INTRODUCTION TO ADAPTIVE ARRAYS, 2ND EDITION

Thorough Modernization of THE Seminal Text-Reference on Adaptive Arrays

This second edition is an extensive modernization of the seminal introduction to the subject of adaptive
array sensor systems. With the number of applications of adaptive array sensor systems growing each
year, this look at the principles and fundamental techniques that are critical to these systems is more
important than ever before.

Introduction to Adaptive Arrays, 2nd Edition is organized as a tutorial, taking the readers by the hand
and leading them through the maze of jargon that often surrounds this highly technical subject. It is
easy to read and easy to follow as fundamental concepts are introduced with examples before more
current developments and techniques are introduced.

Problems at the end of each chapter serve both instructors and professional readers by illustrating and
extending the material presented in the text. Both students and practicing engineers will easily gain
familiarity with the modern contribution that adaptive arrays have to offer practical signal reception
systems.

AUDIENCE
Radar • Sonar • Communications • Seismology • Radio Astronomy

ABOUT THE AUTHORS

Robert Monzingo received the B.S.E.E. degree from Stanford University, the M.S.E.E. degree from
the University of Arizona, and the Sc.D. degree from Washington University, St. Louis. He has been
active in the areas of automatic control, radar, and communications for many years and is presently
associated with the Aerospace Corporation in El Segundo, CA as a Senior Engineering Specialist where
he is engaged in work on adaptive arrays, GPS systems, and satellite communication systems.

Randy Haupt is a Fellow of the IEEE and Applied Computational Electromagnetics Society (ACES) and
is a Senior Scientist at the Penn State Applied Research Lab. From 1999-2003 he was Professor and
Department Head of ECE at Utah State University. He also was a Professor of EE at the Air Force Academy
and the University of Nevada-Reno. He is a retired Lt. Col. from the US Air Force. He has many journal
articles, conference publications, and book chapters to his credit and is the author of Antenna Arrays:
A Computational Approach by Wiley (2010).

Thomas Miller currently works for Raytheon Corporation where he has been cited for leadership
in advanced early warning surveillance radar and adaptive signal processing. He is the author and
coauthor of multiple journal articles as well as contributor to several books.
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 i

Introduction to Adaptive Arrays

Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 ii
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 iii

Introduction to Adaptive Arrays

2nd Edition

Robert A. Monzingo
Randy L. Haupt
Thomas W. Miller

Raleigh, NC
[Link]
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 iv

Published by SciTech Publishing, Inc.

911 Paverstone Drive, Suite B
Raleigh, NC 27615
(919) 847-2434, fax (919) 847-2568
[Link]

Copyright © 2011 by SciTech Publishing, Raleigh, NC. All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any
means, electronic, mechanical, photocopying, recording, scanning or otherwise, except as permitted under Sections
107 or 108 of the 1976 United Stated Copyright Act, without either the prior written permission of the Publisher,
or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, 222 Rosewood
Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 646-8600, or on the web at [Link]. Requests to the
Publisher for permission should be addressed to the Publisher, SciTech Publishing, Inc., 911 Paverstone Drive, Suite
B, Raleigh, NC 27615, (919) 847-2434, fax (919) 847-2568, or email editor@[Link].

The publisher and the author make no representations or warranties with respect to the accuracy or completeness of
the contents of this work and specifically disclaim all warranties, including without limitation warranties of fitness
for a particular purpose.

Editor: Dudley R. Kay

Editorial Assistant: Katie Janelle
Production Manager: Robert Lawless
Typesetting: MPS Limited, a Macmillan Company
Cover Design: Brent Beckley
Printer: Sheridan Books, Inc.

Printed in the United States of America

10 9 8 7 6 5 4 3 2 1

ISBN: 978-1-891121-57-9

Library of Congress Cataloging-in-Publication Data

Monzingo, Robert A., 1938-
Introduction to adaptive arrays / Robert A. Monzingo, Randy L. Haupt,
Thomas W. Miller. – 2nd ed.
p. cm.
Includes bibliographical references.
ISBN 978-1-891121-57-9 (hardcover : alk. paper)
1. Antenna arrays. I. Haupt, Randy L. II. Miller, Thomas W. III. Title.
TK7871.6.M66 2011
621.382’4–dc22
2010047941
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 v

Dedication
Acknowledgements are generally used to thank those people who, by virtue
of their close association with and support of the extended effort required
to produce a book. Occasionally, however, it becomes apparent that rip-
ples across a great span of distance and time should receive their due. I
only recently learned that I am a direct ancestor of a Bantu warrior named
Edward Mozingo who, some 340 years ago, filed a successful lawsuit in the
Richmond County Courthouse thereby gaining his freedom after 28 years of
indentured servitude to become that rarest of creatures in colonial Virginia:
a free black man. While that accomplishment did not directly result in the
production of this volume, the determination and effort that it involved rep-
resent a major human achievement that, it seems to me, deserves to be
recognized, however belatedly.
With deep appreciation to Richard Haupt, a great brother.
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 vi
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 vii

Brief Contents

Preface xiv

PART I Adaptive Array Fundamental Principles: System

Uses, System Elements, Basic Concepts,
and Optimum Array Processing
1 Introduction 3
2 Adaptive Array Concept 29
3 Optimum Array Processing 81

PART II Adaptive Algorithms

4 Gradient-Based Algorithms 153
5 Direct Inversion of the Sample Covariance Matrix 239
6 Recursive Methods for Adaptive Array Processing 273
7 Cascade Preprocessors 303
8 Random Search Algorithms 335
9 Adaptive Algorithm Performance Summary 365

PART III Advanced Topics

10 Compensation of Adaptive Arrays 373
11 Direction of Arrival Estimation and Related Topics 421
12 Recent Developments in Adaptive Arrays 463

Appendix A: Frequency Response Characteristics of Tapped-Delay Lines 507

Appendix B: Complex Envelope Notation 511
Appendix C: Convenient Formulas for Gradient Operations 515
Appendix D: Useful Matrix Relations and the Schwartz Inequality 517
Appendix E: Multivariate Gaussian Distributions 519
Appendix F: Geometric Aspects of Complex Vector Relationships 525
Appendix G: Eigenvalues and Eigenvectors 527
Appendix H: Selected Answers 529
Index 537

vii
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 viii
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 ix

Contents

Preface xiv

PA R T I Adaptive Array Fundamental Principles: System

Uses, System Elements, Basic Concepts,
and Optimum Array Processing

1 Introduction 3
1.1 Motivation For Using Adaptive Arrays 4
1.2 Historical Perspective 5
1.3 Principal System Elements 6
1.4 Adaptive Array Problem Statement 7
1.5 Existing Technology 9
1.6 Organization of the Book 21
1.7 Summary and Conclusions 22
1.8 Problems 23
1.9 References 24

2 Adaptive Array Concept 29

2.1 Signal Environment 30
2.2 Array Element Spacing Considerations 33
2.3 Array Performance 45
2.4 Nulling Limitations Due to Miscellaneous Array Effects 58
2.5 Narrowband and Broadband Signal Processing Considerations 62
2.6 Adaptive Array Performance Measure—Coverage Improvement
Factor (CIF) 66
2.7 Summary and Conclusions 68
2.8 Problems 69
2.9 References 78

ix
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 x

x Contents

3 Optimum Array Processing 81

3.1 Mathematical Preliminaries 82
3.2 Signal Descriptions for Conventional and Signal Aligned Arrays 88
3.3 Optimum Array Processing for Narrowband Applications 90
3.4 Optimum Array Processing for Broadband Applications 103
3.5 Optimum Array Processing for Perturbed Propagation Conditions 121
3.6 Polarization Sensitive Arrays 124
3.7 Summary and Conclusions 130
3.8 Problems 131
3.9 References 146

PART II Adaptive Algorithms

4 Gradient-Based Algorithms 153

4.1 Introductory Concepts 154
4.2 The LMS Algorithm 158
4.3 The Howells–Applebaum Adaptive Processor 171
4.4 Introduction of Main Beam Constraints 191
4.5 Constraint for the Case of Known Desired Signal Power Level 199
4.6 The DSD Algorithm 201
4.7 The Accelerated Gradient Approach (AG) 209
4.8 Gradient Algorithm with Constraints 213
4.9 Simulation Results 224
4.10 Phase-Only Adaptive Nulling Using Steepest Descent 227
4.11 Summary and Conclusions 228
4.12 Problems 230
4.13 References 235

5 Direct Inversion of the Sample Covariance Matrix 239

5.1 The Direct Matrix Inversion (DMI) Approach 239
5.2 Diagonal Loading of the Sample Covariance Matrix 251
5.3 Factorization Methods 253
5.4 Transient Response Comparisons 257
5.5 Sensitivity to Eigenvalue Spread 262
5.6 Summary and Conclusions 262
5.7 Problems 263
5.8 References 270
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 xi

Contents xi

6 Recursive Methods for Adaptive Array Processing 273

6.1 The Weighted Least Squares Error Processor 273
6.2 Updated Covariance Matrix Inverse 277
6.3 Kalman Filter Methods for Adaptive Array Processing 284
6.4 The Minimum Variance Processor 291
6.5 Simulation Results 295
6.6 Summary and Conclusions 299
6.7 Problems 299
6.8 References 301

7 Cascade Preprocessors 303

7.1 Nolen Network Preprocessor 304
7.2 Interference Cancellation with a Nolen Network Preprocessor 311
7.3 Gram–Schmidt Orthogonalization Preprocessor 315
7.4 Simulation Results 324
7.5 Summary and Conclusions 328
7.6 Problems 328
7.7 References 332

8 Random Search Algorithms 335

8.1 Linear Random Search 336
8.2 Accelerated Random Search 341
8.3 Guided Accelerated Random Search 344
8.4 Genetic Algorithm 346
8.5 Comparison of Random Search Algorithms 353
8.6 Summary and Conclusions 358
8.7 Problems 359
8.8 References 362

9 Adaptive Algorithm Performance Summary 365

PA R T I I I Advanced Topics

10 Compensation of Adaptive Arrays 373

10.1 Array Errors 374
10.2 Array Calibration 377
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 25, 2010 21:3 xii

xii Contents

10.3 Broadband Signal Processing Considerations 380

10.4 Compensation for Mutual Coupling 396
10.5 Multipath Compensation 398
10.6 Analysis of Interchannel Mismatch Effects 406
10.7 Summary and Conclusions 415
10.8 Problems 416
10.9 References 418

11 Direction of Arrival Estimation and Related Topics 421

11.1 Periodogram 422
11.2 Capon’s Method 423
11.3 Music Algorithm 423
11.4 The Maximum Entropy Method 426
11.5 Comparing DOA Algorithm Spectra 436
11.6 Sequential Implementation of Bayes Optimal Array Processor 437
11.7 Parameter Estimation via Subspace Fitting 441
11.8 The Maximum Likelihood Estimate 443
11.9 Cramer–Rao Lower Bound on AOA Estimation Error 449
11.10 Fisher Information Matrix and CR Bound for General Cases 451
11.11 Summary and Conclusions 452
11.12 Problems 452
11.13 References 459

12 Recent Developments in Adaptive Arrays 463

12.1 Beam Switching 463
12.2 Space-Time Adaptive Processing 465
12.3 MIMO 473
12.4 Reconfigurable Antennas and Arrays 479
12.5 Performance Characteristics of Large Sonar Arrays 484
12.6 Adaptive Processing for Monopulse Tracking Antennas 486
12.7 Partially Adaptive Arrays 488
12.8 Summary and Conclusions 503
12.9 Problems 503
12.10 References 504

Appendix A: Frequency Response Characteristics of Tapped-Delay Lines 507

A.1 References 510

Appendix B: Complex Envelope Notation 511

B.1 References 513
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 25, 2010 21:3 xiii

Contents xiii

Appendix C: Convenient Formulas for Gradient Operations 515

C.1 References 516

Appendix D: Useful Matrix Relations and the Schwartz Inequality 517

D.1 References 518

Appendix E: Multivariate Gaussian Distributions 519

E.1 Real Gaussian Random Vectors 519
E.2 Complex Gaussian Random Vectors 521
E.3 References 524

Appendix F: Geometric Aspects of Complex Vector Relationships 525

F.1 Reference 526

Appendix G: Eigenvalues and Eigenvectors 527

Appendix H: Selected Answers 529
Index 537
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 xiv

Preface

This book is intended to serve as an introduction to the subject of adaptive array sensor
systems whose principal purpose is to enhance the detection and reception of certain
desired signals. Array sensor systems have well-known advantages for providing flexible,
rapidly configurable, beamforming and null-steering patterns. The advantages of array
sensor systems are becoming more important, and this technology has found applications
in the fields of communications, radar, sonar, radio astronomy, seismology and ultrasonics.
The growing importance of adaptive array systems is directly related to the widespread
availability of compact, inexpensive digital computers that make it possible to exploit
certain well-known theoretical results from signal processing and control theory to provide
the critical self-adjusting capability that forms the heart of the adaptive structure.
There are a host of textbooks that treat adaptive array systems, but few of them take
the trouble to present an integrated treatment that provides the reader with the perspective
to organize the available literature into easily understood parts. With the field of adaptive
array sensor systems now a maturing technology, and with the applications of these systems
growing more and more numerous, the need to understand the underlying principles of
such systems is a paramount concern of this book. It is of course necessary to appreciate
the limitations imposed by the hardware adopted to implement a design, but it is more
informative to see how a choice of hardware “fits” within the theoretical framework of
the overall system. Most of the contents are derived from readily available sources in the
literature, although a certain amount of original material has been included.
This book is intended for use both as a textbook at the graduate level and as a reference
work for engineers, scientists, and systems analysts. The material presented will be most
readily understood by readers having an adequate background in antenna array theory,
signal processing (communication theory and estimation theory), optimization techniques,
control theory, and probability and statistics. It is not necessary, however, for the reader
to have such a complete background since the text presents a step-by-step discussion of
the basic theory and important techniques required in the above topics, and appropriate
references are given for readers interested in pursuing these topics further. Fundamental
concepts are introduced and illustrated with examples before more current developments
are introduced. Problems at the end of each chapter have been chosen to illustrate and
extend the material presented in the text. These extensions introduce the reader to actual
adaptive array engineering problems and provide motivation for further reading of the
background reference material. In this manner both students and practicing engineers
may easily gain familiarity with the modern contributions that adaptive arrays have to
offer practical signal reception systems.
The book is organized into three parts. Part One (Chapters 1 to 3) introduces the
advantages that obtain with the use of array sensor systems, define the principal system
components, and develop the optimum steady-state performance limits that any array
system can theoretically achieve. This edition also includes two new topics that have
practical interest: the subject of a performance index to grade the effectiveness of the
overall adaptive system, and the important theme of polarization sensitive arrays. Part Two

xiv
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 xv

Preface xv

(Chapters 4 through 9) provides the designer with a survey of adaptive algorithms and
a performance summary for each algorithm type. Some important modern developments
in matrix inversion computation and random search algorithms are treated. With this
information available, the designer may then quickly identify those approaches most likely
to lead to a successful design for the signal environment and system constrains that are
of concern. Part Three (Chapters 10, 11, and 12) considers the problem of compensation
for adaptive array system errors that inevitably occur in any practical system, explores the
important topic of direction of arrival (DOA) estimation, and introduces current trends in
adaptive array research. It is hoped that this edition succeeds in presenting this exciting
field using mathematical tools that make the subject interesting, accessible, and appealing
to a wide audience.
The authors would like to thank Northrop Grumman (Dennis Lowes and Dennis
Fortner), the National Electronics Museum (Ralph Strong and Michael Simons), Material
Systems Inc. (Rick Foster), and Remcom Inc. (Jamie Knapil Infantolino) for providing
some excellent pictures.
Monzingo-7200014 monz7200014˙fm ISBN : XXXXXXXXXX November 24, 2010 16:30 xvi
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 1

PART I

Adaptive Array
Fundamental Principles:
System Uses, System
Elements, Basic Concepts,
and Optimum Array
Processing

CHAPTER 1 Introduction
CHAPTER 2 Adaptive Array Concept
CHAPTER 3 Optimum Array Processing:
Steady-State Performance Limits
and the Wiener Solution
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 3

CHAPTER

Introduction
1
' $
Chapter Outline
1.1 Motivation for Using Adaptive Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Principal System Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Adaptive Array Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Existing Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Organization of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.7 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8 Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
& %

An array of sensor elements has long been an attractive solution for severe reception
problems that commonly involve signal detection and estimation. The basic reason for
this attractiveness is that an array offers a means of overcoming the directivity and sen-
sitivity limitations of a single sensor, offering higher gain and narrower beamwidth than
that experienced with a single element. In addition, an array has the ability to control
its response based on changing conditions of the signal environment, such as direction
of arrival, polarization, power level, and frequency. The advent of highly compact, inex-
pensive digital computers has made it possible to exploit well-known results from signal
processing and control theory to provide optimization algorithms that automatically ad-
just the response of an adaptive array and has given rise to a new domain called “smart
arrays.” This self-adjusting capability renders the operation of such systems more flexible
and reliable and (more importantly) offers improved reception performance that would
be difficult to achieve in any other way. This revised edition acquaints the reader with
the historical background of the field and presents important new developments that have
occurred over the last quarter century that have improved the utility and applicability of
this exciting field.

3
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 4

4 CHAPTER 1 Introduction

1.1 MOTIVATION FOR USING ADAPTIVE ARRAYS

An array consists of two or more sensors in which the signals are coherently combined
in a way that increases the antenna’s performance. Arrays have the following advantages
over a single sensor:
1. Higher gain. The gain is higher, because the array gain is on the order of the number
of elements in the array. Higher resolution or narrower main beam follows from the
larger aperture size.
2. Electronic beam scanning. Moving large antennas to steer the main beam is slow. Arrays
with phase shifters at each element are able to steer the beam without mechanical
motion, because the signals are made to add in phase at the beam steering angle.
3. Low sidelobes. If the desired signal enters the main beam while interfering signals
enter the sidelobes, then lowering the sidelobes relative to the main beam improves the
signal to interference ratio.
4. Multiple beams. Certain array feeds allow simultaneous multiple main beams.
5. Adaptive nulling. Adaptive arrays automatically move nulls in the directions of signals
over the sidelobe region.
On the other hand, these advantages are countered by the significant disadvantages of
increased cost and complexity.
Conventional signal reception systems are susceptible to degradation in signal-to-
noise ratio (SNR) [or more generally, the signal-to-interference plus noise ratio (SINR)]
performance because of the inevitable presence in the signal environment of undesired
“noise” signals that enter the sidelobes or main beam of an array. Signals that interfere
with the desired signal include deliberate electronic countermeasures (ECMs), nonhos-
tile radiofrequency interference (RFI), clutter, multipath, and natural noise sources. The
resulting SNR degradation is further aggravated by array motion, poor siting, sensor fail-
ures, and a nonstationary interference environment. As traffic in the spectrum increases,
the suppression of interference becomes even more paramount.
Adaptive arrays improve the reception of desired signals in the presence of interference
signals in radar, sonar, seismic, and communications systems. They automatically sense
the presence of interference and suppress them while simultaneously enhancing desired
signal reception without prior knowledge of the signal–interference environment. Adaptive
arrays are designed to complement other interference suppression techniques, such as low
sidelobes, spread-spectrum techniques, and high directivity.
An adaptive array has a computer algorithm that controls the signal levels at the
elements until a measure of the quality of the array performance improves. It adjusts its
pattern to form nulls, to modify gain, to lower sidelobes, or to do whatever it takes to
improve its performance. An adaptive array offers enhanced reliability compared with
that of a conventional array. When a single sensor element in a conventional array fails,
the sidelobe structure of the array pattern degrades. With an adaptive array, however,
the remaining operational sensors in the array automatically adjust to restore the pattern.
Adaptive arrays are more reliable than conventional arrays, because they fail gracefully.
The reception pattern of an array in place on an aircraft or ship is often quite different
from the array pattern measured in isolation (in an anechoic chamber) as a result of signal
scattering that occurs from vehicle structures located in the vicinity of the antenna. An
adaptive array often yields successful operation even when antenna patterns are severely
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 5

1.2 Historical Perspective 5

distorted by near-field effects. The adaptive capability overcomes any distortions that occur
in the near field (i.e., at distances from the radiating antenna closer than λ/2π where λ is
the wavelength) and merely responds to the signal environment that results from any such
distortion. Likewise, in the far field (at distances from the radiating antenna greater than
2λ) the adaptive antenna is oblivious to the absence of any distortion.
An adaptive array improves the SNR by preserving the main beam that points at the
desired signal at the same time that it places nulls in the pattern to suppress interference
signals. Very strong interference suppression is possible by forming pattern nulls over a
narrow bandwidth. This exceptional interference suppression capability is a principal ad-
vantage of adaptive arrays compared to waveform processing techniques, which generally
require a large spectrum-spreading factor to obtain comparable levels of interference sup-
pression. Sensor arrays possessing this key automatic response capability are sometimes
referred to as “smart” arrays, since they respond to far more of the signal information
available at the sensor outputs than do more conventional array systems.
The capabilities provided by the adaptive array techniques to be discussed in this
book offer practical solutions to the previously mentioned realistic interference problems
by virtue of their ability to sort out and distinguish the various signals in the spatial do-
main, in the frequency domain, and in polarization. At the present time, adaptive nulling
is considered to be the principal benefit of the adaptive techniques employed by adap-
tive array systems, and automatic cancellation of sidelobe jamming provides a valuable
electronic counter–countermeasure (ECCM) capability for radar systems. Adaptive arrays
are designed to incorporate more traditional capabilities such as self-focusing on receive
and retrodirective transmit. In addition to automatic interference nulling and beam steer-
ing, adaptive imaging arrays may also be designed to obtain microwave images having
high angular resolution. It is useful to call self-phasing or retrodirective arrays adaptive
transmitting arrays to distinguish the principal function of such systems from an adaptive
receiving array, the latter being the focus of this book.

1.2 HISTORICAL PERSPECTIVE

The term adaptive antenna was first used by Van Atta [1] and others [2] to describe a
self-phasing antenna system that automatically reradiates a signal in the direction from
which it was received. This “retrodirective” system works without prior knowledge of the
signal direction. Retrodirective arrays in point-to-point satellite communications systems
automatically return a strong pilot signal to overcome the usual beamwidth (and consequent
directivity) limitations.
The development of the phase-lock loop was another major step that made possible the
self-steering (or self-phasing) type of adaptive array [3]. A self-phased array has each of
the array elements independently phased, based on information obtained from the received
signals. For example, several large-aperture receiving antennas with slaved steering can be
self-phased on received signals from satellites or space vehicles so the effective receiving
aperture is the sum of the individual apertures of all participating antennas.
In the early 1960s the key capability of adaptive interference nulling was recog-
nized and developed by Howells [4,5]. Subsequently, Applebaum established the con-
trol law associated with the Howells adaptive nulling scheme by analyzing an algorithm
that maximizes a generalized SNR [6]. Concurrently, the capability of self-training or
self-optimizing control was applied to adaptive arrays by Widrow and others [7-9]. The
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 6

6 CHAPTER 1 Introduction

self-optimizing control work established the least mean square (LMS) error algorithm that
was based on the method of steepest descent. The Applebaum and the Widrow algorithms
are very similar, and both converge toward the optimum Wiener solution.
The use of sensor arrays for sonar and radar signal reception had long been common
practice by the time the early adaptive algorithm work of Applebaum and Widrow was
completed [10,11]. Early work in array processing concentrated on synthesizing a “desir-
able” pattern. Later, attention shifted to the problem of obtaining an improvement in the
SNR [12–14]. Seismic array development commenced about the same period, so papers
describing applications of seismic arrays to detect remote seismic events appeared during
the late 1960s [15–17].
The major area of current interest in adaptive arrays is their application to problems
arising in radar and communications systems, where the designer almost invariably faces
the problem of interference suppression [18]. A second example of the use of adaptive
arrays is that of direction finding in severe interference environments [19,20]. Another
area in which adaptive arrays are proving useful is for systems that require adaptive
beamforming and scanning in situations where the array sensor elements must be organized
without accurate knowledge of element location [21]. Furthermore, large, unstructured
antenna array systems may employ adaptive array techniques for high angular resolution
imaging [22,23]. Adaptive antennas are a subset of smart antennas and include topics
such as multiple input, multiple output (MIMO) [24], element failure compensation [25],
reconfigurable antennas [26], and beam switching [27].

1.3 PRINCIPAL SYSTEM ELEMENTS

Figure 1-1 shows a diagram of an adaptive array. It consists of the sensor array, the
beamforming network, and the adaptive processor that adjusts the variable weights in the
beamforming network. The array design depends on the propagation medium in which
the array operates, the frequency spectrum of interest, and the user’s knowledge of the
operational signal environment.
FIGURE 1-1 z
Functional diagram Signal #2
of an N-element Signal #1 q Signal #3
adaptive array.

Sensor
array
Adaptive
algorithm
w1 w2 w3 … wN

Beamforming
network Σ

Array output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 7

1.4 Adaptive Array Problem Statement 7

The array consists of N sensors designed to receive (and transmit) signals in the
propagation medium. The sensors are arranged to give adequate coverage (pattern gain)
over a desired spatial region. The selection of the sensor elements and their physical
arrangement place fundamental limitations on the ultimate capability of the adaptive array
system. The output of each of the N elements goes to the beamforming network, where
the output of each sensor element is first multiplied by a complex weight (having both
amplitude and phase) and then summed with all other weighted sensor element outputs to
form the overall adaptive array output signal. The weight values within the beamforming
network (in conjunction with the sensor elements and their physical arrangement) then
determine the overall array pattern. It is the ability to shape this overall array pattern that
in turn determines how well the specified system requirements can be met for a given
signal environment.
The exact structure of the adaptive algorithm depends on the degree of detailed in-
formation about the operational signal environment that is available to the array. As the
amount of a priori knowledge (e.g., desired signal location, jammer power levels) concern-
ing the signal environment decreases, the adaptive algorithm selected becomes critical to
a successful design. Since the precise nature and direction of all signals present as well
as the characteristics of the sensor elements are not known in practice, the adaptive algo-
rithm must automatically respond to whatever signal environment (within broad limits)
confronts it. If any signal environment limits are known or can reasonably be construed,
such bounds are helpful in determining the adaptive processor algorithm used.

1.4 ADAPTIVE ARRAY PROBLEM STATEMENT

The fundamental problem facing the adaptive array designer is to improve the reception of
a desired signal in the presence of undesired interfering signals. The terms desired signal
and interfering signals imply that the characteristics of these two signal classes are different
in some respect and that this difference provides the key to improving the desired signal
reception. For example, if the direction of arrival of the desired signal is known (or can be
deduced), then any signals arriving from different directions are suppressed by forming
array pattern nulls in those directions. Likewise, if the interference signals are outside
the desired bandwidth, then the interference signals are eliminated by band-pass filtering.
Certain characteristics of the desired signal distinguish it from interference signals, so it
is reasonable to assume that the nature of the desired signal is known even though certain
signal parameters (e.g., direction of arrival, amplitude, phase) must be estimated. If the
designer were solely concerned with suppressing interfering signals, then desired signal
reception might suffer. Likewise, if desired signal enhancement were the sole focus of
attention, then interference signal reception might also be enhanced. Therefore, the twin
(and sometimes conflicting) objectives of desired signal enhancement and interference
signal suppression are sought so that the overall desired signal reception performance
is improved. In many cases, the overall reception performance is best measured by the
output SNR. For passive sensor systems, however, the basic problem is that of determining
whether a desired signal is present in a background of ambient noise and interfering
signals. Determining signal presence or absence requires a decision that is not provided
simply by maximizing the output SNR, and statistical decision theory provides solutions
to problems of this kind that minimize the risk associated with incorrect decisions. The
optimum processors prescribed by statistical decision theory are closely related to those
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 8

8 CHAPTER 1 Introduction

obtained by maximizing the output SNR, so there is an underlying unity to problems that
initially appear to be quite different.
An adaptive array design includes the sensor array configuration, beamforming net-
work implementation, signal processor, and adaptive algorithm that enables the system to
meet several different requirements on its resulting performance in as simple and inexpen-
sive a manner as possible. The system performance requirements are conveniently divided
into two types: transient response and steady-state response. Transient response refers to
the time required for the adaptive array to successfully adjust from the time it is turned on
until reaching steady-state conditions or successfully adjusting to a change in the signal
environment. Steady-state response refers to the long-term response after the weights are
done changing. Steady-state measures include the shape of the overall array pattern and
the output signal-to-interference plus noise ratio. Several popular performance measures
are considered in detail in Chapter 3. The response speed of an adaptive array depends on
the type of algorithm selected and the nature of the operational signal environment. The
steady-state array response, however, can easily be formulated in terms of the complex
weight settings, the signal environment, and the sensor array structure.
A fundamental trade-off exists between the rapidity of change in a nonstationary
noise field and the steady-state performance of an adaptive system: generally speaking,
the slower the variations in the noise environment, the better the steady-state performance
of the adaptive array. Any adaptive array design needs to optimize the trade-off between
the speed of adaptation and the accuracy of adaptation.
System requirements place limits on the transient response speed. In an aircraft com-
munication system, for example, the signal modulation rate limits the fastest response
speed (since if the response is too fast, the adaptive weights interact with the desired
signal modulation). Responding fast enough to compensate for aircraft motion limits the
slowest speed.
The weights in an adaptive array may be controlled by any one of a variety of different
algorithms. The “best” algorithm for a given application is chosen on the basis of a host
of factors including the signal structures, the a priori information available to the adaptive
processor, the performance characteristics to be optimized, the required speed of response
of the processor, the allowable circuit complexity, any device or other technological limi-
tations, and cost-effectiveness.
Referring to Figure 1-1, the received signal impinges on the sensor array and arrives
at each sensor at different times as determined by the direction of arrival of the signal
and the spacing of the sensor elements. The actual received signal for many applications
consists of a modulated carrier whose information-carrying component consists only of
the complex envelope. If s(t) denotes the modulated carrier signal, then s̃(t) is commonly
used to denote the complex envelope of s(t) (as explained in Appendix B) and is the only
quantity that conveys information. Rather than adopt complex envelope notation, however,
it is simpler to assume that all signals are represented by their complex envelopes so the
common carrier reference never appears explicitly. It is therefore seen that each of the N
channel signals xk (t) represents the complex envelope of the output of the element of a
sensor array that is composed of a signal component and a noise component, that is,

xk (t) = sk (t) + n k (t), k = 1, 2, . . . , N (1.1)

In a linear sensor array having equally spaced elements and assuming ideal propagation
conditions, the sk (t) are determined by the direction of the desired signal. For example, if
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 9

1.5 Existing Technology 9

the desired signal direction is located at an angle θ from mechanical boresight, then (for
a narrowband signal)
 
2πkd
sk (t) = s(t) exp j sin θ (1.2)
λ
where d is the element spacing, λ is the wavelength of the incident planar wavefront, and
it is presumed that each of the sensor elements is identical.
For the beamforming network of Figure 1-1, the adaptive array output signal is written
as

N
y(t) = w k xk (t) (1.3)
k=1

Equation (1.3) can be conveniently expressed in matrix notation as

y(t) = wT x = xT w (1.4)

where the superscript T denotes transpose, and the vectors w and x are given by

wT = [w 1 w 2 . . . w N ] (1.5)
xT = [x1 x2 . . . x N ] (1.6)

Throughout this book the boldface lowercase symbol (e.g., a) denotes a vector, and a
boldface uppercase symbol (e.g., A) denotes a matrix.
The adaptive processor must select the complex weights, w k , to optimize a stipulated
performance criterion. The performance criterion that governs the operation of the adap-
tive processor is chosen to reflect the steady-state performance characteristics that are of
concern. The most popular performance measures that have been employed include the
mean square error [9,28–31]; SNR ratio [6,14,32–34]; output noise power [35]; maximum
array gain [36,37]; minimum signal distortion [38,39]; and variations of these criteria
that introduce various constraints into the performance index [16,40–43]. In Chapter 3,
selected performance measures are formulated in terms of the signal characterizations
of (1.1)–(1.4). Solutions are found that determine the optimum choice for the complex
weight vector and the corresponding optimum value of the performance measure. The
operational signal environment plays a crucial role in determining the effectiveness of
the adaptive array to operate. Since the array configuration has pronounced effects on the
resulting system performance, it is useful to consider sensor spacing effects before pro-
ceeding with an analysis using an implicit description of such effects. The consideration
of array configuration is undertaken in Chapter 2.

1.5 EXISTING TECHNOLOGY

In any echo-ranging system, the maximum detection range Rmax determines the minimum
period between consecutive pulses Tmin [and hence the pulse repetition frequency (PRF)],
according to
2Rmax
Tmin = (1.7)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 10

10 CHAPTER 1 Introduction

FIGURE 1-2 t
Pulse modulated
carrier signal.

Carrier signal

where  is the velocity of propagation of the transmitted signal. For underwater applications
the velocity of sound in water varies widely with temperature, although a nominal value
of 1,500 m/sec can be used for rough calculations. The velocity of electromagnetic wave
propagation in the atmosphere can be taken approximately to be the speed of light or
3 ×108 m/sec.
If the range discrimination capability between targets is to be rd , then the maximum
pulse length tmax (in the absence of pulse compression) is given by
2rd
tmax = (1.8)

It will be noted that rd also corresponds to the “blind range”— that is, the range within
which target detection is not possible. Since the signal bandwidth ∼ = 1/pulse length, the
range discrimination capability determines the necessary bandwidth of the transducers
and their associated electrical channels.
The transmitted pulses form a pulse train in which each pulse modulates a carrier fre-
quency as shown in Figure 1-2. The carrier frequency f 0 in turn determines the wavelength
of the propagated wavefront since

λ0 = (1.9)
f0
where λ0 is the wavelength. For sonar systems, frequencies in the range 100–100,000
Hz are commonly employed [44], whereas for radar systems the range can extend from
a few megahertz up into the optical and ultraviolet regions, although most equipment is
designed for microwave bands between 1 and 40 GHz. The wavelength of the propagated
wavefront is important because the array element spacing (in units of λ) is an important
parameter in determining the array pattern.

1.5.1 Radar Technology

There has been a steady increase in the demand for increased radar system performance
and additional capability for both military and civilian purposes, and the vast number
of applications of modern radar technology precludes anything more than the briefest
mention of the major areas in which radar systems are found [45]. Military applications
may very well involve a number of requirements that in the past would have involved a
separate radar system for each different requirement. For example, a fire control system
radar may be required to search large volumes of space, to detect and track both high-
and low-speed targets ranging from very low to extremely high altitudes, to provide fire
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 11

1.5 Existing Technology 11

Antenna FIGURE 1-3

array Simplified block
diagram of a typical
Circulator radar system.
Signal
Receiver Display
processor

Scan Transmitter and

pattern stable local oscillator Control
generator (STALO)

control for both missiles and guns against both airborne and ground (or sea) targets,
and additionally to provide navigational aid and perform reconnaissance. Current civil
aeronautical needs include air traffic control, collision avoidance, instrument approach
systems, weather sensing, and navigational aids. Additional applications in the fields of
law enforcement, transportation, and Earth resources are just beginning to grow to sizable
proportions [48].
Figure 1-3 is a block diagram of a typical radar system. These major blocks and their
corresponding functions are described in Table 1-1 [46]. The antenna, receiver, and signal
processing blocks are of primary interest for our purposes, and these are now each briefly
discussed in turn.

[Link] Radiating Elements and Antenna Arrays

The vast frequency range and power over which modern radar systems operate have led
to an astonishing variety of radiator elements ranging from parabolic reflectors to horns,
dipoles, bow ties [47], multiturn loops [48], spirals [49], log periodics [50], microstrip
patches [51], and Vivaldis [52]. Large antenna apertures result in narrow beamwidths that
are required for long-range detection and high resolution if targets close to one another
are to be distinguished. Microwave frequencies are by far the most popular for radar
applications since antenna apertures of relatively small physical size (but large in terms
of wavelengths) are reasonable to build.
The antenna type selected for a radar application usually differs from one selected for
a communications system. Shaped beam patterns that can be scanned are most popular for
radar uses, whereas most communication applications require beams designed for omnidi-
rectional coverage or for fixed point-to-point transmission. The earliest radars (developed

TABLE 1-1 Functions of the Radar Blocks in Figure 1-3

Block Function
Transmitter Generates high power RF waveform
Antenna Determines direction and shape of transmit-and-receive beam
Receiver Provides frequency conversion and low-noise amplification
Signal processing Provides target detections, target and clutter tracking,
and target trajectory estimates
Display Converts processed signals into meaningful tactical information
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 12

12 CHAPTER 1 Introduction

FIGURE 1-4 FuG

65 Wurzburg Riese
radar antenna
(Courtesy of the
National Electronics
Museum).

during World War II) operated in the very high frequency (VHF) and ultra high frequency
(UHF) bands. Sometimes, a parabolic dish was used, such as the FuG 65 Wurzburg Riese
radar antenna in Figure 1-4. Its 3 m parabolic dish operated at 560 MHz and was used
to guide German intercept fighters during WWII [53]. Arrays, such as the SCR-270 in
Figure 1-5, were used by the United States in WWII for air defense [54]. It has four rows
of eight dipoles that operate at 110 MHz. After WWII, radar and communications systems
began operating at higher frequencies. Different types of antennas were tried for various
systems. A microwave lens was used in the Nike AJAX MPA-4 radar shown in Figure 1-6
[55]. In time, phased array antennas became small enough to place in the nose of fighter
airplanes. In the 1980s, the AN/APG-68 (Figure 1-7) was used in the F-16 fighter [56]. The
array is a planar waveguide with slots for elements. Active electronically scanned arrays
(AESA) provide fast wide angle scanning in azimuth and elevation and include advanced
transmit/receive modules [57]. An example is the AN/APG-77 array for the F-22 fighter
shown in Figure 1-8.

FIGURE 1-5
SCR-270 antenna
array (Courtesy of
the National
Electronics
Museum).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 13

1.5 Existing Technology 13

FIGURE 1-6
Waveguide lens
antenna for the
Nike AJAX MPA-4
radar (Courtesy of
the National
Electronics
Museum).

The two most common forms of antenna arrays for radar applications are the linear
array and the planar array. A linear array consists of antenna elements arranged in a straight
line. A planar array, on the other hand, is a two-dimensional configuration in which the
antenna elements are arranged to lie in a plane. Conformal arrays lie on a nonplanar
surface. The linear array generates a fan beam that has a broad beamwidth in one plane
and a narrow beamwidth in the orthogonal plane. The planar array is most frequently used
in radar applications where a pencil beam is needed. A fan-shaped beam is easily produced
by a rectangular-shaped aperture. A pencil beam may easily be generated by a square- or
circular-shaped aperture. With proper weighting, an array can be made to simultaneously
generate multiple search or tracking beams with the same aperture.
Array beam scanning requires a linear phase shift across the elements in the array.
The phase shift is accomplished by either software in a digital beamformer or by hardware
phase shifters. A phase shifter is often incorporated in a transmit/receive module. Com-
mon phase-shifter technology includes ferroelectrics, monolithic microwave integrated

FIGURE 1-7
AN/APG-68 array
(Courtesy of
Northrop Grumman
and available at the
National Electronics
Museum).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 14

14 CHAPTER 1 Introduction

FIGURE 1-8
AN/APG-77 array
(Courtesy of
Northrop Grumman
and available at the
National Electronics
Museum).

circuit (MMIC), ferrites, and micro-electro-mechanical systems (MEMS). Some of their

characteristics are shown in Table 1.2.

[Link] Receivers
A receiver design based on a matched filter or a cross-correlator maximizes the SNR in
the linear portion of the receiver. Different types of receivers that have been employed
in radar applications include the superheterodyne, superregenerative, crystal video, and
tuned radio frequency (TRF) [58]. The most popular and widely applied receiver type is
the superheterodyne, which is useful in applications where simplicity and compactness
are especially important. A received signal enters the system through the antenna, then
passes through the circulator and is amplified by a low-noise RF amplifier. Following RF
amplification, a mixer stage is entered to translate the RF to a lower intermediate frequency
(IF) where the necessary gain is easier to obtain and filtering is easier to synthesize. The
gain and filtering are then accomplished in an IF amplifier section.

[Link] Signal Processing

Having maximized the SNR in the receiver section, the next step is to perform two basic
operations by signal processing as follows: (1) detection of the presence of any targets, and

TABLE 1-2 Phase-Shifter Characteristics [57]

Type Ferroelectric MMIC Ferrite MEMS

Cost Low High Very high Low
Reliability Good Very good Excellent Good
Power handling >1W > 10 W kW < 50 mW
Switch speed ns ns 10 to 100 μs 10 to 100 μs
Direct current power consumption Low low High Negligible
Size Small Small Large Small
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 15

1.5 Existing Technology 15

(2) extraction of information from the received waveform to obtain target trajectory data
such as position and velocity. Detecting a signal imbedded in a noise field is treated by
means of statistical decision theory. Similarly, the problem of the extraction of information
from radar signals can be regarded as a problem concerning the statistical estimation of
parameters.

1.5.2 Sonar Technology

Operational active sonar systems may be classified as (1) search light sonar, (2) scanning
sonar, or (3) rotational directional transmission (RDT) sonar [59]. Searchlight sonar has
very narrow transmit and receive beams and provides an azimuth search capability by
mechanically rotating the directional hydrophone array. Since the array is mechanically
scanned and aimed, the data rate is correspondingly low, and the system does not provide
a multiple-target detection and tracking capability. The requirement for mechanical direc-
tional training also limits the array size, so that operational frequencies are usually greater
than 15 kHz, thereby increasing attenuation loss.
Scanning sonar systems overcome the data rate limitation of searchlight sonars by
transmitting an omnidirectional, short-duration pulse and using electronic means to rapidly
rotate a narrow receive beam continuously over a 360◦ azimuthal sector. The receiving
beam output is presented to a panoramic display called a plan position indicator (PPI)
that is used extensively in both radar and sonar systems. Scanning-type sonar systems
thereby provide a multiple-target detection and tracking capability, and lower operating
frequencies can be used, thereby decreasing attenuation losses. The scan speed of the
receive beam is a compromise between the desired target resolution and the maximum
receiver bandwidth (or minimum input SNR) that is permissible.
An RDT sonar system is characterized by RDT and a scanned preformed beam (PFB)
receiver. Consequently, high transmitting directivity and a high data rate are accompanied
by a low operational frequency. An RDT system combines the best features of searchlight
and scanning sonars. A PFB receiver can have a smaller bandwidth than a scanning
receiver, thereby improving the SNR. Furthermore, a PFB receiver can be corrected for
Doppler due to own ship’s motion employing a method called own Doppler nullifying
(ODN), whereas a scanning receiver cannot.
The principal elements of a sonar receiver (Figure 1-9) are the hydrophone array, the
beamformer, the signal processor, and the information processor. Each of these principal
elements (except for the information processor, which involves display formatting and
other command and control functions) is briefly discussed in turn.

[Link] Sonar Transducers and Hydrophones

A hydrophone produces an output voltage proportional to the acoustic signals incident
on it; whereas a transducer generates and receives sound. For underwater applications,
a very wide frequency range is involved—from about 10 Hz to more than 1 MHz [60].

FIGURE 1-9
Control Sonar receiver block
diagram.

Array Beamformer Signal processor Information processor

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 16

16 CHAPTER 1 Introduction

a
c
b

g
d
f

FIGURE 1-10 Various acoustic transducers. a: Seabed mapping—20 kHz multibeam receive
array module; 8 shaded elements per module; 10 modules per array. b: Subbottom profiling
(parametric)—200 kHz primary frequency; 25 kHz secondary frequency. c: Port and harbor
Security—curved 100 kHz transmit/receive array. d. Obstacle avoidance—10 × 10 planar re-
ceive array with curved transmitter. e. ACOMMS—Broadband piezocomposite transducers for
wideband communication signals. f. AUV FLS—high-frequency, forward-looking sonar array.
g. Mine hunting—10 × 10 transmit/receive broadband array; available with center frequen-
cies between 20 kHz to 1MHz h. Side scan—multibeam transmit/receive array (Courtesy of
Materials Systems Inc.).

A transmitting power ranging from a few acoustic watts up to several thousand acoustic
watts at ocean depths up to 20,000 ft can be achieved [61]. Figure 1-10 shows a sampling
of different acoustic transducers manufactured by Materials Systems Inc. for various
applications.
The basic physical mechanisms most widely used in transducer technology include
the following [62]:
1. Moving coil. This is long familiar from use as a loudspeaker in music reproduction
systems and used extensively in water for applications requiring very low frequencies.
2. Magnetorestrictive. Magnetic materials vibrate in response to a changing magnetic
field. Magnetorestrictive materials are rugged and easily handled, and magnetorestric-
tive transducers were highly developed and widely used during World War II.
3. Piezoelectric. The crystalline structure of certain materials results in mechanical vi-
bration when subjected to an alternating current or an oscillating electric field. The
relationship between mechanical strain and electric field is linear. Certain ceramic ma-
terials also exhibit a similar effect and have outstanding electromechanical properties.
Consequently, over the last decade the great majority of underwater sound transducers
have been piezoceramic devices that can operate over a wide frequency band and have
both high sensitivity and high efficiency.
4. Electrostrictive. Similar to piezoelectric but has a nonlinear relationship between me-
chanical strain and electric field.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 17

1.5 Existing Technology 17

5. Electrostatic. These capacitive transducers use the change in force between two charged
parallel plates due to mechanical movement. These have found use with MEMS but
not in underwater acoustics.
6. Variable reluctance and hydroacoustic transducers have also been used for certain
experimental and sonar development work, but these devices have not challenged the
dominance of piezoceramic transducers for underwater sound applications [59].

[Link] Sonar Arrays

Sonar transducer arrays have low sidelobes and beams that scan over wide angular sectors.
Acoustic array configurations include linear, planar, cylindrical, spherical, conformal,
volumetric, reflector, and acoustic lenses [63]. These different array types lend themselves
to towing, conformal mounting on hulls (where the array surface conforms to the shape
of the underwater hull, so that no appendage is required), beam steering, and side-looking
sonar and synthetic-aperture applications. Figure 1-11 is an example of a 10 × 10 acoustic
planar array. A circular acoustic array is shown in Figure 1-12. Sometimes, the array must
be conformal to the surface on which it is mounted. Figure 1-13 shows a towfish with an
acoustic communications (ACOMMS) receive array that operates from 10 to 40 kHz and
a sidescan transducer that transmits and receives at 900 kHz. A towfish is a sidescan sonar
that is towed underwater by a boat.
A simple 2 ft × 4 ft planar array having more than 500 sensor elements for deep
submergence applications is shown in the diagram of Figure 1-14. In the quest for larger
power and lower frequency (with attendant lower attenuation losses), some arrays are
very large. In one case a 35 ft × 50 ft low-frequency planar array weighs 150 tons and
transmits close to 106 watts [63]. Large arrays with many closely spaced elements develop
“hot spots” from mutual coupling. Consequently, the concept of “velocity control” was
developed [64,65] to protect individual transducer elements against extreme impedance
variations.

FIGURE 1-11
Picture of a
100-element receive
acoustic array
manufactured in
four layers:
matching layer,
piezocomposite,
flex circuit, and
absorbing back
(Courtesy of
Materials Systems
Inc.).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 18

18 CHAPTER 1 Introduction

FIGURE 1-12 Left—Port and harbor surveillance piezocomposite array; 100 kHz
transmit/receive array. Right—Forward-looking piezocomposite sonar for AUV; piezocompos-
ite facilitates broad bandwidth, high element count arrays, and curved geometries (Courtesy
of Materials Systems Inc.).

Surface ships often use a bubble-shaped bow dome in which a cylindrical array is
placed like that shown in Figure 1-15. This array uses longitudinal vibrator-type elements
composed of a radiating front end of light weight and a heavy back mass, with a spring
having active ceramic rings or disks in the middle [63]. The axial symmetry of a cylindrical
array renders beam steering fairly simple with the azimuth direction in which the beam
is formed, because the symmetry allows identical electronic equipment for the phasing
and time delays required to form the beam. Planar arrays do not have this advantage,

FIGURE 1-13 ACOMS receive array

The acoustic
communications
(ACOMMS) receive
array operates from
10 to 40 kHz, and
the sidescan
transducer transmits Sidescan
transducer
and receives at 900
kHz (Courtesy of
Materials Systems
Inc.).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 19

1.5 Existing Technology 19

Existing Technology FIGURE 1-14

Rectangular planar
array for searchlight
type sonar system
Sensor elements
having more than
500 elements.

FIGURE 1-15
Cylindrical sonar
array used in
bow-mounted
dome.

since each new direction in space (whether in azimuth or in elevation) requires a new
combination of electronic equipment to achieve the desired pointing.
A spherical array is the ideal shape for the broadest array coverage in all directions.
Spherical arrays like that shown in the diagram of Figure 1-16 have been built with a
diameter of 15 ft and more than 1,000 transducer elements. This spherical arrangement
can be integrated into the bow of a submarine by means of an acoustically transparent
dome that provides minimum beam distortion. For instance, the bow dome of a Virginia
class submarine is a 25 ton hydrodynamically shaped composite structure that houses
a sonar transducer sphere. The bow dome is 21 feet tall and has a maximum diameter
of 26 feet. A two-inch thick, single-piece rubber boot is bonded to the dome to enhance
acoustic performance. Minimal sound energy absorption and reflection properties inherent
in the rubber material minimally reflect and absorb acoustic signals. Figure 1-17 shows
a spherical microphone array that was constructed by placing a rigid spherical array at
the center of a larger open spherical array [66]. Both arrays have 32 omnidirectional
microphones and a relatively constant directivity from about 900 Hz to 16 kHz.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 20

20 CHAPTER 1 Introduction

FIGURE 1-16
Spherical array
having 15 ft diameter
and more than 1,000 Transducer
transducer elements
elements.

FIGURE 1-17 A
dual, concentric
SMA. (A. Parthy,
C. Jin, and A. van
Schaik, “Acoustic
holography with a
concentric rigid and
open spherical
microphone array,”
IEEE International
Conference on
Acoustics, Speech
and Signal
Processing, 2009,
pp. 2173–2176.)

[Link] Beamformer
Beamforming ordinarily involves forming multiple beams from multielement arrays
through the use of appropriate delay and weighting matrices. Such beams may be di-
rectionally fixed or steerable. After that, sonar systems of the 1950s and 1960s consisted
largely of independent sonar sets for each transducer array. More recently, the sophisti-
cated use of multiple sensors and advances in computer technology have led to integrated
sonar systems that allow the interaction of data from different sensor arrays [67,68]. Such
integrated sonar systems have software delays and weighting matrices, thereby general-
izing the structure of digital time domain beamformers. Consequently several units of a
single (programmable) beamformer design may be used for all the arrays in an integrated
system. Furthermore, programmable beamformer matrices make it possible to adapt the
receive pattern to the changing structure of the masking noise background.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 21

1.6 Organization of the Book 21

[Link] Signal Processor

Signal processing involves filtering, spectral analysis, correlation, and the related opera-
tions of dynamic range compression and normalization (DRCN) that “match” the received
array signals to the display/decision functions contained in the information processing
block of Figure 1-9 [59,69]. The DRCN removes some of the spatial and temporal char-
acteristics of the acoustic channel that are impressed on the propagating signal. Whether
a sufficient degree of DRCN is achieved usually determines whether a system design will
succeed or fail since it specifies the integration of a theoretically designed optimum signal
processor to its system interfaces.

1.6 ORGANIZATION OF THE BOOK

1.6.1 Part 1
Chapter 2 introduces the adaptive array concept by first considering the nature of the
signal environment in which an adaptive array is expected to improve the overall reception
of a desired signal. The question of array architecture is considered, and the impact of
the architecture selection on the ultimate system performance that can be realized is
discussed. The potential of an array to enhance the output SNR performance by adjusting
the beamforming network is demonstrated.
An adaptive array manipulates the received signals to meet a defined performance
measure. This performance measure is an objective function to be maximized (or mini-
mized). Several widely used performance measures for both narrowband and broadband
applications are presented in Chapter 3 that are formulated in terms of the elements found
within the adaptive array functional model.
Any adaptive array system must have performance limits:
1. Imposed by the array physical structure.
2. Resulting from the nature of the signal environment
The consideration of performance limits leads to a discussion of the array performance
that results after the automatic adaptation process has been permitted to operate long
enough to reach a steady-state solution. A steady-state analytic solution to the adaptive
array control problem can be found for each performance measure, which enables the
designer to determine ultimate system performance limits. The mathematical foundation
required to develop the steady-state analytic solution for various performance measures is
laid, and the relationships among the solutions obtained to the optimal “Wiener solution”
are presented.

1.6.2 Part 2
The heart of the adaptive capability within an adaptive array system is the adaptive algo-
rithm that adjusts the array pattern in response to the signal information found at the sensor
element outputs. Part 2, including Chapters 4 through 8, introduces different classes of
adaptation algorithms. In some cases adaptation algorithms are selected according to the
kind of signal information available to the receiver:
1. The desired signal is known.
2. The desired signal is unknown, but its direction of arrival is known.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 22

22 CHAPTER 1 Introduction

3. The desired signal is unknown and its direction of arrival is known.

4. No signal information is available at the outset, but as array operation progresses such
information must be “learned” by the adaptive processor. These algorithms are called
blind adaptive algorithms.
The selection of an adaptive algorithm has important consequences for the system transient
performance.
The convergence properties of the various adaptation algorithms are analyzed and
performance comparisons are made. Furthermore, the shortcomings of the algorithms
under certain conditions are discussed. These results are summarized for convenience and
provide the designer with a means for assessing which candidate algorithms are most
appropriate for the signal conditions and system requirements.

1.6.3 Part 3
The adaptive array operating conditions considered so far were nonideal only in that in-
terference signals were present with which the array had to contend. In actual practice,
however, the effects of several other nonideal operating conditions often result in unaccept-
able degradation of array performance unless compensation of such effects is undertaken.
Such nonideal operating conditions include processing of broadband signals, multipath
effects, channel mismatching, and array propagation delay effects. Compensation for these
factors by means of tapped delay-line processing is considered, and the question of how
to design a tapped delay line to achieve a desired degree of compensation is addressed.
Finally, current trends in adaptive array research that provide an indication of the direction
that future developments are likely to take are discussed.

1.7 SUMMARY AND CONCLUSIONS

The motivation for and actual use of adaptive array systems are presented. The principal
elements of an adaptive array system are defined, and the fundamental problems facing
an adaptive array designer are given. Adaptive array design is a compromise among such
factors as [70]:
1. Hardware complexity and cost
2. Data rate
3. Maximum range of detection (for radar and sonar)
4. Resolution in angle (and range and Doppler for radar and sonar)
5. Precision in the measurement of range, bearing, and Doppler (for radar and sonar)
6. Ability of the adaptive array to meet both transient and steady-state system performance
requirements
A suboptimal acoustical array processor known as the DICANNE processor operated
in sea tests against ship-generated interferences and consistently formed cancellation nulls
10–15 dB deep [59]. Use of an optimal wideband processor based on the minimum sig-
nal distortion performance measure in a computer-simulated sonar experiment resulted in
effectively suppressing a strong coherent interfering signal by forming cancellation nulls
50 dB deep [59]. Such deep cancellation nulls were found, however, to be quite sensitive to
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 23

1.8 Problems 23

(1) small changes in interference signal bearing, (2) small errors in the adaptive weight val-
ues, and (3) statistical fluctuations of measured correlations due to finite integration time.
A lightweight four-element adaptive array using hybrid microwave integrated circuitry
and weighing only 1 pound, intended for communication applications, was built and tested
[71]. This unit employed a null-steering algorithm appropriate for a coherent sidelobe
canceller and succeeded in forming broadband nulls over a 60–100 MHz bandwidth having
a cancellation depth of 25–30 dB under weak desired signal and strong interference signal
conditions. To attain this degree of interference signal cancellation, it was essential that
the element channel circuitry be very well matched over a 20% bandwidth.
Another experimental four-element adaptive array system for eliminating interference
in a communication system was also tested [48]. Pattern nulls of 10–20 db for suppressing
interference signals over a 200–400 MHz band were easily achieved so long as the desired
signal and interference signal had sufficient spatial separation (greater than the resolution
capability of the antenna array), assuming the array has no way to distinguish between
signals on the basis of polarization. Exploiting polarization differences between desired
and interference signals by allowing full polarization flexibility in the array, an interference
signal located at the same angle as the desired signal can be suppressed without degrading
the reception of the desired signal. Yet another system employing digital control was
developed for UHF communications channels and found capable of suppressing jammers
by 20–32 dB [72].
In summary, interference suppression levels of 10–20 dB are consistently achieved
in practice. It is more difficult but nevertheless practicable to achieve suppression levels
of 20–35 dB and usually very difficult to form cancellation nulls greater than 35 dB in a
practical operating system.
The rapid development of digital technology is presently having the greatest impact
on signal reception systems. The full adaptation of digital techniques into the processing
and interpretation of received signals is making possible the realization of practical sig-
nal reception systems whose performance approaches that predicted by theoretical limits.
Digital processors and their associated memories have made possible the rapid digestion,
correlation, and classification of data from larger search volumes, and new concepts in the
spatial manipulation of signals have been developed. Adaptive array techniques started
out with limited numbers of elements in the arrays, and the gradual increase in the num-
bers of elements and in the sophistication of the signal processing will likely result in an
encounter with techniques employed in optical and acoustical holography [69,73]. Holog-
raphy techniques are approaching such an encounter from the other direction, since they
start out with a nearly continuous set of spatial samples (as in optical holography) and
move down to a finite number of samples (in the case of acoustic holography).

1.8 PROBLEMS
1. Radar Pulse Waveform Design Suppose it is desired to design a radar pulse waveform that
would permit two Ping-Pong balls to be distinguished when placed only 6.3 cm apart in range
up to a maximum range from the radar antenna of 10 m.
(a) What is the maximum PRF of the resulting pulse train?
(b) What bandwidth is required for the radar receiver channel?
(c) If it is desired to maintain an array element spacing of d = 2 cm where d = λ0 /2, what
pulse carrier frequency should the system be designed for?
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 24

24 CHAPTER 1 Introduction

2. Sonar Pulse Carrier Frequency Selection In the design of an actual sonar system many
factors must be considered—all the sonar parameters (e.g., source level, target strength) and the
environment parameters. The effect of environmental parameters depends largely on frequency.
Suppose in a highly oversimplified example that only the factors of transmission loss (due to
attenuation) and ambient noise are of concern. Let the attenuation coefficient α be given by

1
log10 (α) = [−21 + 5 log10 ( f )]
4
Furthermore, let the ambient noise spectrum level N0 be given by

1
10 log10 (N0 ) = [20 − 50 log10 ( f )]
3
If the cost to system performance is given by J = C1 α + C2 N0 where C1 and C2 denote the
relative costs of attenuation and noise to the system, what value of pulse carrier frequency f
should be selected to optimize the system performance?

1.9 REFERENCES
[1] L. C. Van Atta, “Electromagnetic Reflection,” U.S. Patent 2908002, October 6, 1959.
[2] IEEE Trans. Antennas Propag. (Special Issue on Active and Adaptive Antennas), Vol. AP-12,
March 1964.
[3] D. L. Margerum, “Self-Phased Arrays,” in Microwave Scanning Antennas, Vol. 3, Array
Systems, edited by R. C. Hansen, Academic Press, New York, 1966, Ch. 5.
[4] P. W. Howells, “Intermediate Frequency Sidelobe Canceller,” U.S. Patent 3202990, August
24, 1965.
[5] P. W. Howells, “Explorations in Fixed and Adaptive Resolution at GE and SURC,” IEEE
Trans. Antennas Propag., Special Issue on Adaptive Antennas, Vol. AP-24, No. 5, pp. 575–
584, September 1976.
[6] S. P. Applebaum, “Adaptive Arrays,” Syracuse University Research Corporation, Rep. SPL
TR66-1, August 1966.
[7] B. Widrow, “Adaptive Filters I: Fundamentals,” Stanford University Electronics Laboratories,
System Theory Laboratory, Center for Systems Research, Rep. SU-SEL-66-12, Tech. Rep.
6764-6, December 1966.
[8] B. Widrow, “Adaptive Filters,” in Aspects of Network and System Theory, edited by R. E.
Kalman and N. DeClaris, Holt, Rinehard and Winston, New York, 1971, Ch. 5.
[9] B. Widrow, P. E. Mantey, L. J. Griffiths, and B. B. Goode, “Adaptive Antenna Systems,” Proc.
IEEE, Vol. 55, No. 12, p. 2143–2159, December 1967.
[10] L. Spitzer, Jr., “Basic Problems of Underwater Acoustic Research,” National Research Council,
Committee on Undersea Warfare, Report of Panel on Underwater Acoustics, September 1,
1948, NRC CUW 0027.
[11] R. H. Bolt and T. F. Burke, “A Survey Report on Basic Problems of Acoustics Research,”
Panel on Undersea Warfare, National Research Council, 1950.
[12] H. Mermoz, “Filtrage adapté et utilisation optimale d’une antenne,” Proceedings NATO Ad-
vanced Study Institute, September 14–26, 1964. “Signal Processing with Emphasis on Un-
derwater Acoustics” (Centre d’Etude des Phénomenès Aléatoires de Grenoble, Grenoble,
France), 1964, pp. 163–294.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 25

1.9 References 25

[13] F. Bryn, “Optimum Signal Processing of Three-Dimensional Arrays Operating on Gaussian

Signals and Noise,” J. Acoust. Soc. Am., Vol. 34, March 1962, pp. 289–297.
[14] S. W. W. Shor, “Adaptive Technique to Discriminate Against Coherent Noise in a Narrowband
System,” J. Acoust. Soc. Am., Vol. 39, January 1966, pp. 74–78.
[15] P. E. Green, Jr., R. A. Frosch, and C. F. Romney, “Principles of an Experimental Large Aperture
Seismic Array (LASA),” Proc. IEEE, Vol. 53, December 1965, pp. 1821–1833.
[16] R. T. Lacoss, “Adaptive Combining of Wideband Array Data for Optimal Reception,” IEEE
Trans. Geosci. Electron., Vol. GE-6, May 1968, pp. 78–86.
[17] B. D. Steinberg, “On Teleseismic Beam Formation by Very Large Arrays,” Bull. Seismol. Soc.
Am., August 1971.
[18] IEEE Trans. Antennas Propag. (Special Issue on Adaptive Antennas), Vol. AP-24, September
1965.
[19] A. J. Berni, “Angle of Arrival Estimation Using an Adaptive Antenna Array,” IEEE Trans.
Aerosp. Electron. Syst., Vol. AES-11, No. 2, March 1975, pp. 278–284; see errata IEEE Trans.
Aerosp. Electron. Syst., Vol. AES-11, May and July 1975.
[20] R. C. Davis, L. E. Brennan, and I. S. Reed, “Angle Estimation with Adaptive Arrays in
External Noise Fields,” IEEE Trans. Aerosp. Electron. Syst., Vol. AES-12, No. 2, March 1976,
pp. 179–186.
[21] B. D. Steinberg, Principles of Aperture and Array System Design, Wiley, New York, 1976,
Ch. 11.
[22] B. D. Steinberg, “Design Approach for a High-Resolution Microwave Imaging Radio Camera,”
J. Franklin Inst., Vol. 296, No. 6, December 1973.
[23] E. N. Powers and B. D. Steinberg, “Valley Forge Research Center Adaptive Array,”
USNC/URSI Annual Meeting, Boulder, CO, October 1975.
[24] E. Biglieri, MIMO Wireless Communications, Cambridge, Cambridge University Press,
2007.
[25] M. H. Er, and S. K. Hui, “Beamforming in Presence of Element Failure,” Electronics Letters,
vol. 27, no. 3, pp. 273–275, 1991.
[26] J. T. Aberle, O. Sung-Hoon, D. T. Auckland, et al., “Reconfigurable Antennas for Wireless
Devices,” Antennas and Propagation Magazine, IEEE, vol. 45, no. 6, pp. 148–154, 2003.
[27] C.C. Chang, R.H. Lee, and T.Y. Shih, “Design of a beam switching/steering Butler matrix
for phased array system,” IEEE Transactions on Antennas and Propagation, vol. 58, no. 2,
pp. 367–374, Feb. 2010.
[28] L. J. Griffiths, “A Simple Adaptive Algorithm for Real-Time Processing in Antenna Arrays,”
Proc. IEEE, Vol. 57, No. 10, pp. 1696–1704, October 1969.
[29] J. Chang and F. Tuteur, “Optimum Adaptive Array Processor,” Proceedings of the Sympo-
sium on Computer Processing in Communications, April 8–10, 1969, Polytechnic Institute of
Brooklyn, pp. 695–710.
[30] O. L. Frost, “An Algorithm for Linearly Constrained Adaptive Array Processing,” Proc. IEEE,
Vol. 60, No. 8, pp. 926–935, August 1972.
[31] C. A. Baird, “Recursive Processing for Adaptive Arrays,” Proceedings of Adaptive Antenna
Systems Workshop, Vol. 1, Naval Research Laboratory, Washington, DC, September 27, 1974,
pp. 163–180.
[32] R. T. Adams, “An Adaptive Antenna System for Maximizing Signal-to-Noise Ratio,” Wescon
Conference Proceedings, Session 24, 1966, pp. 1–4.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 26

26 CHAPTER 1 Introduction

[33] L. J. Griffiths, “Signal Extraction Using Real-Time Adaptation of a Linear Multichannel

Filter,” SEL-68-017, Tech. Rep. No. 6788-1, System Theory Laboratory, Stanford University,
February 1968.
[34] L. E. Brennan, E. L. Pugh, and I. S. Reed, “Control Loop Noise in Adaptive Array Antennas,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-7, March 1971, pp. 254–262.
[35] C. A. Baird, Jr. and J. T. Rickard, “Recursive Estimation in Array Processing,” Proceedings of
the Fifth Asilomar Conference on Circuits and Systems, 1971, Pacific Grove, CA, pp. 509–513.
[36] J. J. Faran and R. Hills, Jr., “Wide-Band Directivity of Receiving Arrays,” Harvard University
Acoustics Research Laboratory, Tech. Memo 31, May 1, 1953.
[37] H. H. Az-Khatib and R. T. Compton, Jr., “A Gain Optimizing Algorithm for Adaptive
Arrays,” IEEE Trans. Antennas Propag., Vol. AP-26, No. 2, March 1978, pp. 228–235.
[38] N. Wiener, Extrapolation, Interpretation, and Smoothing of Stationary Time Series, New
York, MIT Press, 1949.
[39] H. L. Van Trees, “Optimum Processing for Passive Sonar Arrays,” IEEE 1966 Ocean
Electronics Symposium, August 29–31, Honolulu, HI, pp. 41–65.
[40] K. Takao, M. Fujita, and T. Nishi, “An Adaptive Antenna Array Under Directional Constraint,”
IEEE Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 662–669.
[41] J. Capon, R. J. Greenfield, and R. J. Kolker, “Multi-Dimensional Maximum-Likelihood Pro-
cessing of a Large Aperture Seismic Array,” Proc. IEEE, Vol. 55, February 1967, pp. 192–217.
[42] J. F. Clearbout, “A Summary, by Illustration, of Least-Square Filter With Constraints,” IEEE
Trans. Inf. Theory, Vol. IT-14, March 1968, pp. 269–272.
[43] A. Booker and C. Y. Ong, “Multiple-Constraint Adaptive Filtering,” Geophysics, Vol. 36,
June 1971, pp. 498–509.
[44] R. L. Deavenport, “The Influence of the Ocean Environment . . . ,” EASCON 1975 Record,
IEEE Electronics and Aerospace Systems Convention, September 29–October 1, Washington,
DC, pp. 66. A - 66E.
[45] E. Brookner, “Now: Phased-array Radars: Past, Astounding Breakthroughs and Future
Trends, January 2008,” Microwave Journal, Vol. 51, no. 1, pp. 31–48, 2008.
[46] W. G. Bruner, W. N. Case, and R. C. Leedom, “Digital Control in Modern Radar,” EASCON
1971 Record, IEEE Electronics and Aerospace Systems Convention, October 6-8, 1971,
Washington, DC, pp. 200–207.
[47] R. T. Compton, Jr., “An Experimental Four-Element Adaptive Array,” IEEE Trans. Antennas
& Prop., Vol AP-25, No. 5, September 1965, pp. 697–706.
[48] W. L. Curtis, “Spiral Antennas,” IRE Trans. Antennas & Prop, Vol AP-8, May, 1960, pp.
298–306.
[49] R. H. DuHamel and D. G. Berry, “Logarithmically Periodic Antenna Arrays,” IRE WESCON
Convention Record, 1958, Vol. 2, Pt. 1, pp. 161–174.
[50] R. Garg, P. Bhartia, I. Bahl, and A. Ittipiboon, Microstrip Antenna Design Handbook,
Norwood, MA, Artech House, 2001.
[51] P. J. Gibson, “the Vivaldi Aerial,” 9th European Microwave Conf. Proc., 19779, pp. 103–105.
[52] National Electronics Museum, Wurzburg Riese Radar Antenna, 2009, Display Information.
[53] National Electronics Museum, SCR-270 Antenna, 2009, Display Information.
[54] National Electronics Museum, Nike AJAX MPA-4 Antenna, 2009, Display Information.
[55] National Electronics Museum, AN/APG-68 Antenna, 2009, Display Information.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 27

1.9 References 27

[56] M. I. Skolnik, Introduction to Radar Systems, McGraw-Hill, New York, 1962, Ch. 2.
[57] C. H. Sherman, “Underwater Sound - A Review: I. Underwater Sound Transducers,” IEEE
Trans. Sonics and Ultrason., Vol. SU-22, No. 5, September 1975, pp. 281–290.
[58] A. A. Winder, “Underwater Sound - A Review: II. SKonar System Technology,” IEEE Trans.
Sonics and Ultrason., Vol. SU-22, No. 5, September 1975, pp. 291–332.
[59] E. A. Massa, “Deep Water trnasducers for Ocean Engineering Applications,” Proceedings
IEEE, 1966 Ocean Electronics Symposium, August 29–31, Honolulu, HI, pp. 31–39.
[60] C. H. Sherman and J. L. Butler, Transducers and Arrays for Underwater Sound, New York,
Springer, 2007.
[61] T. F. Hueter, “Twenty Years in Underwater Acoustics: Generation and Reception,” J. Acoust.
Soc. Am., Vol. 51, No 3 (Part 2), March 1972, pp. 1025–1040.
[62] J. S. Hickman, “Trends in Modern Sonar Transducer Design,” Proceedings of the Twenty-
Second National Electronics Conference, October 1966, Chicago, Ill.
[63] E. L. Carson, G. E. Martin, G. W. Benthien, and J. S. Hickman, “Control of Element
Velocity Distributions in Sonar Projection Arrays,” Proceedings of the Seventh Navy Science
Symposium, May 1963, Pensacola, FL.
[64] A. Parthy, C. Jin, and A. van Schaik, “Acoustic Holography with a Concentric Rigid and
Open Spherical Microphone Array,” IEEE International Conference on Acoustics, Speech
and Signal Processing, 2009, pp. 2173–2176.
[65] J. F. Bartram, R. R. Ramseyer, and J. M Heines, “Fifth Generation Digital Sonar Signal
Processing,” IEEE EASCON 1976 Record, IEEE Electronics and Aerospace Systems
Convention, September 26–29, Washington, D.C., pp. 91A–91G.
[66] T. F. Hueter, J. C. Preble, and G. D. Marshall, “Distributed Architecture for Sonar Systems
and Their Impact on Operability,” IEEE EASCON 1976 Record, IEEE Electronics and
Aerospace Convention, September 16–29, Washington, D.C., pp. 96A–96F.
[67] V. C. Anderson, “The First Twenty Years of Acoustic Signal Processing,” J. Acoust. Soc.
Am., Vol. 51, No. 3 (Part 2), March 1972, pp. 1062–1063.
[68] M. Federici, “On the Improvement of Detection and Precision Capabilities of Sonar Systems,”
Proceedings of the Symposium on Sonar Systems, July 9–12, 1962, University of Birmingham,
pp. 535–540, Reprinted in J. Br. I.R.E., Vol. 25, No.6, June 1963.
[69] G. G. Rassweiler, M. R. Williams, L. M. Payne, and G. P. Martin, “A Miniaturized Lightweight
Wideband Null Steerer,” IEEE Trans. Ant. & Prop., Vol. AP-24, September 1976, pp. 749–754.
[70] W. R. Jones, W. K. Masenten, and T. W. Miller, “UHF Adaptive Array Processor Development
for Naval Communications,” Digest AP-S Int. Symp., IEEE Ant. & Prop Soc., June 2–22,
1977, Stanford, CA., IEEE Cat. No. 77CH1217-9 AP, pp. 394–397.
[71] G. Wade, “Acoustic Imaging with Holography and Lenses,” IEEE Trans. Sonics & Ultrason.,
Vol. SU-22, No. 6, November 1975, pp 385–394.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:5 28
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 29

CHAPTER

Adaptive Array Concept

2
' $
Chapter Outline
2.1 Signal Environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Array Element Spacing Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Array Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.4 Nulling Limitations Due to Miscellaneous Array Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.5 Narrowband and Broadband Signal Processing Considerations . . . . . . . . . . . . . . . . . . 62
2.6 Adaptive Array Performance Measure—Coverage Improvement Factor (CIF). . . . . . 66
2.7 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.8 Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
& %

To understand why an array of sensor elements has the potential to improve the reception
of a desired signal in an environment having several sources of interference, it is neces-
sary to understand the nature of the signals as well as the properties of an array of sensor
elements. Furthermore, the types of elements and their arrangement impact the adaptive
array performance. To gain this understanding the desired signal characteristics, inter-
ference characteristics, and signal propagation effects are first discussed. The properties
of sensor arrays are then introduced, and the possibility of adjusting the array response
to enhance the desired signal reception is demonstrated. Trade-offs for linear and planar
arrays are presented to aid the designer in finding an economical array configuration.
In arriving at an adaptive array design, it is necessary to consider the system constraints
imposed by the nature of the array, the associated system elements with which the designer
has to work, and the system requirements the design is expected to satisfy. Adaptive array
requirements may be classified as either (1) steady-state or (2) transient depending on
whether it is assumed the array weights have reached their steady-state values (assuming
a stationary signal environment) or are being adjusted in response to a change in the signal
environment. If the system requirements are to be realistic, they must not exceed the
predicted theoretical performance limits for the adaptive array system being considered.
Formulation of the constraints imposed by the sensor array is addressed in this chapter.
Steady-state performance limits are considered in Chapter 3. The formulation of transient
performance limits, which is considerably more involved, is addressed in Part 2. For the
performance limits of adaptive array systems to be analyzed, it is necessary to develop a
generic analytic model for the system. The development of such an analytic model will
be concerned with the signal characteristics and the subsequent processing necessary to
obtain the desired system response.
29
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 30

30 CHAPTER 2 Adaptive Array Concept

2.1 SIGNAL ENVIRONMENT

Our goal is to extract useful information from a desired signal received by an array of sen-
sors. The adaptive array designer must exploit significant differences between the desired
signal and any interference signals to distinguish the desired signal from all other signals
received by the array. The signal parameters that may be exploited include direction of
arrival, amplitude, phase, spectral characteristics (e.g., frequency and power), modulation
characteristics, and polarization. It is therefore worthwhile to consider the desired signal
characteristics and the nature of spurious interfering signals in different contexts.

2.1.1 Signals in Active and Passive Sensors

Active sensing devices, such as radar and sonar systems, generate a known pulse (or pulse
train) that propagates through a transmission medium and reflects from a target back to the
original sender. During most of the listening time interval the desired signal is absent in
contrast with communication systems where the desired signal is usually present. Often,
the signal structure and the direction of arrival are known, so the desired signal is easily
recognized when it is present. In addition to the desired signal echo, noise due to clutter
and multipath may also be present [1]. For radar systems, diffuse scattering of multipath
gives rise to spurious signals, and a jammer may deliberately generate an interference
signal. For active sonar systems, the two main types of interference signals come from
ambient noise and reverberation return [2].
Reverberation is analogous to clutter in radar systems and is defined as signal-
generated noise that results when a sonar signal transmitted into the ocean encounters
many kinds of diffuse scattering objects that reradiate part of the signal back to the re-
ceiver. Reverberation returns are classified as surface, bottom, or volume reverberation
depending on whether the unwanted reflections originate from the ocean surface, the
ocean surface, the ocean bottom, or some point in between. Furthermore, multipath sig-
nals arise from signals that reflect from nondiffuse reflectors located at different reflecting
angles and impinge on the receiving array. Other propagation effects that cause sonar sig-
nal distortion are geometric spreading, attenuation, multiple propagation paths, Doppler
frequency shift, finite amplitude, medium coherency, and time dispersion [3].
In the case of passive sensing systems, the target or event of interest generates the
desired signal and in many cases is present for most of the listening interval. A passive
system has difficulty distinguishing the desired target signal from the background noise
[4]. In contrast to active sensing devices, however, the direction of arrival and structure of
the desired signal may not be known beforehand. The bandwidth of the desired signal is the
most common means for distinguishing it from an interference signal. In some cases, the
power level of the desired signal may be known and used as a distinguishing characteristic.
Spread spectrum communication systems commonly employ a known pseudo noise (PN)
code to modulate the transmitted waveform, and this code then provides a convenient
means for distinguishing the desired signal.

2.1.2 Signal Models

Signals in passive sonar systems arise from a variety of sources such as engine noise or
propeller noise [5] and are treated as random processes. Likewise, an unknown commu-
nications signal may often be regarded as random, although (as we will see) a simple
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 31

2.1 Signal Environment 31

signal model possessing some of the signal characteristics of the unknown signal can be
adopted in certain circumstances. Communications signals pass through environments that
randomly add scattering and noise to the desired signal. Thermal sensor noise, ambient
noise, and interference signal sources are also random in nature. These noises typically
arise from the combined effect of many small independent sources, and application of the
central limit theorem of statistics [6] permits the designer to model the resulting noise
signal as a Gaussian (and usually stationary) random process. Quite frequently, the phys-
ical phenomena responsible for the randomness in the signals of concern are such that it
is plausible to assume a Gaussian random process. The statistical properties of Gaussian
signals are particularly convenient because the first- and second-moment characteristics
of the process provide a complete characterization of the random signal.
The statistical properties of the signal are not always known, so a selected deterministic
signal is used instead. This deterministic signal does not have to be a perfect replica of the
desired signal. It needs only to be reasonably well correlated with the desired signal and
uncorrelated with the interference signals.
Sometimes the desired signal is known, as in the case of a coherent radar with a target
of known range and character. In other cases, the desired signal is known except for the
presence of uncertain parameters such as phase and signal energy. The case of a signal
known except for phase occurs with an ordinary pulse radar having no integration and
with a target of known range and character. Likewise, the case of a signal known except
for phase and signal energy occurs with a pulse radar operating without integration and
with a target of unknown range and known character. The frequency and bandwidth of
communication signals are typically known, and such signals may be given a signature by
introducing a pilot signal.
For a receive array composed of N sensors, the received waveforms correspond to N
outputs, x1 (t), x2 (t), . . . , x N (t), which are placed in the received signal vector x(t) where
⎡ ⎤
x1 (t)
⎢ x (t) ⎥
 ⎢ 2 ⎥
x(t) = ⎢ . ⎥ for 0 ≤ t ≤ T (2.1)
⎣ .. ⎦
x N (t)
over the observation time interval. The received signal vector is the sum of the desired
signal vector, s(t), and the noise component, n(t).

x(t) = s(t) + n(t) for 0≤t ≤T (2.2)

where the desired signal vector is represented by

⎡ ⎤
s1 (t)
⎢ s2 (t) ⎥
⎢ ⎥
s(t) = ⎢ . ⎥ for 0 ≤ t ≤ T (2.3)
⎣ .. ⎦
s N (t)
The various signal components either may be known exactly, known only to a rough
approximation, or known only in a statistical sense.
Interference noise fields at best are stationary and unknown, but they usually ex-
hibit variations with time. Adaptive array processors that automatically respond to the
interference environment must cope with such time variations. Most adaptive processing
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 32

32 CHAPTER 2 Adaptive Array Concept

techniques assume slowly varying Gaussian ambient noise fields. Adaptive processors
designed for Gaussian noise are distinguished by the pleasant fact that they depend only
on second-order noise moments. Consequently, when non-Gaussian noise fields must be
dealt with, the most convenient approach is to design a Gaussian-equivalent suboptimum
adaptive system based on the second-order moments of the non-Gaussian noise field.
In general, an adaptive system works best when the variations that occur in the noise
environment are slow.

2.1.3 Ideal Propagation Model

The signal vector, s(t), relates to a scalar signal, s(t), generated at a point source in space
by [7]

s(t) = m(t − τ )s(τ ) dτ (2.4)

where the ith component of m(t) is m i (t) and represents the propagation effects from the
source to the ith sensor as well as the response of the ith sensor. For the ideal case of
nondispersive propagation and distortion-free sensors, then m i (t) is a simple time delay
δ(t − τi ), and the desired signal component at each sensor element is identical except for
a time delay so that (2.4) can be written as
⎡ ⎤
s(t − τ1 )
⎢ s(t − τ2 ) ⎥
⎢ ⎥
s(t) = ⎢ .. ⎥ (2.5)
⎣ . ⎦
s(t − τ N )

When the array is far from the source, then the signal is represented by a plane wave from
the direction α as shown in Figure 2-1 (where α is taken to be a unit vector). In this case,

FIGURE 2-1 z
Three-dimensional
array with plane
Wa

wave signal
ve

propagation.
fro
nt

Sensor s(t)
locations
r1

r2
a
x

r4
r3

y
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 33

2.2 Array Element Spacing Considerations 33

the various time delays are simply given by [8]

α · ri
τi = (2.6)

where  is the propagation velocity, each sensor coordinate is given by the vector ri , and
α ri denotes the dot product

α · ri = α T ri (2.7)

and T denotes transpose.

Measuring the relative time delays experienced at each sensor element therefore pro-
vides a means of determining the unknown direction of arrival of the desired signal s(t).
In the following sections on array properties, the previously described plane wave prop-
agation effects play a fundamental role in determining the nature of the array response.
Furthermore, it should be noted that the relative time-delay properties suggested by (2.5)
and (2.6) work for random and nonrandom signals.

2.2 ARRAY ELEMENT SPACING

CONSIDERATIONS
It was noted earlier that an array of sensors increases the gain and decreases the bandwidth
of a single sensor as well as changes the shape of the array pattern. To demonstrate, a
single pair of identical sensor elements is first considered. The discussion then proceeds
to consider the one-dimensional linear array of equally spaced elements and finally extends
to planar and conformal arrays.
After considering the effects of sensor element arrangement within the array, the
impact on array performance limits that are determined by a selected array configuration
are examined. The position of the elements within the array determines array resolution
and interferometer (grating lobe) [9] effects. In general, resolution increases as the array
dimension (or separation between elements) increases. High array resolution improves
the maximum output signal-to-noise ratio (SNR) when the angular separation between the
desired and the undesired signals is small. High-resolution capability also implies sharp
array pattern nulls, however, thereby reducing the array ability to place a broad null on
clustered interference sources. An N -element linear array can have up to N − 1 degrees
of freedom so that up to N − 1 array beam pattern nulls can be independently adjusted
for array operation.
Typically, there is a limited amount of space over which to distribute antenna elements.
Unfortunately, closely spaced elements yield fields that interact, which results in mutual
coupling between elements, and it is well known [10] that mutual coupling has deleterious
effects on array signal-to-interference plus noise ratio (SINR) performance and transient
response. A discussion of how to account for mutual coupling effects is taken up in
Chapter 10.

2.2.1 Pair of Identical Sensors

Consider the pair (N = 2) of identical isotropic point sources shown in Figure 2-2 spaced
apart by a distance d. An isotropic point source radiates equally in all directions. Let a
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 34

34 CHAPTER 2 Adaptive Array Concept

FIGURE 2-2 Pair Array normal

of isotropic point
sources.

q x(t)

d sin q

Sensor 2 Sensor 1
Array axis

signal x(t) impinge on the two sensor elements in a plane containing the two elements and
the signal source from a direction θ with respect to the array normal. Figure 2-2 shows
that element 2 receives the signal τ after element 1.
d sin θ
τ= (2.8)

Let the array output signal y(t) be given by the sum of the two sensor element signals so
that

y(t) = x(t) + x(t − τ ) (2.9)

If x(t) is a narrowband signal having center frequency f 0 , then the time delay τ
corresponds to a phase shift of 2π(d/λ0 ) sin θ radians, where λ0 is the wavelength corre-
sponding to the center frequency,

λ0 = (2.10)
f0
The overall array response is the sum of the signal contributions from the two array
elements. That is,
2
y(t) = x(t)e j (i−1)ψ (2.11)
i=1

where

ψ = 2π(d/λ0 ) sin θ (2.12)

The directional pattern or array factor of the array (sensitivity to signals vs. angle at
a specified frequency) may be found by considering only the term
2
AF(θ ) = e j (i−1)ψ (2.13)
i=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 35

2.2 Array Element Spacing Considerations 35

The normalized array factor in decibels for the two elements is then given by

|AF(θ )2 |
AF(θ)(decibels) = 10 log10 (2.14)
22

so that the peak value is unity or 0 dB. A plot of the normalized AF(θ ) for this two-
element example is given in Figure 2-3 for d/λ0 = 0.5, 1.0, and 1.5. From Figure 2-3a
it is seen that for d/λ0 = 0.5 there is one principal lobe (or main beam) having a 3 dB
beamwidth of 60◦ and nulls at θ = ±90◦ off broadside. The nulls at θ = ±90◦ occur
because at that direction of arrival the signal wavefront must travel exactly λ0 /2 between

0 0

−5 −5

−10 −10

AF (dB)
AF (dB)

−15 −15

−20 −20

−25 −25

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(a) (b)

−5

−10
AF (dB)

−15

−20

−25

−30
−90 −45 0 45 90
q (degrees)
(c)

FIGURE 2-3 Array beam patterns for two-element example. (a) d/λ0 = 0.5. (b) d/λ0 = 1.
(c) d/λ0 = 1.5.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 36

36 CHAPTER 2 Adaptive Array Concept

the two sensors, which corresponds to a phase shift of 180◦ between the signals appearing
at the two sensors and therefore yields exact cancellation of the resulting phasor sum.
If the element spacing is less than 0.5λ0 , then exact cancellation at θ = ±90◦ does not
result, and in the limit as the element spacing approaches zero (ignoring mutual coupling
effects) the directional pattern becomes an isotropic point source pattern. There is very
little difference in the directional pattern between a single element and two closely spaced
elements (less than λ/4 apart); consequently, arrays employing many elements very closely
spaced are considered “inefficient” if it is desired to use as few array elements as possible
for a specified sidelobe level and beamwidth. If the element spacing increases to greater
than 0.5λ0 , the two pattern nulls migrate in from θ = ±90◦ , occurring at θ = ±30◦ when
d = λ0 , as illustrated in Figure 2-3b. The nulls at θ = ±30◦ in Figure 2-3b occur because
at that angle of arrival the phase path difference between the two sensors is once again
180◦ , and exact cancellation results from the phasor sum. Two sidelobes at θ = ±90◦
having an amplitude equal to the principal lobe at θ = 0◦ . They appear because the phase
path difference between the two sensors is then 360◦ , two phasors exactly align, and
the array response is the same as for broadside angle of arrival. As the element spacing
increases to 1.5λ0 , the main lobe beamwidth decreases still further, thereby improving
resolution, the two pattern nulls migrate further in, and two new nulls appear at ±90◦ ,
as illustrated in Figure 2-3c. Further increasing the interelement spacing results in the
appearance of even more pattern nulls and sidelobes and a further decrease in the main lobe
beamwidth.
When N > 2, the array factor for N isotropic point sources becomes

N
2π
[xn sin θ cos φ+yn sin θ sin φ+z n cos θ ]
AF(θ, φ) = wne j λ (2.15)
n=1

where

(xn , yn , z n ) = position of sensor n in the array

N = number of sensors in the array
(θ, φ) = direction from array phase center

Three common examples of multiple element arrays used for adaptive nulling are [11]
N
2π
xn sin θ cos φ
Linear array along the x-axis: AF(θ, φ) = wne j λ (2.16)
n=1
N
2π
[xn sin θ cos φ+yn sin θ sin φ]
Planar array in the x-y plane: AF(θ, φ) = wne j λ (2.17)
n=1
N
2π
[xn sin θ cos φ+yn sin θ sin φ+z n cos θ ]
3-dimensional array: AF(θ, φ) = wne j λ (2.18)
n=1

The sensors take samples of the signals incident on the array. As long as the sensors
take two samples in one period or wavelength, then the received signal is adequately
reproduced.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 37

2.2 Array Element Spacing Considerations 37

2.2.2 Linear Arrays

For a linear array of N equispaced sensor elements, the overall array response may again be
found by considering the sum of signal contributions from each array element as in (2.11):
N
y(t) = w i x(t)e j (i−1)ψ (2.19)
i=1

The directional pattern in a plane containing the array may therefore be found by consid-
ering the array factor
N
AF(θ ) = w i e j (i−1)ψ (2.20)
i=1

When w i = 1.0, the array is called a “uniform array.” The uniform array has an array
factor given by
Nψ
sin 2
AF(θ ) = (2.21)
ψ
sin 2

Nulls occur when the numerator is zero or

 
mλ
θm = sin−1 ± , m = 1, 2, . . . (2.22)
dN
Sidelobe peaks are approximately halfway between two adjacent nulls
 
(2m + 1) λ
θm = sin−1 ± , m = 1, 2, . . . (2.23)
2d N
The first sidelobe is approximately 13.5 dB below the peak of the main beam.
For nonisotropic sensor elements, it is necessary to introduce an additional factor
Fi ( f 0 , θ) in (2.20) to include the pattern introduced by each sensor element.
N
AF(θ) = Fi ( f 0 , θ )w i e j (i−1)ψ (2.24)
i=1

If mutual coupling is ignored or if the element patterns are averaged, the individual element
patterns are all the same, so the array pattern is the product of an element pattern times
the array factor.
N
AF(θ) = F( f 0 , θ ) w i e j (i−1)ψ (2.25)
i=1

Alternatively, the response of an arbitrary configuration of identical elements having the

same orientation can be derived by using the principle of pattern multiplication [9]. The
directional pattern of an array of identical spatial elements may be found by
1. Replacing each of the elements by a point source
2. Determining the directional array pattern of the resulting array of point sources
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 38

38 CHAPTER 2 Adaptive Array Concept

0 0

−5 −5

−10 −10

AF (dB)
AF (dB)

−15 −15

−20 −20

−25 −25

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(a) (b)

FIGURE 2-4 inear array beam patterns for d/λ0 = 0.5. (a) Three-element array.
(b) Four-element array.

3. Multiplying the array pattern resulting from step 2 by the beam pattern of the individual
elements of the original array
Maintaining the interelement spacing at d/λ0 = 0.5 and increasing the number of
point sources, the normalized array directional (or beam) pattern may be found from
(2.21), and the results are shown in Figure 2-4 for three and four elements. It is seen that
as the number of elements increases the main lobe beamwidth decreases, and the number
of sidelobes and pattern nulls increases.
To illustrate how element spacing affects the directional pattern for a seven-element
linear array, Figures 2-5a through 2.5d show the directional pattern in the azimuth plane
for values of d/λ0 ranging from 0.1 to 1.0. So long as d/λ0 is less than 17 , the beam pattern
has no exact nulls as the −8.5 dB null occurring at θ = ±90◦ for d/λ0 = 0.1 illustrates.
As the interelement spacing increases beyond d/λ0 = 17 , pattern nulls and sidelobes (and
grating lobes) begin to appear, with more lobes and nulls appearing as d/λ0 increases and
producing an interferometer pattern. When d/λ0 = 1 the endfire sidelobes at θ = ±90◦
have a gain equal to the main lobe since the seven signal phasors now align exactly and
add coherently.
Suppose for the linear array of Figure 2-6 that a phase shift (or an equivalent time
delay) of δ is inserted in the second element of the array, a phase shift of 2δ in the third
element, and a phase shift of (n − 1)δ in each succeeding nth element. The insertion of
this sequence of phase shifts has the effect of shifting the principal lobe (or main beam)
by
   
−1 1 λ0
θs = −sin δ (2.26)
2π d

so the overall directional pattern has in effect been “steered” by insertion of the phase-
shift sequence. Figure 2-7 is the array factor in Figure 2-5c steered to θs = −30◦ using
δ = −90◦ .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 39

2.2 Array Element Spacing Considerations 39

0
0
−5
−5

−10
−10
AF (dB)

AF (dB)
−15
−15

−20 −20

−25 −25

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(a) (b)

0 0

−5 −5

−10 −10
AF (dB)
AF (dB)

−15 −15

−20 −20

−25 −25

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(c) (d)

FIGURE 2-5 Seven-element linear array factors. (a) d/λ0 = 0.1. (b) d/λ0 = 0.2. (c) d/λ0 =
0.5. (d) d/λ0 = 1.0.

The gain of an array factor determines how much the signal entering the main beam
is magnified by the array. For a radar/sonar system the received power is given by [12]
Pt G 2 λ2 σ
Pr = (2.27)
(4π )3 R 4
where
Pt = power transmitted
G = array gain
σ = target cross section
R = distance from array to target
and for a communications system is given by Friis transmission formula [13]
Pt G t G r λ2
Pr = (2.28)
(4π R)2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 40

40 CHAPTER 2 Adaptive Array Concept

FIGURE 2-6
Seven-element
q
linear array steered
with phase-shift
elements.

d = l 0 /2 d
Sensor
elements

6d 5d 4d 3d 2d d

Array
output
y(t )

FIGURE 2-7 0
Seven-element
linear array factor
steered to θs = 30◦ −5
with δ = −90◦ .
−10
AF (dB)

−15

−20

−25

−30
−90 −45 0 45 90
q (degrees)

The array gain is the ratio of the power radiated in one direction to the power delivered
to the array. Directivity is similar to gain but does not include losses in the array. As a
result, directivity is greater than or equal to gain. Realized gain includes the mismatch
between the array and the transmission line feeding the array. If gain is not written as a
function of angle, then G is the maximum gain of the array. Gain is usually expressed
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 41

2.2 Array Element Spacing Considerations 41

in dB as
G dB = 10 log10 G = 10 log G (2.29)
The directivity of an array is found by solving
4π |AFmax |2
D= (2.30)
 π
2π
|AF(θ, φ)|2 sin θ dθ dφ
0 0

If the elements are spaced 0.5λ apart in a linear array, then the directivity formula simplifies
to
 2
 N 
 w n 

D = n=1 (2.31)
N
|w n | 2
n=1

The z-transform converts the linear array factor into a polynomial using the substitu-
tion
z = e jψ (2.32)
Substituting z into (2.16) yields a polynomial in z
N
AF = w n z (n−1) = w 1 + w 2 z + · · · + w N z N −1 (2.33)
n=1

that can be factored into the form

AF = w N (z − z 1 )(z − z 2 ) · · · (z − z N −1 ) (2.34)
Roots of the array factor polynomial (z = z n ) correspond to nulls in the array factor. Roots
have a magnitude of 1 and phase of ψn . The roots of (2.34) have a magnitude of 1 and a
phase of ψn and hence can be graphed on a circle in the z-plane as shown in Figure 2-8

p /2 FIGURE 2-8 Unit

circle representation
of an eight-element
uniform array.

p 0

−p /2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 42

42 CHAPTER 2 Adaptive Array Concept

for the case of an eight-element uniform array with d = 0.5λ:

z7 + z6 + z5 + z4 + z3 + z2 + z + 1
       
= z + e jπ/4 z + e− jπ/4 z + e jπ/2 z + e− jπ/2 z + e j3π/4 z + e j3π/4 z + e jπ
(2.35)

2.2.3 Planar Arrays

Consider the rectangular-shaped planar array of discrete sensor elements arranged in the
x–y plane as shown in Figure 2-9, where the coordinate origin is chosen at the central
element. With N x elements in each x-axis parallel column and uniform spacing dx , and N y
elements in each y axis parallel row with uniform spacing d y , the entire array has N x × N y
elements.
The sum of signal contributions from each array element in a single column is the
same as that for a linear array and is therefore given by
Nx
y(t) = w i x(t)e j (i−1)ψx (2.36)
i=1

whereas for a single-row array the output is

Ny
y(t) = w k x(t)e j (k−1)ψ y (2.37)
k=1

FIGURE 2-9 z
Rectangular-shaped
planar array.
P(r, q, f)
r

Projection of
P(r, q, f) onto the
dx x–y plane

dy

Sensor x
elements
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 43

2.2 Array Element Spacing Considerations 43

where
   
dx dy
ψx = 2π sin θ cos φ and ψ y = 2π sin θ sin φ (2.38)
λ0 λ0
The output signal now depends on both the projected azimuth angle, φ, and the elevation
angle, θ. The total sum of signal contributions from all array elements is given by
Nx Ny
y(t) = w i,k x(t)e j (i−1)ψx e j (k−1)ψ y (2.39)
i=1 k=1

If the weights are separable, w i,k = w i × w k , then the planar array factor is

AF(θ, φ) = AFx (θ, φ)AFy (θ, φ) (2.40)

where
⎫
Nx ⎪
⎪
⎪
⎪
AFx (θ, φ) = w i e j (i−1)ψx ⎬
i=1 (2.41)
and ⎪
⎪
Ny ⎪
⎪
⎭
AFy (θ, φ) = w k e j (k−1)ψ y
k=1

Elements in a planar array are usually arranged in either a rectangular lattice or an

equilateral triangle lattice. The y-spacing in an equilateral triangle lattice is found from
the x-spacing by
√
d y = dx 3/2 (2.42)

If w n = 1.0 for all n, then the array factor for rectangular spacing is
N x ψx N y ψy
sin 2
sin 2
AF = ψy
(2.43)
ψx
N x sin 2
N y sin 2

where
2π
ψx = dx sin θ cos φ
λ
2π
ψy = d y sin θ sin φ
λ
Usually, the beamwidth of a planar array is defined for two orthogonal planes. For
example, the beamwidth is usually defined in θ for φ = 0◦ and φ = 90◦ . As with a linear
array, nulls in the array factor of a planar array are found by setting the array factor equal
to zero. Unlike linear arrays, the nulls are not single points.
The directivity of a planar array can be found by numerically integrating (2.24). A
reasonable estimate for the 3 dB beamwidth in orthogonal planes is given by
32400
D= ◦ ◦ (2.44)
θ3dBφ=0 ◦ θ3dBφ=90◦
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 44

44 CHAPTER 2 Adaptive Array Concept

where
◦ ◦
θ3dBφ=0 ◦ = 3 dB beamwidth in degrees at φ= 0

◦ ◦
θ3dBφ=90 ◦ = 3 dB beamwidth in degrees at φ= 90

The array factor for a planar array with rectangular element spacing can be written as

Ny Nx  dy

(n−1) dλx (u−u s )+(m−1) (v−vs )
AF = w mn e j2π λ (w) (2.45)
m=1 n=1

where

u = sin θ cos φ
v = sin θ sin φ
(2.46)
us = sin θs cos φs
vs = sin θs sin φs

when the main beam is steered to (θs , φs ).

2.2.4 Conformal Arrays

Array elements do not have to lie along a line or in a plane. Many times, it is more
convenient if the elements lie on or conform to a surface. The surface may be the side of a
building or the side of an airplane. In any event, the array factor now has element locations
given by (xn , yn , z n ), which are points lying on a defined surface.

N
2π
(xn sin θ cos φ+yn sin θ sin φ+z n cos θ )
AF = wne j λ (2.47)
n=1

The array curvature causes phase errors that distort the array factor unless phase weights at
the elements compensate. As an example, consider a 12-element linear array bent around
a cylinder of radius r = 3.6λ as shown in Figure 2-10. If no phase compensation is
applied, then the array factor looks like the dashed line in Figure 2-11. Adding a phase
delay of yn 2π/λ results in the array factor represented by the solid line in Figure 2-11.
The compensation restores the main beam and lowers the sidelobe level.

FIGURE 2-10 y
Twelve-element
linear array bent
around a cylinder of
radius r = 3.6λ.
(xn, yn)
f
x
r
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 45

2.3 Array Performance 45

10 FIGURE 2-11
Array factor for the
12-element
0 Uncompensated conformal array (with
and without phase
compensation).
− 10
AF (dB)

−20

−30 Compensated

−40
0 45 90 135 180
f (degrees)

2.3 ARRAY PERFORMANCE

The ability to place nulls in the array pattern is highly dependent on the array architecture.
Array designers start with the following requirements specified:
1. Resolution (gain)
2. Scan coverage
3. Sidelobe level
4. Cost (primarily driven by the number of elements)
These design specifications will limit the possible approaches to adaptive nulling. Control
of the main beam gain, sidelobe level, and null placement is possible via the
1. Number of elements in the array
2. Element phase
3. Element amplitude
4. Element positions
5. Size of the array
6. Shape of the array
Trade-offs exist among array size, resolution, and cost. Increased control of the array
pattern comes at a high cost. Increasing the number of elements in an array increases
not only the gain but also the cost. The same is true for adaptive weights. More adaptive
weights allows more control over null placement, but the cost and complexity for that
control are high.

2.3.1 Enhanced Signal Reception by Adjustment

of the Array Response
To demonstrate steering and modifying the array pattern to enhance desired signal re-
ception and simultaneously suppress interference signals by complex weight selection,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 46

46 CHAPTER 2 Adaptive Array Concept

FIGURE 2-12 Desired signal Interference signal

Two-element array p(t) = Pe jw0t I(t) = Ne jw0t
for interference p
suppression 6
example.

d = l 0 /2

w1 + jw2 w1 + jw2

Array output

consider the two element array of isotropic point sources in Figure 2-12 in which a desired
signal arrives from the normal direction θ = 0◦ , and the interference signal arrives from
the angle θ = 30◦ . For simplicity, both the interference signal and the desired pilot signal
are assumed to be at the same frequency f 0 . Furthermore, assume that at the point exactly
midway between the array elements the desired signal and the interference are in phase
(this assumption is not required but simplifies the development). The output signal from
each element is input to a variable complex weight, and the complex weight outputs are
then summed to form the array output.
Now consider how the complex weights can be adjusted to enhance the reception of
p(t) while rejecting I (t). The array output due to the desired signal is

Pe jω0 t {[w 1 + w 3 ] + j[w 2 + w 4 ]} (2.48)

For the output signal of (2.48) to be equal to p(t) = Pe jω0 t , it is necessary that

w1 + w3 = 1
(2.49)
w2 + w4 = 0
The incident interfering noise signal exhibits a phase lead with respect to the array midpoint
when impinging on the element with complex weight w 3 + jw 4 of value 2π( 14 ) sin(π/6) =
π/4 and a phase lag when striking the other element of value −π/4. Consequently, the
array output due to the incident noise is given by

N e j (ω0 t−π/4) [w 1 + jw 2 ] + N e j (ω0 t+π/4) [w 3 + jw 4 ] (2.50)

Now
1
e j (ω0 t−π/4) = √ [e jω0 t (1 − j)]
2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 47

2.3 Array Performance 47

and
1
e j (ω0 t+π/4) = √ [e jω0 t (1 + j)]
2
so for the array noise response to be zero it is necessary that

w1 + w2 + w3 − w4 = 0
(2.51)
−w 1 + w 2 + w 3 + w 4 = 0

Solving (2.49) and (2.51) simultaneously then yields

1 1 1 1
w1 = , w2 = − , w3 = , w4 = (2.52)
2 2 2 2
With the previous weights, the array will accept the desired signal while simultaneously
rejecting the interference.
While the complex weight selection yields an array pattern that achieves the desired
system objectives, it is not a very practical way of approaching adaptive nulling. The
method used in the previous example exploits the facts that there is only one directional
interference source, that the signal and interference sources are monochromatic, and that
a priori information concerning the frequency and the direction of arrival of each signal
source is available. A practical processor must work without detailed information about
the location, number, and nature of signal sources. Nevertheless, this example has demon-
strated that an adaptive algorithm achieves certain performance objectives by adjusting the
complex weights. The development of a practical adaptive array algorithm is undertaken
in Part 2.

2.3.2 Interference Rejection

Improving the reception of a desired signal requires that the gain in the direction of the
desired signal is maximized while the gain in the direction of the interfering signals is
minimized. This goal is accomplished by appropriately modifying the signal magnitude
or phase at all or some of the elements. Amplitude tapering or thinning an array produces
low sidelobes at the requisite angles. An alternative approach is to synthesize nulls in the
directions of the interfering signals.

[Link] Low Sidelobe Amplitude Tapers

The amplitude of interference entering the sidelobes is proportional to the sidelobe level.
A binomial amplitude taper [14] applied to an array results in an array factor with no side-
lobes. Unfortunately, the binomial array is not practical, because it is extremely inefficient
and results in a low-gain main beam.
A better alternative is the Dolph-Chebyshev amplitude taper. All the sidelobes peaks
are equal and are at a level determined by the array designer [15]. The following equation
provides the zeros of the array factor that correspond to a sidelobe level in dB of sll:
⎧ ⎫
⎨ cos (n−0.5)π ⎬
N −1
ψn = 2 cos−1 (2.53)
⎩ cosh πζ ⎭
N −1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 48

48 CHAPTER 2 Adaptive Array Concept

FIGURE 2-13 p /2
Unit circle for the
eight-element 30 dB
Chebyshev taper.

p 0

−p/2

where
1  
ζ = cosh−1 10sll/20 (2.54)
π
The factored polynomial is found using ψn . Once the ψn are known, the polynomial of
degree N − 1 in factored form easily follows. The polynomial coefficients are the array
amplitude weights.
As an example, designing an eight-element 30 dB Chebyshev array starts with calcu-
lating ζ = 1.1807. Substituting into (2.53) results in angular locations (in radians) on the
unit circle shown in Figure 2-13 and given by

ψn = 3.1416, ±2.389, ±1.1985, ±1.6941

Finally, the normalized amplitude weights are shown in Figure 2-14 with the corresponding
array factor in Figure 2-15.
The Chebyshev taper is not practical for large arrays, because the amplitude weights
are difficult to implement in a practical array. The Taylor taper [16] is widely used in the
design of low sidelobe arrays. The first n̄ − 1 sidelobes on either side of the main beam
are sll, whereas the remaining sidelobes decrease as 1/ sin θ . The Taylor taper moves the
first n̄ − 1 nulls on either side of the main beam to
⎧ "
⎨sin−1 ±λn̄ ζ 2 +(n−0.5)22 n < n̄
Nd ζ 2 +(n̄−0.5)
θn =   (2.55)
⎩
sin−1 ±λnNd
n ≥ n̄

As with the Chebyshev taper, ζ is first found from (2.54). Next, the array factor nulls
are found using (2.55) and then substituting into ψn . Finally the factored array polynomial
is multiplied together to get the polynomial coefficients, which are also the Taylor weights.
A Taylor amplitude taper is also available for circular apertures as well [17].
The design of a 16-element 25 dB n̄ = 5 Taylor taper starts by calculating ζ = 1.1366.
Substituting into (2.55) results in angular locations (in radians) on the unit circle shown
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 49

2.3 Array Performance 49

1 FIGURE 2-14
Amplitude weights
for the eight-element
30 dB Chebyshev
0.8 taper.

0.6
Amplitude

0.4

0.2

0
1 2 3 4 5 6 7 8
Element

0 FIGURE 2-15
Array factor for the
eight-element 30 dB
−10 Chebyshev taper.

−20
AF (dB)

−30

−40

−90 −45 0 45 90
q (degrees)

in Figure 2-16 and given by

ψn = 3.1416, ±2.7489, ±2.3562, ±1.9635, ±1.5582, ±1.1628, ±0.79576, ±0.52396

Finally, the normalized amplitude weights are shown in Figure 2-17 with the corresponding
array factor in Figure 2-18.

[Link] Thinned Arrays

Not only can an amplitude taper lower the sidelobes of an array, but tapering the element
density can also mimic an amplitude taper and lower the sidelobes as well. Aperiodic arrays
have element spacings that are not periodic [18–24]. Nonuniformly spacing elements in
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 50

50 CHAPTER 2 Adaptive Array Concept

FIGURE 2-16 p /2
Unit circle for 20 dB
Taylor n̄ = 5 taper.

p 0

−p /2

FIGURE 2-17 1
Amplitude weights
for Taylor 25 dB
n̄ = 5 taper.
0.8

0.6
Amplitude

0.4

0.2

0
0 5 10 15
Element

large arrays is impractical due to the complexity of the feed network and the inconsistency
in the mutual coupling environment of the elements. As a result, density tapering in a large
array is accomplished by “thinning” or removing active elements from the element lattice
in the array aperture.
If the desired amplitude taper function is normalized, then it looks like a probability
density function. A uniform random number generator assigns a random number to each
element. If that random number exceeds the desired normalized amplitude taper for that
element, then the element is passive in the array; otherwise, it is active. An active element
is connected to the feed network, whereas an inactive element is not. The advantages of
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 51

2.3 Array Performance 51

0 FIGURE 2-18
Array factor for 20
dB Taylor n̄ = 5
−10 taper.

−20
AF (dB)

−30

−40

−90 −45 0 45 90
q (degrees)

thinning include the following:

• Simple feed network
• No amplitude weights needed
• Reduced number of active elements for the same beamwidth as a similar fully populated
array
An expression for the root mean square (rms) sidelobe level of a statistically thinned
array is given by [25]
1
sll2 = (2.56)
Nactive

where sll2 is the power level of the average sidelobe level, and Nactive is the number of
active elements out of N elements in the array. An expression for the peak sidelobe level
of linear and planar arrays having half-wavelength spacing is found by assuming all the
sidelobes are within three standard deviations of the rms sidelobe level and has the form
[25]
  N /2
P all sidelobes < sll2p  1 − e−sll p /sll
2 2
(2.57)

Statistical thinning was used until the 1990s when computer technology enabled
engineers to numerically optimize the thinning configuration of an array to find the desired
antenna pattern. Current approaches to array thinning include genetic algorithms [26] and
particle swarm optimization [27]. In addition, array thinning with realistic elements, like
dipoles [28], spirals [29], and thinned planar arrays [30], are used in place of point sources.
As an example, Figure 2-19 shows the array factor for a 100-element thinned linear
array with half-wavelength spacing. The thinning was performed using a 25 dB Taylor
n̄ = 5 amplitude taper as the probability density function. The array factor shown has
a peak sidelobe level of −18.5 dB below the peak of the main beam and has 77% of
the elements turned on. This array factor had the lowest relative sidelobe level of 20
independent runs. The genetic algorithm was used to find the thinning configuration that
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 52

52 CHAPTER 2 Adaptive Array Concept

FIGURE 2-19 0
Statistically thinned
100-element array.
−5

−10

AF (dB)
−15

−20

−25

−30
−90 −45 0 45 90
q (degrees)

FIGURE 2-20 0
Genetic algorithm
thinned 100-element
−5
array.

−10
AF (dB)

−15

−20

−25

−30
−90 −45 0 45 90
q (degrees)

yields the lowest relative sidelobe level. Figure 2-20 shows the optimized array factor with
a −23.2 dB peak relative sidelobe level when 70% of the elements are turned on.

[Link] Null Synthesis Using the Unit Circle

Just as array factor nulls can be moved to synthesize low sidelobes, they can also be moved
in certain directions where interference signals exist. For instance, a null can be placed at
−38◦ of the eight-element uniform array with half-wavelength spacing by moving the zero
at −3π/4 to π sin(−38◦ ) (Figure 2-21(a)), which equates to moving the null at −48.6◦
to −38◦ (Figure 2-21(b)). Moving this zero in (2.35) results in an array polynomial with
complex weights:
     ◦   
z + e jπ/4 z + e− jπ/4 z + e jπ/2 z + e− jπ/2 z + e− jπ sin(38 ) z + e j3π/4 z + e jπ
= z7 + (0.65 + j0.23) z 6 + (1.06 + j0.32) z 5 + (0.83 − j0.04) z 4 (2.58)
+ (0.74 + j0.37) z + (1.09 + j0.15) z + (0.68 + j0.06) z + 0.91 + j0.41
3 2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 53

2.3 Array Performance 53

0
Nulled Quiescent
−10
p /2

−20

AF (dB)
−30
p 0

−40

−90 −45 0 45 90
−p/2 q (degrees)
(a) (b)

0
Nulled
Quiescent
−10
p/2

−20
AF (dB)

−30
p 0
−40

−90 −45 0 45 90
−p /2 q (degrees)
(c) (d)

FIGURE 2-21 Null synthesis with the unit circle. a: Moving one zero from −3π/4 to
π sin(−38◦ ). b: Array factor nulls moves from −48.6◦ to −38◦ . c: Moving a second zero from
3π/4 to π sin(38◦ ). d: Second-array factor null moves from 48.6◦ to 38◦ .

A null can be placed at −38◦ using real weights by also moving the null in the array factor
at 48.6◦ to 38◦ to form complex conjugate pairs:
     ◦  ◦  
z + e jπ/4 z + e− jπ/4 z + e jπ/2 z + e− jπ/2 z + e− jπ sin(38 ) z + e jπ sin(38 ) z + e jπ
= z7 + 0.30z 6 + 1.29z 5 + 0.59z 4 + 0.59z 3 + 1.29z 2 + 0.30z + 1 (2.59)

These examples of using null synthesis with the unit circle illustrate several points
about adaptive nulling:
1. There are an infinite number of ways to place a null at −38◦ . After one of the zeros
is placed at the desired angle, then the other zeros can be redistributed as well. Thus,
placing a null at a given location does not require a unique set of weights unless further
restrictions are applied.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 54

54 CHAPTER 2 Adaptive Array Concept

2. Some weights produce greater distortion to the array factor than other weights. The
weight selection impacts the gain of the array factor as well as the sidelobe levels and
null positions.
3. An N element array with complex weights can uniquely place N − 1 nulls. If the
weights are real, then only N /2 − 1 nulls can be uniquely placed.
4. Adaptive algorithms maneuver the nulls on the unit circle indirectly by changing the
element weights. The root music algorithm presented in Chapter 10 makes use of the
zeros on the unit circle.
The next section shows an approach to null synthesis that results in small changes in the
array weights and minimal distortion to the quiescent array factor.

[Link] Null Synthesis Using Cancellation Beams

Adaptive weights can be written as a perturbation to the quiescent weights

w n = an (1 + n ) (2.59)

where an is the amplitude taper, and n is the complex weight perturbation that causes
the nulls. Substituting (2.59) into the equation for a linear array factor results in a nulled
array factor that is the sum of the quiescent array factor and a cancellation array factor
[31].

N N N
wne jkxn u
= an e jkxn u
+ n an e jkxn u (2.60)
n=1 n=1 n=1
# $% & # $% &
quiescent array factor cancellation array factor

An infinite number of combinations of w n place nulls in the M desired directions. Setting

(2.60) equal to zero and writing the resulting equation in matrix form results in

AWT = BT (2.61)

where
⎡ ⎤
a1 e jkx1 u 1 a2 e jkx2 u 1 · · · a N e jkx N u 1
⎢ a1 e jkx1 u 2 a2 e jkx2 u 2 · · · a N e jkx N u 2 ⎥
⎢ ⎥
A=⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
a1 e jkx1 u M
a2 e jkx2 u M
· · · aN e jkx N u M

W = [1 2 · · ·  N ]
 N 
 
N N
B=− an e jkxn u 1
an e jkxn u 2
··· an e jkxn u M
n=1 n=1 n=1

Since (2.61) has more columns than rows, finding the weights requires a least squares
solution in the form

W = A† (AA† )−1 B (2.62)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 55

2.3 Array Performance 55

Ideally, the quiescent pattern should be perturbed as little as possible when placing
the nulls. The weights that produce the nulls are [32]
M
w n = an − γm cn e− jnkdu m (2.63)
m=1

where
γm = sidelobe level of the quiescent pattern at u m
cn = amplitude taper of the cancellation beam
When cn = an , the cancellation pattern looks the same as the quiescent pattern. When
cn = 1.0, the cancellation pattern is a uniform array factor and is the constrained least
mean square approximation to the quiescent pattern over one period of the pattern. The
nulled array factor can now be written as
N N M N
wne jkxn u
= an e jkxn u
− γm an e jkxn (u−u m ) (2.64)
n=1 n=1 m=1 n=1

The sidelobes of the cancellation pattern in (2.64) produce small perturbations to the
nulled array factor.
As an example, consider an eight-element 30 dB Chebyshev array with interference
entering the sidelobe at θ = −40◦ . When cn = an , the cancellation beam is a Chebyshev
array factor as shown in Figure 2-22(a), whereas when cn = 1.0, the cancellation beam is
a uniform array factor as shown in Figure 2-22(b).
This procedure also extends to phase-only nulling by taking the phase of (2.63). When
the phase shifts are small, then e jφn ≈ 1 + jφn , and the array factor can be written as [33]

N N N N N
w n e jkxn u = an e jδn e jkxn u ≈ an (1 + jδn )e jkxn u = an e jkxn u + j an δn e jkxn u
n=1 n=1 n=1 n=1 n=1
(2.65)

Adapted
0 Cancellation Adapted
0 Cancellation
Quiscent
Quiscent

−20
Gain (dB)

−20
Gain (dB)

−40
−40

−60
−60
−90 −45 0 45 90
−90 −45 0 45 90
q (degrees) q (degrees)
(a) (b)

FIGURE 2-22 Array factors for an eight-element array with a 30 dB Chebyshev taper when
a null is synthesized at θ = −40◦ . (a) cn = an . (b) cn = 1.0.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 25, 2010 21:5 56

56 CHAPTER 2 Adaptive Array Concept

FIGURE 2-23
Array factors for an Adapted
eight-element array 0 Cancellation
with a 30 dB Quiscent
Chebyshev taper
when a phase-only
null is synthesized at −20

Gain (dB)
θ = −40◦ .

−40

−60

−90 −45 0 45 90
q (degrees)

Using Euler’s identity, (2.65) is written as


N 
N
an δn [cos (kxn u m ) + j sin (kxn u m )] e jkxn u = j an [cos (kxn u m ) + j sin (kxn u m )]
n=1 n=1
(2.66)
Equating the real and imaginary parts of (2.66) leads to


N 
N
an δn sin (kxn u m ) e jkxn u = an cos(kxn u m ) (2.67)
n=1 n=1


N 
N
an δn cos (kxn u m ) e jkxn u
=− an sin(kxn u m ) = 0 (2.68)
n=1 n=1

Now, (2.67) is in the form that can be solved via (2.62).

Repeating the previous example of an eight-element 30 dB Chebyshev array with
interference entering the sidelobe at θ = −40◦ results in the nulled array factor and
associated cancellation beam shown in Figure 2-23. An interesting observation is that the
sidelobe at θ = 40◦ increased. This phenomenon occurs in phase-only nulling unless large
phase shifts are used to perform the nulling [34].
The adapted patterns in Figure 2-22 and Figure 2-23 are quite different. If the array
polynomials associated with those patterns are factored, then the unit circle represents can
be graphed as shown in Figure 2-24(a) and Figure 2-24(b). In both cases, a zero appears on
the unit circle at ψ = −115.7◦ . All of the other zeros differ. The zeros in Figure 2-24(a)
are on the unit circle, whereas two of the zeros in Figure 2-24(b) are off the unit circle.
A monopulse phased array has separate sum and difference channels. The sum pattern
has a peak, whereas the difference pattern has a null in the direction of the desired signal.
The sum pattern peak detects the desired signal, and the difference pattern null accurately
locates it. Figure 2-25 is a diagram of a monopulse array with sum and difference channels
that are split at each element. Prior to the split, both channels share beam steering phase
shifters (and amplitude weights) that can be used to simultaneously place nulls in the sum
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 57

2.3 Array Performance 57

p /2 p/2

p 0 p 0

−p /2 −p /2
(a) (b)

FIGURE 2-24 Unit circle representations of the array factors in (a) Figure 2-22 (b) Figure 2-23.

FIGURE 2-25
Diagram of an array
with low sidelobe
w1 w2 … wN sum and difference
channels.

a1 b1 a2 b2 aN bN

Σ − Σ +

and difference patterns. The matrixes in (2.61) are now written as [35]
⎡ ⎤
a1 e jkx1 u 1 · · · a N e jkx1 u M
⎢ .. .. .. ⎥
⎢ . . . ⎥
⎢ ⎥
⎢a1 e jkx N u 1 · · · aN e jkx N u M ⎥
A=⎢ ⎢ b1 e jkx1 u 1 · · ·
⎥
⎢ b N e jkx1 u M ⎥
⎥
⎢ .. .. .. ⎥
⎣ . . . ⎦
b1 e jkx N u 1
··· bN e jkx N u M
' (
W = 1 2 · · ·  N
 N 
 
N 
N 
N
B = −j an e jkxn u 1 ··· an e jkxn u M
bn e jkxn u 1
··· bn e jkxn u M
n=1 n=1 n=1 n=1

Phase-only nulling requires further simplification as noted in (2.66).

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 58

58 CHAPTER 2 Adaptive Array Concept

10 10

−10 −10
Gain (dB)

Gain (dB)
−30 −30

−50 −50

−90 −45 0 45 90 −90 −45 0 45 90

q (degrees) q (degrees)
(a) (b)

FIGURE 2-26 Nulls are simultaneously placed at θ = −40◦ in the sum and difference array
factors for a 16-element monopulse array. (a) 25 dB n̄ = 5 Taylor sum pattern. (b) 25 dB n̄ = 5
Bayliss difference pattern. The dotted lines are the quiescent patterns.

Consider a 16-element array with a 25 dB n̄ = 5 Taylor sum pattern and a 25 dB n̄ = 5

Bayliss difference [11]. Assume that the interference enters the sidelobes at θ = 52◦ . After
synthesizing the shared weights to place the nulls, the resulting sum and difference patterns
are shown in Figure 2-26. The resulting patterns have the desired nulls.

2.4 NULLING LIMITATIONS DUE TO

MISCELLANEOUS ARRAY EFFECTS
A principal measure of the effectiveness of an adaptive array in cancelling an undesired
interference signal is the ratio of the total output noise power, P0 , to internal noise power,
PN , as a function of frequency. To obtain this ratio for a simple example, consider the
two-element adaptive array of Figure 2-27. If the interference signal arriving at element
1 from the direction θ is s(t), then the signal appearing at element 2 is s(t + τ ), where
τ = (d/v) sin θ, and v = wavefront propagation velocity.
The signals in the two element channels are weighted and summed to form the array
output y(t) so that

y(t) = w 1 s(t) + w 2 s(t + τ ) (2.69)

which can be written in the frequency domain as

Y (ω) = S(ω)[w 1 + w 2 e− jωτ ] (2.70)

To exactly cancel the interference signal at the array output at a particular frequency,
f 0 (termed the center frequency of the jamming signal), it is necessary that

w 2 = −w 1 e jω0 τ (2.71)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 59

2.4 Nulling Limitations Due to Miscellaneous Array Effects 59

Pj , f0 FIGURE 2-27
Two-element
s(t) adaptive array with
PN jammer located at θ .
w1

q
d Σ y (t)

w2

s(t + τ)

If we select the two adaptive weights to satisfy (2.69), it follows that the interference signal
component of the array output at any frequency is given by

Y (ω) = S(ω)w 1 [1 − e− jτ (ω−ω0 ) ] (2.72)

and consequently the interference component of output power at any frequency may be
expressed as (assuming momentarily for simplicity that |S(ω)|2 = 1)

|Y (ω)|2 = |w 1 |2 {2 − 2cos[τ (ω − ω0 )]} (2.73)

Let |S(ω)|2 = PJ and denote the constant spectral density of the internal noise power in
each channel by PN , it follows that the total output noise power spectral density, Po (ω),
can be written as

Po (ω) = 2|w 1 |2 [1 − cos(τ (ω − ω0 ))]PJ + 2|w 1 |2 PN (2.74)

If we recognize that the output thermal noise power spectral density is just Pn = 2|w 1 |2 PN
and normalize the previous expression to Pn , it follows that the ratio of the total output
interference plus noise power spectral density, P0 , to the output noise power spectral
density, Pn , is then given by

P0 {1 − cos[τ (ω − ω0 )]}PJ
(ω) = 1 + (2.75)
Pn PN

Where
PJ /PN = jammer power spectral density to internal noise power spectral density per
channel
τ = (d/v) sin θ
d = sensor element spacing
θ = angular location of jammer from array broadside
ω0 = center frequency of jamming signal
At the center frequency, a null is placed exactly in the direction of the jammer so
that P0 /Pn = 1. If, however, the jamming signal has a nonzero bandwidth, then the other
frequencies present in the jamming signal will not be completely attenuated. Figure 2-28
is a plot of (2.75) and shows the resulting output interference power density in decibels as
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 60

60 CHAPTER 2 Adaptive Array Concept

FIGURE 2-28 12
Two-element
cancellation 10 q = 90°
performance: P0 /Pn d = l 0 /2
versus frequency for f0 = 300 MHz
8
PJ /PN = 30 and

(dB)
40 dB.
6
PJ
= 40 dB

Pn
P0
PN
4
30 dB
2

0
295 300 305

a function of frequency for a 10 MHz bandwidth jamming signal located at θ = 90◦ and
an element spacing of λ0 /2 for two values of PJ /PN . The results shown in Figure 2-28
indicate that for PJ /PN = 40 dB about 12 dB of uncanceled output interference power
remains at the edges of the 10 MHz band (with f 0 = 300 MHz). Consequently, the output
residue power at that point is about 40 − 12 = 28 dB below the interfering signal that
would otherwise be present if no attempt at cancellation were made.
It can also be noted from (2.75) and Figure 2-28 that the null bandwidth decreases as
the element spacing increases and the angle of the interference signal from array broadside
increases. Specifically, the null bandwidth is inversely proportional to the element spacing
and inversely proportional to the sine of the angle of the interference signal location.
Interference signal cancellation depends principally on three array characteristics: (1)
element spacing; (2) interference signal bandwidth; and (3) frequency-dependent interele-
ment channel mismatch across the cancellation bandwidth. The effect of sensor element
spacing on the overall array sensitivity pattern has already been discussed. Yet another
effect of element spacing is the propagation delay across the array aperture.
Assume that a single jammer having power PJ is located at θ and that the jamming
signal has a flat power spectral density with (double-sided) bandwidth B Hz. It may be
shown that the cancellation ratio P0 /PJ , where P0 is the (canceled) output residue power
in this case, is given by

P0 sin2 (π Bτ )
=1− (2.76)
Pj (π Bτ )2

where τ is given by (2.8). Note that (2.76) is quite different from (2.75), a result that
reflects the fact that the weight value that minimizes the output jammer power is different
from the weight value that yields a null at the center frequency. Equation (2.76) is plotted
in Figure 2-29, which shows how the interference signal cancellation decreases as the
signal bandwidth–propagation delay product, Bτ , increases. It is seen from (2.76) that the
cancellation ratio P0 /PJ is inversely proportional to the element spacing and the signal
bandwidth.
An elementary interchannel amplitude and phase mismatch model for a two-element
array is shown in Figure 2-30. Ignoring the effect of any propagation delay, the output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 61

2.4 Nulling Limitations Due to Miscellaneous Array Effects 61

Jam FIGURE 2-29

me
r + Cancellation ratio
Cancelled
q output versus jammer
−10 − signal bandwidth–
w1 propagation delay
−20
product Bτ .
(dB)

−30
−40
P0
PJ

−50
−60
0.001 0.01 0.1
Bt

Sensor FIGURE 2-30

element Channel gain = a Cancellation ratio
1 − P0 /PJ versus
Phase error Output signal y (t) amplitude and
= e jφ phase channel
Sensor +
element mismatch for a
−20 2 two-element array.
nly
rso
rro
−30 ee
litud
Amp
(dB)

−40 nly
so
P0
PJ

r
erro
a se
−50 Ph

−60
0.01 0.1 1.0 10
Error: amplitude (dB), phase (degrees)

signal may be expressed as

y(t) = 1 − ae jφ (2.77)

From (2.77) it follows that the canceled output power is given by

|y(t)|2 = |1 − ae jφ |2 = 1 + a 2 − 2a cos φ (2.78)

Figure 2-30 shows a plot of the cancellation ratio P0 /PJ based on (2.78) for amplitude
errors only (φ = 0) and phase errors only (a = 1). In Part 3, where adaptive array
compensation is discussed, more realistic interchannel mismatch models are introduced,
and means for compensating such interchannel mismatch effects are studied. The results
obtained from Figure 2-30 indicates that the two-sensor channels must be matched to
within about 0.5 dB in amplitude and to within approximately 2.8◦ in phase to obtain 25
dB of interference signal cancellation.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 62

62 CHAPTER 2 Adaptive Array Concept

2.5 NARROWBAND AND BROADBAND SIGNAL

PROCESSING CONSIDERATIONS
We have seen that by adjusting the complex weights in the two-element array of Figure 2-12
it is possible to modify the overall array pattern to place a pattern null in the direction of an
undesired interference signal. So long as the interference signal is “narrowband” (i.e., can
be adequately characterized by a single frequency, ω0 ), the resulting array pattern will
succeed in suppressing the undesired interference signal. The most common to implement
a complex weight is a quadrature hybrid circuit (Figure 2-31). Each individual sensor
element is divided into an in-phase and quadrature-phase component by means of a 90◦
phase shifter. Each component is then applied to a variable weight whose output is summed
" resulting complex gain factor is then Ae , where φ =
jφ
to form the output signal. The
− tan−1 (w 2 /w 1 ), and A = w 12 + w 22 . The weights w 1 , w 2 assume a continuum of both
positive and negative values, and the range of magnitude A is limited only by the range of
limitations of the two individual weights. It should be recognized that it is not absolutely
necessary to maintain a phase shift of exactly 90◦ to obtain quite satisfactory processing
with quadrature hybrid circuitry.
In the event that the interference signal is broadband, then the complex weight selection
appropriate for one frequency ω1 will not be appropriate for a different frequency ω2 , since
the array pattern nulls shift as the value of λ0 changes. Figure 2-32 shows the array factor
expanded around the null placed at 52◦ for the center frequency ( f 0 ), 5% below the
center frequency (0.95 f 0 ), and 5% above the center frequency (1.05 f 0 ). This example

FIGURE 2-31
Realization of a w1
complex weight by
Sensor
means of a Σ
quadrature hybrid element
circuit. 90° w2

FIGURE 2-32 The 0

region around the
null at 52◦ over a
10% bandwidth. −10 f0

−20
AF (dB)

1.05 f0
0.95 f0

−30

−40

−50
40 45 50 55 60 65
f (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 63

2.5 Narrowband and Broadband Signal Processing Considerations 63

Time Time Time FIGURE 2-33

delay delay delay Transversal filter
Sensor realized with a
element
Δ Δ Δ wL Σ
tapped-delay line
having L complex
w1 w2 w3 weights.

demonstrates that the desired null is narrowband, because it moves as the frequency
changes. This fact leads to the conclusion that different complex weights are required
at different frequencies if an array null is to be maintained in the same direction for
all frequencies of interest. A simple and effective way of obtaining different amplitude
and phase weightings at a number of frequencies over the band of interest is to replace
the quadrature hybrid circuitry of Figure 2-31 by a transversal filter having the transfer
function h(ω). Such a transversal filter consists of a tapped-delay line having L complex
weights as shown in Figure 2-33 [33,34]. A tapped-delay line has a transfer function
that is periodic, repeating over consecutive filter bandwidths B f , as shown in Appendix
A. If the tap spacing is sufficiently close and the number of taps is large, this network
approximates an ideal filter that allows exact control of gain and phase at each frequency
within the band of interest. An upper limit on the tap spacing is given by the desired
array cancellation bandwidth Ba , since B f ≥ Ba and B f = 1/ for uniform tap spacing.
The transversal filter not only is useful for providing the desired adjustment of gain and
phase over the frequency band of interest for wideband signals but is also well suited for
providing array compensation for the effects of multipath, finite array propagation delay,
and interchannel mismatch effects; these additional uses of transversal filters are explored
further in Part 3. Transversal filters are also enhance the ability of an airborne moving
target indication (AMTI) radar system to reject clutter, to compensate for platform motion,
and to compensate for near-field scattering effects and element excitation errors [35].
To model a complete multichannel processor (in a manner analogous to that of Section
1.4 for the narrowband processing case), consider the tapped-delay line multichannel
processor depicted in Figure 2-34. The multichannel wideband signal processor consists
of N sensor element channels in which each channel contains one tapped-delay line
like that of Figure 2-33 consisting of L tap points, (L − 1) time delays of  seconds
each, and L complex weights. On comparing Figure 2-34 with Figure 1-1, it is seen that
x1 (t), x2 (t), · · · x N (t) in the former correspond exactly with x1 (t), x2 (t) . . . x N (t) in the
latter, which were defined in (1.6) to form the elements of the vector x(t). In like manner,
therefore, define a complex vector x 1 (t) such that

(x 1 )T = (x )T = [x1 x2 · · · x N ] (2.79)

The signals appearing at the second tap point in all channels are merely a time-delayed
version of the signals appearing at the first tap point, so define a second complex signal
vector x 2 (t) where

[x 2 (t)]T = [x (t − )]T

(2.80)
= [x1 (t − )x2 (t − ) · · · x N (t − )]
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 64

64 CHAPTER 2 Adaptive Array Concept

1st Tap l th Tap (L − 1) Tap Lth Tap

Sensor point point point point
Δ Δ Δ
no. 1
x1(t ) x1[t − (l − 1)Δ] x1[t − (L − 2)Δ] x1[t − (L − 1)Δ]

w11 w1l w1(L−1) w1L

Sensor
Δ Δ Δ
no. k
xk [t − (l − 1)Δ] xk [t − (L − 2)Δ] xk [t − (L − 1)Δ]
xk(t )

Processor
wk1 wkl wk (L−1) wkL output
Σ signal
y (t)

Sensor
Δ Δ Δ
no. N
xN [t − (l − 1)Δ] xN [t − (L − 2)Δ] xN [t − (L − 1)Δ]
xN(t )

wN 1 wN l wN (L−1) wN L

FIGURE 2-34 Tapped-delay line multichannel processor for wideband signal processing.

Continuing in the previously given manner for all L tap points, a complete signal vector
for the entire multichannel processor can be defined as
⎡ ⎤ ⎡ ⎤
x1 (t) x (t)
⎢ x2 (t) ⎥ ⎢ 
x (t − ) ⎥
⎢ ⎥ ⎢ ⎥
x(t) = ⎢ . ⎥ = ⎢ .. ⎥ (2.81)
⎣ .. ⎦ ⎣ . ⎦
xL (t) x [t − (L − 1)]

It is seen that the signal vector x(t) contains L component vectors each having dimen-
sion N .
Define the weight vector


(w1 )T = [w 11 w 21 · · · w N 1 ] (2.82)

The weight vector for the entire multichannel processor is then given by
⎡
⎤
w1
⎢w ⎥
 ⎢ 2⎥
w=⎢ . ⎥ (2.83)
⎣ .. ⎦
wL
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 65

2.5 Narrowband and Broadband Signal Processing Considerations 65

It is seen that the weight vector w so defined is a N L-dimensional vector containing L

component vectors, each having dimension N .
As a consequence of the signal and weight vector definitions previously introduced,
the output of the multichannel tapped-delay line processor can be written in the form

L L
y(t) = (wl )T xl (t) = (wl )T x [t − (l − 1)]
l=1 l=1 (2.84)
= w x(t)
T

which is the same form as (1.4). Yet another reason for expressing the signal vector in the
form of (2.81) is that this construction leads to a Toeplitz form for the correlation matrix
of the input signals, as shown in Chapter 3.
The array processors of Figures 2-33 and 2-34 are examples of classical time-domain
processors that use complex weights and delay lines. Fast Fourier transform (FFT) tech-
niques replace a conventional time-domain processor by an equivalent frequency domain
processor using the frequency-domain equivalent of time delay [39–41]. When a tapped
delay-line array and an FFT array use the same time delay between samples and the same
number of samples, their performance is identical. FFT processing has two advantages
compared with its time-domain equivalent: (1) the hardware problem can be alleviated;
and (2) the computational burden can be reduced. The use of phase shifters or delay lines to
form a directional beam in the time domain becomes cumbersome from a hardware stand-
point as the number of delay elements increases, whereas using the frequency-domain
equivalent permits beamforming to be accomplished with a digital computer, thereby sim-
plifying the hardware. Likewise, taking FFTs tends to reduce the correlation between sam-
ples in different frequency subbands. When samples in different subbands are completely
decorrelated, the signal covariance matrix has a block diagonal form, and the optimal
weights can be computed separately in each subband, thus reducing the computational
burden.
As more taps are added to each delay line, the bandwidth performance of the array
improves. The bandwidth cutoff, Bc , of an array is defined as the maximum bandwidth
such that the array SINR is within 1dB of its continuous wave (narrowband) value. It is
better to divide the taps equally among the array elements rather than to employ an unequal
distribution, although higher bandwidth performance can usually be realized by placing
an extra tap behind the middle element in an array. A piecewise linear curve is plotted in
Figure 2-35, giving the optimal number of taps per element (number of samples for FFT
processing) versus signal fractional bandwidth, B, for 3 to 10 elements in a linear array
(B = BW/carrier frequency). Since it is easiest to use 2n samples in an FFT processor,
where n is some integer, one would use the smallest value of n such that 2n is at least
as large as the value shown in Figure 2-35. The time delay between taps, To , can be
described in terms of the parameter r , where r is the number of quarter wavelength delays
found in To . For optimal SINR array performance, the value of r must be in the range
0 < r < 1/B. If r < 1/B, the performance is optimal but results in large array weights.
Therefore, the choice r = 1/B is dictated by good engineering practice. The notion of
the frequency-domain equivalent of time delay is introduced in the Problems section, and
more detailed presentations of the techniques of digital beamforming may be found in
[42–45].
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 66

66 CHAPTER 2 Adaptive Array Concept

FIGURE 2-35 6
Number of taps
versus fractional
bandwidth, B. From 5
Vook and Compton,
IEEE Trans. Aerosp. 10
and Electron. Sys., 4 Number of elements
Vol. 38, No. 3, July Number of taps indicated along
each graph) 6
1992.
7, 8, 9, 10
3

8, 9, 10 4, 5, 6, 7 3, 4, 5, 6
2
3 elements
3, 4, 5, 6, 7
1

0
0 0.05 0.10 0.15 0.20
Fractional bandwidth, B

2.6 ADAPTIVE ARRAY PERFORMANCE

MEASURE—COVERAGE IMPROVEMENT
FACTOR (CIF) [44]
At any point over the array field of view, the value of the jammer-to-signal ratio can be
computed as

n
G i Pi + N0
i=1
J/S = (2.85)
G s Ps
where Pi is the ith jammer’s power at the face of the array, G i is the array gain toward
the ith jammer, N0 is the thermal noise power, Ps is the desired signal power at the array
face, G s is the array gain toward the desired signal, and n is the number of jammers
present.
A typical J/S sky contour plot over a planar array field of view (in this case the upper
hemisphere) is shown in Figure 2-36 The circle edges in this figure occur at an elevation
angle of 0◦ , and the apex of the hemisphere (elevation angle of 90◦ ) occurs at the circle
center with other values of constant elevation angle forming concentric circles inside the
outer circle. The contours of constant J/S values in this figure indicate that four jammers
are present: three near the horizon at azimuth values of about 20◦ , 80◦ , and 110◦ , and one
at about 60◦ elevation and 20◦ azimuth. Plotting the percentage of the sky for which J/S is
less than a specified level then yields the J/S coverage statistic plot of Figure 2-37, which
corresponds to the J/S contours of Figure 2-36 across the sky azimuth and elevation.
Figure 2-37 shows that a J/S ratio of 46 dB is achieved over 72% of the sky.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 67

2.6 Adaptive Array Performance Measure—Coverage Improvement Factor (CIF) 67

68 FIGURE 2-36 J/S

55 70 sky contour plot.
40
38 35 50
60 65 From Weiss,
60
Aerospace Corp.
55
50 50 Technical Report,
45
TOR-95(5474)-1, 20
December 1994
60

45 55
35
40 50
sky elevation

45
50 40
38 45
45 50 55 40
40 60
35

36

35
35

38

sky azimuth

100 FIGURE 2-37 J/S

coverage statistics
90 for Figure 2-36.
From Weiss,
80
Aerospace Corp.
70 Technical Report,
TOR-95(5474)-1, 20
60 December 1994.
% Coverage

50

40

30

20

10
0
20 30 40 50 60 70 80 90 100
J/S (dB)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 68

68 CHAPTER 2 Adaptive Array Concept

The coverage improvement factor (CIF) of the array at each point in the field of view
can be computed as
n 
G P +N0q 
i=1 iq i 
G sq Ps 
CIF = n 
before nulling
(2.86)
G i Pi +N0 
i=1 
G s Ps 
after nulling

where the numerator gains G iq and G sq denote the array gain toward the ith jammer and
the desired signal, respectively, for the quiescent pattern case, and N0q denotes the thermal
noise power before nulling.
The coverage improvement factor statistic (CIFS) can then be obtained by plotting
the percentage of the field of view where the nulling system provides X dB of protection
compared with the quiescent system in the same manner as the J/S coverage statistic plot
was obtained and yields a convenient way of assessing how effective the adaptive array is
in yielding a performance advantage over a nonadaptive array.

