Manikas A. - Differential Geometry in Array Processing-ICP (2004)

DIFFERENTIAL GEOMETRY IN
ARRAY
PROCESSING
This page intentionally left blank
Athanassios Manikas
Imperial College London, UK
DIFFERENTIAL GEOMETRY IN
ARRAY
PROCESSING
Imperial College Press

Published by
Imperial College Press
57 Shelton Street
Covent Garden
London WC2H 9HE
Distributed by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401–402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
DIFFERENTIAL GEOMETRY IN ARRAY PROCESSING

Copyright © 2004 by Imperial College Press
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN 1-86094-422-1
ISBN 1-86094-423-X (pbk)
Typeset by Stallion Press

E-mail: enquiries@stallionpress.com
Printed in Singapore.
34_Differential Geometry.pmd 1 10/4/2005, 6:12 PM

July 6, 2004 9:31 WSPC/Book Trim Size for 9in x 6in fm
To my wife, Eleni
v
Preface
During the past few decades, there has been significant research into
sensor array signal processing, culminating in the development of super-
resolution array processing, which asymptotically exhibits infinite resolu-
tion capabilities.
Array processing has an enormous set of applications and has recently
experienced an explosive interest due to the realization that arrays have a
major role to play in the development of future communication systems,
wireless computing, biomedicine (bio-array processing) and environmental
monitoring.
However, the “heart” of any application is the structure of the employed
array of sensors and this is completely characterized [1] by the array mani-
fold. The array manifold is a fundamental concept and is defined as the locus
of all the response vectors of the array over the feasible set of source/signal
parameters. In view of the nature of the array manifold and its signifi-
cance in the area of array processing and array communications, the role
of differential geometry as the most particularly appropriate analysis tool,
cannot be over-emphasized.
Differential geometry is a branch of mathematics concerned with the
application of differential calculus for the investigation of the properties of
geometric objects (curves, surfaces, etc.) referred to, collectively, as “mani-
folds”. This is a vast subject area with numerous abstract definitions,
theorems, notations and rigorous formal proofs [2,3] and is mainly confined
to the investigation of the geometrical properties of manifolds in three-
dimensional Euclidean space R3 and in real spaces of higher dimension.
However, the array manifolds are embedded not in real, but in
N -dimensional complex space (where N is the number of sensors). There-
fore, by extending the theoretical framework of R3 to complex spaces, the
vii
viii Differential Geometry in Array Processing
underlying and under-pinning objective of this book is to present a sum-

mary of those results of differential geometry which are exploitable and of
practical interest in the study of linear, planar and three-dimensional array
geometries.
Thanassis Manikas — London 2003

a.manikas@imperial.ac.uk
http://skynet.ee.imperial.ac.uk/manikas.html
Acknowledgments
This book is based on a number of publications (presented under a unified

framework) which I had over the past few years with some of my former
research students. These are Dr. J. Dacos, Dr. R. Karimi, Dr. C. Proukakis,
Dr. N. Dowlut, Dr. A. Alexiou, Dr. V. Lefkadites and Dr. A. Sleiman. I wish
to express my pleasure in having had the opportunity in working with them.
I am indebted to my research associates Naveendra and Jason Ng as
well as to my MSc student Vincent Chan for reading, at various stages, the
manuscript and for making constructive suggestions on how to improve the
presentation material.
Finally, I should like to thank Dr. P. Wilkinson who had the patience to
read the final version of the manuscript. His effort is greatly appreciated.
Keywords
linear arrays, non-linear arrays, planar arrays, array manifolds, differential

geometry, array design, array ambiguities, array bounds, resolution and
detection.
Contents
Preface vii
1. Introduction 1
1.1 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Main Abbreviations . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Array of Sensors — Environment . . . . . . . . . . . . . . . 5
1.4 Pictorial Notation . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.1 Spaces/Subspaces . . . . . . . . . . . . . . . . . . . 8
1.4.2 Projection Operator . . . . . . . . . . . . . . . . . . 9
1.5 Principal Symbols . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Modelling the Array Signal-Vector and Array Manifold . . 11
1.7 Significance of Array Manifolds . . . . . . . . . . . . . . . . 17
1.8 An Outline of the Book . . . . . . . . . . . . . . . . . . . . 19
2. Differential Geometry of Array Manifold Curves 22

2.1 Manifold Curve Representation — Basic Concepts . . . . . 22
2.2 Curvatures and Coordinate Vectors in CN . . . . . . . . . . 26
2.2.1 Number of Curvatures and Symmetricity
in Linear Arrays . . . . . . . . . . . . . . . . . . . . 26
2.2.2 “Moving Frame” and Frame Matrix . . . . . . . . . 28
2.2.3 Frame Matrix and Curvatures . . . . . . . . . . . . . 29
2.2.4 Narrow and Wide Sense Orthogonality . . . . . . . . 31
2.3 “Hyperhelical” Manifold Curves . . . . . . . . . . . . . . . 33
2.3.1 Coordinate Vectors and Array Symmetricity . . . . . 37
2.3.2 Evaluating the Curvatures of Uniform Linear Array
Manifolds . . . . . . . . . . . . . . . . . . . . . . . . 38
ix
x Differential Geometry in Array Processing
2.4 The Manifold Length and Number of Windings

(or Half Windings) . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 The Concept of “Inclination” of the Manifold . . . . . . . . 45
2.6 The Manifold-Radii Vector . . . . . . . . . . . . . . . . . . 46
2.7 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.7.1 Proof of Eq. (2.24) . . . . . . . . . . . . . . . . . . . 53
2.7.2 Proof of Theorem 2.1 . . . . . . . . . . . . . . . . . 54
3. Differential Geometry of Array Manifold Surfaces 59

3.1 Manifold Metric . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2 The First Fundamental Form . . . . . . . . . . . . . . . . . 62
3.3 Christoffel Symbol Matrices . . . . . . . . . . . . . . . . . . 64
3.4 Intrinsic Geometry of a Surface . . . . . . . . . . . . . . . . 65
3.4.1 Gaussian Curvature . . . . . . . . . . . . . . . . . . 66
3.4.2 Curves on a Manifold Surface: Geodesic Curvature . 68
3.4.2.1 Arc Length . . . . . . . . . . . . . . . . . . 69
3.4.2.2 The Concept of Geodicity . . . . . . . . . . 69
3.4.3 Geodesic Curvature . . . . . . . . . . . . . . . . . . 69
3.5 The Concept of “Development” . . . . . . . . . . . . . . . . 72
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.7 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.7.1 Proof of Eq. (3.36) — Geodesic Curvature . . . . . . 74
4. Non-Linear Arrays: (θ, φ)-Parametrization of Array

Manifold Surfaces 77
4.1 Manifold Metric and Christoffel Symbols . . . . . . . . . . 78
4.2 3D-grid Arrays of Omnidirectional Sensors . . . . . . . . . 80
4.3 Planar Arrays of Omnidirectional Sensors . . . . . . . . . . 81
4.4 Families of θ- and φ-curves on the Manifold Surface . . . . 83
4.5 “Development” of Non-linear Array Geometries . . . . . . . 87
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.7 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.7.1 Proof that the Gaussian Curvature of an
Omni-directional Sensor Planar Array Manifold is Zero 93
4.7.2 Proof of the Expression of det G for Planar Arrays in
Table 4.2 . . . . . . . . . . . . . . . . . . . . . . . . 94
4.7.3 Proof of “Development” Theorem 4.6 . . . . . . . . 95
Contents xi
5. Non-Linear Arrays: (α, β)-Parametrization 97

5.1 Mapping from the (θ, φ) Parameter Space
to Cone-Angle Parameter Space . . . . . . . . . . . . . . . 97
5.2 Manifold Vector in Terms of a Cone-Angle . . . . . . . . . 100
5.3 Intrinsic Geometry of the Array Manifold Based on
Cone-Angle Parametrization . . . . . . . . . . . . . . . . . 101
5.4 Defining the Families of α- and β-parameter Curves . . . . 104
5.5 Properties of α- and β-parameter Curves . . . . . . . . . . 105
5.5.1 Geodecity . . . . . . . . . . . . . . . . . . . . . . . . 105
5.5.2 Length of Parameter Curves . . . . . . . . . . . . . . 107
5.5.3 Shape of α- and β-curves . . . . . . . . . . . . . . . 107
5.6 “Development” of α- and β-parameter Curves . . . . . . . 110
6. Array Ambiguities 113

6.1 Classification of Ambiguities . . . . . . . . . . . . . . . . . 114
6.2 The Concept of an Ambiguous Generator Set . . . . . . . . 118
6.3 Partitioning the Array Manifold Curve into Segments of
Equal Length . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3.1 Calculation of Ambiguous Generator Sets of Linear
(or ELA) Array Geometries . . . . . . . . . . . . . . 130
6.4 Representative Examples . . . . . . . . . . . . . . . . . . . 132
6.5 Handling Ambiguities in Planar Arrays . . . . . . . . . . . 135
6.5.1 Ambiguities on φ-curves . . . . . . . . . . . . . . . . 136
6.5.2 Ambiguities on α-curves/β-curves . . . . . . . . . . 139
6.5.3 Some Comments on Planar Arrays . . . . . . . . . . 146
6.5.4 Ambiguous Generator Lines . . . . . . . . . . . . . . 149
6.6 Ambiguities and Manifold Length . . . . . . . . . . . . . . 152
6.7 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.7.1 Proof of Theorem 6.1 . . . . . . . . . . . . . . . . . 155
7. More on Ambiguities: Symmetrical Arrays 157

7.1 Symmetric Linear Arrays and det(AN (s)) . . . . . . . . . . 157
7.2 Characteristic Points on the Array Manifold . . . . . . . . 159
7.3 Array Symmetricity and Non-Uniform Partitions
of Hyperhelices . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.4 Ambiguities of Rank-(N − 1) and Array Pattern . . . . . . 167
7.5 Planar Arrays and ‘Non-Uniform’ Ambiguities . . . . . . . 170
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
xii Differential Geometry in Array Processing
8. Array Bounds 174

8.1 Circular Approximation of an Array Manifold . . . . . . . . 174
8.2 Accuracy and the Cramer Rao Lower Bound . . . . . . . . 179
8.2.1 Single Emitter CRB in Terms of
Manifold’s Differential Geometry . . . . . . . . . . . 180
8.2.2 Two Emitter CRB in Terms of Principal Curvature . 184
8.2.2.1 Elevation Dependence of Two Emitters’ CRB 186
8.2.2.2 Azimuth Dependence of Two Emitters’ CRB 188
8.3 “Detection” and “Resolution” Thresholds . . . . . . . . . . 190
8.3.1 Estimating the Detection Threshold . . . . . . . . . 191
8.3.2 Estimating the Resolution Threshold . . . . . . . . . 194
8.4 Modelling of the Uncertainty Sphere . . . . . . . . . . . . . 198
8.5 Thresholds in Terms of (SNR × L) . . . . . . . . . . . . . . 199
8.6 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
8.6.1 Schmidt’s Definition of Resolution . . . . . . . . . . 203
8.6.2 CRB at the Resolution Threshold . . . . . . . . . . 204
8.6.3 Directional Arrays . . . . . . . . . . . . . . . . . . . 205
8.7 Array Capabilities Based on α- and β-curves . . . . . . . . 206
8.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.9 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
8.9.1 Radius of Circular Approximation . . . . . . . . . . 209
8.9.2 “Circular” and “Y” Arrays — Sensor Locations . . . 210
8.9.3 Proof: CRB of Two Sources in Terms of κ1 . . . . . 210
Bibliography 215
Index 217
July 9, 2004 13:51 WSPC/Book Trim Size for 9in x 6in chap01
Chapter 1
Introduction
An array system is a collection of sensors (transducers) which are spatially

distributed at judicious locations in the 3-dimensional real space, with a
common reference point. How the sensors are spatially distributed (array
geometry) is influential not only on the overall array capabilities but also on
its “abnormalities.” The type of the sensors varies with the application and
sensors can take a wide variety of forms. Some common examples of sensors
include electromagnetic devices (such as RF antennas, optical receivers,
etc.) and acoustic transducers (such as hydrophones, geophones, ultrasound
probes, etc.).
The signals at the array elements contain both temporal and spatial
information about the array signal environment which is usually contam-
inated by background and sensor noise. Thus, the main aim of array pro-
cessing is to extract and then exploit this spatio-temporal information to
the fullest extent possible in order to provide estimates of the parameters of
interest of the array signal environment. Depending on the application, typ-
ical parameters of interest associated with emitting sources (i.e. signals that
use the same frequency and/or time-slot and/or code) can be the number
of incident signals, Directions-of-Arrival (DOAs), Times-of-Arrival (TOAs),
ranges, velocities etc. Indeed, with an array system operating in the pres-
ence of a number of emitting sources, and by observing the received array
signal-vectors x(t), the following four general problems are of great interest:
(1) Detection problem — concerned with the determination/estimation of

the number of incident signals. This problem is essentially the spa-
tial analogue of model order selection in time-series analysis. Thus,
the most popular methods for the solution of this problem are based
on “Akaike Information Criterion” (AIC) [4] and the “Minimum
Description Length” (MDL) criterion [5, 6]. Both methods involve the
1
2 Differential Geometry in Array Processing
minimization of a function of the noise eigenvalues of the array output

covariance matrix.
(2) Parameter estimation problem — where various signal and channel
parameters are estimated. One important problem of this type is the
“Direction Finding” (DF). In this case the parameters of interest are
the bearings of emitters/targets (e.g. [7]). This problem is essentially
the spatial analogue of the frequency estimation problem in time-series
analysis.
(3) Interference cancellation (or reception problem) — the acquisition of
one (desired) signal from a particular direction and the cancellation of
unwanted co-channel interfering signals (or jammers), from all other
directions. When the desired signal and the interference occupy the
same frequency band, temporal filtering is inappropriate. However, the
spatial separation of the sources can be exploited using an array of sen-
sors (e.g. an antenna array). This operation falls, in array processing
terms, under the general heading of “beamforming” while, in commu-
nication systems terms, a beamformer is a “linear receiver” (e.g. [8]).
(4) Imaging — here the parameters of interest are the shapes and sizes
of various objects in the environment. These are typically determined
by the generation of two- or three-dimensional maps depicting some
feature of the received signals (e.g. intensity) as a function of their
spatial coordinates (e.g. [9]).
The four types of problem described above are inter-related and the
solution to one problem may result in a partial or complete solution to
another. For example, the successful operation of all parametric parameter
estimation algorithms requires solving firstly the detection problem (i.e.
a priori knowledge of the number of emitters present). Furthermore, once
the number and directions-of-arrival (DOAs) of signals received at the array
site are estimated by solving the detection and direction-finding problems,
nulls may be readily placed along the directions of the unwanted signals,
hence achieving interference cancellation.
The applications of arrays in various scientific disciplines (such as the
ones already mentioned) are extensive and suffice to reveal the multi-
dimensional significance of the array concept. For instance, although array
processing has been extensively used in high frequency communications in
the past, the explosive growth in demand for cellular services in recent years
has placed it at the centre of interest. Spatial diversity is considered to be
one of the most promising solutions for increasing capacity and spectral
Introduction 3
efficiency. Indeed, the integration of array processing and communications

techniques, exploiting the structure of antenna-array systems, has evolved
into a well-established technology. This technology is moving from the con-
ventional direction nulling and phase-arrays to advanced superresolution
spatiotemporal-arrays, MIMO array systems and arrayed wireless sensor
networks, which exploit the spatial and temporal properties of the channel
in their quest to handle multipaths, and to increase capacity and spectral
efficiency. Using these properties, an extra layer of co-channel interfer-
ence (CCI) and inter-symbol-interference (ISI) cancellation is achieved —
asymptotically providing complete interference cancellation.
The performance of array systems, especially the ones with super-
resolution capabilities is, in general, limited by three main factors:
• The presence of inherent background and sensor noise.

• The limited amount of information the sensors can measure due to finite
observation interval (number of snapshots) and array geometry.
• The lack of calibration, modelling errors and system uncertainties
that are embedded in the received array signal-vector x(t), which are
not accounted for. Examples include uncertainties in mutual coupling
between sensors, perturbations in the geometrical and electrical charac-
teristics of the array, the presence of moving emitters, nonplanar wave-
fronts, source angular/temporal spread, etc.
However, the overall quality of the system’s performance is naturally a

function of the array structure in conjunction with the geometrical charac-
teristics of the signal environment, as well as the algorithms employed. An
algorithm would behave differently when used with different array struc-
tures and, vice-versa, a certain array would generate different results when
its output is applied to different algorithms.
1.1 Nomenclature
It is assumed that the reader is familiar with the fundamentals of vector and
matrix algebra. In this book, for typographical convenience, matrices will be
denoted by blackboard bold symbols (e.g. A, T, I) or, in the absence of a cor-
responding blackboard bold symbol, by boldface (e.g. r, k, Γ). Any under-
lined symbol will represent a column vector, e.g. A, a, a. Derivatives with
respect to a general parameter p will be denoted with a “dot” (e.g. ȧ), while
the “prime” symbol (e.g. a ) will be reserved for differentiation with respect
to certain “invariant” parameters. The overall notation to be employed in

this book is as follows:
A, a Scalar
A, a Column vector
A, A Matrix
(·)T Transpose
(·)H Hermitian transpose
(·)† Pseudoinverse
·F Frobenius norm of a matrix
· Norm of a vector
|·| Magnitude
, Hadamard (Schur) product and division respectively
⊗ Kronecker product
exp(A or A) Elementwise exponential of vector A or matrix A
expm(A) Matrix exponential
Tr(A) Trace of matrix A
det(A) Determinant of A
diag(A) Diagonal matrix formed from the elements of A
diag(A) Column vector consisting of the diagonal elements of A
rowi (A) ith row of A
eleij (A) (ith , j th ) element of A
fix(A) Round down to integer
E{·} Expectation operator
Ab Element by element power
0N Zero vector of N elements
1N Column vector of N ones
IN N × N Identity matrix
ON ×d N × d Zero matrix
R Set of real numbers
N Set of natural numbers
Z Set of integer numbers
C Field of complex numbers
1.2 Main Abbreviations
AGS Ambiguous Generator Set

ELA Equivalent Linear Array
CRB Cramer Rao Bound
DF Direction Finding
DOA Directions of Arrival
Introduction 5
FOV Field-of-View
SNR Signal-to-Noise Ratio
RMS Root Mean Square
ULA Uniform Linear Array
UCA Uniform Circular Array
1.3 Array of Sensors — Environment
By distributing, in the 3-dimensional Cartesian space, a number N 2

of sensors (transducing elements, antennas, receivers, etc.) with a common
reference point, an array is formed. In general, the positions of the sensors
are given by the matrix r ∈ R3×N
T
r = r1 , r2 , . . . , rN = rx , ry , rz (1.1)
with rk = [xk , yk , zk ]T ∈ R3×1 denoting the Cartesian coordinates (loca-

tion) of the kth sensor of the array ∀k = 1, 2, . . . , N.
It is common practice to express the direction of a wave impinging on
the array in terms of the azimuth angle θ, measured anticlockwise from the
positive x-axis, and the elevation angle φ, measured anticlockwise from the
x-y plane, as illustrated in Fig. 1.1. Then, the (3 × 1) real unit-norm vector
Fig. 1.1 Relative geometry between a far-field emitting source and an array of sensors.
pointing towards the direction (θ, φ) is

T
u u(θ, φ) = cos θ cos φ, sin θ cos φ, sin φ (1.2)
Note that u = 1. If the velocity, wavelength and frequency of propaga-

tion of the incident wave is denoted by c, λ and Fc , respectively, then the
wavenumber vector in the direction (θ, φ) is defined as


 2πFc · u = 2π · u in meters
k = k(θ, φ) = c λ (1.3)

π · u in λ/2
In the most general case, the parameter space is
Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ) and φ ∈ (−90◦ , 90◦ )} (1.4)
but in most applications, Ω is restricted to only a sector of interest or,

in other words, field-of-view (FOV). For instance, in the case of ground
surveillance radars, only signals in the plane of the array are of interest —
i.e. the system is azimuth-only.
The array configuration is, to a large extent, dictated by the application
of interest. One obvious restriction is the shape and size of the available
site, which might be, to cite just a few examples, an aircraft’s wing, a ship’s
hull, a building rooftop, or simply a terrain. In addition, if the signals to be
intercepted are known to be coplanar and within a 180◦ field-of-view, as in
ground and marine navigation applications, then a linear or 1-dimensional
(1D) array of sensors may be sufficient.
(1) Linear or 1-dimensional (1D) Array.

The linear or 1D array consists of a one-dimensional distribution of
sensors along a line conventionally taken as the x-axis (Fig. 1.2(a)),
with sensor positions in units of half-wavelengths given by the matrix
r = [rx , 0N , 0N ]T ∈ R3×N (1.5)

where
rx = [r1 , r2 , . . . , rN ]T
The most popular array of this type is the standard Uniform

Linear Array (ULA) whose sensors are uniformly spaced at one half-
wavelength apart along the x-axis. For example, a 5-sensor standard
Introduction 7
Fig. 1.2 Examples of array geometries: (a) linear array, (b) uniform circular array,
(c) “cross” array.
ULA is given by rx = [−2, −1, 0, 1, 2]T in half wavelengths. Note that

the phase reference, or origin of the coordinate system, is taken at the
array centroid. By exploiting the regular structure of the ULA, many
DF algorithms can be simplified allowing for significant computational
savings. The FOV of 1D arrays is restricted to
Ω = {(θ, φ) : θ ∈ [0◦ , 180◦ ), φ = 0◦ } (1.6)
It can be easily deduced from the geometry of the problem that a lin-
ear array is incapable of distinguishing directions which are symmetric
with respect to the array line or which have the same elevation. If a
360◦ FOV is instead required, then a planar or 2D array should be
employed.
(2) Planar or 2-dimensional (2D) Array.

If the application of interest demands a full azimuthal FOV, that is,
θ ∈ [0◦ , 360◦ ), and possibly some elevation discrimination capability,
then a planar or 2D array is necessary. A planar array consists of a two
dimensional distribution of sensors over a plane, which is convention-
ally taken as the x-y plane, and whose sensor positions are given by the
matrix
r = [rx , ry , 0N ]T ∈ R3×N (1.7)
where rx and ry , respectively, are the column vectors of length N denot-

ing the x- and y-coordinates of the sensors in units of half-wavelengths.
The FOV of a planar array is the full azimuth space and half the ele-
vation space
Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ), φ ∈ [0◦ , 90◦ )} (1.8)
Note that a planar array cannot distinguish between directions which

are symmetric with respect to the array plane. Practical planar array
structures include the grid, X, Y and L-shaped arrays, but the most
popular is the uniform circular array (UCA) which, due to its symme-
try, exhibits uniform performance over the entire azimuthal space [16].
Fig. 1.2(b,c) illustrates two practical planar array (UCA and cross-
array) configurations used in radio direction-finding systems operating
in the UK in the HF (3 to 30 MHz) band. It is important to point out
that a planar array also permits the discrimination of signals in eleva-
tion φ ∈ [0◦ , 90◦ ) — i.e. the emitters need not be coplanar (as is the
case in airborne surveillance), although its resolving power in the eleva-
tion space is not as good as that of a 3D array. A 3D array has a FOV
spanning the entire parameter space. However, for most applications
1D and 2D arrays prove to be sufficient.
1.4 Pictorial Notation
1.4.1 Spaces/Subspaces
Because the visualization of a space greater than a 3-dimensional real
space is impossible, the notation shown in Fig. 1.3 is employed to pro-
vide some illustrative aid. However, this pictorial representation should be
used with care.
Introduction 9
Fig. 1.3 Illustrative Notation: (a) denotes a N -dimensional complex (or real) obser-
vation space. Note that any vector in this space has N elements; (b) denotes a one-
dimensional subspace/space spanned by the vector a; (c) denotes an M -dimensional
subspace/space (with M ≥ 2) spanned by the columns of the matrix A.
1.4.2 Projection Operator

Consider an (N ×M ) matrix A with M < N (i.e. the matrix has M columns)
and let L[A] represent the linear subspace spanned by the columns of A.
Assuming that the columns of A are linearly independent — that is, a
column of A cannot be written as a linear combination of the remaining
M − 1 columns — the subspace L[A] is of dimensionality M, i.e.
dim {L[A]} = M < N (1.9)
lying in a N -dimensional space H, as shown in Fig. 1.4. The complement

subspace to L[A] is denoted by L[A]⊥ and is of dimensionality N − M , i.e.

⊥
dim L[A] =N −M (1.10)
Then:
(1) any vector x ∈ L[A] can be written as a linear combination of the

columns of the matrix A
(2) any vector x ∈ H can be projected onto L[A] (or onto L[A]⊥ ), as
shown in Fig. 1.5, using the concept of the projection operator defined
as follows:
PA = projection operator onto the subspace L[A]

= A(AH A)−1 AH (1.11)
P⊥
A = the projection operator onto L[A]
⊥
= IN − PA (1.12)


(N × N ) matrix

N.B.: Properties of PA PA PA = PA (1.13)


P = PH
A A
Fig. 1.4 Illustrative representation of the subspaces L[A] and L[A]⊥ lying in an
N -dimensional complex space.
Fig. 1.5 Illustrative representation of the projection of any vector x ∈ H onto L[A] and
onto L[A]⊥ .
Introduction 11
(3) For any vector x ∈ L[A] the following expressions are valid

PA x = x
(1.14)
P⊥
A x = 0N
1.5 Principal Symbols
In this book the following symbols will always denote
p, q generic parameters
θ azimuth angle
φ elevation angle
α alpha — cone angle
β beta — cone angle
s arc length
N number of sensors
r (3 × N ) real matrix with columns the sensor locations
a array manifold vector
lm array manifold length
la array aperture
A array manifold curve
M array manifold surface
H space/subspace
1.6 Modelling the Array Signal-Vector and

Array Manifold
Consider an array of N sensors, with sensor locations r, operating in the

presence of M narrowband point sources and having the same known carrier
frequency Fc . The modelling of the signal due to the ith emitter, received
at the zero-phase reference point (taken to be the origin of the coordinate
system) is determined by whether the source is located in, or close to, the
array’s near-field or in the array’s far-field. In practice, this is determined
according to the value of its range ρi with respect to the array aperture la ,
defined as follows:

la = max
ri − rj
(1.15)
∀i,j
To be more specific, if ρi 2la /λ, where λ is the wavelength of the trans-

mitted signals, the ith source is located close to the array near-field border
(which defines the so-called Fresnel zone) and the spherical wave prop-
agation model has to be considered. For ρi > 2la /λ and especially for
ρi 2la /λ, the ith source is situated in the array far-field (the so-called
Fraunhofer zone) and plane wave propagation is assumed. As a matter of
fact, sources of range ρi 2la /λ or ρi > 2la /λ, but not ρi 2la /λ, are
usually regarded as being located in the near far-field of an array, for which
the spherical wave propagation model has to be utilized.
Based on the above discussion and by considering the M sources in the
far-field of the array, the array signal is the superposition of plane waves
from each individual source. The planewave/signal due to the ith emitter,
received at the zero-phase reference point, can be written as
ith signal at the reference point: mi (t) exp(j2πFc t) (1.16)
where mi (t) represents the complex envelope (message) of the signal and
exp(j2πFc t) represents the carrier. If τik represents the propagation delay
of the ith signal between the phase reference location and the kth sensor
then the signal arriving at the kth sensor will be a delayed version of the
signal given in Eq. (1.16), expressed as shown below
ith signal at the kth sensor: mi (t − τik ) exp(j2πFc (t − τik )) (1.17)

≈mi (t)
This propagation delay τik is a function of the DOA of the ith signal and
the position of the kth sensor with respect to the reference point. Indeed,
with reference to Fig. 1.6, τik can be derived as

ik
Pu r
i k
τik = =
c c
rTk PTui Pui rk
=
c
rk Pui rk
T
= (using Eq. (1.13))
c −1 T
rTk ui uTi ui ui r k
=
c
Introduction 13
Fig. 1.6 Geometry of a travelling plane wave, relative to the kth sensor and the array
reference point.
2
rTk ui uTi rk rTk ui rTk ui
= = =
c c c
i.e.
rTk ui
τik = (1.18)
c
It is reasonable to assume that the envelope (baseband signal mi (t)) does
not change significantly as it traverses the array. Hence τik τmax , ∀i, k
and mi (t − τmax ) mi (t) has been used in Eq. (1.17), where τmax is the
maximum possible time for a signal to traverse the array. This is due to the
narrowband assumption, i.e. the highest frequency in the message signal
is much less than the carrier frequency. By taking into account the above
assumption, as well as the complex response (gain and phase) γk (θi , φi ) of
the kth sensor at frequency Fc in the azimuth-elevation direction (θi , φi ),
the received signal at the kth sensor due to the ith emitting source can be
written as

rT · u
γk (θi , φi ) · mi (t) · exp −j2πFc k i exp(j2πFc t)
c
= γk (θi , φi ) · mi (t) · exp(−jrTk · ki ) exp(j2πFc t) (1.19)
where ki = k(θi , φi ) denotes the wavenumber vector in the direction of the

ith emitter.
Next, the received signal is downconverted to baseband by multiplying
with exp(−j2πFc t). The baseband signal at the output of the kth sensor,
in the presence of M far-field emitters and additive baseband noise nk (t)
can be written as

M
xk (t) = γk (θi ,φi ) mi (t) exp(−j rTk ki ) + nk (t) (1.20)
i=1
Using vector notation, the output from all the N sensors, can be
expressed as
T
x(t) = x1 (t), x1 (t), . . . , xN (t)

M
= mi (t) ai + n(t) (1.21)
i=1
T
where n(t) = n1 (t) n2 (t) · · · nN (t) ∈ CN ×1 denotes the baseband
additive white Gaussian noise of power σ . The (N × 1) complex vector
2
ai is known as the array manifold vector associated with the ith emitting
source (signal), representing the complex array response to a unit amplitude
plane wave impinging from direction (θi , φi ), compactly written as
 
 
ai a(θi , φi ) = γ(θi , φi ) exp −j [rx , ry , rz ]k(θi , φi ) (1.22)

=rT
where
γ(θi , φi ) = [γ1 (θi ,φi ), γ2 (θi , φi ), . . . , γN (θi , φi )]T ∈ CN ×1
Introduction 15
This vector is also referred to as the “array response vector” or “array

steering vector.”
Given that the same array may be represented by an infinite number
of matrices r through a change in the coordinate system reference point, a
choice is made to fix the coordinate reference point (0, 0, 0) to be the array
centroid. This translates into a condition on the location vectors r x , ry and
rz such that
sum(rx ) = sum(ry ) = sum(rz ) = 0 (1.23)

or equivalently,
rTx 1N = rTy 1N = rTz 1N = 0 (1.24)
As can be seen from Eq. (1.22), a shift in the reference point of the array
coordinate system results in a manifold vector ar (θ, φ) which is related to
the initial manifold vector:
ar (θ, φ) = g (θ, φ) · a (θ, φ) (1.25)
where g(θ, φ) is a complex number with |g(θ, φ)| = 1, ∀(θ, φ). In fact,
g(θ, φ) = exp(−jrT0 k(θ, φ)) where r0 is the translation vector of the ref-
erence point from the centroid of the array. For simplicity, all arrays
henceforth will have their centroid as the coordinate reference point, i.e.
g(θ, φ) = 1, ∀(θ, φ).
Note that in the case of a linear array of isotropic sensors Eq. (1.22)
simplifies to
a(θ) = exp(−jπrx cos θ) (1.26)
while for a planar array it takes the following form:
a(θ, φ) = exp(−jπr(θ) cos φ) (1.27)

where
r(θ) = rx cos θ + ry sin θ (1.28)
It can easily be seen that r(θ) in Eq. (1.27) represents the positions of
the projections of the sensor locations of the planar array onto the line of
azimuth θ (see Fig. 1.7), and is known as the “Equivalent Linear Array”
(ELA) along the direction θ. Notice the similarity between the response
Fig. 1.7 Equivalent Linear Array (ELA) associated with the azimuth angle θ.
vectors of the linear and planar arrays in Eqs. (1.26) and (1.27) respectively.
Using Eq. (1.22) the observed array signal-vector x(t) ∈ CN ×1 of Eq. (1.21)
can be written concisely as
x(t) = Am(t) + n(t) (1.29)

where

A = a1 , a2 , . . . , aM
(1.30)
m(t) = [m1 , m2 , . . . , mM ]T
where A ∈ CN ×M is the array manifold (response) matrix and m(t) ∈ CM ×1

is the vector of signal envelopes (messages). Note that in the case of isotropic
sensors the matrix A can be written in a compact way as
A = exp(−jrT k) (1.31)
with k = [k1 , k2 , . . . , kM ] representing the wavenumber matrix.

Based on the above model of Eq. (1.29), the theoretical covariance
matrix Rxx of the array signal-vector x(t) can be formed as
Rxx E{x(t) x(t)H } ∈ CN ×N

= ARmm AH + σ 2 IN (1.32)
Introduction 17
where Rmm E{m(t) m(t)H } ∈ CM ×M is the source covariance matrix,

which is diagonal in the case of uncorrelated signals and σ 2 IN is the addi-
tive white Gaussian noise covariance matrix Rnn E{n(t) n(t)H } ∈ CN ×N .
In practice, only a finite number L data snapshots [x(t1 ), x(t2 ), . . . , x(tL )],
collected by an array of sensors, can be observed. Hence only an esti-
mate of Rxx is computed in the form of the sample covariance matrix
given by
L
xx 1
R
1
x(ti ) x(ti )H = XXH (1.33)
L i=1 L
where
X [x(t1 ), x(t2 ), . . . , x(tL )] ∈ CN ×L
It is important to point out that if the assumption mi (t − τik ) ≈

mi (t), ∀i, k used in Eq. (1.17) is not valid, then the array signal vector
associated with the ith signal should be modelled as

M
x(t) = mτi (t) ai + n(t) (1.34)
i=1
where
mτi (t) = [mi (t − τi1 ), mi (t − τi2 ), . . . , mi (t − τiN )]T
1.7 Significance of Array Manifolds
It is clear from Eqs. (1.22), (1.26) or (1.27) that the manifold vector for
a particular direction contains all the information about the geometry
involved when a wave is incident on the array from that direction. By
recording the locus of the manifold vectors as a function of direction, a
“continuum” (i.e. a geometrical object such as a curve or surface) is formed
lying in an N -dimensional space. This geometrical object (locus of manifold,
or response, vectors) is known as the array manifold. The array manifold
can be calculated (and stored) from only the knowledge of the locations and
directional characteristics of the sensors. Thus, according to Schmidt [1],
“the array manifold completely characterizes any array and provides a rep-
resentation of the real array into N -dimensional complex space.”
The significance of the array manifold concept becomes apparent when it
is realized that all subspace-based parameter estimation algorithms involve
searching over the array manifold for response vectors which satisfy a
given criterion. For example, in the case of the MUSIC algorithm [7],
the manifold is searched for the array manifold vectors which are (nearly)
orthogonal to the estimated noise subspace. In other words the manifold
is searched for vectors belonging to the subspace L[A] i.e. the subspace
spanned by the columns of the matrix A = [a1 , a2 , . . . , aM ] (the signal sub-
space). Figure 1.8 illustrates a single-parameter manifold curve embedded
in N -dimensional complex space together with the subspace L[A] associated
with two sources. These two sources can be asymptotically resolved, irre-
spective of their angular separation — hence the title “superresolution.” It
is clear that the accuracy of the DOA estimates is dependent on how accu-
rately the noise subspace is estimated; which in turn is a function of the
observation interval and the signal-to-noise ratio (SNR). Thus the DOAs
can be exactly estimated from the array output only on the basis of either
an infinite number of snapshots (i.e. L → ∞) or, an infinite signal-to-noise
ratio scenario (i.e. n(t) = 0) — asymptotic conditions.
Another determining factor which might not be initially apparent is
the shape of the array manifold. A little thought reveals that a poten-
tially unresolvable situation might arise if two response (manifold) vectors
corresponding to different DOAs happen to be identical. For instance, this
“abnormality” arises when a single-parameter array manifold crosses upon
itself, as shown in Fig. 1.9. Such phenomena are commonly referred to as
“ambiguities” and are a direct consequence of the array geometry dictating
Fig. 1.8 Illustrative representation of an array manifold curve embedded in N -

dimensional complex space and the subspace spanned by two manifold vectors.
Introduction 19
Fig. 1.9 Illustrative representation of array “abnormality” (ambiguity).
the behavior of the array manifold. As will be discussed below, the behavior
of the array manifold, in particular its local geometry, plays an important
role in handling “abnormalities” and also in defining the capabilities (e.g.
the resolving power) of an array system.
1.8 An Outline of the Book
The theoretical framework associated with curves lying in N -dimensional

complex space, qualifying to be manifolds of linear array structures, is
presented in Chapter 2. More specifically by recording the locus of the
vector a(p) as a function of p, a one-dimensional continuum (i.e. a curve A)
is formed known as the array manifold and embedded in a N -dimensional
space. Thus in this chapter the properties and characteristics of single-
parameter manifolds (i.e. curves) are investigated and supported by a num-
ber of representative examples of symmetric and non-symmetric linear array
geometries [10]. These manifold curves have been found to have a hyper-
helical shape with numerous advantages. For example, all the curvatures
of a hyperhelix are constant (do not vary from point to point) and may
be evaluated recursively. The convenient nature of a hyperhelix’s geometry
will be proven invaluable for the rest of the book, not only for linear but
also for non-linear (2D and 3D) arrays.
If the manifold vector a corresponds to two parameters (p,q) where,

for instance p may be the azimuth and q the elevation angle, then a
2-dimensional continuum, i.e. a surface M, is formed by recording the locus
of the vector a(p, q) as a function of both p and q. Chapter 3 is concerned
with the study of manifold surfaces embedded in a multidimensional com-
plex space and presents the essential differential geometry parameters which
have been grouped into those related to the surface itself, and those related
to curves lying on a surface [11]. Then a manifold surface is treated as
a family of curves on the surface by an appropriate parametrization. This
treatment is very convenient as it permits a unified framework for the analy-
sis of the linear and non-linear array manifolds. To provide a simplified
representation of the analysis, with many potential benefits, the concept
of isometric mapping is also introduced in Chapter 3. Then an isometric
mapping of the manifold surface (embedded in a multidimensional com-
plex space) onto the real plane (two-dimensional space) is presented which
preserves certain differential geometry characteristics of the manifold sur-
face, under certain conditions.
It is common practice, and intuitively appealing, in array processing
to use azimuth and elevation as the directional parameters of a waveform
impinging on an array of sensors. However, this is by no means unique
and, furthermore, this is not the most suitable parametrization for the
study of the behavior of the array manifold [12]. Therefore in Chapters 4
and 5, based on the material presented in Chapters 2 and 3, the following
manifold-surface parametrizations
• the (azimuth, elevation), or (θ, φ), parametrization

• the cone-angles, or (α, β), parametrization
are examined for non-linear arrays (2D and 3D geometries). The significance
of these parametrizations is demonstrated by a number of representative
examples. Furthermore, properties such as Gaussian and geodesic curva-
tures are defined and their implications with regards to isometric mappings
are discussed.
In the next two chapters of the book (Chapters 6 and 7) the funda-
mental effects of the array geometry behavior on the performance of the
system irrespective of the type of algorithm used, are studied. The array
structure is incorporated into the observed array signal-vectors x(t) (and
therefore to the estimation problem) through the array manifold. Thus the
geometry of an array plays a crucial role in dictating the shape, properties
Introduction 21
and “anomalies” of the array manifold, and as a consequence, in dictat-

ing the phenomenon where some manifold vectors can be written as linear
combinations of others [13,25]. This gives rise to the array “ambiguity prob-
lem” where the occurrence of false parameter estimates seriously impairs
the performance of array system irrespective of its other capabilities. In
particular, in Chapters 6 and 7 two general types of ambiguities are pre-
sented based on the partitioning of the array manifold curves into equal or
unequal segments from which the ambiguous generator sets (AGS) can be
constructed — with each AGS representing an infinite number of ambigu-
ous sets of parameter values (e.g. directions). Furthermore, the theoretical
aspects of the investigation provide a sufficient condition for the presence
of ambiguities while the results are then extended to non-linear arrays by
treating surfaces as families of curves [14].
Finally, in Chapter 8, the knowledge of the shape of the array mani-
fold is used to determine the array’s ultimate capabilities according to the
following criteria:
• Accuracy and the Cramer Rao Lower Bound
• Detection threshold
• Resolution threshold
In particular, by approximating, locally, a manifold curve with a circular
arc, the Cramer Rao Lower Bound (CRB) is studied and its relation to
resolution and detection bounds is established [15, 16]. Then by defining
the detection and resolution subspaces, in conjunction with the circular
approximation (locally) of the array manifold, the minimum arc length
separation in order to detect and resolve two sources located close together
is estimated. This is done in terms of the curve’s principal curvature only,
thereby simplifying the analysis considerably.
Chapter 2
Differential Geometry of Array

Manifold Curves
2.1 Manifold Curve Representation — Basic Concepts
Let a a(p) ∈ CN be the manifold vector of an array of N sensors where p

is a generic directional parameter. This is a single-parameter vector function
and as p varies the point a will trace out a curve A, as shown in Fig. 2.1,
embedded in an N -dimensional complex space CN . This was expected, as
vector functions of one parameter are used to define space curves — also
known as single-parameter manifolds.
The curve A, which is formally defined

A a(p) ∈ CN , ∀p : p ∈ Ωp (2.1)
where Ωp denotes the parameter space, is said to be a regular parametrized

differential curve if
ȧ(p) = 0N , ∀p ∈ Ωp (2.2)
where a “dot” at the top of a symbol is used to denote differentiation

with respect to parameter p. As the vector ȧ(p) represents the tangent
vector to the curve, the “regularity” condition of Eq. (2.2) ensures that the
tangent vector exists at all points on the array manifold. The arc length
s(p) along the manifold curve A and its rate-of-change ṡ(p) are formally
defined, respectively, as
p
s(p) ȧ(p) dp (2.3)
0
and
ds
ṡ(p) = ȧ(p) (2.4)
dp
22
Differential Geometry of Array Manifold Curves 23
Fig. 2.1 Manifold curve A embedded in CN .
Fig. 2.2 Manifold parameterization in terms of arc length s.
The array manifold is conventionally parametrized in terms of a generic

bearing parameter p. For instance for a linear array, p may represent the
azimuth angle θ. However, parametrization in terms of the arc length s (see
Fig. 2.2), which is the most basic feature of a curve and a natural parameter
representing the actual physical length of a segment of the manifold curve in
CN , is more suitable. There is a further advantage of using s as a parameter:

the arc length s (in contrast to p) is an “invariant” parameter. This means
that the resulting tangent vector to the curve, expressed in terms of s,
always has unity length. Indeed, it can be seen that

da(s)

da(p)/dp
ȧ(p) ṡ(p)
a (s) =

=

ds

ds/dp
= ṡ(p) = ṡ(p) = 1, ∀s (2.5)
which is a result used next in defining the unit tangent vector u1 (s) (see
Fig. 2.3)
u1 (s) a (s) (2.6)
Unless otherwise stated, differentiation with respect to s will be denoted

by “prime” and differentiation with respect to any other parameter p will
be denoted by “dot.” For example,
da(s) d2 a(s) d3 a(s) da(p)
a ; a ; a ; ȧ (2.7)
ds ds2 ds3 dp
Note that Eqs. (2.3) and (2.4) can be interpreted in physical terms by
considering the curve A to be a “route” travelled by a moving object as
a function of “time” p (see Fig. 2.4). Then ȧ(p) is the tangent vector to
the curve at various points and equals the “velocity” vector of the moving
object. The rate of change of arc length ṡ(p) is then the magnitude of
the “velocity” vector (tangent vector) and represents the “speed” of the
Fig. 2.3 Arc length s is an “invariant” parameter — tangent vector u1 (s) has unit
norm ∀s.
Fig. 2.4 The parameter p is not an invariant parameter — i.e. speed = ȧ(p) = ṡ(p) =
variable = 1.
moving object while the total distance travelled, s(p), can be obtained by
integrating speed with respect to “time” p. Parametrization in terms of s,
i.e. a(s), indicates that the “velocity” vector (tangent vector) a (s) always
has constant length equal to 1 (moving object with unity speed).
Note that, in general, it is not necessary to mention the origin of
the arc length (s = 0) since most concepts are defined only in terms
of the derivatives of a(s). Furthermore, the rate of change of arc length
ṡ(p) is a local property of the curve and, as will be demonstrated in
Chapter 8, plays a crucial role in dictating the resolution/detection capa-
bilities offered by the array manifold and consequently by the geometry of
the array.
As was mentioned before, the tangent vector a (s) is of unity length
∀s, indicating that s is an “invariant” parameter. In this case the norm
of the second derivative of the manifold vector (parametrized in terms of
s) a (s) measures how fast the curve pulls away from the tangent line
at s, in the neighborhood of s (see Fig. 2.5). In other words, it measures
the rate of change of the angle which neighboring tangents make with the
tangent at s. The number a (s) = κ1 (s) is called the first curvature of
the manifold curve at s. As we will see next, although many curvatures can
be defined at a point on the manifold curve A ∈CN , the first curvature is
the most important one and for this reason this is also known as “principal”
curvature.
Fig. 2.5 How fast the curve pulls away from the tangent line, at point s, is measured
by a (s) with a (s) = κ1 (s).
2.2 Curvatures and Coordinate Vectors in CN
We have seen that the arc length s is the invariant parameter of a manifold
curve of an array of N sensors. With the array manifold curve embedded in
an N -dimensional complex space (or equivalently in a 2N real space) and
parametrized in terms of the arc length s, it is essential to attach/define
to the running point a = a(s) on the curve a number of curvatures as well
as a continuous differential and orthonormal system of coordinates u1 (s),
u2 (s), . . . , u2N (s).
2.2.1 Number of Curvatures and Symmetricity

in Linear Arrays
The curvatures of a space curve are of immense value in differential geom-
etry as, according to the fundamental uniqueness theorem of [2, 17], cur-
vatures uniquely define a space curve expressed in terms of arc length s,
except its position in space.
For an array of N sensors (i.e. the manifold vector a(s) ∈ CN may be
described by 2N real components), at most 2N − 1 manifold curvatures
can be defined. However, if the array has some symmetrical sensors with
respect to the array centroid, then the manifold curve is situated wholly
in some subspace of dimensionality d (within the 2N dimensional space).

