcs. (b) As an example, consider four plane
wavefronts A, A´, B, B´ emitted by distant sources in the source array.
Corresponding circular arcs are shown in (a). The object Fourier transform is
scanned on semicircle AO by wavefront A and on OB by wavefront B. Wavefronts A´
and B´scan the object Fourier transform on A´OB´.

© 2001 by CRC Press LLC

 v u'2 = 0

O'

A k0

u
- k0 o k0
O''

u'2 = 0

Figure 7.20: For a fixed u1′ , the object Fourier transform is scanned over a circle A.
For different values of u2′ , that is, for different directions of the beam, a series of
circles will intersect the circle A over an arc O´´OO´. The object Fourier transform
will be sensed over this arc only.

centered on the y-axis at ±k0 . In fig. 7.19 we show different circles (arcs) for
different values of u1′ ; in particular, the thick arcs are for u1′ = 0 and
k
u1′ = −k0 and the thin arcs are for u1′ = ± 0 . Similarly, by eliminating u1′
2
from (7.19) we obtain

[ ] + (u − u′ )
2
v ± k02 − (u2′ )2 2
2
= k02 (7.21)

which describes a circle with radius k0 and centered at (u2′ , ± k02 − (u2′ )2 ) .
For a fixed u2′ the object Fourier transform is scanned over a circle, for
example, A in fig. 7.20.
Let the receiving array (ULA) be beamed to receive the wavefield in
some particular direction, that is, for some fixed u2′ . For u2′ = 0 the circle
described by (7.21) will intersect circle A at two points, namely, O ′ and O (see
fig. 7.20). The object Fourier transform is sensed only at these points of
intersection. For different values of u2′ , that is, for different directions of the
beam, we can draw a series of circles which will intersect the circle A over an

© 2001 by CRC Press LLC

 arc O ′′OO ′ . Some of these circles are shown in fig. 7.20. There will be as
many arcs as the number of sources. In the ideal situation of an infinitely long
source array the arcs over which the object Fourier transform is sensed will fill
k0
two circles shown in fig. 7.19. A few arcs corresponding to u2′ = ±k0 , ±
2
are also shown in this figure. In summary, in cross borehole tomography the
object Fourier transform can be scanned over a pair of circles along a series of
semicircular arcs as shown in fig. 7.19.
7.3.4 Maximum Entropy Extrapolation: The object Fourier transform scanned
in cross borehole tomography (see fig. 7.19) skips a good portion of the
Fourier plane, in particular along the u-axis. Consider a square, 4k0 × 4k0 ,
superscribing the two circles. It is found that just about 39% of the square is
included within the two circles. Consequently, the resolution particularly along
the x-axis will be considerably deteriorated. This is true only for infinitely long
source and sensor arrays. Additionally, since in any borehole tomographic
experiment the array size is bound to be finite, there will be a further reduction
of the scanned area [21]. To overcome the effect of undercoverage of the object
Fourier transform it has been suggested to extrapolate the measured Fourier
transform into an area where no measurements were possible, using the
principle of maximum entropy [21, 22]. We shall briefly review this approach.
The aim of tomographic imaging is to determine the wave speed
function 1 + δc̃(x, y) . We have already shown how from the scattered field we
can obtain the Fourier transform of the wave speed variations over a limited
domain (fig. 7.19), hereafter referred to as Ω , which is completely determined
by the geometry of experiment. Nothing is however known of the wave speed
outside this domain. Of the many possible functions we choose one, which is
consistent with the observed Fourier transform in the specified domain and is
maximally noncommittal with regard to unavailable data. This is the principle
of maximum entropy founded by Burg [23] and widely used in spectrum
analysis [20]. The constrained optimization problem may be formulated as
follows:
minimize:

H = ∫ ∫ (1 + δc̃(x, y)) ln(1 + δc̃(x, y))dxdy (7.22)

Γ
subject to
∆c̃(u,v) = ∆c̃(u,v) measured in Ω (u,v) ∈Ω (7.23)

∂H = ∫ ∫ ∂δc̃(x, y)[1 + ln(1 + δc̃(x, y))]dxdy

Γ

© 2001 by CRC Press LLC

 k0 k0

∫ ∫ ∂∆c̃(u,v)∫ ∫ [1 + ln(1 + δc̃(x, y))]e

1 j (ux + vy)
= 2 dxdydudv
4π − k0 − k0 Γ

H will be minimized when ∂H = 0 . Since ∆c̃(u,v) is already specified in

(u,v) ∈Ω , ∂∆c̃(u,v) = 0 in Ω . Therefore, maximization of entropy
requires

∫ ∫ [1 + ln(1 + δ c̃(x, y))]e

j (ux + vy)
dxdy = 0 in (u,v) ∉Ω (7.24)
Γ

The solution is obtained by requiring to alternatively satisfy (7.23) and (7.24).

The algorithm for the constrained optimization has the following steps:

(1) Compute q(u,v) = ∫ ∫ [1 + ln(1 + δc̃(x, y))]e j (ux + vy) dxdy

Γ
(2) Set q(u,v)=0 for (u,v) ∉Ω
(3) Compute FT {δc̃(x, y) = e IFT { q(u,v)} −1
−1 }
(4) Set∆c̃(u,v) = ∆c̃(u,v) measured for (u,v) ∈Ω
(5) Compute δc̃(x, y) = IFT {∆c̃(u,v)} and go to step 1

The procedure is terminated when the reconstructed function meets some

criterion of convergence. The algorithm is demonstrated through an example. A
square object of size (5x5) with wave speed c (=1+0.1) is embedded in a
background with wave speed, c0=1.0. The Fourier transform of the object was
computed over a grid of 64x64. Let us assume that cross borehole geometry
permits measuring the object Fourier transform over a figure of 8 (see fig.
7.19a) where the radius of the pass disc is four. The filtered object Fourier
transform was then used in the maximum entropy reconstruction algorithm
described above. The reconstructed object (a horizontal cross section) after four
iterations is shown in fig. 7.21, but it did not change much even after 40
iterations. Much of the spectrum lost during filtering remains unrecoverable
except what lies between the upper and lower discs.

§7.4 Reflection Tomography:

The wave field returned by an object may be considered either as a reflection at
the surface of discontinuity in physical parameters (wave speed, density, etc.) or
backscattering from deep inside the object due to variable physical parameters.
In this section we shall deal with the latter situation and reserve the former to

© 2001 by CRC Press LLC

 0.06

0.05

0.04

0.03

0.02

0.01

-0.01
0 10 20 30 40 50 60 70

Figure 7.21: Maximum entropy reconstruction, dashed line (--) filtered object and
solid line (__) maximum entropy reconstruction. There is only a marginal
improvement in the maximum entropy reconstructed object.

be dealt with in chapter 8. The backscatter from an inhomogeneous object may

be related to the inhomogeneity inside the object under the usual assumption of
weak scattering (Born approximation). Indeed, the backscattering coefficient in
frequency domain may be easily derived from (7.15) by setting
χ=−(180−χ 0 ) . We obtain

π
k 2 j ( k 0 R+ 4 ) 2
Ps (R, χ 0 ) = 0 e ∆C (−2k0 cos χ 0 , −2k0 sin χ 0 ) (7.25)
4 πk0 R

where R now stands for distance to the center of the object whose size is
assumed to be much smaller than R, so that the far field approximation hold
good. The experimental setup is shown in fig. 7.22a. The backscatter
coefficient at a fixed frequency, Ps (R, χ 0 ) , is proportional to the object
Fourier transform at spatial frequency (−2k0 cos χ 0 , −2k0 sin χ 0 ) where χ 0
is the angle of illumination (see fig. 7.22b). If we now use a broadband signal
for illuminating the object (keeping χ 0 fixed) and Fourier decompose the

© 2001 by CRC Press LLC

 v

χ0
χ0

u
x
Backscatter

Transceiver

(a) (b)

F igure 7.22: (a) A single transceiver is used to illuminate an object and receive the
backscatter. A plane wavefront is assumed to be incident at an angle χ 0 . (b)
Backscatter coefficients are proportional to the object Fourier transform over a radial
line as shown.

received signal we shall obtain the object Fourier transform over a radial line at
an angle χ 0 (see fig. 7.22b). By illuminating the object repeatedly over 360o,
either by physically taking the transceiver around the object or rotating the
object around its axis, we can cover the entire Fourier plane. This commonly
used experimental setup was suggested by [13]. The above result is akin to the
Fourier slice theorem of nondiffracting tomography (§7.1). Naturally, many of
the reconstruction methods developed for transmission (nondiffracting)
tomography, in particular, backprojection, backpropagation and interpolation
methods, can be used in the present case of reflection tomography.
Additionally, there is further similarity between the transmission tomography
and reflection tomography. As noted in §7.1 the projection of an object is equal
to the line integral of some physical quantity (e.g., absorption in X-ray
tomography) over the ray path. A similar physical insight can be given to
reflection tomography.
7.4.1 Line Integral: Let a broadband plane wavefront be incident on a scattering
object. A series of echoes will be emitted as the wavefront penetrates the
object. At any time instant the receiver will receive echoes from all scattering
elements which lie on a surface (see fig. 7.23). Let f(x,y) be the reflectivity
function

ps (ρ, χ 0 ) = ∫ f (x, y)ds

s

© 2001 by CRC Press LLC

 ∞ 2π

= ∫ ∫ f (r,θ)δ( r 2 + R2 + 2rRcos(θ − χ 0 ) − ρ)rdrdθ (7.26)

0 0

Different variables appearing in (7.26) are illustrated in fig. 7.24 The limits on
integrals suggest that the object is of infinite size. But in practice the object is
finite and the transceiver is placed outside the object. To overcome this
difficulty we shall assume that the reflectivity is zero outside the domain of the
object.
For fixed χ 0 we have a waveform which is a function of time or ρ
ρ
(=ct). We then compute a Fourier transform of the waveform ps ( , χ 0 ) . After
c
simplification we obtain

Ps (k, χ 0 ) =
∞ 2π
(7.27)
∫ ∫ f (r,θ)exp(− jk(
0 0
r 2 + R2 + 2rRcos(θ − χ 0 ))rdrdθ

Under the assumption R>>r the exponent in (7.27) may be expanded in

bionomial series. Retaining the first two terms in the binomial expansion

r 2 + R2 + 2rRcos(θ − χ 0 )
r2
≈R+r cos((θ−χ 0 )+ (1−cos2 (θ−χ 0 )))
2R
we obtain

∞ 2π r2
− jk ( (1− cos 2 (θ− χ 0 )))
∫ ∫ f (r,θ)e
− jkR − jkr cos(θ− χ 0 )
Ps (k, χ 0 ) = e 2R
e rdrdθ
0 0
(7.28)
It is easy to see that

r cos(θ − χ 0 ) = x cos(χ 0 ) + ysin(χ 0 )

and

r 2 (1 − cos2 (θ − χ 0 )) = (x sin(χ 0 ) − y cos(χ 0 ))2

Substituting in (7.28) we obtain

© 2001 by CRC Press LLC

 time
s

Object

Transceiver

Figure 7.23: At any time instant the receiver will receive echoes from all scattering
elements which lie on a surface, s.

s
(r, θ)
θ
x
ρ
R
Transceiver

Figure 7.24: Symbols used in (7.26) are explained in the figure above.

