Electromagnetic Scattering From Random Media

INTERNATIONAL SERIES OF MONOGRAPHS
ON PHYSICS
SERIES EDITORS
J. BIRMAN CITY UNIVERSITY OF NEW YORK

S. F. EDWARDS UNIVERSITY OF CAMBRIDGE
R. FRIEND UNIVERSITY OF CAMBRIDGE
M. REES UNIVERSITY OF CAMBRIDGE
D. SHERRINGTON UNIVERSITY OF OXFORD
G. VENEZIANO CERN, GENEVA
INTERNATIONAL SERIES OF MONOGRAPHS ON PHYSICS
144. T. R. Field: Electromagnetic scattering from random media
143. W. Götze: Complex dynamics of glass-forming liquids – a mode-coupling theory
142. V.M. Agranovich: Excitations in organic solids
141. W.T. Grandy: Entropy and the time evolution of macroscopic systems
140. M. Alcubierre: Introduction to 3 + 1 numerical relativity
139. A. L. Ivanov, S. G. Tikhodeev: Problems of condensed matter physics – quantum
coherence phenomena in electron-hole and coupled matter-light systems
138. I. M. Vardavas, F. W. Taylor: Radiation and climate
137. A. F. Borghesani: Ions and electrons in liquid helium
136. C. Kiefer: Quantum gravity, Second edition
135. V. Fortov, I. Iakubov, A. Khrapak: Physics of strongly coupled plasma
134. G. Fredrickson: The equilibrium theory of inhomogeneous polymers
133. H. Suhl: Relaxation processes in micromagnetics
132. J. Terning: Modern supersymmetry
131. M. Mariño: Chern-Simons theory, matrix models, and topological strings
130. V. Gantmakher: Electrons and disorder in solids
129. W. Barford: Electronic and optical properties of conjugated polymers
128. R. E. Raab, O. L. de Lange: Multipole theory in electromagnetism
127. A. Larkin, A. Varlamov: Theory of fluctuations in superconductors
126. P. Goldbart, N. Goldenfeld, D. Sherrington: Stealing the gold
125. S. Atzeni, J. Meyer-ter-Vehn: The physics of inertial fusion
123. T. Fujimoto: Plasma spectroscopy
122. K. Fujikawa, H. Suzuki: Path integrals and quantum anomalies
121. T. Giamarchi: Quantum physics in one dimension
120. M. Warner, E. Terentjev: Liquid crystal elastomers
119. L. Jacak, P. Sitko, K. Wieczorek, A. Wojs: Quantum Hall systems
118. J. Wesson: Tokamaks, Third edition
117. G. Volovik: The Universe in a helium droplet
116. L. Pitaevskii, S. Stringari: Bose–Einstein condensation
115. G. Dissertori, I.G. Knowles, M. Schmelling: Quantum chromodynamics
114. B. DeWitt: The global approach to quantum field theory
113. J. Zinn-Justin: Quantum field theory and critical phenomena, Fourth edition
112. R.M. Mazo: Brownian motion – fluctuations, dynamics, and applications
111. H. Nishimori: Statistical physics of spin glasses and information processing – an
introduction
110. N.B. Kopnin: Theory of nonequilibrium superconductivity
109. A. Aharoni: Introduction to the theory of ferromagnetism, Second edition
108. R. Dobbs: Helium three
107. R. Wigmans: Calorimetry
106. J. Kbler: Theory of itinerant electron magnetism
105. Y. Kuramoto, Y. Kitaoka: Dynamics of heavy electrons
104. D. Bardin, G. Passarino: The Standard Model in the making
103. G.C. Branco, L. Lavoura, J.P. Silva: CP Violation
102. T.C. Choy: Effective medium theory
101. H. Araki: Mathematical theory of quantum fields
100. L. M. Pismen: Vortices in nonlinear fields
99. L. Mestel: Stellar magnetism
98. K. H. Bennemann: Nonlinear optics in metals
96. M. Brambilla: Kinetic theory of plasma waves
94. S. Chikazumi: Physics of ferromagnetism
91. R. A. Bertlmann: Anomalies in quantum field theory
90. P. K. Gosh: Ion traps
88. S. L. Adler: Quaternionic quantum mechanics and quantum fields
87. P. S. Joshi: Global aspects in gravitation and cosmology
86. E. R. Pike, S. Sarkar: The quantum theory of radiation
83. P. G. de Gennes, J. Prost: The physics of liquid crystals
73. M. Doi, S. F. Edwards: The theory of polymer dynamics
69. S. Chandrasekhar: The mathematical theory of black holes
51. C. Møller: The theory of relativity
46. H. E. Stanley: Introduction to phase transitions and critical phenomena
32. A. Abragam: Principles of nuclear magnetism
27. P. A. M. Dirac: Principles of quantum mechanics
23. R. E. Peierls: Quantum theory of solids
Electromagnetic Scattering from
Random Media
Timothy R. Field
McMaster University
1
3
Great Clarendon Street, Oxford OX2 6DP
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York

c Oxford University Press 2009
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First Published 2009
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Field, Timothy R.
Electromagnetic scattering from random media/Timothy R. Field.
p. cm — (International series of monographs on physics; no. 144)
ISBN 978–0–19–857077–6
1. Stochastic processes. 2. Random fields. 3. Mathematical physics. 4. Electromagnetic
waves–Scattering. I. Title.
QC20.7.S8.F54 2008
530.14 1—dc22 2008042115
Typeset by Newgen Imaging Systems (P) Ltd., Chennai, India
Printed in Great Britain
on acid-free paper by
Biddles Ltd., King’s Lynn, Norfolk
ISBN 978–0–19–857077–6
1 3 5 7 9 10 8 6 4 2
To my family
This page intentionally left blank
PREFACE
The topic covered by the monograph is the scattering of electromagnetic waves