Therefore, in this case up to d − 1 non-zero curvatures can be defined. In
particular, if m denotes the number of sensors in symmetrical pairs then d
is given as follows:

2N − m if sensor at the array centroid
d= (2.8)
2N − m − 1 otherwise
This implies that 0 ≤ m ≤ N and according to m, linear sensor arrays are

divided into the following broad categories:
(1) Symmetric: m = N , i.e. all the sensors occur in symmetric pairs about
the origin.
(2) Non-Symmetric: m < N
(a) Partially symmetric: 0 < m < N , i.e. at least one sensor has a
symmetric counterpart about the origin;
(b) Fully asymmetric: m = 0, i.e. no sensor has a symmetrical coun-
terpart about the origin.
Note that a sensor at the origin is taken to be a symmetrical sensor.

A representative example from each category is illustrated in Figs. 2.6, 2.7
and 2.8.
Fig. 2.6 A symmetric linear array of N = 5 elements (m = 5, d = 2N − m − 1 = 4).
Fig. 2.7 A partially symmetric linear array of N = 5 elements (m = 2, d = 2N −m = 8).
Fig. 2.8 An asymmetric linear array of N = 4 elements (m = 0, d = 2N − m = 8).

2.2.2 “Moving Frame” and Frame Matrix

In the previous section we have seen that to the running point a(s) on the
curve A we attach a frame of orthonormal vectors u1 (s), u2 (s), . . . , u2N (s)
together with d − 1 non-zero curvatures κ1 (s), . . . , κd−1 (s) with N − 1 ≤
d ≤ 2N .
By ignoring the coordinate vectors with indices greater than d (those
corresponding to zero curvatures) a frame of d orthonormal vectors u1 (s),
u2 (s), . . . , ud (s), can be defined and attached to the running point a(s) on
the manifold curve, making a(s) the “origin” of the new coordinate system.
This set of coordinate vectors, known as a “moving frame” (see Fig. 2.9),
forms the matrix
U(s) = [u1 (s), u2 (s), . . . , ud (s)] ∈ CN ×d (2.9)
and is derived from a fixed known frame U(0), (i.e. at s = 0 say) by rotation,
using the transformation matrix F(s) ∈ Cd×d , i.e.
U(s) = U(0) · F(s) where F(0) = Id (2.10)
The matrix F(s) is a continuous differential real transformation matrix

called frame matrix. This is a non-singular matrix with its main properties
listed in Table 2.1.
Fig. 2.9 “Moving frame” U(0) and U(s).

Table 2.1 Properties of the frame matrix F(s).
1st property F(s) ∈ Rd×d

2nd property det(F(s)) = 1
3rd property F−1 (s) = FT (s)
4th property FT (s) · F(s) = Id
5th property eleij (F) = (−1)i+j eleji (F)

d
Tr(F)−ε 0 if d even
6th property eleii (F) = 2
where ε =
i=2 1 if d odd
even
To summarize, the d×d real matrix F(s) is known as the “frame matrix”
while the set of coordinates represented by the N × d complex matrix U(s)
is the “moving frame.”
2.2.3 Frame Matrix and Curvatures

The question is, how the frame matrix F(s), at the running point s, is
related to the curvatures of the manifold attached to this point. The initial
part of the mechanism to answer this question is to define the coordinate
vectors in CN by exploiting and extending the procedure used in real spaces
where a basis of up to N orthonormal vectors is necessary to fully define a
curve in RN [17]. This extension, which provides the d coordinate vectors in
CN blended with the curvatures, is summarized in a structured formation
in Table 2.2.
Now let us focus on the (i + 1)th row of Table 2.2. That is,
ui (s) + κi−1 .ui−1 (s)
ui+1 (s) = (2.11)
κi
By solving Eq. (2.11) with respect to ui (s), it is obvious that the differen-
tiation of the ith coordinate vector, for i 2, can be written as
ui (s) = κi (s)ui+1 (s) − κi−1 (s)ui−1 (s) for i 2 (2.12)

with
u1 (s) = a (s) = κ1 (s)u2 (s)
Note that the coordinate vectors are normalized to unity length.

Equation (2.12) which can be rewritten in a more compact form as
U (s) = U(s) C(s) (2.13)

Table 2.2 Estimation procedure of coordinate vectors and curvatures of

a curve A.
Coordinate Vector Curvature

1st u1 (s) a (s) κ1 (s) = u1 (s)
1
2nd u2 (s) = u (s)
κ1 1
κ2 (s) = u2 (s) + κ1 u1 (s)
u2 (s) + κ1 u1 (s)
3rd u3 (s) = κ3 (s) = u3 (s) + κ2 u2 (s)
κ2
... ... ...
... ... ...
ui−1 (s) + κi−2 .ui−2 (s)
ith ui (s) = κi (s) = ui (s) + κi−1 .ui−1 (s)
κi−1
ui (s) + κi−1 .ui−1 (s)
(i+1)th ui+1 (s) = κi+1 (s) = ui+1 (s) + κi .ui (s)
κi
... ... ...
... ... ...
ud−1 (s) + κd−2 .ud−2 (s)
dth ud (s) = κd (s) = 0
κd−1
where U (s) = [u1 (s), u2 (s), . . . , ud (s)] ∈ CN ×d , d is the dimensionality of
the subspace in which the curve is embedded. The matrix C(s) ∈ Cd×d is
termed the Cartan matrix, which is a real skew-symmetric matrix of the
curvatures defined as follows:
 
0, −κ1 (s), 0, ..., 0, 0
κ (s), −κ2 (s), . . . , 
 1 0, 0, 0 
 
 0, κ2 (s), 0, ..., 0, 0 
C(s)  .. .. .. .. .. ..  (2.14)

 . . . . . . 
 
 0, 0, 0, ..., 0, −κd−1 (s)
0, 0, 0, . . . , κd−1 (s), 0
Starting with Eq. (2.13) and then using Eq. (2.10), we have
U (s) = U(s) C(s)

⇒ (U(0) F(s)) = U(0) F(s) C(s)
⇒ U(0) F (s) = U(0) F(s) C(s)
leading to the following expression
F (s) = F(s) C(s) (2.15)

This is a first order differential equation of the frame matrix F(s) with
initial condition F(0) = Id . Therefore, its solution is
F(s) = expm(s C(s)) (2.16)
where expm(·) denotes the matrix exponential. Equation (2.16) provides the
relationship between the frame matrix F(s) at the running point s and the
curvatures (Cartan matrix) of the manifold attached to this point, which
can also be used to rewrite Eq. (2.10) as follows:
U(s) = U(0) expm(s C(s)) (2.17)
Finally, Eq. (2.15) implies that the Cartan matrix can always be written
as a function of the frame matrix F(s) as follows
C(s) = F−1 (s) F (s) = FT (s) F (s) (2.18)
where the 3rd property of F(s) given in Table 2.1 has been used. Remember
that the Cartan matrix, as a formation of the curvatures, contains all the
information about the local behavior and shape of the manifold curve.
2.2.4 Narrow and Wide Sense Orthogonality

As stated previously, for a general array manifold curve it is essential to
attach, to a running point on the curve, a continuously differentiable and
orthonormal system of coordinates U(s). The orthonormality of this contin-
uously differentiable “moving” frame relies on the vector continuum ui (s)
with constant magnitude and being orthogonal to its tangent. That is,
ui (s) ⊥ ui (s), i.e. ui (s)H ui (s) = 0
This, however, is only true for the broad class of arrays which are symmetric
with respect to their centroid, in which case m = N . In general, if the array
is nonsymmetric (m < N ), then the orthogonality is not valid although the
complex vectors ui (s) have constant (unity) magnitude. In particular, the
inner product of a complex vector of constant norm with its derivative is
purely imaginary, i.e.

uH
i (s)ui (s) = imaginary (2.19)
This leads us to define orthogonality of two complex vectors of unity

magnitude as being “wide-sense” and “narrow-sense” orthogonality,
as follows:
Definition 2.1 Wide sense and narrow sense orthogonality of two com-
plex vectors of unity magnitude:
“wide” sense orthogonality: Re(uH

i (s).uj (s)) = 0, for i = j (2.20)
“narrow” sense orthogonality: uH
i (s) · uj (s) = 0, for i = j (2.21)
Thus, for a general linear array, that is, for a non-symmetric array, the
wide-sense orthogonality should be used, i.e.
Re(UH (s) · U(s)) = Id (2.22)
This, of course, is simplified to narrow-sense orthogonality UH (s)·U(s) = Id

for symmetrical arrays.
Using the “wide sense” orthogonality for a general linear array, and
starting with Eq. (2.18), we get
Re(UH (s) · U (s)) = C(s) (2.23)
Indeed
C(s) = FT (s) · F (s) = FT (s) · Id · F (s) = FT (s) · Re(UH (0) · U(0)) · F (s)

=Id
 
 
= Re FT (s) · UH (0) · U(0)F (s) = Re(UH (s)U (s))

=UH (s) =U (s)
Furthermore, as it is shown in Appendix 2.7.1, the ith derivative of the

manifold vector with respect to the arc length s (denoted by a(i) (s)) is
related to the coordinate vector in a simple way as follows:
&& 'H '
Re a(i) (s) · ui (s) = κ1 (s) · κ2 (s) · · · κi−1 (s)
(
i−1
= κ (s) (2.24)
=1
Based on Eq. (2.24), it is easy to prove that the determinant of the N × N

matrix formed by the first N derivatives of the manifold vector is related
to the curvatures by
|det[a (s), a (s), . . . , a(N ) (s)]|

−1 −2
= κN
1 (s) · κN
2 (s) · · · κ2N −2 (s) · κN −1 (s)
(
N −1
−
= κN
(s) (2.25)
=1
2.3 “Hyperhelical” Manifold Curves
In this section, the shape of the array manifold of a linear array of N

isotropic sensors is investigated using the theory developed in the previous
section. The manifold is parametrized in terms of a directional parameter
p with the sensor locations given by the vector r in half-wavelengths. The
theory is valid for any regular parametrized differential manifold curve A
defined as follows:
A {a(p) ∈ CN , ∀p : p ∈ Ωp } (2.26)
where
a(p) = exp(−j(πr cos p + v)) (2.27)
where r and v are two constant vectors and p is a generic parameter

(parameter of interest). The main results are presented in the form of two
theorems but, firstly, it is easy to show using Eq. (2.3) that the arc length
s(p), and rate of change of arc length ṡ(p), are given by the following expres-
sions respectively:

s
s(p) = π r (1 − cos p) i.e. p(s) = cos−1 1 − (2.28)
πr
and
ṡ(p) = πr sin p (2.29)
where the initial condition s(0) = 0 has been assumed.

It is worth noting that for a linear array of N sensors with locations
r (in units of half-wavelengths) the rate of change of the arc length is
a non-linear function of the directional parameter p and depends on the
norm of the vector of sensor locations. From Eq. (2.29), it can be further
deduced that for a small directional increment ∆p | p2 − p1 |, the cor-
responding change in arc length, to a first order approximation, can be
calculated as:
∆s = |s(p2 ) − s(p1 )| (2.30)

= πr|cosp2 − cos p1 | (2.31)
≈ πr∆p sin p̆; for small ∆p (2.32)
where p̆ (p1 + p2 )/2. This expression reveals crucial properties of the

array manifold of a linear array and provides considerable insight into its
resolving power. Consider a linear array and two impinging emitters with
a directional separation of ∆p. It can easily be seen from Eq. (2.32) that
the distance (arclength ∆s) between the corresponding manifold (response)
vectors in CN is maximum when the emitters are at broadside and minimum
when they are at endfire. Furthermore, if a larger array is employed, then
the corresponding manifold vectors are further apart in CN . These facts
are demonstrated by the variations of the rate of change of arc length
with azimuth for 5-sensor and 7-sensor standard ULAs of half-wavelength
spacings, plotted in Fig. 2.10.
Fig. 2.10 Variations of ṡ(θ) for ULAs of 5 and 7 sensors.

This explains the better resolution of sources at broadside and the better
resolving power of the 7-sensor standard ULA, as evidenced by Fig. 2.10.
It can therefore be concluded that the rate of change of arc length of the
array manifold plays a prominent role in determining the resolving power
of the array, as will be further detailed in Chapter 8.
Another important parameter of the array manifold curve is its total
length,
lm s(π) = 2π r (2.33)
which clearly increases in direct proportion to the sensor spacings. As can be

seen, the manifold length depends on the number of sensors and their posi-
tions. The manifold length is expected to influence the ambiguity properties
of a linear array since it is obvious that longer manifolds are more prone to
“spurious” parameters/results (manifold vectors that can be expressed as
linear combinations of other manifold vectors) unless the arrays are care-
fully designed.
The array manifold can now be parametrized in terms of its arc length
as follows:
A = {a(s) ∈ CN , ∀s : s ∈ [0, lm ]} (2.34)

where
a(s) exp(j()
rs − πr + v)) (2.35)
and
r
)
r (2.36)
r
is the vector of “normalized” sensor positions.
From Eq. (2.35), and bearing in mind that the coordinate vectors are of
unit length, the following expressions for the first three manifold curvatures,
for instance, may be derived from the first three rows of Table 2.2
u1 (s) a (s) = j)r a(s)

1 1 2
u2 (s) = u (s) = − ) r a(s)
κ1 1 κ1
1 j
u3 (s) = (u2 (s) + κ1 u1 (s)) = − r3 − κ21)
() r) a(s)
κ2 κ1 κ2
(2.37)
κ1 (s) = u1 (s) = )
r2
1 3
κ2 (s) = u2 (s) + κ1 u1 (s) = )
r − κ21)r
κ1
1
κ3 (s) = u3 (s) + κ2 u2 (s) = r4 − (κ21 + κ22 ))
) r2
κ 1 κ2
In general, it can be shown that the ith coordinate vector and curvature
of the manifold of a linear array of isotropic sensors can be calculated
according to the following theorem.
Theorem 2.1 Hyperhelical Manifolds: The manifold of any linear

array of N omnidirectional sensors with locations given by r,
(a) is a curve of hyperhelical

√ shape lying on a complex N -dimensional
sphere with radius N .
(b) The coordinate vectors and curvatures of this hyperhelix depend on the
lower order curvatures, the number of sensors and their relative spacing
and can be estimated by the following recursive equations:
fix( i−1
(j)
i
2 )+1
ui (s) = ri−2n+2 a(s) (2.38)

(−1)n−1 bi−1,n )
κ1 κ2 · · · κi−1 n=1

i

fix( 2 )+1
1

κi =
(−1)n−1
b )
r i−2n+3
(2.39)
κ1 κ2 · · · κi−1
i,n

n=1
where


) r
r = r (normalized sensor positions)


κ = ||)2
r ||
1

κi−1 = 0



sum(r) = 0 (i.e. phase reference = array centroid)
Furthermore, the coefficients bi,n are given by:
i−2n+5
i−2n+3
i−1
bi,n = ··· κ2m1 κ2m2 · · · κ2mn−1 ; i, n > 2
m1 =1 m2 =2+m1 mn−1 =2+mn−2
(2.40)
with

bi,1 = 1; i 1
* 2
i−1
bi,n = κm i > 1
m=1
or recursively by
bi,n = bi−1,n + κ2i−1 bi−2,n−1 , i > 2, n > 1 (2.41)
with the initial conditions:

bi,1 = 1, i≥1
(2.42)
b2,2 = κ21
Corollary 2.1 The manifolds of uniform linear arrays are hypercircular

arcs (closed hyperhelices).
The proof of the above theorem can be found in Appendix 2.7.2
(page 54). Note that for notational convenience the second condition (row)
of Eq. (2.8) may be ignored by redefining d 2N − m with N ≤ d ≤ 2N .
This implies that in the special case of an array with a sensor at the array
centroid, we attach to the running point a(s) on the curve A a moving
frame of u1 (s), u2 (s), ..., ud (s) together with d − 2 non-zero curvatures and
one zero curvature (κd−1 (s) = 0). In this special case Eq. (2.38) cannot be
used to calculate the dth coordinate vector ud (s) since the curvature κd−1
vanishes. Using an orthogonalization procedure, the (N × 1) vector ud (s)
can be shown to be given by
T
ud (s) = [ 0, . . . , 0, 1, 0, . . . , 0 ] (2.43)
where the non-zero entry is in the same position as the sensor at the centroid
in the vector of sensor locations )
r.
2.3.1 Coordinate Vectors and Array Symmetricity

A close examination of Eq. (2.39) of Theorem 2.1, reveals some important
properties of the manifold of a linear array of isotropic sensors. In particular,
unlike the real case, the coordinate vectors of the linear array manifold curve
embedded in CN are not in general mutually orthogonal. For example:

uH
j T 3
) 

1 (s)u 2 (s) = 1 r 

κ1 N 

H
u1 (s)u3 (s) = 0 
(2.44)
r3 
j 1 T 5
uH2 (s)u3 (s) = 1N )
r − 1TN ) 

κ2 κ1 2 



uH (s)u (s) = 0
2 4
It can, however, be seen that some of the coordinate vectors (some

columns of matrix U(s)) are mutually orthogonal. Indeed, it can be shown
that the odd-indexed coordinate vectors forming the matrix

Uodd (s) = u1 (s), u3 (s), u5 (s), . . . , udodd (s) (2.45)
are mutually orthogonal and that the same holds for the even-indexed coor-
dinate vectors forming the matrix

Ueven (s) = u2 (s), u4 (s), u6 (s), . . . , udeven (s) (2.46)
where deven = fix( d2 ) and dodd = fix( d−1 2 ) + 1. This can be expressed in a
compact form as
.
UH
odd (s)Uodd (s) = Idodd
(2.47)
UHeven (s)Ueven (s) = Ideven
with .
UHodd (s)Ueven (s) = Ododd ×deven
(2.48)
Re{UH odd (s)Ueven (s)} = Ododd ×deven
Note also that, at broadside (i.e. p = 90◦ ), Uodd (s) is an imaginary matrix
while Ueven (s) is a real matrix. In addition, in the special case of “symmetric
arrays”, the coordinate vectors form a mutually orthogonal frame
symmetric arrays: UH (s)U(s) = IN ; (2.49)
since the sum of odd powers of the sensor locations is equal to zero.
2.3.2 Evaluating the Curvatures of Uniform Linear Array

Manifolds
The manifold curvatures (and hence the Cartan matrix C) of a linear array
of isotropic sensors are constant, or equivalently, independent of parameters
s or p. They depend, however, on the relative rather than the absolute
sensor spacings. For instance, the manifold of a 3-element linear array with
intersensor spacings 1.5 and 1 has the same curvatures with a 3-element
linear array with spacings 0.75 and 0.5 (where the spacings are measured
in half wavelengths) but the length of its manifold is twice as long.
Using Theorem 2.1, all the curvatures of any linear array can be cal-
culated. For instance, in Table 2.3, the first eight curvatures are tabulated
for arrays with 3 to 10 elements chosen from the popular class of uniform
linear arrays with half-wavelength spacing. Furthermore, in Fig. 2.11 the
first four curvatures are shown versus the number of elements in the array.
It is apparent from Fig. 2.11 that the curvatures are monotically decreasing
as the number of elements exceeds twice the order of the curvature.
In Fig. 2.12, the curvatures of a 13-element uniform linear array of
sensors are seen to follow a fading oscillating pattern.
It is important to point out that, because the manifold √

has constant
principal (first) curvature and lies on a sphere with radius N , a lower
limit for κ1 is
1
κ1 > √ (2.50)
N
Table 2.3 The first eight curvatures of the manifold of uniform linear arrays.
N =3 N =4 N =5 N =6 N =7 N =8 N =9 N = 10
r 1.4142 2.2361 3.1623 4.1833 5.2915 6.4807 7.7460 9.0830

κ1 0.7071 0.6403 0.5801 0.5372 0.5000 0.4693 0.4435 0.4214
κ2 0 0.1874 0.2058 0.2047 0.1984 0.1908 0.1832 0.1760
κ3 — 0.2343 0.3430 0.3603 0.3568 0.3469 0.3352 0.3234
κ4 — 0 0 0.1489 0.1783 0.1880 0.1899 0.1884
κ5 — — — 0.1323 0.2270 0.2554 0.2647 0.2660
κ6 — — — 0 0 0.1201 0.1502 0.1637
κ7 — — — — — 0.0895 0.1694 0.1002
κ8 — — — — — 0 0 0.0664
Fig. 2.11 Curvatures of uniform linear arrays as functions of the number of sensors N .
Fig. 2.12 Curvatures of uniform linear array of 13 sensors with half-wavelength

intersensor spacing.
The above relation also stems from the following inner product aH u2
in conjunction with the expression
(aH u1 ) = aH u1 + aH u1

⇒ (aH u1 ) = u1H u1 + aH u1
⇒ aH u1 = (aH u1 ) − u1H u1
That is,
1 H 1 H 1 1
aH u 2 = a u1 = (a u1 ) − u1H u1 = (0 − 1) = −
κ1 κ1 κ1 κ1
1 1
⇒ a .u2 > ⇒ κ1 > √
κ1 N
√
= N =1
Finally, it should be noted that the manifolds of non-omnidirectional lin-

ear arrays are not hyperhelices. The first curvature of the manifold of
such arrays depends both on the relative spacing and the first two deriva-
tives of each elemental pattern γk (p). In the case of the manifold of a
4-element uniform linear array with non-isotropic, but identical, sensors
with directional pattern γk (p) = cos p, the variation of the first curvature
with the azimuth direction is shown in Fig. 2.13. The constant first curva-
ture of the corresponding isotropic array is also shown in the same figure.
It is seen that there exist two bearings where the first curvature becomes
equal to zero. From the variation of the first curvature, it is possible to
deduce the geometrical object to which the manifold best fits. In the case
considered, it is apparent that the effect of the sinusoidal elemental pattern
is to deform the hyperhelix to a geometrical figure resembling an “eight”
with a double point at the origin of the coordinates of CN .
2.4 The Manifold Length and Number of Windings

(or Half Windings)
It is noted that the manifolds of uniform linear arrays with an odd number
of sensors and half-wavelength spacing consist of one winding (round) since
the manifold vectors corresponding to the manifold boundaries coincide
(i.e. a(0◦ ) = a(180◦ )). In addition, the manifolds of uniform linear arrays
with an even number of sensors and half wavelength spacing, consist of one
half winding (half round) since the manifold vectors corresponding to the
manifold boundaries are opposite (i.e. a(0◦ ) = −a(180◦ )). This situation
has forced the separation of a hyperhelix corresponding to a general array
with an odd number of sensors into consecutive equal arcs of one winding
and a hyperhelix corresponding to an array with an even number of sensors
into consecutive equal arcs of half-winding.
The following theorem is concerned with the estimation of the number
of windings (N odd) and the number of half windings (N even) existing in
the manifold of an arbitrary linear array.
Theorem 2.2 Windings of a Hyperhelix: The number of windings

(N odd), or the number of half windings (N even), of the hyperhelix of
Eq. (2.26) is
manifold length=lm

2π r
nw = (2.51)
lw
Fig. 2.13 The principal curvature as a function of azimuth (top) and its polar repre-
sentation (κ, p) (bottom) for a 4-element ULA of directional sensors — γk (p) = cos p.
where lw is the positive number that corresponds to the (N − 1)th root of

the function
ξw (s) = Tr {C expm(s C)} (2.52)
with C denoting the Cartan matrix.
The independence of the curvatures from the absolute spacings (and

hence the independence of the winding length from the absolute spacings)
in conjunction with the dependence of the manifold length on the absolute
spacings, reveals that two linear arrays with sensors at positions given by
the vectors r(1) and r(2) with r(1) = qr(2) , where q is a scalar constant,
• have manifolds that fit upon each other, and

• the associated number of windings (or half windings) satisfy the following
relationship
(1)
nw
(2)
=q (2.53)
nw
It is worth noting that the definition of half winding is restricted to
linear arrays with an even number of sensors and


a(0) = a(lw )

U(0) = U(lw )

 F(0) = Id = F(lw )

⇒ Tr(F(0)) = Tr(F(lw )) = d
where lw represents the arc length of half winding. For an array with an
odd number of sensors, the function ξw (Eq. (2.52)) does not generally have
a root at the point on the manifold that corresponds to half the length of
one manifold winding.
Similarly, the definition of one winding is restricted to linear arrays with
an odd number of sensors and


 a(0) = −a(lw )

U(0) = −U(lw )
 F(0) = Id = −F(lw )


⇒ Tr(F(0)) = −Tr(F(lw )) = d
where lw represents the arc length of one winding. For an array with an
even number of sensors, the function ξw does not generally have a root
at the point on the manifold that corresponds to twice the length of a
Fig. 2.14 Tr(F(s)) and ξw (s), given by Eq. (2.52), for a 5-element and a 6-element ULA
of half wavelength spacing. Note that A1 = 5 = A2 and B1 = 6 = −B2.
half-manifold winding. Figure 2.14 provides a representative example for

two arrays − with even and odd number of sensors.
2.5 The Concept of “Inclination” of the Manifold
In the previous section it has been shown that the shape of the manifold
of a general linear array of omnidirectional sensors is a complex circular
√
hyperhelix lying on the complex N -dimensional sphere with radius N . In
this section, an attempt is made to identify the orientation of the array
manifold (hyperhelix) on the complex N -dimensional space. Therefore, it is
necessary to establish a measure of the orientation of the manifold curve in
CN using the concept of array inclination angle. This is intuitively defined
as follows:
Definition 2.2 Inclination Angle: The inclination angle ζinc of the

manifold of a linear array is the angle formed between any manifold vector
a(s) and the subspace L [Usub (s)], i.e. the subspace spanned by the columns
of the matrix

Usub (s) u2 (s), u4 (s), u6 (s), . . . , uuptoN (s) ,
that is,
ζinc = (a(s), L [Usub (s)])
or equivalently /
1
cos ζinc = aH (s) PUsub a(s) (2.54)
N
where PUsub is the projection operator onto L [Usub (s)].
Having defined the matrix Usub (s) as a submatrix of Ueven (s) and
because of the orthonormality of columns of the matrix Ueven (s) (see
Eq. (2.47)) it is apparent that Eq. (2.54) can be rewritten as follows:
0
1
1 1
N
=2
2
cos ζinc |aH (s) ui (s)| (2.55)
N i=even
It is intuitively expected that the inclination angle should increase with

the degree of non-symmetry of the array, at least in the neighborhood
in which the array is nearly symmetrical. This can, in fact, be verified
from Fig. 2.15 in which the inclination of the manifolds for a number of
Fig. 2.15 Inclination angle for a number of linear arrays with 4 and 6 sensors.
different arrays resulting from uniform linear arrays with 4 and 6 elements,
of half-wavelength spacing, by varying the rightmost (leftmost) spacing
from 0.5 half-wavelengths to 3 half-wavelengths is shown.
2.6 The Manifold-Radii Vector
Equation (2.55) shows that the inclination angle is directly related to the
inner products of a manifold vector aH (s) with the coordinate vectors u2 ,
u4 , u6 , . . . , etc. This demonstrates that the relationship between aH (s) and
its associated coordinate vectors ui (s), ∀i is very important.
For example, the relationships between aH (s) and the first four coordi-
nate vectors, at a point, are:

aH (s)u1 (s) = j1TN )
r 




1 

a (s)u2 (s) = −
H


κ1
j & ' (2.56)
aH (s)u3 (s) = r 
r3 − κ21 1TN )
1TN ) 


κ1 κ2 

κ2 

a (s)u4 (s) = −
H 
κ1 κ3
It is easy to see that the orthogonality between a manifold vector and

the corresponding tangent vector u1 (s) at every point is guaranteed if the
phase reference is taken at the array centroid so that
r) = 0 or, equivalent, 1TN )

sum () r=0 (2.57)
If the array is also symmetric, then at every point along the manifold, the
corresponding response vector is orthogonal to the odd-indexed coordinate
vectors.
To provide a more general expression of the inner product
aH (s)ui (s); ∀i
the following definition is essential.
Definition 2.3 Manifold-Radii Vector: For an N sensor linear array

with m sensors in symmetrical pairs the vector


 [0, −R2 , 0, −R4 , 0, . . . , 0, −Rd ]T

 if sensor at the array centroid
R= (2.58)

[0, −R2 , 0, −R4 , 0, . . . , 0, −Rd−1 , 1]T


otherwise
is defined as the array manifold-radii vector having elements the manifold

radii Ri (for i = even) with
3i−2
1 κn
R2 = and Ri = 3n=even
i−1
for i > 2 (2.59)
κ1 κ
n=odd n
Lemma 2.1 The inner product aH (s)ui (s), ∀i: For a linear array of
N sensors the inner products of a manifold vector a(s) with its d coor-
dinate vectors ui (s) are constant and given by the following expressions:
aH (s)ui (s) = −Ri for i even (2.60)

H 0 for i = d and odd
Re(a (s)ui (s)) =
1 for i = d and odd
or, in a more compact form, as
aH (s)Ueven (s) = [−R2 , −R4 , . . . , −Rdeven ] (2.61)

Re aH (s)U(s) = RT
Note that the inner product of a manifold vector and the corresponding
normal vector (or second coordinate vector) is the negative of the radius
of curvature (i.e. −1/κ1 ) and hence the other interior products can be
considered as higher order “manifold radii.” Using the above lemma the
inclination angle of Eq. (2.55) can be rewritten as

Rsub
ζinc = arccos √ = constant (2.62)
N
revealing that the inclination is constant, depending only on the norm of a

sub-vector of the manifold radii vector and the number of sensors, where
Rsub is the vector formed by the first N elements of vector R.
Definition 2.4 Dual arrays: The dual array of a linear array, non-
symmetrical with respect to the array centroid, is the array with sensors
having positions given by rdual
i = −rN -i+1 for i = 1, . . . , N .
Using the above definition it can be seen that, although the manifolds
of two dual arrays are not identical, they have the same curvatures and
length. Thus, apart from the fact that they have opposite inclinations, they
have the same differential geometry properties.
Based on the above concepts the array manifold vector of Eq. (2.27) or
(2.35) can be written as
a (s) = U (s) R = U (0) F(s)R where F(s) = expm (sC) (2.63)
or, by choosing the N coordinate vectors (i.e. N columns of U(0)) at s = 0

to be the standard orthonormal basis in RN , i.e. setting U(0) = IN , and
expressing s as a function of p using Eq. (2.28),
a (p) = expm (π r (1 − cos p) C) R (2.64)
where p = 0 (s = 0) is taken along the array axis. Equation (2.64) is

associated with a real N -dimensional hyperhelix having the same length
and identical curvatures with those of the complex N -dimensional manifold
and this is formally described below.
Theorem 2.3 Real N -dimensional Hyperhelix of Symmetric

Linear Arrays: For a symmetric linear array of N sensors with loca-
tions r, manifold radii vector R and Cartan matrix C, there is a
real N -dimensional hyperhelix described as the locus of the following
real vector
areal (p) = expm(πr(1 − cos p) C)R ∈ RN (2.65)
having differential geometry properties equivalent to those of the complex

N -dimensional manifold of the array.
The vector areal (p) can be regarded as the real representation of the
manifold of a symmetric array in RN .
Thus, the manifolds of symmetrical arrays are shown to admit real rep-
resentation. This can be seen intuitively from the fact that the compo-
nents of a manifold vector of symmetrical arrays exist in conjugate pairs.
In addition, in the absence of weighting, such arrays have purely real array
patterns.
Example 2.1 3-element linear array: The real N -dimensional curve
having the same differential geometry properties as the manifold of a 3-
element uniform linear array, is a circle. In fact, for N = 3, Eq. (2.65)
of Theorem 2.3 takes a closed analytical form. The second curvature κ2
vanishes and the Cartan and Frame matrices are expressed as
 
0, −κ1 , 0
C = κ1 , 0, 0 (2.66)
0, 0, 0
 
cos(κ1 s), sin(κ1 s), 0
F(s) = expm(s C) = − sin(κ1 s), cos(κ1 s), 0 (2.67)
0, 0, 1
In addition, the vector R is
4 5T
1
R = 0, − , 1 (2.68)
κ1
By applying Eq. (2.65) of Theorem 2.3, the real curve with the same
differential geometry properties with the manifold of a 3-element uniform
linear array, is given by the real vector continuum
4 5T
1 1
areal (s) = − sin (κ1 s) , cos(κ1 s), 1 (2.69)
κ1 κ1
or, as a function of p,
4 5T
1 1
areal (p) = − sin (κ1 π r (1 − cos p)) , cos (κ1 π r (1 − cos p)) , 1
κ1 κ1
while the conventional complex array manifold vector of this array is

T
a(p) = exp (jπ r cos p) , 1, exp (−jπ r cos p) (2.70)
It is clear that Eq. (2.69) is the equation of a closed circle (assuming half-
wavelength spacing) with radius 1/κ1 for −rπ ≤ s ≤ rπ. If the two
spacings of the above array become unequal, κ2 does not vanish and the
manifold becomes a helix.
Example 2.2 MUSIC Algorithm: To test the proposed results and

expressions as well as to demonstrate their use in a DF problem, the MUSIC
algorithm is employed, expressed in terms of the curvatures and coordinate
vectors. Indeed, if Pn is the projection operator on to the noise subspace
spanned by the noise-level eigenvectors of the array covariance matrix, the
cost function of MUSIC can be expressed as a function of the differential
geometry parameters of the array manifold as follows:
ξMUSIC (s) = aH (s)Pn a(s)
= RT FT (s)UH (0)Pn U(0)F(s)R

=aH (s) a(s)
 T

v(s) =v (s)
 H 
= Tr 
U (0)P U(0) F(s)R RT F(s)

n

=Pn =V(s)

= Tr Pn V(s) (2.71)
The matrix Pn is a constant d × d Hermitian matrix corresponding

to a transformation of the second-order statistics of the data collected
and is termed the orientation of the noise subspace. The matrix V(s) is
a real d × d symmetric matrix depending exclusively on the curvatures
and acts as a differential geometry operator on the suitably transformed
data. If the array is symmetric it is apparent that the same expression
holds but the matrices V(s) and Pn are N × N dimensional real matri-
ces and v(s) = areal (s). Hence, there is a significant reduction in the
dimensionality.
In Fig. 2.16, the differential geometry version of MUSIC is simulated
for a seven element uniform linear array which operates in the presence of
signals with directions of arrival 35◦ , 120◦ , and 125◦ . This figure illustrates
that the proper estimation of the three sources is possible by expressing the
Fig. 2.16 MUSIC spectrum based on Eq. (2.71) for a ULA of 7 sensors operating in the
presence of 3 emitting sources.
MUSIC cost function in terms of the differential geometry properties of the

manifold.
Deductions
√
(1) Setting a(s) = N and using the wide sense orthonormality of the
coordinate vectors, the following relation is derived
√
expm(s C)R = N (2.72)
By setting s = 0, it is obvious that

√
R = N (2.73)
(2) If Rsub is the vector formed by the first N components of the manifold
radii vector, then
√
Rsub ≤ R = N (2.74)
However, for symmetrical arrays, Rsub = R, therefore,

√
symmetric arrays: Rsub = N (2.75)
From the above it can be deduced that, for any symmetric sequence of
N real numbers ri , Eq. (2.65) gives an expression
√ for a real hyperhelix
(on a real N -dimensional sphere with radius N ) of latitude


 π − arccos √1 N odd
latitude = 2 N

0 N even
and with curvatures given by Eq. (2.39), Theorem 2.1.

(3) All 3-element non-symmetrical linear arrays have a constant inclination
angle of 35.26◦ . Also, the first curvature of all 3-element linear arrays
is 0.707 so√that, in the non-uniform case, i.e. non-symmetrical case,
Rsub = 2.
(4) The manifold of arrays with symmetry with respect to their centroid do
not exhibit inclination, which is equivalent to saying that these mani-
folds “stand upright” or, in other words, that they admit a representa-
tion entirely in RN .
Finally the Frobenius norm of the Cartan matrix (i.e. CF ) is related
to the array symmetricity as follows:
(i) CF = √ 1 if array = symmetric

(ii) CF = 2 √ if array = fully asymmetric (2.76)
(iii) 1 < CF < 2 if array = partially symmetric
and, at this point, it is convenient to summarize the special properties of

the manifolds of symmetric linear arrays:
Symmetric Linear Arrays of N Sensors
• sum(rn ) = 0, for n odd

• m = N (with d = 2N − m = N )

N − 1 if sensor at the array centroid
• (No. of κi = 0) =
N − 2 otherwise
◦
• inclination angle: ζinc = 0
• UH (s)U(s) = IN (narrow sense orthogonality)
• aH (s)U(s) = RT
• Rsub = R
• areal (p) = expm(π r (1 − cos p) C) R ∈ RN
(2.77)
2.7 Appendices
2.7.1 Proof of Eq. (2.24)

The proof will be done by induction. For i = 2, we have
a (s)H u2 (s) = κ1 u2 (s)H u2 (s) = κ1 (s) (2.78)

a (s)H
Assuming that

Re a(i−1) (s)H ui−1 (s) = κ1 (s)κ2 (s) · · · κi−2 (s) (2.79)
we have to prove that

Re a(i) (s)H ui (s) = κ1 (s)κ2 (s) · · · κi−1 (s) (2.80)
For convenience let us drop the parameter s. The argument of the first part
of Eq. (2.80) can be written as follows:
(a(i) )H ui
= (a(i−1) )H ui
& '
= (a(i−1) )H ui − (a(i−1) )H ui
& '
= (a(i−1) )H ui − (a(i−1) )H (κi ui+1 − κi−1 ui−1 )

=ui (see Eq. (2.12))
& '
= (a(i−1) )H ui − κi (a(i−1) )H ui+1 + κi−1 (a(i−1) )H ui−1

=κ1 κ2 ···κi−2
(2.81)
However, by using the fact that the vectors a(i−1) can be expressed as a
linear combination of coordinate vectors with no higher order than (i − 1),
Eq. (2.81) can be rewritten as follows:
 
 H   H
 i−1 
i−1
 
(2.81) =  c j u j  u i  − κi  cj uj  ui+1 + κ1 κ2 · · · κi−2 κi−1
 
 j=1  j=1

imaginary imaginary
(2.82)
where cj are scalar coefficients. Then, by taking the real part of both sides
and using the property that
j ui } = 0
Re{uH for j = i (2.83)
we have & 'H .
Re a(i) ui = 0 + 0 + κ1 κ2 · · · κi−2 κi−1 (2.84)
and this completes the proof.
2.7.2 Proof of Theorem 2.1

Consider a manifold vector described by the following equation:
a a(p) = exp(−j(πr cos p + v)) (2.85)
The magnitude of the manifold tangent vector is
ȧ(p) = jπ sin p r a = π sin pr = ṡ(p) (2.86)
Hence the length of the manifold is

π
lm s(π) = ȧ(p) dp = 2π r (2.87)
0
The first coordinate vector is merely the tangent given by

da da dp
u1 (s) = a (s) = =
ds dp ds
−1
= (jπ sin pr a) · (π sin p r)

=ȧ(p) =ṡ(p)
= j r̃ a (2.88)
It can be seen that, since the origin of the coordinates is at the array
centroid,
aH u1 (s) = 0 (2.89)
However, the derivative of u1 (s) with respect to arc length s is

du1 dp
u1 (s) =
dp ds
−1
= (j r̃ ȧ) · (π sin pr)

=u̇1 =ṡ(p)
1
= j r̃ (jπ sin pr a) ·
π sin pr
ȧ
= −r̃ a
2
(2.90)
Therefore, by using Table 2.2, the first curvature is estimated as
κ1 = u1 (s) = r̃2 (2.91)
Next, the second coordinate vector is chosen to be
u1 (s)
u2 (s) =
u1 (s)
1 2
=− r̃ a (2.92)
κ1
Furthermore,
du2 dp
u2 (s) =
dp ds

1 −1
= − r̃2 ȧ (π sin p r)
κ1
=ṡ(p)
=u̇2
1 2 1
=− r̃ (jπ sin pr a) ·
κ1 π sin pr
ȧ
1 3
= −j r̃ a (2.93)
κ1
and

1 3
u2 (s) + κ1 u1 (s) = −j r̃ − κ1 r̃ a (2.94)
κ1
Hence, the second curvature (see Table 2.2) and the third coordinate
vector are
κ2 = u2 (s) + κ1 u1 (s)

1
= . r̃3 − κ21 r̃ (2.95)
κ1
and
1
u3 (s) = (u (s) + κ1 u1 (s))
κ2 2

1 1 3
= −j · r̃ − κ1 r̃ a
κ2 κ1
1
= −j · (r̃3 − κ21 r̃) a (2.96)
κ1 κ2
In order to estimate the third curvature it is necessary to estimate the
terms u3 (s) and κ2 u2 (s). Indeed,
du3 dp
u3 (s) =
dp ds
1 −1
= −j · r̃3 − κ21 r̃ ȧ · (π sin pr)
κ1 κ2
=ṡ(p)
=u̇3
1
= −j · (r̃3 − κ21 r̃) (jπ sin pr a)(π sin pr)−1
κ1 κ2
=ȧ
1
= · (r̃4 − κ21 r̃2 ) a (2.97)
κ1 κ2
and
κ2 2
κ2 u2 (s) = − r̃ a (2.98)
κ1
which implies
1 4
u3 (s) + κ2 u2 (s) = r̃ − (κ21 + κ22 )r̃2 a (2.99)
κ1 κ2
By taking the magnitude of the above equation the third curvature is
estimated as
κ3 = u3 (s) + κ2 u2 (s)

1
= r̃4 − (κ21 + κ22 )r̃2 (2.100)
κ1 κ2
and, thus, the fourth coordinate vector is
1
u4 (s) = (u (s) + κ2 u2 (s))
κ3 3
1 4
= r̃ − (κ21 + κ22 )r̃2 a (2.101)
κ1 κ2 κ3
By continuing in the same way as above, it can be shown that
(i) the fourth curvature, for N > 4, is
κ4 = u4 (s) + κ3 u3 (s)

r̃5
=
2 2 2 3
−(κ1 + κ2 + κ3 )r̃
(2.102)
κ1 κ2 κ3
+(κ1 κ3 )2 r̃
so that the fifth coordinate vector is

1
u5 (s) = (u (s) + κ3 u3 (s))
κ4 4
 
r̃5
1  −(κ21 + κ22 + κ23 )r̃3  a
=j (2.103)
κ1 κ2 κ3 κ4
+(κ1 κ3 )2 r̃
(ii) the fifth curvature, for N > 5, is
κ5 = u5 (s) + κ4 u4 (s)

r̃6

−(κ
1 2 2 2 2
1 + κ2 + κ3 + κ4
)r̃
= (2.104)
κ1 κ2 κ3 κ4
2

+ (κ1 κ3 ) + (κ1 κ4 )
2

2 r̃2
+(κ2 κ4 )
so that the sixth coordinate vector is

1
u6 (s) = (u (s) + κ4 u4 (s))
κ5 5
 
r̃6
1  −(κ21 + κ22 + κ23 + κ24 )r̃4 
=−  a
κ1 κ2 κ3 κ4 κ5  (κ1 κ3 )2 + (κ1 κ4 )2 2

+ 2 r̃
+(κ2 κ4 )
(2.105)
The above equations can be generalized for any index i, to give the
curvature κi and the vectors ui by the following expressions:

i−1
i−3
i−1
1
i+1 i−1 i−3
κi =
)
r −)
r · κn + )
2
r · (κn κ ) − · · ·
2
κ1 κ2 · · · κi−1
n=1 n=1 =n+2

(2.106)
or, equivalently,

i

fix( 2 )+1
1

κi =
(−1)n−1
bi,n )
r i−2n+3
(2.107)
κ1 κ2 · · · κi−1

n=1
and
fix( i−1
(j)i 2 )+1
ui (s) = ri−2n+2 a(s)

(−1)n−1 bi−1,n ) (2.108)
κ1 κ2 · · · κi−1 n=1
where bi,n is given by Eq. (2.40).