2.7 SUMMARY AND CONCLUSIONS

The signal environment in which an adaptive array operates is introduced by discussing
the different types of interference and modeling of the desired signal as either a random or
nonrandom process. Signal propagation effects are also important because of the distortion
introduced into the desired signal, and the rate of change of variations in the ambient noise
environment directly affects the ability of an adaptive system to cope with a nonstationary
signal environment by adjusting the pattern-forming network elements.
Sensor element spacing within an array directly affects its ability to respond appro-
priately to any given signal environment and to yield enhanced reception performance.
The potential of an array to offer enhanced reception performance is demonstrated with a
simple example. The presence of undesirable pattern grating lobes may be avoided either
by using a “filled array” for systems having a small number of elements or by using a
“thinned array” by exploiting the theory of randomly spaced elements for large arrays
to achieve an efficient design. In addition to element spacing, interference signal nulling
limitations may also derive from the following:
1. Jammer signal bandwidth
2. Signal propagation delay across the array aperture
3. Interchannel mismatch effects
Narrowband and broadband signal processing requirements differ in that narrowband
processing can be achieved by means of a single constant complex weight in each element
channel whereas a broadband signal requires frequency-dependent weighting that leads
to the introduction of a transversal filter (tapped delay line). Time-domain processors can
be replaced by equivalent frequency-domain processors whose functions are realized by
a digital computer, thereby alleviating the need for cumbersome hardware.
A convenient way of assessing the effectiveness of an adaptive array compared with a
nonadaptive array was discussed by introducing the concept of the CIF and its associated
CIFS.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 69

2.8 Problems 69

The best possible steady-state performance that can be achieved by an adaptive array
system can be computed theoretically, without explicit consideration of the array factors
affecting such performance. The theoretical performance limits for steady-state array
operation are addressed in Chapter 3.

2.8 PROBLEMS
1. Circular Array. For the circular array of Figure 2-38 having N equally spaced elements, select
as a phase reference the center of the circle. For a signal whose direction of arrival is θ with
respect to the selected angle reference, define the angle φk for which φk = θ − ψk .
(a) Show that ψk = (2π/N )(k − 1) for k = 1, 2, . . . , N .
(b) Show that the path length phase difference for any element with respect to the center of
the circle is given by u k = π(R/λ) cos φk .
(c) Show that an interelement spacing of λ/2 is maintained by choosing the radius of the array
circle to be
λ
R=
4 sin(π/N )

2. Linear Array with Spherical Signal Wavefront. Consider the linear array geometry of
Figure 2-39 in which v is the vector from the array center to the signal source, θ is the angle
between v and the array normal, and  is the propagation velocity.

FIGURE 2-38
fk Azimuth plane of
circular array having
N equally spaced
q elements.
Yk

0°
R

y FIGURE 2-39
Linear array
Signal source geometry.

v
z
l
q

x
L L
−
2 2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 70

70 CHAPTER 2 Adaptive Array Concept

(a) Show that for a spherical signal wavefront the time delay experienced by an element
located at position x in the array with respect to the array center is given by
) *     +
||v|| x2 x
τ (x) = 1− 1+ −2 sin θ
 ||v||2 ||v||

(τ is positive for time advance and negative for time delay) where ||v|| denotes the length
of the vector v.
(b) Show that for a spherical signal wavefront the attenuation factor experienced by an element
located at position x in the array with respect to the array center is given by
||v|| 1
ρ(x) = = ,
||l|| x2
1+ ||v||2
−2 x
||v||
sin θ

where ||l|| denotes the distance between the signal source and the array element of concern.
3. Uniformly Spaced Linear Array [11]. Show that for a uniformly spaced linear array having N
elements the normalized array factor is given by
1 sin(N u/2)
AF(u) =
N sin(u/2)
where u = 2π(d/λ) sin θ.
4. Representation of Nonuniformly Spaced Arrays [20]
(a) For nonuniformly spaced elements in a linear array, a convenient “base separation” d can
be selected and the element position then represented by
 
i
di = + εi d
2

where εi is the fractional change from the uniform spacing represented by d. The normal-
ized field pattern for a nonuniformly spaced array is then given by
N N   
1 1 i
AF = cos u i = cos + εi u
N N 2
i=1 i=1

where u = 2π(d/λ) sin θ. Show that AF can be put in the form

1 - u u.
AF = AF u − sin εi u sin i + (1 − cos εi u) cos i
N 2 2
i

where Au is the pattern of a uniform array having element separation equal to the base
separation and is given by the result obtained in Problem 3.
(b) Assuming all εi u are small, show that the result obtained in part (a) reduces to
u u
AF = AF u − εi sin i
N 2
i

(c) The result obtained in part (b) can be rearranged to read

u N
εi sin i = (AF u − AF)
2 u
i
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 71

2.8 Problems 71

which may be regarded as a Fourier series representation of the quantity on the right-hand
side of the previous equation. Consequently, the εi are given by the formula for Fourier
coefficients:
 0
2N 1 u
εi = (AF u − AF) sin i du
π π u 2
Let
L
AF u − AF 1
= ak δ(u − u k )
u u
k=1

and show that

L
N sin[(i/2)u k ]
εi = 2 ak
π uk
k=1

(d) For uniform arrays the pattern sidelobes have maxima at positions approximately given
by
π
u k = (2k + 1)
N
Consequently, u k gives the positions of the necessary impulse functions in the represen-
tation of (AF u − AF)/u given in part (c). Furthermore, since the sidelobes adjacent to the
main lobe drop-off as 1/u, (AF u − AF)/u is now given by

AF u − AF 1
L - π .
= AF 2 (−1)k δ u − (2k + 1)
u u N
k=1

For this representation of (AF u − AF)/u show that

 3 L
N sin(iπ/2N )(2k + 1)
εi = 2AF (−1)k
π (2k + 1)2
k=1

5. Nonuniformly Spaced Arrays and the Equivalent Uniformly Spaced Array [19]. One method
of gaining insight into the gross behavior of a nonuniformly spaced array and making it
amenable to a linear analysis is provided by the correspondence between a nonuniformly
spaced array and its equivalent uniformly spaced array (EUA). The EUA provides a best mean
square representation for the original nonuniformly spaced array.
(a) Define the uniformly spaced array factor A(θ) as the normalized phasor sum of the re-
sponses of each of the sensor elements to a narrowband signal in the uniform array of
Figure 2-40, where the normalization is taken with respect to the response of the center

FIGURE 2-40
Uniformly spaced
q array having an odd
number of sensor
elements.

3′ 2′ 1′ 1 2 3

d
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 72

72 CHAPTER 2 Adaptive Array Concept

FIGURE 2-41
Symmetrical
q
nonuniformly
spaced array having
an odd number of
sensor elements.
3′ 2′ 1′ 1 2 3

d1

d2

d3

element. Show that

(N −1)/2

AF(θ ) = 1 + 2 cos[2πil sin θ ]

i=1

where l = d/λ. Furthermore, show that when d/λ = 2 the array factor has its second
principal maxima occurring at θ = π/6.
(b) Show that the nonuniformly spaced array factor for the array of Figure 2-41 is given by

(N −1)/2

AF(θ) = 1 + 2 cos[2πli sin θ ]

i=1

where li = di /λ.
(c) The result obtained in part (b) may be regarded as AF(θ ) = 1 + 2i cos ωi where
 
2
ωi = li (π R sin θ ), R = arbitrary scale factor
R # $% &
ω1

Regarding ω1 = π g sin θ as the lowest nonzero superficial array frequency in the EUA,
the scale factor R determines the element spacing in the EUA. For example, R = 1
corresponds to half-wavelength spacing, and R = 12 corresponds to quarter-wavelength
spacing. Each term in AF(θ) may be expanded (in a Fourier series representation for each
higher harmonic) into an infinite number of uniformly spaced equivalent elements. Since
the amplitude of the equivalent elements varies considerably across the array, only a few
terms of the expansion need be considered for a reasonable equivalent representation in
terms of EUAs. The mth term in AF(θ) is given by

2
cos ωm = cos lm ω1 = cos μm ω1
R

where μm = (2lm /R)

The mth term may consequently be expanded in an integral harmonic cosine series as
follows:
∞
cos μm ω1 = amv cos(vω1 )
v=0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 73

2.8 Problems 73

where
 πR
2
amv = cos μm ω1 cos vω1 dω1
π 0

Show that when R1 = 1

2 μm (−1)v sin μm π
amv = for μm = 0, ±1, ±2, · · ·
π μ2m − v 2
(d) Applying the harmonic cosine series expansion given in part (c) to each component cos ωi ,
show that the resulting amplitude coefficients may be added term by term to give the EUA
representation:
P
) ∞
+
AF(θ ) = 1 + 2 amv cos(vω1 )
i=1 v=0
P
=1+2 Ar cos r ω1
r =0

where
AF 0 = a10 + a20 + a30 + · · ·
AF 1 = a11 + a21 + a31 + · · ·

Note that the representation given by the previous result is not unique. Many choices are
possible as a function of the scale factor R. From the least mean square error property of
the Fourier expansion, each representation will be the best possible for a given value of ω1 .

6. Array Propagation Delay Effects. Array propagation delay effects can be investigated by
considering the simple two-element array model of Figure 2-42. It is desired to adjust the value
of w 1 so the output signal power P0 is minimized for a directional interference signal source.
(a) Since x2 (ω) = e− jωτ x1 (ω) in the frequency domain, the task of w 1 may be regarded as
one of providing the best estimate (in the mean square error sense) of x1 (ω)e− jωτ , and
the error in this estimate is given by ε(ω) = x1 (ω)(e− jωτ − w 1 ). For the weight w 1 to be
optimal, the error in the estimate must be orthogonal to the signal x1 (ω) so it is necessary
that
E{|x1 (ω)|2 [e− jωτ − w 1 ]} = 0

where the expectation E{·} is taken over frequency. If |x1 (ω)|2 is a constant indepen-
dent of frequency, it follows from the previously given orthogonality condition that w 1

d sin q FIGURE 2-42

t= b Two-element linear
array having a single
complex weight
x2(t) +
R(w, q ) with interelement
propagation delay, τ .
−
d

q
l PJ
gna
nce si x1(t)
rfere w1
Inte
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 74

74 CHAPTER 2 Adaptive Array Concept

must satisfy

E{e− jωτ − w 1 } = 0

If the signal has a rectangular spectrum over the bandwidth −π B ≤ ω ≤ π B, show that
the previous equation results in

sin(π Bτ )
w1 =
π Bτ
(b) The output power may be expressed as
 πB
P0 = φ R R (ω, θ )dω
−π B

where φ R R (ω, θ ) is the output power spectral density. The output power spectral density
is in turn given by

φ R R (ω, θ) = |e− jωτ − w 1 |2 φss (ω)

where φss (ω) is the power spectral density of the directional interference signal. Assuming
that φss (ω) is unity over the signal bandwidth (so that PJ = 2π B), then from part (a) it
follows that
  
 − jωτ sin(π Bτ ) 2
πB
P0 =  −  dw
e π Bτ 
−π B

Show that P0 can be expressed as

 2
P0 sin(π Bτ )
=1−
PJ π Bτ

where PJ is the directional interference signal power.

7. Open-Loop Scanning of Adaptively Formed Beams [47]. To scan an adaptively formed beam
about the direction of a received signal requires open-loop scanning. Open-loop scanning,
however, requires knowledge of the array element locations. An airborne thinned array may
contain elements distributed over the entire air frame. Although element coordinates may be
measured accurately on the ground, an airframe is flexible; consequently, the relative element
positions may vary considerably in flight, with rms deviations of 10−1 ft easily occurring.
For open-loop scanning, four theoretical results have been obtained as follows [47]:
1. The rms array element coordinate tolerance is

λ
σ x = σ y = σz =
4π θs

where λ is the radiated wavelength, and θs is the maximum scan angle from the initial
pointing direction of the array (i.e., θs is half the field of view).
2. The rms tolerance in the estimate of initial pointing direction is

θ
σθ ∼
=
2θs

where θ is the beamwidth ∼

= λ/D, and D is the linear dimension of the array.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 75

2.8 Problems 75

3. The distance beyond which far-field scanning may be employed is given by

θs D 2
Rm ≈
2λ
where D is the linear dimension of the array.
4. The rms tolerance in the range estimate of the target is
R0 θ
σ R0 ∼
=
2θs
where R0 < Rm is the target range.
(a) Assuming λ = σx = σ y = σz = 10−1 ft, what field of view can successfully be scanned
with open-loop techniques?
(b) What is the minimum range at which far-field scanning may be used?
(c) For the field of view found in part (a), determine the rms tolerance in the initial pointing
error.
8. Frequency-Domain Equivalent of Time Delay [45]. The conventional time-domain proces-
sor for beamforming involving sum and delay techniques can be replaced by an equivalent
frequency-domain processor using the FFT.
The wavefront of a plane wave arriving at a line of sensors having a spacing d between
elements will be delayed in time by an amount τ1 between adjacent sensors given by
 
d
τ1 = sin θ1

where θ1 is the angle between the direction of arrival and the array normal. Consequently, for
a separation of n elements

τn1 = nτ1

In classical time-domain array processing, a spatial beam is formed at an angle θ1 by introducing

appropriate delay lines at the output of each sensor. Let xn (t) denote the input signal from the
plane wave at the nth sensor, and let the time delay present in the nth sensor channel be τn .
Since the array output is formed by coherent addition of each of the sensor outputs, the output
y(t) is given by
N −1
y(t) = xn (t − τn )
n=0

for an array having N sensors

(a) Show the following time–frequency domain relationships based on the Fourier transfor-
mation:

Time Domain Frequency Domain

xn (t) X n (ω)
xn (t − τn ) X n (ω)e− jφn
N
−1 N
−1
xn (t − τn ) X n (ω)e− jφn
n=0 n=0

Note that φn represents the phase shift equivalent to the time delay τn .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 76

76 CHAPTER 2 Adaptive Array Concept

(b) The plane wave arriving at the array with arrival angle θ1 can be represented by
- z .
x(t, z) = cos ω t − = cos[ωt − kz]

where k is the wavenumber of the plane wave, and z is the axis of propagation at an angle
θ1 with respect to the array normal. Show that the phase shift φn associated with a time
delay τn is given by

φn = ωτn

where
 
d
τn = n sin θ1


(c) It is convenient to index θ and ω to account for different combinations of the angle θl and
frequency ω so that

θ = θl = lθ, l = 0, 1, . . . , N − 1
ω = ωm = mω, m = 0, 1, . . . , M − 1

Furthermore, index the input time samples in i so that

t = ti = it, i = 0, 1, . . . , M − 1

Let X n (ω) be represented at discrete frequencies (a discrete Fourier transform) by

X n (mω) = X n (ωm ) = X mn

Likewise, the sampled time waveform from the nth array element can be represented by

xn (it) = xn (ti ) = xin

The time–frequency domain equivalence between xin and X mn is then given by

Time Domain Frequency Domain

xin = xn (it) ↔ X n (mω) = X mn

Consequently, the summed expression from part (a) can be written as

N −1 N −1
− jφn
X n (ω)e = (mω)e− jφn
n=0 n=0

Show that φnl = nkd sin θl . Consequently, show that the previously summed expressions
can be rewritten in terms of both frequency and angle as
N −1
X m (1θ ) = X n (mω)e− jφnl
n=0

(d) The phase shift φnl can be expressed as

nl
φnl =
N
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 77

2.8 Problems 77

implying that μ = 1/Nkd where sin(1θ ) = 1μ, so that

N −1
X ml = X (1θ ) = Nn (mω)e− jnl/N
n=0

which is of the form of a discrete Fourier transform (DFT).

Show that X mn = X n (mω) can also be expressed as a DFT
M−1
X mn = xn (it)e− jim/N
i=0

Now expressing xn (it) as

xn (it) = xni

a double DFT can now be written as

N −1 M−1
X ml = xni e− jim/N e− jnl/N
n=0 i=0

The array output has therefore been obtained in the form of spectral sample versus beam
number, which results after a two-dimensional DFT transformation of array input. The
required DFTs can then be implemented using the FFT algorithm.
9. Linear Array Directivity. Calculate the directivity of a 10-element uniform array with d = 0.5λ,
and compare it with a 10-element uniform array with d = 1.0λ. Why are they different?
10. Moving Nulls on the Unit Circle. Plot the unit circle representations of a four-element uniform
array. Now, move the null locations from ψ = ±90◦ to ψ = ±120◦ .
11. Chebyshev Array. Plot the unit circle representation, the amplitude weights, and the array
factor for a six-element 20 dB Chebyshev array.
12. Taylor Array. Plot the unit circle representation, the amplitude weights, and the array factor
for a 20-element 20 dB n̄ = 5 Taylor array.
13. Taylor Array. Start with a Taylor n̄ = 4 sll = 25dB taper, and place a null at u = 0.25 when
d = 0.5λ. Do not allow complex weights.
14. Thinned Array. A 5,000-element linear array has a taper efficiency of 50% and elements spaced
a half wavelength apart. What are the average and peak sidelobe levels of this array?
15. Thinned Array with Taylor Taper. Start with a 100-element array with element spacing d = 0.5,
and thin to a 20 dB n̄ = 5 Taylor taper.
16. Cancellation Beams. A 40-element array of isotropic point sources spaced λ/2 apart has a
30 dB n̄ = 7 Taylor taper. Plot the array factors and cancellation beam for a null at 61o .
17. Cancellation Beams. Repeat the previous example using both uniform and weighted cancel-
lation beams.
18. Cancellation Beams. A 40-element array of isotropic point sources spaced λ/2 apart has a
40 dB n̄ = 7 Taylor taper. Plot the array factors and cancellation beams for nulls at 13◦ and
61◦ . Use phase-only nulling.
19. Simultaneous Nulling in the Sum and Difference Patterns. A 40-element array of isotropic
point sources spaced λ/2 apart has a 30 dB n̄ = 7 Taylor taper and a 30 dB n̄ = 7 Bayliss
taper. Plot the array factors and cancellation beams for nulls at 13◦ and 61◦ . Use simultaneous
phase-only nulling in the sum and difference patterns.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 78

78 CHAPTER 2 Adaptive Array Concept

20. Triangular Spacing in a Planar Array [11]. Show that the uniform array factor for triangular
spacing is
   
sin (N xe ψx ) sin N ye ψ y sin (N xo ψx ) sin N yo ψ y
AF =   + e− j (ψx +ψ y )  
N xe sin (ψx ) N ye sin ψ y N xo sin (ψx ) N yo sin ψ y

2.9 REFERENCES
[1] P. W. Howells, “Explorations in Fixed and Adaptive Resolution at GE and SURC,” IEEE
Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 575–584.
[2] R. L. Deavenport, “The Influence of the Ocean Environment on Sonar System Design,”
IEEE 1975 EASCON Record, Electronics and Aerospace Systems Convention, September
29–October 1, pp. 66.A–66.E.
[3] A. A. Winder, “Underwater Sound—A Review: II. Sonar System Technology,” IEEE Trans.
Sonics and Ultrason., Vol. SU-22, No. 5, September 1975, pp. 291–332.
[4] L. W. Nolte and W. S. Hodgkiss, “Directivity or Adaptivity?” IEEE 1975 EASCON Record,
Electronics and Aerospace Systems Convention, September 29–October 1, pp. 35.A–35.H.
[5] G. M. Wenz, “Review of Underwater Acoustics Research: Noise,” J. Acoust. Soc. Am., Vol.
51, No. 3 (Part 2), March 1972, pp. 1010–1024.
[6] B. W. Lindgren, Statistical Theory, New York, MacMillan, 1962, Ch. 4.
[7] F. Bryn, “Optimum Structures of Sonar Systems Employing Spatially Distributed Receiv-
ing Elements,” in NATO Advanced Study Institute on Signal Processing with Emphasis on
Underwater Acoustics, Vol. 2, Paper No. 30, Enschede, The Netherlands, August 1968.
[8] H. Cox, “Optimum Arrays and the Schwartz Inequality,” J. Acoust. Soc. Am., Vol. 45, No. 1,
January 1969, pp. 228–232.
[9] R. S. Elliott, “The Theory of Antenna Arrays,” in Microwave Scanning Antennas, Vol. 2, Array
Theory and Practice, edited by R. C. Hansen, New York, Academic Press, 1966, Ch. 1.
[10] I. J. Gupta and A. A. Ksienski, “Effect of Mutual Coupling on the Performance of Adaptive
Arrays,” IEEE AP-S Trans., Vol. 31, No. 5, September 1983, pp. 785–791.
[11] R. L. Haupt, Antenna Arrays: A Computational Approach, New York, Wiley, 2010.
[12] M.I. Skolnik, Introduction to Radar Systems, New York, McGraw-Hill, 2002.
[13] H. T. Friis, “A Note on a Simple Transmission Formula,” IRE Proc., May 1946, pp. 254–256.
[14] J. S. Stone, United States Patent No. 1,643,323 and 1,715,433.
[15] C. L. Dolph, “A Current Distribution for Broadside Arrays which Optimizes the Relationship
Between Beam Width and Side-Lobe Level,” Proceedings of the IRE, Vol. 34, No. 6, June
1946, pp. 335–348.
[16] T. T. Taylor, “Design of Line Source Antennas for Narrow Beamwidth and Low Side Lobes,”
IRE AP Trans., Vol. 3, 1955, pp. 16–28.
[17] T. Taylor, “Design of Circular Apertures for Narrow Beamwidth and Low Sidelobes,” IEEE
AP-S Trans., Vol. 8, No. 1, January 1960, pp. 17–22.
[18] F. Hodjat and S. A. Hovanessian, “Nonuniformly Spaced Linear and Planar Array Antennas
for Sidelobe Reduction,” IEEE Trans. Antennas Propag., Vol. AP-26, No. 2, March 1978,
pp. 198–204.
[19] S. S. Sandler, “Some Equivalences Between Equally and Unequally Spaced Arrays,” IRE
Trans. Antennas and Progag., Vol. AP-8, No. 5, September 1960, pp. 496–500.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 79

2.9 References 79

[20] R. F. Harrington, “Sidelobe Reduction by Nonuniform Element Spacing,” IRE Trans. Antennas
Propag., Vol. AP-9, No. 2, March 1961, pp. 187–192.
[21] D. D. King, R. F. Packard, and R. K. Thomas, “Unequally Spaced, Broadband Antenna
Arrays,” IRE Trans. Antennas Propag., Vol. AP-8, No. 4, July 1960, pp. 380–385.
[22] M. I. Skolnik, “Nonuniform Arrays,” in Antenna Theory, Part I, edited by R. E. Collin and F.
J. Zucker, New York, McGraw-Hill, 1969, Ch. 6.
[23] Y. T. Lo, “A Mathematical Theory of Antenna Arrays with Randomly Spaced Elements,”
IEEE Trans. Antennas Propag., Vol. AP-12, May 1964, pp. 257–268.
[24] Y. T. Lo and R. J. Simcoe, “An Experiment on Antenna Arrays with Randomly Spaced
Elements,” IEEE Trans. Antennas Propag., Vol. AP-15, March 1967, pp. 231–235.
[25] E. Brookner, “Antenna Array Fundamentals—Part 1,” in Practical Phased Array Antenna
Systems, Norwood, MA, Artech House, 1991, pp. 2-1–2-37.
[26] R.L. Haupt, “Thinned Arrays Using Genetic Algorithms,” IEEE AP-S Transactions, Vol. 42,
No. 7, July 1994, pp. 993–999.
[27] R. L. Haupt and D. Werner, Genetic Algorithms in Electromagnetics, New York, Wiley, 2007.
[28] R. L. Haupt, “Interleaved Thinned Linear Arrays,” IEEE AP-S Trans., Vol. 53, September
2005, pp. 2858–2864.
[29] R. Guinvarc’h and R. L. Haupt, “Dual Polarization Interleaved Spiral Antenna Phased Array
with an Octave Bandwidth” IEEE Trans. Antennas Propagat., Vol. 58, No. 2, 2010, pp. 397–
403.
[30] R. L. Haupt, “Optimized Element Spacing for Low Sidelobe Concentric Ring Arrays,” IEEE
AP-S Trans., Vol. 56, No. 1, January 2008, pp. 266–268.
[31] H. Steyskal, “Synthesis of Antenna Patterns with Prescribed Nulls,” IEEE Transactions on
Antennas and Propagation, Vol. 30, no. 2, 1982, pp. 273–279.
[32] H. Steyskal, R. Shore, and R. Haupt, “Methods for Null Control and Their Effects on the
Radiation Pattern,” IEEE Transactions on Antennas and Propagation, Vol. 34, No. 3, 1986,
pp. 404–409.
[33] H. Steyskal, “Simple Method for Pattern Nulling by Phase Perturbation,” IEEE Transactions
on Antennas and Propagation, Vol. 31, No. 1, 1983, pp. 163–166.
[34] R. Shore, “Nulling a Symmetric Pattern Location with Phase-Only Weight Control,” IEEE
Transactions on Antennas and Propagation, Vol. 32, No. 5, 1984, pp. 530–533.
[35] R. Haupt, “Simultaneous Nulling in the Sum and Difference Patterns of a Monopulse Antenna,”
IEEE Transactions on Antennas and Propagation, Vol. 32, No. 5, 1984, pp. 486–493.
[36] B. Widrow and M. E. Hoff, Jr., “Adaptive Switching Circuits,” IRE 1960 WESCON Convention
Record, Part 4, pp. 96–104.
[37] J. H. Chang and F. B. Tuteur, “A New Class of Adaptive Array Processors,” J. Acoust. Soc.
Am., Vol. 49, No. 3, March 1971, pp. 639–649.
[38] L. E. Brennan, J. D. Mallet, and I. S. Reed, “Adaptive Arrays in Airborne MTI Radar,” IEEE
Trans. Antennas Propag., Special Issue Adaptive Antennas, Vol. AP-24, No. 5, September
1976, pp. 607–615.
[39] R. T. Compton, Jr., “The Bandwidth Performance of a Two-Element Adaptive Array with
Tapped Delay-Line Processing,” IEEE Trans. Ant. & Prop., Vol. AP-36, No.1, Jan. 1988,
pp. 5–14.
[40] R. T. Compton, Jr., “The Relationship Between Tapped Delay-Line and FFT Processing in
Adaptive Arrays,” IEEE Trans. Ant. & Prop., Vol. AP-36, No. 1, January 1988, pp. 15–26.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:9 80

80 CHAPTER 2 Adaptive Array Concept

[41] F. W. Vook and R. T. Compton, Jr., “Bandwidth Performance of Linear Adaptive Arrays with
Tapped Delay-Line Processing,” IEEE Trans. Aerosp. & Electron. Sys., Vol. AES-28, No. 3,
July 1992, pp. 901–908.
[42] P. Rudnick, “Digital Beamforming in the Frequency Domain,” J. Acoust. Soc. Am., Vol. 46,
No. 5, 1969.
[43] D. T. Deihl, “Digital Bandpass Beamforming with the Discrete Fourier Transform,” Naval
Research Laboratory Report 7359, AD-737-191, 1972.
[44] I. M. Weiss, “GPS Adaptive Antenna/Electronics—Proposed System Characterization Per-
formance Measures,” Aerospace Corp. Technical Report No. TOR-95 (544)-1, 20 December
1994.
[45] L. Armijo, K. W. Daniel, and W. M. Labuda, “Applications of the FFT to Antenna Array
Beamforming,” EASCON 1974 Record, IEEE Electronics and Aerospace Systems Convention,
October 7–9, Washington, DC, pp. 381–383.
[46] A. M. Vural, “A Comparative Performance Study of Adaptive Array Processors,” Proceedings
of the 1977 International Conference on Acoustics, Speech, and Signal Processing, May,
pp. 695–700.
[47] S. H. Taheri and B. D. Steinberg, “Tolerances in Self-Cohering Antenna Arrays of Arbitrary
Geometry,” IEEE Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 733–739.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 81

CHAPTER

Optimum Array Processing:

Steady-State Performance 3
Limits and the Wiener Solution
' $
Chapter Outline
3.1 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.2 Signal Descriptions for Conventional and Signal Aligned Arrays. . . . . . . . . . . . . . . . . . . 88
3.3 Optimum Array Processing for Narrowband Applications . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.4 Optimum Array Processing for Broadband Applications . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.5 Optimum Array Processing for Perturbed Propagation Conditions . . . . . . . . . . . . . . . 121
3.6 Polarization Sensitive Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.7 Summary and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
3.9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
& %

Optimum array processing is an optimum multichannel filtering problem [1–12]. The ob-
jective of array processing is to enhance the reception (or detection) of a desired signal that
may be either random or deterministic in a signal environment containing numerous inter-
ference signals. The desired signal may also contain one or several uncertain parameters
(e.g., spatial location, signal energy, phase) that it may be advantageous to estimate.
Optimum array processing techniques are broadly classified as (1) processing appro-
priate for ideal propagation conditions and (2) processing appropriate for perturbed prop-
agation conditions. Ideal propagation implies an ideal nonrandom, nondispersive medium
where the desired signal is a plane (or spherical) wave and the receiving sensors are distor-
tionless. In this case the optimum processor is said to be matched to a plane wave signal.
Any performance degradation resulting from deviation of the actual operating conditions
from the assumed ideal conditions is minimized by the use of complementary methods,
such as the introduction of constraints. When operating under the aforementioned ideal
conditions, vector weighting of the input data succeeds in matching the desired signal.
When perturbations in either the propagating medium or the receiving mechanism
occur, the plane wave signal assumption no longer holds, and vector weighting the input
data will not match the desired signal. Matrix weighting of the input data is necessary [13]
for a signal of arbitrary characteristics. The principal advantage of such an element space-
matched array processor operates on noncoherent wavefront signals where matching can
be performed in only a statistical sense.
Steady-state performance limits establish a performance measure for any selected
design. Quite naturally, optimum array processing has mostly assumed ideal propagation

81
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 82

82 CHAPTER 3 Optimum Array Processing

conditions, and various approaches to this problem have been proposed for both narrow-
band and broadband signal applications. By far, the most popular and widely reported
approaches involve the adoption of an array performance measure that is optimized by ap-
propriate selection of an optimum weight vector. Such performance measure optimization
approaches have been widely used for radar, communication, and sonar systems and there-
fore are discussed at some length. Approaches to array processing that use the maximum
entropy method (which is particularly applicable to the passive sonar problem) [14,15]
and eigenvalue resolution techniques [16,17] have been reported.
To determine the optimal array weighting and its associated performance limits, some
mathematical preliminaries are first discussed, and the signal descriptions for conventional
and signal-aligned arrays are introduced. It is well known that when all elements in an
array are uniformly weighted, then the maximum signal-to-noise ratio (SNR) is obtained
if the noise contributions from the various element channels have equal power and are
uncorrelated [18]. When there is any directional interference, however, the noise from the
various element channels is correlated. Consequently, selecting an optimum set of weights
involves attempting to cancel correlated noise components. Signal environment descrip-
tions in terms of correlation matrices therefore play a fundamental role in determining the
optimum solution for the complex weight vector.
The problem of formulating some popular array performance measures in terms of
complex envelope signal characterizations is discussed. It is a remarkable fact that the
different performance measures considered here all converge (to within a constant scalar
factor) toward the same steady-state solution: the optimum Wiener solution. Passive de-
tection systems face the problem of designing an array processor for optimum detection
performance. Classical statistical detection theory yields an array processor based on
a likelihood ratio test that leads to a canonical structure for the array processor. This
canonical structure contains a weighting network that is closely related to the steady-state
weighting solutions found for the selected array performance measures. It therefore turns
out that what at first appear to be quite different optimization problems actually have a
unified basis.

3.1 MATHEMATICAL PRELIMINARIES

Our derivation of the optimal weight settings represents the signal envelopes as well
as the adaptive weights in their complex envelope form. The meaning of such complex
representations therefore is briefly reviewed.
From Appendix B, the relationship between real signals and their corresponding
complex envelope representations is

actual signal = Re{(complex envelope representation)e jω0 t } (3.1)

where Re{ } denotes “real part of.” This result is further discussed in Appendix B. Applying
this notation to the array output yields

actual y(t) = Re{xT w∗ e jω0 t } = Re{w† xe jω0 t } (3.2)

where w† x is the complex envelope representation of y(t), ∗ denotes complex conju-

gate, and † denotes complex conjugate transpose [( )∗ ]T . (Boldface lowercase symbols
denote vectors; boldface uppercase symbols denote matrices.) A closely related complex
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 83

3.1 Mathematical Preliminaries 83


representation is the analytic signal ψ(t) for which y(t) = Re{ψ(t)} and


ψ(t) = y(t) + j y̌(t) (3.3)

where y̌(t) denotes the Hilbert transform of y(t). It follows that for a complex weight
representation w = w 1 + jw 2 having an input signal with analytic representation x1 + j x2 ,
the resulting actual output signal is given by

actual ouput = Re{(w 1 − jw 2 )(x1 (t) + j x2 (t))} = w 1 x1 (t) + w 2 x2 (t) (3.4)

where x2 (t) = x̌1 (t). Note that (complex envelope) e jω0 t = ψ(t). As a consequence of
the previously given meaning of complex signal and weight representations, two alternate
approaches to the optimization problems for the optimal weight solutions may be taken
as follows:
1. Reformulate the problem involving complex quantities in terms of completely real
quantities so that familiar mathematical concepts and operations are conveniently car-
ried out.
2. Revise the definitions of certain concepts and operations (e.g., covariance matrices,
gradients) so that all complex quantities are handled directly in an appropriate manner.
Numerous examples of both approaches are found in the adaptive array literature. It is
therefore appropriate to consider how to use both approaches since there are no compelling
advantages favoring one approach over the other.

3.1.1 Problem Formulation in Terms of Real Variables

Let z 1 be the first component of a complex vector z having n components; z may represent,
for example, a weight or signal vector. Furthermore, let the real part of z 1 be denoted by
x1 and the imaginary part of z 1 be denoted by x2 . Continuing in this manner, we have

z k = x2k−1 + j x2k (3.5)

It follows that the n-component complex vector z is completely represented by the 2n-
component real vector x. By representing all complex quantities in terms of corresponding
real vectors, the adaptive array problem is solved using familiar definitions of mathematical
concepts and procedures. If we carry out the foregoing procedure for the single complex
weight w 1 + jw 2 and the complex signal x1 + j x2 , there results
 
x1
w x = [w 1 w 2 ]
T
= w 1 x1 + w 2 x2 (3.6)
x2

which is in agreement with the result expressed by (3.4). If the foregoing approach
is adopted, then certain off-diagonal elements of the corresponding correlation matrix
(defined in the following section) will be zero. This fact does not affect the results
obtained in this chapter; the point is discussed more fully in connection with an example
presented in Chapter 11.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 84

84 CHAPTER 3 Optimum Array Processing

3.1.2 Correlation Matrices for Real Signals

Once the signal and noise processes in an adaptive array system have been described in
terms of statistical properties, the system performance is conveniently evaluated in terms
of its average behavior. The evaluation of average behavior leads directly to an interest
in quantities related to the second statistical moment such as autocorrelation and cross-
correlation matrices. The degree of correlation that exists between the components of two
random vectors is given by the elements of the correlation matrix between the two vectors
[19]. For example, the cross-correlation matrix between the vectors x(t) and y(t) having
stationary statistical properties is defined by

Rx y (τ ) = E{x(t)yT (t − τ )} (3.7)

where E{ } denotes the expected value, and τ is a running time-delay variable. Likewise,
the autocorrelation matrix for the vector x(t) is defined by

Rx x (τ ) = E{x(t)xT (t − τ )} (3.8)

If a signal vector x(t) consists of uncorrelated desired signals and noise components
so that

x(t) = s(t) + n(t) (3.9)

then

Rx x (τ ) = Rss (τ ) + Rnn (τ ) (3.10)

The correlation matrices of principal concern in the adaptive array analysis are those
for which the time-delay variable τ is zero. Rather than write the correlation matrix
argument explicitly as Rx y (0) and Rx x (0), we define

Rx x = Rx x (0) (3.11)

and

Rx y = Rx y (0) (3.12)

It follows that for an N-dimensional vector x(t), the autocorrelation matrix is simply written
as Rx x where
⎡ ⎤
x1 (t)x1 (t) x1 (t)x2 (t) . . . x1 (t)x N (t)
⎢ ⎥
 ⎢ x 2 (t)x 1 (t) x 2 (t)x 2 (t) . . . ⎥
Rx x = E{xx } = ⎢
T ⎢ ⎥ (3.13)
.. ⎥
⎣ . ⎦
x N (t)x1 (t) ... x N (t)x N (t)

where xi (t)xk (t) denotes E{xi (t)xk (t)}.

From the previous definitions, it immediately follows that

RTxx = Rx x and RTxy = R yx (3.14)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 85

3.1 Mathematical Preliminaries 85

so the autocorrelation matrix is symmetric. In general, the autocorrelation matrix Rx x is

only positive semidefinite, and the inverse matrix R−1
x x may not exist. Since in practice only
estimates of the signal environment correlation matrices are available and such estimates
are based on distinct time samples, then a sufficient number of such time samples for
a nonzero bandwidth signal process may guarantee positive definiteness so the inverse
matrix will exist. The signal vectors x(t) and n(t) of (3.9), for example, often contain
uncorrelated amplifier self-noise and therefore may be regarded as having positive defi-
nite correlation matrices. The desired signal vector s(t) may, however, contain correlated
components, as when a uniform plane wave arrives simultaneously at two or more array
sensor elements, thereby producing identical, in-phase signal components. Consequently,
the autocorrelation matrix Rss is only positive semidefinite, and its inverse R−1ss may not
exist.
Now consider the correlation matrix that results for the tapped delay-line processor
of Section 2.5 by examining the (N × N )-dimensional autocorrelation matrix given by

Rx x (τ ) = E{x (t)x (t − τ )}
T
(3.15)

where x (t) is the (real) signal vector defined in (2.80) for a tapped delay-line multichannel
processor. Likewise, for the NL-dimensional vector of all signals observed at the tap points,
the autocorrelation matrix is given by

Rx x (τ ) = E{x(t)xT (t − τ )} (3.16)

Substituting (2.82) for x(t) into (3.16) then yields

⎧⎡  ⎤ ⎫
⎪
⎪ x (t) ⎪
⎪
⎪
⎨⎢x (t − ) ⎥ ⎪
⎬
⎢ ⎥ T  T  T 
Rx x (τ ) = E ⎢. ⎥ x (t − τ )x (t − τ − ) · · · x [t − τ − (L − 1)]
⎪
⎪ ⎣.. ⎦ ⎪
⎪
⎪
⎩  ⎪
⎭
x [t − (L − 1)]
(3.17)
and, using (3.15), Rx x (τ ) then becomes
⎡ ⎤
Rx x (τ ) Rx x (τ + ) ... Rx x [τ + (L − 1)]
⎢ T .. ⎥
⎢ R (τ − ) Rx x (τ ) . ⎥
Rx x (τ ) = ⎢
⎢.
xx
..
⎥
⎥ (3.18)
⎣ .. . ⎦
RT
x x [τ − (L − 1)] ... 
Rx x (τ )

The (NL × NL)-dimensional matrix Rx x (τ ) given by (3.18) has the form of a Toeplitz
matrix—a matrix having equal valued matrix elements along any diagonal [20]. The
desirability of the Toeplitz form lies in the fact that the entire matrix is constructed from
the first row of submatrices; that is, Rx x (τ ), Rx x (τ + ), . . . , Rx x [τ + (L − 1)]. Con-
sequently, only an (N × NL)-dimensional matrix need be stored to have all the information
contained in Rx x (τ ).
Covariance matrices are closely related to correlation matrices since the covariance
matrix between the vectors x(t) and y(t) is defined by

cov[x(t), y(t)] = E{(x(t) − x)(y(t) − y)T } (3.19)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 86

86 CHAPTER 3 Optimum Array Processing

where

x = E{x(t)} and y = E{y(t)}

Thus, for zero-mean processes and with τ = 0, correlation matrices and covariance ma-
trices are identical, and the adaptive array literature frequently uses the two terms inter-
changeably.
Frequency-domain signal descriptions as well as time-domain signal descriptions
are valuable in considering array processors for broadband applications. The frequency-
domain equivalent of time-domain descriptions may be found by taking the Fourier trans-
form of the time domain quantities, (ω) = { f (t)}. The Fourier transform of a signal
correlation matrix yields the signal cross-spectral density matrix

x x (ω) = {Rx x (τ )} (3.20)

Cross-spectral density matrices therefore present the signal information contained in cor-
relation matrices in the frequency domain.

3.1.3 Revised Definitions Required for Complex Vector Quantities

The optimization problems that must be solved to yield the most desirable complex weight
vector choices for different performance measures involve the use of norms, gradients,
and covariance (or correlation) matrices of complex vector quantities. It is therefore useful
to consider the definitions of these terms for both real and complex vectors.
The norm of a vector in Hilbert space, denoted by ||x||, represents the length of the
vector. For a real vector

√
||x|| = xT x (3.21)

whereas for a complex vector


√
||x|| = x† x (3.22)

The gradient of a scalar function, ∇ y , consists of the partial derivatives of f (·) along each
component direction of y. In the case of real variables, the gradient operator is a vector
operator given by
 T
∂ ∂
∇y = ... (3.23)
∂ y1 ∂ yn

so that
∂f ∂f ∂f
∇ y f (y) = e1 + e2 + · · · + en (3.24)
∂ y1 ∂ y2 ∂ yn

where e1 , e2 , . . . , en is the set of unit basis vectors for the vector y. For a complex vector
y each element yk has a real and an imaginary component:

yk = xk + j z k (3.25)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 87

3.1 Mathematical Preliminaries 87

Therefore, each partial derivative in (3.24) now has a real and an imaginary component so
that [21]
∂f ∂f ( j)∂ f
= + (3.26)
∂ yk ∂ xk ∂z k
In the optimization problems encountered in this chapter, it is frequently desired to
obtain the gradient with respect to the vector x of the scalar quantity xT a = aT x and of
the quadratic form xT Ax (which is also a scalar), where A is a symmetric matrix. Note
that xT a is an inner product

(x, a) = xT a = aT x (3.27)

If A has a dyadic structure, the quadratic form xT Ax is regarded as an inner product

squared, or

xT Ax = (x, v)2 (3.28)

where

A = vvT (3.29)

The trace of a matrix product given by


trace[ABT ] = aik bik (3.30)
i k

has all the properties of an inner product, so formulas for the differentiation of the trace
of various matrix products are of interest in obtaining solutions to optimization problems.
A partial list of convenient differentiation formulas is given in Appendix C. From these
formulas it follows for real variables that

∇x (yT Ax) = yT A (3.31)

and

∇x (xT Ax) = 2Ax (3.32)

whereas for complex variables the corresponding gradients are given by

∇x (y† Ax) = y† A (3.33)

and

∇x (x† Ax) = 2Ax (3.34)

3.1.4 Correlation Matrices for Complex Signals

For complex vector quantities, the corresponding correlation matrix definitions must be
revised from (3.7) and (3.8) so that
 
Rx x = E{x∗ xT } and Rx y = E{x∗ yT } (3.35)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 88

88 CHAPTER 3 Optimum Array Processing

An alternative correlation matrix definition also found in the literature is given by

 
Rx x = E{xx† } and Rx y = E{xy† } (3.36)

The definition of (3.36) yields a matrix that is the complex conjugate (or the transpose) of
the definition given by (3.35). So long as one adheres consistently to either one definition
or the other, the results obtained (in terms of the selected performance measure) will turn
out to be the same; therefore, it is immaterial which definition is used. With either of the
aforementioned definitions, it immediately follows that

R†x x = Rx x and R†x y = R yx (3.37)

So the autocorrelation matrix Rx x is Hermitian, whereas the cross-correlation matrix Rx y

is in general not Hermitian (since R†yx = Rx y ). Whether the autocorrelation matrix is
positive definite or positive semidefinite once again is determined by whether the signal
vector of concern is x(t), s(t), or n(t).

3.2 SIGNAL DESCRIPTIONS FOR CONVENTIONAL

AND SIGNAL ALIGNED ARRAYS
Two types of arrays have proven useful in different applications: (1) the conventional array
illustrated in Figure 3-1; and (2) the signal aligned array illustrated in Figure 3-2. The signal
aligned array is useful where the direction of arrival of the desired signal is known a priori,

FIGURE 3-1 x1
Conventional w1
narrowband array.

x2
w2 Σ y(t) = wTx (t)

xN
wN

FIGURE 3-2 x1 z1
Signal aligned f1 w1
Time
narrowband array. delay or
phase
shift
elements x2 z2
f2 w2 Σ y(t) = wTz(t)
zi (t) = e jf i xi (t)

xN zN
fN wN
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 89

3.2 Signal Descriptions for Conventional and Signal Aligned Arrays 89

and this information is used to obtain time-coincident desired signals in each channel.
One advantage of the signal aligned array structure is that a set of weights is found that
is independent of the desired signal time structure, which provides a distortionless output
for any waveform incident on the array from the (assumed) known desired signal direction
[22]. Such a processor is useful for extracting impulse “bursts,” which are present only
during relatively short time intervals.
The outputs of the arrays illustrated in Figures 3-1 and 3-2 are expressed, respectively,
as

y(t) = w† x(t) and y(t) = w† z(t) (3.38)

where x(t) = s(t) + n(t) is the vector of received signals that are complex valued func-
tions. The noise vector n(t) may be assumed to be stationary and ergodic and to have
both directional and thermal noise components that are independent of the signal. The
signal vector s(t) induced at the sensor elements from a single directional signal source is
assumed to be
√
s(t) = S e jω0 t (3.39)

where ω0 is the (radian) carrier frequency, and S represents the signal power. Assuming
identical antenna elements, the resulting signal component in each array element is just a
phase-shifted version (due to propagation along the array) of the signal appearing at the
first array element encountered by the directional signal source. It therefore follows that
the signal vector s(t) is written as
√ jω0 t √ jω0 t+θ1 √ 
sT (t) = Se , Se , . . . , S e jω0 t+θ N −1
= s(t)vT (3.40)

where v is defined to be the array propagation vector

vT = [1, e jθ1 , . . . , e jθ N −1 ] (3.41)

Consequently, the received signal vector for the conventional array of Figure 3-1 is written
as

x(t) = s(t)v + n(t) (3.42)

The received signal vector (after the time-delay or phase-shift elements) for the signal
aligned array of Figure 3-2 is written as

z(t) = s(t)1 + n (t) (3.43)

where now v of (3.42) has been replaced by 1 = (1, 1, . . . , 1)T since the desired signal
terms in each channel are time aligned and therefore identical. The components of the
noise vector then become

n i (t) = n i (t)e jφi (3.44)

In developing the optimal solutions for selected performance measures, four corre-
lation matrices will be required. These correlation matrices are defined as follows for
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 90

90 CHAPTER 3 Optimum Array Processing

narrowband uncorrelated signal processes:


Rss = E {s ∗ (t)sT (t)} = Sv ∗ vT (3.45)

where S denotes the signal power


Rnn = E {n ∗ (t)nT (t)} (3.46)
 ∗ ∗
rxs = E {x (t)s(t)} = Sv (3.47)

and

Rx x = E{x ∗ (t)xT (t)} = Rss + Rnn (3.48)

3.3 OPTIMUM ARRAY PROCESSING FOR

NARROWBAND APPLICATIONS
As previously noted, several different performance measures are adopted to govern the
operation of the adaptive processor that adjusts the weighting for each of the sensor element
outputs of Figure 1-1. Now consider the problem of formulating four popular performance
measures in terms of complex envelope signal characterizations and of determining the
optimum steady-state solution for the adaptive weight vector.
The following performance measures are considered:
1. Mean square error (MSE) criterion
2. Signal-to-noise ratio criterion
3. Maximum likelihood (ML) criterion
4. Minimum noise variance (MV) criterion
These criteria form the basis for adaptive algorithms for narrowband systems. Narrowband
adaptive arrays have complex weights in each channel. For wideband signals, however,
the adaptive array has a set of linear filters that are usually approximated by tapped delay
lines in each channel of the array. Thus, consideration of the narrowband processing case
is slightly easier, although conceptually the wideband processor is regarded as a set of
frequency-dependent complex weights.
The optimum solutions for the complex weight vector for each of the aforementioned
four performance measures are derived first. Since it is desirable to be equally familiar
with real and complex notation, the derivations are carried out using real notation; the
corresponding complex solutions are then given so the student may develop the complex
results using the derivations for real quantities but employing complex notation. Following
the derivation of the desired results for each performance measure, it is shown that each
solution is closely related to a single optimal solution by means of factorization of the
results: this factorization provides the connection between the various solutions obtained
and the form known as the Wiener solution.

3.3.1 The MSE Performance Measure

The MSE performance measure was considered by Widrow et al. [5] for the conventional
array configuration, and additional procedures based on this criterion have been developed
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 91

3.3 Optimum Array Processing for Narrowband Applications 91

and further extended [23–25]. Suppose the desired directional signal s(t) is known and
represented by a reference signal d(t). This assumption is never strictly met in practice
because a communication signal cannot possibly be known a priori if it is to convey in-
formation; hence, the desired signal must be unknown in some respect. Nevertheless, it
turns out that in practice enough is usually known about the desired signal that a suitable
reference signal d(t) is obtained to approximate s(t) in some sense by appropriately pro-
cessing the array output signal. For example, when s(t) is an amplitude modulated signal,
it is possible to use the carrier component of s(t) for d(t) and still obtain suitable operation.
Consequently, the desired or “reference” signal concept is a valuable tool, and one can
proceed with the analysis as though the adaptive processor had a complete desired signal
characterization.
The difference between the desired array response and the actual array output signal
defines an error signal as shown in Figure 3-3:

ε(t) = d(t) − wT x(t) (3.49)

The squared error can therefore be written as

ε 2 (t) = d 2 (t) − 2d(t)wT x(t) + wT x(t)xT (t)w (3.50)

Taking expected values of both sides of (3.50) then yields

E{ε 2 (t)} = d 2 (t) − 2wT rxd + wT Rx x w (3.51)

where
⎡ ⎤
x1 (t)d(t)
⎢ x 2 (t)d(t) ⎥
⎢ ⎥
rxd =⎢ .. ⎥ (3.52)
⎣ . ⎦
x N (t)d(t)

x1 (t)
FIGURE 3-3
w1 Basic adaptive array
structure with known
desired signal.

xi (t)
wi Σ Output y(t)

xN (t)
wN
−
Error signal e (t)
+

d (t)
Reference signal
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 92

92 CHAPTER 3 Optimum Array Processing

Since d(t) = s(t), from (3.39) it follows that d 2 (t) = S so that

E{ε2 (t)} = S − 2wT rxd + wT Rx x w (3.53)

Since (3.53) is a quadratic function of w, its extremum is a minimum. Therefore, the value
of w that minimizes of E{ε 2 (t)} is found by setting the gradient of (3.53) with respect to
the weight vector equal to zero, that is,

∇w (ε 2 ) = 0 (3.54)

Since

∇w (ε 2 ) = −2rxd + 2Rx x w (3.55)

it follows that the optimum choice for the weight vector must satisfy

Rx x wopt = rxd or wopt = R−1

x x rxd (3.56)

Equation (3.56) is the Wiener-Hopf equation in matrix form, and its solution, wopt , is
consequently referred to as the optimum Wiener solution.
If we use d(t) = s(t) and (3.39) and (3.42), it then follows that

rxd = E{xd} = Sv (3.57)

so that

wMSE = SR−1
xx v (3.58)

where it is assumed that Rx x is nonsingular so that R−1

x x exists. Setting the weight vector
equal to wMSE , the resulting minimum MSE is found from (3.53) to be

εmin
2
= S − rTxd R−1
x x rxd (3.59)

For complex quantities, (3.53), (3.58), and (3.59) become

|ε(t)|2 = S − 2Re{w† rxd } + w† Rx x w (3.60)

w MSE = SR−1xx v
∗
(3.61)
|ε|2min = S − r†xd R−1
x x rxd (3.62)

3.3.2 The SNR Performance Measure

Adaptive algorithms based on maximizing the SNR have been developed for the conven-
tional array configuration in communication and detection systems [26–28]. The output
signal from the adaptive array of Figure 3-1 is written as

y(t) = wT x(t) (3.63)

where the input signal vector may be regarded as composed of a signal component s(t)
and a noise component n(t) so that

x(t) = s(t) + n(t) (3.64)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 93

3.3 Optimum Array Processing for Narrowband Applications 93

The signal and noise components of the array output signal may therefore be written as

ys (t) = wT s(t) = sT (t)w (3.65)

and

yn (t) = wT n(t) = nT (t)w (3.66)

where
⎡ ⎤ ⎡ ⎤
s1 (t) n 1 (t)
⎢ s2 (t) ⎥ ⎢ n 2 (t) ⎥
⎢ ⎥ ⎢ ⎥
s(t) = ⎢ . ⎥ and n(t) = ⎢ . ⎥ (3.67)
⎣ . ⎦. ⎣ . ⎦.
s N (t) n N (t)
Consequently, the output signal power may be written as

E{|ys (t)|2 } = |wT s|2 (3.68)

and the output noise power is

E{|yn (t)|2 } = |wT n|2 (3.69)

The output SNR is therefore given by

 
  wT ssT w
s |wT s|2 wT Rss w
= =   = (3.70)
n |wT n|2 wT nnT w wT Rnn w

The ratio given by (3.70) is rewritten as

 
s zT R−1/2 −1/2
nn Rss Rnn z
= (3.71)
n zT z
where

z = R1/2
nn w (3.72)

Equation (3.70) may be recognized as a standard quadratic form and is bounded by the
minimum and maximum eigenvalues of the symmetric matrix R−1/2 nn Rss Rnn
−1/2
(or, more
−1
conveniently, Rnn Rss ) [29]. The optimization of (3.70) by appropriately selecting the
weight vector w consequently results in an eigenvalue problem where the ratio (s/n) must
satisfy the relationship [30]
 
s
Rss w = Rnn w (3.73)
n
in which (s/n) now represents an eigenvalue of the symmetric matrix noted already. The
maximum eigenvalue satisfying (3.73) is denoted by (s/n)opt . Corresponding to (s/n)opt ,
a unique eigenvector wopt represents the optimum element weights. Therefore
 
s
Rss wopt = Rnn wopt (3.74)
n opt
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 94

94 CHAPTER 3 Optimum Array Processing

Substitution of (3.70) corresponding to (s/n)opt into (3.74) yields

T
wopt Rss wopt
Rss wopt = T
Rnn wopt (3.75)
wopt Rnn wopt

Substituting Rss = [ssT ] and noting that sT wopt is a scalar quantity occurring on both
sides of (3.75) that may be canceled, we get the result
T
wopt s
s= T
· Rnn wopt (3.76)
wopt Rnn wopt
 T T

The ratio wopt s/wopt Rnn wopt may be seen as just a complex (scalar) number, denoted
here by c. It therefore follows that
 
1
wopt = R−1
nn s (3.77)
c
√
Since from (3.39) the envelope of s is just S v, it follows that

wSNR = αR−1
nn v (3.78)
where
√
S
α=
c
The maximum possible value that (s/n)opt is derived by converting the original system
into orthonormal system variables. Since Rnn is a positive definite Hermitian matrix,
it is diagonalized by a nonsingular coordinate transformation as shown in Figure 3-4.
Such a transformation is selected so that all element channels have equal noise power
components that are uncorrelated. Denote the transformation matrix that accomplishes
this diagonalization by A so that

s = As (3.79)
and
n = An (3.80)

where a prime ( ) denotes quantities after the transformation.

FIGURE 3-4
s1, n1 s ′1, n ′1
Functional w ′1
representation of
orthonormal
transformation and
adaptive weight s2, n2 Matrix s ′2, n ′2 Array
combiner equivalent transformation w ′2 Σ Output
A
to system of
Figure 3-1.

sN, nN s ′N, n ′N
w ′N
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 95

3.3 Optimum Array Processing for Narrowband Applications 95

The signal component of the array output now becomes

ys = wT s = wT As (3.81)

and the noise component becomes

yn = wT n = wT An (3.82)

For the array output of the system in Figure 3-4 to be equivalent to the output of the
system in Figure 3-1, it is necessary that

w = AT w (3.83)

The output noise power of the orthonormal system is given by

E{|yn (t)|2 } = E{|wT n |2 = wT E{n nT }w (3.84)

Since the transformation matrix A decorrelates the various noise components and equalizes
their powers, the covariance matrix of the noise process n (t) is just the identity matrix,
that is,

E{n nT } = I N (3.85)

It immediately follows from (3.84) and (3.85) that

E{|yn (t)|2 } = w w = ||w ||2

T
(3.86)

The output noise power of the original system of Figure 3-1 is given by

E{|yn (t)|2 } = wT Rnn w (3.87)

Substituting (3.83) into (3.87), it follows that

E{|yn (t)|2 } = w ARnn AT w

T
(3.88)

For the output noise power of the two systems to be equivalent, it is necessary that

ARnn AT = I N (3.89)

or

Rnn = [AT A]−1 (3.90)

Equation (3.90) simply expresses the fact that the transformation A diagonalizes and
normalizes the matrix Rnn .
The signal output of the orthonormal array system is given by (3.81). Applying the
Cauchy-Schwartz inequality (see Appendix D and [4]) to this expression immediately
yields an upper bound on the array output signal power as

|yS (t)|2 ≤ ||w ||2 ||s ||2 (3.91)

where

||s ||2 = sT s and ||w ||2 = wT w (3.92)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 96

96 CHAPTER 3 Optimum Array Processing

From (3.86) and (3.91) it follows that the maximum possible value of the SNR is given by

SNRmax = ||s ||2 (3.93)

Substituting (3.79) into (3.83) and using (3.93) and (3.90), we find it then follows that

SNRopt = sT R−1
nn s (3.94)

For complex quantities, (3.70) becomes

 
  w† s∗ s T w
s w† Rss w
=   = (3.95)
n w† n∗ nT w w† Rnn w

Equation (3.78) is replaced by

wSNR = αR−1
nn v
∗
(3.96)

(3.90) is now

Rnn = [AT A∗ ]−1 (3.97)

whereas (3.94) becomes

SNRopt = sT R−1
nn s
∗
(3.98)

Designing the adaptive processor so that the weights satisfy Rnn w = αv∗ means that
the output SNR is the governing performance criterion, even for the quiescent environment
(when no jamming signal and no desired signal are present). It is usually desirable, however,
to compromise the output SNR to exercise some control over the array beam pattern
sidelobe levels. An alternative performance measure that yields more flexibility in beam
shaping is introduced in the manner described in the following material.
Suppose in the normal quiescent signal environment that the most desirable array
weight vector selection is given by wq (where now wq represents the designer’s most
desirable compromise among, for example, gain or sidelobe levels). For this quiescent
environment the signal covariance matrix is Rnnq . Define the column vector t by

Rnnq wq = αt∗ (3.99)

On comparing (3.99) with (3.96), we see that the ratio being maximized is no longer
(3.95) but is the modified ratio given by

|w† t|2
(3.100)
w† Rnn w
This ratio is a more general criterion than the SNR and is often used for practical appli-
cations. The vector t is referred to as a generalized signal vector, and the ratio (3.100) is
called the generalized signal-to-noise ratio (GSNR). For tutorial purposes, the ordinary
SNR is usually employed, although the GSNR is often used in practice.
The adaptive array of Figure 3-1 is a generalization of a coherent sidelobe canceller
(CSLC). As an illustration of the application of the foregoing SNR performance measure
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 97

3.3 Optimum Array Processing for Narrowband Applications 97

Desired Jamming signals FIGURE 3-5

signal Coherent sidelobe
cancellation (CSLC)
system.

Main Auxiliary channels

0
channel

1 2 N
y1 (t) y2(t) yN (t)

w0 w1 w2 wn

x(t)
Σ

+
−

Array output (residue), r (t) Δ= x(t) − w† y(t)

Residue power, Pr Δ= E {r(t)r* (t)}
zT Δ= E {x(t)y†(t)}
Px Δ= E {x(t)x* (t)}

concepts, it is useful to show how sidelobe cancellation may be regarded as a special case
of an adaptive array.
The functional diagram of a standard sidelobe cancellation system is shown in
Figure 3-5. A sidelobe canceller consists of a main antenna with high gain designated
as channel o and N auxiliary antenna elements with their associated channels. The auxil-
iary antennas have a gain approximately equal to the average sidelobe level of the main
antenna gain pattern [18,31]. If the gain of the auxiliary antenna is greater than the highest
sidelobe, then the weights in the auxiliary channel are less than one. A properly designed
auxiliary antenna provides replicas of the jamming signals appearing in the sidelobes
of the main antenna pattern. These replica jamming signals provide coherent cancel-
lation in the main channel output signal, thereby providing an interference-free array
output response from the sidelobes. Jamming signals entering the main beam cannot be
canceled.
In applying the GSNR performance measure, it is first necessary to select a t column
vector. Since any desired signal component contributed by the auxiliary channels to the
total array desired signal output is negligible compared with the main channel contribution,
and the main antenna has a carefully designed pattern, a reasonable choice for t is the N + 1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 98

98 CHAPTER 3 Optimum Array Processing

component vector
⎡ ⎤
1
⎢0⎥
⎢ ⎥
⎢ ⎥
t = ⎢0⎥ (3.101)
⎢.⎥
⎣ .. ⎦
0

which preserves the main channel response signal. From (3.99) the optimum weight vector
must satisfy

Rnn w = αt (3.102)

where Rnn is the (N + 1) × (N + 1) covariance matrix of all channel input signals (in the
absence of the desired signal), and w is the N + 1 column vector of all channel weights.
Let the N × N covariance matrix of the auxiliary channel input signals (again in the
absence of the desired signal) be denoted by Rnn , and let w be the N-component column
vector of the auxiliary channel weights. Equation (3.102) may now be partitioned to yield
    
†
P0 w0 α
= (3.103)
Rnn w 0

where P0 = noise power output of the main channel and

⎡ ∗ ⎤
x1 x0
⎢ x 2∗ x 0 ⎥ cross-correlation vector between the
 ⎢ ⎥
 = E ⎢ . ⎥ = main channel output and the output of
⎣ .. ⎦
each auxiliary channel
x N∗ x0

Equation (3.103) in turn may be written as two separate equations

(scalar equation) P0 w 0 +  † w = α (3.104)

(matrix equation) Rnn w = −w 0  (3.105)

The solution represented by (3.102) is implemented using the N + 1 weights w 0 ,

w 1 , . . . , w N shown in Figure 3-5. It is also possible to attain the same optimum SNR
using only the N weights in the auxiliary channels, however. To see this, note that, if w is
optimum for a given noise environment (a given Rnn ), then any scalar multiple of w will
also yield the optimum SNR since the output SNR does not depend on the absolute level of
w . Consequently, the scalar weight w 0 in (3.104) is fixed at any convenient nonzero value.
Note that since (w )T t = w 0 , the GSNR represented by (3.100) will never be optimized
by w 0 = 0. Only the weights in the auxiliary channels need adjusted when w 0 is fixed at
some convenient value ŵ 0 , so that

Rnn wopt = −ŵ 0  (3.106)

Since ŵ 0 is fixed and (3.106) optimizes the GSNR ratio of ŵ 0 to the output noise power, it
follows that wopt must be minimizing the output noise power resulting from the sidelobe
response.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 99

3.3 Optimum Array Processing for Narrowband Applications 99

3.3.3 The ML Performance Measure

When the desired signal waveform is completely unknown (e.g., with seismic waves),
then the desired signal is a time function that must be estimated. The derivation of the
maximum likelihood estimator of the desired signal requires the assumption that the noise
components have a multidimensional Gaussian distribution [22,32].
The input signal vector may once again be written as

x(t) = s(t) + n(t) (3.107)

where

s(t) = s(t)v (3.108)

for the conventional array of Figure 3-1, and an estimate of s(t) is desired. Define the
likelihood function of the input signal vector as

 [x(t)] = − ln[P{x(t)/x(t) = s(t) + n(t)}] (3.109)

where P{z/y} is the probability density function for z given the event y. Thus, the likelihood
function defined by (3.109) is the negative natural logarithm of the probability density
function for the input signal vector x(t) given that x(t) contains both the desired signal
and interfering noise.
Now assume that the noise vector n(t) is a stationary, zero-mean Gaussian random
vector with a covariance matrix Rnn . Furthermore, assume that x(t) is also a stationary
Gaussian random vector having the mean s(t)v, where s(t) is a deterministic but unknown
quantity. With these assumptions, the likelihood function is written as

 [x(t)] = c[x(t) − s(t)v]T R−1

nn [x(t) − s(t)v] (3.110)

where c is a scalar constant independent of x(t) and s(t).

The maximum likelihood processor is obtained by solving for the estimate of s(t),
denoted by ŝ(t), which maximizes (3.110). Taking the partial derivative of [x(t)] with
respect to s(t) and setting the result equal to zero yields
∂[x(t)]
0= = −2vT R−1 T −1
nn x + 2ŝ(t)v Rnn v (3.111)
∂s(t)
It immediately follows that the estimate ŝ(t) that maximizes [x(t)] is given by

ŝ(t)vT R−1 T −1
nn v = v Rnn x (3.112)

Since the quantity vT R−1

nn v is a scalar, (3.112) is rewritten as

vT R−1
ŝ(t) = nn
x(t) (3.113)
vT R−1
nn v

which is of the form ŝ(t) = wML

T
x(t). Consequently, the maximum likelihood weight
vector is given by
R−1
nn v
wML = (3.114)
vT R−1
nn v
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 100

100 CHAPTER 3 Optimum Array Processing

For complex quantities (3.110) becomes

 [x(t)] = c[x(t) − s(t)v]† R−1

nn [x(t) − s(t)v] (3.115)

Likewise (3.112) is replaced by

ŝ(t)v† R−1 † −1
nn v = v Rnn x(t) (3.116)

and the result expressed by (3.114) then becomes

R−1
nn v
wML = (3.117)
v R−1
†
nn v

3.3.4 The MV Performance Measure

When both the desired signal s(t) and the desired signal direction are known (as with
a signal aligned array), then minimizing the output noise variance provides a means of
ensuring good signal reception, and methods based on this performance measure have
been developed [33,34]. For the signal aligned array of Figure 3-2, the array output is
given by

N 
N
y(t) = wT z(t) = s(t) wi + w i n i (3.118)
i=1 i=1

where the n i represent the noise components after the signal aligning phase shifts. When
we constrain the sum of the array weights to be unity, then the output signal becomes

y(t) = s(t) + wT n (t) (3.119)

which represents an unbiased output signal since

E{y(t)} = s(t) (3.120)

The variance of the array output may therefore be expressed as

 
var[y(t)] = E wT n (t)n (t)w = wT Rn  n  w
T
(3.121)

The relationship between n(t), the noise vector appearing before the signal aligning phase
shifts, and n (t) is given by

n (t) = n(t) (3.122)

where  is the diagonal unitary transformation described by

⎡ ⎤
e jφ1 0
⎢ ejφ2 ⎥
⎢ ⎥
=⎢ .. ⎥ (3.123)
⎣ . ⎦
0 e jφ N
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 101

3.3 Optimum Array Processing for Narrowband Applications 101

The variance of the array output remains unaffected by such a unitary transformation so
that

var[y(t)] = wT Rn  n  w = wT Rnn w (3.124)

It is now desired to minimize (3.124) subject to the constraint

wT 1 = 1 (3.125)

where

1 = [1, 1, . . . , 1]T (3.126)

To solve this minimization problem, form the modified performance criterion

1 T
MV = w Rnn w + λ[1 − wT 1] (3.127)
2
where the factor λ is a Lagrange multiplier. Since MV is a quadratic function of w, it
follows that the optimal choice for w may be found by setting the gradient ∇w MV equal
to zero. The gradient is given by

∇w MV = Rnn w − λ1 (3.128)

so that

wMV = R−1
nn 1λ (3.129)

The optimum solution wMV satisfies the constraint condition so that

T
wMV 1=1 (3.130)

and on substituting (3.130) into (3.129) there results

1
λ= (3.131)
1T R−1
nn 1

It follows immediately that

R−1
nn 1
wMV = (3.132)
1 R−1
T
nn 1

where wMV satisfies (3.130).

On substituting (3.132) into (3.124), the minimum value of the output noise variance
is found to be
1
varmin [y(t)] = (3.133)
1T R−1
nn 1

If complex quantities are introduced, all the foregoing expressions remain unchanged,
except the definition of the covariance matrix Rnn must be appropriate for complex vectors.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 102

102 CHAPTER 3 Optimum Array Processing

3.3.5 Factorization of the Optimum Solutions

The solutions obtained in the preceding sections all applied to the conventional array, with
the single exception of the minimum variance weights, where it was necessary to use the
signal aligned array to define a well-posed problem. All of the solutions are closely related
to one another since (as will be shown) they differ only by a scalar gain factor. Hence, the
different solutions possess identical output SNRs. This relationship is shown by factoring
the various solutions into a linear matrix filter followed by a scalar processor as described
in the following paragraphs [35].
The optimum weight vector obtained for the minimum MSE performance measure is
written from (3.61) as

wMSE = R−1 ∗ ∗ T −1
x x Sv = [Sv v + Rnn ] Sv
∗
(3.134)

Applying the matrix identity (D.10) of Appendix D to (3.134) results in

 
2 −1 ∗ T −1
S R v v R
wMSE = SR−1 nn −
nn nn
v∗
1 + SvT R−1 nn v
∗
 
S
= R−1
nn v
∗
(3.135)
1 + SvT R−1
nn v ∗

From (3.135), the minimum MSE weights are the product of a matrix filter R−1 ∗
nn v (which
is also common to the other weight vector solutions) and a scalar factor. Since the MV
solution is applied only to a signal aligned array, for the other solutions to pertain to the
signal aligned array it is necessary to replace the v vector only wherever it occurs by the
1 vector.
Now consider the noise power N0 and the signal power S0 that appear at the output
of the linear matrix filter corresponding to w = R−1 ∗
nn v as follows:

N0 = w† Rnn w = vT R−1
nn v
∗
(3.136)

and

S0 = w† Rss w = S N02 (3.137)

The optimal weight vector solution given in Sections 3.3.1 and 3.3.3 for the MSE and the
ML ratio performance measures are now written in terms of the previous quantities as
1 S0
wMSE = · · R−1
nn v
∗
(3.138)
N0 N0 + S0
and
1
wML = · R−1
nn v
∗
(3.139)
N0
Likewise, for the signal aligned array (where v = 1) the ML weights reduce to the unbiased,
MV weights, that is,

R−1
nn 1
wML |v = 1 = = wMV (3.140)
1T R−1
nn 1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 103

3.4 Optimum Array Processing for Broadband Applications 103

Linear matrix filter FIGURE 3-6 The

Scalar processor
factored minimum
MSE processor
^ ^ ^ ^ showing relation-
S SNR S ML (S MV) S0 S MSE ships to the
−1 v * 1
Rnn N0 S0 + N0 maximum SNR, ML,
and MV processors.

The previous expressions show that the minimum MSE processor can be factored into a
linear matrix filter followed by a scalar processor that contains the estimates corresponding
to the other performance measures, as shown in Figure 3-6. The different optimum weight
vector previously derived be expressed by

w = βR−1
nn v
∗
(3.141)

where β is an appropriate scalar gain; hence, they all yield the same SNR, which can then
be expressed as
    ∗ T −1 ∗
s w† Rss w β 2 SvT R−1nn v v Rnn v −1 ∗
= † =   = SvT Rnn v (3.142)
n w Rnn w β 2 vT R−1
nn v ∗

For the case of a wideband processor, it can similarly be shown that the solutions to
various estimation and detection problems are factored into a linear matrix filter followed
by appropriate scalar processing. This development is undertaken in the next section.
The fact that the optimum weight vector solutions for an adaptive array using the
different performance criteria indicated in the preceding section are all given (to within
a constant factor) by the Wiener solution underscores the fundamental importance of the
Wiener-Hopf equation in establishing theoretical adaptive array steady-state performance
limits. These theoretical performance limits provide the designer with a standard for
determining how much any improvement in array implementation can result in enhanced
array steady-state performance, and they are a valuable tool for judging the merit of
alternate designs.

3.4 OPTIMUM ARRAY PROCESSING FOR

BROADBAND APPLICATIONS
An array processor for passive sensing devices must decide whether the random processes
observed at the array element outputs consist of a desired signal obscured by noise or
noise alone. Performance measures like maximum output SNR or minimum MS, are
not adequate for deciding whether a signal is present. It is well known from statistical
detection theory, however, that decisions based on the likelihood ratio test minimize the risk
associated with an incorrect decision. In addition, the likelihood ratio test yields decisions
that are optimum for a wide range of performance criteria [36–38]. This section establishes
the relationship between the likelihood ratio test and some other popular performance
measures for broadband signal applications.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 104

104 CHAPTER 3 Optimum Array Processing

Let the observation vector x consist of elements representing the outputs from each
of the array sensors xi (t), i = 1, 2, . . . , N . The likelihood ratio is then given by the ratio
of conditional probability density functions [39]
 p[x/signal present]
(x) = (3.143)
p[x/signal absent]
If (x) exceeds a certain threshold η then the signal is assumed present, whereas if (x) is
less than this threshold the signal is assumed absent. The ratio (3.143) therefore represents
the likelihood that the sample x was observed, given that the signal is present relative to the
likelihood that it was observed given that the signal is absent. Such an approach certainly
comes far closer to determining the “best” system for a given class of decisions, since
it assumes at the outset that the system makes such decisions and obtains the processor
design accordingly.
It is worthwhile to mention briefly some extensions of the likelihood ratio test repre-
sented by (3.143). In many practical cases, one or several signal parameters (e.g., spatial
location, phase, or signal energy) are uncertain. When uncertain signal parameters (denoted
by θ) are present, one intuitively appealing approach is to explicitly estimate θ (denoted
by θ̂) and use this estimate to form a classical generalized likelihood ratio (GLR) [40]

 p[x/θ̂, signal present]

G (x/θ̂ ) = (3.144)
p[x/signal absent]
Uncertain signal parameters can also be modeled as random variables with any prior
information about them summarized in the form of an a priori density function p(θ).
The likelihood ratio can then be written as the ratio of marginal density functions and is
referred to as the Bayes likelihood ratio [41]

 p[x/θ , signal present] p(θ )dθ
 B (x) = θ (3.145)
p[x/signal absent]
where θ ∈ θ.
It is well known that likelihood ratio tests are regarded as completely equivalent to
matched filter processing in which the output SNR is maximized at a certain observation
time t = T [41–43]. Another goal of this section is to determine the role that matched
filtering plays in optimum array processing.
An additional important concept in likelihood ratio tests is that of a “sufficient statistic”
[44]. A brute force approach to the tests represented by the ratios (3.143) to (3.145) is to
process the observation vector x to actually construct the ratio (x) and compare this ratio
with a threshold η to make a decision. In many cases, although the observation vector is
N-dimensional, the threshold decision occurs along a one-dimensional coordinate in the
N-dimensional space, so that choosing a sufficient statistic simply amounts to picking a
coordinate system in which one coordinate contains all the information required to make
a decision. If such a sufficient statistic is available and is denoted by the scalar l(x), then
it is not necessary to construct an unwieldy ratio (x), since only the relatively simple
quantity l(x) needs be considered in arriving at the required decision.
Detection performance is commonly measured in terms of the receiver operating
characteristic (ROC). The ROC is a figure displaying the probability of detection (given
by Pr[(x) > η/signal present]) versus the probability of false alarm (given by Pr[(x) >
η/signal absent]) with the SNR as an independent parameter [38,44]. Solutions for the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 105

3.4 Optimum Array Processing for Broadband Applications 105

optimum detection processor may also be obtained very simply by working instead with
the “detection index” d at a specified observation time T where

 E{y(T )/signal present} − E{y(T )/signal absent}

d= (3.146)
{var[y(T )/signal absent]}1/2
and y(T) is the processor output at t = T . By maximizing the detection index, the detection
performance is also optimized, since this corresponds to obtaining the greatest possible
expected normalized difference of the processor output between the two signal conditions
it is desired to discriminate.
The unification of a variety of problems occurring in estimation and detection theory
is achieved by considering an observation vector described by

x(t) = Ms(t) + n(t) (3.147)

where x = observation vector

n = additive random noise vector having zero mean and noise covariance matrix Rnn
s = signal vector (known, unknown, or random)
M = known transformation matrix
The treatment of (3.147) for different classes of problems given here generally follows the
development presented by Cox [10]. In most cases, results are obtained using arguments
only by analogy, which sacrifice the mathematical rigor of techniques that transform
continuous time functions into discrete functions and use arguments based on Fourier
series expansions of stationary random processes but indicate how important results come
about with very little effort. Equation (3.147) is interpreted as a complex equation, where
a complex random vector

z = ζ + jγ (3.148)

is required to have the following two properties, which remain invariant under any linear
transformation:
1. The real part ζ and the imaginary part γ are both real random vectors having the same
covariance matrix.
2. All components of ζ and γ satisfy

E{ζl γm } = −E{ζm γl } for all l and m (3.149)

The development that follows involves some useful matrix properties and generalizations
of the Schwartz inequality that are summarized for convenience in Appendix D. Gaussian
random vectors are important in the subsequent analysis, so some useful properties of both
real and complex Gaussian random vectors are given in Appendix E.

3.4.1 Estimation of a Random Signal

Assuming that s in (3.147) is random, then the problem is posed of estimating s given the
observation vector x. Let the mean value of s be given by

E{s} = u (3.150)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 106

106 CHAPTER 3 Optimum Array Processing

and the associated covariance matrix be given by

E{(s − u)(s − u)† } = Rss (3.151)

where u and Rss are both known. The best estimate of s, given x, for a quadratic cost
criterion, is just the conditional mean E{s/x}.

[Link] Gaussian Random Signal

When s and n are both Gaussian and independent, then x and s are jointly Gaussian, and
the conditional mean E{s/x} is obtained by considering the vector formed by combining
s and x into a single vector. From (3.147), it follows that
 !  
s u
E = (3.152)
x Mu
and
   
s Rss Rss M†
cov = (3.153)
x MRss MRss M† + Rnn
Applying (E.14) or (E.42) from Appendix E it follows immediately that
 −1
ŝ = E{s/x} = u + Rss M† MRss M† + Rnn (x − Mu) (3.154)

The associated covariance matrix of ŝ is obtained from (E.15) or (E.43) from Appendix E
so that
 −1
cov(ŝ) = Rss − Rss M† MRss M† + Rnn MRss (3.155)

Applying the matrix identities (D.10) and (D.11) of Appendix D to (3.154) and (3.155)
yields
 −1 † −1
ŝ = u + R−1 † −1
ss + M Rnn M M Rnn (x − Mu) (3.156)

or
 −1  
ŝ = I + Rss M† R−1
nn M u + Rss M† R−1
nn x (3.157)

and
 −1
cov(ŝ) = R−1 †
ss + M Rnn M (3.158)

or
 −1
cov(ŝ) = I + Rss M† R−1
nn M Rss (3.159)

Letting u = 0 yields the result for the interesting and practical zero-mean case, then (3.156)
yields
 −1 † −1
ŝ = R−1 † −1
ss + M Rnn M M Rnn x (3.160)

which will be used later.

Since the mean, the mode, and the maximum likelihood of a Gaussian density function
are equal, the best estimate of s corresponds to the maximum of the a posteriori density
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 107

3.4 Optimum Array Processing for Broadband Applications 107

function. In other words, the value of s is chosen that maximizes

p(x/s) p(s)
p(s/x) = (3.161)
p(x)
From (3.147) it is seen that p(x/s) is Gaussian with mean Ms and covariance matrix Rnn .
Also, p(s) is Gaussian with mean u and covariance matrix Rss . Consequently, from (3.161)
and (E.49) of Appendix E it follows that
 !
α 
p(s/x) = (const.) exp − (x − Ms)† R−1 nn (x − Ms) + (s − u) † −1
R ss (s − u) (3.162)
2
where α = 2 for a complex random vector, α = 1 for a real random vector, and “const.”
denotes a constant of proportionality. Choosing s to maximize (3.162) is equivalent to
choosing s to minimize the following part of the exponent of (3.162)

J = (x − Ms)† R−1 † −1
nn (x − Ms) + (s − u) Rss (s − u) (3.163)

Either maximizing (3.162) or minimizing (3.163) leads to (3.154) as the appropriate

equation for the estimate ŝ. Minimizing the exponent of (3.162) is sometimes used in
nonlinear estimation problems, where the quantity Ms is replaced by a nonlinear function
m(s) that has a conditional mean that is difficult to compute.

[Link] Non-Gaussian Random Signal

In the event that s and n are non-Gaussian random vectors, the conditional mean E{s/x}
remains the best estimate for a quadratic cost criterion. In many cases, convenient expres-
sions for computing the conditional mean do not exist, and only second-order statistics
are available. Sometimes the estimate given by (3.154) or its equivalent forms are used
even though such an estimate no longer represents the conditional mean of p(s/x). Another
approach often used in non-Gaussian problems is to find the linear estimate that minimizes
the MSE. A linear estimate is one that has the form

ŝ = a + Bx (3.164)

The values assigned to a and B are determined by minimizing the expression

e = trace{E[(s − ŝ)(s − ŝ)† ]} = E{(s − ŝ)† (s − ŝ)} (3.165)

On combining (3.147) and (3.164) it follows that

(s − ŝ) = [I − BM](s − u) + {[I − BM]u − a} − Bn (3.166)

Consequently,

E{(s − ŝ)(s − ŝ)† } = [I − BM]Rss [I − M† B† ] + BRnn B†

+ {[I − BM]u − a}{[I − BM]u − a}† (3.167)

By setting the gradient of (3.167) with respect to a equal to zero, the value of a that
minimizes (3.165) is given by

a = [I − BM]u (3.168)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 108

108 CHAPTER 3 Optimum Array Processing

After we substitute (3.168) into (3.167) and complete the square in the manner of (D.9)
of Appendix D, it follows that
 −1
E{(s − ŝ)(s − ŝ)† } = Rss − Rss M† MRss M† + Rnn MRss
  −1  
+ B − Rss M† MRss M† + Rnn MRss M† + Rnn
· {B† − [MRss M† + Rnn ]−1 MRss } (3.169)

The value of B that minimizes (3.165) may easily be found from (3.169)
 −1
B = Rss M† MRss M† + Rnn (3.170)

By substituting the results from (3.168) and (3.170) into (3.164), the expression for ŝ
corresponding to (3.154) is once again obtained. Likewise, by substituting (3.170) into
(3.169) the same expression for the error covariance matrix as appeared in (3.155) also
results. In the Gaussian case the estimate ŝ given by (3.154) is the conditional mean,
whereas in the non-Gaussian case the same estimate is the “best” linear estimate in the
sense that it minimizes the MSE.

[Link] Application of Random Signal Estimation Results to

Optimum Array Processing
The signal vector received by an array consisting of N sensors is expressed as
" t
x(t) = m(t − v)s(v)dv + v(t) (3.171)
−∞

where m(t) is a linear transformation that represents propagation effects and any signal
distortion occurring in the sensor. In the case of ideal (nondispersive) propagation and
distortion-free electronics, the elements of m(t) are time delays, δ(t − Ti ), whereas the
scalar function s(t) is the desired signal.
When the signal and noise processes are stationary and the observation interval is long
(t → ∞), then the convolution in (3.171) is circumvented by working in the frequency
domain using Fourier transform techniques. Taking the Fourier transform of (3.171) yields

(ω) = (ω)(ω) + (ω) (3.172)

where the     denote Fourier transformed variables. Note that (3.172)
is the same form as (3.147); however, it is a frequency-domain equation, whereas (3.147) is
a time-domain equation. The fact that (3.172) is a frequency-domain equation implies that
cross-spectral density matrices (which are the Fourier transforms of covariance matrices)
now fill the role that covariance matrices formerly played for (3.147).
Now apply the frequency-domain equivalent of (3.160) to obtain the solution
 −1 −1 †
ˆ
(ω) = φss (ω) + † (ω)−1
nn (ω)(ω)  (ω)−1
nn (ω)(ω) (3.173)

Note that the quantity φss−1 (ω) + † (ω)−1

nn (ω)(ω) is just a scalar so that (3.173) is
rewritten as

(ω) = | (ω)|2 † (ω)−1

nn (ω)(ω) (3.174)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 109

3.4 Optimum Array Processing for Broadband Applications 109

where

φss (ω)
| (ω)|2 = (3.175)
1 + φss (ω)† (ω)−1nn (ω)(ω)

and φss (ω) is simply the power of s(t) appearing at the frequency ω. In general, the
frequency response given by (3.174) is not realizable, and it is necessary to introduce
a time delay to obtain a good approximation to the corresponding time waveform ŝ(t).
Taking the inverse Fourier transform of (3.174) to obtain ŝ(t) then yields
" ∞
1
ŝ(t) = | (ω)|2 † (ω)−1
nn (ω)(ω)e
jωt
dω (3.176)
2π −∞

Note that (3.174) is rewritten as

ˆ
(ω) = | (ω)|2 (ω)(ω) (3.177)

where

(ω) = † (ω)−1
nn (ω) (3.178)

is a 1 × n row vector of individual filters. The single filter  j (ω) operates on the received
signal component  j (ω) so that it performs the operation of spatial prewhitening and then
matching to the known propagation and distortion effects imprinted on the structure of the
signal. The processor obtained for this estimation problem therefore employs the principle:
first prewhiten, then match. A block diagram of the processor corresponding to (3.174) is
shown in Figure 3-7.

3.4.2 Estimation of an Unknown, Nonrandom Signal

Now assume that s in (3.147) is a vector of unknown parameters rather than a vector
of random variables. The problem, once again, is estimating s given the observation
vector x. An estimate having desirable optimal properties for nonrandom parameters is
the maximum likelihood estimate that corresponds to the value of s that maximizes the
conditional density function p(x/s) for a particular observation x.

x1 (t ) FIGURE 3-7
η1* (w) Optimum array
processor for
x2(t ) estimation of a
η*2 (w) random signal.
Σ la (w) l2 s^ (t )

xN (t )
ηN* (w)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 110

110 CHAPTER 3 Optimum Array Processing

[Link] Gaussian Noise Case

When the noise vector n is Gaussian, the conditional density function takes the form
 !
1
p(x/s) = const. exp − α(x − Ms)† R−1
nn (x − Ms) (3.179)
2
Maximizing (3.179) corresponds to minimizing the exponent, and the estimate that mini-
mizes this exponent is easily shown to be given by
 −1 † −1
ŝ = M† R−1
nn M M Rnn x (3.180)

It is seen from (3.180) that ŝ is obtained from a linear transformation on x, so the distribution
of ŝ is also Gaussian. From (3.180) it immediately follows that since E{x} = Ms, then

E{ŝ} = s (3.181)

and
 −1
cov(ŝ) = M† R−1
nn M (3.182)

It is interesting to note that (3.180) is reached in the limit in (3.156) as u → 0 and

R−1
ss → 0. In other words, the ML estimate corresponds to the estimate obtained with an
a priori Gaussian distribution p(s) having zero mean and infinite variance.

[Link] Non-Gaussian Noise Case

When n is not Gaussian, the likelihood function will not have the simple form of (3.179);
thus, the ML estimate may not be so easily obtained. Furthermore, the available information
is limited to second-order statistics. Consequently, a weighted least square estimate is
popular to use in these circumstances, and it is obtained by minimizing

J = (x − Ms)† R−1
nn (x − Ms) (3.183)

The value of s that minimizes (3.183) is given by (3.180).

Yet another approach uses the minimum variance unbiased linear estimate. A linear
estimate has the form

ŝ = a + Bx (3.184)

and an unbiased estimate requires that

E{ŝ} = s (3.185)

If an estimate is to have minimum variance, then

trace{E [(ŝ − s)(ŝ − s)† ]} = E {(ŝ − s)† (ŝ − s)} (3.186)

must be minimized. Combining (3.147) and (3.184) results in

ŝ = a + BMs + Bn (3.187)

For (3.185) to hold, it is necessary that

a=0 (3.188)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 111

3.4 Optimum Array Processing for Broadband Applications 111

and

BM = I (3.189)

From (3.187) and (3.189), it follows that

E{ŝŝ† } = ss† + BRnn B† (3.190)

and (3.186) is minimized by choosing B to minimize the quantity trace (BRnn B† ) subject
to the constraint (3.189). This constrained minimization problem is solved by introducing
a matrix Lagrange multiplier λ and minimizing the quantity
 
J = trace BRnn B† + [I − BM]λ + λ† [I − M† B† ] (3.191)

On completing the square of (3.191) using the formula (D.9) in Appendix D, the previous
expression is rewritten as

J = trace λ + λ† − λ† M† R−1 Mλ
  nn † 
+ B − λ M Rnn Rnn B − R−1
† † −1
nn Mλ (3.192)

From (3.192), it follows that the minimizing value of B is given by

B = λ† M† R−1
nn (3.193)

To eliminate λ, use (3.189) so that

λ† M† R−1
nn M = I (3.194)

or
 −1
λ† = M† R−1
nn M (3.195)

Hence
 −1 † −1
B = M† R−1
nn M M Rnn (3.196)

Consequently, the estimate (3.184) once again reduces to (3.180), and the mean and
covariance of the estimate are given by (3.181) and (3.182), respectively.

[Link] Application of Unknown, Nonrandom Signal Estimation Results to

Optimum Array Processing
Again, using the frequency-domain equation (3.172) it follows by analogy with (3.147)
and (3.180) that the optimum frequency-domain estimator is given by
 −1 †
(ω) = † (ω)−1
nn (ω)(ω)  (ω)−1
nn (ω)(ω) (3.197)

Nothing that † (ω)−1

nn (ω)(ω) is a scalar, (3.197) is written as
 
1
(ω) = −1
+ 1 · | (ω)|2 (ω)(ω) (3.198)
φss (ω) (ω)nn (ω)(ω)
†

where | (ω)|2 and (ω) are given by (3.175) and (3.178), respectively. Therefore, the
only difference between the optimum array processor for random signal estimation and
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 112

112 CHAPTER 3 Optimum Array Processing

for unknown, nonrandom signal estimation is the presence of an additional scalar transfer
function. The characteristic prewhitening and matching operation represented by (ω) =
† (ω)−1
nn (ω) is still required.

3.4.3 Detection of a Known Signal

The binary detection problem (i.e., whether the signal is present) leads to a likelihood ratio
test in which the likelihood ratio given by (3.143) is compared with a threshold η.

[Link] Gaussian Noise Case

When the noise is Gaussian, then the likelihood ratio (3.143) is written as
 
exp − 12 α(x − Ms)† R−1
nn (x − Ms)
(x) =    (3.199)
exp − 12 α x† R−1
nn x

where α = 2 for complex x and α = 1 for real x. Clearly (3.199) is written in terms of a
single exponential function so that
 !
1  † † −1 † † −1 † −1

(x) = exp − α s M Rnn Ms − s M Rnn x − x Rnn Ms (3.200)
2

Since the term s† M† R−1nn Ms does not depend on any observation of x, a sufficient test
statistic for making a decision is the variable
1  † † −1   † † −1 
y= s M Rnn x + x† R−1
nn Ms = Re s M Rnn x (3.201)
2
The factor α is assumed to be incorporated into the threshold level setting. The distribution
of the sufficient statistic y is Gaussian, since it results from a linear operation on x, which
in turn is Gaussian both when the signal is present and when it is absent.
When the signal is absent, then

E{y} = 0 (3.202)

and from (E.7) and (E.41) of Appendix E, it follows that

s† M† R−1
nn Ms
var(y) = (3.203)
α
Likewise, when the signal is present, then

E{y} = s† M† R−1
nn Ms (3.204)

and the variance of y is the same as when the signal is absent.

[Link] Non-Gaussian Noise Case

In the event that the noise vector n is non-Gaussian and only second-order statistics are
available, a linear transformation of the observation vector x is commonly sought that is
of the form
1 †
y= [k x + x† k] = Re{k† x} (3.205)
2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 113

3.4 Optimum Array Processing for Broadband Applications 113

where the vector k is selected so the output SNR is maximized. The ratio given by
change in mean-squared output due to signal presence
r0 =
mean-squared output for noise alone
or equivalently

E{y 2 /signal present} − E{y 2 /signal absent}

r0 = (3.206)
E{y 2 /signal absent}
is referred to as the signal-to-noise ratio. Using (3.205) and (E.54) of Appendix E, the
ratio is written as
α[Re{k† Ms}]2 αk† Mss† M† k
r0 = ≤ (3.207)
k† Rnn k k† Rnn k
It is convenient to factor Rnn by introducing

Rnn = T† T (3.208)

in which case the upper bound for r0 given in (3.207) is written as

αk† Mss† M† k αk† T† (T† )−1 Mss† M† T−1 Tk

= (3.209)
k† Rnn k k† T† Tk
With the upper bound written as in (3.209), (D.16) of Appendix D is applied and the
resulting terms rearranged to yield

α[Re{k† Ms}]2
r0 = ≤ αs† M† R−1
nn Ms (3.210)
k† Rnn k
where equality results when

k† = s† M† R−1
nn (3.211)

On substituting k† of (3.211) into (3.205), it is found that the test statistic is once again
given by (3.201).
A different approach to the problem of detecting a known signal embedded in non-
Gaussian noise is to maximize the detection index given by (3.146). For y given by (3.205),
it is easily shown that
√
√ Re{k† Ms} αk† Mss† M† k
d= α √ † ≤ √ (3.212)
k Rnn k k† Rnn k
Now applying (D.17) of Appendix D to the upper bound in (3.212) in the same manner as
(D.16) was applied to the upper bound in (3.209), the detection index is shown to satisfy

√ Re{k† Ms} # † † −1
d= α √ † ≤ αs M Rnn Ms (3.213)
k Rnn k
Equality results in (3.213) when (3.211) is satisfied, so (3.201) once again results for the
test statistic.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 114

114 CHAPTER 3 Optimum Array Processing

FIGURE 3-8 x1 (t )
Linear processor 1* (w)
structure for known
signal detection
problem.
x2(t )
2* (w)

Σ y (T )
T

xN (t )
N* (w)

[Link] Application of Known Signal Detection Results

to Optimum Array Processing
When the received signal is known and expressed by (3.171), maximizing the detection
index at a specified observation time, T, is best done using the frequency domain in (3.172)
and applying an appropriate form of the Schwartz inequality to the detection index. With
the linear processor specified in Figure 3-8, then
" ∞
1
E {y(T )/signal present} = † (ω)(ω)(ω)e jωT dω (3.214)
2π −∞

and
" ∞
1
var[y(T )] = † (ω)nn (ω)(ω)dω (3.215)
2π −∞

Since E{y(T )/signal absent} = 0, the detection index becomes

√ ∞
1/ 2π −∞ † (ω)(ω)(ω)e jωT dω
d=  ∞ 1/2 (3.216)
−∞  (ω)nn (ω)(ω)dω
†

Now let nn (ω) = † (ω)(ω), so (3.216) is rewritten as

∞ †
√  (ω)† (ω)[† (ω)]−1 (ω)(ω)e jωT dω
d 2π = −∞  ∞ 1/2 (3.217)
−∞  (ω) (ω)(ω)(ω)dω
† †

The Schwartz inequality in the form of (D.14) of Appendix D may be applied to (3.217)
by letting † (ω)† (ω) play the role of f † and [† (ω)]−1 (ω)(ω)e jωT play the role of g.
It then follows that
√ " ∞ 1/2
d 2π ≤ ∗ (ω)† (ω)−1 nn (ω)(ω)(ω)dω (3.218)
−∞

where equality occurs if and only if

† (ω) = e− jωT  ∗ (ω)† (ω)−1

nn (ω) (3.219)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 115

3.4 Optimum Array Processing for Broadband Applications 115

Once again, the usual spatial prewhitening and matching operation represented by
† (ω)−1
nn (ω) appears in the optimum processor.

3.4.4 Detection of a Random Signal

When s is a random signal vector it will again be assumed that the mean value u and the
covariance matrix Rss are both known. The optimum processor for the binary detection
problem again leads to a likelihood ratio test.

[Link] Gaussian Case

When the vectors s and n are both Gaussian, the likelihood ratio (3.143) is written as
  −1 
exp − (α/2)(x − Mu)† MRss M† + Rnn (x − Mu)
(x) = const.   (3.220)
exp − (α/2)x† R−1
nn x

where “const.” represents a constant of proportionality. Expanding (3.220) by carrying

out indicated multiplications and taking logarithms yields the following sufficient test
statistic:
1   −1 
y= (x − Mu)† R−1 †
nn − MRss M + Rnn (x − Mu)
2
1 1 † −1
+ u† M† R−1
nn x + x Rnn Mu (3.221)
2 2
where the factor α has again been incorporated into the threshold level setting.
Thinking of the known mean value u as the deterministic part of s and the deviation of
the observation from its mean x − Mu as the random part of the observation, we then see
that the test statistic (3.221) is composed of a linear term corresponding to the test statistic
in the known signal case [represented by (3.201)] and a quadratic term involving only the
random part of the observation. Since the known signal case was treated in Section 3.4.3,
only the random part of the problem will be considered here by assuming that the mean
has been subtracted out or, equivalently, by assuming that s has zero mean.
When u = 0, then (3.221) reduces to
1 †  −1  −1 
y= x Rnn − MRss M† + Rnn x (3.222)
2
Applying the matrix identities (D.11) and (D.12) from Appendix D transforms (3.222) as
follows:
1 † −1  −1 −1 † −1
y= x Rnn M Rss + M† R−1
nn M M Rnn x (3.223)
2
or
1 † −1  −1
y= x Rnn MRss M† MRss M† + Rnn x (3.224)
2
A special case of some practical importance arises when s is only a scalar s and M is
a column vector m. Then (3.223) becomes
$ † −1 $2
$x R m$
2y = nn
(3.225)
φss + m† R−1
nn m
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 116

116 CHAPTER 3 Optimum Array Processing

For small signals

 −1
MRss M† + Rnn ≈ R−1
nn (3.226)

In this case, (3.224) becomes

1 † −1
y= x Rnn MRss M† R−1
nn x (3.227)
2
[Link] Gaussian Noise, Non-Gaussian Signal
Whereas the sufficient test statistic for the known signal detection problem involved only
a linear operation on the observation vector, the random signal detection problem of the
preceding section involved a test statistic that was quadratic in the observation vector.
Likewise, when s is non-Gaussian and only second-order statistics are available, then the
“best” quadratic processor is found (in the sense that the detection index is maximized).
When we assume s has zero mean, the test statistic is presumed to have the form

y = x† KK† x (3.228)

where K maximizes the detection index defined by (3.146). Note that since y is quadratic
in x and the variance of y in the denominator of (3.146) involves E{y 2 /signal absent}, then
fourth-order moments of the distribution for x are involved. By assuming the noise field
is Gaussian, the required fourth-order moments are expressed in terms of the covariance
matrix by applying (E.51) of Appendix E.
The numerator of (3.146) when y is given by (3.228) is written as

E{y/signal present} − E{y/signal absent} = trace[K† MRss M† K] (3.229)

By use of (E.51), the denominator of (3.146) becomes

&
% 2
var(y/signal absent) = trace[(K† Rnn K)2 ] · (3.230)
α
Consequently
√
(α/2) trace(K† MRss M† K)
d= % (3.231)
trace[(K† Rnn K)2 ]
It is now desired to maximize d by applying the Schwartz inequality to obtain an upper
bound. It is again convenient to introduce Rnn = T† T before applying (D.16) so that
√  
(α/2) trace (T† )−1 MRss M† T−1 TKK† T†
d= #
trace{(TKK† T† )2 }
#
 2 
≤ (α/2) trace MRss M† R−1 nn (3.232)

Equality obtains in (3.232) when

K† = A† M† R−1
nn (3.233)

where

Rss = AA† (3.234)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 117

3.4 Optimum Array Processing for Broadband Applications 117

Substituting (3.233) into (3.228) then yields the test statistic

y = x† R−1 † −1
nn MRss M Rnn x (3.235)

which is identical to the test statistic (3.227) obtained from the likelihood ratio in the small
signal case.

[Link] Application of Random Signal Detection Results

to Optimum Array Processing
Starting with the frequency-domain equation (3.172), (3.223) leads to a sufficient test
statistic is given by

1  †  † 
(ω) = | (ω)|2 † (ω)−1
nn (ω)(ω)  (ω)−1
nn (ω)(ω)
2
1$ $2
= $ (ω)† (ω)−1nn (ω)(ω)
$ (3.236)
2

where | (ω)|2 is given by (3.175). The structure of the optimum processor corresponding
to (3.176) is shown in Figure 3-9, where Parseval’s theorem was invoked to write (3.236)
in the time domain
" T
1
y(T ) = |z(t)|2 dt (3.237)
2T 0

If the small signal assumption is applicable, then using (3.227) instead of (3.223)
leads to

1 $ $2
(ω) = φss (ω)$† (ω)−1
nn (ω)(ω)
$ (3.238)
2

The result expressed by (3.228) is different from (3.236) only in that the φss (ω) has
replaced the scalar factor | (ω)|2 in (3.236).

x1 (t ) FIGURE 3-9
η1* (w) Optimum array
processor for
x2(t ) detection of a
η*2 (w) T random signal.
∫0 (
z (t) 1
Σ a (w) ( )2 T
) dt

Squarer
Averaging
xN (t ) circuit
ηN* (w)

−1 (w)
η(w) = m† (w) F nn

fss (w)
⎜a (w) ⎜2 =
−1 (w) m† (w)
1 + fss (w) m† (w) F nn
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 118

118 CHAPTER 3 Optimum Array Processing

3.4.5 Detection of an Unknown, Nonrandom Signal

When s is unknown but nonrandom, good results are obtained from the GLR test. The
procedure is to form the ratio

p(x/ŝ, signal present)

G (x) = (3.239)
p(x/signal absent)

where ŝ is selected to maximize the conditional density function p(x/s, signal present) and
is the maximum likelihood estimate of s.

[Link] Gaussian Noise Case

When n is Gaussian the likelihood ratio is given by
 
p(x/s, signal present) exp − (α/2)(x − Ms)† R−1
nn (x − Ms)
=   (3.240)
p(x/signal absent) exp − (α/2)x† R−1
nn x

Substituting the likelihood estimate of ŝ in (3.180) into (3.240) yields the generalized
likelihood ratio
 !
α † −1  † −1 −1 † −1
G (x) = exp x Rnn M M Rnn M M Rnn x (3.241)
2
The sufficient test statistic for (3.241) is obviously
1 † −1  † −1 −1 † −1
y= x Rnn M M Rnn M M Rnn x (3.242)
2
or
1 † † −1
y= ŝ M Rnn x (3.243)
2 1
where
 −1 † −1
ŝ1 = M† R−1
nn M M Rnn x (3.244)

Setting R−1
ss = 0 in (3.223) for the case of a random signal vector reduces (3.223) to
(3.242).

[Link] Application of Unknown, Nonrandom Signal Detection Results to

Optimum Array Processing
Again, using the frequency-domain equation (3.172), it follows from (3.242) that a suffi-
cient test statistic is given by
1 † −1  † †  † 
(ω) =  (ω)−1
nn (ω)(ω)  (ω)−1 −1
nn (ω)(ω) ·  (ω)nn (ω)(ω)
2
1 −1 $ † $
= † (ω)−1 $ (ω)−1 (ω)(ω)$2
nn (ω)(ω) nn (3.245)
2
The result expressed by (3.245) reveals the characteristic prewhitening and matching
operator † (ω)−1
nn (ω).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 119

3.4 Optimum Array Processing for Broadband Applications 119

3.4.6 Array Gain

The array gain is defined as the ratio of the output signal-to-noise spectral ratio to the input
signal-to-noise spectral ratio. The linear processor structure of Figure 3-8 consists of a set
of filters (one for each sensor), followed by a summation device for which the array gain
is given by [17]

† (ω)
˜ ss (ω)(ω)
G(ω) = (3.246)
† (ω)
˜ nn (ω)(ω)

where  ˜ ss (ω) and 

˜ nn (ω) represent the normalized cross-spectral density matrices of the
signal vector and noise vector, that is,

ss (ω) = σs2 (ω)

˜ ss (ω) (3.247)

and

nn (ω) = σn2 (ω)

˜ nn (ω) (3.248)

where σn2 (ω) is the noise power spectral density averaged over the N sensors so that
1
σn2 (ω) = trace[nn (ω)] (3.249)
N
and
1
σs2 (ω) = trace[ss (ω)] (3.250)
N
The array gain corresponding to (3.246) is the ratio of the output signal-to-noise spectral
ratio to the input signal-to-noise spectral ratio.
Whenever the signal vector s(t) is related to a scalar signal s(t) by a known transfor-
mation m(t) such that
" t
s(t) = m(t − τ )s(τ )dτ (3.251)
−∞

˜ ss (ω) is simply given by the dyad [17]

Then 


˜ ss (ω) = (ω)
˜ ˜ † (ω) (3.252)

where (ω)
˜ denotes the normalized Fourier transform of m(t) so that

˜ † (ω)(ω)
 ˜ =N (3.253)

When 
˜ ss (ω) is given by (3.252), then the array gain becomes

|† (ω)(ω)|
˜ 2
G(ω) = (3.254)
† (ω) ˜ nn (ω)(ω)

The quantity † (ω)(ω)

˜ may be regarded as the inner product of the beam steering vector
 and the signal direction 
˜ for plane wave propagation. Define a generalized angle γ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 120

120 CHAPTER 3 Optimum Array Processing

between  and 
˜ as described in Appendix F such that

|† (ω)(ω)|
˜ 2
cos2 (γ ) = (3.255)
(† (ω)(ω))(˜ † (ω)(ω))
˜
In a conventional beamformer the vector  is chosen to be proportional to ,
˜ thus making
γ equal to zero and “matching to the signal direction.” This operation also maximizes the
array gain against spatially white noise as shown subsequently.
1
Substituting (3.252) into (3.246) and using  = ˜ nn
2
 yields
$ † $2
˜ − 2 (ω)(ω)
1
$ (ω) ˜ $
G(ω) = nn
(3.256)
† (ω)(ω)
Applying the Schwartz inequality (D.14) to (3.256) then gives

˜ −1
˜ † (ω)
G(ω) ≤  nn (ω)(ω)
˜ (3.257)

where equality results when the vector † be a scalar multiple of  ˜ †˜ −1

nn . Therefore,
maximizing the array gain with no constraints yields the same prewhitening and match-
ing operation found in the preceding sections for a variety of detection and estimation
problems. Note also that this processor reduces to the conventional direction matching
beamformer when the noise field is uncorrelated from sensor to sensor so that  ˜ nn = I,
the identity matrix [45].
Maximizing the array gain is closely related to minimizing the array output signal
variance under a desired signal response constraint, because this is completely equivalent to
minimizing the denominator of (3.246) subject to a constraint on the numerator. Since G(ω)
is not changed by any scaling of the vector , introducing a constraint like † 
˜ = β does not
affect G(ω) and merely determines the scalar multiple used in selecting . Consequently,
selecting  = β  ˜ −1
nn /(
˜  ˜ −1
˜ † nn )
˜ maximizes G(ω) as well as satisfies the constraint
†
 ˜ = β. Maximizing the array gain and minimizing the output signal distortion yield
exactly the same filters for monochromatic signals [3]. The filter that yields the maximum
likelihood estimate and the Wiener filter are different by only a scalar transfer function
[46]. Furthermore, the likelihood ratio processor, the maximum SNR filter, and the Wiener
filter are known to be equivalent in the case of a narrowband signal that is known except
for phase and corrupted by additive, Gaussian noise [47].

3.4.7 Criterion Invariant Array Processor

Various detection and estimation techniques and performance measures related to optimum
array processing were treated in the preceding sections by means of a unified theory. The
likelihood ratio test for optimum signal detection, several signal estimation problems, and
various performance criteria are all related by a prewhitening and matching operation that
defines a vector operator for the received signal vector that has a scalar output. This scalar
output is a single waveform that is then processed by different scalar operators depending
on the problem of concern.
The results obtained for the various classes of problems may be conveniently summa-
rized in the criterion invariant processor shown in Figure 3-10. This figure illustrates
that the principle “first prewhiten, then match,” represented by the operator † −1 nn ,
is common for a wide variety of different optimization problems. Since the optimum
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 121

3.5 Optimum Array Processing for Perturbed Propagation Conditions 121

x1 (t ) *(w) e−jwT I
η1* (w)

x2(t ) T
∫0 (
η*2 (w) 1
Σ a (w) ( )2 T
) dt II

Averaging
xN (t ) filter
ηN* (w)

a*(w) III

Scalar
−1 (w) Wiener IV
η† (w) = m† (w) F nn
filter
fss (w) b (w) V
⎜a (w) ⎜2 =
−1 (w) m (w)
1 + fss (w) m† (w) F nn

1
b (w) = I – Detection, known signal
−1 (w) m (w)
m† (w) Fnn II – Detection, random signal, and unknown nonrandom signal
III – Estimation (maximum likelihood), random signal, and unknown
nonrandom signal
IV – Estimation (MMSE), random signal, and unknown nonrandom signal
V – Maximum array gain and minimum output variance

FIGURE 3-10 Criterion invariant array processor for broadband applications.

processor depends on the inverse of the noise cross-spectral density matrix, but in practice
the only measurable quantity is the cross-spectral matrix of the sensor outputs (which in
general contains desired signal-plus-noise components), the use of the signal-plus-noise
spectral matrix may result in performance degradation unless provision is made to obtain
a signal-free estimate of the noise cross-spectral matrix or the use of the signal-plus-noise
spectral matrix is specifically provided for. The consequences involved for providing an
inexact prewhitening and matching operation are discussed in reference [45]. It may be
further noted that the minimum mean square error (MMSE) signal estimate is different
from the maximum likelihood (distortionless) estimate (or any other estimate) only by a
scalar Wiener filter. A Wiener processor is therefore regarded as forming the undistorted
signal estimate for observational purposes before introducing the signal distortion result-
ing from the scalar Wiener filter. The nature of the scalar Wiener filter is further considered
in the Problems section.

3.5 OPTIMUM ARRAY PROCESSING FOR

PERTURBED PROPAGATION CONDITIONS
The processor structure of Figure 3-8 is optimum only when the signal cross-spectral
density matrix is a simple dyad [17]

ss (ω) = (ω)† (ω) (3.258)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 122

122 CHAPTER 3 Optimum Array Processing

Perturbations in the propagation process destroys the dyad nature of ss (ω), so the pro-
cessor structure must be more general than that of Figure 3-8. To determine the nature of
a more general processor structure, the optimum processor for a noncoherent narrowband
signal will be found.
Consider the problem of detecting a signal that is narrowband with unknown amplitude
and phase. Such conditions frequently occur when the received signal undergoes unknown
amplitude and phase changes during propagation. When the signal is present, the received
waveform is expressed as

x(t) = Re{ae jθ r (t)e jφ(t) e jω0 t } + n(t) (3.259)

where
a = unknown amplitude
θ = unknown phase
r (t) = known amplitude modulation
φ(t) = known phase modulation
ω0 = known carrier frequency
Using real notation we can rewrite (3.259) as

x(t) = m 1 (t)s1 + m 2 (t)s2 + n(t) (3.260)

where

s1 = a cos θ
s2 = −a sin θ
m 1 (t) = r (t) cos[ω0 t + φ(t)] = f (t) cos(ω0 t) − g(t) sin(ω0 t) (3.261)
m 2 (t) = r (t) sin[ω0 t + φ(t)] = f (t) sin(ω0 t) + g(t) cos(ω0 t) (3.262)

and

f (t) = r (t) cos[φ(t)] (3.263)

g(t) = r (t) sin[φ(t)] (3.264)

where
#
r (t) = f 2 (t) + g 2 (t)
 !
−1 g(t)
φ(t) = tan
f (t)
Equation (3.259) may now be rewritten as

x = ms + n (3.265)

where

s1
m = [m 1 m 2 ], s= , and var(n) = σn2
s2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 123

3.5 Optimum Array Processing for Perturbed Propagation Conditions 123

The results of Section 3.4.5 for the detection of an unknown, nonrandom signal may now
be applied, for which the sufficient test statistic is given by (3.242). For real variables
(3.242) becomes
1 T −1  T −1 −1 T −1
y= x Rnn M M Rnn M M Rnn x (3.266)
2
Note that m 1 (t) and m 2 (t) are orthogonal functions having equal energy so that |m 1 |2 =
|m 2 |2 , and m 1 (t) · m 2 (t) = 0. Consequently, when the functions are sampled in time to
form xT = [x(t1 ), x(t2 ), . . . ], miT = [m i (t1 ), m i (t2 ), . . . ], then (3.266) for the test statistic
is rewritten as
 
1   m1  −1
y = x2 m1T m2T |m1 |2 σn2
2 m2
1 2  −1  2  2 
= σn |m1 |2 m1T x + m2T x (3.267)
2
 −1
Since the scalar factor 12 σn2 |m1 |2 in (3.267) is incorporated into the threshold setting
for the likelihood ratio test, a suitable test statistic is
 2  2
z = m1T x + m2T x (3.268)

For a continuous time observation, the test statistic given by (3.268) is rewritten as
" T 2 " T 2
z= m 1 (t)x(t)dt + m 2 (t)x(t)dt (3.269)
0 0

The test statistic represented by (3.269) is conveniently expressed in terms of the “sine”
and “cosine” components, f (t) and g(t). Using (3.261) and (3.262), we then can rewrite
(3.269) as shown.
" T " T 2
z= f (t) cos(ω0 t)x(t)dt − g(t) sin(ω0 t)x(t)dt
0 0
" T " T 2
+ f (t) sin(ω0 t)x(t)dt + g(t) cos(ω0 t)x(t)dt (3.270)
0 0

The previous test statistic suggests the processor shown in Figure 3-11.
The optimum detector structure for a noncoherent signal shown in Figure 3-11 leads
to the more general array processor structure illustrated in Figure 3-12. This more general
processor structure is appropriate when propagating medium or receiving mechanism
perturbations cause the plane wave desired signal assumption to hold no longer or when it
is desired to match the processor to a signal of arbitrary covariance structure. A matched
array processor like that of Figure 3-12 using matrix weighting is referred to as an element
space-matched array processor [48].
Regarding the quantity Ms in (3.147) as the signal having arbitrary characteristics,
then it is appropriate to define an array gain in terms of the detection index at the output of
a general quadratic processor [17]. For Gaussian noise, the results summarized in (3.231)
may be used to give
 
trace † (ω)ss (ω)(ω)
G=  1/2 (3.271)
trace († (ω)nn (ω)(ω))2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 124

124 CHAPTER 3 Optimum Array Processing

FIGURE 3-11 cos w 0t

Quadrature matched f (T− t)
filter and envelope
detector for received
waveform having
unknown amplitude g (T− t ) Σ ( )2
and phase.
x(t ) Σ y (T )
T
g (T− t ) Σ ( )2

Quadratic
processor
f(T− t )
sin w 0t
Matched filter

FIGURE 3-12 x1 (t )
Structure for general ( )2
array processor. x2(t )

† (w) Σ y (t)

xN (t ) ( )2

Matrix Quadratic
filter processor

It may be seen that (3.271) reduces to (3.246) when  is a column vector. Under perturbed
propagation conditions, ss is given by [13]

ss (ω) = (ω)† (ω) (3.272)

where the matrix  has N rows and r columns, and r denotes the rank of the matrix ss .
For a plane wave signal, the cross-spectral density matrix ss has rank one, and the dyad
structure of (3.258) holds. The array gain given by (3.271) may be maximized in the same
manner as (3.246), now using (D.16) instead of (D.14), to yield [13]
' ( 2 )*1/2
G ≤ trace ss (ω)−1 nn (ω) (3.273)

where equality obtains when the matrix  is chosen to be a scalar multiple of † −1
nn . It
follows that the maximum gain cannot be achieved by any  having less than r columns.

3.6 POLARIZATION SENSITIVE ARRAYS

For radio applications (in contrast to sonar applications), if an array receives and uses more
than one polarization (by using polarization sensitive elements), then its performance is
far superior to one that does not [34]. When two signals arrive from the same direction,
it is possible to null one signal and not the other if their polarizations are different. How
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 125

3.6 Polarization Sensitive Arrays 125

“different” must two polarizations be to be handled differently by an antenna array?

This question concerns the separation between signal polarizations, which is treated by
consideration of the Poincare sphere.
Antennas are designed for maximum directivity and polarization match when the an-
tennas point at each other. Since the gain and polarization of an antenna change with angle,
moving the antennas in a communication system causes a directivity loss coupled with
a polarization mismatch, which reduces the received power. This polarization alignment
problem is the reason that circular polarization is preferred in mobile communications
systems.
To improve the link budget, the polarization and directivity of the receive or transmit
antennas are adaptively modified as the positions of the two antennas change. Adapting
the polarization requires an antenna, such as crossed dipoles, that can modify the major
and minor axes of its polarization ellipse of the transmitted wave. Some applications that
used adaptive crossed dipoles are given in [49–54].
Consider a four-element adaptive array consisting of two pairs of crossed dipoles as
shown in Figure 3-13. The polarization properties of a transverse electromagnetic (TEM)
wave are defined using the orthogonal electric fields shown in Figure 3-14, where that Eφ
represents the horizontal field component, and Eθ represents the vertical field component.
As time progresses, the wave propagates, and the Eφ and Eθ components traverse the
polarization ellipse. The major axis of the ellipse has a tilt given by β (where 0 ≤ β < π

z FIGURE 3-13
Crossed Dipole
Array Geometry.
L From Compton,
z=+
2 IEEE Trans. Ant. &
q Prop., Sept. 1981.

z =– L
2
x

Eq FIGURE 3-14
Polarization Ellipse.
From Compton,
IEEE Trans. Ant. &
a Prop., Sept. 1981

b
Ef
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 126

126 CHAPTER 3 Optimum Array Processing

to avoid ambiguity). The ellipticity angle, α, is then given by

α = tan−1 (r) (3.274)

where r is the axial ratio

minor axis
r= (3.275)
major axis
the angle α is positive when the electric field vector rotates clockwise, and negative when
it rotates counterclockwise (α is always in the range −π/4 ≤ α ≤ π/4 ). Figure 3-14 depicts
the situation where α is positive.
The electric field at any instant is given by the vector

E = Eφ φ + Eθ θ (3.276)

The electric field components for a given state of polarization are related to the polarization
ellipse and are given by (aside from a common phase factor)

Eφ = A cos γ (3.277a)
Eθ = A sin γ ejη (3.277b)

where “A” denotes amplitude, and γ and η are related to α and β by

cos(2γ ) = cos(2α) cos(2β) (3.278a)

tan(η) = tan(2α)csc(2β) (3.278b)

The previous equations relating the four variables α, β, γ , and η have a geometrical
relationship shown on the Poincare sphere in Figure 3-15. For a given point on the sphere,
M, the quantities 2γ , 2β, and 2α form the sides of a right spherical triangle. 2γ is the
side of the triangle between the point M and the point labeled H (H is the point on
the sphere representing horizontal polarization. The point V correspondingly represents
vertical polarization and lies 180◦ removed from H in the equatorial plane of the sphere).
The side 2β lies along the equator and extends to the point where it meets the perpendicular
projection of M onto the equator. The angle η then lies between the sides 2γ and 2β. The
special case α = 0 corresponds to linear polarization in which case the point M lies on the
equator: if in addition β = 0 then M lies at the point H, and only Eφ is nonzero so we have

FIGURE 3-15
Poincaré Sphere.
From Compton, M
IEEE Trans. Ant. &
V
Prop., Sept. 1981
2g
2a
h
H 2b
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 127

3.6 Polarization Sensitive Arrays 127

horizontal polarization; if, however, β = π/2 , then M lies at the point V, and we have vertical
polarization. The poles of the sphere correspond to circular polarization (α = ±45◦ ), with
clockwise circular polarization (α = +45◦ ) at the upper pole. We may conclude that five
parameters characterize the polarization of an electric field: (θ, φ, α, β, and A).
The electric field in Figure 3-16 has an Eφ component with an x-component of
−Eφ sin φ and a y-component of Eφ cos φ. The component of Eθ lying in the x-y plane
is Eθ cos θ, while the z-component is −Eθ sin θ . It immediately follows that the x and y
components of Eθ are given by Eθ cos θ cos φ and Eθ cos θ sin φ. The conversion of the
electric field from spherical to rectangular coordinates is given by

E = (Eθ cos θ cos φ − Eφ sin φ)x̂ + (Eθ cos θ sin φ + Eφ cos φ)ŷ
− (Eθ sin θ)ẑ (3.279)

Substituting (3.277) into (3.279) then yields

E = A[(sin γ cos θ cos φ e jη − cos γ sin φ)x̂

+ (sin γ cos θ cos φ e jη + cos γ cos φ)ŷ − (sin γ sin θ e jη )ẑ] (3.280)

A crossed dipole antenna has two or three orthogonal dipoles that can excite two or
three orthogonal linear polarizations. Adaptive crossed dipoles control the currents fed to
the dipoles to change the polarization and radiation pattern. The array geometry is shown in
Figure 3-16, where the top dipole consists of elements x1 (vertical) and x2 (horizontal along
the x-axis) and the bottom dipole consists of elements x3 (vertical) and x4 (horizontal along
the x-axis). We may then write the response of the four linear elements to the incoming
electric field of Figure 3-16 (now including time and space phase factors) in vector form as

X = Aej(ωt+ψ) U (3.281)

yr FIGURE 3-16
Diagram of an
adaptive crossed
fr dipole
xr
communications
system.

zt
Eq r
qt
qr

zr
Ef

yt

ft

xt
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 128

128 CHAPTER 3 Optimum Array Processing

where
⎡ ⎤
(− sin γ sin θ e jη )e j p
⎢ (sin γ cos θ cos φ e jη − cos γ sin φ)e j p ⎥
U=⎢
⎣
⎥
⎦ (3.282)
(− sin γ sin θ e jη )e− j p
(sin γ cos θ cos φ e jη − cos γ sin φ)e− j p
and ω is the carrier signal frequency, ψ is the carrier phase of the signal at the coordinate
origin when t = 0, and p is the phase shift of the signals at the two dipoles with respect to
the origin as a result of spatial delay where
πL
p= cos θ (3.283)
λ
Now suppose there is a desired signal specified by (Ad , θd , φd , αd , βd ) and an interfer-
ence signal by (Ai , θi , φi , αi , βi ). If thermal noise is present on each signal then the total
signal vector is given by

X = Xd + Xi + Xn (3.284a)
= Ad e j ( t+ϕd ) Ud + Ai e j ( t+ϕi ) Ui + Xn (3.284b)

where Ud , Ui are given by (3.282). The corresponding signal covariance matrix is then
given by
xx = d + i + n (3.285a)
where
 
d = E Xd XdH = A2d Ud UH d (3.285b)
 
i = E Xi XH
i = A 2
i U U
i i
H
(3.285c)

and
n = σ 2 I (3.285d)
A small difference in polarization between two received signals (measured in terms of
the angular separation between two points on the Poincare sphere) is enough for an adaptive
polarized array to provide substantial protection against the interference signal. Compton
[34] shows( that) the signal-to-noise plus interference ratio (SINR) is nearly proportional
to cos2 Md2Mi , where Md Mi represents the (radian) distance on the sphere between the
two polarizations.
The currents fed to the crossed dipoles in an adaptive communications system can be
adjusted to optimize the received signal [54]. Figure 3-16 shows the transmit antenna is
at an angle of (θr , ϕr ) from the receive antenna, and the receive antenna is at an angle of
(θt , ϕt ). Maximum power transfer occurs when θt = 0◦ and θr = 0◦ or when the main
beams point at each other. Controlling the complex signal weighting at the dipoles of the
transmit and receive antennas modifies the directivity and polarization of both antennas.
If the crossed dipole has three orthogonal dipoles, then the transmitted electric field is
written as
ωμe− jkr  
E = −j Ix L x cos θ cos ϕ + I y L y cos θ sin ϕ − Iz L z sin θ θ̂
 4πr  
+ − Ix L x sin ϕ + I y L y cos ϕ φ̂
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 129

3.6 Polarization Sensitive Arrays 129

where
r = distance from the origin to the field point at (x,y,z),
L x,y,z = dipole length in the x, y, and z directions
ω = radial frequency
k = wave number
μ = permeability
Ix,y,z = constant current in the x, y, or z direction
The directivity and polarization loss factors are given by

|E θ (θ, ϕ)|2 + |E ϕ (θ, ϕ)|2

D(θ, ϕ) = 4π  2π  π   (3.286)
0 0 |E θ (θ, ϕ)|2 + |E ϕ (θ, ϕ)|2 sin θ dθ dϕ
E θt E θr E ϕt E ϕr
PLF = # # +# # (3.287)
2
E θt + E ϕt
2 2
E θr + E ϕr
2 E θ2t + E ϕt
2 2
E θr + E ϕr
2

where 0 ≤ PLF ≤ 1 with PLF = 1 a perfect match. The t and r subscripts represent
transmit and receive, respectively.
When the transmitting antenna is a pair of orthogonal crossed dipoles in the x-y plane
that emits a circularly polarized field in the z-direction, then increasing θt causes the
transmitted electric field to transition from circular polarization through elliptical until
linear polarization results at θt = 90◦ . Assume the transmit antenna is a ground station
that tracks a satellite and has currents given by Ix = 1, I y = j, and Iz = 0. The transmit
antenna delivers a circularly polarized signal at maximum directivity to the moving receive
antenna. If the receive antenna has two crossed dipoles, then the maximum receive power
transfer occurs when the receive antenna is directly overhead of the transmit antenna.
If the receive antenna remains circularly polarized as it moves, then the power received
decreases, because the two antennas are no longer polarization matched. The loss in power
transfer comes from a reduction in the directivity and the PLF. If the currents at each dipole
are optimally weighted using an adaptive algorithm, then the power loss is compensated.
The curve in Figure 3-17 results from finding the optimum weights for a range of angles.
The link improvement over the nonadaptive antenna is as much as 3dB at θr = 90◦ .

4 FIGURE 3-17
Power loss as a
3 function of receive
Adapted dipoles angle when two
2 dipoles are adaptive.
Link loss (dB)

0
Unadapted dipoles
−1

−2

−3
0 20 40 60 80
q r (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 130

130 CHAPTER 3 Optimum Array Processing

FIGURE 3-18 4
Power loss as a
function of receive 3
Adapted dipoles
angle when three
dipoles are adaptive. 2

Link loss (dB)

1

0
Unadapted dipoles
−1

−2

−3
0 20 40 60 80
q r (degrees)

A third orthogonal dipole adds another degree of freedom to the receive antenna as
shown in Figure 3-18. Since a tri-dipole antenna is able to produce any θ or ϕ polarization,
the receive antenna adapts to maximize the received signal in any direction, and there is
no change in the link budget as a function of θr . The three orthogonal dipoles compensate
for changes in the receive antenna directivity and polarization as it moves and produces
up to 6 dB improvement in the link budget at θr = 90◦ .

3.7 SUMMARY AND CONCLUSIONS

Optimum array processing techniques for several classes of detection, estimation, and array
performance optimization problems have been shown to be closely related to one another
and, in fact, are different only by virtue of some scalar processing that follows a common
matrix filter and combiner operator. This matrix filter operator embodies the principle
of “first prewhiten, then match” in its realization. For narrowband signals, the matching
operation provides for time-delay (or phase) steering to obtain an undistorted represen-
tation of the signal. For broadband signals it is necessary to provide for signal spectral
matching, which is not equivalent to time-delay steering. Perturbed signal propagation
conditions resulting in noncoherent wavefront signals require element space-matched
processing in which the signal matching can be performed only in a statistical sense.
The optimum processors derived in this chapter for the most part have involved solving
unconstrained optimization problems. Various classes of constrained optimum processors
are introduced in the Problems section of this chapter.
The optimum array processor solutions resulting from the prewhitening and matching
operation are intimately related to the optimal Wiener solution, which thereby provides a
convenient theoretical performance limit against which to compare actual system perfor-
mance. Direct implementation of an optimum processor solution is impractical because
the signal environment statistics are usually unknown. Furthermore, even if the signal en-
vironment statistics were known, an optimal processor designed for a particular noise field
would have limited utility; consequently, adaptive techniques for realizing an optimum
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 131

3.8 Problems 131

processor are extremely important. Since the spatial properties of the steady-state solu-
tions resulting from various different performance measures are either identical or very
similar, the question of which array performance measure to select for a given application
is usually not very significant; rather, it is the temporal properties of the adaptive control
algorithm to be used to reach the steady-state solutions that are of principal concern to the
designer. The characteristics of adaptive control algorithms to be used for controlling the
weights in the array pattern-forming network are therefore highly important, and it is to
these concerns that Part 2 of this book is addressed.
Finally, the concept of a polarization sensitive array using polarization sensitive el-
ements was introduced. Such an array possesses more degrees of freedom than a con-
ventional array and permits signals arriving from the same direction to be distinguished
from one another. The notion of polarization “distance” is introduced using the Poincare
sphere. The polarization and gain of crossed dipoles can be adapted to maximize a dynam-
ically changing communications link. Adaptively adjusting the currents on the dipoles in
a crossed dipole system can significantly improve the link budget.

3.8 PROBLEMS
1. Use of the Maximum SNR Performance Measure in a Single Jammer Encounter [18]
The behavior of a linear adaptive array controlled by the maximum SNR performance measure
when the noise environment consists of a single jammer added to the quiescent environment
is of some practical interest. Consider a linear, uniformly spaced array whose weights are
determined according to (3.96). Assume that the quiescent noise environment is characterized
by the covariance matrix

⎡ ⎤
pq
⎢ 0 ⎥
⎢ ⎥
⎢ pq ⎥
Rnnq =⎢
⎢
⎥ = pq Ik
⎥
⎢ .. ⎥
⎣ 0 . ⎦
pq

where pq = noise power output of each of the K array elements

Ik = identity matrix of order K
In general, the various weight amplitudes are not equal, and the array beam pattern is
given by


K
G q (β) = ak e j (k−1)(β−βs )
k=1

where

2π d
β= sin θ
λ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 132

132 CHAPTER 3 Optimum Array Processing

or, in matrix notation, G q (β) = bT wq , where

⎡ ⎤ ⎡ ⎤
1 a1
⎢ e jβ ⎥ ⎢ a2 e− jβs ⎥
⎢ e j2β ⎥ ⎢ a e− j2βs ⎥
⎢
b=⎢ ⎥, ⎢
wq = ⎢ 3 ⎥
. ⎥ . ⎥
⎣ .. ⎦ ⎣ .. ⎦
j (K −1)β − j (K −1)βs
e ak e

In view of the expressions for wq and G q (β) previously given, it follows that
⎡ ⎤
1
⎢1⎥
UHwq = G q (β J )⎢ ⎥
⎣ .. ⎦
.
1

where U and H are described subsequently.

Consider a single narrowband jammer located at an angle θ J from mechanical boresight.
Denote the jamming signal appearing in the first element channel by J(t); then the jamming
signal appearing in the kth channel will be J(t) exp[− j (k − 1)β J ], where β J = (2π d/λ) sin θ J .
Let p J represent the envelope jamming power in each channel; then the covariance of the
jamming signals in the kth and lth channels is p J exp[− j (k − l)β J ], which represents the klth
element of the jammer covariance matrix RJJ .
(a) Show that RJJ is expressed as

RJJ = p J H∗ UH

where H is the diagonal matrix

⎡ ⎤
1
⎢ e jβ J 0 ⎥
⎢ ⎥
H=⎢
⎢
e j2β J ⎥
⎥
⎣ .. ⎦
.
0 e j (K −1)β J

and U is a K × K matrix of ones.

(b) By use of the control law (3.99), it follows that the weight vector for the quiescent noise
environment should satisfy Rnnq wq = μt∗ , where Rnnq wq = pq wq . If the noise environ-
ment changes so the noise covariance matrix is now Rnn instead of Rnnq , it follows that the
optimum weight vector should then satisfy Rnn w = μt∗ or w = pq R−1 nn wq . The covariance
matrix corresponding to the total noise environment (consisting of the quiescent noise plus
jammer) is the sum Rnn = Rnnq + RJJ or Rnn = pq Ik + p J H∗ UH. To obtain the optimum
weight vector for the total noise environment requires the inverse of Rnn . Using the fact
that H∗ = H−1 , show that
   !
1 PJ
R−1
nn = Ik − H∗ UH
pq pq + Kp J

(c) Using the result from part (b), show that the optimum weight vector is given by
 
pJ
w = wq − H∗ UHwq
pq + Kp J
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 133

3.8 Problems 133

(d) Let the optimum quiescent weight vector for a desired signal located in the direction θs
from mechanical boresight be expressed as
⎡ ⎤
a1
⎢ a2 e− jβs ⎥
⎢ a e− j2βs ⎥
wq = ⎢
⎢.
3 ⎥
⎥
⎣ .. ⎦
− j (K −1)βs
aK e
where
2πd
βs = sin θs
λ
and the ak represent weight amplitudes. If the various ak are all equal, the resulting array
beam pattern is of the form (sinKx)/(sinx).
It then follows from the definition of H in part (a) that
⎡ ⎤
a1
⎢ a2 e j (β J −βs ) ⎥
Hwq = ⎢
⎣ ..
⎥
⎦
.
j (K −1)(β J −βs )
aK e
From the foregoing expressions and the results of part (c), show that the pattern obtained
with the optimum weight vector may be expressed as
 
PJ
G(β) = bT w = bT wq − G q (β J )b∗J
Pq + KP J
where b J is just b with the variable β J replacing β.
(e) Recalling from part (d) that bT wq = G q (β), show that bT b∗J = C(β − β J ) where
   !
(K − 1)x sin K x/2
C(x) = exp j
2 sin x/2

(f) Using the fact that bT wq = G q (β) and the results of parts (d) and (e), show that
 
PJ
G(β) = G q (β) − G q (β J )C(β − β J )
Pq + KP J
This result expresses the fact that the array beam pattern of an adaptively controlled linear
array in the presence of one jammer consists of two parts. The first part is the quiescent
pattern G q (β), and the second part (which is subtracted from the first part) is a (sin Kx)/(sin
x)-shaped cancellation beam centered on the jammer.
(g) From the results of part (e) it may be seen that C(x)|x=0 = K . Using this fact show that
the gain of the array in the direction of the jammer is given by
 
Pq
G(β J ) = G q (β J )
Pq + KP J
With the array weights fixed at wq , the array pattern gain in the direction of the jammer
would be G q (β J ). The foregoing result therefore shows that the adaptive control reduces
the gain in the direction of the jammer by the factor
Pq 1
=
Pq + KP J 1 + K (PJ /Pq )
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 134

134 CHAPTER 3 Optimum Array Processing

(h) The proper measure of performance improvement against the jammer realized by the
introduction of adaptively controlled weights is the cancellation ratio  defined by
 γ
=
1 − γ · (J/N )q
where

 wq† Rnnq R−1

nn RJJ wq
γ = †
wq RJJ wq
and (J/N )q denotes the ratio of the jammer noise output to the quiescent noise output
when the weights are fixed at wq , that is,
 
J  wq† RJJ wq
= †
N q wq Rnnq wq
It is now desired to evaluate γ and  for the single jammer environment. First, show that
 
1
Rnnq R−1 = Ik − RJJ
nn
Pq + KP J
Then show that

RJJ RJJ = KP J RJJ

to obtain the result

 
Pq
Rnnq R−1
nn RJJ = RJJ
Pq + KP J
Substituting these results into the expression for γ immediately yields
1
γ =
1 + (KP J /Pq )
For a main beam of the form (sin Kx)/(sin x), the maximum possible value of output
jammer noise power is Kp J , which will occur only at the peak of the main beam. Conse-
quently, the maximum possible value of (J/N )q is KP J /Pq , in which case  = 1 (thereby
indicating no performance improvement against the jammer). For jammers located in the
sidelobe region, however, (J/N )q  KP J /Pq so that γ · (J/N )q  1; hence,  ≈ γ . For
a jammer located in the sidelobe region, the adaptive control “cancels” the jammer power
by approximately the jammer-to-noise ratio in the cancellation beam.