from random media. The book falls under the general categories of Mathemat-
ical Physics and Stochastic Methods. The treatment is based upon recent de-
velopments in the journal literature by the author, which is provided here with
sufficient background material required for a self-contained account.
The overall objective of the book is to formulate the space-time dynamics of
electromagnetic scattering processes from random media in terms of stochastic
differential equations (SDEs), and to demonstrate the utility of this theoretical
framework in model simulation and experimental data analysis. The results pro-
vide various means for detecting anomalous behaviour in the scattering process,
which should have significant consequences in several important application ar-
eas, including radar signal processing, optical scintillation effects and laser prop-
agation through turbulent media.
The book is divided into three main parts: Part I – Stochastic calculus the
opening chapters of which cover the necessary background to the Ito stochas-
tic calculus, including the Ito stochastic differential and stochastic integral and
related concepts. Part II – Stochastic dynamics of the scattering process discusses
Rayleigh scattering, population models, scattering generating K-distributed noise
resulting from the birth–death–immigration (BDI) population model and scat-
tering from more general populations. SDEs for the various scattering processes
are derived from first principles and their statistical and correlation properties
are analysed in detail. This part contains the main body of results of the book,
which constitute a space-time dynamical model for the scattering process. SDEs
Part III – Simulation and experiment describes how the SDEs for the various
scattering processes can be simulated in discrete time, and provides iterative
schemes for reducing the error in this procedure. The theory is substantiated
through experimental examples of scattering at optical and radar wavelengths,
including laser propagation through a random phase screen, and anomaly de-
tection for coherent radar scattering. The treatment is self-contained so as to
be accessible by those familiar with elementary probability theory and electro-
magnetism. The book addresses, for the first time, the fundamental issue of the
space-time dynamics of electromagnetic-scattering processes from random me-
dia. As such, it represents a significant advance in this area of research, and
should prove indispensable to researchers and practitioners in this field at the
postgraduate level and above. The statistical description of electromagnetic scat-
tering from random media is important in a number of contexts. These include,
for example, radar scattering, sea clutter suppression, and the propagation of
laser light through turbulent atmospheric conditions. An understanding of the
statistical properties of such scattering processes enables their characterization
vii
viii PREFACE
in terms of correlation functions and higher order statistics of various spatial and
temporal physical scattering data. Under certain assumptions concerning the na-
ture of the scattering population, it is possible to deduce that the scattered field
intensity consists of two components, a Rayleigh or Gaussian speckle pattern,
modulated by the local scattering cross-section, which fluctuate over two inde-
pendent correlation timescales. This is the so-called compound K-distribution
model, which has been the extant model for sea-surface radar scattering since its
inception over two decades ago. Since then, although a prominent and intensive
area of research, the problem of understanding the dynamics of the scattering
process has remained unsolved. This problem is of vital significance, since a de-
scription of the scattering dynamics (in terms of propagators) implies the form of
all correlation functions and higher order statistics, and provides a valid prior for
the power spectral density.1 An appropriate dynamical model of these scattering
processes has recently been discovered. From very primitive physical assumptions
it has been possible to provide a detailed description of the stochastic dynamics
of the electromagnetic scattering process that captures all the physical degrees of
freedom in the problem. This discovery and its implications for modelling elec-
tromagnetic scattering and propagation phenomena are the basis of the current
monograph.
The major features of the book can be summarized as follows: this pioneer-
ing study describes electromagnetic scattering from a population whose size is
fluctuating in space and time. Its key findings are to characterize the dynami-
cal laws governing the time evolution of the scattered electromagnetic field; the
physical model enables extraction of all spatio-temporal correlation information
and higher order statistics; this enables the results of radar and laser scatter-
ing experiments to be interpreted, and has implications for real time anomaly
detection; this work is significant also because it illustrates how ideas in the
Black–Scholes theory of financial option pricing can be applied to a physical
problem in which non-Gaussian noise processes play an essential role, revealing
scope for cross-fertilization between these disparate areas.
The primary readership for the book is intended as researchers in academia
and industry, in the areas of electromagnetic scattering, SDEs, diffusion mod-
elling, radar systems engineering, signal processing, applications of non-Gaussian
noise processes, quantum optics, laser propagation, population processes, radar
systems engineering, stochastic volatility models in the context of financial math-
ematics, and those with a general interest in applications of stochastic methods
to physics. The book should also benefit advanced graduate students working in
these areas. Practitioners in these fields should find the book a useful resource in
the design and implementation of the various models in both simulation and ex-
perimental data analysis. Lastly, we have attempted to access a broad readership
1 The reader should compare the remarks in Jakeman and Tough (1988), p. 508: “A full
analysis of the temporal correlation properties of the variables x and z implicit in (5.24) would
require knowledge of its fundamental solution or propagator, which is, as yet, unknown. ...”
et seq.
PREFACE ix
by making, to some extent, the various chapters self-contained so that readers