From Eq. (2.106), it can be seen that the terms inside the norm operator
form a polynomial of the sensor locations with coefficients given by the cur-
vatures. This polynomial is expressed in a more compact form in Eq. (2.39).
Furthermore, note that the numbers bi,n satisfying the recursive Eq. (2.41)
can be easily proven by induction.
Chapter 3
Differential Geometry of Array

Manifold Surfaces
The locus of all array manifold vectors a(p, q) ∀(p, q) forms a surface known
as the array manifold. This is a two-parameter manifold embedded in an
N -dimensional complex space and is formally defined as
M = {a(p, q) ∈ CN , ∀(p, q) : p, q ∈ Ω} (3.1)
where Ω denotes the parameter space. The shape of the surface is very
important. In the same way that the shape of a curve A is uniquely
defined by its curvatures, the shape of a two-parameter manifold (surface)
may be quantitatively expressed in terms of intrinsic geometrical parame-
ters such as the Gaussian curvature KG (p, q) of the surface (see Fig. 3.1)
and the geodesic curvature κg (p, q) of the curves lying on the surface (see
Fig. 3.2).
More specifically, for a point a a(p, q) on the manifold surface, the
parameters of interest, intrinsic to the manifold surface, together with neces-
sary building blocks of surface differential geometry, will be identified in this
chapter and extended from three-dimensional real space to N -dimensional
complex space. In order to achieve this all the ramifications and subtleties
involved in this process have been taken into consideration.
Thus, it is first necessary to introduce the building blocks of surface
differential geometry such as the manifold metric, first fundamental coeffi-
cients and the Christoffel symbols.
However, to proceed, it is essential to start with two definitions which
will aid in the development of the basic theory. These are
• the definition of a constant parameter curve on M,

• the “regularity” condition by defining the array manifold (surface) as a
“vector” function.
59
Fig. 3.1 The Gaussian curvature KG (p, q) provides information about the local shape
of the surface M in the neighbourhood of a(p, q).
Fig. 3.2 The geodesic curvature κg provides information about a curve on a surface.
Definition 3.1 A constant-parameter curve is defined as the curve that

joins all those points on the manifold surface M corresponding to a constant
value of one of the two parameters p, q.
Differential Geometry of Array Manifold Surfaces 61
Definition 3.2 A regular parametric representation of a surface M is

defined as a vector function
a = a(p, q) (3.2)
of parameters p and q in the parameter space Ω, if and only if ∀p, q ∈ Ω, the

derivatives ȧp = da/dp and ȧq = da/dq exist, are continuous and non-zero
with
rank{[ȧp , ȧq ]} = 2 (3.3)
The vectors ȧp and ȧq at a specific point A on the manifold represent
the tangent vectors to the p- and q-parameter curves respectively passing
through A and also form a basis for the tangent plane to the surface at that
point, i.e.
Tangent plane = L [T] where T = [ȧp , ȧq ] ∈ CN ×2 (3.4)
The regularity condition of Eq. (3.3) ensures that a tangent plane exists at
all points on the surface. Note that the basis created by T is, in general,
not orthonormal.
3.1 Manifold Metric
Let a = a(p, q) be the regular parametric representation of a surface M

embedded in an N -dimensional complex space.
Then the (2 × 2) real semipositive definite symmetric matrix G,
defined as
 
ȧ 2
, Re ȧ H
ȧ 4 5
H p p q gpp , gpq
G Re T T = 

2  = (3.5)
Re ȧH ȧ ,
q p

ȧ
q
gqp , gqq
is said to be the manifold metric. The elements gpp , gqq , and gpq (note
gqp = gpq ) of G are known, in Differential Geometry terms, as the first
fundamental coefficients (or metric coefficients) expressing the magnitudes
and inner products of the parameter-curve tangent vectors and entirely
describing the manifold properties of the surface. For instance, the angle
between the two parameter curves of the surface at a point (p, q) can be
expressed as

gpq
ȧp , ȧq = arccos √ (3.6)
gpp gqq
The metric coefficients provide a way of measuring trajectories on non-

Euclidean spaces through the use of weighting coefficients. For example, an
infinitesimal distance on the Euclidean plane is measured by
ds2 = dx2 + dy 2 (3.7)
whereas on a general non-Euclidean surface it is given by
ds2 = gpp dp2 + 2gpq dpdq + gqq dq 2 (3.8)
It is also customary to use superscripts to denote the elements of the

inverse matrix G−1
4 pp 5
g g pq
G−1 = qp (3.9)
g g qq
where clearly
gqq gpp −gpq
g pp = , g qq = , g pq = g qp = (3.10)
det(G) det(G) det(G)
while the following properties may be easily proven:
(i) gpp , gqq ≥ 0

(ii) det(G) ≥ 0 (by Schwarz’s inequality) (3.11)
(iii) gpp , gpq and gqq are variant under parameter transformation
Note that the metric G is, by construction, a semipositive definite

matrix, i.e.
xT Gx 0, ∀x ∈ R2 (3.12)
which implies that at some points on the surface M this matrix can possibly
be singular in order to account for the fact that there may not exist a
unique tangent space (L[T]) at each point. That is, at a point (p, q) with
a non-unique tangent space, the matrix G is singular. These points are the
boundaries of the manifold surface.
3.2 The First Fundamental Form
Again, let a = a(p, q) be the regular parametric representation of a manifold

surface M embedded in N -dimensional complex space. Then the differential
mapping da = ȧp dp + ȧq dq maps increments dp and dq on the (p, q)
parameter plane on to vector ȧp dp + ȧq dq on the tangent plane at a(p, q).
The scalar I, defined as

2
I = da =
ȧp dp + ȧq dq
2
(3.13)
4 5
dp
= dp dq G = dpT Gdp (3.14)
dq
with T
dp = dp, dq
is known as the first fundamental form and essentially represents the dis-
tance between two neighboring points a(p, q) and a(p + dp, q + dq) on the
manifold surface.
Two properties of I which may be readily established are
(i) I ≥ 0 with equality when dp = dq = 0

(3.15)
(ii) I is invariant under a parameter transformation
The first fundamental form I is essential for the evaluation of lengths

of curves, and areas, on the manifold surface. For instance, let a(t)
a(p(t), q(t)) with t1 ≤ t ≤ t2 be a regular curve lying on a surface
a = a(p, q). Then clearly the length of the curve is given by

t2 /
t2
da
I
length of curve =

dt = dt (3.16)

dt
dt2
t1 t1
It may be similarly shown that the area of a segment W on the surface of

a manifold a = a(p, q) is given by
6
area of segment = det(G) dp dq (3.17)
W
The significance of Eqs. (3.16) and (3.17) in relation to the intrinsic

geometry of a surface M will be6 clarified later on in this chapter. However,
it is clear from Eq. (3.17) that det(G) dp dq describes an infinitesimally
small area on the manifold surface which is a function of (p, q) and the
array geometry r. Small uncertainties, or variations, in the sensor posi-
tions
6 will invariably result in small variations in the infinitesimal6 area
det(G) dp dq on the manifold at point (p, q). The parameter det(G)
can hence serve as a tool for detecting the changing shape of the manifold
surface [18].
3.3 Christoffel Symbol Matrices
In order to investigate the differential geometry of a surface a(p, q), it is

necessary to determine how the non-orthonormal basis T of the tangent
plane varies from point to point on the array manifold. This information
can be expressed in terms of the Christoffel symbols of first kind.
By using the matrix T = [ȧp , ȧq ] and its two derivatives Ṫp and Ṫq with
Ṫζ = dT/dζ = [äpζ , äqζ ] with ζ = p or q (3.18)
the Christoffel symbols of first kind represent the inner products between
the tangent vectors ȧp , ȧq (columns of T) and their derivatives (columns of
Ṫp and Ṫq ), and are defined as

Γi,jk Re ȧH i äjk with i, j, k = p or q (3.19)
Thus there are eight Christoffel symbols of first kind forming the two
Christoffel matrices of the first kind Γ1p and Γ1q defined as follows:
4Γ Γp,qζ
5
p,pζ
Γ1ζ Re T Ṫζ =
H
with ζ = p or q (3.20)
Γq,pζ Γq,qζ
where, due to the symmetry inherent in differentiation,
Γp,qp = Γp,pq and Γq,qp = Γq,pq . (3.21)
Although the differential geometry of a surface can be completely described

in terms of the above symbols, the process may be considerably simplified
by the use of two other matrices Γ2p and Γ2q which are related to the
symbols of the first kind Γ1p and Γ1q , respectively, in the following way:
7 8
−1 Γppζ Γpqζ
Γ2ζ G Γ1ζ = q with ζ = p or q (3.22)
Γpζ Γqqζ
These matrices are known as the Christoffel symbol matrices of second

kind and are used as a transformation which provides the infinitesimal vari-
ation of the tangent plane as a point moves on the manifold surface, namely
dT = T(Γ2p dp + Γ2q dq) (3.23)
Once again, due to symmetry,
Γppq = Γpqp and Γqpq = Γqqp (3.24)

It is obvious that all Christoffel symbols are functions only of the first
fundamental coefficients gij (i, j = p, q) and their derivatives. Thus, for
this reason they play a central role in expressing the intrinsic geometry of
a surface, which is described in the next section.
3.4 Intrinsic Geometry of a Surface
Let us assume that
At {a(t) ∈ CN , ∀t : t1 < t < t2 } (3.25)

where
a(t) a(p(t), q(t)) (3.26)
denotes an arbitrary regular curve on a surface M
M = {a(p, q) ∈ CN , ∀(p, q) : p, q ∈ Ω} (3.27)
where Ω denotes the parameter space.

In order to define the intrinsic geometry of a surface, it is first necessary
to appreciate the concept of an isometric mapping.
Definition 3.3 Isometry: A one-to-one mapping of a surface M on to
another surface M̆ is called an isometric mapping, or isometry, if the length
of an arbitrary curve At on M is equal to the length of its image Ăt on M̆,

Ăt {ă(t) ∈ CK , ∀t : t1 < t < t2 }
where
M̆ {ă(p, q) ∈ CK , ∀(p, q) : p, q ∈ Ω}
Hence we may think of isometry as the bending of a surface into a

different shape without changing the distance (along the surface) between
any of its points. Consequently, the “inhabitants” of such a surface would
not be aware of any change at all, as their geometric measurements remain
exactly the same. The combination of the above definition with Eqs. (3.16)
and (3.17) implies that:
Corollary 3.1 A one-to-one mapping of a surface M on to another
surface M̆ is an isometry if and only if at corresponding points a(p, q) and
ă(p, q) the first fundamental coefficients remain unchanged, i.e. gij = ğij
∀i, j = p, q.
Thus, there is no difference in the measurement of lengths, angles, and
areas on isometric surfaces although the surfaces, when viewed from the
embedding space, may have entirely different geometric shapes. An iso-

metric mapping is by necessity both angle-preserving (conformal)
and area-preserving (equiareal).
Definition 3.4 Intrinsic Property and Intrinsic Geometry: A
property of a surface which remains invariant under an isometry is
called an intrinsic property of the surface. The totality of the intrin-
sic properties of a surface is known as the intrinsic geometry of the
surface.
Naturally a property of a surface is an intrinsic property if it only
depends on the first fundamental coefficients (and hence the magnitudes
and inner products of the tangents to the surface). The intrinsic geometry
of a surface is completely independent of the space in which the surface is
embedded and is built from solely isometric invariants.
Next, two important features of intrinsic geometry are presented,
namely:
• Gaussian curvature KG , and,
• geodesic curvature κg .
3.4.1 Gaussian Curvature

By considering the surface as a whole body with intrinsic properties, Gauss,
according to his “Theorema Egregium”, assigned to every point on the
surface a(p, q), independently of any specific curve passing through it, a
real number KG called Gaussian curvature. Research into surfaces has
produced several formulas for KG . For example, the Gaussian curvature
for surfaces embedded in the 3-dimensional Euclidian space R3 is given
below (see [2]).
& ' & ' d 2 a T
äp T · äq T −
T T
T
dp dq
KG (p, q) = (3.28)
det(G)
where T T (p, q) is the unique normal to the surface at point (p, q).
As the array manifold is embedded in CN (and not in R3 ) a unique
normal does not exist to each point and, therefore, Eq. (3.28) cannot be
defined for array manifolds. Here we will only consider the intrinsic for-
mula for the Gaussian curvature which is independent of the normal space.
For instance, an intrinsic formula for the Gaussian curvature, KG (p, q),
of a surface, can be shown to be of the form [17] (also known as the
curvature scalar ):
 2 
1 d gpq d2 gpp d2 gqq
 2 2 − 2
− 
1  dp dq dq
7 dp2 8 
KG (p, q) =  * ij Γi,pq Γi,pp  (3.29)
det(G)  
+ g det
i,j=p,q Γj,qq Γj,qp
The above expression may be more compactly represented in terms of the
Christoffel symbols:
 
d d
Γp,qq − Γp,qp
 dp dq 
1  
KG (p, q) =  + col2 {Γ2p }T col1 {Γ1q }  (3.30)
det (G)  
T
− col2 {Γ2q } col1 {Γ1p }
where coli {matrix} denotes the ith column of the matrix.
Here the following intrinsic formula is used, adapted from [19], and
expressed as a function of the metric G as follows:
 √ √ 
det(G) q det(G) q
 d gpp Γpq d gpp Γ pp 
1   (3.31)
KG (p, q) = − 6  − 
det(G) dp dq
where Γqpq , Γqpp are elements of the 2 × 2 real matrices Γ2p and Γ2q .
The sign of the Gaussian curvature of Eq. (3.31) or (3.29), provides
an indication of the local shape of the surface in that neighborhood.
For instance, the surface around a point (p, q) is locally:
• elliptic, if KG (p, q) > 0 (the whole neighborhood of the surface at the
point considered lies on one side of the tangent hyperplane),
• hyperbolic, if KG (p, q) < 0 (one part of the surface at the point considered
lies on one side of the tangent hyperplane and the other part on the other
side),
• either parabolic or planar (i.e. flat), if KG (p, q) = 0 (there is a straight
line of the surface, lying totally on the tangent space).
Example 3.1 The following lists a number of surfaces in R3 and their
respective Gaussian curvatures:
(a) Sphere of radius R: KG = 1/R2 ⇒ Surface is elliptic at every point.

(b) Cylinder or Cone: KG = 0 ⇒ Surface is parabolic at every point.
(c) 2-D plane: KG = 0 ⇒ Surface is flat at every point.
If the Gaussian curvature at a point on the manifold is a positive num-

ber, the local neighborhood of that point6 is equivalent to the local neigh-
borhood of a point on a sphere of radius 1/KG . Also note that surfaces
which have a constant Gaussian curvature at every point (e.g. Parts (b) and
(c) of Example 3.1 above — a zero Gaussian curvature at every point) are
said to be developable surfaces and will be discussed later in Section 3.5.
Apart from defining the local shape of a surface, the Gaussian curva-
ture has important implications with regards to isometric mappings. For
example,
two surfaces which are related by an isometric mapping must

have the same Gaussian curvature at corresponding points
(since KG is an intrinsic property)
The converse of this is in general not true. However if two surfaces have
the same constant Gaussian curvature, then any two sufficiently small
neighborhoods of the surfaces are related by an isometric mapping. In
particular
• if KG (p, q) = constant, ∀(p, q) then, using Minding’s theorem [19] which

states that two surfaces of the same constant curvature are locally
isometric, we have the following cases:
(1) if KG = 0, the surface M is isometric with the plane,
(2) if KG > 0, the surface M is isometric with a sphere of radius
√
1/ KG , and
(3) if KG < 0, the surface M is isometric with a surface of revolution,
√
called a pseudo-sphere, determined by the value 1/ −KG .
As a result, developable surfaces for which KG = 0 are the only surfaces

that can be mapped isometrically onto a plane. This feature will be used
later in Section 3.5.
3.4.2 Curves on a Manifold Surface: Geodesic Curvature

The Gaussian curvature, discussed in the previous section, and the geodesic
curvatures are the two most important parameters of the intrinsic geometry
of a surface. This means that these two parameters remain invariant under
an “isometry” (see Corollary 3.1, page 65).
In this section we will focus our attention on geodesic curvature κg (s)
which is the final differential geometry parameter of interest to be presented
in this study, and is associated with curves lying on the manifold surface.
However, initially the arc length of a curve on the manifold surface should
be defined and then, in order to understand the significance of κg (s), the

geodecity concept should be introduced.
3.4.2.1 Arc Length
The arc length s can be interpreted in physical terms as the total distance
covered by a person travelling along a certain path on a certain landscape.
For a curve At
At {a(t) ∈ CN , ∀t : t ∈ Ωt } (3.32)
where
a(t) a(p(t), q(t))
on the array manifold surface M, the arc length s is the total distance
travelled along the curve At from an “initial” point (p(t1 ), q(t1 )) to the
“current” point (p(t2 ), q(t2 )) under consideration. Its expression as a func-
tion of the manifold metric is, based on Eq. (3.16):
9
t2
dp T dp
arc length : s= G dt (3.33)
(curve on a surface) t1 dt dt
where
4 5T
T
dp dp dq
p = [p(t), q(t)] and, hence, = , (3.34)
dt dt dt
3.4.2.2 The Concept of Geodicity

It is well known that the curve with the minimum length between two
points in an Euclidian space is a straight line. This concept is extended
to a curve connecting two points on a surface and the curve of minimum
length belonging to the surface is called a geodesic curve (equivalent to a
straight line in a plane). By considering a curve on a surface connecting
two points, the closeness of this curve to a geodesic curve can be assessed
by means of the geodesic curvature, κg , with κg = 0 corresponding to a
geodesic curve. That is,
geodesic curve ⇐⇒ a curve with κg = 0 (3.35)
3.4.3 Geodesic Curvature

We have seen that geodesic curves on an arbitrary geometric surface gen-
eralize the notion of a straight line in Euclidean geometry. It is also known
that a straight line a(t) = x + ty is characterized by zero curvature or
acceleration κ1 = a = 0. Following this line of reasoning, the intrinsic

acceleration, or geodesic curvature, of a curve is defined as follows:
Definition 3.5 Geodesic Curvature: The geodesic curvature κg of a

curve a(t) on a surface M is equal to the component of the curve’s first
curvature κ1 along the tangent plane to the surface at every point along
the curve.
Clearly the geodesic curvature is a function of the first fundamental

coefficients only and is consequently preserved under an isometric map-
ping. This implies that an isometric mapping is by necessity also a
geodesic mapping.
Following the above definition, we reach the same conclusion as that of
Eq. (3.35), i.e.
a geodesic curve on a surface M may be defined as a curve

whose geodesic curvature is zero at every point along its
length
For example, since for a geodesic curve the direction of acceleration is

always orthogonal to the surface, the inhabitants of a surface perceive no
acceleration at all — for them the geodesic is a “straight” line.
It can also be shown that in the neighborhood of a point A on a surface,
there exists a unique geodesic through A in any given direction. Conse-
quently a unique geodesic joins point A with every point B in its neighbor-
hood. Furthermore, the geodesic defines a unique arc of minimum length
between neighboring points A and B. The converse is also true in that if
a(t) is an arc of minimum length between any two points on a surface, a(t)
is a geodesic.
The geodesic curvature can be estimated (for proof, see Appendix 3.7.1)
by the following expression:
4 5
6 0 1
κg (s) = det (G) p (p + (Γ2p p + Γ2q q ) p )
T
(3.36)
−1 0
where
p = [p(s), q(s)]T
where s is the arc length of the curve under consideration. From Eq. (3.36),
a geodesic curve, i.e. a curve for which the geodesic curvature is invariably
zero, must satisfy the equation:
p + (Γ2p p + Γ2q q ) p = 02 ∀(p, q) (3.37)
Equation (3.37) is a system of two simultaneous equations which are the

well-known geodesic differential equations (page 64 [19]).
It is clear that the computation of both the Gaussian and geodesic
curvature parameters (see Eqs. (3.31) and (3.36)), involves the elements of
the Christoffel Symbol matrices of second kind Γ2p and Γ2q .
It is also clear that the general expression for the geodesic curvature
κg of an arbitrary curve on a surface is rather involved but can be con-
siderably simplified for the special case of constant-parameter curves [2]:
Let us denote a “p-parameter curve” (or simply p-curve) on the surface M
corresponding to a constant value of q = qo as
Ap|qo = {a(p, qo ) ∈ CN , ∀p : p ∈ Ωp , qo = constant} (3.38)
and, in a similar fashion, a “q-parameter curve” (or simply q-curve) as
Aq|po = {a(po , q) ∈ CN , ∀q : q ∈ Ωq , po = constant} (3.39)
Different values of qo (or po , accordingly) generate a family of curves having

the same properties covering the whole of the manifold surface M. Thus,
there are two such families of curves providing two alternative ways of
treating the manifold surface. These are
• the family of p-parameter curves

M = Ap|qo , ∀qo : qo ∈ Ωq (3.40)
• the family of q-parameter curves

M = Aq|po , ∀po : po ∈ Ωp (3.41)
Both families of p- and q-curves can be used to describe the manifold

surface M but different parametrizations of the surface provide families of
curves with different properties.
Note that if two families of curves are orthogonal and one family consists
of geodesic curves, then these two families constitute a set of what is known
as “geodesic coordinates.”
Based on the above definitions and using Eq. (3.36), the geodesic cur-
vature κg,p for p-curves (i.e. curves of constant parameter q = qo ) of an
√
array manifold can be shown, using dq/ds = 0 and dp/ds = 1/ gpp , to be
equal to
9
det (G) qo
κg,p κg (p, qo ) = 3
Γpp (3.42)
gpp
Similarly, for the q-curves (i.e. curves of constant parameter p = po ), using

√
dp/ds = 0 and dq/ds = 1/ gqq , the geodesic curvature κg,q can be found
to be
9
det (G) po
κg,q κg (po , q) = − 3
Γqq (3.43)
gqq
Thus an alternative approach is produced by treating the manifold

surface as a family of curves which fully covers and describes the corre-
sponding surface.
In this book (Chapters 4 and 5), two different array manifold
parametrizations, having significant differences, will be studied. Each
parametrization treats the same array manifold M, using two different
families of curves with diverse properties. These are
• the (θ, φ) parametrization (or, azimuth-elevation parametrization),
• the (α, β) parametrization (or, cone-angle parametrization).
3.5 The Concept of “Development”
In Section 3.4.1, we have seen that by examining the sign of the Gaussian
curvature KG of the array manifold it is possible to study its shape. It is
recognized that apart from the fact that the Gaussian curvature offers an
indication about the manifold shape, its importance is enhanced by the fact
that if it satisfies certain stringent conditions (valid for a very broad class
of arrays) the geodesic mappings of the individual manifold curves result in
a consistent mapping of the whole manifold on the parameter plane, called
the development of the manifold (see Fig. 3.3), which is defined as follows:
Definition 3.6 Development: Let At be a differentiable curve lying on
the manifold M expressed in terms of a parameter t as follows:
At = {a(t) a(p(t), q(t)) ∈ CN , ∀ζ : t ∈ Ωt } (3.44)
Furthermore, let s(t) be the arc length along At given by Eq. (3.33). The
development of At on the plane R2 is the plane curve
Ad = {ad (t) ∈ R2 , ∀t : t ∈ Ωζ } (3.45)

Fig. 3.3 The concept of “development.”
with curvature κd (t) and arc length sd (t) given by

κd (t) = κg (t) and sd (t) = κ−1 d (t) dt (3.46)
t
where
κg (t) κg ((p(t), q(t))) = geodesic curvature of At
Thus, the curvature κd (t) of the development of a manifold

curve At is the geodesic curvature of that curve. The motivation
behind the conservation of the geodesic curvatures is to maintain the char-
acteristic that the shortest path between two points on the surface maps
to a straight line on the plane (i.e. the shortest path on the development),
otherwise the development will represent a “distorted” surface.
The relation between the Gaussian curvature and the “development”
is addressed by Beltrami’s theorem [20, 21]. This theorem states that if a
surface is mapped to another surface and the mapping is such that geodesic
curves in one surface are mapped onto geodesic curves in the other sur-
face (geodesic mapping) then if one surface has a constant Gaussian cur-
vature the other must also have a constant Gaussian curvature, in order
to have a distortionless mapping. Surfaces of constant KG are known as
“developable” surfaces [17]. Thus the condition for the development1 to
exist is KG to be constant ∀(p, q), since a manifold surface M with a vary-
ing Gaussian curvature KG cannot be adequately mapped onto the plane.
3.6 Summary
In this chapter regularly parametrized differential surfaces, embedded in

CN , which may represent the locus of array manifold vectors, have been
studied in terms of two generic parameters p and q. Thus for a point a(p, q)
on a manifold surface M in CN the main intrinsic geometry parameters
have been presented with main emphasis given to
• the Gaussian curvature KG of M, and,
• the geodesic curvature κg of a curve A lying on M.
Then a manifold surface was treated as a family of curves on the sur-
face M. This treatment is very convenient as it permits a unified framework
for the analysis of the linear and non-linear array manifolds. Furthermore,
to provide a simplified representation of the analysis, with many potential
benefits, the class of developable surfaces has been identified and the con-
cept of the development has been introduced. This concept will be employed
in the next chapter to map the array manifold surface, embedded in a com-
plex N -dimensional space, on to a real parameter plane whilst preserving its
main intrinsic geometry properties and characteristics. This has potential
to simplify the analysis of array manifold surfaces.
3.7 Appendices
3.7.1 Proof of Eq. (3.36) — Geodesic Curvature

Consider the locus of the manifold vectors a(p, q), ∀(p, q), forming the sur-
face M embedded in CN . Consider a curve A
A {a(s) ∈ CN , ∀s : s ∈ Ωs } (3.47)
1 Note that the “development” is on a plane and planes have constant (equal to zero)
Gaussian curvature.
where
a(s) a(p(s), q(s))
lying on M and let us define p(s) [p(s), q(s)]T where s is the arc length
of the curve. Then the tangent vector at a point (p(s), q(s)) on the curve
A is given as follows:
da(s)
a ≡ = T p
ds
where T = [ȧp , ȧq ] ∈ CN ×2 such that L[T] is the tangent space to
the manifold surface at (p(s), q(s)). Furthermore, let us define the matrix
B = [b1 , b2 ] ∈ CN ×2 as an orthonormal basis for the space L[T] with b1
chosen to be equal to a /a . This gives the expression:
B = [b1 , b2 ] = TEp H (3.48)
where
9 
det (G) gpq
1  −√ 
Ep = 6  gpp gpp  ∈ R2×2 (3.49)
det (G) √
0 gpp
is the Gram-Schmidt orthonormalization matrix, such that the first col-

umn of TEp is the basis vector ȧp /ȧp and the matrix H ∈ R2×2 is the
orthogonal rotation matrix:
4 5
cos ψ − sin ψ
H= (3.50)
sin ψ cos ψ
where ψ is the rotation angle between the tangent vector to the p-curve ap
and the tangent vector to the curve under consideration a :
: ;

Re{ȧH p a }
ψ = ȧp ,a = arccos

ȧp
a (3.51)
By projecting the second derivative
d (a ) dT dp d2 p
a = +T 2 (3.52)
ds ds ds ds
of the manifold vector at (p(s), q(s)) onto the subspace L[B] (note that
L[B] = L[T]), we have
PB a = κ1 b1 + κg b2 (3.53)
−1 H
where PB = B BH B B is the projection operator. The coefficient of b2 ,
that is κg , is defined as the geodesic curvature of the curve at (p(s), q(s)) and
represents the component of a in the direction of b2 . Hence its expression is

2 a }
κg = Re{bH (3.54)
The vector b = [0, 1]T will be used for notational purposes such that
b2 = Bb. Substituting the expressions of b2 and a into Eq. (3.54) gives,
after some manipulation:
<
dp dq dp d2 p
κg (s) = b H Ep G
T T T
Γ2p + Γ2q + 2 . (3.55)
ds ds ds ds
6
In Eq. (3.55) the term bT HT ETp G can be simplified to det(G) · [−p , q ],
providing
4 5T
6 −p
κg (s) = det (G) (Γ2p p + Γ2q q ) p + p (3.56)
q

Chapter 4
Non-Linear Arrays:
(θ, φ)-Parametrization of Array
Manifold Surfaces
It is common practice to express the direction of array signals in terms of

azimuth angle θ and elevation angle φ. In this case the response vector of
an array of N omnidirectional sensors is (see Eq. (1.22))

a (θ, φ) = exp −jrT k (θ, φ) (4.1)
where r is the matrix with columns the sensor location and k(θ, φ) is the
wavenumber vector given by Eq. (1.3). Thus by using the directional param-
eters (θ, φ), instead of the generic parameters (p, q) of the previous chapter,
the two-parameter array manifold M of Eq. (3.1) becomes
M = {a(θ, φ) ∈ CN , ∀(θ, φ) : (θ, φ) ∈ Ω} (4.2)

where
Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ) and φ ∈ [0◦ , 90◦ )}
The aim of this chapter is to demonstrate the feasibility and potential

benefits of the theoretical framework presented in Chapters 2 and 3 by
examining various non-linear arrays of omnidirectional sensors. However,
in order to proceed, the following derivatives of the wavenumber vector
k(θ, φ) are required

k̇θ = π[− sin θ cos φ, cos θ cos φ, 0]T
first derivatives: (4.3)
k̇φ = π[− cos θ sin φ, − sin θ sin φ, cos φ]T

k̈θθ = π[− cos θ cos φ, − sin θ cos φ, 0]
 T
second derivatives: k̈θφ = −k̇θ tan φ (4.4)



k̈φφ = −k(θ, φ)
77
where the following notation has been used
dk (θ, φ) d2 k(θ, φ)
k̇p = and k̈pq = (4.5)
dp dp dq
with
(p and q) = (θ or φ)
4.1 Manifold Metric and Christoffel Symbols
For an azimuth-elevation array system of omnidirectional sensors the func-

tion a = a(θ, φ), ∀(θ, φ), provides a regular parametric representation of
the manifold surface M. With M embedded in an N -dimensional complex
space, the manifold metric G is a function of the array geometry r and
given by the following expression
4 5
gθθ , gφθ
G=
gθφ , gφφ
= [k̇θ k̇φ ]T rrT [k̇θ k̇φ ] (4.6)
Thus, for arrays of omnidirectional sensors, the elements of the metric G

can be computed as
T
gpq = k̇p rrT k̇q where p and q = θ or φ (4.7)
while the Christoffel symbol matrices are

  =Γθ,θθ 

 =Γθ,φθ

  T T 

 

  θ rr k̈θθ ,
k̇ −gθθ tan φ 



 Γ1θ =  

  T T 

 k̇φ rr k̈θθ , −gθφ tan φ 





 =Γφ,θθ =Γφ,φθ
First kind: and (4.8)


  



 =Γθ,θφ =Γθ,φφ


 

  −gθθ tan φ,

  −kT rrT k̇θ 


Γ =  

 1φ
−gθφ tan φ,

  −kT rrT k̇φ 


 
=Γ φ,θφ =Γφ,φφ
Non-Linear Arrays: (θ, φ)-Parametrization of Array Manifold Surfaces 79
  
=Γθφθ



  

 Γθθθ , − tan φ 

  

 Γ2θ =  

 Γφ , 

  θθ 0




 =Γφ

 φθ
Second kind: and (4.9)


  



 =Γθθφ

  

 − tan φ, Γθφφ 

  

Γ =  


2φ
 0 , 

  Γφφφ 


=Γφ
θφ
It should be noted from Eq. (4.9) that Γφθφ = 0. This simplifies the
Gaussian curvature of Eq. (3.31), for a 3D array of N -omnidirectional
sensors, to
:6 ;
1 d det(G) φ
KG (θ, φ) = 6 Γθθ (4.10)
det(G) dφ gθθ
which, in general, is not constant and, as it can be easily proven, can-

not be negative (i.e. KG (θ, φ) 0, ∀(θ, φ)). This leads to the following
theorem.
Theorem 4.1 The manifold surface of an array of N -omnidirectional

sensors of arbitrary geometry is never hyperbolic. It is elliptic and embedded
in an N -dimensional complex space.
In Chapter 3 we have stated that surfaces of constant Gaussian cur-

vature are known as “developable surfaces.” As KG (θ, φ) of Eq. (4.10)
is in general positive or zero, but not constant ∀(θ, φ), the array man-
ifold of omnidirectional sensors is not a developable surface. Therefore,
the existence of the “development” (see Definition 3.6, page 72) is not
guaranteed. However, for some array geometries, the Gaussian curvature
of Eq. (4.10) is constant. Array geometries which have been identified to
satisfy this condition include all planar arrays as well as a special class of
three-dimensional (3D) arrays, known as 3D-grid arrays which are defined
and discussed next.
4.2 3D-grid Arrays of Omnidirectional Sensors
Definition 4.1 — 3D-grid arrays.1 A three-dimensional array geometry

of omnidirectional sensors is said to be a 3D-grid array if and only if the
following expression is satisfied:
rrT = ρ2 I3 where ρ ∈ R (4.11)
Thus, in 3D-grid arrays the vectors rx , ry and rz are not only orthogonal
but also have the same magnitude. The following theorem is concerned
with the shape of the manifolds of 3D-grid arrays. To prove this theorem
it suffices to show that these arrays have constant and positive Gaussian
curvature.
Theorem 4.2 The manifold surface of a 3D-grid array of N -
omnidirectional sensors is spherical with radius πρ (and hence developable)
Example 4.1 The 8-element cube array with all sides equal to one half-
wavelength is a 3D-grid array
√ and has a spherical manifold with Gaussian
curvature (2π)−1 , or radius 2π half-wavelengths.
Table 4.1 summarizes the results of the array manifold parameters for
3D-grid arrays.
Table 4.1 Manifold parameters of 3D-grid arrays.
Intrinsic Parameter Expression

2
cos φ, 0
G ρ2 π 2
(Eq. (3.5)) 0, 1
det(G) ρ4 π 4 cos2 φ

0, −1
Γ1θ ρ2 π 2 cos φ sin φ
1, 0
(Eq. (3.20), (ζ=θ))

2 2 −1, 0
Γ1φ ρ π cos φ sin φ
0, 0
(Eq. (3.20), (ζ=φ))

0, − tan φ
Γ2θ
cos φ sin φ, 0
(Eq. (3.22), (ζ=θ))

− tan φ, 0
Γ2φ
(Eq. (3.22), (ζ=φ))
0, 0
1
KG
(Eq. (3.31)) ρ2 π 2
1 3D-grid arrays are also known as 3D balance symmetric arrays.

4.3 Planar Arrays of Omnidirectional Sensors
A planar (2D) array can be seen as a special case of a 3D-array where all
the elements of the last (3rd) row of the matrix r are zeros (i.e. rz = 0N ).
In this case, the term rrT has the following form

L, 02 c1 , c2
where L = ∈ R2×2 (4.12)
0T2 , 0 c2 , c3
Furthermore, if L = ρ2 I2 with ρ ∈ R, the planar array is said to be a

2D-grid array (or 2D balance symmetric). Fig. 4.1 shows some examples of
2D-grid array geometries.
Theorem 4.3 The Gaussian curvature KG of the two-parameter mani-

fold of a planar array of N -omnidirectional sensors is constant at every
point and equal to zero.
Proof. A proof of this theorem is given in Appendix 4.7.1.
Fig. 4.1 Examples of 2D-grid array geometries.