2. Effect of Adaptive Weight Errors [55]

(a) For the CSLC system depicted in Figure 3-5 with w 0 = 1, show that Pr = Px
− w† z–z† w + w† R yy w.
(b) Since wopt = R−1 ∗ † −1
yy z and Prmin = Px − z R yy z, show that when w = wopt + , then
Pr = Prmin + Padd where Padd = † R yy . Note that since Padd is given by a Hermitian
quadratic form, it is bounded by the smallest and largest eigenvalues of R yy as λmin ||||2 ≤

Padd ≤ λmax ||||2 where ||||2 = † .
(c) Since λmin and λmax are both geometry dependent, it is instructive to use a worst-case design
approach. Assume each auxiliary array element receives the same interference power PIa .
Furthermore, assume that the receiver of each array element has the same effective received
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 135

3.8 Problems 135

noise power Pn . Under these conditions, each diagonal entry of R yy is equal to Pn + PIa ,
and hence trace (R yy ) = N (Pn + PIa ). Since the largest eigenvalue of R yy is less than the
trace of R yy , it follows that Padd ≤ N (Pn + PIa )||||2 . Assume that the weight errors are
due to quantization errors where the quanta size of the in-phase and quadrature channel
is q. Under worst-case conditions, each complex weight component quantization error is
identical and equal to (q/2)(i ± j). Hence, show that

N 2q 2
Padd ≤ (Pn + PIa )
2
(d) The interference-to-noise ratio (PI /Pn )main for the main element is related to the ratio
(PIa /Pn )aux for each auxiliary element by
   
PI PIa
= |α|2
Pn main
Pn aux

where α is the average voltage gain of the main antenna in the sidelobe region. Assume
that α is given by

α = q · 2 B−1

where B represents the number of bits available to implement the quantizer. Show that by
assuming (PI /Pn )main  1 and using the results of part (c), then
 
Padd N2
= R0
Prmin 22B−1

where
 Prmin
R0 =
Pn
(e) The principle use of a worst-case analysis is to determine the minimum number of bits
required to avoid any degradation of the SIR performance. When the number of available
quantization bits is significantly less than the minimum number predicted by worst-case
analysis, a better prediction of SIR performance is obtained using an MSE analysis. The
only change required for an MSE analysis is to treat ||||2 in an appropriate manner.
Assuming the weight errors of the in-phase and quadrature channels are independent and
uniformly distributed, show that
1
||||2average = ||||2max
6
and develop corresponding expressions for Padd and Padd /Prmin .
3. Wiener Linear MMSE Filtering for Broadband Signals [56]
Consider the scalar signal x(t)

x(t) = s(t) + n(t)

where the desired signal s(x) and the noise n(t) are uncorrelated. The signal x(t) is to be passed
through a linear filter h(t) so that the output y(t) will best approximate s(t) in the MMSE sense.
The error in the estimate y(t) is given by

e(t) = y(t) − s(t)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 136

136 CHAPTER 3 Optimum Array Processing

and it is desired to minimize E{e2 (t)} where

" ∞ " ∞
y(t) = h(τ )s(t − τ )dτ = h(τ )s(t − τ )dt
−∞ −∞
+ ,- .
 
= s (t)
" ∞
+ h(τ )n(t − τ )dτ
+ −∞ ,- .
 
= n (t)

(a) Show that

 
E{e2 (t)} = E [s  (t) − s(t)]2 +n  (t)
2
+ ,- . + ,- .
signal noise
distortion component
component

(b) Show that

" ∞ " ∞
E{n 2 (t)} = rnn (v)h(τ )h(τ + v)dvdτ
−∞ −∞

where rnn (v) = autocorrelation function of n(t).

(c) By not imposing realizability or finite memory requirements on h(t), then H (ω) may be
introduced using
" ∞
H (ω) = h(t)e− jωt dt = {h(t)}
−∞

show that
" ∞
E{n 2 (t)} = H (ω)H ∗ (ω)φnn (ω)dω
−∞

where φnn (ω) = {rnn (τ )}, the spectral density function of n(t). Likewise, show that
" ∞

E{[s (t) − s(t)] } = 2
[1 − H (ω)][1 − H ∗ (ω)] · φss (ω)dω
−∞

so that
" ∞  
E{e2 (t)} = H (ω)H ∗ (ω)φnn (ω) + [1 − H (ω)][1 − H ∗ (ω)]φss (ω) dω
−∞

(d) It is desired to minimize E{e2 (t)} by appropriately choosing H (ω). Since the integrand
of the expression for E{e2 (t)} appearing in part (c) is positive for all choices of H (ω), it
is necessary only to minimize the integrand by choosing H (ω). Show that by setting the
gradient of the integrand of E{e2 (t)} with respect to H (ω) equal to zero, there results
φss (ω)
Hopt (ω) =
φss (ω) + φnn (ω)
which is the optimum scalar Wiener filter. Therefore, to obtain the scalar Wiener filter, it
is necessary to know only the signal spectral density and the noise spectral density at the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 137

3.8 Problems 137

point in question. For scalar processes, the foregoing result may be used to obtain the MMSE
signal estimate by introducing the appropriate scalar Wiener filter at the point where a MMSE
signal estimate is desired. It is easy to show that the corresponding result for vector processes
is given by

opt (ω) = [ss (ω) + nn (ω)]−1 ss (ω)

and when x(t) = s(t) + n(t) where s(t) = vd(t) then opt (ω) is expressed as

opt (ω) = −1

x x (ω)φxd (ω)

which represents the broadband signal generalization of (3.56).

(e) Determine the appropriate scalar Wiener filter to insert at point III in Figure 3-10 to obtain
the MMSE estimate of s(t). Likewise, determine the appropriate scalar Wiener filter to insert
at point V to obtain the MMSE estimate of s(t).
4. The Use of Multiple Linear Constraints in Array Gain Maximization [17]
It was found that in the process of maximizing array gain with a single linear constraint of the
form, k† m = β, the constraint could be handled by using the degree of freedom in specifying
the magnitude of k. In the event that a null constraint of the form k† d = 0 is introduced,
however, the situation must be handled as a special case of the more general problem involving
multiple linear constraints. Multiple linear constraints may be introduced to keep the array gain
relatively constant over a range of signal perturbations, thereby reducing any sensitivity to such
perturbations. Likewise, multiple linear constraints may be used to control sidelobes in a specific
neighborhood of directions. The basic problem is stated as that of minimizing the output power
z subject to a constraint of the form H† k = g where each row vector of the constraint matrix H†
†
imposes a constraint of the form hi k = gi . Clearly, then, the constraint matrix H† has a row for
each constraint, and the total number of rows must be less than the number of sensors, or the
problem will be overspecified. Introducing the multiple linear constraints into the expression
for output power by using a vector Lagrange multiplier λ, it is therefore desired to minimize
   
z = k† + λ† H† xx −1 xx xx −1 Hλ + k − λ† H† xx −1 Hλ − λ† g − g† λ

(a) Complete the square of the above expression for ???z to obtain the equivalent result
   
z = k† + λ† H† −1 −1 † †−1 † †
x x x x x x Hλ + k − λ Hx x Hλ − λ g − g λ

Since k appears only in the previously given first quadratic term, the minimizing value of
k is obviously that for which the quadratic term is zero or

kopt = −−1
x x Hλ

(b) Use the constraint equation H† k = g to eliminate λ from the result for kopt obtained in
part (a) and thereby show that
† −1
−1
kopt = −1
x x H H x x H g

With this value of k the output power then becomes

 −1
z = g† H−1
xx H g

(c) For the general array processor of Figure 3-12, the output power is given by
z = trace(† x x ), and the problem of minimizing z subject to multiple linear constraints
of the form H†  = L is handled by using a matrix Lagrange multiplier, , and considering
 
z = trace † x x  + † [H†  − L] + [† H − L† ]
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 138

138 CHAPTER 3 Optimum Array Processing

Complete the square of the previous expression to show that the optimum solution is

† −1
−1
opt = −1
x x H H x x H L

†  −1
for which opt x x opt = L† H† −1
xx H L
5. Introduction of Linear Constraints to the MSE Performance Measure [15]
The theoretically optimum array processor structure for maximizing (or minimizing) some
performance measure may be too complex or costly to fully implement. This fact leads to
the consideration of suboptimal array processors for which the processor structure is properly
constrained within the context of the signal and interference environment.
The K-component weight vector w is said to be linearly constrained if

f = c† w

The number of linear, orthonormal constraints on w must be less than K if any remaining
degrees of freedom are to be available for adaptation.
(a) The MSE is expressed as

E{|y − y A |2 }

where y A = w† x.
Consequently

E{|y − y A |2 } = E{|y|2 } − 2 Re{w† rx y } + w† Rx x w

Append the constraint equation to the MSE by means of a complex Lagrange multiplier
to form

J = E{|y|2 } − 2 Re{w† rx y } + w† Rx x w + λ[ f − c† w] + [ f ∗ − w† c]λ∗

Take the gradient of the foregoing expression with respect to w and set the result equal to
zero to obtain
 
wopt = R−1 ∗
x x rx y + λ c

(b) Apply the constraint f = c† w to the result in part (a) thereby obtaining a solution for λ
 
f ∗ − r†x y R−1
xx c
λ=  
c† R−1
xx c

This solution may be substituted into the result of part (a) to obtain the resulting solution
for the constrained suboptimal array processor.

6. Introduction of Multiple Linear Constraints [15]

A multiple linear constraint on the array processor takes the form

f = C† w

where the matrix C has the vector ci as its ith column, and the set {ci /i = 1, 2, . . . , M} must
be a set of orthonormal constraint vectors.
Append the multiple linear constraint to the expected output power of the array output
signal to determine the optimum constrained solution for w.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 139

3.8 Problems 139

7. Introduction of Quadratic Constraints to an Expected Output Power Performance

Measure [15]
The weight vector w is said to be quadratically constrained if

g = w† Qw

where Q is a K × K Hermitian matrix.

(a) The expected output power of an array output signal is

E{|y|2 } = E{w† xx† w} = w† E{xx† }w = w† Rx x w

Appending the constraint equation to the expected output power with a complex Lagrange
multiplier yields

J = w† Rx x w + λ[g − w† Qw]

Take the gradient of the previous expression with respect to w and set the result equal to
zero (to obtain the extremum value of w) to obtain

R−1 −1
x x Qw = λ w
 
(b) Note that the previous result for w is satisfied when w is an eigenvector of R−1 x x Q and
λ is the corresponding eigenvalue. Therefore,  maximizing
 (minimizing) J corresponds to
selecting the largest (smallest) eigenvalue of R−1
xx Q . It follows that a quadratic constraint
is regarded simply as a means of scaling the weight vector w in the array processor.
(c) Append both a multiple linear constraint and a quadratic constraint to the expected output
power of the array to determine the optimum constrained solution for w. Note that once
again the quadratic constraint merely results in scaling the weight vector w.
8. Introduction of Single-Point, Multiple-Point, and Derivative Constraints to a Minimum
Power Output Criterion for a Signal-Aligned Array [49]
(a) Consider the problem of minimizing the array output power given by

P0 = w† Rx x w

subject to the constraint C† w = f. Show that the solution to this problem is given by

† −1
−1
wopt = R−1
x x C C Rx x C f

(b) A single-point constraint corresponds to the case when the following conditions hold:

C = 1, the N × 1 vector of one’s where N is the number of array elements

so the constraint equation becomes

w† 1 = N

Show that the corresponding optimal weight vector is given by

N R−1
xx 1
wopt =  
1T R−1
xx 1

Under the single-point constraint, the weight vector minimizes the output power in all
directions except in the look (presteered) direction.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 140

140 CHAPTER 3 Optimum Array Processing

(c) A multiple-point constraint is introduced by specifying that C is a matrix of dimension

N × 3 (for a three-point constraint) given by

C = P = [e1 , e, e2 ]

where e1 and e2 are direction vectors referenced to the beam axis on either side, and
⎡ †
⎤
e1 e
f = P† e = ⎣ N ⎦
†
e2 e

where C† w = f. Show that the corresponding optimal constrained weight vector is given by

† −1
−1
wopt = R−1
x x P P Rx x P P† e

(d) Derivative constraints are used to maintain a broader region of the main lobe by specifying
both the response on the beam axis and the derivatives of the response on the beam axis.
The constraint matrix now has dimension N × k and is given by
 
C = D = e0 , e0 , e0 , . . .

where e0 = the array steering vector

e0 = the derivative of e0 with respect to sin θ0 where θ0 is the look direction
e0 = the second derivative of e0 with respect to sin θ0
and fT = δ10T
= [N , 0, 0, . . .] has dimension 1 × k. Show that the corresponding
optimal weight vector is given by
† −1
 −1
wopt = R−1
x x D DRx x D δ 10

9. Maximum Likelihood (ML) Estimates of Target Range and Bearing [39]

Optimal array processors designed for the detection and estimation of desired signals may
only partially satisfy the requirements placed on the system signal processor since the second
stage of the signal processing problem often involves the extraction of information concerning
parameters such as target range and bearing.
A maximum likelihood estimator selects the parameter α, which maximizes the conditional
probability density function (or “likelihood function”) p(x/α). Since it is usually preferable to
work with lnp(·), solutions are sought to the equation
∂
ln p(x/α) = 0
∂α
By forming the likelihood functional
∂
y(α) = ln p(x/α)
∂α
the maximum likelihood estimate α̂ML is found by requiring y(α) to be equal to zero, and the
statistics of α̂ML is related to those of y(α).
When the components of the received signal vector x (using a Fourier coefficients repre-
sentation) are complex Gaussian random processes, then
1
p(x/α) = exp(−x† M −1 x)
Det(π M)
where M = ss (ω) + nn (ω).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 141

3.8 Problems 141

(a) Using the expression for p(x/α), show that the likelihood functional is expressed as
 
∂M −1 ∂M
y(α) = x† M−1 M x − trace M−1
∂α ∂α

Note that ss (ω) is a function of α, whereas nn (ω) is not.

(b) It is shown [40] that the variance associated with the ML estimate α̂ML for the case of
spatially incoherent noise is proportional to the quantity

var(α̂ML ) ∝ [trace(TT† )]−1

where T is a weighting matrix that incorporates all geometrical properties of the array. In
particular, for linear arrays the T matrix has elements given by the following:

sin θ
For bearing estimation: ti j = (z i − z j )

sin2 θ  2 
For range estimation: ti j = − z − z 2j
2r 2 i
where θ = signal bearing with respect to the array normal
z n = position of nth sensor along the array axis (the z-axis)
 = velocity of signal propagation
r = true target range
Show that for an array having length L, and with K  1 equally spaced sensors then

62
For bearing estimation: [trace(TT† )]−1 =
K 2 L 2 sin2 θ
452 r 4
For range estimation: [trace(TT† )]−1 =
2L 4 K 2 sin4 θ
The foregoing results show that the range estimate accuracy is critically dependent on the
true range, whereas the bearing estimate is not (except for the range dependence of the
SNR). The range estimate is also more critically dependent on the array aperture L than
the bearing estimate.
10. Suboptimal Bayes Estimate of Target Angular Location [9]
The signal processing task of extracting information concerning parameters such as target an-
gular location can also be carried out by means of Bayes estimation. A Bayes estimator is just
the expectation of the variable being estimated conditioned on the observed data, that is,
" ∞
û = E{u k /x} = u k p(u k /x)du
−∞

where u k = sin θ denotes the angular location of the kth target, x denotes the observed data
vector, and the a posteriori probability density function p(u/x) is rewritten by application of
Bayes’s rule as

p(x/u) p(u)
p(u/x) =
p(x)

The optimum estimators that result using the foregoing approach are quite complex, re-
quiring the evaluation of multiple integrals that may well be too lengthy for many portable
applications. Consequently, the development of a simple suboptimum estimator that approxi-
mates the optimum Bayes estimator is of practical importance.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 142

142 CHAPTER 3 Optimum Array Processing

(a) The lowest-order nonlinear approximation to the optimum Bayes location estimator is
given by [9]

û = xT B

where the matrix of N target locations is a 1 × N row vector

û = [û 1 , û 2 , . . . , û N ]

Assuming a (2K + 1) element array where each element output is amplified and detected
with synchronous quadrature detectors, then the output of the mth pair of quadrature de-
tectors is

x ym (t) = s ym (t) + n ym (t)

and

x zm (t) = szm (t) + n zm (t)

The data vector x can then be defined by a 2K (2K + 1) × 1 column vector

x = [x ym x zn ]

in which each index pair m, n occupies a separate row. Define a 2K (2K + 1) × 1 column
vector of coefficients

(k)

bk = bmn , m = n

where, again, each index pair m, n occupies a separate row. The full matrix of coefficients
B is then given by the 2K (2K + 1) × N matrix

B = [b1 , b2 , . . . , b N ]

Show that, by choosing the coefficient matrix B so that the resulting estimator is orthogonal
to the estimation error, that is, E[ûT (u − û)] = 0, then B is given by

B = [E{xxT }]−1 E {xu}

With B selected as indicated, the MSE in the location estimator for the kth target is then
given by
 
k = E[u k (u k − û k )] = E u 2k − [E{xu k }]T · [E{xxT }]−1 · [E{xu k }]

(b) Consider a one-target location estimation problem using a two-element array. From the
results of part (a), it follows that

û = b12 x y1 x z2 + b21 x y2 x y1

Let the signal components of the quadrature detector outputs be given by

s ym = α cos mπ u + β sin mπ u

and

szm = β cos mπ u − α sin mπu

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 143

3.8 Problems 143

where the joint probability density function (PDF) for α, β is given by

 !
1 (α 2 + β 2 )
p(α, β) = exp −
2π σ 2 2σ 2
Furthermore, assume the quadrature noise components are independent so the PDF for the
noise originating at the mth antenna element is
/  0
1 n 2ym + n 2zm
p(n ym , n zm ) = exp −
2π σn2 2σn2

The aforementioned signal and noise models correspond to a Rayleigh fading environ-
ment and additive noise due to scintillating clutter. Show that the optimum coefficients for
determining û are given by

(1/σ 2 )E{u sin π u}

b12 = −b21 =
4E{sin2 π u} + 2/γ + 1/γ 2

where γ = σ 2 /σn2 , the SNR. Finally, show that the MSE for this estimator is given by

E 2 {u sin π u}
MSE = E{u 2 } −
2E{sin π u} + 1/γ + 2/2γ 2
2

and therefore depends only on the SNR and p(u).

11. Constrained Minimum Power Criterion for Element Space Matched Array Processing in
Narrowband Application [48]
A minimum variance estimate of the signal power Ps for an element space-matched array
processor may be obtained by solving

minimize : var[x† Kx] = tr[(KRx x )2 ]

subject to the constraint tr[KRss ] = 1.

(a) Show that the solution of the problem is given by

R−1 −1
x x Rss Rx x
Kopt =  2 
tr R−1
x x Rss

so the expected value of the processor output power is then

 
tr R−1
x x Rss
P0 =  2 
tr R−1
x x Rss

(b) Show that when Rss has rank one and is consequently given by the dyad Rss = vv† , then

Kopt = h(ω)h† (ω)

where

h = scalar · v† R−1
nn

This result demonstrates that under plane wave signal assumptions the element space-
matched array processor degenerates into a plane wave matched processor with a quadratic
detector.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 144

144 CHAPTER 3 Optimum Array Processing

FIGURE 3-19
Points Md and Mi on
the Poincare Sphere
From Compton,
Md
IEEE Trans. On Ant.
and Prop, 2yd
September, 1981. nd–ni Mi
2yi
H

12. Signal-to-Interference–plus-Noise Ratio (SINR) for a Two Element Polarized Array [49]
Let Md and Mi be points on the Poincare sphere representing the polarizations of the desired
and interference signals, respectively as shown in Figure 3-20. From the figure it is seen that
2γd , 2γi , and the arc Md Mi form the sides of a spherical triangle. The angle ηd − ηi is the angle
opposite side Md Mi . Using a well-known spherical trigonometric identity, the sides on the
spherical triangle are related by cos 2γd cos 2γi + sin 2γd sin 2γi cos(ηd − ηi ) = cos(Md Mi ).
(a) For the case where both the desired and interference signals arrive from broadside, and
θd = φd = θi = φi = 90◦ , show that
 
Md Mi
|Ud Ui |2 = 2[1 + cos(Md Mi )] = 4 cos2
2

Pd A2 $ $2 A2 $ $2
(b) Using the fact that SINR = , where Pd = d $UTd w$ , Pi = i $UiT w$ , and
Pi + Pn 2 2
σ2 2
Pn = |w| where w denotes the weight vector, use the matrix inversion lemma (m.i.l.)
2
to show that the SINR is written as
 $ H $2 
$Ud Ui $
SINR = ξd UHd Ud −
ξi−1 + UiH Ui

A2d
where ξd = is the desired signal-to-noise ratio and similarly for ξi .
σ2
(c) Using the result of part (a) in the SINR of part (b) show that
  Md Mi  
4 cos2
SINR = ξd 2 − 2
ξi−1 + 2

13. Development of the SINR [57]

The signal covariance matrix is given by

 = σ 2 I + A2d Ud UHd + Ai2 Ui UiH ,

where the U denote the steering vectors for the desired and interference signals, respectively
It is first desired to find −1 .
(a) Use the matrix inversion lemma,

[B − βZZH ]−1 = B−1 − τ B−1 ZH B−1

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 145

3.8 Problems 145

where B is a nonsingular N × N matrix, Z is an N × 1 column vector, and β and τ are

scalars related by

τ −1 + β −1 = ZB−1 ZH

Applying the m.i.l. to the quantity σ 2 I + Ai2 Ui UiH , where B = σ 2 I, Z = Ui , and β = −Ai2 ,
show that
 −1
−1 −1 −1 1 1
τ = (ZB Z − β ) H
= Ui UiH + 2
σ 2 Ai
and hence
 
 2 
H −1 1 Ui UiH
σ I + Ai Ui Ui
2
= 2 I − −1
σ ξi + Ui UiH

where ξi is defined to be Ai2 /σ 2

(b) Use the m.i.l again, this time letting B = σ 2 I + Ai2 Ui UiH , Z = Ud , and β = −A2d to show
that
   
−1 1 τ 1 1 τ Ud UiH Ui UHd
 = 2 I − 4 Ud UHd − −1 + Ui UiH
σ σ ξi + UiH Ui σ 2 σ 2 ξi−1 + UiH Ui

 
τ Ud UHd τ Ui UHd
+ 4 −1 U d UH
+ Ui UHd
σ ξi + UiH Ui i
σ 4 ξi−1 + UiH Ui
where
  
1 1 1 Ui UHd Ud UiH
τ −1 = 2 + 2 Ud UHd − 2
Ad σ σ ξi−1 + UiH Ui

(c) Use the fact that w = −1 S where S = E{X∗ R(t)} = Ar Ad U∗d along with the result of
part (b) to show that
  
Ar A d τ γ Ui UHd
w= 1− U∗d − −1
Ui∗ .
σ2 σ 2 A2d ξi + UiH Ui
where Ar denotes reference signal amplitude, and γ is given by
   
Ui UHd Ud UiH
γ = A2d Ud UHd −
ξi−1 + UiH Ui

(d) Use the result of part (c) to show that

 
Ar A d τ γ γ
UTd w = 1−
σ2 σ 2 A2d A2d
(e) Use the result of part (d) along with the fact that τ from part (a) is expressed as
( γ )−1
τ = A2d 1 +
σ2
to show that
 
Ar γ
UTd w = .
Ad σ2 + γ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 146

146 CHAPTER 3 Optimum Array Processing

A2d $$ T $$2
(f) Use the result of part (e) along with Pd = Ud w to show that
2
 2
A2 γ
Pd = r
2 σ2 + γ
(g) In a manner similar to part (d) show that
   
Ar A d τ γ ξi−1
UiT w = 1− 2 2 Ui UHd −1
σ2 σ Ad ξi + UiH Ui

Ai2 $$ T $$2
(h) Use the result of part (g) along with Pi = Ui w to show that
2
 2  2
Ar2 A2d Ai2 1 ξi−1   
Pi = UiH Ud UHd Ui
2 σ +γ
2
ξi−1 + UiH Ui
(i) Finally, from part (c) show that
 2  
1 γ  H  H  ξi−1
|w| =
2
Ar2 A2d − U Ui U U d  2
σ2 + γ A2d d i
ξ −1 + UH Ui i i

σ 2
(j) From the fact that Pn = |w|2 , use the result of part (i) to show that
2
 2 /  2 0
A2 A2 A2 1 γ ξi−1  H  H  ξi−1
Pn = r d i − Ud Ui Ui Ud −1
2 σ2 + γ A2d ξi + UiH Ui

(k) Combining the results of part (h) with part (j), show that
 2
Ar2 Ai2 1
Pi + Pn = γ ξi−1
2 σ2 + γ
(l) Using the results of part (f) for Pd , show that the SINR reduces to the form
Pd γ
SINR = = 2.
Pi + Pn σ
Using the definition of γ given in part (c), it follows that an alternative expression for the
SINR is given by
 $ H $2 
$Ud Ui $
SINR = ξd UHd Ud −
ξi−1 + UiH Ui

3.9 REFERENCES
[1] P. L. Stocklin, “Space-Time Sampling and Likelihood Ratio Processing in Acoustic Pressure
Fields,” J. Br. IRE, July 1963, pp. 79–90.
[2] F. Bryn, “Optimum Signal Processing of Three-Dimensional Arrays Operating on Gaussian
Signals and Noise,” J. Acoust. Soc. Am., Vol. 34, No. 3, March 1962, pp. 289–297.
[3] D. J. Edelblute, J. M. Fisk, and G. L. Kinneson, “Criteria for Optimum-Signal-Detection
Theory for Arrays,” J. Acoust. Soc. Am., Vol. 41, January 1967, pp. 199–206.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 147

3.9 References 147

[4] H. Cox, “Optimum Arrays and the Schwartz Inequality,” J. Acoust. Soc. Am., Vol. 45, No. 1,
January 1969, pp. 228–232.
[5] B. Widrow, P. E. Mantey, L. J. Griffiths, and B. B. Goode, “Adaptive Antenna Systems,” Proc.
IEEE, Vol. 55, December 1967, pp. 2143–2159.
[6] L. J. Griffiths, “A Simple Algorithm for Real-Time Processing in Antenna Arrays,” Proc.
IEEE, Vol. 57, October 1969, pp. 1696–1707.
[7] N. Owsley, “Source Location with an Adaptive Antenna Array,” Naval Underwater Systems
Center, Rept. NL-3015, January 1971.
[8] A. H. Nuttall and D. W. Hyde, “A Unified Approach to Optimum and Suboptimum Processing
for Arrays,” U.S. Navy Underwater Sound Laboratory Report 992, April 1969, pp. 64–68.
[9] G. W. Young and J. E. Howard, “Applications of Spacetime Decision and Estimation Theory
to Antenna Processing System Design,” Proc. IEEE, Vol. 58, May 1970, pp. 771–778.
[10] H. Cox, “Interrelated Problems in Estimation and Detection I and II,” Vol. 2, NATO Advanced
Study Institute on Signal Processing with Emphasis on Underwater Acoustics, Enchede, The
Netherlands, August 12–23, 1968, pp. 23-1–23-64.
[11] N. T. Gaarder, “The Design of Point Detector Arrays,” IEEE Trans. Inf. Theory, part 1, Vol. IT-
13, January 1967, pp. 42–50; part 2, Vol. IT-12, April 1966, pp. 112–120.
[12] J. Capon, “Applications of Detection and Estimation Theory to Large Array Seismology,”
Proc. IEEE, Vol. 58, May 1970, pp. 760–770.
[13] H. Cox, “Sensitivity Considerations in Adaptive Beamforming,” Proceedings of NATO
Advanced Study Institute on Signal Processing, Loughborough, England, August 1972,
pp. 621–644.
[14] A. M. Vural, “An Overview of Adaptive Array Processing for Sonar Applications,” IEEE 1975
EASCON Record, Electronics and Aerospace Systems Convention, September 29–October 1,
pp. 34.A–34 M.
[15] J. P. Burg, “Maximum Entropy Spectral Analysis,” Proceedings of NATO Advanced Study
Institute on Signal Processing, August 1968.
[16] N. L. Owsley, “A Recent Trend in Adaptive Spatial Processing for Sensor Arrays: Constrained
Adaptation,” Proceedings of NATO Advanced Study Institute on Signal Processing, Lough-
borough, England, August 1972, pp. 591–603.
[17] R. R. Kneiper, et al., “An Eigenvector Interpretation of an Array’s Bearing Response Pattern,”
Naval Underwater Systems Center Report No. 1098, May 1970.
[18] S. P. Applebaum, “Adaptive Arrays,” IEEE Trans. Antennas Propag., Vol. AP-24, No. 5,
September 1976, pp. 585–598.
[19] A. Papoulis, Probability, Random Variables, and Stochastic Processes, New York, McGraw-
Hill, 1965, Ch. 8.
[20] R. A. Wiggins and E. A. Robinson, “Recursive Solution to the Multichannel Filtering Prob-
lem,” J. Geophys. Res., Vol. 70, No. 8, April 1965, pp. 1885–1891.
[21] R. V. Churchill, Introduction to Complex Variables and Applications, New York, McGraw-
Hill, 1948, Ch. 8.
[22] M. J. Levin, “Maximum-Likelihood Array Processing,” Lincoln Laboratories, Massachusetts
Institute of Technology, Lexington, MA, Semiannual Technical Summary Report on Seismic
Discrimination, December 31, 1964.
[23] J. Chang and F. Tuteur, Symposium on Common Practices in Communications, Brooklyn
Polytechnic Institute, 1969.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 148

148 CHAPTER 3 Optimum Array Processing

[24] P. E. Mantey and L. J. Griffiths, “Iterative Least-Squares Algorithms for Signal Extraction,”
Proceedings of the 2nd Hawaii Conference on System Sciences, 1969, pp. 767–770.
[25] R. L. Riegler and R. T. Compton, Jr., “An Adaptive Array for Interference Rejection,” Proc.
IEEE, Vol. 61, No. 6, June 1973, pp. 748–758.
[26] S. P. Applebaum, Syracuse University Research Corporation Report SPL TR 66-1, Syracuse,
NY, August 1966.
[27] S. W. W. Shor, “Adaptive Technique to Discriminate Against Coherent Noise in a Narrow-Band
System,” J. Acous. Soc. Am., Vol. 34, No. 1, pp. 74–78.
[28] R. T. Adams, “An Adaptive Antenna System for Maximizing Signal-to-Noise Ratio,”
WESCON Conference Proceedings, Session 24, 1966, pp. 1–4.
[29] R. Bellman, Introduction to Matrix Analysis, New York, McGraw-Hill, 1960.
[30] R. F. Harrington, Field Computation by Moment Methods, New York, Macmillan, 1968, Ch. 10.
[31] B. D. Steinberg, Principles of Aperture and Array System Design, New York, Wiley, 1976,
Ch. 12.
[32] L. J. Griffiths, “Signal Extraction Using Real-Time Adaptation of a Linear Multichannel
Filter,” SEL-68-017, Technical Report No. 6788-1, System Theory Laboratory, Stanford
University, February 1968.
[33] R. T. Lacoss, “Adaptive Combining of Wideband Array Data for Optimal Reception,” IEEE
Trans. Geosci. Electron., Vol. GE-6, No. 2, May 1968, pp. 78–86.
[34] C. A. Baird, Jr., and J. T. Rickard, “Recursive Estimation in Array Processing,” Proceedings
of the Fifth Asilomar Conference on Circuits and Systems, 1971, pp. 509–513.
[35] C. A. Baird, Jr., and C. L. Zahm, “Performance Criteria for Narrowband Array Process-
ing,” IEEE Conference on Decision and Control, December 15–17, 1971, Miami Beach, FL,
pp. 564–565.
[36] W. W. Peterson, T. G. Birdsall, and W. C. Fox, “The Theory of Signal Detectability,” IRE
Trans., PGIT-4, 1954, pp. 171–211.
[37] D. Middleton and D. Van Meter, “Modern Statistical Approaches to Reception in Communi-
cation Theory,” IRE Trans., PGIT-4, 1954, pp. 119–145.
[38] T. G. Birdsall, “The Theory of Signal Detectability: ROC Curves and Their Character,”
Technical Report No. 177, Cooley Electronics Laboratory, University of Michigan, Ann Arbor,
MI, 1973.
[39] W. J. Bangs and P. M. Schultheiss, “Space-Time Processing for Optimal Parameter Estima-
tion,” Proceedings of NATO Advanced Study Institute on Signal Processing, Loughborough,
England, August 1972, pp. 577–589.
[40] L. W. Nolte, “Adaptive Processing: Time-Varying Parameters,” Proceedings of NATO
Advanced Study Institute on Signal Processing, Loughborough, England, August 1972,
pp. 647–655.
[41] D. O. North, “Analysis of the Factors which Determine Signal/Noise Discrimination in
Radar,” Report PTR-6C, RCA Laboratories, 1943.
[42] L. A. Zadeh and J. R. Ragazzini, “Optimum Filters for the Detection of Signals in Noise,”
IRE Proc., Vol. 40, 1952, pp. 1223–1231.
[43] G. L. Turin, “An Introduction to Matched Filters,” IRE Trans., Vol. IT-6, June 1960,
pp. 311–330.
[44] H. L. Van Trees, Detection, Estimation, and Modulation Theory, Part I, New York, Wiley,
1968, Ch. 2.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 149

3.9 References 149

[45] H. Cox, “Resolving Power and Sensitivity to Mismatch of Optimum Array Processors,”
J. Acoust. Soc. Am., Vol., 54, No. 3, September 1973, pp. 771–785.
[46] H. L. Van Trees, “Optimum Processing for Passive Sonar Arrays,” IEEE 1966 Ocean Elec-
tronics Symposium, August, Hawaii, pp. 41–65, 1966.
[47] L. W. Brooks and I. S. Reed, “Equivalence of the Likelihood Ratio Processor, the Maximum
Signal-to-Noise Ratio Filter, and the Wiener Filter,” IEEE Trans. Aerosp. Electron. Syst.,
Vol. AES-8, No. 5, September 1972, pp. 690–691.
[48] A. M. Vural, “Effects of Perturbations on the Performance of Optimum/Adaptive Arrays,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-15, No. 1, January 1979, pp. 76–87.
[49] R. T. Compton, “On the Performance of a Polarization Sensitive Adaptive Array,” IEEE Trans.
Ant. & Prop., Vol. AP-29, September 1981, pp. 718–725.
[50] A. Singer, “Space vs. Polarization Diversity,” Wireless Review, February 15, 1998,
pp. 164–168.
[51] M. R. Andrews, P. P. Mitra, and R. deCarvalho, “Tripling the Capacity of Wireless Communi-
cations Using Electromagnetic Polarization,” Nature, Vol. 409, January 18, 2001, pp. 316–318.
[52] R. E. Marshall and C. W. Bostian, “An Adaptive Polarization Correction Scheme Using
Circular Polarization,” IEEE International Antennas and Propagation Society Symposium,
Atlanta, GA, June 1974, pp. 395–397.
[53] R. T. Compton, “The Tripole Antenna: An Adaptive Array with Full Polarization Flexibility,”
IEEE Trans. Ant. & Prop., Vol. AP-29, November 1981, pp. 944–952.
[54] R. L. Haupt, “Adaptive Crossed Dipole Antennas Using a Genetic Algorithm,” IEEE Trans.
Ant. & Prop., Vol. AP-52, August 2004, pp. 1976–1982.
[55] R. Nitzberg, “Effect of Errors in Adaptive Weights,” IEEE Trans. Aerosp. Electron. Syst.,
Vol. AES-12, No. 3, May 1976, pp. 369–373.
[56] J. P. Burg, “Three-Dimensional Filtering with an Array of Seismometers,” Geophysics, Vol. 29,
No. 5, October 1964, pp. 693–713.
[57] A. Ishide and R.T. Compton, Jr., “On grating Nulls in Adaptive Arrays,” IEEE Trans. on Ant.
& Prop., Vol. AP-28, No. 4, July 1980, pp. 467–478.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:36 150
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 151

P A R T II

Adaptive Algorithms

CHAPTER 4 Gradient-Based Algorithms

CHAPTER 5 Direct Inversion of the Sample
Covariance Matrix
CHAPTER 6 Recursive Methods for Adaptive
Array Processing
CHAPTER 7 Cascade Preprocessors
CHAPTER 8 Random Search Algorithms
CHAPTER 9 Adaptive Algorithm Performance
Summary
Chapter 3 showed that a variety of popular performance measures led to closely related
optimum weight vector solutions. Consequently, the choice of a specific performance
measure is not as important as the adaptive algorithm that adjusts the array weights,
since the algorithm influences the speed of the array transient response and the hardware
complexity. Part 2 presents a survey of algorithms and discusses the important performance
characteristics of each one. In some cases, algorithms are tailored to particular signal
conditions, whereas in other cases they handle a variety of signal environments. These
algorithm characteristics provide the designer with a means for picking the algorithm
based on convergence speed, operational signal environment, and hardware complexity.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 152
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 153

CHAPTER

Gradient-Based Algorithms
4
' $
Chapter Outline
4.1 Introductory Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.2 The LMS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.3 The Howells–Applebaum Adaptive Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.4 Introduction of Main Beam Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.5 Constraint for the Case of Known Desired Signal Power Level . . . . . . . . . . . . . . . . . . . 199
4.6 The DSD Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
4.7 The Accelerated Gradient Approach (AG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
4.8 Gradient Algorithm with Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
4.9 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
4.10 Phase-Only Adaptive Nulling Using Steepest Descent. . . . . . . . . . . . . . . . . . . . . . . . . . . 227
4.11 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
4.12 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
4.13 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
& %

Gradient algorithms are popular, because they are simple, easy to understand, and solve
a large class of problems. The performance, (w), and adaptive weights determine the
nature of the performance surface. When (w) is a quadratic function of the weight
settings, then it is a bowl-shaped surface with a minimum at the “bottom of the bowl.” In
this case, local optimization methods, such as gradient methods, can find the bottom. In
the event that the performance surface is irregular, having several relative optima or saddle
points, then the transient response of the gradient-based minimum-seeking algorithms get
stuck in a local minimum. The gradient-based algorithms considered in this chapter are
as follows:
1. Least mean square (LMS)
2. Howells–Applebaum loop
3. Differential steepest descent (DSD)
4. Accelerated gradient (AG)
5. Steepest descent for power minimization

153
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 154

154 CHAPTER 4 Gradient-Based Algorithms

Variations of these algorithms come from introducing constraints into the adjustment rule,
and one section develops the procedure for deriving such variations. Finally, changes in
the modes of adaptation are discussed, illustrating how two-mode adaptation enhances the
convergence.

4.1 INTRODUCTORY CONCEPTS

The method of steepest descent is the most common gradient algorithm applied to quadratic
performance surfaces. Any quadratic performance surface has a unique minimum point
that can be found by using a gradient-based algorithm.

4.1.1 The Quadratic Performance Surface

A bowl-shaped quadratic performance surface, like the mean square error (MSE) per-
formance measure for the adaptive array of Figure 3-3, is the ideal playing field for a
gradient-based algorithm. Recall from Chapter 3 that the array output signal is given by
y(t) = wT (t)x(t) (4.1)
Denoting the desired array response by d(t), we may express the error signal as
e(t) = d(t) − y(t) = d(t) − wT (t)x(t) (4.2)
The square of the foregoing error signal is then
e2 (t) = d 2 (t) − 2d(t)xT (t)w(t) + wT (t)x(t)xT (t)w(t) (4.3)
The MSE is just the expected value of e2 (t), or
E{e2 (t)} = ξ [w(t)]
2 (4.4)
= d (t) − 2rTxd (t)w(t) + wT (t)Rx x (t)w(t)
where the overbar denotes expected value, rxd (t) is given by (3.52), and Rx x (t) is given
by (3.13). When the input signals are statistically stationary, then rxd and Rx x are also sta-
tionary, and there is no need to write these quantities as a function of time. In nonstationary
signal environments, however, the notation rxd (t) and Rx x (t) is required.
The MSE in (4.4) is a quadratic function of the weight vector w(t). In the nonstationary
case, the bottom of the bowl as well as its curvature and orientation change. The analysis
of time-varying adaptive performance signal statistics is beyond the scope of this book.

4.1.2 The Method of Steepest Descent

If the statistics of the signal environment are perfectly known, then the gradient at any
point on the performance surface can be calculated. The gradient of (4.4) with respect to
the weight vector is [1]

∇{ξ [w(t)]} = −2rxd + 2Rx x w(t) (4.5)

It was shown in Chapter 3 that the minimum is the Wiener solution

wopt = R−1
x x rxd (4.6)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 155

4.1 Introductory Concepts 155

w2 FIGURE 4-1
Steepest descent
with very small step
size (overdamped
case).

Initial guess
w1

w2 FIGURE 4-2
Steepest descent
with large step size
(underdamped
case).

Initial guess

w1

On substituting (4.6) into (4.4), the minimum MSE is then found to be

2
ξmin = d (t) − wopt
T
rxd (4.7)

The method of steepest descent begins with an initial guess of the weight vector
components. Having selected a starting point, we then calculate the gradient vector and
perturb the weight vector in the opposite direction (i.e., in the direction of the steepest
downward slope). Contour plots of a quadratic performance surface (corresponding to a
two-weight adjustment problem) are shown in Figures 4-1 and 4-2. In these figures the
MSE is measured along a coordinate normal to the plane of the paper. The ellipses in these
figures are contours of constant MSE. The gradient is orthogonal to these constant value
contours (pointing in the steepest direction) at every point on the performance surface. If
the steepest descent uses small steps, it is “overdamped,” and the path taken to the bottom
appears continuous as shown in Figure 4-1. If the steepest descent uses large steps, it is
“underdamped,” and each step is normal to the error contour as shown in Figure 4-2.
The discrete form of the method of steepest descent is [1]
 
w(k + 1) = w(k) − s ∇ e2 (k) (4.8)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 156

156 CHAPTER 4 Gradient-Based Algorithms

where
w(k) = old weight vector guess at time kT
 + 1) = new weight vector guess at time (k + 1)T
w(k
∇ e2 (k) = gradient vector of the MSE determining the direction in which to move
from w(k)
s = step size
Substituting the gradient of (4.5) into (4.8) then yields

w(k + 1) = w(k) − 2s (Rx x w(k) − rxd ) (4.9)

4.1.3 Feedback Model of Steepest Descent

The transient behavior of the method of steepest descent yields valuable insight into the
behavior of the LMS algorithm. The only difference between the two weight adjustment
algorithms is that with steepest descent the signal environment statistics are perfectly
known (so the gradient at any point can be exactly determined), whereas the LMS algo-
rithm signal statistics are unknown (although here they are assumed to be stationary) and
therefore must be estimated. The first step in determining the transient behavior of the
method of steepest descent is to formulate a feedback model of the weight adjustment
relationship.
Figure 4-3 is the feedback flow graph of (4.8) and (4.9). The symbol Z −1 is the
Z -transform representation [2–5] of a unit (one iteration cycle) time delay, and Z −1 I is
the matrix transfer function of a unit delay branch. This flow graph represents a first-order
multidimensional sampled-data control loop.
By setting the initial weight vector w(0) equal to the initial guess in the flow graph, the
resulting sequence of w(k) behaves exactly as in the local minimization weight adjustment
algorithm. Since the “output” of the flow graph model is the current weight vector w(k), the
flow graph model determines the transient behavior of the weight adjustment algorithm.
Each transfer function appearing in the flow graph of Figure 4-3 is a diagonal matrix
except for the feedback branch denoted by 2Rx x . This branch matrix in general has off-
diagonal elements, since the input signals are usually mutually correlated. Consequently,
transients cross-couple from one component of the weight vector to the next, thereby
complicating the study of transient behavior. A remedy is to diagonalize the flow graph and
eliminate such cross-coupling effects from consideration: the diagonalization then enables
one to consider the natural modes of behavior of the flow graph by merely introducing a
coordinate transformation.

FIGURE 4-3 2rxd

Feedback model of
steepest descent. Gradient
+
+ w(k + 1)
Σ Δs I Σ Z −1 I w(k)
− +

2Rxx
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 157

4.1 Introductory Concepts 157

To diagonalize the flow graph of Figure 4-3, consider the expression for the MSE
given by (4.4). Using wopt and ξmin in (4.6) and (4.7), the MSE becomes
   T  
E e2 (k) = ξ(k) = ξmin + w(k) − wopt Rx x w(k) − wopt (4.10)
Since the matrix Rx x is real, symmetric, and positive definite (for real variables), it is
diagonalized by means of a unitary transformation matrix Q so that
Rx x = Q−1 Q (4.11)
where  is the diagonal matrix of eigenvalues, and Q is the modal square matrix of
eigenvectors. If Q is constructed from normalized eigenvectors, then it is orthonormal so
that Q−1 = QT , and the MSE becomes
 T  
ξ(k) = ξmin + w(k) − wopt QT Q w(k) − wopt (4.12)
Now define

Qw(k) = w (k) (4.13)

Qwopt = wopt (4.14)

Equation (4.12) can then be rewritten as

 T  
ξ(k) = ξmin + w (k) − wopt  w (k) − wopt (4.15)
Q projects w(k) into the primed coordinates [resulting in w (k)]. Since  is a diagonal
matrix and ξ(k) is a quadratic performance surface, the primed coordinates comprise the
principal axes of the quadratic performance surface. The feedback model of Figure 4-3
expresses all quantities in terms of the primed coordinate system. The resulting equivalent
feedback model of Figure 4-4 eliminates all cross-couplings that previously existed within
the feedback paths.
The steepest descent algorithm in Figure 4-4 is composed of the natural modes of the
flow graph. The transients of each mode are isolated (since each of the primed coordinates
has its own natural mode), and the natural behavior of steepest descent is completely
explored by considering the behavior of a single primed coordinate.
An isolated one-dimensional feedback model for the pth normal coordinate is shown
in Figure 4-5. The pulse transfer function of this closed-loop feedback system is [1]
wp (z) s Z −1
= (4.16)
r p (z) 1 + (1 − 2s λ p )Z −1

2rxd FIGURE 4-4

Diagonalized version
of feedback model
+ + w′(k + 1) w′(k) of steepest descent
Q Σ Δs I Σ Z −1 I Q−1 using normal
− + (primed)
coordinates.
w(k)

2L
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 158

158 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-5 r p′ (k)

One-dimensional
feedback model for +
the pth normal + w p′ (k +1)
Σ Δs Σ Z −1 w p′ (k)
coordinate of
steepest descent. − +

2l p

where λ p is the pth eigenvalue of Rx x . The impulse response of (4.16) is found by letting
r p (z) = 1 and taking the inverse Z -transform of the resulting output −1 {w p (z)}. It follows
that the impulse response is of the form

w p (kT ) = constant × e−α p (kT )

where
1
α p = − ln(1 − 2s λ p ) (4.17)
T
and T = one iteration period. The time response of (4.17) is a stable system when

|1 − 2s λ p | < 1 (4.18)

Equation (4.18) is satisfied when

 
s > 0 and  s λ p  < 1 (4.19)

Since Rx x is positive definite, λ p > 0 for all p. Consequently, the stability of the mul-
tidimensional flow graph of Figure 4-4 is guaranteed if and only if λ p = λmax in (4.19)
and

s > 0 and |s λmax | < 1 (4.20)

The stability of the steepest descent adaptation process is therefore guaranteed so long as
1
> s > 0 (4.21)
λmax

4.2 THE LMS ALGORITHM

When the operational environment signal statistics are stationary but unknown (a typical
situation), then the gradient of the performance surface at any point must be estimated.
The LMS algorithm introduced by Widrow has proven particularly useful for a quadratic
performance function [6–10]. It is worthwhile noting that the LMS algorithm requires
a reference signal, d(t), to generate the error signal given by (4.2). The desired signal
in a communications system is usually present, so the actual signal is used as the ref-
erence signal. In systems where the desired signal is usually not present (as in radar or
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 159

4.2 The LMS Algorithm 159

sonar systems), it is pointless to try to generate a fictitious desired signal. Thus, the LMS
algorithm described here is usually employed to improve communications system per-
formance. The LMS algorithm is exactly like the method of steepest descent except that
now changes in the weight vector are made in the direction given by an estimated gradient
vector instead of the actual gradient vector. In other words, changes in the weight vector
are expressed as

w(k + 1) = w(k) − s ∇[ξ(k)]

ˆ (4.22)

where
w(k) = weight vector before adaptation step
w(k + 1) = weight vector after adaptation step
s = step size that controls rate of convergence and stability
∇[ξ(k)]
ˆ = estimated gradient vector of ξ with respect to w
The adaptation process described by (4.22) attempts to find a solution as close as
possible to the Wiener solution given by (4.6). It is tempting to try to solve (4.6) directly,
but such an approach has several drawbacks:
1. Computing and inverting an N × N matrix when the number of weights N is large
becomes more challenging as input data rates increase.
2. This method may require up to [N (N + 3)]/2 autocorrelation and cross-correlation
measurements to find the elements of Rx x and rxd . In many practical situations, such
measurements must be repeated whenever the input signal statistics change.
3. Implementing a direct solution requires setting weight values with high accuracy in
open loop fashion, whereas a feedback approach provides self-correction of inaccurate
settings, thereby giving tolerance to hardware errors.
To obtain the estimated gradient of the MSE performance measure, take the gradient
of a single time sample of the squared error as follows:
ˆ k = ∇[ξ(k)] = 2e(k)∇[e(k)]
∇ (4.23)

Since

e(k) = d(k) − xT (k)w (4.24)

it follows that

∇[e(k)] = ∇[d(k) − xT (k)w] = −x(k) (4.25)

so that
ˆ k = −2e(k)x(k)
∇ (4.26)

It is easy to show that the gradient estimate given by (4.26) is unbiased by considering the
expected value of the estimate and comparing it with the gradient of the actual MSE. The
expected value of the estimate is given by
ˆ k } = −2E{x(k)[d(k) − xT (k)w(k)]}
E{∇ (4.27)
= −2[rxd (k) − Rx x (k)w(k)] (4.28)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 160

160 CHAPTER 4 Gradient-Based Algorithms

Now consider the MSE

ξ [x(k)] = E[d 2 (k)] + wT Rx x (k)w − 2wT rxd (k) (4.29)

Differentiating (4.29) with respect to w yields the gradient ∇{ξ [w(k)]} as

∇{ξ [w(k)]} = 2Rx x (k)w(k) − 2rxd (k) (4.30)

Comparing (4.28) and (4.30) reveals that

ˆ k } = ∇{ξ [w(k)]}
E{∇ (4.31)

so the expected value of the estimated gradient equals the true value of the gradient of
the MSE.
Substituting the estimated gradient of (4.26) into the weight adjustment rule of (4.22)
then yields the weight control rule

w(k + 1) = w(k) + 2s e(k)x(k) (4.32)

The LMS algorithm given by (4.32) can be rewritten for complex quantities as
w(k + 1) − w(k)
= 2ks e(k)x∗ (k) (4.33)
t
where t is the elapsed time between successive iterations, and s = ks t. In the limit
as t → 0, (4.33) yields an equivalent differential equation representation of the LMS
algorithm that is appropriate for use in continuous systems as
dw(t)
= 2ks e(t)x∗ (t) (4.34)
dt
Equation (4.34) can also be written as
 t
w(t) = 2ks e(τ )x∗ (τ )dτ + w(0) (4.35)
0

A block diagram representation of the weight adjustment rule represented by (4.35) is

shown in Figure 4-6.
The discrete version of (4.34) is given by (4.33) and is more commonly written as

w(k + 1) = w(k) + 2ks te(k)x∗ (k) (4.36)

A block diagram representation of the weight adjustment rule represented by (4.35) is

illustrated in Figure 4-7.

4.2.1 Convergence to the Wiener Solution

Assume that the time between successive iterations of the LMS algorithm is long enough
so that the signal vectors x(k) and x(k + 1) are uncorrelated. From (4.32) it follows that
w(k) is a function of only x(k − 1), x(k − 2), . . . , x(0), and w(0), where the successive
input signal vectors are uncorrelated so that w(k) is independent of x(k). It will now be
shown that for a stationary input signal process meeting these conditions, the expected
value of the weight vector E{w(k)} converges to the Wiener solution given by (4.6).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 161

4.2 The LMS Algorithm 161

FIGURE 4-6
Complex conjugate
Analog realization
of the LMS weight
adjustment
algorithm.
wi (t) xi (t)
ƒ ith Element

Integrator
Other
2ks Σ element
outputs
e(t)
_
+
d(t)

Array
output

FIGURE 4-7
Complex conjugate
Digital realization
of the LMS weight
adjustment
xi (k) algorithm.
w i (k + 1) Unit time w i (k)
Σ delay (Δt)
ith Element
Δt

Other
2ks Δt Σ element
outputs
e(k)
_
+
d(k)

Array
output

Taking the expected value of both sides of (4.36), there results

E{w(k + 1)} = E{w(k)} + 2ks t · E{x∗ (k)[d(k) − xT (k)w(k)]} (4.37)

Now let

E{x∗ (k)d(k)} = rxd (4.38)

∗
E{x (k)x (k)} = Rx x
T
(4.39)

Consequently, (4.37) is rewritten as

E{w(k + 1)} = E{w(k)} − 2ks tRx x E{w(k)} + 2ks trxd

= [I − 2ks tRx x ]E{w(k)} + 2ks trxd (4.40)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 162

162 CHAPTER 4 Gradient-Based Algorithms

Starting with an initial guess w(0), the (k + 1)th iteration of (4.40) yields

E{w(k + 1)} = [I − 2ks tRx x ](k + 1) w(0)


k
+ 2ks t [I − 2ks tRx x ]i rxd (4.41)
i=0

Diagonalizing (4.41) by using (4.11) to obtain the normal form results in

E{w(k + 1)} = Q−1 [I − 2ks t](k+1) Qw(0)


k
+ 2ks tQ−1 [I − 2ks t]i Qrxd (4.42)
i=0

When the magnitude of all the terms in the diagonal matrix [I − 2ks t] are less than
one, then

lim [I − 2ks t](k+1) → 0 (4.43)

k→∞

Therefore, the first term of (4.42) vanishes after a sufficient number of iterations, and the
summation factor in the second term of (4.42) becomes


k
1
lim [I − 2ks t]i = −1 (4.44)
k→∞
i=0
2ks t

Therefore, after a sufficient number of iterations, (4.42) yields

1
lim E{w(k + 1)} = 2ks tQ−1 −1 Qrxd
k→∞ 2ks t
= R−1
x x rxd (4.45)

This result shows that the expected value of the weight vector in the LMS algorithm does
converge to the Wiener solution after a sufficient number of iterations.
Since all the eigenvalues in  are positive, it follows that all the terms in the afore-
mentioned diagonal matrix, I − 2ks t, have a magnitude less than one provided that

|1 − 2ks tλmax | < 1

1
> ks t > 0 (4.46)
λmax
where λmax is the maximum eigenvalue of Rx x . The convergence condition (4.46) is
exactly the same as the stability condition (4.21) for the noise-free steepest descent feed-
back model.
The foregoing condition on ks for convergence of the mean value of the LMS algorithm
relates to the total input signal power PIN , as described henceforth. Since λmax satisfies
the inequality

λmax ≤ trace[Rx x ] (4.47)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 163

4.2 The LMS Algorithm 163

where

N

trace[Rx x ] = E{x† (k)x(k)} = E{|xi |2 } = PIN (4.48)
i=1

then the convergence condition (4.46) is assured if

1
> ks t > 0 (4.49)
PIN
These LMS algorithm convergence results assumed that successive input signal sam-
ples are independent. This independence assumption is probably overly restrictive, since
Griffiths [11] presented experimental results that show that adaptation using highly cor-
related successive samples also converges to the Wiener solution, although the resulting
steady-state MSE is slightly higher than what results for statistically independent suc-
cessive samples. For some applications, mean squared convergence and its associated
stability properties may be of concern, in which case more stringent conditions on ks must
be satisfied [12].

4.2.2 Transient Response Characteristics for LMS Adaptation

In normal coordinates, the adaptive weight transients consist of sums of exponentials with
time constants given by
1
τp = , p = 1, 2, . . . , N (4.50)
2(ks t)λ p
where λ p is the pth eigenvalue of the correlation matrix Rx x . Since τ p is inversely pro-
portional to λ p , the transient response is fastest for strong signals (large λ p ) and slowest
for weak signals (small λ p ). Thus, the LMS algorithm convergence time depends on the
eigenvalue spread in Rx x . The exponential having the longest time constant (correspond-
ing to the smallest normal coordinate signal power) determines the transient response of
the LMS algorithm. Often there is no choice for the value of the constant ks that repre-
sents a good compromise between the various eigenvalues that will yield a desirably short
transient period of operation.
Figure 4-8 shows a contour plot of a quadratic performance surface corresponding to
two widely diverse eigenvalues. The highly elongated MSE contours in Figure 4-8 result
in many adaptive iterations before weight values become acceptably close to the desired
Wiener solution. In the event that all the eigenvalues are equal, then all the time constants
are equal, and
1
τ= (4.51)
2(ks t)λ
A “learning curve” that plots the expected value of the performance measure at each
stage of the learning process as a function of the number of adaptation iterations provides
a convenient way of monitoring the progress of an adaptive process. It has just been shown
that the underlying transient behavior of the adaptive weights has an exponential nature.
Since the MSE is a quadratic function of the weight values, the transients in the MSE
function are also exponential.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 164

164 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-8 w2
Steepest descent
transient response
with widely diverse
eigenvalues.

w1

Initial guess

Since the square of an exponential function is an exponential having half the time
constant of the original exponential function, it follows that when all the time constants
are equal the MSE learning curve is an exponential having the time constant
τ 1
τMSE = = (4.52)
2 4(ks t)λ
In general, of course, the eigenvalues of Rx x are unequal so that
τp 1
τ pMSE = = (4.53)
2 4(ks t)λ p
where τ pMSE is the time constant for the MSE learning curve, τ p is the time constant in the
weights, and λ p is the eigenvalue of the pth normal mode. The adaptive process uses one
signal data sample/iteration, so the time constant expressed in terms of the number of data
samples is

T pMSE = τ pMSE (4.54)

Plots of actual experimental learning curves look like noisy exponentials—an effect
due to the inherent noise that is present in the adaptation process. A slower adaptation
rate (i.e., the smaller the magnitude of ks ) has a smaller noise amplitude that corrupts the
learning curve.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 165

4.2 The LMS Algorithm 165

4.2.3 Weight Misadjustment During LMS Adaptation

Adaptation speed of the LMS algorithm depends on two factors: (1) the weight adjustment
step size and (2) the statistical averages sample size. If a large step size is selected for
the weight adjustment process, then the excursions in successive weight values will be
large, resulting in transient behavior like the underdamped case shown in Figure 4-2. If a
small number of samples is used to estimate the statistical averages, then the time elapsed
in obtaining such averages is small, but the quality of the resulting estimates is low. In
general, a fast adaptive algorithm has a poor steady-state performance.
Since the input signal statistics are not known a priori, the minimum MSE is not the
Wiener solution. The “misadjustment” M compares the actual MSE with the optimum
Wiener solution
 [ξactual − ξmin ]
M= (4.55)
ξmin

where ξ = E{e2 }. The LMS algorithm misadjustment can be evaluated for a specified
value of ks t by considering the noise associated with the gradient-estimation process.
Assume that the adaptive process converged to a steady state in the neighborhood of
the MSE surface minimum point. The gradient-estimation noise of the adaptive algorithm
at the minimum point (where the true gradient is zero) is just the gradient estimate itself.
Therefore, the gradient noise vector g is given by

g(k) = ∇(k)
ˆ = −2e(k)x(k) (4.56)

The covariance of this estimation noise is given by

cov[g(k)] = E{g(k)gT (k)} = 4E{e2 (k)x(k)xT (k)} (4.57)

When the weight vector is optimized (w(k) = wopt ), then the error e(k) is uncorrelated
with the input vector x(k). If e(k) and x(k) are Gaussian processes, then not only are
they uncorrelated at the minimum point of the MSE surface, but they are also statistically
independent. With these conditions (4.57) becomes

cov[g(k)] = 4E{e2 (k)}E{x(k)xT (k)} = 4ξmin Rx x (4.58)

In the primed normal coordinates, the previous covariance can be written as

cov[g (k)] = Qcov[g(k)]Q−1 = 4ξmin  (4.59)

Adaptation based on noisy gradient estimates results in noise in the weight vector.
Recall that the noise-free method of steepest descent is described by the iterative relation

w(k + 1) = w(k) + s [−∇(k)] (4.60)

where s is the constant that controls stability and rate of convergence, and ∇(k) is the
gradient at the point on the performance surface corresponding to w = w(k). Following

Widrow and McCool [13], subtract wopt from both sides of (4.60), and define v(k) =
w(k) − wopt to obtain

v(k + 1) = v(k) + s [−∇(k)] (4.61)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 166

166 CHAPTER 4 Gradient-Based Algorithms

With estimated gradients instead of exact gradients, (4.61) can be rewritten as

v(k + 1) = v(k) + s (−∇(k))

ˆ = v(k) + s [−∇(k) − g(k)] (4.62)

Now since ∇(k) is given by (4.5), it follows that

∇(k) = −2rxd + 2Rx x [wopt + v(k)] = 2Rx x v(k) (4.63)

Consequently, (4.62) can be written as

v(k + 1) = (I − 2s Rx x )v(k) − s g(k) (4.64)

which represents a first-order vector difference equation with a stochastic driving function—
s g(k). Multiplying (4.64) by Q produces

v (k + 1) = (I − 2s )v (k) − s g (k) (4.65)

After initial transients have died out and the steady state is reached, v (k) responds to
the stationary driving function −s g (k) in the manner of a stationary random process.
The absence of any cross-coupling in the primed normal coordinate system means that the
components of both g (k) and v (k) are mutually uncorrelated, and the covariance matrix
of g (k) is therefore diagonal. To find the covariance matrix of v (k) consider

v (k + 1)vT (k + 1) = (I − 2s )v (k)vT (k)(I − 2s )

+ 2s g (k)gT (k) − s (I − 2s )v (k)gT (k)
− s g (k)vT (k)(I − 2s ) (4.66)

Taking expected values of both sides of (4.66) (and noting that v (k) and g (k) are un-
correlated since v (k) is affected only by gradient noise from previous iterations), we
find

cov [v (k)] = (I − 2s )cov [v (k)](I − 2s ) + 2s cov [g (k)]
 −1
= 2s 4s  − 42s 2 cov [g (k)] (4.67)

In practical applications, the LMS algorithm uses a small value for s , so that

s   I (4.68)

With (4.68) satisfied, the squared terms involving s  in (4.67) may be neglected, so
s −1
cov [v (k)] =  cov [g (k)] (4.69)
4
Using (4.59), we find
s −1
cov [v (k)] =  (4ξmin ) = s ξmin I (4.70)
4
Therefore, the covariance of the steady-state noise in the weight vector (near the minimum
point of the MSE surface) is

cov [v(k)] = s ξmin I (4.71)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 167

4.2 The LMS Algorithm 167

Without noise in the weight vector, the actual MSE experienced would be ξmin . The
presence of noise in the weight vector causes the steady-state weight vector solution
to randomly meander about the minimum point. This random meandering results in an
“excess” MSE— that is, an MSE that is greater than ξmin . Since
2
ξ(k) = d (k) − 2rTxd w(k) + wT (k)Rx x w(k) (4.72)

where
2
ξmin = d (k) − wopt
T
rxd (4.73)
wopt = R−1
x x rxd (4.74)

It follows that (4.72) can be rewritten as (also see (4.10))

ξ(k) = ξmin + vT (k)Rx x v(k) (4.75)

In terms of the primed normal coordinates, (4.75) can be rewritten as

ξ(k) = ξmin + vT (k)v (k) (4.76)

It immediately follows from (4.76) that the average excess MSE is


N
 2 
E{vT (k)v (k)} = λp E vp (k) (4.77)
p=1

Using (4.70) to recognize that E{[vp (k)]2 } is just s ξmin for each p, we see it then
follows that

N
E{vT (k)v (k)} = s ξmin λp
p=1
= s ξmin tr (Rx x ) (4.78)

The misadjustment in the LMS algorithm is therefore given by

E{vT (k)v (k)}

M= = s tr (Rx x ) (4.79)
ξmin
Since s = ks t, (4.79) emphasizes the fact that the degree of misadjustment experienced
with the LMS algorithm can be controlled merely by adjusting ks . When the step size is
decreased, the time required to reach the steady-state condition increases, so there is a
trade-off between the misadjustment and the adaptation speed.
The LMS algorithm misadjustment can also be expressed in a manner that gives
insight into the relationship between misadjustment and adaptation speed. From (4.53) it
follows that
1
s λ p = (4.80)
4τ pMSE
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 168

168 CHAPTER 4 Gradient-Based Algorithms

Furthermore

N 
N
1 N 1
s tr (Rx x ) = s λp = = (4.81)
p=1 p=1
4τ pMSE 4 τ pMSE av

where
1 
N
1  1
= (4.82)
τ pMSE av N p=1 τ pMSE

Consequently, the misadjustment can be written as

N 1 N 1
M= = (4.83)
4 τ pMSE av 4 T pMSE av
where T pMSE is the learning curve time constant in units of the number of data samples.

4.2.4 Practical Considerations for LMS Adaptation

Generation of the error signal in LMS adaptation requires an appropriate desired signal. If
the desired signal is the signal itself, then the adaptive array output reproduces the signal
in the best MSE sense and nearly eliminates the noise. As a practical matter, the signal is
not available for adaptation purposes—indeed, if it were available there would be no need
for a receiver and a receiving array.
An LMS algorithm artificially injects a known signal termed the reference signal or
pilot signal for the desired signal. The pilot signal has the same (or similar) directional
and spectral characteristics as those of the desired signal. These directional and spectral
characteristics may sometimes be known a priori, but usually only estimates of these
parameters are available. Many practical communication systems derive the reference
signal from the array output—a practice that requires a high degree of compatibility
between the signaling waveforms and the adaptive array. In general, it is not feasible
to simply put an adaptive array in any arbitrary communication system because of the
following reasons [14]:
1. The adaptive array weights are random processes that modulate the desired signal;
consequently, either the desired signal waveforms or the adaptive algorithm must be
chosen so this modulation does not impair the communication system effectiveness.
2. The desired signal and interference signal waveforms must be different in some respect,
so this known difference can be exploited to enable the adaptive array to distinguish
these two signal classes.
3. A practical method for reference-signal generation must be available.
The reference signal needs to satisfy only the following criteria [14]:
1. The reference signal must be highly correlated with the desired signal at the array
output.
2. The reference signal must be uncorrelated with any interference signal components
appearing at the array output.
If these two correlation properties are satisfied, then the adaptive array behaves in the
desired manner, since only the correlation between the reference signal and the element
signals xi (t) affects the adaptive weights. The impact of any phase shift occurring in
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 169

4.2 The LMS Algorithm 169

the network responsible for generating the reference signal (when the reference signal is
derived from the array output) is discussed in [15].
LMS algorithm adaptation with an injected pilot signal causes the array to form a
beam in the pilot-signal direction. This array beam has a flat spectra response and linear
phase shift characteristic within the passband defined by the spectral characteristic of the
pilot signal. Furthermore, directional noise incident on the array manifests as correlated
noise components that the array will respond by producing beam pattern nulls in the noise
direction within the array passband.
Since injection of the pilot signal could “block” the receiver (by rendering it insensitive
to the actual signal of interest), mode-dependent adaptation schemes have been devised
to overcome this difficulty. Two such adaptation algorithms are discussed in the following
section.

4.2.5 One-Mode and Two-Mode LMS Adaptation

Figure 4-9 illustrates a practical two-mode method [10] for providing a pilot signal to form
the array beam and then switching the pilot signal off to adapt to the inputs to eliminate
noise. The ideal time delays δ1 , δ2 , . . . , δ N are selected to produce a set of input signals
that appear to be a radiated plane wave from the desired direction. The adaptive processor
inputs are connected either to the actual sensor element outputs (during adaptation to
eliminate noise) or to the set of delayed signals obtained from the pilot signal generator
and the selected time-delay elements (to preserve the main lobe in the desired direction).
During adaptation, all signals delivered to the adaptive processor are sensor element
outputs derived from the actual noise field. The adaptation process in this mode tends to
eliminate all received signals since the desired response signal has been set to zero.
To preserve the main beam in the desired direction during adaptation, the input signals
to the adaptive processor are derived from the pilot signal. For example, if a sinusoidal
pilot signal having frequency f 0 is used, then minimizing the MSE forces the array gain

Pilot FIGURE 4-9

signal Two-mode LMS
generator _ adaptation for beam
Desired response + preservation and
noise elimination.

Beam preservation mode

d1

Noise elimination mode

d2
Adaptive
Output
processor

dN
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 170

170 CHAPTER 4 Gradient-Based Algorithms

in the desired look direction to have a specific amplitude and phase shift at that frequency.
On the other hand, if the pilot signal is chosen to be the sum of several sinusoids having
different frequencies, then the adaptation process forces the array gain and phase in the
desired look direction to have specific values at each one of the pilot-signal frequencies.
Finally, if several pilot signals corresponding to different look directions are added to-
gether, then the array gain is simultaneously constrained at the various frequencies and
angles corresponding to the different pilot signals selected. In summary, the two-mode
adaptation process minimizes the total power of all signals received that are uncorrelated
with the pilot signals while constraining the gain and phase of the array beam to values
corresponding to the frequencies and angles dictated by the pilot-signal components.
Figure 4-10 illustrates a practical one-mode method for simultaneously eliminating all
noises uncorrelated with the pilot signal and forming a desired array beam. The circuitry
of Figure 4-10 circumvents the difficulty of being unable to receive the actual signal, while
the processor is connected to the pilot-signal generator by introducing an auxiliary adaptive
processor. For the auxiliary adaptive processor, the desired response is the pilot signal,
and both the pilot signal and the actual received signals enter the processor. A second
processor performs no adaptation (its weights are slaved to the weights of the adaptive
processor) and generates the actual array output signal. The slaved processor inputs do
not contain the pilot signal and can therefore receive the transmitted signal at all times.
In the one-mode adaptation method, the pilot signal is on continuously so the adap-
tive processor that minimizes the MSE forces the adaptive processor output to closely
reproduce the pilot signal while rejecting all signals uncorrelated with the pilot signal.

FIGURE 4-10 Pilot

One-mode LMS signal
adaptation for generator
simultaneous beam _
+
preservation and
noise elimination.
d1
+
x1 +
Auxiliary
adaptive
d2 processor
+
x2 +

Second Array
(slaved) output
processor

dN
+
xN +
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 171

4.3 The Howells–Applebaum Adaptive Processor 171

The adaptive processor therefore preserves the desired array directivity in the look direc-
tion (over the pilot-signal passband) while placing nulls in the directions of noise sources
(over the noise frequency bands).

4.3 THE HOWELLS–APPLEBAUM

ADAPTIVE PROCESSOR
The key capability of adaptive nulling was developed for an intermediate frequency (IF)
radar sidelobe canceller as represented by the patent of Howells [16]. An analysis of this
approach by Applebaum [17] established the control-law theory governing the operation
of an adaptive control loop for each array element. The Applebaum algorithm maximizes
a generalized signal-to-noise ratio (SNR) with the assumptions that the desired signal is
absent most of the time (as in a pulsed radar or sonar system) and the direction of arrival
of the desired signal is known. Because the Howells–Applebaum processor is practical
to implement, it has been applied extensively to the problem of clutter and interference
rejection in radar systems [18–21]. Unless otherwise noted, the analysis in this section
follows the treatment of this subject given by Gabriel [22].
A six-element linear array with Howells–Applebaum loops at each element is shown
in Figure 4-11. This figure shows the close relationship between an LMS loop and a

Sensor FIGURE 4-11

elements Six-element linear
adaptive array
having six
x1 x6 Howells–Applebaum
x2 x5 control loops with
x3 x4 beam steering
signals. From
Gabriel, Proc. IEEE,
Σ February 1976.

+ + + + + +
_ _ _ _ _ _
b*1 b*2 b*3 b*4 b*5 b*6

Filter Filter Filter Filter Filter Filter

G G G G G G

x*1 x*2 x*3 x*4 x*5 x*6

6
Array output = Σ wi xi
i=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 172

172 CHAPTER 4 Gradient-Based Algorithms

Howells–Applebaum loop. The beam steering vector, b∗ , in the Howells–Applebaum

loop plays the same role as the cross-correlation vector, rxd , in the LMS loop. Define an
element signal vector x in which the kth component, xk , consists of the quiescent receiver
channel noise voltage n k and a summation of voltage terms associated with I external,
narrowband interference sources:
xT = [x1 , x2 , . . . , x N ] (4.84)
where

I
xk = n k + Ji e jψi (2k−N −1)/2 (4.85)
i=1
and
2π d
ψi = sin θi (4.86)
λ
The interference sources are assumed to be statistically independent where Ji is the element
channel voltage amplitude associated with the ith source at θi .
Beam steering signals steer the receive beam to θs . For quiescent conditions (when
only receiver noise is present), the adaptive weights settle to steady-state values denoted
by the quiescent weight vector wq , where

wqT = [w q1 , w q2 , . . . , w q N ] (4.87)

and

w qk = ak e− jψ0 (2k−N −1)/2 (4.88)

πd
ψ0 = sin φ0 (4.89)
λ
where ak are the element amplitude weights. The resulting quiescent array factor is
expressed as

N
AFq (θ ) = (sT wq ) = ak exp[ j (ψ − ψs )(2k − N − 1)/2] (4.90)
k=1

where

sT = [s1 , s2 , . . . , s N ] = element signals (4.91)

sk = exp[ jψ(2k − N − 1)/2] (4.92)
2π d
ψ= sin θ (4.93)
λ
The components of the input beam steering vector b∗
 
b∗T = b1∗ , b2∗ , . . . , b∗N (4.94)

are directly related to the components of wq by the relation

bk∗ = ck w qk (4.95)

where the constants ck are evaluated in the section immediately following.

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 173

4.3 The Howells–Applebaum Adaptive Processor 173

4.3.1 Adaptive Weight Equations

The weight w k associated with the kth sensor element is given by
w k = bk∗ − z k (4.96)
where z k = output voltage from the kth integrating RC filter. Each correlation mixer
voltage is given by the product of the signal vk∗ with the summed array output

N
vk = k 2 xk∗ w i xi (4.97)
i=1

where
dz k 
N
τ0 + zk = γ xk∗ w i xi (4.98)
dt i=1

γ = k2G (4.99)

The constant γ represents a conversion-factor gain constant that is assumed to be the same
for all the loops. It is convenient to use (4.96) to convert from z k to w k , so that (4.98) now
becomes

dw k N
∗ ∗
τ0 + w k = bk − γ xk w i xi (4.100)
dt i=1

Using matrix notation, we may write the complete set of N differential equations corre-
sponding to (4.100) as
dw  
τ0 + w = b∗ − γ x∗ wT x (4.101)
dt
N
Since (wT x) = (xT w) = i=1 w i xi , the bracketed term in (4.101) can be rewritten as
[x∗ wT x] = [x∗ xT ]w (4.102)
The expected (averaged) value of x∗ xT yields the input signal correlation matrix

Rx x = E{x∗ xT } (4.103)
The averaged values of the correlation components forming the elements of Rx x are
given by
⎧
⎪
⎪ I
⎪
⎨ |J i |2 exp[ jψi (l − k)] l =  k (4.104)
∗
xk xl = i=1  I
⎪
⎪
⎩|x k | = |n k | + |J i |2 l = k
2 2
⎪ (4.105)
i=1

Since the correlation matrix in the absence of the desired signal is the sum of the quies-
cent receiver noise matrix Rnnq and the individual interference source matrixes Rnni , it
follows that

I
Rnn = Rnnq + Rnni (4.106)
i=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 174

174 CHAPTER 4 Gradient-Based Algorithms

where Rnnq can be expressed as

⎡ ⎤
|n 1 |2 0 0 ···
⎢ 0 |n |2 0 ··· ⎥
⎢ 2 ⎥
Rnnq =⎢
⎢ ..
⎥
⎥ (4.107)
⎣ . ⎦
0 ··· · · · |n N | 2

and
⎡ ⎤
1 e jψi e j2ψi · · ·
⎢ e− jψi 1 e jψi · · · ⎥
⎢ ⎥
⎢ − j2ψi ⎥
Rnni = |J i |2 ⎢e e− jψi 1 · · · ⎥ (4.108)
⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
1
Substituting Rnn of (4.106) into (4.101) and rearranging terms, the final expression for the
adaptive weight matrix differential equation becomes
dw
τ0 + [I + γ Rnn ]w = b∗ (4.109)
dt
where I is the identity matrix.
In general, Rnn is not diagonal, so multiplying Rnn by a nonsingular orthonormal
model matrix, Q, results in a simple transformation of coordinates that diagonalizes Rnn .
The resulting diagonalized matrix has diagonal elements that are the eigenvalues of the
matrix Rnn . The eigenvalues of Rnn are given by the solutions of the equation

|Rnn − λi I| = 0, i = 1, 2, . . . , N (4.110)

Corresponding to each eigenvalue there is an associated eigenvector ei that satisfies

Rnn ei = λi ei (4.111)

These eigenvectors (which are normalized to unit length and are orthogonal to one another)
make up the rows of the transformation matrix Q, that is,
⎡ ⎤
e11 e12 e13 · · · ⎡ ⎤
⎢ e21 e22 e23 · · ·⎥ ei1
⎢ ⎥ ⎢ ei2 ⎥
⎢ ⎥ ⎢ ⎥
Q = ⎢ e31 e32 e33 · · ·⎥ , where ei = ⎢ . ⎥ (4.112)
⎢ . ⎥ ⎣ .. ⎦
⎣ .. ⎦
ei N
eN 1 eN 2 eN 3 · · ·
Once Rnn is diagonalized by the Q-matrix transformation, there results
⎡ ⎤
λ1 0 0
⎢0 λ 0⎥
⎢ 2 ⎥
⎢ ⎥
∗
[Q Rnn Q ] = ⎢
T ⎢ 0 0 λ 0 ⎥ (4.113)
3 ⎥
⎢ .. .. ⎥
⎣ . . ⎦
··· ··· · λN
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 175

4.3 The Howells–Applebaum Adaptive Processor 175

Now since Rnn = E{x∗ xT }, it follows that (4.113) may be written as

[Q∗ Rnn QT ] = [Q∗ x∗ xT QT ] = [x∗ xT ] =  (4.114)
where
x = Qx (4.115)
The Q matrix transforms the real signal vector x into the orthonormal signal vector x .
Furthermore, the components of x are determined by the eigenvectors of Rnn , that is,
 
xk = ekT x (4.116)
Note that the orthonormal signal vector components xk have two special characteristics:
1. They are uncorrelated so
 
E xk∗ xl = 0 for l = k (4.117)
2. Their amplitudes are given by the square root of the corresponding eigenvalue so that
 
E xk∗ xk = λk (4.118)
The transformation matrix Q yields the same signal components as an appropriately
selected orthogonal beamforming network.
Just as the signal vector x was transformed into x by (4.115), the beam steering vector
b may likewise be transformed to define a new beam steering vector b∗ as
∗

b = Qb (4.119)
where the kth component of b is determined by the kth eigenvector appearing in Q.
The Q-coordinate transformation operating on both x and b∗ suggests an equivalent
circuit representation for the system that is illustrated in Figure 4-12b, where an equivalent
“orthonormal adaptive array” system is shown alongside a simplified representation of the
real system in Figure 4-12a. There are a set of weights forming the weight vector w in the
orthonormal system, and the adaptive weight matrix equation for the equivalent system is
dw  
τ0 + I + γ Rnn w = b∗ (4.120)
dt
where
Rnn = E{x∗ xT } =  (4.121)
This diagonalization results in an orthonormal system, a set of independent linear dif-
ferential equations, each of which has a solution when the eigenvalues are known. Each
of the orthonormal servo loops in the equivalent system responds independently of the
other loops, because the xk input signals are orthogonalized and are therefore completely
uncorrelated with one another. The weight equation for the kth orthonormal servo loop
can therefore be written as
dw k
τ0 + (1 + γ λk )w k = bk∗ (4.122)
dt
Note that the equivalent servo gain factor can be defined from (4.122) as

μk = γ λk (4.123)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 176

176 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-12
Equivalent circuit x1 x2 x3 x4 x5 x6
representations for a
six-element adaptive
Σ
array system.
a: Real adaptive
array system.
b: Equivalent w1 w2 w3 w4 w5 w6
orthonormal
adaptive array Amp Amp Amp Amp Amp Amp
system. From int int int int int int
Gabriel, Proc. IEEE,
b*1 b*2 b*3 b*4 b*5 b*6
February 1976.

Output
(a)

x1 x2 x3 x4 x5 x6

Q − Transformation network
x′5
x′1 x′2 x′3 x′4 x′5 x′6

w′1 w′2 w′3 w′4 w′5 w′6

Amp Amp Amp Amp Amp Amp

int int int int int int

b′1* b′2* b′3* b′4* b′5* b′6*

Output
Q* − Transformation network

b*1 b*2 b*3 b*4 b*5 b*6

(b)

so the equivalent servo gain factors for the various orthonormal loops are now determined
by the eigenvalues of the input signal covariance matrix. The positive, real eigenvalues
λk correspond to the square of a signal voltage amplitude, and any given eigenvalue is
proportional to the power appearing at the orthonormal network output port.
For the input beam steering vector b∗ , the output desired signal power is given by

Ps = |wT b|2 (4.124)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 177

4.3 The Howells–Applebaum Adaptive Processor 177

Likewise, the array output noise power can be written as

Pn = |wT x|2 (4.125)

where the signal vector x is assumed to be composed only of quiescent receiver channel
noise plus the directional noise signal components due to external sources of interference.
The signal-to-noise performance measure is therefore just a ratio of the aforementioned
two quadratic forms
s |wT b|2 w† [b∗ bT ]w
= = (4.126)
n |wT x|2 w† Rnn w
The optimum weight vector (see Chapter 3) that yields the maximum SNR for (4.126) is
1
wopt = R−1 b∗ (4.127)
(constant) nn
On comparing (4.127) with (4.45), it is seen that both the LMS and maximum SNR
algorithms yield precisely the same weight vector solution (to within a multiplicative con-
stant when the desired signal is absent) provided that rxd = b∗ , since these two vectors
play exactly the same role in determining the optimum weight vector solution. Conse-
quently, adopting a specific vector rxd for the LMS algorithm is equivalent to selecting b∗
for the maximum SNR algorithm, which represents direction of arrival information—this
provides the relation between a reference signal and a beam steering signal for the LMS
and maximum SNR algorithms to yield equivalent solutions.
From the foregoing discussion, it follows that the optimum orthonormal weight is
1
w k opt = bk∗ (4.128)
μk
Substitute (4.123) and (4.128) into (4.122) results in
dw k
τ0 + (1 + μk )w k = μk w k opt (4.129)
dt
For a step-function change in the input signal the solution may be written as follows:
 
w k (t) = w k (0) − w k (∞) e−αk t + w k (∞) (4.130)

where
μk
w k (∞) = w k opt (4.131)
1 + μk
1 + μk
αk = (4.132)
τ0
In the foregoing equations w k (∞) represents the steady-state weight, w k (0) is the initial
weight value, and αk is the transient decay factor. The adaptive weight transient responses
can now be determined by the eigenvalues. The kth orthonormal servo loop may be
represented by the simple type-0 position servo illustrated in Figure 4-13.
To relate the orthonormal system weights w k to the actual weights w k note that the
two systems shown in Figure 4-12 must be exactly equivalent so that

wT x = wT x = wT Qx (4.133)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 178

178 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-13
Type-O servo model
for kth orthonormal
adaptive control _
+ R
loop. From Gabriel, ′
wkopt mk w′k
Proc. IEEE, February
1976.
mk = glk
C

Consequently

w = QT w (4.134)

From (4.134) it follows that the solution for the kth actual weight can be written as
 
w k = e1k w 1 + e2k w 2 + · · · + e Nk w N (4.135)

where en k is the kth element of the nth eigenvector.

In the quiescent state, only receiver noise is present in any channel so that various
channel signals are uncorrelated, and Q is an identity matrix, and there is no difference
between w and w. With Qq = I, the quiescent noise covariance matrix Rnnq is already
diagonalized, and if it is further assumed that the receiver noise power in all channels is
equal (and denoted by |n 0 |2 ), then from (4.113) it follows that

Qq∗ Rnnq QqT = [λ0 δi j ] (4.136)

where

λ0 = |n 0 |2 (4.137)

so the smallest eigenvalue is simply equal to the receiver channel noise power. This smallest
eigenvalue then defines the minimum servo gain factor μmin as

μmin = γ λ0 (4.138)

Since the quiescent steady-state weight w(∞) must by definition be equal to wq , (4.131),
(4.132), and (4.95) can be applied to yield
1 ck
w qk = b∗ = w qk
1 + μmin k 1 + μmin
or

ck = (1 + μmin ) (4.139)

From (4.130)–(4.131) and (4.123), it follows that the effective time constant with
which the kth component of w converges to its optimum value is τ0 /(1 + γ λk ). In effect,
λmin determines how rapidly the adaptive array follows changes in the noise environment.
Equation (4.135) shows that each actual weight can be expressed as a weighted sum of
exponentials, and the component that converges most slowly is the λmin component.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 179

4.3 The Howells–Applebaum Adaptive Processor 179

When the adaptive array in Figure 4-11 operates with a distributed external noise field,
the loop convergence is very slow for some angular noise distributions [23]. Furthermore,
if γ is increased or τ0 is decreased to speed the weight convergence, the loop becomes
“noisy.” Slow weight convergence occurs whenever trace(Rnn )/λmin is large, and in these
cases there is no choice of γ and τ0 that yields rapid convergence without excessive
loop noise. These facts suggest that the effects of noise on the solutions represented by
(4.128)–(4.132) are important.
Griffiths provides a discrete form of the Howells–Applebaum weight update formula
given by [24]
 
w(k + 1) = w(k) + γ μb∗ − x∗ (k)x† (k)w(k) (4.140)

where γ and μ are constants. The weights converge if γ is less than one over the largest
eigenvalue. Compton shows that this is equivalent to [25]
1
0<γ < (4.141)
PIN
where PIN is the total received power in (4.48). If γ is close to 1/PIN then convergence is
fast, but weight jitter is large. The weight jitter causes SNR fluctuations of several dB at
steady state. If γ is small, then weight jitter is small, but the convergence is slow. A gain
constant of [26]
1
γ = (4.142)
2.5PIN
was found to provide a reasonably stable steady-state weights and rapid conversion.

4.3.2 Loop Noise Considerations

The random variations in the adaptive element weights of a Howells–Applebaum control
loop result in an additional noise component in the array output signal. In this section
expressions are given for the variance of the element weights and for the resulting additional
noise in the array output [23].
Let ξ denote the noise component of the adaptive weight vector w, and let  denote
the random component of Rnn , so that

w=w+ξ (4.143)
Rnn = Rnn +  (4.144)

where now w and Rnn denote average values. The adaptive weights must satisfy
dw
τ0 + (I + γ Rnn )w = b∗ (4.145)
dt
Substitute the values w and Rnn into (4.145) and subtract the result from the equation
resulting with (4.143) and (4.144) substituted into (4.145) to give
dξ  
τ0 + I + γ Rnn ξ = −γ w (4.146)
dt
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 180

180 CHAPTER 4 Gradient-Based Algorithms

Premultiplying (4.146) by the transformation matrix Q∗ and using the fact that Q∗ QT = I,
then
dζ 1
+ (I + γ )ζ = −βQ∗ w = u (4.147)
dt τ0
where

ζ = Q∗ ξ (4.148)
γ
β= (4.149)
τ0
Equation (4.147) represents a system of N independent linear differential equations of
which the nth component can be written as
dζn + σn ζn dt = u n dt (4.150)
where
1 + γ λn
σn = (4.151)
τ0
u n = u n (τ, w, ζ ) = (−βQ∗ w)n (4.152)

Multiplying (4.150) by the factor eσn t and integrating each term from t0 to t then yields
t
ζn (t) = ζn (t0 ) exp[−σn (t − t0 )] + t0 e−σn (r −τ )
(4.153)
· u n (τ, w, ζ )dτ
If only the steady-state case is considered, then the weights are near their mean steady-
state values. The steady-state solution for variations in the element weights can be obtained
from (4.153) by setting t0 = −∞ and ignoring any effect of the initial value ζn (t0 ) to give
 ∞
ζn (t) = e−σn τ u n (t − τ )dτ (4.154)
0
One important measure of the noise present in the adaptive loops is the variance of
the weight vector denoted by var(w):
 N 

var(w) = E |wn − wn | = E{ξ † ξ }
2
(4.155)
n=1

where N is the dimension of the weight vector (or the number of degrees of freedom in
the adaptive array system). Now since ζ = Q∗ ξ , (4.155) becomes
var(w) = E{ζ † ζ } (4.156)
The elements of the covariance matrix of ζ (t) in (4.156) are obtained from (4.154) and
the definition of u n
 ∞  ∞
∗

E{ζ j ζk } = β 2
dτ1 E [Q∗ (t − τ1 )w]∗j
0 0
· exp(−σ j τ1 − σk τ2 )[Q ∗ (t − τ2 )w]k } dτ2
 ∞  ∞

+β 2
dτ1 E [Q∗ (t − τ1 )ξ (t − τ1 )]∗j (4.157)
0 0
· exp(−σ j τ1 − σk τ2 )[Q∗ (t − τ2 )ξ (t − τ2 )]k } dτ2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 181

4.3 The Howells–Applebaum Adaptive Processor 181

where the cross-product terms do not appear since E{ξ (t)} = 0 and (t) and ξ (t) are
independent noise processes. A useful lower bound for (5.157) is given by [23].

β 2  1  ∗
N

var(w) ≥ E |(Q w)n |2 (4.158)
2 n=1 σn

where  represents the time interval between successive independent samples of the input
signal vector. For a pulse radar,  is approximately the same as the pulse width. For a
communications system,  is approximately 1/B, where B is the signal bandwidth.
The bound in (4.158) is useful in selecting parameter values for the Howells–Applebaum
servo loops. If this bound is not small, then the noise fluctuations at the output of the adap-
tive loops are correspondingly large. For cases of practical interest [when var(w) is small
compared with w† w], the right-hand side of (4.158) is an accurate estimate of var(w).
Equation (4.158) simplifies (after considerable effort) to yield the expression

β β 
N
1
var(w) ≥ − w† Rnn w (4.159)
2 2γ n=1 λn + 1/γ

where λn represents the nth eigenvalue of Rnn

N 1
= − trace(μ) (4.160)
β γ
μ = (βH)−1 (4.161)
I
H =  nn (4.162)
γ
Since the total output noise power is the noise power without noisy weights w† Rnn w
plus the additional noise due to the random weight components, it can be shown that the
total output noise power is given by
   
E |wT x|2 = w† Rnn w + E ξ † Rnn ξ

β 
N
∼ †
= w Rnn w 1 + λn (4.163)
2 n=1

when γ λn 1 for n = 1, 2, . . . , N . The quantity β occurs both in (4.163) and in

(4.159) is the ratio γ /τ0 , which is the gain divided by the loop time constant where the
time constant is measured in intervals of the independent-sample rate of the system.
When loop noise is present in the system, the total noise power output increases by
the factor (1 + K n ), where from (4.163)

γ 
N
γ
Kn ≥ λn = trace(Rnn ) (4.164)
2τ0 n=1 4Bτ0

where  = 1/2B (i.e., B is the bandwidth of the input signal process), so that K n is a
direct measure of algorithm misadjustment due to noise in the weight vector. Recalling the
solution to (4.129), we see that the effective time constant of the normal weight component
w k  having the slowest convergence rate is
τ0 ∼ τ0
τeff = = (4.165)
1 + γ λmin γ λmin
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 182

182 CHAPTER 4 Gradient-Based Algorithms

where γ λmin ≥ 1 to avoid a steady-state bias error in the solution. On combining (4.164)
and (4.165) there results
τeff 1 N
trace(Rnn )
≥ λn = (4.166)
 2K n λmin n=1 2K n λmin
Equation (4.166) shows that, when the smallest eigenvalue λmin is small compared with
trace(Rnn ), many independent samples of the input signal are required before the adaptive
array settles to a near-optimum set of weights without excessive loop noise; no set of loop
parameters yields both low loop noise and rapid convergence in this case. Berni [27] gives
an analysis of steady-state weight jitter in Howells–Applebaum control loops when there
is no statistical independence between the input signal and weight processes. Steady-state
weight jitter is closely related to the statistical dependence between the weight and signal
processes.

4.3.3 Adaptive Array Behavior in Terms of Eigenvector Beams

The Q-matrix transformation defined by (4.112) is composed of normalized and mutu-
ally orthogonal eigenvectors. The components of these eigenvectors are interpreted as
array element weights, associated with normalized orthogonal eigenvector beams. The
kth eigenvector beam is expressed as

N
gk (θ ) = (sT ek ) = eki si (4.167)
i=1

where s and its components si for a linear N -element array are defined by

sT = [s2 , s2 , . . . , s N ] (4.168)
si = e jψ(2i−N −1)/2 (4.169)

By defining the variable z related to the spatial angle θ as


z = e jψ (4.170)
then the eigenvector beam may be conveniently rewritten as
N −1  
1
gk (θ) = √ ek1 + ek2 z + ek3 z 2 + · · · + ek N z N −1 (4.171)
z
As shown in Chapter 2, the array polynomial can also be expressed in the factored form
N −1
1
gk (θ) = √ [a N −1 (z − z 1 )(z − z 2 ) · · · (z − z N −1 )] (4.172)
z
where the roots z 1 , z 2 , . . . , z N −1 are nulls in the eigenvector beam pattern.
If one narrowband interference source is at θ1 , then Rnn contains one unique eigenvalue
and the corresponding unique eigenvector that produces a retrodirective eigenvector beam
centered on the source at θ1 as illustrated in Figure 4-14. The Rnn matrix in this case also
contains nonunique eigenvalues having arbitrary nonunique eigenvectors; these arbitrary
nonunique eigenvector beams are not essential to array operation, and array pattern per-
formance is characterized solely in terms of the unique retrodirective eigenvector beams.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 183

4.3 The Howells–Applebaum Adaptive Processor 183

Unadapted steered–beam FIGURE 4-14

pattern Retrodirective beam
principle illustrating
subtraction of
retrodirective beam
from unadapted
pattern to obtain
q adapted pattern with
q0 one interference
source. From
Retrodirective beam Gabriel, Proc. IEEE,
February 1976.
q

Resulting adapted pattern

q
q0 Null
qJ
Jammer
azimuth

The overall array beam pattern is most easily derived by considering the output of the
orthonormal system represented in Figure 4-12b for the input signal vector s, defined in
(4.169). Since the output for the real orthonormal systems are identical, it follows that

N 
N
AF(θ, t) = w i si = w i si = w T s (4.173)
i=1 i=1
where
s = Qs (4.174)

Now the ith component of s is given by
   N
si = eiT s = eik sk (4.175)
k=1
but this summation defines the ith eigenvector beam [as can be seen from (4.167)], so that
 
si = eiT s = gi (θ ) (4.176)
Consequently, the overall array factor can be expressed as

N
AF(θ, t) = w i gi (θ ) (4.177)
i=1
which shows that the output array factor is the summation of the N eigenvector beams
weighted by the orthonormal system adaptive weights.
Since the kth component of the quiescent orthonormal weight vector is given by
 
w q k = e†k wq (4.178)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 184

184 CHAPTER 4 Gradient-Based Algorithms

the steady-state solution for the kth component of the orthonormal weight vector given by
(4.131) can be rewritten using (4.132), (4.95), and (4.139) to yield
1 + μmin
w k (∞) = w q k (4.179)
1 + μk
Assume as before that quiescent signal conditions up to time t = 0 consist only of
receiver noise and that the external interference sources are switched on at t = 0; then

w k (0) = w q k (4.180)

and the solution for w k expressed by (4.131) is rewritten in the more convenient form
 !
  −αk t μk − μmin
w k = w qk − (1 − e ) w q k (4.181)
1 + μk
It is immediately apparent that at time t = 0 (4.177) results in

N
   
AF(θ, 0) = w q i gi (θ ) = wqT s = wqT Qs (4.182)
i=1

From (4.134) it is seen that wqT = wqT Q so that

 
AF(θ, 0) = wqT s = AFq (θ ) (4.183)
where the quiescent pattern AF q (θ ) was previously defined by (4.90).
Finally, by substituting (4.183) and (4.181) into (4.177), there results

N  !
−αi t μi − μmin
AF(θ, t) = AF q (θ ) − (1 − e ) w q i gi (θ ) (4.184)
i=1
1 + μi

where it will be recalled that

1 + μi
αi =
τ0
μi = γ λi
μmin = γ λmin

The foregoing result emphasizes that the adaptive array factor consists of two parts:
1. The quiescent beam pattern AFq (θ )
2. The summation of weighted orthogonal eigenvector beams that is subtracted from
AFq (θ)
Note also from (4.184) that the weighting associated with any eigenvector beams corre-
sponding to eigenvalues equal to λ0 (the quiescent eigenvalue) is zero since the numerator
(μi − μmin ) is zero for such eigenvalues. Consequently, any eigenvector beams associated
with λ0 is disregarded, leaving only unique eigenvector beams to influence the resulting
pattern. The transient response time of (4.184) is determined by the value of αi , which in
turn is proportional to the eigenvalue. Therefore, a large eigenvalue yields a fast transient
response for its associated eigenvector beam, whereas a small eigenvalue results in a slow
transient response.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 185

4.3 The Howells–Applebaum Adaptive Processor 185

FIGURE 4-15 Two

eigenvector beam
−10 Gq (q ) patterns g1 (θ ) and
g2 (θ) for
−15 two-jammer
g′1 (q )
Output power (dB)
example. From
−20 Gabriel, Proc. IEEE,
February 1976.
−25 g′2 (q )

−30

−35

−40
−10° 0° 10° 20° 30° 40° 50°
Azimuth angle
Jammer
positions

The foregoing eigenvector beam interpretation of adaptive array behavior is illus-

trated by considering an eight-element linear array having λ/4 element spacing and two
narrowband interference sources having nearly equal power ratios of P R1 = 1,250 and
P R2 = 1,200 located at θ1 = 18◦ and θ2 = 22◦ , respectively. Forming the covariance ma-
trix using (4.106) for this case and solving for the eigenvalues yields two unique solutions:
λ1 = 18,544.4 and λ2 = 1,057.58. These widely different eigenvalues result despite the
nearly equal jammer powers, because the interference sources are close together compared
with the array quiescent beamwidth.
Solving for the two (normalized) eigenvectors associated with the unique eigenvalues
then permits the two eigenvector beam patterns to be found g1 (θ ) and g2 (θ ), which are
both illustrated in Figure 4-15. Beam g1 (θ ) covers both interference sources in the same
manner as a centered beam pattern, and its total output power is equal to the first eigenvalue
λ1
= 1 + P1 g1 (θ1 ) + P2 g2 (θ2 ) = 18,544
2 2
(4.185)
λ0
The second eigenvector beam g2 (θ ) splits the interference sources in the manner of a
difference beam, and its total output power is equal to the second eigenvalue
λ2
= 1 + P1 g2 (θ1 ) + P2 g2 (θ2 ) = 1057
2 2
(4.186)
λ0
Although both eigenvector beams contain power from both sources, their respective
output signals are decorrelated. The cross-correlation product of the two eigenvector beam
outputs is

E{x1∗ x2 } = E{|J1 |2 }g1 (θ1 )g2 (θ1 ) + E{|J2 |2 }g1 (θ2 )g2 (θ2 ) (4.187)

This cross-correlation product can be zero if the product [g1 (θ )g2 (θ )] is positive when
θ = θ1 and negative when θ = θ2 , thereby resulting in decorrelation between the two
eigenvector beam signals. Figure 4-16 shows the overall quiescent beam pattern and the
resulting steady-state adapted pattern for this two-source example. Figure 4-17 illustrates
the transient response (in terms of increase in output noise power) of the adaptive array for
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 186

186 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-16 0
Steady-state
adapted array −5
pattern and

Output power (dB)

quiescent array Quiescent
−10 pattern
pattern for two- Adapted
jammer example. pattern
−15
From Gabriel, Proc.
IEEE, February 1976. −20

−25

−30
−90° −60° −30° 0° 30° 60° 90°
Azimuth angle
Jammer
positions

FIGURE 4-17
Increase in output noise power (dB)

Array transient Time constant a 1

response for 25
two-jammer
example. From 20
Gabriel, Proc. IEEE,
February 1976. 15

10 Time constant a 2

0
0 2 4 6 8 10 12 14 16 18 20 22
Time (msec)

this two-interference source example, where it is seen that the response has two distinct
slopes associated with the two distinct (and widely different) eigenvalues.

4.3.4 Example of N-Element Linear Adaptive Array Performance

This section presents results from an eight-element linear array with Howells–Applebaum
loops.

[Link] Assumptions and Initial Conditions

Only the quiescent receiver noise is present in each channel up to time t = 0, when all the
external interference sources are switched on in a single step function. The quiescent RMS
noise voltage in all channels equals the constant n 0 , such that the quiescent eigenvalue
λ0 is |n 0 |2 . This defines the quiescent servo gain factor μmin by way of (4.138). For the
configuration of Figure 4-11, it is convenient to choose μmin = 1, so the amplifier gains G
are set accordingly. Once μmin is selected, it is convenient to express μi from (4.123) as
a ratio of eigenvalues,
μi λi
= (4.188)
μmin λ0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 187

4.3 The Howells–Applebaum Adaptive Processor 187

The quiescent steered-beam pattern AFq (θ ) and its associated quiescent weight vector
wq are given by (4.87)–(4.90). The eight-element linear array has an element spacing
λ/2, μ = π/2 sin θ, and ak = 1. The quiescent weights and array factor are given by
w qk = e− jψ0 (2k−9)/2 (4.189)
sin [8(ψ − ψ0 )/2]
AFq (θ ) = (4.190)
sin [(ψ − ψ0 )/2]
The coefficients of the input beam steering vector b∗ are found from (4.140) and (4.88)

ck = (1 + μmin ) = 2 (4.191)
bk∗ = ck w qk = 2e− jψ0 (2k−9)/2 (4.192)

The maximum power condition for each of the orthonormal loops of Figure 4-12b is
λmax π Bc τ0
μmax = μmin = −1 (4.193)
λ0 10
where λmax represents the maximum eigenvalue. The channel bandwidth Bc and filter time
constant τ0 are the same for all element channel servo loops. Solving for τ0 from (4.193)
yields
 ! 
10 λmax 10  R
τ0 = 1 + μmin = 1 + μmin + μmin Pr gm (θr )
2
π Bc λ0 π Bc r =1

(4.194)
The maximum power (maximum eigenvalue) is much larger than the jammer-to-receiver-
noise power ratios, because the various Pr are multiplied by the power gain of the eigen-
vector beams.

[Link] Output Noise Power and SNR Degradation

The output SNR of the adaptive array compared with the output SNR of a conventional
array under the same interference conditions is the performance characteristic of ultimate
interest. Instead of forming the actual SNR, it is sufficient to consider the output noise
power by itself to illustrate the system transient behavior. Since the receiver noise and
external interference sources are statistically independent, the total output noise power is
the sum of the two separate output noise powers.
The receiver noise output power can be expressed as

N 
N
|y0n (t)|2 = |w k n k |2 = |w i n 0 |2 (4.195)
k=1 i=1

Substituting for w i from (4.181) then yields


k
|y0n (t)|2 = |n 0 |2 [1 − Ai (t)]2 |w q i |2 (4.196)
i=1

where
 !
μi − μ0
Ai (t) = (1 − e−αi t ) (4.197)
1 + μi
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 188

188 CHAPTER 4 Gradient-Based Algorithms

From (4.197) it is seen that Ai (t) is zero for t = 0 and for μi = μ0 (for nonunique
eigenvalues). Therefore, for quiescent conditions at t = 0, it follows that


N 
N
|y0n (0)|2 = |n 0 |2 |w q i |2 = |n 0 |2 |w qk |2 (4.198)
i=1 k=1

since the output noise power must be the same for either the real system or the equivalent
orthonormal system. Consequently, (4.196) is rewritten as


N 
N
|y0n (t)|2 = |n 0 |2 |w qk |2 − [2 − Ai (t)]Ai (t)|w q i |2 (4.199)
k=1 i=1

Equation (4.199) is a particularly convenient form because the w q i associated with nonunique
eigenvalues need not be evaluated since A(t) = 0 for such eigenvalues.
The output noise power contributed by R external interference sources is given by the
sum of their output power pattern levels:


R
|y0 j (t)|2 = |n 0 |2 Pr AF 2 (θr , t) (4.200)
r =1

where Pr is the r th source power ratio, θr is its angular location, and AF(θr , t) is given
by (4.184).
The total output noise power is the sum of (4.199) and (4.200), and the increase in the
output noise power (with interference sources turned on) is this sum over the quiescent
noise (4.198).
⎧ ⎫
⎪
⎪  R 
N ⎪
⎪
⎨ Pr AF (θr , t) − [2 − Ai (t)]Ai (t)|w qi | ⎪
2  2 ⎪
⎬
|y0 (t)|2 r =1 i=1
= 1 + (4.201)
|y0n (0)|2 ⎪
⎪ N ⎪
⎪
⎪
⎩ |w qk | 2 ⎪
⎭
k=1

The output noise power increase in (4.201) indicates the system transient behavior. An
increase in output noise power indicates the general magnitude of the adapted (steady-
state) weights.
The degradation in the SNR, Dsn , enables one to normalize the effect of adapted-
weight magnitude level. This degradation is the quiescent SNR divided by the adapted
SNR.

AFq2 (θs ) |y0 (t)|2

Dsn = (4.202)
AF 2 (θs , t) |y0n (0)|2

where the ratio in the second factor is just (4.201), the increase in output noise power.

[Link] Eigenvalues and Eigenvectors of the Noise Covariance Matrix

For computational convenience, receiver noise is unity and all noise powers expressed as
ratios to receiver noise power. Adopting this convention, the quiescent noise matrix Rnnq
is an identity matrix, and with R narrowband interference sources the noise covariance
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 189

4.3 The Howells–Applebaum Adaptive Processor 189

Power FIGURE 4-18

Interference source
Interference bandwidth
power spectrum
having uniform
Δ fm amplitude spectral
∋
mth Spectral line
lines spaced ε apart.
From Gabriel, Proc.
IEEE, February 1976.

Frequency

f0

matrix becomes

R
Rnn = I + Pr Mr (4.203)
r =1

where Mr now represents the covariance matrix due to the r th interference source.
Wideband interference sources are represented by dividing the jammer power spec-
trum into a series of discrete spectral lines. A uniform amplitude spectrum of uncorrelated
lines spaced apart by a constant frequency increment ε is once again assumed as illus-
trated in Figure 4-18. If Pr is the power ratio of the entire jammer power spectrum, then
the power ratio of a single spectral line (assuming a total of L r spectral lines) is
Pr
Prl = (4.204)
Lr
Furthermore, if Br (Br < element channel receiver bandwidth, Bc ) denotes the percent
bandwidth of the jamming spectrum, then the frequency offset of the lth spectral line is
 !
 fl Br 1 l −1
= − + (4.205)
f0 100 2 Lr − 1
The covariance matrix with R broadband interference sources is written as

R 
Lr
Rnn = I + Prl Mrl (4.206)
r =1 l=1

The mnth component (mth row and nth column) of the matrix Mrl is in turn given by

(Mrl )mn = e jψrl (n−m) (4.207)

where
fl  fl
ψrl = π sin θr = 1 + π sin θr (4.208)
f0 f0

[Link] Performance Characteristics for Various Signal Conditions

Four narrowband sources located in the sidelobe region of the quiescent beam pattern yield
four distinct eigenvalues and require four degrees of freedom to provide the eigenvector
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 190

190 CHAPTER 4 Gradient-Based Algorithms

beams required to place nulls at the jammer locations. If the adaptive weight adjustments
are large, there may be appreciable main beam distortion in the overall adapted pattern.
The Howells–Applebaum adaptive loop has one adaptive weight in each element
channel of the array; this configuration works interference sources with a bandwidth of
up to about 20%. Gabriel [22] gives two examples as follows: a 2% bandwidth source
in the sidelobe region for which two degrees of freedom (two pattern nulls) are required
to provide proper cancellation; and a 15% bandwidth source in the sidelobe region for
which three degrees of freedom are required. Broadband interference sources require a
transversal equalizer in each element channel (instead of a single adaptive weight) for
proper compensation, with a Howells–Applebaum adaptive loop then required for every
tap appearing in the tapped delay line.
The adapted pattern for main beam nulling exhibits severe distortions. For interference
sources located in the main beam, the increase in output noise power is an unsatisfactory
indication of array performance, because there is a net SNR degradation due to the resulting
main beam distortion in the adapted pattern. Main beam constraints for such cases can be
introduced.

4.3.5 Hard Limiter Modification for N Adaptive Loops

It was shown in Section 4.3.2 that the adaptive array performance depends on the external
noise field as well as on the parameters of the adaptive control loops. The power level
and angular location of the external noise field determine the noise covariance matrix and
therefore its eigenvalues. The eigenvalues, in turn, directly affect the array performance,
since both the transient response of the adaptive array and the control loop noise depend
explicitly on these eigenvalues. For a nonstationary signal environment, wide variations in
array performance may occur, ranging from excessive control loop noise (when the inter-
ference is strong) to very slow convergence (when the interference is weak). Introducing
a hard limiter into the adaptive control loop reduces the effects of varying noise intensity,
and the dynamic range of signals in the control loops are reduced without degrading ar-
ray performance [28]. Figure 4-19 shows a six-element linear array with a hard limiter
introduced in the conjugate signal branches.
With the signal envelopes hard limited, the input to the correlation mixers changes
from xk∗ to u ∗k = xk∗ /|xk |. This normalization removes amplitude variations in the conjugate
signals but retains the phase variations. The correlation mixer voltage vk is now given by


N
vk =k 2
u ∗k w i xi (4.209)
i=1

On comparing (4.209) with (4.97), it is seen that u ∗k has simply replaced xk∗ , so the
resulting adaptive weight matrix differential equation now becomes
dw
τ0 + [I + γ  M]w = b∗ (4.210)
dt
which is analogous to (4.109) with M replacing Rnn and γ  = k 2 G  where M is the
modified noise covariance with envelope limiting having elements given by
% ∗ &
x m xl
Mml = E (4.211)
|xm |
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 191

4.4 Introduction of Main Beam Constraints 191

FIGURE 4-19
x1 x6 Hard limiter
x2 x5 modification of linear
x3 x4 six-element adaptive
array system. From
Gabriel, Proc. IEEE,
Σ February 1976.

w1 w2 w3 w4 w5 w6
+ + + + + +
_ _ _ _ _ _
b*1 b*2 b*3 b*4 b*5 b*6

Filter Filter Filter Filter Filter Filter

G G G G G G

u*1 u*2 u*3 u*4 u*5 u*6

6
Output = Σ wi xi
i=1

Assuming the quadrature components of each signal xk are zero-mean Gaussian random
variables having variance σ 2 , we can then compute the elements of the covariance matrix
M directly from the elements of Rnn by using the relation [28]
'
π 1
Mml = (Rnn )ml (4.212)
8σ
It follows
√ that the elements of M differ from the elements of Rnn by a common factor
(1/σ ) (π/8). Consequently, the effective time constants that determine the rate of con-
vergence and control loop noise are changed by this same factor, thereby reducing the
dependence of array performance on the strength of the external noise field.
It is worthwhile noting that limiting does not change the relative values of the sig-
nal covariance matrix elements or the relative eigenvalue magnitudes presuming identical
channels. Thus, for widely different eigenvalues, limiting does reduce the eigenvalue
spread to provide rapid transient response and low control loop noise. Nevertheless, limit-
ing always reduces the dynamic range of signals in the control loops, thereby simplifying
the loop implementation.

4.4 INTRODUCTION OF MAIN

BEAM CONSTRAINTS
As a result of introducing beam pattern nulls, the main beam may become distorted, result-
ing in a degradation in SNR performance. To prevent such array performance degradation
from occurring, it is possible to introduce constraints so the adaptive processor maintains
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 192

192 CHAPTER 4 Gradient-Based Algorithms

desired mainlobe signals while realizing good cancellation of interference in the sidelobes.
The constraint methods discussed here follow the development that is given by Applebaum
and Chapman [29].
Techniques for applying main beam constraints to limit severe array pattern degrada-
tion include the following:
1. Time domain: The array adapts when the desired signal is not present in the main beam.
These weights are kept until the next adaptation or sampling period. This approach does
not protect against main beam distortion resulting from main beam jamming and is
also vulnerable to blinking jammers.
2. Frequency domain: When the interference sources have much wider bandwidths than
the desired signal, the adaptive processor is constrained to adapt to signals only out-
side the desired signal bandwidth. This approach somewhat degrades the cancellation
capability and distorts the array factor.
3. Angle domain: Three angle domain techniques provide main beam constraints in the
steady state: (a) pilot signals; (b) preadaptation spatial filters; and (c) control loop
spatial filters. These techniques are also helpful for constraining the array response
to short duration signals, since they slow down the transient response to main beam
signals. The angle domain techniques provide the capability of introducing main beam
constraints into the adaptive processor response.

4.4.1 Pilot Signals

To illustrate the use of pilot signal techniques, consider the multiple sidelobe canceller
(MSLC) adaptive array configuration shown in Figure 4-20, where an integrator with
feedback structure is taken to represent the integrating filter in the Howells–Applebaum

FIGURE 4-20
Beam
Multiple sidelobe
steering
canceller (MSLC)
adaptive array f1 f2 f3 f4
configuration with
beam steering pilot x1 x2 x3 x4
signals and main Pilot
beam control. Σ ms 1 Σ ms2 Σ ms3 Σ ms4
signals

w1 w2 w3 w4

Low
∫ pass ∫ ∫ ∫
filter
Reference − − − −
pilot
+ + + +
signal
ms0
u*1 u*2 u*3 u*4
+
_ Residue, ε
Output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 193

4.4 Introduction of Main Beam Constraints 193

adaptive loop. The “pilot signals” shape the array beam and maintain the main beam gain
(avoiding SNR degradation). The pilot signals are continuous wave (CW) tones injected
into each element channel at a frequency that is easily filtered out of the signal bandwidth.
It is not necessary to use the beam steering phase shifters shown in Figure 4-20, since if
they are not present the pilot signals may be injected with the proper phase relationship
corresponding to the desired main beam direction instead of in phase with each other
as shown. The amplitudes and phases of the injected pilot signals s1 , . . . , s4 may be
represented by the vector μs, where s has unit length, and μ is a scalar amplitude factor.
The reference channel (or main beam) signal is represented by the injected pilot signal s0 .
For the adaptive control loops shown in Figure 4-20, it follows that the vector differ-
ential equation for the weight vector is written as
dw
= u∗ (t)ε(t) − w(t) (4.213)
dt
Since ε = μs0 − xT w, it follows (4.213) and the results of Section 4.3.5 that
dw
= gμrxs0 − [I + gRx x ]w (4.214)
dt
where g is a gain factor representing the correlation mixer gain and the effect of the limiter.
The steady-state solution of (4.214) is given by

wss = [I + gRx x ]−1 gμrxs0 (4.215)

In the absence of any desired signal, then

x = n + μs (4.216)

where n is the noise signal vector, and μs is the injected pilot signal vector. Consequently,

Rx x = E{x∗ xT } = Rnn + μs∗ sT (4.217)
 ∗ ∗
rxs0 = E{x s0 } = μs s0 (4.218)

On substituting (4.217) and (4.218) into (4.215) it can be shown that

K−1 s∗ gμ2 s0
wss = (4.219)
1 + gμ2 sT K−1 s∗
where K = I + gRnn . Substituting the expression for K−1 in (4.219) then yields
(I + gRnn )−1 s∗ gμ2 s0
wss = (4.220)
1 + gμ2 sT (I + gRnn )−1 s∗
For large pilot signals μ2 → ∞, and (4.220) becomes

−1 ∗
∼ (I + gR nn ) s
wss = s0 (4.221)
sT (I + gRnn )−1 s∗

If s has equal amplitude components, then the main beam response from (4.221) is

sT wss ∼
= s0 (4.222)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 194

194 CHAPTER 4 Gradient-Based Algorithms

which is a constant, independent of Rnn (and hence independent of any received wave-
forms).
The array configuration of Figure 4-20 uses one set of pilot signals for a single main
beam constraint. Multiple constraints require multiple sets of pilot signals, with each set at
a different frequency. Pilot signals are inserted close to the input of each element channel
to compensate for any amplitude and phase errors. Strong pilot signals require channel
elements with a large dynamic range, so they must be filtered to avoid interfering with the
desired signal.

4.4.2 Preadaption Spatial Filters

Preadaption spatial filtering forms two beams following the beam steering phase shifters
(Figure 4-21). The quiescent pattern main beam is formed with fixed weights s∗ (forming
a unit length weight vector). The second beam (termed a cancellation beam) is formed
adaptively by an MSLC whose input channels are obtained from spatial filtering repre-
sented by the matrix transformation A. A has one less output channel than the number of
sensor elements and is selected to maintain a constant response in the main beam direction
so that As = 0.
From the signal vector definitions of Figure 4-21 it follows that
em = s† x = xT s∗ (4.223)
e 0 = em − y T u (4.224)
u = Ax (4.225)
Therefore
e0 = s† x − yT Ax = (s∗ − AT y)T x = wT x (4.226)

FIGURE 4-21
General structure of
preadaption spatial
filtering. f1 f2 f3 f4 Beam steering

s* A As = 0

Main beam em = s†x

u = Ax
+
Cancellation beam
_ MSLC
yTu

Residue signal

e0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 195

4.4 Introduction of Main Beam Constraints 195

The composite weight vector for the entire system can therefore be written as
w = s∗ − AT y (4.227)
Consequently
sT w = sT (s∗ − AT y) = s2 − sT AT y (4.228)
Since A was selected so that As = 0, it follows that
sT w = s2 = 1 (4.229)
Denote the covariance matrix associated with u by

Ruu = E{u∗ uT } = E{A∗ x∗ xT AT } = A∗ Rx x AT (4.230)

The MSLC unit generates a weight vector y that satisfies (4.213).

dy
= gu∗ e0 − y (4.231)
dt
In the steady state [where (dy/dt) = 0] it follows that

[I + gRuu ]y = g E{u∗ em } (4.232)

where g is a gain factor, and the right side of (4.232) represents the cross-correlation vector
of em with each component of u. Using (4.223), (4.225), and (4.230) in (4.232), we find
that
 
I + gA∗ Rx x AT y = gA∗ Rx x s∗ (4.233)
Premultiply (4.233) by AT and use (4.227); it then follows that the composite weight
applied to the input signal vector satisfies the steady-state relation
 
I + gAT A∗ Rx x wss = s∗ (4.234)

when x does not contain a desired signal component, then Rx x may be replaced by Rnn .
Now allow g to become very large so that (4.233) yields

A∗ Rx x AT y = A∗ Rx x s∗ (4.235)
A∗ Rx x (s∗ − AT y) = A∗ Rx x wss = 0 (4.236)

Since As = 0 and the rank of the transformation matrix A is N − 1, (4.236) implies that
Rx x w is proportional to s∗ so that

Rx x wss = μs∗ (4.237)

wss = μR−1
xx s
∗
(4.238)

where μ is a proportionality constant that may be evaluated using (4.229). Substituting

μ = (sT R−1 ∗ −1
xx s ) in (4.238) then yields
R−1
xx s
∗
wss = (4.239)
sT R−1
xx s
∗

as the solution that the composite weight vector approaches when g becomes very large.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 196

196 CHAPTER 4 Gradient-Based Algorithms

Preadaption spatial filtering avoids dynamic range problems, so it does require the
implementation of multiple beams. The accuracy of the beam steering phase shifters limits
the effectiveness of the constraints, but this limit is true of all three methods considered
here. Two realizations of preadaption spatial filtering represented by Figure 4-21 include
[29]: (1) the use of a Butler matrix to obtain orthogonal beams, one of which is regarded
as the “main” beam; and (2) the use of an A matrix transformation obtained by fixed
element-to-element subtraction.

4.4.3 Control Loop Spatial Filters

The Howells–Applebaum adaptive control loop with constraints applied directly in the
loop by means of a spatial matrix filter is illustrated by the configuration of Figure 4-22.
The spatial matrix filter removes any components of the signal vector v pointing in the
direction of the unit length beam steering vector b∗ by means of a projection operator.
The successful removal of such signal components then constrains the array response in
the direction of b.
The amplified output of the correlation mixer configuration of Figure 4-22 is given by
v(t) = u∗ (t)xT (t)w(t) (4.240)
Taking expected values of (4.243) results in the steady-state values given by
v = gRx x wss (4.241)

FIGURE 4-22
Adaptive processor
with control loop x1 x2 xk xk + 1 xN
spatial filtering.

wk

+
b*k
_
zk

Spatial
∧
matrix P
filter

vk

u*k

wTx
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 197

4.4 Introduction of Main Beam Constraints 197

where g is a gain factor. Now

z = P̂v = g P̂Rx x wss (4.242)

Substituting w = b∗ − z into (4.245) means the steady-state weight values must satisfy

(I + g P̂Rx x )wss = b∗ (4.243)

When the beam steering vector is uniformly weighted, the projection performed by
the spatial filter to remove signal components in the direction of b is

P̂ = I − b∗ bT (4.244)

Substituting (4.244) into (4.243) results in

(I + gRx x − gb∗ bT Rx x )wss = b∗ (4.245)

[(I + gRx x )R−1
xx
∗ T
− gb b ]Rx x wss = b ∗
(4.246)

Rewrite (4.246) in the form

Rx x wss = [Q − gb∗ bT ]−1 b∗ (4.247)


where Q = (I + gRx x )R−1
x x , and apply a matrix inversion identity to obtain the result

Rx x (I + gRx x )−1 b∗
[Q − gb∗ bT ]−1 b∗ = (4.248)
1 − gbT Rx x (I + gRx x )−1 b∗
The denominator of (4.243) may be simplified as

gbT Rx x (I + gRx x )−1 b∗ = bT (I + gRx x − I)(I + gRx x )−1 b∗

= bT b∗ − bT (I + gRx x )−1 b∗ (4.249)
−1 ∗
= 1 − b (I + gRx x ) b
T

Substituting (4.248) and (4.249) into (4.247) then yields

(I + gRx x )−1 b∗
wss = (4.250)
bT (I + gRx x )−1 b∗

The introduction of P̂ = I − b∗ bT constrains the array response only in the direction

of b. Additional constraints are required to constrain the response over a finite region of
the main lobe instead of only one direction. Likewise, when the steering vector is not
uniform, an additional constraint is also required to obtain the desired quiescent pattern.
It is furthermore desirable to transform the constraints to an orthogonal set, thereby min-
imizing the accuracy requirements of the spatial matrix filter. To illustrate a projection
filter constructed from an orthogonal set, consider the representation of a projection filter
with M + 1 orthogonal constraints:

M
P̂ = I − cm c†m (4.251)
m=0

where m n = δmn , δmn = the Kronecker delta, and the m are the constraint vectors.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 198

198 CHAPTER 4 Gradient-Based Algorithms

The constraint that maintains the array response at the peak of the beam is the “zero-
order” constraint. The weight vector solution obtained with a zero-order constraint differs
from the unconstrained solution (P̂ = I) by a multiplicative scale factor. Multiple con-
straints are typically used to increase the beam constraint zone by controlling the first few
derivatives of the pattern function in the direction of interest. A constraint that controls
the mth derivative is referred to as an “mth-order” constraint.
To synthesize a m constraint vector corresponding to the mth derivative of the pattern
function, note that the pattern function of a linear array can be written as


N
AF(θ ) = w k e jkθ (4.252)
k=1

The mth derivative of AF(θ ) is


N
AF m (θ ) = ( jk)m w k e jkθ (4.253)
k=1

Consequently, the elements of m (for m = 0, 1, and 2) are given by

0i = d0 (4.254)
1i = e0 + e1 i (4.255)
2i = f 0 + f 1 i + f 2 i 2 (4.256)

The constants defining the m elements are made unit length and mutually orthogonal.
Consider how to establish a beam having nonuniform weighting as well as zero-,
first-, and second-order constraints on the beam shape at the center of the main beam. First
expand wq in terms of the constraint vectors m (for m = 0, 1, and 2) and a remainder
vector r as

wq = a0 0 + a1 1 + a2 2 + ar r (4.257)

where

ai = wqT i for i = 0, 1, 2 (4.258)


2
ar r = w q − ai i (4.259)
i=0

Now, construct the complementary projection matrix filter according to

P̂ = I − 0 †0 − 1 †1 − 2 †2 − · · · − r r† (4.260)

The subspace spanned by the constraint vectors in the N -dimension space of the adaptive
processor is preserved by the foregoing construction. The spatial matrix filter constructed
according to (4.263) results in a signal vector z containing no components in the direction
of wq or its first and second derivatives. The vector wq is now added back in (at the point
in Figure 4-21 where b∗ is inserted) to form the final weight vector w.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 199

4.5 Constraint for the Case of Known Desired Signal Power Level 199

4.5 CONSTRAINT FOR THE CASE OF KNOWN

DESIRED SIGNAL POWER LEVEL
If there is no a priori knowledge of the desired signal structure or the direction of arrival,
acquiring a weak desired signal in the presence of strong jamming or interference is
accomplished by placing a constraint on the adaptation algorithm to prevent suppression
of all signals (including interference) below a specified input power level. When a signal
power threshold is greater than the desired signal input power, the weak desired signal will
not be suppressed, whereas all interference signals above the threshold are suppressed.
The most common method of obtaining a power discrimination capability was for-
mulated by Compton [30] and is based on the use of proportional feedback control in a
Howells–Applebaum adaptive loop. To accomplish the same result Zahm [31] proposed
another technique that uses a combination of a steering command vector and a bias signal.
The weight adjustment control loops for the adaptive null-steering array in Figure 4-23
are governed by the differential equation
dw
= α{x(t)[x0 (t) − x† (t)w] − aw} (4.261)
dt
where a is a real scalar constant. The additional feedback path around the integrator
provides the means for setting a power threshold.
For α sufficiently small, (4.261) can be approximated by expected values so that
dw
= α(rx x0 − [Rx x + aI]w) (4.262)
dt
The steady-state weight vector is then given by

w = [Rx x + aI]−1 rx x0 (4.263)

x1 (t) FIGURE 4-23

Aux. 1 Adaptive null
steering with N
Main auxiliary sensor
channel elements. From
∫ dt
Compton, Ohio
a
State Univ. Quarterly
Report 3234-1,
a
December, 1971.
_ _ +
Σ Output e(t)
+ y(t)

xN(t)
Aux. N
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 200

200 CHAPTER 4 Gradient-Based Algorithms

If a desired signal is present then the output signal-to-interference plus noise ratio (SINR) as
1
SNR = (4.264)
(Pe /|s0 − w† s|2 ) − 1
where Pe represents the total output power, s is the desired signal direction vector, and s0 is
the main channel desired signal component. It will be convenient to define the parameter
|s0 − w† s|2
SN  = (4.265)
Pe
so that
1
SNR = (4.266)
(1/S N  ) − 1
It can be shown that
 2
 
N (Qrx x0 )i∗ (Qs)i 
s0 − 
 λi +a 
SN = i=1
(4.267)

N |(Qrx x0 )i |2
a2 λi (λi +a)2
+ Pe0
i=1

where Q is the unitary transformation that diagonalizes Rx x , λi are the eigenvalues of Rx x ,

and Pe0 represents the minimum output power of e(t) when a = 0. It can be shown from the
previous results and (4.261) that a may be selected to prevent cancellation of the desired
signal while suppressing high-level jammers. The output signal-to-interference plus noise
ratio is therefore higher than when a pure integrator is used in the feedback loop.
To maximize the output SNR, the parameter a is selected to maximize S N  . To illus-
trate how a is selected, consider the case of one interfering jammer so that

Rxx = J v J v J + Ps vs vs (4.268)
( (
rxx0 = J0 J v J e jφ J + Ps0 Ps vs e jφs (4.269)

where
J0 = main channel jammer power
J = auxiliary channel jammer power (assumed equal in all auxiliary channels)
v J = jammer direction delay vector
Ps0 , Ps , and vs are similarly defined for the desired signal. φ J and φs represent the relative
phase between the main and auxiliary channel signals for the jamming and desired signals,
respectively.
If the desired signal and the interference signal angles of arrival are such that vs and
v J are orthogonal (which simplifies the discussion for tutorial purposes), then (4.270)
reduces to
2
σ2 + a
Ps0

σ 2 + a + NPs
SN =  !
Ps0 Ps J0 J
a2 N + + Pe0
(NPs + σ 2 + a)2 (NPs + σ 2 ) (NJ + σ 2 + a)2 (NJ + σ 2 )
(4.270)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 201

4.6 The DSD Algorithm 201

where σ 2 = auxiliary channel thermal noise power. For a = 0 (which corresponds to the
conventional LMS null-steering algorithm) SN  becomes
2
σ2
Ps0

NPs + σ 2
SN = ; a=0 (4.271)
Pe0
This result shows that S N  decreases as the input desired signal power in the auxiliary
channels NPs increases above the thermal noise level σ 2 . When NPs σ 2 in (4.271), SN 
is inversely proportional to the input desired signal power, which is the power inversion
characteristic of the minimum MSE performance criterion. When a NPs + σ 2 and
a NJ + σ , (4.270) becomes
2

Ps0
SN  = (4.272)
NPs NJ
Ps + J0 + Pe0
NPs + σ 2 0 NJ + σ 2
Suppression of the main channel signal is prevented by selecting a to be sufficiently
large. However, a is too large in this example, because jammer suppression has also been
prevented, as indicated by the presence of the term NJJ 0 /(NJ + σ 2 ) in the denominator
of (4.272).
Next, assume that the main channel jammer power J0 is nominally equal to the aux-
iliary channel jammer power, and choose a = NPs . Then S N  becomes
0.25Ps0
SN  = (4.273)
Ps0 (NPs − σ 2 )2 (NPs − σ 2 )2
+ + Pe0
4NPs (NPs + σ 2 ) N [1 + (Ps /J )]2 (NJ + σ 2 )
For J Ps , Ps σ 2,
0.25Ps0
SN  ≈
0.25Ps0 + Pe0
An approximation for the output signal-to-interference plus noise ratio in (4.269) is
1 Ps0
SNR ∼ = (4.274)
4 Pe0
Thus, the output signal-to-interference plus noise ratio is now proportional to the main
channel signal power divided by the output residue power Pe0 (recall that Pe0 is the
minimum output residue power obtained when a = 0). Equation (4.274) shows that when
J Ps and Ps σ 2 , the output signal-to-interference plus noise ratio can be significantly
improved by selecting the weight feedback gain as
a ≈ NPs (4.275)
This value of a (when J Ps and Ps σ ) then prevents suppression of the relatively
2

weak desired signal while strongly suppressing higher power level jamming signals.

4.6 THE DSD ALGORITHM

We have seen that, if there are perfect gradient measurements on each iteration, the adap-
tive weight vector converges to the optimal (Wiener) weight vector. In practice, gradi-
ent vector estimates are obtained from a limited statistical sample. The DSD algorithm
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 202

202 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-24
One-dimensional
x
gradient estimation
by way of direct
measurement.

d d
w
w(k)

obtains gradient vector estimates by direct measurement and is straightforward and easy
to implement [13].
The parabolic performance surface representing the MSE function of a single variable
w is defined by

ξ [w(k)] = ξ(k) = ξmin + αw 2 (k) (4.276)

Figure 4-24 represents the parabolic performance surface as a function of a single com-
ponent of the weight vector w. The first and second derivatives of the MSE are
 !
dξ(k)
= 2αw(k) (4.277)
dw w=w(k)

d 2 ξ(k)
= 2α (4.278)
dw 2
w=w(k)

These derivatives are numerically estimated by taking the “symmetric differences”

 !
dξ(k) ξ [w(k) + δ] − ξ [w(k) − δ]
= (4.279)
dw w=w(k) 2δ

d 2 ξ(k) ξ [w(k) + δ] − 2ξ [w(k)] + ξ [w(k) − δ]
= (4.280)
dw 2 δ2
w=w(k)

The procedure for estimating the first derivative illustrated in Figure 4-24 requires
that the weight adjustment be altered to two distinct settings while the gradient estimate
is obtained. If K data samples are taken to estimate the MSE at the two weight settings
w(k) + δ and w(k) − δ, then the average MSE experienced (over both settings) is greater
than the MSE at w(k) by an amount γ . Consequently, a performance penalty is incurred
that results from the weight alteration used to obtain the derivative estimate.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 203

4.6 The DSD Algorithm 203

4.6.1 Performance Penalty Incurred by Gradient Measurement

Figure 4-24 shows that γ for the one-dimensional case is

α[w(k) + δ]2 + α[w(k) − δ]2 + 2ξmin

γ = − αw 2 (k) − ξmin = αδ 2 (4.281)
2
Thus, the value of γ depends only on α and not on w(k). A dimensionless measure of the
system perturbation each time the gradient is measured is defined by

 γ αδ 2
P= = (4.282)
ξmin ξmin
The perturbation is the average increase in the MSE normalized with respect to the mini-
mum achievable MSE.
A two-dimensional gradient is needed for the input signal correlation matrix
 !
r11 r12
Rx x = (4.283)
r21 r22

The MSE corresponding to this correlation matrix is then

ξ = r11 w 12 + r22 w 22 + 2r12 w 1 w 2 + ξmin (4.284)

Measuring the partial derivative of the previous performance surface along the coordinate
w 1 yields a perturbation

r11 δ 2
P= (4.285)
ξmin
Likewise, the perturbation for the measured partial derivative along the coordinate w 2 is

r22 δ 2
P= (4.286)
ξmin
If we allot equal time for the measurement of both partial derivatives (a total of 2K data
samples are used for both measurements), the average perturbation experienced during
the complete measurement process is given by

δ 2 r11 + r22
Pav = · (4.287)
ξmin 2
For N dimensions, define a general perturbation as the average of the perturbations
experienced for each of the individual gradient component measurements so that

δ 2 tr(Rx x )
P= · (4.288)
ξmin N
where “tr” denotes trace, which is defined as the sum of the diagonal elements of the
indicated matrix. When we convert the Rx x matrix to normal coordinates, the trace of Rx x
is the sum of its eigenvalues. Since the sum of the eigenvalues divided by N is the average
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 204

204 CHAPTER 4 Gradient-Based Algorithms

of the eigenvalues (λav ), the perturbation for the N -dimensional case is

δ2
P= λav (4.289)
ξmin
Alternative means of measuring the gradient have also been used in practical systems.
By perturbing or dithering a single weight sinusoidally, the cross-correlation between the
weight value and the performance measure can be measured to determine the derivative of
the performance surface. Likewise, all weights can be dithered simultaneously at distinct
individual frequencies and the gradient components then obtained by cross-correlation.
The procedure for determining the derivative illustrated in Figure 4-24 corresponds to
square-wave dithering.