with specific interests should be able to focus on those relevant parts of the text.
An alternative, and well developed, approach to the statistical description of
scattered radiation, is based on a more detailed description of both the scatter-
ing medium and its interaction with the electromagnetic field. Much of this work
has been motivated by problems in microwave and optical sensing and commu-
nication. The Russian contribution to this subject has been significant and is
reviewed in detail in the following references: V.I. Tatarskii ‘Wave propagation
in a Turbulent Medium’, McGraw-Hill, 1961; S.M. Rytov, Yu. A. Kravtsov, and
V.I. Tatarskii, ‘Principles of Statistical Radiophysics’, Vols 1–4, Springer Verlag,
1986; F.G. Bass and I.M. Fuks, ‘Wave Scattering from Statistically Rough Sur-
faces’, Pergamon Press, 1979; B.R. Levin, ‘Fondements Theoretique de la Radio
technique Statistique’, MIR, Moscow Vols 1–2, 1973; Vol. 3, 1979. This approach
has been extended, with particular attention being paid to the non-Gaussian
statistical properties of the scattered radiation, by Jakeman and others in the
UK. The recent book E. Jakeman and K. Ridley ‘Modelling the Fluctuations
in Scattered Waves’ describes these developments in authoritative detail. An
extensive review, which also contains some discussion of the relationship of this
physical model based approach to an analysis based on SDEs, is provided by
E. Jakeman and R.J.A. Tough, ‘Non-Gaussian models for the statistics of scat-
tered waves.’ Adv. Phys. 37, 471, 1988.
The idea for writing the book arose from a suggestion of Simon Haykin,
to whom we express sincere appreciation, following a two-day series of invited
lectures by the author at McMaster University in October 2002.
T. R. F.
June 2007
ACKNOWLEDGEMENTS
We are grateful for academic support from the Departments of Electrical and
Computer Engineering and Mathematics at McMaster University and the Brain
Body Institute. Amongst those colleagues who deserve special thanks are John
Bienenstock, Alexander Bain, Simon Haykin FRSC, Thomas Hurd, Eric Jakeman
FRS, Vikram Krishnamurthy, John McWhirter FRS, Jose Principe, David Sher-
rington FRS, and Kon (Max) Wong FRSC, and my graduate students Patrick
Fayard and Tao (Stephen) Feng. We especially thank Robert Tough for collabo-
ration at the early stages and inspiring much of this scientific development, and
for contributing the majority of Chapter 11. His constant mentoring and encour-
agement has been an essential part of the success of establishing this research
area. We also thank former colleagues Brian Bramson, Richard Glendinning,
Stephen Luttrell, and John O’Loghlen for valuable discussions, and Samantha
Lycett and Kevin Ridley for supplying experimental data studied in Chapter 12.
The author acknowledges the award of a Discovery Grant from the Natural
Sciences and Engineering Research Council of Canada.
x
CONTENTS
I STOCHASTIC CALCULUS
1 Heat equation and Brownian motion 3
1.1 Einstein’s 1905 derivation 3
1.1.1 Green’s function solution 5
1.2 Brownian motion 5
2 Ito calculus 7
2.1 Ito stochastic integral 7
2.1.1 Ito isometry 9
2.2 Ito differential 9
2.2.1 Ito’s formula 10
2.2.2 Ito product rule 10
3 Stochastic differential geometry 12
3.1 Diffusions on manifolds 12
3.2 Kinematics of diffusion 16
4 Examples of stochastic differential equations 20
4.1 Geometric Brownian motion 21
4.2 Bessel process 21
4.3 Solution to Laplace’s equation 23
II DYNAMICS OF THE SCATTERING PROCESS

5 Diffusion models of scattering 27
5.1 Non-Gaussian statistical models 28
5.2 Dynamics of the vector scattering process 30
5.2.1 Natural boundaries 32
5.3 Correlation in the vector scattering process 35
6 Rayleigh scattering 40
6.1 Quadrature components 40
6.1.1 Rayleigh intensity 41
6.1.2 Statistical properties 42
6.2 Random walk model 43
6.2.1 Ornstein–Uhlenbeck process 44
7 Population dynamics 46
7.1 Master equations and the Kramers–Moyal expansion 46
7.2 Birth–death–immigration processes 47
7.3 Continuum diffusion limit 48
xi
xii CONTENTS
8 Dynamics of K-scattering 52
8.1 Stochastic dynamics of K-amplitude process 53
8.1.1 Intensity 55
8.1.2 Phase 57
8.2 Geometry of K-amplitude fluctuations 59
8.3 Asymptotic behaviour 60
8.3.1 Equilibrium distribution 60
8.3.2 Detailed balance 62
8.4 Correlation and spectra 63
8.4.1 Intensity autocorrelation 63
8.4.2 Power spectral density 64
8.5 Interpretation and implications 67
9 Models of weak scattering 69
9.1 Weak scattering amplitudes 70
9.2 Stochastic dynamics 71
9.3 Geometry of amplitude fluctuations 74
9.4 Asymptotic behaviour 79
9.4.1 Detailed balance 85
10 Scattering from general populations 88
10.1 Extended random walk model 88
10.2 Generalized dynamics 91
III SIMULATION AND EXPERIMENT

11 Simulation of K-scattering 99
11.1 Iterative schemes 99
11.2 Rayleigh and gamma processes 105
11.3 Compound K-model 109
11.4 Influence of Doppler on volatilities 111
11.5 Coherent clutter 113
11.5.1 Target returns in clutter 114
11.6 Second-order algorithms 115
12 Experimental tests 124
12.1 Scattering at optical wavelength 125
12.1.1 Random phase screen 125
12.2 Scattering at radar wavelength 126
12.2.1 Coherent sea clutter 127
12.2.2 Anomaly detection 132
13 Non-linear dynamics of sea clutter 138
13.1 Hybrid AM/FM model of sea clutter 140
13.2 Non-linear dynamics from SDE theory 141
13.3 Radar parameters 144
CONTENTS xiii
13.3.1 Superposition 144