Ps (k, χ 0 ) = e − jωR ×
∞ ∞ ( x sin( χ 0 )− y cos( χ 0 )) 2
− jk (7.29a)
∫∫ e − jk ( x cos( χ 0 )+ y sin( χ 0 )) dxdy
2R
f (x, y)e
−∞ −∞

If we assume that the object size is much smaller than R, the middle term in
(7.29a) may be set to 1, leading to a simple Fourier transform relation,

Ps (k, χ 0 ) = F(k cos(χ 0 ), k sin(χ 0 )) (7.29b)

© 2001 by CRC Press LLC

 where F(.) is the Fourier transform of f(x,y). When this assumption is not
acceptable the presence of a middle term can be accounted for through an
iterative procedure described by [24].
7.4.2 Spherical Wavefronts: Plane wave illumination is not always practical.
Often point sources at a finite distance are used giving rise to spherical
wavefronts penetrating a scattering object. While the basic results derived in
relation to a simplified plane wave model hold good, the spherical wavefront
model, though mathematically more complex, yields more accurate results. We
shall derive the exact result in one case where the transceivers are placed on a
horizontal plane above the scattering object. Consider a uniform two
dimensional array of transceivers and a three dimensional object with arbitrary
speed variation (see fig. 7.25). The transceivers are sequentially fired (in any
order) and all returned signals are arranged with common zero time, that is, as if
all transceivers were fired at the same time instant. Interestingly in this setup it
is practical to think of using a single transceiver and move it from place to
place. Consider a weakly inhomogeneous (speed fluctuations only) object
illuminated by a point source with the scattered field being received by a sensor
kept close to the source. From every scattering volume element a backscatter
radiation reaches the detector as modeled. The first order back scatter may be
derived from equation (1.77), which we reproduce here for convenience,

+∞
1 2k02δc̃e j ( k 0 r − r ′ )
4π ∫− ∫∞ ∫
f 1 (r, t) = f 0 (r' , t)dx ′dy ′dz ′
r − r′

where f 0 (r' , t) is an illuminating wavefield, which, for a point source

emitting a sinusoid, is given by

j (k 0 r s − r ′ )
e
f 0 (r' , t) = e − jω 0 t
r s − r′

where r s is source position vector. Since the source and detector are at the
same location, r = r s . Using the point source illumination expression given
above in (1.77) we obtain

+∞
1 − jω 0 t 2k02δ c̃e j 2( k 0 r − r ′ )
f 1 (r, t) =
4π
e ∫− ∫∞ ∫ r − r′ 2 dx ′dy′dz ′ (7.30a)

The scattered field on a z=0 surface in rectangular coordinates may be written as

© 2001 by CRC Press LLC

 r-r m,n

Object

Figure 7.25: Uniform planar array of transceivers above a three dimensional object.

f 1 (x,y,z=0, ω 0 )=
1

+∞
2k δc̃( x ′, y ′, z ′ )e
2 [
j 2( k 0 ( x − x ′ ) 2 +( y − y ′ ) 2 +( z − z ′ ) 2 ] 2 )
(7.30b)
1
∫ ∫
4π − ∞
∫ 0

[(x − x ′)2 + (y − y′)2 + ( z′)2 ]

dx ′dy ′dz ′

which we shall express in a form that enables us to use a result (1.78), derived
in chapter 1 on page 61,

∂  f 1 (x, y, z = 0, ω 0 ) 
∂ω 0  =
ω 20 
1
(7.31)
j
+∞
δc̃( x ′, y ′, z ′ )e [
j 2( k 0 ( x − x ′ ) 2 +( y − y ′ ) 2 +( z − z ′ ) 2 ] 2 )

πc 3 ∫∫∫ 1 dx ′dy ′dz ′

−∞ [(x − x ′) 2
+ (y − y ′ ) 2
+ ( z ′ )2 2 ]
Now using the result (1.78) in (7.31) we go into the frequency domain,

∂  f 1 (x, y, z = 0, ω 0 ) 
 =
∂ω 0  ω 20 
1 + ∞ ∆c̃(u,v, 4k0 − u − v ) j (ux + vy)
2 2 2

∫∫ e dudv (7.32)
2π 2 c 3 − ∞ 4k02 − u 2 − v 2

© 2001 by CRC Press LLC

 From (7.32) we can get the Fourier transform of the speed fluctuations in terms
of the Fourier transform of the wavefield observed on the surface,

∆c̃(u,v, 4k02 − u 2 − v 2 ) =
c3 ∂  F1 (u,v, ω 0 ) 
4k02 − u 2 − v 2   (7.33)
2 ∂ω 0  ω 20 

Thus, the Fourier transform of the speed variations is derived from the Fourier
transform of the backscatter measured on a plane surface. It is interesting to
observe that the Fourier transform thus computed actually corresponds to the
Fourier transform of the object on a sphere centered at the origin and with
radius equal to 2k0 (see fig. 7.26). A broadband signal will be necessary to
cover the entire Fourier transform of the speed variation function.

§7.5 Object Shape Estimation:

If the boundary of an object is piecewise linear, the corner points are sufficient
for pattern recognition, image compression and coding, shape analysis, etc.,
[25]. The corner detection algorithms work on spatial image data in the form of
a photograph. The sensor arrays are used for corner detection from the scattered
wavefield (acoustic or electromagnetic). When an object, whose refractive index
is slightly different with respect to that of the surrounding medium, is
illuminated with a plane wave the scattered field measured around the object is
proportional to the Fourier transform of the object. Thus, the shape
information is buried in the scattered field. It is of some interest in medical
diagnosis, in subsurface imaging, and in nondestructive testing to be able to
recognize the shape of the buried object from the scattered acoustic or
electromagnetic field, particularly when only a few limited views are permitted.
We shall show that when the object is binary, convex and having a
nondegenerate polygonal cross section, the scattered field is a sum of sinusoids,
a function of wave number and corners of the polygon. The object is
illuminated with a broadband plane wave and the scattered field is measured as a
function of wavenumber. The frequencies of the sinusoids are estimated from
the scattered field using an algorithm described in [20].
It is shown in §7.2 that the scattered field measured by a circular array
is proportional to the 2D Fourier transform of the object profile taken on the
circumference of a circle of radius equal to the wave number and centered at
(−k0 cosχ 0 ,−k0 sin χ 0 ) where χ 0 is angle illumination (see fig. 7.11). By
changing the direction of illumination (0 to 360o) the object Fourier transform
is scanned over a disk of radius k0. When the object is binary (i.e., refractive
index is constant throughout the object) the interest is in the shape of the
object. The shape information may be directly obtained from the scattered field

© 2001 by CRC Press LLC

 w

Figure 7.26: The Fourier transform of a reflected signal (echo) corresponds to the
Fourier transform of the object on a sphere centered at the origin and radius equal to
2k 0 .

or 2D Fourier transform of the object. This approach was taken by Milanfar and
co-workers [26] in the context of ray tomography where the input data are
projections of object.
7.5.1 Fourier Transform of Binary Convex Polygonal Object: Consider the
evaluation of the 2D Fourier transform over a p-sided binary and convex
polygonal domain (see fig. 7.27). Take any point inside the polygon and join it
to all corners forming p triangles which lie entirely inside the polygon and
make this as the origin of the coordinate system

∫∫ e
j (ux + vy)
F(u,v) = dxdy
over polygon
p (7.34)
=∑ ∫∫ e
j (ux + vy)
dxdy
n =1 over n th triangle

To evaluate the integral over nth triangle refer to fig. 7.27b where we show the
integration along a narrow strip under the rotated coordinate system such that
the new the x-axis is perpendicular to nth side. Note that this is valid only for
convex objects (for nonconvex objects, it's not possible to drop a perpendicular
from the origin to at least one edge, such that the perpendicular lies entirely
within the object) that are nondegenerate. The triangle is then covered by a
series of strips. Equation (7.34) reduces to

 − j((u cos θ n + v sin θ n ) x ′ 

ρ x ′ tanφ  
p n 1n  +(v cos θ n − u sin θ n ) y ′ ) 
F(u, v) = ∑ ∫   dx ′dy ′
∫ e (7.35)
n = 1 0 − x ′ tanφ
2n

© 2001 by CRC Press LLC

 where x ′ =(x cosθ k +ysinθ k ) and y ′ =(ycosθ k −xsinθ k ) . Evaluate the
integral in (7.35) first with respect to y′ followed by integration with respect
to x ′ . We obtain
[e − j ( u ′ − v ′ tan φ1n )ρ n − 1] [e − j ( u ′ + v ′ tan φ 2n )ρ n − 1]
p
F(u,v) = ∑ − (7.36)
n =1 v ′(u ′ − v ′ tan φ1n ) v ′(u′ + v ′ tan φ 2n )

where u′ = (u cos θ n + vsin θ n ) and v ′ = (v cos θ n − usin θ n ) . We shall

now rewrite (7.36) by replacing θ n , φ1n , and φ 2n in terms of the
coordinates of the two corners corresponding to the nth side, namely,
(an , bn ) and (an+1 , bn+1 ) . The following relations are used for this
purpose:

ρ = a cosθ + b sin θ = a cosθ + b sin θ

n n n n n n +1 n n +1 n

a = ρ (cosθ + sin θ tan φ )

n n n n 1n

b = ρ (sin θ − cosθ tan φ )

n n n n 1n

a = ρ (cosθ − sin θ tan φ )

n +1 n n n 2n

b = ρ (sin θ + cosθ tan φ )

n +1 n n n 2n
We obtain

 [e − j (uan + vbn ) − 1] [e − j (uan+1 + vbn+1 ) − 1] 

p
F(u,v) = ∑ ρn  −
v ′(uan +1 + vbn +1 ) 
(7.37)
n =1  v ′(uan + vbn )

Our goal is to determine(an , bn ) and (an+1 , bn+1 ) from (7.37). This may
be achieved by expressing (7.37) on the k y = 0 and k x = 0 axes. We get
the following equations

© 2001 by CRC Press LLC

 y
(ak+1 , b k+1)

th
k triangle x'
y'
ρ
k
φ1k
θk ( a k, b k )
φ
2k
x

(a) Object cross section th

(b ) k triangle

Figure 7.27: To evaluate the Fourier transform of a polygonal object we consider

each triangle (From [27] with permission from IEEE ( c 1998 IEEE).)

e − juan − 1 e − juan+1 − 1
p
u F(u,v = 0) = − ∑ ρn {
2
− }
n =1 an sin θ n an +1 sin θ n
(7.38)
e − jvbn − 1 e − jvbn+1 − 1
p
v 2 F(u = 0,v) = ∑ ρn { − }
n =1 bn sin θ n bn +1 sin θ n

The above equations may be solved by modeling them as a sum of sinusoids

and using the well known Prony’s algorithm or its more modern versions [20].
From the coefficients in the exponents of the complex sinusoids we obtain
(an , an +1 ) and (bn ,bn +1 ) but we are yet to pair them, that is, select the right
x and y coordinate pair which will form a valid corner. We note that
F(u, v = 0)and F(u = 0, v) represent backscatter due to a broadband
illumination along the x- and y-axes, respectively (see (7.22b)).
7.5.2 Pairing Algorithm: In the previous section we saw how to obtain the x-
and y-coordinates of the corners of the polygon. This alone will not suffice to
define a unique convex polygon. We need some additional information on how
to pair a given x-coordinate with the right y-coordinate from the list of
estimated y-coordinates. This problem is resolved by using an additional
illumination at an angle θ

k 2 F(k cosθ, k sin θ) =

© 2001 by CRC Press LLC

  − jk(an cosθ + bn sin θ) 
[e − 1] 
p  (a cosθ + b sin θ) 
 n n 
∑ Γn   (7.39)
n = 1  [e − jk(an + 1 cosθ + bn + 1 sin θ) − 1] 
− 
 (a cosθ + b sin θ) 
n +1 n +1