Fig. 4.2 Manifold Surface M of a Planar Array parametrized in terms of (θ,φ). This is
a developable manifold-conoid generated by rotating a φ-curve (geodesic curve) around
the apex.
Thus according to Theorem 4.3, the Gaussian curvature KG of a planar

array is identically zero (wherever defined). Hence, the manifold of a planar
array is, what is called in differential geometry, flat or parabolic of conoid
shape with the apex at point φ = ±90◦ (see Fig. 4.2). It is important to
note that the “flatness” does not imply that there exist straight lines, as in
the case of a surface in R3 . It means that such surfaces can be generated by
rotating a passing curve, which should have the special property of being
geodesic, around an apex point. In this case the φ-curves are shown to
compose a family of geodesic curves passing from a singularity point, i.e.
the apex of the developable manifold-conoid as illustrated in Fig. 4.2.
Furthermore, because KG never takes positive values, the manifold can-
not be closed. According to Beltrami’s theorem, since KG is constant, the
manifold is a developable surface. Based on the above discussion the fol-
lowing theorem, concerned with the shape of the manifold of an array of
omnidirectional sensors, can be easily proven.
Theorem 4.4 The manifold surface of a planar array of √N -

omnidirectional sensors is a conoid lying on a hypersphere of radius N
Table 4.2 summarizes the results of the differential geometry parameters

of planar arrays of omnidirectional sensors together with their correspond-
ing simplified expressions for the case of 2D-grid arrays.
Table 4.2 Manifold parameters of planar arrays.
Intrinsic Parameter 2D General Arrays 2D-grid Arrays

2
gθθ , gφθ cos φ, 0
G ρ2 π 2
(Eq. (3.5)) gθφ , gφφ 0, sin2 φ
π4
det(G) det(L) sin2 (2φ) ρ4 π 4 cos2 φ sin2 φ
4
 g 
θφ
, −gθθ tan φ
 tan φ  0, −1
Γ1θ  gφφ  ρ2 π 2 cos φ sin φ
(Eq. (3.20), (ζ=θ)) , −gθφ tan φ 1, 0
tan φ
 gθφ 
−gθθ tan φ,
 tan φ  −1, 0
Γ1φ  gφφ  ρ2 π 2 cos φ sin φ
(Eq. (3.20), (ζ=φ)) −gθφ tan φ, 0, 1
tan φ

0, − tan φ 0, − tan φ
Γ2θ 1 1
(Eq. (3.22), (ζ=θ)) tan φ
, 0 tan φ
, 0

− tan φ, 0 − tan φ, 0
Γ2φ 1 1
(Eq. (3.22), (ζ=φ))
0, tan φ
0, tan φ
KG 0 0
(Eq. (3.31))
Based on the results presented in Tables 4.1 and 4.2 a number of com-
ments can be made. For instance, the Gaussian curvature of 3D-grid arrays
is always positive and constant whereas for planar arrays it is always zero
(constant). This implies that the array manifold of a 3D-grid array is iso-
metric with a sphere of radius ρπ while the manifold of a planar array is
isometric with a plane. In addition, for all 2D-grid and 3D-grid arrays, the
manifold metric G is a diagonal matrix. The determinant of this matrix is
given by the 2nd row of Tables 4.1 and 4.2 and can be easily proven for grid
4
arrays. For planar arrays, the proof of det(G) = π4 det(L) sin2 (2φ) is given
in Appendix 4.7.2. All these expressions indicate that the determinant of
the manifold metric G is independent of the azimuth angle θ.
4.4 Families of θ- and φ-curves on the Manifold Surface
The array manifold can also be described and analyzed by treating this
surface as one of the following two families of curves:
(1) the family of θ-parameter curves defined as
{Aθ|φo , ∀φo : φo ∈ Ωφ } (4.13)

where Aθ|φo denotes the “θ-curve” on the surface M corresponding to

a constant value of φ = φo represented as follows:
Aθ|φo = {a(θ, φo ) ∈ CN , ∀θ : θ ∈ Ωθ , φo = constant} (4.14)
(2) the family of φ-parameter curves defined in a similar fashion as

Aφ|θo , ∀θo : θo ∈ Ωθ (4.15)
where
Aφ|θo = {a(θo , φ) ∈ CN , ∀φ : φ ∈ Ωφ , θo = constant} (4.16)
◦
with Ωθ [0 to 360◦ ) and Ωφ [0◦ to 90◦ ). Note that

M = Aθ|φo , ∀φo : φo ∈ Ωφ = Aφ|θo , ∀θo : θo ∈ Ωθ (4.17)
Figure 4.3 shows an illustrative representation of one “θ-curve” and one

“φ-curve” lying on the array manifold surface M.
These two curves lying on M can be analyzed using the geodesic cur-
vature. The geodesic curvature is one of the most important parameters
of the intrinsic geometry of a surface. In particular this parameter is asso-
ciated with curves lying on the surface and remains invariant under an
“isometry.”
Fig. 4.3 An illustrative example of a θ-curve and a φ-curve on the manifold surface M.
It can be proven that all 3D and all planar arrays have zero geodesic
curvature for the φ-curves. This implies that
κg,φ = 0 ⇐⇒ φ-curves = geodesic curves (4.18)
Furthermore, the geodesic curvature κg,θ for the θ-curves for 3D-grid and
2D-grid arrays is constant (depending on φo ) but is not zero. Hence
κg,θ = 0 ⇐⇒ θ-curves = geodesic curves (4.19)
while constant κg,θ implies that the development of the θ-curves of grid
arrays are circles whose radius is a function of φo .
Using Eq. (3.42) it can be proven that for a general planar array the
geodesic curvature κg,θ is given as follows
9
det (G) 1
κg,θ = 3 (4.20)
gθθ tan φo
and this may not even be constant. The above results are summarized in
Table 4.3.
Table 4.3 Geodesic curvature.
Curve κg 3D-grid Arrays Planar Arrays 2D-grid Arrays

tan φo det (G) 1 1
Aθ|φo κg,θ 3
ρπ gθθ tan φo ρπ cos φo
Aφ|θo κg,φ 0 0 0
(hyperhelix)
For the 2D-grid and 3D-grid arrays, the off-diagonal elements of the
manifold metric G (i.e. gθφ , gφθ ) are zero (see Tables 4.1, 4.2) and hence
the θ-curves and φ-curves are orthogonal and constitute geodesic coordi-
nates i.e.
.
(θ-curves) ⊥ (φ-curves)
=⇒ geodesic coordinates (4.21)
(φ-curves) = geodesic curve
Figure 4.4 shows an illustrative representation of a planar array manifold
surface and its families of θ- and φ-curves as geodesic coordinates.
Theorem 4.5 The geodesic curves (φ-curves) on the manifold surface
are complex hyperhelices embedded in CN .
Fig. 4.4 Families of θ-curves and φ-curves on the manifold M of a planar array.
Proof. Eq. (4.1), for constant θ (of value θ0 ), can be expressed as
a(θo , φ) = exp(−jπ(rx cos θo + ry sin θo ) cos φ)

= exp(−jπr(θo ) cos φ) (4.22)
where
r(θo ) = rx cos θo + ry sin θo (4.23)
which matches Eq. (2.27) for r = r(θo ), p = φ and v = 0. However,

hyperhelical curves, such as φ-curves, have all their curvatures constant,
independent of the parameters (θ, φ), as shown in Chapter 2.
Hence hyperhelical curves on the manifold surface M are not only ana-
lytically “convenient” but also geodesic curves. Overall it may be concluded
that, although both families of curves can be used to describe the manifold
surface M, the θ-curves are the more complicated of the two families.
Corollary 4.1 All members of the family of φ-curves of planar arrays are
identical apart from their length lm and position in complex N-dimensional
space.
It is immediately apparent that
a(φ | θo ) = a(φ | θo + 180◦ ) (4.24)

Fig. 4.5 Illustration of composite curves on the array manifold surface.
which implies that the two φ-curves at θ = θo and θ = θo + 180◦ have the
same length and curvatures. As illustrated in Fig. 4.5, these two φ-curves
can be considered as a continuation of one another, forming a composite
φ-curve having a hyperhelical shape of double the length of the original
curves. That is
Aφ|θo + Aφ|θo +180◦ = composite curve (4.25)

(lm of Aφ|θo ) = (lm of Aφ|θo +180◦ )
= (s(90◦ ), Eq. (2.28)) = π r (θo ) (4.26)
(lm of composite curve) = 2π r (θo ) (4.27)
4.5 “Development” of Non-linear Array Geometries
In Section 3.5, “development” has been described as a local mapping which

represents a complicated surface (embedded in a multidimensional complex
space) as a simpler surface. This representation would simplify the analysis
of array manifold surfaces and consequently of array systems. The simplest
surface, from a conceptual point of view, is the real plane. In order to be use-
ful, this mapping should preserve the main characteristics of the manifold
surface. For instance, a manifold surface with a varying Gaussian curva-

ture cannot be adequately mapped onto a plane. Therefore KG should be
constant. Also, the transformation should map geodesic parameter curves
on the manifold surface onto straight lines on the real plane which are
the geodesic curves of the plane (geodesic mapping). Based on the above
discussion, such a mapping is provided below as a theorem.
Theorem 4.6 Consider an arbitrary point, (θ, φ), on the manifold sur-
face of an array of N sensors. The image of the point (θ, φ) on the real
parameter plane is given by
7 =θ 8
x0 (φ) + 0 κ−1
g,θ (ϑ, φ) cos(ϑ)dϑ
D(θ, φ) = =θ ∈ R2 (4.28)
y0 (φ) + 0 κ−1
g,θ (ϑ, φ) sin(ϑ)dϑ
where κg,θ (θ, φ) is the geodesic curvature of the θ-curves on the manifold
and x0 (φ) and y0 (φ) ∈ R are the integration constants such that:
=π =
x0 = − 12 0 κ−1 1 π −1
g,θ (ϑ, φ) cos(ϑ)dϑ and y0 = − 2 0 κg,θ (ϑ, φ) sin(ϑ)dϑ
The locus of the images D(θ, φ) ∀(θ, φ) is said to be the development2 of

the manifold and has the following properties:
(i) It exists if and only if the Gaussian curvature of the manifold surface
is constant.
(ii) The arc length sd (θ, φo ) of θ-curves on the development is given by

sd (θ, φo ) = κ−1
g,θ (ϑ, φo ) dϑ (4.29)
p
(iii) The curvature of the θ-curves on the development is equal to the

geodesic curvature κg,θ (θ, φo ) of the θ-curves on the manifold surface.
Proof. The proof is given in Appendix 4.7.3.
Note that the expressions for the integration constants assume the prop-
erty κg,θ (θ, φo ) = κg,θ (θ + π, φo ) (which is generally valid), and have the
effect of shifting the curves to a common centre, i.e.
D(θ, φ) + D(θ + π, φ) = 0 (4.30)
The first condition of the theorem (i.e. KG = constant) is valid for 3D-grid
and all planar arrays, which implies that these arrays are “developable,”
with their “development” expressions summarized in Table 4.4.
2 More on “development” and “developables” can be found in [17].

Table 4.4 Development of Grid and Planar Arrays

ρπ sin θ
3D-grid D (θ,φ) =
tan φ − cos θ
π
θ
π cos φ f (ϑ) cos ϑdϑ − 1
f (ϑ) cos ϑdϑ
D (θ,φ) = 0θ 2 0π
det (L) 0 f (ϑ) sin ϑdϑ − 1
2 0 f (ϑ) sin ϑdϑ
 
c1 c2
 
Planar r 2 , rT r 
 x x y2 
with L = 
 rT r , 
 x y r y 
 
c2 c3
3
and f (ϑ) c3 cos2 ϑ − 2c2 cos ϑ sin ϑ + c1 sin2 ϑ

sin θ
2D-grid D (θ,φ) = ρπ cos φ
− cos θ
Example 4.2 Figures 4.6, 4.7 and 4.8 provide the development of rep-
resentative 3D-grid, planar and 2D-grid array geometries. For instance the
development of a 3D-grid array with sensors located on the eight vertices
of a cube of side one half-wavelength and with reference point the centre
of the cube (array centroid) is shown in Fig. 4.6. For the planar array with
the positions of the sensors, in half-wavelengths, as given below:
 
−2.1 −1.1 0.4 0.9 1.9
r = −2.4 1.1 0.6 −0.4 1.1 (4.31)
0 0 0 0 0
the development is shown Fig. 4.7, while the development for a 5-sensor
uniform circular array (2D-grid array) of radius one half wavelength is pre-
sented in Fig. 4.8.
It is clear that the manifold M of a planar array is mapped on the

parameter plane (θ, φ) as a system of homocentric ellipses while if the array
is a grid array (2D-grid or 3D-grid) this mapping has a circular shape
and is “geodesic” mapping. In the case of 2D this is also an “isometric”
mapping. In other words the grid arrays have circular development for the
θ-curves implying that the geodesic curvature of these curves is constant.
The development of the φ-curves, which are geodesic curves on the surface,
are straight lines. The development of θ-curves for general planar arrays
Fig. 4.6 θ-curves and φ-curves development of the manifold of a 3D-grid array with
sensors located on the eight vertices of a cube of side one half-wavelength and with
reference point the centre of the cube (array centroid).
have an ellipsoidal shape with the two main axes at angles ψ0 and ψ0 + π/2
with respect to the x-axis where:

1 2c2
ψ0 = tan−1 (4.32)
2 c1 − c3
! "
with ci denoting the elements of the matrix L, i.e. L = cc12 ,, cc23 . Indeed the
angle ψ0 can be found by examining the points of extremum curvatures.
Since the curvature of the development is equal to κgθ , the geodesic cur-
vature of the θ-curves of the planar array manifold is differentiated with
respect to θ and equated to zero:
dκgθ
=0 (4.33)
dθ
d(c3 cos2 ψ0 − 2c2 cos ψ0 sin ψ0 + c1 sin2 ψ0 )−3/2
⇔ =0
dψ0
Fig. 4.7 θ-curves and φ-curves development of the manifold of a planar array of 5-sensor
with positions given by Eq. (4.31).
⇔ (c3 cos2 ψ0 − 2c2 cos ψ0 sin ψ0 + c1 sin2 ψ0 )−5/2

×((c1 − c3 ) sin(2ψ0 ) − 2c2 cos(2ψ0 )) = 0
to yield two equations as follows:


c3 cos ψ0 − 2c2 cos ψ0 sin ψ0 + c1 sin ψ0 = 0
2 2
⇔ or (4.34)

(c1 − c3 ) sin(2ψ0 ) − 2c2 cos(2ψ0 ) =0
Considering the first equation of Eq. (4.34):
c3 cos2 ψ0 − 2c2 cos ψ0 sin ψ0 + c1 sin2 ψ0 = 0
the discriminant gives −4 det(L) < 0. Hence the first equation has no solu-
tion. The second equation gives

2c2 1 2c2
tan(2ψ0 ) = and hence ψ0 = tan−1 (4.35)
c1 − c3 2 c1 − c3
Fig. 4.8 θ-curves and φ-curves development of the manifold of a 5-sensor uniform cir-
cular array (2D-grid array) of radius one half-wavelength.
Fig. 4.9 Orientation of the “development” of θ-curves — axes at angles ψ0 and ψ0 +90◦ .
4.6 Summary
A number of differential geometry parameters have been estimated for

3D and planar arrays of omnidirectional sensors parametrized in terms of
azimuth and elevation angles. The main results are summarized as follows:
• The manifolds of 3-dimensional arrays of arbitrary geometry are locally

elliptic (KG 0, ∀(θ, φ)),
• The manifolds of 3D-grid arrays are spherical (KG = constant, ∀(θ, φ)),
• All planar arrays have a manifold of conoidal shape (KG = 0, ∀(θ, φ)),
• The family of the φ-curves is a family of geodesic curves for all grid and
planar arrays,
• The θ-curves are not geodesic curves,
• The φ-curves have a hyperhelical shape,
• θ-curves and the φ-curves constitute a system of geodesic coordinates,
• For planar and grid arrays the development of φ-curves are straight lines
while the development of θ-curves have an ellipsoidal shape.
4.7 Appendices
4.7.1 Proof that the Gaussian Curvature of an

Omni-directional Sensor Planar Array Manifold is
Zero
Equation (4.10) gives the expression for the Gaussian curvature of an array
of omnidirectional sensors, which can be written in an equivalent way as
follows:
1 1
KG = 6
det(G) gθθ tan2 φ
2
 & ' 
dgφφ dgθφ
2 gθθ dφ + gφφ dφ − 2gθφ dφ
1 dgθθ
 6 
×
 6 ×(& det(G))−1 gθθ tan φ 
'
− det(G) dφ tan φ + gθθ (1 + tan φ)
dgθθ 2
dgφφ dgθφ
gθθ dφ + gφφ dgdφθθ − 2gθφ dφ
=
2 det(G)gθθ tan φ
dgθθ
dφ tan φ + gθθ (1 + tan2 φ)
− 2 tan2 φ
(4.36)
gθθ
By using the Christoffel symbols of the second kind, Γθθφ and Γφφφ , which
can be expressed in terms of the elements of the manifold metric G as
 & ' 
 Γθθφ = 2 det(G)
1 dg
gφφ dgdφθθ − gθφ dφφφ 
& ' (4.37)
Γφ = 1
g
dgφφ
+ g
dgφφ
− 2g
dgθφ 
φφ 2 det(G) θθ dφ θφ dφ θφ dφ
Eq. (4.36) can be rewritten as

dgθθ

Γθθφ + Γφφφ dφ tan φ + gθθ 1 + tan2 φ
KG = − 2 tan2 φ
(4.38)
gθθ tan φ gθθ
Using the previous expression in conjunction with Table 4.2, which provides
the Christoffel symbols Γθθφ and Γφφφ for a planar array as a function of φ,
we have

− tan φ + tan1 φ dgθθ
dφ tan φ + gθθ 1 + tan φ
2
KG = − 2 tan2 φ
gθθ tan φ gθθ
dgθθ
dφ + 2gθθ tan φ
=− 2 tan φ (4.39)
gθθ
By using the property:
k̈ θφ = −k̇ θ tan φ (4.40)
KG is simplified to
−2gθθ tan φ + 2gθθ tan φ
KG = − 2 tan φ =0 (4.41)
gθθ

4.7.2 Proof of the Expression of det(G) for Planar Arrays

in Table 4.2
> ?
L, 0
For planar arrays, the term rrT has the following form 0T , 02 where
> ? 2
c1 , c2
L = c2 , c3 ∈ R 2×2
and the elements of the manifold metric G for an
array of omnidirectional sensors can be explicitly written as a function of
the elements of L as

 gθθ = π cos φ c3 cos θ − 2c2 cos θ sin θ + c1 sin θ
2 2 2 2
g = π 2 sin2 φ c1 cos

2
θ + 2c2 cos θ sin θ + c3 sin2 θ
(4.42)
 φφ
gθφ = π cos φ sin φ c2 sin θ − cos θ + (c1 − c3 ) cos θ sin θ
2 2 2
Using Eq. (4.42) the determinant of the manifold metric G is found to be

 
(c3 cos2 θ − 2c2 cos θ sin θ + c1 sin2 θ)
det (G) = π 4 cos2 φ sin2 φ  ×(c1 cos2 θ + 2c2 cos θ sin θ + c3 sin2 θ) 
−(c2 (sin2 θ − cos2 θ) + (c1 − c3 ) cos θ sin θ)2
Let the first term inside the braces be denoted by X and the second by
Y , i.e.
(c3 cos2 θ − 2c2 cos θ sin θ + c1 sin2 θ)

X =
×(c1 cos2 θ + 2c2 cos θ sin θ + c3 sin2 θ) (4.43)
Y = (c2 (sin2 θ − cos2 θ) + (c1 − c3 ) cos θ sin θ)2
The above expressions can be expanded to



 X = c1 c3 χ4 − 2c1 c2 χ3 σ + c21 χ2 σ 2



 + 2c2 c3 χ3 σ − 4c22 χ2 σ 2 + 2c1 c2 χσ 3



 + c23 χ2 σ 2 − 2c2 c3 χσ 3 + c1 c3 σ 4


Y = c22 χ4 − c1 c2 χ3 σ + c2 c3 χ3 σ

 − c22 χ2 σ 2 − c1 c2 χ3 σ + c21 χ2 σ 2



 − c1 c3 χ2 σ 2 + c1 c2 χσ 3 + c2 c3 χ3 σ



 − c1 c3 χ2 σ 2 + c23 χ2 σ 2 − c2 c3 χσ 3


− c22 χ2 σ 2 + c1 c2 χσ 3 − c2 c3 χσ 3 + c22 σ 4
where χ = cos θ and σ = sin θ and hence χ2 + σ 2 = 1.

Hence:
det(G) = π 4 cos2 φ sin2 φ(X − Y )

= π 4 cos2 φ sin2 φ
×((c1 c3 − c22 )χ4 + 2(c1 c3 − c22 )χ2 σ 2 + (c1 c3 − c22 )σ 4 )
= π 4 cos2 φ sin2 φ(c1 c3 − c22 )(χ2 + σ 2 )2
= π 4 cos2 φ sin2 φ(c1 c3 − c22 )
π4
= det(L) sin2 (2φ)
4
4.7.3 Proof of “Development” Theorem 4.6

(i) According to Beltrami’s theorem [20, 21], because the plane has a
constant Gaussian curvature (equal to zero), the condition for the
existence of a distortionless mapping is that the array manifold surface
should have a constant Gaussian curvature.
(ii) Considering the development of the manifold θ-curves for a constant

φo , D(θ, φo ) ∀θ, its arc length is given by

dD(ϑ, φo )
sd (θ, φo ) =

dϑ (4.44)
dϑ
Using the Leibnitz rule of integral differentiation on Eq. (4.28):

4 5
dD(θ, φo ) cos θ
= κ−1g (θ, φ 0 )
dθ sin θ

o)
and hence
dD(θ,φ
dθ
= κ−1
gθ (θ, φo ) proving Eq. (4.29).
(iii) The curvature of a curve is the magnitude of the derivative of the
normalized tangent vector with respect to its arc length. Hence, con-
sidering the development of the manifold θ-curves for a constant φo ,
let u(θ) be its normalized tangent vector given by
dD(θ,φo ) 4 5
cos θ
u(θ, φo ) =
dθ
=±

dD(θ,φo )
(4.45)

dθ
sin θ
Differentiating Eq. (4.29) leads to

dsd
= κ−1
g (θ, φo ) (4.46)
dθ
and hence,
dθ
= κg (θ, φ0 ) (4.47)
dsd
The magnitude of the derivative of the normalized tangent vector with
respect to its arc length hence becomes

du (θ, φo )

du (θ, φo ) dθ

dsd

dθ dsd

4 5

− sin θ
=
κg (θ, φo )
cos θ
= κg (θ, φo ) (4.48)
The curvature of the development is thus equal to the geodesic curva-

ture of the manifold curve.
Chapter 5
Non-Linear Arrays:
(α, β)-Parametrization
In the previous chapter the array manifold surface M was defined in terms
of azimuth θ and elevation φ angles and its intrinsic geometrical properties
were highlighted. Furthermore, the surface M was also considered based on
the families of θ-curves and φ-curves, whose characteristics were examined
using the differential geometry framework presented in Chapters 2 and 3.
In Chapter 4 it was observed that while the φ-curves were hyperhelical and
geodesic, the θ-curves possessed neither property. It was also pointed out
that since the θ-curves are not hyperhelical, their curvatures depend on θ
and so analytical evaluation of curvatures of order greater than two can
become exceedingly laborious and impractical.
In spite of the intuitive appeal of using azimuth and elevation angles
as the medium of array parametrization, this choice of angles is by no
means unique. In this chapter an alternative parametrization, known as
“cone-angle”, is presented. This parametrization is particularly suitable for
planar arrays resulting in two families of hyperhelical curves which can
provide a great deal of additional insight into the nature of planar array
behavior.
5.1 Mapping from the (θ, φ) Parameter Space

to Cone-Angle Parameter Space
For an array of N -omnidirectional sensors with positions given by the 3×N

matrix r = [rx , ry , rz ]T , using the directional parameters (θ, φ) the array
manifold M was defined in the previous chapter as
M = {a(θ, φ) ∈ CN , ∀(θ, φ) : (θ, φ) ∈ Ω} (5.1)
97
where
Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ) and φ ∈ [0◦ , 90◦ )}
with
a(θ, φ) = exp(−jrT k(θ, φ)) (5.2)
An alternative parametrization of M is based on two new angles, known

as the cone angles α-β, and defined as follows. If the x- and y-axes of the
Cartesian frame (x-y-z) are rotated on the (x, y) plane around the z-axis by
an angle Θ, then the frame (x̌-y̌-z) is obtained. The cone-angle α is defined
as the angle between the wavenumber vector k(θ, φ) and the positive side
of the x̌-axis, while β is the angle between k(θ, φ) and the positive side of
y̌-axis.
The relation between the two conventions/parametrization is
cos α = cos φ cos(θ − Θ) 0◦ ≤ α ≤ 180◦ (5.3)

◦ ◦
cos β = cos φ sin(θ − Θ) 0 ≤ β ≤ 180 (5.4)
and is illustrated in Fig. 5.1. These angles α and β are called cone-angles
because the loci of wavenumber vectors of constant α = α0 (or equivalently
constant β = β0 ) form cones about the x̌ (equivalently the y̌) axis. In other
words, all possible incident signals from a fixed angle α, or β, form a cone
in space. Fig. 5.2 shows this for a random choice of cone-angles α0 and β0 .
The mapping
(θ, φ) −→ (α, β) (5.5)
Fig. 5.1 Azimuth-elevation and cone-angles conventions.

Non-Linear Arrays: (α, β)-Parametrization 99
Fig. 5.2 Half-cones formed by wavenumber vectors of constant α and constant β.
Fig. 5.3 Mapping from the (θ, φ) parameter space on to the (α, β) parameter space.
is one-to-one and an illustrative representation of this mapping is given in

Fig. 5.3. For a given pair of values of cone-angles α and β and a given
rotation Θ of the x-y axes, the expressions that calculate θ and φ can be
derived from Eqs. (5.3) and (5.4) and are given by

cos β
θ = Θ + arctan (5.6)
cos α

cos β
φ = arccos
sin(θ − Θ)

cos β
= arccos cos β (5.7)
sin arctan cos α
These equations are only valid for β = 90◦ . If β = 90◦ then

Θ if α ≤ 90◦
θ= (5.8)
Θ + 180◦ if α > 90◦

α if α ≤ 90◦
φ= (5.9)
180◦ − α if α > 90◦
From the above equations it is obvious that, unlike in the θ-φ parametriza-
tion, α and β angles are not completely independent.
α and β angle parameters can vary independently

(5.10)
if and only if φ = 0.
Furthermore, not all combinations of α ∈ [0, 180◦ ] and β ∈ [0, 180◦ ] are
acceptable/possible. The acceptable limits for α, when β = βo , and the
limits of β, when α = αo , are independent of the value of the rotation of
the x-y frame and may easily be derived from the limitations imposed by
Eqs. (5.6) and (5.7). These limits are given below:
if α = αo then


 βmin βmax


90◦ − α β 90◦ + α for αo 90◦
o o
(5.11)

 αo − 90◦ β 270◦ − αo for 90◦ < αo 180◦



βmin βmax
if β = βo then


 αmin αmax


90◦ − βo α 90◦ + βo for βo 90◦
(5.12)
βo − 90◦ α 270◦ − βo
 for 90◦ < βo 180◦



αmin αmax
5.2 Manifold Vector in Terms of a Cone-Angle
Assume the array sensor positions are measured in half-wavelengths and

represented by the matrix r = [rx , ry , rz ]T ∈ R3×N with respect to the
Cartesian frame (x-y-z). If (x̌-y̌-z) is another Cartesian frame obtained by
rotating the (x-y) axes on the (x,y) plane around the z-axis with a rotation
angle Θ, then the sensor positions ř ∈ R3×N with respect to the Cartesian
frame (x̌-y̌-z) can be expressed as follows:
  =r
cos Θ sin Θ 0

ř = cos(Θ + 90◦ ) sin(Θ + 90◦ ) 0 [rx , ry , rz ]T
0 0 1
= [r(Θ), r(Θ + 90◦ ), rz ]

T (5.13)
with
r(Θ) rx cos Θ + ry sin Θ (5.14)
r(Θ + 90◦ ) rx cos(Θ + 90◦ ) + ry sin(Θ + 90◦ ) (5.15)
Based on the above, the wavenumber vector expressed in terms of the cone
angles α and β, and for a rotation angle Θ, can be written as
> 6 ?T
ǩ(α, β) = π cos α, cos β, 1 − cos2 α − cos2 β (5.16)
and thus, for isotropic sensors, the manifold vector of Eq. (1.22) can be
expressed as a function of α and β parameters as follows:
a(α, β) = exp(−jřT ǩ(α, β)) (5.17)
It should be emphasized here that Θ is a fixed scalar and not a direc-

tion of arrival parameter. Angles Θ and Θ + 90◦ simply define the lines of
azimuth with respect to which α and β are respectively measured. Further-
more from Eqs. (5.14) and (5.15) it is easy to deduce that the roles of α
and β can be interchanged by simply replacing Θ with Θ + 90◦ .
5.3 Intrinsic Geometry of the Array Manifold Based on

Cone-Angle Parametrization
Using the cone-angle directional parameters (α, β) instead of the generic

parameters (p, q) of Chapter 3, the two-parameter array manifold M of
Eq. (3.1) can be expressed as
M = {a(α, β) ∈ CN , ∀(α, β) : (α, β) ∈ Ω} (5.18)

where
Ω = {(α, β) : α ∈ [αmin , αmax ) and β ∈ [βmin , βmax )}
with a(α, β) given by Eq. (5.17). As might be expected, the intrinsic shape
of a manifold surface, as defined by its Gaussian curvature KG , is invariant
under any parameter transformation. For instance, using the general frame-
work presented in Chapter 3 and replacing p = α and q = β, the Gaussian
curvature of the manifold surface for a 3D grid array can be found to be
KG = (ρπ)−2 (5.19)
confirming (see Table 4.1) that the manifold is “developable” and that the
shape of the manifold is independent of parametrization.
Next we will focus on the differential geometry properties of the manifold
surfaces of planar arrays of omnidirectional sensors. Results for the 2D-grid
arrays can easily be deduced from the following results and those for 3D-grid
arrays can be analyzed with no difficulty.
Using Eq. (5.14) the manifold vector of Eq. (5.17) for a planar array is
simplified to
a(α, β) = exp(−jπ(r(Θ) cos α + r(Θ + 90◦ ) cos β)) (5.20)
while its manifold surface M is represented by Eq. (5.18).

As usual, a study of intrinsic geometry begins with the specification of
a metric tensor G and the evaluation of the first fundamental coefficients
and the Christoffel symbols. It can be shown that the matrix G is given by
7 8 7 2
8
gαα gβα ȧα Re(ȧH
α ȧβ )
G= =

2 (5.21)
gαβ gββ Re(ȧH ȧ )
ȧ
α β β
7 8
π 2 r(Θ)T r(Θ + 90◦ ) sin α sin β
2
π 2 r(Θ) sin2 α,
=
π 2 r(Θ)T r(Θ + 90◦ ) sin α sin β, π 2 r(Θ + 90◦ ) sin2 β
2
where, as usual,
∂a
ȧα = (5.22)
∂α
and
∂a
ȧβ = (5.23)
∂β
Whenever α or β equal 0 or π, the matrix G of the first fundamental coef-
ficients becomes singular. At such singularities, one or both of the tangent
vectors ȧα or ȧβ vanish and so the tangent plane is no longer defined. Fortu-
nately, the requirement φ = 0 (for the independence of α and β) established
at the outset (see Eq. (5.10)) avoids these singularities. Also note that the
cone parametrization avoids the singularity at the apex (φ = π2 ) which
exists with θ-φ parametrization and hence is more suitable for the study of
the manifold in that region.
The other variables required for the study of intrinsic geometry are the
Christoffel symbols. From the definitions of Chapter 3, it can be shown
that if
T = [ȧα , ȧβ ] ∈ CN ×2 (5.24)
then the Christoffel symbol matrices of first kind are

4 5 
Γα,αα Γα,βα 

Γ1α = Re(T Ṫα ) =
H


Γ Γβ,βα 

4 5 β,αα 

g 0 

= cot α αα 

gβα 0 
7 8 (5.25)
Γα,αβ Γα,ββ 
Γ1β = Re(T Ṫβ ) =
H 



Γ Γβ,ββ 

4 5 β,αβ 

0 gαβ 

= cot β 

0 gββ
while the matrices of second kind are given below:

7 8
αα Γβα 
Γα α
−1
Γ2α = G Re(T Ṫα ) = β
H 

Γαα Γββα 


4 5 

1 0 

= cot α 

0 0 7 8 (5.26)
Γα Γαββ 

Γ2β = G−1 Re(TH Ṫβ ) = αβ 

Γαβ Γββ 
β β


4 5 

0 0 

= cot β 

0 1
Note that the Christoffel matrices are quite sparse (even more so in the case
of balanced-symmetric arrays where gαβ = gβα = 0), which is a desirable
feature since it simplifies the differential geometry considerably.
We have seen that the manifold surface of any planar array (see
Table 4.2) has a zero Gaussian curvature. This is a property which is
independent of the parametrization. Indeed following any of the Eqs.
(3.31), (3.29) or (3.30), in conjunction with the Christoffel symbol matri-
ces defined above (see Eqs. (5.25) and (5.26)), the Gaussian curvature of
the array manifold can be found to be zero. Indeed, by using for instance
Eq. (3.30), we have:
 ∂ 
∂α Γα,ββ − ∂β Γα,βα
∂
1  
Kgauss = +col2 {Γ2α }T col1 {Γ1β }
det[G]
−col2 {Γ2β }T col1 {Γ1α }

1 ∂gαβ
= cot β − 0 + 0 − gβα cot β cot α
det[G] ∂α
1
= (gαβ cot α cot β − gβα cot α cot β) (5.27)
det[G]
=0 (5.28)
The above result implies that the manifold is locally parabolic in CN ,

in the same way that a circular cylinder, or cone, is locally parabolic in
Euclidean space R3 .
This also indicates that the array is developable which implies that the
mapping of the manifold surface onto a real plane is distortionless.
Thus the array manifold, by virtue of its zero Gaussian curvature, is a
developable surface and hence can be isometrically mapped onto any other
surface of zero curvature, say a plane in R2 .
5.4 Defining the Families of α- and β-parameter Curves
Based on the above discussion, and by keeping the frame-rotation angle

Θ constant (with value Θ0 ), the array manifold can also be described and
analyzed by treating this surface as one of the following two families of
curves:
(1) the family of α-parameter curves defined as
{Aα|βo , ∀βo : βo ∈ Ωβ } (5.29)
where Aα|βo denotes the “α-curve” on the surface M corresponding to

a constant value of β = βo and is represented as follows:
Aα|βo = {a(α, βo ) ∈ CN , ∀α : α ∈ Ωα , βo = constant} (5.30)
(2) the family of β-parameter curves defined in a similar fashion as
{Aβ|αo , ∀αo : αo ∈ Ωα } (5.31)

where
Aβ|αo = {a(αo , β) ∈ CN , ∀β : β ∈ Ωβ , αo = constant} (5.32)
Fig. 5.4 An example of an α-curve and a β-curve lying on the array manifold M.
In the previous equations Ωα [αmin to αmax ) and Ωβ [βmin to βmax )

with αmin , αmax , βmin and βmax given by Eqs. (5.12) and (5.11). Note that
% & % &
M = Aα|βo , ∀βo : βo ∈ Ωβ = Aβ|αo , ∀αo : αo ∈ Ωα (5.33)
Fig. 5.4 shows an illustrative representation of one “α-curve” and one “β-
curve” lying on the array manifold surface M.
The differential geometry of the α- and β-curves on the array manifold
surface are next investigated. Naturally, due to the symmetry of Eq. (5.20),
the parameter curves are expected to exhibit similar characteristics.
5.5 Properties of α- and β-parameter Curves
5.5.1 Geodecity
Following the definitions of Chapter 3, the geodesic curvatures of the α- and
β-curves (see Eqs. (3.42) and (3.43)) are given by

 Γβ
κg,α = det[G] (g αα
)3/2
αα
(5.34)

κg,β Γα
= − det[G] (g ββ
)3/2
ββ
However, Eq. (5.26) clearly indicates that Γβαα = Γα

ββ = 0. Hence
κg,α = κg,β = 0 ∀α, β (5.35)
which is a very interesting result, implying that the two parameter curves
are of minimum length on the array manifold surface, and leads to the
following theorem.
Theorem 5.1 Both the α- and β-curves are geodesic curves for all val-
ues of Θ.
To check if, in addition to being geodesic curves, the α- and β-curves
are also “orthogonal” (constituting a geodesic coordinate system), the angle
between the tangent vectors ȧα and ȧβ should be found. Indeed
r(Θ)T r(Θ + 90)

(ȧα , ȧβ ) = arccos
r(Θ)r(Θ + 90)
= (r(Θ)T r(Θ + 90)) (5.36)
and is, hence, constant. If rx ⊥ry then
r(Θ)T r(Θ + 90◦ ) = 12 (ry 2 − rx 2 ) sin(2Θ)

and the tangent vectors ȧα and ȧβ (and therefore α- and β-curves) are
strictly orthogonal and constitute geodesic coordinates if Θ = 0, or
rx = ry . Thus, in summary,

(i) (α-curves) and (β-curves) : geodesic curves 

i.e. κg,α = 0 = κg,β ∀Θ 


Θ = 0
and
rx ⊥ry , or (5.37)
(ii) if



rx = ry and rx ⊥ry 


then (α-curves) ⊥ (β-curves) =⇒ geodesic coordinates
It is a simple matter to confirm that for 2D-grid (balanced-
symmetric1 ) arrays
r(Θ)T r (Θ + 90◦ ) = 0, ∀Θ (5.38)
and so the tangent vectors ȧα and ȧβ are strictly orthogonal. For such arrays
matrix G is diagonal and the hyperhelical α-β curves form an orthogonal
coordinate system ∀Θ.

1• balanced-symmetric arrays: r x = r y and r x ⊥ ry

• unbalanced-symmetric arrays: r x = r y and r x ⊥ ry
5.5.2 Length of Parameter Curves

As was stated before, the most basic features of a curve a(p) are its arc
length s and rate of change of arc length ṡ(p)
ds
ṡ(p) = = |ȧ(p)| (5.39)
dp
In the case of the α- and β-parameter curves (assuming φ = 0) for a given
rotation Θo , the arc length of an α-curve Aα|βo and of a β-curve Aβ|αo are
respectively
s(α) = πr(Θo )(1 − cos α) (5.40)

◦
s(β) = πr(Θo + 90 )(1 − cos β) (5.41)
while their rate of change of arc length can be expressed as
ṡ(α) = πr(Θo ) sin α; ∀α ∈ Ωα|βo = [αmin , αmax ] (5.42)

◦
ṡ(β) = πr(Θo + 90 ) sin β; ∀β ∈ Ωβ|αo = [βmin , βmax ] (5.43)
Then, by integrating Eq. (5.42) over its limits αmin and αmax , the length
lm of an α-curve can be found as
Aα|βo : lm = 2πr(Θo ) sin β0 (5.44)
which indicates that α-curves, although identical in shape, have different

lengths, with the α-curve associated with βo = 90◦ having the maximum
length and their lengths gradually reducing as βo increases from 90◦ to 180◦
or decreases from 90◦ to 0◦ .
In a similar fashion the length of a β-curve, is as follows
Aβ|αo : lm = 2πr(Θo + 90◦ ) sin αo (5.45)
Clearly the α- (or β-) curve degenerates into a single point (i.e. a curve
of zero length) whenever β (or α) approaches either 0 or π radians.
5.5.3 Shape of α- and β-curves

The shapes of the parameter curves can be identified through evaluation
of their curvatures. It can be found that the curvatures of the α- and
β-parameter curves are constant and can be found from the recursive
Eq. (2.39) with the difference that r(Θ) and r(Θ + π2 ), respectively, have
taken on the role of the sensor location vector r of the linear array. This
is apparent from Eq. (2.39) which indicates that curvatures are indepen-
dent of α and β respectively and are only functions of the array geometry
projected along the lines of azimuth at Θ and Θ + π2 . Consequently, the
curvatures and hence the shapes of the α, β-curves depend strongly on Θ,
reminiscent of the drastic variations in array performance which can be
observed along different lines of azimuth. Thus, Theorem 2.1 is the centre-
piece of the investigation of the α/β-curves and, as will be demonstrated
next, has some far-reaching implications.
Corollary 5.1 The α- and β-curves are both hyperhelical.
For instance for an α-curve, of constant β (β = βo ) and constant Θ

(Θ = Θo ), Eq. (5.20) becomes identical to Eq. (2.27) with
r = r(Θo ), p = α, v = r(Θo + 90◦ ) cos βo = constant (5.46)
From the above it is clear that all α-curves have the same ELA r(Θo )
but different “visible” areas specified by the minimum and maximum per-
missible values of α, i.e. Ωα|βo = [αmin , αmax ]. The same is true for the
β-curves but with r(Θo + 90◦ ) as their ELA with Ωβ|αo = [βmin , βmax ].
ELA of the family {Aα|βo ∀βo : βo ∈ Ωβ } : r(Θo )

(5.47)
ELA of the family {Aβ|αo ∀αo : αo ∈ Ωα } : r(Θo + 90◦ )
Figure 5.5 provides an illustrative picture of an α-curve highlighting its

hyperhelical shape and the fact that its length increases for values of
βo approaching 90◦ .
Fig. 5.5 An illustrative diagram of the length of the hyperhelical α-curve Aα|βo .
Corollary 5.2 For a given frame rotation angle Θo , all members of the
family of α- (or β-) curves of a planar array of N sensors are identi-
cal (congruent) in shape apart from their length and position in complex
N -dimensional space.
The above result means that the whole of the manifold surface may be
fully covered by the simple translation of either of the two individual curves
with displacement values given by v = r(Θo + 90◦ ) cos β0 for α-curves and
v = r(Θo ) cos α0 for β-curves. Note that α-curves are the same as β-curves
for rotation Θo + 90◦ .
Corollary 5.3 The α- and β-curves are identical to the θ-parameter
manifold of a linear array of isotropic sensors with locations r(Θ) and
r(Θ + π2 ) respectively.
Figure 5.6 illustrates the α-β parametrization of the array manifold in
the special case of Θ = 0. Notice that for Θ = 0, cone-angles are mea-
sured with respect to the x and y axes and consequently, according to
Corollary 5.3, the α- and β-curves are identical to the θ-parameter man-
ifolds of linear arrays with sensor locations r(0) = rx and r( π2 ) = ry
respectively.
Although the hyperhelical α (or β) family of curves shown above covers
the whole of the manifold surface and is sufficient to fully define the man-
ifold shape, it is not unique. In fact, to understand the behavior of the
array for all possible directions of arrival, it is necessary to consider the
hyperhelical coordinates corresponding to all values of Θ from 0 to π2 .
Fig. 5.6 (θ, φ) and (α, β) parametrizations of the same manifold surface M with the
frame rotation for the cone-angles Θo = 0.
Fig. 5.7 Families of α and β curves for a frame rotation Θo = 0◦ . and for Θo < 90◦ .
The above concept is more easily appreciated by observing Fig. 5.7

which shows the effect of a non-zero value of Θ. It should be stressed that,
despite the appearance of Fig. 5.7, the use of various values of Θ is not
equivalent to a trivial rotation of the parameter curves, since their differ-
ential geometry is a function of Θ (Corollary 5.3).
Figures 5.6 and 5.7 also reveal a clear relationship between the α (or β)
curves and the hyperhelical φ-curves, which will be stated as the final two
corollaries:
Corollary 5.4 Each member of the α-curve family is identical in shape

to the combination of the φ-curves corresponding to θ = Θ and θ = Θ + π.
Similarly each member of the β-curve family is identical to the combination
of the φ-curves at θ = Θ + π2 and θ = Θ + 3π
2 .
Corollary 5.5 The α-curve with β = 90◦ for a frame rotation Θo and
the composite φ curve Aφ|θo + Aφ|(θo +π) are one and the same curve if and
only if θo = Θo .
Similarly the β-curve with α = 90◦ for a frame rotation Θo is one and
the same curve with the composite φ curve Aφ|θo + Aφ|(θo +π) if and only if
θo = Θo + π2 .
5.6 “Development” of α- and β-parameter Curves
The most significant implication of the property that the array manifold
M ∈CN of a planar array has a zero Gaussian curvature, is that M is a
developable surface and hence can be isometrically mapped on to any other

surface of zero curvature, say a plane in R2 . The result of such a mapping
would be a consistent two-dimensional representation (development) of the
manifold in real space. A development may be readily evaluated by noting
that if Aα|βo ,Aβ|αo ∈ CN are arbitrary α and β curves on the manifold
surface and A(d)
α|βo
,A(d)
β|αo
∈ R2 are their respective images on the develop-
ment, then since an isometric mapping is by necessity both geodesic and
conformal:
• Length of A(d)α|βo
∈ R2 = Length of Aα|βo ∈ CN ,
• Principal curvature of A(d) α|βo
∈ R2 = Geodesic curvature of Aα|βo ∈ CN ,
• (A(d)
α|βo
, Aβ|α
(d)
o
) in R2 = (Aα|βo , Aβ|αo ) in CN .
Fig. 5.8 (a) Square-grid array structure, ry ⊥r x and r y = r x (balanced-symmetric).