13.3.2 Sea state and polarization 145
14 Observability of scattering cross-section 150
14.1 Simulated data 150
14.2 Experimental applications 151
A Stability and infinite divisibility 153
B Ito versus Stratonovich stochastic integrals 155
C Filtrations, conditional probability, and Markov property 158
D Girsanov’s theorem 159
D.0.1 Relation to mathematical finance 160
E Partition function solution to BDI model 161
F Summary of K-scattering 165
F.1 Rayleigh scattering 165
F.2 K-distributed noise 166
F.2.1 Amplitude 167
F.2.2 Intensity 167
F.2.3 Phase 168
F.2.4 Geometry of fluctuations 168
G Iterative solution for vector processes 170
H Open problems 172
I Suggested further reading 174
References 177
Index 181
Part I
Stochastic calculus
The part introduces the concepts and techniques of stochastic calculus that en-
able us to describe dynamical (scattering) processes in terms of stochastic dif-
ferential equations (SDEs). Chapter 1 begins with an historical account of the
heat (diffusion) equation in terms of Einstein’s 1905 derivation and the corre-
sponding Green’s function solution. This leads naturally to the concept of the
Wiener process, which describes the trajectory of an individual particle amongst
a cloud whose density evolves according to the heat equation. Equipped with
the Wiener process as an essential ingredient for the random driving force, in
Chapter 2 we introduce the reader to the basic concepts and techniques of the
Ito (and, to some extent, Stratonovich) stochastic calculus that we shall require
in the description of scattering, and highlight its key aspects that differ from
notions of the classical calculus.
Since we shall be concerned with physical scattering processes that are rep-
resented as vectors, and also in part for mathematical edification, we develop
the subject further from the point of view of stochastic differential geometry in
Chapter 3. The part concludes in Chapter 4 with an account of some specific
examples of SDEs and their applications, which serve to illustrate the various
concepts and techniques of the stochastic calculus introduced in the previous two
chapters.
1
1
HEAT EQUATION AND BROWNIAN MOTION
This chapter explores the interrelationship between the theory of the heat pro-
cess, in terms of the probability density function for a cloud of diffusing particles,
and the underlying stochastic process pertaining to individual particles, which
generates such an ensemble average behaviour that is called Brownian motion.
1.1 Einstein’s 1905 derivation

Of considerable historical interest and significance is Albert Einstein’s famous
original 1905 derivation of the heat equation from considerations of Brownian
motion (Einstein 1905), which we shall recall here. The argument, which is ele-
gant yet mathematically relatively unsophisticated, contains many of the seeds
of modern stochastic calculus.2
We consider a cloud of particles (pollen grains) suspended in a (stationary)
fluid that undergo random ‘thermal’ motion due to the impacts they experience
with molecules in the fluid. Let ρ(x, t) denote the density of the Brownian par-
ticles at location x and time t. Suppose we consider an individual particle and
the displacements ∆ it experiences in some small time τ , and let the probability
density of ∆ be the function φ(∆) which is assumed to have the following two
properties: 1) φ is a symmetric function, 2) φ decays very rapidly for |∆| > ,
0 < 1. The first property arises since a particle without ‘drift’ is equally
likely to be displaced in either direction, while the second is due to the very
small chance that a particle experiences a large displacement in a short time.
Now, the number of particles lying in the interval (x, x + dx), which experience
displacement ∆ is
dn = nφ(∆)d∆. (1.1)
The heat equation concerns the evolution of the density ρ in space and time. To
this end we write
∂ρ
ρ(x, t + τ ) = ρ(x, t) + τ + o(τ ). (1.2)
∂t
On the other hand, we can express the left-hand side as an integral over space,
via the following lucid reasoning. The events that a Brownian particle finds itself
2 The reader may also be interested to compare the (broadly) contemporaneous work of Louis
Bachelier on random walks, in the context of finance, which appears in Bachelier’s doctoral
thesis (orig. Bachelier 1900) and thus predates Einstein’s famous 1905 paper on the Brownian
movement.
3
4 HEAT EQUATION, WIENER PROCESS
at location x at time t + τ is logically equivalent to the event that it is location

x − ∆ at the earlier time t and experiences displacement ∆ in the subsequent
time τ , for all possible choices of ∆. This can be represented mathematically as
the relation
∞
ρ(x, t + τ ) = ρ(x − ∆, t)φ(∆) d∆, (1.3)
−∞
where, to be precise, we have used the following three facts: any equivalent
pair of events, such as those identified above, have equal probabilities; distinct
choices of ∆ yield a set of disjoint events and so the probability of their union is
additive (reflected in integral on the right-hand side above); the displacements
are independent of space-time location, and so the joint probability of these
attributes factorizes. Now, equating the two expressions for ρ(x, t + τ ) above is
the essence of Einstein’s argument. Expanding (1.3) in powers of ∆ we obtain
∞
∂ρ 1 2 ∂ 2 ρ
ρ(x, t + τ ) = ρ−∆ + ∆ + o(∆ 2
) (x,t) φ(∆) d∆. (1.4)
∂x 2 ∂x2
−∞
Terms of o(∆2 ) can be neglected since we assume that φ decays rapidly for
|∆| > , 1 and so ∆n φ ≈ 0 (since ∆n ≈ 0 for |∆| < ), n > 2. Also, ∆φ
vanishes since φ is
assumed to be an even function, and since it is a probability
density function φ = 1. Thus we are left with
∂2ρ
ρ(x, t + τ ) = ρ(x, t) + τ D (1.5)
∂x2
in which we define the diffusion coefficient D as

1
D := lim ∆2 φ. (1.6)
τ →0 2τ
Remark. In this derivation of the diffusion coefficient we can see both the
mathematical and physical origins of the identity for the quadratic variation of
the Wiener process, namely the non-vanishing integral of dWt2 = dt, that is
familiar in the modern Ito calculus.
Now comparing (1.5) with (1.2) and taking the limit as τ → 0 we obtain the
heat equation
∂ρ ∂2ρ
= D 2. (1.7)
∂t ∂x
A generalization of Einstein’s argument that incorporates a drift (correspond-
ing to a displacement density φ that is non-symmetric) and retains all higher
derivatives in space is the Kramers–Moyal expansion (see e.g. Risken 1989) that
we shall encounter later in the context of population models in Section 7.1.
BROWNIAN MOTION 5
1.1.1 Green’s function solution

The fundamental Green’s function solution to this equation, i.e. the solution to
the initial value problem ρ(x, 0) = δ(x) (where δ is a Dirac delta function) is the
familiar Gaussian
2
1 −x
ρ(x, t) = √ exp (1.8)
4πDt 4Dt
which has the variance property
Vart [x] = 2Dt. (1.9)
This corresponds to an individual particle motion following a stochastic process

Xt with stochastic differential equation (SDE):
√
dXt = 2DdWt (1.10)
in which Wt is the familiar Wiener process.