4.6.2 Gradient Measurement Noise and Misadjustment

in the Weight Vector
Gradients measured by taking finite differences as in Figure 4-23 are noisy because the
MSE measurements on which the differences are based are noisy. Each MSE measurement
is an estimate ξ̂ of the actual MSE ξ based on K data samples:

1  2
K
ξ̂ = e (k) (4.290)
K k=1

It is well known that the variance of a sample average estimate of the mean square obtained
from K independent samples is given by the difference between the mean fourth and the
square of the mean square all divided by K. Consequently the variance of ξ̂ may be
expressed as [32]
E{e4 (k)} − [E{e2 (k)}]2
var[ξ̂ ] = (4.291)
K
If the random variable e(k) is normally distributed with zero mean and variance σ 2 ,
then its mean fourth is 3σ 4 , and the square of its mean square is σ 4 . Consequently, the
variance in the estimate of ξ is given by
1 2σ 4 2ξ 2
var[ξ̂ ] = (3σ 4 − σ 4 ) = = (4.292)
K K K
From (4.292) we find that the variance of ξ̂ is proportional to the square of ξ and inversely
proportional to the number of data samples. In general, the variance can be expressed as
ξ2
var[ξ̂ ] = η (4.293)
K
where η has the value of 2 for an unbiased Gaussian density function. In the event that
the probability density function for ξ̂ is not Gaussian, then the value of η is generally less
than but close to 2. It is therefore convenient to assume that the final result expressed in
(4.292) holds for the analysis that follows.
The derivatives required by the DSD algorithm are measured in accordance with
(4.279). The measured derivative involves taking finite differences of two MSE estimates,
so the error in the measured derivative involves the sum of two independent components
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 205

4.6 The DSD Algorithm 205

[since the error samples e(k) are assumed to be independent]. The variance of each compo-
nent to the derivative error is given by (4.292). Assume that we are attempting to measure
the derivative at a point on the performance surface where the weight vector is near the
minimum point of the MSE surface and that the perturbation P is small, then the two
components of measured derivative error will have essentially the same variances. The
total variance of the measured derivative error will then be the sum of the variances of the
two components. From (4.279) and (4.292) it follows that the variance in the estimate of
the derivative is given by
 ! 
dξ 1 2ξ 2 [w(k) + δ] 2ξ 2 [w(k) − δ]
var = 2 +
dw w=w(k) 4δ K K

∼ ξ2
= min2 (4.294)
Kδ
When an entire gradient vector is measured, then the errors in each component are
independent. It is convenient to define a gradient noise vector g(k) in terms of the true
gradient ∇(k) and the estimated gradient ∇(k):
ˆ

∇(k)
ˆ = ∇(k) + g(k) (4.295)
where g(k) is the gradient noise vector. Under the previously assumed conditions, the
covariance of the gradient noise vector can be expressed as
ξmin
2
cov [g(k)] = I (4.296)
K δ2
Transforming the gradient noise vector into normal coordinates, we have
g (k) = Qg(k) (4.297)
We see from (4.296) that the covariance matrix of g(k) is a scalar multiplying the identity
matrix, so projecting into normal coordinates through the orthonormal transformation Q
yields the same covariance for g (k):
  ξ2
cov [g (k)] = E Qg(k)gT (k)Q−1 = min2 I (4.298)
Kδ
This result merely emphasizes that near the minimum point of the performance surface
the covariance of the gradient noise is essentially a constant and does not depend on w(k).
The fact that the gradient estimates are noisy means that weight adaptation based on
these gradient estimates will also be noisy, and it is consequently of interest to determine
the corresponding noise in the weight vector. Using estimated gradients, the method of
steepest descent yields the vector difference equation

v(k + 1) = v(k) + s (−∇(k))

ˆ = v(k) + s [−∇(k) − g(k)] (4.299)

where v(k) = w(k) − wopt . Since the true gradient from (4.63) is given by

∇(k) = 2Rx x v(k) (4.300)

(4.299) can be rewritten as

v(k + 1) = [I − 2s Rx x ]v(k) − s g(k) (4.301)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 206

206 CHAPTER 4 Gradient-Based Algorithms

which is a first-order difference equation having a stochastic driving function –s g(k).
Projecting the previous difference equation into normal coordinates by premultiplying by
Q then yields

v (k + 1) = [I − 2s ]v (k) − s g (k) (4.302)

After initial adaptive transients have died out and the steady state is reached, the
weight vector v (k) behaves like a stationary random process in response to the stochastic
driving function –s g (k). In the normal coordinate system there is no cross-coupling
between terms, and the components of g (k) are uncorrelated; thus, the components of
v (k) are also mutually uncorrelated, and the covariance matrix of g (k) is diagonal. The
covariance matrix of v (k) describes how noisy the weight vector will be in response to the

stochastic driving function, and we now proceed to find this matrix. Since cov[v (k)] =
E{v (k)vT (k)}, it is of interest to determine the quantity v (k + 1)vT (k + 1) by way of
(4.302) as follows:

v (k + 1)vT (k + 1) = (I − 2s )v (k)vT (k)(I − 2s )

+ 2s g (k)gT (k) − s (I − 2s )v (k)g (k)vT (k)
− s g (k)vT (k) (4.303)

Taking expected values of both sides of (4.303) and noting that v (k) and g (k) are uncor-
related since v (k) is affected only by gradient noise from previous adaptive cycles, we
obtain for the steady state

cov[v (k)] = (I − 2s )cov[v (k)](I − 2s ) + 2s cov[g (k)]

 
= I − 4s  + 42s 2 cov[v (k)] + 2s cov[g (k)] (4.304)

Combining like terms in (4.304) then yields

 −1
cov[v (k)] = 2s 4s  − 42s 2 cov[g (k)] (4.305)

In practice, the step size in the method of steepest descent is selected so that

s   I (4.306)

As a result of (4.306), squared terms occurring in (4.305) can be neglected so that

s −1
cov[v (k)] ∼
=  cov[g (k)] (4.307)
4
Since cov[g (k)] is given by (4.298), we now have
s ξmin
2
cov[v (k)] ∼
= −1 (4.308)
4K δ 2
The covariance of the weight vector in the operational coordinate system can be obtained
−1 −1
from (4.308) by recalling that R−1 
x x = Q  Q and v = Qv so that
 
cov[v(k)] = E Q−1 v (k)vT (k)Q
s ξmin
2
= R−1 (4.309)
4K δ 2 x x
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 207

4.6 The DSD Algorithm 207

Without any noise in the weight vector, the method of steepest descent converges to a
steady-state solution at the minimum point of the MSE performance surface (the bottom
of the bowl). The MSE would then be ξmin . The noise present in the weight vector causes
the steady-state solution to randomly wander about the minimum point. The result of this
wandering is a steady-state MSE that is greater than ξmin and hence is said to have an
“excess” MSE. We will now consider how severe this excess MSE is for the noise that is
in the weight vector.
We have already seen in Section 4.1.3 that the MSE can be expressed as

ξ(k) = ξmin + vT (k)v (k) (4.310)

where v (k) = w (k) − wopt . Consequently, the average excess MSE is

  N
 2 
E vT (k)v (k) = λ p E v p (k) (4.311)
p=1

But from (4.308) we may write

 2  s ξmin
2
1
E v p (k) = (4.312)
4kδ 2 λp

Thus, (4.311) can be rewritten as

N s ξmin
2
E{vT (k)v (k)} = (4.313)
4K δ 2
Recalling that the misadjustment M is defined as the average excess MSE divided by
the minimum MSE there results for the DSD algorithm
N s ξmin
M= (4.314)
4K δ 2
The foregoing result is more usefully expressed in terms of time constants of the learning
process and the perturbation of the gradient estimation process as developed next.
Each measurement to determine a gradient component uses 2K samples of data. Each
adaptive weight iteration involves N gradient component measurements and therefore
requires a total of 2KN data samples. From Section 4.2.3 it may be recalled that the MSE
learning curve has a pth mode time constant given by
1 τp
τ pMSE = = (4.315)
4s λ p 2

in time units of the number of iterations. It is useful to define a new time constant T pMSE
whose basic time unit is the data sample and whose value is expressed in terms of the
number of data samples. It follows that for the DSD algorithm

T pMSE = 2KNτ pMSE (4.316)

The time constant T pMSE relates to real time units (seconds) once the sampling rate is
known.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 208

208 CHAPTER 4 Gradient-Based Algorithms

By using the perturbation formula (4.282) to substitute for ξmin in (4.314), the misad-
justment for the DSD algorithm is rewritten as
N s λav
M= (4.317)
4KP
The time constant defined by (4.316) is rewritten using (4.315) as
NK
T pMSE = (4.318)
2s λ p
from which one can conclude that
NK 1
λp = (4.319)
2s T pMSE
so that
NK 1
λav = (4.320)
2s TMSE av

Combining (4.26) and (4.320) then yields the misadjustment as

N2 1
M= (4.321)
8P TMSE av

Equation (4.321) shows that for the DSD algorithm, misadjustment is proportional
to the square of the number of weights and inversely proportional to the perturbation. In
addition, the misadjustment is also inversely proportional to the speed of adaptation (fast
adaptation results in high misadjustment). Since the DSD algorithm is based on steepest
descent, it suffers from the disparate eigenvalue problem discussed in Section 4.2.2.
It is appropriate here to compare the misadjustment for the DSD algorithm given by
(4.321) with the misadjustment for the LMS algorithm given by (4.83). With a specified
level of misadjustment for the LMS algorithm, the adaptive time constants increase linearly
with the number of weights rather than with the square of the number of weights as
is the case with the DSD algorithm. Furthermore, with the LMS algorithm there is no
perturbation. As a result, in typical circumstances much faster adaptation is possible with
the LMS algorithm than with the DSD algorithm.
M is defined as a normalized performance penalty that results from noise in the
weight vector. In an actual adaptive system employing the DSD algorithm, the weight
vector is not only stochastically perturbed due to the presence of noise but in addition
is deterministically perturbed so the gradient can be measured. As a consequence of
the deterministic perturbation, another performance penalty accrues as measured by the
perturbation P, which is also a normalized ratio of excess MSE. The total excess MSE is
therefore the sum of the “stochastic” and “deterministic” perturbation components. The
total misadjustment can be expressed as

Mtot = M + P (4.322)

Adding the previous two components then yields

N2 1
Mtot = +P (4.323)
8P TMSE av
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 209

4.7 The Accelerated Gradient Approach (AG) 209

Since P is a design parameter given by (4.282), it can be selected by choosing the deter-
ministic perturbation size δ. It is desirable to minimize the total misadjustment Mtot by
appropriately selecting P. The result of such optimization is to make the two right-hand
terms of (4.323) equal so that
1
Popt = Mtot (4.324)
2
The minimum total misadjustment then becomes
1/2
N2 1 N2 1
(Mtot )min = = (4.325)
4Popt TMSE av 2 TMSE av

Unlike the LMS algorithm, the DSD algorithm is sensitive to any correlation that
exists between successive samples of the error signal e(k), since such correlation has the
effect of making the effective statistical sample size less than the actual number of error
samples in computing the estimated gradient vector. Because of such reduced effective
sample size, the actual misadjustment experienced is greater than that predicted by (4.325),
which was derived using the assumption of statistical independence between successive
error samples.

4.7 THE ACCELERATED GRADIENT

APPROACH (AG)
Algorithms based on the steepest descent method exhibit an undesirable degree of sen-
sitivity of the convergence speed to the eigenvalue spread in the input signal covariance
matrix. The conjugate gradient method [33–39] is faster than the steepest descent ap-
proach. This reluctance to apply accelerated gradient methods to adaptive array problems
is due principally to the following reasons:
1. There is increased hardware complexity associated with implementing the algorithm.
2. The enhanced convergence speed realized is not as fast as what can be obtained with
other methods (to be discussed in later chapters).
3. The various accelerated gradient methods are all susceptible (although with different
degrees) to signals that are noise corrupted.
4. There is a required increase in computation and memory space over steepest descent
methods.
Despite these objections, it is worthwhile to apply one of the accelerated gradi-
ent methods to determine what improvement in convergence speed reasonably might
be expected by recourse to such techniques. Powell’s method [35] is tolerant of noise-
induced errors, although other methods theoretically may yield faster convergence speeds.
Powell’s method assumes that if a performance measure (w) is quadratic in the inde-
pendent variables, then any line that passes through the minimum point of the quadratic
performance surface intersects the family of constant performance contours at equal an-
gles. This property is illustrated in Figure 4-25 for two dimensions where it is seen that
the line AC connecting point A with the minimum point C intersects the constant perfor-
mance contours at equal angles. As a consequence of the equal angle property, the line
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 210

210 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-25
Two-dimensional
diagram showing Negative gradient
A direction for point B
directional
relationships for the
Powell descent
method. D C

B Desired direction
for point A

Negative gradient
direction for point A

joining the point A with the point D in Figure 4-25 passes through the point C where the
derivative of the performance measure (w) with respect to distance along the line AD
is zero.
Given an initial estimate w0 at point A, first find the gradient direction that is normal
to the tangent of the constant performance measure contour. Proceed along the line defined
by the negative gradient direction to the point B where the derivative of (w) with respect
to distance along the line is zero. The point B may in fact be any arbitrary point on the
line that is a finite distance from A; however, by choosing it in the manner described the
convergence of the method is assured.
Having found point B, the negative gradient direction that is parallel to the original tan-
gent at (w0 ). Traveling in this new normal direction, we find a point where the derivative
of (w) with respect to distance along the line is zero (point D in Figure 4-25). The line
passing through the points A and D also passes through the point C. The desired point C
is the point where the derivative of (w) with respect to distance along the line AD is zero.
The generalization of the foregoing procedure to an N -dimensional space can be
obtained by recognizing that the directional relationships (which depend on the equal
angle property) given in Figure 4-25 are valid only in a two-dimensional plane. The
first step (moving from point A to point B) is accomplished by moving in the negative
gradient direction in the N -dimensional space. Having found point B, we can construct
(N − 1) planes between the original negative gradient direction and (N − 1) additional
mutually orthogonal vectors, thereby defining points C, D, E, . . . , until (N −1) additional
points have been defined. The last three points in the N -dimensional space obtained in
the foregoing manner may now be treated in the same fashion as points A, B, and D of
Figure 4-25 by drawing a connecting line between the last point obtained and the point
defined two steps earlier. Traveling along the connecting line one may then define a new
point C of Figure 4-25. This new point may then be considered as point D in Figure 4-25,
and a new connecting line may be drawn between the new point and the point obtained
three steps earlier.
The steps corresponding to one complete Powell descent cycle for five dimensions
are illustrated in Figure 4-26. The first step from A to B merely involves traveling in the
negative gradient direction v1 with a step size α1 chosen to satisfy the condition
d
{[w(0) + α1 v1 ]} = 0 (4.326)
dα1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 211

4.7 The Accelerated Gradient Approach (AG) 211

A FIGURE 4-26
Illustration of Powell
1 descent method
steps required in five
J dimensions for one
B
complete cycle.
9
2

I
C

8
3

n s)
N Steps

io
ct
in ps
(orthogonal H

ire
D

g l te
ed
gradient

in 1 S
directions)

−
7

N
4

ct
ne
G

on
E

6 (c
5

so that

w(1) = w(0) + α1 v1 (4.327)

Having determined point B, we determine point C by traveling in the negative gradient

direction v2 (a direction that is also orthogonal to v1 ) from point B with step size α2
selected to satisfy
d
{[w(1) + α2 v2 ]} = 0 (4.328)
dα2
Point D is determined from point C by using the aforementioned procedure, and this
process continues until a total of five points (B through F in Figure 4-26) are defined.
A descent direction v6 is defined by drawing a connecting line between points D and F
(analogous to points A and D of Figure 4-25) and traveling along this line with step size
α6 selected to satisfy
d
{[w(5) + α6 v6 ]} = 0 (4.329)
dα6
thereby arriving at point G. A new descent direction v7 is defined by drawing a connecting
line between points C and G and traveling along this line with step size α7 selected to
satisfy
d
{[w(6) + α7 v7 ]} = 0 (4.330)
dα7
thereby arriving at point H . This process continues until the solution point J for the cycle
is found from the descent direction v9 (defined by the connecting line between points A
and I ) and the step size α9 . In general, one Powell descent cycle in N -dimensional space
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 212

212 CHAPTER 4 Gradient-Based Algorithms

therefore requires N + N − 1 = 2N − 1 steps. On completion of one descent cycle, the

entire cycle are repeated using the last weight vector obtained as the initial weight vector
for the new cycle.

4.7.1 Algorithm Based on the Powell Accelerated Gradient Cycle

Each step in a Powell descent cycle involves traveling from a weight vector w along a
direction v with step size α such that

d
[(w + αv)] = 0 (4.331)
dα
For complex weights the MSE performance measure is given by

ξ(w) = E{d 2 } + w† rxd + r†xd w + w† Rx x w (4.332)

The gradient of ξ(w + αv) with respect to α is then given by

∇α [ξ(w + αv)] = v† rxd + r†xd v + v† Rx x w + w† Rx x v + 2αv† Rx x v (4.333)

and this gradient is equal to zero when the step size is

v† rxd + v† Rx x w
α=− (4.334)
v† Rx x v
Since rxd and Rx x are unknown, some estimate of the numerator is employed to obtain an
appropriate step size estimate. Noting that rxd + Rx x w is one-half the gradient of ξ(w),
it follows that the numerator of (4.334) are approximated by v† Av{e(k)x(k)}. Note that
the quantity v† x is regarded as the output of a processor whose weights correspond to v
and that Av{(v† x)(x† v)} is an approximation of the quantity v† Rx x v, where the average
Av{ } is taken over K data samples. The simultaneous generation of the estimates ∇ ˆ w and
† †
Av{v xx v} requires parallel processors: one processor with weight values equal to w(k)
and another processor with weight values equal to v(k). Having described the procedure
for determining the appropriate step size along a direction v, we may now consider the
steps required to implement an entire Powell descent cycle.
The steps required to generate one complete Powell descent cycle are as follows.
Step 1 Starting with the initial weight setting w(0), estimate the negative gradient
direction v(0) using K data samples then travel in this direction with the appropriate step
size to obtain w(1). The step size determination requires an additional K data samples to
obtain by way of (4.334).
Steps 2 →N Estimate the negative gradient direction at w(k) using K data samples.
If the gradient estimates and the preceding step size were error free, the current gradient
is automatically orthogonal to the previous gradient directions. Since the gradient esti-
mate is not error free, determine the new direction of travel v(k) by requiring it to be
orthogonal to all previous directions v(0), v(1), . . . , v(k − 1) by employing the Schmidt
orthogonalization process so that


k−1
[v† (i)∇(k)]
ˆ
v(k) = ∇(k)
ˆ − · v(i) (4.335)
i=0
[v† (i)v(i)]
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 213

4.8 Gradient Algorithm with Constraints 213

Travel in the direction −v(k) using the appropriate step size (which requires an additional
K data samples to obtain) to arrive at w(k + 1).
Steps N + 1 → 2N − 1 Determine the new direction of travel at w(k) by forming

v(k) = w(k) − w[2(N − 1) − k] (4.336)

Travel in the direction −v(k) from w[2(N − 1) − k] using the appropriate step size to
arrive at w(k + 1). These steps require only K data samples since now the direction of
travel does not require that a gradient estimate be obtained.

4.8 GRADIENT ALGORITHM WITH CONSTRAINTS

The early applications of sidelobe cancellation to radar antennas neglected the effects
of signals in the main beam on the adapted response. Such neglect was amply justified
because adaptive processors would not respond to low level reflected target signals, and
the small number of degrees of freedom then available to the adaptive processor limited the
response to large targets or extended clutter. Current adaptive arrays with large numbers
of degrees of freedom are explicitly constrained to avoid degradation of the main beam.
Adaptive arrays having large numbers of degrees of freedom and fast response times
operating with high-energy, long-duration waveforms may have reflected signal returns
that are large enough to elicit undesirable responses from the adaptive processor. Such
undesirable responses may produce signal cancellation and signal waveform distortion.
Furthermore, jammer power level affects the array response in the main beam direction,
thereby allowing blinking jammers to modulate the signal response and consequently
degrade the performance of any subsequent coherent processing. The Frost [40] algorithm
imposes constraints on the adaptive weights such that certain main beam properties are
preserved. It turns out that the resulting constrained optimization system has two parts:
(1) a preprocessing part called a “spatial correction filter,” which compensates the signals
for the misalignment between the plane wave front and the sensor array geometry; and
(2) a signal processor that includes the adaptive weights and accounts for the primary
function of the adaptive array.
The constrained LMS algorithm requires that the direction of arrival and a frequency
band of interest be specified a priori for the appropriate constraint conditions to be imposed.
Because the look direction frequency response relates to the adaptive weights, the algo-
rithm maintains a selected frequency response in the look direction while simultaneously
minimizing output noise power. If the look direction is perpendicular to the line of sensors,
then identical signal components appear at the first taps [so x1 (t) = x2 (t) = · · · = x N (t)
in Figure 4-27] and propagate down the tapped delay lines following each sensor [so
x N +1 (t) = x N +2 (t) = · · · = x2N (t), and x(J −1)N +1 (t) = x(J −1)N +2 (t) = · · · = xNJ (t)].
Noise component waveforms arriving at the sensors from any direction other than the
look direction will not usually produce equal voltage components at any vertical column
of taps. Consequently, as far as the signal is concerned, the adaptive processor appears
as an equivalent single tapped delay line in which each adaptive weight equals the sum
of the weights in the corresponding vertical column of the original processor. These J
summation weights in the equivalent tapped delay line are assigned a value to give the
desired frequency response characteristic in the look direction, thereby giving rise to J
constraint conditions. In the event that the look direction is other than that perpendicular
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 214

214 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-27 Spatial

Broadband adaptive correction Time delays of t seconds
Sensors
array having N filter
sensors and J x1 xN +1 x(J − 1)N + 1
d t t t
adjustable weights 1
per sensor. Adaptive weights

Delay settings
Σ Σ Σ

x2 xN +2 x(J − 1)N + 2
2d t t t
2

Σ Σ Σ Σ Array output

xN x2N xNJ
t t t
N
Nd

Σ Σ Σ

to the line of sensors (as the previous discussion has assumed), then the time delays in
the spatial correction filter are adjusted so the signal components of each channel at the
output of the preprocessor are in phase.
The adaptive signal processor of Figure 4-27 has N sensors and J taps per sensor
for a total of NJ adjustable weights. Using J constraints to determine the look direction
frequency response leaves NJ − J degrees of freedom to minimize the total array output
power. Since the J constraints fix the look direction frequency response, minimizing the
total output power is equivalent to minimizing the nonlook direction noise power (provided
the signal voltages at the taps are uncorrelated with the corresponding noise voltages at
these taps). If signal-correlated noise in the array is present, then part or all of the signal
component of the array output may be cancelled. Although signal-correlated noise may
not occur frequently, sources of such noise include multiple signal-propagation paths, and
coherent radar or sonar “clutter.”
It is desirable for proper noise cancellation that the noise voltages appearing at the
adaptive processor taps be correlated among themselves (although uncorrelated with the
signal voltages). Such noise sources may be generated by lightning, “jammers,” noise
from nearby vehicles, spatially localized incoherent clutter, and self-noise from the struc-
ture carrying the array. Noise voltages that are uncorrelated between taps (e.g., amplifier
thermal noise) are partially rejected by the adaptive array either as the result of incoherent
noise voltage addition at the array output or by reducing the weighting applied to any taps
that may have a disproportionately large uncorrelated noise power.

4.8.1 Optimum Constrained Weight Vector Solution

The voltages appearing at each array tap in Figure 4-27 are sampled every  seconds
(where  is a multiple of the delay τ between taps). The vector of tap voltages at the kth
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 215

4.8 Gradient Algorithm with Constraints 215

sample is defined by

xT (k) = [x1 (k), x2 (k), . . . , xNJ (k)] (4.337)

At any tap the voltages that appear may be regarded as the sums of voltages due to look
direction signals s and nonlook direction noises n, so that

x(k) = s(k) + n(k) (4.338)

where the NJ-dimensional vector of look direction signals is defined by

⎡ ⎤⎫
s(k) ⎪
⎬
⎢ . ⎥
⎢ .. ⎥ N taps
⎢ ⎥⎪
⎢ ⎥⎭
⎢ s(k) ⎥⎫
⎢ s(k − τ ) ⎥
⎢ ⎥⎪
⎢ .. ⎥⎬
⎢ ⎥ N taps
 ⎢ . ⎥⎪
s(k) = ⎢ ⎥⎭ (4.339)
⎢ s(k − τ ) ⎥
⎢ .. ⎥
⎢ ⎥
⎢ . ⎥
⎢ ⎥⎫
⎢s[k − (J − 1)τ ]⎥ ⎪
⎢ ⎥⎬
⎢ .. ⎥ N taps
⎣ . ⎦⎪
⎭
s[k − (J − 1)τ ]

and the vector of nonlook direction noises is defined by


nT (k) = [n 1 (k), n 2 (k), . . . , n NJ (k)] (4.340)

The weight vector appearing at each tap is denoted by w, where


wT = [w 1 , w 2 , . . . , w NJ ] (4.341)

We assume that the signals and noises are zero-mean random processes with unknown
second-order statistics. The covariance matrices of x, s, and n are given by

E{x(k)xT (k)} = Rx x (4.342)

E{s(k)s (k)} = Rss
T
(4.343)
E{n(k)nT (k)} = Rnn (4.344)

Since the vector of look direction signals is assumed uncorrelated with the vector of
nonlook direction noises

E{n(k)sT (k)} = 0 (4.345)

Assume that the noise environment is such that Rx x and Rnn are positive definite and
symmetric.
The adaptive array output (which forms the signal estimate) at the kth sample is
given by

y(k) = wT x(k) = xT (k)w (4.346)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 216

216 CHAPTER 4 Gradient-Based Algorithms

From (4.346), it follows that the expected array output power is

E{y 2 (k)} = E{wT x(k)xT (k)w} = wT Rx x w (4.347)

Suppose that the weights in the jth vertical column of taps sums to a selected number
f j . This constraint may be expressed by the relation

Tj w = f j , j = 1, 2, . . . , J (4.348)

where the NJ-dimensional vector  j is given by

⎡ ⎤
⎢0⎥ ⎫
⎢ ⎥⎪
⎢.⎥ ⎬
⎢ .. ⎥ N
⎢ ⎥⎪
⎢0⎥ ⎭
⎢ ⎥
⎢.⎥
⎢.⎥
⎢.⎥ ⎫
⎢ ⎥
⎢0⎥ ⎪
⎢.⎥ ⎬
⎢.⎥ N
⎢.⎥ ⎪
⎢ ⎥⎭
⎢0⎥
⎢ ⎥⎫
⎢1⎥ ⎪
⎢ ⎥⎬
⎢ ⎥
 j = ⎢ ... ⎥ jth column of N elements (4.349)
⎢ ⎥⎭ ⎪
⎢1⎥
⎢ ⎥⎫
⎢0⎥ ⎪
⎢ ⎥⎬
⎢.⎥
⎢ .. ⎥ N
⎢ ⎥⎪
⎢0⎥ ⎭
⎢ ⎥
⎢.⎥
⎢.⎥
⎢.⎥ ⎫
⎢ ⎥
⎢0⎥ ⎪
⎢ ⎥⎬
⎢ .. ⎥ N
⎢.⎥ ⎪
⎢ ⎥⎭
⎣0⎦

Now consider the requirement of constraining the entire weight vector to satisfy all J
equations given by (4.348). Define a J × NJ constraint matrix C having  j as elements.


C = [1 · · ·  j · · ·  J ] (4.350)

Furthermore, define f as the J -dimensional vector of summed weight values for each of
the j vertical columns that yield the desired frequency response characteristic in the look
direction as
⎡ ⎤
f1
⎢ f2 ⎥
 ⎢ ⎥
f=⎢ . ⎥ (4.351)
⎣ .. ⎦
fJ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 217

4.8 Gradient Algorithm with Constraints 217

It immediately follows by inspection that the full set of constraints (4.348) can be written
in matrix form as

CT w = f (4.352)

Now that the look direction frequency response is fixed by the constraint equation
(4.352), minimizing the nonlook direction noise power is equivalent to minimizing the
total output power given by (4.347). The constrained optimization problem reduces to

Minimize wT Rx x w (4.353)
w
subject to C w = f
T
(4.354)

Lagrange multipliers are used to find wopt that satisfy (4.353) and (4.354) [41]. Ad-
joining the constraint equation (4.354) to the cost function (4.353) by a J -dimensional
vector λ, whose elements are undetermined Lagrange multipliers (and including a factor
of 12 to simplify the arithmetic), then yields
1 T
Minimize (w) = w Rxx w + λT [CT w − f] (4.355)
w 2
The gradient of (4.355) with respect to w is given by

∇w (w) = Rx x w + Cλ (4.356)

A necessary condition for (4.355) to be minimized is that the gradient be equal to zero so
that

Rx x w + Cλ = 0 (4.357)

Therefore, the optimal weight vector is given by

wopt = −R−1
x x Cλ (4.358)

where the vector λ remains to be determined. The vector of Lagrange multipliers may now
be evaluated from the constraint equation
 
CT wopt = f = CT − R−1 x x Cλ (4.359)

It then follows that the vector λ is given by

 −1
λ = − CT R−1
xx C f (4.360)

where the existence of [CT R−1

x x C]
−1
is guaranteed by the fact that Rx x is positive definite
and C has full rank. Combining (4.358) and (4.360) then yields the optimum constrained
weight vector
 T −1 −1
wopt = R−1 x x C C Rx x C f (4.361)

If we substitute wopt into (4.346), it follows that the constrained least squares estimate of
the look direction signal provided by the array is

yopt (k) = wopt

T
x(k) (4.362)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 218

218 CHAPTER 4 Gradient-Based Algorithms

If the vector of summed weight values f is selected so the frequency response char-
acteristic in the look direction is all-pass and linear phase (distortionless), then the output
of the constrained LMS signal processor is the maximum likelihood (ML) estimate of
a stationary process in Gaussian noise (provided the angle of arrival is known) [42]. A
variety of other optimal processors can also be obtained by a suitable choice of the vector
f [43]. It is also worth noting that the solution (4.361) is sensitive to deviations of the
actual signal direction from that specified by C and to various random errors in the array
parameters [44].

4.8.2 The Adaptive Algorithm

Assume the adaptive processor must determine the correlation matrix Rx x . It follows that
in stationary environments during learning and in time-varying environments an estimate
of the optimum adaptive processor weights must be periodically recomputed. The initial
guess of an appropriate weight vector satisfies (4.354), so a good starting point is

w(0) = C[CT C]−1 f (4.363)

where the quantity C[CT C]−1 represents the pseudo-inverse of the singular matrix CT [45].
For a gradient type algorithm, after the kth iteration the next weight vector is given by

w(k + 1) = w(k) − s ∇w [w(k)]

= w(k) − s [Rx x w(k) + Cλ(k)] (4.364)

where s is the step size constant, and  denotes the performance measure. Requiring
w(k + 1) to satisfy (4.352) then yields

f = CT w(k + 1) = CT {w(k) − s [Rx x w(k) + Cλ(k)]} (4.365)

Consequently, the Lagrange multipliers are given by

1
λ(k) = −[CT C]−1 CT Rx x w(k) − [CT C]−1
s
· [f − CT w(k)] (4.366)

Substituting (4.366) into (4.364) then gives the iterative relation

w(k + 1) = w(k) − s [I − C(CT C)−1 CT ]Rx x w(k)

+ C(CT C)−1 [f − CT w(k)] (4.367)

It is convenient to define the NJ-dimensional vector


 = C(CT C)−1 f (4.368)
and the NJ × NJ matrix

P = I − C(CT C)−1 CT (4.369)
Then the iterative relation (4.369) may be rewritten as
w(k + 1) = P[w(k) − s Rx x w(k)] +  (4.370)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 219

4.8 Gradient Algorithm with Constraints 219

In the actual system the input correlation matrix is not known, and it is necessary to adopt
some estimate of this matrix to insert in place of Rx x in the iterative weight adjustment
equation. An approximation for Rx x at the kth iteration is merely the outer product of the
tap voltage vector with itself: x(k)xT (k). Substituting this estimate of Rx x into (4.370)
and recognizing that y(k) = xT (k)w(k) then yields the constrained LMS algorithm

w(0) = 
(4.371)
w(k + 1) = P[w(k) − s y(k)x(k)] + 
If it is merely desired to ensure that the complex response of the adaptive array system
to a normalized signal input from the look direction is unity, then the spatial correction
filter is dispensed with and the compensation for phase misalignment incorporated directly
into the variable weight selection as suggested by Takao et al. [46]. Denote the complex
response (amplitude and phase) of the array system by Y (θ ), where θ is the angle measured
from the normal direction to the array face. The appropriate conditions to impose on the
adaptive weights are found by requiring that e{Y (θ )} = 1 and Im{Y (θ )} = 0 when
θ = θc , the look direction.

4.8.3 Conditions Ensuring Convergence to the Optimum Solution

The weight vector w(k) obtained by employing (4.371) is a random vector. Convergence
of the mean value of the weight vector to the optimum is shown by considering the length
of the difference vector between the mean of the actual weight vector and the optimum
weight vector: convergence is assured if the length of the difference vector asymptotically
approaches zero.
If we start with the weight adjustment equation

w(k + 1) = P[w(k) − s x(k)y(k)] +  (4.372)

and recognize that y(k) = xT (k)w(k), then taking the expected value of both sides of
(4.372) yields

E[w(k + 1)] = P{E[w(k)] − s Rx x E[w(k)]} +  (4.373)

Define the difference vector v(k + 1) by


v(k + 1) = E [w(k + 1)] − wopt (4.374)

Substitute (4.373) into (4.374) and use  = (I − P)wopt and PRx x wopt = 0 [which may be
verified by direct substitution of (4.361) and (4.369)], then the difference vector satisfies

v(k + 1) = Pv(k) − s PRx x v(k) (4.375)

Note from (4.369) that P is idempotent (i.e., P2 = P), then premultiplying (4.375) by
P reveals that Pv(k + 1) = v(k + 1) for all k, so (4.375) can be rewritten as

v(k + 1) = [I − ∇s PRx x P]v(k)

= [I − ∇s PRx x P](k+1) v(0) (4.376)

From (4.376) it follows that the matrix PRx x P determines both the rate of convergence of
the mean weight vector to the optimum solution and the steady-state variance of the weight
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 220

220 CHAPTER 4 Gradient-Based Algorithms

vector about the optimum. The matrix PRx x P has J zero eigenvalues (corresponding to
the column vectors of the constraint matrix C) and NJ − J nonzero eigenvalues σi , i =
1, 2, . . . , NJ − J [48]. The values of the NJ − J nonzero eigenvalues are bounded by the
relation

λmin ≤ σmin ≤ σi ≤ σmax ≤ λmax (4.377)

where λmin and λmax denote the smallest and largest eigenvalues of Rx x , respectively, and
σmin and σmax denote the smallest and largest nonzero eigenvalues of PRx x P, respectively.
The initial difference vector v(0) =  − wopt can be expressed as a linear combination
of the eigenvectors of PRx x P corresponding to the nonzero eigenvalues [47]. Consequently,
if v(0) is equals an eigenvector ei of PRx x P corresponding to the nonzero eigenvalue
σi , then

v(k + 1) = [I − s PRx x P](k+1) ei

= [1 − s σi ](k+1) ei (4.378)

From (4.378) it follows that along any eigenvector of PRx x P the mean weight vector
converges to the optimum weight vector geometrically with the geometric ratio (1−s σi ).
Consequently, the time required for the difference vector length to decay to 1/e of its initial
value is given by the time constant
t
τi =
ln(1 − s σi )
∼ t
= if s σi  1 (4.379)
s σi
where t denotes the time interval corresponding to one iteration.
If the step size constant s is selected so that
1
0 < s < (4.380)
σmax
then the length (given by the norm) of any difference vector is bounded by

(1 − s σmax )(k+1) v(0) ≤ v(k + 1)

≤ (1 − s σmin )(k+1) v(0) (4.381)

It immediately follows that if the initial difference vector length is finite, then the mean
weight vector converges to the optimum so that

lim E{w(k)} − wopt  = 0 (4.382)

k→∞

where the convergence occurs with the time constants given by (4.379).
The LMS algorithm is designed to cope with nonstationary noise environments by
continually adapting the weights in the signal processor. In stationary environments, how-
ever, this adaptation results in the weight vector exhibiting an undesirable variance about
the optimum solution thereby producing an additional (above the optimum) component
of noise to appear at the adaptive array output.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 221

4.8 Gradient Algorithm with Constraints 221

The optimum (minimum) output power level is given by

 2 
E yopt (k) = wopt Rx x wopt
 −1
= fT CT R−1 xx C f (4.383)

The additional noise caused by adaptively adjusting the weights can be compared with
(4.383) to determine the penalty incurred by the adaptive algorithm. A direct measure of
this penalty is the “misadjustment” M defined by (4.55). For a step size constant satisfying
1
0 < s < (4.384)
σmax + 1
2
tr(PRx x P)
The steady-state misadjustment has been shown to be bounded by [48]
s tr(PRx x P) s tr(PRx x P)
· ≤M≤ ·
2 1 − (s /2)[tr(PRx x P) + 2σmin ] 2 1 − (s /2)[tr(PRx x P) + 2σmax ]
(4.385)

If s is chosen to satisfy
2
0 < s < (4.386)
3tr(Rx x )
then it will automatically also satisfy (4.384). It is also worth noting that the upper bound in
(4.383) can be easily calculated directly from observations since tr(Rx x ) = E{xT (k)x(k)},
the sum of the powers of the tap voltages.

4.8.4 A Useful Geometrical Interpretation

The constrained LMS algorithm (4.372) has a simple geometrical interpretation [40] that
is useful for visualizing the error correcting property that prevents the weight vector
from deviating from the constraint condition. Even unavoidable computational errors due
to roundoff, truncation, or quantization are prevented from accumulating by the error
correcting property, which continuously corrects for any errors that may occur, whatever
their source may be.
In an error-free algorithm, the successive values of the NJ-dimensional weight vector
w all exactly satisfy the constraint equation (4.354) and therefore all lie on a constraint
plane  defined by
 = {w : CT w = f} (4.387)

This constraint plane [which is (NJ − J )-dimensional] may be indicated diagramatically

as shown in Figure 4-28.
Any vectors that point in a direction normal to the constraint plane are linear combi-
nations of the constraint matrix column vectors and therefore all have the form Ca, where
a is a constant vector whose components determine the linear combination. Consequently,
the initial weight vector in the algorithm (4.371),  = C(CT C)−1 f, points in a direction
normal to the constraint plane. In addition, the initial weight vector terminates exactly on
the constraint plane since CT  = f. As a result,  is the shortest vector that can terminate
on the constraint plane, as illustrated in Figure 4-28.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 222

222 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-28
Representation of
the constraint plane,
the constraint
subspace plane, and
initial weight vector
in the w-space. Initial weight vector
 = C(CTC)−1 f

Constraint plane
Λ = {w : CTw = f}

Constraint subspace plane

S = {w : CTw = 0}

FIGURE 4-29 Weight vector, w

Matrix P projects
vectors onto the
constraint subspace
Pw
plane.

Constraint subspace plane

S = {w : CTw = 0}

By setting the constraint weight vector f equal to zero, the homogeneous form of the
constraint equation

CT w = 0 (4.388)

defines a second plane [that is also (NJ−J )-dimensional] that passes through the coordinate
space origin. This constraint subspace is depicted in Figure 4-28.
The constrained LMS algorithm (4.371) premultiplies a certain vector in the W-space
by the matrix P, a projection operator. Premultiplication of any weight vector bythe matrix
P results in the elimination of any vector components perpendicular to the plane , thereby
projecting the original weight vector onto the constraint subspace plane as illustrated in
Figure 4-29.
The only factor in (4.371) remaining to be discussed is the vector y(k)x(k), which
is an estimate of the unconstrained gradient of the performance measure. Recall from
(4.355) that the unconstrained performance measure is 12 wT Rx x w and from (4.356) that
the unconstrained gradient is given by Rx x w. Since the covariance matrix Rx x is unknown
a priori, the estimate provided by y(k)x(k) is used in the algorithm.
The constrained optimization problem posed by (4.353) and (4.354) is illustrated
diagramatically in the w-space as shown in Figure 4-30. The algorithm must succeed
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 223

4.8 Gradient Algorithm with Constraints 223

FIGURE 4-30
Diagrammatic
representation of
wopt the constrained
optimization
w(k) problem showing
Contours of constant  = C(CTC)−1 f contours of constant
output power output power, the
wTRxxw constraint plane ,
the initial weight
vector , and the
optimum
constrained weight
vector wopt that
minimizes the output
power.

Λ = {w : CTw = f}

in moving from the initial weight vector  to the optimum weight vector wopt along the
constraint plane . The operation of the constrained LMS algorithm (4.371) in solving
the previously given constrained optimization problem is considered.
In Figure 4-31 the current value of the weight vector, w(k), is to be modified by
taking the unconstrained negative gradient estimate −y(k)x(k), scaling it by s , and
adding the result to w(k). In general, the resulting vector lies somewhere off the constraint
plane. Premultiplying the vector [w(k) − s y(k)x(k)] by the matrix P, the projection
onto the constraint subspace plane is obtained. Finally, adding  to constraint subspace
plane projection produces a new weight vector that lies on the constraint plane. This new

FIGURE 4-31
Operation of the
constrained LMS
w(k) − Δsy(k) x(k)
w(k+1) algorithm:
w(k + 1) = P[w(k) −
s y(k)x(k)]+ .
w(k)


P[w(k) − Δsy(k) x(k)]

S Λ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 224

224 CHAPTER 4 Gradient-Based Algorithms

weight vector w(k + 1) satisfies the constraint to within the numerical accuracy of the
computations. This error-correcting feature of the constrained LMS algorithm prevents
any computational errors from accumulating.
The convergence properties of the constrained LMS algorithm are closely related to
those for the unconstrained LMS algorithm and have been previously discussed. Likewise,
the same procedures that increased convergence speed for the LMS algorithm also work
for the constrained LMS algorithm.

4.9 SIMULATION RESULTS

The fundamental misadjustment versus speed of adaptation trade-off is less favorable for
the DSD algorithm than for the LMS algorithm [13]. Thus, it remains to determine the
improvement in this fundamental trade-off that can be realized using the Powell accelerated
gradient (PAG) algorithm compared with the LMS algorithm where eigenvalue spread in
the input signal covariance matrix is present.
Figure 4-32 depicts a four-element Y array having d = 0.787λ element spacing with
the desired signal located at 0◦ and three distinct narrowband Gaussian jamming signals
located at 15◦ , 90◦ , and 165◦ . The received signal covariance matrix is therefore given by

1 s 3
Ji  † 
Rx x = (uu† ) + vi vi + I (4.389)
n n i=1
n

where n denotes the thermal noise power (taken to be unity), s/n denotes the signal-to-
thermal noise ratio, and Ji /n denotes the jammer-to-thermal noise ratios for each of the
three jammers (i = 1, 2, 3). The elements of the signal steering vector u and the jammer
steering vectors vi are easily defined from the array geometry and the signal arrival angles.
The desired signal is a biphase modulated signal having a phase angle of either 0◦ or 180◦
with equal probability at each sample.
Two signal conditions were simulated corresponding to two values of eigenvalue
spread in the received signal covariance matrix. The first condition represents a respectable

FIGURE 4-32 y
Four-element
Y-array geometry
with signal and (0.394l, 0.682l)
jammer locations for
selected example.
J2 /n

J3 /n J1 /n

60° 60°
(−0.787l, 0) x
s/n = 10
l
87
0.7

(0.394l, −0.682l)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 225

4.9 Simulation Results 225

eigenvalue spread of λmax /λmin = 2440, whereas the second condition represents a more
extreme eigenvalue spread of λmax /λmin = 16, 700. Choosing the jammer-to-thermal noise
ratios to be J1 /n = 500, J2 /n = 40, and J3 /n = 200 together with s/n = 10 yields the
corresponding eigenvalues λ1 = 2.44 × 103 , λ2 = 4.94 × 102 , λ3 = 25.62, and λ4 = 1.0
for which the optimum output SNR is SNRopt = 15.0 (11.7 dB). Likewise, choosing
the jammer-to-thermal noise ratios to be J1 /n = 4000, J2 /n = 40, and J3 /n = 400
along with s/n = 10 yields the eigenvalues λ1 = 1.67 × 104 , λ2 = 103 , λ3 = 29, and
λ4 = 1.0 for which the optimum output SNR is also SNRopt = 15.0. In all cases the
initial weight vector setting was taken to be wT (0) = [0.1, 0, 0, 0]. Figures 4-33 and 4-34
show the convergence results for the LMS and PAG algorithms, respectively, plotted as
output SNR in decibels versus number of iterations for an eigenvalue spread of 2,440
(here output SNR means output signal-to-jammer plus thermal noise ratio). The expected
value of the gradient and v† Rx x v required by the PAG algorithm was taken over K = 9
data samples, and one iteration of the PAG algorithm occurred every nine data samples,

15.00 FIGURE 4-33

Output SNR versus
10.00 Optimum number of iterations
for LMS algorithm
5.00 with eigenvalue
spread = 2440 and
Output SNR (dB)

0 α L = 0.1.
−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3 2 3 4 5
Number of iterations

15.00 FIGURE 4-34

Output SNR versus
10.00 Optimum number of iterations
for PAG algorithm
5.00 with eigenvalue
spread = 2,440 and
Output SNR (dB)

0 K = 9.
−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3 2 3 4 5
Number of iterations
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 226

226 CHAPTER 4 Gradient-Based Algorithms

even though a weight update does not occur on some iterations. The loop gain of the
LMS loop was selected in accordance with (4.49), which requires that s tr(Rx x ) < 1 for
stability. Letting s tr(Rx x ) = α L and choosing α L = 0.1 therefore ensures stability while
giving reasonably fast convergence with an acceptable degree of misadjustment error. As
a consequence of the manner in which an iteration was defined for the PAG algorithm,
the time scale for Figure 4-34 is nine times greater than the time scale for Figure 4-33.
In Figure 4-34 the PAG algorithm is within 3 dB of the optimum after approximately
80 iterations (720 data samples), whereas in Figure 4-33 the LMS algorithm requires
approximately 1500 data samples to reach the same point. Furthermore, it may be seen
that the steady-state misadjustment for the two algorithms in these examples is very
comparable so the PAG algorithm converges twice as fast as the LMS algorithm for a
given level of misadjustment in this example.
Figures 4-35 and 4-36 show the convergence of the LMS and PAG algorithms for the
same algorithm parameters as in Figures 4-33 and 4-34 but with the eigenvalue spread =
16,700. In Figure 4-36 the PAG algorithm is within 3 dB of the optimum after approx-

FIGURE 4-35 15.00

Output SNR versus
number of iterations 10.00 Optimum
for LMS algorithm
with eigenvalue 5.00
spread = 16,700
Output SNR (dB)

and α L = 0.1. 0

−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3 2 3 4 5
Number of iterations

FIGURE 4-36 15.00

Output SNR versus
number of iterations 10.00 Optimum
for PAG algorithm
with eigenvalue 5.00
spread = 16,700
Output SNR (dB)

and K = 9. 0

−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3 2 3 4 5
Number of iterations
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 227

4.10 Phase-Only Adaptive Nulling Using Steepest Descent 227

imately 200 iterations (1,800 data samples), whereas the LMS algorithm in Figure 4-35
does not reach the same point even after 4,500 data samples. The degree of convergence
speed improvement that is attainable therefore increases as the degree of eigenvalue spread
increases.
A word of caution is needed concerning the expected convergence when using the PAG
algorithm. The simulation results given here were compiled for an array having only four
elements; as the number of array elements increases, the number of consecutive steps in or-
thogonal gradient directions also increases, thereby yielding significant direction errors in
the later steps (since estimation errors accumulate over the consecutive step directions). Ac-
cordingly, for a given level of misadjustment the learning curve time constant does not in-
crease linearly with N (as with LMS adaptation), but rather increases more rapidly. In fact,
when N > 10, the PAG algorithm actually converges more slowly than the LMS algorithm.

4.10 PHASE-ONLY ADAPTIVE NULLING

USING STEEPEST DESCENT
A phased array may or may not have variable amplitude weights but always has phase
shifters for beam steering and calibration. Since the phase shifters are already in place for
beam steering purposes, they can also serve as adaptive elements to adjust the antenna
pattern. The authors in [49] presented a beam-space algorithm for low sidelobe arrays that
relies on small phase shifts. When the direction of arrival for all the interfering sources is
known, then cancellation beams in the directions of the sources are subtracted from the
original pattern. Adaptation consists of matching the peak of the cancellation beam with
the culprit sidelobe and subtracting [50].
The steepest descent algorithm can also find phase settings that minimize the output
power [51]. Multiplying the change in output power due to a change in phase at each ele-
ment by a step size, μ, produces a gradient vector for the phase weights having components
given by

P(κ) − P(κ − 1)
δn (κ + 1) = δn (κ) + μ (4.390)
(κ)

where
P(κ) = array output power at time step κ
δn (κ) = phase shift at element n
(κ) = small phase increment
2
μ= )

N
[P(κ) − P(−1)]2
n=1

This algorithm worked well for phase-only simultaneous nulling of the sum and difference
patterns of an 80-element linear array of H plane sectoral horns [50]. A diagram of the
array appears in Figure 2-25 of Chapter 2. The sum channel has a 30 dB low sidelobe
Taylor taper and the difference channel has a 30 dB low sidelobe Bayliss taper. These
channels share eight-bit beam steering phase shifters. Experiments used a CW signal
incident on a sidelobe but no signal incident on the main beam. The cost function takes
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 228

228 CHAPTER 4 Gradient-Based Algorithms

FIGURE 4-37 0
Adapted sum
pattern for −10
simultaneous

Array Pattern (dB)

phase-only nulling in
−20
the sum and
difference channels.
−30 Quiescent

−40

−50
Adapted
−60
0 10 20 23 30 40
q (degrees)

FIGURE 4-38 0
Adapted difference
pattern for −10
simultaneous
Array Pattern (dB)

phase-only nulling in
−20
the sum and
difference channels.
−30 Quiescent

−40

−50
Adapted
−60
0 10 20 23 30 40
q (degrees)

into consideration both the sum and difference channel output powers; otherwise, a null
will not appear in both patterns. Minimizing the output power of both channels when
an interfering signal appears at θ = 23◦ results in the patterns shown in Figure 4-37 and
Figure 4-38. The desired nulls are place with relatively small deviations from the quiescent
patterns.

4.11 SUMMARY AND CONCLUSIONS

The LMS algorithm applies the method of steepest descent to the MSE performance
measure to obtain a simple implementation that is particularly well suited to continuous
signal communication systems. The LMS algorithm requires a reference signal that is
compared with the array output to form an error signal. This technique is useful for
adaptive arrays that are expected to distinguish between desired and undesired signals
on the basis of differences in modulation characteristics. The heart of an LMS loop is
the correlator (multiplier), which forms the product e(t)xi (t) that is required to obtain
the estimated gradient. For an N -element array, N correlators are therefore required to
implement the LMS algorithm to control each array element.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 229

4.11 Summary and Conclusions 229

The Howells–Applebaum adaptive processor is generally employed in situations

where the desired signal is usually absent (in contrast to the LMS algorithm, which re-
quires that the desired signal be present) and makes use of a beam steering vector instead
of a reference signal. We find that the Howells–Applebaum processor behavior is char-
acterized by a quiescent mode (when the desired signal is absent) and an adapted mode
(when the desired signal is present). The transient behavior of the algorithm is most eas-
ily described in terms of eigenvector beams, which can be analyzed by introduction of
a transformation to a normal coordinate system that diagonalizes the signal covariance
matrix. The processor exhibits the characteristic of sensitivity to eigenvalue spread, so
that strong interference sources are canceled rapidly while weak interference sources are
suppressed slowly. The dynamic range requirements of the circuitry used to implement
the Howells–Applebaum processor can be reduced by introduction into the control loop
of a hard limiter that modifies the effective signal covariance matrix but does not affect
the eigenvalue spread.
Different methods for constraining the maximum SNR algorithm to maintain a favor-
able desired signal response are discussed:
1. The use of pilot signals
2. Preadaptation spatial filters
3. Control loop spatial filters
4. Discrimination constraint for known desired signal power level
The close relationship that exists between the Howells–Applebaum maximum SNR pro-
cessor and the LMS algorithm makes the similar transient behavior characteristics of
these two algorithms hardly surprising. The susceptibility of the algorithm performance
to eigenvalue spread in the signal covariance matrix leads to a consideration of ways this
susceptibility can be reduced. One way of reducing this susceptibility and maintaining
fast transient response for all eigenvalue conditions is to employ a direct matrix inversion
(DMI) algorithm, which is introduced in the next chapter.
For some practical applications it may be undesirable to require N correlators as
the LMS algorithm does. In such cases, the alternative presented by the DSD algorithm,
which requires only direct performance index measurements (error power measurements
in the case of the MSE criterion), may be attractive. The DSD algorithm does not have
as favorable a convergence speed versus misadjustment trade-off as the LMS algorithm,
and both the DSD and LMS algorithms exhibit the same degree of convergence speed
sensitivity to eigenvalue spread in the input signal covariance matrix.
One way of reducing the convergence speed sensitivity to eigenvalue spread is to
employ an algorithm based on an accelerated gradient approach, provided the number of
degrees of freedom of the array processor is not too high. An algorithm based on the Powell
descent cycle was presented illustrating the improvement in the speed of convergence that
can be realized. Accelerated gradient approaches have certain implementation drawbacks,
however, and other methods (discussed in later chapters) may be preferred to obtain the
desired reduction in convergence speed sensitivity to eigenvalue spread. In applications
involving high energy, long-duration waveforms, it is often desirable to constrain the main
beam of the array so that undesirable signal waveform distortion will not occur.
A nice summary of the LMS and Howells–Applebaum algorithms is given by [52]

(I − γ Rn )wn + γ μb∗ Howells–Applebaum
wn+1 = (4.391)
(I − γ Rn )wn + γ rxd LMS
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 230

230 CHAPTER 4 Gradient-Based Algorithms

4.12 PROBLEMS
1. Misadjustment-Speed of Adaptation Trade-off for the LMS and DSD Algorithms [13] For the
LMS algorithm the total misadjustment in the steady state is given by (4.83), whereas the total
(minimum) misadjustment for the DSD algorithm is given by (4.317).
(a) Assuming all eigenvalues are equal so that (T pMSE )av = TMSE and that M = 10% for the
LMS algorithm, plot TMSE versus N for N = 2, 4, 8, . . . , 512.
(b) Assuming all eigenvalues are equal so that (T pMSE )av = TMSE and that (Mtot )min = 10% for
the DSD algorithm, plot TMSE versus N for N = 2, 4, 8, . . . , 512 and compare this result
with the previous diagram obtained in part (a).
2. Reference Signal Generation for LMS Adaptation Using Polarization as a Desired Signal
Discriminant [53] LMS adaptation requires a reference signal to be generated having properties
sufficiently correlated either to the desired signal or the undesired signal to permit the adaptive
system to preserve the desired signal in its output. Usually, the desired signal waveform properties
(e.g., frequency, duration, type of modulation, signal format) are used to generate the reference
signal, but if the signal and the interference can be distinguished by polarization, then polarization
may be employed as a useful discriminant for reference signal generation.
Let s denote a linearly polarized desired signal having the known polarization angle θ , and
let n denote a linearly polarized interference signal having the polarization angle α (where it is
known only that α = θ). Assume that the desired signal and interference impinge on two linearly
polarized antennas (A and B) as shown in Figure 4-39 where the antennas differ in orientation
by the angle β. The two signals va and vb may then be expressed as
va = s cos θ + n cos α
vb = s cos(β − θ ) + n cos(β − α)

(a) Show that by introducing the weight w 1 as illustrated in Figure 4-39, then the signal vb =
vb − w 1 va can be made to be signal free (have zero desired signal content) by setting
cos(β − θ )
w1 =
cos θ
so that
sin β
vb = n sin(α − θ ) = n f (α, β, θ )
cos θ
(b) From the results of part (a), show that

v0 = va − w 2 vb = s cos θ + n[cos α − w 2 f (α, β, θ )]

FIGURE 4-39 A va
Adaptive array
configuration for Linealy +
w1 Fixed v0
interference rejection polarized weight Output
on the basis of antennas _ __
vb + v′b
polarization using w2
B
LMS adaptation.

Adaptive
control
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 231

4.12 Problems 231

Since the output signal v0 contains both desired signal and interference components, corre-
lating it with the signal-free voltage vb yields a measure of the interference remaining in the
output signal, and the adaptive weight w 2 can then be adjusted to reduce the interference
content in the output.
(c) The error in the output signal v0 is the interference signal component that is still present
after w 2 bb is subtracted from va . Assume that the interference and the desired signal are
uncorrelated; then
 
E v02 = E{s 2 } cos2 θ + E{n 2 }[cos α − w 2 f (α, β, θ )]2
 
If the rate of change of w2 isproportional to ∂ E v02 /∂w 2 , show that the final value of the
weight occurs when ∂ E v02 /∂w 2 = 0 so that

cos α
w2 =
f (α, β, θ )

(d) With w 2 set to the final value determined in part (c), show that the steady-state system output
is given by

v0 = s cos θ

thereby showing that the system output is free of interference under steady-state conditions.
The previous result assumes that (1) knowledge of θ and the setting of w 1 are error free;
(2) the circuitry is noiseless; And (3) the number of input signals equals the number of
antennas available. These ideal conditions are not met in practice, and [53] analyzes the
system behavior under nonideal operating conditions.
3. Relative Sensitivity of the Constrained Look-Direction Response Processor to Perturbed
Wavefronts [44] The solution to the problem of minimizing the expected output power of
†
an array η = E{w† xx† w} subject to x0 w = f (or equivalently, η = f 2 ) is given by (4.366).
†
Since the look direction response is constrained by x0 w = f where x0 denotes a plane wave
signal arriving from the angle θ0 , the rationale behind this constraint is to regard the processor
as a filter that will pass plane waves from the angle θ0 but attenuate plane waves from all other
directions.
Let a perturbed plane wave be represented by x, having components

xk = Ak0 (1 + αk ) exp[ j (φk0 + ξk )]

where αk represents amplitude deviations, and ξk represents phase deviations from the nominal
plane wave signal x0 . Assume that αk , ξk are all uncorrelated zero-mean Gaussian random
variables with variances σα2 , σξ2 at each sensor of the array.
(a) Using η = w† E{xx† }w and the fact that
 
E{xi x ∗j } = xi0 x ∗j0 exp −σξ2 for i = j

and
   
E xi x ∗j = |xi0 |2 1 + σα2 for i = j

show that
  †    
η = exp −σξ2 w† x0 x0 w + 1 − exp −σξ2 + σα2 w† w
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 232

232 CHAPTER 4 Gradient-Based Algorithms

 
or η ∼
= f 2 + σξ2 + σα2 w† w for small values of σξ2 , σα2 assuming that |xi0 |2 = 1 (which is
the case for a planar wave).
(b) The result in part (a) can be rewritten as
  
η = f2 1+ σξ2 + σα2

where
 w† w
=
f2
Consequently, the ratio can be regarded as the relative sensitivity of the processor to the
perturbations whose variances are σξ2 , σα2 . Using the weights given by (4.169), show that
†
x0 R−2
x x x0
=  † 2
x0 Rx−1
x x0

The previous relative sensitivity can become large if the eigenvalues of Rx x have a large
spread, but if the eigenvalues of Rx x have a small spread then cannot become large.
4. MSLC Relationships [29] Show that (4.219) results from (4.215) by the following:
(a) Substitute (4.217) and (4.218) into (4.219).
(b) Let K = I + gRnn .
(c) Apply the matrix inversion lemma [(D.10) of Appendix D] to the resulting expression.
5. MSLC Relationships [29] Show that (4.234) follows from the steady-state relationship given
by (4.232).
6. Control Loop Spatial Filter Relationships [29] Apply the matrix inversion identity
Q−1 e
[Q + efT ]−1 e =
1 + fT Q−1 e
where Q is a nonsingular N × N matrix and e and f are N × 1 vectors to (4.247), and show that
(4.248) results.
7. Control Loop Spatial Filter Relationships [29] By substituting the relationships expressed by
(4.248) and (4.249) into (4.247), show that the steady-state weight vector relationship given by
(4.250) results.
8. Control Loop Spatial Filter Relationships [29] To show that (4.267) can be developed from
(4.265), define the ratio

 w† s s† w
SN =
w† Rx x w
where
* −w +  !
  s
w = , s =
1 s0
and
 !
Rx x rx x0
Rx x =
r†x x0 P0

(a) Show that |s0 − w† s|2 = w† s s† w .

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 233

4.12 Problems 233


(b) Show that Pe = w† Rx x w = w† Rx x w − w† rx x0 − r†x x0 w + P0 .
(c) Since w = [Rx x + aI]−1 rx x0 from (4.263) show that

w = wopt + w

where wopt = R−1 −1

x x rx x0 (the Wiener solution), and w = −aRx x w.

(d) Substitute w = wopt + w into Pe from part (b) and show that

Pe = Pe0 + w† Rx x w

where
† †
Pe0 = P0 − wopt rx x0 − r†x x0 wopt + wopt Rx x wopt
= P0 − r†x x0 R−1
x x rx x0
†
= P0 − wopt Rx x wopt

Hint: Note that

†
w† Rx x wopt + wopt Rx x w
− w† rx x0 − r†x x0 w = 0

because

w† Rx x wopt = w† Rx x R−1

x x rx x0

= w† rx x0

(e) Show that

N
|(Qrx x0 )i |2
†
w Rx x w = a 2
λi (λi + a)2
i=1

by using w = −aR−1
x x w.

Hint: Note that

r†x x0 Q−1 Q[Rx x + aI]−1 QQ−1 R−1

x x QQ
−1
· [Rx x + aI]−1 QQ−1 rx x0

is composed entirely of diagonalized matrices since

QR−1
xx Q
−1
=  and QQ−1 = I

9. Performance Degradation Due to Errors in the Assumed Direction of Signal Incidence [54]
The received signal vector can be represented by


m
x(t) = s(t) + gi (t) + n(t)
i=2

where

s(t) = desired signal vector = s(t)v1

gi (t) = directional noise sources = gi (t)vi
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 234

234 CHAPTER 4 Gradient-Based Algorithms

and

n(t) = thermal noise vector comprised of narrowband Gaussian noise

components independent from one sensor element to the next.

The vectors vi , i = 1, . . . , m can be regarded as steering vectors where

viT = [exp(− jωc τi1 ), exp(− jωc τi2 ), . . . , exp(− jωc τi N )]

and τik represents the delay of the ith directional signal at the kth sensor relative to the geometric
center of the array; ωc is the carrier signal frequency.
The optimum weight vector should satisfy

wopt = R−1
x x rxd

where Rx x is the received signal covariance matrix, and rxd is the cross-correlation vector
between the desired signal s and the received signal vector x. Direction of arrival information is
contained in rxd , and if the direction of incidence is assumed known, then rxd can be specified
and only R−1
x x need be determined to find wopt . If the assumed direction of incidence is in error,
however, then w = R−1 x x r̃xd where r̃xd represents the cross-correlation vector computed using
the errored signal steering vector ṽ1 .
(a) For the foregoing signal model, the optimum weight vector can be written as wopt =
†
[Sv1 v1 + Rnn ]−1 · (Sv1 ), where Rnn denotes the noise covariance matrix, and S denotes
the desired signal power per sensor. If v1 is in error, then r̃xd = (S ṽ1 ). Show that the
resulting weight vector computed using r̃xd is given by
S  †  † 
w= †
1 + Sv1 R−1 −1 −1 −1
nn v1 Rnn ṽ1 − Sv1 Rnn ṽ1 Rnn v1
1+ Sv1 R−1
nn v1

(b) Using the result obtained in part (a), show that the output signal-to-noise power ratio (when
only the desired signal and thermal noise are present) from the array is given by
 †
S w† E{ss† }w Sw† v1 v1 w
= =
N out
w† Rnn w w† Rnn w
†
S|v1 R−1
nn ṽ1 |
2
= † † † † † † †
nn ṽ1 − 2S|v1 Rnn ṽ1 | + v1 Rnn v1 [S {(v1 Rnn v1 ) × (ṽ1 Rnn ṽ1 ) − |v1 Rnn ṽ1 | } + 2S ṽ1 Rnn ṽ1 ]
ṽ1 R−1 −1 2 −1 2 −1 ∗ −1 −1 2 −1

(c) Use the fact that Rnn = σ 2 I and the result of part (b) to show that
 † 2
N v1 ṽ1 
S
S σ2 N2
=  † 2   † 2 
N out NS
2 v1 ṽ1  NS v1 ṽ1 
1+ 2 1− +2 1−
σ N 2 σ2 N2

(d) Show for a uniform linear array that

 †  | sin[(N π d/λc ) sin θ̃ ]|2
v1 ṽ1  =
| sin[(π d/λc ) sin θ̃]|2

where d represents the separation between sensors, and θ̃ represents the angular uncertainty
from boresight.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 235

4.13 References 235

4.13 REFERENCES
[1] B. Widrow, “Adaptive Filters,” in Aspects of Network and System Theory, edited by R. E.
Kalman and N. DeClaris, New York, Holt, Rinehart and Winston, 1970, pp. 563–587.
[2] F. R. Ragazzini and G. F. Franklin, Sampled-Data Control Systems, New York, McGraw-Hill,
1958.
[3] E. I. Jury, Sampled-Data Control Systems, New York, Wiley, 1958.
[4] J. T. Tou, Digital and Sampled-Data Control Systems, New York, McGraw-Hill, 1959.
[5] B. C. Kuo, Discrete-Data Control Systems, Englewood Cliffs, NJ, Prentice-Hall, 1970.
[6] B. Widrow, “Adaptive Filters I: Fundamentals,” Stanford Electronics Laboratories, Stanford,
CA, Rept. SEL-66-126 (Tech. Rept. 6764-6), December 1966.
[7] J. S. Koford and G. F. Groner, “The Use of an Adaptive Threshold Element to Design a Linear
Optimal Pattern Classifier,” IEEE Trans. Inf. Theory, Vol. IT-12, January 1966, pp. 42– 50.
[8] K. Steinbuch and B. Widrow, “A Critical Comparison of Two Kinds of Adaptive Classifi-
cation Networks,” IEEE Trans. Electron. Comput. (Short Notes), Vol. EC-14, October 1965,
pp. 737–740.
[9] F. W. Smith, “Design of Quasi-Optimal Minimum-Time Controllers,” IEEE Trans. Autom.
Control, Vol. AC-11, January 1966, pp. 71–77.
[10] B. Widrow, P. E. Mantey, L. J. Griffiths, and B. B. Goode, “Adaptive Antenna Systems,” Proc.
IEEE, Vol. 55, No. 12, December 1967, pp. 2143–2159.
[11] L. J. Griffiths, “Signal Extraction Using Real-Time Adaptation of a Linear Multichannel
Filter,” Ph.D. Disseration, Stanford University, December 1967.
[12] K. D. Senne and L. L. Horowitz, “New Results on Convergence of the Discrete-Time LMS
Algorithm Applied to Narrowband Adaptive Arrays,” Proceedings of the 17th IEEE Confer-
ence on Decision and Control, San Diego, CA, January 10–12, 1979, pp. 1166–1167.
[13] B. Widrow and J. M. McCool, “A Comparison of Adaptive Algorithms Based on the Methods
of Steepest Descent and Random Search,” IEEE Trans. Antennas Propag., Vol. AP-24, No. 5,
September 1976, pp. 615–638.
[14] R. T. Compton, Jr., “An Adaptive Array in a Spread-Spectrum Communication System,” Proc.
IEEE, Vol. 66, No. 3, March 1978, pp. 289–298.
[15] D. M. DiCarlo and R. T. Compton, Jr., “Reference Loop Phase Shift in Adaptive Arrays,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-14, No. 4, July 1978, pp. 599–607.
[16] P. W. Howells, “Intermediate Frequency Side-Lobe Canceller,” U.S. Patent 3202990, August
24, 1965.
[17] S. P. Applebaum, “Adaptive Arrays,” Syracuse University Research Corp., Report SPL TR
66-1, August 1966.
[18] L. E. Brennan and I. S. Reed, “Theory of Adaptive Radar,” IEEE Trans. Aerosp. Electron.
Syst., Vol. AES-9, No. 2, March 1973, pp. 237–252.
[19] L. E. Brennan and I. S. Reed, “Adaptive Space-Time Processing in Airborne Radars,” Tech-
nology Service Corporation, Report TSC-PD-061-2, Santa Monica, CA, February 24, 1971.
[20] A. L. McGuffin, “Adaptive Antenna Compatibility with Radar Signal Processing,” in
Proceedings of the Array Antenna Conference, February 1972, Naval Electronics Labora-
tory, San Diego, CA.
[21] L. E. Brennan, E. L. Pugh, and I. S. Reed, “Control Loop Noise in Adaptive Array Antennas,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-7, No. 2, March 1971, pp. 254–262.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 236

236 CHAPTER 4 Gradient-Based Algorithms

[22] W. F. Gabriel, “Adaptive Arrays—An Introduction,” Proc. IEEE, Vol. 64, No. 2, February
1976, pp. 239–272.
[23] L. E. Brennan, E. L. Pugh, and I. S. Reed, “Control Loop Noise in Adaptive Array Antennas,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-7, No. 2, March 1971, pp. 254–262.
[24] L. J. Griffiths, “A Simple Adaptive Algorithm for Real-Time Processing in Antenna Arrays,”
Proc. of IEEE, Vol. 57, No. 10, October 1969, pp. 1696–1704.
[25] R. T. Compton, Jr., Adaptive Antennnas Concepts and Performance, Englewood Cliffs, NJ,
Prentice-Hall, Inc., 1988.
[26] M. W. Ganz, “Rapid Convergence by Cascading Applebaum Adaptive Arrays,” IEEE Trans.
AES, Vol. 30, No. 2, April 1994, pp. 298–306.
[27] A. J. Berni, “Weight Jitter Phenomena in Adaptive Control Loops,” IEEE Trans. Aerosp.
Electron. Syst., Vol. AES-14, No. 4, July 1977, pp. 355–361.
[28] L. E. Brennan and I. S. Reed, “Effect of Envelope Limiting in Adaptive Array Control Loops,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-7, No. 4, July 1971, pp. 698–700.
[29] S. P. Applebaum and D. J. Chapman, “Adaptive Arrays with Main Beam Constraints,” IEEE
Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 650– 662.
[30] R. T. Compton, Jr., “Adaptive Arrays: On Power Equalization with Proportional Control,”
Ohio State University, Columbus, Quarterly Report 3234-1, Contract N0019-71-C-0219,
December 1971.
[31] C. L. Zahm, “Application of Adaptive Arrays to Suppress Strong Jammers in the Presence
of Weak Signals,” IEEE Trans. Aerosp. Electron. Syst., Vol. AES-9, No. 2, March 1973,
pp. 260–271.
[32] B. W. Lindgren, Statistical Theory, New York, Macmillan, 1960, Ch. 5.
[33] M. R. Hestenes and E. Stiefel, “Method of Conjugate Gradients for Solving Linear Systems,”
J. Res. Natl. Bur. Stand., Vol. 29, 1952, p. 409.
[34] J. D. Powell, “An Iterative Method for Finding Stationary Values of a Function of Several
Variables,” Comput. J. (Br.), Vol. 5, No. 2, July 1962, pp. 147–151.
[35] R. Fletcher and M. J. D. Powell, “A Rapidly Convergent Descent Method for Minimization,”
Comput. J. (Br.), Vol. 6, No. 2, July 1963, pp. 163–168.
[36] R. Fletcher and C. M. Reeves, “Functional Minimization by Conjugate Gradients,” Comput.
J. (Br.), Vol. 7, No. 2, July 1964, pp. 149–154.
[37] D. G. Luenberger, Optimization by Vector Space Methods, New York, Wiley, 1969, Ch. 10.
[38] L. Hasdorff, Gradient Optimization and Nonlinear Control, New York, Wiley, 1976.
[39] L. S. Lasdon, S. K. Mitter, and A. D. Waren, “The Method of Conjugate Gradients for
Optimal Control Problems,” IEEE Trans. Autom. Control, Vol. AC-12, No. 2, April 1967,
pp. 132–138.
[40] O. L. Frost, III, “An Algorithm for Linearly Constrained Adaptive Array Processing,” Proc.
IEEE, Vol. 60, No. 8, August 1972, pp. 926–935.
[41] A. E. Bryson, Jr. and Y. C. Ho, Applied Optimal Control, Waltham, MA, Blaisdell, 1969,
Ch. 1.
[42] E. J. Kelly and M. J. Levin, “Signal Parameter Estimation for Seismometer Arrays,”
Massachusetts Institute of Technology, Lincoln Laboratories Technical Report 339, January
1964.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 237

4.13 References 237

[43] A. H. Nuttall and D. W. Hyde, “A Unified Approach to Optimum and Suboptimum Processing
for Arrays,” U.S. Navy Underwater Sound Laboratory, New London, CT, USL Report 992,
April 1969.
[44] R. N. McDonald, “Degrading Performance of Nonlinear Array Processors in the Presence of
Data Modeling Errors,” J. Acoust. Soc. Am., Vol. 51, No. 4, April 1972, pp. 1186–1193.
[45] D. Q. Mayne, “On the Calculation of Pseudoinverses,” IEEE Trans. Autom. Control, Vol.
AC-14, No. 2, April 1969, pp. 204–205.
[46] K. Takao, M. Fujita, and T. Nishi, “An Adaptive Antenna Array Under Directional Constraint,”
IEEE Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 662–669.
[47] O. L. Frost, III, “Adaptive Least Squares Optimization Subject to Linear Equality Con-
straints,” Stanford Electronics Laboratories, Stanford, CA, DOC. SEL-70-053, Technical
Report TR6796-2, August 1970.
[48] J. L. Moschner, “Adaptive Filtering with Clipped Input Data,” Stanford Electronics Labora-
tories, Stanford, CA, Dec. SEL-70-053, Technical Report TR6796-1, June 1970.
[49] C. Baird, and G. Rassweiler, “Adaptive Sidelobe Nulling Using Digitally Controlled Phase-
Shifters,” IEEE Transactions on Antennas and Propagation, Vol. 24, No. 5, 1976, pp. 638–649.
[50] H. Steyskal, “Simple Method for Pattern Nulling by Phase Perturbation,” IEEE Transactions
on Antennas and Propagation, Vol. 31, No. 1, 1983, pp. 163–166.
[51] R. L. Haupt, “Adaptive Nulling in Monopulse Antennas,” IEEE Transactions on Antennas
and Propagation, Vol. 36, No. 2, 1988, pp. 202–208.
[52] D. S. De Lorenzo, J. Gautier, J. Rife, P. Enge, and D. Akos, “Adaptive Array Processing for
GPS Interference Rejection,” Proc. ION GNSS 2005, pp. 618–627.
[53] H. S. Lu, “Polarization Separation by an Adaptive Filter,” IEEE Trans. Aerosp. Electron. Sys.,
Vol. AES-9, No. 6, November 1973, pp. 954–956.
[54] C. L. Zahm, “Effects of Errors in the Direction of Incidence on the Performance of an Adaptive
Array,” Proc. IEEE, Vol. 60, No. 8, August 1972, pp. 1008–1009.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:47 238
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 239

CHAPTER

Direct Inversion of the Sample

Covariance Matrix 5
' $
Chapter Outline
5.1 The Direct Matrix Inversion (DMI) Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
5.2 Diagonal Loading of the Sample Covariance Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
5.3 Factorization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
5.4 Transient Response Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
5.5 Sensitivity to Eigenvalue Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
5.6 Summary and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
5.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
5.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
& %

The usefulness of an adaptive array often depends on its convergence rate. For example,
when adaptive radars simultaneously reject jamming and clutter while providing auto-
matic platform motion compensation, then rapid convergence to steady-state solutions is
essential. Convergence of adaptive sensor arrays using the popular maximum signal-to-
noise ratio (SNR) or least mean squares (LMS) algorithms depend on the eigenvalues
of the noise covariance matrix. When the covariance matrix eigenvalues differ by orders
of magnitude, then convergence is exceedingly long and highly example dependent. One
way to speed convergence and circumvent the convergence rate dependence on eigenvalue
distribution is to directly compute the adaptive weights using the sample covariance matrix
of the signal environment [1–3].

5.1 THE DIRECT MATRIX INVERSION

(DMI) APPROACH
The signals impinging on the receiving elements of an N-element adaptive array are
represented by the N-dimensional signal vector x, whose associated covariance matrix is
given by

Rx x = E{xx† } (5.1)

When the desired signal is absent, then only noise and interference are present and

Rx x = Rnn (5.2)
239
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 240

240 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

When the desired signal is present, then from Chapter 3 the optimal weight vector solution
is given by

wopt = R−1
x x rxd (5.3)

where rxd is the cross-correlation between the random vector x(t) and the reference signal
d(t). When the desired signal is absent, then the optimal weight vector solution is given by

wopt = R−1 −1 ∗
nn rxd = Rnn b (5.4)

where b∗ is the vector of beam steering signals matched to the target Doppler frequency
and angle of incidence. Note that specifying rxd is equivalent to specifying b∗ .
If the signal, clutter, and interference situation are known a priori, then the covariance
matrix is evaluated and the optimal solution for the adaptive weights is computed directly
using either (5.3) or (5.4). In practice the signal, clutter, and interference situation are
not known a priori, and furthermore the interference environment frequently changes
due to the presence of moving near-field scatterers, antenna motion, interference, and
jamming. Consequently, the adaptive processor continually updates the weight vector to
respond to the changing environment. In the absence of detailed a priori information, the
weight vector is updated using estimates of Rx x or Rnn , and rxd from a finite observation
interval and substituting into (5.3) or (5.4). This method for implementing the adaptive
processor is referred to as the DMI or sample matrix inversion (SMI) technique. The
estimates of Rx x , Rnn , and rxd are based on the maximum likelihood (ML) principle,
which yields unbiased estimates having minimum variance [4]. Although an algorithm
based on DMI theoretically converges faster than the LMS or maximum SNR algorithms,
the covariance matrix could be ill conditioned, so the degree of eigenvalue spread also
affects the practicality of this approach.
It is worth noting that when the covariance matrix to be inverted has the form of a
Toeplitz matrix (a situation that arises when using tapped-delay line channel processing),
then the matrix inversion algorithm of W. F. Trench [5] can be exploited to facilitate the
computation. The convergence results discussed in this chapter assume that all compu-
tations are done with sufficient accuracy to overcome the effects of any ill-conditioning
and therefore represent an upper limit on how well any DMI approach can be expected to
perform.

5.1.1 Use of the Sample Covariance Matrix

Suppose that the cross-correlation vector rxd (or equivalently that the beam steering vector
b∗ ) is known. The optimal weight vector estimate ŵ is formed by using

ŵ1 = R̂−1
x x rxd (5.5)

and assuming x(t) contains the desired signal, where R̂x x is the sample covariance estimate
of Rx x , or by using

ŵ2 = R̂−1 −1 ∗
nn rxd = R̂nn b (5.6)

if we assume x(t) does not contain the desired signal where R̂nn is the sample covariance
estimate of Rnn .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 241

5.1 The Direct Matrix Inversion (DMI) Approach 241

The array output SNR using ŵ1 or ŵ2 is written as

 
s ŵ†1 ss† ŵ1 r†xd R̂−1 † −1
x x ss R̂x x rxd
= = (5.7)
n 1 ŵ†1 Rnn ŵ1 r†xd R̂−1 −1
x x Rnn R̂x x rxd
 
s r†xd R̂−1 † −1
nn ss R̂nn rxd
= (5.8)
n 2 r†xd R̂−1 −1
nn Rnn R̂nn rxd

where s denotes the desired signal vector component of x [it will be recalled from Chapter 3
that s(t) = s(t)v]. The SNR (s/n)2 has meaning only during time intervals when a desired
signal is present; the weight adjustment in this case takes place when the desired signal
is absent. The “rate of convergence” of the two algorithms (5.5) and (5.6) depends on the
output SNR normalized to the optimum output SNR, SN o , compared with the number of
independent signal samples K used to obtain the required sample covariance matrices.
Assuming that all signals present at the array input are modeled as sample functions
from zero-mean Gaussian processes, then an ML estimate of Rx x (or Rnn when the desired
signal is not present) is formed using the sample covariance matrix given by

1
K
R̂xx = x( j)x† ( j) (5.9)
K j=1

where x( j) denotes the jth time sample of the signal vector x(t). Note that the assumption
of independent zero-mean samples implies E[x(i)x† ( j)] = 0 for i = j.
Since each element of the matrix R̂x x is a random variable, the output SNR is also a
random variable. It is instructive to compare the actual SNR obtained using ŵ1 and ŵ2 of
(5.5) and (5.6) with the optimum SNR obtained using (5.3) and (5.4) (S No = s† R−1 nn s), by
forming the normalized SNR as follows:
(s/n)1
ρ1 = (5.10)
SN o
(s/n)2
ρ2 = (5.11)
SN o
It can be shown [2] that the probability distribution of ρ2 is described by the incomplete
beta distribution given by
y
K!
Pr(ρ2 ≤ y) = (1 − u) N −2 u K +1−N du (5.12)
(N − 2)!(K + 1 − N )!
0

where
K = total number of independent time samples used in obtaining R̂nn
N = number of adaptive degrees of freedom
The probability distribution function of (5.12) contains important information concerning
the convergence of the DMI algorithm that is easily seen by considering the mean and the
variance of ρ2 . From (5.12) it follows that the average value of ρ2 is given by
K +2− N
E{ρ2 } = ρ 2 = (5.13)
K +1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 242

242 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

The variance of ρ2 is given by

(K + 2 − N )(N − 1)
var(ρ2 ) = (5.14)
(K + 1)2 (K + 2)

For fixed K and N, (5.14) suggests that var(ρ2 ) is independent of the amount of noise
the system must contend with and the eigenvalue spread of the noise covariance matrix.
Recalling that ρ is a normalized SNR, however, we see that both the actual SNR (s/n)
and the optimum SNR SN o are affected in the same way by any noise power increase, so
the normalized ratio remains the same, and the variance of the normalized ratio likewise
remains unchanged. Eigenvalue spread has no effect on (5.14), since this expression as-
sumes that the sample matrix inversion is computed exactly. As a result, (5.14) contains
only the sample covariance matrix estimation errors. The effect of eigenvalue spread on
the matrix inversion computation is addressed in a later section.
A plot of (5.13) in Figure 5-1 (we assume that N is significantly larger than 2) shows
that so long as K ≥ 2N , the loss in ρ 2 due to nonoptimum weights is less than 3 dB. This
result leads to the convenient rule of thumb that the number of time samples required to
obtain a useful sample covariance matrix (when the desired signal is absent) is twice the
number of adaptive degrees of freedom.
Next, consider the convergence behavior when the signal is present while estimating
w with (5.5). Rather than attempt to derive the probability distribution function of ρ1
directly, it is more convenient to exploit the results obtained for ρ2 by defining the random
variable

r†xd R̂−1 † −1
x x ss R̂x x rxd
ρ1 = (5.15)
r†xd R̂−1 −1 † −1
x x Rx x R̂x x rxd s Rx x s

which has the same probability distribution function as that of ρ2 . Knowing the statistical
properties of ρ1 and the relationship between ρ1 and ρ1 then enables the desired information
about ρ1 to be easily obtained. It can be shown that the relationship between ρ1 and ρ1 is

FIGURE 5-1 Plot 1

of ρ 2 versus K of
(5.13). We assume
N  2.

r 2 0.5

0
N 2N 3N 4N 5N 6N 7N 8N
K
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 243

5.1 The Direct Matrix Inversion (DMI) Approach 243

given by [3]
ρ1
ρ1 = (5.16)
S No (1 − ρ1 ) + 1
Since ρ1 has the same probability distribution function as ρ2 , it immediately follows that

E{ρ1 } < E{ρ2 } (5.17)

and

lim E{ρ1 } = E{ρ2 } (5.18)

S No →0

The inequality expressed by (5.17) implies that, on the average, the output SNR
achieved by using ŵ1 = R̂−1 −1
x x rxd is less than the output SNR achieved using ŵ2 = R̂nn rxd
(except in the limit as K → ∞, in which case both estimates are equally accurate). This
behavior derives from the fact that the presence of the desired signal increases the time
required (or number of samples required) to obtain accurate estimates of Rx x from the
sample covariance matrix compared with the time required to obtain accurate estimates
of Rnn when the desired signal is absent. The limit expressed by (5.18) indicates that for
S N o < 1, the difference in SNR performance obtained using ŵ1 or ŵ2 is negligible.
The mean of ρ1 in (5.16) can be expressed as the following infinite series [3]:
 ∞     
a  b b + 1 i + b − 1
E{ρ1 } = 1+ (−S No )i ··· ·
a+b i=1
a+b+1 a+b+2 a+b+i
(5.19)

where a = K − N + 2 and b = N − 1. The manner in which E{ρ1 } depends on S N o is

illustrated in Figure 5-2 for N = 4 with K as an independent parameter.
Since ŵ2 significantly outperforms ŵ1 when SN o  1, it is advantageous to remove
the signal components from x(t) before forming R̂x x . When the desired signal vector
component s(t) is completely known, then subtracting s(t) from the received signal vector

1.0 FIGURE 5-2 The

average normalized
K = 80 output SNR versus
0.8 optimum output
SNR for several
= 40 sample sizes.
0.6
E {r1}

= 20
0.4
−1 r = 10
ŵ1 = R̂xx xd
0.2 N=4

0
0.1 1.0 10 100
SNo
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 244

244 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

x(t) enables a desired signal-free noise vector to be formed that can be used to generate
the sample noise covariance matrix R̂nn . An improper procedure that is occasionally
suggested for eliminating the desired signal component is to form the sample covariance
matrix D̂, where
D̂ = R̂x x − ss† (5.20)
so that
E{D̂} = Rnn (5.21)
The procedure suggested by (5.20) is unsatisfactory for obtaining the fast convergence
associated with ŵ2 because even though E {D̂} = Rnn , the weight vector estimate obtained
using ŵ3 = D̂−1 rxd results only in an estimate that is a scalar multiple of ŵ1 = R̂−1
x x rxd
and therefore has the associated convergence properties of ŵ1 . This fact may easily be
seen by forming

ŵ3 = D̂−1 rxd

= [R̂x x − ss† ]−1 rxd (5.22)
1
= R̂−1
x x rxd
1 + s† R̂−1
xx s

The coefficient of R̂−1

x x rxd in (5.22) is a scalar, and therefore the output SNR resulting
from the use of ŵ3 is identical to that obtained using ŵ1 , so no transient performance
improvement results. If any transient response improvement were to be possible, it would
be necessary for ŵ3 of (5.22) to be a scalar multiple of R̂−1 −1
x x rxd . Since ŵ3 = α R̂x x rxd ,
however, the transient response of ŵ3 is the same as that of w1 , and a priori knowledge of
s as employed in (5.20) does not improve the DMI algorithm response.

5.1.2 Use of the Sample Covariance Matrix and the Sample

Cross-Correlation Vector
In many practical radar and communications systems the cross-correlation vector rxd (or
the beam-steering vector b∗ ) is not known a priori. An alternative approach determines
the optimal weight vector using

ŵ4 = R̂−1
x x r̂xd (5.23)

where r̂xd is the sample cross-correlation vector given by

1
K
r̂xd = x( j)d ∗ ( j) (5.24)
K j=1

The transient behavior of the DMI algorithm represented by ŵ4 of (5.23) is different from
that found for ŵ1 and ŵ2 of (5.5) and (5.6), respectively.
As an example, consider a four-element uniform array with λ/2 spacing. The desired
signal is incident at 0◦ and one interference signal is incident at 45◦ with σnoise = 0.01.
The amplitude and phase of the weights as a function of sample are shown in Figure 5-3
and Figure 5-4. The adapted pattern after 500 samples is shown in Figure 5-5. The weights
do not vary much after 50 samples. A null appears where a sidelobe peak used to be.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 25, 2010 21:9 245

5.1 The Direct Matrix Inversion (DMI) Approach 245

1 FIGURE 5-3
Magnitude of the
DMI weights versus
0.8 sample.

0.6
|wn|

0.4

0.2

0
0 100 200 300 400 500
Sample

200 FIGURE 5-4

Phase of the DMI
150 weights versus
sample.
100
∠wn(degrees)

50

−50

−100

−150

−200
0 100 200 300 400 500
Sample

FIGURE 5-5 The

DMI adapted pattern
5 for the four-element
array after 500
Adapted samples.
Directivity (dB)

−5

−15

Quiescent
−25

−90 −45 0 45 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 246

246 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

FIGURE 5-6
Adaptive array x1 (t) x2(t) xN(t)
configuration for
considering transient
w1 w2 wN
behavior of ŵ4 .

Σ
Reference
signal + −
w† x(t)
x0 (t)

Error
signal
e (t)

The transient response characteristics of ŵ4 are determined by considering the least
squares estimate of wopt based on K independent samples of the input vector x. It is
convenient in posing this problem to assume the adaptive array configuration shown in
Figure 5-6. This configuration represents two important adaptive array structures in the
following manner. If x0 (t) = d(t), the reference signal representation of the desired signal,
then the minimum mean square error (MMSE) and the maximum output SNR are both
given by the Wiener solution wopt . Likewise, if x0 (t) represents the output of a reference
antenna (usually a high-gain antenna pointed in the direction of the desired signal source),
then the configuration represents a coherent sidelobe canceller (CSLC) system, for which
obtaining the MMSE weight vector solution minimizes the output error (or residue) power
and hence minimizes the interference power component of the array output.
The transient response of ŵ4 given by (5.23) is characterized in terms of the output SNR
versus the number of data samples K used to form the estimates R̂x x and r̂xd . An alternate
way of characterizing performance that is appropriate for CSLC applications considers
the output residue power versus the number of data samples. System performance in
terms of the output residue power is considered. The output residue power (or MSE) is
given by

ξ(ŵ) = E|e(t)|2 = σ02 − ŵ† rxd − r†xd ŵ + ŵ† Rx x ŵ† (5.25)

and the estimated MSE (or sample MSE) may be expressed as

1
k
ξ̂ (ŵ) = | e(i) |2 = σ̂02 − ŵ† r̂xd − r̂†xd ŵ + ŵ† R̂x x ŵ (5.26)
K i=1

where σ02 is the signal power in the reference antenna, and

1
K
σ̂02 = x0 (i)x0∗ (i) (5.27)
K i=1

The transient behavior of the system is characterized by evaluating the statistical properties
of ξ(ŵ) for a given number of data samples K. Assume that x(i) and x0 (i) are sample
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 247

5.1 The Direct Matrix Inversion (DMI) Approach 247

functions from an (N + 1) variate, zero-mean, Gaussian random vector x having the

density function
 † −1
p(x ) = π − (N +1) |R xx |−1 exp − x R xx x (5.28)

where R xx is the (N + 1) × (N + 1) covariance matrix

 

x(t) † ∗
R xx = E x (t)|x0 (t)
x0 (t)
Rxx rxx0
= (5.29)
r†xx0 σ02

where rx x0 plays the role of rxd so that R̂x x is given by the estimates R̂x x , r̂xd , and σ̂02 .
The sample covariance matrix R̂x x has the following important properties:
1. The elements of R̂x x are jointly distributed according to the complex Wishart probability
density function [6]
|A| K −N −1 exp[−tr(R −1
x x A)]
p(A) = (5.30)
π 1/2(N +1)N (K )(K − 1) · · · (K − N )|R x x | K
where
A = K R̂x x
(k) = (k − 1)!

2. R̂xx is the ML estimate of R x x [6]. Therefore R̂x x , r̂xd , and σ̂02 are the ML estimates
of Rx x , rxd , and σ02 , respectively.
By using a series of transformations on the partitioned matrix R̂x x , the following
important results are obtained [7]:
1. The mean and variance of the sample MSE ξ̂ , realized using ŵ4 of (5.23) is given by
 
N
E{ξ̂ } = 1 − ξmin (5.31)
K
 
1 N
var{ξ̂ } = 1− ξmin
2
(5.32)
K K
where ξmin = σ02 − r†xd R−1
x x rxd .
2. The difference between the output residue power ξ(ŵ4 ) and the minimum output residue
power ξmin is a direct measure of the quality of the array performance relative to
the optimum. The normalized performance quality parameter (or “misadjustment”)
M = r 2 is defined as
ξ(ŵ4 ) − ξmin
r2 = (5.33)
ξmin
The parameter r is a random variable having the density function
K! r 2N −1
p(r ) = 2 · , 0<r <∞ (5.34)
(K − N )!(N − 1)! (1 + r 2 ) K +1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 248

248 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

The statistical moments of the misadjustment r 2 are easily obtained by recognizing

that the variable y defined by
1
y= (5.35)
1 + r2
is governed by an incomplete beta function distribution with parameters a = K − N +1
and b = N . Consequently, the mean and variance of r 2 are given by
N
E{r 2 } = (5.36)
K−N
and
NK
var {r 2 } = (5.37)
(K − N )2 (K − N − 1)
3. For the case when x0 (t) = d(t), define a normalized SNR as in (5.10), (5.11) for ŵ4
according to

(s/n)3 ŵ† ss† ŵ4

ρ3 = = † 4 (5.38)
S No ŵ4 Rnn ŵ4 s† R−1
nn s

The probability density function of ρ3 is difficult to evaluate in closed form, but the
mean and variance of ρ3 can be determined numerically using the relations given as
follows:
1
ρ3 =  2
 (5.39)
C + sin φ1
(1 + S No ) 2 − S No
C+sin φ1 cos2 φ2
√
where C = (1/r ) (s/n)3 and where the joint density function of r, φ1 , and φ2 is
given by
2 K! r 2N −1
p(r, φ1 , φ2 ) = · (sin φ1 )2N −2 (sin φ2 )2N −3
π (K − N )!(N − 2)! (1 + r 2 ) K +1
(5.40)

for 0 < r < ∞, 0 ≤ φ1 < π , 0 ≤ φ2 < π . Note that r , φ1 , and φ2 are statistically
independent and that r has the same density as (5.34). With the foregoing expressions
the numerical evaluation of E{ρ3 } can be obtained from
  
E{ρ3 } = P(r ) P(φ1 ) ρ3 P(φ2 )dφ2 dφ1 dr (5.41)
R 1 2

E {ρ32 } can likewise be obtained from (5.41) with ρ32 replacing ρ3 . The variance of ρ3
is then given by var{ρ3 } = E{ρ32 } − E 2 {ρ3 }.
4. The normalized MSE performance measure defined by

2K ξ̂
ξ̂ N = (5.42)
ξmin

is statistically independent of both ŵ and R̂x x .

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 249

5.1 The Direct Matrix Inversion (DMI) Approach 249

The results just summarized place limits on the transient performance of the DMI
algorithm. Let us first consider the results expressed by (5.36) and (5.37). From (5.36)
it is seen that the output residue power is within 3 dB of the optimum value after only
2N distinct time samples, or within 1 dB after 5N samples, thereby indicating rapid con-
vergence independent of the signal environment or the array configuration. We see this
rapid convergence property, however, applies directly to the interference suppression of
a sidelobe canceller (SLC) system assuming no desired signal is present when form-
ing R̂x x .
In communications systems the desired signal is usually present, and the SNR perfor-
mance measure is the primary quantity of interest rather than the MSE, ξ . Furthermore, in
radar systems the SLC is often followed by a signal processor that rejects clutter returns,
so only that portion of the output residue power due to radiofrequency (RF) interference
(rather than clutter) must be suppressed by the SLC system. Let us now show that the
presence of either the desired signal or clutter returns in the main beam of an SLC system
acts as a disturbance that tends to slow the rate of convergence of a DMI algorithm.
Consider the radar SLC configuration shown in Figure 5-7. The system consists of an
SLC designed to cancel only interference followed by a signal processor to remove clutter.
To simplify the discussion, assume the clutter power ξc2 received in the main antenna is
much larger than clutter power entering the low-gain auxiliary antennas so that clutter in
the auxiliary channels is neglected and ξmin = σc2 + ξ N0 where ξ N0 represents the minimum
output interference plus thermal noise power. Furthermore, assume that the clutter returns
are represented as a sample function from a stationary, zero-mean Gaussian process. It
then follows from (5.25), (5.33), and (5.36) that

E[ξ(ŵ) − ξmin ] N σ2
= 1+ c (5.43)
ξ N0 K−N ξ N0

From (5.43) it is evident that the presence of main beam clutter prolongs conver-
gence of the DMI algorithm by an amount that is approximately proportional to the main
antenna clutter power divided by the minimum output interference power. This slower
convergence is due to the presence of clutter-jammer cross-terms in r̂xd , which results in
noisier estimates of the weights. For rapid convergence, adaptation should be confined
to time intervals that are relatively clutter free, or a means for minimizing clutter terms
present in the estimate of r̂xd must be found.
The foregoing result is also applicable to communications systems where the main
antenna is pointed toward an active desired signal source. Let σs2 represent the desired
signal power received in the main channel, and assume the desired signal entering the
auxiliary channels can be neglected. Then, the SLC output signal to noise ratio SNR can

FIGURE 5-7 SLC

Main beam for interference
clutter Main suppression in a
antenna radar system.
1 Signal
SLC for e(t) processor
Sidelobe interference
interference for clutter
interference 2 suppression
minimized suppression

N
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 250

250 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

easily be shown to be given by

σs2
SNR = (5.44)
σs2r 2 + ξ N0 (1 + r 2 )

The maximum value of SNR is denoted by S N o where S No = σs2 /ξ N0 occurs when r 2 = 0

or ŵ = wopt . Normalizing (5.44) to S N o yields the expression for the normalized output
signal-to-interference plus noise ratio:

SNR 1
ρ4 = = (5.45)
SN o 1 + [1 + (SN o )]r 2
Note that 0 ≤ ρ4 ≤ 1 so the optimum value of ρ4 is unity. When the output signal to
interference plus noise ratio is small so that S No 1, then the probability distribution of
ρ4 is approximated by the following beta probability density function:
K!
P(ρ4 ) = (1 − ρ4 ) N −1 ρ4K −N (5.46)
(N − 1)!(K − N )!
On comparing (5.46) with the probability density function contained in (5.12), it imme-
diately follows that the two density functions are identical provided that N in (5.12) is
replaced by N + 1. It then follows from (5.13) for the CSLC system of Figure 5-6 that
K − N +1
E{ρ4 } = ; SN o 1 (5.47)
K +1
Hence, for small S N o only K = 2N − 1 independent samples are needed to converge
within 3 dB of S N o . For large SNR ( K /N ), the expected SNR (unnormalized) is
approximated by
K K
(SNR) ∼
= − 1; SN o  ; K >N (5.48)
N N
The presence of a strong desired signal in the main channel therefore slows convergence
to the optimum SNR but does not affect the average output SNR after K samples under
the conditions of (5.48).
Finally, consider the results given in (5.39) and (5.40) for the case x0 (t) = d(t)
(continuous reference signal present). For large S N o , the distribution of ρ3 in (5.39) is
approximated by the density function of (5.46) so that
K − N +1
E{ρ3 } ∼
= ; SN o  1 (5.49)
K +1
However, the rate of convergence decreases as S N o decreases below zero decibels. This
effect is illustrated by the plot of E{ρ3 } versus K in Figure 5-8.
The behavior described for a reference signal configuration is just the converse of that
obtained for the SLC configuration with the desired signal present. The presence of the
desired signal in the main channel of the SLC introduced a disturbance that reduced the
accuracy of the weight estimate and slowed the convergence. For the reference signal
configuration, however, the estimates r̂xd and R̂x x are highly correlated under strong
desired signal conditions, and the errors in each estimate tend to compensate each other,
thereby yielding an improved weight estimate and faster convergence.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 251

5.2 Diagonal Loading of the Sample Covariance Matrix 251

1.0 FIGURE 5-8

0.9 SNo = 10 dB E{ρ3 } of (5.38)
versus number of
0.8
independent
0.7 = −10 dB samples for N = 8
= 0 dB
0.6 and selected values
E {r 3}

0.5 of SN o.
0.4
= −20 dB
0.3
0.2
0.1

10 30 50 100 150 200 250

Number of independent samples (K)

5.2 DIAGONAL LOADING OF THE SAMPLE

COVARIANCE MATRIX
Diagonal loading augments the diagonal elements of the signal covariance matrix with
either a positive or a negative term and in effect alters the background noise level in which
the array is operating [8,9]. Positive loading speeds the weight convergence (although to
a suboptimal solution) and avoids very high pattern sidelobes that result from inadequate
estimation of the ambient noise; the price paid for these advantages is the reduced ability to
suppress weak interference signals. Negative loading creates deeper nulls in the direction of
interfering signals (which is useful in television reception from geosynchronous satellites
where weak interference from adjacent satellites can produce “ghosts”); negative loading
produces a suboptimal SINR value yet also produces a higher signal-to-interference ratio
(SIR) value than that of the unloaded result because the resulting weights will produce
deeper nulls on weak interfering signals [10].
Consider the example of a four-element linear array with half-wavelength spacing
where there are two weak signals incident on the array: a desired signal having a 10 dB
SNR incident from the broadside direction (θd = 0◦ ) and an interfering signal having a 0 dB
interference-to-noise ratio (INR) arriving from θi = 35◦ . Figure 5-9 shows the results of

50 FIGURE 5-9 SINR

and SIR vs. K, the
SIR = 40.778
40 number of samples.
Input SNR = 10 dB,
SIR and SINR in dB

30
Input IRN = 0 dB,
No Diagonal 3 dB
Loading. From Ganz
20 SINR = 15.930 et al., IEEE Trans.
Ant. & Prop., March
10 1990.
SIR-theory SINR-asymptotic value
0 SINR-theory SIR-100 run simulation
SIR-asymptotic value SINR-100 run simulation
−10
5 10 100 1000 10000
Sample size = K
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 252

252 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

FIGURE 5-10 50
SINR and SIR vs. K,
the number of 40 SIR = 36.364
samples. Input SNR

SIR and SINR in dB

= 10 dB, Input IRN 30
= 0 dB, Positive
Diagonal 3 dB
20
Loading. From Ganz SINR = 15.942
et al., IEEE Trans.
Ant. & Prop., March 10
1990. SIR-theory SINR-asymptotic value
0 SINR-theory SIR-100 run simulation
SIR-asymptotic value SINR-100 run simulation

−10
5 10 100 1000 10000
Sample size = K

FIGURE 5-11 50
SINR and SIR vs. K,
SIR = 45.867
the number of 40 SIR-theory
samples. Input SNR SINR-theory
SIR and SINR in dB

= 10 dB, Input IRN 30

SIR-asymptotic value

= 0 dB, Negative
Diagonal 3 dB
20
Loading. From Ganz SINR = 15.946
et al., IEEE Trans.
Ant. & Prop., March 10
1990. SINR-asymptotic value
0 SIR-100 run simulation
SINR-100 run simulation

−10
5 10 100 1000 10000
Sample size = K

the SIR and SINR versus number of samples for the case when there is no diagonal loading.
Figure 5-10 shows the same results when 3 dB of positive loading is added to the covariance
matrix (i.e., σ 2 is added to each diagonal element). The SIR asymptote is approximately
5.8 dB lower than it was without loading, but the SINR asymptote is essentially unchanged.
The positive loading has decreased the number of samples required for convergence of the
SINR curves to their asymptotic value. Figure 5-11 shows the results of the SIR and SINR
versus number of samples for the case when 0.5σ 2 is subtracted (negative loading) from
each diagonal element of the covariance matrix. The SINR asymptote remains essentially
unchanged, but the SIR asymptote is 5.9 dB higher than it was without loading. Negative
loading achieves a better SIR but a slower convergence to the asymptotic value compared
with the unloaded case. Notice, however, that with negative loading both the output SIR
and SINR curves are quite erratic when a small number of samples are taken; this behavior
is due to the variance of the output powers is large for small sample sizes.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 253

5.3 Factorization Methods 253

5.3 FACTORIZATION METHODS

For the system of equations, Rw = v, where R is N × N, and w and v are N × 1 vectors.
The vector w is computed by factoring the matrix R rather than directly computing R−1
since the direct inversion process can be unstable. Three R factorization approaches for
obtaining w are presented: triangularization, Cholesky factorization, and spectral (or U-D,
L-D) factorization.

5.3.1 Triangularization of R
Suppose there is a transformation matrix, Q, such that QR = U where U is an upper
triangular matrix. In this case it is possible to solve the system Uw = Qv = y for w
directly without finding a matrix inverse by means of a simple back-substitution algorithm
[11]. The following back-substitution algorithm computes the elements of w directly.
For j = N, N − 1, . . . , 1, compute
⎛ ⎞
N
w( j) = ⎝y( j) − u( j, k)w(k)⎠ u( j, j) (5.50)
k= j+1

The algorithm begins with w(N) = y(N)/U(N,N) and progresses backward until
computing element w(1). Suppose instead that QR = L where L is a lower triangular
matrix; then a simple forward-substitution algorithm (which is an analog of the forward-
substitution algorithm) yields the desired solution for the vector w.
Sometimes there is a need to explicitly have the inverse Z = U−1 . A simple algorithm
for obtaining the elements of Z (which is also upper triangular) is given by

z(1, 1) = 1/u(1, 1) (5.51)

Then for j = 2, . . . , N, evaluate the following two equations:

z( j, j) = 1/u( j, j) (5.52)
 j−1 

z(k, j) = − z(k, m)u(m, j) u( j, j), k = 1, . . . , j − 1 (5.53)
m=k

We now need to determine a transformation matrix, Q, which succeeds in triangularizing

the original matrix, R. There are two methods for obtaining such a transformation matrix:
(1) Givens rotations; and (2) Householder transformations.

[Link] Givens Rotations [12]

The original R matrix is N × N, and it is desired to find a transformation Q such that the
product QR is an upper triangular matrix, U. One way of accomplishing this result is to
annihilate the elements of R one element at a time; first progress up the first column of R
starting with rN1 and ending with r21 and continue across N − 1 columns. Therefore, there

N −1 
γ N, N odd
are a total of (N − 1) + (N − 2) + · · · + 2 + 1 = k= ,
γ (N − 1), N even
  k=1
N
where γ = Int annihilations and Int[ ] denotes the integer remaining after rounding
2
the number in the bracket to the next lower integer.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 254

254 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

Givens rotations annihilate one element at a time using the planar rotation transfor-
mation
 ∗ 
c s
G= (5.54)
−s ∗ c
 
v
With the complex vector v = 1 , we find that
v2
 ∗ 
c v1 + sv2
Gv = (5.55)
−s ∗ v1 + cv2

and require that s ∗ v1 = cv2 with the unitary condition |c|2 + |s|2 = 1. It follows that
v1
c=  (5.56)
|v1 |2 + |v2 |2
v∗2
s=  (5.57)
|v1 |2 + |v2 |2

  v1 
Gv = v with v1 = (5.58)
0

where v1  = |v1 |2 + |v2 |2 (5.59)

Consider the complex column vector v = [v1 . . . vm , . . . vn , . . . , vK ]T . For a K × K

matrix, A, to eliminate the a(m,n) element requires altering the elements in the mth row
and the nth column using the following Givens transformation:
⎡ ⎤
1 0 0
⎢0 1 0⎥
⎢ ⎥
⎢0 ∗
cmm smn 0 ⎥
G(m, n) = ⎢⎢ ⎥, (5.60)
⎢0 1 0⎥ ⎥
⎣ −snm∗
cnn ⎦
0 1
∗ vm
where cmm = cnn =  (5.61)
|vm |2 + |vn |2
∗ v∗n
and snm = smn = . (5.62)
|vm |2 + |vn |2

Then G(m, n)v = [v1 . . . |vm |2 + |vn |2 . . . 0, . . . vK ]T (5.63)
⎡ ⎤
a11 a12 a13
As an example, let A = ⎣ a21 a22 a23 ⎦. To annihilate element a31 requires affecting
a31 a32 a33
elements in the first and third rows (rotate a31 into a11 ) so that
⎡ ∗ ⎤ ⎡   ⎤
c1 0 s1 r1 a12 a13
G1 = ⎣ 0 1 0 ⎦ and G1 A = ⎣ a21 a22 a23 ⎦ (5.64)
∗  
−s1 0 c1 0 a32 a33
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 255

5.3 Factorization Methods 255

a11 a∗
where the 3,1 product element becomes zero since c1 a31 = s1∗ a11 , c1 = , s1 = 31
 r1 r1
and r1 = |a11 | + |a31 | . The next step is to annihilate element a21 , which is done by
2 2

rotating element a21 into element r1 using the Givens rotation

⎡ ∗ ⎤ ⎡ ∗    ⎤
c2 s2 0 c2 a11 + s2 a21 c2∗ a12 + s2 a23 c2∗ a13 + s2 a23
⎢ ⎥
G2 = ⎣−ss∗ c2 0⎦ so that G2 A = ⎣−s2∗ a11 + c2 a21 −s2∗ a12 + c2 a22 −s2∗ a13 + c2 a23 ⎦
  

 
0 0 1 0 a32 a33
(5.65)

which requires that


r1 a∗   2
|r1 |2 + a12  .

c2 a21 = s2∗ a11 , where c2 = , s2 = 21 and r11 = (5.66)
r11 r11

The last step is then to annihilate element a32 with Givens rotation G3 by rotating element
 
a32 into element a22 . The complete triangularization QR = U is then accomplished with
the matrix Q, where Q is the product of all the successive Givens rotation matrices,
i.e., Q = Gp Gp−1 , . . . , G2 G1 where G1 corresponds to the matrix required to annihilate
element rN,1 , and Gp corresponds to the matrix required to annihilate element rN,N−1 .
Obviously, annihilating the upper right off-diagonal entries in R will result in the lower
triangular analog, L, of the upper triangular result.

[Link] Householder Transformations [12]

Householder transformations annihilate a complete column of entries to upper triangular-
ize R. The basic idea is shown as follows:
 
s1 r1T
T1 R1 = (5.67)
0 R2
where R1 is N × N, s1 is a scalar, R2 is (N − 1) × (N − 1), “0” is a (N − 1) × 1 column
 T
r1
vector, and r1 is a 1 × (N − 1) row vector. The matrix defined by
T
is denoted
R2
by R̃1 . The transformation T1 is not explicitly computed, although it can be obtained if
desired. Having annihilated all the elements in the first column (except for “s1 ”), the next
step applies a second transformation to annihilate the elements in the second column as
follows:
 
s2 r2T
T2 R2 = (5.68)
0 R3
This process continues until the triangularization of the original matrix is complete. The
overall Householder transformation can then be written as
    
I 0 I N −3 0 I 0
Th = N −2 ··· 1 T (5.69)
0 T N −1 0 T N −2 0 T2 1
where Ik denotes the k × k identity matrix. Clearly, the TN−1 transformation is just a 2 × 2
Givens rotation that annihilates the N, N − 1 element of the R matrix. The various Tk
transformations are then orthogonal (both symmetric and orthogonal so that T−1 k = Tk ).
T

The operation for computing the triangularization of a matrix, R, with elements r(i,j)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 256

256 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

involves a process described by the matrix operation

s
Tu R = R̃ (5.70)
0

The algorithms for computing s and R̃ are then given by [12]

⎛ 2
⎞
N
s = −sgn(r(1, 1)) ⎝ [r(i, 1)] ⎠ (5.71)
i=1
u(1) = r(1, 1) − s (5.72)
u(i) = r(i, 1), for i = 2 . . . N (5.73)
β = 1/(su(1)) (5.74)

For j = 2, . . . , N, evaluate (5.75) and (5.76) (apply T to the successive columns of R)


N
γ =β · u(i)R(i, j) (5.75)
i=1
R̃(i, j − 1) = R(i, j) + γ u(i), i = 1, . . . , N (5.76)

where R̃(i,j) may be replaced by R(i,j). This completes the reduction of an N × N matrix
to a triangular form.

5.3.2 Cholesky Factorization

To compute the lower triangular Cholesky factorization R = LLT , the original problem
Rw = v is written as LLT w = v. Letting LT w = u yields a pair of rectangular matrix
equations to solve:

Lu = v is solved for u by a forward substitution algorithm

LT w = u is solved for w by a backward substitution algorithm

The Cholesky factorization of R with positive diagonal elements is given by the following
algorithm [11,13]:

For j = 1, 2, . . . , N − 1
L( j, j) = R( j, j)1/2 (5.77)
For k = j + 1, . . . , N
L(k, j) = R(k, j)/L(j, j) (5.78)
For i = k, . . . , N
R(i, k) = R(i, j) − L(i, j)L(k, j) (5.79)
end for
L(N, N) = P(N , N )1/2 (5.80)

An analogous algorithm starting with U(N, N) obviously applies for the upper triangular
Cholesky factorization.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 257

5.4 Transient Response Comparisons 257

5.3.3 Spectral (U-D or L-D) Factorization [11,14]

Consider the spectral factorization R = MMT (or LDLT , UDUT ), where  is diagonal
(consisting of the eigenvalues of the matrix R arranged in descending order), the columns
of M are the normalized eigenvectors corresponding to each of the eigenvalues, and U is
upper triangular with ones along the principal diagonal where the elements of D and U
are given by the following algorithm:
The diagonal elements of D are denoted by d(j)

for j = N, N − 1, . . . , 2
d(j) = R(j, j) (5.81)
U(j, j) = 1
for k = 1, 2, . . . , j − 1
U(k, j) = R(k, j)/d(j) (5.82)
for i = 1, 2, . . . , k
R(i, k) = R(i, k) − U(i, j)U(k, j)d(j) (5.83)
end for
U(1, 1) = 1 and d(1) = R(1, 1) (5.84)

The U-D factorization given by (5.81)–(5.84) as well as the L-D factorization is different
from the spectral factorization involving eigenvalues and normalized eigenvectors. Spec-
tral factorization renders finding the desired inverses particularly simple, thereby making
the additional computation required by the accompanying eigenvalue and eigenvector
computations worthwhile. Note that the U-D and L-D factorizations are also different
from each other.
The matrix R−1 is formed by inspection from the spectral factorization as follows.
Since R−1 = [MMT ]−1 = (MT )−1 −1 M−1 where the elements of −1 are merely
1/λ(i) for i = 1, . . . , N , MH = M−1 and (MT )−1 = M since M is unitary. The extra
computation involved in finding the eigenvalues and associated eigenvectors is rewarded
by the ease in finding the desired inverses.
The inverses required by the U-D factorization are U−1 , D−1 , where the elements of
−1
D are merely the inverse of the elements of D, and the inverse of U is given by equations
(5.51)–(5.53).

5.4 TRANSIENT RESPONSE COMPARISONS

The transient response characteristics of the DMI algorithms corresponding to ŵ1 of (5.5)
and ŵ4 of (5.23) may be obtained by examining
E{output SNR}
ρ=
S No

as a function of K (the number of independent samples used in obtaining R̂x x and r̂xd ) for
selected signal conditions. Assuming an input interference-to-signal ratio of 14 dB and an
input signal-to-thermal noise ratio of 0 dB, ρ was obtained for ŵ1 and ŵ4 by averaging
50 independent responses for a four-element linear array for which d/λ = 12 . A single
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 258

258 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

FIGURE 5-12 1.0

Transient response
of ŵ1 and ŵ4 for
SN o = 3.8 and one
ŵ1 LMS with rxd given
interference signal
located 75◦ away
from the desired
signal. LMS with d(t) given
r 0.1
ŵ4

0.01
1 10 100 400
Number of samples (K )

directional interference source was assumed that was separated from the desired signal
location by 75◦ . The resulting transient response is given in Figure 5-12.
The transient response obtained using ŵ1 and ŵ4 can be compared with the transient
response that would result using the LMS algorithm having an iteration period equal to the
intersample collection period of the DMI algorithm by considering the behavior of ρ(K ).
For the DMI approach, K represents the number of independent signal samples used to
compute the sample covariance matrix, whereas for the LMS algorithm K represents the
number of iterations completed. This comparison of the DMI and LMS algorithms is not
satisfactory for the following reasons:
1. The transient response of the LMS algorithm depends on the selection of the step size,
which can be made arbitrarily small (thereby resulting in an arbitrarily long transient
response time constant).
2. The transient response of the LMS algorithm depends on the starting point at which the
initial weight vector guess is set. A good initial guess may result in excellent transient
response.
3. Whereas the variance of ρ(K) decreases as K increases with the DMI approach, the
variance of ρ(K) remains constant (in the steady state) as K increases for the LMS
algorithm since the steady-state variance is determined by the step size selection.
Nevertheless, a comparison of ρ(K ) behavior does yield an indication of the output SNR
response speed and is therefore of some value.
The LMS algorithm convergence condition (4.49) is violated if the step size s exceeds
1/PIN , where PIN = total array input power from all sources. Consequently, by selecting
s = 0.4/PIN , the step size is only 4 dB below the convergence condition limit, the LMS
control loop is quite noisy, and the resulting speed of the transient response is close to the
theoretical maximum. Since the maximum time constant for the LMS algorithm transient
response is associated with the minimum covariance matrix eigenvalue

1
τmax ∼
= (5.85)
ks tλmin
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 259

5.4 Transient Response Comparisons 259

and λmin ≥ σ 2 , the thermal noise power variance,1 it follows that

PIN
τmax ∼
= (5.86)
0.4σ 2
With a four-element array, PIN = 4(PD + σ 2 + PI ),
where
PD = desired signal power
σ 2 = thermal noise power
PI = interference signal power
Therefore, if PI /PD = 14 dB and PD /σ 2 = 0 dB, then PIN /σ 2 = 108, and τmax = 270
samples.
The transient response time constant of the desired signal power is proportional to
the eigenvalue λ1 associated with the desired signal power. Since in the current example
λ1 ∼
= 4PD , the time constant associated with the desired signal power is approximately
1 PIN
τ1 ∼
= = = 67 samples (5.87)
λ
s 1 (0.4)4PD

A convenient starting point for the weight vector with the LMS algorithm is w0T =
[1, 0, 0, 0], since this corresponds to an omnidirectional array pattern. For the foregoing
conditions, the behavior of ρ(K ) resulting from the use of ŵ3 , ŵ4 and the two versions of
the LMS algorithm given by
w(k + 1) = w(k) + ks t [rxd − x(k)x† (k)w(k)], rxd given (5.88)
and
w(k + 1) = w(k) + ks t [x(k) d ∗ (k) − x(k)x† (k)w(k)], d(k) given (5.89)
was determined by simulation. The results are illustrated in Figure 5-12 where for the
specified conditions S N o = 3.8.
The results of Figure 5-12 indicate that the response resulting from the use of the DMI
derived weights ŵ1 and ŵ4 is superior to that obtained from the LMS derived weights.
Whereas the initial response of the LMS derived weights indicated improved output SNR
with increasing K, this trend reverses when the LMS algorithm begins to respond along
the desired signal eigenvector, since any decrease in the desired signal response with-
out a corresponding decrease in the thermal noise response causes the output signal to
thermal noise ratio to decrease. Once the array begins to respond along the thermal noise
eigenvectors, then ρ again begins to increase.
The undesirable transient behavior of the two LMS algorithms in Figure 5-12 can
be avoided by selecting an initial starting weight vector different from that chosen for
the foregoing comparison. For example, by selecting w(0) = rxd the initial LMS algo-
rithm response is greatly improved since this initial starting condition biases the array
pattern toward the desired signal direction thereby providing an initially high output SNR.
Furthermore, by selecting w(0) = αrxd where α is a scalar constant the initial starting
condition can also result in a monotonically increasing ρ(K ) as K increases for the two

1 It is assumed that each weighted channel contains thermal noise (with noise power σ 2 ) that is uncorrelated

between channels.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 260

260 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

LMS algorithms since by appropriately weighting rxd , the magnitude of the initial array
response to the desired signal can be made small enough so that as adaptation proceeds,
the resulting ρ(K ) always increases.
Improvement of the transient response through judicious selection of the initial starting
condition can also be introduced into the DMI derived weight vectors as well. For example,
by selecting
⎡ ⎛ ⎞⎤−1
1 K
ŵ1 = ⎣ ⎝ x( j)x†( j) + αI⎠⎦ rxd (5.90)
K j=1

then even before forming an estimate of Rx x the weight vector is biased toward the
desired signal direction, and the transient responses corresponding to ŵ1 and ŵ4 in Fig-
ure 5-12 is greatly improved. Nevertheless, even the improved transient response for an
LMS algorithm obtained by appropriately biasing the initial weight vector is slower than
the convergence speed of DMI.
It is also interesting to obtain the transient response of ρ(K ) with two interference sig-
nals present. Assume one interference-to-signal ratio of 30 dB and a second interference-
to-signal ratio of 10 dB, where the stronger interference signal is located 30◦ away from the
desired signal, the weaker interference signal is located 60◦ away from the desired signal,
with all other conditions the same as for Figure 5-12. The resulting transient response for
the two DMI derived weight vectors and for the two LMS algorithms [with d(t) given and
with rxd given] with initial starting weight vector = [1, 0, 0, 0] is illustrated in Figure 5-13.
The presence of two directional interference sources with widely different power levels
results in a wide eigenvalue spread and a consequent slow convergence rate for the LMS
algorithms. Since PIN /λ1 is now 40 times larger than for the conditions of Figure 5-12, the
time constant τ1 is now 40 times greater than before. The DMI derived weight transient
response, however, is virtually unaffected by this eigenvalue spread.
The principal convergence results for DMI algorithms under various array configura-
tions and signal conditions can be conveniently summarized as shown in Table 5-1. The
derivation of these results may be found in [2,3].

FIGURE 5-13 1.0

Transient response
of ŵ1 and ŵ4 for
SN o = 0.4158 and ŵ4
two interference ŵ1
signals located 30◦
and 60◦ away from
the desired signal.
r 0.1 LMS algorithms

0.01
1 10 100 400
Number of samples (K )
 Monzingo-7200014
book

TABLE 5-1 Comparison of DMI Algorithm Convergence Rates for Selected Array Configurations and Signal Conditions

Approximate Number of Independent Samples

Required for Convergence (to within 3 dB
Array Configuration ŵopt Signal Conditions Performance Measure of Optimum)
Desired signal absent Minimum MSE 2N ,N = number of array elements
 
S No
Desired signal present in main Maximum SNR, SN o 2N 1 + , SN o  1
2
beam only
ISBN : XXXXXXXXXX

2N , SN o 1
Sidelobe canceller R̂−1
x x r̂x x0  
σc2
Clutter returns in main beam only Minimum output 2N 1 + , σc2 = main channel clutter power
2ξ N0
(interference + noise)
power, ξ N0
5.4

Known desired signal present 2N , Maximum SNR, S N o SN o  1

 
R̂−1
x x r̂xd d(t) = s(t) 1
2N 1 + , SN o 1
November 24, 2010

2SN o

Fully adaptive array Desired signal direction of arrival Maximum SNR, SN o 2N − 3

known but desired signal absent
18:51

R̂−1
x x rxd Desired signal direction of arrival Maximum SNR, SN o 2N , SN o 1
 
known and desired signal present S No
2N 1 + , SN o  1
261

2
Transient Response Comparisons
261
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 262

262 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

FIGURE 5-14 12
Output SNR 11 4 Bits
degradation versus
10 6 Bits
eigenvalue spread

dB Degradation (output SNR

for a specified 9

averaged over 50 trials)

number of available 8
bits. 7
6
5 10 Bits
4
3
2
1
0
0 10 20 30 40 50 60
Eigenvalue spread (dB)

5.5 SENSITIVITY TO EIGENVALUE SPREAD

Despite the fact that the convergence speed of a DMI algorithm is insensitive to eigenvalue
spread in Rx x , the accuracy of the steady-state solution will exhibit sensitivity to eigenvalue
spread once that spread exceeds a certain critical amount. This steady-state accuracy
sensitivity arises in the following manner.
If a sufficient number of independent time samples have been collected to ensure
that the sample covariance matrix estimate is arbitrarily close to Rx x (or Rnn ), then the
exact solution for the adaptive weight vector results unless the sample matrix inversion has
insufficient accuracy due to matrix ill conditioning (as measured by the eigenvalue spread).
The eigenvalue spread can always reach a point where the computer cannot accurately
invert the matrix and output SNR degradation results.
Consequently, there is a trade-off between the accuracy of the inversion algorithm
and the allowable eigenvalue spread in the input signal covariance matrix for any DMI
algorithm. This trade-off was examined by simulating the desired matrix inversion for a
prescribed number of available bits, and Figure 5-14 illustrates the nature of this trade-off
for a four-element linear array in a two-jammer interference signal environment where the
eigenvalue spread is defined by λmax /λmin , the ratio of maximum to minimum eigenvalues
in the input signal covariance matrix. It is seen that so long as the sample covariance
matrix has an eigenvalue spread less than a critical value (that depends on the number of
bits available in the computer to accomplish the required matrix inversion), then a DMI
algorithm is insensitive to eigenvalue spread. Once this critical value of eigenvalue spread
is exceeded, however, very rapid degradation in the output SNR results with any additional
increase in eigenvalue spread.

5.6 SUMMARY AND CONCLUSIONS

DMI algorithms yield convergence speeds that surpass the LMS algorithm. The resulting
weight vector depends only on the signal covariance matrix (through the Wiener solution)
and knowledge of the signal covariance matrix comes from forming the sample covariance
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 263

5.7 Problems 263

matrix. When the desired signal is absent from the signal environment, then use of the beam
steering vector rxd yields the most desirable transient response characteristics. When the
desired signal is strong, however, formation of the estimate r̂ xd yields transient response
characteristics that are superior to the use of apriori information represented by rxd alone.
DMI algorithms are insensitive to eigenvalue spread until a certain critical level is
exceeded that depends on the number of bits available in the computer to perform the matrix
inversion computation. The practicality of the DMI approach is restricted by the number
of degrees of freedom of the adaptive processor. When the feasibility of the DMI approach
is not precluded, however, the additional complexity introduced by directly obtaining the
sample covariance matrix is rewarded by the rapid convergence and insensitivity to the
previously noted eigenvalue spread.
The difficult problem of the direct computation of a matrix inverse can be circum-
vented by recourse to factorization methods that boast superior accuracy and numerical
stability properties. The three methods that have been presented here (triangularization
of the covariance matrix and solution of the triangular system by forward and backward
substitution, Cholesky factorization, and spectral factorization [U-D or L-D]) offer attrac-
tive alternatives to direct matrix inversion computations. Furthermore, algorithms based
on these methods are extremely fast.

5.7 PROBLEMS
1. Development of a Recursion Formula. The DMI algorithm requires a matrix inversion
each time a new weight is to be calculated (as, e.g., when the signal environment
changes). By applying the matrix inversion lemma
[A + u† Ru]−1 = A−1 − A−1 u† [R−1 + uA−1 u† ]−1 uA−1
show that the weights can be updated at each sample using the recursive formulas
P(k − 1)x(k) ε∗ (k)
W (k) = W (k − 1) +
1 + x† (k)P(k − 1)x(k)
where
ε∗ (k) = −x† (k)P(k − 1) b∗
and
P(k − 1)x(k)x† (k)P(k − 1)
P(k) = P(k − 1) −
1 + x† (k)P(k − 1)x(k)
Hint: Apply the lemma to the inversion of
[R̂x x (k − 1) + x(k)x† (k)]
2. Similarity between Array Element Outputs and Tapped-Delay Line Outputs.
Consider an adaptive filter consisting of a tapped delay line with a complex weight at
each tap.
(a) Show that the DMI algorithm for minimizing the MSE between the filter output
and the reference signal is of the same form as for an adaptive array if the tap
outputs are taken as analogous to the array inputs.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 264

264 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

(b) What restrictions on sampling are imposed if we desire independent samples of

the inputs?
(c) Apply the analytical expressions derived for the adaptive array to determine the
transient behavior of the DMI algorithm applied to adaptive filters. Assume in-
dependent samples.
3. Statistical Relations [7]. Starting with the probability density for r 2 = [ξ(ŵ) − ξmin ]/ξmin ,
determine the density of the weight vector w where w = ŵ − wopt . Show that the
covariance of w depends on the covariance matrix Rx x .
4. Statistical Relations [7]. Starting with the following expression for ρ (for the reference
signal case)
ŵ† ss† ŵ 1
ρ=
ŵ (Rx x − ss )ŵ S No
† †

show that
1
ρ=

N
| qi |2
1 + (1 + SN o ) √
i=2
| SN o + q1 |2
where
⎡√ ⎤
SN o
⎢ 0 ⎥
√ ⎢ ⎥
⎢ 0 ⎥
q= SN o Ĝ − ⎢ ⎥
⎢ .. ⎥
⎣ . ⎦
0
!
SN o
Ĝ = P−1 R1/2
x x ŵ
1 + SN o

and where P is the unitary transformation that gives

⎡ ⎤
! 1
⎢ ⎥
SN o ⎢0⎥
x x w0 = P⎢ .. ⎥
R1/2
1 + SN o ⎣.⎦
0
The variable q has the probability density function
k! 1
p ( q) =
π N (k − N )! (1 + q† q)k+1
5. Statistical Relations [7]. Using the density for q in Problem 4 and the following
transformation:
Re{ q1 } = r cos φ1
Im{ q1 } = r sin φ1 cos φ2
Re{ q2 } = r sin φ1 sin φ2 cos φ3
Im{ q2 } = r sin φ1 sin φ2 sin φ3 cos φ4
..
.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 265

5.7 Problems 265

Re{ qm } = r sin φ1 · · · sin φ2N −2 cos φ2N −1

Im { qm } = r sin φ1 · · · sin φ2N −2 sin φ2N −1

where

0 ≤ r < ∞; i + 1, 2, · · · , 2N − 1
0 ≤ φi < π

derive the density p(r , φ1 , φ2 ) of (5.40) and the expression for ρ(r , φ1 , φ2 ).
Hint: The Jacobian of the transformation is

J = r 2N −1 (sin φ1 )2N −2 (sin φ2 )2N −3 · · · sin φ2N −2

and
 π [(n + 1)/2] √
(sin φ)n dφ = π
0 [(n + 2)/2]
6. Statistical Relations [7]. Derive the result in (5.44) using (5.33) and the fact that for a
CSLC ξ(ŵ) = σs2 + σ N2 and ξmin = ξ N0 + σs2 , and σ N2 is the output noise plus jammer
power.
7. Statistical Relations [7]. Define the following transformation of R x x in (5.29):
" #
x = K σ̂02 − r̂†x x0 R̂−1
x x r̂x x0

Y = K R̂x x
ŵ = R̂−1
x x r̂xd

Then the joint density of x, Y, and ŵ is given by

p(x, Y, ŵ) = P(x) P(ŵ, Y)

where
|Y |k−N +1 exp{−tr[I + (1/ξ0 ) w w† Rx x ] R−1 x x Y}
p(ŵ, Y) = (N −1)
π π
N 1/2N (k) · · · (k − N + 1)|Rx x | |ξ0 | N
k

and where w = ŵ − w0 . Derive the density of r 2 = (1/ξ0 ) w† Rx x w [see (5.34)]

from the foregoing density function.
8. Statistical Relations [7]. The random variable x defined in Problem 7 has the proba-
bility density function
 
1 1 K −N −1 x
p(x) = |x| exp −
(ξ0 ) K −N (K − N ) ξ0
(a) Show that 2x/ξ0 is χ 2 distributed with 2(K − N ) degrees of freedom.
(b) Show that ξ(ŵ) = (1/K )x has mean and variance [see (5.31), (5.32)].
 
N
E{ξ(ŵ)} = 1 − ξ0
k
K−N 2
var [ξ(ŵ)] = ξ
K2 0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 266

266 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

9. Inversion of Complex Matrices Using Real Arithmetic [14,15]

Techniques for inverting complex matrices using only real matrix operations are of
considerable value since although complex inversion routines are generally superior
(in terms of accuracy and computing time) to real matrix approaches; nevertheless,
mathematical packages for small computers generally include matrix inversion rou-
tines that apply only to real matrices.
(a) To invert the complex n × n matrix M = A + jB to obtain M −1 = E + jF is
equivalent to solving the two simultaneous equations

AE − BF = I
AF + BE = 0

for the unknown matrices E and F. Premultiply the two previous equations by B
and A, respectively, and by subtracting show that

[AB − BA] E + [A2 + B 2 ] F = −B

Similarly, premultiply the original equation pair by A and B, respectively, and by

adding show that

[BA − AB] F + [A2 + B 2 ] E = A

If A and B commute (so that AB = BA), show that M −1 = [A2 + B 2 ]−1 [A − jB].
The foregoing result involves the inverse of a real n × n matrix to obtain M −1
but is restricted by the requirement that A and B commute.
(b) Let C = A + B and D = A − B. Show that the original equation pair in part (a)
reduce to

CE + DF = I
−DE + CF = −I

From the results expressed in the equation pair immediately preceding, show that
either

M −1 = [C + DC−1 D]−1 [(DC −1 + I ) + j (DC−1 − I )]

or

M −1 = [D + CD−1 C]−1 [(CD−1 + I ) + j (−CD−1 + I )]

provided the indicated inverses exist. The foregoing equation pair represents
alternate ways of obtaining M −1 by inverting real n × n matrices without the
restriction that A and B commute.
(c) An isomorphism exists between the field of complex numbers and a special set
of 2 × 2 matrixes, that is,
 
a b
a + jb ∼
−b a
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 267

5.7 Problems 267

Consider the 2n × 2n real matrixes defined by

   
A B E F
G= and H =
−B A −F E

Show that H is the inverse of G only if the original equation pair in part (a) is
satisfied. Therefore, one way of obtaining M −1 is to compute G−1 and identify
the n × n submatrices E and F appearing in G −1 . Then M −1 = E + jF. This
approach does not involve the restrictions that beset the approaches of (a) and
(b) but suffers from the fact that it requires the inversion of a 2n × 2n matrix
that drastically increases computer storage requirements, and therefore should be
used only as a last resort.
10. Development of the Cholesky Decomposition Algorithm [11]
Consider the quadratic form xT Px for the case n = 3, where P is positive definite and
symmetric.
(a) Express the quadratic form as a difference of squares plus a remainder where the
two squares involve only P(1, 1), P(1, 2), and P(1, 3) and the remainder involves
P(2, 2), P(2, 3) and P(3, 3).
(b) Now set L(1, 1) = P(1, 1)1/2 , L(2, 1) = P(2, 1)/L(1, 1), and L(3, 1) = P(3, 1)/L(1, 1).
$
3
Furthermore let y1 = L(i, 1)xi . With this notation rewrite xT Px as
i=1

y21 + [P(2, 2) − L(2, 1)L(2, 1)]x22 + 2[P(2, 3) − L(2, 1)L(3, 1)]x2 x3

+ [P(3, 3) − L(3, 1)L(3, 1)]x23 = y21 + P(2, 2)x22 + 2P(2, 3)x2 x3 + P(3, 3)x23

where P(i, j) = P(i, j) − L(I, 1)L(j, 1) for i = 2, 3 and j = i, 3.

(c) Now complete the square on the x2 variables, that is,

P(2, 2)x22 + 2P(2, 3)x2 x3 + P(3, 3)x23 = [P(2, 2)1/2 x2 + (P(2, 3)/P(2, 2)1/2 )x3 ]2
−[(P(2, 3)/P(2, 2)1/2 )x3 ]2 + P(3, 3)x23

(d) Consistent with part (b), set L(2, 2) = P(2, 2)1/2 , L(3, 2) = P(3, 2)/L(2, 2),
and y2 = L(2, 2)x2 + L(3, 2)x3 .
Now by setting L(3, 3) = [P(3, 3) − L(3, 2)L(3, 2)]1/2 and y3 = L(3, 3)x3 , we
now have xT Px = yT y = xT LLT x.
11. Development of the Inversion of a Triangular Matrix [11]
The following identity for triangular matrices is easily verified:
 −1
Rj y R −1 −1 −1
j −R j yσ j+1
= −1 = R −1
0 σ j+1 0 σ j+1 j+1

This relation enables one to compute recursively the inverse of an ever larger
matrix, that is, if R−1
j = Uj , where Rj is the upper left j × j partition of R, then show
that
 
U j −U j (R(1, j + 1), . . . , R( j, j + 1))T σ j+1
Uj+1 =
0 σ j+1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 268

268 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

where σj+1 = 1/R(j + 1, j + 1). This result is readily transformed into the desired
algorithmic form given by (5.57), (5.58), and (5.59).
12. Structure of the Householder Transformation [16]
The Householder transformation can be written as
2uu H
H=I− = I − 2u[uH u]−1 uH = I − 2Pu
u2
where Pu = u[uH u]−1 uH
(a) Using the fact that H is Hermitian unitary (H−1 = HH = H), show that

Hv = v (length of v is preserved where v = [v1 v2 . . . vN ]T )

(b) Define vin = |vvii | where vi denotes the ith element of v.

We can now proceed to show that H can be used to eliminate all the elements of
v except the ith element.
Let u = v + vin vei where ei = [0, . . . , 0, 1, 0, . . . 0]T , and 1 is the ith element
of ei .
Show that uH v = (vH + (v H + v∗in veiT )v = v2 + |vi |v
(c) Using the result of part (b), show that
 
2uu H
Hv = I − v = v − u = −vin vei , which is the desired result.
u2
13. Equivalence of Diagonal Loading and Omnidirectional Jamming [17]
Consider a narrowband, uniform linear array consisting of n elements with spacing
d where each element has a cosine element pattern. Let the external interference
environment consist of a uniform continuum of jammers from all azimuth angles, θ,
as measured from broadside to the array.
(a) Show that the correlation element of the covariance matrix for the pair of elements
with indexes n and m is given by
π
2  
2π
Rn,m = cos(θ ) exp j d(n − m) sin(θ ) dθ
λ
− π2

(b) Making the change of variable u = sin(θ), show that Rn,m of part (a) becomes
1  
2π
Rn,m = exp j d(n − m)u du
λ
−1

(c) Part (b) becomes

1
Rn,m = exp( jπ(n − m)) − exp(− jπ(n − m))
jπ(n − m)
sin(π(n − m))
=2
π(n − m)
= 2 for n − m = 0
= 0 for n − m = 0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 269

5.7 Problems 269

14. Computer Simulation Problem An eight-element uniform array with λ/2 spacing
has the desired signal incident at 0◦ and two interference signals incident at −21◦ and
61◦ . Use the DMI algorithm to place nulls in the antenna pattern. Assume σnoise =
0.01. In MATLAB, represent the signal by cos (2π (1 : K ) /K ) exp ( j rand) and the
interference by sign(randn(1, K )).
15. Development of the Inverse of a Toeplitz Matrix [5]
(a) A Toeplitz matrix has the form
⎡ ⎤
τ0 τ−1 τ−2 τ−3
⎢ τ1 τ0 τ−1 τ−2 ⎥
⎢ ⎥
⎣ τ2 τ1 τ0 τ−1 ⎦
τ3 τ2 τ1 τ0

As a result, a normalized Toeplitz matrix can be written as

 
1 at
L n+1 = λn
r Ln

where at is a 1xn row vector

r is an nx1 column vector
L n is an nxn Toeplitz matrix

Postulate that the inverse of L n+1 , denoted as Bn+1 , can be written as

 
1 1 et
Bn+1 =
λn g Mn

By taking the product Bn+1 L n+1, show that λn = 1 + et r and g = −Mn r

(b) There now remain two terms in the product Bn+1 L n+1 to be examined. Multiply
each of these terms by Bn = L −1  thatt e = −a Bn and Mn = λn Bn + ge
n to show
t t t
t
1 e 1 e
Therefore λn Bn+1 = = , which shows that all
g Mn g λn Bn + get
elements of Bn+1 can be expressed as a function of Bn .
⎡ ⎤
0 ··· 0 1
⎢ .. ⎥
⎢. ... 1 0⎥
(c) Define a persymmetric Exchange Matrix as E = ⎢ ⎢
⎥
. ⎥ where
⎣ 0 . . . . . . .. ⎦
1 0 ... 0
persymmetric means symmetry about the cross-diagonal.
Show that EPt E = P where P is persymmetric.
and (P t )−1 P t = I
Likewise show that E E = I and E(P t )−1 E = P −1
(d) Use the fact that
 
1 gt
λn Bn+1 = E n+1 E
e λn Bn + eg t n+1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 270

270 CHAPTER 5 Direct Inversion of the Sample Covariance Matrix

to show that
 
λ B + ê ĝ t ê
λn Bn+1 = n n t
ĝ 1
where Eg = ĝ and Ee = ê. This results shows that given an element of Bn+1, all
the remaining elements along the same diagonal are given if we know λn , gn , and
en (the elements of the 1st row and column of Bn+1 .
(e) Since Bi+1 is persymmetric, we can write

ei+1 = −(E i+1 Bi+1 E i+1 )ai+1 = −E i+1 Bi+1 âi+1

Writing this result out explicitly then yields

  
1 1 eit ρ−(i+1)
ei+1 = − E i+1 , 1≤i <n
λi gi λi Bi + gi eit âi
where âi+1 has been expressed as âi augmented by the term ρ−(i+1) . The above
result allows one to write a recursion relationship for ei+1 :
   
e ρ−(i+1) + eit âi ĝi
ei+1 = i − , 1≤i <n
0 λi 1
Symmetry relations then allow a corresponding recursion relationship for gi+1 to
be written.