ρn
where Γ n = . From the back scatter due to an illumination at angle
sin(θ − θ n )
π
θ (≠ 0 or ) , we can estimate as described in [20, 27] the coefficients in the
2
exponents of the complex sinusoids. Thus, we get the additional information in
the form of linear combination of the x- and y-coordinates of the corners,
(an cos θ + bn sin θ ) . . n = 1,2, ... p . The steps in the pairing algorithm are
as below:
1) Generate a list of x-coordinates, y-coordinates and the linear combination of
the x- and y-coordinates. It is presumed that the list is not in the same order as
the indexed corners.
2) Take the first element from the x-coordinate list and any one element from
the y-coordinate list and form a linear combination, (a1 cosθ + bn sin θ)
n = 1,2, ... p
3) Compare the result of the linear combination with those estimated with
π
θ (≠ 0 or ) illumination. The best match (within the limits of estimation
2
error) will indicate the correct choice of bn .
4) Take the next element from the x-coordinate list and go to step (2).
For the purpose of illustration we consider a square object of size (6m,6m),
rotated by 30 deg. and shifted away from the origin by (5m,5m). It is
illuminated from three directions, 0, π and π/6. The x- and y-coordinates got
from the noiseless scattered field in the first two directions and their linear
combination (θ=30 deg) are shown in table 7.4 and those estimated from the
scattered field got in the third direction are shown in table 7.5. The application
of the pairing algorithm is illustrated in table 7.6. The best match with the
estimated coefficients is shown in column three in bold figures and the
corresponding y-coordinate is shown in the last column.
We may encounter the problem of repeated x- or y-coordinates or their
projections. The projections of two corners may overlap or come very close to
each other depending upon the orientation of the object. As shown in fig. 7.28,
for a square object depending upon the orientation, the adjacent projections
(e.g., x1 and x2) may come close to each other or overlap. The problem of

© 2001 by CRC Press LLC

 x y x cos θ + y sin θ

9.0981 6.0981 10.9282

3.9019 9.0981 7.9282
0.9019 3.9019 2.7321
6.0981 0.9019 5.7321
Table 7.4: The x- and y-coordinates and their linear combination (θ=30 deg) are
shown in the above table.(Source: [27] with permission from IEEE ( c 1998 IEEE).)

x 9.0981, 6.0981, 3.9019, 0.9019

y 9.0981, 6.0981, 3.9019, 0.9019
x cos θ + y sin θ 10.9282, 5.7321, 7.9282, 2.7321

Table 7.5: The estimated projections from the scattered field (noise free) are shown
(Source: [27] with permission from IEEE ( c 1998 IEEE).)

x y x cos θ + y sin θ Best match for

x
9.0981 9.0981 12.4282 6.0981
6.0981 10.9282
3.9019 9.8302
0.9019 8.3301
6.0981 9.0981 9.8302 0.9019
6.0981 8.3301
3.9019 7.2321
0.9019 5.7321
3.9019 9.0981 7.9282 9.0981
6.0981 6.4290
3.9019 5.3301
0.9019 3.8302
0.9019 9.0981 5.3301 3.9019
6.0981 3.8302
3.9019 2.7321
0.9019 0.5000
Table 7.6: A numerical illustration of the pairing algorithm. Q=30 deg.(Source: [27]
with permission from IEEE ( c 1998 IEEE).)

repeated projection can be resolved by selecting another direction of

illumination whenever the number of sinusoids estimated in x, y differs. The

© 2001 by CRC Press LLC

 y
4

y
3

y
2

y
1

x x x x
1 2 3 4

Figure 7.28: A square object and the projections of its corners on the x- and y-axes.
Note that, depending upon the orientation, the adjacent projections (e.g., x1 and
x 2 ) may come close to each other(From [27] with permission from IEEE ( c 1998
IEEE).).

number of sinusoids observed in all three illuminations must be equal. In

practice it may be necessary to illuminate an unknown object along several
directions and estimate the sinusoids along each direction. From this set choose
three directions, preferably two orthogonal directions, having an equal number
of sinusoids. The number of sinusoids which may be determined from a finite
data in the presence of noise is indeed a complex problem and hence it is
outside the scope of this book. The reader may like to refer to a book, for
example [20], or current literature on this topic.
When there are two or more objects, as the center of coordinates would
lie outside all but one object, it is necessary to modify (7.31), which was
derived under the assumption that the center lies inside the object. Also, there
may be some ambiguity in the process of constructing the object shape even
when all x- and y-coordinates are correctly paired. The amplitude of the sinusoid
corresponding to a corner can then be used to resolve such an ambiguity in
addition to the fact that the objects are convex and the number of objects is
known (see [27] for more details).
7.5.3 Performance Analysis: The performance of the object shape estimation
procedure has been investigated through numerical experiments [27]. For this
we have considered a square object (refractive index contrast equal to 0.01) of
2
size 8x8m and rotated by 9 degrees with respect to the x-axis (see fig. 7.28).
The object was illuminated with a broadband plane wavefront whose wave

© 2001 by CRC Press LLC

 0.10

0.08

0.06
Amplitude

0.04

0.02

0.00
0.0 0.4 0.8 1.2 1.6 2.0 2.4 2.8 3.2
Wavenumber
Figure 7.29: Computed backscatter from a weakly scattering square object (shown in
fig 7.28) illuminated by a broadband signal.(From [27] with permission from IEEE
( c 1998 IEEE).)

π π
number is varied from to π in steps of along three directions,
64 64
namely, x-axis, y-axis and a radial direction at an angle θ =30 degrees. The
backscatter at each wavenumber was computed using the Fourier transform
approach described in [1]. A typical example of backscatter caused by x-axis
illumination is shown in fig. 7.29. To this scattered field sufficient white
Gaussian noise was added so that the snr became equal to a specified figure.
Here the snr is defined as ten times the logarithm (base 10) of the ratio of the
average scattered energy to noise variance. The corners and also the Tn ' were
estimated using the procedure given in [27]. The mean square error (MSE) in
the estimation of coordinates of the corners was studied as a function of snr.
The results, obtained by averaging over fifty independent experiments, are
shown in fig. 7.30a. Notice that MSE rises very rapidly for snr below 8 dB.
This is largely on account of the fact that the projections of two adjacent
corners (e.g., x1 and x4, and x2 and x3 in fig. 7.29) are close to each other; in
this example they are 1.2515 meters apart. For a different orientation, say, at
6o when the separation becomes 0.8362 meters, the MSE rapidly rises for snr
below 15 dB. The estimation error (MSE) also depends upon the number of
corners in a polygonal object. The numerical results are shown in fig 7.30b.

§7.6 Exercises:
1. What is the essential difference between an array of sensors used for DOA
estimation (chapters 2 and 5) or for signal waveform estimation (chapters 3 and
6) and the array used in nondiffracting radiation tomography?

© 2001 by CRC Press LLC

 100

1
mse

.01

3 6 9 12 15
snr (in dB)
.1

.01
mse

.001

.0001
4 5 6 7 8
Number of corners
Figure 7.30: (a) Mean square error (m2 ) in the estimation x- and y-coordinates as a
function snr. (b) Mean square error (m2 ) as a function of the number of corners
(snr=20dB). The dotted lines show the error bounds for 95% confidence.

2. Show that the filter function used in the backpropagation algorithm,

equation (7.14c), reduces to that used in the backprojection algorithm (page
368) for the radiation of wave length much smaller than the scale of
speed/density fluctuations.
3. A rectangular object (2D) is illuminated by a plane wavefront as shown in
fig. 7.31 below. An array of sensors measures the relative delays. Assume that
the speed of propagation inside the object is constant and that there are no
diffraction effects. (a) Sketch the delay profile (delay as a function of sensor
position) for different object orientations, θ =0o (as seen in the figure above),
θ =45o, and θ =90o. (b) How would you estimate the size of the object?
4. Consider a simple arrangement of sources and sensors in two borehole as
shown in fig. 7.32. Compute the sampling points in the Fourier plane. See
[28] on how to compute the sampling points in a more general case.

© 2001 by CRC Press LLC

 ULA
Wavefront

Figure 7.31: A rectangular object (2D) is illuminated by a plane wavefront

20m

50m

Figure 7.32: A cross borehole tomography experimental setup.

References

1. A. C. Kak, Tomographic Imaging with diffracting and non diffracting

sources, in Array Signal Processing, Ed. S. Haykin, Prentice-Hall, Englewood
Cliffs, NJ, pp. 351-428, 1985.
2. S. Ivansson, Seismic borehole tomography- theory and computational
methods, Proc. of IEEE, vol. 74, pp. 328-338, 1986.
3. M. Gustavsson, S. Ivansson, P. Moren, and J. Pihl, Seismic borehole
tomography- measurement system and field studies, Proc. IEEE, vol. 74, pp.
339-346, 1986.

© 2001 by CRC Press LLC

 4. K. Iwata and R. Nagata, Calculation of refractive index distribution from
interferograms using the Born and Rytov's approximation, Japan, J. Appl.
Phy., vol. 14, Suppl. 14-1, pp. 379-383, 1974.
5. R. K. Mueller, M. Kaveh, and G. Wade, Reconstructive tomography and
applications to ultrasonics, Proc. IEEE, vol. 67, pp. 567-587, 1979.
6. E. Wolf, Three dimensional structure determination of semi-transparent
objects from holographic data, Optics Cummunication, vol. 1, pp. 153-156,
1969.
7. A. J. Devaney, Diffraction Tomography, in Inverse Methods in
Electromagnetic Imaging, Part 2, W. M. Boerner et. al. (Eds), pp. 1107-1135
D. Reidel Pub Co., 1985.
8. A. J. Devaney, Geophysical diffraction tomography, IEEE Trans., vol. GE-
22, pp. 3-13, 1984.
9. Prabhakar S. Naidu, A. Vasuki, P. Satyamurthy, and L. Anand, Diffraction
tomographic imaging with circular arra, IEEE Trans., UFFC-42, pp. 787-789,
1995.
10. R. Porter, Determination of structure of weak scatterers from holographic
images, Opt. Commun. vol. 39, pp. 362-364, 1981.
11. A. J. Devaney and G. Beylkin, Diffraction tomography using arbitrary
transmitter and receiver surfaces, Ultrasonic Imaging, vol. 6, pp. 181-193
1984.
12. S. J. Norton, Tomographic reconstruction of acoustic reflectivity using
circular array: Exact solution, J. Acoust. Soc. Amer. vol 67, pp. 1266-1273,
1980.
13. S. J. Norton and M. Linzer, Ultrasonic reflectivity tomography with
circular transducer arrays, Ultrasonic Imaging, vol. 1, pp. 154-184, 1979.
14. Z. Qin, et al., Circular array ultrasound holographic imaging using the
linear array approach, IEEE Trans., vol. UFFC-36, pp. 485, 1989.
15. J. M. Ruis, M. Ferrando, L. Jofre, E. De Los Reyes, A. Elias, and A.
Broquet, Microwave tomography: An algorithm for cylindrical geometries,
Electronics Letters, vol. 23, pp. 564-565, 1987.
16. P. Tamarkin, Scattering of an underwater ultrasonic beam from liquid
cylindrical obstacles, J. Acoust. Soc. Am., vol. 21, pp. 612-616, 1949.
17. S. J. Bezuszka, Scattering of underwater plane ultrasonic waves by liquid
cylindrical obstacles, J. Acoust. Soc. Am., vol. 25, pp. 1090-1095, 1953.
18. A. Vasuki and P. S. Naidu, Broadband tomographic imaging with circular
array, Acoustic Letters (UK), vol. 21, pp. 144-150, 1998.
19. J. L. Sanz, On the reconstruction of band-limited multidimensional signals
from algebraic contours, IBM Research Report, RJ 4351, (47429), 1984.
20. P. S. Naidu, Modern Spectrum Analysis of Time Series, CRC Press, Boca
Raton, Fl, 1996.
21. Tien-when Lo, G. L. Duckworth, and M. N. Toksoz, Minimum cross
entropy seismic diffraction tomography, J. Acoust. Soc. of Am., vol. 87, pp.
748-756, 1990.