(b) Rectangular-grid array structure ry ⊥r x and r y = 1.5 rx (unbalanced-symmetric).
(c) and (d) Developments of the manifolds of arrays in Figs. (a) and (b) respectively when
the manifolds are parametrized by cone-angles α and β with frame rotation Θ = 0◦ .
Table 5.1 Cone-parametrization

manifold parameters for 3D-grid
and planar arrays.
Parameter 3D Planar
1
KG 0
(Eq. (3.31)) ρ2 π 2
cot β
κg,α 0
ρπ
cot α
κg,β 0
ρπ
The above features may be demonstrated by the following example.

Example 5.1 Consider the developments of the manifolds of two array
configurations when parametrized with cone-angles α and β for Θ = 0◦
(recall that r(0) = rx and r(π/2) = ry ). The array structures are indicated
in Figs. 5.8(a) and (b) with the respective developments in Figs. 5.8(c) and
(d). Cone-angles α and β are considered in steps of 10◦ from 0◦ to 180◦ .
Since both arrays are symmetric (i.e. rTx ry = 0), the α- and β-curves
are orthogonal and geodesic, as a result of which their images on the

devel-

opment
are
orthogonal straight lines (geodesics) and of length 2π
rx
sinβ
and 2π
ry
sinα respectively. Note that the second array geometry is unbal-

anced (
ry
> rx ) and consequently the envelope of the development is
elliptical rather than circular.
Finally, Table 5.1 summarizes the results associated with the Gaussian
and geodesic curvatures indicating that for planar arrays both α and β
curves are geodesic curves. This indicates that the study of manifold geom-
etry is considerably simplified as a result of (α, β) parametrization, once
again confirming the advantages of cone-angles over azimuth and elevation.
Chapter 6
Array Ambiguities
In this chapter ambiguous sets of parameter values (e.g. directions) that

exist in any hyperhelical curve A embedded in an N dimensional complex
space, are modelled, identified and calculated. The curve A, which is defined
as follows
A = {a(p) ∈ CN , ∀p : p ∈ Ωp } (6.1)
with
a(p) = exp(−j(πr cos p + v)) (6.2)
could be the manifold of a linear array or a p-curve (with p = φ, α or β)

lying on the manifold surface of a planar array (using the concept of the
ELA). In these cases the manifold vector a(p) represents the complex
array response to a unity power signal impinging on the array from direc-
tion p and is a function of the locations r ∈ RN of the array elements,
defined as
sensor locations : r r(Ψ0 ) = rx cos Ψ0 + ry sin Ψ0 (6.3)
The parameter Ψ0 , shown in Fig. 6.1 and used in Eq. (6.3), is defined as
 ◦

 0 for any linear array on the x-axis

θo for one φ-curve Aφ|θo
Ψ0 (6.4)
Θo
 for all α-curves

Θo + 90◦ for all β-curves
where θo is the constant azimuth (φ-curves) and Θo is the frame rotation

angle (α-curves). Table 6.1 provides the main parameters of Eq. (6.2) for
the various hyperhelical array manifold curves.
113
Fig. 6.1 Equivalent linear array r(Ψ0 ).
Table 6.1 Array manifold vector parameters (Eq. (6.2)).
Parameter Linear Arrays Planar Arrays
p (generic) θ φ α β
Ωp ∈ R 1 [0, π) [0, π/2) [0, π) [0, π)

r rx r(θo ) r(Θo ) r(Θo + 90◦ )
h(p) = cos(p) cos(θ) cos(φ) cos(α) cos(β)
v 0 0 r(Θo + 90◦ ) cos(βo ) r(Θo ) cos(αo )
curve A Aφ|θo Aα|βo Aβ|αo
6.1 Classification of Ambiguities
The array manifold A of Eq. (6.1) is essentially a mapping from the param-
eter space Ωp ∈ R1 to the complex N -dimensional space CN
array geometry
p ∈ Ωp −→ a(p) ∈ CN (6.5)
When finding the directions-of-arrival (DOA’s) of signals using array sensor

measurements, it is important to be certain that the problem has a unique
solution. This is guaranteed if the above mapping is one-to-one.
Array Ambiguities 115
However, in this study we will consider the problem which arises when
the mapping of Eq. (6.5) is not one-to-one. In this case the array cannot dis-
tinguish between two (or more) different signal environments. For instance
two different sets of signals impinging on the array can provide identical
responses at the array output, i.e. the same measurements. In such a case
any estimation algorithm may be unable to distinguish the true parameter
values (e.g. directions) from the false ones. This is termed as the ambiguity
problem.
Definition 6.1 — Array Ambiguity: It is the inability of an array of a

given geometry to distinguish a set of parameter values
(p1 , p2 , . . . , pM ) with pi = pj ∈ Ωp for i = j (6.6)
from at least one of its subsets (or one subset from another).
Array ambiguities can be classified as “trivial” and “non-trivial.” We

will start with trivial ambiguities.
Definition 6.2 — Trivial Ambiguities: An array ambiguity is said to

be classified as ‘trivial’ if there exist at least one pair p1 , p2 ∈ Ωp with
p1 = p2 such that
array geometry
p1 −→ a(p1 ) ∈ CN
array geometry
(6.7)
p2 −→ a(p2 ) ∈ CN
with
a(p1 ) = ka(p2 ) (6.8)
and
k = scalar
In this case the two parameter values p1 , p2 are indistinguishable by the

array.
Trivial ambiguities are the simplest type of ambiguity and can be easily
identified/detected. For instance it is impossible to distinguish whether a
signal is impinging on a linear array from direction-of-arrival p1 or from the
direction p2 = 360◦ − p1 , i.e. the mirror image of p1 with respect to 180◦ .
In this case
a(p1 ) = a(p2 ) (6.9)

and for this reason, in order to avoid this ambiguity problem, the parameter
space in the case of linear arrays is confined to Ωp = [0◦ , 180◦ ).
However, Eq. (6.9) is a special case of Eq. (6.8) for k = 1. Eq. (6.8) indi-
cates that there exists at least one manifold vector a(p2 ) which is colinear
to another manifold vector a(p1 ). Then the two waves with bearings p1 , p2
are indistinguishable by the array even if p1 , p2 ∈ Ωp = [0◦ , 180◦ ).
Non-trivial ambiguities are a more complicated type of ambiguity arising
when a manifold vector can be written as a linear combination of two or
more different manifold vectors.
Definition 6.3 — Non-Trivial Ambiguities: An array ambiguity is

said to be classified as “non-trivial” if there exist at least a set of values
p1 , p2 , . . . , pM ∈ Ωp with M > 2 and pi = pj for i = j such that
array geometry
pi ∈ Ωp −→ a(pi ) ∈ CN ; ∀i = 1, 2, . . . , M (6.10)
with
a(pM ) = c1 a(p1 ) + c2 a(p2 ) + · · · + cM −1 a(pM −1 ) (6.11)
Then this set (p1 , p2 , . . . , pM ) and all its subsets of M − 1 elements are
indistinguishable by the array.
In such a case the array will have identical responses for the sets of
bearings (p1 , p2 , . . ., pM −1 ), (p2 , p3 , . . ., pM ), (p1 , p3 , p4 , . . ., pM ), etc.
Fig. 6.2 illustrates a non-trivial ambiguous situation where the manifold
vector a(p3 ) belongs to the subspace L [A] where A = [a(p1 ), a(p2 )].
This type of ambiguity is much more difficult to identify since,
unlike a trivial ambiguity, it cannot be detected by a simple search
of the manifold. For example consider the non-uniform linear array
of four sensors with sensor locations (measured in half wavelengths)
given by
r = [−2, −1, 1, 2]T (6.12)
One non-trivial ambiguity of this array is related to the set of directions
0◦ , 60◦ , 90◦ , 120◦
as the array is unable to distinguish between the following signal

environments:
(0◦ , 60◦ , 90◦ ), (0◦ , 60◦ , 120◦ ), (60◦ , 90◦ , 120◦ ), (0◦ , 60◦ , 90◦ , 120◦ )
Fig. 6.2 Illustrative representation of array ambiguity (the vectors a(p1 ), a(p2 ) and
a(p3 ) are linearly dependent).
For instance, for this array, the above four different signal environments
provide identical MUSIC spectra. Note that for this specific array,
rank ([a(0◦ ), a(60◦ ), a(90◦ ), a(120◦ )]) = 3 (6.13)
and hence these four manifold vectors are linearly dependent.

In general we can say that every array suffers from ambiguities in some
way or another as a result of the array geometry and thus different array
geometries have different sets of ambiguous directions. This implies that
an array which is unambiguous for a given set of directions might become
ambiguous for the same set if we change its sensor locations even slightly
to form a new array geometry. A typical example can be seen using the
following linear array,
r = [−3, −1.5, 0, 1.0

←→, 3]
T
(6.14)
operating in the presence of two sources from unambiguous directions of 5◦

and 35◦ . If the fourth sensor located at +1 half-wavelengths is now moved
to the location +1.5 half-wavelengths, so that the uniform linear array
r = [−3, −1.5, 0, 1.5

←→, 3]
T
(6.15)
is obtained, then these two sources at 5◦ and 35◦ become “ambiguous” as

these two directions are now a subset of the “non-trivial” type ambiguous
set of directions 5◦ , 35◦ , 109.70◦ , 120.94◦ . This can also be seen, for instance,
Fig. 6.3 MUSIC spectrum in an “ambiguous” situation.
from Fig. 6.3 where in the MUSIC spectrum, four nulls appear at directions
5◦ , 35◦ , 109.70◦ , 120.94◦ rather than two at 5◦ and 35◦ .
Note that in this study the array centroid will be taken as the reference
point, implying that sum(r)= rT 1N = 0. However, it can be proved that
the ambiguities are independent of the choice of the array reference point.
6.2 The Concept of an Ambiguous Generator Set
Before continuing it is necessary to present the following definitions which

will be extensively used.
Definition 6.4 — Ambiguous Set and Rank of Ambiguity
(1) Ambiguous Set:

An ordered set of arc lengths s = [s1 , s2 , . . . , sc ]T , is said to be an
ambiguous set of arc lengths
if c ≤ N and the matrix A(s) ∈ CN ×c with columns the manifold vec-
tors a(s1 ), a(s2 ), . . . , a(sc ) has rank less than c, i.e. rank(A(s)) < c
or
if c > N and all subsets si of s which contain exactly N elements
are themselves ambiguous sets of arc lengths,
i.e. rank(A(si )) < N, ∀i (6.16)

(2) Rank of Ambiguity:

For an ambiguous set of arc lengths s = [s1 , s2 , . . . , sc ]T its rank of
ambiguity is defined as the integer ρa

rank(A(s)) if c N
where ρa = min {rank (A(s ))} if c > N (6.17)
i
∀i
with si ∈ RN , ∀i.
Note that the previously introduced definitions can be directly applied

to sets of directional parameters by simply substituting (using Eq. (2.28))
the arc length s with p.
The following theorem, which is a very important result, essentially
states that if all the elements of an ambiguous set of arc lengths are rotated
on the array manifold by the same value, then the resulting set is also an
ambiguous set of arc lengths.
Theorem 6.1 — Arc length Rotation: For a linear array of N sen-

sors, if
s = [s1 , s2 , . . . , sc ]T (6.18)
is an ambiguous set of arc lengths with rank of ambiguity ρa , then any set
š = [s1 + ∆s, s2 + ∆s, . . . , sc + ∆s]T

= s + ∆s.1c , (6.19)
with s1 + ∆s ≥ 0 and sc + ∆s ≤ lm , is also an ambiguous set of arc lengths

with the same rank of ambiguity ρa .
The proof of this theorem can be found in Appendix 6.7.1 while its
essential features are illustrated in the following example.
Example 6.1 Consider a linear array having sensor locations given, in

half-wavelengths, by the following vector:
r = [−3, −1, −0.5, 2, 2.5]T
and having a manifold length lm = 2πr = 28.4483. It can easily be seen

that the set
p = [0◦ , 53.13◦ , 78.46◦ , 101.54◦ , 126.87◦ ]T

is an ambiguous set of directions. Using Eq. (2.28) the corresponding

ambiguous set of arc lengths is computed to be
s = [0, 5.6897, 11.3793, 17.0690, 22.7857]T
If this set s is rotated by ∆s = 2, then a new set of arc lengths
š = [2, 7.6897, 13.3793, 19.0690, 24.7857]T
is obtained, which corresponds to the set of directions
p̌ = [30.75◦ , 62.65◦ , 86.59◦ , 109.91◦ , 137.78◦ ]T
which is also an ambiguous set. This is also true for any ∆s so long as the
last element of š is smaller than, or equal to, the manifold length lm . Note
that the rotation should be carried out in the “arc length” domain and not
in the “directions-of-arrival” domain (parameter p domain). For instance,
if the directions of the original sources are rotated by ∆p = 20◦ so that the
new set of directions
p + ∆p15 = [20◦ , 73.13◦ , 98.46◦ , 121.54◦ , 146.87◦ ]T
is obtained, then this set is not ambiguous.

Furthermore any subset of 4 elements of p or p̌ are also ambiguous sets of
directions. If the array operates in the presence of 4 sources with directions
any subset of 4 elements of p will give the same MUSIC spectrum.
It becomes clear that if one ambiguous set is identified, then by simple
rotation in arc lengths, an infinite number of ambiguous sets can be gen-
erated and therefore two different ambiguous sets may in fact be just a
rotation of each other. It is an impossible task to try to identify all the
ambiguous sets and, since all these sets can be generated from a single set,
the idea of the ambiguous generator set arises, defined as follows:
Definition 6.5 — Ambiguous Generator Set: An ordered set s =
[s1 , s2 , . . . , sc ]T of c arc lengths, where 2 ≤ c ≤ N , is said to be an ambigu-
ous generator set (AGS) of arc lengths if and only if:
(a) s1 = 0 and sj = 0 with 2 ≤ j ≤ c. That is, all the elements of the set
but the first element are non-zero.
(b) the rank of the N × c matrix A(s) with columns the manifold vectors
associated with the elements of the set is less than c, i.e.
rank(A(s)) ρa < c (6.20)

(c) for any subset si of m elements of s (including s1 = 0) with ρa ≤ m < c,

the rank of A(si ) is equal to ρa
According to the previous definition, a set of arc lengths s =

[s1 , s2 , s3 , s4 ]T with s1 = 0, s2 , s3 , s4 = 0 and a(s3 ) = c1 a(s1 ) + c2 a(s2 )
is not an ambiguous generator set. This is because
rank([a(s1 ), a(s2 ), a(s3 ), a(s4 )]) = 3 (6.21)

but
rank([a(s1 ), a(s2 ), a(s3 )]) = 2 (6.22)
(i.e. third condition is not satisfied). On the other hand, the subset
[s1 , s2 , s3 ]T is an ambiguous generator set since it satisfies all the three
conditions of Definition 6.5.
The next section is concerned with the identification of ambiguous gen-
erator sets (AGS) existing in the manifold of a linear array of arbitrary
geometry.
6.3 Partitioning the Array Manifold Curve into Segments

of Equal Length
In this section a framework is proposed for the calculation of a type of

ambiguous generator sets of directions existing in linear arrays of any geom-
etry, as well as their associated rank of ambiguity. The framework is based
on the uniform partitioning of the array manifold hyperhelix, obtained by
dividing the manifold length lm by the difference between any two array
sensor locations. Note that from now on the vector r with elements the
sensor locations (in half-wavelengths) will be assumed to be ordered in the
sense that ri < rj ⇔ i < j. Furthermore, the absolute difference between
the i-th and j-th sensor locations, measured in half-wavelengths, will be
denoted by ∆rij > 0 and hence the aperture of the array will be ∆r1,N .
Note that the term “intersensor spacing” will refer to the spacing between
any two sensors, not necessarily adjacent.
Thus one type of array ambiguity can be identified by dividing a hyper-
helical manifold curve into equal segments, according to the following
theorem.
Theorem 6.2 — Uniform-Partition Type Ambiguities: Consider a
linear array (or an ELA) of N sensors, with locations r = [r1 , r2 , . . . , rN ]T
and let lm denote the length of its hyperhelical manifold curve embedded in
CN . By uniformly partitioning this hyperhelix into segments of arc length

∆rij where ∆rij = |ri − rj | with i = j, a set of c points on the manifold
lm
curve can be identified, forming the vector
4 5T
lm lm lm
s∆rij = 0, ,2 , . . . , (c − 1) (6.23)
∆rij ∆rij ∆rij
where

1 + fix(∆rij ) if ∆rij ∈
/ N+
c=
∆rij if ∆rij ∈ N +
and fix (•) denotes the integer part of a number
(a) If c ≥ N then any subset of s∆rij ∀i, j with N elements is an ambiguous

set.
(b) if c < N then s∆rij ∀i, j might be an ambiguous set.
Proof. The proof of this theorem is as follows. Consider a linear array of

N sensors with locations r. If the difference between the ith and jth sensor
locations, i.e. ∆rij = |ri − rj |, satisfies c ≥ N , then we have to prove that
any subset of N elements of the set of arc lengths s∆rij of Eq. (6.23) is an
ambiguous set.
Using the manifold vector expressed in terms of arc length, i.e.
rs − πr + v))
a(s) = exp(j() (6.24)
in conjunction with
r
lm = 2π r and )
r= (6.25)
r
the matrix A(s∆ri,j ) ∈ CN ×c with columns the associated manifold vectors

can be written as follows:
> & & ''
A(s∆rij ) = exp (j (−πr + v)) , exp j ∆r 2π
ij
r − πr + v ,...,
& & '' ?
exp j (c − 1) ∆r
2π
ij
r − πr + v (6.26)
Since c N the number of columns of A(s∆rij ) is equal to, or greater than,

N. Hence, by taking a submatrix of AN of A with exactly N elements, e.g.
the N first columns of A, we have the following square submatrix denoted
by AN
> & & ''
AN = exp (j (−πr + v)) , exp j ∆r 2π
ij
r − πr + v , . . . ,
& & '' ?
exp j (c − 1) ∆r
2π
ij
r − πr + v
)N
= diag(exp (j (−πr + v)) A (6.27)
where
4 5
) N = 1 , exp j2π 1 r , . . . , exp j(2π(N − 1) 1 r
A (6.28)
N
∆rij ∆rij
Consider now the ith and jth rows of A) N with i > j. By using the
property that

∆rij
exp −j2πk = 1, ∀k ∈ N + (6.29)
∆rij
) N can be written as follows

the ith row of A
) N ) = IN · rowT (A
rowTi (A )N )
i
 
1 0 0 ... 0

0 exp −j2π ∆rij 
 0 ... 0 
 ∆rij

= .
∆rij
 0 0 exp −j2π2 ∆rij . . . 0 
0 
 0 0 ...

0 
∆rij
0 0 0 . . . exp −j2π(N − 1) ∆rij
> & ' & '?T
1N , exp j2πri ∆r1ij , . . . , exp j2πri (N − 1) ∆r1ij
> & ' & '?T
r −∆r r −∆r
= 1N , exp j2π i∆rij ij , . . . , exp j2π(N − 1) i∆rij ij
> & ' & '?T
r r
= 1N , exp j2π ∆rjij , . . . , exp j2π(N − 1) ∆rjij
)N )
= rowTj (A
i.e.
) N ) = rowj (A
rowi (A )N )
& '
=⇒ det A ) N = 0 =⇒ det (AN ) = 0 (6.30)
It is easy to see that the fact that AN was chosen to consist of the first N
columns of A, and not any N columns, is not restrictive in the least since
the ith and the jth rows of any N × N submatrix of A will be equal. Since
all submatrices AN of A are singular, it results that A is rank deficient and
therefore, any subset of s∆ri,j with exactly N elements is an ambiguous set.

Note that

1 + fix(∆rij ) if ∆rij ∈
/ N+ lm
c= =⇒ c < 1 + ∆rij =⇒ (c − 1) < lm
∆rij if ∆rij ∈ N + ∆r ij
(6.31)
indicating that the last element of the vector s∆rij of Eq. (6.23), is smaller
that the manifold length lm . The vector s∆rij with elements c ∈ N + man-
ifold points will be known as the “uniform basic set” (UBS) of arc lengths
associated with the intersensor spacing ∆rij as it is constructed by uni-
formly partitioning the manifold length by the spacing ∆rij between the
ith and jth sensors.
According to Theorem 6.2, all subsets of N elements from a set s∆rij ,
of the form of Eq. (6.23) are ambiguous sets but not necessarily ambiguous
generator sets. In order for such a subset to be an ambiguous generator
set it must have its first element equal to zero and also satisfy the third
condition of the ambiguous generator set in Definition 6.5.
From the investigation of the precise cause of the ambiguous nature
of the set of arc lengths s∆rij constructed by uniformly partitioning the
manifold length by the spacing ∆rij , it is clear that the most common
AGSs are of rank N − 1. However, AGSs of rank less than (N − 1) can exist
if more pairs of rows of the determinant matrix A )N s
∆rij are identical.
This is possible if the array geometrical structure satisfies certain special
conditions and for this reason we need the following definition.
Definition 6.6 — A “Distinct” Set of Intersensor Spacings: In a

linear array of N sensors, a set of k intersensor spacings is said to be a
“distinct” set if no intersensor spacing is a combination of the other spacings
in the set. There are N (N − 1)/2 intersensor spacings but at most (N − 1)
elements in a distinct set of intersensor spacings (i.e. k N − 1)).
For instance, if ∆rij and ∆rjk are members of a distinct set with i =
j = k then ∆rik is not a member.
Based on the above definition, the following theorem states that if there
are distinct sets of intersensor spacings then AGSs of rank less than (N − 1)
may be found.
Theorem 6.3 — Number of AGSs of a given rank: Consider an N -
sensor linear array. If k∆rij denotes the elements of a distrinct set of inter-
sensor spacings which are integer multiples of the smallest (∆rij ) amongst
them, and

∆rij N − k∆rij (6.32)
then the Uniform Basic Set s∆rij of c elements is an ambiguous set from
which
min(N −1,c−1)
c−1
(6.33)

=N −k∆rij

ambiguous generator sets of rank- N − k∆rij can be constructed.
The implication of the above theorem to the number of ambiguous gen-
erator sets of a linear array is twofold. Firstly, it reveals the existence of
additional ambiguous generator sets corresponding to intersensor spacings
which are less than (N − 1), provided that they occur in integer multiples of
each other, which includes the special case of repeated occurrences. In par-
ticular, if a spacing,
say ∆r
ij , occurs k∆rij times then the size of this spacing
need only be N − k∆rij to give rise to ambiguous generator sets. Secondly,
the larger the k∆rij is, the more rank- N − k∆rij ambiguous generator sets
can be constructed. This point can be illustrated from the intersensor spac-
T
ings ∆r13 = ∆r24 = 4.1 of the linear array r = [ −3 , −2.2 , 1.1 , 1.9 , 2.2 ]
where k∆r13 = 2 and ∆r13 (N − k∆r13 ) = 3. Thus from s∆r1,3 , five
rank-3 ambiguous generator sets can be constructed (instead of a single
rank-4 set).
From the array’s
geometrical
perspective, the following statement with
respect to rank- N − k∆rij ambiguity can be made:
Corollary 6.1 If k intersensor spacings of an N -sensor linear array
are integer multiples of the smallest among them, with the latter being at
least (N − k) half-wavelengths, then the array suffers from rank-(N − k)
ambiguity.
In summary the reason for the existence of these lower than N − 1 rank
ambiguities can easily be deduced from the definition of the ambiguous
generator set and especially its third condition (see Definition 6.5). More
analytically, it can be easily derived from the proof of Theorem 2.1, that
if two or more intersensor spacings of a linear array (or ELA) are integer
multiples of each other (which is the case for either of the above conditions)
and if s is an ambiguous generator set associated with the smallest of these
intersensor spacings, then the manifold matrix A(s) may have more than
two identical rows. This implies that its rank will be smaller than N − 1.
As a direct result of Theorem 6.3 and of the previous corollary it can
be said that the uniform linear array (or ELA) with intersensor spacing
∆r1,2 1 suffers from trivial ambiguities. Indeed, the intersensor spacing
∆r = ∆r1,2 occurs N −1 times. That is ∆r1,2 = ∆r2,3 = · · · = ∆rN −1,N , i.e.
k∆rij = N − 1 =⇒ ρa = N − N + 1 = 1 (6.34)
This well known result is rediscovered by observing that in such a case the
set s∆ri,j will contain at least two elements and the matrix A will be rank
deficient since all its rows will be equal.
By setting the value of ∆rij = |ri −rj | in the above equation to be equal
to the biggest possible difference between two array sensor locations, which
is obviously the aperture la = ∆r1,N , and by combining the above with the
Theorem 6.2(a), the following sufficient condition can be obtained.
Corollary 6.2 A SUFFICIENT condition for the presence of ambiguities
in ALL linear arrays of N sensors, is
la > N − 1 where la = max ∆rij = ∆r1,N (6.35)

∀i,j
It should be clear however, that the above provides a sufficient but

by no means necessary condition for the presence of ambiguities. This
means that a linear array (or ELA) can possibly suffer from ambiguities
even if la ≤ N − 1 half wavelengths. Only for the specific case of uniform
linear (or ELA) arrays, has it been proven that no ambiguities exist if
la ≤ N − 1 or, equivalently, if the intersensor spacing is not greater than
half a wavelength.
Theorem 6.2 has an implication that needs to be stressed. It is well
known that an increase in the aperture results in better resolution capabili-
ties. However, from the point of view of ambiguities, increasing the aperture
may not be a very good idea. By increasing the aperture, the number of
elements in the set of Eq. (6.23) increases, and once this number becomes
greater than N − 1, the set becomes ambiguous. Furthermore, increasing
all the intersensor spacings may be considered an even more serious issue
since an array with sensor locations given by ř = br, with b > 1, might
have many intersensor spacings ∆rij which result in ambiguous sets of the
form given in Eq. (6.23).
Example 6.2 Consider an array with sensor locations given by r =

[−2.1, −0.9, 0.9, 2.1]T . The manifold length lm of this array is calculated
to be 20.3016. Furthermore, the only absolute difference between sensor
locations satisfying c N is the aperture ∆r1,4 = 4.2. This provides the
following uniform basic set of arc lengths:
s∆r1,4 = [0, 4.8337, 9.6674, 14.5011, 19.3348]T
According to Theorem 6.2, any subset of 4 elements (N = 4) from the above

set will be an ambiguous set and the matrix A with columns the manifold
vectors corresponding to such a subset will be singular. Furthermore, it can
be shown that no matrix A corresponding to any three elements of s∆r1,4
is rank deficient. This means that four different ambiguous generator sets
can be identified from s∆r1,4 which are all the subsets of s∆r1,4 with their
first element zero and three non-zero elements of s∆r1,4 . As expected from
Theorem 6.2, the set s∆r1,4 is ambiguous since c ≥ N . However, ambiguous
sets might also be defined from Theorem 6.2 when c < N . To see this,
consider the uniform basic set for the difference ∆r1,3 = ∆r2,4 = 3 (i.e.
k∆r1,3 = 2), which is c = 3 < N = 4
T
s∆r1,3 = s∆r2,4 = [0, 6.7672, 13.5344]
=⇒ θ∆r1,3 = θ∆r2,4 = [0◦ , 70.53◦ , 109.47◦ ]T
If the matrix A with columns the corresponding manifold vectors is

obtained, then rank(A) = 2 and hence the above set is an ambiguous
set with rank of ambiguity equal to 2. It should also be noted that not all
the uniform basic sets s∆ri,j resulting from a difference ∆rij between two
sensor locations are ambiguous if c < N . To see this, consider the differ-
ences ∆r1,2 = ∆r3,4 = 1.2 (or ∆r2,3 = 1.8) of the previous array which
T T
results in the set s∆r1,2 = [0, 16.9180] (or s∆r2,3 = [0, 11.2787] ) (i.e.
c = 2 < N = 4) which is unambiguous.
So far it has been shown that Theorem 6.2 can be used in order to
identify ambiguous sets inherent in linear arrays of any geometry. When
this theorem is focused on specific array geometries it can produce some
more useful results as the following corollaries indicate.
Corollary 6.3 All the ambiguous sets that exist in a two-element array
can be calculated from Eq. (6.23).
Proof. The proof is as follows. Consider the two-element array r =
[−r1 , r1 ]. Let s = [s1 , s2 ]T with s1 = 0 being an ambiguous set of arc
lengths. The matrix A with columns the manifold vectors corresponding
to s is
A = [a(0), a(s2 )]
 
s2
 exp(jπr 1 ) exp jπr 1 1 − 
= π r  (6.36)
 s2 
exp(−jπr1 ) exp −jπr1 1 −
π r
Since it is assumed that s is an ambiguous set, it results that A is singular
and therefore

s2
det(A) = exp(jπr1 ). exp −jπr1 1 −
π r

s2
− exp(−jπr1 ). exp jπr1 1 − =0
πr

s2 s2
=⇒ exp jπr1 − exp −jπr1 =0
πr πr

s2
=⇒ sin r1 =0
r
s2
=⇒ r1 = kπ with k = 1, 2, 3, . . . (since s2 r1 = 0)
r
πr 2πr 2πr lm
=⇒ s2 = k =k =k =k
r1 2r1 ∆r1,2 ∆r1,2
That is
lm
s1 = 0 and s2 = k (6.37)
∆r1,2
with
k = 1, 2, 3, . . .
Therefore the only ambiguous set of arc lengths that can possibly exist for
a two-element array are of the form of Eq. (6.23). As a result, a necessary
and sufficient condition for the presence of ambiguities in a two element
linear array is
lm
< lm =⇒ ∆r12 > 1 (6.38)
∆r12
√ √ √
Corollary 6.4 The set of arc lengths s = [0, π 2, 2π 2, . . . , (c−1)π 2]T
is an ambiguous Uniform Basic Set for all three-element symmetric linear
arrays, as long as ∆r1,3 > 2.
Proof. For all three-element symmetric linear arrays with sensor pos-
itions given by r = [−r1 , 0, r1 ]T , the manifold length is equal to
6 √
lm = 2πr = 2π (−r1 )2 + (r1 )2 = 2πr1 2 (6.39)
The Uniform Basic Set s∆r1,3 associated with the difference between the
first and the third sensors is
> ?T
s = 0, ∆rlm1,3 , 2 ∆rlm1,3 , . . . , (c − 1) ∆rlm1,3
> √ √ √ ?T
= 0, 2πr1 2
2r1 , 2 2πr2r11 2 , . . . , (c − 1) 2πr2r112
√ √ √ T
= 0, π 2, 2π 2, . . . , (c − 1)π 2 (6.40)

Corollary 6.5 Let two arrays of N sensors have a common difference
∆r between two sensor locations. The Uniform Basic Set of arc lengths
associated with the common ∆r is different for these two arrays but the
corresponding ambiguous sets of DOA’s are the same.
Proof. Consider two different linear arrays (or ELAs) each having N
sensors with sensor locations r1 and r2 . Let there be a common dif-
ference ∆r between two sensor locations, which, for both arrays, satisfy
the conditions of Theorem 6.2. In this case, the two Uniform Basic Sets
will be:
First array:
> ?T
s1 = 0, 2πr 1 2πr 1 2πr 1
∆r , 2 ∆r , . . . , (c − 1) ∆r
(6.41)
Second array:
> ?T
s2 = 0, 2πr 2 2πr 2 2πr 2
∆r , 2 ∆r , . . . , (c − 1) ∆r
(6.42)
with c N , which implies that the above two sets are ambiguous. Using
Eq. (2.28) the corresponding set of DOA’s for the first array is
4 2πr 1
2πr 1
5T
θ1 = 0, arccos 1 − πr , . . . , (c − 1) arccos 1 − πr
∆r ∆r
1 1
T
= 0, arccos 1 − 2
∆r , . . . , (c − 1) arccos 1 − 2
∆r
(6.43)
whereas, for the second array

4 2πr 2
2πr 2
5T
θ2 = 0, arccos 1 − πr , . . . , (c − 1) arccos 1 −
∆r ∆r
πr 2
2
T
= 0, arccos 1 − 2
∆r , . . . , (c − 1) arccos 1 − 2
∆r
(6.44)
The two sets are obviously identical.
6.3.1 Calculation of Ambiguous Generator Sets of Linear

(or ELA) Array Geometries
For a given linear array (or ELA) with sensor locations given by r and
based on Theorem 6.2, the following technique (presented in a step by step
form) provides a table with rows all the ambiguous generator sets, and their
associated rank of ambiguity:
(1) For a given array with sensor locations r, calculate the manifold length
lm = 2πr.
(2) Calculate the intersensor spacings by taking the Kronecker difference
between the vector r (with elements the sensor locations) and itself,
i.e. r r, which results in a (N 2 × 1)-dimensional real vector. Create a
new vector ∆r by keeping all the positive elements of r r. Due to the
properties of the Kronecker difference N elements of r r will be equal
to zero and half of the remaining elements will be negative. Therefore,
∆r ∈ Rd×1 , where d = N2 (N − 1).
(3) Redefine ∆r by eliminating all elements which are smaller than one, i.e.
d ≤ N2 (N − 1). Note that the elements of ∆r which are smaller than
one provide
c = 1 + fix(∆rij ) = 1 + 0 = 1 (6.45)
and therefore cannot possibly give rise to ambiguous generator sets.

(4) For each element of ∆r, i.e. for each difference ∆rij = |ri − rj | > 1
between two sensor locations, construct the “Uniform Basic Set” vector
s∆rij using Eq. (6.23). Note that there are d different vectors s∆rij and
that each s∆ri,j corresponds to a different uniform partition of the array
manifold.
(5) Identify all the vectors s∆rij , with c < N and then eliminate those for
which:
rank(A(s∆rij )) = c < N
Form the “Table of Ambiguous Uniform Basic Sets” with rows all the
remaining vectors s∆rij . All these vectors are ambiguous sets and pro-
duce ambiguous generator sets. It is obvious that the set with the max-
imum number of elements is the set s∆r1,N .
(6) For each s∆rij (i.e. for each row of the “Table of Uniform Basic Sets”)
construct the AGSs (and associated rank of ambiguity), according to
the following rules:
Rule (a) If the non-zero elements of s∆rij cannot be found in any
other “Uniform Basic Set” vector, then

c−1
ambiguous generator sets
N −1
can be produced by this s∆rij . These ambiguous genera-
tor sets are all the possible subsets of N elements of s∆ri,j
with their first element zero and N − 1 non-zero elements
of s∆rij .
All ambiguous generator sets constructed

in this way have rank of ambiguity
ρa = N − 1
Note that such rows must definitely have c ≥ N (see

Theorem 6.2), otherwise they would have been eliminated
in Step 5.
Rule (b) If the non-zero elements of s∆ri,j can be found in other rows,
then ambiguous generator sets with rank of ambigu-
ity less than N − 1 might be obtained. This means
that all subsets of s∆ri,j with their first element 0 and with
lengths 2, 3, . . . , min(N −1, c−1) must be considered. These
subsets are classified as ambiguous generator sets if the three
conditions of the ambiguous generator set definition (see
Definition 6.5) are satisfied. Furthermore, for each ambigu-
ous generator set s, rank of ambiguity ρa = rank(A(s∆ri,j ))
is estimated. Note that this step clarifies why in Step 5 it
is incorrect to eliminate two s∆ri,j vectors which are the
same, although they will result in the same ambiguous gen-
erator sets.
(7) Create the Ambiguous Generator Table of the array whose rows are
all the different ambiguous generator sets (and their associated rank
of ambiguity) found in Step 6. Also, eliminate rows of s∆ri,j which are

duplicates.
6.4 Representative Examples
An indicative example of the above technique is presented below.

Example 6.3 The steps of the previously described technique for an
array of N = 5 sensors with sensor positions r = [−3, −2.2, 1.1, 1.9, 2.2]T
are as follows:
(1) The manifold length of this array is computed to be lm = 30.4589.

(2) The Kronecker difference between r and itself, reshaped in the form of
an N × N matrix, is
 
0, −0.8, −4.1, −4.9, −5.2,
0.8, 0, −3.3 −4.1 −4.4,
 
4.1, 3.3, 0, −0.8, −1.1 
 
4.9, 4.1, 0.8, 0, −0.3 
5.2, 4.4, 1.1, 0.3, 0
(3) Eliminating those entries of r r which are smaller than unity results in
 
∆r1,3 = 4.1
∆r = 4.9
 1,4 
∆r = 5.2
 1,5 
 
∆r = ∆r2,3 = 3.3 (6.46)
 
∆r2,4 = 4.1
 
∆r2,5 = 4.4
∆r3,5 = 1.1
(4) The seven Uniform Basic Sets (vectors) s∆ri,j are given in Table 6.2.
(5) The Uniform Basic Sets s∆rij which have less than N = 5 elements (i.e.
c < N ) are examined. These are the sets s∆r2,3 and s∆r3,5 .
• The number of elements in s∆r2,3 is 4 < N = 5 and, furthermore,
rank (A(s∆r2,3 )) = 4 = c and hence this row is eliminated.
• Also, s∆r3,5 is eliminated since c = 2 < N = 5 and
rank(A(s∆r2,3 )) = 2 = c. (6.47)
Thus the remaining sets are s∆r1,3 , s∆r1,4 , s∆r1,5 , s∆r2,4 and s∆r2,5 .
(6) In this step, the AGSs arising from the remaining Uniform Basic Sets
are calculated based on the rules (6a) and (6b).
Table 6.2 Uniform Basic Sets of ∆rij > 1.
s1 s2 s3 s4 s5 s6
sT
∆r1,3 0 7.4290 14.8580 22.2870 29.7160 —
sT
∆r1,4 0 6.2161 12.4322 18.6483 24.8644 —
sT
∆r1,5 0 5.8575 11.7150 17.5725 23.4300 29.2875
sT
∆r2,3 0 9.2300 18.4600 27.6900 — —
sT
∆r2,4 0 7.4290 14.8580 22.2870 29.7160 —
sT
∆r2,5 0 6.9225 13.8450 20.7675 27.6900 —
sT
∆r3,5 0 27.6900 — — — —
The non-zero elements of s∆r1,4 , s∆r1,5 and s∆r2,5 cannot be found

in any other s∆ri,j and hence should be investigated based on rule 6a.
The non-zero elements of s∆r1,3 can also be found in s∆r2,4 and vice-
versa and hence should be investigated based on rule 6b. The above are
summarized in Table 6.3.
Table 6.3 Rules for examining the UBSs of Table 6.2.
c ρa rule
sT
∆r1,3 5 3 (6b)
sT
∆r1,4 5 4 (6a)
sT
∆r1,5 6 4 (6a)
sT
∆r2,3 4 4 removed
sT
∆r2,4 5 3 (6b)
sT
∆r2,5 5 4 (6a)
sT
∆r3,5 2 2 removed
Based on Rule (6a)
• s∆r1,4 : Since the number of the elements of s∆r1,4 is 5 (c = 5),

4
we have Nc−1 −1 = 4 = 1 which implies that only one ambiguous
generator set (which is the same as the Uniform Basic Set s∆r1,4 )
can be identified, with ρa = 4. 5
• s∆r1,5 : This time c = 6 and therefore Nc−1
−1 = 4 = 5 ambiguous
generator sets can be found by taking all the subsets of N elements
of s∆r1,5 (with one zero and N − 1 non-zero elements). Each of these
sets has ρa = 4.
4
• s∆r2,5 : Finally, s∆r2,5 has c = 5 and it generates Nc−1
−1 = 4 = 1
ambiguous generator set. This AGS is the same as set s∆r2,5 and has
ρa = 4.
Based on Rule (6b)
• s∆r1,3 : The non-zero elements of s∆r1,3 can also be found in s∆r2,4

and hence s∆r1,3 is investigated based on rule 6b. The ambiguous
generator set definition is initially checked for the set s∆r1,3 . This
set is certainly ambiguous and the matrix with columns the corre-
sponding manifold vectors has
& '
rank A(s∆r1,3 ) = 3 < c = 5 (6.48)
However, any one of the four subsets š with four elements of s∆r1,3
involving one zero and three non-zero elements, is also an ambiguous
set. The matrix with columns the corresponding manifold vectors has
rank(A(š)) = 3 < c = 5. Therefore, the set s∆r1,3 satisfies the third
condition of the definition of the ambiguous generator set. Similarly,
the four subsets š are ambiguous generator sets since all their sub-
sets with three elements are unambiguous. Hence, the following five
ambiguous generator sets can be defined from s∆r1,3 , all of which
have ρa = 3. Indeed
min(N −1,c−1) min(4,4)

c−1 5−1
=

=N −k∆rij =5−2
4

4 4 4
= = + = 5 (6.49)

3 4
=3
• s∆r2,4 : Since s∆r2,4 is the same as s∆r1,3 , it follows that the ambigu-
ous generator sets obtained from this row are the ones which were
already obtained.
(7) The ambiguous generator table, Table 6.4, with rows the AGSs pro-
duced in Step-6 is now constructed. Their associated rank of ambiguity
is also shown.
In conclusion, twelve ambiguous generator sets can be identified for

this array. Five of these have ρa = 3, while the remaining have ρa = 4,
constructed from UBSs as shown below:
AGS UBS
1st to 5th s∆r1,3 ; s∆r2,4

6th s∆r1,4
7th to 11th s∆r1,5
12th s∆r2,5
Table 6.4 List of AGSs and their associated rank of

ambiguity.
s1 s2 s3 s4 s5 ρa
(1) 0 7.4290 14.8580 22.2870 29.7160 3

(2) 0 7.4290 14.8580 22.2870 — 3
(3) 0 7.4290 14.8580 29.7160 — 3
(4) 0 7.4290 22.2870 29.7160 — 3
(5) 0 14.8580 22.2870 29.7160 — 3
(6) 0 6.2161 12.4322 18.6483 24.8644 4
(7) 0 5.8575 11.7150 17.5725 23.4300 4

(8) 0 5.8575 11.7150 17.5725 29.2875 4
(9) 0 5.8575 11.7150 23.4300 29.2875 4
(10) 0 5.8575 17.5725 23.4300 29.2875 4
(11) 0 11.7150 17.5725 23.4300 29.2875 4
(12) 0 6.9225 13.8450 20.7675 27.6900 4
6.5 Handling Ambiguities in Planar Arrays
The material presented in this chapter is valid for any hyperhelical curve
(of the form of Eq. (6.2)) embedded in CN . However, for p-curves Ap|qo
(p = φ, α or β) some extra steps should be taken to accommodate the
property that these curves of hyperhelical shape are lying on the manifold
surface of a planar array. These curves can be seen as the array manifolds
of an ELA with location given by r(Ψ0 ) represented by Eqs. (6.3) and (6.4).
Therefore, the Uniform Basic Sets s∆rij (see Eq. (6.23)), and consequently
the AGSs, can be constructed according to the procedure described in the
previous section with
s∆rij s∆rij (Ψ0 )

> ?T
2πr(Ψ ) 2πr(Ψ ) 2πr(Ψ )
= 0, ∆rij0 , 2 ∆rij0 , . . . , (c − 1) ∆rij0 (6.50)
∀i,j with i = j
where ∆rij represents the intersensor spacing between the ith and jth sen-
sors of the ELA r(Ψ0 ).
Before we proceed it is important to point out that
(1) ambiguous sets of directions with a common azimuth are located on

φ-curves, while
(2) ambiguous sets of directions of different azimuth and elevation angles
are located on the α-curves (or β-curves) for a rotation Θ.
6.5.1 Ambiguities on φ-curves

Equation (4.24) of Chapter 4 which is reshown below
a(φ | θo ) = a(φ | θo + 180◦ ) (6.51)
implies that the two φ-curves Aφ|θo and Aφ|θo +180◦ have the same length
and curvatures and can be considered as a continuation of one another
forming a composite φ-curve having a hyperhelical shape. The family of
these composite curves (from now on known as ϕ-curves) can be described
with an alternative but equivalent parameter space as
Ω = {(ϑo , ϕ) : ϑo ∈ [0◦ , 180◦ ), ϕ ∈ [0◦ , 180◦ )} (6.52)

i.e.
Ωϑ = Ωϕ = [0◦ , 180◦ ) (6.53)
and is defined as

Aϕ|ϑo , ∀ϑo : ϑo ∈ Ωϑ (6.54)
where
Aϕ|ϑo = {a(ϑo , ϕ) ∈ CN , ∀ϕ : ϕ ∈ Ωϕ , ϑo = constant} (6.55)
It is obvious now that a ϕ-curve represents directly the array manifold of

an Equivalent Linear Array (ELA) with sensor locations given by
r(Ψ0 ), with Ψ0 ϑ0 (6.56)
having a manifold length
(lm of Aϕ|ϑo ) = 2πr(ϑo ) (6.57)
This implies that every member of the family of ϕ-curves of a planar

array has its own ELA. Based on the above discussion the transformation
of the parameters from
{(θo , φ) : θo ∈ [0◦ , 360◦ ), φ ∈ [0◦ , 90◦ )} (6.58)

to
{(ϑo , ϕ) : ϑo ∈ [0◦ , 180◦ ), ϕ ∈ [0◦ , 180◦ )} (6.59)
is essential before using the technique described in Section 6.3.1 and is given
as follows:

ϕ=φ
if θo < 180◦ then (6.60)
ϑ o = θo

ϕ = 180◦ − φ
if θo 180◦ then (6.61)
ϑo = θo − 180◦
Furthermore, after the construction of the Ambiguous Generator Table

of arc lengths, the associated composite directions (azimuth, elevation) can
be estimated using Eq. (2.28). Then the (ϑo , ϕ) should be mapped to the
original space of the family of the φ-curves as follows:

φ=ϕ
if ϕ < 90◦ then (6.62)
θo = ϑo

φ = 180◦ − ϕ
if ϕ 90◦ then (6.63)
θo = 180◦ + ϑo
The following example illustrates these additional steps.