The reader should also compare the Remarks on time symmetry below in
relation to the discussion of the current and osmotic velocities for a pure heat
process.
Finally, we remark that it is intriguing from a historical perspective that the
heat equation exhibits non-causal propagation (as can be seen explicitly through
the Green’s function solution provided in (1.8), observing that ρ is positive for
arbitrarily small times at arbitrarily large distances). It is therefore incompatible
with the advent of Einstein’s special theory of relativity that occurred in the same
year 1905, Einstein’s Annus Mirabilis.
1.2 Brownian motion

A Brownian motion or Wiener process Wt is defined, consistently with Einstein’s
considerations above, as that which satisfies the independent increments property
E [(Wv − Wu )(Wt − Ws )] = 0, (1.11)
where s < t ≤ u < v (so the intervals (s, t), (u, v) are non-overlapping) and such
that the distribution of Wt − Ws is the Gaussian N (0, |t − s|), and is subject to
the initialization W0 = 0.
The autocorrelation function of the Wiener process E [Ws Wt ], s ≤ t, can be
calculated by writing the second factor under the expectation as (Wt − Ws ) + Ws
and then applying the above independent increments property, which leads to
E [Ws Wt ] = s. (1.12)
Therefore, the Wiener process, although it has constant mean zero, fails to be
wide sense stationary.
6 HEAT EQUATION, WIENER PROCESS
This process can also be constructed from the continuum limit of a discrete
random walk, as follows. Consider a discrete random walk in discrete time, start-
ing at the origin, with step size and time step δt. Suppose that the displacement
at each step is drawn from the discrete set {, −} with equal probability. Let
the step size and time step scale relative to each other according to (cf. the dis-
cussion of Einstein’s derivation above) 2 = δt, for all δt > 0. Then, from the
central limit theorem applied to the displacement after n steps in the interval
[0, t], where t is fixed and n → ∞, δt → 0, we find that this is N (0, t) distributed
and, by construction, satisfies the independent increments property. The process
so constructed therefore coincides (in probability) with the above Wiener process
defined in the standard (continuum) manner.
2
ITO CALCULUS
This chapter develops the basic tools of the Ito calculus that flow from the devel-
opment of the Brownian motion or Wiener process in Chapter 1. We emphasize
those particular techniques that are relevant to the detailed analysis of the scat-
tering processes encountered in the chapters that follow.
2.1 Ito stochastic integral

Recall, from the discussion of Brownian motion in the previous chapter, that
a Wiener process Wt is defined as a continuous time stochastic process which
satisfies the following two properties: first, the independent increments property
E [(Wv − Wu )(Wt − Ws )] = 0, (2.1)
where s < t ≤ u < v, so the intervals (s, t) and (u, v) are non-overlapping;
second, the distribution of the increment Wt − Ws is the Gaussian with mean
zero and variance |t − s|, i.e.,
Wt − Ws ∼ N (0, |t − s|); (2.2)
third, the process is defined subject to the initialization W0 = 0.

Equipped with this Wiener process, one is able to extend the ordinary motions
of classical calculus to include a certain family of stochastic integrals, which differ
from the ordinary Riemannian integral in that the integration measure becomes
a random quantity, the increment in the Wiener process. Thus, in place of the
standard Riemannian integral,
t
f (s) ds (2.3)
0
for some integrable function f (s) (e.g. a continuous one), we shall consider re-
placing the measure ds by the stochastic differential dWs . In general, the inte-
grand will be considered to be a stochastic process σs , the so-called (stochastic)
volatility. As we shall see, in the sections that immediately follow, this concept
of stochastic integration leads to corresponding modifications in the classical no-
tion of derivative, that will be quite essential to our development of scattering
dynamics.
7
8 ITO CALCULUS
Fig. 2.1. Ito stochastic integral – the volatility σs is evaluated at the extreme
left point of each subinterval, before multiplying by the Wiener measure dWs .
Now, in light of the above expository remarks, a stochastic integral is defined

as a quantity of the form
b
σs dWs , (2.4)
a
wherein the measure of integration (the increment of the Wiener process) is itself
random. Such an integral differs in nature from the classical Riemann integral,
and is illustrated in Fig. 2.1. It is important (unlike in the case of the classical
Riemannian integral, where it is irrelevant) that in the discrete sum approxi-
mating the Ito integral, the integrand is evaluated at the left-most point of each
subinterval. Thus, in the Ito interpretation,
b j
max
σs dWs = lim σa+jδs (Wa+(j+1)δs − Wa+jδs ), (2.5)

a δs→0
j=0
where jmax = (b − a)/δs − 1. A detailed comparison of the above Ito integral and
the (intimately related) Stratonovich integral, which evaluates the integrand at
the midpoint of each subinterval, is provided in Appendix B.
Remarks. On expectations of stochastic integrals. Owing to the distributive
property of the expectation functional E over addition, the definition (2.5), and
the symmetry of the fluctuations δWt with respect to sign (cf. the associated
symmetry of the displacement distribution function φ(∆) discussed in Section
1.1), the expectation of a stochastic integral is zero, i.e.
ITO DIFFERENTIAL 9

b
E σs dWs = 0. (2.6)
a
In comparison with the above remarks, an important related property of such

integrals, when they arise as products, that we shall require is the following.
2.1.1 Ito isometry