© 2001 by CRC Press LLC

22. A. Mohammad-Djafari and G. Demoment, Maximum entropy Fourier
synthesis with application to diffraction tomopraphy, Applied Optics, vol. 26,
pp. 1745-1754, 1987.
23. J. P. Burg, Maximum entropy spectrum analysis, Presented at the
International Meeting of Exploration Geophysicists, Orlando, Fl, 1967.
24. C. Q. Lang and W. Xiong, An iterative method of ultrasonic reflection
mode tomography, IEEE Trans on Medical Imaging, vol. 13, pp. 419-425,
1994.
25. G. Medioni and Y. Yasumoto, Corner detection and curve representation
using cubic B-spline,'Computer Vision Graphics and Image Processing, Vol
39, pp. 267-278, 1987.
26 P. Milanfar, W. C. Karl and A. S. Willsky, Reconstructing binary
polygonal objects from projections: A statistical view. Graphical Models and
Image Processing, vol. 56, pp. 371-391, 1994.
27. A. Buvaneswari and P. S. Naidu, Estimation of shape of binary polygonal
object from scattered field, IEEE Trans. on Med Imaging, vol. 7, pp. 253-257,
1998.
28. D. Nahamoo, S. X. Pan, and A. C. Kak, Synthetic aperture diffraction
tomography and its interpolation free computer implementation, IEEE Trans.,
vol. SU-31, pp. 218-226, 1984.

© 2001 by CRC Press LLC

 Chapter Eight

Imaging by Wavefield Extrapolation

We have seen in chapter 2 that the wavefield on a horizontal plane can be

derived from the wavefield observed on another horizontal plane given that the
space lying between the two planes is homogeneous and free from any sources.
This operation is called extrapolation of wavefield; forward extrapolation is
when we go away from the sources and backward extrapolation is when we go
toward the sources (see fig. 8.1). Wavefield extrapolation enables us to map the
source distribution provided it is known that all sources are confined to a layer.
This problem is known as inverse source problem [1]. A slightly different
situation arises in the scattering problem. An external source induces a field on
the surface of a scatterer which in turn will radiate wavefield, known as scattered
field, back into the space. This scattered field contains information about the
scatterers. The inverse scattering problem pertains to extraction of information
about the scatterers from the scattered field. The tomographic imaging covered
in chapter 7 falls in the realm of the inverse scattering problem. In the present
chapter we seek a means of reconstructing a layered (but not necessarily
horizontally layered) medium using the reflected wavefield. This problem is of
great significance in seismic exploration where it is commonly known as
migration. An image of subsurface reflectors can also be achieved through
focused beamformation, which gives an estimate of the reflected energy received
from a subsurface point. The focused beamformation is based on a ray theoretic
description of the wavefield, as in optics, but the migration is based on
diffraction properties of the wavefield. Both approaches lead to similar results.
For imaging, an essential input is the wave speed which, fortunately, has to be
estimated from the observed wavefield only.

§8.1 Migration:
The interface between two homogeneous layers may be considered as a thin
layer of point sources (scatterers). This forms the basis of the exploding
reflector model [2]. The wavefield observed on the surface of earth can be
extrapolated downward into the earth. The interfaces separating homogeneous
layers reflect or scatter wave energy. Such an interface may be modeled as a
surface with point scatterers whose density is proportional to the impedance
discontinuity. To image an interface, that is, to map the impedance
discontinuity, it is necessary to compute the distribution of the wave energy on
an interface. This problem has been treated as a boundary value problem [3, 4]
or an initial value problem [5]. As a boundary value problem we solve the
wave equation in homogeneous half space with a boundary condition that the
wavefield is given on the surface of the earth (z=0).

© 2001 by CRC Press LLC

 Forward Extrapolation

us
Measurement plane

eo
ium en
ed g
m omo
H
Backward Extraploation

Sheet of point sources

Figure 8.1: The wavefield measured on a horizontal plane can be extrapolated upward
(forward extrapolation) or downward toward the source (backward extrapolation).

The boundary value problem has been solved in the time domain by
Claerbout [3] by solving a finite difference equation with a time varying
boundary condition and in the frequency domain by Stolt [4] by expressing the
extrapolation as a filtering problem. Extrapolation in the z direction may also
be expressed as propagation in backward time. Note that in the wave equation
the double derivative with respect to z differs from the double derivative with
respect to time only in a scale factor given by -c2. Thus, starting from some
time instant, the wavefield observed at the surface (z=0) is propagated backward
in time, a process that is equivalent to extrapolation of the wavefield to a lower
level (z<0). The boundary value problem may be reformulated as a source
problem with zero boundary condition but driven by an external source which is
given as a time reversed output of each receiver [6]. In another approach
extrapolation is posed as an initial value problem but marching backward [5].
In this approach the time axis is scaled by the wave speed which converts a
recorded seismic section into a wavefield throughout the subspace as it might
have appeared at the latest recording time. Thus the converted wavefield is next
propagated backward in time.
8.1.1 Imaging Conditions: Imaging requires two steps, namely, (i)
extrapolation in space or reverse propagation in time and (ii) an imaging
condition, that is, how to decide when an image has been formed. In optical
imaging convergence of all rays emerging from a point to another point (image
point) is the imaging condition. In seismic or acoustic imaging, the imaging
condition commonly used is when the depropagated field reaches the starting
time which is the time when the scatterer was illuminated or excited. This
information can be found given the wave speed and the distance from the source
[see fig. 8.2]. It is also possible to set the excitation time to zero provided the
scattered wavefront and illuminating wavefront travel along the same path but

© 2001 by CRC Press LLC

 Source Array

Scatterer

Figure 8.2: Depropagate a wavefront to a time instant when the illuminating

wavefront hits the scatterer. When this happens, the scattered wave field is found to
be concentrated around the scattering point.

in opposite directions. This forms the basis of the popular migration principle
called exploding reflector which we shall describe in detail in the next section.
Qualitatively speaking, imaging is focusing of wave energy. An imaging
condition based on how well the wave energy is focused at a point is also a
likely candidate as an imaging condition. Indeed, in seismic imaging, it has
been suggested that when p-waves and s-waves are focused at the same point, an
image of the point is obtained [7].
8.1.2 Downward Continuation of Sources and Sensors: The source and sensor
arrays are normally placed on the same surface. It is possible to analytically
compute the field when both source and sensor arrays are relocated onto another
plane given the field on the observation plane. For simplicity we shall assume
that both the observation plane and the plane onto which the source and sensor
arrays are relocated are horizontal. We are already familiar with continuation of
wavefield from one plane to another (see chapter 1 and also later in this
chapter). Continuation of a source array requires an additional concept of
reciprocity which states that when the positions of an omnidirectional source
and an omnidirectional sensor are interchanged the observed field remains
unchanged [8]. To extrapolate the source array keeping the sensor array fixed we
need to interchange the source and the sensor arrays and then apply the
wavefield extrapolation algorithm. By virtue of the principle of reciprocity the
result of the above approach will be same as that of actual relocation of the
source array. In actual application the source and the sensor arrays are relocated
alternatively in small steps. As the source and the sensor arrays are continued
downwards towards a reflector at some stage the two arrays will completely
coincide when they reach the reflector after a lapse of time equal to the one way
travel time. Occurrence of such a coincidence of source and sensor may be used
as a condition for imaging. This phenomenon is illustrated in fig. 8.3.

© 2001 by CRC Press LLC

 Observation plane

Continuation plane

Continuation plane

Reflector

Figure 8.3: The wavefield measured in one plane is continued to another plane as if
the source and the sensor arrays are located on that plane. It may be recalled that as
the wavefield is continued the wave energy actually propagate along a ray.

§8.2 Exploding Reflector Model:

The wavefield, measured on a horizontal surface, is propagated backward in time
in order to arrive at the reflector location giving an estimate of the wavefield as
existing at the reflector. The seismic traces are first stacked with proper delays
so that the output corresponds to a hypothetical sensor kept close to the source.
A set of such stacked seismic traces may be modeled as a wavefield observed in
an imaginary experiment in which small charges are placed on a reflector and all
of them are exploded at the same time instant. The wavefield is assumed to
propagate upwards and reach the surface at time to. Conversely, the wavefield
when propagated backwards will reach the reflector point after to time units. It
is assumed that there are no multiple reflections, surface waves, refractions, etc.
Indeed, during the process of stacking, since the array is focused downward,
much of the interference would be attenuated. The exploding reflector model,
also known as the Loewenthal model [2], consists of tiny charges placed on a
reflector and fired at the same time instant (fig. 8.4). The quantity of charge
placed at a given point on the interface is proportional to the reflection
coefficient at that point.
Let the interface be described by a function, z = g(x, y) , where the
acoustic impedances above and below the interface are constant but different (see
fig. 8.4). The reflection coefficient for vertical incidence is given by
(ρ2 c2 − ρ1c1 )
r0 = . The reflectivity function may be written as
(ρ2 c2 + ρ1c1 )

© 2001 by CRC Press LLC

 Array

ray path

c1 ρ 1
Tiny charges are fired at the
same time instant
z=g(x,y)
c2 ρ 2

Figure 8.4: Exploding reflector model, also known as the Loewenthal model. Tiny
charges are placed on a reflector and fired at the same time instant. The ray paths are
perpendicular to the reflector. The wavefront, at the time of firing, that is, t=0,
coincides with the reflector.

r(x, y, z) = r0δ(z − g(x, y)) (8.1)

The wavefield in a medium bounded by the earth’s surface above and the
interface below will satisfy the following boundary condition and initial
conditions:

Boundary condition: f (x, y, z,t) z = 0 = f 0 (x, y,t) , that is, pressure field
observed on the surface.
Initial condition: f (x, y, z,t) t = 0 = r (x, y, z)δ(t) , that is, pressure
field generated by the exploding charges on the
interfaces.

The wavefront at time instant t=0 is the interface itself. As the time progresses
the wavefront travels upward toward the surface. The wavefield observed at the
surface acts as a boundary condition, and the shape of the wavefront at t=0 is
the initial condition which, in practice, is not known. The boundary condition
(when given over an infinite plane) is enough to solve the wave equation. The
wavefield thus obtained at all points within the space bounded from above by
the observation plane and from below by the initial wavefront and for all t in
the range 0 ≤ t < ∞ . Observe that the wavefield at t=0 is actually the pressure
field generated by setting off charges on the interface and everywhere else it is
zero. Conversely, if the wavefield observed on the surface is propagated

© 2001 by CRC Press LLC

 backward in time till we reach the time instant t=0 we shall indeed reach the
surface where the charges were set off. This is the rationale for imaging a
reflector by propagation backward in time.
8.2.1 Initial Value Problem: Alternatively, an equivalent description of
imaging is through a solution of initial value problem which may also be
expressed as an inverse source problem [9]. The wave equation with the right
hand side equal to a source creating a wavefield is

1 d2 f
∇2 f = + r (x, y, z)δ ′(t)
c02 dt 2

where r(x,y,z) as before stands for the reflectivity function and δ ′(t) is the
derivative of δ(t) . At each point an explosive charge proportional to r(x,y,z) is
set off at t=0. The waves propagate unhindered by other reflecting interfaces (no
multiple reflections). The solution of the inhomogeneous wave equation on the
z=0 surface is given by

f (x, y, z = 0,ω)
+∞
e jk ( x − x ′ ) +( y − y ′ ) +( z ′ )
2 2 2
jω (8.2)
= ∫ ∫
4π − ∞ ∫ r ( x ′ , y ′ , z ′ )
(x − x ′ ) + (y − y ′ ) + ( z ′ )
2 2 2
dx ′dy ′dz ′

which may be further simplified following the procedure used in obtaining

(1.81)

+∞
ω R(u,v, w) j (ux + vy)
f (x, y, z = 0,ω) =
4π 2 ∫∫
−∞
w
e dudv (8.3a)

R(u,v, w)
F0 (u,v,ω) = ω (8.3b)
w

where w = −sgn(ω) k − u − v for upgoing waves. From (8.3b) we can

2 2 2

obtain the unknown reflectivity from the surface pressure field

w
R(u,v, w) = F0 (u,v,ω) (8.4)
ω
Next we shall show how the same result (that is, (8.4)) can be obtained as a
boundary value problem (see (8.7)).