Example 6.4 Consider the following planar array of 6 sensors (in half-
wavelengths)
 T
−2.5, −2, −2.5, 2.5, 2, 2.5
rx , ry , 0 =  1.5, 0, −1.5, 1.5, 0, −1.5 (6.64)
0, 0, 0, 0, 0, 0
In order to estimate the ambiguous generator sets associated with the
φ-curve of constant θ equal to 5◦ we will use its corresponding composite
curve (i.e. ϕ-curve for ϑ0 = 5◦ ). This implies that
{(θo , φ) : θo ∈ [0◦ , 360◦ ), φ ∈ [0◦ , 90◦ )} (6.65)

will be
{(ϑo , ϕ) : ϑo ∈ [0◦ , 180◦ ), ϕ ∈ [0◦ , 180◦ )} (6.66)
while the ELA associated with this ϕ-curve is the following:
r(ϑ0 ) = [−2.6212, −2.3598, −1.9924, 1.9924, 2.3598, 2.6212]T

(6.67)
having a manifold length equal to
(lm of Aϕ|5◦ ) = 35.9943 (6.68)
By using Eq. (6.50) in conjunction with the technique described in

Section 6.3.1 four AGSs of arc lengths are obtained as shown in Table 6.5.
These four AGSs, for this specific ELA of ϑ = 5◦ , have been constructed
from the UBS as shown below:
AGS UBS
1st s∆r1,6
2nd s∆r1,5 ;s∆r2,6
3rd s∆r1,4 ;s∆r3,6
4th s∆r2,4 ;s∆r3,5
Note that the directions in degrees associated with the above ambiguous
generator sets of arc lengths can be estimated using Eq. (2.28) of Chapter 2
and are given in Table 6.6.
Table 6.5 Ambiguous generator sets (arc lengths).
s1 s2 s3 s4 s5 s6 ρa
(1) 0 6.8659 13.7319 20.5978 27.4638 34.3297 5

(2) 0 7.2264 14.4527 21.6791 28.9054 — 4
(3) 0 7.8018 15.6035 23.4053 31.2071 — 4
(4) 0 8.2705 16.5410 24.8114 33.0819 — 4
Table 6.6 Ambiguous generator sets (in degrees) corresponding to Table 6.5.
(θ1 , φ1 ) (θ2 , φ2 ) (θ3 , φ3 ) (θ4 , φ4 ) (θ5 , φ5 ) (θ6 , φ6 ) ρa
(1) (5◦ , 0◦ ) (5◦ , 51.79◦ ) (5◦ , 76.29◦ ) (185◦ , 81.69◦ ) (185◦ , 58.26◦ ) (185◦ , 24.84◦ ) 5
(2) (5◦ , 0◦ ) (5◦ , 53.24◦ ) (5◦ , 78.64◦ ) (185◦ , 78.20◦ ) (185◦ , 52.69◦ ) — 4
(3) (5◦ , 0◦ ) (5◦ , 55.49◦ ) (5◦ , 82.36◦ ) (185◦ , 72.51◦ ) (185◦ , 42.78◦ ) — 4
(4) (5◦ , 0◦ ) (5◦ , 57.29◦ ) (5◦ , 85.36◦ ) (185◦ , 67.75◦ ) (185◦ , 33.05◦ ) — 4
6.5.2 Ambiguities on α-curves/β-curves

Next let us examine a curve Aα|βo belonging to the family of α-curves
{Aα|βo , ∀βo : βo ∈ Ωβ } for a fixed rotation Θ = Θo . In other words let us
examine the locus of all manifold vectors over the whole α parameter space
Ωα ≡ [αmin , αmax ] at a particular β0 , i.e.
Aα|βo = {a(α, βo ) ∈ CN , ∀α : α ∈ Ωα , βo = constant} (6.69)
with αmin and αmax given by Eq. (5.12). As we have seen in Chapter 5
this family covers and represents the whole of the manifold surface M of a
planar array while all its members are associated with the same ELA
r(Ψ0 ) with Ψ0 Θo (6.70)
but with different “visible” areas specified by the minimum and maximum
permissible values of α, i.e. the parameter space is Ωa = [αmin , αmax ].
The same is true for the β-curves but with r(Θo + 90◦ ) as their ELA with
parameter space Ωβ = [βmin , βmax ].
A few other points must be emphasized here. Firstly, from Eq. (6.3),
the following relation can be derived:
r(Θ) = −r(Θ + 180◦ ) (6.71)
which implies that for a given rotation Θo of the x–y frame and a given
βo , the resulting α-curve Aα|βo (Θo ) has the same length as the α-curve
Aα|βo (Θo + 180◦ ) resulting from rotation Θo + 180◦ . This is the reason
why only frame rotations Θo in the region [0◦ , 180◦ ] should be considered.
Secondly, as already mentioned, the roles of α and β can be interchanged
by simply replacing the rotation Θo of the x–y frame with Θo + 90◦ and
therefore the β-curves are effectively transposed versions of the α-curves.
This means that the α-curves for some rotation Θo , are the same as the
β-curves for rotation Θo +90◦ . This is the reason the β-curves are neglected
throughout this analysis and only the α-curves are considered.
Based on the above discussion the ambiguous generator sets of arc
lengths of the α-curve of βo = 90◦ , i.e. Aα|90◦ , can be estimated by par-
titioning this curve according to the elements (arc lengths) of Eq. (6.50)
using Ψ0 Θo .
Note that the α cone-angle parameter, corresponding to the arc lengths
sk of the α-curve, can be calculated by the following equation:

sk
αk = arccos sin βo − (6.72)
πr(Θo )
and that the α-curve with βo = 90◦ for a frame rotation Θo has identical
AGSs with the composite φ-curve associated with the composite azimuth
ϑo = Θo .
Indeed if βo = 90◦ then

Θo if α ≤ 90◦
θo = (6.73)
Θo + 180 if α > 90◦
◦

α if α ≤ 90◦
φ= (6.74)
180◦ − α if α > 90◦
Now, based on Eq. (6.50), let us assume that all AGSs of arc lengths of
the α-curve with βo = 90◦ , have been constructed. Then consider another
α-curve of βo different to 90◦ . For this second curve let smin and smax
be the arc lengths corresponding to the parameter angles αmin and αmax
(given by Eq. (5.12)), respectively. This curve has ambiguous sets but not
ambiguous generator sets, since smin = 0 for any β = 90◦ , which implies
that the first condition of the definition of the ambiguous generator set (see
Definition 6.5) is not satisfied. Let us define the ambiguous sets with their
first element equal to smin as the first permissible ambiguous sets. These can
be found by rotating all the ambiguous generator sets of the α-curve with
βo = 90◦ by smin , subject to the condition that the maximum element of
each set cannot exceed smax . Thus it is clear that the ambiguous generator
sets for the whole family of α-curves, for a given frame rotation Θo , can be
provided by examining only the α-curve of βo = 90◦ . The ambiguities of
any other α-curve can be generated by a simple rotation (Theorem 6.1) of
those ambiguous generator sets, making sure that its maximum element sc
is smaller than the manifold length lm of Aα|βo .
Based on the above discussion it is clear that the Ambiguous Generator
Table in arc lengths should be constructed for the α-curve with βo = 90◦ ,
using the technique of Section 6.3.1. Then, for a specific α-curve with
βo < 90◦ only the AGSs should be kept whose last element sc is smaller
than the length of the α-curve given by lmα|βo = 2πr(Θo ) sin βo . These
sets can be expressed from arc lengths to degrees using Eq. (6.72) but using
an additional term smin which takes into account the fact that for βo = 90◦
the minimum value of α is not 0◦ . That is
 smin 

 sk + πr(Θo )(1 − sin βo ) 
αk = arccos 1 −

 (6.75)
πr(Θo )
Note that s = 0 must correspond to αk = arccos(sin βo ) = 90◦ − βo , ∀k.

Furthermore, some of these sets may also be AGSs of various other

α-curves with βo = 90◦ . However the AGS of arc lengths for an α-curve
with βo = 90◦ , although numerically identical to those of an α-curve with
βo = 90◦ (or equivalently to those of φ-curve ϑo = Θo ), correspond to
different sets of directions having different azimuth and elevation angles.
A number of these properties are shown in the following example.
Example 6.5 Consider a planar array whose sensor locations (in half-
wavelengths) are given below:
 T
−2.6, −1.3, 1.3, 2.6
rx , ry , 0 =  2, −0.5, −1.5, 0  (6.76)
0, 0, 0, 0
The Equivalent Linear Array of the family of α-curves {Aα|βo , ∀βo : βo ∈

Ωβ } for a fixed rotation of the x-y frame Θo = 0◦ is (see Eq. (6.3))
 
−2.6
−1.3
r(Θo ) = rx cos Θo + ry sin Θo = 
 1.3
 (6.77)
2.6
The manifold length of this ELA r(Θo ) is lm = 2πr(Θo ) = 25.8299.

By employing the technique of Section 6.3.1, Table 6.7 of the AGSs
can be obtained where the relation between AGSs and UBSs is given in
Table 6.8.
Table 6.7 AGSs (arc lengths) of the ELA of

Eq. (6.77).
s1 s2 s3 s4 ρa
(1) 0 19.8692 — — 1
(2) 0 6.6231 13.2462 — 2
(3) 0 4.9673 24.8365 — 2
(4) 0 9.9346 14.9019 24.8365 3
(5) 0 4.9673 9.9346 14.9019 3
Table 6.8 UBS associated with the AGSs of Table 6.7.
AGS UBS
1st s∆r1,2 ; s∆r2,3 ; s∆r2,4 ; s∆r1,3 ; s∆r1,4 ; s∆r3,4

2nd s∆r1,3 ; s∆r2,4
3rd to 5th s∆r1,4
The sets of Table 6.7 are the ambiguous generator sets for the whole
family of α-curves (since ambiguities of any α-curve with βo < 90◦ will
relate to a subset of the above AGSs). To see this property let us consider
βo = 90◦ . In this case, all the AGSs of Table 6.7 are certainly included
since the length lm of this α-curve Aα|90◦ is equal to the manifold length
of the ELA lm = 25.8299. The values in Table 6.7, transformed in degrees,
are shown in Table 6.9. If a different α-curve of βo < 90◦ is chosen then its
length is
lm = 2πr(Θo ) sin βo = 25.8299 sin βo (6.78)
In this case only the rows in Table 6.7 with the last element less than lm
should be kept. Indeed for the α-curve with βo = 74◦ , i.e. Aα|74◦ , then
lm = 24.8293 and therefore the AGSs will be as shown in Table 6.10.
The α-curve Aα|50◦ , having a length lm = 19.7869 will be associated
with only two out of the five AGS of the family. These are the ones with
their last element smaller than, or equal to, the length of the curve 19.7869
(i.e. s∆r1,3 and s∆r1,3 ) and are given in Table 6.11.
Table 6.9 AGSs (in degrees) of the ELA of Eq. (6.77).
(α1 , βo ) (α2 , βo ) (α3 , βo ) (α4 , βo ) ρa

or or or or
(φ1 , θo ) (φ2 , θo ) (φ3 , θo ) (φ4 , θo )
(1) (0◦ , 90◦ ) (122.58◦ , 90◦ ) — — 1

(2) (0◦ , 90◦ ) (60.84◦ , 90◦ ) (91.47◦ , 90◦ ) — 2
(3) (0◦ , 90◦ ) (52.02◦ , 90◦ ) (157.38◦ , 90◦ ) — 2
(4) (0◦ , 90◦ ) (76.65◦ , 90◦ ) (98.85◦ , 90◦ ) (157.38◦ , 90◦ ) 3
(5) (0◦ , 90◦ ) (52.02◦ , 90◦ ) (76.65◦ , 90◦ ) (98.85◦ , 90◦ ) 3
Table 6.10 AGSs of the ELA of Eq. (6.77) for βo = 74◦ .
(α1 , βo ) (α2 , βo ) (α3 , βo ) (α4 , βo ) ρa
(1) (16◦ , 74◦ ) (125.25◦ , 74◦ ) — — 1

(2) (16◦ , 74◦ ) (63.36◦ , 74◦ ) (93.69◦ , 74◦ ) — 2
(5) (16◦ , 74◦ ) (54.79◦ , 74◦ ) (78.93◦ , 74◦ ) (101.10◦ , 74◦ ) 3
(φ1 , θ1 ) (φ2 , θ2 ) (φ3 , θ3 ) (φ4 , θ4 ) ρa
(1) (16◦ , 0◦ ) (154.47◦ , 50.24◦ ) — — 1

(2) (16◦ , 0◦ ) (31.58◦ , 58.24◦ ) (103.15◦ , 73.56◦ ) — 2
(5) (16◦ , 0◦ ) (25.55◦ , 50.27◦ ) (55.14◦ , 70.37◦ ) (124.94◦ , 70.35◦ ) 3
(α1 , βo ) (α2 , βo ) (α3 , βo ) (α4 , βo ) ρa
(2) (40◦ , 50◦ ) (75.33◦ , 50◦ ) (105.05◦ , 50◦ ) — 2

(5) (40◦ , 50◦ ) (67.58◦ , 50◦ ) (90.18◦ , 50◦ ) (112.82◦ , 50◦ ) 3
(φ1 , θ1 ) (φ2 , θ2 ) (φ3 , θ3 ) (φ4 , θ4 ) ρa
(2) (40◦ , 0◦ ) (68.50◦ , 46.30◦ ) (112◦ , 46.11◦ ) — 2

(5) (40◦ , 0◦ ) (59.32◦ , 41.63◦ ) (90.28◦ , 50◦ ) (121.10◦ , 41.35◦ ) 3
(α1 , βo ) (α2 , βo ) (α3 , βo ) (α4 , βo ) ρa
(2) (55◦ , 35◦ ) (86.52◦ , 35◦ ) (116.88◦ , 35◦ ) — 2
(φ1 , θo ) (φ2 , θo ) (φ3 , θo ) (φ4 , θo ) ρa
(2) (55◦ , 0◦ ) (85.76◦ , 34.77◦ ) (118.89◦ , 20.67◦ ) — 2
Finally, Table 6.12 gives the only AGS associated with the α-curve
Aα|35◦ (curve length lm = 14.8154), while no ambiguity can be identified
for any α-curve with βo < 30◦ or βo > 150◦ .
All these are shown in Fig. 6.4 which plots the length of the α-curves
versus the value of βo where the darker the shade of the area, the larger
the number of ambiguous sets that can be identified.
Note that in order to have a set of M (azimuth, elevation) directions
belonging to the same α-curve, there should be a frame rotation Θ = Θo
such that
arccos(cos φ sin(θ − Θ1M )) = k1M (6.79)
where k is a constant. Then these directions which satisfy Eq. (6.79) belong
to the same α-curve Aα|βo , i.e correspond to a common βo = k for this
rotation Θo . For instance this can be seen for the following planar array:
 T
−1, 2, −4, 3
rx , ry , 0 =  0, 0.5, 1.5, −2 (6.80)
0, 0, 0, 0
Fig. 6.4 Manifold length for the family of α-curves. In the region of the darker shade
the number of AGS is 5 while no AGS can be identifed in the white region (0 AGSs).
operating in the presence of four signals with directions:
1st DOA 2nd DOA 3rd DOA 4th DOA
θ 47◦ 54.72◦ 129.57◦ 203.36 (6.81)

φ 0◦ 59.09◦ 82◦ 58.91◦
Indeed in this case Eq. (6.79) is satisfied for Θ = Θo = 39◦ with constant
k = 82◦ (= βo ) as can be seen from Fig. 6.5.
This indicates that these directions are on the same α-curve which is
the manifold of the ELA r(Ψ0 ) with Ψ0 = Θo , i.e.
r(Θo ) = rx cos Θo + ry sin Θo (6.82)
This α-curve (manifold) can be found using Eqs. (5.3) and (5.4) to trans-
form the set of (θ, φ) to cone-angles for the frame rotation Θo . Thus for
Θo = 39◦ the directions (θ, φ) are transformed to cone-angles (α, β) given
Fig. 6.5 Plot of the four equations given by Eq. (6.79) for different rotation Θ.
in the following table:
1st DOA 2nd DOA 3rd DOA 4th DOA
α 8◦ 60.37◦ 90.08◦ 119.82◦

β 82◦ 82◦ 82◦ 82◦
This implies that the common βo is 82◦ for this rotation Θo (i.e. belong to
the same α-curve Aα|βo ). This curve is the manifold of the following ELA
r(Ψ0 ) with Ψ0 = Θo :
T
r(39◦ ) = rx cos 39◦ + ry sin 39◦ = −2.1646, −0.7771, 1.0728, 1.8690
(6.83)
Note that if there exists an intersensor difference ∆rij = |ri (Θo ) − rj (Θo )|
for which all elements of α satisfy

2
α = arccos sin β0 1M − [1, 2, . . . , M ] T
(6.84)
∆rij
then this set is an ambiguous set.

6.5.3 Some Comments on Planar Arrays

The concept of the ELA has proved to be extremely useful in analyzing
planar array ambiguities by allowing the re-use of the results derived for lin-
ear arrays. However, one situation which does not normally occur in linear
arrays but is nevertheless common in the ELA is the collocation of projected
sensors, resulting from sensors lying along the same perpendicular to the
line of the ELA. Such an ELA can be analyzed by first eliminating the col-
located sensors and identifying the ambiguities of the underlying structure.
Consider for instance the case of an ELA of N -sensors consisting of

collocated sensor pairs. The first step is to analyze the N ) -sensor array
)
(with N = N −
) obtained by removing the collocations. Suppose this
array is found to suffer from uniform rank-(N ) − k) ambiguity, i.e. the ELA
suffers from rank-(N −
− k) ambiguity. Recall that uniform rank-(N ) − k)
ambiguity is caused by the occurrence of identical rows in the determinant
matrix. Now,
collocated sensor pairs will give rise to
identical rows in
the determinant matrix, so that the N -sensor ELA will also suffer from
rank-(N ) − k) or, equivalently, rank-(N −
− k) ambiguity.
Note that an ELA has two or more collocated sensors when for one
or more different pairs of sensors the parameter Ψ0 satisfies the following
equation:
rxi − rxj
tan Ψ0 = − , i = j (6.85)
ryi − ryj
where Ψ0 is given by Eq. (6.4), Ψ0 ∈ [0, 180◦ ) and (rxi , ryi ), (rxj , ryj )
represent the locations of the ith and j th sensors. It can be proven that, for
a given array geometry, the maximum number nmax of ELAs with collocated
sensors, is given by:

N2 − N m
nmax (r) = −2 δ1 − 2mδ2 (6.86)
2 2
with m being the number of symmetric pairs with respect to the origin in
the planar array,

1 if m > 1
δ1 = (6.87)
0 otherwise
and

1 if the array has a sensor at the origin
δ2 = (6.88)
0 otherwise
To better illustrate the condition of collocated sensors, consider

for instance a uniform circular array (UCA) with 8 elements and 1.5
half-wavelengths intersensor spacing. Equation (6.86) will give us that the
maximum number of ELAs with collocated sensors is nmax = 8. Indeed the
array has eight ELAs with collocated sensors corresponding to Ψ0 = 0◦ ,
22.5◦ , 45◦ , 67.5◦ , 90◦ , 112.5◦ , 135◦ and 157.5◦ . In Fig. 6.6, the ELA for
Ψ0 = 45◦ is presented (with the circles) and has only one ambiguous gen-
erator set of rank two. Furthermore in this figure the number of ambiguous
generator sets over the whole parameter space are shown grouped by rank.
The above example is a characteristic one which illustrates the importance
of the array geometry in the ambiguity problem and in the presence of
lower rank ambiguities. Note that all symmetric array geometries (grid,
X, Y shaped arrays, etc.) have similar properties to those of the uniform
circular array.
Within this new framework, the results reported so far concerning ambi-
guities present in linear array manifolds are directly applicable to the
p-curves of planar array manifolds via the concept of the equivalent linear
Fig. 6.6 ELA with co-located sensors.

array. Based on similar arguments to the linear array case the following
statements about ambiguities inherent in planar arrays can be made:
Proposition 6.1
(a) A planar array is bound to suffer from rank-(N − 1) ambiguity if two

of its sensors are at least (N − 1) half-wavelengths apart.
(b) Furthermore, rank-(N − k) ambiguity will exist if there exists an ELA,
where k intersensor spacings are integer multiples of the smallest among
them, and the latter is at least (N − k).
Let us now investigate the implications of condition (b) on the planar

array configuration,
r = [rx , ry , 0N ]
= [r1 , r2 , . . . , rN ]T (6.89)
where ri = [xi , yi , 0]T is the position of the ith sensor in half-wavelengths.

Consider the ELA r(Ψ0 ) rx cos Ψ0 + ry sin Ψ0 for the case when k = 2,
say ∆rpq = n∆rij , n ∈ Z + and ∆rij (N − 2). Then the condition
∆rpq = n∆rij implies that

rp − rq = ± n ri − rj , n ∈ Z + (6.90)
Geometrically, this is equivalent to the line joining the pth and qth sensors
and that joining the ith and jth sensors, being parallel, with the length of
the former being an integer multiple of the latter. A special case is when
∆rpq = n∆rqj ; n ∈ Z + , resulting in

rp − rq = ± n rq − rj , n ∈ Z + (6.91)
In geometrical terms, this means that the three sensors are colinear, with
the spacing between the pth and qth sensors being an integer multiple of
that between the ith and jth sensors. Therefore, in general (k 2).
Proposition 6.2 A planar array will suffer from rank-(N − k) ambiguity
if either
(a) the spacings between k pairs of sensors are integer multiples of the
smallest among them, the latter being at least (N − k) half-wavelengths,
and the lines joining the sensor pairs are parallel,
or
(b) (k+1) sensor locations (vectors) are colinear and the k intersensor spac-
ings between adjacent sensor pairs are integer multiples of the smallest
among them, with the latter being at least (N − k) half-wavelengths.
Note that the same factors which contribute to the growth of the number
of ambiguous generator sets in linear arrays are applicable to planar arrays
via the equivalent linear array.
Corollary 6.6 The causes of ambiguities in planar arrays are namely
the colinearity of the sensor locations, uniformity in the sensor spacings
and the size of the common spacing.
6.5.4 Ambiguous Generator Lines

We have seen that for a given ELA r(Ψ0 ) the AGSs can be constructed
using the technique presented in Section 6.3.1. It is also evident from the
previous sections that in a planar array, the whole family of α-curves is
represented with only one ELA, with Ψ0 being the frame rotation Θo , i.e.
ELA of the family of α-curves = r(Θo ) (6.92)
while an infinite number of ELAs are associated with the family of φ-curves

set of ELAs of the family of φ-curves = r(θo ) ∈ RN , ∀θo ∈ Ωθ
(6.93)
By varying Ψ0 (θo or Θo ) over its parameter space ΩΨ0 = [0◦ , 180◦ ) we

will get the set of ambiguous generator lines of the whole family of p-curves
to which the elements of the ambiguous generator sets belong.
Example 6.6 Let us consider the array of Example 6.4. The first ambigu-
ous generator set of this array, for this specific ELA of ϑ = 5◦ , has rank of
ambiguity equal to 5 and is given below:
AGS = [0, 6.8659, 13.7319, 20.5978, 27.4638, 34.3297]
while its elements are shown as dots in Fig. 6.7. Furthermore, in the same
figure the locus of the manifold lengths of all ELAs, i.e. {lm (Ψ0 ), ∀Ψ0 } is
also shown.
By varying Ψ0 over its parameter space ΩΨ0 = [0◦ , 180◦ ) we will get
the set of ambiguous generator lines of the φ-curves to which the elements
of the ambiguous generator sets of arc lengths belong.
In these lines the AGS (in arc length) are associated with all ELA
r(Ψ0 ), ∀Ψ0 with Ψ0 = Θo or ϑo . This implies that, for variable Θo in
the region 0◦ to 180◦ , and for the array of Eq. (6.64) the ambiguous
generator lines of all families of α-curves are numerically, in arc lengths,
equal to those of the family of φ-curves. Thus for the array of Eq. (6.64),
Fig. 6.7 The locus of the manifold lengths of ELA of Example 6.4. The dots represent
the 1st AGS (rank 5) given in Table 6.5.
Fig. 6.8 The set of ambiguous generator lines of rank 5 in which the elements of the
1st AGS of Table 6.5 (Example 6.4) belong.
Figs. 6.8–6.11 not only represent the set of ambiguous generator lines of
the φ-curves but also of the α and β-curves by generalizing the polar axes
to (Ψ0 , s).
2nd AGS of Table 6.5 (Example 6.4) belong.
3rd AGS of Table 6.5 (Example 6.4) belong.
Thus, Fig. 6.8 shows the set of ambiguous generator lines of rank 5 in
which the first row of Table 6.5 belongs. As it can be seen in this figure,
the set of ambiguous generator lines ceases to exist at θo = 0 and for values
of θo in the region 61◦ to 119◦ while for θo in the region 119◦ to 180◦ its
4th AGS of Table 6.5 (Example 6.4) belong.
values are the mirror image of the values from 0◦ to 61◦ . Note that the set
of ambiguous generator lines is defined only in those areas of the parameter
space Ωθ, at which the last line (e.g. the 6th line in Fig. 6.8) in which the
largest element of the associated AGS is located, is below the locus of the
manifold length of all ELAs.
In addition, in Figs. 6.9–6.11, the sets of ambiguous generator lines in

which the 2nd, 3rd and 4th rows of ambiguous generator table belong, are
shown. Finally, in Fig. 6.12 three “discrete” ambiguous generator sets are
shown as “squares.” Two of these discrete AGSs are of rank 2 and one of
rank 1, appear at θo = 45◦ , 135◦ and 90◦ having the following values:
θo = 45◦ : [0, 13.5732, 27.465]; ρa = 2

θo = 135◦ : [0, 13.5732, 27.1465]; ρa = 2
θo = 90◦ : [0, 12.5664]; ρa = 1
6.6 Ambiguities and Manifold Length
It must be apparent by now that the existence of ambiguities is intimately

related to the topology of the array manifold and especially to the length of
the manifold curve. Intuitively, longer manifolds are more prone to ambigu-
ous situations. Consider the set of arc lengths resulting from the uniform
partitions of the manifold length lm by the spacing between the ith and jth
Fig. 6.12 “Discrete” AGSs of corresponding to ELA of 45◦ , 90◦ and 135◦ for the array
of the Example 6.4.
sensors of an N -sensor linear array:

4 5
lm 2lm (c − 1) lm
s = 0, , ,..., (6.94)
∆rij ∆rij ∆rij
Now lm = 2πr and hence the set s can be rewritten as
4 5
2π 4π 2π(c − 1)
s∆r̃ij = 0, , ,..., (6.95)
∆r̃ij ∆r̃ij ∆r̃ij
where
ri
∆r̃ij = |r̃i − r̃j | with r̃i
r
For s∆r̃ij to be a rank-(N − 1) ambiguous generator set, it must consist of
at least N terms and therefore
2(N − 1)π
lm (6.96)
∆r̃ij
Considering the largest value of ∆r̃ij , a sufficient condition for the N -sensor
array to suffer from rank-(N − 1) ambiguity is then given by
2 (N − 1) π
lm (6.97)
∆r̃1,N
Recall that the eigenvalues of the Cartan matrix are equivalent to j times
the normalized sensor locations of the corresponding array and mirror array.
Fig. 6.13 All the sets of ambiguous generator lines of the array of Example 6.4 shown
together with the “discrete” AGSs.
In the case of a symmetric linear array, the normalized sensor locations are
simply given by j times the eigenvalues of the corresponding Cartan matrix
(since the array and its mirror image about the centroid are identical).
Therefore, in the symmetric linear array case, the denominator ∆r̃1,N in
Eq. (6.97) is equal to the eigenvalue spread of the Cartan matrix-spread(C).
Note that the spread of a matrix A with eigenvalues denoted by λ, is
defined as:
spread (A) max |λi − λj | (6.98)
i,j
and it satisfies the following inequalities [22]:

√
spread(A) 2AF (6.99)
spread(A) max (|aij | + |aji |), if A is normal (6.100)
i,j
Recall that the Cartan matrix is a real normal matrix (CCT = CT C), which
in the case of a symmetric array, is given by
 
0 −κ1 0 ... 0
κ1 0 −κ2 ... 0
 
0 κ2 0 ... 0
C=  (6.101)
. .. .. .. .. 
 .. . . . . 
0 0 0 κN −1 0
where κ1 , known as the first or principal curvature, is the largest of all

the curvatures. Also, from Eq. (2.76) (page 52), CF = 1 in the symmet-
ric array case. Hence, from Eq. (6.100), the spread of the Cartan matrix
satisfies the following inequality:
√
2κ1 ≤ spread (C) ≤ 2 (6.102)
or equivalently,
√
2κ1 ≤ ∆r̃1,N ≤ 2 (6.103)
Then from Eq. (6.97), it can be deduced that:
Proposition 6.3 If the length lm and first curvature κ1 of the manifold

of an N -sensor symmetric LA or ELA satisfy the following condition:
(N − 1) π
lm (6.104)
κ1
then, rank-(N −1) ambiguity, resulting from uniform partitions of the mani-
fold, is bound to exist. If instead,
√
lm < 2 (N − 1) π (6.105)
then, uniform rank-(N − 1) ambiguity cannot exist along the manifold.
It is however important to note that the condition set out in Eq. (6.105)
does not guarantee the absence of other types of rank-(N − 1) ambiguity,
for instance, those corresponding to non-uniform partitions of the array
manifold.
6.7 Appendices
6.7.1 Proof of Theorem 6.1

Since
s = [s1 , s2 , . . . , sc ]T (6.106)
is an ambiguous set the matrix A(s) = [a(s1 ), a(s2 ), . . . , a(sc )] ∈ CN ×c is

rank deficient. This means that any submatrix Ac (s) of A(s) which has
exactly c rows, is singular. That is
det(Ac (s)) = 0 (6.107)

Consider now the set
š = [s1 + ∆s, s2 + ∆s, . . . , sc + ∆s]T

= s + ∆s · 1c (6.108)
The manifold vector corresponding to the i-th element of š is
a(si + ∆s) = a(si ) a(∆s) (6.109)
Therefore, the matrix Ǎ with columns the manifold vectors corresponding

to the set š is
Ǎ = A(š) = A(s + ∆s1c ) = A(s) A(∆s1c ) (6.110)
The determinant of any submatrix Ǎc of Ǎ which has exactly c rows, is

det(Ǎc ) = det(Ac (s) Ac (∆s1c ))
 
 
= det(Ac (s)) · det diag( Ac (∆s) )

c×1 vector
=0 (6.111)
and is hence singular. This implies that Ǎ is rank deficient and therefore
the set š is also an ambiguous set. Furthermore the ranks of ambiguity of
s and of š are the same. This is because the submatrices of A(s) with less
than c rows are non-singular which implies that the submatrices of Ǎ with
less than c rows are also non-singular.
Chapter 7
More on Ambiguities:
Symmetrical Arrays
In contrast with the type of ambiguities identified in the previous chapter

which are based on partitioning of a manifold curve into equal segments, this
chapter is concerned with the partitioning of the manifold curve into unequal
segments from which a new type of ambiguity existing only in symmetrical1
array structures can be identified. That is, this type of ambiguity exists in
all symmetric linear and planar arrays as well as non-symmetric planar
arrays which may have at least one p-curve (p = ϕ, α, β) corresponding to
a symmetric ELA.
7.1 Symmetric Linear Arrays and det(AN (s))
Definition 7.1 Symmetric Linear Array: A linear array of N sensors

is said to be symmetric, if there is a real number b, such that the elements
of the vector )r r + b1N satisfy the following equation
rn ) = 0, ∀n odd
sum() (7.1)
Consider a set s of N arc lengths
T
s = [s1 , s2 , ..., sN ] (7.2)
on the manifold of a symmetric linear array of N sensors.

Let AN AN (s) ∈ CN ×N be the matrix with columns the manifold
vectors corresponding to the elements of s:
AN (s) = [a(s1 ), a(s2 ), . . . , a(sN )] (7.3)
1 Every sensor has a symmetric counterpart with respect to the the array centroid.
157
As we have seen in Chapter 2, at every point along the manifold curve of

a symmetric linear array, a set of N unity-norm coordinate vectors U(s) =
[u1 (s), . . . , uN (s)] ∈ CN ×N and (N − 1) curvatures κ1 , . . . , κN −1 can be
defined according to the first order matrix differential equation
U (s) = U(s)C (7.4)

whose solution is given by
F(s)

U(s) = U(0) expm(sC) (7.5)
where C ∈ RN ×N is the Cartan matrix, which is a real skew-symmetric

matrix of the curvatures. Using Eq. (2.63) in conjunction with Eq. (7.5),
the array manifold matrix AN can be rewritten as
AN (s) = [U(0)F(s1 )R, U(0)F(s2 )R, . . . , U(0)F(sN )R]

= U(0)[F(s1 )R, F(s2 )R, . . . , F(sN )R]
The matrix AN (s) is a square matrix and, therefore, its determinant can
be used to examine the linear dependence of its columns. An expression for
the determinant of AN (s) can be provided as follows:
det(AN (s)) = det(U(0)) · det([ F(s1 )R, F(s2 )R, . . . , F(sN )R]) (7.6)
Noting that for a symmetric linear array, the columns of U(s) are mutually
orthogonal, that is
UH U = IN and det(U(s)) = ±1 (7.7)
and therefore the determinant of AN becomes
det(AN (s)) = ±det[ F(s1 )R, F(s2 )R, . . . , F(sN )R] (7.8)
Because of the special structures of the radii vector R and of the Frame
matrix F, the det(A(s)) can be further written as
 
RT F(s1 )T
 RT F(s2 )T 
det(AN (s)) = ± det 



...
R F(sN )
T T
More on Ambiguities: Symmetrical Arrays 159
 
RT F(−s1 )
 RT F(−s ) 
 2 
= ±det  
 ... 
RT F(−sN )
 
RT F(s1 )
 RT F(s ) 
 2 
= ±det   (7.9)
 ... 
RT F(sN )
From Eq. (2.61) evaluated at s = 0 and for a symmetric linear array:
RT = aH (0)U(0) (7.10)
and therefore,
 
a(0)H U(0)F(s1 )
 a(0)H U(0)F(s ) 
 2 
det(AN (s)) = ±det 
 ..


 . 
a(0)H U(0)F(sN )
 
a(0)H U(s1 )
 a(0)H U(s ) 
 2 
= ±det 
 ..
 (using Eq. (7.5))

 . 
a(0) U(sN )
H
 
a(0)H u1 (s1 ) a(0)H u2 (s1 ) ... a(0)H uN (s1 )
 
 a(0)H u1 (s2 ) a(0)H u2 (s2 ) ... a(0)H uN (s2 ) 
= ±det 
 .. .. .. .. 