5.8 REFERENCES
[1] I. S. Reed, J. D. Mallett, and L. E. Brennan, “Sample Matrix Inversion Technique,” Proceedings
of the 1974 Adaptive Antenna Systems Workshop, March 11–13, Vol. 1, NRL Report 7803,
Naval Research Laboratory, Washington, DC, pp. 219–222.
[2] I. S. Reed, J. D. Mallett, and L. E. Brennan, “Rapid Convergence Rate in Adaptive Arrays,”
FIEEE Trans. Aerosp. Electron. Sys., Vol. AES-10, No. 6, November 1974, pp. 853–863.
[3] T. W. Miller, “The Transient Response of Adaptive Arrays in TDMA Systems,” Ph.D. Disser-
tation, Department of Electrical Engineering, The Ohio State University, 1976.
[4] H. L. Van Trees, Detection, Estimation and Modulation Theory, Part 1, New York, Wiley,
1968, Ch. 1.
[5] S. Zohar, “Toeplitz matrix inversion: the algorithm of W. F. Trench,” J. Assoc. Comput. Mach,
Vol 16, Oct 1969, pp. 592–601.
[6] N. R. Goodman, “Statistical Analysis Based on a Certain Multivariate Gaussian Distribution,”
Ann. Math. Stat., Vol. 34, March 1963, pp. 152–177.
[7] A. B. Baggeroer, “Confidence Intervals for Regression (MEM) Spectral Estimates,” IEEE
Trans. Info. Theory, Vol. IT-22, No. 5, September 1976, pp. 534–545.
[8] B. D. Carlson, “Covariance Matrix Estimation Errors and Diagonal Loading in Adaptive
Arrays,” IEEE Trans. on Aerosp. & Electron. Sys., Vol. AES-24, No. 4, July 1988, pp. 397–
401.
[9] M. W Ganz, R. L. Moses, and S. L. Wilson, “Convergence of the SMI and the Diagonally
Loaded SMI Algorithms with Weak Interference,” IEEE Trans. Ant. & Prop., Vol. AP-38,
No. 3, March 1990, pp. 394–399.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 271

5.8 References 271

[10] I. J. Gupta, “SMI Adaptive Antenna Arrays for Weak Interfering Signals,” IEEE Trans. Ant.&
Prop., Vol. AP-34, No. 10, October 1986, pp. 1237–1242.
[11] G. J. Bierman, Factorization Methods for Discrete Sequential Estimation, New York, Aca-
demic Press, 1977.
[12] P. E. Gill, G. H. Golub, W. Murray, and M. A. Saunders, “Methods for Modifying Matrix
Factorizations,” Mathematics of Computation, Vol. 28, No. 126, April 1974, pp. 505–535.
[13] J. E. Gentle, “Cholesky Factorization,” Section 3.2.2 in Numerical Linear Algebra for Appli-
cations in Statistics, Berlin: Springer-Verlag, 1998, pp. 93–95.
[14] H. L. Van Trees, Optimum Array Processing, Part IV of Detection, Estimation and Modulation
Theory, New York, Wiley-Interscience, 2002.
[15] W. W. Smith, Jr. and S. Erdman, “A Note on Inversion of Complex Matrices,” IEEE Trans.
Automatic Control, Vol. AC-19, No. 1, February 1974, p. 64.
[16] M. E. El-Hawary, “Further Comments on ‘A Note on Inversion of Complex Matrices,’ ” IEEE
Trans. Automatic Control, Vol. AC-20, No. 2, April 1975, pp. 279–280.
[17] B. D. Carlson, “Equivalence of Adaptive Array Diagonal Loading and Omnidirectional Jam-
ming,” IEEE Trans. Ant. & Prop., Vol. Ap-43, No. 5, May 1995, pp. 540–541.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:51 272
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 273

CHAPTER

Recursive Methods for

Adaptive Array Processing 6
' $
Chapter Outline
6.1 The Weighted Least Squares Error Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
6.2 Updated Covariance Matrix Inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
6.3 Kalman Filter Methods for Adaptive Array Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
6.4 The Minimum Variance Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
6.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
6.6 Summary and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
6.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
6.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
& %

The least mean squares (LMS) and maximum signal-to-noise ratio (SNR) algorithms
avoid the computational problems associated with the direct calculation of a set of adap-
tive weights. Chapter 8 shows that random search algorithms also circumvent computa-
tional problems. These algorithms have the advantage that the required calculations are
usually much simpler than the corresponding direct calculation, are less susceptible to
hardware inaccuracy, and are continually updated to compensate for a time-varying signal
environment.
Recursively inverting the matrix circumvents many computational problems [1–4].
The recursive algorithms exhibit a steady-state sensitivity to eigenvalue spread in the signal
covariance matrix as found for direct matrix inversion (DMI) algorithms. Furthermore,
since the principal difference between the recursive methods and the DMI algorithms lies
in the manner in which the matrix inversion is computed, their rates of convergence are
comparable. The recursive algorithms are based on least square estimation techniques and
are closely related to Kalman filtering methods [5]. For stationary environments, these
recursive procedures compute the best possible selection of weights (based on a least
squares fit to the data received) at each sampling instant, whereas in contrast the LMS,
maximum SNR, and random search methods are only asymptotically optimal.

6.1 THE WEIGHTED LEAST SQUARES

ERROR PROCESSOR
Consider the conventional N -element array of Figure 6-1 having a sequence of (real
or complex) weights multiplying the received signals to form the adaptive processor.
273
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 274

274 CHAPTER 6 Recursive Methods for Adaptive Array Processing

FIGURE 6-1
Conventional x1 x2 xN
N-element adaptive
array processor.
w1 w2 wN

y = wTx

Assume that the received signals xi (t) contain a directional desired signal component
si (t) and a purely random component n i (t), due to both directional and thermal noise
so that xi (t) = si (t) + n i (t). Collecting the received signals xi (t) and the multiplicative
weights w i (t) as components in the N -dimensional vectors x(t) and w(t), we write the
adaptive processor output signal y(t) as

y(t) = wT (t)x(t) (6.1)

The narrowband processor model of Figure 6-1 has been chosen instead of the more general
tapped delay line wideband processor in each element channel because the mathematical
manipulations are simplified. The weighted least squares error processor extends to the
more general tapped delay line form.
Consider the weighted least squares performance measure based on k data samples
following Baird [6]

1
k
1
(w) = αi [wT x(i) − d(i)]2 = [X(k)w − d(k)]T A−1
k [X(k)w − d(k)] (6.2)
2 i=1 2

where the elements of X(k) are received signal vector samples, and the elements of d(k)
are desired (or reference) signal samples as follows:
⎡ ⎤
xT (1)
⎢ xT (2) ⎥
 ⎢ ⎥
X(k) = ⎢ .. ⎥ (6.3)
⎣ . ⎦
xT (k)
⎡ ⎤
d(1)
⎢ d(2) ⎥
 ⎢ ⎥
d(k) = ⎢ .. ⎥ (6.4)
⎣ . ⎦
d(k)

Both (6.4) and (6.2) presume that the desired array output signal, d(t), is known and
sampled at d(1), d(2), . . . , d(k). In the performance measure of (6.2), Ak is a diagonal
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 275

6.1 The Weighted Least Squares Error Processor 275

weighting matrix that deemphasizes old data points and is of the form
⎡ k−1 ⎤
α 0 ... . 0
⎢ 0 α k−2
. . . . .⎥
⎢ ⎥
⎢ .. ⎥
Ak = ⎢ .⎥ (6.5)
⎢ ⎥
⎣ 0 ... α 0⎦
0 0 1
where 0 < α ≤ 1, so that older data have increasingly less importance. If the signal envi-
ronment is stationary so that all data samples are equally important, then Ak = I, the iden-
tity matrix. The performance measure given by (6.2) is minimized by selecting the weight
vector to yield the “best” (weighted least squares) estimate of the desired signal vector d(k).
To minimize the weighted least squares performance measure (6.2), set the derivative
of (w) with respect to w equal to zero, thereby yielding the optimum weight setting as
−1
wls (k) = XT (k)A−1
k X(k) XT (k)A−1
k d(k) (6.6)
When an additional data sample is taken, the foregoing weight vector solution is updated
in the most efficient manner. The updated signals X(k + 1) and d(k + 1) as well as the
updated matrix Ak+1 can each be partitioned as follows:
X(k)
X(k + 1) = (6.7)
xT (k
+ 1)
d(k)
d(k + 1) = (6.8)
d(k + 1)
and
⎡ ⎤
..
⎢ . 0
⎢ ·. . ⎥ ⎥
⎢ αAk .. .. ⎥
Ak+1 ⎢
=⎢ · .. ⎥ (6.9)
⎢ . 0⎥ ⎥
⎣ · · · · · ··· · · · ⎦
..
0···0 . 1
With this partitioning, the updated weight vector can be written as
−1
wls (k + 1) = XT (k + 1)A−1
k+1 X(k + 1) XT (k + 1)A−1
k+1 d(k + 1)
−1
= XT (k + 1)A−1
k+1 X(k + 1)

· αXT (k)A−1
k d(k) + x(k + 1) d(k + 1) (6.10)
From (6.10) it is seen that the updated weight vector solution requires the inversion of
the matrix [XT (k + 1)A−1k+1 X(k + 1)], which can also be expanded by the partitioning
previously given as
⎧ ⎡ ⎤ ⎫−1
⎪ .. ⎪
⎪
⎪ . 0 ⎪
⎪
⎪
⎪ ⎢ ·. . ⎥ ⎪
⎪
⎪
⎨ ⎢ −1 . . ⎥ ⎪
⎬
−1  ⎢ αA . . ⎥ X(k)
−1
X (k + 1)Ak+1 X(k + 1)
T 
= [X (k) x(k + 1)] ⎢
T ⎢ k
·.. ⎥
⎪
⎪
⎪ ⎢ . 0⎥ ⎥ xT (k + 1) ⎪
⎪
⎪
⎪
⎪ ⎣ ·
· · · · · ·.· · · · ⎦ ⎪
⎪
⎪
⎩ . ⎪
⎭
0···0 . 1
 −1
= α XT (k)A−1 k X (k) + x(k + 1)x (k + 1)
T T
(6.11)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 276

276 CHAPTER 6 Recursive Methods for Adaptive Array Processing

Now define

P−1 (k) = XT (k)A−1
k X(k) (6.12)

Likewise, define

P−1 (k + 1) = XT (k + 1)A−1
k+1 X(k + 1)

= α XT (k)A−1
k X(k) + x(k + 1)x (k + 1)
T

1
= α P−1 (k) + x(k + 1)xT (k + 1) (6.13)
α
Inverting both sides of (6.13) and applying the matrix inversion lemma [(D.10) of
Appendix D] then yields
 
1 P(k)x(k + 1)xT (k + 1)P(k)
P(k + 1) = P(k) − (6.14)
α α + xT (k + 1)P(k)x(k + 1)

By our use of (6.14) in (6.10) and recognition that wls (k) = P(k)X(k)A−1
k d(k), the updated
solution for the weight vector now becomes
P(k)x(k + 1)
wls (k + 1) = wls (k) +
α+ xT (k+ 1)P(k)x(k + 1)
· d(k + 1) − wlsT (k)x(k + 1) (6.15)

Equations (6.14) and (6.15) are the iterative relations of the recursive least squares algo-
rithm. Equations (6.14) and (6.15) are started by adopting an initial guess for the weight
vector w(0) and the initial Hermitian matrix P(0). It is common practice to select as an
initial weight vector w(0) = [1  0◦ , 0, 0, . . . , 0], thereby obtaining an omnidirectional
array pattern (provide the sensor elements each have omnidirectional patterns) and to
select P(0) as the identity matrix.
Equations (6.14) and (6.15) yield the updated weight vector in a computationally
efficient manner that avoids calculating the matrix inverses present in (6.6) and (6.10). It
is instructive to consider (6.6) in more detail for the additional insight to be gained into the
mechanics of the processor. Since the trace of Ak , tr [Ak ], is a scalar, (6.6) can be rewritten
as
 −1  
XT (k)A−1
k X(k) XT (k)A−1k d(k)
wls (k) = (6.16)
tr [Ak ] tr [Ak ]

The first bracketed term on the right-hand side of (6.16) is an estimate of the autocorrelation
matrix R̂xx based on k data samples, that is,


k
R̂xx i, j
(k) = α k−n xi (k)x j (k) (6.17)
n=1

Equation (6.17) is an expression for forming exponentially deweighted estimates of the

matrix Rxx also used by Mantey and Griffiths [2]. Similarly, the second bracketed term
on the right-hand side of (6.16) is an estimate for the cross-correlation vector r̂xd (k).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 277

6.2 Updated Covariance Matrix Inverse 277

Consequently, the P matrix defined by (6.12) is directly related to the autocorrelation

matrix since

R̂−1
xx (k) = tr [Ak ]P(k) = (1 + α + α + · · · + α
2 k−1
)P(k)
 
1−α k
= P(k) (6.18)
1−α

The form of the algorithm given by (6.14) and (6.15) requires that the desired signal be
known at each sample point, which is an unrealistic assumption. An estimate of the desired
signal is used in practice, Consequently, replacing d(k + 1) by d(k
ˆ + 1) in (6.15) results
in a practical weighted least square error recursive algorithm.
Other useful forms of (6.15) arise from replacing certain instantaneous quantities by
their known average values [3,7]. To obtain these equivalent forms, rewrite (6.15) as

P(k)
w(k + 1) = w(k) +
α + x T (k + 1)P(k)x(k + 1)
· x(k + 1)d(k + 1) − x(k + 1)y(k + 1) (6.19)

where y(k + 1) is the array output. The product x(k + 1) · d(k + 1) is replaced by its
average value, which is an estimate of the cross-correlation vector r̂xd . Since the estimate
r̂xd does not follow instantaneous fluctuations of x(t) and d(t), it may be expected that
the convergence time would be greater using r̂xd than when using x(k) d(k) as shown in
Chapter 5.
In the event that only the direction of arrival of the desired signal is known and the
desired signal is absent, then the cross-correlation vector rxd (which conveys direction of
arrival information) is known, and the algorithm for updating the weight vector becomes

P(k)
w(k + 1) = w(k) + rxd − x(k + 1)y(k + 1) (6.20)
α+ xT (k + 1)P(k)x(k + 1)
Equation (6.20) may of course also be used when the desired signal is present, but the
rate of convergence is then slower than for (6.19) with the same desired signal present
conditions.
For stationary signal environments α = 1, but this choice is not practical. As long as
0 < α < 1, (6.14) and (6.15) lead to stable numerical procedures. When α = 1, however,
after many iterations the components of P(k) become so small that round-off errors have a
significant impact. To avoid this numerical sensitivity problem, both sides of (6.14) are mul-
tiplied by the factor (k + 1) to yield numerically stable equations, and (6.18) then becomes

R̂−1
xx (k) = kP(k) (6.21)

6.2 UPDATED COVARIANCE MATRIX INVERSE

The weighted least squares error processor of the previous section was based on weighting
the current received signal vector data compared with past data according to

P−1 (k + 1) = αP−1 (k) + x∗ (k + 1)xT (k + 1) (6.22)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 25, 2010 21:10 278

278 CHAPTER 6 Recursive Methods for Adaptive Array Processing


where 0 ≤ α ≤ 1 and P−1 (k) = XT (k)A−1 k X(k). A closely related alternative data
weighting scheme uses the sample covariance matrix, R̂xx , for summarizing the effect of
old data, so that
R̂xx (k + 1) = α R̂xx (k) + x∗ (k + 1)xT (k + 1) (6.23)
where 0 ≤ α ≤ 1. Inverting both sides of (6.24) yields
 −1
1 1 ∗
R̂−1 (k + 1) = R̂ xx (k) + x (k + 1)x T
(k + 1) (6.24)
xx
α α
Applying the matrix inversion lemma to (6.24) then results in
 
−1 1 −1 R̂−1 ∗
xx (k)x (k + 1)x (k + 1)
T
−1
R̂xx (k + 1) = R̂xx (k) − R̂xx (k) (6.25)
α α + xT (k + 1)R̂−1 ∗
xx (k)x (k + 1)

In the absence of a desired signal

wopt = R−1
xx b
∗
(6.25)
(or R−1 ∗
xx rxd ) so that each side of (6.25) can be postmultiplied by b to give
 
1 R̂−1 ∗
xx (k)x (k + 1)
ŵ(k + 1) = ŵ(k) − x (k + 1)ŵ(k)
T
(6.26)
α α + xT (k + 1)R̂−1 ∗
xx (k)x (k + 1)

This is known as the recursive least squares (RLS) algorithm, because it recursively updates
the correlation matrix such that more recent time samples receive a higher weighting than
past samples [19].
As an example, consider a four-element uniform linear array with λ/2 spacing. The
desired signal is incident at 0◦ , and one interference signal is incident at 45◦ with σnoise =
0.01. After K = 25 iterations and α = 0.9, the antenna pattern appears in Figure 6-2
with a directivity of 5.6 dB, which is less than the quiescent pattern directivity of 6 dB.
Figure 6-3 and Figure 6-4 are the RLS weights as a function of time. They converge in
about 20 iterations.
The data weighting represented by (6.26) implies that past data [represented by R̂xx (k)]
is never more important than current data [represented by x∗ (k + 1)xT (k + 1)]. An

FIGURE 6-2 RLS

adapted pattern. 5
Directivity (dB)

−5

−15

−25

−90 −45 0 45 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 279

6.2 Updated Covariance Matrix Inverse 279

3.5 FIGURE 6-3

Magnitude of the
3 RLS weights as a
function of iteration.
2.5

2
|wn|

1.5

0.5

0
0 5 10 15 20 25 30
Iteration

100 FIGURE 6-4

Phase of the RLS
weights as a
50
function of iteration.

0
∠wn(degrees)

−50

−100

−150

−200
0 5 10 15 20 25 30
Iteration

alternative data weighting scheme that permits past data to be regarded either as less
important or more important than current data is to use

R̂xx (k + 1) = (1 − β)R̂xx (k) + βx∗ (k + 1)xT (k + 1), 0≤β≤1 (6.27)

The data weighting scheme represented by (6.27) has been successfully employed [using
β = 1/(k +1) so that each sample is then equally weighted] to reject clutter, to compensate
for platform motion, and to compensate for near-field scattering effects in an airborne
moving target indication (AMTI) radar system [8]. Inverting both sides of (6.27) and
applying the matrix inversion lemma results in [9]
1 β
R̂−1
xx (k + 1) = R̂−1 (k) −
(1 − β) xx (1 − β)
R̂−1 ∗ −1
xx (k)x (k + 1) x (k + 1)R̂xx (k)
T
· (6.28)
(1 − β) + β xT (k + 1)R̂−1 ∗
xx (k)x (k + 1)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 280

280 CHAPTER 6 Recursive Methods for Adaptive Array Processing

To obtain the updated weight vector R̂−1 xx (k + 1) must be postmultiplied by the beam
steering vector b∗ . The beam steering vector for AMTI radar systems is matched to a
moving target by including the expected relative Doppler and polarization phase factors,
thereby minimizing the effects of main beam clutter due to (for example) stationary targets
[8,10]. Carrying out the postmultiplication of (6.28) by b∗ then yields
 
1 R̂−1 ∗
xx (k)x (k + 1)
ŵ(k + 1) = ŵ(k) − β x (k + 1)ŵ(k)
T
(1 − β) (1 − β) + β xT (k + 1)R̂−1 ∗
xx (k)x (k + 1)]
(6.29)

The updating computation represented by (6.28) requires N 2 complex multiplications

to form R̂−1 ∗
xx (k)x (k + 1), where N is the number of degrees of freedom present in the
adaptive processor. Furthermore, (6.28) also requires an additional 0.75N 2 and 2.25N
multiplications to complete the computation of R̂−1 xx (k + 1). On comparing (6.29) and
(6.26) with (6.20), it may appear that the direction-of-arrival information contained in rxd
(or b∗ ) is missing in (6.26) and (6.29). We realize that the initial starting weight selected
in either (6.26) or (6.29) reflects any direction-of-arrival information, however, so that
(6.20), (6.26), and (6.29) merely represent different data weighting versions of the same
basic weight update equation.

6.2.1 Recursive Algorithms Based on Factorization Methods

The recursive update equation (6.29) is not the most accurate and numerically stable equa-
tion possible using the factorization methods introduced in Chapter 5. In [11,12], recursive
update equations are described that possess highly desirable accuracy and stability prop-
erties superior to those of (6.29). We will consider the solution to a linearly constrained
optimization problem like that of Chapter 4. It will then be easy to see how the various
algorithms based on factorization methods relate to the development given here.
An array output is described by

y(n) = w† (n)x(n), (6.30)

where x(n) is an N × 1 data vector describing the outputs of each array element at time
nT and † denotes complex conjugate transpose. A linear constraint for (6.30) is given by

C† w(n) = f (6.31)

where C is an N × K constraint matrix, and f is a K × 1 response column vector.

  √ 
Let 1/2 = diag μn−1 μn−2 . . . μ 1 (6.32)

where 1/2 is an n × n matrix that describes the exponential weighting to be applied to

the data (0 < μ ≤ 1). By defining the n × N data matrix
†
X(n) = x(1)x(2) . . . x(n) (6.33)

where x(i) is a data vector of length N at time iT, then the weighted output power is
given by
! !
ξ = !1/2 X(n)w(n)! (6.34)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 281

6.2 Updated Covariance Matrix Inverse 281

where   denotes the Euclidian norm, and X(n) is the data matrix of (6.33). The term
“Q-R decomposition” is usually employed to describe the problem of obtaining an upper
triangular transformation of an arbitrary matrix, A, by applying an orthogonal matrix,
Q, obtained by a series of Givens rotations as described in Section [Link]. When this
procedure is applied to the n × N matrix 1/2 X(n), the result can be described by
R(n)
Q(n)1/2 X(n) = (6.35)
0
where R(n) is an N × N upper triangular matrix, and 0 is an (n − N ) × N null matrix.
In keeping with notation introduced earlier, we retain U(n) to denote the upper triangular
matrix in (6.35), even though R(n)—called the Cholesky factor—is what the terminology
“Q-R decomposition” refers to. The weight vector that minimizes ξ of (6.34) is then
given by
 −1
w(n) = [U† (n)U(n)]−1 C C† [U† (n)U(n)]−1 C f (6.36)

The similarity of equation (6.36) to equation (4.169) is duly noted. The Q-R decomposition
of the input data matrix using Givens rotations enables the weight vector to be obtained
by using back substitution and can be implemented in parallel and systolic structures.
Back substitution is a costly operation to perform in an algorithm and impedes a parallel
implementation, so the inverse Q-R decomposition, which uses the inverse Cholesky factor,
U−1 (n), is used instead.
To develop a recursive implementation of (6.36), a recursive update of the upper
triangular matrix U(n) is implemented [11]
√
U(n) μ U(n − 1)
= T(n) (6.37)
0T x H (n)
where T(n) is an (N + 1) × (N + 1) orthogonal matrix that annihilates the row vector
√
x† (n) by rotating it into μ U(n − 1). By premultiplying both sides of (6.37) by their
respective Hermitian forms and recognizing that T† (n)T(n) = I, it follows that

U† (n)U(n) = μU† (n − 1)U(n − 1) + x(n)x† (n) (6.38)

Using the matrix inversion lemma on (6.38) then yields

1 −1 U−1 (n − 1)z(n)z H (n)U−H (n − 1)
U−1 (n)U−† (n) = U (n − 1)U−† (n − 1) − (6.39)
μ μ t 2 (n)
where
U−1 (n − 1)x(n)
z(n) = √ (6.40)
μ
and the scalar is defined by

t (n) = 1 + z H (n)z(n) (6.41)

Equation (6.39) can now be shown to be equivalent to

1 −1
U−1 (n)U−† (n) = U (n − 1)U−† (n − 1) − g(n)g† (n) (6.42)
μ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 282

282 CHAPTER 6 Recursive Methods for Adaptive Array Processing

where
U−1 (n − 1)z(n)
g(n) = √ (6.43)
μ t (n)

On comparing (6.42) and (6.38), it is apparent that there is a relationship corresponding

to (6.37) using an orthogonal matrix P(n), which can be written as

U−H (n) μ−1/2 U−H (n − 1)

= P(n) (6.44)
g (n)
H
0T

Equation (6.44) may easily be verified by forming the product of each side of (6.44) with its
respective complex conjugate transpose and verifying that (6.43) results. There is a major
difference, however, between the orthogonal matrix P(n) in (6.44) and the orthogonal
matrix T(n) in (6.37). The derivation of the orthogonal matrix P(n) is given in [13], and
the result is
z(n) 0
P(n) = (6.45)
1 t (n)

In other words, P(n) is a rotation matrix that successively annihilates the elements of the
vector z(n), starting from the top, by rotating them into the last element at the bottom.
To relate these results to equation (6.36), it will be convenient to define a new N × N
matrix S(n), given by

S(n) = U−1 (n)U−† (n) or S−1 (n) = U† (n)U(n) (6.46)

Since it can easily be shown that S−1 (n) = X† (n)(n)X(n), S−1 (n) is referred to as a
“correlation matrix” of the exponential weighted sensor outputs averaged over n samples.
It is convenient to define (n) = S(n)C and (n) = C† (n) = C† S(n)C. Then we may
rewrite (6.36) as

w(n) = (n)−1 (n) f (6.47)

Substituting S(n) of (6.46) into (6.39) and using the matrix inversion lemma, it follows
that
1 1
S(n) = S(n − 1) − k(n)x† (n)S(n − 1) (6.48)
μ μ

where
1
μ
S(n − 1)x(n)
k(n) = = S(n)x(n). (6.49)
1+ x (n)S(n − 1)x(n)
1 H
μ

Using the previously given definition for (n) and right multiplying both sides of (6.48)
by C yields

1 1
(n) = (n − 1) − k(n)x† (n)(n − 1) (6.50)
μ μ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 283

6.2 Updated Covariance Matrix Inverse 283

Likewise, premultiplying both sides of (n) = S(n)C by C† and using the previously
given definition for (n) gives the result
1 1
(n) = (n − 1) − k(n)x† (n)(n − 1) (6.51)
μ μ
Postmultiplying both sides of (6.43) by C then yields

(n) = μ−1 (n − 1) − g(n)α(n) (6.52)

where

α(n) = g† (n)C (6.53)

A similar development can be given for (n) starting with

(n) = C† U−1 (n)U−† (n)C = C† [μ−1 U−1 (n − 1)U−† (n − 1) − g(n)g† (n)]C

= μ−1 (n − 1) − α † (n)α(n) (6.54)
√
Setting γ (n) = μ α † (n), (6.54) can be rewritten as

(n) = μ−1 [(n − 1) − γ (n)γ † (n)] (6.55)

Applying the matrix inversion lemma to (6.55) then yields the recursive relation
√
−1 (n) = μ[I + μ q(n)α(n)]−1 (n − 1) (6.56)

where
√
μ−1 (n − 1)α † (n)
q(n) = = μ−1/2 [−1 (n)α † (n)] (6.57)
1 − μα(n)−1 (n − 1)α † (n)
Finally, to apply the previous recursive relationships to the weight vector of (6.47), we
have
√
w(n) = w(n − 1) − μ[g(n) − μ(n)q(n)]α(n)−1 (n − 1)f (6.58)

The recursive relationship of (6.58) can be further simplified by applying the definitions
of (6.40) and (6.42) to obtain

w(n) = w(n − 1) − ρ(n)ξ(n, n − 1) (6.59)

where
√
μ
ρ(n) = k(n) − (n)q(n) (6.60)
t (n)
ξ(n, n − 1) = x† (n)w(n − 1) (6.61)

A summary of the previous development for a recursive beamforming algorithm is given

in Table 6-1.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 284

284 CHAPTER 6 Recursive Methods for Adaptive Array Processing

TABLE 6-1 Recursive Algorithm Based on Inverse QR Updating

• Initialization: U−1 (0) = δ −1 I, δ = small positive constant

(0) = U−1 (0)U−† (0)C

w(0) = (0)[C† (0)]−1 f
• For n = 1, 2, . . . , do the following:
– Compute the intermediate vector z(n)

U−† (n − 1)x(n)
z(n) = √
μ

– Evaluate the rotations that define P(n), which annihilate z(n) and compute the scalar t (n) from

z(n) 0
P(n) =
1 t (n)

– Update the lower triangular matrix U−† (n), and compute the vector g(n) and α(n) = g† (n) C from

μ−1/2 U−† (n − 1) U−† (n)

P(n) =
0T g† (n)

– Update (n), q(n), and −1 (n) using

(n) = μ−1 (n − 1) − g(n)α(n)

√ −1
μ (n − 1)α † (n)
q(n) =
1 − μα(n)−1 (n − 1)α † (n)
√
−1 (n) = μ[I + μ q(n)α(n)]−1 (n − 1)
– Update the weight vector
w(n) = w(n − 1) − ρ(n)ξ(n, n − 1)
where
√
μ
ρ(n) = k(n) − (n)q(n)
t (n)

and ξ(n, n − 1) = x† (n)w(n − 1)

6.3 KALMAN FILTER METHODS FOR ADAPTIVE

ARRAY PROCESSING
The adaptive algorithms for small communications and data collection arrays are often
based on relatively simple gradient and random search methods. More complex problems,
such as command and control of remote vehicles or rapid angular tracking in radar sys-
tems, require sophisticated processing, such as Kalman filtering [5,14,15]. These methods
include more a priori signal environment data than other methods and open the possibility
of constructing processors that integrate adaptive array control with other system functions
such as position location and navigation.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 285

6.3 Kalman Filter Methods for Adaptive Array Processing 285

6.3.1 Development of a Kalman-Type Array Processor

Consider the simple N -element linear narrowband array model of Figure 6-1 for which
the (sampled) signal vector is x(k), the adaptive processor weight vector (at corresponding
sample times) is w(k), and the array output is y(k) = wT (k)x(k). To accommodate wider
bandwidth signals, each channel of the array has a tapped delay line filter, but the tap
weights and delayed signals are represented as vectors w(k) and x(k) so the mathematical
development for this case remains the same as for the narrowband case.
Let the dynamic behavior of the optimal array weights be represented by
wopt (k + 1) = (k + 1, k)wopt (k), wopt (0) = w0 (6.62)
where (k + 1, k) is a transition matrix. If the array signal environment is stationary, then
the optimal weights are fixed and (k + 1, k) is the identity matrix. For time-varying
environments, a more complex model for (k + 1, k) must be developed that reflects how
the optimal array weights change in response to the varying environment. Let the system
measurements be represented by a noise corrupted version of the optimal (not the actual)
array output:
d(k) = xT (k)wopt (k) + v(k) (6.63)
where v(k) is a member of a white, Gaussian noise sequence having zero mean and
variance given by
E{v(k)v( j)} = σ 2 (k)δk j (6.64)
The selection of a value for σ 2 (k) is discussed later.
For the dynamical model (6.62) having state vector wopt (k) and measurement from
(6.63), Kalman filter theory leads to a minimum mean square error (MMSE) estimator
for the array weights (rather than for the desired signal). It immediately follows that the
optimum filtered estimate of the optimal array weight vector ŵopt (k/k) is given by the
relationship [12]

ŵopt (k/k) = ŵopt (k/k − 1) + K(k) d(k) − xT (k)ŵopt (k/k − 1) (6.65)

where (k/k) denotes a filtered quantity at sample time k based on measurements through
(and including) k, (k/k − 1) denotes a predicted quantity at sample time k based on
measurements through k − 1, and K(k) is the Kalman gain matrix. For complex quantities
(6.65) is rewritten as

ŵopt (k/k) = ŵopt (k/k − 1) (6.66)

+ K(k) d ∗ (k) − x† (k)ŵopt (k/k − 1)

Now
ŵopt (k/k − 1) = (k, k − 1)ŵopt (k − 1/k − 1) (6.67)

and the quantity in brackets of (6.66) is the difference between the optimal array output (or
desired reference signal) and the actual array output. The Kalman-type processor based on
the foregoing equations is shown in Figure 6-5 where the Kalman gain vector is given by
P(k/k − 1)x(k)
K(k) = T (6.68)
[x (k)P(k/k − 1)x(k) + σ 2 (k)]
286
Monzingo-7200014
book

v(k)
ISBN : XXXXXXXXXX

Unit wopt (k) d(k) + e (k) ŵopt (k/k) Unit ŵopt (k −1/k − 1)
Desired signal
Σ time xT (k) Σ K(k) Σ time
approximation
delay − delay

Φ(k +1, k) xT (k) Φ(k, k− 1)

ŵopt (k /k− 1)
November 24, 2010

Dynamical model Estimator for optimal array weights

for optimal array
weights

FIGURE 6-5 Kalman-type processor for optimal array weights.

18:54
286
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 287

6.3 Kalman Filter Methods for Adaptive Array Processing 287

or for complex quantities

P(k/k − 1)x∗ (k)
K(k) = (6.69)
[xT (k)P(k/k − 1)x∗ (k) + σ 2 (k)]
The predicted error covariance matrix is given by

P(k/k − 1) = (k, k − 1)P(k − 1/k − 1)T (k, k − 1) (6.70)

The filtered error covariance matrix is defined by

  T
P(k/k) = E wopt (k) − ŵopt (k/k) wopt (k) − ŵopt (k/k) (6.71)

Equation (6.71) can be expressed in the equivalent form

P(k/k) = P(k/k − 1) − K(k)xT (k)P(k/k − 1) (6.72)

On substituting (6.68) into (6.72) there results

P(k/k − 1)x(k)xT (k)P(k/k − 1)

P(k/k) = P(k/k − 1) − (6.73)
[σ 2 (k) + xT (k)P(k/k − 1)x(k)]
Comparing (6.73) with (6.14) reveals significant similarity between the recursive
equations for P(k + 1) of (6.14) and P(k/k) of (6.73). Equation (6.73) is rewritten by
application of the matrix inversion lemma [(D.14) of Appendix D] as
1
P−1 (k/k) = P−1 (k/k − 1) + x(k)xT (k) (6.74)
σ 2 (k)
Comparing (6.74) with (6.13) shows that the Kalman error covariance matrix with
σ 2 (k) = 1 corresponds to P(k) for the weighted least square error processor with α = 1. In
this case, (6.21) leads to the error covariance matrix P(k/k) being related to the finite-time
average estimate for the received signal autocorrelation matrix R̂xx (k) by the relationship

R̂−1
xx (k) = kP(k/k) (6.75)

To begin the recursive equations (6.65)–(6.73), the initial values for ŵ(0/0) and
P(0/0) must be specified. It is desirable to select ŵopt (0/0) = E{wopt } and P(0/0) =
E{w(0)wT (0)} [5] where w(0) = ŵ(0/0) − wopt . From Chapter 3, wopt is given by
the Wiener solution.

wopt = R−1
xx (k)rxd (k) (6.76)

In general, the signal statistics represented by the solution (6.76) are unknown, so a different
procedure (discussed subsequently) is employed to initialize the recursive equations. In
the event that a priori environment data are available, then such information is used to
form a refined initial estimate of wopt using (6.76) as well as to construct a dynamical
system model by way of (6.62).
In situations where a priori information concerning the signal environment is not
available, then ŵopt (0) is generally chosen to yield an omnidirectional array beam pattern,
and P(0/0) can merely be set equal to the identity matrix. Furthermore, some means of
selecting a value for the noise statistic σ 2 (k) must be given.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 288

288 CHAPTER 6 Recursive Methods for Adaptive Array Processing

As can be seen from (6.65), the bracketed quantity represents the difference between
the actual array output (using the estimated weights) and a noise corrupted version of the
optimal array output derived from an identical array employing the optimal weight vector
on the received signal set. This optimal output signal d(t) is interpreted as a reference
signal approximation for the actual desired signal since the optimal array weights are
designed to provide an MMSE estimate of the desired signal. Such an interpretation of the
optimal array output signal in turn suggests a procedure for selecting a value for the noise
statistic σ 2 (k). Since d(t) is an approximation for the actual desired signal, s(t), one may
write [5]

d(k) = xT (k)wopt (k) + v(k) = s(k) + η(k) (6.77)

where η(k) indicates the error in this approximation for the desired signal s(k). Conse-
quently,

v(k) = s(k) + η(k) − xT (k)wopt (k) (6.78)

If s(k), η(k), and x(k) are all zero mean processes, then E{v(k)} = 0. Furthermore, if the
noise sequence η(k) is not correlated with s(k) and the noise components of x(k), then

σ 2 (k) = E{v(k)v(k)} = wopt

T
Rxx wopt − 2wopt
T
rxs + E{s 2 } + E{η2 }
= MMSE + E{η2 } (6.79)

It follows that if the “measurement” d(k) is generated by using a realistic approximation

for the desired signal (possibly obtained from a modem connected to the array output), then
the value of σ 2 (k) is chosen based on the MMSE that can be achieved by an optimal array
and the quality (as measured by the error variance) of the desired signal approximation.
If we rewrite (6.65) as
P(k/k − 1)
ŵ(k/k) = ŵ(k/k − 1) + [x(k)d(k) − x(k)y(k)] (6.80)
[xT (k)P(k/k − 1)x(k) + σ 2 (k)]
where y(k) = ŵT (k/k −1)x(k), it is apparent that the quantity x(k)d(k) can be replaced by
its average value, in exactly the same manner as with the weighted least squares recursive
algorithm of the preceding section. Consequently, the Kalman filter weight update equation
is also modified to accommodate two additional forms involving the use of either r̂xd (k)
(when either the desired signal or a reference signal model is available) or rxd (k) (when
only direction of arrival information is available).
Yet another alternative for handling a nonstationary environment with a Kalman-type
processor exists when the time-varying nature of the signal statistics are unknown. Rather
than attempt to construct an accurate dynamical model representation by way of (6.62),
simply use

wopt (k + 1) = wopt (k) + ξ (k) (6.81)

where ξ (k) is a zero-mean white noise process with

cov[ξ (k)] = Q(k) (6.82)

The elements of Q represent the degree of uncertainty associated with adopting the sta-
tionary environment assumption represented by using the identity state transition matrix
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 289

6.3 Kalman Filter Methods for Adaptive Array Processing 289

in (6.62). Equation (6.70) then becomes

P(k/k − 1) = (k, k − 1)P(k − 1/k − 1)T (k, k − 1) + Q (6.83)

The practical effect of the previous modification is to prevent the Kalman gains in
K(k) from decaying to values that are too small, so when variations in the environment
occur sufficient importance is attached to the most recent measurements. The estimate
ŵopt then “follows” variations in the actual value of wopt , although the resulting optimal
weight vector estimates are more “noisy” than when the Q matrix was absent.

6.3.2 Speed of Convergence

The fact that the value of σ 2 (k) is related to the MMSE and to the quality of the desired
signal approximation makes it possible to obtain an expression from which the rate of
convergence for the algorithm can be deduced under certain conditions [16]. Equation
(6.68) can be rewritten as

K(k)xT (k)P(k/k − 1)x(k) + K(k)σ 2 (k) = P(k/k − 1)x(k) (6.84)

or

K(k)σ 2 (k) = [I − K(k)xT (k)]P(k/k − 1)x(k) (6.85)

Equation (6.85) may now be substituted into (6.72) to obtain

P(k/k)x(k) = K(k)σ 2 (k) (6.86)

so that
P(k/k)x(k)
K(k) = (6.87)
σ 2 (k)
When we substitute the result expressed by (6.87) into (6.72) there results
P(k/k)x(k) T
P(k/k) = P(k/k − 1) − x (k)P(k/k − 1) (6.88)
σ 2 (k)
or
[P(k/k)x(k)xT (k)P(k/k − 1)]
P(k/k) = P(k/k − 1) − (6.89)
σ 2 (k)

Premultiplying both sides of (6.89) by P−1 (k/k) and postmultiplying both sides of (6.90)
by P−1 (k/k − 1), we see it follows that [also see (6.74)]

x(k)xT (k)
P−1 (k/k) = P−1 (k/k − 1) + (6.90)
σ 2 (k)
Equation (6.90) can be rewritten as
1
P−1 (k/k) = [σ 2 (k)P−1 (k/k − 1) + x(k)xT (k)] (6.91)
σ 2 (k)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 290

290 CHAPTER 6 Recursive Methods for Adaptive Array Processing

so that

P(k/k) = σ 2 (k)[σ 2 (k)P−1 (k/k − 1) + x(k)xT (k)]−1 (6.92)

The recursive relationship expressed by (6.92) can be repeatedly applied beginning with
P−1 (0/−1) to obtain
 −1

k
−1
P(k/k) = σ (k) σ (k)P (0/ − 1) +
2 2
x(i)x (i)
T
(6.93)
i=1

For cases where the desired signal approximation is quite good σ̂ 2 (k) ≈ MMSE, and
"k
the diagonal matrix σ̂ 2 (k)P−1 (0/−1) can be neglected in comparison with x(i)xT (i)
i=1
so that
 −1

k
P(k/k) ∼
= σ̂ (k) 2
x(i)x (i)
T
(6.94)
i=1

When we use the result (6.87), it follows immediately that

 −1

k
K(k) ∼
= x(i)x (i)
T
x(k) (6.95)
i=1

which is independent of σ 2 (k). The arithmetic average is given by

1
k
x(i)xT (i) → Rxx (k) as k → ∞ (6.96)
k i=1

The MSE at the kth sampling instant, ξ 2 (k), can be written as [16]

ξ 2 (k) = trace[P(k/k)Rxx (k)] + MMSE (6.97)

From (6.94) and (6.96) it follows that

trace[P(k/k)Rxx (k)] ∼
= σ 2 (k)N k −1 (6.98)

where N is the dimension of x(k) so the MSE at the kth sampling instant becomes

ξ 2 (k) ∼
= MMSE[1 + N k −1 ] (6.99)

The result in (6.99) means that convergence for this Kalman-type algorithm is theoretically
obtained within 2N iterations, which is similar to the convergence results for the DMI
algorithm.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 291

6.4 The Minimum Variance Processor 291

6.4 THE MINIMUM VARIANCE PROCESSOR

It is useful at this point to consider the application of the concepts used in obtaining a
recursive algorithm to a broadband signal aligned array processor to illustrate the slight
modifications involved in handling a multichannel tapped delay line processor. If the
direction of arrival of the desired signal is known, this information may be used to construct
a signal-aligned array as shown in Figure 6-6. The signal-aligned array processor uses
spatial correction filters (SCF) and knowledge of the desired signal’s direction of arrival
to align the desired signal components in each channel of the array by properly selecting
the time delays τi , that is [17]

⎡ ⎤
z 1 (t − τ1 )
⎢ z 2 (t − τ2 ) ⎥
⎢ ⎥
x(t) = ⎢ .. ⎥ = d(t)1 + n(t) (6.100)
⎣ . ⎦
z N (t − τN )

where 1 = [1, 1, . . . , 1]T , d(t) is the desired reference signal, and n(t) is the vector of
the interference terms after the time delays. Collecting the steered received signal vector
x(t) and its delayed components along the tapped delay line into a single (M + 1)N × 1

FIGURE 6-6
z1 z1 zN Broadband signal
z(t)
aligned array
t1 SCF t2 SCF SCF tN processor.
x1 (t) x2(t) xN(t)
x (t)

Δ Δ Δ

x (t −Δ)

Δ Δ Δ

Δ Δ Δ
x(t −MΔ)

w01 w11 wM
1 w02 w12 wM
2 wN0 wN1 wNM

y (t) = wT (t) ~
x(t)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 292

292 CHAPTER 6 Recursive Methods for Adaptive Array Processing

dimensional vector gives

⎡ ⎤ ⎡ ⎤
x(t) d(t)1
⎢ ⎥ ⎢
x(t − ) ⎥ d(t − )1
⎢ ⎥ ⎢ ⎥
x (t) = ⎢ ⎥=⎢
.. ⎥ + n (t) .. (6.101)
⎣ ⎦ ⎣
. ⎦ .
x(t − M) d(t − M)1
In what follows it will be convenient to drop the prime notation and simply remember
that all signals under consideration are the collection of terms in the tapped delay line filter.
Furthermore, collecting the N -dimensional weight vectors w0 (t), w 1 (t), . . . , w M (t), as
⎡ 0 ⎤
w (t)
⎢ w1 (t) ⎥
⎢ ⎥
w(t) = ⎢ . ⎥ (6.102)
⎣ .. ⎦
w M (t)
the array output can be written in the form


N
y(t) = wT x(t) = d(t) w i0 (6.103)
i=1


N 
N 
N
= d(t) w i0 + d(t − ) w i1 + L · · · + d(t − M) w iM + wT n(t)
i=1 i=1 i=1

T

N
since wl 1 = w il . If the adaptive weights are constrained according to
i=1


N 
1, l = 0
w il = (6.104)
0, l = 1, 2, · · · , k
i=1

then the output signal can be written as

y(t) = d(t) + wT n(t) (6.105)

The output signal of (6.105) is unbiased since

E{y(t)} = E{d(t) + wT n(t)} = d(t) (6.106)

if we assume the noise vector has zero mean. The variance of the output signal is then
given by

var [y(t)] = E{wT n(t)nT (t)w} = wT Rnn w (6.107)

Define the N (M + 1) × (M + 1) matrix,

⎡ ⎤
1 0 ··· 0
⎢0 1 · · · 0⎥
 ⎢ ⎥
I1 = ⎢ . . ⎥ (6.108)
⎣ .. .. ⎦
0 0 ··· 1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 293

6.4 The Minimum Variance Processor 293

where 0 is an N × 1 vector with all zero components. If it is desired to minimize the output
noise variance, then the noise variance performance measure can be defined by

mv = wT Rnn w (6.109)

It is now desired to minimize (6.109) subject to the constraint (6.104), which can be
rewritten as
⎡ ⎤
1
⎢ 0 ⎥
⎢ ⎥
I1T w = ⎢ .. ⎥ = c (6.110)
⎣ . ⎦
0
The weight vector w that minimizes (6.109) subject to (6.110) can be chosen by using a
vector Lagrange multiplier to form the modified performance measure
1 T
mvm = w Rnn w + λ c − I1T w (6.111)
2
Setting the derivative of mvm with respect to w equal to zero to obtain wmv , requiring
wmv to satisfy (6.110) to evaluate λ, and substituting the resulting value of λ into wmv
gives the minimum variance weight vector solution
−1
wmv = R−1 T −1
nn I1 I1 Rnn I1 c (6.112)

For a signal-aligned array like that of Figure 6-6, it can also be established (as was done
in Chapter 3 for a narrowband processor) that the minimum variance estimator resulting
from the use of (6.112) is identical to the maximum likelihood estimator [18].
The weight vector computation defined by (6.112) requires the measurement and
inversion of the noise autocorrelation matrix Rnn . The noise autocorrelation matrix can
be obtained from the data that also include desired signal terms, and use of a recursive
algorithm will circumvent the necessity of directly inverting Rnn .
The difficulty of measuring the noise autocorrelation matrix when desired signal terms
are present can be avoided by reformulating the optimization problem posed by (6.109)
and (6.110). The minimization of (6.109) subject to the constraint (6.110) is completely
equivalent to the following problem:

Minimize = wT Rxx w (6.113)

subject to the constraint

I1 w = c (6.114)

The solution of (6.113), (6.114) may be found by once again using Lagrange multipliers
with the result that

wopt = R−1 T −1 −1
xx I1 [I1 Rxx I1 ] c (6.115)

To show the equivalence between the problem (6.113) and (6.114) and the original problem
(6.109) and (6.110), expand the matrix Rxx as follows:

Rxx = E{x(t)xT (t)} = Rdd + Rnn (6.116)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 294

294 CHAPTER 6 Recursive Methods for Adaptive Array Processing

where
⎧⎡ ⎤ ⎫
⎪
⎪ d(t)1 ⎪
⎪
⎪
⎨⎢ d(t − )1 ⎥ ⎪
⎬
⎢ ⎥
Rdd =E ⎢ .. ⎥ [d(t)1d(t − )1, . . . , d(t − M)1] (6.117)
⎪
⎪⎣ . ⎦ ⎪
⎪
⎪
⎩ ⎪
⎭
d(t − M)1

Substituting (6.116) into (6.113), the problem now becomes

Minimize  = wT Rxx w = wT Rnn w + wT Rdd w (6.118)

With the constraint (6.104), however, it follows that

 = wT Rnn w + E{d 2 (t)} (6.119)

Now the minimization of (6.119) subject to (6.114) must give exactly the same solution
as the optimization problem of (6.109) and (6.110), since E{d 2 (t)} is not a function of w,
and the two problems are therefore completely equivalent.
The received signal correlation matrix at the kth sample time Rxx (k) can be measured
using the exponentially deweighted finite time average

1 
k
R̂xx (k) = # $ α k−n x(n)xT (n) (6.120)
"
k
α k−n n=1
n=1

where 0 < α ≤ 1. By defining the matrices

⎡ ⎤
xT (1)
⎢xT (2)⎥
 ⎢ ⎥
X(k) = ⎢ .. ⎥ (6.121)
⎣ . ⎦
xT (k)

and
⎡ ⎤
α k−1 0 ··· 0
⎢ .. ⎥
 ⎢ 0 α k−2 .⎥
A(k) = ⎢
⎢ ..
⎥
⎥ (6.122)
⎣ . α 0⎦
0 ··· 0 1

then R̂xx (k) can be rewritten as

1 1
R̂xx (k) = XT (k)A(k)X(k) = P−1 (k) (6.123)
tr[A(k)] tr[A(k)]

where P−1 (k) = XT (k)A(k)X(k). Partitioning x(k + 1) and A(k + 1) as in (6.7) and
(6.9) and applying the matrix inversion lemma immediately leads to (6.14). The desired
updated weights may then be written in terms of P(k + 1) from (6.115) as
−1
wopt (k + 1) = P(k + 1)I1 I1T P(k + 1)I1 c (6.124)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 295

6.5 Simulation Results 295

The inverse of I1T P(k + 1)I1 required in (6.124) can be efficiently computed by application
of the matrix inversion lemma to yield

P(k)
wopt (k + 1) = I − − P(k + 1)
α + x (k + 1)P(k)x(k + 1)
T

· x(k + 1)xT (k + 1) I−11 c (6.125)

The minimum variance recursive processor for the narrowband case takes a particu-
larly simple form. It was found in Chapter 3 for this case that
R−1
nn 1
wmv = (6.126)
1 R−1
T
nn 1

The use of (6.126) presents an additional difficulty since the received signal vector x(t)
generally contains signal as well as noise components. This difficulty can be circumvented,
though.
The input signal covariance matrix is given by

Rxx = E{x∗ (t)xT (t)} = E{d 2 (t)}11T + Rnn (6.127)

where E{d 2 (t)} = β, a scalar quantity. Inverting both sides of (6.127) and applying the
matrix inversion lemma yields
βR−1 T −1
nn 11 Rnn
R−1
xx = [β11 + Rnn ]
T −1
= R−1
nn − (6.128)
1 + β1T R−1
nn 1

Substituting (6.128) into the ratio

R−1
xx 1
1 R−1
T
xx 1
the following matrix identity results:
R−1
xx 1 R−1
nn 1
≡ (6.129)
1T R−1
xx 1 1 T R−1 1
nn

Exploit (6.21) for the case when α = 1; it then follows from (6.129) and (6.126) that
(k + 1)P(k + 1)1 P(k + 1)1
wmv (k + 1) = = T (6.130)
(k + 1)1T P(k + 1)1 1 P(k + 1)1
where P(k + 1) is given by (6.14). Note that when there is no desired signal, wmv of
(6.130) converges faster than when the desired signal is present, a result already found in
Chapter 5.

6.5 SIMULATION RESULTS

The recursive processor defined by (6.66), (6.70), and (6.73) represents both the Kalman
and the weighted least squares error processor, since the Kalman error covariance matrix
with σ 2 (k) = 1 corresponds to P(k) for the weighted least square error processor with
α = 1. The parameter σ 2 (k) for the Kalman processor is selected to be equal to the MMSE
in accordance with (6.79). To simulate this recursive algorithm, it is necessary to define
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 296

296 CHAPTER 6 Recursive Methods for Adaptive Array Processing

FIGURE 6-7 Array y

geometry with signal
and jammer
locations for (0.394l, 0.682l)
selected example.
J2 /n

J3 /n J1 /n

60° 60°
(−0.787l, 0) x
s/n = 10

l
87
0.7
(0.394l, −0.682l)

FIGURE 6-8 20.00

Output SNR versus
number of iterations
for Kalman 15.00
processor with
Output SNR (dB)

eigenvalue spread of Optimum

10.00
16,700. Input
s/n = 10 for which
the output SNRopt = 5.00
15 (11.76 dB) with
algorithm
parameters 0
σ 2 = MMSE,
w(0) = 0, and
P(0) = I. −5.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5
Number of iterations

an array geometry and signal environment as in Figure 6-7 (which duplicates Figure 4-30
and is repeated here for convenience). This figure depicts a four-element Y array having
d = 0.787λ element spacing with one desired signal at 0◦ and three distinct narrowband
Gaussian jamming signals located at 15◦ , 90◦ , and 165◦ .
The desired signal in each case was taken to be a biphase modulated signal having
a phase angle of either 0◦ or 180◦ with equal probability at each sample. Two signal
environments are considered corresponding to eigenvalue spreads of 16,700 and 2,440.
Figures 6-8 to 6-11 give convergence results for an eigenvalue spread of 16,700, where the
jammer-to-thermal noise ratios are J1 /n = 4000, J2 /n = 400, and J3 /n = 40, for which
the corresponding noise covariance matrix eigenvalues are given by λ1 = 1.67 × 104 ,
λ2 = 1 × 103 , λ3 = 29.0, and λ4 = 1.0. The input signal-to-thermal noise ratio is
s/n = 10 for Figures 6-8 and 6-9, s/n = 0.1 for Figure 6-10, and s/n = 0.025 for
Figure 6-11. The performance of the algorithm in each case is recorded in terms of the
output SNR versus number of iterations, where one weight iteration occurs with each new
independent data sample.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 297

6.5 Simulation Results 297

15.00 FIGURE 6-9

Output SNR versus
10.00 Optimum number of iterations
for weighted least
5.00 squares error
Output SNR (dB)

processor with
0 eigenvalue spread of
16,700. Input
−5.00 s/n = 10 for which
output SNRopt = 15
−10.00 with algorithm
parameters α = 1,
−15.00 wT (0) = [1, 0, 0, 0],
and P(0) = I.
−20.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4
Number of iterations

−5.00

−10.00 Optimum
Output SNR (dB)

−15.00

−20.00

−25.00

−30.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3
Number of iterations

FIGURE 6-10 Output SNR versus number of iterations for Kalman processor with
eigenvalue spread of 16,700. Input s/n = 0.1 for which output SNRopt = 0.15 (−8.24 dB)
with algorithm parameters σ 2 = MMSE, w(0) = 0, and P(0) = I.

−10.00 FIGURE 6-11

Output SNR versus
number of iterations
−15.00 Optimum for Kalman
processor with
−20.00
Output SNR (dB)

eigenvalue spread
of 16,700. Input
−25.00 s/n = 0.025 for
which output
−30.00 SNRopt =
0.038 (−14.2 dB)
with algorithm
−35.00
parameters
σ 2 = MMSE,
−40.00 w(0) = 0, and
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3
P(0) = I.
Number of iterations
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 298

298 CHAPTER 6 Recursive Methods for Adaptive Array Processing

FIGURE 6-12 −5.00

Output SNR versus
number of iterations −10.00 Optimum
for Kalman
−15.00
processor with

Output SNR (dB)

eigenvalue spread −20.00
of 2440. Input
s/n = 0.1 for which −25.00
output SNRopt =
0.15(−8.24 dB) with −30.00
algorithm −35.00
parameters
σ 2 = MMSE, −40.00
w(0) = 0, and
P(0) = I. −45.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3
Number of iterations

FIGURE 6-13 15.00

Output SNR versus
number of iterations 10.00 Optimum
for Kalman
5.00
processor with
Output SNR (dB)

eigenvalue spread of 0
2440. Input s/n = 10
for which output −5.00
SNRopt =
−10.00
15 (11.76 dB) with
algorithm −15.00
parameters
σ 2 = MMSE, −20.00
w(0) = 0, and
P(0) = I. −25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 8 1E3
Number of iterations

Figures 6-12 and 6-13 give convergence results for an eigenvalue spread of 2,440,
where the jammer-to-thermal noise ratios are J1 /n = 500, J2 /n = 200, and J3 /n = 40,
for which the corresponding noise covariance matrix eigenvalues are given by λ1 =
2.44 × 103 , λ2 = 4.94 × 102 , λ3 = 25.62, and λ4 = 1.0. The input signal-to-thermal noise
ratio is s/n = 0.1 for Figure 6-12 and s/n = 10.0 for Figure 6-13.
The simulation results shown in Figures 6-8 to 6-13 illustrate the following important
properties of the recursive algorithms:
1. Recursive algorithms exhibit fast convergence comparable to that of DMI algorithms,
especially when the output SNRopt is large (approximately five or six iterations when
SNRopt = 15.0 for the examples shown). On comparing Figure 6-8 (for an eigenvalue
spread of 1.67 × 104 ) with Figure 6-13 (for an eigenvalue spread of 2.44 × 103 ), it is
seen that the algorithm convergence speed is insensitive to eigenvalue spread, which
also reflects the similar property exhibited by DMI algorithms.
2. Algorithm convergence is relatively insensitive to the value selected for the parameter
σ 2 (k) in (6.69). On comparing the results obtained in Figure 6-8 where σ 2 = MMSE
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 299

6.7 Problems 299

with the results obtained in Figure 6-9 where σ 2 = 1, it is seen that virtually the same
number of iterations are required to arrive within 3 dB of SNRopt even with different
initial weight vectors.
3. The convergence speed of the recursive algorithm for the examples simulated here is
slower for small values of SNRopt and faster for large values of SNRopt , as seen in
Figures 6-8, 6-10, and 6-11. This behavior is also exhibited by the DMI algorithm
(and to some degree by all algorithms that do not assume direction-of-arrival informa-
tion). When no direction-of-arrival information is assumed, such information can be
“learned” from a strong desired signal component.

6.6 SUMMARY AND CONCLUSIONS

The specific form selected for a recursive processor should reflect the data weight scheme
that is appropriate for the desired application. The various recursive algorithms may
be developed by applying the matrix inversion lemma to the same basic weight update
equation.
Since the recursive algorithms are different from a DMI algorithm primarily because
the required matrix inversion is accomplished in a recursive manner, it is hardly surprising
that many of the desirable properties found to apply to DMI algorithms also hold for
recursive algorithms. Rapid convergence rates and insensitivity to eigenvalue spread are
characteristics that make recursive processors attractive algorithm candidates provided
sufficient computational power and accuracy are available to carry out the required calcu-
lations.

6.7 PROBLEMS
1. The Minimum Variance Weight Vector Solution
(a) Show that setting the derivative of mvm of (6.111) equal to zero yields

wmv = R−1
nn I1 λ

(b) Show that requiring wmv obtained in part (a) to satisfy (6.110) results in

−1
λ = I1T R−1
nn I1 c

(c) Show that substituting the result obtained in part (b) into wmv of part (a) results in (6.112).
2. Equivalence of the Maximum Likelihood and Minimum Variance Estimates [15]. In some
signal reception applications, the desired signal waveform is completely unknown and cannot
be treated as a known waveform or even as a known function of some unknown parameters.
Hence, no a priori assumptions regarding the signal waveform are made, and the waveform is
regarded as an unknown time function that is to be estimated. One way of obtaining an undistorted
estimate of an unknown time function with a signal-aligned array like that of Figure 6-6 is to
employ a maximum likelihood estimator that assumes that the noise components of the received
signal have a multidimensional Gaussian distribution. The likelihood function of the received
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 300

300 CHAPTER 6 Recursive Methods for Adaptive Array Processing

signal at the SCF output can then be written as

 
1 
k
1
= exp − (xm − sm ))T ρ −1 (xn − sn )
(2π ) N /2(2k+1) |ρ|1/2 2
m,n=−k

where n, m denote distinct sample times, and ρ is the noise covariance matrix that is a matrix
of N × N submatrices corresponding to the various tap points along the tapped delay line.
Differentiate the logarithm of the likelihood function with respect to sn and equate the result to
zero to obtain ŝm , the maximum likelihood estimator for sm . Show that this result corresponds
to the signal estimate obtained from the processor defined by the result in Problem 1.
3. Derivation of Optimum Weights via Lagrange Multipliers [18]. Show that the solution to the
optimization problem posed by (6.113) and (6.114) is given by (6.115).
4. Development using the M.I.L [18]. Show that (6.124) leads to (6.125) by means of the following
steps:
(a) Pre- and postmultiply (6.14) by I1T and I1 , respectively, to obtain

1 T IT P(k)x(k + 1)xT (k + 1)P(k)I1

I1T P(k + 1)I1 = I1 P(k)I1 − 1
α α + xT (k + 1)P(k)x(k + 1)

(b) Apply the matrix identity (D.4) of Appendix D to the result obtained in part (a) to show that
−1 −1
I1T P(k + 1)I1 = α I1T P(k)I1 + I−1 −T
1 x(k + 1)x (k + 1)I1
T

(c) Using the result of part (b) in (6.124), show that

wopt (k + 1) = P(k + 1)[αP−1 (k) + x(k + 1)xT (k + 1)]I−T

1 c

(d) Noting that

1 P(k)x(k + 1)xT (k + 1)
P(k + 1)P−1 (k) = I−
α α + xT (k + 1)P(k)x(k + 1)

show that the result obtained in part (c) leads to (6.125).

5. Derivation using the M.I.L. [18]. Show that the matrix identity (6.129) results from the substi-
tution of (6.128) into the ratio

R−1
xx 1
1 R−1
T
xx 1

6. Derivation of the Orthogonal Matrix Relation of Equation (6.77) [5]

(a) Equation (6.76) suggests that P(n) may be partitioned in the following manner:

E ψ
P(n) =
ηH h

where E is N × N , ψ and η are both N × 1, and h is a scalar.

Show that substituting the aforementioned P(n) of part (a) into (6.76) gives the following
relations:

EU−H (n − 1) η H U−H (n − 1)
√ = U−H (n) and √ = g H (n)
μ μ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 301

6.8 References 301

(b) Recognizing that P(n) is orthogonal (P(n)P H (n) = I), show that substituting P(n) of part
(a) into the orthogonal condition yields the following relationships:

EE H + H
=I
Eη + h=0
h2 + η H η = 1

(c) Show that substituting (6.75) into the corresponding result of part (a) and defining

U−H (n − 1)x(n)
a(n) = √
μ

results in η = a(n)
t (n)
(d) Show that substituting η from part (c) into the third relationship of part (b) gives
1
h=
t (n)
(e) The unknown vector ψ may be eliminated by using h of part (d) and η of part (c) in the
second relationship of part (b) to give

ψ = −Ea(n)

(f) Form the product of P(n) with the augmented vector [z H (n)1] H and use the partitioning of
part (a) for P(n) to produce

z(n) E z(n)
P(n) =
1 η h 1

(g) Finally, show that substituting the results from parts (c) and (d) into the result of part (f)
yields the desired result

z(n) 0
P(n) =
1 t (n)

7. RLS algorithm. An eight-element uniform array with λ/2 spacing has the desired signal incident
at 0◦ and two interference signals incident at −21◦ and 61◦ . Use the RLS algorithm to place
nulls in the antenna pattern. Assume that σn = 0.01.
8. RLS algorithm. Plot the received signal as a function of iteration for the RLS and LMS algorithms
when a 0 dB desired signal is incident on the eight-element array at 0◦ while 12 dB interference
signals are incident at −21◦ and 61◦ .

6.8 REFERENCES
[1] C. A. Baird, “Recursive Algorithms for Adaptive Arrays,” Final Report, Contract No. F30602-
72-C-0499, Rome Air Development Center, September 1973.
[2] P. E. Mantey and L. J. Griffiths, “Iterative Least-Squares Algorithms for Signal Extraction,”
Second Hawaii International Conference on System Sciences, January 1969, pp. 767–770.
[3] C. A. Baird, “Recursive Processing for Adaptive Arrays,” Proceedings of the Adaptive Antenna
Systems Workshop, March 11–13, 1974, Vol. I, NRL Report 7803, Naval Research Laboratory,
Washington, DC, pp. 163–182.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 18:54 302

302 CHAPTER 6 Recursive Methods for Adaptive Array Processing

[4] R. D. Gitlin and F. R. Magee, Jr., “Self-Orthogonalizing Adaptive Equalization Algorithms,”

IEEE Trans. Commun., Vol. COM-25, No. 7, July 1977, pp. 666–672.
[5] C. A. Baird, “Kalman-Type Processing for Adaptive Antenna Arrays,” IEEE International
Conference on Communications, June 1974, Minneapolis, Minnesota, pp. 10G-1–10G-4.
[6] C. A. Baird, Jr. and J. T. Rickard, “Recursive Estimation in Array Processing,” Fifth Asilomar
Conference on Circuits and Systems, Pacific Grove, CA, November 1971, pp. 509–513.
[7] L. J. Griffiths, “A Simple Adaptive Algorithm for Real-Time Processing in Antenna Arrays,”
Proc. IEEE, Vol. 57, No. 10, October 1969, pp. 1695–1704.
[8] L. E. Brennan, J. D. Mallet, and I. S. Reed, “Adaptive Arrays in Airborne MTI Radar,” IEEE
Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976, pp. 607–615.
[9] J. M. Shapard, D. Edelblute, and G. Kinnison, “Adaptive Matrix Inversion,” Naval Undersea
Research and Development Center, NUC-TN-528, May 1971.
[10] L. E. Brennan and I. S. Reed, “Theory of Adaptive Radar,” IEEE Trans. Aerosp. Electron.
Syst., Vol. AES-9, No. 2, March 1973, pp. 237–252.
[11] S. J. Chern and C.Y. Chang, “Adaptive Linearly Constrained Inverse QRD-RLS Beamforming
Algorithm for Moving Jammers Suppression,” IEEE Trans. Ant. & Prop., Vol. AP-50, No. 8,
August 2002, pp. 1138–1150.
[12] R. Schreiber, “Implementation of Adaptive Array Algorithms,” IEEE Trans. Acoustics.,
Speech, & Signal Processing, Vol. ASSP-34, No. 5, October 1986, pp. 205–212.
[13] S. T. Alexander and A. L. Ghirnikar, “A Method for Recursive Least Squares Filtering
Based Upon an Inverse QR Decomposition,” IEEE Trans. Signal Processing, Vol. 41, No. 1,
January 1993, pp. 20–30.
[14] R. E. Kalman, “A New Approach to Linear Filtering and Prediction Problems,” Trans. ASME,
J. Basic Eng., Series D, Vol. 82, 1960, pp. 35–45.
[15] R. E. Kalman and R. S. Bucy, “New Results in Linear Filtering and Prediction Theory,” Trans.
ASME, J. Basic Eng. Series D, Vol. 83, March 1961, pp. 95–108.
[16] D. Godard, “Channel Equalization Using a Kalman Filter for Fast Data Transmission,” IBM
J. Res. Dev., May 1974, pp. 267–273.
[17] C. A. Baird, Jr., “Recursive Minimum Variance Estimation for Adaptive Sensor Arrays,”
Proceedings of the IEEE 1972 International Conference on Cybernetics and Society,
October 9– 12, Washington, DC, pp. 412–414.
[18] J. Capon, R. J. Greenfield, and R. J. Kolker, “Multidimensional Maximum-Likelihood Pro-
cessing of a Large Aperture Seismic Array,” Proc. IEEE, Vol. 55, No. 2, February 1967,
pp. 192–211.
[19] F. B. Gross, Smart antennas for wireless communications: with MATLAB, New York:
McGraw-Hill, 2005.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 303

CHAPTER

Cascade Preprocessors
7
' $
Chapter Outline
7.1 Nolen Network Preprocessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
7.2 Interference Cancellation with a Nolen Network Preprocessor . . . . . . . . . . . . . . . . . . . 311
7.3 Gram–Schmidt Orthogonalization Preprocessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
7.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
7.5 Summary and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
7.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
7.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
& %

The least mean squares (LMS) and maximum signal-to-noise ratio (SNR) algorithms
converge slowly whenever there is a wide spread in the eigenvalues of the input signal
correlation matrix. A wide eigenvalue spread occurs if the signal environment includes a
very strong source of interference together with other weaker but nevertheless potent in-
terference sources. This condition also happens when two or more very strong interference
sources arrive at the array from closely spaced but not identical directions.
It was shown in Chapter 4 that by appropriately selecting the step size and moving in
suitably chosen directions an accelerated gradient procedure offers marked improvement
in the convergence rate over that obtained with an algorithm that moves in directions deter-
mined by the gradient alone. Another approach for obtaining rapid convergence rescales
the space in which the minimization is taking place by appropriately transforming the input
signal coordinates so that the constant cost contours of the performance surface (repre-
sented by ellipses in Chapter 4) are approximately circular and no eigenvalue spread is
present in the rescaled space. If such a rescaling is done, then in principle it would be possi-
ble to correct all the error components in a single step by choosing an appropriate step size.
With a method called scaled conjugate gradient descent (SCGD) [1], this philosophy
is followed with a procedure that employs a CGD cycle of N iterations and uses the
information gained from this cycle to construct a scaling matrix that yields very rapid
convergence on the next CGD cycle.
The philosophy behind the development of cascade preprocessors is similar to that of
the SCGD method. The cascade preprocessor introduced by White [2,3] overcomes the
problem of (sometimes) slow convergence by reducing the eigenvalue spread of the input
signal correlation matrix by introducing an appropriate transformation (represented by
the preprocessing network). Used in this manner, the cascade network resolves the input

303
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 304

304 CHAPTER 7 Cascade Preprocessors

signals into their eigenvector components. Equalizing the resolved signals with automatic
gain control (AGC) amplifiers reduces the eigenvalue spread, thereby simplifying the task
of any gradient algorithm.
By modifying the performance measure governing the control of the adaptive ele-
ments in a cascade preprocessor, a cascade network performs the complete task of array
pattern null steering without any need for a gradient type processor [4]. Using a cascade
preprocessor in this manner reduces the complexity and cost of the overall processor and
represents an attractive alternative to conventional gradient approaches.
Finally, we introduce the use of a cascade preprocessor developed by Brennan et al.
[5–7] to achieve adaptive null steering based on the Gram–Schmidt orthogonalization pro-
cedure [8–10]. A Gram–Schmidt cascade preprocessor is simpler than the other cascade
networks discussed and possesses very fast convergence properties that make it a most ap-
pealing practical alternative [11]. The discussion of eigenvector component preprocessing
networks gives perspective to the development and use of cascade preprocessors.

7.1 NOLEN NETWORK PREPROCESSOR

Suppose that it were possible to construct a lossless transformation network to insert
between the sensor elements and the adaptive weights that resolves the input signals into
their eigenvector components, so the correlation matrix of the transformed signals was
diagonal with the eigenvalues {λ1 , λ2 , . . . , λ N } at the elements. The insertion of such a
transformation network ahead of the adaptive weights to resolve the signals into orthogonal
normalized eigenvector beams was suggested by Gabriel [12]. If a second transformation
equalizes the various eigenvalues, then in principle it should be possible to select a step
size that convergences in a single step. If the necessary a priori eigenvalue information
were available to the designer, then a five-element array would have the transformation
matrix and equalizing network shown in Figure 7-1.
Since the eigenvalues of the input signal correlation matrix are unknown, an approx-
imation of the eigenvector network must be constructed. Otherwise, when the correlation
matrix is ill conditioned (i.e., the eigenvalues are widely diverse), then small errors made
in estimating certain matrix components become magnified enough that the signals cannot
be resolved into eigenvector components. Fortunately, the method presented here works
quite well with only a rough approximation. A preprocessor excels under the conditions
that are the most difficult for gradient-type algorithms.

7.1.1 Single-Stage Nolen Network

Consider the lossless, passive, reflectionless network shown in Figure 7-2. This network
has N input ports (one for each array sensor output) and N output ports. The network
elements consist of N − 1 variable phase shifters and N − 1 variable directional couplers
that are connected in a series feed configuration. This network was first described by Nolen
in an unpublished Bendix memorandum [13] concerned with the problem of synthesizing
multiple beam antennas [14] and is therefore termed a Nolen transformation network.
Since the network is passive, lossless, and reflectionless, the total output power is the
same as the total input power, and the overall transformation matrix is unitary.
The signals in the transformation network are denoted by vnk , where k indicates the
level in the processor, and n indicates the element channel to which the signal corresponds.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 305

7.1 Nolen Network Preprocessor 305

FIGURE 7-1
Five-element
x1 x2 x3 x4 x5
adaptive array
with eigenvector
5 × 5 Eigenvector component
transformation matrix transformation
and eigenvalue
e1 e2 e3 e4 e5 Eigenvector equalization
components network.

Equalization
AGC AGC AGC AGC AGC
network

d1 d2 d3 d4 d5

Adaptive
weights

Array output

FIGURE 7-2
Single-stage
x1 = v11 x2 = v12 x3 = v13 x4 = v14 x5 = v15
lossless, passive,
reflectionless Nolen
Phase Phase Phase Phase transformation
shifter shifter shifter shifter network for N = 5.

Directional Directional Directional Directional

coupler coupler coupler coupler

v22 v23 v24 v25 v12 = e1

The output signal in Figure 7-2 is expressed as


N
v12 = an vn1 (7.1)
n=1

where the weighting factors are constrained by the unitary condition n |an |2 = 1 and
the condition that a1 must be real.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 306

306 CHAPTER 7 Cascade Preprocessors

Now suppose we want to maximize the output power resulting from the single-level
Nolen network. The output power is expressed as

N 
N
P1 = E{|v12 |2 } = E{v12 v12∗ } = an E{vn1 vl1∗ }al∗
n=1 l=1
(7.2)

N 
N
= an m 1nl al∗
n=1 l=1

where

m knl = E{vnk vlk∗ } (7.3)

is an element
 of the correlation matrix of the input signals. To introduce the unitary
constraint n |an |2 = 1 while maximizing P1 , employ the method of Lagrange multipliers
by maximizing the quantity
 

N
Q = P1 + λ 1 − |an |2 (7.4)
n=1

where λ denotes a Lagrange multiplier. Setting an = u n + jvn , the partial derivatives of

Q with respect to u L and v L is found to be
∂Q  
= m 1Ln (u n − jvn ) + (u l + ivl )m l1L − 2λu L
∂u L n l
 
= an m 1n L + m 1Ll al∗ − λ(a L + a L∗ ) (7.5)
n l
∂Q  
= j m 1Ln (u n − jvn ) − j (u l + jvl )m l1L − 2λv L
∂v L n l
 
= −j an m n L + j
1
m Ll al + jλ(a L − a L∗ )
1 ∗
(7.6)
n l

Setting both ∂ Q/∂u L and ∂ Q/∂v L = 0, the solution for a L must satisfy the following
conditions:
  ⎫
λ(a L + a L∗ ) = an m 1n L + m 1Ll al∗ ⎪
⎬
n l
  (7.7)
λ(a L − a L∗ ) = an m 1n L − m 1Ll al∗ ⎪
⎭
n l

The conditions represented by (7.7) immediately simplify to


N
λa L = an m 1n L (7.8)
n=1

which is a classical eigenvector equation. Nontrivial solutions of (7.8) exist only when λ
has values corresponding to the eigenvalues of the matrix M1 (of which m 1n L is the n Lth
element).
Substituting (7.8) into (7.2) yields

N
P1 = λ an an∗ (7.9)
n=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 307

7.1 Nolen Network Preprocessor 307


In view of the unitary constraint |an |2 = 1, (7.9) becomes
P1 = λ (7.10)
Consequently, the largest value of P1 (which is the maximum power available at the
right-hand output port) is just the largest eigenvalue of the matrix M1 .
When the phase shifters and directional couplers are adjusted to maximize P1 , then
the signal v12 in Figure 7-2 is orthogonal to all the other signals vk2 . That is,
E{v12 vk2∗ } = 0 for k = 2, 3, . . . , N (7.11)
This orthogonality condition reflects the fact that no other unitary combination of v12 and
any of the other vk2 signals delivers more power than is obtained with v12 alone—such a
result would contradict the fact that P1 has been maximized.
The result of (7.11) shows that all the off-diagonal elements in the first row and the
first column of the covariance matrix of the signals vk2 are set equal to zero. Consequently,
the transformation introduced by the single-stage network of Figure 7-2 is step one in
diagonalizing the covariance matrix of the input signals. To complete the diagonalization,
additional networks are introduced as described in the next section.

7.1.2 Cascaded Networks

Now let the single-stage transformation network of Figure 7-2 be followed by a cascade
of similar networks as shown in Figure 7-3. The second transformation network operates
only on the signals v22 through v52 , leaving v12 = ε1 undisturbed. Because of the reflec-
tionless character of the transformation networks, the parameters of the second network
are adjusted without affecting the prior maximization of P1 . Adjust the parameters of the
second transformation network to maximize P2 —the output power from the right-hand
port of Level 2. The maximum power now available equals the largest eigenvalue of the
covariance matrix of the input set {v22 , v32 , . . . , v 2N }. If P1 was truly maximized in the first
level, then the largest eigenvalue of the submatrix equals the second largest eigenvalue of
the complete covariance matrix.
The signals emerging from network 2 form the signal set {v22 , v32 , . . . , v 2N } and ad-
justing the parameters of network 2 to maximize P2 results in setting all the off-diagonal
elements in the first row and first column of the submatrix for the signal set {v22 , v32 , . . . , v 2N }
to zero. As far as the complete covariance matrix is concerned, the first two transformation
networks have diagonalized the first two rows and the first two columns. Each succeeding
transformation network likewise diagonalizes one row and one column at a time. Once
the output of the last transformation in the cascaded network system is reached, the entire
matrix diagonalization is complete. Since the adjustment of parameters in each network
leaves previous networks in the cascade undisturbed, there is no need to iterate the process
unless the input signal statistics change.
Figure 7-1 shows that an equalization network that consists of AGC amplifiers through
which the signals are now passed follows the eigenvector transformation matrix. As a result,
the signal powers on the various output leads {d1 , d2 , . . . , d N } are equalized. Since the
covariance matrix corresponding to the eigenvector component signals {ε1 , ε2 , . . . , ε N }
is diagonalized, the eigenvalues are consequently equalized. Therefore, the covariance
matrix of the output signals from the equalization network is a scalar constant times the
identity matrix.
The transformation resulting from the entire cascade of Figure 7-3 is unitary, since
each of the networks in the cascade is unitary. Consequently, not only is the total output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 308

308 CHAPTER 7 Cascade Preprocessors

FIGURE 7-3
Nolen cascade x1 x2 x3 x4 x5
network for
five-element array.
f12 f13 f14 f15 Phase
shifter
Level 1
Directional
coupler v21
y12 y13 y14 y15

v22 v32 v42 v25

f23 f24 f25

Level 2

v32
y23 y24 y25

v33 v34 v35

f34 f35
Level 3
v43
y34 y35

v44 v45

f45

Level 4
v54
y45

Level 5

e5 e4 e3 e2 e1

power strictly equal to the total input power, but also the eigenvalues of the covariance
matrix are unchanged. If the element G J represents the overall transfer matrix of the first J
stages, then G J is a product of factors in which each factor represents a single stage, that is,

G J = F J · F J −1 · · · F2 · F1 (7.12)

where
⎡ ⎤
ε1
⎢ ε2 ⎥
⎢ ⎥ ⎡ ⎤
⎢ .. ⎥ x1
⎢ . ⎥
⎢ ⎥ ⎢ x2 ⎥
⎢ ε ⎥ ⎢ ⎥
⎢ J ⎥ = GJ ⎢ .. ⎥ (7.13)
⎢ J +1 ⎥ ⎣ . ⎦
⎢v J +1 ⎥
⎢ . ⎥ xN
⎢ . ⎥
⎣ . ⎦
v NJ +1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 309

7.1 Nolen Network Preprocessor 309

Furthermore, the second and lower stages of the network have no effect on the signal ε1 .
It follows that the first row of G J is the same as the first row of
G1 = F 1 (7.14)
Likewise, since the third and lower stages of the network have no effect on F2 , the second
row of G J is identical to the second row of
G2 = F2 · F1 (7.15)
Continuing in the foregoing manner, the entire G N matrix is implemented in the Nolen
form using a step-by-step process. Since the first row of G J is identical to the first row
of G1 , the second row of G J is identical to the second row of G2 , . . ., and the (J − 1)st
row of G J is identical to the (J − 1)st row of G J −1 , it follows that the first J − 1 rows
and J − 1 columns of F J are the same as those of an identity matrix. Furthermore, since
there are no phase shifters between the leftmost input port at any stage and the right-hand
output port, the J th diagonal element of F J is real. These constraints in addition to the
unitary constraint on the transfer matrix define the bounds of the element values (phase
shift and directional coupling) of the J th row.
If the output power maximization at each stage of the cascade transformation net-
work is only approximate, the off-diagonal elements of the covariance matrix are not
completely nulled. However, with only a rough approximation, the off-diagonal elements
are at least reduced in amplitude, and, although the equalization network will no longer
exactly equalize the eigenvalues, it reduces the eigenvalue spread.

7.1.3 Control of the Eigenvector Transformation Networks

Consider the problem of adjusting the parameters in each stage of an eigenvector transfor-
mation network to realize the desired covariance matrix diagonalization. It turns out that
the signal environment conditions that prove the most difficult for the LMS algorithm are
the same conditions that prove to be easy for the preprocessor [3].
For the single-stage network of Figure 7-2, the power at the right-hand output port
is maximized by appropriately adjusting each phase shifter and directional coupler in the
network. Denote the signal flowing downward into the phase shifter directional coupler
combination located in the lth row and the kth column of the cascade of Figure 7-3 by vkl .
Likewise, denote the signal flowing into this same coupler from the left by ykl . The signal
coming out of the right-hand port (which is equivalent to y lN +1 ) is then εl .
Regard the directional coupler as equivalent to a goniometer having shaft angle ψlk ;
the signals into and out of one phase shifter directional coupler combination are written as
l
yk+1 = ykl cos ψl,k + ykl e jφl,k sin ψl,k (7.16)
vkl+1 = ykl sin ψl,k + vkl e jφl,k cos ψl,k (7.17)

Now Pl = E{|εl |2 } is the power out of the lth stage where


N
εl = akl vkl (7.18)
k=l
N N
Pl = ail m lik akl∗ (7.19)
i=l k=l
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 310

310 CHAPTER 7 Cascade Preprocessors

where m lik is given by (7.3) and

⎧ N
⎪ 
⎪
⎨ cos ψlk for i = 1
ail = k=l+1 (7.20)
⎪
⎪ 
N
⎩e jφli sin ψli cos ψlk for i > 1
k=i+1
 N −1 l l
Substituting the expression k=l ak vk = εl − a N vlN and (7.20) into (7.19) and taking
partial derivatives with respect to φlk and ψlk results in the following:
∂ Pl
= c1 Im [E{εl vkl∗ }e− jφlk ] (7.21)
∂φlk
∂ Pl
= c2 Re [E{εl vkl∗ }e jφlk cos ψlk − E{εl ykl∗ } sin ψlk ] (7.22)
∂ψlk
where c1 does not depend on φlk , and c2 does not depend on ψlk . Setting both of the
aforementioned partial derivatives equal to zero then yields

φlk = arg [E{εl vkl∗ }] (7.23)

 
l∗ − jφlk
−1 Re(E{εl vk }e )
ψlk = tan (7.24)
Re(E{εl ykl∗ })
where the arctangent function is a multiple valued function. It turns out that by taking the
value lying in the range 0 → π radians, the output power is maximized. From (7.23) and
(7.24) it is seen that measuring the correlations E{εl vlk∗ } and E{εl ykl∗ } yields an indication
of the correct settings for φlk and ψlk . Starting with the first phase shifter coupler pair at
the extreme left of any stage and proceeding with each successive pair until the last phase
shift coupler pair is reached, φlk and ψlk are adjusted in accordance with (7.23) and (7.24).
However, each time a new setting of these parameters is made while proceeding toward
the output port, the value of εl changes; therefore, the values of the correlations change, so
new settings are required and the entire process is applied recursively. There is presently
no proof of convergence for this recursive adjustment process, but experience indicates
that convergence in five or six complete row adjustments is commonplace [3]. When the
eigenvalues of the input signal covariance matrix are widely separated, then convergence
is even faster.
The eigenvector transformation network parameter adjustment time is reduced, be-
cause an exact eigenvector decomposition is not required to realize substantial eigenvalue
spread reduction. This may be done by resorting to a piecemeal maximization technique
in which each phase shift coupler combination is set to produce maximum power at its
horizontal output port without regard to power at the entire stage output port and consid-
ering only the signals present at the two inputs of each individual combination. With this
piecemeal procedure, the phase shift and directional coupler are now set according to

φlk = arg(E{ykl vkl∗ }) (7.25)

 
1 −1 2|E{ykl vkl∗ }|
ψlk = tan (7.26)
2 |E{[ykl ]2 }| − |E{[vkl ]2 }|

where 0 ≤ tan−1 (·) ≤ π .

Each phase shift coupler combination is set from left to right along each row. At
the end of a row, drop down to the next stage and repeat the process until the bottom
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 311

7.2 Interference Cancellation with a Nolen Network Preprocessor 311

of the cascade is reached. The eigenvector beams produced by the piecemeal adjustment
procedure are shown by White [3] to be surprisingly close to those obtained by the com-
plete recursive adjustment procedure. If the adjustment of each phase shifter directional
coupler combination is one iteration, then a total of N (N − 1)/2 iterations completes the
piecemeal adjustment procedure.

7.2 INTERFERENCE CANCELLATION WITH

A NOLEN NETWORK PREPROCESSOR
Instead of using a Nolen eigenvector component preprocessor in cascade with a gradient-
type processor to alleviate the eigenvalue spread problem and speed convergence, the
entire task of interference cancellation and desired signal preservation are accomplished
using a Nolen cascade network alone by selecting a different performance measure from
that adopted in the preceding section [4]. Each stage in the Nolen network of Figure 7-3
introduces one null in the radiation pattern corresponding to the bottom output port. Errors
in parameter settings in the lower stages do not disturb the nulls set by the upper stages, but
the inverse statement is not true, so a nonadaptive Nolen network is vulnerable to errors in
the early stages. When adaptive control is applied to the parameter adjustment, however,
a portion of the lower stage adjustment capability partially compensates for upper stage
parameter setting errors. By virtue of the properties of the Nolen network, an appropriate
set of parameter adjustments for interference suppression are determined by means of a
step-by-step process in which each step involves only a single stage.

7.2.1 Problem Formulation

Consider the representation of a Nolen beamforming network shown in Figure 7-4 in which
a general (passive, lossless, and matched) network having N input ports and N output ports
follows an array having N elements. Under these conditions, the total output power equals
the total input power, and the transfer function of the network is a unitary matrix.
Let x be a complex vector representing the input signal envelopes at each sensor
element, and let ε be a complex vector representing the output signal envelopes, that is,

⎡ ⎤ ⎡ ⎤
x1 ε1
⎢ x2 ⎥ ⎢ ε2 ⎥
⎢ ⎥ ⎢ ⎥
x = ⎢ . ⎥, ε=⎢ . ⎥ (7.27)
⎣ .. ⎦ ⎣ .. ⎦
xN εN

FIGURE 7-4
x1 x2 xN Nolen beamforming
network.
e1
e2
Passive lossless matched
Nolen network
en

eN
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 312

312 CHAPTER 7 Cascade Preprocessors

The vectors x and ε are related by

ε = Gx (7.28)
where G is the transfer matrix of the Nolen network. Since G is unitary it follows that
ε † ε = x† x (7.29)
G† G = I (7.30)
Equation (7.30) is referred to as the “unitary constraint” that G must satisfy. Define the
covariance matrices of the input and the output signals by
Rx x = E{x∗ xT } (7.31)
Rεε = E{ε∗ ε T } = G∗ Rx x GT (7.32)
The total output power appearing at the nth output port is the nth diagonal element of
Rεε . Each output signal consists of a desired signal component and an interference signal
component (interference here denotes the sum of external jamming and internal thermal
noise). Assume that the desired signal is not correlated with any of the interference signal
sources; then Rεε are divided into signal and interference components
Rεε = G∗ [Rss + RII ]GT (7.33)
where
Rss = E{x∗s xsT } (7.34)
RII = E{x∗I xTI } (7.35)

The vectors xs and x I represent the signal and interference components, respectively, of
the input signal envelopes. At the nth output port denote the desired signal power by n
and the interference power by n so that
n = [Rss ]nn (7.36)
n = [RII ]nn (7.37)
That is, the output power of interest is the nth diagonal element of the corresponding
covariance matrix.
The goal is to maximize the signal power n and to minimize the interference power
n at the output port n. Since these two objectives conflict, a trade-off between them
is necessary. One approach to this trade-off selects the unitary transfer matrix, G, that
maximizes n = n /n . A more convenient way of attacking the problem is to adopt as
the performance measure
n = n − n (7.38)
where is a fixed scalar constant that reflects the relative importance on minimizing
interference compared with maximizing the signal. If = 0, then the desired signal is
maximized, while if → ∞, then only the interference is minimized.
There is a value = opt for which maximizing n produces exactly the same result
as maximizing the ratio n . The value opt is a function of the environment and is not
ordinarily known in advance. By setting = min where min is the minimum acceptable
signal-to-interference ratio that provides acceptable performance, then maximizing n
ensures that n is maximized under the conditions when it is most needed. If the signal
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 313

7.2 Interference Cancellation with a Nolen Network Preprocessor 313

environment improves, then n also improves, although not quite as much as if n were
maximized directly.
It is useful to define the matrix
Z = G∗ Rss GT − G∗ RII GT (7.39)
The performance measure n is then the nth diagonal element of Z. Furthermore
Z = G∗ MGT (7.40)
where
M = Rss − RII (7.41)
A consequence of (7.41) is that the matrix M (and hence Z) have both positive and negative
eigenvalues. Selecting the transfer matrix, G, that maximizes the nth diagonal element of
Z subject to (7.30) maximizes the element n . Using the method of Lagrangian multipliers
to maximize the element Z nn with the unitary constraint on G and setting the resulting
gradient equal to zero yield the relation

Mn L G Lm = λG mn (7.42)
L

Equation (7.42) is precisely the form of an eigenvector equation. Consequently, the nth
column of GT is the eigenvector of the matrix M corresponding to the largest eigenvalue.
When the nth column of GT is so constructed, then the element Z nn equals this largest
eigenvalue, and all other elements of the nth row and the nth column of Z vanish.
It is desired to obtain a unitary transfer matrix G for which the elements of the nth
row are the elements of the eigenvector of the matrix M corresponding to the maximum
eigenvalue (resulting in maximizing n ). If the first stage of the cascade Nolen network
is adjusted to minimize 1 , if there is no conflict with maximizing n , and if the first
diagonal element of Z is set equal to the most negative eigenvalue of M, then the off-
diagonal elements of the first row and first column of Z will disappear, and the first row of
G will correspond to the appropriate eigenvector. Likewise, proceed to adjust the second
stage to minimize 2 . This second adjustment results in the diagonalization of the second
row and the second column of Z, and the second row of G corresponds to the second
eigenvector. Continue in this manner adjusting in turn to minimize the corresponding 
until reaching stage n. At this point (the nth stage) it is desired to maximize n so the
adjustment criterion must be reversed.
The physical significance of the foregoing adjustment procedure is that the upper
stages of the Nolen network are adjusted to maximize the interference and minimize the
desired signal observed at the output of each stage. This process is the same as maximizing
the desired signal and minimizing the interference that proceeds downward to the lower
stages. When the nth stage is reached where the useful output is desired, then the desired
signal should be maximized and the interference minimized so the adjustment criterion is
reversed.

7.2.2 Recognition of the Desired Signal

The performance criterion must have the ability to distinguish between the desired signal
and interference. If the desired signal comes from an unknown direction (as would usually
be the case), it is necessary to recognize some distinguishing characteristics of the signals.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 314

314 CHAPTER 7 Cascade Preprocessors

Since communication systems designed to operate in a jamming environment commonly

employ some form of spread spectrum modulation, then the signal recognition scheme is
based on the characteristics associated with spread spectrum signals.
Spread spectrum modulation applies some form of pseudo-random coding at the
transmitter to widen the transmitted signal spectrum. When the receiver applies appropriate
decoding to the spread spectrum signal, the original (unspread) desired signal is recovered
with a narrow spectrum, whereas any interference still emerges from the decoder having
a wide spectrum. The demodulation process for spread spectrum signals uses narrowband
filters to at least partially separate the desired signal from the surrounding interference.
A first step toward determining whether the parameter adjustments in each stage of
the Nolen network are succeeding in the diagonalization of the matrix Z is to define a
generalized correlation product [4]

A∗ ⊗ B = E{A∗s Bs } − E{A∗I B I } (7.43)

where A is the complex envelope of one waveform having desired signal and interference
components As and A I . Likewise, B represents the complex envelope of a second wave-
form having desired signal and interference components Bs and B I . If we assume that
the desired signal is modulated with pseudo-random phase reversals and that a synchro-
nized key generator is available for demodulation at the receiver, Figure 7-5 shows the
block diagram of a generalized correlator that forms an estimate of A∗ ⊗ B. After passing
the received signal through a synchronized demodulator to obtain the original unspread
signal, narrow passband filters extract the desired signal while band reject filters extract
the interference. The narrow passband outputs are applied to one correlator (the “signal
correlator”) that then forms estimates of E{A∗s Bs }. The band reject outputs likewise are
applied to a second correlator (the “interference correlator”) that then forms estimates of
E{A∗I B I }. A weighted combination of the outputs (signal component weighted by K s and
interference component weighted by K I ) then forms an estimate of the complex quantity
A∗ ⊗ B. The ratio of K s to K I determines the effective value of .
If the correlation product operator ⊗ is taken to include operations on vector quantities,
then we may write

M = x∗ ⊗ xT (7.44)
Z = h∗ ⊗ h T (7.45)

FIGURE 7-5 Bandpass and

Generalized bandreject
correlator. filters
Balanced
A
modulator I +
Signal
Ks Re (A∗ ⊗ B)
correlator Q −
Ks

Synchronized
key generator
I +
Interference
Kl lm (A∗ ⊗ B)
correlator Q −
Balanced Kl
B
modulator
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 315

7.3 Gram–Schmidt Orthogonalization Preprocessor 315

The generalized correlator network of Figure 7-5 provides a basis for estimating the
elements of the matrix Z and determining whether the diagonalization of this matrix is
complete.

7.2.3 Piecemeal Adjustment of the Nolen Network Stages

The piecemeal adjustment procedure for Nolen network parameters described by (7.25)
and (7.26) can no longer employ the correlation products of ykl vkl∗ , (ykl )2 and (vkl )2 if the
modified performance measure l is to be maximized (or minimized) for the lth stage,
since it is now the generalized correlation product that is related to the elements of the
matrix Z [4]. Therefore, (7.25) and (7.26) should now be modified to yield

φlk = arg{ykl ⊗ vkl∗ } (7.46)

 
1 −1 2|ykl ⊗ vkl∗ |
ψlk = tan (7.47)
2 |ykl ⊗ ykl∗ | − |vkl ⊗ vkl∗ |

The value of the arctangent function lying between −π and zero radians minimizes l ,
whereas the value lying between zero and π radians maximizes l . The complete piece-
meal adjustment of a full Nolen cascade network for an N -element array using (7.46)
and (7.47) requires N (N − 1)/2 iterations, making the practical use of a Nolen eigenvec-
tor component cascade processor less attractive when compared with the Gram–Schmidt
cascade described in the next section.

7.3 GRAM–SCHMIDT ORTHOGONALIZATION

PREPROCESSOR
Section 7.1 shows that using a unitary transformation to obtain an orthogonal signal
set accelerates convergence of a gradient-based algorithm by circumventing the eigen-
value spread problem. The Gram–Schmidt orthogonalization procedure [15] is another
way to obtain an orthogonal signal set. While a normalized Gram–Schmidt preproces-
sor may be followed by a Howells–Applebaum adaptive processor to realize acceler-
ated convergence as illustrated in Figure 7-6, canceling the interference is also possible
by using the preprocessor alone in a coherent sidelobe canceller (CSLC) configuration.

FIGURE 7-6
Gram–Schmidt
x1 x2 x3 x4 x5
orthogonalization
network with
Gram − Schmidt orthogonalization Howells–Applebaum
preprocessor adaptive processor
for accelerated
y1 y2 y3 y4 y5 convergence.

Final stage Howells − Applebaum

adaptive MSNR processor

Array output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 316

316 CHAPTER 7 Cascade Preprocessors

FIGURE 7-7
Gram–Schmidt x x1 = v11 x2 = v12 x3 = v13 x4 = v14 x5 = v15
transformation to
obtain independent
variables. Level 1

v12 v22 v32 v24 v25

Level 2

v13 v32 v33 v34 v35

Level 3

v41 v42 v43 v44 v45

Level 4

y y1 = v51 y2 = v25 y3 = v53 y4 = v45 y5 = v55 Level 5

The Gram–Schmidt orthogonalization cascade preprocessor is very easily implemented,

has excellent transient response characteristics, and therefore presents an attractive alter-
native to the eigenvector component preprocessor of Section 7.1.
To understand the coordinate transformation based on the Gram–Schmidt orthogonal-
ization procedure, consider the five-element array of Figure 7-7 in which a transformation
(indicated by a square) is introduced at each node in the cascade network to achieve in-
dependence (orthogonality) between the transformed output signal and a reference input
signal [5].
Each transformation in the network has two input signals vkk = yk (which is the
reference signal for level k) and vnk where n ≥ k + 1. Every transformation achieves
independence between vkk and vnk . At the last level the output signals y1 , y2 , . . . , yk are
mutually independent so that E{ym yn∗ } = 0 for all m = n. Having achieved mutual
independence among the output signals, a single Howells–Applebaum loop is applied to
each signal where the loop gain is high for small signals and low for large signals, thereby
equalizing the time response for the various signals in the interference environment. The
first level of the network provides the necessary transformation of vn1 , n ≥ 2, to achieve
independence from v11 = y1 for the signal set vn2 , n ≥ 2. The signal set vn2 , n ≥ 3 is
then transformed in the second level of the network to achieve independence from the
signal v22 = y2 . This process continues until a complete transformation G is obtained,
where
y = Gx (7.48)
The Gram–Schmidt transformation, G, is found by ensuring that the appropriate
transformations take place at each node indicated in Figure 7-7. Recalling from Chapter 4
that a single Howells–Applebaum adaptive loop yields an output signal that in the steady
state is uncorrelated with a selected reference signal, we see that each transformation in
Figure 7-7 may be realized with a single Howells–Applebaum adaptive control loop as
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 317

7.3 Gram–Schmidt Orthogonalization Preprocessor 317

FIGURE 7-8
x1 x2 x3 x4 x5 Gram–Schmidt
orthogonalization for
+ + + +
− − − − five-element array
y1
realized with
u11 u12 u13 u14 Howells–Applebaum
Filter Filter Filter Filter
adaptive loops.

* * * *
y2
+ + +
− − −

u22 u23 u24

Filter Filter Filter

* * *
y3
+ +
− −

u33 u34
Filter Filter

* *
y4
+
−

u44
Filter

y5

shown in Figure 7-8. The transformation occurring at each node in Figure 7-7 (using the
weight indexing of Figure 7-8) are expressed as
vnk+1 = vnk − u k(n−1) vkk , k+1≤n ≤ N (7.49)
where N = number of elements in the array. In the steady state, the adaptive weights have
values given by
 k∗ 
v vk
u k(n−1) =  k n  (7.50)
vkk∗ vkk
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 318

318 CHAPTER 7 Cascade Preprocessors

where the overbars denote expected values. The transformation represented by (7.48) and
(7.49) is a close analog of the familiar Gram–Schmidt orthogonalization equations. For
an N -element array, N (N − 1)/2 adaptive weights produce the Gram–Schmidt orthogo-
nalization. Since an adaptive Howells–Applebaum maximum SNR processor requires N
adaptive weights, the configuration of Figure 7-6 requires a total of N (N + 1)/2 adaptive
weights.
The transformation of the input signal vector x into a set of independent output
signals y is not unique. Unlike the eigenvector transformation, which yields output signals
in a normal coordinate system in which the maximum power signal is the first output
component, the Gram–Schmidt transformation adopts any component of x as the first
output component and any of the remaining components of x as the signal component vnk
to be transformed by way of (7.49).
The fact that the transformation network of Figure 7-7 yields a set of uncorrelated
output signals suggests that this network functions in the manner of a CSLC system whose
output signal (in the steady state) is uncorrelated with each of the auxiliary channel input
signals. Recall from the discussion of the SNR performance measure in Chapter 3 that
an N − 1 element CSLC is equivalent to an N -element adaptive array with a generalized
signal vector given by
⎡ ⎤
1
⎢0⎥
⎢ ⎥
t = ⎢ .. ⎥ (7.51)
⎣.⎦
0
It is shown in what follows that by selecting x5 of Figure 7-7 as the main beam channel
signal b (so that tT = [0, 0, . . . , 0, 1]) and z = y5 as the output signal, then the cascade
preprocessor yields an output that converges to z = b − wT x. Here x is the auxiliary
channel signal vector, and w is the column vector of auxiliary channel weights for the
equivalent CSLC system of Figure 7-9 that minimizes the output noise power. The CSLC
system of Figure 7-9 and the cascade preprocessor of Figure 7-7 are equivalent in the

FIGURE 7-9 Auxillary channels

Equivalent CSLC
system for the
cascade x1 x2 x3 x4 x5 = b
preprocessor
system of w1 w2 w3 w4
Figure 7-7.

+
−

Array
output z
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 319

7.3 Gram–Schmidt Orthogonalization Preprocessor 319

sense that the cascade network output is the same as the CSLC system output once the
optimum solutions are reached. Recall from Chapter 3 that the optimum weight vector for
the CSLC system is given by

wopt = R−1 ∗
x x (x b) (7.52)

If the main beam channel signal is replaced by a locally generated pilot signal p(t), then

wopt = R−1 ∗
x x (x p) (7.53)

and the array output is given by

z(t) = p(t) − w† x(t) (7.54)

7.3.1 Convergence of the Gram–Schmidt Cascade Preprocessor

Let us first demonstrate that the cascade network of Figure 7-7 converges to the solution
provided by a conventional CSLC system. Complete steady-state equivalence is shown
for the three-element CSLC system of Figure 7-10 (in which two Howells–Applebaum
SLC control loops yield weights w 1 and w 2 ) and the three-element cascade system of
Figure 7-11 (in which three Howells–Applebaum control loops yield weights u 11 , u 12 ,
and u 22 ) [5]. The analysis for this case may then be easily extended to any arbitrary
number of elements.
For the conventional three-element CSLC system of Figure 7-10, the steady-state
solution for the optimum weight vector (whose components are w 1 and w 2 ) is given by
(7.52), where Rx x = Rnn in the absence of a desired signal. From (7.52) it follows that
the optimum steady-state weights of Figure 7-10 are given by
   ∗ −1  
w1 x1 x1 x1∗ x2 x1∗ b
=
w2 x2∗ x1 x2∗ x2 x2∗ b
  
x2∗ x2 −x1∗ x2 x1∗ b
−x2∗ x1 x1∗ x1 x2∗ b
= (7.55)
(x1∗ x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )

FIGURE 7-10
Conventional
x1 x2 x3 = b three-element CSLC
system having two
Howells–Applebaum
SLC control loops.
SLC SLC
w1 w2
+
−
Σ
z z

z
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 320

320 CHAPTER 7 Cascade Preprocessors

FIGURE 7-11
Cascade x1 x2 x3 = b
Gram–Schmidt
three-element CSLC
system having three
Howells–Applebaum + +
SLC control loops. SLC − SLC −
x1 = y1 u11 u12

v22 = y2 v23

+
SLC −
u22

v33 = y3 = z

so that

(x2∗ x2 )(x1∗ b) − (x1∗ x2 )(x2∗ b)

w1 = (7.56)
(x1∗ x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )
(x ∗ x1 )(x2∗ b) − (x2∗ x1 )(x1∗ b)
w 2 = ∗1 (7.57)
(x1 x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )

For the cascade system of Figure 7-11, it follows from (7.49) and (7.50) that

y2 = x2 − u 11 x1 (7.58)
v32 = b − u 12 x1 (7.59)
z = v32 − u 22 y2 (7.60)

The weight elements have steady-state values given by

(x1∗ x2 )
u 11 = (7.61)
(x1∗ x1 )
(x ∗ b)
u 12 = ∗1 (7.62)
(x1 x1 )
(y2∗ v32 )
u 22 = (7.63)
(y2∗ y2 )

From (7.58) to (7.60) it follows that

z = b − [(u 12 − u 11 u 22 )x1 + u 22 x2 ] (7.64)

Comparing (7.64) with Figure 7-10, we see that w 2 and w 1 fill the roles taken by u 22 and
u 12 − u 11 u 22 , respectively, in the cascade system.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 321

7.3 Gram–Schmidt Orthogonalization Preprocessor 321

If we substitute (7.58) and (7.59) into (7.63) it then follows that

(x1∗ x1 )(x2∗ b) − (x2∗ x1 )(x1∗ b)

u 22 = (7.65)
(x1∗ x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )
which is identical with the solution (7.57) for w 2 . Likewise, if we substitute (7.61), (7.62),
and (7.63) into u 12 − u 11 u 22 it follows that

(x1∗ b)(x2∗ x2 ) − (x1∗ x2 )(x2∗ b)

u 12 − u 11 u 22 = (7.66)
(x1∗ x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )
which is identical with the solution (7.56) for w 1 . Therefore, the steady-state solutions
reached by the two CSLC of Figures 7-10 and 7-11 are identical with weight element
equivalences given by

w 2 = u 22
(7.67)
w 1 = u 12 − u 11 u 22
Additional weight element equivalences may easily be obtained for an arbitrary num-
ber of sensor elements by considering the transformation (7.48). By induction, the trans-
formation G is a lower triangular matrix whose elements gij are given by the following
iterative relationships:

gij = 0 if i < j (7.68)

gij = 1 if i = j (7.69)

i−1
gij = − u ∗jk · gi(k+1) for i > j (7.70)
k= j

The jth row of the transformation matrix G contains the desired weight equivalences for
a j-element CSLC system where the auxiliary channel weights are given by
⎫
w j−1 = −g j ( j−1) ⎪
⎪
w j−2 = −g j ( j−2) ⎪⎬
.. ⎪
(7.71)
. ⎪
⎪
⎭
w 1 = −g j1
Having found the steady-state weight element equivalence relationships, it is now
appropriate to consider the transient response of the networks of Figures 7-10 and 7-11.
The transient responses of the two CSLC systems under consideration are investigated
by examining the system response to discrete signal samples of the main beam and the
auxiliary channel outputs. The following analysis assumes that the adaptive weights reach
their steady-state expected values on each iteration and therefore ignores errors that would
be present due to loop noise. Let xkn denote the nth signal sample for the kth element
channel. The main beam channel samples are denoted by bn . For a system having N
auxiliary channels and one main beam channel, after N independent samples have been
collected, a set of auxiliary channel weights are computed using


N
w k xkn = bn for n = 1, 2, . . . , N (7.72)
k=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 322

322 CHAPTER 7 Cascade Preprocessors

The foregoing system of equations yields a unique solution since the matrix defined
by the signal samples xkn is nonsingular provided that either thermal receiver noise
or N directional interference sources are present. Equation (7.72) is closely related to
∗
(7.52), since multiplying both sides of (7.72) by xmn and summing over the index n
yields

N 
N 
N
∗ ∗
w k xmn xkn = xmn bn (7.73)
n=1 k=1 n=1

which are written in matrix form as

∧
R̂x x w = (x∗ b) (7.74)

Comparing (7.74) with (7.52) reveals that these two equations yield similar solutions for
the weight vector w in a stationary signal environment.
The iterative weight correction procedure for a single Howells–Applebaum SLC loop
is modeled as shown in Figure 7-12. On receipt of the ith signal sample, the resulting
change in the weight u k(n−1) is computed by applying (7.50) and assuming that the signal
samples are approximately equal to their expected values to yield
vkk∗ [vnk (i) − u k(n−1) (i − 1) · vkk (i)]
u k(n−1) (i) = (7.75)
(vkk∗ vkk )
The element weight value is then updated in accordance with

u k(n−1) (i) = u k(n−1) (i − 1) + u k(n−1) (i) (7.76)

FIGURE 7-12 vkk(i) vnk (i)

Iterative weight
correction model for
+
Howells–Applebaum −
SLC loops.

uk(n−1)(i−1) vnk+ 1(i)

z−1

uk(n−1)(i)

Δuk(n−1)(i)

uk(n −1)(i) = uk(n −1)(i −1) + Δuk(n −1)(i)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 323

7.3 Gram–Schmidt Orthogonalization Preprocessor 323

With all weights in the cascade Gram–Schmidt network initially set to zero, it follows that
on receipt of the first signal sample v22 = x21 , and v32 = v33 = b1 so the change in weight
settings after the first iteration results in
∗ 1 ∗
x11 v2 x11 x21
u 11 (1) = = (7.77)
|x11 | 2 |x11 |2
∗ 1
x v x ∗ b1
u 12 (1) = 11 32 = 11 2 (7.78)
|x11 | |x11 |
and
∗
v22∗ v32 x21 b1
u 22 (1) = = (7.79)
|v2 |
2 2
|x21 |2

On receipt of the second signal sample v22 = x22 − u 11 (1)x12 , v32 = b2 − u 12 (1)x12 ,
and v33 = v32 − u 22 (1)v22 so that
∗
x12 [x22 − u 11 (1)x12 ]
u 11 (2) = (7.80)
|x12 |2
x ∗ [b2 − u 12 (1)x12 ]
u 12 (2) = 12 (7.81)
|x12 |2
(x22 − u 11 (1)x12 ) ∗ [b2 − u 12 (1)x12 − u 22 (1)(x22 − u 11 (1)x12 )]
u 22 (2) = (7.82)
|x22 − u 11 (1)x12 |2
Substitute (7.77) to (7.79) into (7.80) to (7.82), it then follows that
∗
x12 x22
u 11 (2) = u 11 (1) + u 11 (2) = (7.83)
|x12 |2
x ∗ b2
u 12 (2) = u 12 (1) + u 12 (2) = 12 2 (7.84)
|x12 |
x11 b2 − x12 b1
u 22 (2) = u 22 (1) + u 22 (2) = (7.85)
x11 x22 − x12 x21
By use of the weight equivalence relationships of (7.67) it follows that w 2 (2) = u 22 (2)
and
x22 b1 − x21 b2
w 1 (2) = u 12 (2) − u 11 (2)u 22 (2) = (7.86)
x11 x22 − x12 x21
Note, however, that w 1 (2) and w 2 (2) of (7.85) and (7.86) are the weights that satisfy (7.72)
when N = 2. Therefore, for a two-auxiliary channel CSLC, the cascade Gram–Schmidt
network converges to a set of near-optimum weights in only two iterations. The foregoing
analysis can also be carried out for the case of an N -auxiliary channel CSLC and leads to
the conclusion that convergence to a set of near-optimum weights in the noise-free case
occurs after N iterations. Averaging the input signals mitigates the effects of noise-induced
errors.
Note that using (7.75) to update the weight settings results in values for u 11 and u 12 (the
weights in the first level) that depend only on the current signal sample. The update setting
for u 22 , however, depends on the last two signal samples. Therefore, an N -level cascade
Gram–Schmidt network always updates the weight settings based on the N preceding
signal samples.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 324

324 CHAPTER 7 Cascade Preprocessors

The conventional two-weight element CSLC does not converge to a near-optimum

weight solution with only two signal samples. To see this result note that for the configu-
ration of Figure 7-10 after receipt of the first signal sample that
∗
x11 b1
w 1 (1) = (7.87)
|x11 |2
x ∗ b1
w 2 (1) = 21 2 (7.88)
|x21 |
Following the second signal sample
∗
x12 (b2 − w 1 (1)x12 − w 2 (1)x22 )
w 1 (2) = w 1 (1) + (7.89)
|x12 |2
∗
x (b2 − w 1 (1)x12 − w 2 (1)x22 )
w 2 (2) = w 2 (1) + 22 (7.90)
|x12 |2
The weights given by (7.89) and (7.90) are not the same as the weights given by (7.85)
and (7.86) and consequently do not satisfy (7.72). Simulation results show that the con-
figuration of Figure 7-10 yields a transient response that is considerably slower (even
with a higher level of loop noise) than the transient response of a Gram–Schmidt cascade
preprocessor.

7.4 SIMULATION RESULTS

A four-element network configuration like that of Figure 7-6 was simulated with a digital
LMS adaptive loop used in place of each Howells–Applebaum loop in the final maximum
SNR stage. The four loop gains of the LMS adaptive loops in the final maximum SNR
stage of the selected configuration were set inversely proportional to an estimate of the
input signal power to each loop p̂i obtained by averaging the squared loop input signal
over K samples so that si = α L / p̂i . Likewise, the loop gain of the digital version of
each Howells–Applebaum loop in the cascade preprocessor was selected as G i = α/ p̂i .
Selecting the adaptive loop gains in this manner equalizes the transient response time to
the signals y1 , . . . , y4 . Furthermore, the correlations (products) performed in the analog
processor of Figure 7-8 are represented in the digital version as products of individual
sample values averaged over K samples. The averaging interval also permits signal power
estimates to be obtained by averaging the instantaneous signal powers appearing at the
loop inputs of the final processing stage. The array configuration and signal environment
correspond to that given in Figure 7-4, and the desired signal was again selected to be
a biphase modulated signal having a 0◦ or 180◦ reference phase equally likely at each
sample.
The algorithm performance in each case is given in terms of the resulting output SNR
versus the number of iterations. The signal environment conditions are specified to repre-
sent two values eigenvalue spread: λmax /λmin = 16,700 and 2,440. The convergence results
given in Figures 7-13 to 7-17 are for an eigenvalue spread of 16,700 and five values of α L
ranging over 0.1, 0.25, 0.3, 0.5, and 0.65. Likewise the convergence results displayed in
Figures 7-18 to 7-20 are for An eigenvalue spread of 2,440 and various conditions for the
gain, K, and the initial weight vector setting. All other conditions are specified by the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 325

7.4 Simulation Results 325

15 .00 FIGURE 7-13

Output SNR versus
10.00 Optimum number of iterations
for Gram–Schmidt
5.00
Output SNR (dB)
cascade
preprocessor with
0.00
eigenvalue spread =
−5.00 16,700. Algorithm
parameters are K =
−10.00 3, α = α L = 0.5,
and w(0) = 0 with
−15.00 s/n = 10 for which
SNRopt = 15
−20.00
1 2 3 4 5 6 78 1E1 2 3 4 5 6 78 1E2 2 3 4 5 6 78 1E3 (11.76 dB).
Number of iterations

15 .00 FIGURE 7-14

Output SNR versus
10.00 Optimum number of iterations
for Gram–Schmidt
5.00
Output SNR (dB)

cascade
preprocessor with
0.00
eigenvalue spread =
−5.00 16,700. Algorithm
parameters are K =
−10.00 3, α = α L = 0.1,
and w(0) = 0 with
−15.00 s/n = 10 for which
SNRopt = 15.
−20.00
1 2 3 4 5 6 78 1E1 2 3 4 5 6 78 1E2 2 3 4 5 6 78 1E3
Number of iterations

15 .00 FIGURE 7-15

Output SNR versus
10.00 Optimum number of iterations
for Gram–Schmidt
5.00
Output SNR (dB)

cascade
preprocessor with
0.00
eigenvalue spread =
−5.00 16,700. Algorithm
parameters are
−10.00 K = 9, α = 0.3, α L =
0.25α, and w(0) = 0
−15.00 with s/n = 10 for
which SNRopt = 15.
−20.00
1 2 3 4 5 6 78 1E1 2 3 4 5 6 78 1E2 2 3 4 5 6 78 1E3
Number of iterations

number of measurement samples averaged per iteration K , the Howells–Applebaum

loop gain α, the LMS loop gain α L , the desired signal-to-thermal noise ratio s/n, and
the initial weight vector selection w(0). For s/n = 10, the optimum output SNR is
SNRopt = 15(11.76 dB), whereas for s/n = 0.1 the optimum output SNR is SNRopt =
0.15(−8.2 dB).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 326

326 CHAPTER 7 Cascade Preprocessors

FIGURE 7-16 15 .00

Output SNR versus
number of iterations 10.00 Optimum
for Gram–Schmidt
5.00

Output SNR (dB)

cascade
preprocessor with
0.00
eigenvalue spread =
16,700. Algorithm −5.00
parameters are K =
9, α = α L = 0.3, −10.00
and w(0) = 0 with
s/n = 10 for which −15.00
SNRopt = 15.
−20.00
1 2 3 4 5 6 78 1E1 2 3 4 5 6 78 1E2 2 3 4 5 6 78 1E3
Number of iterations

FIGURE 7-17 15 .00

Output SNR versus
number of iterations 10.00 Optimum
for Gram–Schmidt
5.00
Output SNR (dB)

cascade
preprocessor with
0.00
eigenvalue spread =
16,700. Algorithm −5.00
parameters are
K = 9, α = α L = −10.00
0.65, and w(0) = 0
with s/n = 10 for −15.00
which SNRopt = 15.
−20.00
1 2 3 4 5 6 78 1E1 2 3 4 5 6 78 1E2 2 3 4 5 6 78 1E3
Number of iterations

FIGURE 7-18 1 5.00

Output SNR versus
10 .00 Optimum
number of iterations
for Gram–Schmidt 5.00
Output SNR (dB)

cascade Cascade preprocessor

preprocessor and 0.00
LMS processor with −5.00
eigenvalue spread =
2,440. Algorithm −10.00
parameters are −15.00
K = 3, α = α L = LMS curve
0.1, and −20.00
wT (0) = [1, 0, 0, 0]
−25.00
with s/n = 10 for 1 2 3 4 5678 1E1 2 3 4 5678 1E2 2 3 4 5678 1E3 2 3 4 5
which SNRopt = 15. Number of iterations

Observations from Figures 7-13 to 7-20 include the following:

1. A large spread in covariance matrix eigenvalues has little effect on the speed of the
algorithm response (compare Figure 7-14 with Figure 7-20).
2. Comparable steady-state output variations (reflecting the same level of loop noise) are
obtained for K = 3, α = 0.3/ p and K = 9, α = 0.1/ p, thereby indicating that the
product K · α determines the loop noise level (compare Figure 7-14 with Figure 7-15).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 327

7.4 Simulation Results 327

−5.00 FIGURE 7-19

Output SNR versus
−10.00 Optimum number of iterations
Output SNR (dB) −15.00 for Gram–Schmidt
cascade
−20.00 preprocessor with
−25.00
eigenvalue spread =
2,440. Algorithm
−30.00 parameters are
K = 3, α = α L =
−35.00
0.1, and w(0) = 0
−40.00 with s/n = 0.1 for
which SNRopt =
−45.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2 0.15(−8.2 dB).
Number of iterations

15.00 FIGURE 7-20

Output SNR versus
10 .00 Optimum number of iterations
5.00 for Gram–Schmidt
Output SNR (dB)

cascade
0.00 preprocessor with
−5.00 eigenvalue spread =
2,440. Algorithm
−10.00 parameters are
−15.00
K = 3, α = α L =
0.1, and w(0) = 0
−20.00 with s/n = 10 for
which SNRopt = 15.
−25.00
1 2 3 4 5678 1E1 2 3 4 5678 1E2 2 3 4 5678 1E3 2 3 4 5
Number of iterations

3. Only a slight increase in the loop gain α causes the weights to become excessively
noisy (compare Figure 7-16 with Figure 7-13 and Figure 7-17).
4. The appropriate value to which the product K · α should be set for an acceptable level of
loop noise depends on the value of SNRopt (compare Figures 7-19 and 7-20). Smaller
values of SNRopt require smaller values of the product K · α to maintain the same
loop noise level. It may also be seen that the adaptation time required with these
parameter values for the Gram–Schmidt orthogonalization preprocessor is greater than
that required for either a recursive or direct matrix inversion (DMI) algorithm. The
question of relative adaptation times for the various algorithms is pursued further in
Chapter 10.
5. The degree of transient response improvement that are obtained with a cascade pre-
processor is shown in Figure 7-18 where the Gram–Schmidt cascade preprocessor
response is compared with the LMS processor response. The LMS curve in this fig-
ure was obtained simply by removing the cascade preprocessing stage in Figure 7-6,
thereby leaving the final maximum SNR stage consisting of four LMS adaptive loops.
The greater the eigenvalue spread of the Rxx matrix, then the greater is the degree of
improvement that are realized with a cascade preprocessor compared with the LMS
processor.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 328

328 CHAPTER 7 Cascade Preprocessors

7.5 SUMMARY AND CONCLUSIONS

The sensitivity of the convergence rate of the LMS and Howells–Applebaum adaptive
processors to eigenvalue spread are greatly reduced by using a cascade preprocessor that
resolves the input signal vector into orthogonal components. A Nolen cascade network
resolves the input signal vector into orthogonal eigenvector beam components but re-
quires an excessive number of iterations to adjust the network parameters before rapid
convergence is achieved. The Gram–Schmidt cascade preprocessor resolves the input sig-
nal vector into orthogonal components (although these components are not eigenvector
beam components) and succeeds in greatly reducing the sensitivity of the convergence
rate to eigenvalue spread while requiring a relatively small number of iterations to ad-
just the preprocessor parameters to the desired steady-state values. The Gram–Schmidt
preprocessor therefore is a highly attractive candidate for achieving fast algorithm conver-
gence rates while retaining the implementation simplicity associated with the LMS and
Howells–Applebaum adaptive processors.

7.6 PROBLEMS
Problems 1 through 12 all concern Cascade SLC Control Loops
1. Consider the cascade control loop configuration of Figure 7-21.
(a) Show that the weight equivalences between the configuration of Figure 7-21 and the
standard CSLC configuration are given by

w1 = u1 + u3 w 2 = −u 2 u 3

(b) Show that the steady-state weight values of Figure 7-21 are given by

x1∗ b x2∗ x1
u1 = u 2 =
x1∗ x1 x ∗ x2
 ∗ 2 ∗ 
(x1 x2 ) (x1 x1 )(x2 b) − (x2∗ x1 )(x1∗ b)
∗
u3 = ∗
(x1 x1 ) (x2∗ x2 )(x1∗ x1 ) − (x1∗ x2 )(x2∗ x1 )

FIGURE 7-21
Cascade x2 x1 b
arrangement of three
SLC +
Howells–Applebaum −
SLC control weight
u1
loops yielding SLC +
−
convergence to weight
incorrect solution. u2

SLC +
−
weight
u3
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 329

7.6 Problems 329

Note that these steady-state weights do not correspond to the correct solution for the
standard CSLC configuration.
Note: Problems 2 through 12 concern derivations that may be found in [7].
2. Consider the three-element array cascade configuration of Figure 7-11.
(a) Show that the steady-state weight values for this configuration are given by (7.61), (7.62),
and (7.65).
(b) Using (7.64) for the array output, show that

(x1∗ x3 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x3 )

z = x3 − u 22 x2 − x1
(x1∗ x1 )(x2∗ x2 ) − (x2∗ x1 )(x1∗ x2 )

(c) The average noise output power is defined by N0 = |z|2 = zz ∗ . For notational simplicity
let u 1 = u 11 , u 2 = u 12 , u 3 = u 22 , and show that

N0 = (x3∗ x3 ) + |u 3 |2 (x2∗ x2 ) + |u 2 − u 1 u 3 |2 (x1∗ x1 )

− u 3 (x3∗ x2 ) − u ∗3 (x2∗ x3 )
− (u 2 − u 1 u 3 )(x3∗ x1 ) − (u ∗2 − u ∗1 u ∗3 )(x1∗ x3 )
+ u 3 (u ∗2 − u ∗1 u ∗3 )(x1∗ x2 ) + u ∗3 (u 2 − u 1 u 3 )(x2∗ x1 )

3. Using the weight notation of Problem 2(c) for the configuration of Figure 7-11, let

u n = u n + δn

where δn represents the fluctuation component of the weight. The total output noise power is
then given by

N0TOT = N0 + Nu
= |x3 − (u 3 + δ3 )x2 − [(u 2 + δ2 ) − (u 1 + δ1 )(u 2 + δ3 )]x1 |2

where Nu represents the excess weight noise due to the fluctuation components. Note that
δ n = 0 and neglect third- and fourth-order terms in δn to show that

Nu = |δ3 |2 [(x2∗ x2 ) − (x2∗ x1 )u 1 ] + (x1∗ x1 )|δ2 − u 3 δ1 |2

4. The weight u 1 = u 11 of Figure 7-11 satisfies the differential equation

   
u̇ 1 u1
τ1 + = x1∗ (x2 − u 1 x1 )
G1 G1

where G 1 denotes amplifier gain, and τ1 denotes time constant of the integrating filter in the
Howells–Applebaum control loop for u 1 . The foregoing equation is rewritten as
 
1 1 G1
u̇ 1 + x1∗ x1 + u 1 = x1∗ x2 where α1 =
α1 G1 τ1

so that
 
1 1
u̇ 1 + x1∗ x1 + u 1 = x1∗ x2
α1 G1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 330

330 CHAPTER 7 Cascade Preprocessors

Subtracting the mean value differential equation from the instantaneous value differential
equation and recalling that u 1 = u 1 + δ1 then yields
 
1 1
δ̇1 + + x1∗ x1 δ1 + (x1∗ x1 − x1∗ x1 )u 1 = (x1∗ x2 − x1∗ x2 )
α1 G1

provided that the second-order term δ1 (x1∗ x1 − x1∗ x1 ) is ignored. In the steady state where

x1∗ x2
u1 =
x1∗ x1

show that
   
1 1 (x1∗ x2 ) ∗ 1
δ̇1 + x1∗ x1 + δ1 = x1∗ x2 − x1 x1 = f 1 (t)
α1 G1 (x1∗ x1 ) α1

where f 1 (t) may be regarded as a random variable.

5. The solution of the differential equation obtained in Problem 4 is given by
 t    
1
δ1 (t) = f 1 (τ ) exp −α1 x1∗ x1 + (t − τ ) dτ
0 G1

so that
 
1
δ̇1 (t) = f 1 (t) − α1 x1∗ x1 + δ1 (t)
G1

Substitute δ̇1 (t) into the differential equation for δ̇1 (t) obtained in Problem 4, and show that this
reduces the resulting expression to an identity, thereby proving that δ1 (t) is indeed a solution
to the original differential equation.
6. The second moment of the fluctuation δ1 is given by δ1∗ δ1 .
(a) Use the expression for δ1 (t) developed in Problem 5 to show that
    t  t    
1 1
δ1∗ δ1 ∗
= exp −2α1 x1 x1 + ∗ ∗
f 1 (τ ) f 1 (u) exp α1 x1 x1 + (τ + u) dτ du
G1 0 0 G1

(b) Assume that the random variable

 
(x1∗ x2 ) ∗
f 1 (t) = α1 x1∗ x2 − x1 x1
(x1∗ x1 )

has a correlation interval denoted by ε, such that values of f 1 (t) separated by more than ε
are independent. Let
 t
f 1∗ (τ ) f 1 (u) du = f 1∗ f 1 ε for t > τ + ε.
0

Show that δ1∗ δ1 then reduces to

ε f 1∗ f 1
δ1∗ δ1 = for t > τ + ε
2α1 [(x1∗ x1 ) + 1/G 1 ]
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 331

7.6 Problems 331

7. Show that the second moment of f 1 defined in Problem 6 (b) is given by

 
(x ∗ x2 )
f 1∗ f 1 = α12 x1∗ x2 − 1∗ x1∗ x1
(x1 x1 )
 
(x2∗ x1 ) ∗
× x1 x2∗ − x1 x1
(x1∗ x1 )
= α12 [(x1∗ x1 )(x1∗ x2 ) − (x1∗ x2 )(x2∗ x1 )]

8. Substitute f 1∗ f 1 obtained in Problem 7 into δ1∗ δ1 obtained in Problem 6 (b) to show that

εα1 [(x1∗ x1 )(x2∗ x2 ) − (x2∗ x1 )(x1∗ x2 )]

δ1∗ δ1 =
2[(x1∗ x1 ) + 1/G 1 ]

Note that to obtain δ2∗ δ2 , merely replace x2 with x3 , α1 by α2 , and G 1 by G 2 in the previous
expression.
9. From Figure 7-11 it may be seen that the inputs to u 3 (u 22 of the figure) are y1 and y2 . Conse-
quently, replace x1 by y2 , x2 by y1 , and G 1 by G 3 in the expression obtained in Problem 8. This
replacement requires the second moments y1∗ y1 and y2∗ y2 and y2∗ y1 to be computed. Show that

(x1∗ x3 )(x3∗ x1 )
y1∗ y1 = |x3 − u 3 x1 |2 = (x3∗ x3 ) −
(x1∗ x1 )
and
 2
 (x1∗ x2 ) 

y2∗ y2 = |x2 − u 1 x1 | = x2 − ∗ 2
x1 
 (x1 x1 ) 
(x1∗ x2 )(x2∗ x1 )
= (x2∗ x2 ) −
(x1∗ x1 )
Furthermore
 ∗  
(x1∗ x2 ) (x ∗ x3 )
y2∗ y1 = x2 − ∗
x1 x3 − 1∗ x1
(x1 x1 ) (x1 x1 )
∗ ∗
(x x1 )(x x3 )
= (x2∗ x3 ) − 2 ∗ 1
(x1 x1 )

10. The computation of Nu found in Problem 3 requires the second moment δ2∗ δ1 . Corresponding
to the expression for δ1 (t) in Problem 5 we may write
 t  
1
δ2 (t) = f 2 (τ ) exp − α2 x1∗ x1 + (t − τ )dτ
0 G2
where
 
(x1∗ x3 ) ∗
f 2 (t) = α2 x1∗ x3 − x1 x1
(x1∗ x1 )
With the previous results, show that
 t t
δ2∗ δ1 = f 2∗ (τ ) f 1 (u)
0
0
   
∗ 1 ∗ 1
· exp −α2 x1 x1 + (t − τ ) − α1 x1 x1 + (t − u)dτ du
G2 G1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 332

332 CHAPTER 7 Cascade Preprocessors

Now assuming that

 t
f 2∗ (τ ) f 1 (u) du = f 2∗ f 1
0

show that
ε f 2∗ f 1
δ2∗ δ1 =
[α2 (x1∗ x1 ) + 1/G 2 ] + α1 [(x1∗ x1 ) + 1/G 1 ]
11. Evaluate f 2∗ f 1 by retaining only terms of the form xm∗ xn in the expansion
  
(x ∗ x1 ) (x ∗ x2 )
f 2∗ f 1 = α2 α1 x1 x3∗ − 3∗ x1∗ x1 x1∗ x2 − 1∗ x1∗ x1
(x1 x1 ) (x1 x1 )

show that the resulting expression for f 2∗ f 1 is given by

f 2∗ f 1 = α2 α1 [(x1∗ x1 )(x3∗ x2 ) − (x3∗ x1 )(x1∗ x2 )]

so that
εα1 α2 [(x1∗ x1 )(x3∗ x2 ) − (x3∗ x1 )(x1∗ x2 )]
δ2∗ δ1 =
α2 [(x1∗ x1 ) + 1/G 2 ] + α1 [(x1∗ x1 ) + 1/G 1 ]
12. Substitute the results obtained in Problem 11 into the expression for Nu , assuming that α1 = α2
and neglecting all 1/G terms, show that

Nu = {[(x1∗ x1 )(x2∗ x2 ) − (x1∗ x2 )(x2∗ x1 )][(x1∗ x1 )(x3∗ x3 ) − (x3∗ x1 )(x1∗ x2 )]

 
∗ ∗ ∗ ∗ α3 ε α1 ε
− |(x1 x1 )(x2 x3 ) − (x2 x1 )(x1 x2 )| } 2
+
2(x1∗ x1 )2 2(x1∗ x1 )(x2∗ x2 ) − (x2∗ x1 )(x1∗ x2 )
Note that the portion of Nu proportional to α3 is the component due to fluctuations in u 3 . The
remaining part of Nu that is proportional to α1 is the contribution due to fluctuations in both
u 1 and u 2 . If τ1 , τ2 , and τ3 are adjusted so that all three circuits have the same effective time
constant,
(x1∗ x1 )2
α1 = α2 = α3
(x1 x1 )(x2 x2 ) − (x2∗ x1 )(x1∗ x2 )
∗ ∗

then the previous expression for Nu shows that the control loop noise contribution due to
fluctuations in u 3 is just equal to the sum of the contributions from fluctuations in u 1 and u 2 .

7.7 REFERENCES
[1] L. Hasdorff, Gradient Optimization and Nonlinear Control, New York, Wiley, 1976, Ch. 3.
[2] W. D. White, “Accelerated Convergence Techniques,” Proceedings of the 1974 Adaptive
Antenna Systems Workshop, March 11–13, Vol. 1, Naval Research Laboratory, Washington,
DC, pp. 171–215.
[3] W. D. White, “Cascade Preprocessors for Adaptive Antennas,” IEEE Trans. Antennas Propag.,
Vol. AP-24, No. 5, September 1976, pp. 670–684.
[4] W. D. White, “Adaptive Cascade Networks for Deep Nulling,” IEEE Trans. Antennas Propag.,
Vol. AP-26, No. 3, May 1978, pp. 396–402.
[5] L. E. Brennan, J. D. Mallett, and I. S. Reed, “Convergence Rate in Adaptive Arrays,” Tech-
nology Service Corporation Report No. TSC-PD-A177-2, July 15, 1977.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 333

7.7 References 333

[6] R. C. Davis, “Convergence Rate in Adaptive Arrays,” Technology Service Corporation Report
No. TSC-PD-A177-3, October 18, 1977.
[7] L. E. Brennan and I. S. Reed, “Convergence Rate in Adaptive Arrays,” Technology Service
Corporation Report No. TSC-PD-177-4, January 13, 1978.
[8] C. C. Ko, “On the Performance of Adaptive Gram–Schmidt Algorithm for Interference Can-
celling Arrays,” IEEE Transactions on Antennas and Propagation, Vol. 39, No. 4, 1991,
pp. 505–511.
[9] H. Liu, A. Ghafoor, and P. H. Stockmann, “Application of Gram–Schmidt Algorithm to Fully
Adaptive Arrays,” IEEE Transactions on Aerospace and Electronic Systems, Vol. 28, No. 2,
1992, pp. 324–334.
[10] R. W. Jenkins and K. W. Moreland, “A Comparison of the Eigenvector Weighting and Gram–
Schmidt Adaptive Antenna Techniques,” IEEE Transactions on Aerospace and Electronic
Systems, Vol. 29, No. 2, 1993, pp. 568–575.
[11] C. C. Ko, “Simplified Gram–Schmidt Preprocessor for Broadband Tapped Delay-Line
Adaptive Array,” IEE Proceedings G Circuits, Devices and Systems, Vol. 136, No. 3, 1989,
pp. 141–149.
[12] W. F. Gabriel, “Adaptive Array Constraint Optimization,” Program and Digest 1972 G-AP
International Symposium, December 11–14, 1972, pp. 4–7.
[13] J. C. Nolen, “Synthesis of Multiple Beam Networks for Arbitrary Illuminations,” Bendix
Corporation, Radio Division, Baltimore, MD, April 21, 1960.
[14] N.J.G. Fonseca, “Printed S-Band 4 4 Nolen Matrix for Multiple Beam Antenna Applications,”
IEEE Transactions on Antennas and Propagation, Vol. 57, No. 6, June 2009, pp. 1673–1678.
[15] P. R. Halmos, Finite Dimensional Vector Spaces, Princeton, NJ, Princeton University Press,
1948, p. 98.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:1 334
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 335

CHAPTER

Random Search Algorithms

8
' $
Chapter Outline
8.1 Linear Random Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
8.2 Accelerated Random Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
8.3 Guided Accelerated Random Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
8.4 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
8.5 Comparison of Random Search Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
8.6 Summary and Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
8.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
8.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
& %

Gradient-based algorithms assume the performance measures are either quadratic or uni-
modal. For some classes of problems [1–3], the mathematical relation of the variable
parameters to the performance measure is either unknown or is too complex to be useful.
In yet other problems, constraints are placed on the variable parameters of the adap-
tive controller with the result that the performance surface is no longer unimodal. When
the performance surface of interest is multimodal and contains saddlepoints, then any
gradient-based algorithms finds only a local minimum. A random search algorithm has
the ability to jump out of one valley with a local minimum into another valley with a
potentially lower local minimum. Random algorithms have global search capabilities that
work for any computable performance measure [4–14]. Random search algorithms tend to
have slow convergence, especially in unimodal applications. They do, however, have the
advantages of being simple to implement in logical form, of requiring little computation,
of being insensitive to discontinuities, and of exhibiting a high degree of efficiency where
little is known about the performance surface.
Systematic searches exhaustively survey the parameter space within specified bounds,
making them capable of finding the global extremum of a multimodal performance mea-
sure. As a practical matter, however, this type of search is very time-consuming and incurs
a high search loss, since most of the search period occurs in regions of poor performance.
Random searches are classified as either guided or unguided, depending on whether
information is retained whenever the outcome of a trial step is learned. Furthermore, both
the guided and unguided varieties of random search are given accelerated convergence
by increasing the adopted step size in a successful search direction. Four representative
examples of random search algorithms used for adaptive array applications are considered

335
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 336

336 CHAPTER 8 Random Search Algorithms

in this chapter: linear random search (LRS), accelerated random search (ARS), guided
accelerated random search (GARS), and genetic algorithm (GA).