© 2001 by CRC Press LLC

§8.3 Extrapolation in ω−k Plane:
Recall the integral representation of a wavefield in a homogeneous medium (see
chapter 1, (1.19)) which we reproduce here for quick reference,

−∞
1
3 ∫ ∫ ∫ 0
f (x, y, z,t) = F (u,v,ω)e + j k 2 −u 2 − v 2 z − j (ux + vz −ωt )
e dudvdω
8π − ∞

± j k 2 −u 2 − v 2 z
We have chosen the positive sign in e as the wavefield is
propagated from surface to the interface where charges are placed, that is,
propagation is towards the source; hence, as per our convention (chapter 1), +ve
is chosen.
8.3.1 Downward Continuation: The wavefield measured on the surface may be
continued downwards to any depth and for all times (see (1.26)). Using the
initial condition in the exploding charge model the wavefield at time t=0 is
equal to the reflectivity function,

r(x, y, z) = f (x, y, z,t) t = 0

−∞
1 (8.5)
3 ∫ ∫ ∫ 0
= F (u,v,ω)e + j k 2 −u 2 − v 2 z − j (ux + vy)
e dudvdω
8π − ∞

ω
Note that in (8.2) k= where c is the wave speed in the medium above the
c
interface and it is assumed to be known. Further, we relate the temporal
ω s2
frequency to the vertical spatial frequency, w. Since = −w 1 + 2 , k=
c w
ω sc
where s = u + v , we can express w = − 1 − ( )2 for an upgoing
2 2

c ω
c
wave and dω = − dw . Using these results in (8.5) we obtain [10]
s2
1+ 2
w

r(x, y, z) =
∞
c 1 s 2 − j (ux + vy + wz )
8π 3 ∫ −∫∞ ∫
F0 (u,v, −c w 1 + 2 )e dudvdw (8.6a)
s2 w
1+ 2
w

© 2001 by CRC Press LLC

 Equation (8.6a) is applicable only when the sources actually replace the
reflecting interfaces, but in practice the wave excitation is done on the surface
and the wave propagates into the medium and it is then reflected at an interface
back to the surface. In this process, since the wavefield travels first down and
then up, the travel time is doubled; equivalently, the wave speed may be halved.
Hence, we have

r(x, y, z) =
∞
c 1 c s 2 − j (ux + vy + wz )
16π 3 ∫ −∫∞ ∫
F0 (u,v, − w 1 + )e dudvdw (8.6b)
s2 2 w2
1+ 2
w

2ω sc
where w=− 1 − ( )2 . This result agrees with that given in [9].
c 2ω
Computing the inverse Fourier transform on both sides of (8.6) we obtain

c s2
R(u,v, w) = F0 (u,v, − cw 1 + ) (8.7a)
s2 w2
1+ 2
w

or

1 c c s2
R(u,v, w) = F0 (u,v, − w 1 + 2 ) (8.7b)
s2 2 2 w
1+ 2
w

The wavefield in a homogeneous (also in horizontally layered) medium

has a radial symmetry in the (x,y) plane. For this case the appropriate
extrapolation equation in polar coordinates is (1.27b), reproduced here for
convenience,

∞ ∞
1
∫ F (s,ω)e dω ∫ sJ0 (sr)e ± j ( k 2 −s2 z)
f (r, z,t) = 2 0
jωt
ds
4π −∞ 0

where F0 (s,ω) is the Fourier transform of the surface wavefield having a

radial symmetry. Let us rewrite (1.27) in terms of plane wave decomposition

© 2001 by CRC Press LLC

(PWD). We map s domain to γ domain where s = k sin(γ ) and rewrite
(1.27b) as
f (r, z,t) =
∞ ∞
1
∫ke dω ∫ F0 (sin(γ ),ω)e ± j (kz cos( γ )) cos(2γ )J0 (k sin(γ )r)dγ
2 jωt

8π 2 −∞ 0
(8.8)

F0 (sin(γ ),ω) in (8.8) or more specifically its inverse Fourier transform,

F0 (sin(γ ),t) , may be obtained from (1.32), that is, by slant stacking or
Radon transform. Recall that F0 (sin(γ ),t) is a plane wave incident at angle
γ , a result of plane wave decomposition of a point source. F0 (sin(γ ),ω) is
also known as the angular spectrum of the wavefield on surface. The angular
spectrum at depth z is given by

F(sin(γ ), z,ω) = F0 (sin(γ ),ω)e ± j (kz cos( γ )) (8.9)

8.3.2 Layered Medium: We can extend the wavefield extrapolation problem

from a single layer to a two layer medium. Each layer is separated by a plane
interface, either horizontal as in fig. 8.5a or inclined as in fig. 8.5b. In each
layer the density and wave speed are constant. We shall assume that zero offset
processing has removed all multiply reflected waves.
Let f 1 (x, y, z.t) and f 2 (x, y, z.t) be the wavefields produced by the
exploding charges placed on the interface I and interface II respectively. The
total field is given by f (x, y, z.t) = f 1 (x, y, z.t) + f 2 (x, y, z.t) . The
wavefield observed on the surface z=0 is given by
f 0 (x, y,0,t) = [ f 1 (x, y, z.t) + f 2 (x, y, z.t)] z = 0 . The wavefield generated
by exploding charges on two interfaces is given by

f (x, y, z,t) t = 0 = r1δ (z − g1 (x, y)) + r2δ (z − g2 (x, y)) (8.10)

where r1 and r2 are reflection coefficients and g1 (x, y) and g2 (x, y) surfaces
separating the two layers. It may be noted that by removing the multiple
reflections we have decoupled the two interfaces; in effect we have linearized the
propagation effects. Extension to the N-layer medium is straightforward when
all layers are decoupled.
For extrapolation it is necessary that the correct propagation speed of
the material in each layer is used. First, extrapolation to the first interface is
carried out using (1.26). The Fourier transform of the wavefield observed on the
surface is multiplied with the propagation filter function

© 2001 by CRC Press LLC

 ω
exp(− j k12 − u 2 − v 2 z) , 0 ≤ z ≤ z1 where k1 = . Extrapolation to the
c1
second interface is obtained by multiplying with a filter function,
ω
exp(− j k22 − u 2 − v 2 z) , z1 ≤ z ≤ z2 where k2 = .
c2

f 2 (x, y, z, t) t = 0
−∞
1
∫ ∫ ∫ F (u, v, ω)e
− j k12 −u 2 − v 2 z1 − j k 22 −u 2 − v 2 ( z − z1 ) − j (ux + vy)
= 0 e e dudvdω
8π 3
−∞

c2 w
−∞ F0 (u, v, ±c2 sgn(ω) w 2 + u 2 + v 2 )
1 w + u2 + v2
2
= ∫∫∫
8π 3
−∞
e
−j { }
k12 −u 2 − v 2 − k 22 −u 2 − v 2 z1
e − j (ux + vy − wz ) dudvdw
= r2 (x, y, z)
(8.11)

Note that r1 (x, y, z) is zero in the second layer.

8.3.3 Sloping Interface: We shall now look at an example of a single sloping
reflector as shown in fig. 8.6 [10]. We shall assume a ULA of transceivers
oriented along the slope. For simplicity let us assume that the transceiver
radiates a spike and receives, after a delay, the same spike. The seismic data are
assumed to have been stacked so that we have a zero-offset data corresponding
to a field from hypothetical charges placed on the interface (exploding reflector
model described in §8.2). The zero-offset data may be expressed as

f (md,t) = δ(t − m τ 0 − t0 ), m = 0,1,... M − 1

d z
where τ0 = sin θ 0 , and t0 = 0 cosθ 0 . Note that z0 is a depth to the
c c
interface below sensor m=0. The Fourier transform of the zero-offset data may
be obtained assuming the array size is very large,

F0 (u,ω) = e − jωt 0 δ(ud + ωτ 0 ) (8.12)

Using (8.12) in (8.6a), that is, its 2D equivalent, we obtain after simplification

© 2001 by CRC Press LLC

 Source/sensor array

z=0
ρ1c 1

z=z1
ρ2c 2

z=z 2
ρ3c 3

(a)

Source/sensor array

z=0
ρ c1
1
∆ z1

ρ2c 2
∆ z2

ρ 3c 3

(b)

Figure 8.5: Wavefield extrapolation in two layer medium. (a) Horizontal layers and
(b) inclined layers. represents transceiver.

© 2001 by CRC Press LLC

 r(x, z)
1 w 2 +u 2 t 0
−∞ 2
e jc δ(ud + c τ 0 w 2 + u 2 )
c s
=
4π 2 ∫∫ −∞
1+
w2
e − j (ux + wz ) dudw (8.13)
−∞ c t0
c −j w

2π −∫∞
cos θ1 − j ( − tan θ 0 x + z )w
= cosθ 0 e e dw

= c cosθ 0δ(− tan θ 0 x − z0 + z)

From (8.13) the reflector is given by

z = tan θ 0 x + z0 (8.14)

It is interesting to note that the slope of the interface is not equal to the slope
of the line joining the spike arrivals in the zero-offset seismic data (see
exercise 3 at the end of this chapter).
8.3.4 Depropagation of Wavefield: Extrapolating a wavefield backward in time
is known as depropagation; when continued it is possible to home onto the
starting point. In the exploding reflector model all scatter points are
simultaneously fired. We shall now show that the field observed on the surface
is equal to the field generated at scatter points and then propagated to the
surface. By depropagating, we hope to obtain the field generated at the
exploding reflectors. The zero offset data provides only a blurred image of the
reflectors, but by depropagating the blurring can be reduced [5]. Consider a 2D
model with uniform wave speed with reflecting facets. The wavefield in a
uniform medium is given, in the frequency domain, by

+∞
1
∫ ∫ F(u,ω)e
j ( −ux + k 2 −u 2 z ) jωt
f (x, z,t) = 2 e dudω
4π −∞
+∞
c 1
∫∫ F(u,ω)e − j (ux + vz )e − j u 2 + v 2 ct
= 2 dudv (8.15)
4π −∞ u2
1+ 2
v

where v = −sgn(ω) k 2 − u 2 . The field on the surface, z=0, is given by

© 2001 by CRC Press LLC

 Source/Sensor array

Exploding
charges θ1

Figure 8.6: A sloping reflector and source/sensor (transceiver) array on the surface.
By a process of stacking, the seismic field is reduced to a field generated by
hypothetical exploding charges on the sloping interface.

+∞
c 1
∫∫ F(u,ω)e − j (ux + u 2 + v 2 ct )
f (x, z = 0,t) = dudv (8.16)
4π 2 −∞ u2
1+ 2
v

The mathematical framework for imaging by wavefield extrapolation is covered

in great detail in [11, Berkhout].