 . . . . 
H
a(0) u1 (sN ) a(0)H u2 (sN ) . . . a(0)H uN (sN )
(7.11)
7.2 Characteristic Points on the Array Manifold
An initial partition of a hyperhelical curve into N − 1 (or more) unequal

segments can be achieved using the concept of characteristic points, defined
as follows:
Definition 7.2 Characteristic Points: If A is the manifold of an array

of N sensors with locations r, a point s on A, with
0 ≤ s ≤ lm = 2πr (7.12)
is a characteristic point if and only if
Re{aH (0)u1 (s)} = 0 (7.13)
where u1 (s) denotes the tangent vector at the manifold point s, i.e.
u1 (s) = a (s).
For a general linear array the above definition indicates an important

property of characteristic points. The tangent at any characteristic point
is orthogonal (wide sense orthogonality since only the real part is zero) to
the manifold vector at s = 0. This orthogonality, in the case of a symmetric
linear, becomes “narrow” sense. Thus, if the manifold of a symmetric linear
array has N (or more) characteristic points sc1 , sc2 , . . . , scN , where sc1 = 0,
then this (denoted by scp ) set of characteristic points is ambiguous, as the
first column of Eq. (7.11) is the all zero vector, implying that
rank(AN (scp )) = N − 1 and det(AN (scp )) = 0 (7.14)
This is the starting point for the identification of a new class of ambiguous
generator set having a rank of ambiguity ρa equal to (N − 1).
The existence, however, of at least N characteristic points within the
manifold length lm is very important for the identification of ambiguous sit-
uations and this, in turn, relates to the number of windings. In particular we
have seen in Chapter 2 (Theorem 2.2, page 41) that the manifold of a linear
array of N sensor has a hyperhelical shape with the number of windings
(N = odd) or half-windings (N = even) nw given by the following equation
lm
nw = (7.15)
lw
where lw is the length of one winding or one half-winding and is given by
the (N − 1)th positive root of the following expression:
Tr(C expm(sC)) = 0 (7.16)
This expression and its derivatives with respect to s are the pivots for
identifying ambiguities which are only present in symmetric linear arrays
(or symmetric ELAs).
Equation (7.15) indicates that when nw > 1, then there is more than
one winding (or half-winding) contained in the manifold length of the curve.
It must be noted that uniform linear arrays with intersensor spacing equal
to one half-wavelength have lm = lw =⇒ nw = 1. A spiral of a spring could
be used to provide a visualization of the concept of a winding.
Corollary 7.1 A sufficient condition for the presence of nontrivial

ambiguities in the manifold A of a symmetric linear array (or ELA) is
lm > lw i.e. nw > 1
The above implies that if 0◦ ≤ p1 < pN ≤ 180◦ and
lw 2
nw = (7.17)
2πr | cos p1 − cos pN |
then there exist N − 2 bearings p2 , . . . , pN −1 in the region (p1 , pN ) such

that the manifold vectors corresponding to p1 , p2 , . . . , pN −1 are linearly
dependent.
Example 7.1 Consider a symmetric linear array of N = 4 sensors with

locations given by the following vector in half-wavelengths:
r = [−1.9, −0.6, 0.6, 1.9]T (7.18)
The length of the hyperhelical manifold curve of this array is lm = 2πr =

17.7048 and its Cartan matrix
 
0 −0.6462 0 0
0.6462 0 −0.1819 0 
C =
0
 (7.19)
0.1819 0 −0.2222
0 0 0.2222 0
Figure 7.1 is the plot of Eq. (7.16) with respect to the arc length. From
this plot it can be seen that the manifold curve consists of more than one
half-winding. Actually, the length of one half-winding is equal to lw =
14.0476, i.e. the arc length at the (N − 1)th positive characteristic point.
The number of half-windings of the hyperhelical curve is also calculated
using Eq. (7.15) and is equal to nw = 1.2603. The first four roots (arc
lengths) of Eq. (7.16) are
s = [0, 5.0780, 8.8664, 14.0476]T (7.20)

Fig. 7.1 Plot of Tr(C expm(sC)) with respect to s for the array of Eq. (7.18).
corresponding to the following set of azimuth angles (i.e. the directions of

arrival)
θ = [0◦ , 64.7614◦ , 90.0919◦ , 125.9361◦ ]T (7.21)
Note that the above set s is an ambiguous set. i.e.
rank(A(s)) = N − 1 and det(A(s)) = 0 (7.22)
Furthermore the set s has its first element equal to zero and also satisfies
all the other conditions of the definition of the ambiguous generator set
(see Definition 6.5, page 120). Hence s is also an ambiguous generator set
of parameters.
It is clear from the above discussion that if the array is symmetric, in

addition to the AGSs constructed from the Uniform Basic Sets of Eq. (6.23),
more AGSs can be found using a non-uniform partition of the array mani-
fold based on the roots of Eq. (7.16). However, the roots of this equation are
not the only ones producing AGSs. In fact we will see in the next section
that the derivatives of the above equation produce an infinite number of
AGSs and this complicates the representation and identification of this type
of ambiguity.
7.3 Array Symmetricity and Non-Uniform Partitions

of Hyperhelices
Before we proceed let us generalize the concept of characteristic points to

“characteristic points of order n” by rewriting the inner product aH (0)u1 (s)
as follows:
d
a(0)H u1 (s) = (a(0)H a(s))
ds
d
= (RT U(0)H U(0)F(s)R)
ds  
d  
= Tr U(0)H U(0)F(s)R RT 
ds
=IN
d
= Tr(F(s)R RT )
ds
d
= Tr(F(s))
ds
= Tr(F (s))
i.e.
a(0)H u1 (s) = Tr(C expm(sC)) (7.23)
where the expression F(s) = Cexpm(sC) has been used. Then by differen-
tiating Eq. (7.23) we get
(n−1)
a(0)H u1 (s) = Tr(Cn expm(sC)), n = 1, 2, . . . (7.24)
(n)
where u1 denotes the nth derivative of the tangent vector at the manifold
point s. This leads to the following definition:
Definition 7.3 Consider a symmetric linear array of N sensors having a

manifold A of length lm and a Cartan matrix C. A point s on A with
0 ≤ s ≤ lm = 2πr (7.25)
is said to be a characteristic point of order n if and only if
Tr(Cn expm(sC)) = 0 (7.26)

The characteristic points of order 1 (n = 1), i.e.
a(0)H u1 (s) = Tr(C expm(sC)) = 0 (7.27)
will simply be referred to as “characteristic points.”

Theorem 7.1 Non Uniform Partition Type Ambiguities: Con-
sider a symmetric linear array (or an ELA) of N sensors. If there are
c N characteristic points of order n on the array manifold A, forming
the vector
T
s(n)
cp = [s1 , s2 , . . . , sc ] (7.28)
(n)
that is, the elements of scp are the first n roots of the function
Tr(Cn expm(sC)) = 0 (7.29)

(n)
then scp , as well as any of its subsets with N elements, constitute a rank
ρa = (N − 1) ambiguous set of arc lengths.
Theorem 7.1 can be easily proven for n N, using the recursive equa-
tion Eq. (2.12) while for n > N , n ∈ Z + it can be proven using the
Cayley–Hamilton Theorem, which states that “every matrix satisfies its
own characteristic equation,” as applied to the Cartan matrix.
However not all the existing ambiguous generator sets are calculated by
using the set of roots, between zero and the manifold length. Roots that
are bigger than the manifold length should also be taken into account due
to the non uniform segmentation of the manifold. Furthermore the roots of
Eq. (7.29) are ambiguous for any real value of n greater than 0, as shown
by the following example
Example 7.2 For an array with sensor positions
r = [−2.1, −0.9, 0.9, 2.1]T (7.30)
the following table gives the AGSs produced from the Uniform Basic Sets
s1 s2 s3 s4 ρa
0 6.7672 13.5344 — 2
0 4.8337 9.6674 14.5011 3
0 4.8337 9.6674 19.3349 3
0 4.8337 14.5011 19.3349 3
0 9.6674 14.5011 19.3349 3
Furthermore the Cartan Matrix C of this array is

 
0 −0.6074 0 0
0.6074 0 −0.2056 0 
C =
0
 (7.31)
0.2056 0 −0.2981
0 0 0.2981 0
and since it is a symmetric array, the roots of Eq. (7.29), (for n = 1) i.e.
the set of characteristic points,
s(1)
cp = [0, 5.5147, 9.3228, 14.0360, 19.8000]
T
provide the four “non-uniform” AGS of rank ρa = 3 as shown below:
s1 s2 s3 s4 ρa
0 5.5147 9.3229 14.0360 3

0 5.5147 9.3229 19.8000 3
0 5.5147 14.0360 19.8000 3
0 9.3229 14.0360 19.8000 3
However for n = 1.9 the first four roots of Eq. (7.29) for s ∈ [0, lm ] forming
the vector
s(1.9)
cp = [2.9041, 7.6098, 12.0167, 17.1851]T (7.32)
is also an ambiguous set and of rank N − 1 = 3. Eq. (7.29) has been plotted
in Fig. 7.2 for n = 1.9.
From Example 7.2 it can be observed that the first root of Eq. (7.29) is
not equal to zero for all values of n. This means that although in the case
of n = 1 the AGSs can be simply identified as all the combinations of three,
(1)
or in general N − 1, non-zero elements of scp with a zero prepended, the
same cannot be done for n = 1.9. In order to find the AGSs in this case, the
first element of the ambiguous set must be subtracted from all its elements.
Note that the first root of Eq. (7.29) is equal to zero only if n = 1, 3, 5,
etc. This can be seen from Eq. (7.29) which for s = 0 becomes equal to
Tr(Cn ) that is zero only if n is odd, since in that case its structure is
identical with the structure of C (zero diagonal, more precisely only its
two secondary diagonals are non-zero). This result is confirmed by Fig. 7.3
Fig. 7.2 Plot of Tr(Cn expm(sC)) for n = 1.9.
Fig. 7.3 Plot of smallest root of Tr(Cn expm(sC)) versus n for the array in Example 7.2.
where the first root of Tr(Cn expm(sC)) is plotted for n = 0.1 to 10 in steps
of 0.1 for the Cartan matrix of the array in Example 7.2.
From the previous discussion it is clear that an infinite number of AGSs
exist in symmetric linear arrays because n takes infinite values. However
for large values of n there exists a relationship between Theorems 6.2 and
7.1. This is stated in the following lemma:
Lemma 7.1 For n → ∞ (large), the roots of Tr(Cn expm(sC)) = 0,
n > 0 are a shifted version of the set of arc lengths created by partitioning
the manifold in segments equal to lm /la , where la = ∆r1,N is the aperture
of the array.
Example 7.3 Consider the array
r = [2.6870, 1.2728, 0, −1.2728, −2.6870]T (7.33)
Using the aperture intersensor spacing ∆r1,5 = 5.3740 the UBS s∆r1,5 can
be found as
s∆r1,5 = [0, 4.9162, 9.8322, 14.7483, 19.6645, 24.5807]T (7.34)

(n)
The set of characteristic points scp of order n > 6, given by using Eq. (7.29)
of Theorem 7.1, has approximately the same elements as s∆r1,5 . For instance
for n = 13
s(13)
cp = [0, 4.9162, 9.8322, 14.7483, 19.6645, 24.5807]
T
(7.35)
indicating that the two sets are almost identical, i.e.
cp s∆r1,5
s(13) (7.36)
7.4 Ambiguities of Rank-(N − 1) and Array Pattern
In the previous section we have seen that the function Tr(Cexpm(sC)) is

central for the identification and generation of non-uniform rank-(N − 1)
ambiguous sets of arc lengths in symmetric linear arrays. This can be also
rewritten as:
d
Tr(C expm(sC)) ≡ a(0)H u1 (s) = a(0)H a(s) (7.37)
ds
H
where the term a (0) a (s) is purely imaginary and represents the array
pattern of an array steered towards endfire, as a function of the arc length s.
It is immediately apparent that the roots of Tr(C expm(sC)) = 0

are equal to the arc lengths corresponding to the stationary points of
the gain pattern steered towards endfire and this leads to the following
proposition.
Proposition 7.1 If the array pattern of an N -sensor symmetric linear

array, parametrized in terms of arc length, steered towards endfire exhibits
more than N stationary points (usually lobes), then the corresponding direc-
tions are rank-(N − 1) ambiguous.
The following example illustrates the main characteristics of the above

proposition:
Example 7.4 Consider a symmetric linear array with sensor locations in

half-wavelengths given by
r = [−1.45, −0.8, 0.8, 1.45]T (7.38)
and with manifold length lm = 14.7153. The function Tr(C expm(sC)) is

plotted in Fig. 7.4 for s ∈ [0, lm ], which is equivalent to the azimuth space
θ ∈ [0◦ , 180◦ ). As can be seen, the equation Tr(C expm(sC)) = 0 has 4 roots
within the manifold length given by
s(1)
cp = [0, 5.9142, 10.5499, 14.2941] (7.39)
which constitutes a rank-3 AGS. The corresponding set of rank-3 ambiguous

set of directions is given by
◦ ◦ ◦ ◦
θ(1)
cp = [0 , 78.69 , 115.71 , 160.52 ] (7.40)
Furthermore, as can be seen from Fig. 7.5, the gain pattern of the array
steered towards endfire exhibits stationary points at
s = [0, 5.9142, 10.5499, 14.2941] (7.41)
when parametrized in terms of the arc length, and at
θ = [0◦ , 78.69◦ , 115.71◦ , 160.52◦ ] (7.42)
when parametrized with respect to the azimuth. It is easy to verify that

s and θ are equivalent. It might however be argued that a source close
to endfire is unlikely; but it must be recalled that any displacement of
the ambiguous set of arc lengths by an arbitrary amount will result in
another ambiguous set, provided that the manifold length is not exceeded.
Fig. 7.4 The function Tr(Cexpm(sC)) and the array pattern, plotted as a function of
the arc length, for a symmetric array with locations r = [−1.45, −0.8, 0.8, 1.45 ]T . Note
that the arc length space s ∈ [0, lm = 14.7153) is equivalent to the azimuth space
θ ∈ [0◦ , 180◦ ), while the black dots constitute a rank-3 AGS of arc lengths.
For instance a displacement of the set s by ∆s = 0.3, corresponding to a

new set of directions given by
θ = [16.42◦ , 81.06◦ , 118.34◦ , 169.59◦ ] (7.43)
is also ambiguous.
Fig. 7.5 The results of Fig. 7.4 expressed as a function of azimuth.
7.5 Planar Arrays and ‘Non-Uniform’ Ambiguities
The technique presented in the previous section to identify ambiguities

arising from non-uniform partitions of the symmetric linear array manifold
can be readily applied to any hyperhelical curve (e.g. a composite φ-curve)
of the planar array manifold if the corresponding equivalent linear array

(ELA) is symmetric.
Note that if the planar array itself is symmetric, then all ELAs of the
family of p-curves are symmetric too and this is formally stated as follows:
Corollary 7.2 All equivalent linear arrays r(Ψ0 ),∀Ψ0 (see Eq. 6.4) asso-
ciated with the families of p-curves (p = ϕ, α, β) of a symmetric planar
array are symmetric and therefore can suffer from both uniform and non-
uniform types of ambiguity.
Note that ambiguities resulting from non-uniform partitions of the array

manifold are present only in symmetric linear arrays and hence this class of
ambiguities is applicable only when an ELA is symmetric. However, the pre-
requisite symmetry of the ELA does not impose any special restriction on
the planar array configuration, and an arbitrary planar array may exhibit
symmetric ELAs along certain Ψo angles. However, if the planar array is
symmetric about both the x- and y-axes or any two orthogonal directions,
then the ELA along any Ψo angle is symmetric. Hence symmetric planar
arrays, like symmetric linear arrays, are more prone to ambiguities. If the
planar array is asymmetric, then it is more likely to suffer from ambigu-
ities along Ψo angles where the ELA is symmetric than those where it
is not.
Consider the issue of collocation of projected sensors in a symmetric
ELA. For instance an N -sensor ELA consisting of an underlying symmetric
(N −
)-sensor linear array and
sensors collocated with other sensors. If
the underlying (N −
)-sensor array suffers from rank-(N −
− 1) ambiguity
resulting from non-uniform partitions of the array manifold, then it is easy
to see that the N -sensor ELA will also suffer from the same rank-(N −
−1)
ambiguity, with the same ambiguous directions. In fact, the N -sensor ELA
will suffer from any ambiguity inherent to the underlying (N −
)-sensor
linear array. For example, consider a planar array whose sensor locations
in half-wavelengths are given by
 T
−0.9548 −0.6623 0.1077 0.5174 0.6168
r = −1.2453 1.0057 −1.5300 −0.9669 1.3359
0 0 0 0 0
The ELA along θ◦ = 75◦ is given by
r(θ◦ ) = [−1.45, −1.45, −0.8, 0.8, 1.45]T (7.44)

Fig. 7.6 Two-dimensional MUSIC in an ambiguous signal environment.
with two collocated sensors at −1.45. Discarding one of the collocated sen-
sors, the resulting array is found to be symmetric and given by
r = [−1.45, −0.8, 0.8, 1.45]T (7.45)

Note that this is precisely the array examined in Example 7.4 and hence
the following set of directions will constitute an ambiguous set of directions
for the corresponding planar array:
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
cp = {(75 , 16.42 ), (75 , 81.06 ), (255 , 61.66 ), (255 , 10.41 )}
(θ, φ)(1)
Figure 7.6 illustrates the contour plot of the MUSIC spectrum generated
using 100 data snapshots collected from the planar array, operating in the
presence of three equipower uncorrelated sources with SNR = 10 dB and
impinging from directions (75◦ , 16◦ ), (75◦ , 81◦ ) and (255◦ , 62◦ ) respectively.
A spurious peak around the (255◦ , 10◦ ) direction is also observed.
7.6 Conclusions
Although all “uniform” ambiguities can be found using the Uniform Basic
Sets of arc lengths of Eq. (6.23), there is no single coherent framework
at this moment to identify and estimate all “non-uniform” ambiguities.

However, this chapter provides a starting point by presenting a class of rank-
(N −1) non-uniform ambiguities. Since the range of n is infinite it can hence
be concluded that an infinite number of AGSs exist in symmetric linear
and symmetric planar arrays. This is also true for non-symmetric planar
arrays if at least one of its ELAs is “symmetric.” Note that each AGS
represents an infinite number of ambiguous sets — rotated in arc length
(see Theorem 6.1). Having an infinite rather than a finite number of AGSs
is currently a fundamental limitation of the theory for handling ambiguities
in symmetric arrays. Thus further research effort should be devoted to the
generalization of the concept of AGSs in order to reduce their number
to finite values. Furthermore the exploitation of any inherent relationship
between the various AGSs, constructed from the characteristic points of
different order n, may play a crucial role in dictating the identification,
classification and estimation of this huge number of ambiguities inherent in
symmetric array structures.
Chapter 8
Array Bounds
It was explained in Chapter 2 that, according to standard differential geom-

etry, a d-dimensional curve may be fully described via d − 1 curvatures.
However, it is intuitively apparent that a sufficiently small segment of a
multi-dimensional curve A may in fact be accurately approximated by a
circular arc. Despite the fact that visualization of a curve of dimensionality
greater than three is impossible, circular approximation provides an
informative notion of the curve’s shape in the local neighborhood
of an arbitrary point s̆ and is defined in terms of the curve’s princi-
pal curvature only, thereby simplifying the analysis considerably. In this
chapter the circular approximation of a curve A is used to determine the
array’s ultimate accuracy, detection and resolution capabilities. Note that
the curve A could be the manifold of a linear array or a p-curve (Ap|qo with
p = θ, φ, α or β), lying on the manifold surface M of a planar array and
associated with an ELA.
8.1 Circular Approximation of an Array Manifold
A sufficiently small neighborhood of a point s̆ on a curve in real N -

dimensional space (where the coordinate vectors are strictly orthogonal)
can be approximated by a circular arc. This circular arc lies on
H12 L([u1 (s̆), u2 (s̆)]) (i.e. a plane) (8.1)
174
Array Bounds 175
with the vector u1 (s̆) as its tangent, and has radius R which is equal to the
inverse of the curvature of the curve at s̆. This implies that

∆si ∆si
u1 (s̆ ± ∆si ) = cos u1 (s̆) ± sin u2 (s̆) (8.2)
R R

∆s2i ∆si
1−
2 R
2R
Here the above concept is applied in order to derive a “circular approx-

imation” to the manifold curve A which is embedded in an N -dimensional
complex space and is defined as follows:
A = {a(p) ∈ CN , ∀p : p ∈ Ωp } (8.3)
Let us consider an arbitrary point s̆ (arc length) on the manifold and

form the Frame matrix F at s̆−∆s1 and s̆+∆s2 corresponding to parameter
values p̆ − ∆p1 and p̆ + ∆p2 . Since F(s̆) = expm(s̆C), by using Taylor series
expansion and retaining terms up to second order:

1
F(s̆ ± ∆si ) F(s̆) Id ± (∆si )C + (∆si )2 C2 for i = 1, 2 (8.4)
2
However the matrix of the manifold coordinate vectors U at s̆ ± ∆si can
be expressed (see Eq. (2.10), page 28) as
U(s̆ ± ∆si ) = U(0)F(s̆ ± ∆si ) (8.5)
Combining Eqs. (8.4) and (8.5), and using the special skew symmetric form
of the Cartan matrix (Eq. (2.14)), the following approximation to the mani-
fold tangent vector at s̆ ± ∆si (i.e. the first column of U(s̆ ± ∆si )) may be
obtained:

(∆si )2 2 (∆si )2
u1 (s̆ ± ∆si ) 1 − κ1 u1 (s̆) ± ∆si κ1 u2 (s̆) + κ1 κ2 u3 (s̆)
2 2
(8.6)
However, the 3rd term of Eq. (8.6) can be ignored so long as a tolerance
factor T OL 1 is defined such that:
@ @
@ 3rd term @ 1
@ @
@ 2nd term @ = 2 ∆si κ2 ≤ T OL (8.7)
where κ2 can be found using Eq. (2.37), i.e.

1 3
κ2 = )
r − κ21 r) (8.8)
κ1
Thus, Eq. (8.7) provides the following condition
2κ1
∆si ≤ 3 T OL (8.9)
r − κ21 r

under which Eq. (8.6) is simplified to

(∆si )2 2
u1 (s̆ ± ∆si ) 1− κ1 u1 (s̆) ± ∆si κ1 u2 (s̆) (8.10)
2
matching Eq. (8.2) for R = κ−1 1 .

The above indicates that a circular arc can be adopted as an excellent
approximation to the array manifold in the region of interest neighboring
s̆. This approximation also provides a 2D representation where, at a local
level, a curve has its main components along the first two coordinate vectors
at point s̆, as illustrated in Fig. 8.1, and has not yet had the opportunity
to “move into” higher dimensions (subspace L{[u3 , u4 , . . . , ud ]}).
Note that by writing the incremental arc length as ∆si ≈ ṡ(p̆)∆pi and
recalling that ṡ(p) = πr sin p, the condition of Eq. (8.9) provides an upper
limit to the angular separation ∆pi for which the above approximation is
Fig. 8.1 H12 subspace used to approximate the array manifold at a local level.
Array Bounds 177
valid. That is, for i = 1, 2 we have

2κ1
∆pi ≤ T OL (8.11)
r3 − κ21 r)
π r sin p̆)

=ṡ(p̆)
It should be emphasized that √ although the array manifold is a vector con-

tinuum with a constant
√ norm N , the radius of the circular approximation
is different from N .
Consider the Fig. 8.2 which illustrates a manifold vector a(s̆ + ∆si ) in
the local neighborhood of s̆ with ∆si equal to −∆s1 or +∆s2 corresponding
to two signals arriving from bearings p1 and p2 . If u1 (s̆) ⊥ u2 (s̆) then the
radius of this circular arc is equal to κ−11 (s̆). However, u1 (s̆) and u2 (s̆) are
not, in general, strictly orthogonal. Therefore, the radius of the correspond-
ing circular arc at a point s̆ is equal to κ̂−1
1 (s̆) which takes into account the
“orientation” of the array manifold curve. The parameter κ̂1 (s̆) is the cur-
vature (inverse of radius) of the circular approximation of the manifold A
and in Appendix 8.9.1 is proved to be
κ̂1 = κ1 sin(ζ) (8.12)
where ζ is the angle between u1 (s̆) and u2 (s̆). Indeed

sin(ζ) = 1 − |uH 1 (s)u2 (s)|
2 (8.13)
9 @ @
@ j T 3 @2
@
= 1 − @ 1N () r )@@ (using Eq. (2.44)) (8.14)
κ1
Thus Eq. (8.12) becomes
/ @ @2
@ @
κ̂1 = κ21 − @1TN r)3 @ (8.15)
where r) = r/ r
Since the manifold can be approximated by a circular arc, it is clear that
∆si = κ̂−1
1 ∆ψi (8.16)
where ∆ψi (with reference to Fig. 8.2) is estimated as follows:

κ̂−1
1 − µxi
cos(∆ψi ) = (8.17)
−1
κ 1
⇒ ∆ψi = arccos(1 − κ̂1 µxi ), for i = 1, 2 (8.18)

Fig. 8.2 Two alternative ways of visualizing the ‘circular approximation’ of array man-
ifold curve at an arbitrary point s̆: (a) view showing H12 plane, (b) view orthogonal to
L ([H12 , a(s̆)]).
or equivalently, as
µyi
sin(∆ψi ) = (8.19)
κ̂−1
1
⇒ ∆ψi = arcsin(κ̂1 µyi ), for i = 1, 2 (8.20)

Array Bounds 179
Hence, substituting ∆ψi given by Eq. (8.19) back into Eq. (8.16) we have
∆si = κ̂−1
1 ∆ψi
arccos(1 − κ̂1 µxi )
= (8.21)
κ̂1
arcsin(κ̂1 µyi )
(or, equivalently) =
κ̂1
It must be noted that if the inclination angle ζinc (see Eq. (2.55)) of the
manifold is 0◦ then ζ = 90◦ and κ̂1 = κ1 , which is an important feature of
symmetric linear arrays.
8.2 Accuracy and the Cramer Rao Lower Bound
The most popular bound in array processing is a well-known statistical

result called the Cramer Rao lower bound (CRB) [23]. The CRB sets a
lower limit on the error covariance matrix of any unbiased estimate, p̂, of
the true parameter vector p ∈ RM in the array signal model of Eq. (1.29).
In the case of an array of N sensors receiving M narrowband signals with
additive sensor noise of power σ 2 , and for a sufficiently large number of
snapshots (L 1), the expression for the deterministic CRB has been
shown to be as follows [24]:
σ2 & '−1
CRB[p] = Re(H RTm ) ∈ RM ×M (8.22)
2L
where
H
H = Ȧ P⊥
A Ȧ ∈ C
M ×M
(8.23)
RTm = E{m(t)m(t)H } ∈ CM ×M = source covariance matrix
with

A = [a1 , a2 , . . . , aM ]

Ȧ = [ȧ1 , ȧ2 , . . . , ȧM ] (8.24)

 ⊥ H −1
PA = IN − A(A A) AH
based on the following assumptions:
• N > M and the manifold vectors are independent,

• Noise is a zero-mean, temporally white Gaussian process,
• Noise is spatially white from sensor to sensor; i.e.
E{n(t) nH (t)} = σ 2 IN ,
• Parameters other than p are known a priori.
In this section the nature of the CRB and its relationship with the
differential geometry of the array manifold (and hence the array geometry)
is clarified by focusing on the special cases of one emitter as well as two
emitters located close together.
8.2.1 Single Emitter CRB in Terms of

Manifold’s Differential Geometry
Assume that the array receives a single signal m(t) of power
P = E{m(t)m∗ (t)}
from bearing p. Then, since Rm = P and A = a(p), Eq. (8.22) implies that
the CRB may be expressed as
σ2 1
CRB[p] = (8.25)
2LP ȧH (p)P⊥a ȧ(p)
Recalling that ȧ(p) = u1 (s) ṡ(p) and u1 (s) = 1 the above equation
becomes
σ2 1
CRB[p] = (8.26)
2LP ṡ(p) u1 (s)P⊥
2 H
a u1 (s)
However, u1 (s) is orthogonal to a(p). Hence P⊥

a u1 (s) = u1 (s) and Eq. (8.26)
may be further simplified as follows:
σ2 1
CRB[p] = =
2LP ṡ(p) 2 2(SNR × L)ṡ(p)2 (8.27)
2
where SNR = P/σ
Note that the above expressions are valid for any manifold curve, not
necessarily of hyperhelical shape. For instance Eq. (8.27) can be written so
as to provide information about the accuracy of p estimates in terms of the
Array Bounds 181
Table 8.1 Rate-of-change of arc length for linear and planar arrays. The
planar arrays are parametrized in terms of (θ, φ) and (α, β).
Array Curve A Rate of change of arc length s Parameter p
Linear A ṡ(θ) = πr x sin θ p=θ
Planar (θ, φ) Aφ|θo ṡ(φ) = πr(θo ) sin φ p=φ
Aθ|φo ṡ(θ) = πṙ(θ) cos φo p=θ
Planar (α, β) Aα|βo ṡ(α) = πr(Θo ) sin α p=α
Aβ|αo ṡ(β) = πr(Θo + 90◦ ) sin β p=β
sensor locations of a linear array for p = θ, or of a planar array for p = θ,

φ, α or β by replacing its corresponding ṡ(p) as shown in Table 8.1.
For an emitter at a bearing (p, q) the p-estimates correspond to points
on the curve Ap|q lying on the manifold surface M as the parameter q is
assumed constant and known. Thus the notation
CRB[p|Ap|qo ] (8.28)
is adopted to represent the CRB for the p-estimates and show its depen-
dence on the properties of the Ap|qo curve, with the subscript ‘o’ indicating
that q is assumed constant and known. Similarly, the q-estimates are asso-
ciated with the curve Aq|p on M as the parameter p is assumed constant
and known, with the CRB of q-estimates represented by
CRB[q|Aq|po ] (8.29)
Using this notation in conjunction with Eq. (8.27) and Table 8.1 for an
emitter at a direction (θ, φ), the CRB for the φ-estimates and θ-estimates
are expressed as follows

 1

 CRB[φ|Aφ|θo ] =
 2 (SNR × L) (π r(θo ) sin φ)
2
(8.30)

 1

 CRB[θ|A θ|φo ] = 2
2 (SNR × L) (π ṙ(θ) cos φo )
dr(θ)
where as usual r(θ) = rx cos θ + ry sin θ and ṙ(θ) = dθ .
Based on the above discussion a number of observations can be made

as follows:
(1) For all planar array geometries:

 1

CRB[φ|Aφ|θo ] ∝

sin2 φ
(8.31)

 1
 CRB[θ|Aθ|φo ] ∝
cos2 φo
The above expressions imply that the θ estimates are more accurate
for φo → 0◦ while the φ estimates are more accurate for φ → 90◦ .
Furthermore, the variations of CRB[φ|Aφ|θo ] and CRB[θ|Aθ|φo ] with
respect to elevation φ are both independent of the array geometry,
monotonic and 90◦ out of phase.
(2) Since ṙ(θ) = r(θ + 90◦ ), consequently for all planar array geometries
CRB[θ|Aθ|φo ] = CRB[(90◦ − φo ) |Aφ|(θo +90◦ ) ] (8.32)
In other words, the CRB for the θ-estimates of an emitter at bearings

(θ, φ) equals the CRB for the φ-estimates of a similar emitter at bearing
(θ + 90◦ , 90◦ − φ). Furthermore, the variations of CRB[θ|Aθ|φo ] and
CRB[φ|Aφ|θo ] with respect to azimuth θ are both functions of the array
geometry, and are 90◦ out of phase.
(3) For balanced-symmetric arrays
ṙ(θ)2 = r(θ)2 = rx 2 (independent of θ) (8.33)
therefore:
 1

 CRB[φ|Aφ|θo ] =

 2

 2L × SNR (π rx sin(φ))

 ⇒ φ accuracy is independent of θ
(8.34)

 1

 CRB[θ|Aθ|φo ] =

 2L × SNR (π rx cos (φo ))
2


⇒ θ accuracy is independent of θ
Example 8.1 Figure 8.3 shows values of the CRB[θ|Aθ|φo ] and

CRB[φ|Aφ|θo ] for a single emitter at bearing (θ, φ), as functions of θ and
φ. The two equation of Eq. (8.30) are evaluated for the 24-element Y-
shaped and circular arrays given in Appendix 8.9.2. The figures are based
on the assumption of L = 100 snapshots and a Signal-to-Noise ratio of
Array Bounds 183
Fig. 8.3 CRB for a single emitter with SNR×L = 1000.
10 dB. Both arrays are balanced-symmetric, and so the accuracy of the

parameter estimates will be independent of azimuth. On the other hand,
while the accuracy of the parameter estimates varies dramatically with ele-
vation, the performances of the two arrays remain unchanged in relative
terms.
At this stage it is also interesting to consider the CRB corresponding

to the estimates of the manifold arc length parameter s. Replacement of
the bearing parameter p with the arc length s requires that the manifold
derivative ȧ = da/dp in Eq. (8.25) be replaced by a = da/ds = u1 (s).
Again since P⊥a u1 (s) = u1 (s), it follows that
σ2 1
CRB[s] = = (8.35)
2LP 2(SNR × L)
In other words,
the lower bound on the variance of the estimates

of arc length parameter (s) is independent of the
array geometry and depends only on the (SNR × L)
product.
8.2.2 Two Emitter CRB in Terms of Principal Curvature

Expressions for the CRB (on the variance of unbiased parameter estimates)
become progressively more complicated with increasing numbers of emit-
ters, M , since the accuracy of the bearing estimates is not only a function
of the additive sensor noise but also depend on the interactions between
the various emitters.
Consider a multiple-emitter scenario involving two correlated emitters
and M − 2 uncorrelated emitters. It is easy to show that if the two cor-
related signals arrive from bearings p1 and p2 (corresponding to manifold
vectors a1 a(p1 ) and a2 a(p2 )), have powers P1 and P2 , and a correla-
tion coefficient ρ, then Eq. (8.22) for the CRB corresponding to the signal
arriving from p1 may be written as
1 1 1
CRB[p1 ] = (8.36)
2(SNR1 × L) ȧ1 P⊥
H
A ȧ1 1 −
Re2 [ρȧH ⊥
2 PA ȧ1 ]
⊥ ȧ )(ȧH P⊥ ȧ )
P1 P2 (ȧH P
1 A 1 2 A 2
where SNR1 = P1 /σ 2 and P⊥ A = IN −A(A A)A with A being an (N ×M )

H H
matrix with columns the M manifold vectors. To make Eq. (8.36) more
tractable, consider the scenario where all the M emitters are uncorrelated
Array Bounds 185
(emissions from independent sources). Setting ρ = 0 in Eq. (8.36) we have

1 1
CRB[p1 ] = (8.37)
2(SNR1 × L) ȧH P ⊥
1 A ȧ1
The dependence of the CRB on the rate of change of manifold arc length,
ṡ(p), become apparent by considering a1 and a2 as two points on the
same manifold curve A. Then using the expression ȧ(p1 ) = u1 (s1 ) ṡ(p1 ),
Eq. (8.37) can be written as follows:
1 1
CRB[p1 |A] = (8.38)
2(SNR1 × L) ṡ(p1 ) u1 (s1 )P⊥
2 H
A u1 (s1 )
Further interpretation of Eq. (8.38) in terms of the principal curvature of A

is possible when M = 2 (i.e. two-emitter scenario) and the two emitters are
closely spaced at bearings p1 and p2 = p1 + ∆p , corresponding to manifold
arc lengths s1 = s̆ − ∆s/2 and s2 = s̆ + ∆s/2 respectively. Under such cir-
cumstances, circular approximation can be applied to a local neighborhood
⊥
of a(s̆) in order to evaluate the term uH 1 (s1 ) PA u1 (s1 ). As shown in the
Appendix 8.9.3, this hence leads to the following expression:
1 2
CRB[p1 |A] = (8.39)
(SNR1 × L) ṡ(p1 ) (∆s) κ̂21 (p̆) −
2 2 1
N

∆s = ṡ(p̆) ∆p and
where (8.40)
κ̂1 (p) = κ1 (p) sin(ζ(p))
Note that the bearing p̆ corresponds to the point with arc length s̆, and
to a first-order approximation, equals (p1 + p2 )/2 or equivalently p̆ =
p1 + ∆p2 .
It is important to point out that, since ∆s = ṡ(p̆)∆p, Eq. (8.39) indicates
that the CRB is inversely proportional to the square of the emitters’ angular
separation,
1
CRB[p1 |A] ∝ 2 (8.41)
(∆p)
i.e. the accuracy of the system deteriorates rapidly as the emitters approach
one another.
The technique deployed for the derivation of Eq. (8.39) can also be
directly applied to scenarios involving more than two closely spaced emit-
ters. However, due to the proliferating number of manifold vector inner
Table 8.2 Principal curvature for linear and planar arrays parametrized —
planar arrays are in terms of (θ, φ) and (α, β).
Array Curve A Principal curvature κ1 Hyperhelix
Linear A κ1 = r̃2x , ∀s Yes
Planar (θ, φ) Aφ|θo κ1 = r̃2 (θo ), ∀s Yes

1 2 j
Aθ|φo κ1 (θ) = ṙ (θ) + P ⊥
r(θ) No
ṙ(θ)2 π cos φo ṙ(θ)

ṙ(θ)2
Planar (α, β) Aα|βo κ1 = r̃2 (Θo ), ∀s Yes
Aβ|αo κ1 = r2 (Θo + 90◦ ) Yes
= R and Ṙ = dR
where R
R dp
products, the expression for the CRB can become very cumbersome.
Nevertheless, it can be shown that so long as the additional manifold points
s3 , s4 , . . . , sM are not in the neighborhood of s̆ (i.e. in the neighborhood of
s1 and s2 ), the value of CRB[p1 |A] is primarily dominated by the presence
of the source at bearing p2 .
Example 8.2 In Fig. 8.4 the value of CRBθ [θ1 ], as defined in [24], is
compared to that given by Eq. (8.39) for two unit-power emitters located
at bearings θ1 = 20◦ and θ2 = θ1 + ∆θ, and both at a common elevation of
φ = 20◦ . The expressions are evaluated for the 24-element Y-shaped array,
assuming the availability of L = 100 snapshots and a signal-to-noise ratio
of 10 dB (for each emitter).
As can be seen, for small emitter separations, Eq. (8.39) provides an
excellent estimate of the CRB. However for increasing separations the
circular approximation breaks down and Eq. (8.39) is no longer valid. Note
that for large emitter separations, the exact CRB settles down to the value
corresponding to a single emitter.
8.2.2.1 Elevation Dependence of Two Emitters’ CRB

For a planar array with two emitters on the same φ-curve Aφ|θo , the
dependence of the CRB on elevation φ may be deduced by recalling from
Table 8.1 (page 181) that
ṡθ (θ) ∝ cos φ and ṡφ (φ) ∝ sin φ (8.42)

Array Bounds 187
Fig. 8.4 Plots of the CRB Eqs. (8.37) (exact) and (8.39) (appr.) for two emitters with
angular separation ∆θ (Y-shaped array of 24 sensors). The CRB for a single source is
also shown for comparison.
where the subscripts θ and φ have been added for reference to θ-curves and
φ-curves respectively.
Also since the φ-curves are hyperhelical, κ̂1,φ (φ) is independent of φ.
Furthermore, for large arrays, κ̂1,θ (θ) is independent of φ so long as elevation
is not close to 90◦ . The combination of these results, along with Eq. (8.39)
reveal the following dependencies on elevation φ:
• For emitters corresponding to two points on the same θ-curve Aθ|φo , that
is having the same elevation angle, i.e. (θ1 ,φo ) and (θ2 , φo ),
1
CRB[θ1 |Aθ|φo ] ∝ (8.43)
ṡ2θ (θ1 )ṡ2θ (θ̆)
1
∝ (8.44)
cos4 φo
θ1 + θ 2
where θ̆ = (8.45)
2
• For emitters with the same azimuth angle, i.e. (θo ,φ1 ) and (θo , φ2 ), hence
on the same φ-curve Aφ|θo
1
CRB[φ1 |Aφ|θo ] ∝ (8.46)
ṡ2φ (φ1 )ṡ2φ (φ̆)
1
∝ (8.47)
sin φ1 sin2 φ̆
2
φ1 + φ2
where φ̆ = (8.48)
2
Consequently, the variations of CRB[θ1 |Aθ|φo ] and CRB[φ1 |Aφ|θo ] with
respect to φ are independent of the array geometry and 90◦ out of phase.
8.2.2.2 Azimuth Dependence of Two Emitters’ CRB

The dependence of the CRB on azimuth θ is a rather complex function of
the array geometry. However, it can be proven that
ṡθ (θ, φ) = ṡφ (θ + 90◦ , 90◦ − φ) (8.49)

and
κ̂1θ (θ, φ) κ̂1φ (θ + 90◦ , φ) = (independent of φ) (8.50)
Incorporating these results into Eq. (8.39), and on the grounds that for
closely spaced emitters
ṡθ (θ1 ) ṡθ (θ̆) and ṡφ (φ1 ) ṡφ (φ̆), (8.51)
one may deduce that:
• if ∆θ = ∆φ, then the CRB on the azimuth estimates of two emitters

equally distributed about azimuth θ̆ and with common elevations φ̆, is
equal to the CRB on the elevation estimates of two similar emitters
equally distributed about elevation (90◦ − φ̆) and with common azimuth
θ̆ + 90◦ . This is somewhat similar to the result derived for the single-
emitter case (see Eq. (8.32)) in that the variations of CRB[θ1 |Aθ|φo ] and
CRB[φ1 |Aφ|θo ] with respect to θ are 90◦ out of phase.
Example 8.3 Figures 8.5(a) and (b) show values of CRB[θ1 |Aθ|φ ]
(Eq. (8.39)) as a function of θ and φ, in the case of two unit-power
emitters of common elevation φ but separated by ∆θ = 1◦ about azimuth θ.
The bound is evaluated for the 24-element Y-shaped and circular arrays
assuming the availability of L = 100 snapshots and signal-to-noise ratio
Array Bounds 189
Fig. 8.5 CRL for two emitters with SNR×L = 1000.
of 10 dB (for each emitter). In Fig. 8.5(c) and (d) the exercise is repeated
for CRB[φ1 |Aφ|θ ] with emitters of common azimuth θ but separated by
∆φ = 1◦ about elevation φ. While the circular array again exhibits uni-
form accuracy for all azimuths, the Y-shaped array shows fluctuations with
a period of 60◦ due to its special shape. Again, the performances of the
two arrays remain unchanged in relative terms as a function of elevation

(true for CRB[θ1 |Aθ|φ ] for elevations not too close to 90◦ ) as can be readily
confirmed from the graphs.
8.3 “Detection” and “Resolution” Thresholds
In practice, the availability of only a limited number of snapshots, L, pre-

vents the full elimination of the noise and can result in a poor estimation
of array manifold vectors.
For a source corresponding to point si on the array manifold curve A, the
RMS value σei of the uncertainty due to noise which remains in the system
after L snapshots can be represented as an N -dimensional hypersphere of
radius σei centered at the manifold vector a(si ) (see Fig. 8.6). This sphere
will be known as the “uncertainty sphere” and helps to examine the effects
of noise and finite snapshots on the ability of a system to detect and resolve
the bearings of two closely spaced sources. In particular in this section
concepts from the differential geometry of array manifold curves will be
used to develop global measures of detection and resolution capabilities.
The detection capability of an array system is determined by its ability
to estimate the number of sources, M , contributing to the signal at the
Fig. 8.6 View of the H12 subspace (plane): An illustrative representation of the geom-
etry at ‘resolution threshold’ (circular approximation)
Array Bounds 191
array output. Once this information is available, the parameters of interest

can be estimated, provided that all the M sources present are resolved.
The array resolution is the ability to obtain distinct, albeit inaccurate,
parameter estimates for each individual source present. In other words, the
sources present are resolved when the corresponding points on the array
manifold can be estimated (this, for instance, may correspond to distinct
peaks- or nulls-observed in the spectrum of the DF algorithm employed).
However both problems become very difficult if there are sources located
“close together” and the determination of “how close” two sources can be,
provides the ultimate detection and resolution capabilities of the array.
In order to analytically determine these ultimate array capabilities it
is first necessary to define the detection and resolution “threshold sub-
spaces” (hyperplanes) and then to provide the conditions under which
the “thresholds” of detection and resolution occur. Thus consider two
closely located sources of powers P1 and P2 arriving from bearings p1 and
p2 = p1 + ∆p corresponding to the manifold points s1 and s2 = s1 + ∆s
respectively. The detection subspace Hdet and resolution subspaces Hres are
defined for a point s̆ between s1 and s2 as follows:

detection: Hdet L [a(s̆), PA a(s̆)] (8.52)

resolution: Hres L [a(s̆), u1 (s̆)] (8.53)
where
H −1
PA = A(A A) AH with A = [a(s1 ), a(s2 )] (8.54)
These two subspace are illustrated in Figs. 8.7 and 8.8 together with
the two manifold vectors a(s1 ) and a(s2 ), and their associated uncertainty
spheres. Note that in these figures the manifold curve may be the manifold
of a linear array or a p-curve lying on the manifold surface of a planar
array.
8.3.1 Estimating the Detection Threshold

Two sources corresponding to points s1 and s2 = s1 + ∆s on the array
manifold A are detected if and only if the uncertainty spheres do not make
contact with the detection threshold subspace Hdet defined in Eq. (8.52).
When, however, the subspace Hdet becomes the tangent plane to the uncer-
tainties spheres the threshold of detection occurs. To a first order approxi-
mation, this is equivalent to the following definition.
Fig. 8.7 Detection threshold subspace — an illustrative visualization.
Fig. 8.8 Resolution threshold subspace — an illustrative visualization.
Definition 8.1 Detection Threshold: Two sources are detected if and

only if the uncertainty spheres do not make contact. The threshold of detec-
tion occurs when the two uncertainty spheres just make contact.
The two sources remain undetectable if their associated uncertainty
spheres overlap. This implies that the arc length separation ∆s = |s2 − s1 |
between two points s1 and s2 associated with two sources should be greater
Array Bounds 193
Fig. 8.9 An illustrative representation of the geometry at “detection threshold” on the

H12 subspace (circular approximation).
than, or equal to, ∆sdet-thr in order to be detected. That is
∆s ∆sdet-thr (8.55)
Based on this definition and once again using the circular approxima-
tion/representation of the array manifold, the scenario of Fig. 8.2, at the
detection threshold, pertains:
µy1 = σe1,d and µy2 = σe2,d (8.56)
as shown in Fig. 8.9. Hence Eq. (8.21) becomes
∆sdet-thr = κ̂−1
1 ∆ψ
∆ψ1 ∆ψ2

arcsin(κ̂1 σe1,d ) + arcsin(κ̂1 σe2,d )
= (8.57)
κ̂1
Noting that arcsin(x) x for x 1, one may write
∆ψ1 = arcsin(κ̂1 σe1,d ) κ̂1 σe1,d (8.58)
∆ψ2 = arcsin(κ̂1 σe2,d ) κ̂1 σe2,d (8.59)

which implies that Eq. (8.57) can be rewritten as
∆sdet-thr = σe1,d + σe2,d (8.60)
providing a lower limit to the angular separation threshold as
1
∆pdet-thr = (σe + σe2,d ) (8.61)
π r sin p̆ 1,d
where

σei,d =
P⊥
H ⊥
Hdet a(si ) = a (si )PHdet a(si ) for i = 1, 2 (8.62)
It should be emphasized that detecting the presence of two sources does

not necessarily mean that their bearings have been resolved. In fact for a
sufficiently low SNR or a sufficiently small number of snapshots, a typical
super-resolution algorithm, operating in the presence of two closely located
sources, will always provide a single estimate (e.g. a spectrum with a single
null) even if it has been given the true number of sources a priori, that
is equal to two. That is even if the two sources are assumed to have been
detected.
8.3.2 Estimating the Resolution Threshold

Definition 8.2 Resolution Threshold: Two emitters corresponding
to points s1 and s2 on the manifold are resolved if and only if the uncer-
tainty spheres around these two points do not make contact with the
resolution threshold subspace, Hres L([a(s̆), u1 (s̆)]). The threshold of
resolution ∆sres-thr occurs when Hres is a tangent plane to the uncer-
tainty spheres (i.e. the uncertainty spheres just make contact with the
subspace Hres ).
This implies that the arc length separation ∆s = |s2 − s1 | between two
points s1 and s2 , corresponding to two already detected sources, should be
greater than, or equal to, ∆sres-thr in order to be resolved. That is
∆s ∆sdet-thr (8.63)
or remain unresolvable otherwise.