8.1 LINEAR RANDOM SEARCH

The LRS algorithm discussed by Widrow and McCool [15] attempts to “learn” when a trial
change in the adaptive weight vector is rejected. After adding a random change, wk , to
the weight vector at the beginning of each iteration, the corresponding change in the per-
formance measured is observed. A permanent weight vector change is proportional to the
estimated performance measure change times the tentative weight vector change, that is,
ˆ
wk+1 = wk + μs {[w ˆ
k ] − [wk + wk ]}wk (8.1)

where [·] denotes the selected array performance measure, and μs is a step size constant.
The random vector wk has components generated from a normal probability density
function with zero mean and variance σ 2 . The constants μs and σ 2 are selected to ensure
a fast, stable algorithm convergence. The LRS algorithm is “linear” because the weight
change is proportional to the change in the performance measure.
The true change in the performance measure resulting from adding wk to wk is

()k = [wk + wk ] − [wk ] (8.2)

When the performance measure value is estimated, then the corresponding estimated
change in the performance measure is given by
ˆ =  ˆ ˆ
()k [wk + wk ] − [wk] (8.3)

The error in the estimated change of the performance measure is then

 ˆ
γk = ()k − ()k (8.4)

and the variance in this error is given by

ˆ ]
var[γk ] = var[() k
ˆ ˆ
= var{[wk + wk ]} + var{[w k ]} (8.5)

To determine the variance of the estimate [wˆ

k ], it is necessary to consider a specific
performance measure and the estimate of that measure to be employed.
The mean square error (MSE) is the performance measure of interest so that [w] =
ξ [w]. The estimate of the MSE comes from averaging K independent samples as given by
(4.98) of Section 4.3.2. With this choice of performance measure and its corresponding
estimate, it follows that
2 2
var[γk ] = {ξ [wk + wk ] + ξ 2 [wk ]} (8.6)
K
where K = the number of independent samples on which the estimate ξ̂ [w] is based. In
the steady state when the weight adjustment process operates near the minimum point of
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 337

8.1 Linear Random Search 337

the performance surface, then (8.6) is very nearly

4 2
var[γk ] ≈ ξ (8.7)
K min
The tentative random changes in the weight vector produced by the LRS algorithm
result in MSE perturbations. K data samples each iteration result in ξ̂ [wk + wk ]. Parallel
processors (with one processor using wk and the other processor using wk +wk ) decrease
the convergence time, because the samples required to obtain ξ̂ [wk ] and ξ̂ [wk + wk ]
are collected simultaneously. The value of the adaptive weight vector at the beginning of
the next iteration is selected after the two ξ̂ estimates are obtained. For any given iteration
the average excess MSE resulting from the perturbation in the weight vector is given by
 
ξ [wk ] + ξ [wk + wk ] 1
E ξ [wk ] − = E{ξ [wk ] − ξ [wk + wk ]} (8.8)
2 2
The random weight perturbation vector wk has zero mean and is uncorrelated with wk .
Let wk denote the random weight perturbation vector in normal coordinates (in which
the covariance matrix is diagonal so Rx x = ), then cov[wk ] = cov[wk ] = σ 2 I.
Consequently, the average excess MSE can also be expressed as
1   1   1 2
E wkT Rx x wk = E wT 
k wk = σ tr(Rx x ) (8.9)
2 2 2
Define the perturbation P as the ratio of the average excess MSE (resulting from the
random perturbations in the weight vector) to the minimum MSE, then
σ 2 tr(Rx x )
P= (8.10)
2ξmin

8.1.1 LRS Algorithm Stability

The weight adjustment equation (8.1) are rewritten in terms of the definitions given by
(8.2), (8.3), and (8.4) for the MSE performance measure as
wk+1 = wk + μs {−(ξ )k + γk }wk (8.11)

Recalling that vk = wk − wopt , then we can rewrite (8.11) as
vk+1 = vk + μs {−(ξ )k + γk }wk (8.12)
Specify that σ 2 be small so that wk is always small, and then
(ξ )k = wkT ∇k (8.13)
where ∇k is the gradient of the performance surface evaluated at wk . Since ∇k = 2Rxx vk ,
(8.13) is written as
(ξ )k = 2wkT Rxx vk (8.14)
Consequently, (8.12) is rewritten as
 
vk+1 = vk + μs wk −2wkT Rxx vk + γk

= I − 2μs wk wkT Rxx vk + μs γk wk (8.15)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 338

338 CHAPTER 8 Random Search Algorithms

The alternate (but equivalent) form of (8.1) represented by (8.15) is more useful for analysis
even though the algorithm is implemented in the form suggested by (8.1). Equation (8.15)
emphasizes that the adaptive weight vector is regarded as the solution of a first-order linear
vector difference equation having a randomly varying coefficient I − 2μs wk wkT Rxx
and a random driving function μs γk wk .
Premultiplying both sides of (8.15) by the transformation matrix Q of Section 4.1.3
converts the foregoing linear vector difference equation into normal coordinates

vk+1 = I − 2μs wk wT 
k  vk + μs γk wk

(8.16)

Although (8.16) is somewhat simpler than (8.15), the matrix coefficient of vk still contains
cross-coupling and randomness, thereby rendering (8.16) a difficult equation to solve.
Stability conditions for the LRS algorithm are obtained without an explicit solution to
(8.16) by considering the behavior of the adaptive weight vector mean.
Taking the expected value of both sides of (8.16) and recognizing that wk is a random
vector that is uncorrelated with γk and vk , we find that
 
E{vk+1 } = E I − 2μs wk wT  
k  vk + μs E{γk wk }
  
= I − 2μs E wk wT 
k  E{vk } + 0
= (I − 2μs σ 2 )E{vk } (8.17)

The solution to (8.17) is given by [15]

E{vk } = (I − 2μs σ 2 )k v0 (8.18)

For the initial conditions v0 , (8.18) gives the expected value of the weight vector’s transient
response. If (8.18) is stable, then the mean of vk must converge. The stability condition
for (8.18) is
1
> μs σ 2 > 0 (8.19)
λmax
If we choose μs σ 2 to satisfy (8.19), it then follows that

lim E{vk } = 0 (8.20)

k→∞

Since the foregoing transient behavior is analogous to that of the method of steepest
descent discussed in Section 4.1.2, it is argued by analogy that the time constant of the
pth mode of the expected value of the weight vector is given by
1
τp = (8.21)
2μs σ 2 λ p
Furthermore, the time constant of the pth mode of the MSE learning curve is one-half the
aforementioned value so that
1
τ pmse = (8.22)
4μs σ 2 λ p
Satisfying the stability condition (8.19) implies only that the mean of the adaptive
weight vector will converge according to (8.20); variations in the weight vector about
the mean value may be quite severe, however. It is therefore of interest to obtain an
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 339

8.1 Linear Random Search 339

indication of the severity of variations in the weight vector when using the LRS algorithm
by deriving an expression for the covariance of the weight vector. In obtaining such an
expression, it will simply be assumed that the weight vector covariance is bounded and
that the weight vector behaves as a stationary stochastic process after initial transients
have died out. Assuming that a bounded steady-state covariance matrix exists, we may
calculate an expression for such a covariance by multiplying both sides of (8.16) by their
respective transposes to obtain
 
vk+1 vT  T  T
k+1 = I − 2μs wk wk  vk vk I − 2μs wk wk
 T

+μ2s γk2 wk wTk


+ I − 2μs wk wT 
k  vk μs γk wk
T

+μs γk wk vT
k I − 2μs wk wk
 T
(8.23)

Now we take expected values of both sides of (8.23) recalling that γk and wk are zero-
mean uncorrelated stationary processes so that
    
E vk+1 vT
k+1 = E I − 2μs wk wT  T 
k  vk vk I − 2μs wk wk
T
   
+μ2s E γk2 E wk wT k +0 (8.24)

Since var[γk ] ∼
= (4/K )ξmin
2
and cov[wk ] = σ 2 I, it follows that (8.24) is expressed
as
  
E vk+1 vT
k+1 = E I − 2μs wk wT
k 
  4 2 2
·vk vT 
k I − 2μs wk wk
T
+ μ2s ξmin σ I (8.25)
K
In the steady state vk is also a zero-mean stationary random process that is uncorrelated
with wk so (8.25) is written as
     T  
E vk+1 vT
k+1 = E I − 2μs wk wT
k  E vk vk I − 2μs wk wT
k (8.26)
4
+μ2s ξmin 2
σ 2I
K
Consequently, the steady-state covariance of the adaptive weight vector is
   4 2 2
cov[vk ] = E I − 2μs wk wT 
k  cov[vk ] I − 2μs wk wk
 T
+ μ2s ξmin σ I
  N
= cov[vk ] − 2μs E wk wTk cov[vk ]

 
− 2μs cov [vk ]E wk wT k
  4 2 2
+ 4μs E wk wk cov[vk ]wk wT
2  T 
k + μ2s ξmin σ I
N
  
= cov[vk ] − 2μs σ cov[vk ] − 2μs σ cov[vk ]
2 2

  4 2 2
+ 4μ2s E wk wT 
k cov[vk ]wk wk
 T
+ μ2s ξmin σ I (8.27)
K
Equation (8.27) is not easily solved for the covariance of vk , because the matrices
appearing in the equation cannot be factored. It is likely (although not proven) that the
steady-state covariance matrix of vk is diagonal. The results obtained with such a simplify-
ing assumption do, however, indicate that there is some merit in the plausibility argument.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 340

340 CHAPTER 8 Random Search Algorithms

The random driving function appearing in (8.16) consists of components that are
uncorrelated with each other and uncorrelated over time. Furthermore, the random coeffi-
cient I − 2μs wk wTk  is diagonal on the average (though generally not for every value
of k) and uncorrelated both with vk and with itself over time. Thus, it is plausible that the
covariance of vk is a diagonal matrix.
Assuming that cov[vk ] is in fact diagonal, then (8.27) can immediately be rewritten
by merely rearranging terms as
  4 2 2
4μs σ 2 cov[vk ] − 4μ2s E wk wT   T
k cov[vk ]wk wk = μ2s ξmin σ I (8.28)
K
Of greatest interest is the case for which adaptation is slow, in which
μs σ 2   I (8.29)
Furthermore, it may be noted that
 
T ∼
μ2s E wk wT  
k cov[vk ]wk wk = (μs σ 2 )2 cov[vk ] (8.30)
and from (8.29) it follows that
(μs σ 2 )2 cov[vk ]  μs σ 2 cov[vk ] (8.31)
With the result of (8.31) it follows that the term −4μ2s E{ } appearing in (8.28) is neglected.
Consequently, (8.28) is rewritten as
μs 2 −1
cov[vk ] = ξ  (8.32)
K min
The steady-state covariance matrix of vk given by (8.32) is based on a plausible assumption,
but experience indicates that the predicted misadjustment obtained using this quantity
generally yields accurate results [15].
The misadjustment experienced using the LRS algorithm is obtained by considering
the average excess MSE due to noise in the weight vector, which is given by
N
 
E vT 
k vk = λ p E{(v pk )2 } (8.33)
p=1

where N is the number of eigenvalues of . If we use (8.32), it follows that for the LRS
algorithm
N
  μs 2 1 N μs 2
E vT 
k vk = λp ξ = ξ (8.34)
p=1
K min λ p K min

Since the misadjustment M is defined to be the average excess MSE divided by the
minimum MSE
 T 

 E vk vk
M= (8.35)
ξmin
It follows that for the LRS algorithm
N μs
M= ξmin (8.36)
K
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 341

8.2 Accelerated Random Search 341

The result given by (8.36) can be expressed in terms of the perturbation of the LRS process
as
N μs σ 2 tr(Rxx ) N 2 μs σ 2 λav
M= = (8.37)
2KP 2KP
Now recall that the time constant of the pth mode of the learning curve for the LRS
algorithm (in terms of the number of iterations required) is given by (8.21). Since one
iteration of the weight vector requires two estimates of ξ̂ , 2K samples of data are used per
iteration, and the learning curve time constant expressed in terms of the number of data
samples is
 K
T pmse = 2K τ pmse = (8.38)
2μs σ 2 λ p
From (8.38) it follows immediately that
K 1
λp = (8.39)
2μs σ 2 T pmse
and
K 1
λav = (8.40)
2μs σ 2 T pmse av

Substituting (8.40) into (8.37) then yields

N2 1
M= (8.41)
4P T pmse av

Since the total misadjustment consists of a stochastic component M and a deterministic

component P from (8.41), we may write
N2 1
Mtot = +P (8.42)
4P T pmse av

If the deterministic component of the total misadjustment is optimally chosen, then both
M and P are equal and P is one-half the total misadjustment so that
1/2
N2 1 1
(Mtot )min = =N (8.43)
2Popt T pmse av T pmse av

It is informative to compare this result with the corresponding result (4.83) for the LMS
algorithm.

8.2 ACCELERATED RANDOM SEARCH

Suppose we want to minimize the performance measure, (w), over some range of values
for the complex weight vector w. Let the adaptive processor change the complex weights
using a simplified version of accelerated random search reported by Baird and Rassweiler
[16] as follows:

w(k + 1) = w(k) + μs (k)[w(k)] (8.44)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 342

342 CHAPTER 8 Random Search Algorithms

where μs (k) is the step size initially set at μs (0) = μ0 , and w(k) is a random vector
whose components are given by

w i (k) = cos θi + j sin θi , i = 1, 2, . . . , m (8.45)

where θi is a uniformly distributed random angle on the interval {0, 2π} so that
|w i (k)| = 1 and w(k) controls the direction of the weight vector change while μs
controls the magnitude.
Initially the weight vector, w(0), and the corresponding performance measure, [w(0)],
ˆ
(or an estimate thereof [w(0)]) is evaluated. The weight vector changes in accordance
with (8.44) using μs (0) = μ0 . The performance index [w(1)] is evaluated and com-
pared with [w(0)]. If this comparison indicates improved performance, then the weight
direction change vector w is retained, and the step size μs is doubled (resulting in
“accelerated” convergence). If, however, the resulting performance is not improved, then
the previous value of w is retained as the starting point, a new value of w is selected,
and μs is reset to μ0 . As a consequence of always returning to the previous value of w as
the starting point for a new weight perturbation in the event the performance measure is
not improved, the ARS approach is inherently stable, and stability considerations do not
play a role in step size selection. A block diagram of this simplified version of accelerated
random search is given in Figure 8.1.
Consider a single component of the complex weight vector for which vi = w i − w opti .
If vi (k) lies within μ0 /2 of w opt , then any further perturbation in that component of the
weight vector of step size μ0 will result in vi (k + 1) ≥ vi (k) as shown in Figure 8.2.
Consequently, if all components of the weight vector lie within or on the best performance
surface contour contained within the circle of radius μ0 /2 about w opt , then no further
improvement in the performance measure can possibly occur using step size μ0 . The
condition where all weight vector components lie within this best performance surface
contour therefore represents a lower limit on the possible improvement that is achieved
with the ARS procedure, and this is the ultimate condition to which the weight vector is
driven in the steady state.

FIGURE 8-1
Block diagram for
ARS algorithm. Calculate
wi(k+1) = wi(k) + ms(k +1) Δwi(k+1) Array weighting
i = 1, 2, …, N
Array output

Δwi(k +1) = cos qi + j sin qi

i = 1, 2, …, N
Evaluate b [Δw(k+1)]
ms = m0

Reset w = w(k)
≥0 b [w(k+1)] − <0 Δw(k+1) = Δw(k)
choose qi, i = 1, 2, …, N
b [w(k)] ms(k+1) = 2ms(k)
qi uniform on {0, 2p}
Random Deterministic
phase phase
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 343

8.2 Accelerated Random Search 343

Im(vi) vi = wi − wopt FIGURE 8-2

Complex
perturbation of step
size μ0 for a single
weight vector
component.

0
m
3
Re(vi)
wopt m
0
2
m0

To simplify the development, assume it is equally likely that the weight vector com-
ponent w i (k) lies anywhere within or on the circle for which vi = μ0 /2, it then follows
that the steady-state expected value of vi is vss (k) = 0, and the average excess MSE in
this steady-state condition is (adopting  = ξ and noting that E{|vi |2 } = μ20 /8)
μ20
E{vT (k)v (k)} = ξmin tr(Rxx ) (8.46)
8
The average misadjustment for this steady-state condition is therefore
μ20
Mav = tr(Rxx ) (8.47)
8
On each successive iteration the weight vector components are perturbed by μ0 from
their steady-state values. Assume that the perturbation is taken from vss = ρ as shown in
Figure 8.3, then E{v p } = μ0 , E{|v p |2 } = 9μ20 /8, and the average total misadjustment for
the random search perturbation is therefore
9μ20
Mtot = tr(Rxx ) (8.48)
8
From the foregoing discussion, it follows that in the steady state, as long as a correct
decision is made concerning ξ [w(k + 1)] − ξ [w(k)], the average total misadjustment is
given by (8.48).
ˆ
In practice, the ARS algorithm examines the statistic [w(k ˆ
+ 1)] − [w(k)] instead
of [w(k + 1)] − [w(k)], and the measured statistic contains noise that may yield a
misleading indication of the performance measure difference. The performance measure
difference due to the weight vector perturbation must be significantly larger than the
standard deviation of the error in the estimated change in the performance measure: this
is done by selecting  = [w(k + 1)] − [w(k)] > σγ where σγ2 is given by (8.7).
Selecting K and μs so that  > σγ results in the average steady-state misadjustment
approximated by (8.48). Furthermore, the performance measure difference due to the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 344

344 CHAPTER 8 Random Search Algorithms

FIGURE 8-3
vp = √ r 2 + m 20 − 2rm0 cos q vmax = m0 + r
Weight vector
component
perturbation

m0
resulting from step
size μ0 starting from
vss = μ0 /3.

m0

r
wopt 2

vmin = m0 − r

weight vector perturbation should be less than E{[w(k)]}. Therefore, for  selected to
be ξ , in the steady state the constants K and μs should be selected to satisfy
9 2 2ξmin
ξmin > μ tr(Rxx ) > √ (8.49)
8 0 K
Even with K and μs selected to satisfy  > σγ , it is still possible for noise present
in the measurements of system performance to produce deceptively good results for any
one experiment. Such spurious results will on the average be corrected on successive trials.

8.3 GUIDED ACCELERATED RANDOM SEARCH

The GARS introduced by Barron [17–19] consists of two phases: a random phase and
a deterministic phase. Control of the parameter space search passes back and forth be-
tween these two phases as the search finds parameter space regions having better or worse
performance. In the initial random phase (which is an information-gathering phase), the
adaptive weights are randomly perturbed according to a multivariate probability density
function (PDF). Once a direction is found in which performance improvement occurs, the
deterministic phase is entered, and the information acquired from the initial random phase
is exploited as larger step sizes are taken in the direction of improved performance. When-
ever an accelerated step in the deterministic phase produces an unsuccessful outcome, the
random phase is reentered, but now the PDF governing the random search assigns smaller
excursions to the parameter values than occurred in the initial random phase.
A block diagram representation of a simplified version of GARS is shown in Figure 8.4.
Starting with an initial weight vector w0 a corresponding performance measure [w0 ] is
evaluated. At all times the minimum value of the performance measure attained is stored
and denoted by ∗ , so initially ∗ = [w0 ]. The GARS algorithm begins in its random
phase by generating a random weight vector perturbation w, each of whose elements
are drawn from a normal probability density function having zero mean and a variance
σ 2 . The variance is selected according to

σ 2 = K 1 + K 2 ∗ (8.50)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 345

8.3 Guided Accelerated Random Search 345

FIGURE 8-4
Block diagram for
GARS algorithm.
Calculate
w(k+1) = w(k) + Δw(k) Array weighting
Array output
Start

Generate w0

Evaluate b [w(k +1)]

Evaluate b [w0]
b ∗ = b [w0]

≥0 b [w(k+1)] − <0
b [w(k)]

Random phase: Deterministic phase:

retain old w(k ) retain old Δw(k)

s 2 = K1 + K2 b ∗ Δw(k ) = 2{old Δw(k)]

Generate random components: b ∗ = b [w(k+1)]

Δwi = n{0, s}
i = 1, 2, …, m

where K 1 and K 2 are design constants for the GARS algorithm selected so the step size
is small enough when the optimum performance is realized yet big enough to gain useful
performance surface information when the trial weight vector is far from optimum.
The next trial adaptive weight vector is then computed using
w(k + 1) = w(k) + w(k) (8.51)
and the corresponding performance measure, [w(k + 1)], is evaluated. If no improve-
ment in the performance measure is realized, the algorithm remains in the random phase
for the next trial weight vector, returning to the previous value of w as the starting point
for the next weight perturbation. Once a direction in which to move for improved perfor-
mance is determined, the deterministic phase of the algorithm is entered, and convergence
is accelerated by continuing to travel in the direction of improved performance with twice
the previous step size. The weight vector step w is continually doubled as long as per-
formance measure improvements are realized. Once the performance measure begins to
degrade, the search is returned to the random phase where the adaptive weight vector
perturbations w are considerably smaller than before due to the smaller value of σ used
in generating new search directions.
From the foregoing description of the simplified version of GARS, it is seen that the
principal difference between GARS and ARS lies in how the random phase of the search
is conducted. Not only is the search direction random (as it was before), but the step size
is also random and governed by the parameter σ whose assigned value depends on the
minimum value that the selected performance measure has attained. As a result, the search
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 346

346 CHAPTER 8 Random Search Algorithms

step reduces as performance measure improvements are realized. The observations made
in the previous section for the ARS algorithm with minimum step size μ0 now apply in a
statistical sense to the GARS algorithm. When  is selected to be ξ the condition expressed
by (8.49) should also be satisfied where now the expected change in performance measure
due to weight perturbations is given by
E{ξ } = σ 2 tr(Rxx ) (8.52)
Noise in the measurements of the performance measure can produce deceptively
good results for any one experiment, thereby creating the risk of spurious measurements
locking the search to a false solution. Local minima are avoided in the random phase of the
algorithm by periodically reexamining the performance measure of the best weight vector
found so far. Furthermore, the algorithm handles nonstationary operating conditions by
periodically using large step sizes and conducting the exploration perturbations uniformly
throughout the parameter space.
The general GARS algorithm incorporates features that are not simulated here. The
most important such feature is the provision for a long-term memory due to employing a
nonuniform multivariate probability distribution function [20] to generate search directions
and thereby to guide the search to increase the probability that future trials will yield better
performance scores than past trials. This multivariate PDF is shaped according to the
results of a series of initial trials conducted during the opening stage of the search when
no preference is given to any search direction. During the middle stage of the search,
the multivariate PDF formed during the opening stage guides the search by generating
new search directions. In the final search stage, the dimensionality of the parameter space
search is reduced by converting from a simultaneous search involving all the parameters to
a nearly sequential search involving only a small fraction of the parameters at any step. This
selected fraction of the parameters to search is chosen randomly for each new iteration.

8.4 GENETIC ALGORITHM

A GA uses genetic and natural selection rules to maximize the fitness (i.e., minimize the
cost) of a fitness (cost) function. The GA has proven very useful in complex antenna design
and adaptive nulling. Some of the advantages of a GA for adaptive nulling use include the
following:
• It optimizes with continuous or discrete variables. Thus, the GA interfaces with digitally
controlled hardware like phase shifters and attenuators without having to quantize
variables.
• It has constraints inherent in the formulation for limiting adaptive weights to a reason-
able range.
• It does not require derivative information.
• It deals with a large number of variables, so it works with large arrays.
• It is well suited for parallel computers.
• It is able to avoid getting stuck in the valley of a local minimum.
A diagram of a GA is shown in Figure 8.5, and its operation is explained in the following
paragraphs. Details for a practical implementation of a genetic algorithm are found in
[21,22].
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 347

8.4 Genetic Algorithm 347

1. Form population and evaluate cost FIGURE 8-5 GA

flowchart.
chrom1 cost1
chrom2 cost2
⇒ antenna ⇒

chromNpop costNpop

2. Natural selection
chrom1 cost1
⇒
chromNmate costNmate

3. Mating
chroma
chromb ⇒ chromc
chrom1
chromc chromNmate +1
⇒ ⇒ binary mask ⇒
chromd chromNmate +2
chromNmate
chrome ⇒ chrome
chromf

4. Mutation
chrom1 100011 10001 100011 10001
111000 01010 101000 01010
001110 01100 001110 01100
chromNmate
⇒ 100111 00111 ⇒ 100111 01111
chromNmate +1

chromNpop 000011 11100 010011 11100

No
Converged?

Yes

1. Form population and evaluate cost. A GA starts with a random population matrix
with N pop rows . Each row is a chromosome and contains the adaptive weights for all the
array elements. Since the adapted weights are normally digital, the population matrix is
binary. If the adaptive weights have Nb bits and there are Na adaptive elements, then each
chromosome contains Nb × Na bits. Thus, the population matrix is given by
⎡ ⎤
⎢ b ⎥ ⎡ chrom ⎤
⎢ 1,1,1 b1,1,2 · · · b1,1,Nbits · · · b1,Na ,1 b1,Na ,2 · · · b1,Na ,Nbits ⎥ 1
⎢ ⎥ ⎢
⎢ b2,1,1 b2,1,2 b2,Na ,1 b2,Na ,2 ⎥ ⎢ chrom2 ⎥ ⎥
P=⎢
⎢ .. .. .. .. .. .. ⎥=⎢
⎥ ⎣ .. ⎥
⎢ . . . . . . ⎥ . ⎦
⎢ ⎥
⎣ b Npop ,1,1 b Npop ,1,2 · · · b Npop ,Nbits · · · b Npop ,Na ,1 b Npop ,Na ,2 · · · b Npop ,Na ,Nbits ⎦ chrom Npop
     
Element 1 Element Na
(8.53)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 348

348 CHAPTER 8 Random Search Algorithms

A cost function generates the cost or output from the chromosomes in the population
matrix and places them in a corresponding cost vector. Each row is sent to the cost function
for evaluation, so that cost, cn , of chromn is given by
cn = f (chromn ) (8.54)
Then, all the costs are placed in a vector.
C = [ c1 c2 · · · c N pop ]T (8.55)
In the case of an adaptive array, the cost is the array performance measure.
2. Natural selection. Once all of the costs are determined, then natural selection occurs
each generation (iteration) and discards unacceptable chromosomes that have a high cost.
Of the N pop chromosomes in a generation, the Nmate chromosomes with the lowest cost
survive and form the mating pool, whereas the bottom Npop − Nmate are discarded to make
room for the new offspring.
3. Mating. Mating is the process of selecting two chromosomes from the mating pool
of Nmate chromosomes to produce two new offspring. Chromosomes from the mating
population are selected and paired to create Npop −Nmate offspring that replace the discarded
chromosomes. Tournament selection is a popular way to select chromosomes for mating.
It randomly selects a subset of chromosomes from the mating pool, and the chromosome
with the lowest cost in this subset becomes a parent. The tournament repeats Npop − Nmate
times.
Mating combines the characteristics of two chromosomes to form two new chro-
mosomes that replace two chromosomes discarded in the natural selection step. Uniform
crossover is a general procedure that selects variables from each parent chromosome based
on a mask and then places them in a new offspring chromosome. First, a random binary
mask is created. A 1/0 in the mask column means the offspring receives the variable value
from chromm/n . If it has a 0/1, then the offspring receives the variable value in chromn/m .

chromm = bm,1 bm,2 bm,3 bm,4 bm,5 bm,6 bm,7 bm,8

chromn = bn,1 bn,2 bn,3 bn,4 bn,5 bn,6 bn,7 bn,8
mask = 0 1 1 0 0 1 0 1 (8.56)
offspring = bn,1 bm,2 bm,3 bn,4 bn,5 bm,6 bn,7 bm,8
offspring = bm,1 bn,2 bn,3 bm,4 bm,5 bn,6 bm,7 bn,8
4. Mutation. The final step induces random mutations to alter a certain percentage of the
variables in the list of chromosomes. A mutation changes a “1” to a “0” or vice versa in
the Npop × Nbits total number of bits in the population matrix. The best chromosome is
usually not mutated. Mutation is specified as a percentage of the bits in the population
matrix.
Particle swarm optimization (PSO) is another possible adaptive algorithm that mini-
mizes the total output power [23]. It starts with the same population matrix as the GA. The
rows in the matrix are called particles instead of chromosomes. The particles move on the
cost surface with a velocity. The particles update their velocities and positions based on
the local and global best solutions.
   
vm,n
new
= vm,n
old
+ 1 × r1 × pm,nlocal best
− pm,n
old
+ 2 × r2 × pm,n
global best
− pm,n
old
(8.57)
new
pm,n = pm,n
old
+ vm,n
new
(8.58)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 349

8.4 Genetic Algorithm 349

where
vm,n = particle velocity
pm,n = particle variables
r1 , r2 = independent uniform random numbers
1 = cognitive parameter

2 = social parameter
local best
pm,n = best local solution
global best
pm,n = best global solution

PSO updates the velocity vector for each particle and then adds that velocity to the particle
position. Velocity updates depend on the estimate of the global solution found thus far
and the best local solution in the present population. If the best local solution has a cost
less than the cost of the estimate of the global solution, then the best local solution replaces
the global solution estimate.
As an example, consider a 40-element array along the x-axis with elements spaced
d = λ/2 apart and a 30 dB Chebyshev amplitude taper (an ) [24]. Elements have a sin φ
element pattern and six-bit phase shifters. Assume there are two interference sources at
φ = 43.9◦ and 51.7◦ that are 60 dB stronger than the desired signal power in the main
beam. The cost function assumes the phase shifts are antisymmetric about the center of
the array.
 2   
 20
 2π 
cost = 20 log10 si sin φi  an cos (n − 1) du i + δn  (8.59)
 λ 
i=1 n=1

where si is the signal strength of the two interference signals, u i = cos φi , and δn are the
adapted quantized phases.
Only two least significant bits are needed to perform the nulling (11.25◦ and 5.625◦ ).
The adapted phase settings are shown in Figure 8.6. Figure 8.7 shows the adapted pattern

FIGURE 8-6
15 Adapted phase
weights that place
nulls in the far-field
10
pattern at u = 0.62
and u = 0.72.
5 From R. L. Haupt,
“Phase-only
Phase

0 adaptive nulling with

a genetic algorithm,”
IEEE Transactions on
−5
Antennas and
Propagation, Vol. 45,
−10 No. 6, 1997,
pp. 1009–1015.
−15

5 10 15 20 25 30 35 40
Element
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 350

350 CHAPTER 8 Random Search Algorithms

FIGURE 8-7 0
Adapted (solid line)
and quiescent
(dotted line) array
−10
patterns for two
60 dB interference

Relative pattern in dB
sources at u = 0.62
and u = 0.72. From −20
R. L. Haupt,
“Phase-only
adaptive nulling with −30
a genetic algorithm,”
IEEE Transactions on
Antennas and
Propagation, Vol. 45, −40
No. 6, 1997,
pp. 1009–1015.
−50
−1 −0.5 0 0.5 1
u

FIGURE 8-8 Null −35

depths at each null
u = 0.62 (solid line) −40
and u = 0.72
(dashed line). From −45
R. L. Haupt,
“Phase-only
Null depth in dB

−50
adaptive nulling with
a genetic algorithm,” −55
IEEE Transactions on
Antennas and −60
Propagation, Vol. 45,
No. 6, 1997, pp. −65
1009–1015.
−70

−75
0 5 10 15 20 25
Iteration

superimposed on the quiescent pattern. Convergence occurred in eight iterations as shown

by the plots in Figure 8.8, Figure 8.9, and Figure 8.10. The solid line represents the
best chromosome of the population, and the dashed line is the average of the popula-
tion. After eight iterations, the algorithm reaches a set of adaptive weights that have
effectively suppressed the two interferers. At this point, the algorithm required 68 power
measurements.
Amplitude and phase adaptive nulling with a genetic algorithm has been successfully
verified on an experimental antenna (Figure 8.11) that has 128 vertical columns with 16
dipoles per column equally spaced around a cylinder 104 cm in diameter [25]. Figure 8.12 is
a cross sectional view of the antenna. The signals from the 16 dipoles are added together
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 351

8.4 Genetic Algorithm 351

FIGURE 8-9
30
Output power of the
best chromosome
25 (solid line) and
average
chromosome of the
Output power in dB

20
population (dashed
line). From R. L.
15 Haupt, “Phase-only
adaptive nulling with
a genetic algorithm,”
10
IEEE Transactions on
Antennas and
5 Propagation, Vol. 45,
No. 6, 1997,
0
pp. 1009–1015.

0 5 10 15 20 25
Iteration

10

−5
SNR in dB

−10

−15

−20

−25

−30

−35
0 5 10 15 20 25
Iteration

FIGURE 8-10 Signal-to-noise ratio (SNR) of the array as a function of the number of iter-
ations of the genetic algorithm. The solid line is the SNR of the best chromosome, and the
dashed line is the SNR of the average chromosome of the population. From R. L. Haupt,
“Phase-only adaptive nulling with a genetic algorithm,” IEEE Transactions on Antennas and
Propagation, Vol. 45, No. 6, 1997, pp. 1009–1015.

to form a fixed elevation main beam pointing 3◦ above horizontal. Eight consecutive
elements are active at a time and with the elements spaced 0.42λ apart at 5 GHz. Each
element has an eight-bit phase shifter (least significant bit equal to 0.0078125π radians)
and eight-bit attenuator (least significant bit equal to .3125 dB). The antenna has a 25 dB
n = 3 Taylor amplitude taper.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 352

352 CHAPTER 8 Random Search Algorithms

FIGURE 8-11 The

cylindrical array has
128 elements with
8 active at a time.
From R. L. Haupt,
“Adaptive nulling
with a cylindrical
array,” AFRL-SN-
RS-TR-1999-36,
March 1999.

FIGURE 8-12
Cross section of q
experimental
cylindrical array.
Elements
Active
elements

A 5 GHz continuous source served as the interference. Only the four least significant
bits of the phase shifters and attenuators were used to perform the adaptive nulling, so
minimal distortion occurs to the main beam. The genetic algorithm had a population size
of 16 chromosomes, and only one bit in the population was mutated every generation (mu-
tation rate of 0.1%). The algorithm placed a deep null in less than 30 power measurements
as shown in Figure 8.13 when the interference was at 45◦ . Figure 8.14 is the convergence
plot for placing the null in the antenna pattern in Figure 8.13.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 353

8.5 Comparison of Random Search Algorithms 353

0 FIGURE 8-13
Measured adapted
and quiescent
−10 far-field patterns for
an interference
source at 45◦ .
Far field pattern (dB)

−20

−30

−40
Quiescent
−50
Adapted

−60
−60 −30 0 30 60
q (degrees)

FIGURE 8-14 GA
convergence for
−30 placing a null at 45◦ .
Sidelobe level (dB)

−40

−50

−60
0 1 2 3 4 5
Generation

8.5 COMPARISON OF RANDOM

SEARCH ALGORITHMS
A four-element array was simulated having the array configuration and signal environment
represented in Figure 4.30. The random search algorithms have convergence speeds that are
much slower than for the LMS algorithm when the MSE performance measure is used, and
the selected three-jammer scenario was simulated under two different jamming conditions:
one condition having a moderate eigenvalue spread of λmax /λmin = 153.1; and another
condition with a more severe eigenvalue spread of λmax /λmin = 2440. When we select
jammer-to-thermal noise ratios of J1 /n = 25, J2 /n = 4, and J3 /n = 20 and a signal-
to-thermal noise ratio of s/n = 10, the corresponding eigenvalues are λ1 = 153.1, λ2 =
42.6, λ3 = 3.34, and λ4 = 1 for which SNRopt = 15.9 (12 dB). Figures 8.15–8.18
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 354

354 CHAPTER 8 Random Search Algorithms

FIGURE 8-15 15.00

Output SNR versus
10.00 Optimum
number of iterations
for the LRS 5.00

Output SNR (dB)

algorithm with
eigenvalue 0
spread = 153.1. −5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5678 1E1 2 3 4 5678 1E2 2 3 4 5678 1E3 2 3 4 5
Number of iterations

FIGURE 8-16 15.00

Output SNR versus
10.00 Optimum
number of iterations
for the ARS 5.00
Output SNR (dB)

algorithm with
eigenvalue 0
spread = 153.1. −5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2
Number of iterations

FIGURE 8-17 15.00

Output SNR versus
10.00 Optimum
number of iterations
for the GARS 5.00
Output SNR (dB)

algorithm with
eigenvalue 0
spread = 153.1. −5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2
Number of iterations

give typical convergence results for the case where λmax /λmin = 153.1. Likewise, when
we select jammer-to-thermal noise ratios of J1 /n = 500, J2 /n = 40, and J3 /n = 200
and a signal-to-thermal noise ratio of s/n = 10, the corresponding eigenvalues are then
λ1 = 2440, λ2 = 494, λ3 = 25.6, and λ4 = 1 for which SNRopt = 15.08 (11.8 dB).
Figures 8.19–8.24 then give typical convergence results for the case where λmax /λmin =
2, 440.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 355

8.5 Comparison of Random Search Algorithms 355

15.00 FIGURE 8-18

Output SNR versus
10.00 Optimum
number of iterations
Output SNR (dB) 5.00 for the LMS
algorithm with
0 eigenvalue
−5.00 spread = 153.1.

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2
Number of iterations

15.00 FIGURE 8-19

Output SNR versus
10.00 Optimum number of iterations
5.00 for the LRS
Output SNR (dB)

algorithm with
0 eigenvalue
−5.00 spread = 2,440.

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5678 1E1 2 3 4 5678 1E2 2 3 4 5678 1E3 2 3 4 5
Number of iterations

15.00 FIGURE 8-20

Output SNR versus
10.00 Optimum number of iterations
5.00 for the ARS
Output SNR (dB)

algorithm with
0 eigenvalue
−5.00 spread = 2,440.

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2
Number of iterations

The ARS weight adjustment scheme in Figure 8.1 needs modification, because the
farther w(k + 1) is from wopt the greater is the variance in the estimate ξ̂ [w(k + 1)].
Consequently, if the step size μ0 is selected to obtain an acceptable steady-state error in
the neighborhood of wopt , it may well be that the changes in ξ [w(k + 1)] occurring as a
consequence of the perturbation w(k + 1) are overwhelmed by the random fluctuations
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 356

356 CHAPTER 8 Random Search Algorithms

FIGURE 8-21 15.00

Output SNR versus
10.00 Optimum
number of iterations
for the GARS 5.00

Output SNR (dB)

algorithm with
eigenvalue 0
spread = 2,440. −5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 678 1E1 2 3 4 5 678 1E2 2 3 4 5 678 1E3 2
Number of iterations

FIGURE 8-22 15
Output SNR versus
number of iterations Optimum
for the GA with 10
eigenvalue
spread = 2,440. 5
SNR (dB)

−5

−10

−15
100 101 102 103
Iteration

FIGURE 8-23 90
Adapted pattern 120 60
after 1,000 iterations
of the GA. 10 dB
150 30
0 dB
−10 dB
180 0

210 330

240 300
270
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 357

8.5 Comparison of Random Search Algorithms 357

1 5.00 FIGURE 8-24

Output SNR versus
10.00 Optimum
number of iterations
Output SNR (dB) 5.00 for the LMS
algorithm with
0 eigenvalue
−5.00 spread = 2,440.

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5678 1E1 2 3 4 5678 1E2 2 3 4 5678 1E3 2 3 4 5
Number of iterations

experienced in ξ̂ [w(k + 1)] when w(k + 1) is far from wopt . When this situation occurs,
the adjustment algorithm yields a succession of weights that slowly meander aimlessly
with step size μ0 . As a result, the step size μ0 should reflect the changes in the variance
of ξ̂ [w(k + 1)] that occur when ! w(k + 1) is far from wopt . This correction involves
incorporating a step size μs = K 1 + K 2 ∗ into the ARS algorithm in accordance with
the philosophy expressed! by the GARS algorithm in Figure 8.4. Of course, it would be
preferable to use μs = K 1 + K 2 (∗ − min ), but in general min is unknown.
The LRS, ARS, and GARS algorithms were all simulated using K = 90 to obtain
the estimate of MSE, which was the performance measure used in all cases. To satisfy the
condition imposed by (8.49), the GARS algorithm was simulated using

σ 2 tr(Rxx ) = 1 + 2 ξ ∗ (8.60)

where 1 = 1
160
and 2 = 0.1. Likewise, the ARS algorithm was simulated using

μ2s tr(Rxx ) = 1 + 2 ξ ∗ (8.61)

with 1 and 2 assigned the same values as for the GARS algorithm. The LRS algorithm was
simulated using the constants μs = 1.6 and σ 2 tr(Rxx ) = 0.05, thereby yielding a greater
misadjustment error than either the ARS or GARS algorithms. The LMS algorithm was
also simulated for purposes of comparison with step size corresponding to μs tr(Rxx ) = 0.1
and using an estimated gradient derived from the average value of three samples of e(k)x(k)
so K = 3 instead of the more common K = 1. In all cases the initial weight vector was
taken to be wT (0) = [0.1, 0, 0, 0].
The results of Figures 8.15–8.17 show that both the ARS and GARS algorithms are
within 3 dB of the optimum SNR after about 800 iterations, whereas the LRS algorithm
does not reach this point after 4,000 iterations, even though the misadjustment is more
severe than for the ARS and GARS algorithms. This result indicates that the misadjustment
versus speed of adaptation trade-off favors the ARS and GARS algorithms more than the
LRS algorithm. The LMS algorithm by contrast is within 3 dB of the optimum output
SNR after only 150 iterations with only a small degree of misadjustment. The extreme
disparity in speed of convergence between the LMS algorithm and the three random search
algorithms is actually more pronounced than the comparison of number of iterations
indicates because each iteration in the random search algorithms represents 90 samples,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 358

358 CHAPTER 8 Random Search Algorithms

whereas each iteration in the LMS algorithm represents only three samples. Consequently,
the time scale on Figures 8.15–8.17 is 30 times greater than the corresponding time scale
on Figure 8.18.
The results given in Figures 8.18–8.24 for the case where eigenvalue spread = 2,440
confirm the previous results obtained with an eigenvalue spread of only 153.1. These
results also show that both the random search algorithms and the LMS algorithm are
sensitive to eigenvalue spread in the Rxx matrix.
Figure 8.22 is a plot of the signal to noise ratio versus iteration when a GA controls the
complex element weights of the Y -array. The GA has a population size of 8 and a mutation
rate of 15%. The best result for each iteration appears in the plot. The number of power
measurements per iteration is less than 7. The GA optimized the SNR without knowledge
of the signal and jammer directions and powers. Figure 8.23 shows the adapted pattern
after 1,000 iterations of the GA. Phase-only adaptive nulling is not a good alternative in
this case, because there are not enough degrees of freedom to null all the jammers. It is
remarkable that the convergence results of the GA in Figure 8-22 are very close to the
results obtained for the LMS algorithm shown in Figure 8-24. The GA algorithm is the
only random search algorithm that is actually competitive with the LMS algorithm in this
extreme eigenvalue spread condition.

8.6 SUMMARY AND CONCLUSIONS

Random search algorithms search irregular and multimodal performance surfaces and
require only a direct evaluation of the selected performance measure to implement. The
weight adjustment computation for the LRS, ARS, GARS, and GA algorithms is extremely
simple, requiring only modest computational power. More elaborate and complicated
random search algorithms have been applied to adaptive control and pattern recognition
systems [15], but the introduction of more sophisticated measures into random searches
removes their simplicity (which is a primary virtue for adaptive array applications).
The price to be paid for simple computation and implementation requirements is
longer convergence time to reach the optimal weight vector solution, although both the
ARS and GARS algorithms have more favorable misadjustment versus speed of conver-
gence trade-offs than the LRS algorithm. Nevertheless, the convergence speed realized
with the LMS algorithm on unimodal performance surfaces is orders of magnitude faster
than that realized by a random search algorithm. Furthermore, these three random search
algorithms exhibited the same degree of convergence speed sensitivity to eigenvalue spread
as the LMS algorithm. Finally, these random searches exhibit relatively slow convergence
in high-dimensional spaces, a characteristic that reflects the fact that as the number of
possible directions to be searched increases, then the convergence time also increases.
This performance characteristic leads to the suggestion that reducing the dimensionality
of the parameter space search over the final stages of convergence is helpful. Note that
the brief discussion given here is not a complete survey of search algorithms that employ
only measurements of the output power. In particular, consideration should be given to
the “directed search” techniques like that proposed by Hooke and Jeeves [26] which may
well exhibit superior convergence properties under certain conditions.
The genetic algorithm and other evolutionary and nature-based algorithms have proven
useful in overcoming the deficiencies of the other random search algorithms. They easily
escape from local minima, handle large numbers of parameters, and have been experi-
mentally demonstrated.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 359

8.7 Problems 359

8.7 PROBLEMS
1. Misadjustment versus Speed of Adaptation Trade-off for the LRS Algorithm [15]
Assuming all eigenvalues are equal so that (T pmse )av = Tmse , plot Tmse versus N for
the LRS algorithm assuming (Mtot )min = 10% in (8.43), and compare this result with
the corresponding plots obtained for the LMS and DSD algorithms in Problem 1 of
Chapter 4.
2. Search Loss for a Simple Random Search with Reversing Step [27]
Consider the simple random search algorithm described by

x(i + 1) = x(i) + x(i + 1)

where x is a random displacement vector satisfying |x|2 = 1 and


ξ if performance improvement is observed
x(i + 1) =
−x(i) if no performance improvement is observed
where ξ is the random direction of the displacement vector.
(a) Consider the parameter space of Figure 8.25 in which a performance improve-
ment is realized for any angle in the range –π/2 < φ < π/2 and no performance
improvement is realized for any angle in the range π/2 ≤ φ ≤ 3π/2. The mean
displacement in the direction of a successful random step is given by
" π/2
U (n) = cos φp(φ)dφ
0

where n is the number of degrees of freedom and p(φ) is the probability density
function of the angle φ for a uniform distribution of directions of the random step
in the n-dimensional space. Show that
sinn−2 φ (n − 1)
p(φ) = # π/2 =   n−1 2 sin
n−2
φ
2 0
n−2
sin φdφ 2n−2  2
where (·) is the gamma function.
Hint: Note that the area of a ring-shaped zone on the surface of an n-dimensional
sphere corresponding to the angle dφ is An−2 × sinn−2 φdφ. Consequently, the area

FIGURE 8-25
Parameter space
section showing
displacement vector
X and direction φ.
f
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 360

360 CHAPTER 8 Random Search Algorithms

of the surface of the hypersphere included in the hypercone with vertical angle 2φ
is given by
" φ
S(φ) = An−2 sinn−2 φdφ
0

The probability that a random vector lies in this cone for a uniform probability
choice of the random direction is equal to the ratio of the “areas” S(φ) and S(π).
The desired probability density is then the derivative of this ratio with respect to the
angle φ.
(b) Show that U (n) defined in part (a) is given by
# π/2
cos φ sinn−2 φdφ (n − 1)
U (n) = 0
# π/2 =   2
0 sin n−2
φdφ 2n−3 (n − 1)  n−1
2

Since the probability of successful and unsuccessful steps is the same for the pa-
rameter space of Figure 8.25, then on the average for one successful step there is
one unsuccessful step and a corresponding reverse step—three steps in all. There-
fore, the mean displacement for one successful step is reduced by two-thirds and is
only 13 U (n).
Defining search loss to be the number of steps required by the search such
that the vector sum of these steps has the same length as one operating step in the
successful direction, it follows that the mean search loss for the aforementioned
algorithm is just 3/U (n).
3. Relative Search Efficiency Using the Search Loss Function [27]
Consider a fixed step size gradient search defined by
(i)
x(i + 1) = x(i) − a(i)μs
|(i)|
where

μs = step size
 
∂F ∂F
(i) = ,...,
∂ x1 ∂ xn x(i)

1 if F[x(i + 1)] < F[x(i)]
a(i) =
0 otherwise

and where F[·] denotes a known performance measure. Likewise, consider a fixed
step size random search defined by

x(i + 1) = x(i) − b(i)x(i) + x(i + 1)

where x(i) is a random vector having a uniform distribution on a hypersphere of

radius μs and centered at the origin, and

1 if F[x(i)] ≥ F[x(i − 1)]
b(i) =
0 otherwise
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 361

8.7 Problems 361

Define the search loss function to be the performance measure F[x] divided by the
ratio of performance improvement per performance evaluation, that is,

 F(x)
SL(x) =  
F(x)−F(x+x)
N

where N is the total number of performance measure evaluations required at each

search point.
(a) For the quadratic performance measure F(x) = ρ 2 , show that for the gradient search
algorithm F(x + x) = ρ 2 − 2ρμs + μ2s .
(b) To empirically evaluate the gradient  at some point x requires the evaluation of
F[·] at x and at points along each coordinate direction separated from x by some
distance ε. Therefore, a total of n + 1 function evaluations are required (where n =
number of degrees of freedom). Show for the previously defined gradient search
that

ρ 2 (n + 1)
S L(x) = 
2ρμs − μ2s

(c) For a given base point x(i), the successor trial state x(i + 1) for the random search
defines an angle φ with respect to a line connecting x(i) with the extremum point
of the performance measure for which the probability density function p(φ) was
obtained in Problem 2. Show that the expected value of performance measure im-
provement using the previously defined random search is
" φ0
E{−F} = E{ρ } = 2
ρ 2 p(φ)dφ
0

where
 
−1 μs
φ0 = cos and ρ 2 = 2ρμs cos φ − μ2s
2ρ

(d) Show for the previous random search algorithm that

# π/2
2ρ 2 sinn−2 φdφ
S L(x) = # φ0 0
#φ
0 2μs ρ cos φ sin n−2
φdφ − 0 0 μ2s sinn−2 φdφ

The search loss function of parts (b) and (d) is compared for specific values of
ρ, μs , and n to determine whether the gradient search or the random search is more
efficient.
4. Search Loss Function Improvement Using Step Reversal [28]
The relative efficiency of the fixed step size random algorithm introduced in Problem
3 is significantly improved merely by adding a “reversal” feature to the random search.
The fixed step size random search algorithm with reversal is described by

x(i + 1) = x(i) + c(i)x(i + 1) + 2[c(i) − 1]x(i)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 362

362 CHAPTER 8 Random Search Algorithms

where

1 if F[x(i)] ≤ F[x(i − 1)]
c(i) =
0 otherwise

The previous modification of the random search in effect searches at y − y if the

initial search at y+y failed to produce an improvement in the performance measure.
(a) Show that
# φ0
3 ρ 2 sinn−2 φdφ
E{−F} = E{ρ } = 2 0
#π
2 0 sinn−2 φdφ
where
 
−1 μs
φ0 = cos and ρ 2 = 2ρμs cos φ − μ2s
2
(b) Show that
# π/2
4μ2s sinn−2 φdφ
S L(x) = # φ0 0
#φ
0 6μs ρ cos φ sin n−2
φdφ − 0 0 3μ2s sinn−2 φdφ
This search loss is compared with that obtained in Problem 3(d) for specified ρ, μs ,
and n to determine exactly the improvement that is realized by addition of the
reversal feature.
5. Genetic Algorithm. A 20-element array of point sources spaced 0.5λ apart has six-bit
amplitude and phase weights and a 20 dB, n = 3 low sidelobe Taylor amplitude taper.
The desired signal is incident on the peak of the main beam and is normalized to 1 or
0 dB. Two 30 dB jammers enter the sidelobes at 111◦ and 117◦ . The genetic algorithm
has a population size of 8 and a mutation rate of 10%. Plot the power levels received
by the array versus generation, the ratio of the signal power to the jammer versus
generation, and the adapted antenna pattern superimposed on the quiescent pattern.

8.8 REFERENCES
[1] M. K. Leavitt, “A Phase Adaptation Algorithm,” IEEE Trans. Antennas Propag., Vol. AP-24,
No. 5, September 1976, pp. 754–756.
[2] P. A. Thompson, “Adaptation by Direct Phase-Shift Adjustment in Narrow-Band Adaptive
Antenna Systems,” IEEE Trans. Antennas Propag., Vol. AP-24, No. 5, September 1976,
pp. 756–760.
[3] G. J. McMurty, “Search Strategies in Optimization,” Proceedings of the 1972 International
Conference on Cybernetics and Society, October, Washington, DC, pp. 436–439.
[4] C. Karnopp, “Random Search Techniques for Optimization Problems,” Automatica, Vol. 1,
August 1963, pp. 111–121.
[5] G. J. McMurty and K. S. Fu, “A Variable Structure Automation Used as a Multimodal Search-
ing Technique,” IEEE Trans. Autom. Control, Vol. AC-11, July 1966, pp. 379–387.
[6] R. L. Barton, “Self-Organizing Control: The Elementary SOC—Part I,” Control Eng., February
1968.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 363

8.8 References 363

[7] R. L. Barton, “Self-Organizing Control: The General Purpose SOC— Part II,” Control Eng.,
March 1968.
[8] M. A. Schumer and K. Steiglitz, “Adaptive Step Size Random Search,” IEEE Trans. Autom.
Control, Vol. AC-13, June 1968, pp. 270–276.
[9] R. A. Jarvis, “Adaptive Global Search in a Time-Variant Environment Using a Probabilistic
Automaton with Pattern Recognition Supervision,” IEEE Trans. Syst. Sci. Cybern., Vol. SSC-6,
July 1970, pp. 209–217.
[10] A. N. Mucciardi, “Self-Organizing Probability State Variable Parameter Search Algorithms
for the Systems that Must Avoid High-Penalty Operating Regions,” IEEE Trans. Syst., Man,
Cybern., Vol. SMC-4, July 1974, pp. 350–362.
[11] R. A. Jarvis, “Adaptive Global Search by the Process of Competitive Evolution,” IEEE Trans.
Syst., Man, Cybern., Vol. SMC-5, May 1975, pp. 297–311.
[12] R. L. Haupt, “Beamforming with Genetic Algorithms,” Adaptive Beamforming, S. Chandran,
ed., pp. 78–93, Berlin, Springer-Verlag, 2004.
[13] R. L. Haupt, “Genetic Algorithms for Antennas,” Modern Antenna Handbook, C. Balanis,
New York, Wiley, 2008.
[14] R. L. Haupt, Antenna Arrays: A Computational Approach, New York, Wiley, 2010.
[15] B. Widrow and J. M. McCool, “A Comparison of Adaptive Algorithms based on the Methods
of Steepest Descent and Random Search,” IEEE Trans. Antennas Propag., Vol. AP-24, No. 5,
September 1976, pp. 615–637.
[16] C. A. Baird and G. G. Rassweiler, “Search Algorithms for Sonobuoy Communication,” Pro-
ceedings of the Adaptive Antenna Systems Workshop, March 11–13, 1974, NRL Report 7803,
Vol. 1, September 27, 1974, pp. 285–303.
[17] R. L. Barron, “Inference of Vehicle and Atmosphere Parameters from Free-Flight Motions,”
AIAA J. Spacecr. Rockets, Vol. 6, No. 6, June 1969, pp. 641–648.
[18] R. L. Barron, “Guided Accelerated Random Search as Applied to Adaptive Array AMTI
Radar,” Proceedings of the Adaptive Antenna Systems Workshop, March 11–13, 1974, Vol. 1,
NRL Report 7803, September 27, 1974, pp. 101–112.
[19] A. E. Zeger and L. R. Burgess, “Adaptive Array AMTI Radar,” Proceedings of the Adaptive
Antenna Systems Workshop, March 11–13, 1974, NRL Report 7803, Vol. 1, September 27,
1974, pp. 81–100.
[20] A. N. Mucciardi, “A New Class of Search Algorithms for Adaptive Computation,” Proc.
1973 IEEE Conference on Decision and Control, December, San Diego, Paper No. WA5-3,
pp. 94–100.
[21] R. L. Haupt and S. E. Haupt, Practical Genetic Algorithms, 2d ed., New York, John Wiley &
Sons, 2004.
[22] R. L. Haupt and D. Werner, Genetic Algorithms in Electromagnetics, New York, Wiley, 2007.
[23] J. Kennedy and R. C. Eberhart, Swarm Intelligence, San Francisco, Morgan Kaufmann
Publishers, 2001.
[24] R. L. Haupt, “Phase-Only Adaptive Nulling with a Genetic Algorithm,” IEEE Transactions
on Antennas and Propagation, Vol. 45, No. 6, 1997, pp. 1009–1015.
[25] R. L. Haupt and H. L. Southall, “Experimental Adaptive Nulling with a Genetic Algorithm,”
Microwave Journal, Vol. 42, No. 1, January 1999, pp. 78–89.
[26] R. Hooke and T. A. Jeeves, “Direct Search Solution of Numerical and Statistical Problems,”
J. of the Assoc. for Comput. Mach., Vol. 8, April 1961, pp. 212–229.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:10 364

364 CHAPTER 8 Random Search Algorithms

[27] L. A. Rastrigin, “The Convergence of the Random Search Method in the External Con-
trol of a Many-Parameter System,” Autom. Remote Control, Vol. 24, No. 11, April 1964,
pp. 1337–1342.
[28] J. P. Lawrence, III and F. P. Emad, “An Analytic Comparison of Random Searching and
Gradient Searching for the Extremum of a Known Objective Function,” IEEE Trans. Autom.
Control, Vol. AC-18, No. 6, December 1973, pp. 669–671.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:44 365

CHAPTER

Adaptive Algorithm
Performance Summary 9

Chapters 4 through 9 considered the transient response characteristics and implementation

considerations associated with different classes of algorithms that are widely used for
adaptive array applications. This chapter summarizes the principal characteristics of each
algorithm class before considering some practical problems associated with adaptive array
system design.
In each chapter of Part 2 the convergence speed of an algorithm representing a distinct
adaptation philosophy was compared with the convergence speed of the least mean squares
(LMS) algorithm. The convergence speeds of the various algorithms are compared for a
selected example in this chapter. Since the misadjustment versus rate of adaptation trade-
offs for the random search algorithms—linear random search (LRS), accelerated random
search (ARS), and guided accelerated random search (GARS)—and for the differential
steepest descent (DSD) algorithm of Chapter 4 are unfavorable compared with the LMS
algorithm, recourse to these methods would be taken only if the meager instrumentation
required was regarded as a cardinal advantage or nonunimodal performance surfaces were
of concern. Furthermore, the Howells–Applebaum maximum signal-to-noise ratio (SNR)
algorithm has a misadjustment versus convergence speed trade-off that is nearly identical
with the LMS algorithm. Attention for the direct comparison consequently is focused on
the following adaptive algorithms:
1. LMS error algorithm (Section 4.2 of Chapter 4)
2. Powell’s accelerated gradient (PAG) algorithm (Section 4.4.1 of Chapter 4)
3. Direct matrix inversion (DMI) [Version (6.24) from Section 5.1.2 of Chapter 5]
4. Recursive (R) algorithm (Section 6.5 of Chapter 6)
5. Gram–Schmidt cascade preprocessor (GSCP) (Section 7.4 of Chapter 7)
Each algorithm starts with the same initial weight vector wT (0) = [0.1, 0, 0, 0],
for the array geometry and signal configuration of Figure 4.19 with signal conditions
corresponding to an eigenvalue spread of 2440. The convergence results averaged over 10
random trials for each algorithm appear in Figures 9.1–9.5.
The results of Figures 9.1–9.5 show that the LMS algorithm requires about 1,750
data samples to converge within 3 dB of the optimum output SNR, the PAG algo-
rithm requires 110 iterations (990 data samples), the DMI and recursive algorithms both
require 8 data samples, and the GSCP algorithm requires 40 iterations (120 data sam-
ples). Algorithm parameters were selected so the degree of steady-state misadjustment
for the LMS, PAG, and GSCP algorithms is comparable. Misadjustment of the DMI and
365
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:44 366

366 CHAPTER 9 Adaptive Algorithm Performance Summary

FIGURE 9-1 15.00

Output SNR versus
10.00 Optimum
number of iterations
for LMS algorithm 5.00
with α L = 0.1.

Output SNR (dB)

0

−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5678 1E1 2 3 4 56 8 1E2 2 3 4 56 8 1E3 2 3 4 5
Number of iterations

FIGURE 9-2 15.00

Output SNR versus
10.00 Optimum
number of iterations
for PAG algorithm 5.00
with K = 9 samples
Output SNR (dB)

per iteration. 0

−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 56
Number of iterations

FIGURE 9-3 15.00

Output SNR versus
10.00 Optimum
number of iterations
for DMI algorithm. 5.00
Output SNR (dB)

−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 56
Number of iterations

recursive algorithms decreases as the number of iterations increases and cannot be mod-
ified by altering algorithm parameters. These results indicate that the DMI and recursive
algorithms offer by far the best misadjustment versus speed of convergence trade-off, fol-
lowed (in order) by the GSCP algorithm, the PAG algorithm, and the LMS algorithm for
this moderate eigenvalue spread condition of λmax /λmin = 2,440.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:44 367

CHAPTER 9 Adaptive Algorithm Performance Summary 367

15.00 FIGURE 9-4

Output SNR versus
10.00 Optimum number of iterations
5.00 for recursive
algorithm with α = 1
Output SNR (dB)

0 and P(0) = I.
−5.00

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 56
Number of iterations

15.00 FIGURE 9-5

Output SNR versus
10.00 Optimum number of iterations
5.00 for GSCP with
α = α L = 0.1 and
Output SNR (dB)

0 K = 3 samples per
−5.00 iteration.

−10.00

−15.00

−20.00

−25.00
1 2 3 4 5 6 7 8 1E1 2 3 4 5 6 7 8 1E2 2 3 4 5 6 7 1E3
Number of iterations

Table 9.1 summarizes the principal operational characteristics associated with the
adaptive algorithms considered throughout Part 2. The algorithms that achieve the simplest
possible instrumentation by requiring only direct measurement of the selected performance
measure pay a severe penalty in terms of the increased convergence time required to reach
the steady-state solution for a given degree of misadjustment. Accepting the instrumenta-
tion necessary to incorporate one correlator for each controlled array element enables the
misadjustment versus convergence speed trade-off for the LMS and Howells–Applebaum
interference suppression loops to be achieved. Further improvement in the misadjustment
versus convergence speed trade-off is obtained where sensitivity to eigenvalue spread is
a concern by paying the price of additional instrumentation or additional computational
power. As the shift to digital processing continues, algorithms requiring more sophisti-
cated computation become not only practicable but also preferable in many cases. Not
only can high performance be achieved that was impractical before, but also the low cost
of the increased computational power may in some cases render a sophisticated algorithm
more economical.
The genetic algorithm (GA) is a “global” minimum seeker, so it is less likely to
get stuck in a local minimum like the steepest descent algorithm. It minimizes the total
output power, if the number of adaptive elements or the adaptive weight range are limited.
 TABLE 9-1 Operational Characteristics Summary of Selected Adaptive Algorithms

368
LMS MSNR PAG DSD DMI R GSCP RS GA
Monzingo-7200014

Algorithm Steepest descent Similar Conjugate Perturbation Direct estimate Data weighting Orthogonalize Trial and error Minimize total
philosophy to LMS gradient descent technique of covariance; similar to input signals power or
open loop Kalman filter; maximize SNR
book

closed loop
Transient Misadjustment Nearly More favorable Unfavorable Achieves the Same as DMI More favorable Unfavorable Limited by the
response vs. convergence the misadjustment misadjustment fastest misadjustment misadjustment hardware
characteristic speed trade-off same as vs. convergence vs. convergence convergence vs. convergence vs. convergence
is acceptable for LMS speed trade-off speed trade-off with most speed trade-off speed trade-off
numerous than LMS compared to favorable than PAG. compared to
applications LMS misadjustment Convergence LMS;
vs. convergence speed accelerated
speed trade-off approaches DMI steps improve
ISBN : XXXXXXXXXX

speed
Algorithm Easy to Same as Convergence Easy to Very fast Same as DMI; Convergence Can be applied Minimal
strengths implement, LMS speed less implement; convergence different data speed enjoys to any directly hardware, fast,
requiring N sensitive to requires only speed weighting reduced measurable does not get
correlators and eigenvalue instrumentation independent of schemes are sensitivity to performance stuck in local
integrators; spread than to directly eigenvalue easily eigenvalue index; easy to minimum,
tolerant of LMS; fast measure the spread incorporated spread implement, with independent of
hardware errors convergence for performance compared with meager eigenvalue
November 24, 2010

small number index LMS; tolerant of instrumentation spread

of degrees of hardware errors and computation
freedom requirements
19:44

Algorithm Convergence Same as Relatively Rate of Requires Requires Requires a Convergence Some
weaknesses speed sensitive LMS difficult to convergence N (N + 1)/2 N (N + 1)/2 respectable speed sensitive measurements
to eigenvalue implement and sensitive to correlators to correlators and amount of to eigenvalue are bad each
368

spread requires parallel eigenvalue implement; heavy hardware— spread and the iteration
processors; spread, with matrix inversion computational N (N + 1)/2 slowest of all
convergence speed requires load adaptive loops algorithms
speed sensitive comparable to adequate to implement considered
to number of that of RS with precision and
degrees of accelerated step N 3 /2 + N 2
freedom complex
multiplies
Chapter 4 4 4 4 5 6 7 8 8
reference
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:44 369

CHAPTER 9 Adaptive Algorithm Performance Summary 369

Since it works with standard phased array hardware, it can be implemented on existing
arrays and is much cheaper than requiring a receiver at each element. The convergence
is relatively independent of the interference and signal power levels. Each iteration, the
GA must evaluate the entire population, which can be small. As a result, there are bad
measurements in addition to the improved power and SNR measurements. The transient
response is limited by the hardware switching speed and settling time. It may be necessary
to average a few power measurements to get the desired results.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:44 370
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 371

P A R T III

Advanced Topics

CHAPTER 10 Compensation of Adaptive Arrays

CHAPTER 11 Direction of Arrival Estimation and
Related Topics
CHAPTER 12 Recent Developments in Adaptive
Arrays
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 372
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 373

CHAPTER

Compensation of
Adaptive Arrays 10
' $
Chapter Outline
10.1 Array Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
10.2 Array Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
10.3 Broadband Signal Processing Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
10.4 Compensation for Mutual Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
10.5 Multipath Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
10.6 Analysis of Interchannel Mismatch Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
10.7 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
10.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
10.9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
& %

Narrowband adaptive arrays need only one complex adaptive weight in each element chan-
nel. Broadband adaptive arrays, however, require tapped delay lines (transversal filters)
in each element channel to make frequency-dependent amplitude and phase adjustments.
The analysis presented so far assumes that each element channel has identical electron-
ics and no reflected signals. Unfortunately, the electrical characteristics of each channel
are slightly different and lead to “channel mismatching” in which significant differences
in frequency-response characteristics from channel to channel may severely degrade an
array’s performance without some form of compensation. This chapter starts with an anal-
ysis of array errors and then addresses array calibration and frequency-dependent mis-
match compensation using tapped delay line processing, which is important for practical
broadband adaptive array designs.
The number of taps used in a tapped delay line processor depends on whether the
tapped delay line compensates for broadband channel mismatch effects or for the effects
of multipath and finite array propagation delay. Minimizing the number of taps required
for a specified set of conditions is an important practical design consideration, since each
additional tap (and associated weighs) increases the cost and complexity of the adaptive
array system.

373
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 374

374 CHAPTER 10 Compensation of Adaptive Arrays

10.1 ARRAY ERRORS

Array errors result from the manufacturing tolerances defined by the materials, processes,
and construction of the components in an array. These small errors are random, because the
manufacturing techniques employed have very tight tolerances. The random differences
between any components distort the signal path by adding phase and amplitude errors
as well as noise to each signal. These types of errors are static, because once measured
they remain relatively unchanged over the life of the component. Higher frequencies have
tighter tolerances for phase distortion than lower frequencies, because the errors are a
function of wavelength. Not only are the accuracy of the dimensions of the components
important, but the accuracy of the values of the constitutive parameters of the components
are also important. For instance, the dielectric constant determines the wavelength and
hence the phase of the signal passing through it, so an error in the dielectric constant
produces a phase error.
Dynamic errors change with time and are primarily due to changes in temperature.
Online calibration corrects for these dynamic errors also takes care of any drift in the static
errors. The dynamic errors are also frequency dependent. The effects of temperature are
smallest at the center frequency and increase as the frequency migrates away from the
center frequency.

10.1.1 Error Analysis

Random errors that affect arrays fall into four categories:
1. Random amplitude error, δna
2. Random phase error, δnp
3. Random position error, δns 
1 element functioning properly
4. Random element failure, Pne =
0 element failure
The first three types of random errors fit into the array factor as perturbations to the array
weights and element locations


N
 
an + δna e j ( pn +δn ) e jk (sn +δn )u
p s
AF err = (10.1)
n=1

Element failures result when an element no longer transmits or receives. The probability
that an element has failed, 1 − Pe , is the same as a root mean square (rms) amplitude error,
δna 2 . Position errors are not usually a problem, so a reasonable formula to calculate the
rms sidelobe level of the array factor for amplitude and phase errors with element failures
is [1]
p2
(1 − Pe ) + δna 2 + Pe δn
sllrms =   (10.2)
p2
Pe 1 − δn ηt N

Figure 10-1 is an example of a typical corporate-fed array. A random error that occurs
at one element is statistically uncorrelated with a random error that occurs in another
element in the array as long as that error occurs after the last T junction and before an
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 375

10.1 Array Errors 375

1 2 3 4 5 6 7 8 FIGURE 10-1
Corporate-fed array
with random errors.

A
B
C

10 FIGURE 10-2
Array factor
5 with random,
uncorrelated errors
0 Errors superimposed on
the error-free array
factor.
Directivity (dB)

−5
Error free
−10

−15

−20

−25

−30
−50 0 50
q (degrees)

element. If a random error occurs prior to A, for instance, then the random error becomes
correlated between the elements that share the error. For instance, a random error between
A and B results in a random correlated error shared by elements 1 and 2. Likewise, a
random error between B and C results in a random correlated error shared by elements 1,
2, 3, and 4.
As an example, consider an eight-element, 20 dB Chebyshev array that has elements
spaced λ/2 apart. If the random errors are represented by δna = 0.15 and δnp = 0.15, then
an example of the array factor with errors is shown in Figure 10-2. Note that the random
errors lower the main beam directivity, induce a slight beam-pointing error, increase the
sidelobe levels, and fill in some of the nulls.

10.1.2 Quantization Errors

Phase shifters and attenuators have Nbp control bits with the least significant bits given by

a = 2−Nba (10.3)
 p = 2π × 2−Nbp (10.4)

If the difference between the desired and quantized amplitude weights is a uniformly dis-
tributed random number with the bounds being√ the maximum amplitude error of ±a /2,
then the rms amplitude error is δna = a / 12. The quantization error is random only
when no two adjacent elements receive the same quantized phase shift. The difference
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 376

376 CHAPTER 10 Compensation of Adaptive Arrays

between the desired and quantized phase shifts is treated as uniform random variables
between ±√ p /2. As with the amplitude error, the random phase error formula in this case
is δnp =  p / 12. Substituting this error into (b) yields the rms sidelobe level.
The phase quantization errors become correlated when the beam steering phase shift
is small enough that groups of adjacent elements have their beam steering phase quantized
to the same level. This means that N /N Q subarrays of N Q elements receive the same
phase shift. The grating lobes due to these subarrays occur at [2]

mλ m (N − 1) 2 Nbp  
sin θm = sin θs ± = sin θs 1 ±  sin θs 1 ± m2 Nbp (10.5)
N Q de N

The approximation in (10.5) assumes that the array has many elements. For large scan
angles, quantization lobes do not form, because the element-to-element phase difference
appears random. The relative peaks of the quantization lobes are given by [1]

√
1 1 − sin θ 2
AFNQ L = Np (10.6)
2 1 − sin θs2

Figure 10-3 shows an array factor with a 20 dB n = 3 Taylor amplitude taper for a
20-element, d = 0.5λ array with its beam steered to θ = 3◦ when the phase shifters
have three bits. Four quantization lobes appear. The quantization lobes decrease when
higher-precision phase shifters are used and when the beam is steered to higher angles.
Significant distortion also results from mutual coupling, variation in group delay
between filters, differences in amplifier gain, tolerance in attenuator accuracy, and aperture
jitter in a digital beamforming array. Aperture jitter is the timing error between samples
in an analog-to-digital (A/D) converter. Without calibration, beamforming or estimation
of the direction of arrival (DOA) of the signal is difficult, as the internal distortion is
uncorrelated with the signal. As a result, the uncorrelated distortion changes the weights
at each element and therefore distorts the array pattern.

FIGURE 10-3 0
Array factor steered
to 3 degrees with
three-bit phase
shifters compared −10
with phase shifters
Array factor (dB)

with infinite
3 bit phase shifter
precision.
−20

−30

−90 −45 0 45 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 377

10.2 Array Calibration 377

10.2 ARRAY CALIBRATION

A phased array needs calibrated before it can generate an optimum coherent beam. Calibra-
tion involves tuning, for example, the phase shifters, attenuators, or receivers to maximize
the gain and to create the desired sidelobe response. Offline calibration takes care of the
static errors and is done at the factory or on deployment. Narrowband calibration is applied
at the center frequency of operation. Broadband calibration is applied over the whole oper-
ating bandwidth of the array. The calibrated phase settings are stored for all beam steering
angles. Temperature causes drift in the component characteristics over time, so the array
requires periodic recalibration. The gain of the radiofrequency (RF) channels must be ac-
curately controlled to avoid nonlinearities arising from saturation of components, because
these nonlinearities cannot be removed.
The top vector in Figure 10-4 shows the resulting uncalibrated array output when the
individual five-element vectors have random amplitude and phase errors. When the array is
calibrated (bottom vector in Figure 10-4), then the individual element vectors are the same
length and align. As a result, the calibrated array output vector magnitude is maximized,
and its phase is zero. Methods for performing array calibration use a calibrated source,
signal injection, or near-field scanning. These approaches are discussed in the following
sections.

10.2.1 Calibrated Source

A known calibration source radiates a calibration signal to all elements in the array [3].
Figure 10-5 shows a calibration source in the far field of an array. At regular intervals, the
main beam is steered to receive the calibration source signal. Alternatively, a multibeam
antenna can devote one beam to calibration. Calibration with near-field sources requires
that distance and angular differences be taken into account. If the calibration source is in
the far field, then the phase shifters are set to steer the beam in the direction of the source.
In either case, each element toggles through all of its phase settings until the output signal
is maximized. The difference between the steering phase and the phase that yields the
maximum signal is the calibration phase.

Uncalibrated array output

Element 1 Element 2 Element 3 Element 4 Element 5

Calibrated array output

FIGURE 10-4 The uncalibrated array output is less than the calibrated array output, because
errors in the uncalibrated array do not allow the signal vectors from the elements to align.

Calibration source FIGURE 10-5

Far-field calibration.

Target
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 378

378 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-6
Layout of the smart
antenna test bed.

Making power measurements for every phase setting at every element in an array
is extremely time-consuming. Calibration techniques that measure both amplitude and
phase of the calibrated signal tend to be much faster. Accurately measuring the signal
phase is reasonable in an anechoic chamber but difficult in the operational environment.
Measurements at four orthogonal phase settings yield sufficient information to obtain
a maximum likelihood estimate of the calibration phase [4]. The element phase error is
calculated from power measurements at the four phase states, and the procedure is repeated
for each element in the array. Additional measurements improve signal-to-noise ratio, and
the procedure can be repeated to achieve desired accuracy within resolution of the phase
shifters, since the algorithm is intrinsically convergent.
Another approach uses amplitude-only measurements from multiple elements to find
the complex field at an element [5]. The first step measures the power output from the
array when the phases of multiple elements are successively shifted with the different
phase intervals. Next, the measured power variation is expanded into a Fourier series to
derive the complex electric field of the corresponding elements. The measurement time
reduction comes at the expense of increased measurement error.
Transmit/receive module calibration is an iterative process that starts with adjusting
the attenuators for uniform gain at the elements [6]. The phase shifters are then adjusted to
compensate for the insertion phase differences at each element. Ideally, when calibrating
the array, the phase shifter’s gain remains constant as the phase settings are varied, but the
attenuator’s insertion phase can vary as a function of the phase setting. This calibration
should be done across the bandwidth, range of operating temperatures, and phase settings.
If the phase shifter’s gain varies as a function of setting, then the attenuators need to be
compensated as well. After iterating over this process, all the calibration settings are saved
and applied at the appropriate times.
Figure 10-6 shows an eight-element uniform circular array (UCA) in which a cen-
ter element radiates a calibration signal to the other elements in the array [7]. Since the
calibration source is in the center of the array, the signal path from the calibration source
to each element is identical. As previously noted, random errors are highly dependent
on temperature [8]. An experimental model of the UCA in Figure 10-6 was placed in-
side a temperature-controlled room and calibrated at 20◦ C. The measured amplitude and
phase errors at three temperatures are shown in Figure 10-7 and Figure 10-8, respectively.
Increasing the temperature of the room to 25◦ C then to 30◦ C without recalibration in-
creases the errors shown in Figure 10-7 and Figure 10-8. This experiment demonstrates
the need of dynamic calibration in a smart antenna array.

10.2.2 Signal Injection

Calibrating with a radiating source is difficult, because the calibration signal transmission/
reception depends on the environment. One technique commonly used in digital
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 379

10.2 Array Calibration 379

FIGURE 10-7
Amplitude error for
0 20° C the UCA antenna
as the system
temperature
changes from 20◦ C
Phase error (degrees)

with calibration to
−20 25° C 25◦ C without
recalibration and to
30° C 30◦ C without
recalibration.
−40

2 3 4 5 6 7 8
Signal path

FIGURE 10-8
Phase error for the
UCA antenna as the
system temperature
1.0 changes from 20◦ C
20° C
Amplitude error (Vrms)

with calibration to
25° C 25◦ C without
recalibration and
0.8 to 30◦ C without
30° C recalibration.

0.6

2 3 4 5 6 7 8
Signal path

beamforming arrays is injecting a calibration signal into the signal path of each element
in the array behind each element as shown in Figure 10-9 [9]. This technique provides
a high-quality calibration signal for the circuitry behind the element. Unfortunately, it
does not calibrate for the element patterns that have significant variations due to mutual
coupling, edge effects, and multipath.

10.2.3 Near-Field Scan

A planar near-field scanner positioned very close to the array moves a probe directly in
front of each element to measure the amplitude and phase of all the elements [10]. The
measured field is transformed back to the aperture to recreate the field radiated at each
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 380

380 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-9 Calibration

Inserting a signal
calibration signal
into the signal paths Receiver A/D
in a digital

Array elements
beamformer.

Computer
Receiver A/D

Receiver A/D

FIGURE 10-10 Alignment results (measured phase deviation from desired value).
a: Unaligned. b: After single alignment with uncorrected measurements. c: After alignment
with fully corrected measurements. From W. T. Patton and L. H. Yorinks, “Near-field alignment
of phased-array antennas,” IEEE Transactions on Antennas and Propagation, Vol. 47, No. 3,
March 1999, pp. 584–591.

element. The calibration algorithm iterates between the measured phase and the array
weights until the phase at all the elements is the same. Figure 10-10 shows the progression
of the phase correction algorithm from left to right. The picture on the left is uncalibrated,
the center picture is after one iteration, and the picture on the right is after calibration
is completed. This techniques is exceptionally good at correcting static errors prior to
deploying an antenna is not practical for dynamic errors.

10.3 BROADBAND SIGNAL PROCESSING

CONSIDERATIONS
Broadband arrays use tapped delay lines that have frequency-dependent transfer functions.
Array performance is a function of the number of taps, the tap spacing, and the total delay in
each channel. The minimum number of taps required to obtain satisfactory performance
for a given bandwidth may be determined as discussed in Section 2.5. The discussion
of broadband signal processing considerations given here follows the treatment of this
subject given by Rodgers and Compton [11–13]. The ideal (distortionless) channel transfer
functions are derived; adaptive array performance using quadrature hybrid processing and
two-, three-, and five-tap delay line processing are considered; and results and conclusions
for broadband signal processing are then discussed.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 381

10.3 Broadband Signal Processing Considerations 381

10.3.1 Distortionless Channel Transfer Functions

The element channels of the two-element array in Figure 10-11 are represented by the
transfer functions H1 (ω) and H2 (ω). Let the desired signal arrive at θs , measured relative
to the array face normal. The array carrier frequency is ω0 , and the point sources spacing
is d = λ0 /2 = π b/ω0 , where  is the wavefront propagation velocity.
From the point of view of the desired signal, the overall transfer function encountered
in passing through the array of Figure 10-11 is
 
ωd
Hd (ω) = H1 (ω) + H2 (ω) exp − j sin θs (10.7)

and the overall transfer function seen by the interference signal is
 
ωd
HI (ω) = H1 (ω) + H2 (ω) exp − j sin θi (10.8)

Now require that
Hd (ω) = exp(− jωT1 ) (10.9)
and
HI (ω) = 0 (10.10)
By choosing Hd (ω) according to (10.9), the desired signal is permitted to experience a
time delay T1 in passing through the array but otherwise remains undistorted. Choosing
HI (ω) = 0 results in complete suppression of the interference signal from the array output.

Signal FIGURE 10-11

Two-element array.
qi Interference
qs

qs
sin
d

qi
in
ds

H2 (w) H1(w)

Array
output
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 382

382 CHAPTER 10 Compensation of Adaptive Arrays

To determine whether it is possible to select H1 (ω) and H2 (ω) to satisfy (10.9) and (10.10),
solve (10.9) and (10.10) for H1 (ω) and H2 (ω). Setting H1 (ω) = |H1 (ω)| exp[ jα1 (ω)] and
H2 (ω) = |H2 (ω)| exp[ jα2 (ω)] results in
  
πω
|H1 (ω)| exp[ jα1 (ω)] + |H2 (ω)| exp j α2 (ω) − sin θs = exp(− jωT1 ) (10.11)
ω0
  
πω
|H1 (ω)| exp[ jα1 (ω)] + |H2 (ω)| exp j α2 (ω) − sin θi =0 (10.12)
ω0
To satisfy (10.9) and (10.10), it follows from (10.11) and (10.12) (as shown by the devel-
opment outlined in the Problems section) that
1
H1 (ω) = H2 (ω) =     (10.13)
2 1 − cos πω ω0
(sin θi − sin θs )
 
π ω π
α2 (ω) = [sin θs + sin θi ] ∓ n − ωT1 (10.14)
2 ω0 2
 
π ω π
α1 (ω) = [sin θs − sin θi ] ± n − ωT1 (10.15)
2 ω0 2
where n is any odd integer. This result means that the amplitude of the ideal transfer
functions are equal and frequency dependent. Equations (10.14) and (10.15) furthermore
show that the phase of each filter is a linear function of frequency with the slope dependent
on the spatial arrival angles of the signals as well as on the time delay T1 of the desired
signal.
Plots of the amplitude function in (10.13) are shown in Figure 10-12 for two choices
of arrival angles (θs = 0◦ and θs = 80◦ ), where it is seen that the amplitude of the
distortionless transfer function is nearly flat over a 40% bandwidth when the desired signal
is at broadside (θs = 0◦ ) and the interference signal is 90◦ from broadside (θi = 90◦ ).
Examination of (10.13) shows that whenever (sin θ I −sin θs ) is in the neighborhood of ±1,

FIGURE 10-12 100

Distortionless
transfer function 90
amplitude versus θs = 80°
80
normalized θi = 90°
frequency for 70
d = λ0 /2. From
Rodgers and 60
Compton, Technical
| H(w) |

Bandwidth
Report ESL 3832-3, 50
40%
1975 [12]. 40

30

20

10 θs = 0°
θi = 90°
0
0 0.5 1 1.5 2
w/w 0
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 383

10.3 Broadband Signal Processing Considerations 383

then the resulting amplitude function will be nearly flat over the 40% bandwidth region.
If, however, both the desired and interference signals are far from broadside (as when
θd = 80◦ and θi = 90◦ ), then the amplitude function is no longer flat.
The degree of “flatness” of the distortionless filter amplitude function is interpreted in
terms of the signal geometry with respect to the array sensitivity pattern. In general, when
the phases of H1 (ω) and H2 (ω) are adjusted to yield the maximum undistorted response to
the desired signal, the corresponding array sensitivity pattern will have certain nulls. The
distortionless filter amplitude function is then the most flat when the interference signal
falls into one of these pattern nulls.
Equation (10.13) furthermore shows that singularities occur in the distortionless chan-
nel transfer functions whenever (ω/ω0 )π(sin θi − sin θs ) = n2π where n = 0, 1, 2, . . ..
The case when n = 0 occurs when the desired and interference signals arrive from exactly
the same direction, so it is hardly surprising that the array would experience difficulty
trying to receive one signal while nulling the other in this case. The other cases when
n = 1, 2, . . ., occur when the signals arrive from different directions, but the phase shifts
between elements differ by a multiple of 2π radians at some frequency ω in the signal
band.
The phase functions α1 (ω) and α2 (ω) of (10.14) and (10.15) are linear functions of
frequency. When T1 = 0, the phase slope of H1 (ω) is proportional to sin θs − sin θi ,
whereas that of H2 (ω) is proportional to sin θi + sin θs . Consequently, when the desired
signal is broadside, α1 (ω) = −α2 (ω). Furthermore, the phase difference between α1 (ω)
and α2 (ω) is also a linear function of frequency, a result that would be expected since this
allows the interelement phase shift (which is also a linear function of frequency) to be
canceled.

10.3.2 Quadrature Hybrid and Tapped Delay Line Processing

for a Least Mean Squares Array
Consider a two-element adaptive array using the least mean squares (LMS) algorithm. If
w is the column vector of array weights, Rx x is the correlation matrix of input signals to
each adaptive weight, and rxd is the cross-correlation vector between the received signal
vector x(t) and the reference signal d(t), then as shown in Chapter 3 the optimum array
weight vector that minimizes E{ε2 (t)} (where ε(t) = d(t)—array output) is given by

wopt = R−1
x x rxd (10.16)

If the signal appearing at the output of each sensor element consists of a desired signal, an
interference signal, and a thermal noise component (where each component is statistically
independent of the others and has zero mean), then the elements of Rx x can readily be
evaluated in terms of these component signals.
Consider the tapped delay line employing real (instead of complex) weights shown
in Figure 10-13. Since each signal xi (t) is just a time-delayed version of x1 (t), it follows
that
⎫
x2 (t) = x1 (t − ) ⎪
⎪
⎪
⎬
x2 (t) = x1 (t − 2)
.. ⎪
(10.17)
. ⎪
⎪
⎭
x L (t) = x1 [t − (L − 1)]
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 384

384 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-13 Sensor

Tapped delay line element
processor for a
single-element
x1 (t) x2 (t) x3 (t) x4 (t) xL (t)
channel having real Δ Δ Δ Δ
adaptive weights.

w1 w2 w3 w4 wL

Channel
output

Now since the elements of Rx x are given by


r xi x j = E{xi (t)x j (t)} (10.18)
it follows from (10.17) that
r xi x j = r x1 x1 (τij ) (10.19)
where r x1 x1 (τij ) is the autocorrelation function of x1 (t), and τij is the time delay between
xi (t) and x j (t). Furthermore, r xi xi (τij ) is the sum of three autocorrelation functions—those
of the desired signal, the interference, and the thermal noise so that
r x1 x1 (τij ) = rdd (τij ) + rII (τij ) + rnn (τij ) (10.20)
For the elements of Rxx corresponding to xi (t) and x j (t) from different element channels,
r xi x j consists only of the sum of the autocorrelation functions of the desired signal and the
interference signal (with appropriate delays) but not the thermal noise since the element
noise from channel to channel is uncorrelated. Thus, for signals in different element
channels
r xi x j (τij ) = rdd (τdij ) + rII (τ Iij ) (10.21)
where τdij denotes the time delay between xi (t) and x j (t) for the desired signal, and τ Iij
denotes the time delay between xi (t) and x j (t) for the interference signal (these two time
delays will in general be different due to the different angles of arrival of the two signals).
Only when xi (t) and x j (t) are from the same array element channel will τdij = τ Iij (which
may then be denoted by τi j ).
Next, consider the quadrature hybrid array processor depicted in Figure 10-14. Let
x1 (t) and x3 (t) denote the in-phase signal components and x2 (t) and x4 (t) denote the
quadrature-phase signal components of each of the elements output signals. Then the
in-phase and quadrature components are related by

x2 (t) = x̌1 (t)
(10.22)
x4 (t) = x̌3 (t)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 385

10.3 Broadband Signal Processing Considerations 385

FIGURE 10-14
Quadrature hybrid
processing for a
two-element array.
Quadrature Quadrature
hybrid hybrid

x1 (t) x2 (t) x3 (t) x4 (t)

w1 w2 w3 w4

Array output

The symbol ˇ denotes the Hilbert transform

 ∞
 1 x(τ )
x̌(t) = dτ (10.23)
π −∞ t − τ
where the previous integral is regarded as a Cauchy principal value integral. The various
elements of the correlation matrix

r xi x j = E{xi (t)x j (t)} (10.24)

can then be found by making use of certain Hilbert transform relations as follows [14,15]:

E{x̌(t) y̌(s)} = E{x(t)y(s)} (10.25)

E{x̌(t)y(s)} = −E{x(t) y̌(s)} (10.26)

so that

E{x̌(t)x(t)} = 0 (10.27)
E{x(t) y̌(s)} = Ě{x(t)y(s)} (10.28)

where Ě{x(t)y(s)} denotes the Hilbert transform of r x y (τ ) where τ = s − t. With the

previous relations and from (10.22) it then follows that

r x1 x1 = E{x1 (t)x1 (t)} = r x1 x1 (0) (10.29)

r x1 x2 = E{x1 (t)x2 (t)} = E{x1 (t)x̌1 (t)} = 0 (10.30)
r x2 x2 = E{x2 (t)x2 (t)} = E{x̌1 (t)x̌1 (x)} (10.31)
= E{x1 (t)x1 (t)} = r x1 x1 (0)

where r x1 x1 (τ ) is the autocorrelation function of x1 (t) given by (10.20).

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 386

386 CHAPTER 10 Compensation of Adaptive Arrays

When two different sensor element channels are involved [as with x1 (t) and x3 (t), for
example], then
E{x1 (t) x3 (t)} = rdd (τd13 ) + rII (τ I13 ) (10.32)
where τd13 and τ I13 represent the spatial time delays between the sensor elements of Fig-
ure 10-14 for the desired and interference signals, respectively. Similarly

E{x1 (t)x4 (t)} = E{x1 (t)x̌3 (t)} = Ě{x1 (t)x3 (t)}

= řdd (τd13 ) + řII (τ I13 ) (10.33)
E{x2 (t)x3 (t)} = E{x̌1 (t)x3 (t)} = −E{x̌1 (t)x3 (t)}
= − Ě{x1 (t)x3 (t)} = −řdd (τd13 ) − řII (τ I13 ) (10.34)
E{x2 (t)x4 (t)} = E{x̌1 (t)x̌3 (t)} = x{x1 (t)x3 (t)}
= rdd (τd13 ) + rII (τ I13 ) (10.35)

Now consider the cross-correlation vector rxd defined by

⎡ ⎤
x1 (t)d(t)
⎢ x 2 (t)d(t) ⎥
 ⎢ ⎥
rxd = E ⎢ . ⎥ (10.36)
⎣ .. ⎦
x2N (t)d(t)
where N is the number of sensor elements. Each element of rxd , denoted by r xi d , is
just the cross-correlation between the reference signal d(t) and signal xi (t). Since the
reference signal is just a replica of the desired signal and is statistically independent
of the interference and thermal noise signals, the elements of rxd consist only of the
autocorrelation function of the desired signal so that
r xi d = E{xi (t)d(t)} = rdd (τdi ) (10.37)
where τdi represents the time delay between the reference signal and the desired signal
component of xi (t). For an array with tapped delay line processing, each element of rxd
is the autocorrelation function of the desired signal evaluated at a time-delay value that
reflects both the spatial delay between sensor elements and the delay line delay to the tap of
interest. For an array with quadrature hybrid processing, the elements of rxd corresponding
to an in-phase channel yield the autocorrelation function of the desired signal evaluated
at the spatial delay appropriate for that element as follows:

r xi d (in-phase channel) = E{xi (t)d(t)} = rdd (τdi ) (10.38)

The elements of rxd corresponding to quadrature-phase channels can be evaluated using

(10.27) and (10.28) as follows:

r xi+1 d (quadrature-phase channel) = E{xi+1 (t)d(t)}

= E{x̌i (t)d(t)} = −E{xi (t)ď(t)} (10.39)
= − Ě{xi (t)d(t)} = −ř xi d (τdi )

Once Rxx and rxd have been evaluated for a given signal environment, the optimal LMS
weights can be computed from (10.16), and the steady-state response of the entire array
can then be evaluated.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 387

10.3 Broadband Signal Processing Considerations 387

The tapped delay line in the element channel of Figure 10-13 has a channel transfer
function given by

H1 (ω) = w 1 + w 2 e− jω + w 3 e− j2ω + . . . + w L e− j (L−1)ω (10.40)

Likewise, the quadrature hybrid processor of Figure 10-14 has a channel transfer function

H1 (ω) = w 1 − jw 2 (10.41)

The array transfer function for the desired signal and the interference accounts for the
effects of spatial delays between array elements. A two-element array transfer function
for the desired signal is

Hd (ω) = H1 (ω) + H2 (ω)e− jωτd (10.42)

whereas the transfer function for the interference is

HI (ω) = H1 (ω) + H2 (ω) e− jωτ I (10.43)

The spatial time delays associated with the desired and interference signals are represented
by τd and τ I , respectively, between element 1 [with channel transfer function H1 (ω)] and
element 2 [with channel transfer function H2 (ω)]. With two sensor elements spaced apart
by a distance d as in Figure 10-11, the two spatial time delays are given by
d
τd = sin θs (10.44)

d
τ I = sin θ I (10.45)

The output signal-to-total-noise ratio is defined as

 Pd
SNR = (10.46)
PI + Pn
where Pd , PI , and Pn represent the output desired signal power, interference signal power,
and thermal noise power, respectively. The array output power for each of the foregoing
three signals may now be evaluated. Let φdd (ω) and φII (ω) represent the power spectral
densities of the desired signal and the interference signal, respectively; then the desired
signal output power is given by
 ∞
Pd = φdd (ω)|Hd (ω)|2 dω (10.47)
−∞

where Hd (ω) is the overall transfer function seen by the desired signal, and the interference
signal output power is
 ∞
PI = φII (ω)|H1 (ω)|2 dω (10.48)
−∞

where HI (ω) is the overall transfer function seen by the interference signal. The thermal
noise present in each element output is statistically independent from one element to the
next. Let φnn (ω) denote the thermal noise power spectral density; then the noise power
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 388

388 CHAPTER 10 Compensation of Adaptive Arrays

contributed to the array output by element 1 is

 ∞
Pn 1 = φnn (ω)|H1 (ω)|2 dω (10.49)
−∞

whereas that contributed by element 2 is

 ∞
Pn 2 = φnn (ω)|H2 (ω)|2 dω (10.50)
−∞

Consequently, the total thermal noise output power from a two-element array is
 ∞
Pn = φnn (ω)[|H1 (ω)|2 + |H2 (ω)|2 ] dω (10.51)
−∞

The foregoing expressions may now be used in (10.46) to obtain the output signal-to-total-
noise ratio.

10.3.3 Performance Comparison of Four Array Processors

In this subsection, four adaptive arrays—one with quadrature hybrid processing and three
with tapped delay line processing (using real weights)—are compared for signal band-
widths of 4, 10, 20, and 40%. Tapped delay lines use real weights to preserve as much
simplicity as possible in the hardware implementation, although this sacrifices the avail-
able degrees of freedom with a consequent degradation in tapped delay line performance
relative to combined amplitude and phase weighting. The results obtained will neverthe-
less serve as an indication of the relative effectiveness of tapped delay line processing
compared with quadrature hybrid processing for broadband signals.
The four array processors to be compared are shown in Figure 10-15, where each
array has two sensor elements and the elements are spaced one-half wavelength apart at
the center frequency of the desired signal bandwidth. Figure 10-15a shows an array having
quadrature hybrid processing, whereas Figure 10-15b–10-15d exhibit tapped delay line
processing. The processor of Figure 10-15b has one delay element corresponding to one-
quarter wavelength at the center frequency and two associated taps. The processor of
Figure 10-15c has two delay elements, each corresponding to one-quarter wavelength at
the center frequency, and three associated taps. The processor of Figure 10-15d has four
delay elements, each corresponding to one-eighth wavelength at the center frequency,
and five associated taps. Note that the total delay present in the tapped delay line of
Figure 10-15d is the same as that of Figure 10-15c, so the processor in Figure 10-15d may
be regarded as a more finely subdivided version of the processor in Figure 10-15c.
Assume that the desired signal is biphase modulated of the form

sd (t) = A cos[ω0 t + φ(t) + θ ] (10.52)

where φ(t) denotes a phase angle that is either zero or π over each bit interval, and θ is
an arbitrary constant phase angle (within the range [0, 2π ]) for the duration of any signal
pulse. The nth bit interval is defined over T0 + (n − 1)T ≤ t ≤ T0 + nT , where n is any
integer, T is the bit duration, and T0 is a constant that determines where the bit transitions
occur, as shown in Figure 10-16.
Assume that φ(t) is statistically independent over different bit intervals and is zero
or π with equal probability and that T0 is uniformly distributed over one bit interval;
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 389

10.3 Broadband Signal Processing Considerations 389

x5 x1
w3 w1
−90° 0° −90° 0°
QH QH
x4 x3 x2 x1 λ0
Δ Δ
4

w1 w2 w3 w4 w4 w2
x4 x2

Σ Σ

(a) (b)

x6 x1
w6 w1

λ0
Δ Δ
8
x7 x2
w7 w2
x4 x1
w4 w1 λ0
Δ Δ
8
x8 x3
λ0 w8 w3
Δ Δ
4
x5 x2 λ0
w5 w2 Δ Δ
8
λ0 x9 x4
Δ Δ w9 w4
4
λ0
x6 x3 Δ Δ
w6 w3 8
x10 x5
w10 w5

Σ Σ

(c) (d)

FIGURE 10-15 Four adaptive array processors for broadband signal processing
comparison. a: Quadrature hybrid. b: Two-tap delay line. c: Three-tap delay line. d: Five-tap
delay line. From Rodgers and Compton, IEEE Trans. Aerosp. Electron. Syst., January 1979 [13].

then, sd (t) is a stationary random process with power spectral density given by
 
A2 T sin(T /2)(ω − ω0 ) 2
φdd (ω) = (10.53)
2 (T /2) (ω − ω0 )
This power spectral density is shown in Figure 10-17.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 390

390 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-16 f (t)

Bit transitions for
biphase modulated T T T T T
signal. p

t
T0

FIGURE 10-17 Sd(w)

Desired signal power
spectral density.

ω
w0 − w1 w0 w0 + w1

The reference signal equals the desired signal component of x1 (t) and is time aligned
with the desired component of x2 (t). The desired signal “bandwidth” will be taken to be
the frequency range defined by the first nulls of the spectrum given by (10.53). With this
definition, the fractional bandwidth then becomes
2ω1
desired signal bandwidth = (10.54)
ω0
where ω1 is the frequency separation between the center frequency ω0 and the first null
2π
ω1 = (10.55)
T
Assume that the interference signal is a Gaussian random process with a flat, bandlim-
ited power spectral density over the range ω0 − ω1 < ω < ω0 + ω1 ; then the interference
signal spectrum appears in Figure 10-18. Finally, the thermal noise signals present at
each element are statistically independent between elements, having a flat, bandlimited,
Gaussian spectral density over the range ω0 − ω1 < ω < ω0 + ω1 (identical with the
interference spectrum of Figure 10-18).
With the foregoing definitions of signal spectra, the integrals of (10.48) and (10.51)
yielding interference and thermal noise power are taken only over the frequency range ω0 −
ω1 < ω < ω0 + ω1 . The desired signal power also is considered only over the frequency
range ω0 − ω1 < ω < ω0 + ω1 to obtain a consistent definition of signal-to-noise ratio
(SNR). Therefore, the integral of (10.47) is carried out only over ω0 − ω1 < ω < ω0 + ω1 .

FIGURE 10-18 SI (w)

Interference signal
power spectral
density.

ω
w0 − w1 w0 w0 + w1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 391

10.3 Broadband Signal Processing Considerations 391

To compare the four adaptive array processors of Figure 10-15, the output SNR
performance is evaluated for the aforementioned signal conditions. Assume the element
thermal noise power pn is 10 dB below the element desired signal power ps so that ps / pn =
10 dB. Furthermore, suppose that the element interference signal power pi is 20 dB stronger
than the element desired signal power so that ps / pi = −20 dB. Now assume that the
desired signal is incident on the array from broadside. The output SNR given by (10.46) can
be evaluated from (10.47), (10.48), and (10.49) by assuming the processor weights satisfy
(10.16) for each of the four processor configurations. The resulting output signal-to-total
noise ratio that results using each processor is plotted in Figures 10-19–10-22 as a function
of the interference angle of arrival for 4, 10, 20, and 40% bandwidth signals, respectively.
In all cases, regardless of the signal bandwidth, when the interference approaches
broadside (near the desired signal) the SNR degrades rapidly, and the performance of

15 FIGURE 10-19
Pi Output signal-to-
= 20 dB
Ps 3 & 5 Taps interference plus
Pn noise ratio
= −10 dB
10 Ps interference angle
for four adaptive
processors with 4%
3 Taps bandwidth signal.
(dB)

5 5 Taps From Rodgers and

P I + PN

2 Taps Compton, IEEE

Pd

Quadrature hybrid Trans. Aerosp.

Electron. Syst.,
Output

0 January 1979 [13].

20° 40° 60° 80°
Interference angle

−5

−10

15 FIGURE 10-20
Pi Output signal-to-
= 20 dB
Ps 3 & 5 Taps interference plus
Pn noise ratio versus
= −10 dB
10 Ps interference angle
for four adaptive
(dB)

processors with
10% bandwidth
PI + PN

5 3 Taps signal. From

Pd

5 Taps Rodgers and

2 Taps
Output

Compton, IEEE
Quadrature hybrid Trans. Aerosp.
0 Electron. Syst.,
20° 40° 60° 80°
January 1979 [13].
Interference angle

−5
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 392

392 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-21 15
Output signal-to- Pi
= 20 dB 3 & 5 Taps
interference plus Ps
noise ratio versus Pn
= −10 dB
interference angle 10 Ps
3 Taps
for four adaptive
5 Taps
processors with
2 Taps
20% bandwidth

(dB)
Quadrature hybrid
signal. From 5

PI + PN
Rodgers and
Compton, IEEE Pd
Trans. Aerosp. Output
Electron. Syst., 0
20° 40° 60° 80°
January 1979 [13].
Interference angle

−5

−10

FIGURE 10-22 15
Output signal-to- Pi
= 20 dB
interference plus Ps
noise ratio versus Pn
= −10 dB
interference angle 10 Ps
3 Taps
for four adaptive
5 Taps
processors with
2 Taps
40% bandwidth
(dB)

Quadrature hybrid
signal. From 5
PI + P N

Rodgers and
Pd

Compton, IEEE
Trans. Aerosp.
Output

Electron. Syst., 0
20° 40° 60° 80°
January 1979 [13].
Interference angle

−5

−10

all four processors becomes identical. This SNR degradation is expected since, when the
interference approaches the desired signal, the desired signal falls into the null provided
to cancel the interference, and the output SNR consequently falls. Furthermore, as the
interference approaches broadside, the interelement phase shift for this signal approaches
zero. Consequently, the need to provide a frequency-dependent phase shift behind each
array element to deal with the interference signal is less, and the performance of all four
processors becomes identical.
When the interference signal is widely separated from the desired signal, then the
output SNR is different for the four processors being considered, and this difference
becomes more pronounced as the bandwidth increases. For 20 and 40% bandwidth signals,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 393

10.3 Broadband Signal Processing Considerations 393

for example, neither the quadrature hybrid processor nor the two-tap delay line proces-
sor provides good performance as the interference signal approaches endfire. The per-
formance of both the three- and five-tap delay line processors remains quite good in the
endfire region, however. If 20% or more bandwidth signals are accommodated, then tapped
delay line processing becomes a necessity. Figure 10-22 shows that there is no significant
performance advantage provided by the five-tap processor compared with the three-tap
processor, so a three-tap processor is adequate for up to 40% bandwidth signals in the case
of a two-element array.
Figures 10-21 and 10-22 show that the output SNR performance of the two-tap
delay line processor peaks when the interference signal is 30◦ off broadside, because
the interelement delay time is λ/4 (since the elements are spaced apart by λ/2). Conse-
quently, the single-delay element value of λ/4 provides just the right amount of time delay
to compensate exactly for the interelement time delay and to produce an improvement in
the output SNR.
The three-tap and five-tap delay line processors both produce a maximum SNR of
about 12.5 dB at wide interference angles of 70◦ or greater. For ideal channel processing,
the interference signal is eliminated, the desired signal in each channel is added coherently
to produce Pd = 4 ps , and the thermal noise is added noncoherently to yield PN = 2 pn .
Thus, the best possible theoretical output SNR for a two-element array with thermal noise
10 dB below the desired signal and no interference is 13 dB. Therefore, the three-tap and
five-tap delay line processors are successfully rejecting nearly all the interference signal
power at wide off-boresight angles.

10.3.4 Processor Transfer Functions

Ideally, the array transfer function for the desired signal should be constant across the
desired signal bandwidth, thereby preventing desired signal distortion. The interference
transfer function should be a low array response over the interference bandwidth.
The transfer functions for the four processors and the two-element array are evalu-
ated using (10.40)–(10.45). Using the same conditions adopted in computing the SNR
performance, Figures 10-23–10-26 show |Hd (ω)| and |HI (ω)| for the four processors of

90° FIGURE 10-23

0
10°
Quadrature hybrid
Desired signal
−15 transfer functions at
Interference signal Interference angle 4% bandwidth.
−30 From Rodgers and
90°
Compton, Technical
−45
20° Report ESL 3832-3,
Amplitude (dB)

−60 1975 [12].

10°
−75

−90

−105

−120

−135
0.98 0.984 0.988 0.992 0.996 1.001 1.004 1.008 1.012 1.016 1.02
Frequency (w/w 0)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 394

394 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-24 0 90°

Two-tap delay line 10°
transfer functions −15 Desired signal Interference angle
at 4% bandwidth. Interference signal
−30 90°
From Rodgers and 20°
10°
Compton, Technical −45

Amplitude (dB)
Report ESL 3832-3,
1975 [12]. −60

−75

−90

−105

−120

−135
0.98 0.984 0.988 0.992 0.996 1.001 1.004 1.008 1.012 1.016 1.02
Frequency (w/w 0)

Figure 10-15 with a 4% signal bandwidth and various interference signal angles. The
results shown in these figures indicate that for all four processors and for all interference
angles the desired signal response is quite flat over the signal bandwidth. As the interfer-
ence approaches the desired signal angle at broadside, however, the (constant) response
level of the array to the desired signal drops because of the desired signal partially falling
within the array pattern interference null.
The results in Figure 10-23 for quadrature hybrid processing show that the array
response to the interference signal has a deep notch at the center frequency when the
interference signal is well separated (θi > 20◦ ) from the desired signal. As the interference
signal approaches the desired signal (θi < 20◦ ), the notch migrates away from the center
frequency, because the processor weights must compromise between rejection of the
interference signal and enhancement of the desired signal when the two signals are close.
Migration of the notch improves the desired signal response (since the desired signal power
spectral density peaks at the center frequency) while affecting interference rejection only
slightly (since the interference signal power spectral density is constant over the signal
band).
The array response for the two-tap processor is shown in Figure 10-24. The response
to both the desired and interference signals is very similar to that obtained for quadrature
hybrid processing. The most notable change is the slightly different shape of the transfer
function notch presented to the interference signal by the two-tap delay line processor
compared with the quadrature hybrid processor.
Figure 10-25 shows the three-tap processor array response. The interference signal
response is considerably reduced, with a minimum rejection of the interference signal
of about 45 dB. When the interference signal is close to the desired signal, the array
response has a single mild dip. As the separation angle between the interference signal
and the desired signal increases, the single dip becomes more pronounced and finally
develops into a double dip at very wide angles. It is difficult to attribute much significance
to the double-dip behavior since it occurs at such a low response level (of more than
75 dB attenuation). The five-tap processor response of Figure 10-26 is very similar to the
three-tap processor response except slightly more interference signal rejection is achieved.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 395

10.3 Broadband Signal Processing Considerations 395

90° FIGURE 10-25

0
Three-tap delay line
Desired signal 10°
−15 transfer functions
Interference angle
at 4% bandwidth.
−30 From Rodgers and
Interference signal 10° Compton, Technical
−45
20° Report ESL 3832-3,
Amplitude (dB)

−60 70° 1975 [12].

−75 90°

−90

−105

−120

−135
0.98 0.984 0.988 0.992 0.996 1.001 1.004 1.008 1.012 1.016 1.02
Frequency (w/w 0)

90° FIGURE 10-26

0
Five-tap delay line
Desired signal 10°
−15 transfer functions
at 4% bandwidth.
Interference angle
−30 From Rodgers and
Interference signal 10° Compton, Technical
−40
Report ESL 3832-3,
Amplitude (dB)

20°
−60 1975 [12].
50°
−75
90°
−90

−105

−120

−135
0.98 0.984 0.988 0.992 0.996 1.001 1.004 1.008 1.012 1.016 1.02
Frequency (w/w 0)

As the signal bandwidth increases, the processor response curves remain essentially
the same as in Figures 10-23–10-26 except the following:
1. As the interference signal bandwidth increases, it becomes more difficult to reject the
interference signal over the entire bandwidth, so the minimum rejection level increases.
2. The desired signal response decreases because the array feedback reduces all weights
to compensate for the presence of a greater interference signal component at the array
output, thereby resulting in greater desired signal attenuation.
The net result is that as the signal bandwidth increases, the output SNR performance
degrades, as confirmed by the results of Figures 10-19–10-22.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 396

396 CHAPTER 10 Compensation of Adaptive Arrays

10.4 COMPENSATION FOR MUTUAL COUPLING

In many applications, the limited space available for mounting an antenna motivates the
use of a small array. As the array size decreases, the array element spacing becomes less
than a half-wavelength, and mutual coupling effects become more of a factor in degrad-
ing the array performance. When an array consists of single-mode elements (meaning that
the element aperture currents may change in amplitude but not in shape as a function of
the signal angle of arrival), then it is possible to modify the element weights to compensate
for the pattern distortion caused by the mutual coupling at a particular angle [16]. These
weight adjustments may work for more than one angle.
Let the vector v denote the coupling perturbed measured voltages appearing at the
output of the array elements, and let vd represent the coupling unperturbed voltages that
would appear at the array element outputs if no mutual coupling were present. The effect
of mutual coupling on single-mode elements is written as

v(u) = C vd (u) (10.56)

where u = sin θ, θ is the angle of arrival, and the matrix C describes the effects of mutual
coupling and is independent of the signal scan angle. If the array is composed of multimode
elements, then the matrix C would be scan angle dependent.
It follows that the unperturbed signal vector, vd can be recovered from the perturbed
signal vector by introducing compensation for the mutual coupling

vd = C−1 v (10.57)

Introducing the compensation network C−1 as shown in Figure 10-27 then allows all
subsequent beamforming operations to be performed with ideal (unperturbed) element
signals, as are customarily assumed in pattern synthesis.
This mutual coupling compensation is applied to an eight-element linear array having
element spacing d = 0.517 λ consisting of identical elements. Figure 10-28(a) shows the
effects of mutual coupling by displaying the difference in element pattern shape between
a central and an edge element in the array.
Figure 10-28 displays a synthesized 30 dB Chebyshev pattern both without (a) and with
(b) mutual coupling compensation. It is apparent from this result that the compensation
network gives about a 10 dB improvement in the sidelobe level.

FIGURE 10-27 Coupling

Coupling Array perturbed Coupling Unperturbed Complex
elements measured compensation voltages weights
Compensation and
voltages
Beamforming in an W1
Array Antenna. From 1 υ1 υd1
Steyskal & Herd,
IEEE Trans. Ant & W2 Output
2 C υ2 C −1 υd2 Σ
Prop., Dec. 1995. beam

WN
N υN υNd
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 397

10.4 Compensation for Mutual Coupling 397

0
Measured
Theory

−10

−20
Power, dB

−30

−40

−80 −60 −30 0 30 60 80

Angle, deg
(a)

0
Measured
Theory

−10

−20
Power, dB

−30

−40

−80 −60 −30 0 30 60 80

Angle, deg
(b)

FIGURE 10-28 30 dB Chebyshev pattern (a) without and (b) with Coupling Compensation
with a Scan Angle of 0◦ . From Steyskal & Herd, IEEE Trans. Ant. & Prop. Dec. 1995.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 398

398 CHAPTER 10 Compensation of Adaptive Arrays

10.5 MULTIPATH COMPENSATION

In many operating environments, multipath rays impinge on the array shortly after the
direct path signal arrives at the sensors. Multipath distorts any interference signal that
may appear in the various element channels, thereby severely limiting the interference
cancellation. A tapped delay line processor combines delayed and weighted replicas of the
input signal to form the filtered output signal and thereby has the potential to compensate
for multipath effects, since multipath rays also consist of delayed and weighted replicas
of the direct path ray.

10.5.1 Two-Channel Interference Cancellation Model

Consider an ideal two-element adaptive array with one channel’s (called the “auxiliary”
channel) response adjusted so that any jamming signal entering the other channel through
the sidelobes (termed the “main” channel) is canceled at the array output. A system de-
signed to suppress sidelobe jamming in this manner is called a coherent sidelobe canceller
(CSLC), and Figure 10-29 depicts a two-channel CSLC system in which the auxiliary
channel employs tapped delay line compensation involving L weights and L − 1 delay
elements of value  seconds each. A delay element of value D = (L − 1)/2 is included
in the main channel so the center tap of the auxiliary channel corresponds to the output
of the delay D in the main channel, thereby permitting compensation for both positive
and negative values of the off-broadside angle θ . This ideal two-element CSLC system
model exhibits all the salient characteristics that a more complex system involving several
auxiliary channels would have, so the two-element system serves as a convenient model
for performance evaluation of multipath cancellation [17].
The system performance measure is the ability of the CSLC to cancel an undesired
interference signal through proper design of the tapped delay line. In actual practice, an
adaptive algorithm adjusts the weight settings. To eliminate the effect of algorithm selec-
tion from consideration, only the steady-state performance is evaluated. Since the steady-
state solution can be found analytically, it is necessary to determine only the resulting
solution for the output residue power. This residue power is then a direct measure of the
interference cancellation ability of the two-element CSLC model.

FIGURE 10-29 x0 (t) L −1 Δ Main channel

Ideal two-element D=
2
CSLC model with
auxiliary channel +
q
compensation Residue
involving L weights −
xL (t)
and L − 1 delay x2 (t) Auxiliary channel
elements. x1 (t)
Δ Δ Δ wL Σ

w2

w1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 399

10.5 Multipath Compensation 399

Let x0 (t), x1 (t), and e(t) represent the complex envelope signals of the main channel
input signal, the auxiliary channel input signal, and the output residue signal, respectively.
Define the complex signal vector

xT = [x1 (t), x2 (t), . . . , x L (t)] (10.58)
where

x2 (t) = x1 (t − )
..
.

x L (t) = x1 [t − (L − 1)]

Also, define the complex weight vector


wT = [w 1 , w 2 , . . . , w L ] (10.59)
The output of the tapped delay line may then be expressed as

L
filter output = x1 [t − (i − 1)]w i∗ = w† x(t) (10.60)
i=1

The residue (complex envelope) signal is given by

e(t) = x0 (t − D) + w† x(t) (10.61)
The weight vector w minimizes the residue signal in a mean square error (MSE) sense.
For stationary random processes, this is equivalent to minimizing the expression

Ree (0) = E {e(t)e∗ (t)} (10.62)

From (10.61) and the fact that

E{x0 (t − D)x0∗ (t − D)} = r x0 x0 (0) (10.63)

E{x(t)x0∗ (t − D)} = r x x0 (−D) (10.64)
E{x(t)x† (t)} = Rxx (0) (10.65)

it follows that

Ree (0) = r x0 x0 (0) − r†x x0 (−D)R−1

xx (0)rx x0 (−D)
+ [rx x0 (−D) + w Rxx (0)] • R−1
† †
xx (0) [rx x0 (−D) + Rxx (0)w] (10.66)

Minimize (10.66) by appropriately selecting the complex weight vector w. Assume the
matrix Rxx (0) is nonsingular: the value of w for which this minimum occurs is given by

wopt = R−1
xx (0)rx x0 (−D) (10.67)

The corresponding minimum residue signal power then becomes

Ree (0)min = r x0 x0 (0) − r†x x0 (−D)R−1

xx (0)rx x0 (−D) (10.68)

Interference cancellation performance of the CSLC model of Figure 10-27 is determined

by evaluating (10.66) using selected signal environment assumptions.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 400

400 CHAPTER 10 Compensation of Adaptive Arrays

10.5.2 Signal Environment Assumptions

Let s1 (t, θ1 ) represent the interference signal arriving from direction θ1 , and let
sm (t, ρm , Dm , θm+1 ) for m = 2, . . . , M represent the multipath structure associated with the
interference signal that consists of a collection of M – 1 correlated plane wave signals of
the same frequency arriving from different directions so that θm+k = θ1 and θm+k = θm+l
for k = l. The multipath rays each have an associated reflection coefficient ρm and a time
delay with respect to the direct ray Dm . The structure of the covariance matrix for this
multipath model can then be expressed as [18]
Rss = Vs AV†s (10.69)
where Vs is the N × M signal matrix given by
 
| | |
Vs = s1 s2
v v · · · v (10.70)
| | | sM
whose components are given by the N × 1 vectors
⎡ ⎤
1
⎢ exp [ j2π(d/λ0 ) sin θm ] ⎥
⎢ ⎥
⎢ ⎥
vsm = Psm ⎢ exp [ j2π(d/λ0 )2 sin θm ] ⎥ (10.71)
⎢. ⎥
⎣ .. ⎦
exp [ j2π(d/λ0 ) (N − 1) sin θm ]
where Psm = ρm2 denotes the power associated with the signal sm , and A is the multipath
correlation matrix. When A = I, the various signal components are uncorrelated whereas
for A = U (the M × M matrix of unity elements) the various components are perfectly
correlated. For purposes of numerical evaluation the correlation matrix model may be
selected as [18]
⎡ ⎤
1 α α 2 · · · α M−1
⎢α 1 α · · · α M−1 ⎥
⎢ ⎥
A = ⎢. ⎥ 0≤α≤1 (10.72)
⎣ .. ⎦
α M−1 · · ··· 1
Note that channel-to-channel variations in θm , Dm , and ρm cannot be accommodated by
this simplified model. Consequently, a more general model must be developed to handle
such variations, which tend to occur where near-field scattering effects are significant. The
input signal covariance matrix may be written as
Rxx = Rnn + Vs AV†s (10.73)
where Rnn denotes the noise covariance matrix.
If only a single multipath ray is present, then s(t, θ1 ) denotes the direct interference
signal, and sm (t, ρm , Dm , θ2 ) represents the multipath ray associated with the direct
interference signal. The received signal at the main channel element is then given by
x0 (t) = s(t, θ1 ) + sm (t, ρm , Dm , θ2 ) (10.74)
Denote s(t, θ1 ) by s(t); then sm (t, ρm , Dm , θ2 ) can be written as ρm s(t − Dm ) ×
exp(− jω0 Dm ) so that

x0 (t) = s(t) + ρm s(t − Dm ) exp(− jω0 Dm ) (10.75)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 401

10.5 Multipath Compensation 401

where ω0 is the center frequency of the interference signal. It then follows that

x1 (t) = s(t − τ12 ) exp(− jω0 τ12 )

+ ρm s(t − Dm − τ22 ) exp [− jω0 (Dm + τ22 )] (10.76)

where τ12 and τ22 represent the propagation delay between the main channel element
and the auxiliary channel element for the wavefronts of s(t, θ1 ) and sm (t, ρm , dm , θ2 ),
respectively.
Assuming the signals s(t, θ1 ) and sm (t, ρm , Dm , θ2 ) possess flat spectral density
functions over the bandwidth B, as shown in Figure 10-30a, then the corresponding
auto- and cross-correlation functions of x0 (t) and x1 (t) can be evaluated by recognizing
that

Rxx (τ ) = −1 {xx (ω)} (10.77)

where −1 {·} is the “inverse Fourier transform,” and xx (ω) denotes the cross-spectral
density matrix of x(t).
From (10.74), (10.76), and (10.77) it immediately follows that
sin π BDm  
r x0 x0 (0) = 1 + |ρm |2 + ρm e− jω0 Dm + ρm∗ e jω0 Dm (10.78)
π BDm
sin π B [ψ + sgn1 · (i − 1) + sgn2 · D]

Likewise, defining f [ψ, sgn1, sgn2] =
π B [ψ + sgn1 · (i − 1) + sgn2 · D]
 sin π B [ψ + sgn · (i − k)]
and g[ψ, sgn] = , then
π B[ψ + sgn · (i − k)]
r xi x0 (−D) = f [τ12 , +, −] exp{− jω0 [τ12 + (i − 1)]}
+ f [Dm + τ22 , +, −]ρm exp{− jω0 [τ22 + (i − 1) + Dm ]} (10.79)
∗
+ f [Dm − τ12 , −, +]ρm exp{− jω0 [τ12 + (i − 1) − Dm ]}
+ f [τ22 , +, −]|ρm |2 exp{− jω0 [τ22 + (i − 1)]}

Φxx (w ) FIGURE 10-30

Flat spectral density
function and
corresponding
w = 2π f autocorrelation
−π B 0 πB function for
(a) interference signal.
a: Spectral density
Rxx (τ)
function.
b: Autocorrelation
function.

t
0

1 2
B B
(b)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 402

402 CHAPTER 10 Compensation of Adaptive Arrays

r xi xk (0) = g[0, +][1 + |ρm |2 ] exp[− jω0 (i − k)]

+ g[τ12 − τ22 − Dm , −]ρm exp{ jω0 [τ12 − τ22 − Dm − (i − k)]} (10.80)
+ g[τ12 − τ22 − Dm , +]ρm∗ exp{− jω0 [τ12 − τ22 − Dm + (i − k)]}
The vector rx x0 (−D) is then given by
⎡ ⎤
r x1 x0 (−D)
⎢ r x2 x0 (−D) ⎥
⎢ ⎥
rx x0 (−D) = ⎢ .. ⎥ (10.81)
⎣ . ⎦
r x N x0 (−D)
and the matrix Rxx (0) is given by
⎡ ⎤
r x1 x1 (0) r x1 x2 (0) ··· r x1 x N (0)
⎢ .. ⎥
⎢ . r x2 x2 (0) ⎥
Rxx (0) = ⎢⎢ ⎥ (10.82)
.. ⎥
⎣ . ⎦
r x1 x N (0) ··· r x N x N (0)
To evaluate (10.68) for the minimum possible value of output residue power (10.78),
(10.79), and (10.80), show that it is necessary to specify the following parameters:

N = number of taps in the transversal filter

ρm = multipath reflection coefficient
ω0 = (radian) center frequency of interference signal
Dm = multipath delay time with respect to direct ray
τ12 = propagation delay between the main antenna element and the auxiliary antenna
element for the direct ray
τ22 = propagation delay between the main antenna element and the auxiliary antenna
element for the multipath ray
 = transversal filter intertap delay
B = interference signal bandwidth
D = main channel receiver time delay

The quantities τ12 and τ22 are related to the CSLC array geometry by
⎫
d
τ12 = sin θ1 ⎪
⎬
 (10.83)
d ⎪
τ22 = sin θ2 ⎭

where
d = interelement array spacing
 = wavefront propagation speed
θ1 = angle of incidence of direct ray
θ2 = angle of incidence of multipath ray

10.5.3 Example: Results for Compensation of Multipath Effects

An interference signal has a direct ray angle of arrival is θ1 = 30◦ , the multipath ray angle
of arrival is θ2 = −30◦ , and the interelement spacing is d = 2.25λ0 . Some additional
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 403

10.5 Multipath Compensation 403

signal and multipath characteristics are

center frequency f 0 = 237 MHz
signal bandwidth B = 3 MHz (10.84)
multipath reflection coefficient ρm = 0.5

Referring to (10.76), (10.79), and (10.80), we see that the parameters ω0 , τ12 , τ22 ,
Dm , and  enter the evaluation of the output residue power in the form of the products
ω0 τ12 , ω0 τ22 , ω0 Dm , and ω0 . These products represent the phase shift experienced at the
center frequency ω0 as a consequence of the four corresponding time delays. Likewise, the
parameters B, D, Dm , τ12 , τ22 , and  enter the evaluation of the output residue power in
the form of the products BD, BDm , Bτ12 , Bτ22 , and B; these time–bandwidth products
are phase shifts experienced by the highest frequency component of the complex envelope
interference signal as a consequence of the five corresponding time delays. Both the intertap
delay  and the multipath delay Dm are important parameters that affect the CSLC system
performance through their corresponding time–bandwidth products; thus, the results are
given here with the time–bandwidth products taken as the fundamental quantity of interest.
Since for this example θ1 = −θ2 , the product ω0 τ12 is specified as
π ⎫
ω0 τ12 = ⎪
⎪
4 ⎬
then the product (10.85)
π⎪⎪
ω0 τ22 = − ⎭
4
Furthermore, let the products ω0 Dm and ω0  be given by

ω0 Dm = 0 ± 2kπ, k any integer
(10.86)
ω0  = 0 ± 2lπ, l any integer
For the element spacing d = 2.25λ0 and θ1 = 30◦ , then specify
1
Bτ12 = −Bτ22 = , P = 72 (10.87)
P
Finally, specifying the multipath delay time to correspond to 46 meters yields

BDm = 0.45 (10.88)

Since
N −1
D=  (10.89)
2
Only N and B need to be specified to evaluate the output residue power by way of
(10.68).
To evaluate the output residue power by way of (10.68) resulting from the array
geometry and multipath conditions specified by (10.84)–(10.89) requires that the cross-
correlation vector rx x0 (−D), the N × N autocorrelation matrix Rxx (0), and the autocorre-
lation function r x0 x0 (0) be evaluated by way of (10.78)–(10.80). A computer program to
evaluate (10.68) for the multipath conditions specified was written in complex, double-
precision arithmetic.
Figure 10-31 shows a plot of the output residue power where the resulting minimum
possible value of canceled power output in dB is plotted as a function of B for various
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 404

404 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-31 0
Decibel cancellation N = number of taps
N=1
versus B for
multipath. −10

N=3

Cancellation (dB)
−20

−30 N=5
Evaluated
with
1
Bτ =
−40 78
N=7 BDm = 0.45
Rm = 0.5
−50
0.1 0.2 0.3 0.4 0.6 0.8 1
BΔ

specified values of N. It will be noted in Figure 10-31 that for N = 1 the cancellation
performance is independent of B since no intertap delays are present with only a single
tap. As explained in Appendix B, the transfer function of the tapped delay line transversal
filter has a periodic structure with (radian) frequency period 2π B f , which is centered at
the frequency f 0 . It should be noted that the transversal filter frequency bandwidth B f
is not necessarily the same as the signal-frequency bandwidth B. The transfer function
of a transversal filter within the primary frequency band (| f − f 0 | < B f /2) may be
expressed as


N
F( f ) = [Ak e jφk ] exp[− j2π(k − 1)δ f ] (10.90)
k=1

where Ak e jφk represents the kth complex weight, δ f = f − f 0 , f 0 = center frequency,

and the transversal filter frequency bandwidth is
1
Bf = (10.91)

Since the transversal filter should be capable of adjusting the complex weights to achieve
appropriate amplitude and phase values over the entire signal bandwidth B, it follows that
B f should satisfy

Bf ≥ B (10.92)

Consequently, the maximum intertap delay spacing is given by

1
max = (10.93)
B
It follows that values of B that are greater than unity should not be considered for
practical compensation designs; however, values of B f > B (resulting in 0 < B < 1)
are sometimes desirable.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 405

10.5 Multipath Compensation 405

Figure 10-31 shows that, as B decreases from 1, for values of N > 1 the cancellation
performance rapidly improves (the minimum canceled residue power decreases) until
B = BDm (0.45 for this example), after which very little significant improvement
occurs. As B becomes very much smaller than BDm (approaching zero), the cancellation
performance degrades since the intertap delay is effectively removed. The simulation could
not compute this result since as B approaches zero the matrix Rxx (0) becomes singular
and matrix inversion becomes impossible. Cancellation performance of −30 dB is virtually
assured if the transversal filter has at least five taps and  is selected so that  = Dm .
Suppose for example that the transversal filter is designed with B = 0.45. Using the
same set of selected constants as for the previous example, we find it useful to consider what
results would be obtained when the actual multipath delay is different from the anticipated
value corresponding to BDm = 0.45. From the results already obtained in Figure 10-31, it
may be anticipated that, if BDm > B, then the cancellation performance would degrade.
If, however, B Dm B, then the cancellation performance would improve since in the
limit as Dm → 0 the system performance with no multipath present would result.

10.5.4 Results for Compensation of Array Propagation Delay

In the absence of a multipath ray, the analysis presented in the preceding section includes all
the features necessary to account for array propagation delay effects. When we set ρm = 0
and let τ12 = τ represent the element-to-element array propagation delay, (10.78)–(10.80)
permit (10.68) to be used to investigate the effects of array propagation delay on cancella-
tion performance. On the basis of the behavior already found for multipath compensation,
it would be reasonable to anticipate that with B = Bτ then maximum cancellation per-
formance would obtain, whereas if B > Bτ then the cancellation performance would
degrade. Figure 10-32 gives the resulting cancellation performance as a function of B
for fixed Bτ . The number of taps N is an independent parameter, and all other system
constants are the same as those in the example of Section 10.4.3. It is seen that the results
confirm the anticipated performance noted already.

−20 FIGURE 10-32

Decibel cancellation
N = number of taps versus B for array
−30 N=1 propagation delay.
Cancellation (dB)

−40 Evaluate with

1
Bτ =
78
no multipath
−50
3
=
N

5
N=

−60

−70
0.1 0.2 0.3 0.4 0.6 0.8 1
BΔ
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 406

406 CHAPTER 10 Compensation of Adaptive Arrays

10.6 ANALYSIS OF INTERCHANNEL

MISMATCH EFFECTS
Any adaptive array processor is susceptible to unavoidable frequency-dependent vari-
ations in gain and phase between the various element channels. Additional degrees of
freedom provided by a tapped delay line compensate for such frequency-dependent “chan-
nel mismatch” effects. Since a simple two-element CSLC system exhibits all the salient
characteristics of channel mismatching present in more complex systems, the two-element
model is again adopted as the example for performance evaluation of channel mismatch
compensation.
Figure 10-33 is a simplified representation of a single auxiliary channel CSLC system
in which the single complex weight is a function of frequency. The transfer function
T0 (ω, θ) reflects all amplitude and phase variations in the main beam sidelobes as a
function of frequency as well as any tracking errors in amplitude and phase between the
main and auxiliary channel electronics. Likewise, the equivalent transfer function for the
auxiliary channel (including any auxiliary antenna variations) is denoted by T1 (ω, θ ). The
spectral power density of a wideband jammer is given by φJJ (ω). The signal from the
auxiliary channel is “multiplied” by the complex weight w 1 = αe jφ , and the “cancelled”
output of residue power spectral density is represented by φrr (ω, θ).
The objective of the CSLC is to minimize the residue power, appropriately weighted,
over the bandwidth. Since the integral of the power spectral density over the signal fre-
quency spectrum yields the signal power, the requirement to minimize the residue power
is expressed as
 ∞
Min φrr (ω, θ )dω (10.94)
w1 −∞
where
φrr (ω, θ ) = |T0 (ω, θ ) − w 1 T1 (ω, θ )|2 φJJ (ω) (10.95)
Now replace the complex weight w 1 in Figure 10-33 by a tapped delay line having
2N + 1 adaptively controlled complex weights each separated by a time delay  as in
Figure 10-34. A delay element of value N  is included in the main channel (just as in the
preceding section) so that compensation for both positive and negative angles of arrival
is provided. The main and auxiliary channel transfer functions are written in terms of the
output of the main channel, so no delay terms occur in the resulting main channel transfer
FIGURE 10-33 Main channel
Simplified model of + R (w, q )
T0 (w, q )
single-channel
CSLC. −

d w1 Adaptive
electronics
q

S(w)

Auxiliary channel
T1 (w, q )
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 407

10.6 Analysis of Interchannel Mismatch Effects 407

A (ω) FIGURE 10-34

Single-channel
+ CSLC having main
Main
channel T0 (w) e−jw NΔ Residue channel distortion
− and tapped delay
line auxiliary channel
F(w) = Â(w)
compensation.

w1 w1 wN+ 1 w 2N +1

Auxiliary
channel e−jwΔ e−jw Δ e−jw Δ

N delay elements N delay elements

N
F(ω) =
k = −N
Σ
wN+ 1 + k e−jwkΔ

function, A(ω). Assume for analysis purposes that all channel distortion is confined to the
main channel and that T1 (ω, θ) = 1. The transversal filter transfer function, F(ω), can be
expressed as

N
F(ω) = w N +1+k e− jωk (10.96)
k=−N

where the w N +1+k ’s are nonfrequency-dependent complex weights.

We want to minimize the output residue power over the signal bandwidth by appro-
priately selecting the weight vector w. Assuming the jammer power spectral density is
constant over the frequency region of interest, then minimizing the output residue power
is equivalent to selecting the F(ω) that provides the “best” estimate (denoted by Â(ω))
of the main channel transfer function over that frequency range. If the estimate Â(ω) is to
be optimal in the MSE sense, then the error in this estimate e(ω) = A(ω) − F(ω) must
be orthogonal to  (ω) = F(ω), that is,
E{[A(ω) − F(ω)]F ∗ (ω)} = 0 (10.97)
where the expectation E{·} is taken over frequency and is therefore equivalent to
 πB
1
E{·} = { } dω (10.98)
2π B −π B
where all frequency-dependent elements in the integrand of (10.98) are reduced to base-
band. Letting A(ω) = A0 (ω)e− jφ0 (ω) , substituting (10.96) into (10.97), and requiring the
error to be orthogonal to all tap outputs to obtain the minimum MSE estimate Â(ω) then
yields the condition

E{[A0 (ω) exp[− jφ0 (ω)] − F(ω)] exp( jωk )} = 0 for k = −N , . . . , 0, . . . , N
(10.99)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 408

408 CHAPTER 10 Compensation of Adaptive Arrays

Equation (10.99) can be rewritten as

"

N
E{A0 (ω) exp[ j (ωk  − φ0 (ω)]} − E W N +1+l exp(− jωl)
l=−N
#
· exp( jωk) =0 for k = −N , . . . , 0, . . . , N

(10.100)
Note that
sin [π B(l − k)]
E{exp[− jω(l − k)]} = (10.101)
π B(l − k)
it follows that
" N #
 N
sin[π B(l − k)]
E W N +1+l exp(− jωl) exp( jωl) = W N +1+l
l=−N l=−N
π B(l − k)
(10.102)
so that (10.100) can be rewritten in matrix form as
v = Cw (10.103)
where
vk = E{A0 (ω) exp[ j (ωk − φ0 (ω))]} (10.104)
sin[π B(l − k)]
Ck,l = (10.105)
π B(l − k)
Consequently, the complex weight vector must satisfy the relation

w = C−1 v (10.106)

Using (10.106) to solve for the optimum complex weight vector, we can find the output
residue signal power by using
 πB
1
Ree (0) = |A(ω) − F(ω)|2 φJJ (ω)dω (10.107)
2π B −π B
where φJJ (ω) is the constant interference signal power spectral density. Assume the inter-
ference power spectral density is unity across the bandwidth of concern; then the output
residue power due only to main channel amplitude variations is given by
 πB
1
Ree A = |A0 (ω) − F(ω)|2 dω (10.108)
2π B −π B
Since A(ω) − F(ω) is orthogonal to F(ω), it follows that [15]

E{|A(ω) − F(ω)|2 } = E{|A(ω)|2 } − E{|F(ω)|2 } (10.109)

and hence
 πB
1
Ree A = [A20 (ω) − |F(ω)|2 ] dω (10.110)
2π B −π B
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 409

10.6 Analysis of Interchannel Mismatch Effects 409

It likewise follows from (10.107) that the output residue power contributed by main
channel phase variations is given by
 πB
1
Ree p = |e− jφ0 (ω) − F(ω)|2 φJJ (ω) dω (10.111)
2π B −π B

where φ0 (ω) represents the main channel phase variation. Once again assuming that
the input signal spectral density is unity across the signal bandwidth and noting that
[e− jφ0 (ω) − F(ω)] must be orthogonal to F(ω), it immediately follows that
 πB
1
Ree p = [1 − |F(ω)|2 ] dω
2π B −π B

N N
sin[π B(k − j)]
=1− w k w ∗j (10.112)
j=−N k=−N
π B(k − j)

where the complex weight vector elements must satisfy (10.103)–(10.106).

If it is desired to evaluate the effects of both amplitude and phase mismatching
simultaneously, then the appropriate expression for the output residue power is given
by (10.107), which (because of orthogonality) may be rewritten as
 πB
1
Ree (0) = {|A(ω)|2 − |F(ω)|2 }φJJ (ω) dω (10.113)
2π B −π B

where the complex weights used to obtain F(ω) must again satisfy (10.102)–(10.106),
which now involve both a magnitude and a phase component and it is assumed that φJJ (ω)
is a constant.

10.6.1 Example: Effects of Amplitude Mismatching

To evaluate (10.110) it is necessary to adopt a channel amplitude model corresponding to
A(ω). One possible channel amplitude model is given in Figure 10-35 for which

1 + R cos ωT0 for |ω| ≤ π B
A(ω) = (10.114)
0 otherwise

A(w ) FIGURE 10-35

Channel amplitude
model having 3 12
R cycles of ripple for
1 evaluation of
amplitude mismatch
effects.