The wavefield at the exploding reflector at a depth z and backpropagated to the

z
surface at time t= is given by
c
+∞
z c 1
∫∫ F(u,ω)e − j (ux + vz )e − j u2 +v2 z
f (x, z,t = ) = 2 dudv (8.17)
c 4π −∞ u2
1+ 2
v

Since the wavefield is by and large vertically propagating, u << v (8.17) may
be approximated as

+∞
z 1
∫ ∫ cF(u,ω)e
− j (ux + 2vz )
f (x, z,t = ) ≈ 2 dudv (8.18)
c 4π −∞

The wavefield observed on the surface (z=0) is mapped into the (x,z) plane by
z
substituting for t. Note that the time axis of the recorded seismic data
c
2

© 2001 by CRC Press LLC

 refers to two way time and hence wave speed is halved as is common in seismic
data processing. We obtain

+∞
z 1
∫ ∫ cF(u,ω)e
j ( −ux + 2vz )
f (x, z = 0,t = )≈ 2 dudv (8.19)
c 4π
2 −∞

Thus, from (8.18) and (8.19) we have an important approximate result which
provides an initial wavefield for depropagation,

z z
f (x,0,t = ) ≈ f (x, z,t = ) (8.20)
c c
2
In words, the wavefield observed on the surface with its time axis mapped into
z
a z-axis ( t = ) is approximately equal to the wavefield at the reflector which
c
2
is then propagated to the surface.
z z
Now given f (x,0,t = f (x, z,t = ) we need to
) or
c c
2
depropagate till we reach the point of reflection. To depropagate by ∆t time
units we shall use (8.15) with approximation u << v

f d (x, z,t − ∆t)

+∞
1
∫ ∫ cF(u,ω)e
− j (ux − vz ) − jv c∆t
≈ 2 e dudv (8.21)
4π −∞

c
+∞  z  − j (ux − vz ) − jv c∆t
=
4π 2 ∫ ∫ FT  f (x,0,t = c e

e dudv
−∞  2

where the subscript d stands for depropagated field (see fig. 8.7)
8.3.5 Relation to Diffraction Tomography: We shall now show that a close
relation exists between imaging by extrapolation and diffraction tomography
which we have dealt with in the previous chapter. Consider a weakly
inhomogeneous object (speed fluctuations only) illuminated by a point source
and the scattered field being received by a sensor kept close to the source. We
have shown in chapter 7 (pp. 400-402) that the Fourier transform of the speed
variations evaluated on a spherical surface of radius 2k is related to the Fourier
transform of the field received by an array of transceivers (see fig. 7.25). The
relationship (7.33) is reproduced here for convenience:

© 2001 by CRC Press LLC

 0 x

∆t

t=z/c
ddDvvvv

t ∆t

Figure 8.7: The time axis of the wavefield measured on the surface is mapped onto
the z-axis (the wave speed is assumed to be known). The wavefield thus obtained in
the (x, z) plane is now depropagated in steps of ∆t. The depropagated segment
corresponds to the desired reflectivity function over a segment of depth, ∆t c.

c3 ∂  F1 (u,v, ω 0 ) 
∆c̃(u,v, 4k02 − u 2 − v 2 ) = 4k02 − u 2 − v 2  
2 ∂ω 0  ω 02 

It may be observed that the zero offset field measured in seismic exploration is
essentially an output of an array of transceivers which may be considered as a
mathematical model of the zero offset field measurement setup.
To relate ∆c̃(u,v, 4k02 − u 2 − v 2 ) to the reflectivity, we shall use (8.7a)
where we have shown the relationship between the Fourier transform of the
reflectivity and the wavefield on the surface (zero offset). It can be shown that
the partial derivative appearing in (7.33) may be given by

© 2001 by CRC Press LLC

 ∂  F1 (u,v, ω)  1 ∂  R(u,v, w)  w 2 + s 2
=−   (8.22)
∂ω  ω2  2 ∂w  w 2 + s 2 w  w

where R(u,v, w) is the spatial Fourier transform of r(x,y,z). Substituting

(8.22) in (7.33) we obtain

∆c̃(u,v, 4k02 − u 2 − v 2 )
c3 ∂  R(u,v, w)  (8.23)
=− w2 + s2  
4 ∂w  w 2 + s 2 w 

After carrying out the required differentiation we obtain the following result:

∆c̃(u,v, w)
c 3  R(u,v, w) R(u,v, w) Rw (u,v, w)  (8.24)
=  + − 
4  (w 2 + s 2 ) w2 w 

where Rw (u,v, w) stands for the derivative of R(u,v, w) with respect to w.

8.3.6 Continuation of Sources and Sensors: Earlier in §8.1 we had mentioned
that imaging can be realized by continuing both the source and sensor arrays
down to the reflector. This is particularly useful for imaging when zero offset
data are not available [12]. We shall now derive filters for the downward
continuation of the source and sensor arrays. The transfer function of a filter to
continue the wavefield to a plane ∆z below the observation plane may be
obtained from (1.18)

H(ur ,vr , ∆z) = exp(+ j (k 2 − ur2 − vr2 ∆z) (8.25a)

where the subscript r refers to the receiver coordinate. In the first step the
sources are held fixed on the z=0 plane (observation plane) and sensor array is
displaced downward. Next we interchange the source array with the displaced
receiver array. By virtue of the reciprocity theorem the wavefield at the new
position of the sensor array due to the source array also at its new position will
be equal to the previously downward continued field. In the second step we
downward continue the previously continued field but the continuation is done in
source coordinate space. The transfer function for downward continuation is
given by

H(us ,vs , ∆z) = exp(+ j (k 2 − us2 − vs2 ∆z) (8.25b)

© 2001 by CRC Press LLC

 Sensor Array

Source/Sensor
array outputs

Figure 8.8: M2 outputs source/receiver ULAs are arranged over a square grid. All
outputs in a row are known as source gathers and all outputs in a column are known
as sensor gathers.

where the subscript s refers to source coordinate. The downward continuation is

alternatively done in the source and sensor spaces. In practice the downward
continuation is carried out as follows. Assume that we have two source and
sensor ULAs each having M devices. There will be M2 outputs which we like
to arrange on a square grid as shown in fig. 8.8. All outputs in a row (known
as source gathers) are continued downwards in the sensor space and all outputs
in a column (known as sensor gathers) are continued in the source space.

§8.4 Focused Beam:

When a source is in the far field, the directions of arrival (DOA), azimuth and
elevation are of interest. We have already seen how a beam is formed in a given
direction (chapters 4 and 5). On the other hand, when a source is in the near
field region, a beam may be formed to receive energy not only from a given
direction but also from a given point. This is akin to focusing in an optical
system. In seismic exploration the array size is of the same order as the depth
of reflectors; therefore, it is often inappropriate to assume a far field or plane
wavefront condition which is required for the purpose of imaging, that is,

© 2001 by CRC Press LLC

 sources receivers

(a)

reflector

(b)

Figure 8.9: Common depth point (CDP) setup. (a) The source and receiver arrays are
arranged in such a manner that a signal is always reflected from the same reflecting
element. (b) Source-receiver pair.

to form a focused beam to receive reflected energy from a given reflector.

8.4.1 Zero-offset Wavefield: In reflection seismic exploration it is often desired
to position both source and detector at the same location, although in practice
this cannot be achieved and some amount of offset is always present. The main
advantage of this arrangement is that the returned signal is a normally reflected
signal; consequently, there is no conversion of wave energy into s-waves (see
chapter 1, p. 28). The zero offset wavefield can however be obtained through
the array processing approach. For this we must focus a beam at the foot of a
perpendicular drawn from the array mid-point to the reflector. A typical example

© 2001 by CRC Press LLC

of a source and receiver setup is shown in fig. 8.9. All signals reaching the
sensor array emanate from the same reflecting element. In seismic parlance
these signals are called common depth point (CDP) gathers.
Common Depth Point: A linear array of sources and detectors are so arranged
that the received signal always comes from the same reflecting element. This
arrangement is illustrated in fig. 8.9a, and is commonly known as a common
depth point (CDP) setup. The total time of travel from the source to the
receiver is a function of separation or offset between the source and the receiver.
When the reflecting element is horizontal the travel time is given by

x2
T x2 = T02 + (8.26)
c2

2l
where x stands for separation between the source and receiver , T0 = , and l
c
is depth to the reflecting element (see fig 8.9b). A plot of T x vs x is very
useful for it enables us to estimate T 0 and the wave speed.
Let us now consider a sloping reflecting element but the source and
receiver arrays are, as before, on a horizontal surface (see fig. 8.10). Let the
slope of the reflecting element be α . To compute the travel time to a
downslope or an upslope sensor we consider the image of the source and
compute the distance from the image to the sensor. For a downslope sensor,

(ir+ )
Tx+ =
c
(8.27)
(x+ )2 + 4l 2 − 4lx+ cos(∠isr+ )
=
c

and squaring on both sides of (8.27) we obtain

2
=  +  + T02 + 2T0 + sin(α)
2 x x
T (8.28)
x+
 c c

Similarly, we have for an upslope receiver

2
T x2− =  −  + T02 − 2T0 − sin(α)
x x
(8.29)
 c c

The upslope and downslope travel times are shown in fig. 8.11.

© 2001 by CRC Press LLC

 r- source r+
s

i image

Figure 8.10: Inclined reflector. There are two sensors, one sensor, r+ , is downslope
and another sensor, r- , is upslope. To compute the travel time we consider an image
of the source. The time of travel to the downslope sensor is equal to (i r+ ) /c and
similarly for the upslope sensor.

7.5
downslope
7

6.5
Time in sec

5.5
upslope
5

4.5
0 10 20 30 40 50 x150
Distance in meters

Figure 8.11: Downslope and upslope travel times (eqs. 8.21 & 8.22) in seconds as a
function of distance in meters. Wave speed =1500m/s, T0 =5sec. Solid curve is for
horizontal reflector (8.19).

© 2001 by CRC Press LLC

 2l
where T0 = and l is the perpendicular depth to the reflector below the
c
source position (see fig. 8.9). Note that (8.28) and (8.29) reduce to (8.26) for
α =0 and for x → 0 , T x → T0 . Rearranging terms in (8.28) and (8.29) we
can express them in a compact form

2 2
T x2± =  T0 ± ± sin(α) ±  ±  cos(α)
x x
(8.30)
 c   c

Let the downslope and the upslope sensor be equidistant from the source, that
is, x+ = x− = x . The average of the squares of downslope and upslope travel
times turns out to be independent of the slope of the reflector,

T x2+ + T x2− 2
=   + T02
x
(8.31)
2  c

Similarly, the difference of the squares of downslope and upslope travel times is
given by

T x2+ − T x2− x
= T0 sin(α) (8.32)
4 c
8.4.2 Layered Medium: We now consider a layered medium overlying a
reflector. The overlying medium is modeled as a stack of uniform layers (see
fig. 8.12). The round trip travel time T x and the receiver position x are the two
parameters of interest. They are given by

N
2∆zk
Tx = ∑ (8.33a)
k =1 ck (1 − p 2 ck2 )

N
2∆zk ck
x = p∑ (8.33b)
k =1 (1 − p 2 ck2 )

© 2001 by CRC Press LLC

 kth layer : ∆ zk ck

reflector

Figure 8.12: Layered medium overlying a reflector. Each layer is homogeneous with
constant thickness and speed. The ray path consists of a series of linear segments.
At an interface between two layers the Snell’s law (see p. ) must be satisfied.

sin(θ k )
where p is the ray parameter ( p = ) where θ k is the angle of
ck
incidence in the kth layer (see page for more information on the ray parameter)
and N stands for the number of layers. Let us expand T x as a function of x in
Taylor’s series,

d 2T x x 2 d 4T x x4
T x = T0 + + + (8.34)
dx 2 x=0
2! dx 4 x=0
4!

Since, for a horizontally layered medium, T x is a symmetric function of x,

only even terms are retained in the Taylor’s series expansion (8.27). Further,
d 2T x 1
the second derivative can be shown to be equal to = N .
dx 2 x=0
∑ ∆z c
k =1
k k

Equation (8.34) simplifies to

x2
T x ≈ T0 + N (8.35)
2 ∑ ∆zk ck
k =1

Upon squaring on both sides of (8.35) and retaining only the second order terms
we obtain

© 2001 by CRC Press LLC

 x 2T0
T x2 ≈ T02 + N

∑ ∆z c
k =1
k k

which after rearranging reduces to a form identical to (8.26),

x2 x2
T x2 ≈ T02 + = T 2
+ (8.36)
1 N ∆zk 2 0 2
crms
∑ k
T0 k =1 ck
c

1 N ∆zk 2
where
2
crms = ∑ ck is known as a root mean square speed of the
T0 k =1 ck
layered medium. Indeed we may replace a layered medium by a uniform medium
having a speed equal to crms . For the purpose of focused beamformation we
may use the root mean square speed.