The objective here is to estimate the resolution threshold ∆sres-thr . That
is, to determine the minimum ∆s which is just sufficient to allow the uncer-
tainty spheres to make contact with the subspace Hres L([u1 (s̆), a(s̆)])
Array Bounds 195
Fig. 8.10 View of the H12 subspace (plane): an illustrative representation of the geom-
etry at “resolution threshold” on the H12 subspace (circular approximation).
at the point s̆. Once again using the circular representation of the array
manifold the situation of Definition 8.2 may be illustrated as in Figs. 8.10
and 8.11. With reference to these two figures, it can be seen that σe2,r (see
Fig. 8.11) and µx2 (see Fig. 8.10) as well as σe1,r and µx1 may be given as
follows:
σe1,r = P⊥ ⊥
Hres a(s1 ) and σe2,r = PHres a(s2 ) (8.64)
σe σe
µx1 = 1,r and µx2 = 2,r (8.65)
sin γ sin γ
Hence Eq. (8.21) becomes
∆sres-thr = ∆s1 + ∆s2 = κ̂−1

1 ∆ψ
∆ψ1 ∆ψ2

σe1,r σe2,r
arccos 1 − κ̂1 + arccos 1 − κ̂1
sin γ sin γ
= (8.66)
κ̂1
Fig. 8.11 View orthogonal to Hres and H12 subspaces (planes): an illustrative view of
the geometry at “resolution threshold”.
Recalling that arccos(1 − 12 x2 ) x for x 1, one may write

σ e1
∆ψ1 = arccos 1 − κ̂1
sin γ

σe
2κ̂1 1 (8.67a)
sin γ

σe
∆ψ2 = arccos 1 − κ̂1 2
sin γ

σ e2
2κ̂1 (8.67b)
sin γ
Upon substitution of Eq. (8.67) into (8.66), it follows that

2 √ √
∆sres-thr = σe1,r + σe2,r (8.68)
κ̂1 sin γ
However, (see Fig. 8.11)
P⊥
H1,2 a(s̆) 1/κ̂1 1
cos γ = = √ ⇒ sin2 γ = 1 − cos2 γ = 1 −
a(s̆) N N κ̂21
Array Bounds 197
where
H1,2 = L{[u1 (s̆), u2 (s̆)]}
Therefore,
9
4 √ √
∆sres-thr = 4 σe1,r + σe2,r (8.69)
κ̂21 − 1
N
or equivalently, the lower limit to the angular separation is

9
1 4 √ √
∆pres-thr = 4
( σe1,r + σe2,r ) (8.70)
π r sin p̆ κ̂21 − 1
N

=∆sres-thr
where

σei,r =
P⊥
H ⊥
Hres a(si ) = a (si )PHres a(si ), for i = 1, 2 (8.71)
Having defined the required performance thresholds and summarizing the

results in Table 8.3, it is apparent that the array detection–resolution
capabilities at p are functions of:
• the manifold arc length ∆s corresponding to the interval ∆p

• the manifold curvatures κ1 corresponding to bearing p
Table 8.3 Resolution and detection arc length as a function of uncertainty

spheres.
Resolution and detection thresholds
arc length: ∆s directional: ∆p ṡ(p

∆s
o)
with ṡ(po ) = πr sin p̆
∆sdet-thr = σe1,d + σe2,d ∆pdet-thr ≈ 1

πr sin p̆
(σe1,r + σe2,r )

4 1
∆sres-thr = 4 4 ∆pres-thr = 4
1 πr sin p̆
2 1 2
κ 1− N
κ1− N

√ √ √ √
× σe1,d + σe2,d × σe1,r + σe2,r

σei,d = PHdet a(si ), i = 1, 2
with
 σe = ⊥ a(s ),
PH i i = 1, 2
i,r res
• the number of sensors N

• the radius of the uncertainty spheres σe
where the fourth factor is dependent on the available SNR and the number
of array output snapshots L, and will be modelled next.
8.4 Modelling of the Uncertainty Sphere
As described above, the uncertainty spheres (in arc lengths) represent

the effects of additive sensor noise of power σ 2 on the system perfor-
mance. Furthermore, as was noted earlier, it is possible to asymptotically
eliminate the effects of noise by increasing the number of snapshots L.
That is
1
σe2 ∝ (8.72)
L
and as L → ∞ ⇒ σe → 0, i.e. zero estimation error. In addition, as the
power of the noise in the system tends to zero (σ → 0), i.e. the SNR =
P/σ 2 tends to infinity (SNR → ∞), the uncertainty sphere shrinks with an
uncertainty radius σe → 0. That is
1
σe2 ∝ (8.73)
SNR
This implies that the uncertainty sphere shrinks, or equivalently, that there
would be a gradual decrease in its radius σe , thus providing an effective
radius σe according to the following model:
/ 9
σ2 1
σe = = ; (in arc lengths) (8.74)
2LP 2(SNR × L)
It should be pointed out that the factor “2” appearing in the denominator
is a direct result of using complex numbers to represent signal envelopes in
the array model.
The shrinking rate of the uncertainty sphere depends also on the esti-
mation algorithm deployed. Some algorithms exhibit better properties than
others thereby “masking” the overall array performance and this makes it
difficult to identify the full detection-resolution capabilities inherent in the
array structure. Thus we may introduce the factor C (with 0 < C ≤ 1) to
Array Bounds 199
represent the reduction in performance due to the employment of a specific

practical estimation algorithm. That is
9
1
σe =
2(SNR × L)C
For C = 1 the above expression becomes a “lower bound” and forms a

benchmark against which any practical algorithm can be compared. One
interpretation of the value C = 1 may be that it is a theoretical limit
achieved by an “ideal” algorithm which does not introduce extra “uncer-
tainties” and eliminates any dependency which may exist between the var-
ious parameters of the received signals (for instance it decorrelates any
correlated signals, etc).
It is interesting to note at this point the relationship between the uncer-
tainty sphere model of Eq. (8.74) and the variance of the estimate of the
manifold’s arc length as predicted by the Cramer Rao lower bound of
Eq. (8.25). That is,
σe2 = CRB[s] (8.75)
The equality of the single source Cramer Rao lower bound with the square
of the uncertainty sphere radius can be interpreted as follows:
The uncertainty sphere represents the smallest achievable

uncertainty (optimal accuracy) due to the presence of noise
after L snapshots, when all the effects of the presence of
other sources have been eliminated by an “ideal” parameter
estimation algorithm.
8.5 Thresholds in Terms of (SNR × L)
Based on the previous discussion and the model of Eq. (8.74), at the detec-
tion and resolution “thresholds” the uncertainty spheres can be expressed as
9 @
@
1 @
σei,d = aH (si )P⊥
Hdet a(si ) = @ (8.76a)
2(SNR × L)C @
det
9 @
@
1 @
H
σei,r = a (si )P⊥
Hres a(si ) = @ (8.76b)
2(SNR × L)C @
res
Substituting Eq. (8.76a) back into Eqs. (8.60) and (8.69), with C = 1, it
follows that:
(a) Detection threshold:

6 6
∆sdet-thr = CRB[s1 |A] + CRB[s2 |A]

1 1 1
= √ √ +√
2 SNR1 × L SNR × L
/ 2
1 P1
= 6 1+
2(SNR1 × L) P2
(b) Resolution threshold:

9
4
∆sres-thr = 4 2 (8.77)
κ̂ − 1
& 61 N 6 '
× 4 CRB[s1 |A] + 4 CRB[s2 |A]
9
4 1
= 4 2 √
κ̂1 − N1 4 2

1 1
× 4 √ √
+ 4
SNR1 × L SNR2 × L
9 /
2 P1
= 4 2 1 + 4
(SNR1 × L) κ̂1 − N1 P2
The incremental manifold length ∆s may also be written in terms of

bearing separation ∆p as ∆s ∆p ṡ(p̆) where ṡ(p̆) = πr sin p̆ is the rate
of change of manifold arc length and bearing p̆ corresponds to the point s̆
on the manifold curve (to a first order approximation — p̆ = (p1 + p2 )/2).
This implies that the minimum angular separation ∆pdet-thr required for
the detection of two sources of powers P1 and P2 may be written as:

1 1 1
∆pdet-thr = √ √ +√
2ṡ(p̆) SNR1 × L SNR2 × L
(b) Resolution threshold: (8.78)

9
1 4 2
∆pres-thr = 2
ṡ(p̆) κ̂1 − N1

1 1
× √ + √
4
SNR1 × L 4
SNR2 × L
Array Bounds 201
Since ṡ(p̆) = πr sin p̆, Eq. (8.78) indicates that the resolution capabilities
of a linear array (or ELA) are maximum at broadside (p = 90◦ ) and zero
along the endfire directions (p = 0◦ or 180◦ ). Zero resolution along end-fire
directions is expected since, due to cylindrical symmetry, two signals at
directions which are symmetrical to the array axis cannot be distinguished,
irrespective of their angular separation.
For the special case of equi-powered sources (P1 = P2 = P ) the ultimate
product (signal-to-noise ratio times L) required to detect and resolve two
sources at bearings p1 and p2 = p1 + ∆p may be written as
√
2 1
∆pdet-thr ≈ √
ṡ(p̆) SNR × L
9 (8.79)
1 4 32 1
∆pres-thr ≈ 2 √
ṡ(p̆) κ̂1 − N1 4 SNR × L
revealing the following square-root law for detection threshold :
∆pdet-thr ∝ (SNR × L)−1/2 (8.80)
and fourth-root law for resolution:
∆pres-thr ∝ (SNR×L)−1/4 (8.81)
In other words the detection and resolution ultimately achievable by a lin-

ear array is inversely proportional to the square-root and fourth-root law
respectively, of
• the signal-to-noise ratio

• the number of snapshots
Furthermore, in the case of equi-powered sources (P1 = P2 ) a relation-

ship between the ultimate resolving and detection capabilities of an array
can be determined by dividing ∆pres-thr by ∆pdet-thr . That is
9
∆pres-thr 8 √
= 4
SNR × L (8.82)
∆pdet-thr κ̂1 − N
2 1
which indicates that resolution is a more demanding operation compared to

detection, since the right hand side of the above equation is always greater
than unity (1). In addition, Eq. (8.79) reveals that the detection threshold
falls more rapidly than the resolution threshold as the number of sensors
N increases for a constant SNR × L.
It is interesting to note that the resolution threshold ∆pres-thr of a linear
array is a function of sin(p). This indicates that an algorithm which per-
forms an exhaustive search of the manifold (e.g. MUSIC) should evaluate
its cost function along a non-uniform grid of azimuths whose values follow
a sinusoidal distribution. Such a non-uniform angular grid corresponds to a
uniform arc length grid along the manifold and would allow full use of the
array’s resolution capabilities.
Finally Eq. (8.77) can be used to derive other useful system require-
ments. For example, the number of snapshots, Ldet-thr , required to detect
two sources at bearings p1 and p2 = p1 + ∆p or Lres-thr to resolve them,
may be written as shown below.

2
1 1 1
Ldet-thr = √ +√
2∆s2 SNR1 SNR2
/ 2
1 P1
(SNR1 × L)det-thr = 2
1+
2∆s P2
(b) Resolution threshold:

4
2 1 1
Lres-thr = √ + √
1 4
SNR1 4
SNR2
∆s4 κ̂21 −
N
/ 4
2 P1
(SNR1 × L)res-thr = 1+ 4
1 P2
∆s4 κ̂21 −
N

∆s = πr| cos p2 − cos p1 |
where
∆s1 κ̂1 1, ∆s2 κ̂1 1
(8.83)
Above, the expressions are also provided for the minimum product
(SNR1 × L) required for a source of power P1 to be detected and then
resolved in the presence of a source of power P2 after L snapshots when the
two sources are separated by ∆s.
Finally Table 8.4 summarizes the main expressions for the special case
of two equi-power sources.
Array Bounds 203
Table 8.4 Detection and resolution expressions for equipowered

sources (P1 = P2 = P ).
Detection Resolution
√
∆pdet-thr ≈ 2 √ 1
∆pres-thr ≈ 1 4 32
1 √
4
1
ṡ(p̆) SNR×L ṡ(p̆) (κ21 − N ) SNR×L
(SNR × L)det-thr = 2
∆s2
(SNR × L)res,thr = 32
1
1− N )
∆s4 (κ2
2 1 32 1
Ldet-thr = ∆s2 SNR
Lres-thr = 1
1− N )
∆s4 (κ2 SNR
where ṡ(p̆) = π r sin p̆
8.6 Comments
8.6.1 Schmidt’s Definition of Resolution

An alternative definition to the one presented in the previous section (see
Definition 8.2 — page 194) for the resolution threshold was proposed by
Schmidt [1] stating that two sources corresponding to points s1 and s2
on the manifold are resolved if, and only if, the uncertainty spheres do
not make contact with the tangent u1 (s̆) to the manifold at s̆. Using this
definition, which was given without any analysis, the resolution threshold
subspace Hres can be defined as
Hres L ([u1 (s̆)]) (8.84)
having dimensionality one. Schmidt’s definition, in conjunction with the

circular approximation of the array manifold is illustrated in Fig. 8.12.
Based on this figure, it can be proven using the approach presented in the
previous sections that the resolution ∆s is given by
9
4 &6 6 '
∆sres-thr = 4 2 4 CRB[s1 |A] + 4 CRB[s2 |A] (8.85a)
κ1
9
4 1 1 1
= 4 2√ √ + √ (8.85b)
κ1 4 2 4 SNR1 × L 4
SNR2 × L
9 : / ;
2 P1
= 4 1+ 4 (8.85c)
(SNR1 × L) κ21 P2
The above expression is an approximation to Eq. (8.77) where the term

κ̂21 − N1 has been replaced by κ21 . These two expressions become identical
Fig. 8.12 Illustrative representation of Schmidt’s definition.
for large N and inclination angle ζ = 90◦ . That is for symmetric arrays
with a large number of sensors.
8.6.2 CRB at the Resolution Threshold

If the CRB is evaluated at the threshold level (SNR1 × L)res , then the
estimation error (accuracy of the estimate) at the resolution threshold is
given as follows:
1
CRB[p1 |A]res = (ȧH Pȧ1 )−1 (8.86)
2(SNR1 × L)res 1
1 (∆p × ṡ(p))2
4 (8.87)
ṡ2 (p1 )
1+ 4 P 1
P2
It is important to note that for two equi-powered sources the above expres-
sion is simplified to
∆p ṡ(p̆)
CRB[p1 ]res .
4 ṡ(p1 )
∆p
→ (8.88)
4
[as (p̆−p1 )→0]
Array Bounds 205
Fig. 8.13 Simulation results of the CRB (exact and approximate expressions) as a
function of azimuth separation, averaged over 50 trials. The resolution threshold and the
single-source CRB are also shown for comparison.
which is an expected result indicating the generality and significance of

the resolution threshold. Figure 8.13 illustrates the variation, with azimuth
separation, for the CRB given by Eqs. (8.37) (exact) and (8.39) (approx.)
together with the results of simulation studies and the resolution threshold.
Remember that for large separations the exact CRB converges to the single-
source CRB while the differential geometry version wrongly tends to zero —
although this is not obvious in Fig. 8.13.
8.6.3 Directional Arrays

The above discussion was carried out under the assumption that the pla-
nar array consists of antennas that are isotropic (with gain of unity) in
both azimuth θ and elevation φ. This assumption might seem unrealis-
tic since many practical antennas (like the elevated-feed monopoles) are
Table 8.5 Expressions for directional array of sensors.
1 −1
CRB[p1 |A] = 1 P ȧ1 )
(ȧH
2 SNR1 × L|g(p1 )|2
1 2

(SNR1 × L)|g(p1 )|2 (∆s)22 (p1 )(
κ21 (p) − 1/N )
!2
1 |g(p1 )|2 P1
(SNR1 × L)det = 1+
2(∆p × (p))2 |g(p1 )|2 |g(p2 )|2 P2
!4
2 |g(p1 )|2 P1
(SNR1 × L)res = 1+ 4
(∆p × (p))4 (
κ21 (p) − 1
N
)|g(p1 )|2 |g(p2 )|2 P2
non-isotropic and exhibit a complex gain response γk ∈ C1 as a function of

one or both bearing parameters. In this case the array manifold is given by
Eq. (1.22) of Chapter 1.
It is easy to show that the results of the previous section can be trans-
formed for directional sensors to the expressions presented in Table 8.5. It
is clear from Table 8.5 that the directional pattern g(p) of a directional
sensor behaves simply as a “voltage gain” term boosting or deteriorating
the effective Signal-to-Noise Ratio at the output of the array. Consequently,
it can be stated that the presence of directional sensors does not affect the
relative merits of one array geometry over the other, although of course,
performance is affected in absolute terms.
8.7 Array Capabilities Based on α- and β-curves
In the previous sections the estimation accuracy, detection and resolution

capabilities of an array of sensors and their dependence on the differential
geometry of manifold parameter curves were investigated. It was seen that
for sufficiently close emitters, the array performance is a function of the local
shape of a manifold curve as specified by its length and first curvature. It
was consequently deduced that while performance varies with azimuth θ in
accordance with the array configuration, its variation with elevation obeys
simple sinusoidal laws irrespective of the array. While these deductions are
completely valid, examination of array performance in terms of cone-angles
can provide an alternative, and perhaps a clearer picture of the operation
of a planar array.
Array Bounds 207
Theorem 8.1 The capability of a planar array to estimate, detect and

resolve cone-angles α1 and α2 of rotation angle Θo for two emitters at
bearings (α1 , βo ) and (α2 , βo ) is
(i) independent of the value of βo ,

(ii) identical to its capability to estimate, detect and resolve the elevations
of two emitters at (azimuth, elevation) bearings:
◦
(θ = Θo , φ = α1 ) and (θ = Θo , φ = α2 ) if α1 , α2 < 90 (8.89)
or
◦
(θ = Θo , φ = 180◦ −α1 ) and (θ = Θo , φ = 180◦ −α2 ) if α1 , α2 > 90
(8.90)
Proof.
(i) According to Corollary 5.2 and Theorem 5.1, all members of the family
of α-curves are identical, i.e. their differential geometry is independent
of β.
(ii) From Corollary 5.5 and Theorem 5.1, each member of the α-curve
family is identical to the combination of the φ-curves corresponding
to θo = Θo and θo = Θo + 180◦ . This implies that α-estimation and
φ-estimation performance must by necessity be equivalent.

Naturally, the roles of α and β may be interchanged in Theorem 8.1 by

simply replacing Θo with Θo + 90◦ .
Part (i) of the above theorem suggests an underlying “conic” behavior in
the performance of a planar array which can not be observed by considering
azimuth and elevation angles alone. Figure 8.14 illustrates this behavior
by showing the loci of two directions of arrival separated by ∆α = 2◦
(α1 = 68◦ , α1 = 70◦ ) for all possible values of β and the corresponding θ-φ
values.
Part (ii) of Theorem 8.1 indicates that the properties of φ-estimation
array performance conducted in Section 8.2 is directly applicable to all cor-
responding cone-angles. In particular, one might immediately deduce that
the α- and β-estimation performances of a planar array of omnidirectional
sensors are at their peaks when α = 90◦ and β = 90◦ respectively, and
degrade sinusoidally as α and β approach either 0◦ or 180◦ .
Fig. 8.14 The loci of two DOAs separated by ∆α = 2◦ for all values of β and the
corresponding DOA loci on the (θ, φ) parameter plane revealing the “conic” nature of
the array performance.
8.8 Summary
In this chapter resolution and detection thresholds/bounds and estimation

accuracy have been studied by approximating, locally, a manifold curve with
a circular arc. In particular, by defining the detection and resolution sub-
spaces, in conjunction with the circular approximation of the array mani-
fold, the minimum arc length separations for detecting and resolving two
sources located close together have been estimated. This is done in terms
of the curve’s principal curvature, thereby simplifying the analysis consid-
erably. Furthermore, the relationship between Cramer Rao Lower Bound
(CRB) and resolution and detection thresholds has been established. These
thresholds were based on differential geometry properties of the manifolds
and can be used as a “figure of merit” for comparing array geometries.
Array Bounds 209
8.9 Appendices
8.9.1 Radius of Circular Approximation

Based on Fig. 8.2(b) the radius of the corresponding circular arc is

κ̂−1
1 = PH12 a(s̆) = aH (s̆)PH12 a(s̆) (8.91)
where PH12 is the projection operator on to the subspace
H12 = L ([u1 (s̆), u2 (s̆)]) (8.92)
However,
aH (s̆)PH12 a(s̆) (below the variance s̆ is suppressed for convenience)

4 5−1
u1 2 , uH
1 u2
H
= aH u1 , u2 u1 , u2 a
uH
2 u1 , u2 2
4 5−1 4 5
H
1, cos ζ 0
= 0, a u2
cos ζ, 1 uH
2 a
4 5 4 5
1
1, −cos ζ 0
= 0, −κ−1
sin2 ζ −cos ζ,
1
1 −κ−1
1
4 5
1 −1 0
= +κ1 cos ζ, −κ−1 −1
2
sin ζ 1
−κ 1
2
κ−1
1
= (8.93)
sin ζ
Therefore from Eqs. (8.91) and (8.93) we have
κ̂−1
1 = PH12 a(s̆)
κ−1
1
= (8.94)
sin ζ
which implies that
κ̂1 = κ1 sin ζ (8.95)

8.9.2 “Circular” and “Y” Arrays — Sensor Locations
“Circular” “Y”
Sensor
number x y x y
1st 75.00 0.00 8.00 0.00

2nd 72.44 19.41 22.00 0.00
3rd 64.95 37.50 38.00 0.00
4th 53.03 53.03 57.00 0.00
5th 37.50 64.95 79.00 0.00
6th 19.41 72.44 105.00 0.00
7th 0.00 75.00 136.00 0.00
8th −19.41 72.44 170.00 0.00
9th −37.50 64.95 −4.00 6.93
10th −53.03 53.03 −11.00 19.05
11th −64.95 37.50 −19.00 32.91
12th −72.44 19.41 −28.50 49.36
13th −75.00 0.00 −39.50 68.42
14th −72.44 −19.41 −52.50 90.93
15th −64.95 −37.50 −68.00 117.78
16th −53.03 −53.03 −85.00 147.22
17th −37.50 −64.95 −4.00 −6.93
18th −19.41 −72.44 −11.00 −19.05
19th 0.00 −75.00 −19.00 −32.91
20th 19.41 −72.44 −28.50 −49.36
21st 37.50 −64.95 −39.50 −68.42
22nd 53.03 −53.03 −52.50 −90.93
23rd 64.95 −37.50 −68.00 −117.78
24th 72.44 −19.41 −85.00 −147.22
8.9.3 Proof: CRB of Two Sources in Terms of κ1

Let u11 u1 (s1 ) and A [a1 , Ar ] where Ar = [a2 , . . . , aM ]. Then the term
⊥
11 PA u11 of Eq. (8.38) can be rewritten as follows:
uH
⊥ ⊥ ⊥ H ⊥ −1 H ⊥
11 PA u11 = u11 (Pa1 − Pa1 Ar (Ar Pa1 Ar )
uH Ar Pa1 )u11
H
⊥ H ⊥ H ⊥ −1 H ⊥
11 Pa1 u11 − u11 Pa1 Ar (Ar Pa1 Ar )
= uH Ar Pa1 u11
H ⊥ −1 H
= 1 − uH
11 Ar (Ar Pa1 Ar ) Ar u11
−1
1 H
= 1 − u11 Ar Ar Ar − Ar a1 a1 Ar
H H H
AH
r u11 (8.96)
N
where P⊥a1 u11 = u11 has been used.

Substituting the above equation back into Eq. (8.38) and then express-
ing the columns of the matrix Ar as ai = a1 + ∆ai (for i = 2, . . . , M ), the
Array Bounds 211
CRB on the bearing p1 can be expressed as a function of the array manifold

parameters.
However, it is informative to evaluate the CRB for two emitters (M = 2)
which are closely spaced at bearings p1 and p2 = p1 + ∆p, corresponding
to arc lengths s1 = s̆ − ∆s/2 and s2 = s̆ + ∆s/2 respectively.
In this case the matrix Ar becomes the manifold vector a2 and Eq. (8.96)
is simplified as follows
⊥ N |uH11 a2 |
2
11 PA u11 = 1 −
uH (8.97)
N 2 − |aH1 a2 |
2
The physical proximity of the arriving signals allows the use of local
differential geometry for the evaluation of Eq. (8.38) (or, of Eq. (8.36) for
the general case of M > 2).
11 a2 | = |u1 (s1 ) a(s2 )| . Clearly:

(a) Evaluation of the term |uH H
u1 (s1 )H a(s2 ) = u1 (s1 )H (a(s1 ) + ∆a) = u1 (s1 )H ∆a (8.98)
However, the manifold in the neighborhood of s̆ may be interpreted

κ1 (s̆) as depicted in Fig. 8.15. In such
as a circular arc of radius 1/
circumstances the inner product may be written as
u1 (s1 )H ∆a ∆a cos ∆ψ

∆a 1 − sin2 ∆ψ
/
1
∆a 1 − ∆a2 κ̂21 (s̆) (8.99)
4
Fig. 8.15 Circular approximation of the manifold in the neighbourhood of s̆.

or
/
1 2
u1 (s1 )H a(s2 ) ∆a 1− ∆a κ̂21 (s̆) (8.100)
4
(b) Evaluation of the term aH H
1 a2 = a(s1 ) a(s2 ). Clearly:
a(s1 )H a(s2 ) = a(s1 )H (a(s1 ) + ∆a) = N − a(s1 )H ∆a (8.101)
The array manifold has constant norm and hence lies on the surface
of a hypersphere. Consequently, for sufficiently small ∆s, vectors a(s̆)
and ∆a are strictly orthogonal as indicated in Fig. 8.16. Using simple
trigonometry:
√
a(s1 )H ∆a N ∆a cos(90◦ + γ) (8.102)
√
− N ∆a sin(γ) (8.103)
√ ∆a 1
− N ∆a √ = − ∆a2 (8.104)
2 N 2
Therefore
a(s1 )H a(s2 ) N − 12 ∆a2 (8.105)
Substituting Eqs. (8.104) and (8.105) back into Eq. (8.97) and then
using Eq. (8.38):
: ;−1
1 N ∆a2 1 − 14 ∆a2 κ̂21 (s̆)
CRB[p|A] = 1−
2(SNR1 × L)ṡ(p1 )2 N 2 − (N − 12 ∆a2 )2
Fig. 8.16 For sufficiently small ∆s, the vectors a(s̆) and ∆a are strictly orthogonal.
Array Bounds 213
−1
1 N (1 − 14 ∆a2 κ̂21 (s̆))
= 1−
2(SNR1 × L)ṡ(p1 )2 N − 14 ∆a2

1 4N − ∆a2
=
2(SNR1 × L)ṡ(p1 )2 ∆a2 (N κ̂21 (s̆) − 1
So finally:
1 2
CRB[p|A] = (8.106)
(SNR1 × L) ṡ(p1 )2 (∆s)2 κ̂21 (p̆) − 1
N
where |∆a| ∆s = ∆p ṡ(s̆) and it has been assumed that 4N |∆a|2 .

It is important to mention that a slight variation to the above expres-
sion may be obtained via a second-order Taylor expansion of a(s2 ) about
s1 . However simulations indicate that Eq. (8.39) provides a more accurate
approximation.
July 6, 2004 9:31 WSPC/Book Trim Size for 9in x 6in bib
Bibliography
[1] R. Schmidt, A Signal Subspace Approach to Multiple Emitter Location and

Spectral Estimation. PhD thesis, Stanford University USA, November 1981.
[2] U. Lipschutz, Theory and Problems of Differential Geometry. Shaum’s Out-
line Series, McGraw Hill, 1969.
[3] M. D. Carmo, Differential Geometry of Curves and Surfaces. Prentice Hall
Inc., New Jersey, 1976.
[4] H. Akaike, “A new look at the statistical model identification,” IEEE Trans.
on Autom. Control, Vol. 19, pp. 716–723, December 1974.
[5] M. Wax and T. Kailath, “Detection of signals by information theoretic cri-
teria,” IEEE Trans. on Acoustics, Speech and Signal Processing, Vol. 33,
pp. 387–392, April 1985.
[6] M. Wax and I. Ziskind, “Detection of the number of coherent signals by the
MDL principle,” IEEE Trans. on Acoustics, Speech and Signal Processing,
Vol. 33, pp. 1190–1196, August 1989.
[7] R. Schmidt, “Multiple emitter location and signal parameter estimation,”
IEEE Trans. on Antennas and Propagation, Vol. 34, pp. 276–280, March
1986.
[8] P. Karaminas and A. Manikas, “Superresolution broad null beamforming for
co-channel interference cancellation in mobile radio networks,” IEEE Trans.
on Vehicular Technology, Vol. 49, pp. 689–697, May 2000.
[9] B. Steinberg, Microwave Imaging with Large Antenna Arrays. Wiley, 1983.
[10] I. Dacos and A. Manikas, “Estimating the manifold parameters of one-
dimensional arrays of sensors,” J. Franklin Institute (Engineering & Applied
Mathematics), Vol. 332B, pp. 307–332, May 1995.
[11] A. Manikas, A. Sleiman and I. Dacos, “Manifold studies of nonlinear antenna
array geometries,” IEEE Trans. on Signal Processing, Vol. 49, pp. 497–506,
March 2001.
[12] H. Karimi and A. Manikas, “Cone-angle parametrization of the array man-
ifold in DF systems,” J. Franklin Institute (Engineering & Applied Mathe-
matics), Vol. 335B, No. 2, pp. 375–394, 1998.
215
July 6, 2004 9:31 WSPC/Book Trim Size for 9in x 6in bib
[13] A. Manikas and C. Proukakis, “Modelling and estimation of ambiguities in

linear arrays,” IEEE Trans. on Signal Processing, Vol. 46, pp. 2166–2179,
August 1998.
[14] A. Manikas, V. Lefkadites and C. Proukakis, “Investigative study of planar
array ambiguities based on ‘hyperhelical’ parametrisation,” IEEE Trans. on
Signal Processing, Vol. 47, pp. 1532–1542, June 1999.
[15] H. R. Karimi and A. Manikas, “Manifold of a planar array and its effects
on the accuracy of direction-finding systems,” IEE Proc. Radar, Sonar and
Navigation, Vol. 143, pp. 349–357, December 1996.
[16] A. Manikas, A. Alexiou and H. R. Karimi, “Comparison of the ultimate
direction-finding capabilities of a number of planar array geometries,” IEE
Proc. Radar, Sonar and Navigation, Vol. 144, pp. 321–329, December 1997.
[17] H. Guggenheimer, Differential Geometry. McGraw-Hill (or Dover Edition),
1963 (1977).
[18] A. Sleiman and A. Manikas., “The impact of sensor positioning on the array
manifold,” IEEE Trans. on Antennas and Propagation, Vol. 51, pp. 2227–
2237, September 2003.
[19] T. Willmore, An Introduction to Differential Geometry. Oxford University
Press (UK), 1959.
[20] M. Spivak, A Comprehensive Introduction to Differential Geometry. Publish
or Perish, Berkley, CA, 1979.
[21] L. Eisenhart, Riemannian Geometry. Princeton University Press, Princeton,
NJ, 1949.
[22] M. Marcus and H. Minc, A Survey of Matrix Theory and Matrix Inequalities.
Dover, 1992.
[23] H. Van Trees, Detection, Estimation and Modulation Theory Part-1. John
Wiley and Sons, US, 1968.
[24] P. Stoica and A. Nehorai, “MUSIC, maximum likelihood and Cramer-Rao
bound,” IEEE Trans. on Acoustics, Speech and Signal Processing, Vol. 37,
pp. 720–741, May 1989.
[25] Y. I. Abramovich, N. K. Spencer and A. Y. Gorokhov, “Detection-estimation
of more uncorrelated gaussian sources than sensors in nonuniform lin-
ear antenna arrays—Part III: detection-estimation nonidentifiability,” IEEE
Trans. on Signal Processing, Vol. 51, pp. 2483–2494, October 2003.
July 6, 2004 9:32 WSPC/Book Trim Size for 9in x 6in INDEX
Index
2D-grid array, 81 cone angles, 98

3D array, 8 conoid, 82
3D-grid array, 80, 89 constant-parameter curve, 60
coordinate vectors, 29
Ambiguous Generator Set (AGS), 21, CRB, 21, 179
130 at the resolution threshold, 204
ambiguities, 18, 113 curvatures, 26, 29, 30, 38–40
non-trivial, 116 manifold curvatures, 35
trivial, 115 first (principal) curvature, 21, 25
ambiguities on α-curves/β-curves, 139
ambiguities on φ-curves, 136
derivatives of the wavenumber vector,
ambiguous generator lines, 149 77
ambiguous set, 118
detection threshold, 192
arc length, 22
developable surfaces, 68, 82
array aperture, 11
development, 72, 87
array manifold, 17
distinct set of intersensor spacings,
array manifold matrix, 16
124
array manifold vector, 14
dual arrays, 48
in terms of cone angle, 100–101
array response vector, 15
array symmetricity, 37 Equivalent Linear Array (ELA), 4,
15, 16, 114, 130, 149, 171
balance symmetric, 81
family of curves
Cartan matrix, 30 φ-parameter curves, 84
characteristic points, 159, 160, 163 p-parameter curves, 71
Christoffel matricies, 103 q-parameter curves, 71
Christoffel symbols, 59 θ-parameter curves, 83
first kind, 64 first curvature, 21, 25
second kind, 64 first fundamental coefficients, 59, 61
circular approximation, 21, 174 first fundamental form, 63
composite curves, 87 frame matrix, 28
217
July 6, 2004 9:32 WSPC/Book Trim Size for 9in x 6in INDEX
Gaussian curvature, 20, 59, 66 normalized sensor positions, 35

geodesic coordinates, 71, 85 number of AGSs, 125
geodesic curvature, 20, 59, 66, 68–70 number of half-windings, 41
geodesic curve, 69 number of windings, 41
hyperhelical curves (or hyperhelical orthogonality

manifold curves), 33, 86 narrow sense, 31, 32
hyperhelix, 48 wide sense, 31, 32
inclination angle, 45, 48 parameter curve, 59

intrinsic geometry of a surface, 65 planar array, 8, 89
based on cone angles, 101 principal curvature, 21, 25
invariant parameter, 24 projection operator, 9
isometric mapping, 20, 65
isometry, 65 rank of ambiguity, 118
rate of change of arc length, 33
law regular parametric representation, 61
fourth-root, 201 regular parametrized differential
square-root, 201 curve, 22, 33
linear arrays, 6 regularity condition, 22
fully asymmetric, 27 resolution threshold, 194
non-symmetric, 27
partially symmetric, 27 sufficient condition for the presence
symmetric, 27, 157 of ambiguities, 126
superresolution, 18
manifold curve, 33
manifold length, 41 tangent plane, 61
manifold metric, 59, 61 threshold
manifold radii vector, 47 detection, 21, 192
mapping resolution, 21, 194
conformal, 66 subspaces, 191
geodesic, 66, 70, 89
isometric, 20, 65, 66, 89 uncertainties spheres, 191, 198
metric coefficients, 62 Uniform Basic Set (UBS), 124
moving frame, 28
MUSIC, 18, 50 wavenumber matrix, 16
wavenumber vector, 6, 14
narrow sense orthogonality, 32 derivatives, 77
narrowband assumption, 13 wide sense orthogonality, 32
non-linear arrays windings
balanced-symmetric, 106 half-windings, 41
unbalanced-symmetric, 106 number of windings, 41
of a hyperhelix, 41

Manikas A. - Differential Geometry in Array Processing-ICP (2004)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Manikas A. - Differential Geometry in Array Processing-ICP (2004)

Uploaded by

Copyright:

Available Formats

DIFFERENTIAL GEOMETRY IN

Imperial College Press

British Library Cataloguing-in-Publication Data

DIFFERENTIAL GEOMETRY IN ARRAY PROCESSING

Typeset by Stallion Press

34_Differential Geometry.pmd 1 10/4/2005, 6:12 PM

viii Diﬀerential Geometry in Array Processing

underlying and under-pinning objective of this book is to present a sum-

Thanassis Manikas — London 2003

This book is based on a number of publications (presented under a uniﬁed

linear arrays, non-linear arrays, planar arrays, array manifolds, diﬀerential

2. Diﬀerential Geometry of Array Manifold Curves 22

x Diﬀerential Geometry in Array Processing

2.4 The Manifold Length and Number of Windings

3. Diﬀerential Geometry of Array Manifold Surfaces 59

4. Non-Linear Arrays: (θ, φ)-Parametrization of Array

5. Non-Linear Arrays: (α, β)-Parametrization 97

6. Array Ambiguities 113

7. More on Ambiguities: Symmetrical Arrays 157

xii Diﬀerential Geometry in Array Processing

8. Array Bounds 174

An array system is a collection of sensors (transducers) which are spatially

(1) Detection problem — concerned with the determination/estimation of

2 Diﬀerential Geometry in Array Processing

minimization of a function of the noise eigenvalues of the array output

eﬃciency. Indeed, the integration of array processing and communications

• The presence of inherent background and sensor noise.

However, the overall quality of the system’s performance is naturally a

4 Diﬀerential Geometry in Array Processing

to certain “invariant” parameters. The overall notation to be employed in

1.2 Main Abbreviations

AGS Ambiguous Generator Set

1.3 Array of Sensors — Environment

By distributing, in the 3-dimensional Cartesian space, a number N 2

with rk = [xk , yk , zk ]T ∈ R3×1 denoting the Cartesian coordinates (loca-

6 Diﬀerential Geometry in Array Processing

pointing towards the direction (θ, φ) is

Note that u = 1. If the velocity, wavelength and frequency of propaga-

In the most general case, the parameter space is

Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ) and φ ∈ (−90◦ , 90◦ )} (1.4)

but in most applications, Ω is restricted to only a sector of interest or,

(1) Linear or 1-dimensional (1D) Array.

r = [rx , 0N , 0N ]T ∈ R3×N (1.5)

The most popular array of this type is the standard Uniform

ULA is given by rx = [−2, −1, 0, 1, 2]T in half wavelengths. Note that

Ω = {(θ, φ) : θ ∈ [0◦ , 180◦ ), φ = 0◦ } (1.6)

8 Diﬀerential Geometry in Array Processing

(2) Planar or 2-dimensional (2D) Array.

r = [rx , ry , 0N ]T ∈ R3×N (1.7)

where rx and ry , respectively, are the column vectors of length N denot-

Ω = {(θ, φ) : θ ∈ [0◦ , 360◦ ), φ ∈ [0◦ , 90◦ )} (1.8)

Note that a planar array cannot distinguish between directions which

1.4 Pictorial Notation

1.4.2 Projection Operator

dim {L[A]} = M < N (1.9)

lying in a N -dimensional space H, as shown in Fig. 1.4. The complement

(1) any vector x ∈ L[A] can be written as a linear combination of the

PA = projection operator onto the subspace L[A]

10 Diﬀerential Geometry in Array Processing

1.5 Principal Symbols

In this book the following symbols will always denote

1.6 Modelling the Array Signal-Vector and

Consider an array of N sensors, with sensor locations r, operating in the

To be more speciﬁc, if ρi 2la /λ, where λ is the wavelength of the trans-