ω
−π B 0 πB

Array bandwidth
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 410

410 CHAPTER 10 Compensation of Adaptive Arrays

where
2n + 1
T0 = for n = 0, 1, 2, . . .
2B
and the integer n corresponds to (2n + 1)/2 cycles of amplitude mismatching across the
bandwidth B. Letting the phase error φ0 (ω) = 0, it follows from (10.104) that
 πB
1
vk = [1 + R cos ωT0 ]e jωk dω (10.115)
2π B −π B
or
 
sin(π Bk) R sin(π B[T0 + k]) sin(π B[T0 − k])
vk = + +
π Bk  2 π B [T0 + k] B[T0 − k]
for k = −N , . . . , 0, . . . , N (10.116)

Evaluation of (10.116) permits the complex weight vector to be found, which in turn may
be used to determine the residue power by way of (10.110).
Now

|F(ω)|2 = F(ω) F ∗ (ω) = w† ββ † w (10.117)

where
⎡ ⎤
e jωN 
⎢ e jω(N −1)  ⎥
⎢ ⎥
β = ⎢. ⎥ (10.118)
⎣ .. ⎦
e− jωN 
Carrying out the vector multiplications indicated by (10.117) then yields

2N +1 2N
 +1
|F(ω)|2 = w i w k∗ e jω(k−i) (10.119)
i=1 k=1

The output residue power is therefore given by [see equation (10.110)]

 πB  
π B 2N +1 2N
 +1
Ree A = [1 + R cos ωT0 ] dω −
2
w i w k∗ e jω(k−i) dω (10.120)
−π B −π B i=1 k=1

Equation (10.120) may be evaluated using the following expressions:

 πB $ %
1 R2 sin π [(2n + 1)/2]
[1 + R cos ωT0 ] dω = 1 +
2
+ 2R
2π B −π B 2 π [(2n + 1)/2]
R 2 sin π (2n + 1)
+ (10.121)
2 π (2n + 1)
 
π B 2N +1 2N
 +1 
2N +1 2N
 +1
1 sin π(k − i)B
w i w k∗ e jω(k−i) dω = w i w k∗
2π B −π B i=1 k=1 i=1 k=1
π(k − i)B
(10.122)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 411

10.6 Analysis of Interchannel Mismatch Effects 411

10.6.2 Results for Compensation of Selected

Amplitude Mismatch Model
The evaluation of (10.120) requires knowing the ripple amplitude R, the number of cycles
of amplitude mismatching across the bandwidth, and the product of B (where B is the
cancellation bandwidth and  is the intertap delay spacing). The results of a computer
evaluation of the output residue power are summarized in Figures 10-36–10-39 for B =
0.25, 0.5, 0.75, and 1, and R = 0.09. Each of the figures presents a plot of the decibel
cancellation (of the undesired interference signal) achieved as a function of the number
of taps in the transversal filter and the number of cycles of ripple present across the
cancellation bandwidth. No improvement (over the cancellation that can be achieved with
only one tap) is realized until a sufficient number of taps is present in the transversal
filter to achieve the resolution required by the amplitude versus frequency variations in

For BΔ = 0.25 and Rm = 0.09 FIGURE 10-36

−20 Decibel cancellation
1
3 Cycles versus number of
2
taps for selected
−30 2
1
Cycles amplitude mismatch
2
models with
B = 0.25.
Cancellation (dB)

−40

−50

−60
1
1 1 Cycles
Cycle 2
2
−70
0 5 10 15
N = number of taps

−10 FIGURE 10-37

For BΔ = 0.5 and Rm = 0.09 Decibel cancellation
versus number of
−20 taps for selected
1 amplitude mismatch
3 Cycles
2 models with
−30
Cancellation (dB)

B = 0.5.

−40

−50
1
2 Cycles
2
−60
1
1 1 Cycles
Cycle 2
2
−70
0 5 10 15
N = number of taps
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 412

412 CHAPTER 10 Compensation of Adaptive Arrays

FIGURE 10-38 −10

Decibel cancellation For BΔ = 0.75 and Rm = 0.09
versus number of
taps for selected −20
amplitude mismatch 1
3 Cycles
models with 2
−30

Cancellation (dB)
B = 0.75.

−40

−50
1
2 Cycles
2
−60
1 1
Cycle 1 Cycles
2 2
−70
0 5 10 15
N = number of taps

FIGURE 10-39 −10

Decibel cancellation For BΔ = 1.0 and Rm = 0.09
versus number of
taps for selected −20
amplitude mismatch
models with −30
Cancellation (dB)

B = 1.0. 1
3 Cycles
2
−40
1
2 Cycles
2
1
−50 1 Cycles
2

−60
1
Cycle
2
−70
0 5 10 15
N = number of taps

the amplitude mismatch model. The sufficient number of taps for the selected amplitude
mismatch model was found empirically to be given by
 
Nr − 1
Nsufficient ≈ [7 − 4(B)] + 1 (10.123)
2
where Nr is the number of half-cycles of ripple appearing in the mismatch model.
If there are a sufficient number of taps in the transversal filter, the cancellation perfor-
mance improves when more taps are added depending on how well the resulting transfer
function of the transversal filter matches the gain and phase variations of the channel
mismatch model. Since the transversal filter transfer function resolution depends in part
on the product B, a judicious selection of this parameter ensures that providing addi-
tional taps provides a better match (and hence a significant improvement in cancellation
performance), whereas a poor choice results in very poor transfer function matching even
with the addition of more taps.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 413

10.6 Analysis of Interchannel Mismatch Effects 413

0 FIGURE 10-40
−10 For N = 3 and Rm = 0.9 Decibel cancellation
−20 versus B for
one-half-cycle
amplitude mismatch
−50 model.
Cancellation (dB)

−100

−150

0 0.5 1 1.5
BΔ

Taking the inverse Fourier transform of (10.114) −1 {A(ω)} yields a time function
corresponding to an autocorrelation function f (t) that can be expressed as
f (t) = s(t) + Ks(t ± T0 ) (10.124)
The results of Section 10.5.3 and equation (10.124) imply that  = T0 (or equivalently,
B = number cycles of ripple mismatch) if the product B is to “match” the amplitude
mismatch model. This result is illustrated in Figure 10-40 where decibel cancellation is
plotted versus B for a one-half-cycle ripple mismatch model. A pronounced minimum
occurs at B = 12 for N = 3 and Rm = 0.9.
When the number of cycles of mismatch ripple exceeds unity, the foregoing rule
of thumb leads to the spurious conclusion that B should exceed unity. Suppose, for
example, there were two cycles of mismatch ripple for which it was desired to compensate.
By setting B = 2 (corresponding to B f = 12 B), two complete cycles for the transversal
filter transfer function are found to occur across the cancellation bandwidth. By matching
only one cycle of the channel mismatch, quite good matching of the entire mismatch
characteristic occurs but at the price of sacrificing the ability to independently adjust the
complex weights across the entire cancellation bandwidth, thereby reducing the ability
to appropriately process broadband signals. Consequently, if the number of cycles of
mismatch ripple exceeds unity, it is usually best to set B = 1 and to accept whatever
improvement in cancellation performance can be obtained with that value, or increase the
number of taps.

10.6.3 Example: Effects of Phase Mismatching

Let φ(ω) corresponding to the phase error be characterized by

A cos ωT0 for |ω| ≤ π B
φ(ω) = (10.125)
0 otherwise
where A represents the peak number of degrees associated with the phase error ripples.
This model corresponds to the error ripple model of (10.112) (with zero average value
present).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 414

414 CHAPTER 10 Compensation of Adaptive Arrays

Since
 πB
1
vk = exp( j{A cos ωT0 + ωk}) dω for k = −N , . . . , 0, . . . , N
2π B −π B
(10.126)
it can easily be shown by defining

 sin π [K + sgn · (i − (N + 1))B]

f (K , sgn) =
π [K + sgn · (i − (N + 1))B]

and g(K ) = f (K , +) + f (K , −) that
 
2n + 1
vi = J0 (A) f (0, +) + j J1 (A)g (10.127)
2
 ∞    
2n + 1
+ (−1) J2k (A)g[k(2n + 1)] + j J2k+1 (A)g (2k + 1)
k

k=1
2

where Jn (·) denotes a Bessel function of the nth order for i = 1, 2, . . . , 2N + 1.

10.6.4 Results for Compensation of Selected Phase

Mismatch Model
The computer evaluation of the output residue power resulted in the performance sum-
marized in Figures 10-41–10-43 for B = 0.2, 0.45, and 1.0 and A = 5◦ . These
figures present the decibel cancellation achieved as a function of the number of taps
in the transversal filter and the number of cycles of phase ripple present across the can-
cellation bandwidth. The general nature of the curves appearing in Figures 10-41–10-43
is the same as that of Figures 10-36–10-39 for amplitude mismatching. Furthermore, just
as in the amplitude mismatch case, a better channel transfer function fit can be obtained
with the transversal filter when the mismatch characteristic has a fewer number of ripples.

FIGURE 10-41 −10

Decibel cancellation For BΔ = 0.2 and A = 5°
versus number of 1 1
taps for selected −20 6 & 10 Cycles
2 2
phase mismatch
models with −30
Cancellation (dB)

1
B = 0.2. 2
2
Cycles

−40

−50

−60
1
Cycle
2
−70
0 1 3 5 7
N = number of taps
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 415

10.7 Summary and Conclusions 415

−10 FIGURE 10-42

For BΔ = 0.45 and A = 5° Decibel cancellation
1 1 versus number of
−20 6 & 10 Cycles taps for selected
2 2
phase mismatch
−30 models with
Cancellation (dB) 1 B = 0.45.
2 Cycles
2
−40

−50

−60
1
Cycle
2
−70
0 1 3 5 7
N = number of taps

−10 FIGURE 10-43

For BΔ = 1 and A = 5° Decibel cancellation
1 1 versus number of
6 & 10 Cycles taps for selected
−20 2 2
phase mismatch
models with
Cancellation (dB)

−30
B = 1.0.

1
2 Cycles
−40 2

−50
1
Cycle
2
−60
0 1 3 5 7
N = number of taps

10.7 SUMMARY AND CONCLUSIONS

Array errors due to manufacturing tolerances distort the array pattern. To minimize these
errors, the array must be calibrated at the factory and at regular intervals once deployed.
The transversal filter consisting of a sequence of weighted taps with intertap delay
spacing offers a practical means for achieving the variable amplitude and phase weighting
as a function of frequency that is required if an adaptive array system is to perform
well against wideband interference signal sources. The distortionless channel transfer
functions for a two-element array were derived. It was found that to ensure distortion-free
response to a broadband signal the channel phase is a linear function of frequency, whereas
the channel amplitude function is nearly flat over a 40% bandwidth. Quadrature hybrid
processing provides adequate broadband signal response for signals having as much as
20% bandwidth. Tapped delay line processing is a practical necessity for 20% or more
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 416

416 CHAPTER 10 Compensation of Adaptive Arrays

bandwidth signals. A transversal filter provides an attractive means of compensating the

system auxiliary channels for the undesirable effects of the following:
1. Multipath interference
2. Interchannel mismatch
3. Propagation delay across the array
For multipath interference, the value of the intertap delay is in the neighborhood of the
delay time associated with the multipath ray. If the intertap delay time exceeds the mul-
tipath delay time by more than about 30% and the multipath delay time is appreciable,
a severe loss of compensation capability is incurred. If the intertap delay is too small,
then an excessive number of taps will be required for effective cancellation to occur. Since
multipath delay having “small” values of associated time delay do not severely degrade the
array performance, it is reasonable to determine the most likely values of multipath delay
that will occur for the desired application and base the multipath compensation design on
those delay times (assuming B ≤ 1). For reflection coefficients of 0.5 and BDm = 0.45,
the use of five taps will ensure a −30 dB cancellation capability.
The results shown in Figures 10-31 and 10-32 indicate that array propagation delay
effects are usually much easier to compensate than are multipath effects. This result
occurs because multipath in effect introduces two (or more) signals in each channel (that
are essentially uncorrelated if BDm  1), which require more degrees of freedom to
adequately compensate.
The problem presented by interchannel mismatch is to obtain a transfer function with
the transversal filter that succeeds in matching the amplitude and phase error charac-
teristics exhibited among the various sensor channels. As might be expected, the more
severe the mismatching between channels, the more difficult it is to achieve an acceptable
degree of compensation. In particular, it is highly undesirable for more than 2 12 cycles of
mismatch ripple to occur over the cancellation bandwidth; even this degree of mismatch
requires seven taps on the transversal filter before a truly effective degree of compensation
can be achieved. It may very well result that the best choice of intertap delay spacing
for the interchannel mismatch characteristic of concern is far different from the optimum
choice of intertap delay selected for multipath compensation; should this actually occur,
it is necessary to adopt a compromise value for the intertap delay spacing. Such a com-
promise value for the intertap delay spacing hopefully results in an acceptable degree of
compensation for both multipath and interchannel mismatch effects.

10.8 PROBLEMS
Distortionless Transfer Functions
1. From (10.11) and (10.12) it immediately follows that |H1 (ω) = |H2 (ω)|, thereby yielding the
pair of equations
f 1 {|H1 |, α1 , α2 , θs } = exp(− jωT1 )
and
f 2 {|H1 |, α1 , α2 , θi } = 0
(a) Show from the previous pair of equations that α1 (ω) and α2 (ω) must satisfy
πω
α2 (ω) − α1 (ω) = sin θi ± nπ
ω0
where n is any odd integer.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 417

10.8 Problems 417

(b) Since the magnitude of exp(− jωT1 ) must be unity, show using f 1 { } = exp(− jωT1 ) that
(10.13) results.
(c) Show that the angle condition associated with f 1 { } = exp(− jωT1 ) yields (10.14).
(d) Show that substituting (10.14) into the results from part (a) yields (10.15).
2. For a three-element linear array, the overall transfer function encountered by the desired signal
in passing through the array is
   
ωd ω2d
Hd (ω) = H1 (ω) + H2 (ω) exp − j sin θs + H3 (ω)e − j sin θs
c c

and the overall transfer function seen by the interference signal is

   
ωd ω2d
HI (ω) = H1 (ω) + H2 (ω) exp − j sin θi + H3 (ω)e − j sin θi
c c

What does imposing the requirements (10.9) and (10.10) now imply for the three-channel
transfer functions?
Hilbert Transform Relations
3. Prove the Hilbert transform relations given by (10.25)–(10.28).
4. Using (10.61), (10.62), and the results of (10.63)–(10.65), show that Ree is given by (10.66).
5. Derive the correlation functions given by (10.78)–(10.80) for the signal environment assump-
tions (10.75) and (10.76)
6. Show that as the time–bandwidth product B approaches zero, then the matrix Rxx (0) [whose
elements are given by (10.80)] becomes singular so that matrix inversion cannot be accom-
plished.
Compensation for Channel Phase Errors
7. For the phase error φ(ω) given by (10.125), show that vk given by (10.127) follows from the
application of (10.126).
8. Let φ(ω) correspond to the phase error model be given by
 & '
A 1 − cos 2ω for |ω| ≤ π B
φ(ω) = B
0 otherwise

Show that vk of (10.126) is given by

 πB       
2 2
vk = cos A 1 − cos ω + j sin A 1 − cos ω exp( jωkdω )
−π B B B

Use the trigonometric identities

     
2 2 2
cos A − A cos ω = cos A cos A cos ω + sin A sin A cos ω
B B B
     
2 2 2
sin A − A cos ω = sin A cos A cos ω − cos A sin A cos ω
B B B
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 418

418 CHAPTER 10 Compensation of Adaptive Arrays

and the fact that


∞
cos(A cos ωT0 ) = J0 (A) + 2 (−1)k · J2k (A) cos[(2k)ωT0 ]
k=1

∞
sin(A cos ωT0 ) = 2 (−1)k J2k+1 (A) · cos[(2k + 1)ωT0 ]
k=0

where Jn (·) denotes a Bessel function of the nth order and define

 sin π[n + sgn · (i − (N + 1))B]

f (n, sgn) =
π [n + sgn · (i − (N + 1))B]

g(n) = f (n, +) + f (n, −)

to show that

vi = J0 (A) · f (0, +)[cos A + j sin A] + J1 (A) · g(2)[sin A − j cos A]


∞
+ (−1)k {J2k (A) · g(4k)[cos A + j sin A]
k=1
+ J2k+1 (A) · g[(2k + 1)2][sin A − j cos A]}
for i = 1, 2, . . . , 2N + 1

9. Let φ(ω) corresponding to the phase error model be given by


bω2 (π B − |ω|) for |ω| ≤ π B
φ(ω) =
0 otherwise

As before, it follows that

 πB
1
vi = exp{ j[bω2 (π B − |ω|) + ωi]} dω
2π B −π B

Letting u = ω/π B, applying Euler’s formula, and ignoring all odd components of the resulting
expression, show that
 1   
27A 2
vi = exp j u (1 − u) cos π[u(i − (N + 1))B]du
0 4

where A = 4b(π B/3)3 for i = 1, 2, . . . , 2N +1. The foregoing equation for vi can be evaluated
numerically to determine the output residue power contribution due to the previous phase error
model.
Computer Simulation Problems
10. A 30-element linear array (d = 0.5λ) has a 20 dB, n = 2 Taylor taper applied at the elements.
Plot the array factor when δna = 0.1 and δna = 0.1.
11. A 30-element linear array (d = 0.5λ) has a 30 dB, n = 7 low sidelobe taper. Plot the array
factors for a single element failure at (1) the edge and (2) the center of the array.
12. Find the location and heights of the quantization lobes for a 20-element array with d = 0.5λ
and the beam steered to θ = 3◦ when the phase shifters have three, four, and five bits.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 419

10.9 References 419

10.9 REFERENCES
[1] E. Brookner, Antenna Array Fundamentals—Part 2, Practical Phased Array Antenna Systems,
E. Brookner, ed., Norwood, MA, Artech House, 1991
[2] R.J. Mailloux, “Array Grating Lobes Due to Periodic Phase, Amplitude, and Time Delay
Quantization,” IEEE AP-S Trans., Vol. 32, December 1984, pp. 1364–1368.
[3] A. J. Boonstra, and A. J. van der Veen, “Gain Calibration Methods for Radio Telescope
Arrays,” IEEE Signal Processing Trans., Vol. 51, No. 1, 2003, pp. 25–38.
[4] R. Sorace, “Phased Array Calibration,” IEEE AP-S Trans., Vol. 49, No. 4, 2001, pp. 517–525.
[5] T. Takahashi, Y. Konishi, S. Makino, et al., “Fast Measurement Technique for Phased Array
Calibration,” IEEE AP-S Trans., Vol. 56, No. 7, 2008, pp. 1888–1899.
[6] M. Borkowski, “Solid-State Transmitters,” Radar Handbook, M. I. Skolnick, ed.,
pp. 11.1–11.36, New York, McGraw Hill, 2008.
[7] N. Tyler, B. Allen, and H. Aghvami, “Adaptive Antennas: The Calibration Problem,” IEEE
Communications Magazine, Vol. 42, No. 12, 2004, pp. 114–122.
[8] N. Tyler, B. Allen, and A. H. Aghvami, “Calibration of Smart Antenna Systems: Measurements
and Results,” IET Microwaves, Antennas & Propagation, Vol. 1, No. 3, 2007, pp. 629–638.
[9] R. L. Haupt, Antenna Arrays: A Computational Approach, New York, Wiley, 2010.
[10] W. T. Patton and L. H. Yorinks, “Near-Field Alignment of Phased-Array Antennas,” IEEE
AP-S Trans., Vol. 47, No. 3, 1999, pp. 584–591.
[11] W. E. Rodgers and R. T. Compton, Jr., “Tapped Delay-Line Processing in Adaptive Arrays,”
Report 3576-3, April 1974, prepared by The Ohio State University Electro Science Labora-
tory, Department of Electrical Engineering under Contract N00019-73-C-0195 for Naval Air
Systems Command.
[12] W. E. Rodgers and R. T. Compton, Jr., “Adaptive Array Bandwidth with Tapped Delay-
Line Processing,” Report 3832-3, May 1975, prepared by The Ohio State University Electro
Science Laboratory, Department of Electrical Engineering under Contract N00019-74-C-0141
for Naval Air Systems Command.
[13] W. E. Rodgers and R. T. Compton, Jr., “Adaptive Array Bandwidth with Tapped Delay-
Line Processing,” IEEE Trans. Aerosp. Electron. Syst., Vol. AES-15, No. 1, January 1979,
pp. 21–28.
[14] T. G. Kincaid, “The Complex Representation of Signals,” General Electric Report No.
R67EMH5, October 1966, HMED Publications, Box 1122 (Le Moyne Ave.), Syracuse, NY,
13201.
[15] A. Papoulis, Probability, Random Variables, and Stochastic Processes, New York,
McGraw-Hill, 1965, Ch. 7.
[16] H. Steyskal & J. Herd, “Mutual Coupling Compensation in Small Array Antennas,” IEEE
Trans. Ant. & Prop., Vol. AP-38, No. 12, December 1995, pp. 603–606.
[17] R. A. Monzingo, “Transversal Filter Implementation of Wideband Weight Compensation
for CSLC Applications,” unpublished Hughes Aircraft Interdepartmental Correspondence
Ref. No. 78-1450.10/07, March 1978.
[18] A. M. Vural, “Effects of Perturbations on the Performance of Optimum/Adaptive Arrays,”
IEEE Trans. Aerosp. Electron. Syst., Vol. AES-15, No. 1, January 1979, pp. 76–87.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:46 420
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 421

CHAPTER

Direction of Arrival Estimation

and Related Topics 11
' $
Chapter Outline
11.1 Periodogram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
11.2 Capon’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
11.3 MUSIC Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
11.4 The Maximum Entropy Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
11.5 Comparing DOA Algorithm Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
11.6 Sequential Implementation of Bayes Optimal Array Processor. . . . . . . . . . . . . . . . . . 437
11.7 Parameter Estimation via Subspace Fitting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
11.8 The Maximum Likelihood Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
11.9 Cramer–Rao Lower Bound on AOA Estimation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
11.10 Fisher Information Matrix and CR Bound for General Cases . . . . . . . . . . . . . . . . . . . . 451
11.11 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
11.12 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
11.13 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
& %

An array’s ability to resolve signals depends on the beamwidth of the array, so high-
resolution algorithms have been developed, in order for small arrays can resolve closely
spaced signals by using narrow nulls in place of the wide main beam. A linear array with
elements separated by half-wavelength spacing has N − 1 nulls in which to locate up to
N − 1 signals.
In this chapter, we apply maximum likelihood (ML) estimation methods to estimate the
direction of arrival (DOA), or angle of arrival (AOA), of one or more signal sources, using
data received by the elements of an N-element antenna array. The Cramer–Rao (CR) lower
bound on angle estimation error is derived under several different signal assumptions. The
CR bound helps determine system performance versus signal-to-noise ratio (SNR) and
array size.
The advantage of optimal array estimation methods (processing the antenna element
signals in an optimum fashion) lies in its application to multiple signal environments
and in conditions where interfering signals are present. If only one signal is present in
a white noise background, conventional monopulse processing achieves the same AOA
estimation accuracy. High angular resolution of desired signals is achieved beyond the one

421
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 422

422 CHAPTER 11 Direction of Arrival Estimation and Related Topics

beamwidth resolution limit of conventional monopulse processing. Angular accuracy and

angular resolution depends not only on array beamwidth but also on the SNR of the desired
signals and the number of snapshots of the data vector. Fractional beamwidth resolution,
or “super-resolution,” of two or more signal sources incident on an array offers a potential
solution to a wide variety of radar and communications problems, including fractional
beamwidth resolution of multiple interference sources, reduction of multipath-induced
angle of arrival estimation error, improved angle tracking accuracy in the presence of
main beam and sidelobe interference, spatial isolation of two closely spaced signals, and
improved ability to separate slow-moving and ground clutter in a radar that is moving, as
in an airborne radar.
The ML AOA estimation method is one of several useful super-resolution techniques
developed in the literature. While the ML methods come very close to theoretical bounds
on AOA estimation error (e.g., the Cramer–Rao bound), they are computationally intensive
when the AOAs of multiple signals are estimated at the same time. As a result, compu-
tationally simpler methods have emerged that perform nearly as well. Further drawbacks
of the ML method are that the desired signal sources (sources whose AOAs are to be esti-
mated) are represented as point sources and that the number of sources is known. In cases
where the spatial spectrum contains spatially distributed sources (e.g., heavy multipath
environment) or where an accurate estimate of the number of sources is difficult (dense
signal environment, heavy multipath), some of the minimum variance methods are more
appropriate, at the expense of AOA accuracy and resolution.
The AOA techniques presented here apply to a wide range of parameter estimation
problems. Examples include time of arrival estimation (e.g., target distance in a radar),
frequency estimation (e.g., target Doppler frequency, estimation of sinusoids), and signal
intensity.

11.1 PERIODOGRAM
The simplest approach to finding the direction of a signal is to scan the main beam of
the array by adjusting the steering vector until the signal is detected. The relative output
power of a linear array lying along the x-axis is given by

P(θ ) = A† (θ )Rx x A(θ ) (11.1)

and RT is the signal plus noise correlation matrix where the subscript “T ” denotes where
the uniform array steering vector is given by
 T
A(θ ) = e jkx1 cos θ · · · e jkx N cos θ , θmin ≤ θ ≤ θmax (11.2)

A periodogram is a plot of the output power versus angle, where a window function
that is independent of the data being analyzed must be adopted. By weighting all angles
equally, a rectangular window function is in effect being adopted. Peaks in the periodogram
correspond to signal locations. Large arrays have a narrower beamwidth than smaller arrays
and can resolve closely spaced signals better. Figure 11-1 shows the periodigram for a
12-element array with λ/2 spacing and three sources incident at θ = −50◦ , 10◦ , and 20◦ .
The source at θ = −50◦ is easy to distinguish, but the sources at θ = 10◦ , and 20◦ appear
to be a single source, because the beamwidth is too wide. This example demonstrates the
need for super-resolution techniques.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 423

11.3 MUSIC Algorithm 423

10 FIGURE 11-1 Plot

of the periodogram
of a 12-element
uniform array when
0 three sources are
incident at
θ = −50◦ , 10◦ ,
and 20◦ .
|P (dB)|

−10

−20

−90 −45 0 45 90
q (degrees)

11.2 CAPON’S METHOD

The array beamwidth, which is inveresely proportional to the array size, limits peri-
odogram resolution. Nulls are better suited to locate a signal than the main beam. The
maximum likelihood estimate of the power arriving from a desired direction, while all the
other sources are considered interference, is known as Capon’s method (or the Maximum
Likelihood Method [MLM])[1]. Both the MLM and Maximum Entropy Method (to be
discussed later) do not have fixed window functions associated with them, a fact that ame-
liorates the windowing problem (in effect, the window function adapts itself to the data
under analysis). The desired signal remains constant while the output power is minimized.
The array weights that maximize the signal-to-interference ratio are

R−1
xx A
w= (11.3)
A R−1
†
xx A

The Capon spectrum is the denominator of (11.3)

1
P(θ ) = (11.4)
A† (θ)R−1
x x A(θ )

Figure 11-2 shows the periodigram for a 12-element array with λ/2 spacing and three
sources incident at θ = −50◦ , 10◦ , and 20◦ . Capon’s method distinguishes between two
closely spaced sources much better than the periodogram.

11.3 MUSIC ALGORITHM

MUltiple SIgnal Classification (MUSIC) assumes the noise is uncorrelated and the signals
have little correlation [2,3]. The MUSIC spectrum is
A† (θ )A(θ )
P(θ ) = (11.5)
A† (θ )Vλ V†λ A(θ )
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 424

424 CHAPTER 11 Direction of Arrival Estimation and Related Topics

FIGURE 11-2 Plot 10

of the Capon
spectrum of a
12-element uniform
array when three 0
sources are incident
at θ = −50◦ , 10◦ ,
and 20◦ .

|P (dB)|
−10

−20

−90 −45 0 45 90
q (degrees)

FIGURE 11-3 Plot 10

of the MUSIC
spectrum of a
12-element uniform
array when three 0
sources are incident
at θ = −50◦ , 10◦ ,
and 20◦ .
|P (dB)|

−10

−20

−90 −45 0 45 90
q (degrees)

The signal plus noise correlation matrix in the denominator of (11.3) is replaced by Vλ V†λ
where the columns of Vλ are the eigenvectors of the noise subspace. In the numerator,
A† (θ) replaces R−1
x x and corresponds to the N − Ns smallest eigenvalues of the correlation
matrix. Figure 11-3 shows the MUSIC spectrum for a 12-element array with λ/2 spacing
and three sources incident at θ = −50◦ , 10◦ , and 20◦ . The MUSIC spectrum is similar
to the Capon spectrum, except the floor between the peaks is much lower for the MUSIC
spectrum.
The root-MUSIC algorithm is a more robust alternative that accurately locates the
direction of arrival by finding the roots of the array polynomial that corresponds to the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 425

11.3 MUSIC Algorithm 425

denominator of (11.5) [4].

†

M−1
†
A (θ)Vλ Vλ A(θ ) = cn z n (11.6)
n=M+1

where
z = e j λ nd sin θ
2π


cn = Vλ V†λ
r −c=n

The cn results from summing the diagonal n = r − c of Vλ Vλ† in (11.6) with r and c
indicating the row and column, respectively, of the matrix. Polynomial roots, z m , on the
unit circle (Chapter 2) are the poles of the MUSIC spectrum. The phase of the polynomial
roots in (11.6) are given by
 
−1 arg(z m )
θm = sin (11.7)
kd
Roots on the unit circle correspond to the signals. Roots off the unit circle are spurious.
The 2N − 1 diagonals of Vλ Vλ† form a polynomial with 2N − 2 roots. Table 11-1 contains
the roots of the polynomial for a 12-element array with λ/2 spacing and three sources
incident at θ = −50◦ , 10◦ , and 20◦ . Roots on the unit circle correspond to signals and have
a “yes” in column 3. Spurious roots are off the unit circle and have a “no” in column 3.
All roots appear in the unit circle plot in Figure 11-4. Note that each root on the unit circle
is actually a double root (see Table 11-1), so it appears that there are only 19 roots in
Figure 11-4 when there are actually 22 roots.
A unitary (real-valued) root-MUSIC algorithm reduces the computational complexity
of the root-MUSIC algorithm by exploiting the eigen decomposition of a real-valued
correlation matrix. Unitary root MUSIC improves threshold and asymptotic performances
relative to conventional root MUSIC.

Imaginary FIGURE 11-4 Unit

circle representation
of all the roots found
using root MUSIC.

Real
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 426

426 CHAPTER 11 Direction of Arrival Estimation and Related Topics

TABLE 11-1 The roots found using root MUSIC. The ones
close to the unit circle represent the correct signal directions.

m Roots Signal Present

1 −1.6766 +j0.6959 no
2 0.5468 −j1.4237 no
3 1.1541 −j0.9681 no
4 1.3909 −j0.2378 no
5 −0.2116 −j1.4237 no
6 −1.1328 +j1.1311 no
7 −1.5093 −j0.0508 no
8 −0.3246 +j1.3931 no
9 0.8585 +j0.5181 yes
10 0.8538 +j0.5153 yes
11 0.4787 +j0.8843 yes
12 0.4735 +j0.8746 yes
13 −0.7461 −j0.6749 yes
14 −0.7372 −j0.6668 yes
15 0.6986 −j0.1194 no
16 0.5086 −j0.4266 no
17 0.2351 −j0.6121 no
18 −0.1021 −j0.6872 no
19 0.1587 +j0.6809 no
20 −0.6618 −j0.0223 no
21 −0.4421 +j0.4414 no
22 −0.5088 +j0.2112 no

11.4 THE MAXIMUM ENTROPY METHOD

The maximum entropy method (MEM) is sometimes called the all poles model or the au-
toregressive model [5–8]. MEM spectral estimation is widely used in geophysics, speech
processing, sonar, and radar [9–12]. In addition to spectrum analysis, MEM applies to
bearing estimation problems for signals received by an array of sensors [13,14]. High
resolution comes from extrapolating the partially known autocorrelation function beyond
the last known lag value in a manner that maximizes the entropy of the corresponding
power spectrum at each step of the extrapolation [5,15]. If the autocorrelation function
is unknown, then the power spectrum is estimated directly from the available time series
data using a method devised by Burg [16]. Excellent power spectrum estimates are ob-
tained from relatively short time series data record lengths, and the approach has a rapid
convergence rate [17].
In its most elementary form the maximum entropy principle for estimating the power
spectrum of a single-channel, stationary, complex time series can be stated as a problem
of finding the spectral density function φxx ( f ) that maximizes
 W
entropy = ln φxx ( f )d f (11.8)
−W

under the constraint that φxx ( f ) satisfies a set of N linear measurement equations
 W
φxx ( f )G n ( f )d f = gn , n = 1, . . . , N (11.9)
−W
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 427

11.4 The Maximum Entropy Method 427

10 FIGURE 11-5 Plot

of the 12-element
uniform array MEM
spectrum with three
0 signals present.
|P (dB)|

−10

−20

−90 −45 0 45 90
q (degrees)

where the time series is sampled with the uniform period t so the Nyquist fold-over
frequency is W = 1/2t, and the power spectrum of the time series is bandlimited to
±W . The functions G n ( f ) in the measurement equations are known test functions, and
the gn are the observed values resulting from the measurements.
MEM is based on a rational function model of the spectrum that has only poles and
not zeros [18]. The MEM spectrum is given by [19]

1
P(θ) = (11.10)
A† (θ )R−1 [:, n]R†−1 [:, n]A(θ )
xx xx

where n corresponds to the nth column of the inverse correlation matrix. Results depend
on which n is chosen. Figure 11-5 shows the MEM spectrum for a 12-element array with
λ/2 spacing and three sources incident at θ = −50◦ , 10◦ , and 20◦ . Very sharp peaks in
the spectrum occur in the signal directions.
Two cases can now be considered: the first where the autocorrelation function is
partially known; and the second where the autocorrelation function is unknown.

11.4.1 Partially Known Autocorrelation Function

Let x(t) represent the time series of a stationary (scalar) random process with an associated
autocorrelation function r (τ ) for which N discrete lag values {r (0), r (1), . . . , r (N − 1)}
are known. An estimate of r (N ) lies outside the interval of known autocorrelation function
values. The basic autocorrelation function theorem states that r (N ) must have a value such
that the (N + 1) × (N + 1) Hermitian Toeplitz autocorrelation matrix given by
⎡ ⎤
r (0) r (−1) . . . r (−N )
 ⎢ r (1) r (0) . . . r (1 − N ) ⎥
Rnn =⎢⎣·
⎥
⎦ (11.11)
... ... ·
r (N ) r (N − 1) . . . r (0)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 428

428 CHAPTER 11 Direction of Arrival Estimation and Related Topics

is positive semidefinite (i.e., all subdeterminants of Rnn must be nonnegative). Since

det(R N ) is a quadratic function of r (N ), two values of r (N ) make the determinant equal
to zero. These two values of r (N ) define boundaries within which the predicted value of
r (N ) must fall. The MEM procedure seeks to select the value of r (N ) that maximizes
det(R N ). For a Gaussian random process this procedure is equivalent to maximizing (11.8)
subject to the constraint equations [16]:
 W
r (n) = φxx ( f ) exp( j2π f nt)d f, −N ≤ n ≤ N (11.12)
−W

Maximizing det(R N ) is equivalent to finding coefficients for a prediction error filter,

and these coefficients play an important role in finding the MEM spectral estimate. Before
finding these coefficients, it is instructive to examine the role played by a prediction
error filter in obtaining the MEM spectral estimate for a partially known autocorrelation
function. Suppose there are N samples of x(t) denoted by x0 , x1 , . . . , x N −1 where each
of the samples is taken t seconds apart. A linear predicted estimate of x N based on the
previous sampled values of x(t) is obtained from an (N + 1)-point prediction filter as
follows:

N
x̂ N = − a(N , i)x N −i (11.13)
i=1

The error associated with x̂ N is then given by


N
ε N = x N − x̂ N = x N + a(N , i)x N −i (11.14)
i=1

Equation (11.14) can be written in matrix form as

ε N = aTN x (11.15)

where xT = [x N , x N −1 , . . . , x0 ], and aTN = [1, a(N , 1), a(N , 2), . . . , a(N , N )], where
the coefficient a(N , N ) = C N is called the reflection coefficient of order N . The error
ε N is regarded as the output of an Nth order prediction error filter whose coefficients are
given by the vector a N and whose power output is
 
PN = E ε2N (11.16)

It is desirable to minimize the MSE of (11.16) by appropriately selecting the prediction

filter coefficients contained in a N . To obtain the minimum mean square estimate x̂ N , the
error ε N must be orthogonal to the past data so that

E{xi ε N } = 0 for i = 0, 1, . . . , N − 1 (11.17)

Furthermore, the error ε N must be uncorrelated with all past estimation errors so that

E{ε N ε N −k } = 0 for k = 1, 2, . . . , N − 1 (11.18)

Equations (11.16) and (11.8) are the conditions required for a random process to have
a white power spectrum of total power PN (or a power density level of PN /2W where
W = 1/2t). The prediction error filter is regarded as a whitening filter that operates
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 429

11.4 The Maximum Entropy Method 429

FIGURE 11-6
Input time series Whitening Output time series Whitening filter
[Spectrum fxx ( f )] filter (white spectrum) representation of
prediction error filter.

on the input data {x0 , x1 , . . . , x N −1 } to produce output data having a white power density
spectrum of level PN /2W as indicated in Figure 11-6. It immediately follows that the
estimate φ̂xx ( f ) of the input power spectrum φxx ( f ) is given by
PN /2W
φ̂xx ( f ) =  2 (11.19)
 
N 
1 + a(N , n) exp(− j2πfnt)
 
n=1

where the denominator of (11.19) is recognized as the power response of the predic-
tion error filter. Equation (11.19) yields the MEM estimate of φxx ( f ) provided that the
coefficients a(N , n), n = 1, . . . , N , and the power PN can be determined.
A relationship among the coefficients a(N , n), n = 1, . . . , N , the power PN , and
the autocorrelation function values r (−N ), r (−N + 1), . . . , r (0), . . . , r (N − 1), r (N ) is
provided by the well-known prediction error filter matrix equation [14]
⎡ ⎤⎡ ⎤ ⎡ ⎤
r (0) r (−1) ... r (−N ) 1 PN
⎢ r (1) r (0) . . . r (−N + 1) ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ a(N , 1) ⎥ ⎢ 0 ⎥
⎢ .. .. .. ⎥⎢ .. ⎥ = ⎢ .. ⎥ (11.20)
⎣ . . . ⎦⎣ . ⎦ ⎣ . ⎦
r (N ) r (N − 1) . . . r (0) a(N , N ) 0
Equation (11.20) can be derived [as done in [16] by maximizing the entropy of (11.8)]
subject to the constraint equations (11.12). If we know the autocorrelation values {r (−N ),
r (−N + 1), . . . , r (−1), r (0), r (1), . . . , r (N − 1), r (N )}, the coefficients a(N , n) and the
power PN may then be found using (11.20). Equation (11.20) can be written in matrix
form as
⎡ ⎤ ⎡ ⎤
PN PN
⎢ 0 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
Rnn a N = ⎢ . ⎥ or a N = R−1 nn ⎢ .. ⎥ (11.21)
.
⎣ . ⎦ ⎣ . ⎦
0 0
Let
⎡ ⎤
z 11 z 12 ...
⎢ z 21 z 22 ⎥
⎢ ⎥
R−1 =⎢ . ⎥ (11.22)
nn
⎣ .. ⎦
zN1
It then follows that
 
z 21 z 31 zN1
aTN = 1, , ,..., (11.23)
z 11 z 11 z 11
Ulrych and Bishop [6] also give a convenient recursive procedure for determining the
coefficients in a N .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 430

430 CHAPTER 11 Direction of Arrival Estimation and Related Topics

Having determined the prediction error filter coefficients and the corresponding MEM
spectral estimate, we must now consider how the autocorrelation function can be extended
beyond r (N ) to r (N + 1) where the autocorrelation values r (0), r (1), . . . , r (N ) are all
known. Suppose for example that r (0) and r (1) are known, and it is desired to extrapolate
the autocorrelation function to the unknown value r (2). The prediction error filter matrix
equation for the known values r (0) and r (1) is given by
    
r (0) r (−1) 1 P1
= (11.24)
r (1) r (0) a(1, 1) 0

where r (−1) = r ∗ (1). To determine the estimate r̂ (2), append one more equation to the
previous matrix equation by incorporating r̂ (2) into the autocorrelation matrix to yield [6]
⎡ ⎤⎡ ⎤⎡ ⎤
r (0) r (−1) r̂ (−2) 1 P1
⎣ r (1) r (0) r (−1) ⎦ ⎣ a(1, 1) ⎦ ⎣ 0 ⎦ (11.25)
r̂ (2) r (1) r (0) 0 0

Solving (11.25) for r̂ (2) then yields r̂ (2) + a(1, 1)r (1) = 0 or

r̂ (2) = −a(1, 1)r (1) = −C1r (1) (11.26)

Continuing the extrapolation procedure to still more unknown values of the autocorrelation
function simply involves the incorporation of these additional r (n) into the autocorrelation
matrix along with additional zeros appended to the two vectors to give the appropriate
equation set that yields the desired solution. Since the prediction error filter coefficients
remain unchanged by this extrapolation procedure, it follows that the spectral estimate
given by (11.19) remains unchanged by the extrapolation as well.

11.4.2 Unknown Autocorrelation Function

The discussion of the previous section assumed that the first N lag values of the autocorre-
lation function were precisely known. Quite often, however, any information concerning
the autocorrelation function is estimated from the time series data. One estimate of the
autocorrelation function is the average given by
N −τ
1 ∗
r̂ (τ ) = xi x + τ (11.27)
N i=1 i

With finite data sets, however, (11.27) implicitly assumes that any data outside the finite
data interval are zero. The application of Fourier transform techniques to a finite data
interval likewise assumes that any data that may exist outside the data interval are peri-
odic with the known data. These unwarranted assumptions about unknown data represent
“end effect” problems that may be avoided using the MEM approach, which makes no
assumptions about any unmeasured data.
The MEM approach to the problem of spectral estimation when the autocorrelation
function is unknown estimates the coefficients of a prediction error filter that never runs
off the end of a finite data set, thereby making no assumptions about data outside the data
interval. The prediction error filter coefficients are used to estimate the maximum entropy
spectrum. This approach exploits the autocorrelation reflection–coefficient theorem, which
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 431

11.4 The Maximum Entropy Method 431

FIGURE 11-7
a (1, 1) 1 Forward filter
Two-point prediction
error filter operating
x1 x2 x3 x4 xN−3 xN−2 xN−1 xN forward and
backward over an
Backward filter 1 a (1, 1) N-point data set.

states that there is a one-to-one correspondence between an autocorrelation function and

the set of numbers {r (0), C1 , C2 , . . . , C N } [16].
To obtain the set {r (0), C1 , C2 , . . . , C N } first consider estimating r (0) as the average
square value of the data set, that is,

1 
N
r̂ (0) = |xi |2 (11.28)
N i=1

Now consider how a two-point prediction error filter coefficient can be estimated from an
N-point long data sample. The problem is to determine the two-point filter (having a first
coefficient of unity) that has the minimum average power output where the filter is not
run off the ends of the data sample. For a two-point filter running forward over the data
sample as shown in Figure 11-7, the average power output is given by
N −1
f 1 
P1 = |xi+1 + a(1, 1)xi |2 (11.29)
N − 1 i=1

Since a prediction filter operates equally well running backward over a data set as well as
forward, the average power output for a backward running two-point filter is given by
N −1
1 
P1b = |xi + a ∗ (1, 1)xi+1 |2 (11.30)
N − 1 i=1

Since there is no reason to prefer a forward-running filter over a backward-running filter and
since (11.29) and (11.30) represent different estimates of the same quantity, averaging the
two estimates should result in a better estimator than either one alone [and also guarantees
that the estimate of the reflection coefficient a(1, 1) is bounded by unity—a fact whose
significance will be seen shortly] so that
1 f 
P1 = P1 + P1b
2  N −1 
 
N −1
1 ∗
= |xi+1 + a(1, 1)xi | +
2
|xi + a (1, 1)xi+1 | 2
(11.31)
2(N − 1) i=1 i=1

Now, minimize P1 by selecting the coefficient a(1, 1). Setting the derivative of P1
with respect to a(1, 1) equal to zero shows that the minimizing value of a(1, 1) is given
by [18]
N
−1
−2 xi∗ xi+1
i=1
a(1, 1) = N
−1
(11.32)
(|xi |2 + |xi+1 |2 )
i=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 432

432 CHAPTER 11 Direction of Arrival Estimation and Related Topics

Having r̂ (0) and a(1, 1), we can find the remaining unknown parts of the prediction error
filter matrix equation using (11.20), that is,
    
r̂ (0) r̂ (−1) 1 P1
= (11.33)
r̂ (1) r̂ (0) a(1, 1) 0
so that
r̂ (1) = −a(1, 1)r̂ (0) (11.34)
and the output power of the two-point filter is given by
P1 = r̂ (0)[1 − |a(1, 1)|2 ] (11.35)
The result expressed by (11.35) implies that |a(1, 1)| ≤ 1, which is also the necessary and
sufficient condition that the filter defined by {1, a(1, 1)} be a prediction error filter.
Next, consider how to obtain the coefficients for a three-point prediction error filter
from the two-point filter just found. The prediction error filter matrix equation takes the
form
⎡ ⎤⎡ ⎤ ⎡ ⎤
r̂ (0) r̂ (−1) r̂ (−2) 1 P2
⎣ r̂ (1) r̂ (0) r̂ (−1) ⎦ ⎣ a(2, 1) ⎦ = ⎣ 0 ⎦ (11.36)
r̂ (2) r̂ (1) r̂ (0) a(2, 2) 0
From the middle row it follows that

r̂ (1) + a(2, 1)r̂ (0) + a(2, 2)r̂ ∗ (1) = 0 (11.37)

On substituting (11.34) into (11.37) it immediately follows that

a(2, 1) = a(1, 1) + a(2, 2)a ∗ (1, 1) (11.38)

The corresponding output power P2 is then

P2 = P1 (1 − |a(2, 2)|2 ) (11.39)

Consequently, the coefficient vector for the three-point filter takes the form

a2T = [1, a(1, 1) + a(2, 2)a ∗ (1, 1), a(2, 2)] (11.40)

An equation for the estimatedpower output

 of a three-point filter can be written corre-
f
sponding to (11.31) as P2 = 12 P2 + P2b where now
N −2
f 1 
P2 = |xi+2 + a(2, 1)xi+1 + a(2, 2)xi |2 (11.41)
N − 2 i=1
N −2
1 
P2b = |xi + a ∗ (2, 1)xi+1 + a ∗ (2, 2)xi+2 |2 (11.42)
N − 2 i=1

Since a(2, 1) is given by (11.38) and a(1, 1) is already known, it follows that the minimiza-
tion of P2 is carried out by varying only a(2, 2) = C2 . As was the case with the two-point
filter, the magnitude of the three-point filter coefficient a(2, 2) must not exceed unity. On
minimizing of P2 , it turns out that |a(2, 2)| ≤ 1, which is the necessary and sufficient
condition that the filter defined by {1, a(1, 1) + a(2, 2)a ∗ (1, 1), a(2, 2)} be a prediction
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 433

11.4 The Maximum Entropy Method 433

error filter. The corresponding Hermitian Toeplitz autocorrelation matrix is nonnegative

definite as required by the basic autocorrelation function theorem.
The four-point filter can likewise be formed from the three-point filter by carrying
out the minimization of P3 with respect to a(3, 3) only. The aforementioned procedure
continues until all the coefficients a(N , n), n = 1, . . . , N , are found, and (11.19) then
yields the MEM spectral estimate. The general solution for the reflection coefficients
C N = a(N , N ) is given in [18]. Equations (11.38) and (11.39) may be expressed in
general terms as

a(N , k) = a(N − 1, k) + a(N , N )a ∗ (N − 1, N − k) (11.43)

PN = PN −1 [1 − |a(N , N )|2 ] (11.44)

Since |a(N , N )| ≤ 1, it follows that 0 ≤ PN ≤ PN −1 , so the error decreases with increasing

filter order N. The choice of N is determined by the desired resolution of the estimated
spectrum.
Burg [8] showed that the inverse of the covariance matrix Rxx is determined from
the prediction error filter coefficients a(m, k) and their corresponding error powers Pk
for k = 1, 2, . . . , L, where L ≤ N − 1 is the filter length. The resulting MEM estimate
of R−1xx differs from the inverse of the sample covariance matrix used in Chapter 6, and
some evidence suggests that the MEM estimate R̂−1 −1
xx converges to Rxx more rapidly than
the direct matrix inversion (DMI) estimate. This fast convergence feature makes the Burg
algorithm particularly attractive in situations where only a small number of independent
data samples are available.

11.4.3 Extension to Multichannel Spectral Estimation

Since the processing of array sensor outputs involves multichannel complex signals, it is
important to generalize the single-channel Burg process to multiple channels to exploit the
MEM procedure for adaptive array applications. Several multichannel generalizations of
the scalar MEM approach have been proposed [20–26], and the generalization provided
by Strand [27,28] is outlined here.
Suppose there are N observations of the p-channel vector x(t) denoted by {x0 , x1 , . . . ,
x N −1 }. A linear prediction of x N based on the previous observations of x(t) can be
obtained as

N
x̂ N = − A† (N , i)x N −i (11.45)
i=1

where A(N , i) now denotes the matrix of N-long forward prediction filter coefficients.
The error associated with x̂ N is then given by


N
ε N = x N − x̂ N = x N + A† (N , i)x N −i = A†N x f (11.46)
i=1

where

A†N = [I, A† (N , 1), A† (N , 2), . . . , A† (N , N )]

 
xTf = xTN , xTN −1 , . . . , x0T
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 434

434 CHAPTER 11 Direction of Arrival Estimation and Related Topics

A prediction error filter running backward over a data set in general will not have the
same set of matrix prediction filter coefficients as the forward-running prediction filter, so
the error vector associated with a backward-running prediction filter is denoted by


N
b N = x0 − x̂0 = x0 + B† (N , i)xi = B†N xb (11.47)
i=1

The fact that A(N , i) = B† (N , i) reflects the fact that the multichannel backward predic-
tion error filter is not just the complex conjugate time reverse of the multichannel forward
prediction error filter (as it was in the scalar case).
The matrix generalization of (11.20) is given by
⎡ ⎤⎡ ⎤
R(0) R(−1) ... R(−N ) I
⎢ R(1) . . . R(−N + 1) ⎥ ⎢ ⎥
⎢ R(0) ⎥ ⎢ A(N , 1) ⎥
R f AN = ⎢ . ⎥ ⎢ . ⎥
⎣ .. ⎦⎣ .. ⎦
R(N ) R(N − 1) . . . R(0) A(N , N )
⎡ f ⎤
PN
⎢ 0 ⎥
⎢ ⎥
=⎢ . ⎥ (11.48)
⎣ .. ⎦
0
where the p × p block submatrices R(k) are defined by

R(k) = E{x(t)x† (t − kt)}, k = 0, 1, . . . , N (11.49)

so that

R(−k) = R† (k) (11.50)

f
The forward power matrix P N for the prediction error filter satisfying (11.48) is given
by

P N = E{ε N ε †N } = A†N R f A N
f
(11.51)

The optimum backward prediction error filter likewise satisfies

⎡ ⎤⎡ ⎤
R(0) R(1) ... R(N ) I
⎢ R(−1) . . . R(N − 1) ⎥ ⎢ ⎥
⎢ R(0) ⎥ ⎢ B(N , 1) ⎥
Rb B N = ⎢ .. ⎥⎢ .. ⎥
⎣ . ⎦⎣ . ⎦
R(−N ) R(−N + 1) . . . R(0) B(N , N )
⎡ b ⎤
PN
⎢ 0 ⎥
⎢ ⎥
=⎢ . ⎥ (11.52)
⎣ .. ⎦
0

The backward power matrix PbN for the prediction error filter satisfying (11.52) is then

PbN = E{b N b†N } = B†N Rb B N (11.53)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 435

11.4 The Maximum Entropy Method 435

The matrix coefficients A(N , N ) and B(N , N ) are referred to as the forward and backward
reflection coefficients, respectively, as follows:
f
C N = A(N , N ) and CbN = B(N , N ) (11.54)
The maximum entropy power spectral density matrix can be computed either in terms
of the forward filter coefficients using
  †   
1 1
xx ( f ) = t A−1 P N A−1
f
(11.55)
z z
where
A(z) = I + A(N , 1)z + · · · + A(N , N )z N (11.56)

and z = e− j2π f t , or in terms of the backward filter coefficients using

xx ( f ) = t[B−1 (z)]† PbN [B−1 (z)] (11.57)

where

B(z) = I + B(N , 1)z + · · · + B(N , N )z N (11.58)

The forward and backward reflection coefficients are furthermore related by

f  −1 † f  f −1 f †
C N = PbN −1 CbN P N −1 or CbN = P N −1 C N PbN −1 (11.59)

If an observation process has collected Nd consecutive vector samples {x1 , x2 , . . . , x Nd },

then a prediction error filter of length N will have M = Nd − N consecutive (N +1)-tuples
from the data set to operate on. It is convenient to array each (N + 1)-tuple of data as an
extended vector defined by
⎡ ⎤
xm+N
⎢ xm+N −1 ⎥
⎢ ⎥
N  ⎢ .. ⎥ N = 0, 1, . . . , Nd − 1
xm = ⎢ . ⎥, (11.60)
⎢ ⎥ m = 1, 2, . . . , M = Nd − N
⎣ xm+1 ⎦
xm
The residual outputs of the forward error filter (denoted by um ) and the backward
error filter (denoted by vm ) when these filters are applied to the mth (N + 1)-tuple of data
are given by
 .  f † .  f†
um = A†N −1 .. 0 xmN + C N 0 .. B†N −1 xmN = ε mN + C N bmN (11.61)
† .   . 
vm = CbN A†N −1 .. 0 xmN + 0 .. B†N −1 xmN = CbN εmN + bmN
†
(11.62)

Using (11.59), however, we can rewrite (11.62) as

vm = PbN −1 C N P−1
f
N −1 ε m + bm
N N
(11.63)

Equations (11.61) and (11.63) show that the forward and backward prediction error
f
filter residual outputs depend only on the forward reflection coefficient C N . The coefficient
f
C N are chosen to minimize a weighted sum of squares of the forward and backward residual
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 436

436 CHAPTER 11 Direction of Arrival Estimation and Related Topics

 f
outputs of the filter of length N; that is, minimize SS C N where

 f 1 M
  f −1  −1 
SS C N = K m u†m P N −1 um + v†m PbN −1 vm (11.64)
2 m=1

and K m is a positive scalar weight = 1/M. The equation that yields the optimum value of
f
C N for (11.64) is then
f f  f −1
BC N + PbN −1 C N P N −1 E = −2G (11.65)

where



M
†
B= K m bmN bmN (11.66)
m=1


M
†
E= K m ε mN εmN (11.67)
m=1


M
†
G= K m bmN εmN (11.68)
m=1

f
After obtaining C N from (11.65) (which is a matrix equation of the form AX + XB = C)
then CbN can be obtained from (11.59), and the desired spectral matrix estimate can be
computed from (11.55) or (11.57).

11.5 COMPARING DOA ALGORITHM SPECTRA

The various high-resolution DOA approaches discussed so far have much better resolution
than a periodogram. Figure 11-8 compares the DOA algorithms when the three signals
incident on the 12-element uniform array at θ = −50◦ , 10◦ , and 20◦ have relative signals

FIGURE 11-8 10
Comparison of the MUSIC
DOA algorithms
Periodogram
when the three
signals incident on Capon
0
the 12-element MEM
uniform array at θ =
−50◦ , 10◦ , and 20◦
|P (dB)|

have relative signals −10

strengths of 16, 1,
and 4, respectively.

−20

−90 −45 0 45 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 437

11.6 Sequential Implementation of Bayes Optimal Array Processor 437

10 FIGURE 11-9
MUSIC
Comparing four DOA
techniques when the
Periodogram
standard deviation
Capon of the noise is 2.
0
MEM
|P (dB)|

−10

−20

−90 −45 0 45 90
q (degrees)

strengths of 16, 1, and 4, respectively. The lowest amplitude signal is almost lost in the
periodogram and is off by several degrees from the true signal position.
Increasing the noise variance causes the noise floor of the periodogram and Capon
spectrum to rise as shown in Figure 11-9. The Capon spectrum barely distinguishes be-
tween the two closely spaced signals, and the peaks no longer accurately reflect the signal
strengths. The noise floor of the MUSIC spectrum goes up, but it has stronger peaks than
the MEM spectrum.

11.6 SEQUENTIAL IMPLEMENTATION OF BAYES

OPTIMAL ARRAY PROCESSOR
It was noted in Chapter 3 that when the overall goal of good signal detection is the principal
concern then array processors based on the likelihood ratio are optimum in the sense of
minimizing the risk associated with an incorrect decision concerning signal presence
or absence. Furthermore, when the signal to be detected has one or several uncertain
parameters (due to, e.g., location uncertainty) then the sufficient test statistic for decision
can be obtained from the ratio of marginal probability density functions in the following
manner.
Let the observation vector x consist of the observed random processes appearing at
the array elements (or the Fourier coefficients representing the random processes). If any
uncertain signal parameters are present, model these parameters as additional random
variables, and summarize any prior knowledge about them with an a priori probability
density function p(θ ). The likelihood ratio is written as the ratio of the following marginal
probability density functions:

p(x/θ , signal present) p(θ)dθ
(x) = (11.69)
p(x/signal absent)
where θ ∈ . Equation (11.69) can be implemented by means of a suboptimal “one-shot
array processor” [29] for which a block diagram is given in Figure 11-10.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 438

438 CHAPTER 11 Direction of Arrival Estimation and Related Topics

FIGURE 11-10 x1 (t)

Block diagram for
one-shot array x2 (t)
processor. Array
elements Decision
signals Λ(x)
statistic

xN (t)

p(q )

11.6.1 Estimate and Plug Array Processor

An intuitively appealing approach to the array detection problem when uncertain signal
parameters exist is to directly estimate these parameters and plug them into the conditional
likelihood ratio as though they were exactly known. A block diagram of such an estimate
and plug array processor is given in Figure 11-11. Quite naturally, the merit of any estimate
and plug structure depends on the accuracy of the estimates generated by the parameter
estimation scheme, and the question of the processor’s sensitivity to mismatch between
the true parameter values and those assumed (or estimated) has received some attention
[30,31].
Two popular approaches to obtaining estimates of random signal parameters for use
in an estimate and plug array processor are the ML approach and the Bayes optimum
approach. In the ML approach, the maximum likelihood estimate (MLE) of the unknown
signal parameters is formed by solving

∂ p(x/θ , signal present) 
 =0 (11.70)
∂θ θ =θ̂ MLE

and then using the resulting estimate in the likelihood ratio test statistic as if it were known
exactly. The Bayes approach to the parameter estimation problem incorporates any a priori
knowledge concerning the signal parameters in the form of a probability density function

FIGURE 11-11 x1 (t)

Block diagram of
estimate and plug x2 (t)
array processor. Decision
Delay Λ(x/q )
statistic

xN (t)

qˆ

Parameter
estimation
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 439

11.6 Sequential Implementation of Bayes Optimal Array Processor 439

p(θ /signal present). To obtain a signal parameter estimate for an estimate and plug array
processor, note that

(x) p(x/θ, signal present) p(θ /signal present)dθ
= (11.71)
(x/θ̂) p(x/θ̂ , signal present)
The optimal Bayes processor explicitly incorporates a priori knowledge about the unknown
signal parameters θ into the likelihood ratio (x) through the averaging process expressed
by the numerator of (11.71). To obtain an estimate θ̂ to use in a suboptimal estimate and
plug structure, require
(x)
=1 (11.72)
(x/θ̂ )

Having evaluated (x) using the averaging process, we find θ̂ as the solution to (11.72),
and it is referred to as a “pseudo-estimate” θ̂ PSE [32]. The performance of the Bayes optimal
processor for the case of a signal known except for direction (SKED) was investigated
by Gallop and Nolte [33]. A comparison between the two estimate and plug structures
obtained with a MLE and a Bayes pseudo-estimate is given in [34] for the case of target
location unknown. The results indicated that the ML detector performs the same as the
Bayes pseudo-estimate detector when the a priori knowledge about the uncertain parameter
is uniformly distributed. When the a priori knowledge available is more precise, however,
the performance of the Bayes pseudo-estimate detector improves whereas that of the
ML detector does not, and this difference between the two processors becomes more
pronounced as the array size becomes larger.

11.6.2 Sequential Optimal Array Processor

When implemented in the form of an estimate and plug structure, the Bayes optimal array
processor processes all the observed data at the same time. By implementing the same
processor sequentially, the resulting array processor will exhibit adaptive (or learning)
features naturally.
To see how to implement a Bayes optimal processor sequentially, let xi denote the
vector of observed outputs (or the Fourier coefficients thereof) from the array elements for
the ith sample period. The sequence {x1 , x2 , . . . , x L } then represents L different observation
vector samples. The joint probability density function of the observed vector sequence can
be written as

L
p(x1 , x2 , . . . , x L ) = p(xi /xi−1 , . . . , x1 ) (11.73)
i=1

When unknown signal parameters are present, application of the averaging process
described in the previous section to p(θ ) yields
 
L
P(x , x , . . . , x ) =
1 2 L
p(xi /xi−1 , . . . , x1 , θ ) p(θ )dθ (11.74)
i=1

Now assume parameter conditional independence of the various xi so that

p(xi /xi−1 , . . . , x1 , θ ) = p(xi /θ) (11.75)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 440

440 CHAPTER 11 Direction of Arrival Estimation and Related Topics

and then it follows that (11.74) can be rewritten as

 
L
P(x1 , x2 , . . . , x L ) = p(xi /θ) p(θ )dθ (11.76)
i=1

According to Bayes’s rule it follows that

p(x1 /θ) p(θ )
p(θ/x1 ) = (11.77)
p(x1 )
p(x2 /θ) p(θ /x1 )
p(θ/x1 , x2 ) = (11.78)
p(x2 /x1 )
..
.
p(xi−1 /θ) p(θ /xi−2 , . . . , x1 )
p(θ /xi−1 , . . . , x1 ) = (11.79)
p(xi−1 /xi−2 , . . . , x1 )
so that
L 

p(x , . . . , x ) =
1 L
p(xi /θ) p(θ /xi−1 , . . . , x1 )dθ (11.80)
i=1

where p(θ/xi−1 , . . . , x1 ) represents the updated form of the a priori information contained
in p(θ ).
A sequential processor may now be implemented using (11.79) and (11.80) to form
the marginal density functions required in the likelihood ratio as follows:
p(x1 , . . . , x L /signal present)
(x1 , . . . , x L ) = (11.81)
p(x1 , . . . , x L /signal absent)
A block diagram of the resulting sequential array processor based on (11.79)–(11.81) is
given in Figure 11-12. The sequential Bayesian updating of p(θ) represented by (11.79)
results in an optimal processor having an adaptive capability.
Performance results using an optimal sequential array processor were reported in [29]
for a detection problem involving a signal known exactly imbedded in Gaussian noise
where the noise has an additive component arising from a noise source with unknown
direction. The adaptive processor in this problem must both succeed in detecting the
presence or absence of the desired signal and in “learning” the actual direction of the

x1i
L

∫Θ p(x /q, signal present) p(q/x

i i− 1, …, x1, signal present) dq Π
i=1
p(xi /xi − 1, …, x1, signal present)
Λ(x1, …, xL)
L
xNi p(x i/signal absent) Π p(xi /xi − 1, …, x1, signal absent)
i=1

p(q/xi− 1, …, x1, signal present)

FIGURE 11-12 Bayes optimal sequential array processor.

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 441

11.7 Parameter Estimation via Subspace Fitting 441

noisy signal source. The results obtained indicated that even though the directional noise to
thermal noise ratio was relatively low, the optimal sequential processor could nevertheless
determine the directional noise source’s location.

11.7 PARAMETER ESTIMATION

VIA SUBSPACE FITTING
The estimation of signal parameters via rotational invariance techniques (ESPRIT) [35–37]
is a DOA estimation exploiting a known array structure. The method described here is a
total least square (TLS) version of the ESPRIT approach to the problem, and extensions
of this method may readily be seen when the problem formulation is cast in a subspace
fitting framework [38]. It has been found that this approach provides estimates for which
the parameter estimate variance coincides with the Cramer–Rao bound.
The output of the ith sensor of an array can be represented by

d
xi (t) = ai (θ j )s j (t) + n i (t), i = 1, . . . , k (11.82)
j=1

where ai (θ j ) is a complex scalar representing the sensor response to the jth emitter signal
(there are a total of “d” signals). The jth emitter signal is denoted by s j (t), and the additive
noise is n i (t). In matrix notation, (11.82) can be written as

x(t) = [a(θ1 ) · · · a(θd )]s(t) + n(t) = A(θ () )s(t) + n(t) (11.83)

where θ () is a d-dimensional parameter vector corresponding to the actual DOAs, and

A(θ() ) is a k × d response matrix where the vector a(θ j ) = [a1 (θ j ) . . . ak (θ j )]T contains
the sensor responses to a unit wavefront from the direction θ j . The array output is sampled
at N time instances, and this collection of samples are arranged in the columns of a k × N
data matrix, X N , given by

X N = [x(1), . . . , x(N )] = A(θ () )s N + N N . (11.84)

The emitter covariance matrix is give by S = E[s(t)s H (t)], where “E” denotes the expected
value. Likewise, the output signal covariance matrix is given by

Rxx = E[x(t)x H (t)] = A(θ () )SA(θ () ) H + σ 2 I (11.85)

The eigendecomposition of Rxx is given by

Rxx = Es s EsH + En n EnH (11.86)

where λ1 > · · · > λd > λd+1 = · · · = λk = σ 2 . The matrix Es = [e1 , . . . , ed ] contains

the d eigenvectors corresponding to the distinct signals. The range space of Es is called the
signal subspace, and its orthogonal complement is the noise subspace, which is spanned
by the columns of En = [ed+1 , . . . , ek ]. The eigendecomposition of the sample covariance
matrix estimate of Rxx then corresponds to (11.86) with estimates for each of the matrices
replacing the actual matrix values.
Equations (11.84)–(11.86) show that the parameter estimation problem is regarded
as a subspace fitting problem in which the subspace spanned by A(θ ) is fitted to the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 442

442 CHAPTER 11 Direction of Arrival Estimation and Related Topics

measurements X N in a least squares sense. The basic subspace fitting problem is defined
by
min
Â, T̂ = arg M − AT2F (11.87)
A, T

min
where A2F = trace(A H A) is the Frobenius norm, and θ̂ = arg V(θ) is the mini-
θ
mizing argument of V(θ ) The k × q matrix M in (11.87) represents the data, whereas T
is any p × q matrix. For a fixed A, the minimum with respect to T is a measure of how
well the range spaces of A and M match. The subspace fitting estimate selects A so these
subspaces are as close as possible. The estimate of θ is then obtained from the parameters
of Â. It is of some practical interest to note that the subspace fitting problem is separable in
A and T. By substituting the pseudo-inverse solution T̂ = A p M back into (11.87) where
A p = (A H A)−1 A H , one obtains the following equivalent problem
max
 = arg tr{P A MM H } (11.88)
A
where P A = AA p is a projection matrix that projects onto the column space of A. The
subspace fitting problem then resolves into a familiar parameter optimization problem
described by
max
θ̂ = arg tr{P A (θ )R̂xx } (11.89)
θ

where R̂xx is the sample covariance matrix. Notice from (11.88) that the same result could
be obtained by simply taking R̂xx = MM H . This approach is the deterministic maximum
likelihood method for obtaining direction-of-arrival estimates.
The ESPRIT algorithm assumes that the array is composed of two identical subarrays,
each having k/2 elements (so the total number of array elements is even). The subarrays
are displaced from each other by a known displacement vector so the propagation between
the subarrays can be described by the diagonal matrix  = diag[e jωτ1 e jωτ2 . . . e jωτe ],
where τi is the time delay in the propagation of the ith emitter signal between the two
subarrays, and ω is the center frequency of the emitters. The time delay is then related to
the angle of arrival by τi = || sin θi /c, where c is the speed of propagation, and  is the
displacement vector between the two subarrays. The output of the array is then modeled as
   
 n (t)
x(t) = s(t) + 1 (11.90)
 n 2 (t)
where the k/2 × d matrix  contains the common array manifold vectors of the two
subarrays, and  is the diagonal matrix discussed already. Since the matrices defined by
[E1T E2T ]T (eigenvectors for the two subarrays) and [ T T  T ] have the same range space,
there is a full rank d × d matrix T such that
   
E1 
= T (11.91)
E2 
Eliminating  in (12.91) then yields

E2 = E1 T−1 T = E1  (11.92)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 443

11.8 The Maximum Likelihood Estimate 443

Using the eigendecomposition

 
Ê1H   V V  VH VH 
1 2 =
11 12
Ê Ê L 11H 21H (11.93)
Ê2H V21 V22 V12 V22

where L = diag[l1 , l2 , . . . , l2d ]. Since  = T−1 T, the elements of  are estimated by
the eigenvalues of ˆ TLS . The principal angles of these eigenvalues give estimates of the
time delays τi , which in turn give the DOA estimates. Solving for  and equating its
eigenvalues to  results in the signal angle estimates of
 
arg(λm)
θm = sin−1 (11.94)
kd
where λ
m = eigenvalues of .
⎡ ⎤
−0.0816 +j0.0000 −0.3214 −j0.0000 0.3239 −j0.0000
⎢ −0.0214 +j0.1413 0.1728 −j0.1651 0.2693 −j0.3461 ⎥
⎢ ⎥
⎢ 0.0775 +j0.2593 −0.0428 +j0.3203 −0.0252 −j0.2482 ⎥
⎢ ⎥
⎢ 0.3061 +j0.0678 −0.1822 −j0.1796 −0.2547 −j0.1985 ⎥
⎢ ⎥
⎢ 0.3649 −j0.1551 0.2694 +j0.1027 −0.0855 +j0.0202 ⎥
⎢ ⎥
⎢ ⎥
⎢ 0.0744 −j0.3899 −0.2220 +j0.1963 −0.0774 +j0.0945 ⎥
Vs = ⎢ ⎥
⎢ −0.1733 −j0.3619 0.1385 −j0.2273 −0.0089 −j0.1292 ⎥
⎢ ⎥
⎢ −0.3708 −j0.0374 0.1865 +j0.2484 −0.0916 −j0.0330 ⎥
⎢ ⎥
⎢ −0.2907 +j0.1344 −0.2014 −j0.1534 −0.3170 +j0.0412 ⎥
⎢ ⎥
⎢ −0.0037 +j0.2384 0.2845 −j0.1651 −0.0919 +j0.2447 ⎥
⎢ ⎥
⎣ 0.0553 +j0.1390 −0.1668 +j0.1730 0.0681 +j0.4281 ⎦
0.0542 −j0.0455 −0.1122 −j0.2981 0.3375 +j0.1231
Next, compute  to get
⎡ ⎤
0.6322 −j0.6792 −0.1035 +j0.2722 0.2247 −j0.0474
 = ⎣ −0.2434 −j0.0108 −0.6666 +j0.5904 −0.3027 +j0.3979 ⎦
−0.1525 +j0.2374 −0.2235 +j0.1595 0.6221 −j0.6479
The eigenvalues of ψ are found and substituted into (11.94) to find an estimate of the
angle of arrival.
 
θm = −50.15◦ 9.99◦ 20.02◦

11.8 THE MAXIMUM LIKELIHOOD ESTIMATE

The ML estimator has importance in estimation theory because it provides signal parameter
estimates that are in some sense optimal; that is, under certain conditions, it provides the
“best” (most accurate) estimate of a particular signal parameter. Examples of parameters
of interest include signal AOA, amplitude, phase, frequency, and time delay of arrival. The
following treatment is developed for AOA, although an estimate for the signal intensity
(received signal amplitude) is a by-product.
The precise form of the ML estimate depends on certain assumptions regarding the
signal model. Two signal models are considered: (1) the first model assumes the desired
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 444

444 CHAPTER 11 Direction of Arrival Estimation and Related Topics

signals to be Gaussian zero mean (“stochastic” signal model); and (2) the second model
assumes that signals are “deterministic” but unknown. We start with the stochastic model.

11.8.1 Mathematical Preliminaries

The maximum likelihood estimator and the derivation of the Cramer–Rao bound are based
on the Gaussian probability density function. First, the notion of the Gaussian density
function for the stochastic signal model is established.
Let xk be an N-dimensional complex vector representing the kth data sample of the
signals received by an N-element antenna array. The sample consists of a desired signal
component si and a noise component ni so that

xk = sk + nk (11.95)

The noise nk is assumed to be a sample function from a zero-mean N -variate complex

Gaussian process, with full rank covariance matrix
 
Rnn = E nk n†k (11.96)

For the stochastic signal model, the desired signal sk is also assumed to be a sample
function from a zero-mean N -variate complex Gaussian process, with covariance
 
Rss = E sk s†k (11.97)

Under these assumptions, the probability density for the data sample xk is given by [7]
 
p(xk ) = (π )−N |R|−1 exp −x†k R −1 xk (11.98)

where R is the N × N covariance matrix

R = E{xk x†k } (11.99)

It is assumed that the desired signal and noise are uncorrelated, so that

R = Rss + Rnn (11.100)

Our objective is to estimate the angle of arrival of the signal sk in the presence of
internal and external interference denoted by nk , based on K independent data samples
(“snapshots”) of xk , k = 1, . . . , K . The ML estimation procedure is based on determining
the value of the unknown parameters (parameters to be estimated) that maximizes the
conditional joint density function of the K independent data samples, which from (11.97)
has the form
 K 
 †
−N K −K −1
p(x1 , x2 , . . . x K |θ, S ) = (π ) |R| exp − xk R xk (11.101)
k=1

The density in (11.100) is conditional on the values of the unknown signal parameters to be
estimated, namely, the AOA (θ ) and the signal intensity (S). To maximize (11.100) with
respect to θ and S, Equation (11.100) must be reformulated to show explicit dependence
on these variables. For narrowband signals, where the element-to-element time delay
experienced by the desired signal can be represented as phase shifts of the signal, this
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 445

11.8 The Maximum Likelihood Estimate 445

equation reduces to
 K  
1 S K
† −1 † −1
P(x1 , x2 , . . . , x K |θ, S ) = C exp x R d(θ )d (θ )Rnn xk
1 + SB 1 + S B k=1 k nn
(11.102)
where
 

K
†
−N K −K −1
C = (π) |Rnn | exp − xk Rnn xk (11.103)
k=1
Rss = Sd(θ )d † (θ ) (11.104)
−1
B = d† (θ )Rnn d(θ ) (11.105)

Here the scalar C is a constant, the scalar B depends on θ , and the N × N covariance
matrix Rss depends on both S and θ. Equation (11.102) was derived under the narrowband
signal assumption, which permits the signal covariance matrix to be written as in (11.104).
The N -component vector d(θ) is the (unknown) vector of phase delays corresponding to
the signal angle of arrival. For the case of a linear array of identical isotropic antenna
elements (received signal power is the same in each element), S represents the (unknown)
received signal power at each element so that d(θ ) can be written as

d = [e jβ1 μ e jβ2 μ · · · e jβ N μ ]T (11.106)

where
μ = sin θ (11.107)
βn = 2π αn /λ (11.108)
The parameter αn is the relative location of the nth element along the line array, and
λ is the wavelength of the desired signal. Here, μ is the angle of arrival in sine space and
θ is the AOA relative to broadside of the linear array. A more general definition for d(θ )
will be given later in this section.
Equation (11.102) is the conditional joint density of the observations xk , k = 1, . . . , K ,
given the unknown parameters S and θ. When viewed as a function of the unknown
parameters, it is known as the likelihood function. The maximum likelihood estimate is
the value of the parameters that maximize the likelihood function, so the objective is to
obtain the values for S and θ that maximize the density p in (11.102). As will be shown,
closed-form solutions for θ and S are difficult to obtain in general, so numerical search
methods must often be used. Although one could search all possible values of θ and
S in Equations (11.102), (11.104), and (11.105) to find those values that maximize p,
a computationally intensive multidimensional search would be required. Fortunately, it
turns out that in this case the dependence of p on θ and S is separable, reducing to a search
only over θ . This result is developed subsequently.
Before proceeding further with the derivation, the treatment will now be generalized
to allow for multiple desired signals. The subsequent development follows that of Jaffer
[38]. Define M narrowband signal sources with angles of arrival denoted by θ1 , θ2 , . . . θ M .
The received data vector xk can be written as the sum of the M signals plus the noise term

M
xk = bkm dm + nk (11.109)
m=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 446

446 CHAPTER 11 Direction of Arrival Estimation and Related Topics

where bkm is the (complex) amplitude of the kth sample of the mth signal, and dm = dm (θm )
is the direction delay vector for the mth signal. The general form of d for the mth narrowband
signal source is given by [38]
 
dm (θm ) = g1 (θm )e− jφ1 (θm ) , g2 (θm )e− jφ2 (θm ) . . . g N (θm )e− jφ N (θm ) (11.110)

where θm is the angular location (angle of arrival) of the mth source, gn is the complex
gain of the nth antenna element (generally a function of θ ), and φn (θm ) is the phase delay
of the m th mth source at the nth antenna element relative to a suitable reference point (e.g.,
the array phase center).
It is convenient to rewrite (11.107) in vector/matrix notation as

xk = D(θ )bk + nk (11.111)

The M-dimensional complex vector bk denotes the signal amplitude and phase in complex
notation for each of the M signals for the kth snapshot. Under the stochastic signal model,
bk is assumed to be a zero-mean Guassian random vector with covariance

Rbb = E{bk b†k } (11.112)

The direction vectors for each of the M signals, dm (θm ), m = 1, . . . M, form the columns
of the N × M matrix D(θ ).

11.8.2 Maximum Likelihood Estimation of the Direction of Arrival

for Stochastic Signals
In the multiple signal case, Equation (11.100) retains its form
 

K
†
−NK −K −1
p(x1 , x2 , . . . x K |θ1 , θ2 , . . . , θ M , Rbb ) = (π ) |R| exp − xk R xk (11.113)
k=1

Maximizing (11.111) with respect to θ and Rbb is equivalent to maximizing ln( p),
denoted the log likelihood function

  
K
ln L (θ 1 , θ 2 , . . . , θ M , Rbb ) = C − K log |R| − x†k R −1 xk (11.114)
k=1

where the dependencies on θ and Rbb are contained in R

R = D(θ)Rbb D † (θ ) + Rnn (11.115)

The likelihood function in Equation (1.18) can be simplified by dropping the terms that
do not depend on either θ or Rbb and by rearranging terms

ln L(θ, Rbb ) = −K log |R| − K Tr(R −1 R̂) (11.116)

where R̂ is the N × N sample covariance matrix

1 
K
R̂ = x x† (11.117)
K k=1 k k
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 447

11.8 The Maximum Likelihood Estimate 447

As discussed already for single-signal case, maximization of (11.114) directly is com-

putationally intractable for all but the simplest cases, because it requires an M 2 + M
dimensional search to jointly optimize the components of θ and Rbb . Jaffer [38] reduced
this to an M-dimensional search of θ by showing that the maximum likelihood estimate
θ = θ̂ M L that maximizes L(θ, Rbb ) in Equation (11.114) can be obtained by maximizing
Jsto (θ ) with respect to θ1 , θ2 , . . . θ M
1    
Jsto (θ ) = Tr G(θ) R̂ − ln G(θ) R̂G(θ ) + σn2 (I N − G(θ )) (11.118)
σ 2

where

G(θ ) = D(θ )[D † (θ)D(θ)]−1 D † (θ ) (11.119)

As noted in [38], G(θ) is the orthogonal projection matrix. For the purpose of simplifica-
tion, the derived result in (11.116) made the assumption that the noise covariance matrix
Rnn is known and is given by Rnn = σn2 I N ; that is, the only noise terms are due to thermal
noise. The result in (11.116) is generalized to handle an arbitrary positive definite matrix
Rnn by transforming xk into a new coordinate system so that the noise covariance matrix
in the new coordinate system is σ 2 I N , that is,
−1/2
yk = Rnn xk (11.120)

Once θ̂ M L is determined, Jaffer [38] wrote the maximum likelihood estimate of Rbb as
      −1 †         −1
R̂bb θ̂ M L = D † θ̂ M L D θ̂ M L D θ̂ M L R̂ − σn2 I N D θ̂ M L D † θ̂ M L D θ̂ M L
(11.121)

In summary to this point, we have shown that the maximum likelihood estimate of the
angles of arrival of M signal sources is determined by finding those values of θ1 , θ2 , . . . , θ M
that maximize J (θ ) in Equation (11.116). This requires an M-dimensional search over the
angular regions of interest. To evaluate J (θ ), it is assumed that the following parameters
are known a priori: (1) antenna element gain versus θ; (2) antenna element location; and
(3) the noise covariance matrix Rnn = σn2 I N . The sample covariance matrix is computed
from the data samples (11.115). Note the result in (11.116) assumes that the signals are
narrowband and propagate as plane waves. Also note that equation (11.116) assumes that
Rnn is known a priori: in most practical situations, the external interference is not known
beforehand and must be estimated from the data samples. In radar, it is often possible
to estimate Rnn by averaging the sample covariance matrix using adjacent range cells
that do not contain the target (desired signal). Such an estimate is more difficult in a
communications system in which the signal is present in all the data samples. In this case,
it may be necessary to use an a priori estimate of Rnn . For example, if it is assumed that
Rnn is made up of only internal noise (Rnn = σn2 I N ), then all external sources must be
estimated using Equation (11.116).
The M-dimensional search required to find the values of θm that maximize J (θ ) can
become computationally intensive for large M. In practice, M is limited to two or three
sources by limiting the angular search region, typically one or two beamwidths in extent.
Any sources outside the search region (e.g., in the sidelobes) are treated as interference and
therefore must be included in Rnn , which as noted already must be known or estimated.
Methods for obtaining an accurate estimate of Rnn depend on the particular situation
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 448

448 CHAPTER 11 Direction of Arrival Estimation and Related Topics

(e.g., radar vs. communications vs. pulsed signals, amplitude of the desired signals),
which is beyond the scope of the current discussion.

11.8.3 Maximum Likelihood Estimation of the Direction of Arrival

for Deterministic Signals
The preceding section developed the maximum likelihood estimate of the AOAs of multiple
signals under the assumption that the signals are zero-mean Gaussian distributed. This
assumption is most appropriate when the received signal amplitude and phase tend to vary
sample to sample in a random manner, as in radar or signal monitoring systems.
In this section, we consider the case where the signal is modeled as “deterministic.”
This assumption is more appropriate for communications signals where sample-to-sample
variations are to be estimated and for preamble signals where the signal is known a priori.
As in the stochastic case, the data vector xk is given by Equation (11.111)
xk = D(θ )bk + nk
where the noise component n k is, as in the preceding section, assumed to be a sample
function from a zero-mean, Gaussian process with covariance matrix Rnn = E{nk n†k }. No
assumption is made about the statistics of sk = D(θ)bk . The conditional joint density of
K independent data samples xk , k = 1, 2, . . . K , is then expressed as

p(x1 , . . . x K |b1 , . . . b M , θ 1 , . . . θ M ) = (π )−MK |Rnn |−K

 K 

† −1
exp − (xk − D(θ )bk ) Rnn (xk − D(θ)bk ) (11.122)
k=1

In this case, the likelihood function p is maximized by minimizing the exponent, which
is a real scalar given by
1 
K
H= Hk (11.123)
K k=1
where
Hk = [bk − T −1 V xk ]† T [bk − T −1 V xk ] + x†k Rnn
−1
xk + x†k V † T −1 V xk (11.124)
−1
T = D † (θ)Rnn D(θ ) (11.125)
† −1
V = D (θ)Rnn

V is an M × N matrix, and the M × M matrix T is assumed positive definite so that Hk

is minimized with respect to bk when
−1
bk (θ ) = (D † (θ )Rnn D(θ))−1 D † (θ )Rnn
−1
xk (11.126)

Substituting bk into Equation (11.123) and deleting terms that do not depend on θ, one
obtains an expression for H that depends only on θ

1  † −1
K
−1
Jdet (θ ) = x R D(θ )(D † (θ )Rnn D(θ ))−1 D † (θ )Rnn
−1
xk (11.127)
K k=1 k nn

The ML estimates of the AOA of M narrowband signal sources received by an N -element

array, under the deterministic signal model, are the values of θ1 , θ2 , . . . θ M that maximize
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 449

11.9 Cramer–Rao Lower Bound on AOA Estimation Error 449

Jdet (θ). bk is then found by substituting the ML estimates of θk , k = 1, . . . , M into

Equation (11.126). Note that the noise covariance matrix Rnn assumed in the derivation
of Jdet (θ ) allows for external interference and thus is a more general result than Jsto (θ).
The result for deterministic signals, Jdet (θ) in Equation (11.127), can be compared
with the stochastic case, Jsto (θ) in Equation (11.116). For Rnn = σn2 I N , which was the
assumption in deriving Jsto (θ) in (11.116), Jdet (θ ) in (11.117) is the same as the first
term in Jsto (θ). The only difference between Jdet (θ ) and Jsto (θ), under the assumption
that Rnn = σn2 I N , is the second term on the right side of Equation (11.116). Experience
has shown that the second term in (11.116) is small relative to the first term for moderate
to high SNR, so accuracy of the two methods is very nearly the same. In general, AOA
accuracy is relatively insensitive to sample-to-sample signal variations.
It is instructive to reformulate Jdet (θ) as follows. Let
−1/2
u(θ ) = Rnn D(θ) (11.128)
and
−1/2
yk = Rnn xk (11.129)
Then Equation (11.127) becomes

1  †
K
Jdet (θ) = y U (θ )(U † (θ)U (θ))−1 U † (θ )yk (11.130)
K k=1 k

This has the same form as the first term in the expression for Jsto (θ ), and indicates how
the first term in (11.116) can be transformed to generalize it to any positive definite Rnn .

11.9 CRAMER–RAO LOWER BOUND ON AOA

ESTIMATION ERROR
The Cramer–Rao bound is derived for the general case of estimating the parameters of a
signal corrupted by Gaussian noise and interference. We will derive a generalized result and
then will specialize it to the estimation of the AOA. The CR bound, along with other bounds
such as Barankin and Ziv–Zakai, is useful for determining system design requirements
(e.g., antenna size, SNR, number of antenna elements N, number of snapshots K) necessary
to achieve a certain level of estimation accuracy. The CR bound is the best known in that
it often leads to closed-form solutions and provides a relatively tight bound under a broad
range of practical conditions [39,40], most notably in cases where the optimum array
SNR exceeds a certain level. The development will follow that of Ballance and Jaffer [39],
starting with a generalized version of the N-variate Gaussian probability density function
derived under the deterministic signal model.
Define the kth snapshot of the N-element antenna array as
xk = sk (υ) + nk (11.131)
where υ is an M-dimensional vector of real-valued, unknown, nonrandom parameters. The
Cramer–Rao lower bound on the variance of any unbiased estimate of the signal parameter
υm , m = 1, . . . M is given by the m, mth component of the inverse of the M × M Fisher
information matrix (FIM) [40]
συ2m ≥ [ −1 ]m,m (11.132)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 450

450 CHAPTER 11 Direction of Arrival Estimation and Related Topics

The M × M Fisher information matrix  is a symmetric matrix given by the expected

value of the second partial derivatives of the log likelihood function with respect to the
unknown parameters υ 1 , υ 2 , . . . υ M
 
∂ 2 ln L(υ1 , . . . υ M )
i j = E − (11.133)
∂υi ∂υ j

From Equation (11.100), the log likelihood function for the stochastic signal model is
given by


K
ln L(υ) = −NK ln(π ) − K ln |R| − x†k Rnn
−1
xk (11.134)
k=1

Now to simplify the development (a more general case will be derived later), assume
a single narrowband source with N × 1 direction vector d(μ), as in Equation (11.95).
Then the likelihood function reduces to Equation (11.101) and the log likelihood function
reduces to
S  †
−1 −1
ln L = ln C − K ln(1 + SB) + xk Rnn d(μ)d† (μ)Rnn xk (11.135)
(1 + SB)
where C is given by (11.103), B is given by (11.105), and S is the received signal power
as defined in (11.102). Taking the partial derivatives in Equation (11.102) with respect to
μ and S, one obtains for the 2 × 2 Fisher information matrix
 
∂ 2 ln L KS2  2 
11 = −E = F + 2(1 + SB)B 2 D
∂μ 2 (1 + SB) 2
 
∂ ln L
2
KSB
21 = −E = F = 12
∂ S∂μ (1 + SB)2
 
∂ 2 ln L KB2
22 = −E =
∂S 2 (1 + SB)2

Inverting , we obtain the final results for  −1

−1 1 + SB
σμ2 ≥ 11 = (11.136)
2KS2 B 2 Q
−1 −1 S −1
12 = 11 F = 21 (11.137)
B
2  
−1 −1 S
σ S2 ≥ 22 = 11 2
F 2 + 2(1 + SB)B 2 Q (11.138)
B
where Q and F are real scalar quantities given by
!  2 "
1 † † −1  † † −1 
Q = 2 Bd (μ)A Rnn Ad(μ) − d (μ)A Rnn d(μ)
B
−1 −1
F = d† (μ)Rnn Ad(μ) + d† (μ)A† Rnn d(μ)

and A is defined as the M × M diagonal matrix with diagonal elements given by Ann =
βn = 2π αn /λ; n = 1, . . . N .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 451

11.10 Fisher Information Matrix and CR Bound for General Cases 451

Equation (11.136) gives the CR lower bound on the variance of any unbiased estimate
of the angle of arrival θ, and Equation (11.137) gives the lower bound on the variance of
any unbiased estimate of the signal power S.
For an N-element line array of identical and equally spaced antenna elements, each
with unity gain, and assuming Rnn = σn2 I N (so S/σn2 = the signal-to-noise ratio of the
signal received by each element), the lower bound on the variance of the AOA estimate
about its true value is given by
   2
−1 6 1 1 + NS/σn2 1 λ
σμ ≥ E(μ̂ − μtr ue ) = 11 =
2 2
 2 (11.139)
(2π ) K NS/σn
2 2 (N − 1) α
2

where α is the separation between antenna elements. Note that NS/σn2 is the array SNR.
Equation (11.101) applies to the deterministic signal case, where the signal is unknown
but nonrandom. Ballance and Jaffer [39] showed that for the stochastic signal model
 2
−1 6 1 1 1 λ
σθ ≥ E(μ̂ − μtr ue ) = 11 =
2 2   (11.140)
(2π ) K NS/σn (N − 1) α
2 2 2

which is smaller than Equation (11.139) by the ratio

 
stochastic bound NS/σn2
=  (11.141)
deterministic bound 1 + NS/σn2

For moderate to large SNR (greater than approx 10 dB), the two bounds are nearly the
same.

11.10 FISHER INFORMATION MATRIX AND CR

BOUND FOR GENERAL CASES
This section gives an overview of results derived by Ballance and Jaffer [39] that give the
FIM for the time-varying, nonlinear case (i.e., sk may be time varying and be a nonlinear
transformation of υ). The FIM can then be inverted to determine the Cramer–Rao bound.
The specific result given here applies to the deterministic signal model. Starting with
the general equation for the data vector xk = sk (υ) + nk , and the general form for the
likelihood function in Equation (11.122), the log likelihood function is given by


K
−1
ln L = −NK ln(π) − K ln |Rnn | − (xk − sk (υ))† Rnn (xk − sk (υ)) (11.142)
k=1

Note the difference between the previous deterministic case and the stochastic case in
Equation (11.114). The deterministic case is easier to deal with because only the exponent
depends on the unknown parameter vector υ. Taking the partial derivatives defined in
Equation (11.133) and following the derivation in [39], the general M × M symmetric
FIM is given by


K  
=2 Re Hk† Rnn
−1
Hk (11.143)
k=1
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 452

452 CHAPTER 11 Direction of Arrival Estimation and Related Topics

where Hk is the N × M matrix

 
∂sk ∂s
Hk = ··· k (11.144)
∂υ 1 ∂υ M

The AOA is determined from one of terms of  −1 . Consider the case of a single emitter,
−1
with Rnn = σn2 I N . Let the kth snapshot xk = d(θ )bk where bk is a 1 × 1 complex vector,
and d(θ) is the N × 1 vector defined as
 
d(θ) = g1 (θ )e− jϕ1 (θ ) , g2 (θ )e− jϕ2 (θ ) . . . g N (θ )e− jϕ N (θ ) (11.145)

where ϕn (θ) = 2π τ (θ ). τn (θ ) is the time delay experienced by the desired signal relative
λ n
to fixed reference delay. The components of the FIM are determined by letting υ1 = θ,
υ2 =  b1 , υ3 = |b1 |, υ4 =  b2 , υ5 = |b2 | , . . . υ2K +1 = |b K |, and then taking the partial
−1
derivatives in (11.144). Following [39], 11 is found to be given by
⎡   ⎤
 † 2
1 # #
⎢#  #2  d d  ⎥
−1
σμ̂2 ≥ 11 = ⎣ d − 2 ⎦
(11.146)
2K (SNR1 ) d

1  |bk |2
N
where SNR1 = is the average array signal-to-noise ratio that would be received
K k=1 σn2
by antenna elements with unity gain.

11.11 SUMMARY AND CONCLUSIONS

Using the main beam of an array to locate signals generates the periodogram. The pe-
riodogram does not accurately locate targets or resolve closely spaced targets. Super-
resolution techniques, such as Capon’s method, MUSIC, root MUSIC, and MEM, use nulls
to accurately determine a target’s position as well as to resolve closely spaced targets. The
MEM approach to spectral estimation was introduced for obtaining high-resolution power
spectral density estimates of stationary time series from limited data records. Extension of
Burg’s scalar MEM algorithm to the multiple-channel problem is required for array pro-
cessing problems, and the extension provided by Strand was outlined for this purpose. The
Problems section introduces a modification of the Burg algorithm introduced by Marple
that alleviates the line-splitting problem occasionally occurring with this algorithm. When
both signal detection and parameter estimation must be accomplished, then the Bayes op-
timal array processor can be implemented sequentially to provide adaptive capabilities
in a natural manner. The notions of subspace fitting and eigenspace beamformers were
introduced using the concept of eigendecomposition as the point of departure.

11.12 PROBLEMS
1. Computer Simulation of Direction of Arrival Estimation Algorithms.
a. Use a periodogram to demonstrate the effect of separation angle between two sources using
an 8 element uniform array with λ/2 spacing when θ1 = −30◦ , 10◦ , 20◦ and θ2 = 30◦ .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 453

11.12 Problems 453

b. An 8 element uniform array with λ/2 spacing has 3 signals incident upon it: s1 (−60◦ ) = 1,
s2 (0◦ ) = 2, and s3 (10◦ ) = 4. Find the Capon spectrum.
c. An 8 element uniform array with λ/2 spacing has three signals incident upon it: s1 (−60◦ ) =
1, s2 (0◦ ) = 2, and s3 (10◦ ) = 4. Find the MEM spectrum.
d. An 8 element uniform array with λ/2 spacing has three signals incident upon it: s1 (−60◦ ) =
1, s2 (0◦ ) = 2, and s3 (10◦ ) = 4. Find the MUSIC spectrum.
e. An 8 element uniform array with λ/2 spacing has three signals incident upon it: s1 (−60◦ ) =
1, s2 (0◦ ) = 2, and s3 (10◦ ) = 4. Find the location of the signals using the root MUSIC
algorithm.
2. Prediction Error Filter Equations The prediction error filter matrix equation for a scalar
random process may be developed by assuming that two sampled values of a random process
x0 and x1 are known, and it is desired to obtain an estimate of the next sampled value x̂2 using
a second-order prediction error filter

x̂2 = a(2, 2)x0 + a(2, 1)x1

so that ε = x2 − x̂2 = x2 − a(2, 1)x1 − a(2, 2)x0

a. Using the fact that the optimal linear predictor must provide estimates for which the error
is orthogonal to the data (i.e., x0 ε = 0 and x1 ε = 0), show that
⎡ ⎤
  1  
r (1) r (0) r (2) ⎣ ⎦ 0
−a(2, 1) =
r (2) r (1) r (0) 0
−a(2, 2)


where r (n) = xi x j , |i − j| = n.

b. If P2 = ε 2 = (x2 − x̂2 )(x2 − x̂2 ), use the fact that the error in the estimate x̂2 is orthogonal
to the estimate itself (i.e., (x2 − x̂2 )x̂2 = 0) and the previously given expression for x̂2 to
show that

P2 = r (0) − a(2, 1)r (1) − a(2, 2)r (2)

The results of part (a) combined with the result from part (b) then yield the prediction error
filter matrix equation for this case.
3. The Relationship between MEM Spectral Estimates and ML Spectral Estimates [19]
Assume the correlation function of a random process x(t) is known at uniformly spaced,
distinct sample times. Then the ML spectrum (for an equally spaced line array of N sensors) is
given by

N x
MLM(k) =
v† (k)R−1
xx v(k)

where
k = wavenumber (reciprocal of wavelength)
x = spacing between adjacent sensors
v(k) = beam steering column vector where vn (k) = e− j2π nkx , n = 0, 1, . . . , N − 1
Rxx = N × N correlation matrix of x(t)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 454

454 CHAPTER 11 Direction of Arrival Estimation and Related Topics

a. First, define the lower triangular matrix L by

⎡ ⎤
1 0 ··· 0
⎢ c(2, N ) 1 ··· 0⎥
 ⎢ c(3, N ) c(2, N − 1) ··· 0⎥
⎥
L=⎢⎢ ⎥
⎣ .
.. ⎦
c(N , N ) c(N − 1, N − 1) ··· 1

where 1, c(2, M), . . . , c(M, M) are the weights of the M-long prediction error filter whose
output power is P(M). Note that
⎡ ⎤
P(N ) −−− −−− −−− −−−
⎢ 0 P(N − 1) −−− −−− −−− ⎥
⎢ 0 −−−
⎥
−−− ⎥
Rxx L = ⎢
⎢ .
0
⎥
⎣ .. .. .. ⎦
. .
0 0 P(1)

The maximum entropy spectrum estimate corresponding to the M-long prediction error
filter is given by

P(M)x
MEM(k, M) =  2
M 
 c(i, M) exp( j2π k(i − 1)x)
 
i=1

where c(1, M) ≡ 1. Define the matrix P according to

P ≡ L† Rxx L

Show that P is an N × N diagonal matrix whose diagonal elements are given by P(N ),
P(N − 1), . . . , P(1).
−1 †
b. Using the fact that R−1
xx = LP L , show that


N
x
† † † −1 †
v R−1
xx v = (L v) P (L v) =
MEM(k, n)
n=1

this result then gives the desired relationship

1 
N
1 1
=
MLM(k) N MEM(k, n)
n=1

Therefore, the reciprocal of the ML spectrum is equal to the average of the reciprocals of
the maximum entropy spectra obtained from the one-point up to the N-point prediction
error filter. The lower resolution of the ML method therefore results from the “parallel
resistor network averaging” of the lowest- to the highest-resolution maximum entropy
spectra.
4. Equivalence of MEM Spectral Analysis to Least-Squares Fitting of a Discrete-Time All-Pole
Model to the Available Data [15] Assume that the first (N + 1) points {r (0), r (1), . . . , r (N )}
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 455

11.12 Problems 455

of the autocorrelation function of a stationary Gaussian process are known exactly, and it is
desired to estimate r (N + 1). Consider the Toeplitz covariance matrix
⎡ ⎤
r (0) r (1) · · · r (N ) r (N + 1)
⎢ r (1) r (0) · · · r (N − 1) r (N ) ⎥
R N +1 =⎣ ⎦
··· ··· ···
r (N + 1) r (N ) r (1) r (0)

The basic autocorrelation function theorem states that R N +1 must be semipositive definite
if the quantities r (0), r (1), . . . , r (N + 1) are to correspond to an autocorrelation function.
Consequently det[R N +1 ] must be nonnegative.
MEM spectral analysis seeks to select that value of r (N + 1) that maximizes det[R N +1 ].
The entropy of the (N + 2) dimensional probability density function with covariance matrix
R N +1 is given by

entropy = 1n(2π e)[N +2/2] det[R N +1 ]1/2

and the choice for r (N +1) maximizes this quantity. To obtain r (N +2), the value of r (N +1) just
found is substituted into R N +2 to find det[R N +2 ], and the corresponding entropy is maximized
with respect to r (N + 2). Likewise, substituting the values of r (N + 1) and r (N + 2) found
already into det[R N +3 ] and maximizing yields r (N + 3). The estimates for additional values
r (N + 4), r (N + 5), · · · may then be evaluated by following the same procedure.
a. Show that maximizing det[R N +1 ] with respect to r (N + 1) is equivalent to the relation
⎡ ⎤
r (1) r (0) · · · r (N − 1)
⎢ r (2) r (1) · · · r (N − 2) ⎥
det ⎢
⎣ .. .. .. ⎥=0
⎦
. . .
r (N + 1) r (N ) ··· r (1)

b. Consider the all-pole data prediction error model given by

y(n) + a1 y(n − 1) + · · · + a N y(n − N ) = e(n)

or

yT a = e(n)

where
aT = [1, a1 , a2 , . . . , a N ]
yT = [y(n), y(n − 1) · · · y(n − N )]
N = order of all-pole model
n = number of data samples, n > N

and where e(n) is a zero-mean random variable with E{e(i)e( j)} = 0 for i = j. Assuming
that E{e(n)y(n − k)} = 0 for k > 0, show that multiplying both sides of the previous
equation for e(n) by y(n − k) and taking expectations yields

r  (k) + a1 r  (k − 1) + a2 r  (k − 2) + · · · + a N r  (k − N ) = 0, for k > 0


where r  (k) = E{y(n)y(n − k)}.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 456

456 CHAPTER 11 Direction of Arrival Estimation and Related Topics

c. Using the results of part (b) and the fact that r (τ ) = r (−τ ), it follows that

r  (1) + a1 r  (0) + · · · + a N r  (N − 1) = 0

r (2) + a1 r  (1) + · · · + a N r  (N − 2) = 0
..
.
r  (N + 1) + a1 r  (N ) + · · · + a N r  (1) = 0

or RN +1 a = 0 in matrix notation. Use this result to show that

⎡ ⎤
r  (1) r  (0) · · · r  (N − 1)

⎢ r (2) r  (1) · · · r  (N − 2) ⎥
det ⎢
⎣ .. .. .. ⎥=0
⎦
. . .
r  (N + 1) r  (N ) · · · r  (1)

If the first N + 1 exact values {r (0), r (1), . . . , r (N )} of any autocorrelation function are
available, then substituting these values into the first N of the simultaneous linear equations
corresponding to R N +1 a = 0 yields a unique solution for the coefficients {a1 , a2 , . . . , a N }.
Consequently, the value for r (N + 1) for a discrete-time all-pole model having the coeffi-
cients {a1 , a2 , . . . , a N } is uniquely determined by
⎡ ⎤
r (1) r (0) · · · r (N − 1)
⎢ r (2) r (1) · · · r (N − 2) ⎥
det ⎢
⎣ .. .. ⎥=0
⎦
. .
r (N + 1) r (N ) · · · r (1)

This result is identical to the relation obtained in part (a); hence, the same solution would
have been obtained from maximum entropy spectral analysis.
5. Angle of Arrival Estimation [37] The MEM technique has superior capability for resolving
closely spaced spectral peaks that may be exploited for estimating the angular distribution of
received signal power.
a. Using a time–space dualism, reformulate equation (11.19) to give a spatial spectrum φ̂xx (μ),
where μ = cos θ , and θ is the angle from array endfire for an N + 1 element linear array.
Assume narrowband signals with spacing d between elements.
b. Reformulate equation (11.20) for the spatial estimation problem of part (a). What correspon-
dence exists between the prediction error filter coefficients and the weights of a coherent
sidelobe canceller with N auxiliary antennas?
6. The Marple Algorithm [41] A new autoregressive (AR) spectral analysis algorithm has been
proposed that yields spectral estimates with no apparent line splitting (the occurrence of two or
more closely spaced peaks in the AR spectral estimate where only one peak should be present)
and reduced spectral peak frequency estimation biases. It exploits forward and backward linear
prediction and therefore is closely related to the Burg algorithm. With the Burg algorithm the
forward linear prediction error, f M,k , in the single-channel case is given by


M 
M
f M,k = xk+M + a M,i xk+M−i = a M,i xk+M−i for 1 ≤ k ≤ N − M
i=1 i=0

where M is the order of the all-pole AR model, xk is the kth sample output of the AR model,
and a M,m is the AR parameter m of the Mth order process. Note that an,0 is defined as unity.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 457

11.12 Problems 457

Likewise, the backward linear prediction error is given by


M
∗
b M,k = a M,i xk+i also for 1 ≤ k ≤ N − M
i=0

Since stationarity is assumed, the backward AR coefficients are the conjugates of the forward
AR coefficients. To obtain estimates of the AR parameters, Burg minimized the sum of the
backward and forward prediction error energies


N −M
  
N −M
 
eM =  f M,k 2 + b M,k 2
k=1 k=1

Substituting f M,k and b M,k into e M and setting the derivatives of e M with respect to the param-
eters a M,1 through a M,M to zero, one obtains


M
2 a M, j r M (i, j) = 0 for i = 1, . . . , M(a M,0 = 1 by definition)
j=0

 N −M ∗ ∗
where r M (i, j) = k=1 (x k+M− j x k+M−i + x k+ j x k+i ) for 0 ≤ i, j ≤ M. The minimum
prediction energy is then given by


M
eM = a M, j r M (0, j).
j=0

a. Show that the previous three expressions can be written in matrix form as

RM AM = EM
⎡ ⎤ ⎡ ⎤
1 eM
⎢ a M,1 ⎥ ⎢ 0 ⎥
where A M =⎢ ⎥ ⎢ ⎥
⎣ .. ⎦ , E M = ⎣ .. ⎦ ,
. .
a M,M 0
⎡ ⎤
r M (0, 0) · · · r M (0, M)
⎢ .. .. ⎥
and R M =⎣ . . ⎦
r M (M, 0) · · · r M (M, M)

b. The matrix expression found in part (a) has a structure that can be exploited to produce an
algorithm requiring a number of operations ∝ M 2 rather than M 3 . R M has both Hermitian
∗ ∗
symmetry [r M (i, j) = r M ( j, i)] and Hermitian persymmetry [r M (i, j) = r M (M −i, M − j);
it does not have Toeplitz symmetry [r M (i, j) = r M (i − j)], as the covariance matrix does.
However, R M is composed of Toeplitz matrices. Show that

R M = (T M ) H T M + (TνM ) H TνM
⎡ ⎤
x M+1 x M · · · x1
⎢ x M+2 x M+1 · · · x2 ⎥
where T M = ⎢
⎣ .. .. ⎥⎦
. .
x N x N −1 · · · x N −M
⎡ ⎤
x1∗ · · · x M+1
∗

⎢ .. .. ⎥
and TνM =⎣ . . ⎦ (conjugate and reversed matrix)
x N∗ −M · · · x N∗
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 458

458 CHAPTER 11 Direction of Arrival Estimation and Related Topics

c. To exploit this structure, introduce two new prediction error energy terms
−M−1 
N
    −M−1 
N
   
eM =  f M,k+1 2 + b M,k 2 and eM =  f M,k 2 + b M,k+1 2
k=1 k=1

which represent time-index-shifted variants of the definition for e M . As a result of these

new prediction error energy terms, show that
RM AM = EM and RM AM = EM where the elements of RM and RM are now given by
N
−M−1

 ∗ ∗

rM (i, j) = xk+M+1− j xk+M+1−i + xk+ j x k+i
k=1
N
−M−1

 ∗ ∗

and r M (i, j) = xk+M− j xk+M−i + xk+1+ j x k+1+i
k=1

As a result of these expressions, show that a persymmetry relationship exists:

 
rM (i, j) = [r M (M − i, M − j)]∗

d. Show that the following relationships exist among the correlation matrices R M , RM , and
RM :
∗
⎡ ⎤ ⎡ ⎤
x M+1 x N −M
⎢ .. ⎥   ⎢ . ⎥ ∗ 
RM = R M − ⎣ . ⎦ x M+1 , · · ·, x1 − ⎣ .. ⎦ x N −M , · · · , x N∗
x1∗ xN
⎡ ⎤ ⎡ ⎤
x N∗ x1
⎢ .. ⎥   ⎢ ⎥ 
RM = R M − ⎣ . ⎦ x N , · · · , x N −M − ⎣ ... ⎦ x1∗ , · · · , x M+1
∗

x N∗ −M x M+1
⎡ ⎤
RM |
r M+1 (0, M + 1)
⎢ .. ⎥
R M+1 = ⎣ −− −− −− −− −− | . ⎦
r M+1 (M + 1, 0) · · · r M+1 (M + 1, M + 1)
⎡ ⎤
r M+1 (0, 0) · · · r M+1 (0, M + 1)
⎢ .. ⎥
R M+1 =⎣ . | −− −− −− −− −− ⎦
r M+1 (M + 1, 0) | RM

Now define the (M + 1) element auxiliary column vectors

⎡ ⎤ ⎡ ⎤
c M,0 d M,0
⎢ .. ⎥ ⎢ .. ⎥
CM = ⎣ . ⎦ , DM = ⎣ . ⎦
c M,M d M,M
with analogous definitions for CM and DM . The auxiliary column vector elements are
defined by the matrix–vector products
∗
⎡ ⎤ ⎡ ⎤
x M+1 x N −M
⎢ .. ⎥ ⎢ .. ⎥
RM CM = ⎣ . ⎦ , RM DM = ⎣ . ⎦
x1∗ xN
with analogous products for RM CM and RM DM . Introduce the notation A M
I
to denote the
I
vector formed by reversing the element order, and introduce the notation A M to denote the
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 459

11.13 References 459

vector formed by reversing the element order and conjugating

⎡∗ ⎤
a M,M
⎢ .. ⎥
I
AM =⎢
⎣ ∗ ⎦
. ⎥
a M,1
1
I I I
Likewise, define E M , CM , and D M . From the expressions of forward and backward prediction
errors given for the Burg algorithm, show that these errors can be expressed in vector notation
as
   
f M,1 = x M+1 · · ·, x1 A M and b M,N −M = x M+1 , · · · , x N A∗M .

At this point, we are approaching the development given by Burg but now using time-shifted
AR parameters rather than the AR parameters employed before. The complete derivation
is rather lengthy and will not be pursued here. A complete block diagram of the Maple
algorithm is given in [41] along with a similar diagram for the Burg algorithm for ease in
comparison.
7. Computer Simulation Problem An eight-element uniform array with λ/2 spacing has three
signals incident upon it: s1 (−60◦ ) = 1, s1 (−0◦ ) = 2, and s3 (−10◦ ) = 4. Estimate the incident
angles using ESPRIT.

11.13 REFERENCES
[1] J. Capon, “High-Resolution Frequency-Wavenumber Spectrum Analysis,” Proceedings of the
IEEE, Vol. 57, No. 8, 1969, pp. 1408–1418.
[2] R. Schmidt, “Multiple Emitter Location and Signal Parameter Estimation,” IEEE Transactions
on Antennas and Propagation, Vol. 34, No. 3, 1986, pp. 276–280.
[3] R. Schmidt and R. Franks, “Multiple Source DF Signal Processing: An Experimental System,”
IEEE Transactions on Antennas and Propagation, Vol. 34, No. 3, 1986, pp. 281–290.
[4] A. Barabell, “Improving the Resolution Performance of Eigenstructure-Based Direction-
Finding Algorithms.” pp. 336–339.
[5] R. T. Lacoss, “Data Adaptive Spectral Analysis Methods,” Geophysics, Vol. 36, August 1971,
pp. 661–675.
[6] T. J. Ulrych and T. N. Bishop, “Maximum Entropy Spectral Analysis and Autoregressive
Decompositions,” Rev. Geophys. Space Phys., Vol. 13, 1975, pp. 183–200.
[7] J. P. Burg, “Maximum Entropy Spectral Analysis,” Paper presented at the 37th Annual Meeting
of the Society of Exploration Geophysicists, October 31, 1967, Oklahoma City, OK.
[8] J. P. Burg, “A New Analysis Technique for Time Series Data,” NATO Advanced Study Institute
on Signal Processing with Emphasis on Underwater Acoustics, Vol. 1, Paper No. 15, Enschede,
The Netherlands, 1968.
[9] T. J. Ulrych and R. W. Clayton, “Time Series Modelling and Maximum Entropy,” Phys. Earth
Planet. Inter., Vol. 12, 1976, pp. 188–200.
[10] S. B. Kesler and S. Haykin, “The Maximum Entropy Method Applied to the Spectral Analysis
of Radar Clutter,” IEEE Trans. Inf. Theory, Vol. IT-24, No. 2, March 1978, pp. 269–272.
[11] R. G. Taylor, T. S. Durrani, and C. Goutis, “Block Processing in Pulse Doppler Radar,” Radar-
77, Proceedings of the 1977 IEE International Radar Conference, October 25–28, London,
pp. 373–378.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 460

460 CHAPTER 11 Direction of Arrival Estimation and Related Topics

[12] G. Prado and P. Moroney, “Linear Predictive Spectral Analysis for Sonar Applications,” C. S.
Draper Report R-1109, C. S. Draper Laboratory, Inc., 555 Technology Square, Cambridge,
MA, September 1977.
[13] R. N. McDonough, “Maximum-Entropy Spatial Processing of Array Data,” Geophysics,
Vol. 39, December 1974, pp. 843–851.
[14] D. P. Skinner, S. M. Hedlicka, and A. D. Matthews, “Maximum Entropy Array Processing,”
J. Acoust. Soc. Am., Vol. 66, No. 2, August 1979, pp. 488–493.
[15] A. Van Den Bos, “Alternative Interpretation of Maximum Entropy Spectral Analysis,” IEEE
Trans. Inf. Theory, Vol. IT-17, No. 4, July 1971, pp. 493–494.
[16] J. P. Burg, “Maximum Entropy Spectral Analysis,” Ph.D. Dissertation, Stanford University,
Department of Geophysics, May 1975.
[17] J. H. Sawyers, “The Maximum Entropy Method Applied to Radar Adaptive Doppler Filtering,”
Proceedings of the 1979 RADC Spectrum Estimation Workshop, October 3–5, Griffiss AFB,
Rome, NY.
[18] S. Haykin and S. Kesler, “The Complex Form of the Maximum Entropy Method for Spectral
Estimation,” Proc. IEEE, Vol. 64, May 1976, pp. 822–823.
[19] J. P. Burg, “The Relationship Between Maximum Entropy Spectra and Maximum Likelihood
Spectra,” Geophysics, Vol. 37, No. 2, April 1972, pp. 375–376.
[20] R. H. Jones, “Multivariate Maximum Entropy Spectral Analysis,” Paper presented at the
Applied Time Series Analysis Symposium, May 1976, Tulsa, OK.
[21] R. H. Jones, “Multivariate Autoregression Estimation Using Residuals,” Paper presented at
Applied Time Series Analysis Symposium, May 1976, Tulsa, OK; also published in Applied
Time-Series Analysis, edited by D. Findley, Academic Press, New York, 1977.
[22] A. H. Nutall, “Fortran Program for Multivariate Linear Predictive Spectra Analysis Employing
Forward and Backward Averaging,” Naval Underwater System Center, NUSC Tech. Doc.
5419, New London, CT, May 9, 1976.
[23] A. H. Nutall, “Multivariate Linear Predictive Spectral Analysis Employing Weighted Forward
and Backward Averaging: A Generalization of Burg’s Algorithm,” Naval Underwater System
Center, NUSC Tech. Doc. 5501, New London, CT, October 13, 1976.
[24] A. H. Nutall, “Positive Definite Spectral Estimate and Stable Correlation Recursion for Mul-
tivariate Linear Predictive Spectral Analysis,” Naval Underwater System Center, NUSC Tech.
Doc. 5729, New London, CT, November 14, 1976.
[25] M. Morf, A. Vieira, D. T. Lee, and T. Kailath, “Recursive Multichannel Maximum
Entropy Spectral Estimation,” IEEE Trans. Geosci. Electron., Vol. GE-16, No. 2, April 1978,
pp. 85–94.
[26] R. A. Wiggins and E. A. Robinson, “Recursive Solution to the Multichannel Filtering
Problem,” J. Geophys. Res., Vol. 70, 1965, pp. 1885–1891.
[27] O. N. Strand, “Multichannel Complex Maximum Entropy Spectral Analysis,” 1977 IEEE
International Conference on Acoustics, Speech, and Signal Processing, pp. 736–741.
[28] O. N. Strand, “Multichannel Complex Maximum Entropy (Autoregressive) Spectral Analysis,”
IEEE Trans. Autom. Control, Vol. AC-22, No. 4, August 1977, pp. 634–640.
[29] L. W. Nolte and W. S. Hodgkiss, “Directivity or Adaptivity?” EASCON 1975 Record,
IEEE Electronics and Aerospace Systems Convention, September 1975, Washington, DC,
pp. 35.A–35.H.
[30] H. Cox, “Sensitivity Considerations in Adaptive Beamforming,” Proceedings of the NATO
Advanced Study Institute on Signal Processing with Emphasis on Underwater Acoustics,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 461

11.13 References 461

August 21–September 1, 1972, Loughborough, UK; also in Signal Processing, edited by

J. W. R. Griffiths, P. L. Stocklin, and C. van Schooneveld, Academic Press, New York, 1973.
[31] H. Cox, “Resolving Power and Sensitivity to Mismatch of Optimum Array Processors,”
J. Acoust. Soc. Am., Vol. 54, No. 3, September 1973, pp. 771–785.
[32] D. Jaarsma, “The Theory of Signal Detectability: Bayesian Philosophy, Classical Statistics,
and the Composite Hypothesis,” Cooley Electronics Laboratory, University of Michigan, Ann
Arbor, Technical Report TR-200, 1969.
[33] M. A. Gallop and L. W. Nolte, “Bayesian Detection of Targets of Unknown Location,” IEEE
Trans. Aerosp. Electron. Syst., Vol. AES-10, No. 4, July 1974, pp. 429–435.
[34] W. S. Hodgkiss and L. W. Nolte, “Bayes Optimum and Maximum-Likelihood Approaches
in an Array Processing Problem,” IEEE Trans. Aerosp. Electron. Syst., Vol. AES-11, No. 5,
September 1975, pp. 913–916.
[35] R. H. Roy, “ESPRIT Estimation of Signal Parameters via Rotational Invariance Techniques,”
Ph.D. Dissertation, Stanford University, Stanford, CA, August 1987.
[36] R. Roy, A. Paulraj, and T. Kailath, “ESPRIT—A Subspace Rotation Approach to Estimation
of Parameters of Cissoids in Noise,” IEEE Trans. Acoust., Speech, & Signal Processing, Vol.
ASSP-34, No. 4, October 1986, pp. 1340–1342.
[37] R. Roy and T. Kailath, “ESPRIT—Estimation of Signal Parameters via Rotation Invariance
Techniques,” IEEE Trans. Acoust., Speech, & Signal Processing, Vol. ASSP-37, No. 7, July
1989, pp. 984–995.
[38] A. G. Jaffer, “Maximum Likelihood Direction Finding of Stochastic Sources: A Separable
Solution,” 1988 IEEE International Conf. Acoustics, Speech, and Signal Processing (ICASSP),
New York, April 1988, pp. 2893–2896.
[39] W. P. Ballance and A. G. Jaffer, “The Explicit, Analytic Cramer-Rao Bound on Angle Estima-
tion,” Proc. Twenty-Second Asilomar Conf. Signals, Systems and Computers, Pacific Grove,
CA, November 1988, pp. 345–351.
[40] Harry L. Van Trees, Detection, Estimation & Modulation Theory, New York, Wiley, 1986,
p. 66.
[41] L. Marple, “A New Autoregressive Spectrum Analysis Algorithm,” IEEE Trans. Acoustics,
Speech & Signal Processing, Vol. ASSP-28, No. 4, August 1980, pp. 441–454.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:52 462
Monzingo-7200014 book ISBN : XXXXXXXXXX November 25, 2010 21:6 463

CHAPTER

Recent Developments in
Adaptive Arrays 12
' $
Chapter Outline
12.1 Beam Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
12.2 Space-Time Adaptive Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
12.3 MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
12.4 Reconfigurable Antennas and Arrays. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
12.5 Performance Characteristics of Large Sonar Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
12.6 Adaptive Processing for Monopulse Tracking Antennas . . . . . . . . . . . . . . . . . . . . . . . . 486
12.7 Partially Adaptive Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
12.8 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
12.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
12.10 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
& %

This chapter presents several innovations that have taken place since the first edition of this
book. Wireless communication applications often resort to very simple beam switching,
in which multiple beams simultaneously exist and the one with the best signal reception
is selected. Moving radars or sonars must deal with clutter as well as interfering signals.
Space-time adaptive processing (STAP) combines a spatial adaptive array with a temporal
adaptive array to improve clutter cancellation and null placement. Another relatively recent
development is multiple input, multiple output (MIMO) antenna array systems where
an adaptive array is used for both transmit and receive to increase channel capacity.
Reconfigurable antennas change their physical layout using switches to adapt for example
the pattern, frequency response, and polarization response to match the desired signal.
Partial adaptivity is of interest when only a portion of the total number of elements is
controlled, thereby reducing the number of processors required to achieve an acceptable
level of adaptive array performance.

12.1 BEAM SWITCHING

Multiple overlapping beam patterns pointing in slightly different directions can be formed
in an array using hardware or software. For instance, Figure 12-1 shows a 10-element
uniform array with elements spaced λ/2 apart with five orthogonal beams. Beams are
orthogonal when the angle at which one beam has a peak and the rest have nulls (indicated
463
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 464

464 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-1 1
Five orthogonal
beams from a
10-element uniform
0.8
array with half-
wavelength spacing.

0.6

AF
0.4

0.2

0
−50 0 50
q (degrees)

FIGURE 12-2 Terminations Array elements

Rotman lens with
multiple beams
(Courtesy of
Beam ports

Remcom, Inc.). The

lens is designed
for microstrip
implementation
using the Rotman
Lens Designer [2].
Terminations

by dashed arrows in Figure 12-1). An algorithm continuously evaluates each beam and
selects the one that maintains the highest signal quality. The system scans each beam output
and selects the beam with the largest output power as well as suppresses interference.
Hardware configurations, such as the Rotman lens [1,2] (Figure 12-2) and Butler
matrix [3] (Figure 12-3), have physical beam ports that correspond to a beam pointing
in a direction determined by the passive feed network, element spacing, and number of
elements. Each beam port in the Rotman lens receives a signal from all the elements. The
different path lengths from the elements to the beam ports account for the phase shift
that steers the beams. The Rotman lens in Figure 12-2 has a 16-element array with 11
beam ports. Each beam port has a pattern of a 16-element array steered to predetermined
directions based on the geometry. The Butler matrix is a hardware version of a fast Fourier
transform (FFT) [4]. Each port receives the signal from one of the beams. A switch
selects the desired beam formed by these hardware beamformers. If an array has a digital
beamformer, then the beams are formed and selected in software. Usually, the beams cover
a desired azimuth range.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 465

12.2 Space-Time Adaptive Processing 465

FIGURE 12-3
Butler matrix with
four beams.

1 2 3 4 Elements

Quadrature
hybrid couplers

45° phase shifts

Quadrature
hybrid couplers

1R 2L 2R 1L

12.2 SPACE-TIME ADAPTIVE PROCESSING

Not only environmental noise, jamming, and unintentional interfering signals but also
clutter entering the sidelobes and main beam prevent detection of a radar signal. Airborne
phased array radar data normally has moving target signals embedded in undesired signals,
such as clutter, broadband noise jamming, and thermal noise. A moving target indicator
(MTI) radar detects a moving target by isolating the Doppler frequency given by [5]

2va cos φ
fD = (12.1)
λ
where va is the velocity of the aircraft, and φ is the angle measured from the velocity
vector. Figure 12-4 is a plot of the received signal power as a function of azimuth angle
and Doppler frequency. The peak of the Doppler clutter occurs normal to the direction of
velocity and at zero Doppler frequency. The interference occurs at a single angle but over
FIGURE 12-4 A
plot of the target,
ce

clutter, and
feren

interference returns
Clutter as a function of
Inter

angle and frequency.

fD

Target

Azimuth angle
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 466

466 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-5
Linear array types. a:
Spatial. b: Temporal. w1
d
T
w2
Σ T
w1 w2 wN

Σ wM T

(a) (b)

all frequencies. The motion of a radar platform spreads the clutter in Doppler frequency.
The Doppler frequency from clutter at a specific point on the ground depends on the angle
of the clutter position relative to the heading of the platform. Interference from a discrete
source appears at one angle but is spread over all Doppler frequencies.
A spatial adaptive array weights and combines the signals received by the array
elements at the same instant in time but at different spatial locations separated by distance
d (Figure 12-5a). A temporal adaptive array combines signals received at the same spatial
location but sampled at different instances in time separated by time T (Figure 12-5b).
A displaced phase center antenna (DPCA) cancels clutter induced by platform motion
in an MTI radar [6]. The idea is to make the antenna appear stationary over the transmitted
pulse train by electronically shifting the phase center of the receive aperture backward to
compensate for the forward motion of the moving platform. The DPCA was first envisioned
for a rotating monopulse radar [7]. By adding and subtracting the output from the azimuth
difference channel to the output of the azimuth sum channel, a fore and aft beam are
formed. If the output from the aft beam is subtracted from the output from the fore beam
at the same pointing angle, then the clutter return would be canceled.
A better implementation is based on synthetic aperture radar (SAR). When the velocity
vector of the platform is parallel to the linear array axis, then the pulse repetition frequency
(PRF) is adjusted to the platform velocity so that the first, second, and subsequent elements
at the current pulse appear to move to the respective positions of the second, third, and
subsequent elements at the previous pulse [8]. Figure 12-6 shows a four-element array
split into two three-element arrays: fore and aft. The full aperture transmits a pulse train
with Nt pulses at t = 0. Both the fore and aft arrays receive one pulse, then the array
moves in space a distance d, and the fore and aft arrays receive another pulse. Once the
array moves forward by (M − 1)d, then the full aperture transmits another pulse train of
Nt pulses. Figure 12-7 shows the phase centers of the full, fore, and aft apertures. They

FIGURE 12-6 The d

four-element array is
split into fore and aft Space
arrays. The array
moves a distance Fore
T
equal to the element
spacing in time T . Aft

Time
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 467

12.2 Space-Time Adaptive Processing 467

Aft aperture phase center FIGURE 12-7 The

Full aperture phase center phase centers
d Fore aperture phase center associated with the
fore, aft, and full
arrays.

Aft Fore
Beamformer Beamformer

T
+ −
Σ

Doppler
Filter Bank

are equally spaced by d/2. Delaying the fore aperture output by T effectively moves the
fore aperture phase center to correspond to the aft aperture phase center. Thus, the array
moves a distance d/2 in one pulse repetition interval (PRI), so [8]
d
va = (12.2)
2PRI
The voltage output from the aft and fore arrays are given by
π π
AF A (φ) = e− j f d cos φ
+ e− j c f d cos φ f d cos φ
3π
Aft beam: c + ej c
π π
AF F (φ) = e− j f d cos φ f d cos φ f d cos φ
3π
Fore beam: c + ej c + ej c (12.3)

Assuming that the clutter does not change from pulse to pulse, when the fore beam is
delayed by T, then it is identical to the aft beam as shown by
  d 
− j2π 2va λcos φ
AF F (φ)e− j2π fd T = AF F (φ)e 2va

= AF F (φ)e− j f d cos φ
2π
c (12.4)
−j 3π
f d cos φ − j πc f d cos φ j πc f d cos φ
=e c +e +e
= AF A (φ)

Thus, subtracting the fore aperture output delayed by T from the aft aperture output should
cancel the clutter, which does not change much from pulse to pulse.
DCPA works perfectly when there are no errors. Some practical problems with DCPA
are as follows [9]:
1. The platform velocity does not perfectly match the PRI of the radar.
2. Clutter can change from position to position.
3. Error tolerances in the antenna components cause the signal path at each element to be
different.
4. Antenna calibration is necessary to compensate for thermal noise and aging of
components.
5. Unwanted platform motion causes deviations from the desired velocity vector.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 468

468 CHAPTER 12 Recent Developments in Adaptive Arrays

An adaptive DCPA algorithm (ADPCA) [10] is not an optimal processor, so it can have
significant signal-to-interference plus noise ratio (SINR) loss. DCPA was implemented
on an advanced development model called Pave Mover and then was followed by the
development of the Joint Surveillance and Target Attack Radar System (Joint STARS)
[11]. Joint STARS uses a 24 ft. long, 2 ft. high phased array antenna (Figure 12-8) mounted
on the forward underfuselage of an Air Force E-8A (Figure 12-9).
STAP enables radars and sonars to detect targets obscured by clutter and jamming. It
was first envisioned by Brennan and Reed [12,13] as an adaptive array for MTI radar. STAP
is an improvement over DCPA, because it integrates the spatial adaptive array with the
temporal adaptive array. It improves clutter cancellation performance and integrates spatial
processing (sidelobe control and null placement) with clutter cancellation. STAP also has

FIGURE 12-8
Photograph of the
phased array used
for Pave Mover and
Joint STARS
(Courtesy of the
National Electronics
Museum).

FIGURE 12-9
Picture of array
mounted beneath
the USAF E-8A
(Courtesy of the
USAF Strategic Air
Command).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 469

12.2 Space-Time Adaptive Processing 469

d FIGURE 12-10
Diagram of a STAP
antenna.

T T T

T T T

T T T

w11 w12 w1M w21 w22 w2M wN1 wN2 wNM

application to sonar reverberation suppression and simultaneous position and Doppler

estimation using high-resolution space-time matched field-processing techniques [8].
STAP is a two-dimensional (angle–frequency) adaptive filter that maximizes the out-
put SINR by weighting the space and time samples of the array as shown in Figure 12-10.
The element locations are the filter’s spatial dimension, and the PRI is the temporal di-
mension. STAP forms a beam in the direction of the target at the same time it creates
a frequency filter about the target spectrum and puts nulls in the directions and at the
frequencies of interferers. It was originally developed to detect slow-moving targets from
airborne radars. The radar transmits Nt coherent pulses that are received by Nr elements of
an array and sampled at L range gates over several pulse repetition intervals. A snapshot is
the M × N sample matrix collected at each range gate. The snapshots from each range gate
are stacked into a data cube. STAP adaptively weights input signal samples to eliminate
the noise, clutter, and interference received from the environment. These weights, at one
instant of time (at the same time delay), form an antenna pattern in space with nulls placed
in the directions of interfering sources. Applied to the signal samples at one antenna array
element over the entire dwell, the STAP weights define a system impulse response. The
clutter spectrum for ground-based radars has a peak at zero Doppler, whereas the clutter
spectrum for moving platform radars depends on frequency. STAP adapts the frequency
response to the clutter spectrum to receive the desired signal while rejecting clutter.
Figure 12-11 is an example of simulated radar returns as a function of angle and
Doppler when both clutter and a jammer are present. The jammer appears stationary at
one angle but exists over a wide frequency range. Clutter, on the other hand, has a maximum
at zero Doppler and extends over all angles as a linear function of frequency. Applying a
STAP filter lowers the clutter and jammer returns while increasing the noise and target as
shown by the picture on the right in Figure 12-11.
STAP has the following advantages:
• Increased detection of low-velocity targets by better suppression of main beam clutter
• Increased detection of small cross section targets that might otherwise be obscured by
sidelobe clutter
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 470

470 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-11

Jammer

Jammer
STAP reduces
clutter and jammer
power as seen by
the before (left) and
after (right) Doppler

Doppler
r r
angle–Doppler plots. te te
ut ut
Cl Cl

Target

STAP

Angle Angle

• Increased detection in combined clutter and jamming environments

• Robust to radar system errors
• Means to handle nonstationary interference
STAP’s big disadvantage is the high computational cost [14]. There are two major parts
of the computation: (1) formation of the covariance matrix; and (2) the inversion of that
matrix. Data from a large number of adjacent range cells are required to compute the
covariance matrix of the undesired signals. Assuming that the structure of the covariance
matrix is known helps speed the calculation of the covariance matrix. Inverting the covari-
ance matrix is the dominant computation in STAP, because the inversion is a nonlinear
transform and is not conducive to parallel processing. Numerous algorithms that reduce
the data sample size required or the computational cost have been proposed. Most of these
approaches are only partially adaptive.
The STAP processor finds the weights in real time by solving an NM × NM system of
linear equations. The new weights require on the order of (NM)3 operations [14]. Airborne
radars constantly scan space searching for targets, so computational throughputs on the
order of hundreds of billions of floating-point operations per second with execution speeds
of fractions of a second are needed. As a result, current research work on STAP focuses on
developing algorithms that decompose the fully adaptive problem into reduced-dimension,
real-time adaptive problems capable of being implemented in real time on reasonably
sized processors. Space and power on an aircraft are premium commodities, so processing
performance per unit size, power, and weight are important [14].
The following analysis follows that in [8]. An estimate of the covariance matrix from
a set of radar data is called the training set. The covariance matrix is estimated from
R̂ = X† X (12.5)
where X, the training set matrix, is a subset of the input data. The weights are computed
using sample matrix inversion (SMI) and finds the weights using direct matrix inver-
sion or a factorization approach that computes the Cholesky decomposition of R via QR
decomposition of X (see Chapter 6).
If X can be factored so that X = UA, where A is upper triangular, then the covariance
matrix is written as
R̂ = A† U† UA = A† A (12.6)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 471

12.2 Space-Time Adaptive Processing 471

Since U is an orthogonal, unitary matrix, U† U = I, where I is the identity matrix. Since

A is triangular, The weights are easily computed from by back-solving
A† u = s
Aw = u (12.7)
where s is the target response at one angle and Doppler frequency, and u is an intermediate
computation vector. Each member of the STAP filter bank has a different s. Beamforming
is the product of the weight vector and the input data for a specific range gate.
z = w† y (12.8)
The covariance matrix of a fully adaptive STAP algorithm, in which a separate adaptive
weight is applied to Nt pulse-repetition intervals as well as Nr elements, has dimensions
NM × NM where 200 ≤ NM ≤ 105 .
The noise interference vector, q, is a zero mean random vector with a multivariate
complex Gaussian distribution represented by the sum of the uncorrelated noise and the
correlated interference (e.g., jammer, clutter).
q=c+n (12.9)
The noise interference vector covariance matrix for a STAP array is given by [8]
⎡ ⎤
Q0 Q1 · · · Q M−2 Q M−1
⎢ Q1 Q0 Q M−2 ⎥
⎢ ⎥
⎢ .. . .. .. ⎥
Rqq = ⎢ . . ⎥ (12.10)
⎢ ⎥
⎣ Q∗ Q0 Q1 ⎦
M−2
Q∗M−1 Q∗M−2 · · · Q∗1 Q0
where Qm are N × N spatial covariance submatrixes measured between pulses separated
in time by mT [8].
⎡ ⎤
q11 (mT ) · · · q1N (mT )
⎢ .. .. .. ⎥
Qm = E{q(nT )q∗ ((n + m)T )} = E{q(0)q∗ (mT )} = ⎣ . . . ⎦
q N 1 (mT ) · · · qNN (mT )
(12.11)
As with spatial adaptive nulling, the optimum weights are found from
−1
wopt = γ Rqq s (12.12)

The initial nonadaptive filtering can be either a transformation into the frequency
domain (e.g., by performing an FFT, over pulses in each channel) or a transformation into
beam space (e.g., by performing nonadaptive beamforming in each pulse). We can perform
both spatial and temporal transformations, if desired, or we can eliminate nonadaptive
filtering altogether.
The nonadaptive filtering determines the domain (frequency or time, element or beam)
in which adaptive weight computation occurs. Figure 12-12 is a diagram representing the
data domain for a single range gate after a different type of nonadaptive transform [14,15].
For example, the upper right quadrant represents PRI data that have been transformed
into Doppler space. Thus, each sample is a radar return for a specific Doppler frequency
and receiver element. The lower left quadrant represents element data that have been
transformed into beam space. Each sample is a radar return for a specific PIR and look
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 472

472 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-12 Time Domain Frequency Domain

STAP algorithm PRI Temporal FFT Doppler Bin
transformations

Element space
involve either
Element space Element space
one-dimensional
pre-Doppler pre-Doppler
space and time
transforms or
two-dimensional
space–time Spatial 2D Space- Spatial
transform. FFT Time FFT FFT

Beam space

Beam space Beam space

pre-Doppler pre-Doppler

Temporal FFT

direction. For example, a STAP kernel that is adaptive in the frequency domain falls
into the element-space post-Doppler quadrant of the taxonomy. A Doppler filter-bank
FFT transforms the signals from each element. Low-sidelobe Doppler filtering localizes
competing clutter in angle to reduce the extent of clutter to be adaptively canceled. The
adaptation occurs over all elements and a number of Doppler bins. The number of Doppler
bins is a parameter of the element-space post-Doppler algorithm. Factored post-Doppler
algorithms perform spatial adaptation in a single bin and are not adaptive in time.
A partially adaptive STAP algorithm breaks down the fully adaptive STAP prob-
lem into a number of independent, smaller, and more computationally tractable adaptive
problems while achieving near-optimum performance [8,14]. Figure 12-13 shows that a

FIGURE 12-13
STAP enhances the
target while N Original
Element

suppressing the data cube

clutter and L
interference. The te
ga
picture to the left is 1 PRI M n ge
before convergence, Ra
and the picture to
the right is after
convergence.
Transformed
data cube
Beam

te
ga
n ge
Doppler Ra

Subarrayed
data cube
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 473

12.3 MIMO 473

partially adaptive algorithm starts by nonadaptive filtering of the input signal data to reduce
dimensionality. Once the input data are transformed and bins and beams (or channels and
pulse repetition intervals) are selected to span the target and interference subspaces, mul-
tiple separate adaptive sample matrix inversion problems are solved, one for each Doppler
frequency bin or pulse repetition interval, across either antenna elements or beams, de-
pending on the domain of the adaptation.
Some current STAP research focuses on the following [16]:
1. Bistatic configurations where the transmit and receive platforms move separately to
keep the receiving platform covert
2. Conformal arrays
3. Nonstationary received signals that arise from bistatic STAP, conformal arrays, terrain
with different reflection coefficients, and clutter motion like vegetation blowing in the
wind.
4. Knowledge-aided STAP, which attempt to remove as much of the heterogeneity from
the snap shots prior to using conventional estimation methods. This is done by using a
priori knowledge, typically stored in databases.
5. STAP applications in sonar, telecommunications, and detection of plastic landmines.