The Ito isometry expresses the expectation of a product of a pair of stochastic
integrals as an ordinary 1-dimensional integral, as follows:
t t t
E φs dWs ψu dWu = E [φs ψs ] ds. (2.7)
0 0 0
Thus, in terms of distribution valued functions the expected product stochas-

tic differential E [dWs dWu ] behaves like a Dirac delta function weighted by the
product time measure, δ(s − u)dsdu. This feature is a reflection of the inde-
pendent increments property of the Wiener process together with the familiar
identity for the quadratic variation dWt2 = dt.
Remarks. On products of stochastic differentials. It is worth remarking further
in this type of context on two other instances of product differential relations for
the Wiener process – this shall avoid confusion and also highlight some of the
care that is needed in manipulating such expressions. For a pair of uncorrelated
Wiener processes W and W̃ it is consistent to write dWs dW̃s = 0 (differentials
with equal time indices) without the need for expectation, since this quantity
equals zero upon integration.3
However, it is not consistent to set the product stochastic differential dξs dξ˜u
(with distinct time indices) equal to zero if ξ, ξ˜ are independent – this fact is
evident from considering the associated double stochastic integral
 t  t 
t
φs ψu dξs dξ˜u =  φs dξs   ψu dξ˜u  (2.8)
0
0 0
which does not vanish.
2.2 Ito differential

According to the prescription for taking the Ito stochastic integral in Section
2.1, we consider a (vector-valued) stochastic (Ito) process Xti with differential
3 Consider
N
the integral as the limit of the discrete sum i=1 nsi ñsi δs where t = N δs and
N → ∞ and n, ñ are N (0, 1) random variables arising in δWs = ns δs1/2 – the limit then yields
the mean of the random variable nñ (scaled by t), which vanishes. If this product
differential is
considered against a continuous (Fs -adapted) process φs then consider lim∆→0 φ(∆) dWs dW̃s
where φ(∆) is an approximation to φ that is piecewise constant over successive sub-intervals of
length ∆, and apply the preceding argument (to each subinterval).
10 ITO CALCULUS
dXti = bi dt + dξti (2.9)
in which ξt satisfies dξti dξtj = σ ij dt. The meaning of this ‘differential’ statement,
which is infinitesimal in nature, can be understood as a mathematical shorthand
for a corresponding integral equation, in which the integrals that arise, where
relevant, are understood in the Ito sense, as explained in Section 2.1.
2.2.1 Ito’s formula
It is often natural and relevant to consider prescribed functions of a stochastic
process such as in (2.9), together with their stochastic differentials. Accordingly,
let us define a new related stochastic process
Ft := f (Xti , t) (2.10)
wherein the function f is known. For example, in the description of random
phasor dynamics in Chapter 6, it is the complex exponential function that is
relevant. Ito’s formula then provides us with a convenient explicit expression for
the stochastic differential of Ft in terms of f and dXt . The key to the result, which
constitutes a departure from the classical rules of calculus, is that the underlying
process Xt has positive quadratic variation, i.e. that its squared fluctuations
integrate to a non-zero amount. Consistently, retaining second-order terms, we
write
∂f ∂f 1 ∂2f
dFt = dt + i
dXti + dXti dXtj + o(dt). (2.11)
∂t ∂X 2 ∂X i ∂X j
Substituting (2.9) into the above we deduce that

∂f ∂f 1 ij ∂ 2 f ∂f
dFt = + bi + σ dt + dξ i (2.12)
∂t ∂X i 2 ∂X i ∂X j ∂X i t
which statement is Ito’s formula.
2.2.2 Ito product rule
The origin of the Ito product rule is similar in essence to the existence of a
quadratic variation that enters into the Ito formula, where now we consider a
pair of processes. Essentially, the result says that we must take into account
product differentials when calculating the differential of a product, and thus the
result constitutes a modification of the classical Leibnitz rule. Explicitly, for the
product of a pair of stochastic processes Ut Vt we write the incremental product
δ(Ut Vt ) = Ut δVt + Vt δUt + δUt δVt + o(δt) (2.13)
which takes the infinitesimal form
d(Ut Vt ) = Ut dVt + Vt dUt + dUt dVt (2.14)
in which it is essential to observe that the third term on the right-hand side
above is non-zero.
ITO DIFFERENTIAL 11
Remarks. Observe that in the product expression dUt dVt = E[dUt dVt ], i.e.
the expectation E is superfluous; in the case that the Wiener processes driving
U and V are independent this product is zero, corresponding to a zero (U, V )
entry in the diffusion tensor σ ij .
3
STOCHASTIC DIFFERENTIAL GEOMETRY
The chapter further develops the theory of (non-degenerate) diffusion processes,

considered as existing on an arbitrary manifold M, from the point of view of
stochastic differential geometry. In this description the diffusion tensor plays an
essential underlying role in supplying the metrical structure of the manifold. We
develop the kinematics of various quantities that arise naturally in this geomet-
rical context, and explain their potential significance in a physical (scattering)
context. This more abstract geometrical interpretation should prove illuminating
for the pure mathematical reader, in terms of our treatment of the dynamics of
the vector scattering processes encountered in subsequent chapters.
Our analysis is carried out using the natural metrical geometry supplied by
the inverse of the (non-degenerate) contravariant diffusion tensor. The notion
of stochastic velocity (and therefore acceleration) is not unique for a stochastic
process, owing to the distinct notions of forward and backward drift velocities
and time derivatives that are conditioned with respect to the past and future.
In this respect the specification of velocity dynamics of a process requires some
care, and we illustrate this concept with the heat process.
The chapter is organized as follows. We begin in Section 3.1 with an ex-
position of the elements of stochastic differential geometry for non-degenerate
diffusion processes on a manifold M, which are necessary for our discussion.
Section 3.2 discusses the kinematics of diffusion and illustrates the distinction
between the various stochastic notions of velocity, providing the heat process as
a concrete example. The relation to the notion of detailed balance in this con-
text is also explained, as it will be required in the description of complex-valued
(vector) scattering processes. We explain the concept of a conserved fluid cur-
rent in a stochastic setting, as distinct from the notion of forward drift in the
Fokker–Planck equation (FPE).
The reader should be familiar with the basic elements of differential geometry,
diffusion theory and the Ito stochastic calculus. Tensorial indices shown in roman
bold type denote the coordinate basis representation, while those in plain italic
are abstract indices that serve to indicate to which space the relevant geometrical
object belongs. We adopt the Einstein summation convention throughout.
3.1 Diffusions on manifolds