8.4.3 Focusing: To form a focused beam we must have many CDP gathers
which are obtained by means of a specially designed source-receiver array. For
example, consider a linear array where every location is occupied either by a
source or a sensor. The array is fired as many times as the number of locations.
From every reflecting element we obtain a number of gathers equal to the
number of receivers. This is illustrated in fig. 8.13. A ULA of four sensors is
headed by a source. In position (A) the source is fired and the reflected signal is
received by sensor #1. The entire receiver-source array is moved laterally by half
sensor spacing as in position (B) and the source is fired once again. The
reflected signal is received by sensor #2. This procedure is continued as in (C)
and (D). Thus, we get four CDP gathers.
Let T1 ,T 2 ,...T N be the round-trip travel time from the source to sensor #1,
#2,...#N, respectively. Let f 0 (t) be the signal transmitted by the source and
f i (t), i = 1,2,... N be the reflected signals received by four receivers. Since
these are the delayed versions of f 0 (t) we can express them as

f i (t) = f 0 (t − Ti ), i = 1,2,... N (8.37)

Focusing involves firstly correcting for delayed reception and secondly

summing coherently after correction. Let the delay correction be given by

© 2001 by CRC Press LLC

 A

Figure 8.13: A reflecting element is illuminated by a source at different angles. The

reflected signal is received by one of the sensors as shown in the figure. The entire
receiver-source array is moved laterally by half sensor spacing.

T̂1 , T̂ 2 ,..., T̂ N which are computed using (8.26) for an assumed depth to the
reflector,

© 2001 by CRC Press LLC

 ∆tn 2
T̂n = T̂0 (1 + ( ) ) (8.38)
T̂0

2lˆ ˆ d
where T̂0 = , l is assumed depth to the reflector and ∆t = . Recall that
c c
d stands for sensor spacing. We assume that the wave speed c is known. The
CDP gathers are coherently summed after correction for the delays computed
from (8.38). As before we shall assume that the source emits a broaband signal.
The coherently summed output may be expressed as follows:

 ∆tn 2 ∆tn 2 
∞
1 1 N −1 − jω  T 0 (1+( ) ) − T̂ 0 (1+( ) )

2π −∫∞
∑e 
T0 T̂ 0
g(t) = F(ω) e jωt dω
N n=0
(8.39a)
∞
1
2π −∫∞
= F(ω)H N (ω)e jωt dω

where

 ∆tn 2 ∆tn 2 
1 N −1 − jω  T 0 (1+( ) ) − T̂ 0 (1+( ) )
H N (ω) = ∑ e 
T0 T̂ 0
(8.39b)
N n=0

is the filter transfer function. A numerical example of the transfer function is

shown in fig. 8.14. A horizontal reflector is assumed at a depth corresponding
to round-trip time equal to 5 seconds. The sensor spacing, measured in units of
d
propagation time, ∆t = = 0.1 seconds.
c
8.4.4 Depth of Focus: The response function has a finite width. A sharp
reflector now appears as a diffused zone whose width is known as the depth of
focus, analogous to that in optics. Ideally one would like the depth of focus to
be as narrow as possible. For the purpose of quantitative measure we shall
define the depth of focus as a distance between two 3 dB points on the response
function. The depth of focus, measured in the units of round-trip propagation
2∆l
time (= , where ∆l is depth of focus), and the array aperture, also
c
x
measured in terms of the propagation time (= ), are shown in fig. 8.15.
c
Notice that the minimum occurs when the aperture size is about four times the
round-trip propagation time, in this case five seconds. With further increase in

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

0.4

0.2

0
3 4 5 6 7
Round-trip time in sec.

Figure 8.14: Response function of focused beamformation. The reflector is at a

depth corresponding to 5 sec round-trip time. The array aperture is measured in
x
propagation time ( ). Solid line is for array aperture of 2.5 seconds and dashed line
c
for 5.0 seconds. Further the angular frequency is assumed to be 100 radians.

the aperture size the depth of focus rapidly deteriorates. The minimum depth of
focus appears to be independent of depth to the reflector; however, the required
array aperture increases rapidly as the reflector depth increases. The dependence
of the depth of focus on frequency is significant, as shown in fig. 8.16. Notice
that the depth of focus becomes very narrow beyond about 50Hz.
8.4.5 Inclined Reflector: We consider a sloping reflector. A small segment of
the reflector is illuminated at different angles by means of a source-receiver
array similar to the one used for a horizontal reflector. As the array is laterally
shifted the point of reflection changes, but only slightly, depending upon the
slope (see fig. 8.17). When the slope is zero all CDP gathers emanate from the
same point on the reflecting element. For gentle slope the spread of the
reflecting points is assumed to be small, small enough to allow the assumption
of a constant slope.

© 2001 by CRC Press LLC

 2.0

1.5
Depth of focus

1.0

0.5

0.0
1 10 100 1000
Array Aperture
Figure 8.15: Depth of focus in units of round-trip propagation time (sec) as a
function of array aperture, also measured in units of propagation time (sec).

0.7

0.6

0.5
Depth of focus

0.4

0.3

0.2

0.1

0.0
0 200 400 600 800 1000
Angular Frequency
Figure 8.16: Depth of focus vs angular frequency. The reflector is at a depth of 5
seconds (round-trip travel time). The array aperture is held fixed at 20 seconds
(propagation time).

© 2001 by CRC Press LLC

 source

l'

Figure 8.17: Common depth point (CDP) gathers from a sloping reflector. The
2l'
reflecting element is at a depth of l' (round trip travel time= ). Two positions of
c
source-sensor array are shown; position #1: circles and position #2: squares. Notice
the displacement of the reflecting point.

The response of the focused beamformer as a function of the slope and the
2l'
return travel time is shown in fig. 8.18. We have assumed that = 5 and
c
x
array aperture, in units of propagation time, is equal to = 20 seconds. There
c
are 200 CDP gathers spaced at an interval equal to 0.1 sec. The result shown in
fig. 8.18 is perhaps the best one can expect as we have assumed the optimum
array size.
8.4.6 Relation Between Focusing and Downward Extrapolation: Focusing
appears to achieve what the downward extrapolation does in migration. Indeed
both are related in the sense that focusing is a simplified version of
extrapolation. In order to see this relationship let us examine the impulse
response function of the downward extrapolation filter whose transfer function
is given by

ω
H(ω,u,v) = exp( j∆z (( )2 − u 2 − v 2 )) (8.40a)
c

To get the impulse response function we compute the 2D inverse Fourier

transform of (8.40a). The result is given in [13]

© 2001 by CRC Press LLC

 1

0.8

0.6
Response

0.4

0.2

0
6

Ro 5.5 14
un 12
dt 5
rip 10
tim eg
e i 4.5 8 e in d
ns op
ec 4 6 Sl

Figure 8.18: A mesh plot of the response function of the focused beamformer as a
function of slope of the reflector and round-trip time. Array aperture is 20 seconds
(propagation time). Angular frequency assumed is equal to 100 radians.

 exp( j ω r)  ω
j r +1
1 ∂  c  1 c ω
h(r,ω) =   = exp( j r) (8.40b)
2π ∂z r 2π r 3
c
 
 

where r = x 2 + y 2 + ∆z 2 . Let us rewrite the exponential term in (8.40b) as

ω ρ2 ∆z 2 ρ2
exp( j r) = exp( jω 2 + 2 ) = exp( jω T02 + 2 ) (8.40c)
c c c c

© 2001 by CRC Press LLC

 where ρ = x + y . The phase delays introduced by the downward
2 2 2

extrapolation filter are identical to those used in the focused beamformer. The
difference, however, lies in the amplitude term. In place of a variable amplitude,
we use a constant amplitude in the focused beamformer.
8.4.7 Focused Beamformation for Imaging: A focused beam enables us to
estimate the amount of scattered or reflected power from a scattering volume
element or a reflecting surface element located at a point in space. Consider a
volume element illuminated by a point source with the scattered waves being
received by an ULA (see fig. 8.19). To form a focused beam we need to
compute the travel time from the source to the volume element and from the
volume element to a sensor. For simplicity let us assume that the background
wave speed is constant so that ray paths are straight lines. In any realistic
problem, however, the ray paths are more likely to be curved. It will be
necessary to trace the rays before we are able to compute the travel times. Such
a situation was considered in [14]. Let f n (t), n = 1,2,... N be the sensor
outputs. The focused beam output is given by

1 N
g(t) = ∑ f n (t − t(so) − t(orn ))
N n =1

where t(so) stands for the travel time from the source to the volume element
and t(orn ) stands for the travel time from the volume element to the nth

∫ g(t)
2
sensor. The scattered power from the volume element is given by dt
Ts
where T s stands for signal duration. The process is repeated at every point in
the space (x-z plane). The resulting map provides an image of the scattering
strength.
The question of resolution needs to be answered. The size of the
volume element is controlled, along the x-axis, by the array aperture (inversely
proportional to array aperture), along the z-axis, by depth of focus and finally
along the y-axis, by the bandwidth of the signal (inversely proportional to
bandwidth) [14]. An interesting application of the focusing by means of back
propagation of wavefield to detection of an acoustic source in a room without
considering the reflections from the walls is reported in [15]. An array of
microphones is deployed in a 3D space surrounding the source(s) of sound
energy. The output from each microphone is back propagated to a point in

© 2001 by CRC Press LLC

 source array

s r1 r2

z
o

Figure 8.19: The array output is focused to a point o where a scattering volume
element is presumed to be located. The power output is proportional to the
scattering strength of the volume element.

3D space. Since we assume that no reflections from the walls or scattering

exist, the back propagation is simply equal to delaying the microphone output
by an amount equal to the travel time along a straight line path from the sensor
to the assumed position of the source. The delayed output is cross-correlated
with similarly delayed outputs from other sensors. An average of the cross-
correlation is a measure of the acoustic energy present at the selected point.
Ideally, the averaged cross-correlation function will have a sharp peak at the true
sound source position. In practice, it is shown in [15] that a broad peak stands
at the true sound source position.
§8.5 Estimation of Wave Speed:
We have assumed that the average background wave speed is known and this
information was used for the purpose of imaging. We shall now look into how
the wave speed can be estimated from the recorded wavefield itself. Firstly, the
medium is assumed to be homogenous for which one needs to estimate just one
unknown, namely, the wave speed of the homogenous medium. Secondly, the
medium is a stack of layers for which we need to estimate the root mean square
speed defined earlier (8.36). It turns out that the approach remains the same for
both.
8.5.1 Wave Speed from CDP Gathers: In the context of CDP gathers the round-
trip travel time is given (8.26) for a homogeneous medium and by (8.36) for a
layered medium. The only difference is that in place of wave speed of a
homogenous medium we have root mean square speed, crms . To estimate the
wave speed or rms wave speed consider the travel time at two distinct sensor
locations and compute the difference of the squares

© 2001 by CRC Press LLC

 x12 − x22
T x21 − T x22 = (8.41)
c2
from which the wave speed can be obtained as

x12 − x22
c= (8.42)
T x21 − T x22

For estimation of crms we use an identical approach. When the reflector is

sloping, we use the relation between the round-trip travel time and the source
sensor separation given by (8.28 & 8.29). Consider two sensors, equidistant
from the source, one located downslope and the other upslope. From (8.31) it is
clear that the slope of the reflector does not affect the average round-trip travel
time. We can now use (8.31) in place of (8.26) for the estimation of wave
speed.
Interval Speed: There is an interesting possibility of estimating the wave speed
of a particular layer. Consider crms at two different depths having N and N+1
layers. From (8.36) we have

1 ∆zk 2
N
2
crms N
=
T0 N
∑
k =1 ck
ck

N +1
1 ∆zk 2
2
crms N +1
=
T0 N +1
∑
k =1 ck
ck

By subtraction we obtain an estimate of the wave speed in the N+1st layer

2
crms T
N +1 0 N +1
− crms
2
T
N 0 N
cN +1 = (8.43)
∆z N +1

Now consider an interval containing p layers. Again from (8.36) we obtain after
subtraction

∑c
k =1
N +k ∆z N + k = crms
2
T
N+p 0 N+p
− crms
2
T
N 0 N
(8.44a)

which may be rewritten in a form,

© 2001 by CRC Press LLC

 1 ∆z N + k 2
p 2
crms T − crms
2
T
T0 N + p − T0 N
∑
k =1 c N + k
c N +k = N+p 0 N+p