12.3 MIMO
A communications system falls into one of four categories shown in Figure 12-14 based
on the number of antennas used for transmit and receive:
1. A single input, single output (SISO) system has one antenna for transmit and one
antenna for receive. SISO is the most common communications system category and
offers no spatial diversity.
2. A single input, multiple output (SIMO) system has one antenna for transmit and multi-
ple antennas for receive. This configuration is common when an antenna array is used
on receive.
3. A multiple input, single output (MISO) system has multiple antennas for transmit and
one antenna for receive.
4. A MIMO system has multiple antennas for transmit and multiple antennas for receive.
This configuration offers the greatest spatial diversity and control but also has the
greatest hardware cost.
Adaptive antennas have traditionally been SIMO. Some work has been done in MISO
systems, but today considerable attention has fallen on MIMO systems because they have

SISO MISO SIMO MIMO FIGURE 12-14

Transmit Antenna Communications
system categories
based on number of
antennas used for
transmit and receive.

Receive Antenna
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 474

474 CHAPTER 12 Recent Developments in Adaptive Arrays

the greatest promise to deliver the highest capacity. MIMO has also found use in acoustic
arrays [17].
In a SISO system with two isotropic point sources separated by R and no multipath,
if the transmitted signal is s, then the received signal, r, is given by

se− j λ
2π
R
r= (12.13)
R
If the signal also takes one bounce from an object with a reflection coefficient, , then the
received signal includes a second path of distance R1 .

se− j λ se− j λ
2π 2π
R R1
r= + (12.14)
R R1
As the number of scattering objects increase, the number of paths increases until the
multipath formulation is given by

se− j λ m se− j
2π M 2π
R λ Rm
r= + (12.15)
R m=1
Rm

Multiple paths turn the channel transfer function, h, into a random Gaussian variable, so
(12.15) becomes
r = hs (12.16)
When the transmit and receive antennas are arrays, then each of the Nt transmit
elements send a signal to each of the Nr receive elements. Figure 12-15 shows a diagram
of a MIMO system [18]. To take interactions between the elements in the transmit and
receive arrays into account, (12.16) is written into matrix form
r = Hs + N (12.17)
where N is a noise vector, and H is the channel matrix given by (Figure 12-16)
⎡ ⎤
h 11 h 12 · · · h 1N
⎢ .. ⎥
⎢ h 21 h 22 . ⎥
H=⎢ ⎢ ..
⎥
⎥ (12.18)
⎣ . ⎦
h M1 · · · hMN
The SVD of H is given by

H = USVD DV†SVD (12.19)

⎡√ ⎤
λ1 0 0
⎢ .. ⎥
D=⎣ 0 . 0 ⎦
√
0 0 λM
λm = singular values
USVD , VSVD = singular vectors

The columns of USVD are the receive weight vectors corresponding to the singular values,
and columns of VSVD are the transmit weight vectors corresponding to the singular values
[19]. Since the channel is reciprocal, exactly the same weights may be used for transmission
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 475

12.3 MIMO 475

Symbols FIGURE 12-15

Diagram of a MIMO
system.
Space/time encoder

Pulse shaping

RF up-conversion
filtering and matching

I Nt Transmit array
Channel
I Nr Receive array

RF down-conversion
filtering and matching

Matched filtering
and sampling

Space/time decoder

Symbols

h11 FIGURE 12-16

1 1 Signal paths
h12
h13 between transmit
h21 and receive
antennas in a MIMO
2 2
h22 system.
h23

hM1
hM2

Nt Nr
hMN

as for reception. Beamforming using the singular vectors as array weights produces eigen-
patterns that create independent (spatially orthogonal) parallel communication channels
in the multipath environment.
Alternatively, (12.17) can be written in terms of power such that the power received
by the array is given by

P = r† r = s† HH† s (12.20)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 476

476 CHAPTER 12 Recent Developments in Adaptive Arrays

An eigenvalue decomposition of the Nr × Nr correlation matrix HH† is written as

⎡ ⎤
λ1 0 0
⎢ .. ⎥
HH† = R H = Vλ ⎣ 0 . 0 ⎦ V† λ (12.21)
0 0 λM

where Vλ are the eigenvectors, and λm are the eigenvalues. The off-diagonal elements
of R H represent the correlation between the transmitted signal streams, with increased
correlation resulting in decreased capacity. The eigenvalue represents the received signal
power level in the the eigenchannel. Once an accurate estimate of H is established, then
the transmitted data, s, is calculated from the received data by inverting the channel matrix.

s = H−1 r (12.22)

In free space with no multipath, the elements of H are the free-space Greens function
⎡ − jk R11 ⎤
e e− jk Rmn e− jk Rmn
⎢ R ···
⎢ 11 Rmn Rmn ⎥ ⎥
⎢ − jk R21 ⎥
⎢e e− jk Rmn ⎥
⎢ ⎥
⎢ R21 Rmn ⎥
H=⎢ ⎢ ⎥ (12.23)
. .. . ⎥
⎢ .. . .
. ⎥
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎣e − jk R mn
e − jk R mn ⎦
···
RM N Rmn
The number of data streams supported must be less than or equal to the rank of H. The
rank of a matrix is the number of nonzero singular values. H is ill conditioned as presented
in (12.23). Increasing the separation between array elements or adding random components
to the matrix elements decreases the matrix condition number and improves the accuracy
of inverting H. Multipath adds random components, so it significantly improves the ability
of inverting H. Low-rank MIMO channels are associated with minimal multipath or large
separation distance between the transmit and receive antennas [20]. The low-rank MIMO
channel is equivalent to a SISO channel with the same total power. High-rank MIMO
channels are associated with a high multipath environment, and the separation distance
between transmit and receive antennas is small. As we pack more antennas into our array,
the capacity per antenna drops due to higher correlation between adjacent elements. MIMO
systems perform best with a full rank channel matrix, which means a low correlation
between signals on the different antennas.
The instantaneous eigenvalues have limits defined by [19]
2 2
Nt − Nr < λn < Nt + Nr (12.24)

When Nt is much larger than Nr , then all the eigenvalues cluster around Nt . Each eigen-
value is nonfading due to the high-order diversity. Thus, the uncorrelated asymmetric
channel with many antennas has a very large theoretical capacity of Nr equal, constant
channels with gains of Nt .
Each element of a MIMO transmit array uses the same frequency spectrum but may
transmit symbols from different modulation schemes and carry independent data streams.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 477

12.3 MIMO 477

The propagation channel is assumed to be Rayleigh and unknown to the transmitter. Each
burst contains a training sequence that allows the receiver to get an accurate estimate of
the propagation conditions. The channel may change from one burst to the next due to
motion of the transmitter or receiver.
When the channel is unknown at the transmitter, the power is uniformly distributed
over the antennas

Pn = P/Nt (12.25)

When the channel is known, then water filling determines the power allocation to the
channels [21]. In water filling, more power is allocated to better subchannels with higher
SNR to maximize the sum of data rates in all subchannels. Since the capacity is a logarith-
mic function of power, the data rate is usually insensitive to the exact power allocation,
except when the SNR is low. Assuming all noise powers to be the same, water filling is the
solution to the maximum capacity, where each channel is filled up to a common level 
1
+ Pn =  where Pn = P (12.26)
λn
The capacity difference between the known and unknown channels is small for large P.
Thus, the highest-gain channel receives the largest share of the power. In the limit where
the SNR is small ( p < 1/λ2 − 1/λ1 ), only one eigenvalue, the largest, is left.
The capacity of a MIMO system depends on the propagation environment, array
geometry, and antenna patterns. If the sources are uncorrelated and equal power and the
channel is random, then the ergodic (mean) capacity is given by [18]
   
P †
C = E log2 det I N + HH bits/s/Hz (12.27)
Nt σn2
MIMO capacity increases linearly with the number of elements when the number of
transmit and receive antennas are the same. Winters suggests that Nt should be of the
order 2Nr [22]. When Nt and Nr are large and Nt > Nr , the capacity is [19]

C = Nr log2 (1 + Nt P/Nr ) (12.28)

As long as the ratio of Nt /Nr is constant, then the capacity is a linear function of Nr .
The vertically layered blocking structure (V-BLAST) algorithm was developed at Bell
Laboratories for spatial multiplexing [20]. It splits a single high data rate data stream into
Nt lower rate substreams in which the bits are mapped to symbols. These symbols are
transmitted from Nt antennas. Assuming the channel response is constant over the system
bandwidth (flat fading), the total channel bandwidth is a fraction of the original data
stream bandwidth. The receive array has an adaptive algorithm where each substream is a
desired signal while the rest are nulled. As a result, the receive array forms Nt beams while
simultaneously creating nulls. The received substreams are then multiplexed to recover
the original transmitted signal. V-BLAST transmits spatial uncoded data streams without
equalizing the signal at the receiver. VB cannot separate the streams and suffers from
multistream interferences (MSI). Thus, the transmission is unsteady, and forward error
coding is not always able to resolve this issue.
Space-time codes deliver orthogonal and independent data streams. Orthogonal fre-
quency division multiplexing (OFDM) is commonly used in MIMO systems [21]. It splits
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 478

478 CHAPTER 12 Recent Developments in Adaptive Arrays

the high-speed serial data to be transmitted into many much lower-speed serial data signals
that are sent over multiple channels. The bit or symbol periods are longer, so multipath
time delays are not as deleterious. Increasing the subcarriers and bandwidth makes the
signal immune from multipath. When the subcarrier spacing equals the reciprocal of the
symbol period of the data signals, they are orthogonal. The resulting sinc frequency spectra
have their first nulls at the subcarrier frequencies on the adjacent channels.
The channel matrix is found through measurement or computer model [18]. A switched
array design employs a single transmitter and single receiver to measure H by using high-
speed switches to sequentially connect all array combinations of array elements. Switching
times range from 2 to 100 ms, so the measurement of all antenna pairs is possible before
the channel appreciably changes for most environments. Virtual array instruments displace
or rotate a single antenna element. This method eliminates mutual coupling effects, but a
complete channel matrix measurement takes several seconds or minutes, so the channel
must remain stationary over that time. As a result, virtual arrays work best for fixed indoor
measurements when activity is low.
Channel models compute H based on a statistics. It is common practice to assume
that the transfer function between one transmit and one receive antenna has Rayleigh
magnitude and uniform phase distributions for a non-line-of-sight (NLOS) propagation
environment. Ray tracing programs can be used to calculate H. Figure 12-17 shows an

(a) (b)

R R

T
T

(c) (d)

R R

T T

FIGURE 12-17 Ray tracing in a MIMO system (Courtesy of Remcom, Inc.).

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 479

12.4 Reconfigurable Antennas and Arrays 479

example of ray tracing from a transmit antenna to a receive antenna in a city when the
transmit antenna and receive antenna are not line of sight [22]. Figure 12-17a is an example
when only two bounces occur between the transmit and receive antenna. The signal paths
in Figure 12-17b take two very different routes. Figure 12-17c and Figure 12-17d are
examples of paths that take many bounces. Signal paths are highly dependent on the
environment and the positions of the transmitters and receivers. Small changes in position
as shown in Figure 12-17 create large changes in H.

12.4 RECONFIGURABLE ANTENNAS

AND ARRAYS
Antennas reconfigure or alter the current flow paths using switches to change, for example,
frequency and polarization. For example, two ultra wide band (UWB) monopoles with
reconfigurable band notch in the wireless LAN frequency range (5.150–5.825 GHz) are
shown in Figure 12-18 [23]. The antenna is an elliptical patch fed with coplanar waveguide
line. The U-shaped slot is approximately λ/2 long and has a frequency notch when the
micro electro mechanical systems (MEMS) switch is open but not when the MEMS switch
is closed. When the switch is open at 5.8 GHz, the currents in the inner and outer side
of the slot flow in opposite directions and cancel each other. When the MEMS switch is
closed, the slot is shorted at its center point, so the total length of the slot is cut in half.
Consequently, the slot no longer supports the resonating currents and radiation occurs as
if the slot was not present. The second antenna has two inverted L-shaped open stubs that
connect and disconnect the stubs from the elliptical patch via MEMS switches. Shorting
the stubs to the patch creates a rejection band. At the resonance frequency, the direction
of the current on the stub flows in the opposite direction to the current along the nearby
edge of the radiator, so the radiated fields cancel. When the stubs are not shorted, the
antenna operates over the whole UWB range (3.1–10.6 GHz) without any rejection band.
The MEMS switches actuate through the radiofrequency (RF) signal path, without any
direct current (DC) bias lines, that might complicate the switch fabrication and integration
while degrading the radiation performance due to RF leakage through the bias.
Reconfigurable antennas can also change the antenna polarization. For instance, plac-
ing MEMS switches in the feeds of microstrip antennas provides the ability to switch

MEMS switches FIGURE 12-18

Diagram of the
reconfigurable
Slot Stub U-slot and L-stub
antennas.

U-slot L-stub
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 480

480 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-19 Antennas

Reconfigurable
antenna that change
polarization using
MEMS.

MEMS switch

Feed

FIGURE 12-20 A B
Patch with edges Patch
made from variable
conductivity
material.

Feed

Edges with variable conductivity

from one linear polarization to the orthogonal linear polarization or to circular polariza-
tion (Figure 12-19) [24].
A patch resonates at several frequencies if conducting extensions to the edges are
added or removed [26]. The patch antenna in Figure 12-20 lies on a 3.17 mm thick
polycarbonate substrate (εr = 2.62), and its edges at A and B made from a material with
variable conductivity, such as silicon. The size of the copper patch as well as the width of
A and B are optimized along with the feed location to produce a 50 input impedance at
1.715, 1.763, 1.875, and 1.930 GHz, depending on whether an edge is conductive. Graphs
of the reflection coefficient, s11 , from 1.6 to 2.0 GHz for the four cases are shown in
Figure 12-21. Each combination has a distinct resonant frequency.
The five-element array in Figure 12-22 has rectangular patches made from a perfect
conductor that is 58.7 × 39.4 mm [25]. The 88.7 × 69.4 × 3 mm optically transparent
fused quartz substrate (εr = 3.78) is backed by a perfectly conducting ground plane.
The patch has a thin strip of silicon 58.7 × 2 mm with εr = 11.7 that separates the
narrow right conducting edge of the patch (58.7 × 4.2 mm) from the main patch. A
light-emitting diode (LED) beneath the ground plane illuminates the silicon through small
holes in the ground plane or by making the ground plane from a transparent conductor.
The silicon conductivity is proportional to the light intensity. A graph of the amplitude
of the return loss is shown in Figure 12-23 when the silicon is 0, 2, 5, 10, 20, 50, 100,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 481

12.4 Reconfigurable Antennas and Arrays 481

0 0

−5 −5
Experimental
Computed
s11 (dB)

s11 (dB)
−10 −10

Computed
−15 −15
Experimental

−20 −20
1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2 1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2
Frequency (GHz) Frequency (GHz)
(a) (b)

0 0
Computed Computed

−5 −5
Experimental Experimental
s11 (dB)

s11 (dB)

−10 −10

−15 −15

−20 −20
1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2 1.6 1.65 1.7 1.75 1.8 1.85 1.9 1.95 2
Frequency (GHz) Frequency (GHz)
(c) (d)

FIGURE 12-21 Reflection coefficient (s11 ) versus frequency for the following: a: No edge.
b: Back edge. c: Front edge. d: Both edges are conductive.

Substrate Patch extension FIGURE 12-22

A five-element
Silicon linear array of
photoconductive
1 2 3 4 5
patches.

Patch

200, and 1,000 S/m. There is a distinct resonance at 2 GHz when the silicon has no
conductivity. Increasing the conductivity gradually changes the resonance to 1.78 GHz.
As the silicon conductivity increases, the resonant frequency of the patch changes, so the
photoconductive silicon acts as an attenuator. The element spacing in the array is 0.5λ.
Carefully tapering the illumination the LEDs creates so that the silicon conductivity at the
five elements is [16 5 0 5 16] S/m results in the antenna pattern in Figure 12-24. It has a
gain of 10.4 dB with a peak relative sidelobe level 23.6 dB.
An experimental adaptive array was constructed with photoconductive attenuators
integrated with broadband monopole antennas (Figure 12-25) [26]. A close-up of one
of the array elements appears in Figure 12-26. The attenuator consists of a meandering
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 482

482 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-23 1
Plots of the
magnitude of s11 for
silicon conductivities 0.8
of 0, 1, 2, 5, 10, 20,
30, 50, 75, 100, 200,
500, and 1,000 S/m.
0.6

|s11| 10 S/m
0.4

0.2 1000 S/m

0 S/m
0
1.75 1.8 1.85 1.9 1.95 2 2.05 2.1
Frequency (GHz)

FIGURE 12-24 15
The quiescent
pattern is the Quiescent
10 G = 10.14 dB
dashed line and has
all the conductivities
set to 0. The 5
adapted pattern is
s11 = 24 dB
the solid line and Adapted
Gain (dB)

0
has the silicon
conductivities set to
−5
[16 8 0 8 16] S/m.

−10

−15

−20
−80 −60 −40 −20 0 20 40 60 80
q (degrees)

center conductor in coplanar waveguide mounted on a silicon substrate [27]. This attenu-
ator is flip mounted over a hole in the coplanar waveguide feed of the broadband element
(Figure 12-26). An infrared (IR) LED illuminates the meandering line in the center rect-
angular section to increase in conductivity of the silicon and to create a short between the
center and outer conductors of the coplanar waveguide. Connecting this attenuator to an
array element attenuates the signal received by the element. The element bandwidth was
measured from 2.1 to 2.5 GHz (S11 < −10 dB). Figure 12-27 shows the array patterns the
uniform array (all LEDs off) and for two Chebyshev amplitude tapers found by relating
the LED current to the signal amplitude at the elements.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 483

12.4 Reconfigurable Antennas and Arrays 483

FIGURE 12-25
Picture of
experimental array
6 7 8 with numbered
3 4 5
1 2 elements.

FIGURE 12-26
The attenuator on
the left is flipped and
soldered to the
antenna element on
the right.

0 FIGURE 12-27
Measured low
sidelobe antenna
patterns for the
eight-element array
−10 Uniform
when the diode
Arrray pattern (dB)

currents are set

for uniform and 20
20 dB Chebyshev and 25 dB Dolph–
Chebyshev tapers.
−20

−30
25 dB Chebyshev

0 15 30 45 60 75 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 484

484 CHAPTER 12 Recent Developments in Adaptive Arrays

12.5 PERFORMANCE CHARACTERISTICS OF

LARGE SONAR ARRAYS
Acoustic waves have much longer wavelengths and slower velocity of propagation than
electromagnetic waves, so extracting information from the signals received by a sonar
array is considerably different from extracting it from a radar. Large aperture sonar arrays
contain hundreds of sensors. Such large arrays have some unexpected problems associated
with adaptive algorithms and acoustic propagation. Since the dominant noise source in
cluttered sonar environments is shipping, the signal environment is nonstationary with an
associated time scale that has important implications for the ability to form array pattern
nulls around moving ships [28].
Most adaptive algorithms either explicitly or implicitly form the sample covariance
matrix
K
1
R̂x x = xm xmH (12.29)
K m=1

There are limits determining the number of time samples (snapshots), K, which are avail-
able in forming the estimate R̂x x : (1) the time duration limit; and (2) the bandwidth limit,
over which frequency averaging can be done. At broadside the main lobe of a sonar
resolution cell has a cross-range extent given by

λr
x≈ (12.30)
L
where r is the range to the source, L is the array aperture length, and λ is the wavelength
of the transmitted sonar signal. For a source traveling with speed v, the associated bearing
rate is
v
φ̇ = (12.31)
r
Hence, the time spent within a single resolution cell is
 λr 
x λ
T= = L
= (12.32)
v v L φ̇

The transit time, Ttransit , of an acoustic wave across the face of the array at endfire is
given by

L
Ttransit = (12.33)
v
where v is the acoustic wave velocity in water (nominally 1,500 m/sec). The estimate of
the phase in the cross-spectra will be smeared if one averages over more than one-eighth
the transmit time bandwidth, so

1 v
B< = (12.34)
8Ttransit 8L
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 485

12.5 Performance Characteristics of Large Sonar Arrays 485

The time–bandwidth product given by the product of (12.33) and (12.34) yields the
approximate number of time samples (snapshots) available to form R̂x x , or
 2  2
λ 1 1 v f λ
K< T∗B= × = = (12.35)
L φ̇ 8Ttransit 8 f φ̇ L 8φ̇ L
where λ = vf .
Equation (12.35) expresses the inverse square law dependence on array lengths and
thereby forces L to be quite small (e.g., a 200 Hz source moving at 20 knots and a 10 km
distance transiting a 100 wavelength array limits the number of time samples to only 3).
This time–bandwidth product limit dependence on propagation speed illustrates a big
difference between sonar and radar.
The transit of a signal source across a resolution cell introduces eigenvalue spread into
the signal source spectrum. To see how this occurs, introduce the parameter μ, representing
the fraction of motion relative to a beamwidth:
L (cos(φ))
μ= (12.36)
λ
where denotes the beamwidth extent in radians, and φ is the angle off-broadside. When
the source is within a beamwidth, there is a single large eigenvalue associated with R̂x x ;
however, as soon as the motion occupies approximately one resolution cell, the second
eigenvalue becomes comparable. Splitting the source into two sources separated at one-half
the distance traveled approximates this behavior. Solving this two-component eigenvalue
problem leads to
  
1 πμ (π μ)2
λ̂2 = 1 − sin c2 ≈ (12.37)
4 2 48
The resulting eigenvalue distribution for a 10λ array with L = 500 is illustrated in Fig-
ure 12-28. A linear approximation for λ2 , denoted by “A”, is about 1 dB below the actual
value. A linear fit, denoted by “B,” is also indicated.

SNR = 300 dB, Monte Carlo runs N = 20, L = 500 (10l) FIGURE 12-28
5 Eigenvalues as a
A
Function of Motion
0 for a Moving Source
l1 B
A = −6.87 + 20 Log10 (m) Transiting a Sonar
−5
l2
10 Log10 (lK) - 10 Log10 (N * SNR)

B = −3.5 + 20 Log10 (m) Resolution Cell.

−20 From Baggeroer &
l3 Cox, 1999 Asilomar
−15 l4 Conference on
l5 Signals, Systems,
−20 and Computers,
pp. 103–08
−25

−30

−35

−40

−45
−20 −15 −10 −5 0 10 15
10 Log10 (m), m is motion in beam widths
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 486

486 CHAPTER 12 Recent Developments in Adaptive Arrays

Most array processing algorithms assume that plane waves impinge on the array. For
large apertures, however, wavefront curvature becomes significant, and plane wave models
are no longer appropriate. The signal source range below which wavefront curvature must
be taken into account is given by the Fresnel range,
 2
L2 λ L
RFresnel = = (12.38)
2λ 2 λ

Near-field effects become important even for modest frequencies (i.e., f > 100 Hz), and
range-dependent processing must be introduced. The general solution to this problem,
which is beyond the scope of this text, is known as matched-field processing, where the
full-field propagation model occurs in the array processing [29].
One common way of dealing with the problem of a scarce number of samples in form-
ing the sample covariance matrix is to introduce diagonal loading (discussed in Chapter 5).
The introduction of diagonal loading improves the poor conditioning of the covariance
matrix (although it hinders the detection of weak signals and introduces biases into the
resulting direction-of-arrival estimates).

12.6 ADAPTIVE PROCESSING FOR MONOPULSE

TRACKING ANTENNAS
Monopulse array patterns are formed by synthesizing sum ( ) and difference ( ) beams.
While it is straightforward to form adapted sum and difference beams, the resulting adapted
monopulse pattern, / , has a highly distorted slope, rendering it ineffective for target
angular location. To avoid this problem it is possible to use an approach that yields a
controlled monopulse pattern [30]. We will consider this problem for a linear antenna
array for which the angle off-boresight defines the tracking angle. The recommended
procedure first forms the sum beam using the classic weight vector

w = −1 s (12.39)

where  is the interference signal covariance matrix, and s is the steering vector for the
target of concern. The next step forms the difference beam (θ0 , f 0 ), where θ0 is the
target azimuth angle, and f 0 is the target Doppler frequency. The basic idea is to form the
difference beam such that the received interference in minimized while satisfying three
distinct constraints as follows:
1. (θ0 , f 0 ) = 0 (the difference beam response is zero at the target location)
(θ0 + θ, f 0 )
2.  = ks θ (maintain a constant slope where ks is a slope constant, and
(θ0 + θ, f 0 )
θ is the angle excursion from θ0 where the constraint holds)
(θ0 + θ, f 0 )
3.  = −ks θ (negative angle excursion)
(θ0 + θ, f 0 )
The steering vector can is the NM × 1 vector defined as

sT = [(s11 s12 . . . s1M ) · · · (sN1 sN2 . . . sNM )] (12.40)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 487

12.6 Adaptive Processing for Monopulse Tracking Antennas 487

reflecting an N -element array having a tapped delay line at each element and each tapped
delay line has M time taps with a T sec time delay between each tap. The elements of s are
snm = e j2 p[(n−1) λ d(sin θ0 )−(m−1) f0 T ]
2π
(12.41)
It is convenient to define the NM × 1 vector g(θ0 , f 0 ) as the previous vector s. With
these vectors, the constraints can be written in matrix notation as
HT w = ρ (12.42)
where
⎡ ⎤
gT (θ0 + θ, f 0 )
⎢ ⎥
HT = ⎣ gT (θ0 , f 0 ) ⎦ (12.43)
gT (θ0 − θ, f 0 )
⎡ T ⎤
w g(θ0 + θ, f 0 )
ρ = ks ⎣ 0 ⎦ θ (12.44)
−w g(θ0 − θ, f 0 )
T

The weight vector w that minimizes the difference beam interference power, w H w ,
subject to the constraint (12.42) is

w = −1 H∗ (HT −1 H∗ )−1 ρ (12.45)

To illustrate these results, consider a 13-element linear array with 14 taps in each
tapped delay line. Assume that the clutter-to-noise ratio is 65 dB per element. The sum
beam weight vector given by (12.1) produces a beam that has an interference plus noise
power after adaptation that is close to the noise floor for all target speeds V such that
0.05 < V/Vb < 0.95, where Vb is the radar blind speed. Likewise, the weight vector given
by (12.47) also produces a difference beam with an adapted, interference plus noise power
close to the noise floor for all target speeds. Figure 12-29 shows the adapted monopulse
pattern, / , for two different target speeds where θ = 0.644×(3 dB -beam angle).

0.7 FIGURE 12-29

= Constraint Positive angle
V portion of adapted
0.6 = 0.0625
Vblind monopulse for three
constraints. From
0.5 V Fante, IEEE Trans.
= 0.5
Vblind Ant. & Prop. May
Delta/Sigma

0.4 1999, pp. 773–74

0.3 Desired

0.2

0.1 3 dB angle
for Σ beam

0
0 0.5 1
Normalized Azimuth Angle
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 488

488 CHAPTER 12 Recent Developments in Adaptive Arrays

12.7 PARTIALLY ADAPTIVE ARRAYS

Consider the block diagram of an interference cancelling adaptive array system given in
Figure 12-30. For this system, a main array output is formed by merely combining the N
element outputs. An adaptive processor combines only M of the element signals (where
M ≤ N ) to form an adaptive output that is then subtracted from the main array output.
In the event that M = 1, a simple sidelobe canceller system results. On the other hand
if M = N , then the array is said to be fully adaptive. Cases that fall in between where
1 < M < N are referred to as a partially adaptive array (or as a multiple sidelobe canceller
for the given configuration).
A fully adaptive array in which every element is individually adaptively controlled is
obviously preferred since it affords the greatest control over the array pattern. For large
arrays containing thousands of elements, individually controlling every element can prove
a prohibitive implementation expense. Furthermore, signal processor implementation be-
comes much more difficult and costly as the dimensionality of the signal vector to be
processed increases. Consequently, it is highly desirable to reduce the dimensionality of
the signal processor while maintaining a high degree of control over the array response
by adopting one of the following adaptive control philosophies:
1. Judiciously select only a fraction of the array elements to adaptively control, thereby
resulting in elemental-level adaptivity [31,32].
2. Combine the N elements in the entire array into a collection of M subarrays by means of
a subarray beamformer transformation Q, and adaptively control each of the resulting
subarray outputs [33,34].

FIGURE 12-30
Interference
x1 x2 x3 x4 xN y1
cancelling adaptive
array block diagram. y2
Adaptive
processor
yM

Σ
Main
array m(t)
output +
− Adaptive element array output a(t)

Total
array
output e0 (t)
P0 = Pm − rym
T r*
ym
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 489

12.7 Partially Adaptive Arrays 489

Each of these approaches will now be discussed to determine the characteristics that typify
these partially adaptive array concepts.
As shown in Chapter 2, null synthesis can be applied to analyzing partial adaptive
nulling. A first step in picking the appropriate elements for a partially adaptive array is to
look at null synthesis using a subset of the elements. The synthesized array factor when a
subset of the elements has variable weights is given by

N Na
j (kxe(n) cos φ+δe(n) )
AF = an e jkxn cos φ − e(n) an e (12.46)
n=1 n=1

where
0 ≤ an ≤ 1.0
0 ≤ δn < 2π
Na = number of adaptive elements
e = vector containing indexes of the adaptive elements
Equation (12.46) can be placed in matrix form

A  = b (12.47)

where
⎡ ⎤
ae(1) e jk(e(1)−1)d cos φ1 · · · ae(Na ) e jk(e(Na )−1)d cos φ1
⎢ .. .. .. ⎥
A = ⎣ . . . ⎦
jk(e(1)−1)d cos φ M jk(e(Na )−1)d cos φ M
ae(1) e · · · ae(Na ) e

 T
 w = e(1) · · · e(Na )
 N N
T
b= w n e jk(n−1)d cos φ1 ··· w n e jk(n−1)d cos φ M
n=1 n=1
The Na adaptive weights in an N element partial adaptive array are written as

an (1 − n )e jδn if element n is adaptive
wn = (12.48)
an if element n is not adaptive

Four different subsets of four adaptive elements are considered:

(a) 3, 4, 5, 6
(b) 1, 2, 7, 8
(c) 1, 3, 5, 7
(d) 2, 4, 5, 8
All the configurations have a cancellation pattern peak at θ = −21◦ , but the rest of the
cancellation patterns are very different. The choice of the adaptive elements determines
the shape of the cancellation pattern thus the distortion to the adapted pattern. Consider an
eight-element array with λ/2 spacing. If all the elements are adaptive, then the synthesized
array factor and cancellation beam in Figure 12-31. Four contiguous adaptive elements
have the cancellation beam shown in Figure 12-32a. This cancellation pattern is just a
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 490

490 CHAPTER 12 Recent Developments in Adaptive Arrays

FIGURE 12-31
Adapted and 10 Adapted
cancellation patterns Cancellation
superimposed on Quiscent
the quiescent 0
pattern when a null
is synthesized at

Gain (dB)
θ = −21◦ with all
−10
elements adaptive.

−20

−30

−90 −45 0 45 90
f (degrees)

10 Adapted 10 Adapted
Cancellation Cancellation
Quiscent Quiscent
0 0
Gain (dB)

Gain (dB)

−10 −10

−20 −20

−30 −30

0 45 90 135 180 0 45 90 135 180

f (degrees) f (degrees)
(a) Elements 3, 4, 5, and 6 adaptive. (b) Elements 1, 2, 7, and 8 adaptive.

10 Adapted 10 Adapted
Cancellation Cancellation
Quiscent Quiscent
0 0
Gain (dB)

Gain (dB)

−10 −10

−20 −20

−30 −30

0 45 90 135 180 0 45 90 135 180

f (degrees) f (degrees)
(c) Elements 1, 3, 5, and 7 adaptive. (d) Elements 2, 4, 5, and 8 adaptive.

FIGURE 12-32 Adapted and cancellation patterns superimposed on the quiescent pattern
when a null is synthesized at θ = −21◦ with four of the eight elements adaptive.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 491

12.7 Partially Adaptive Arrays 491

four-element uniform array factor with its main beam steered to θ = −21◦ . When two
elements on each end of the array are adaptive, then the cancellation pattern has many
lobes of approximately the same height (Figure 12-32b). Making every other element in the
array adaptive induces grating lobes in the cancellation pattern as shown in Figure 12-32c.
Random spacing of the adaptive elements produces the high sidelobe but narrow main beam
cancellation pattern. Figure 12-32d shows the synthesized array factor and cancellation
beam when the random elements are 2, 4, 5, and 8.
The cancellation pattern for any experimental or computed adapted array pattern is
found by subtracting the quiescent electric field, E Quiescent , from the adapted electric field,
EAdapted .

ECancellationPattern = EAdapted − EQuiescent (12.49)

To plot the magnitude of the cancellation pattern correctly, ECancellationPattern is normalized

and adjusted relative to the peak gain of the adapted pattern.
To test the usefulness of (12.49), a 24-element uniform array with λ/2 spacing has
adaptive elements controlled by a genetic algorithm to minimize the total output power. A
desired signal of 0 dB is incident at θ = 0◦ , whereas an interference signal is incident at
θ = −21◦ . Figure 12-33a shows the case where the adaptive elements are four contiguous

Adapted Adapted
Cancellation Cancellation
10 Quiscent 10 Quiscent
Gain (dB)

Gain (dB)

0 0

−10 −10

−20 −20

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(a) Elements 11, 12, 13, and 14 (b) Elements 1, 2, 7, and 8 adaptive
adaptive complex weights. complex weights.

Adapted Adapted
Cancellation Cancellation
10 Quiscent 10 Quiscent
Gain (dB)

Gain (dB)

0 0

−10 −10

−20 −20

−30 −30
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(c) Elements 11, 12, 13, and 14 (d) Elements 1, 2, 7, and 8
adaptive phase weights. adaptive phase weights.

FIGURE 12-33 Adapted and cancellation patterns for a 24-element array superimposed
on the uniform quiescent pattern when a null is placed at θ = −22◦ .
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 492

492 CHAPTER 12 Recent Developments in Adaptive Arrays

elements in the center of the array. The nulling is accomplished with the side of the
main beam of the cancellation beam rather than the peak. As a result, the loss in main
beam gain and sidelobe distortion is significant but better than in the eight-element case.
When the tow edge elements on either end of the array are adaptive, then the problems
are similar to the previous case, but the distortion is different, because the cancellation
pattern has many narrow lobes of nearly the same height rather than a uniform cancellation
pattern (Figure 12-33b). Phase only adaptive nulling does not fare any better as shown in
Figure 12-33c and Figure 12-33d.
Partial adaptive nulling experiments were done using the array in Figure 12-25. The
output from the receiver goes to a computer with a genetic algorithm controller. The
genetic algorithm varies the current fed to the LEDs to control signal attenuation at the
adaptive elements. At least 15 dB of attenuation is available at each element. The largest
possible decrease in gain occurs when the LEDs of the four adaptive elements are fed with
250 mA of current. The array becomes a four-element uniform array, so the main beam
should decrease by 6 dB. Figure 12-34 shows a 5.2 dB decrease in the measured main
beam of the far-field pattern. Thus, the adaptive array reduces the desired signal entering
the main beam but cannot place a null in the main beam.
Figure 12-35a is the adapted pattern when one signal is incident on the array at −19◦
and elements 1, 2, 7, and 8 are adaptive. The adaptation lowered the main beam by 3.9 dB
and the sidelobe level by 15.7 dB at −19◦ . Putting a −10 dBm desired signal at 0◦ and a
15 dBm signal incident at −19◦ produces the adapted pattern in Figure 12-35b. Its main
beam goes down by −3.7 dB, and the sidelobe level at −19◦ goes down by 17.1 dB. These
results show that adaptation with the desired signal present was approximately the same
as when it was absent. Nulls can be placed in any sidelobe of the array pattern.
The adapted pattern when two 15 dBm signals are incident at −35◦ and −19◦ is
shown in Figure 12-36. The main beam is reduced by 3.6 dB, while the sidelobe at
−35◦ goes from −14 dB to −23 dB, and the sidelobe at −19◦ goes from −11.8 dB
to −24.3 dB. Figure 12-37 shows the convergence for six independent random runs.
Significant improvement occurs after five iterations.

FIGURE 12-34
Antenna pattern
when the four 0
Quiescent
adaptive elements
are turned off.

−10
Far field (dB)

−20

Adapted
−30

−90 −45 0 45 90
q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 493

12.7 Partially Adaptive Arrays 493

0 Quiescent 0
Quiescent

Far field (dB)

Far field (dB)

−10 −10

−20 −20

−30 −30 Adapted

Adapted

−90 −45 0 45 90 −90 −45 0 45 90

q (degrees) q (degrees)
(a) One signal at −19°. (b) Two signals at −19° and −0°.

FIGURE 12-35 Adapted pattern (b) with and (a) without a signal present in the main beam.

FIGURE 12-36
Adapted pattern
0 when signals are
Quiescent
incident on the array
at −35◦ and −19◦ .

−10
Far field (dB)

−20

Adapted
−30

−90 −45 0 45 90
q (degrees)

1.0 FIGURE 12-37

Convergence of the
adaptive algorithm
when the signals are
.75 incident on the array
at −35◦ and −19◦
Relative output power

for six independent

random runs.
.5

.25

0 5 10 15 20 25 30
Generation
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 494

494 CHAPTER 12 Recent Developments in Adaptive Arrays

The adapted pattern for a signal is incident at 35◦ when elements 1 and 8 are adaptive
is shown in Figure 12-38a. The main beam is reduced by 0.9 dB, whereas the sidelobe at
35◦ is reduced by 22 dB. A −10 dBm signal incident at 0◦ and a 15 dBm signal incident at
−18◦ with elements 1 and 7 adaptive result in the adapted pattern in Figure 12-38b. The
main beam reduction is 2.7 dB, and the sidelobe level at −18◦ is reduced by 13.1 dB. The
main beam gain reduction is less when only two elements are adaptive compared with
when four elements are adaptive.

12.7.1 Adaptive Subarray Beamforming

The fundamental concept of adaptive subarray beamforming discussed by Chapman [33]
is to reduce the required dimensionality of the signal processor by introduction of an
N × M transformation matrix Q (called a subarray beamformer) as shown in Figure 12-39.
The subarray signals resulting from the elements so combined in subarrays are then all
adaptively controlled to produce the total array response. From Figure 12-39 it follows that

y = QT x (12.50)

Consequently, the covariance matrix of the subarray signal vector is given by

R yy = Q† Rx x Q (12.51)

0 Quiescent 0
Quiescent
Far field (dB)

Far field (dB)

−10 −10

−20 −20

−30 −30 Adapted

Adapted
−90 −45 0 45 90 −90 −45 0 45 90
q (degrees) q (degrees)
(a) Elements 1 and 8 are adaptive. (b) Elements 1 and 7 are adaptive.

FIGURE 12-38 Adapted pattern when a signal is incident on the array at 35◦ and two
elements are adaptive.

FIGURE 12-39 x1 x2 xN Element signals

Use of a subarray
beamformer
to reduce the
signal vector
dimensionality. Q Transformation matrix
(subarray beamformer)

y1 y2 yM Subarray signals
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 495

12.7 Partially Adaptive Arrays 495

If the adaptive processor employs the Howells–Applebaum SNR performance measure

algorithm, then the optimum weight vector solution is given by

w yopt = αR−1 ∗
yy v y (12.52)

where the beam steering vector for the subarray v y is related to the beam steering vector
for the total array vx by

v y = QT vx (12.53)

Consequently, the resulting array beam pattern can be computed by using the implied
relationship

wx = Qw yopt (12.54)

Note that wx of (12.78) is not the same as wxopt = αR−1 ∗

x x vx for the fully adaptive array but
is rather a suboptimum solution that is constrained by the subarray configuration.
Now consider the effect of element-level amplitude and phase errors in a fully adaptive
array and in a partially adaptive array. Amplitude and phase errors between any two
channels is characterized by a mean offset signal level (over the bandwidth of interest)
plus a variation about the mean as a function of frequency. The adjustment of a single
complex weight in one channel removes any mean error that exists between a pair of
channels, so a fully adaptive array is sensitive only to the error variance about the mean.
A subarray configuration, however, does not have as many degrees of freedom as the
fully adaptive array and consequently may not succeed in removing all the mean errors
that may exist among all the elements of the entire array. Thus, the performance of an
array that is partitioned into subarrays is more susceptible to element-level errors than a
fully adaptive array.
The effect of element random phase errors on a subarray structure are determined
by appropriately modifying the Q subarray beamformer matrix. Define an “errored” Q
matrix Qε , which is given by
Qε = Eρ Q (12.55)
where Eρ is an N × N diagonal matrix with elements given by
ρ
eii = exp[ j2π(1 − ρe )z i ] (12.56)
The parameter ρe (0 ≤ ρe ≤ 1) represents the severity of the error (with no error corre-
sponding to ρe = 1), whereas z i is a uniformly distributed random variable (−0.5 ≤ z i ≤
0.5).
This errored Q matrix model results in a random sidelobe structure (RSL) for the total
array whose mean level (with respect to isotropic) is approximately given by [35]
π 2 (1 − ρe )2
|gRSL (θ)|2 ∼
= |ge (θ )|2 (12.57)
3
where ge (θ ) denotes the directional voltage gain of an array element.
Choosing subarrays by merely grouping physically contiguous elements is termed
a simple subarray approach. Any simple subarray configuration has the subarray phase
centers separated by several wavelengths produce grating lobes that cannot be altered by
the adaptive processor [36]. A “beam-space” subarray, in which the full array aperture
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 496

496 CHAPTER 12 Recent Developments in Adaptive Arrays

is used for each subarray so the resulting structure can be regarded as a multibeam pro-
cessor, avoids the grating lobe problem. Vural [37] likewise showed that a beam-based
processor realizes superior performance in partially adaptive operations under diverse in-
terference conditions. To introduce constraints into beam-space adaptive algorithms, the
only difference compared with element-space data algorithms is in the mechanization of
the constraint requirements [34].
Subarray groups for planar array designs are chosen by combining row (or column)
subarrays. This choice is adaptive in only one principal plane of the array beam pattern
and may therefore be inadequate. A more realistic alternative to the row–column subarray
approach is a configuration called the row–column precision array (RCPA) [38], in which
each element signal is split into two paths: a row path and a column path. All the element
signals from a given row or column are summed, and all the row outputs and column
outputs are then adaptively combined. The number of degrees of freedom in the resulting
adaptive processor equals the number of rows plus the number of columns in the actual
array.
When ideal operating conditions are assumed with perfect element channel matching,
simulation studies have shown that the subarray configurations previously discussed yield
array performance that is nearly the same as that of fully adaptive arrays [33]. When the
array elements have independent random errors, however, the resulting random sidelobe
structure severely deteriorates the quality of the array response. This performance dete-
rioration results from the need for precision in the subarray beamforming transformation
since this precision is severely affected by the random element errors.

12.7.2 Element-Level Adaptivity

Adaptive subarray beamforming is a highly attractive solution to the partially adaptive
array problem, particularly if the application inherently requires multiple beamforming.
However, the beamforming matrix transformation introduces an additional expense into
the realization that can be avoided if multiple beams are not required by simply directly
controlling only a fraction of the array elements in an element-level partial adaptivity
scheme. With this approach, the question arises as to which elements of the original array
should be adaptively controlled to achieve the best possible array performance. To gain
some insight into the behavior of the system depicted in Figure 12-30, the approach taken
by Morgan [32] is followed, and an explicit solution for a narrowband two-jammer problem
is obtained.
Using the MMSE performance measure, the optimum weight vector for the adaptive
processor is given by

W = R∗−1
yy r ym (12.58)

where R yy = E{yy† } is the M × M covariance matrix of the adaptive element signal

vector, and r ym = E{my∗ } is the cross-correlation vector of the main array signal and
the adaptive element signal vector. It is worth noting that the solution for w in (12.58)
results instead of the more familiar expression w = R−1 yy r ym , because Morgan defined
R yy = E{yy† } instead of the more definition R yy = E{y∗ yT }. The minimum total array
output power that results when (12.58) is used is then

P0 = Pm − rTym R−1 ∗
yy r ym (12.59)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 497

12.7 Partially Adaptive Arrays 497

where Pm is the main array output power. The straightforward evaluation of (12.59) does
not yield much insight into the relationship between adaptive performance and the array
configuration and jamming environment. A concise mathematical expression of (12.59)
is accordingly more desirable to elucidate the problem.
Consider two narrowband jammers of frequencies ω1 and ω2 . The composite signal
vector y is written as

y(t) = n(t) + J1 exp( jω1 t)v1 + J2 exp( jω2 t)v2 (12.60)

where J1 and J2 represent the reference amplitude and phase, respectively, of the jammers,
and n is the additive noise vector with independent components of equal power Pn . The
spatial vectors v1 , v2 have components given by
 
2π
Vk, i = exp j (ri · uk ) , i ∈ A, k = 1, 2 (12.61)
λ
where A denotes the subset of M adaptive elements, ri is the ith element location vector,
and uk is a unit vector pointed in the direction of arrival of the kth jammer. For a linear
array having elements aligned along the x-axis, (12.61) reduces to
 
2π
Vk, i = exp j xi sin θk (12.62)
λ
where θk is the angle of arrival of the kth jammer measured from the array boresight.
The main array output signal can likewise be written as
N
m(t) = n i (t) + J1 e jω1 t h 1 + J2 e jω2 t h 2 (12.63)
i=1

where
N
hk = vk, i , k = 1, 2 (12.64)
i=1

is the main array factor, which is computed by forming the sum of all spatial vector
components for all the array elements i = 1, . . . , N . It follows that the array output signal
is written as

e0 (t) = m(t) − wT y(t) (12.65)

where w is the vector of adaptive weights.

The minimum total array output power given by (12.59) requires the covariance matrix
R yy and the cross-correlation vector r ym , both of which can be computed from (12.60)
and (12.61), to yield

R yy = E{yy† } = Pn I + P1 v1 v†1 + P2 v2 v†2 (12.66)

r ym = E{y∗ m} = Pn 1 + P1 h 1 v∗1 + P2 h 2 v∗2 (12.67)

where P1 and P2 denote the jammer power levels, and 1 is an M × 1 vector of ones.
Since (12.58) requires the inverse of R yy , this inverse is explicitly obtained by the
twofold application of the matrix inversion identity (D.10) of Appendix D to (12.66).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 498

498 CHAPTER 12 Recent Developments in Adaptive Arrays

Substitution of (12.67) along with R−1

yy into (12.58) then yields
   
∗ Pn ∗ Pn
wopt = 1 + γ F1 − ρ F2 + F1 v1 + γ F2 − ρ F1 + F2 v∗2 (12.68)
M P2 M P1
where

Fk = vk, i (12.69)
i∈ A

denotes the complementary array factor that is computed by summing the spatial vector
components for all the elements that are not adaptively controlled, the factor
1 ∗
ρ= v1, i v2, i (12.70)
M i⊂A

is the complex correlation coefficient of the adaptive element spatial vectors, and
 −1
Pn Pn Pn2
γ = M(1 − |ρ| ) + + 2
+ (12.71)
P1 P2 M P1 P2

For a fully adaptive array where N = M, then (12.58) reduces to wopt = 1 in the
absence of any constraints on the main lobe. If it is assumed that both jammers are much
stronger than the noise so that P1 , P2 Pn , then (12.58) becomes
F1 − ρ ∗ F2 ∗ F2 − ρ F1 ∗
Wopt | P1 ,P2 →∞ → 1 + v + v (12.72)
M(1 − |ρ|2 ) 1 M(1 − |ρ|2 ) 2
The previous expression shows that the adaptive weights are ill conditioned whenever
|ρ| ≈ 1. This condition physically corresponds to the situation that occurs when the
adaptive array pattern cannot resolve the two jammers. Such a failure to resolve occurs
either because the two jammers are too close together or because they simultaneously
appear on distinct grating lobes.
Recognizing that the output residue power is given by

P0 = Pm − rTym w∗opt (12.73)

we see that it follows that the normalized output residue power is expressed as
  
P0 ∗ Pn |F1 |2 |F2 |2
= N − M + γ |F1 | + |F2 | − 2Re(ρ F1 F2 ) +
2 2
+ (12.74)
Pn M P2 P1

When N = M (a fully adaptive array) then, in the absence of main lobe constraints,
P0 = 0. Equation (12.64) immediately yields an upper bound for the maximum residue
power as

P1 P2 (|F1 | + |F2 |)2 + Pn

(P2 |F1 |2 + P1 |F2 |2 )
P0 ≤ P0max = M
(12.75)
P1 + P2 + PMn

and P0 = P0max when ρmax = − exp[arg(F1∗ F2 )].

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 499

12.7 Partially Adaptive Arrays 499

In the case of strong jammers for |ρ| < 1, then (12.64) reduces to

P0  |F1 |2 + |F2 |2 − 2Re(ρ F1 F2∗ )
 →N−M+ , |ρ| < 1 (12.76)
Pn P1 ,P2 →∞ M(1 − |ρ|2 )
This expression emphasizes the fact that the residue power takes on a large value whenever
|ρ| = 1. Furthermore, (12.66) surprisingly does not depend on the jammer power levels P1
and P2 but instead depends only on the geometry of the array and the jammer angles. The
maximum residue power upper bound (12.65) does depend on P1 and P2 , however, since
 P1 P2
P0max  P1 ,P2 →∞ → (|F1 | + |F2 |)2 (12.77)
P1 + P2
The foregoing expressions for output residue power demonstrate the central role the
correlation coefficient plays between the adaptive element spatial vectors in characterizing
the performance of a partially adaptive array. In simple cases, this correlation coefficient
ρ can be related to the array geometry and jammer angles of arrival thereby demonstrat-
ing that it makes a significant difference which array elements are chosen for adaptive
control in a partially adaptive array. In most cases, the nature of the relationship between
ρ and the array geometry is so mathematically obscured that only computer solutions
yield meaningful solutions. For the two-jammer case, it appears that the most favorable
adaptive element locations for a linear array are edge-clustered positions at both ends of
the array [32].
The previous results described for a partially adaptive array and a two-jammer envi-
ronment did not consider the effect of errors in the adaptive weights, even though they
are very important [31]. The variation in the performance improvement achieved with a
partially adaptive array with errors in the nominal pattern weights is extremely sensitive
to the choice of adaptive element location within the entire array. Consequently, for a
practical design the best choice of adaptive element location depends principally on the
error level experienced by the nominal array weights and only secondarily on the optimum
theoretical performance that can be achieved. It has been found that spacing the elements
in a partially adaptive array using elemental-level adaptivity so the adaptively controlled
elements are clustered toward the center is desirable since then the adapted pattern in-
terference cancellation tends to be independent of errors in the nominal adaptive weight
values [31].

12.7.3 Eigenspace Beamformers

An eigendecomposition used by an eigenspace beamformer projects the input onto a
reduced-rank subspace called the eigensubspace containing the signal and interference
[39]. The projected input is then processed to form the beam. This technique is especially
useful in designing a partially adaptive array whose response is nearly the same as the
response of a fully adaptive array.
The eigendecomposition of an N × N signal covariance matrix may be described by
N
Rx x = λi i iH = EE H (12.78)
i=1

where λi are the eigenvalues of Rx x , i are the corresponding eigenvectors, E is an N × N

matrix of eigenvectors, and  is a diagonal matrix of ordered eigenvalues. We may now
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 500

500 CHAPTER 12 Recent Developments in Adaptive Arrays

select Dr of these eigenvalues to form an N × Dr matrix, Er , which is a “reduced”

subspace since Dr < N . How are the eigenvectors selected to be used in Er ? There
are two generally accepted approaches. One way postulates that there are D plan wave
signals in the signal environment; then select the D + 1 largest eigenvalues. The signal
plus interference subspace is then spanned by the eigenvector matrix E S+I , where

E S+I = [1 2 . . .  D+1 ] (12.79)

Using this approach, the number “D” must be estimated.

An alternate approach using the same eigendecomposition (12.78) does not assume
any structure for the interference and merely selects the Dr largest eigenvalues. Various
tests can be used for determining if an eigenvalue is “large” enough to be included in Dr
as follows:
(i) Consider the % of total power contained in λi
λi
> n1
tr (Rx x )
(ii) Consider the % of total power contained in Dr eigenvalues
 Dr
i=1 λi
> n2
tr (Rx x )
(iii) Consider the ratio of adjacent eigenvalues
λi+1
> n3
λi
Having selected Dr , define Er = [1 2 . . .  Dr and  = diag[λ1 λ2 · · · λ Dr ]. Then

Rx x = E S+I  S+I E S+I

H
+ E N  N E NH (12.80)

where E N is orthogonal to E S+I (E NH E S+I = 0) and, hence,

−1 −1 H
R−1
x x = E S+I  S+I E S+I + E N  N E N
H
(12.81)

Introducing the desired signal steering vector, vd , which satisfies w H vd = 1, the minimum
power distortionless response (MPDR) beamformer is then given by
R−1
x x vd E S+I −1 H
S+I E S+I vd
wMPDR = = (12.82)
vdH R−1x x vd vdH E S+I −1 H
S+I E S+I vd

since vd is in E S+I and hence is orthogonal to E N . A corresponding result for the model
of (12.78) yields
Er r−1 ErH vd
WMPDR = (12.83)
vdH Er r−1 ErH vd
Several variations on the theme of eigendecomposition have been suggested [40–42], but
the previously given development expresses the core ideas about the various schemes that
are developed.
An interesting application of eigendecomposition to the interference nulling problems
facing radio telescopes is discussed in [43]. In radio astronomy applications, detection of
the desired signal cannot take advantage of traditional demodulation–detection algorithms,
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 501

12.7 Partially Adaptive Arrays 501

and the signal of interest is so minute that it is many orders of magnitude weaker than the
noise power spectral density, leading to minutes or even hours of integration to achieve
a positive detection. Radio frequency interference (RFI) is a growing problem for ra-
dio astronomy; hence, there is strong motivation to exploit adaptive nulling techniques.
Common sources of RFI in radio astronomy include nongeosynchronous satellites and
land-mobile radio. To deal effectively with such dynamic signals, it is necessary to use a
weight update period that is on the order of 10 msec. On the other hand, a self-calibration
process updated about once per minute is highly important for radio astronomy, since it
removes environmental and instrumental errors. Self-calibration methods require nearly
stationary adapted array patterns between self-calibration updates; hence, most standard-
weight update algorithms, which are subject to “weight jitter,” are inappropriate for radio
astronomy. To circumvent the weight jitter problem, the eigendecomposition of (12.80) is
introduced, and the further decomposition of (12.80) is used. Since the desired signal in
radio astronomy is miniscule, the signal term in (12.80) can safely be ignored, leaving

Rx x = E I  I E IH + E N  N E NH (12.84)

Equation (12.84) is another way of saying that only interference and noise are in the
observations, since the signals are not visible within the timeframe of a weight update.
The term E I  I E IH completely describes the interference, and the column span of E I
is referred to as the interference subspace. Likewise, column span of E N is the noise
subspace. Rather than form an estimate of Rx x as is done with most pattern nulling
algorithms, an attractive approach is to use subspace tracking, in which an estimate of the
interference subspace is formed directly using rank-ordered estimates of the eigenvectors
and associated eigenvalues. This approach has been found to be able to form the desired
interference nulls without experiencing any associated weight (or pattern) jitter.
The estimation of E I is accomplished by recourse to the projection approximation
subspace tracking with deflation (PASTd) method [44] in combination with the Gram–
Schmidt orthonormalization procedure to ensure that the resulting eigenvectors are or-
thonormal. It is beyond the scope of this chapter to present the details of this approach;
suffice it to say that accurate interference subspace identification has been achieved for an
interference-to-noise ratio of unity.

12.7.4 The J Subbeam Approach

A large array with N elements suffers certain disadvantages when operated as a fully
adaptive array—such as an excessively large number of computations and poor sidelobe
performance in directions away from jammer locations. Whereas a fully adaptive array
may significantly degrade the array sidelobes as well as the main lobe peak gain, a partially
adaptive array consisting of (J +1) beams, where J denotes the number of jammers present
[45]. By forming an equivalent small array having (J + 1) ports, the output of each port is
treated as an “element” output to which any standard adjustment algorithm can be applied.
Whereas N may be very large (10,000 or more for large arrays), (J + 1) for the equivalent
array are very small, resulting in significant performance improvement.
The technique of forming an equivalent small array uses a two-step process. First, the
number of interfering jammers and their locations are estimated using such techniques as
a spatial discrete Fourier transform of the array outputs, by maximum entropy spectral
estimation techniques [44,45], or just by a search in angle with an auxiliary beam. Once
the first step is completed, auxiliary beams are formed pointing at each of the jammers,
and one auxiliary beam forms pointing at the target, and an adaptive algorithm applied to
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 502

502 CHAPTER 12 Recent Developments in Adaptive Arrays

Jammer
No. 1 Jammer
No. 2
Target

Jammer
No. J

Main
array
……
1 2 N

J + 1 output for
adaptive
algorithm input
Σ Σ Σ Σ
0
Adaptive-
1
adaptive
2 SMI
array

…
…
output
J

Main Auxiliary Auxiliary Auxiliary

beam beam beam beam
pointing pointing pointing
at jammer at jammer at jammer
number 1 number 2 number J

FIGURE 12-40 Adaptive–adaptive array configuration. N-element array is reduced to

J + 1 element array where J is equal to the number of jammers. From Brookner and Howell,
Proc. IEEE, April 1986, pp. 602–604

the signals appearing at the (J + 1) ports. The implementation of this approach is shown
in Figure 12-40. The number of elements in the equivalent small array is only (J + 1), so
the equivalent signal covariance matrix is only (J + 1) × (J + 1).
The reason the adaptive–adaptive approach does not degrade the antenna sidelobes is
that the equivalent small array subtracts one auxiliary beam pointing at the jammer from
the main signal channel beam. The gain of the auxiliary beam in the direction of the jammer
equals the gain of the main channel beam sidelobe in the direction of the jammer. As a
result, the subtraction produces a null at the jammer location in the main channel sidelobe.
Further variations of this basic scheme are discussed in the previously noted reference.

12.7.5 The Subarray Design Approach

One promising avenue to the design of a partially adaptive array is the subarray design
approach [46]. The basic idea is to divide the original array into a series of tapered subarrays
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 503

12.9 Problems 503

resulting in a series of fixed beams using fixed weight vectors. Each of the resulting fixed
beams may then be treated as a single element by weighting it with a single adaptive
weight. Whereas the original array contains N elements, following the division into K
fixed beams, there will be only K adaptive weights where K < N . It is not surprising
that relatively good interference cancellation performance was observed to occur with
subarrays comprised of elements clustered about the edge of the original array.

12.8 SUMMARY AND CONCLUSIONS

Beam switching is a very simple and cheap adaptive process that uses multiple beam arrays.
In contrast, other techniques requiring extensive hardware and software complexity, such as
DPCA, STAP, and MIMO, have received significant attention over the past 20 years. These
techniques promise to overcome significant problems with clutter and multipath as well as
rejecting interference. Reconfigurable antennas adapt by physically changing the antenna.
The use of eigendecomposition in combination with modern spectral analysis techniques is
yielding useful results in both sonar and radiotelescope applications, both of which experi-
ence unique technical problems not found in most radiofrequency applications. Compensa-
tion for mutual coupling in small arrays and the use of adaptive array concepts in monopulse
tracking antennas are two areas that also offer promising results. The notions of subspace
fitting and eigenspace beamformers were introduced using the concept of eigendecom-
position as the point of departure. Partially adaptive array concepts are highly important
because they offer the possibility of realizing near optimal array performance with only a
fraction of the control elements (and hence the cost) required for a fully adaptive array.

12.9 PROBLEMS
1. Multiple Beams. Write a program in MATLAB to generate Figure 12-1.
2. MIMO. A MIMO system has a three-element array of isotropic point sources spaced d apart on
transmit and receive. The system operates at 2.4 GHz, and the arrays are 100 m apart and face
each other. Show how the condition number of H changes as the element spacing increases.
3. Partial Adaptive Nulling. Plot the cancellation patterns associated with placing a null at θ =
16.25◦ in the array factor of a 32-element uniform array with λ/2 element spacing. Four different
configurations of eight adaptive element are considered:
a. 1,2,3,4,29,30,31,32
b. 13,14,15,16,17,18,19,20
c. 1,5,9,13,17,21,25,29
d. 2,8,13,16,18,23,24,30
4. Element Level Adaptivity Recognizing that

A−1 b† bA−1
(A + bb† )−1 = A−1 −
(1 + b† A−1 b)
†
let A = Pn I + P1 v1 v1 in (12.66), and apply the previously given matrix identity once. Finally,
apply the previous matrix identity to the resulting expression and show that (12.68) results from
these operations on substitution of (12.66) into (12.58).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 504

504 CHAPTER 12 Recent Developments in Adaptive Arrays

12.10 REFERENCES
[1] W. Rotman and R. Turner, “Wide-Angle Microwave Lens for Line Source Applications,” IEEE
Transactions on Antennas and Propagation, Vol. 11, No. 6, 1963, pp. 623–632.
[2] “Rotman Lens Designer,” Remcom, Inc., 2009.
[3] J. Butler and R. Lowe, “Beam-Forming Matrix Simplifies Design of Electronically Scanned
Antennas,” Electronic Design, Vol. 9, April 1961, pp. 170–173.
[4] J. P. Shelton, “Fast Fourier Transforms and Butler Matrices,” Proceedings of the IEEE, Vol. 56,
No. 3, 1968, pp. 350–350.
[5] G. W. Stimson, Introduction to Airborne Radar, Raleigh, NC, SciTech Pub., 1998.
[6] C. E. Muehe and M. Labitt, “Displaced-Phase-Center Antenna Technique,” Lincoln Lab Jour-
nal, Vol. 12, No. 2, 2000, pp. 281–296.
[7] J. K. Day and F. M. Staudaher, “Airborne MTI,” Radar Handbook, M. I. Skolnik, ed., New York,
McGraw Hill, 2008, pp. 3.1–3.34.
[8] R. Klemm and Institution of Engineering and Technology, Principles of Space-Time Adaptive
Processing, 3rd ed., London, Institution of Electrical Engineers, 2006.
[9] Y. Dong, “A New Adaptive Displaced Phase Centre Antenna Processor,” 2008 International
Conference on Radar, September 2–5, 2008, pp. 343–348.
[10] R. S. Blum, W. L. Melvin, and M. C. Wicks, “An Analysis of Adaptive DPCA,” Proceedings
of the 1996 IEEE National Radar Conference, 1996, pp. 303–308.
[11] H. Shnitkin, “Joint STARS Phased Array Radar Antenna,” IEEE Aerospace and Electronic
Systems Magazine, Vol. 9, No. 10, October 1994, pp. 34–40.
[12] L. E. Brennan, and L. S. Reed, “Theory of Adaptive Radar,” IEEE Transactions on Aerospace
and Electronic Systems, Vol. AES-9, No. 2, 1973, pp. 237–252.
[13] L. Brennan, J. Mallett, and I. Reed, “Adaptive Arrays in Airborne MTI Radar,” IEEE Trans-
actions on Antennas and Propagation, Vol. 24, No. 5, 1976, pp. 607–615, 1976.
[14] K. Gerlach and M. J. Steiner, “Fast Converging Adaptive Detection of Doppler-Shifted,
Range-Distributed Targets,” IEEE Transactions on Signal Processing, Vol. 48, No. 9, 2000,
pp. 2686–2690.
[15] J.O. McMahon, “Space-Time Adaptive Processing on the Mesh Synchronous Processor,”
Lincoln Lab Journal, Vol. 9, No. 2, 1996, pp. 131–152.
[16] W. Dziunikowski, “Multiple-Input Multiple-Output (MIMO) Antenna Systems,” Advances
in Direction of Arrival Estimation, S. Chandran (Ed.), Boston, MA, Artechhouse, 2006,
pp. 3–19.
[17] R. Klemm, “STAP Architectures and Limiting Effects,” in ROT Lecture Series 228: Military
Application of Space-Time Adaptive Processing, April 2003, pp. 5-1–5-27.
[18] V. F. Mecca and J. L. Krolik, “MIMO STAP Clutter Mitigation Performance Demonstration
Using Acoustic Arrays,” in 42nd Asilomar Conference on Signals, Systems and Computers,
2008, pp. 634–638.
[19] M. Rangaswamy, “An Overview of Space-Time Adaptive Processing for Radar,” Proceedings
of the International Radar Conference, 2003, pp. 45–50.
[20] J. B. Andersen, “Array Gain and Capacity for Known Random Channels with Multiple Element
Arrays at Both Ends,” IEEE Journal on Selected Areas in Communications, Vol. 18, No. 11,
November 2000, pp. 2172–2178.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 505

12.10 References 505

[21] Wireless InSite, 2.5, Remcom, Inc. ([Link]), 2009.

[22] M. A. Jensen and J. W. Wallace, “A Review of Antennas and Propagation for MIMO Wireless
Communications,” IEEE Transactions on Antennas and Propagation, Vol. 52, No. 11, 2004,
pp. 2810–2824.
[23] J. H. Winters, “On the Capacity of Radio Communication Systems with Diversity in a Rayleigh
Fading Environment,” IEEE J. Select. Areas Commun., Vol. 5, June 1987, pp. 871–878.
[24] S. Nikolaou, N. D. Kingsley, G. E. Ponchak, J. Papapolymerou, and M. M. Tentzeris, “UWB
Elliptical Monopoles With a Reconfigurable Band Notch Using MEMS Switches Actuated
Without Bias Lines,” IEEE Transactions on Antennas and Propagation, Vol. 57, No. 8, August
2009, pp. 2242–2251.
[25] R. L. Haupt, “Reconfigurable Patch with Switchable Conductive Edges,” Microwave and
Optical Technology Letters, Vol. 51, No. 7, pp. 1757–1760, July 2009.
[26] R. L. Haupt, “An Adaptive Rectangular Microstrip Patch Antenna Array Element Using
Photonic Controls,” IEEE Aerospace Conference, Big Sky, MT, March 2008.
[27] R. L. Haupt, J. R. Flemish, and D. W. Aten, “Broadband Linear Array with Photoconductive
Weights,” IEEE AWPL, Vol. 8, 2009, pp. 1288–1290.
[28] A. B. Baggeroer and H. Cox, “Passive Sonar Limits upon Nulling Multiple Moving Ships
with Large Aperture Arrays,” Conference Record of the Thirty-Third Asilomar Conference
on Signals, Systems, & Computers, IEEE (Cat. No. CH37020), Vol. 1, Piscataway, NJ, 1999,
pp. 103–08.
[29] A. Baggeroer, W. Kuperman, and H. Schmidt, “Matched Field Processing: Source Localization
in Correlated Noise as an Optimum Parameter Estimation Problem,” J. Acoust. Soc. of Amer.,
Vol. 83, February 1988, pp. 571–587.
[30] R. L. Fante, “Synthesis of Adaptive Monopulse Patterns,” IEEE Trans. Ant. & Prop., Vol
AP-47. No. 5, May 1999, pp. 773–774.
[31] R.L. Haupt, “Adaptive nulling with weight constraints,” Progress In Electromagnetics Re-
search B, Vol. 26, 2010, pp. 23–38.
[32] D. R. Morgan, “Partially Adaptive Array Techniques,” IEEE Trans. Antennas & Prop., Vol.
AP-26, No. 6, November 1978, pp. 823–833.
[33] D. J. Chapman, “Partial Adaptivity for the Large Array,” IEEE Trans. Antennas & Prop., Vol.
Ap-24, No. 5, September 1976, pp. 685–696.
[34] A. M. Vural, “An Overview of Adaptive Array Processing for Sonar Applications,” EASCON
1975 Record, IEEE Electronics and Aerospace Systems Convention, September 29–October 1,
Washington, DC, pp. 34.A–34.M.
[35] M. Viberg and B. Ottersten, “Sensor Array Processing Based on Subspace Fitting,” IEEE
Trans. Signal Processing, Vol. 39, No. 5, May 1991, pp. 1110–1121.
[36] J. R. Flemish, H. W. Kwan, R. L. Haupt, and M. Lanagan, “A New Silicon-Based Photo-
conductive Microwave Switch,” Microwave and Optical Technology Letters, Vol. 51, No. 1,
January 2009.
[37] A. M. Vural, “A Comparative Performance Study of adaptive Array Processors,” presented
at the 1977 IEEE International Conference on Acoustics, Speech, and Signal Processing,
May 9–11, Hartford, CT., paper 20.6.
[38] P. W. Howells, “High Quality Array Beamforming with a Combination of Precision and
Adaptivity,” prepared by Syracuse University Research Corp., SURC TN74-150, June 1974
(Unclassified).
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 19:55 506

506 CHAPTER 12 Recent Developments in Adaptive Arrays

[39] B. D. Van Veen, “Eigenstructure Based Partially Adaptive Array Design,” IEEE Trans. Ant.
& Prop., Vol. AP-36, No. 3, March 1988, pp. 357–362.
[40] C. Lee and J. Lee, “Eigenspace-Based Adaptive Array Beamforming with Robust Capabili-
ties,” IEEE Trans. Ant. & Prop., Vol. AP-45, No. 12, December 1997, pp. 1711–1716.
[41] N. L. Owsley, “Sonar Array Processing,” Array Signal Processing, S. Haykin, ed., Englewood
Cliffs, NJ, Prentice-Hall, 1985.
[42] I. P. Kirstens and D. W. Tufts, “On the Probability Density of Signal-to-Noise Ratio in an
Improved Adaptive Detector,” Proc. ICASSP, Tampa, FL, Vol. 1, 1985, pp. 572–575.
[43] S. W. Ellingson and G.A. Hampson, “A Subspace-Tracking Approach to Interference Nulling
for Phased Array-Based Radio Telescopes,” IEEE Trans. Ant. & Prop., Vol. AP-50, No. 1,
January 2002, pp. 25–30.
[44] B. Yang, “Projection Approximation Subspace Tracking,” IEEE Trans. Signal Processing,
Vol. 43, No. 1, January 1995, pp. 95–107.
[45] E. Brookner and J. Howell, “Adaptive-Adaptive Array Processing,” Proc. IEEE, Vol. 74, No.
4, April 1986, pp. 602–604.
[46] H. Yang and M. A. Ingram, “Subarray Design for Adaptive Interference Cancellation,” IEEE
Trans. Ant. & Prop., Vol. AP-50, No. 10, October 2002. pp. 1453–1459.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:1 507

Appendix A: Frequency
Response Characteristics
of Tapped-Delay Lines

The frequency response characteristics of the tapped-delay line filter shown in Figure A-1
can be developed by first considering the impulse response h(t) of the network. For the
input signal x(t) = δ(t) it follows that


2n+1
h(t) = w i δ[t − (i − 1)] (A.1)
i=1

where w i , i = 1, . . . , 2n + 1 denotes the various complex weights located at the taps in

the tapped-delay line having intertap delay spacing equal to .
Taking the Laplace transform of (A.1) yields


2n+1
{h(t)} = H (s) = w i e−s(i−1) (A.2)
i=1

Equation (A.1) represents a sequence of weighted impulse signals that are summed to
form the output of the tapped-delay line. The adequacy of the tapped-delay line structure
to represent frequency dependent amplitude and phase variations by way of (A.2) depends
on signal bandwidth considerations.
Signal bandwidth considerations are most easily introduced by discussing the con-
tinuous input signal depicted in Figure A-2. With a continuous input signal, the signals
appearing at the tapped-delay line taps (after time t = t0 + 2n has elapsed where t0
denotes an arbitrary starting time) are given by the sequence of samples x(t0 + (i − 1)),
i = 1, 2, . . . , 2n + 1. The sample sequence x(t0 + (i − 1)), i = 1, . . . , 2n + 1, uniquely
characterizes the corresponding continuous waveform from which it was generated pro-
vided that the signal x(t) is band-limited with its highest frequency component f max less
than or equal to one-half the sample frequency corresponding to the time delay, that is,
1
f max ≤ (A.3)
2
Equation (A.3) expresses the condition that must be satisfied in order for a continuous
signal to be uniquely reconstructed from a sequence of discrete samples spaced  seconds
apart, and it is formally recognized as the “sampling theorem” [1]. Since the total (two-
sided) bandwidth of a band-limited signal x(t) is BW = 2 f max , it follows that a tapped-
delay line can uniquely characterize any continuous signal having BW ≤ 1/ (Hz), so
1/ can be regarded as the “signal bandwidth” of the tapped-delay line.
Since the impulse response of the transversal filter consists of a sequence of weighted
impulse functions, it is convenient to adopt the z-transform description for the filter transfer

507
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:1 508

508 Appendix A: Frequency Response Characteristics of Tapped-Delay Lines

FIGURE A-1
Transversal filter Input, x(t)
∇ ∇ ∇
Σ Output, y(t)
having 2n + 1
complex weights. w2n + 1
w1 w2 w3

FIGURE A-2 x(t)

Continuous input Signals appearing at delay–line taps
signal.

∇ t
0 t0 t0 + 2n
∇

function instead of the (more cumbersome) Laplace transform description of (A.2) by

defining

z = es (A.4)

so that

2n+1
{h(t)} = H (z) = w i z −(i−1) (A.5)
i=1

The frequency response of the transversal filter can then be obtained by setting s = jω
and considering how H ( jω) behaves as ω varies. Letting s = jω corresponds to setting
z = e jω , which is a multiple-valued function of ω since it is impossible to distinguish
ω = +π from ω = −π, and consequently
 
2π
H ( jω) = H jω ± k (A.6)

Equation (A.6) expresses the fact that the tapped-delay line transfer function is a
periodic function of frequency having a period equal to the signal bandwidth capability
of the filter. The periodic structure of H (s) is easily seen in the complex s-plane, which
is divided into an infinite number of periodic strips as shown in Figure A-3 [2]. The strip
located between ω = −π/ and ω = π/ is called the “primary strip,” while all other
strips occurring at higher frequencies are designated as “complementary strips.” Whatever
behavior of H ( jω) obtains in the primary strip, this behavior is merely repeated in each
succeeding complementary strip. It is seen from (A.5) that for 2n + 1 taps in the tapped-
delay line, there will be up to 2n roots of the resulting polynomial in z −1 that describes
H (z). It follows that there will be up to 2n zeros in the transfer function corresponding to
the 2n delay elements in the tapped-delay line.
We have seen how the frequency response H ( jω) is periodic with period determined
by the signal bandwidth 1/ and that the number of zeros that can occur across the signal
bandwidth is equal to the number of delay elements in the tapped-delay line. It remains
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:1 509

Appendix A: Frequency Response Characteristics of Tapped-Delay Lines 509

jw FIGURE A-3
Complex s – plane Periodic structure of
H (s) as seen in the
5p
complex s-plane.
Δ
s1 + j 4p 0
Δ Complementary strip
3p
Δ
s1 + j 2p 0
Δ Complementary strip
p
s1 0 Δ
s Primary strip
p
−
Δ
s1 − j 2p 0
Δ Complementary strip
3p
−
Δ
s1 − j 4p 0
Δ Complementary strip
5p
−
Δ

to show that the resolution associated with each of the zeros of H ( jω) is approximately
1/N , where N = number of taps in the tapped-delay line. Consider the impulse response
of a transversal filter having N taps and all weights set equal to unity so that

 −1
1 − e− jωN 
N
H ( jω) = e− jωi = (A.7)
i=0
1 − e− jω

Factoring e− jω(/2)N from the numerator and e− jω(/2) from the denominator of (A.7)
then yields
 
 sin [ω(/2)N ]
H ( jω) = exp jω (1 − N ) (A.8)
2 sin [ω(/2)]
 
  sin [ω(/2)N ]
 [ω(/2)N ]
= N exp jω (1 − N ) (A.9)
2 sin[ω(/2)]
[ω(/2)]

The denominator of (A.8) has its first zero occurring at f = 1/, which is outside the range
of periodicity of H ( jω) for which 12 BW = 1/2. The first zero of the numerator of (A.8),
however, occurs at f = 1/N  so the total frequency range of the principal lobe is just
2/N ; it follows that the 3 dB frequency range of the principal lobe is very nearly 1/N 
so the resolution (in frequency) of any root of H ( jω) may be said to be approximately
the inverse of the total delay in the tapped-delay line. In the event that unequal weighting
is employed in the tapped-delay line, the width of the principal lobe merely broadens, so
the above result gives the best frequency resolution that can be achieved.
The discussion so far has assumed that the frequency range of interest is centered
about f = 0. In most practical systems, the actual signals of interest are transmitted with
a carrier frequency component f 0 as shown in Figure A-4. By mixing the actual transmitted
signal with a reference oscillator having the carrier frequency, the transmitted signal can
be reduced to baseband by removal of the carrier frequency component, thereby centering
the spectrum of the information carrying signal component about f = 0. By writing all
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:1 510

510 Appendix A: Frequency Response Characteristics of Tapped-Delay Lines

FIGURE A-4 BW
Signal bandwidth Spectrum of baseband
BW centered about reduced signal
carrier frequency f0 .

f
0 BW f0 BW
f0 − f0 +
2 2

signals as though they had been reduced to baseband, no loss of generality results since this
merely assumes that the signal spectrum is centered about f = 0; the baseband reduced
information-carrying component of any transmitted signal is referred to as the “complex
envelope,” and complex envelope notation is discussed in Appendix B.

A.1 REFERENCES
[1] R. M. Oliver, J. R. Pierce, and C. E. Shannon, “The Philosophy of Pulse Code Modulation,”
Proc. IRE, Vol. 36, No. 11, November 1948, pp. 1324–1331.
[2] B. C. Kuo, Discrete-Data Control System, Prentice-Hall, Englewood Cliffs, NJ, 1970, Ch. 2.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:2 511

Appendix B: Complex Envelope

Notation

Complex envelope notation provides a convenient means of characterizing the information-

carrying component of modulated carrier signals so the carrier frequency component of
the signal (which does not carry information) does not appear in the signal description.
The properties of complex signal descriptions outlined in this appendix are discussed in
some detail in references [1] and [2].
We are interested in characterizing signals that propagate toward the array elements as
uniform plane waves through a nondispersive transmission medium that only introduces
propagation delay to the transmitted signals. Consequently the output of any array element
can be modeled by time-advanced and time-delayed versions of the signal arriving at the
array phase center with an additive thermal noise component.
For band-limited real signals whose spectrum is located between zero frequency and
twice the array center frequency 2 f c , then the signals present at the array phase center
may be represented as amplitude- and phase-modulated carrier signals of the form

vi (t) = αi (t) cos(ωc t + φi (t) + i ) (B.1)

where the subscript i denotes the signal from the ith source. There is an analytic signal
associated with each real signal that is defined by

z i (t) = [vi (t) + j v̌i (t)] (B.2)

where v̌i is the Hilbert transform [B.3] of vi (t), that is,

 ∞
 1 vi (τ ) 1
v̌i (t) = dτ = vi (t)∗ (B.3)
π −∞ t − τ πt
where * here denotes the convolution operation. Clearly then, v̌i (t) can be regarded as the
output of a quadrature filter having the input signal vi (t).
As a consequence of (B.3) and (B.1) it follows that

v̌i (t) = αi (t) sin(ωc t + φi (t) + i ) (B.4)

The baseband-reduced complex envelope of vi (t), denoted by ṽi (t), is then defined in
terms of the analytic signal by

ṽi (t) = z i (t) exp(− jωc t) = αi (t)exp{ j[φi (t) + i ]} (B.5)

Consequently

vi (t) = Re{ṽi (t)exp( jωc t)} (B.6)

511
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:2 512

512 Appendix B: Complex Envelope Notation

gives the relationship for real (in-phase) signal components in terms of their corresponding
complex envelopes. Likewise
v̌i (t) = Im{ṽi (t) exp( jωc t)} (B.7)
gives the relationship for imaginary (quadrature) signal components in terms of their
corresponding complex envelopes. Let the output of the kth array element in response to
several signal sources be denoted by xk (t) so that

xk (t) = vi (t − τik ) + n k (t) (B.8)
     
i
delayed real thermal
signal noise
component component
of ith signal

The complex envelope of xk (t) can then be written as


x̃k (t) = ṽi (t − τik ) exp{− jωc τik } + ñ k (t) (B.9)
i

where τik represents the differential delay of the ith signal to the kth array element relative
to the array phase center, and ṽi (t − τik ) denotes the complex envelope ṽi (t), time delayed
by τik .
The power spectral density relationships that exist among v(t), z(t), and ṽ(t) are
illustrated in Figure B-1. These relationships are readily derived from the power spectral
density relationships between input and output signals for linear filters [3].
It is also instructive to consider two signals v1 (t) and v2 (t) and determine the cross
product relationships that exist between them for the implications this has for covariance
matrix entries. Using the notation of (B.1), we see that it follows
α1 (t)α2 (t)
v1 (t)v2 (t) = {cos[φ1 (t) − φ2 (t) + 1 − 2 ]
2
+ cos[2ωc t + φ1 (t) + φ2 (t) + 1 + 2 ]} (B.10)
α1 (t)α2 (t)
v1 (t)v̌2 (t) = {sin[φ1 (t) − φ2 (t) + 1 − 2 ]
2
+ sin[2ωc t + φ1 (t) + φ2 (t) + 1 + 2 ]} (B.11)
α1 (t) α2 (t)
v̌1 (t)v̌2 (t) = {cos[φ1 (t) − φ2 (t) + 1 − 2 ]
2
− cos[2ωc t + φ1 (t) + φ2 (t) + 1 + 2 ]} (B.12)
∗
ṽ1 (t)ṽ2 (t) = v1 (t)v2 (t) + v̌1 (t)v̌2 (t) + j[v̌1 (t)v2 (t) − v1 (t)v̌2 (t)]
= α1 (t)α2 (t){cos[φ1 (t) − φ2 (t) + 1 − 2 ]
− j sin[φ1 (t) − φ2 (t) + 1 − 2 ]} (B.13)

If the products of the real valued waveforms are now low-pass filtered to remove the double
frequency terms, then

2 v1 (t)v2 (t) = Re{ṽ1 (t)ṽ2∗ (t)} = Re{z 1 (t)z 2∗ (t)} (B.14)

and
2 v1 (t)v̌2 (t) = −Im {ṽ1 (t) ṽ2∗ (t)} = −Im {z 1 (t)z 2∗ (t)} (B.15)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:2 513

B.1 References 513

FIGURE B-1
Spectrum
relationships among
the real signal
representation,
the analytic signal
representation,
and the complex
envelope
representation of
the signal v(t).
(a) Spectrum of the
real signal v(t).
(b) Spectrum of z(t),
the analytic signal
associated with v(t).
(c) Spectrum of ṽ(t),
the complex
envelope
representation
of v(t).

B.1 REFERENCES
[1] K. L. Reinhard, “Adaptive Antenna Arrays for Coded Communication Systems,” M. S. Thesis,
The Ohio State University, 1973.
[2] T. G. Kincaid, “The Complex Representation of Signals,” General Electric Company Report
No. R67EMH5, October 1966, and HMED Technical Publications, Box 1122, Le Moyne Ave.,
Syracuse, NY, 13201.
[3] A. Papoulis, Probability, Random Variables, and Stochastic Processes, McGraw-Hill,
New York, 1965, Ch. 10.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:2 514
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:3 515

Appendix C: Convenient
Formulas for Gradient
Operations

The scalar function trace [ABT ] has all the properties of an inner product, so the formulas
for differentiation of the trace of various matrix products that appear in reference [1] are of
interest for computing the gradients of certain inner products that appear in adaptive array
optimization problems. For many problems, the extension of vector and matrix derivative
operations to matrix functions of vectors and matrices is important; this generalization is
described in reference [2].
∂
tr [X] = I (C.1)
∂X
∂
tr [AX] = AT (C.2)
∂X
∂
tr [AXT ] = A (C.3)
∂X
∂
tr [AXB] = AT BT (C.4)
∂X
∂
tr [AXT B] = BA (C.5)
∂X
∂
tr [AX] = A (C.6)
∂XT
∂
tr [AXT ] = AT (C.7)
∂XT
∂
tr [AXB] = BA (C.8)
∂XT
∂
tr [AXT B] = AT BT (C.9)
∂XT
∂
tr [Xn ] = n(Xn−1 )T (C.10)
∂X
∂
tr [AXBX] = AT XT BT + BT XT AT (C.11)
∂X
∂
tr [AXBXT ] = AT XBT + AXB (C.12)
∂X
∂
tr [XT AX] = AX + AT X = 2AX if A is symmetric (C.13)
∂X

515
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:3 516

516 Appendix C: Convenient Formulas for Gradient Operations

C.1 REFERENCES
[1] M. Athans and F. C. Schweppe, “Gradient Matrices and Matrix Calculations,” M.I.T. Lincoln
Laboratory, Lexington, MA, Technical Note 1965-53, November 1965.
[2] W. J. Vetter, “Derivative Operations on Matrices,” IEEE Trans. Autom. Control, Vol. AC-15,
No. 2, April 1970, pp. 241–244.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:3 517

Appendix D: Useful Matrix

Relations and the Schwartz
Inequality

This appendix summarizes some matrix properties that are especially useful for solving
optimization problems arising in adaptive array processing. Derivations for these relations
as well as a treatment of the basic properties of matrixes may be found in the references
[1]–[3] for this appendix.
The trace of a square matrix is defined to be the sum of its diagonal elements so that


trace(A) = aii (D.1)
i

From the definition (D.1) it is easily verified that

trace(AB) = trace(BA) (D.2)
and
trace(A + B) = trace(A) + trace(B) (D.3)
The length or norm of a vector x is
√  
||x|| = x x = trace(xx ) =
1/2 † † |xi |2 (D.4)
i

The norm of a matrix is in turn defined by

 
†
||A|| = trace(AA ) =
1/2
|ai j |2 (D.5)
i, j

A square matrix is called Hermitian if

A† = A (D.6)
The Hermitian matrix A is said to be positive definite if the Hermitian form satisfies

x† Ax = ai j xi∗ x j > 0 for all x = 0 (D.7)
i, j
†
If, however, x Ax ≥ 0 for all x = 0, then the Hermitian form is merely positive semidefinite.
For any arbitrary matrix D the product D† D is Hermitian and positive semidefinite. If
in addition D is square and nonsingular (so that D−1 exists) then D† D is positive definite.
Any positive definite Hermitian matrix A may be factored into the form
A = DD† (D.8)
In certain optimization problems involving the trace of a matrix, the following formula
for completing the square is very useful:
x† Ax − x† y − y† x = −y† A−1 y + [x − A−1 y]† A[x − A−1 y] (D.9)
517
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:3 518

518 Appendix D: Useful Matrix Relations and the Schwartz Inequality

The following matrix inversion identities are very useful in many of the derivations.
First,

[P−1 + M† Q−1 M]−1 ≡ P − PM† [MPM† + Q]−1 MP (D.10)

is easily shown by direct multiplication. The second identity is

[P−1 + M† Q−1 M]−1 M† Q−1 ≡ PM† [MPM† + Q]−1 (D.11)

The third identity is

Q−1 − [MPM† + Q]−1 ≡ Q−1 M[P−1 + M† Q−1 M]−1 M† Q−1 (D.12)

The Schwartz inequality for two scalar functions f (t) and g(t) is as follows:
 2  
 
 f ∗ (t)g(t)dt  ≤ | f (t)|2
dt | g(t)|2 dt (D.13)
 

Equality in (D.13) obtains if and only if f (t) is a scalar multiple of g(t).

For two vector functions f(t) and g(t), the inequality may be written in the form
 2  
 
 f † (t)g(t) dt  ≤ f †
(t)f(t) dt g† (t)g(t) dt (D.14)
 

Equality in (D.14) obtains if and only if f(t) is a scalar multiple of g(t). Note that (D.14)
is analogous to

|x† y|2 ≤ (x† x) (y† y) (D.15)

where x and y represent two vectors.

Equation (D.14) may also be generalized to matrix functions by introducing the trace
of a matrix [defined in (D.1)]. Let the complex matrixes F(t) and G(t) be m × n matrixes
so that F† (t)G(t) is square and n × n. Then the Schwartz inequality can be written as
 2  
 
 trace [F† (t)G(t)] dt  ≤ trace [F† (t)F(t)] dt trace [G† (t)G(t)] dt (D.16)
 

Equality in (D.16) obtains if and only if F(t) is a scalar multiple of G(t). Note that (D.16)
is analogous to

|trace (A† B)|2 ≤ trace (A† A)trace (B† B) (D.17)

Yet another useful generalization of the Schwartz inequality for matrixes is

trace (B† A† AB) ≤ trace (A† A)trace (B† B) (D.18)

D.1 REFERENCES
[1] R. Bellman, Introduction to Matrix Analysis, McGraw-Hill, New York, 1960.
[2] E. Bodewig, Matrix Calculus, North-Holland Publishing Co., Amsterdam, 1959.
[3] K. Hoffman and R. Kunze, Linear Algebra, Prentice-Hall, Englewood Cliffs, NJ, 1961.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 519

Appendix E: Multivariate
Gaussian Distributions

Useful properties that apply to real Gaussian random vectors are briefly discussed and
extensions to complex Gaussian random vectors are presented in this appendix. Detailed
discussions of these properties will be found in the references [1]–[4] for this appendix.

E.1 REAL GAUSSIAN RANDOM VECTORS

Let x denote a real random vector of dimension n × 1 whose components are all scalar
random variables and having a mean value
E{x} = u (E.1)
and an associated covariance matrix

cov(x) = E{(x − u)(x − u)T } = V (E.2)
The covariance matrix V is then symmetric and positive definite (or at least positive
semidefinite). A random vector x is said to be Gaussian if it has a characteristic function
given by
 
 1
C x ( j) = E{exp( jT x)} = exp jT u − T V (E.3)
2
It follows from (E.3) that the probability density function for x has the form
 
1
p(x) = [2π ]−n/2 (det V)−1/2 exp − (x − u)T V−1 (x − u) (E.4)
2
Consider a linear transformation of the random vector x as follows:

y = Ax + b (E.5)

where A and b are nonrandom. Then the vector y is also Gaussian and

E{y} = Au + b (E.6)
cov{y} = AVAT (E.7)

A vector w is referred to as a “normalized Gaussian random vector” if it is Gaussian

with zero mean and its components are all independent having unit variance. In other
words,

E{w} = 0 (E.8)
519
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 520

520 Appendix E: Multivariate Gaussian Distributions

and

E{wwT } = I (E.9)

Since any positive semidefinite matrix V can be factored into the form

V = ZZT (E.10)

it follows that a Gaussian vector x having mean u and covariance matrix V can be obtained
from a normalized Gaussian vector w by means of the following linear transformation:

x = Zw + u (E.11)

Let  
x1
x2
be a random vector x comprised of two real jointly Gaussian random vectors with mean
   
x1 u
E = 1 (E.12)
x2 u2
and
   
x1 V11 V12
cov = (E.13)
x2 V21 V22
Then the conditional distribution for x1 given x2 is also Gaussian with

E {x1 /x2 } = u1 + V12 V−1

22 (x2 − u2 ) (E.14)

and

cov {x1 /x2 } = V11 − V12 V−1

22 V21 (E.15)

If x1 , x2 , x3 , x4 are all jointly Gaussian scalar random variables having zero mean,
then

E {x1 x2 x3 x4 } = E {x1 x2 } E {x3 x4 } + E {x1 x3 } E {x2 x4 } + E {x1 x4 } E {x2 x3 }

(E.16)

The result expressed by (E.16) can be generalized to jointly Gaussian random vectors (all
having zero means). Let the covariance matrix E{wxT } be denoted by

E {wxT } = Vw x (E.17)

and denoted similarly for the other covariances. The desired generalization can then be
expressed as

E {wT xyT z} = (trace Vw x )(trace V yz ) + trace(Vw y Vzx ) + trace(Vw z V yx ) (E.18)

Quadratic forms are important in detection problems. Let x be a real Gaussian random
vector with zero mean and covariance matrix V. The quadratic form defined by

y = xT KKT x (E.19)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 521

E.2 Complex Gaussian Random Vectors 521

then has the associated characteristic function


C y ( j) = E{e jy } = E{exp[ jxT KKT x]} (E.20)

The expectation in (E.20) can be expressed as the integral

  
−n/2 −1/2 1 T −1
C y ( j) = (2π ) (det V) exp − x V [I − 2 jVKK ]x dx
T
(E.21)
2
Noting that
  
−n/2 1 T −1
(det V) 1/2
= (2π ) exp − x V x dx (E.22)
2
it immediately follows from (E.21) that

C y ( j) = [det(I − 2 jVKKT )]−1/2 (E.23)

The corresponding distribution of y is referred to as “generalized chi-squared.” The first

two moments of y are then given by

E{y} + E{xT KKT x} = trace{E [KT xxT K]} = trace(KT VK) = trace(KKT V) (E.24)

and

E{y 2 } = E{xT KKT xxT KKT x} (E.25)

Applying (E.18) to the above expression yields

E{y 2 } = [trace(KKT V)]2 + 2trace[(KKT V)2 ] (E.26)

Combining the results of (E.24) and (E.26) then yields


var(y) = E {y 2 } − E 2 {y} = 2trace[(KKT V)2 ] (E.27)

E.2 COMPLEX GAUSSIAN RANDOM VECTORS

A complex Gaussian multivariate distribution can be defined in such a way that many of
the useful properties of real Gaussian random vectors also hold in the complex case. To
develop some of the properties of complex Gaussian random vectors, first consider two
real random n-dimensional vectors x and y having zero mean and, in addition, having the
special properties

V
cov(x) = cov(y) = (E.28)
2
and
wi j
cov(xi y j ) = −cov(yi x j ) = − (E.29)
2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 522

522 Appendix E: Multivariate Gaussian Distributions

As a consequence of (E.28) and (E.29), the joint covariance matrix has the special form
    
x 1 V −W
E [xT yT ] = (E.30)
y 2 W V

where WT = −W. A one-to-one relationship (known as an isomorphism) exists between

multiplication of real matrixes of the form
 
V −W
(E.31)
W V
and multiplication of complex matrixes of the form

C = V + jW (E.32)

For example, if one defines the complex vector

z = x + jy (E.33)

then it is easy to verify that

   
V −W x
[xT yT ] = z† Cz (E.34)
W V y
and
 −1  
V −W x
T T
[x y ] = z† C−1 z (E.35)
W V y
A complex random vector z as defined in (E.33) will be defined to be a complex
Gaussian random vector if its real part x and imaginary part y are jointly Gaussian and the
joint covariance matrix has the special form given by (E.30).
When z has zero mean, its covariance matrix is given by

E {zz† } = E {(x + jy) (xT − jyT )} = V + jW = C (E.36)

If z 1 , z 2 , z 3 , z 4 are all complex jointly Gaussian scalar random variables having zero mean,
then

E {z 1 z 2∗ z 3∗ z 4 } = E {z 1 z 2∗ } E {z 3∗ z 4 } + E {z 1 z 3∗ } E {z 2∗ z 4 } (E.37)

The probability density function for a complex Gaussian random vector z having mean u
and covariance matrix C can be written as

p(z) = π −n (det C)−1 exp{−(z − u)† C−1 (z − u)} (E.38)

Any linear transformation of a complex Gaussian random vector also yields a complex
Gaussian random vector. For example, if

s = Ax + b (E.39)

then s is also complex Gaussian with mean

E{s} = AE{x} + b (E.40)

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 523

E.2 Complex Gaussian Random Vectors 523

and covariance

cov(s) = ACA† (E.41)

If z1 and z2 are two jointly Gaussian complex random vectors with mean values u1
and u2 and joint covariance matrix
   
z1 C11 C12
cov = (E.42)
z2 C21 C22
then the conditional distribution for z1 given z2 is also complex Gaussian with

E {z1 /z2 } = u1 + C12 C−1

22 [z2 − u2 ] (E.43)

and

cov(z1 /z2 ) = C11 − C12 C−1

22 C21 (E.44)

Hermitian forms are the complex analog of quadratic forms for real vectors and play
an important role in many optimization problems. Suppose that z is a complex Gaussian
random vector with zero mean and covariance matrix C. Then the Hermitian form

y = z† KK† z (E.45)

has the following characteristic function:


C y ( j) = E {e jy } = [det(I − jCKK† )]−1 (E.46)

The mean value of y is given by

E{y} = E{z† KK† z} = trace[E{K† zz† K}] = trace(K† CK)

= trace (KK† C) (E.47)

The second moment of y is given by

E{y 2 } = E{z† KK† zz† KK† z} = trace[(KK† C)2 ] + [trace(KK† C)]2 (E.48)

From (E.47) and (E.48) it immediately follows that

var(y) = trace[(KK† C)2 ] (E.49)

It is convenient to introduce a parameter α such that α = 1 for real random vectors

and α = 2 for complex random vectors. Common expressions that are valid for both real
and complex Gaussian random vectors can then be written using the parameter α. For
example, the probability density function of x, the characteristic function of the quadratic
form y, and the variance of the quadratic form can be written respectively as
   
2π −nα/2 1
p(x) = (det V)−α/2 exp − α(x − u)† V−1 (x − u) (E.50)
α 2
−α/2
VKK† 2
C y ( j) = det I − j (E.51)
α
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 524

524 Appendix E: Multivariate Gaussian Distributions

and
 
2
var(y) = trace{(KK† V)2 } (E.52)
α
One result that applies to certain transformations of non-Gaussian complex random
variables is also sometimes useful. Let z = x + jy be a complex random vector that is
not necessarily Gaussian but has the special covariance properties (E.28), (E.29). Now
consider the real scalar,

r = Re{k† z} (E.53)

where Re{·} denotes “real part of” and

k† = aT − jbT (E.54)

It follows that the scalar r can be written as

r = aT x + bT y (E.55)

and

E{r } = aT E{x} + bT E{y} = Re[k† E{z}] (E.56)

The variance of r can also be found as


var(r ) = E{[r − E(r )]2 }
1 T
= [a Va + bT Vb − aT Wb + bT Wa] (E.57)
2
1
= k† Ck
2

E.3 REFERENCES
[1] T. W. Anderson, An Introduction to Multivariate Statistical Analysis, Wiley, New York, 1959.
[2] D. Middleton, An Introduction to Statistical Communication Theory, McGraw-Hill, New York,
1960.
[3] N. R. Goodman, “Statistical Analysis Based on a Certain Multivariate Complex Distribution
(An Introduction),” Ann. Math. Stat., Vol. 34, March 1963, pp. 152–157.
[4] I. S. Reed, “On a Moment Theorem for Complex Gaussian Processes,” IEEE Trans. Info.
Theory, Vol. IT-8, April 1962.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 525

Appendix F: Geometric
Aspects of Complex Vector
Relationships

For purposes of interpreting array gain expressions, which initially appear to be quite
complicated, it is very convenient to introduce certain geometric aspects of various com-
plex vector relationships and thereby obtain quite simple explanations of the results. The
discussion in this appendix is based on the development given by Cox [1].
The most useful concept to be explored is that of a generalized angle between two
vectors. Let the matrix C be positive definite Hermitian, and let a and b represent two
N -component complex vectors. A generalized inner product between a and b may be
defined in the complex vector space H (C) as

inner product (a, b) = a† Cb (F.1)

In this complex vector space the cosine-squared function of the generalized angle can be
defined as

 |a† Cb|2
cos2 (a, b; C) = (F.2)
{(a† Ca) (b† Cb)}

Applying the Schwartz inequality to (F.2) yields the bounds

0 ≤ cos2 (a, b; C) ≤ 1 (F.3)

In the complex vector space H (C), the length of the vector b is (b† Cb)1/2 . The vectors
a and b are orthogonal in H (C) when cos2 (a, b; C) = 0. Likewise the vectors a and
b are in exact alignment (in the sense that one is a scalar multiple of the other) when
cos2 (a, b; C) = 1. Furthermore the sine-squared function can be naturally defined using
the trigonometric identity

sin2 (·) = 1 − cos2 (·) (F.4)

Cases that are of special interest for adaptive array applications are C = I, the identity
matrix, and C = Q−1 (where the matrix Q represents the normalized noise cross-spectral
matrix). When C = I it is convenient to simply write cos2 (a, b) in place of cos2 (a, b; I ).
To obtain a clearer understanding of the effect of the matrix Q−1 on the generalized
angle between a and b, it is helpful to further compare cos2 (a, b) and cos2 (a, b; Q−1 ).
To obtain such a comparison involves a consideration of the eigenvalues and correspond-
ing orthonormal eigenvectors of Q denoted by {λ1 , . . . , λ N } and {e1 , . . . , e N }, respec-
tively. Since Q is normalized (and its trace is consequently equal to N ), i λi = N . The

525
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:4 526

526 Appendix F: Geometric Aspects of Complex Vector Relationships

two vectors a and b can be represented in terms of their projections on the eigenvectors
of Q as:

N
a= h i ei where h i = ei† a (F.5)
i=1

and

N
b= gi ei where gi = ei† b (F.6)
i=1

Substituting (F.5) and (F.6) into (F.2) and using the orthonormal properties of the ei vectors
then yields
 2
 N 
 i=1 h i∗ gi 
cos2 (a, b) =    (F.7)
N
i=1 |h i | 2 N
i=1 |gi | 2
 2
 N 
 i=1 h i∗ gi /λi 
−1
cos (a, b; Q ) = 
2   (F.8)
N
i=1 |h i | 2 /λ
i
N
i=1 |g i | 2 /λ
i

It will be noted from equations (F.7) and (F.8) that the only effect of the matrix Q−1
is to scale each of the factors h i and gi by the quantity (1/λi )1/2 . This scaling weights
more heavily those components of a and b corresponding to small eigenvalues of Q and
weights more lightly those components corresponding to large eigenvalues of Q. Since
the small eigenvalues of Q correspond to components having less noise, it is natural that
these components would be emphasized in an optimization problem.

F.1 REFERENCE
[1] H. Cox, “Resolving Power and Sensitivity to Mismatch of Optimum Array Processors,”
J. Acoust. Soc. Am., Vol. 54, No. 3, September 1973, pp. 771–785.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 527

Appendix G: Eigenvalues
and Eigenvectors

Let A be a N × N hermitian matrix. Then the eigenvectors of matrix A, denoted by v,

satisfy:

Av = λv or [λI − A]v = 0 (G.1)

The scalar λ is called an eigenvalue of A.

In order to have a solution, we must have

det[λI − A] = 0 (G.2)

The characteristic polynomial of A is f (λ) = det[λI − A], and the roots of f (λ) are
the eigenvalues of A:

N 
N
f (λ) = (λi − λ) = f k λk (G.3)
i=1 k=0

There is an eigenvector associated with each λ. For distinct eigenvalues, the eigenvectors
are orthonormal vectors. For eigenvalues of multiplicity M, we can construct orthonormal
eigenvectors. The eigenvalues of a Hermitian matrix are all real. Furthermore,

λmax = λ1 ≥ λ2 ≥ . . . ≥ λ N = λmin (G.4)

The eigenvectors associated with the larger eigenvalues are called principal
eigenvectors.
There are several important properties of eigenvalues and eigenvectors as follows:
1. the highest coefficient in (G.3) is

f N = (−1) N (G.5)

2. the lowest coefficient in (G.3) is


N
f 0 = det(A) = λi (G.6)
i=1

3. the sum of the eigenvalues equals tr(A),


N
tr(A) = λi (G.7)
i=1

4. if k is a scalar, then the eigenvalues of kA are kλi .

527
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 528

528 Appendix G: Eigenvalues and Eigenvectors

5. the eigenvalues of Am are λi m , m a positive integer.

6. if A is non-singular, then the eigenvalues of A−1 are λi −1 and the eigenvectors are the
eigenvectors of A.
7. the identity matrix I has N eigenvalues equal to unity and any set of N orthonormal
vectors can be used as the eigenvectors.
8. if |λmax (A)| <1, then

(I + A)−1 = I − A + A2 − A3 + · · · (G.8)

9. the eigenvalues of A + σ 2 I are λi + σ 2 and the eigenvectors are the eigenvectors

of A.
10. If we define the N × N matrix V A = [v1 v2 . . . v N ] where the vi are orthonormal
eigenvectors and the diagonal matrix  = diag[λ1 λ2 . . . λ N ], then

A = V A V HA (G.9)
and
A−1 = V A −1 V A if A is nonsingular (G.10)

In addition,

N 
N
1
A= λi vi viH and A−1 = vi viH (G.11)
i=1 i−1
λi

11. If A is positive definite, then all λi are positive.

12. If A is positive semidefinite, the number of positive λi equals rank(A), and the
remaining λi are zero.

D
13. If A consists of a weighted sum of D outer products, A = σi2 u i u iH , then rank(A) =
i=1
D and the first D eigenvectors are linear combinations of the u i and the eigenvalues
are denoted by λi . There are N − D eigenvalues equal to zero and the corresponding
eigenvectors are any orthonormal set that are also orthogonal to the u i , I = 1, . . . , D.
D
14. If A is an N × N matrix that can be written as A = σ i2 u i u iH + σw2 I, (corresponding
i=1
to interference and noise components, respectively), then the eigenvectors are still
given by the formulation in (13) and

λi + σ 2w , i = 1, . . . , D
λi (A) = (G.12)
σ 2w , i = D + 1, . . . , N
The space for A can then be subdivided into a signal subspace and a noise subspace:

A = Vs s Vs H + V N  N V N H (G.13)
where s = diag[λ1 + σ w 2 , . . . , λ D + σ w 2 ]
 N = σ w 2 I(N −D)
Vs = [v1 , v2 , . . . , v D ]
V N = [v D+1 , v D+2 , . . . , v N ]
and viH v j = 0, i = D + 1, . . . , N , j = 1, . . . , D.
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 529

Appendix H: Selected Answers

Chapter 1
1. a. 15 MHz
b. 230 MHz
c. 750 MHz
  1/2.4166
2. f = CC12 0.909 × 10−6

Chapter 2
9. They are the same: D = 10.
11.
p/2
1 0
array factor in dB

−20
0.8
p 0 −40
0.6
−60

0.4 −80
0 2 4 6 −1 0 1
Element −p/2 u

15.
0

−5

−10

−15
Thinned
−20
AF (dB)

Desired

−25

−30

−35

−40

−45
−1 −0.5 0 0.5 1
u

529
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 530

530 Appendix H: Selected Answers

17.
0
−10
−20 Adapted
−30 Cancellation Quiescent

AF (dB)
−40
−50
−60
−70
−80
−90
−90 −45 0 45 90
q (degrees)

19.
0 0
−10 −10
Adapted
−20 −20 Adapted
−30 Quiescent −30 Quiescent
Cancellation
AF (dB)

AF (dB)
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80 cancellation
−90 −90
−50 0 50 −50 0 50
q (degrees) q (degrees)

Chapter 3
7.
10
1
5
−21° 61°
Directivity (dB)

0 0.8
−5 0.6
|wn|

−10
0.4
−15
−20 0.2

−25 0
−50 0 50 0 5 10 15 20
q (degrees) Iteration

80
60
40
∠wn (degrees)

20
0
−20
−40
−60
−80
0 5 10 15 20
Iteration
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 531

Appendix H: Selected Answers 531

Chapter 4
2. a. We know that
vb = vb − w 1 va
where
va = S cos θ + n cos α
and
vb = S cos(β − θ ) + n cos(β − α)
In other words, vb = S[cos(β − α) − w 1 cos θ ] + n[cos(β − α) − w 1 n cos α]
Setting the bracketed quantity associated with S equal to zero, then yields

w 1 cos θ = cos(β − θ )

which yields the desired result.

Chapter 5
1. Applying the Matrix Inversion Lemma, we have

[Rx x (k − 1) + x(k)x ↑ (k)]−1

= Rx−1 −1(k−1)
x (k − 1) − R x x x(k)[I + x ↑ (k)Rx−1 −1 ↑ −1
x (k − 1)x(k)] x (k)R x x (k − 1)

Noting that the quantity inside the square brackets on the right hand side is just a scalar,
it can immediately be placed in the denominator.
Now post-multiply everything by b∗ and noting that Rx−1 ∗
x (k − 1)b = w(k − 1) and
x ↑ (k)Rx−1 ∗ ∗
x (k − 1)b = −ε (k) yields the desired result by merely placing P(k − 1) =
−1
Rx x (k − 1).

Chapter 6
1 t
1. a. Perf Index = w Rnn w + λ[c – It1 w]
2
Taking the derivative with respect to w then yields

Rnn w + λ[−I1 ] = 0

so that wmv = R−1

nn I1 λ
3. Following the same development that is given in (6.111) and (6.112)
It immediately follows that wopt = R−1
x x I1 λ
Solving for λ then yields wopt = R−1
x x I1 [It1 R−1
x x I1 ]
−1
c

Chapter 7
1. a. For the standard CSLC network (Figure 7-10), it follows that the output is given by
Ouput = b − w 1 x1 − w 2 x2
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 532

532 Appendix H: Selected Answers

For the configuration of Figure 7-21, it follows that

Output = b − u 1 x1 − (x1 − u 2 x2 )u 3

Carrying out the indicated multiplication and collecting like terms then yields

Output = b − (u 1 + u 3 )x1 − (−u 2 u 3 )x2

Comparing the outputs for Figure 7-10 and Figure 7-21 then yields the desired
result.
2. a. From Figure 7-11 we find that
y2 = x2 − u 11 x1
v32 = b − u 12 x1
z = v32 − u 22 y2

Use the relationship (7.50) where now v11 = x1 , v21 = x2 , v31 = b, v22 = y2 to
directly obtain (7.61), (7.62) and (7.65).

Chapter 8
3. a. Consider F(x + x)2 where (x + x)2 = (x − sx/ρ)2 = x 2 − 2sx 2 /ρ + s 2 x 2 /ρ 2
but since F(x 2 ) = ρ 2 , we are left with ρ 2 − 2sρ 2 /ρ + s 2 = ρ 2 − 2ρs + s 2
F(x)
b. SL(x) ≡  
F(x)−F(x+x)
N
If each search over n + 1 function evaluations yields a function improvement of
(2ρs − s 2 )
2ρs − s 2 , then the function improvement per function evaluation is just
(n + 1)
ρ2
so that SL(x) per function evaluation becomes
2ρs − s 2
(n + 1)
(n + 1)ρ 2
Hence SL(x) =
(2ρs − s 2 )
c. We are interested in showing that ρ02 − ρ12 = 2sρ cos(φ) − s 2

From the relationship for any plane triangle, a 2 = b2 + c2 − 2bc cos(φ)

Put c = s, b = ρ thereby yielding 2ρs cos(φ) − s 2 = ρ02 − ρ12 = ρ 2
d.
 π/2
2ρ 2 sinn−2 (φ)dφ
F(x)
S L(x) = =  φ0 0
F(x) − F(x + x)
ρ 2 sinn−2 (φ)dφ
N 0

but ρ 2 = 2ρs cos(φ) − s 2 thereby yielding the desired expression

Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 533

Appendix H: Selected Answers 533

5.
0 dB 30 dB
30 40

Received signal (dB)

Array factors (dB) 20 30 Total power
Adapted
10 Quiescent 20 Desired signal
Jammer 2
0 10

−10 0 Jammer 1

−20 −10
0 50 100 150 0 10 20 30 40 50
f (degrees) Generation

15
Signal to jammer ratio (dB)

10
5
0
5
−10
−15
0 10 20 30 40 50
Generation

Chapter 10
1. a. Using |H1 (ω)|{exp j (α1 (ω) + exp j[α2 (ω) − πω ω0
sin θi ]} = 0 implies that the two
angles above must be 180 degrees out of phase with each other. Therefore we may
write
πω
α2 (ω) − sin θi = α1 (ω) ± nπ, n odd
ω0
or
πω
α2 (ω) − α1 (ω) = sin θi ± nπ, n odd
ω0
b. Using |H1 (ω)|{exp jα1 (ω) + exp j[α2 (ω) − πω ω0
sin θs ]} = 1
Since |H1 (ω)|2 = H1 (ω)H∗1 (ω) = 1, it immediately follows that | H1 (ω)|2 AA∗ =
1 where A is just the bracketed expression above. Carrying out the multiplication
πω
of the complex conjugates and recognizing that α2 (ω) − α1 (ω) = sin θi ± nπ
ω0
− jx
e +e
jx
and = cos(x) then yields
2
1
|H1 (ω)|2 =  
πω
2 − 2 cos (sin θi − sin θ S )
ω0
from which the desired result follows. Parts (c) and (d) may be shown by using
the considerations given above.
3. A signal and its Hilbert Transform are orthogonal to each other. i.e., E{x(t)x̌(t)} = 0.
This is most easily approached from the point of view of Fourier Transforms.
∞ ∞
∗
x(t)y (t)dt = X ( f )Y ∗ ( f )d f
−∞ −∞
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 534

534 Appendix H: Selected Answers

Hence,
∞ ∞ ∞
∗ ∗
x(t)x̌ (t)dt = X ( f ) X̌ ( f )dt = X ( f )[ jsgn( f )X ∗ ( f )]dt
−∞ −∞ −∞
∞
= j sgn( f )|X ( f )|2 d f
−∞

but sgn(f) is an odd function of f , while |X ( f )|2 is an even function, so the integrand
is an odd function and the integral is just zero!
Once again we have
∞
E{x̌(t) y̌(s)} = X ( f ) j sgn( f ) Y ∗ ( f )[− j sgn( f )]d f
=∞
∞
= X ( f )Y ∗ ( f )[+sgn2 ( f )]d f
−∞

but sgn2 ( f ) is just 1, so we are left with E{x(t)y(s)}.

The other relationships in Problem 3 can be handled in an analogous manner.
11.
20
Edge element failed
10 Center element failed
AF (dB)

0 No failures

−10

−20
0 0.2 0.4 0.6 0.8 1
u

Chapter 11
1. a. b.
0 0
q1 = 10°

q1 = −30° −10
q1 = 20°
P (dB)

P (dB)

−10

−20

−20 −30
−90 −45 0 45 90 −90 −60 −30 0 30 60 90
q (degrees) q (degrees)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 535

Appendix H: Selected Answers 535

c. d.
0 0

−10 −10
P (dB)

P (dB)
−20 −20

−30 −30
−90 −60 −30 0 30 60 90 −90 −60 −30 0 30 60 90
q (degrees) q (degrees)

e.
p/2

p 0

−p/2

2. ε = x2 − x̂2 = x2 − a(2, 1)x1 − a(2, 2)x0

Now require that x̄0 ε̄ = 0 and x̄1 ε̄ = 0
Multiplying x0 times ε and x1 times ε then yields the pair of equations:

x0 ε = x0 x2 − a(2, 1)x0 x1 − a(2, 2)x02 = 0

x1 ε = x1 x2 − a(2, 1)x12 − a(2, 2)x1 x0 = 0

Recognizing that x̄0 x̄2 = r (2), x̄0 x̄1 = r (1), and x̄02 = x̄12 = r (0), the above pair of
equations can immediately be put in the desired matrix form.

Chapter 12
3. a. b.
20 x 20 x

0 106.25 0 106.25
Gain (dB)

Gain (dB)

−20 −20

−40 −40

−60 −60
0 45 90 135 180 0 45 90 135 180
f (degrees) f (degrees)
(a) (b)
Monzingo-7200014 book ISBN : XXXXXXXXXX November 24, 2010 20:6 536

536 Appendix H: Selected Answers

c. d.
20 x 20 x

0 106.25 0 106.25
Gain (dB)

Gain (dB)
−20 −20

−40 −40

−60 −60
0 45 90 135 180 0 45 90 135 180
f (degrees) f (degrees)
(c) (d)
Monzingo-7200014 monz7200014˙Ind ISBN : XXXXXXXXXX November 25, 2010 21:32 537

Index

A uniformly spaced, 70, 71 Cancellation ratio, see Performance

Accelerated gradient (AG), 209–212, 229 volumetric, 17 measure, cancellation ratio
Accelerated random search (ARS), Array errors, 374–376 Canonical structure, 82
341–344 Array polynomial, 48, 52, 56, 182, Capon’s method, 423, 424, 437
Acoustic communications (ACOMMS) 424–425 Channel capacity, 476–477
receive array, 17, 18 Attenuator, 351, 352, 375, 376, 378, Channel matrix, 474, 476, 478
Active electronically scanned arrays 481–483 Channel mismatch, see Inter-channel
(AESA), 12 Autocorrelation function, 428 mismatch
Adaptive DCPA algorithm (ADPCA), 468 extrapolation procedure, 430 Chapman, D. J., 192, 494
Adaptive subarray beamforming, 494–496 theorem, 427 Chebyshev amplitude taper, 47–49, 55,
Adaptive weight errors, 135 Autocorrelation matrix, 84–85, 88, 56, 349, 482–483
Amplitude mismatch, 410–413 276–277, 287 Cholesky factorization, 256, 263, 281
AMTI radar system, 63, 279, 280 Axial ratio, 125 Clutter, 3, 6, 30, 61, 148, 214, 249, 280,
Analytic signal, 83, 511 295, 296, 306, 307, 327, 465–466,
B 467, 468, 469–470, 472
AN/APG-68, 12, 13
AN/APG-77, 12, 14 Baird, C. A., 274, 341 Coherent sidelobe canceller (CSLC), 23,
Angle of arrival (AOA) estimation, Bandwidth cutoff, 65 96, 134, 246, 265, 315, 318–319,
421, 456 Bandwidth effects, 56–58, 65, 238–241, 397–398, 399, 405–407
Applebaum, S. P., 6, 171, 192 429–451 for radar applications, 305
Array calibration, 377–380 Barron, R. L., 344 Complex correlation coefficient of spatial
Array, circular, 15, 46, 65 Bayes liklihood ratio, 104 vectors, 498
conformal, 13, 17 optimal array processor, 437–441 Complex envelope, 8, 82, 314, 398, 399,
conventional, 88, 89, 90 Bayes optimal array processor, 437–441 511–513
cylindrical, 17, 18 Beamforming, 6, 20, 65, 75, 120 Complex matrix inversion using real
directional sensitivity pattern, see Beamforming network, 6–7, 8, 9, 30 arithmetic, 266–267
Directional pattern Beam scanning, 4, 13 Complex random vector, 105
elliptical, 46 sensitivity pattern, 60, 303 Complex Wishart density function, 247
filled, 68 space processor, 64 Compton, R. T., 179, 199, 380
gain, 54, 119–120, 134, 137 space subarray, 495 Conformal array, 13, 33, 44–45
imaging, 5 steering, 5, 17, 18 Conjugate gradient descent (CGD), 303
lens, 12, 14 steering signal, 172, 177, 240 Constrained weight vector solution,
linear, 13, 17, 37–42, 65 steering vector, 229, 240, 244, 263, 214–218
nonuniformly spaced, 70, 71 280, 495 Constraints, derivative, 139, 198
optimization techniques for, 51 Beam switching, 463–465 linear, 137–138
planar, 13, 17, 18, 42–44 Bishop, T. N., 429 main beam, 191–198
propagation delay, 60, 61, 63, 73, 373, Bode plot, 232, 233 multiple linear, 139
405, 416 Brennan, L. E., 304, 468 multiple point, 139
random, 51–55, 65 Broadband processing, 62–66, 68, power, 138–139
rectangular, 42–44 103–121, 380–395. see also quadratic, 138
reflector, 11, 17 Signal processing, broadband single point, 139
retrodirective, 5 Burg, J. P., 426, 439 unitary, 306, 307
self-phasing, 5 Butler matrix, 196, 464–465 Constraint subspace, 222, 223
signal-aligned, 88–90, 100, 102, Control loop spatial filters, see Spatial
C filter, control loop
291, 293
spherical, 17, 19 Calibrated source, 377–378 Convergence speed vs. misadjustment
square, 13 Cancellation beam, 54–58, 77, 194, 227, trade-off, 8, 167, 229, 358, 365,
thinned, 50–52, 68 489–492 366, 367

537
Monzingo-7200014 monz7200014˙Ind ISBN : XXXXXXXXXX November 25, 2010 21:32 538

538 Index

Corporate-fed array, 374–375 resolution, 82 Gallop, M. A., 439

Correlation matrix, see Matrix, correlation spread, 224, 227, 229, 242, 262, Gaussian random process, 31, 106–107,
Covariance matrix, 85–86, 95, 96, 98, 99, 296–298, 315, 324–325, 353–358 141, 390, 428
106, 107, 116, 128, 131–132, 144, Eigenvector, 93, 257, 304, 305, 306, Generalized angle, 119
166, 188–189, 224, 229, 239–270, 527–528 correlator, 314, 315
277–284, 287, 295, 307–308, beams, 182–186, 304, 311 likelihood ratio, 104, 120
339–340, 400, 433, 441, 444, 446, component signals, 307, 383 signal-to-noise ratio, see Performance
447, 455, 470–471, 484, 494, 499 Electronic countermeasures (ECM), 4 measure, signal-to-noise ratio
Coverage Improvement Factor (CIF), counter-countermeasures (ECCM), 5 Generalized signal-to-noise ratio (GSNR),
66–68 Element-level adaptivity, 496–499 96, 171
Coverage improvement factor statistic Element pattern, 37, 268, 349, 379, 396 Genetic algorithm (GA), 51–52, 346–353,
(CIFS), 68 Element space processor 362, 367, 492
Cox, H., 105, 485 matched array processor, 81, 123, Geometric aspects of complex vectors,
Cramer-Rao (CR) bound, 444, 449–451 130–131, 143 525–526
Criterion invariant array processor, Element spacing, 33–45, 60, 185, 187, interpretation of constrained
120–121 296, 396, 403, 466, 503 adaptation, 221–224
Crossed dipoles, 125, 127–129 Equivalence of maximum likelihood and Givens rotation, 253–255, 281
Cross-correlation interferometer, 218–221 minimum variance estimates, 299 Gradient, 86, 515
Cross-spectral density matrix, see Matrix, of MEM to least-squares fitting, Gram-Schmidt pre-processor, 315–324
cross-spectral density 454–455 orthogonalization equations, 318
Cylindrical array, 18, 352 Error covariance matrix, see Matrix, error transformation elements, 316
covariance Grating lobes, 33, 38, 68, 376, 491,
D Estimate and plug array processor, 495–496
Data rate, 15, 22 438–439 Griffiths, L. J., 163, 179, 276
Degradation in SNR, 4, 187–188, 392 pseudo-estimate for, 439 Guided accelerated random search
Degrees of freedom, 138, 189–190, 213, Estimation of angle-of-arrival, 456 (GARS), 344–346, 365
214, 242, 280, 358, 359, 405, 416, of random signal, 105–109
495, 496 of unknown, non-random signal, H
Detection index, 105, 113, 114, 116, 123 109–112 Hard limiter, 190–191
of known signal, 114–115 Estimation of signal parameters via Hermitian matrix, see Matrix, Hermitian
of unknown signal, 118 rotational invariance techniques High resolution imaging, 11
of random signal, 115–117 (ESPRIT), 441, 442 Hilbert transform, 83, 385, 417
Diagonal Loading, 251–252, 268 Excess mean squared error, 167, 207 Householder transformation, 255–256
Digital beamformer, 13, 380, 464 Howells-Applebaum MSNR processor,
Differential steepest descent (DSD) F 171–191
algorithm, 151, 201–209 Factored MMSE array processor, hard-limiter modification of,
Directional pattern, 34, 36, 37, 38 102–103 190–191, 229
normalized, 38 Factorization, 102–103, 253–257, iterative weight correction model for,
Direction finding, 6 280–283 322–324
Direction of arrival (DOA) estimation, Factorization Methods, 253–257 servo-loop model of, 175–179
421–459 Fading, 143 Howells, P. W., 171
Direction of incidence errors, 233–234 Fan beam, 13 Householder transformations, 255–256
Directivity, 40–41, 43, 125, 129, 245, Fisher information matrix (FIM), Hydrophone, 15
278, 375 449–450, 451–452
Directivity linear array, 77 Frequency domain equivalent of time I
Directivity planar array, 43–44 delay, 65, 75–77 Incomplete Beta distribution function,
Direct matrix inversion (DMI) algorithm, processor, 65, 75 241, 248
239–251 Friis transmission formula, 39 Integrating filter, 192, 329
Displaced phase center antenna (DPCA), Frost, O. L., 213 Inter-channel mismatch, 61, 63
466 FuG 65, 12 compensation of, 373, 405–415
Distortionless transfer functions, 381–383 Interference nulling, 5, 500
Dyad, 119, 121, 122, 124, 145 G cancellation performance, 399, 503
Dynamic errors, 374, 380 Gabriel, W. F., 171, 190, 304 Isotropic point source, 33–34, 36, 46
Gain, 4, 39–41, 45, 54–56, 58, 66, 68,
E 97, 119–120, 134, 137–138, J
Eigenspace beamformers, 499–501 169–170, 175–176, 193, 201, Jamming, 5, 59, 60, 97, 132, 200, 268
Eigenvalue, 93, 257, 273, 303, 527–528 481, 490–492 bandwidth, effect of, 189
Monzingo-7200014 monz7200014˙Ind ISBN : XXXXXXXXXX November 25, 2010 21:32 539

Index 539

Joint Surveillance and Target Attack Misadjustment, 165, 167–168, 204–209, Partially known autocorrelation function,
Radar System (Joint STARS), 468 221, 340, 343, 357, 359, 365–367 427–430
J/S sky contour plot, 66–67 Monolithic microwave integrated circuit Particle swarm optimization (PSO), 51,
J Subbeam approach, 501–502 (MMIC), 13–14 348–349
Monopulse array, 56, 58, 486 Pattern forming network, 4–8, 21, 23
K Monopulse tracking antenna, 486–487 multiplication, principle of, 37–38
Kalman filter methods, 284–290 Morgan, D. R., 496 Pave Mover, 468
Kmetzo, J. L., 55 Moving target indicator (MTI), 465–466, Pencil beam, 13
468 Performance comparison of four
L Multichannel processor, 63, 64 processors, 388–393
Laplace transform, 507, 508 Multipath, 4, 30 Performance limits, 21, 29, 81–82, 103
Learning curve, 163, 338, 341 compensation of, 63, 397–405 Performance measure, array gain, 9, 21,
Least mean square (LMS) algorithm, 6, correlation matrix of, 400 66–68, 81–82, 130–134, 137–138
158–169 Multiple input, multiple output (MIMO), Bayes likelihood ratio, 104
with constraints, 213–224 6, 473–479, 503 cancellation ratio, 134
convergence condition for, 162–163 Multiple sidelobe canceller (MSLC), see detection index, 105, 113, 114,
Likelihood function, 99 Coherent sidelobe canceller 116, 123
ratio, 112, 115, 117 Multiple signal classification (MUSIC), generalized likelihood ratio,
Linear random search (LRS), 336–341 423–426 104, 118
Look direction, 170–171, 213, 215, Multivariate Gaussian distributions, generalized signal-to-noise ratio,
231–232 519–524 96, 171
Loop noise, 179–182, 321 MUSIC algorithm, 423–426 likelihood ratio, 103
Lo, Y. T., 51 Mutual coupling, 33, 396–397 maximum likelihood, 90, 99
mean squared error, 9, 90, 91, 121,
M N 154, 159
Main beam constraints, 191–198 Narrowband processing, 62–66, 90–103, minimum noise variance, 90,
Mantey, P. E., 276 143, 295, 445–446 100–101, 293
Marple Algorithm, 456–459 Narrowband signal processing, see Signal minimum signal distortion, 22, 120
Matched filtering, 104 processing, narrowband output power, 138, 139, 143, 185, 187,
Matrix, autocorrelation, 84, 85, 88 Natural modes, 156, 157 188, 200, 214
Butler, 196, 464, 465 Near field, 5, 379–380, 486 signal-to-interference ratio, 251,
correlation, 84–86, 87, 88, 219 Near-field scattering, 4, 63, 279, 400 312, 423
covariance, see Covariance matrix Nike AJAX MPA-4, 13 signal-to-noise ratio, 33, 82, 90, 113,
cross-spectral density, 86, 108 Noise subspace, 424, 441, 501 144, 303
error covariance, 287, 295 Nolen, J. C., 304 Performance surface, 154–158, 165,
Hermitian, 88, 94, 139, 276, 427, 433 Nolen network, 304–315 202–205
quadratic form, 135 Nolte, L. W., 439 quadratic, 154–157, 163
relations, 300–301 Non-Gaussian random signal, 107–108 Periodogram, 422–423, 436, 437
Toeplitz, 65, 85, 240, 269, 427, 457 Nonstationary environment, 4, 154, 190, Perturbation, 122, 203–204, 205,
trace of, 87, 203 220, 288 208–209, 337, 341–346
Maximum entropy method (MEM), 82, Nulling tree, 50 Phase conjugacy, 190
426–436 Null placement, 45, 463 Phase lock loop, 5
extension to multichannel case, Null synthesis, 52–58, 489 Phase mismatch, 60, 409, 413–415
433–436 Phase-only adaptive nulling,
relation to maximum likelihood O 227–228, 358
estimates, 453–454 One-shot array processor, 437, 438 Phase shifters, 307, 309–310, 311, 349,
Maximum likelihood estimates, 109, 118, Open-loop scanning, 74 351, 352, 375–376, 377, 378
120, 140–141, 218, 299–300, Orthogonal frequency division Phase shift sequence, 38
443–449, 453–454 multiplexing (OFDM), 477–478 shifters, 13
McCool, J. M., 165, 336 Orthonormal system, 94–95, 175, Piecemeal adjustment procedure,
Mean squared error, see Performance 177, 183 310–311, 315
measure, mean square error Output noise power, 93, 95, 177, 181, Pilot signals 5, 168–171, 192–194
Micro-electro-mechanical systems 187–188 Planar array, 13, 17, 18, 29, 42–44, 51, 66,
(MEMS), 14, 479–480 78, 496
Minimum noise variance (MV), 90, P Poincare sphere, 126, 131, 144
100–101, 102, 103 Partially adaptive array, 488–503 Polarization, 3, 5, 23, 124–130, 479–480
Minimum variance processor, 291–295 subarray beamforming, 494–496 as a signal discriminant, 230–231
Monzingo-7200014 monz7200014˙Ind ISBN : XXXXXXXXXX November 25, 2010 21:32 540

540 Index

Polarization loss factor, 129 updated covariance matrix inverse, non-Gaussian random, 32
Polarization sensitive arrays, 124–130 277–284 non-random unknown, 31
Polar projection diagram, 46 weighted least squares, 273–277 Signal parameters, 7, 30, 31, 74–75, 104
Powell descent method, 209–211 Recursive equations for prediction error Signal processing, broadband, 62–66,
algorithm, 212–213, 229 filter, 429–430 103–121, 380–395
Powell, J. D., 209–210 Recursive least squares (RLS) algorithm, narrowband, 62–66
Power centroid, 240 276, 278, 301 Signal spectral matching, 129
ratio, 232, 242 Recursive matrix inversion, 263, 267 Signal-to-interference ratio, see
Preadaption spatial filter, 194–196 Reed, I. S., 468 Performance measure,
Prediction error filter, 428–435, 454, 456 Reference signal, 31, 91, 145, 168–169, signal-to-interference ratio
matrix equation of, 453, 509 230–231, 246, 250, 288, 386 Signal-to-noise ratio, see Performance
whitening filter representation of, Reflection coefficients, 399, 402, 428, measure, signal-to-noise ratio
428–429 430–431 Singular values, 474, 476
Prewhitening and matching operation, Relative search efficiency, 360–361 Sonar, 15–19, 484–486
109, 112, 115, 118, 120, 121, 131 Resolution, 22, 23, 33, 45, 422 Sonar arrays, 17–20, 484–486
Propagation delay, 60–61, 63, Retrodirective array, 5 Sonar technology, 15–21
73–74, 405 beam, 183 operating frequencies, 11, 15
compensation of, 405 transmit, 5 Sonar transducers, 15–17
effects, 22 Reverberation, 30, 469 Space-time adaptive processing (STAP),
ideal model of, 32–33, 81, 108 RF interference (RFI), 4, 501 465–473
perturbed, 81, 121–124 Rodgers, W. E., 391–395 Spatial filter, control loop, 196–198,
vector, 86–87 root-MUSIC algorithm, 424–425 232
Pugh, E. L., 256 Rotman lens, 464 correction, 192, 229
Pulse repetition frequency (PRF), 466 preadaptation, 194–196
Pulse repetition interval (PRI), 467, 469, S Spectral factorization, 257
471–472 Sample covariance matrix, 240–244 Speed of light, 10
Sample cross-correlation vector, of sound, 10
Q 244–251 Spherical array, 19
Quadratic form, 87, 93, 177, 267 Sample matrix inversion (SMI) see Direct Spread spectrum, 30, 314
Quadratic performance surface, 154, matrix inversion (DMI) algorithm Spread spectrum modulation, 4
155, 157 Scaled conjugate gradient descent systems, 30
Quadrature hybrid weighting, 62–63 (SCGD), 303 Static errors, 374, 377, 380
processing, 62, 383–388, 394 Schwartz inequality, 105, 114, 116, 120, Steady-state response, 8, 9, 21, 22, 29
Quantization errors, 375–376 517–518 requirements, 29
Quiescent noise covariance matrix, 132, SCR-270, 12 Steepest descent, feedback model of,
178, 188–189 Search loss function, 360, 361 156–158
Seismology, 11 method of, 154–156
R Self-phased array, 5 stability condition for, 158
Radar technology, 10–15 Sensitivity of constrained processor to Steepest descent for power minimization,
operating frequencies, 11 perturbations, 60, 209, 229, 227–228
Radiating elements, 11–14 231–232 Step reversal, 361–362
Radio astronomy, 500–501 of convergence rate to eigenvalue Strand, O. N., 433, 452
Random errors, 374–375 spread, 163, 209, 227, 229, 328 Subarray beamforming, 494–496
Random search algorithm, accelerated of steady-state accuracy to eigenvalue beamspace, 495
(ARS), 341–344 spread, 273 simple, 495
guided accelerated (GARS), 344–346 Sensitivity to eigenvalue spread, Subarray Design, 502–503
linear (LRS), 336–341 229, 262 Subspace Fitting, 441–443
Range discrimination, 10 Sidelobes, 4, 36, 47, 48, 51, 55, 71, 97 Sufficient statistic, 104, 112, 115, 116,
Rassweiler, G. G., 341 Signal aligned array, see Array, signal 117, 118, 123
Receiver operating characteristic aligned Super-resolution, 422
(ROC), 104 Signal bandwidth-propation delay Synthetic aperture radar (SAR), 466
Receivers, 14 product, 60, 61
Reconfigurable antennas, 479–483 Signal distortion, 9, 22, 30, 108, T
Rectilinear coordinate diagram, 47 120, 393 Takao, K., 219
Recursive algorithm, Kalman type, Signal injection, 378, 379 Tapped delay line, 63–65
284–290 Signal model, Gaussian random, 31 processing, 22, 373, 383–388
minimum variance, 291–295 known, 31 frequency response of, 405–407
Monzingo-7200014 monz7200014˙Ind ISBN : XXXXXXXXXX November 25, 2010 21:32 541

Index 541

Target range and bearing estimates, Trench, W. F., 240 W

140–141 Triangularization, 253–256 Wavefront, planar, 9, 35, 70, 75, 81
Taylor amplitude taper, 48, 351, Triangular matrix inversion, 267–268 spherical, 69, 70
362, 376 Wavelength, 5, 9, 10, 34, 374, 388
Thinned array, 49–52, 77 U
Weight adjustment, 156, 160, 165–168,
Time constant, 163, 164, 178, 181, 207, Ulrych, T. J., 429 204, 219, 336–337, 355, 358
208, 258–259, 338, 341 Ultrasonics, 11 Weighted least squares, 273–277
Time domain processor, 65, 68, 75 Uniform array, 37, 52, 55, 70, 71, Weight jitter, 179, 182
Toeplitz matrix inversion, 240, 269 422, 491 White, W. D., 303, 311
Towfish, 17 Uniform circular array (UCA), 378 Widrow, D., 5, 6, 90, 158, 165, 336
Trace of a matrix, 87, 203, 442 Wiener filtering, 120, 121, 136
Transducer, 15–17 V
broadband generalization of, 137
Transient response, 8, 29, 33 Van Atta, L.C., 5 solution, 6, 21, , 82, 90, 92, 103, 130,
requirements, 29 Velocity control, 17 154, 159–163, 165, 233, 246, 287
Transmit/receive (T/R) module, 12, 13, of propagation, 10, 33, 58, 69
378 Vertically layered blocking structure Z
Transversal filter, 63, 403–405, 406, 412, (V-BLAST) algorithm, 477 Zahm, C. L., 199
413, 415 Vural, A. M., 496 Z-transform, 41, 156, 158
INTRODUCTION TO ADAPTIVE ARRAYS, 2ND EDITION
Thorough Modernization of THE Seminal Text-Reference on Adaptive Arrays
This second edition is an extensive modernization of the seminal introduction to the subject of adaptive
array sensor systems. With the number of applications of adaptive array sensor systems growing each
year, this look at the principles and fundamental techniques that are critical to these systems is more
important than ever before.

Introduction to Adaptive Arrays, 2nd Edition is organized as a tutorial, taking the readers by the hand
and leading them through the maze of jargon that often surrounds this highly technical subject. It is
easy to read and easy to follow as fundamental concepts are introduced with examples before more
current developments and techniques are introduced.

Problems at the end of each chapter serve both instructors and professional readers by illustrating and
extending the material presented in the text. Both students and practicing engineers will easily gain
familiarity with the modern contribution that adaptive arrays have to offer practical signal reception
systems.

AUDIENCE
Radar • Sonar • Communications • Seismology • Radio Astronomy

ABOUT THE AUTHORS

Robert Monzingo received the B.S.E.E. degree from Stanford University, the M.S.E.E. degree from
the University of Arizona, and the Sc.D. degree from Washington University, St. Louis. He has been
active in the areas of automatic control, radar, and communications for many years and is presently
associated with the Aerospace Corporation in El Segundo, CA as a Senior Engineering Specialist where
he is engaged in work on adaptive arrays, GPS systems, and satellite communication systems.

Randy Haupt is a Fellow of the IEEE and Applied Computational Electromagnetics Society (ACES) and
is a Senior Scientist at the Penn State Applied Research Lab. From 1999-2003 he was Professor and
Department Head of ECE at Utah State University. He also was a Professor of EE at the Air Force Academy
and the University of Nevada-Reno. He is a retired Lt. Col. from the US Air Force. He has many journal
articles, conference publications, and book chapters to his credit and is the author of Antenna Arrays:
A Computational Approach by Wiley (2010).

Thomas Miller currently works for Raytheon Corporation where he has been cited for leadership
in advanced early warning surveillance radar and adaptive signal processing. He is the author and
coauthor of multiple journal articles as well as contributor to several books.