We introduce the diffusion process Xti on a manifold M and the associated con-
travariant diffusion tensor σ ij , which in the non-degenerate case has inverse σij
that provides a natural metrical structure on M. The Levi-Civita connection
of this metric is introduced for the purpose of differential geometric operations,
12
DIFFUSIONS ON MANIFOLDS 13
and has the simplifying property that it annihilates the diffusion tensor. In this
geometry it is natural to work with the Ito drift of the process, whose transfor-
mation properties are tensorial with respect to (M, σij ). The situation in this
regard is contrasted with the case of the Kolmogoroff forward drift obtained by
regarding the process as existing on Rn (Karatzas and Shreve 1988). The cur-
rent treatment has the advantage over previous work (cf. Nelson 1967a, 1985)
that each geometrical operation may be expressed in the coordinate free abstract
index notation, as elucidated for example, in Penrose and Rindler (1984).
Consider a continuous time diffusion process Xti on a manifold M of dimen-
sion n taken to satisfy the (time independent) Ito stochastic differential equation
(SDE)

dXti = β i (Xti )dt + i
(Xti )dWt
(j)
σ(j) (3.1)
j
(j)
where {Wt } are a collection of n independent Wiener processes.4 The (con-
travariant) diffusion tensor σ ij is then determined by dXti dXtj = σ ij dt, so that
i j
consistently r σ(r) σ(r) = σ ij . We distinguish between two notions of drift on
the manifold as follows (cf. Nelson 1985). Regarding Xti as a process on a coor-
dinate patch of Rn we have the Kolmogoroff mean forward drift (Karatzas and
Shreve 1988) defined as

i
i Xt+δt − Xti
β = lim Et (3.2)
δt→0 δt
and the corresponding mean backward drift

i i
X t − X t−δt
β̃ i = lim Et (3.3)
δt→0 δt
for δt > 0 in a coordinate chart {xi } that is normal with respect to the Euclidean
metrical geometry δij of Rn (e.g. Nelson 1967b). Denoting the Levi-Civita con-
nection of δij by ∂i the probability density ρE , with respect to the Euclidean
volume measure on Rn , satisfies the FPE (e.g. Risken 1989)
∂ρE ∂ 1
+ i (ρE β i ) = ∂i ∂j (σ ij ρE ) (3.4)
∂t ∂x 2
and the corresponding equation for the backward drift
∂ρE ∂ 1
+ i (ρE β̃ i ) = − ∂i ∂j (σ ij ρE ). (3.5)
∂t ∂x 2
Now, for a non-degenerate diffusion we write σ ij σjk = δki , with σ := det{σij } =
0. In this way the inverse σij supplies the natural metrical geometry of M from
4 The status of the parenthesized index is conceptually distinct from that of a tensor or its
coordinate representation, and the index serves merely to label an individual member amongst
the collection of Wiener processes.
14 STOCHASTIC DIFFERENTIAL GEOMETRY
a diffusion point of view, in that no extra metrical ingredient is required in the

construction of a diffusion on M. In the following sections we shall work mainly
with the Levi-Civita connection ∇i of the diffusion ‘metric’ σij , with associated
components of the covariant derivative given by
∇i V j = ∂i V j + Γijk V k (3.6)
with the Christoffel symbols defined in the standard way as

1 ip
Γijk = σ (∂j σpk + ∂k σpj − ∂p σjk ). (3.7)
2
Thus our chosen ∇i annihilates the diffusion tensor, which shall be a useful
property in the calculations that follow. In the spirit of the σij metrical geometry
it is more convenient to define the Ito forward and backward drifts in the same
way as in (3.2), (3.3), except that {xi } now refers to a coordinate chart that
is normal with respect to the Levi-Civita connection ∇i of the σij metrical
geometry. This drift quantity has the virtue that its components change in a
tensorial way under coordinate transformations. The probability density ρ (for
the same process) with respect to the invariant σ volume measure σ 1/2 dn xi
is related to the previous density, that is tied to the volume measure of the
Euclidean metric, by
√
ρE = ρ σ (3.8)
and satisfies the covariant FPEs

∂ρ 1
+ ∇i (ρbi ) = σ ij ∇i ∇j ρ
∂t 2 (3.9)
∂ρ 1
+ ∇i (ρb̃ ) = − σ ij ∇i ∇j ρ.
i
∂t 2
Adding these two equations yields the equation of continuity familiar from fluid
dynamics (e.g. Batchelor 1967)
∂ρ
+ ∇i (ρv i ) = 0. (3.10)
∂t
Observe that, since the connection ∇ is designed to annihilate σ, the positioning
of the σ ij term on the right-hand side above has no effect, unlike in (3.4), (3.5).
This feature will prove useful in the calculations that follow. Comparison of the
two forms of the FPE yields5
1 √
bi = β i − √ ∂j ( σσ ij ). (3.11)
2 σ
5 The comparison in fact shows that the ρ-weighted left- and right-hand sides of the equation
below have equal divergences. Since this holds for arbitrary ρ independently of b, β, σ the
identity follows.
DIFFUSIONS ON MANIFOLDS 15
This relationship can be re-expressed in√terms of the Christoffel symbol Γ as

follows. From the identity Γkkj = ∂j (log σ) we obtain for the second term on
the right-hand side
1 √
√ ∂j ( σσ ij ) = ∂j σ ij + σ ij Γkkj . (3.12)
σ
We then define the trace of the Christoffel symbol (on its lower indices)
Γi := σ jk Γijk = −∂j σ ij − σ ij Γkkj (3.13)
and thus (3.11) can be written as