T0 N + p − T0 N
N 0 N

By definition, the right hand side is the interval rms wave speed in the interval
containing p layers
1 p
∆z N + k 2
c2
rms p =
T0 N + p − T0 N
∑
k =1 c N + k
cN + k

(8.44b)
2
crms T
N+p 0 N+p
− crms
2
T
N 0 N
=
T0 N + p − T0 N

8.5.2: Estimation in Presence of Errors: Accurate knowledge of the wave speed

is essential not only for imaging but also for characterization of the medium.
The geophysical literature is proliferated with references on wave speed
estimation. Several different approaches have been tried to estimate this
important parameter. Briefly, correlation between the sensor outputs was used
to compute the so-called velocity spectrum in [16, 17]; minimum entropy
approach was suggested in [18], wavefield extrapolation in [19] and maximum
likelihood estimate of rms wave speed was suggested in [20]. Our interest shall
be limited to just two approaches, namely, maximum likelihood and focusing
with optimum wave speed.
Maximum Likelihood: We model the measurement error in T x , as an additive
random variable, that is,

xi2
T̂ x i = T02 + 2
+ τi (8.45)
crms

where τ i is a measurement error in T̂ x i , assumed to be a zero mean Gaussian

random variable. We introduce the following vectors:

{
T̂ x = col T̂ x1 , T̂ x 2 ,..., T̂ x N }
 x2 x2 x2 
T m = col  T02 + 21 , T02 + 22 ,..., T02 + 2N 
 crms crms crms 
τ = col{τ i , i = 1,2,..., N }

© 2001 by CRC Press LLC

 2
Further, let crms be a random variable with probability density function
2
pdf( crms ). We would like to maximize conditional probability density function
2
pdf (crms T̂ x ). Since

2 2
pdf (T̂ x crms ) pdf (crms )
pdf (c 2
rms T̂ x ) = (8.46)
pdf (T̂ x )

it is sufficient to maximize the numerator in (8.46). For this note that

2
pdf (T̂ x crms ) = pdf (τ)
(8.47a)

exp − (T̂ x − T m ) H C−1 (T̂ x − T m )

1 1
= N
2  2 
(2π) C
2
where C is covariance matrix of measurement errors. Further, crms is also
2
assumed to be a Gaussian random variable with mean equal to crms and standard
deviation σ

2
1 1  (c 2 − c 2 ) 
pdf (c ) =2
exp(−  rms rms  ) (8.47b)
σ
rms
2πσ 2 

Substitute (8.47a) and (8.47b) in the numerator of (8.46), which is then

2
maximized with respect to crms . We shall maximize the logarithm (natural) of
the numerator

 −1  (crms
2
− crms
2
)  
2

max (T̂ x − T m ) C (T̂ x − T m ) + 

H
  (8.48)
  σ  w.r.t.c 2
rms

2
Differentiate (8.48) with respect to crms and set the derivative to zero to obtain

∂T m H −1 ∂T m
2
(crms − crms
2
)
3T̂ Hx C−1 − 3T C − 2 =0 (8.49)
∂(crms ) ∂(crms ) σ
2 m 2 2

where

© 2001 by CRC Press LLC

 ∂T m 1
=− 4 ×
∂(crms )
2
2crms
 
  (8.50)
 x12 x22 x N2  −1
col  2
, ,..., = 4 x
 T 2 + x1 x22 x N2  2crms
T 2
+ T 2
+
 0 crms 2 0 2
crms
0 2
crms 
 

Substituting for the derivative in (8.50) we obtain

[ ]
− crms
2 2
3 H −1 H −1 (crms )
T C x − T̂ C x = (8.51)
σ
4 m x 2
4crms

We shall now introduce a few approximations. Let the measurement errors be

uncorrelated. Then, the covariance function becomes C =
{ τ1 τ2 τN }
diag σ 2 ,σ 2 ,...,σ 2 where σ 2 is the variance of the error in it h
τi

measurement. With σ → ∞ the right hand side of (8.51) becomes zero.

2

Equation (8.51) may be simplified to

 
 
N
xi2  T̂ x i =0
∑ 2 
i =1 σ τ i
1 −
x2 
(8.52)
 T02 + 2i 
 crms 

which is equal to the minimum mean square error estimate [20]. Numerical
simulations presented in [20] indicate that the minimum mean square error
estimate is very close to the maximum likelihood estimate which is
2
computationally far more involved. Since the unknown quantity crms in (8.52)
occurs inside the square root term, it is not possible to explicitly solve for it.
x2
To overcome this, we shall introduce an approximation T02 << 2
, which
crms
1 xi2
enables us to replace the square root term by T0 (1 + ) and then solve
2 T02 crms
2

for the unknown,

© 2001 by CRC Press LLC

 Measurement Mean wave std of wave speed
errors (std) speed estimate
0.01 sec 1532.2 m/s 16.3 m/s
0.05 sec 1535.1 m/s 102.8 m/s
Table 8.1: Mean and standard deviation of wave speed estimates given by (8.46).
Following parameters were assumed: T0 =5.0 sec, c=1500 m/s, Maximum source-
receiver separation = 1500 meters, Number of source-receiver pairs=100.

N T̂ x i xi2 xi2
∑ 3 2
1 i =1 T0 σ τ i
2
ĉrms ≈ (8.53)
2 N x 2  T̂ x 
∑ 2 
i

i =1 σ τ i 
i
− 1
 T0 

The wave speed estimate given by (8.53), on account of approximation

x2
T 02 << 2 , suffers from a large error even for small measurement error. This
crms
is demonstrated in table 8.1. The denominator in (8.53) is strongly influenced
by the measurement errors for limited array aperture. As noted in the context of
focused beamformation, for best depth of focus the array aperture was required
to be four times the depth to the reflector. It is understood that a similar
requirement exists for wave speed estimation.
8.5.3 Focusing with Optimum Wave Speed: In the previous section (§ 8.3) we
saw how CDP gathers are focused at the point of reflection. We were required to
know the correct wave speed. We shall now assume that the wave speed is
unknown. In fact, we wish to find an optimum speed for which the focus is the
best, in the sense that the magnitude of the focused signal is the highest
possible. The response function of a focused beamformer, when the wave the
speed is unknown, may be obtained from (8.32b)

 ∆tn 2 ∆tˆn 2 
1 N −1 − jω  T 0 (1+( ) ) − T̂ 0 (1+( ) )
H N (ω) = ∑ e  
T0 T̂ 0
(8.54)
N n=0

∆x
where ∆tˆ = , ĉ is the assumed speed and ∆x is the basic unit of source
ĉ
∆x
and receiver separation. Note that ∆t = is the actual propagation time.
c

© 2001 by CRC Press LLC

 1
0.9
0.8
0.7
0.6
Response

0.5

0.4
0.3
0.2

0.1
0
1400 1450 1500 1550 1600
Wave speed in m/s
Figure 8.20: Response (magnitude square) of focused beamformer as a function of
wave speed. Maximum source sensor separation: 8.5 km (solid line) and 4.5 km
(dashed line). Angular frequency=62.84.

5.4

5.3

5.2
Peak position in sec

5.1

5.0
4.9

4.8

4.7

4.6
1450 1470 1490 1510 1530 1550
Wave speed in m/s

Figure 8.21: Focused beam position as a function of wave speed in m/s. All other
parameters are as in fig. 8.20. The maximum sensor separation is 4.5 km.

© 2001 by CRC Press LLC

 The maximum value of the response function (4.54) is one and it is achieved only
when the exponent is equal to zero, which requires T̂0 = T0 and ∆tˆ = ∆t or
2
ĉ = c . For the present we assume that T̂0 = T0 and evaluate H N (ω) for a
series of assumed wave speeds. The position of the peak changes with the
assumed speed. The correct position is obtained only when the assumed wave
speed is equal to the actual speed. The variation of the peak position is
illustrated in fig. 8.21.

§8.6 Exercises
1. A point scatterer is located at a depth l unit below the observation surface. A
transceiver, placed on the surface, records the round trip-travel time. Show that
this travel time is exactly the same as in the CDP experiment with a reflector
replacing the scattering point.
2. Common depth point (CDP) seismic gathers containing an echo from a
horizontal reflector are cross-correlated. Show that the maximum of the cross-
correlation function lies on a hyperbolic time-distance curve. This property is
the basis for a technique of seismic speed estimation known as velocity
spectrum [21].
3. Show that the slope of the line joining all reflected signals in a zero-offset
seismic data over an incline reflector is related to the slope of the reflector
through the following relation, tan α = sin θ 0 where α is the slope of the
line joining reflected signals and θ 0 is slope of the reflector.

References

1. A. J. Devaney, Inverse source and scattering problem in ultrasonics, IEEE

Trans., SU-30, pp. 355-64, 1983.
2. D. Loewenthal, L. Lu, R. Robertson, and J. Sherwood, The wave
propagation applied to migration, Geophy. Prospecting, vol. 24, pp. 380-399,
1976.
3. J. F. Claerbout, Toward unified theory of reflector mapping, Geophysics,
vol. 31, pp. 467-481, 1971.
4. R. H. Stolt, Migration by Fourier transforms, Geophysics, vol. 43, pp. 23-
48, 1978.
5. D. Lowenthal and I. D. Mufti, Reverse time migration in spatial frequency
domain, Geophysics, vol. 48, pp. 627-635, 1983.
6. G. A. McMechan, Migration by extrapolation of time dependent boundary
values, Geophysical Prospecting, vol.31, pp. 413-420, 1983.
7. R. Sun and G. A. McMechan, Pre-stack reverse time migration, Proc. IEEE,
vol. 74, pp. 457-467, 1989.
8. J. Claerbout, Fundamentals of Geophysical Data Processing, McGraw Hill,
New York, 1972.

© 2001 by CRC Press LLC

9. G. Cheng and S. Coen, The relationship between Born inversion and
migration for common midpoint stack data, Geophysics, vol. 49, pp. 2117-
2131, 1984.
10. E. A. Robinson, Frequency-domain analysis of multodimensional time
series data, in Handbook of Statistics, Eds: D. R. Brillinger and P. R.
Krishnaiah, vol. 3, pp. 321-342, Elsevier Science Pub. B. V., Amsterdam,
1983.
11. A. J. Berkhout, Seismic Migration- Imaging of Acoustic Energy by Wave
Field Extrapolation, Elsevier Amsterdam, 1980.
12. P. S. Schultz and J. W.C. Sherwood, Depth migration before stack,
Geophysics, vol. 45, pp. 376-393, 1980.
13. W. A. Schneider, Integral formulation for migration in two and three
dimensions, Geophysics, vol. 43, pp. 49-76, 1978.
14. R. A. Phinney and D. M. Jurdy, Seismic Imaging of deep crust, Geophysics,
vol. 44, pp.1637-1660, 1979.
15. K. Kido, H. Noto, A. Shima, and M. Abe, Detection of acoustic signal by
use of cross spectra between every pair of sensors distributed in space. IEEE,
ICASSP-86, pp. 2507-2510, 1986.
16. M. T. Taner and F. Koehler, Velocity spectra Digital computer derivation
and applications of velocity functions, Geophy, vol. 34, pp. 859-881, 1969.
17. W. A. Schneider and M Backus, Dynamic correlation analysis, Geophy,
vol. 33, pp. 105-126, 1968.
18. D. De Vries and A. J. Berkhout, Velocity analysis based on minimum
entropy, Geophysics, vol. 49, pp. 2132-2142, 1984.
19. O. Yilmaz and R. Chambers, Migration velocity analysis by wave field
extrapolation, vol. 49, pp. 1664-1674, 1984.
20. R. L. Kirlin, L.A.Dewey, and J. N. Bradely, Optimum seismic velocity
estimation, Geophysics, vol. 49, pp. 1861-1661, 1984.
21. L. C. Wood and S. Treitel, Seismic signal processing, Proc. IEEE, vol.
63, pp. 649-661, 1975.

© 2001 by CRC Press LLC