1
bi = β i + Γi (3.14)
2
(cf. Nelson 1967a, 1967b, 1985). We deduce thetransformation law for β i from
the Ito equation (3.1) which, on setting W i := j σ(j)
i
dW (j) , implies that

∂xi i 1 ∂ 2 xi
dX i = (β dt + dW i
) + dW i dW j + o(dt). (3.15)
∂xi 2 ∂xi ∂xj
Taking Et of this equation and using the definition (3.2), in general coordinates,
yields

∂xi i 1 ij ∂ 2 xi
βi = β + σ . (3.16)
∂xi 2 ∂xi ∂xj
Recalling the transformation for the components of the Christoffel symbol

Γij k = Pii Pjj Pkk Γijk − Pjj Pkk ∂j Pki , (3.17)

where Pii = ∂xi /∂xi , it follows that

i ∂xi i 2 i
jk ∂ x
Γ = Γ − σ . (3.18)
∂xi ∂xj ∂xk
Hence the non-tensorial terms in the transformations for the components β i
and Γi cancel in (3.14), and therefore the components bi transform vectorially.
Observe that, for a non-degenerate diffusion on M, a sufficient condition for
the Kolmogoroff and Ito drifts to coincide at a given point P is that ∂i σjk (or
equivalently Γijk ) vanish there. The converse does not hold however, since it is
possible for Γi to vanish with ∂i σjk not equivalently zero. By contrast, a similar
type of argument leads to the result that the components of the volatility and
diffusion tensor transform tensorially.
The situation in respect of the abstract index notation can be summarized
as follows (cf. eqns 7.1, 7.2 in Hughston 1996). The components of the abstract
covariant derivative ∇b ξ a of a vector ξ a in the tangent space to T M can be

expressed via
∂ξ a
δbb δaa (∇b ξ a ) = + Γabc ξ c , (3.19)
∂xb
where δaa is a coordinate basis on a patch of M and δaa its dual. The coordinate
stochastic differential (3.1) can be expressed via (3.14) as
1
dxa + Γabc σ bc dt − a
σ(i) dW (i) = ba dt. (3.20)
2 i
This enables us to introduce the abstract Ito differential according to

1
dxa = δaa dxa + Γabc σ bc dt (3.21)
2
and hence (3.1) becomes the abstract geometrical equation6

dxa = ba dt + a
σ(i) dW (i) . (3.22)
i
The vector β i = δii β i then becomes a genuine tensorial object on M, but one
whose definition via (3.2) is tied to the choice of a normal coordinate chart
with respect to the Levi-Civita connection ∂i of the Euclidean metric δij on Rn .
Henceforth in abstract index expressions, indices will be raised and lowered with
σ ij , σij , respectively in accordance with the Einstein summation convention.
This geometrical characterization of the diffusion tensor will be relevant to
the treatment of the amplitude fluctuations for the various scattering processes
we consider in Part III. In expressions that arise hereafter we shall use abstract
index or coordinate basis notation as appropriate to the context and clarity of
exposition, and the summation convention shall be assumed throughout.
3.2 Kinematics of diffusion

We begin by introducing some additional kinematic quantities that will be rel-
evant to our treatment of scattering from a stochastic point of view. We define
the forward and backward generators of the process Xti respectively via

Df (t, x) = limδt→0 Et,x [f (t + δt, X(t + δt)) − f (t, X(t))] /δt
(3.23)
D̃f (t, x) = limδt→0 Et,x [f (t, X(t)) − f (t − δt, X(t − δt))] /δt
for an arbitrary function f (t, x). Since Xt is taken to satisfy (3.1) we deduce
from Ito’s formula applied to f (t, Xt ) that these equations may be expressed in
6 We recommend that the interested mathematical reader consult Hughston (1996) for an
eloquent and thorough exposition of the abstract index notation in the context of stochastic
differential geometry.
KINEMATICS 17
operator form as
∂ 1 ij
D= + bi ∇i + σ ∇ i ∇j ,
∂t 2 (3.24)
∂ 1 ij
D̃ = + b̃i ∇i − σ ∇i ∇j .
∂t 2
We further define the ‘current’ and ‘osmotic’ velocities v i , ui respectively accord-

ing to
1 i
v i := (b + b̃i ), (3.25)
2
1 i
ui := (b − b̃i ). (3.26)
2
To calculate these kinematical quantities for specific dynamical processes we

shall require the osmotic equation (Nelson 1967a, 1985) which relates the osmotic
velocity ui to the probability density via
1 ij
ui = σ ∇j log ρ. (3.27)
2
It is illuminating to study the situation for pure diffusion (the heat equation) in
the context of these kinematical quantities. The pure diffusion process has vanish-
ing (forward) drift, as we have seen from the elements of Einstein’s derivation.
Nevertheless, it has non-zero current, v = 0. This (stochastic) current corre-
sponds to the intuitive notion of a ‘flow’ of diffusing particles which, as we shall
see, is quite distinct from the notion of forward drift. The explicit form of this
current can be derived from the Green’s function solution (1.8) and the osmotic
equation (3.27) in combination with the defining equation for the stochastic cur-
rent (3.25). Since the drift vanishes, the current and osmotic velocities are equal
and opposite. According to the osmotic equation and Green’s function solution,
we find
x
v= , t>0 (3.28)
2t
which is notably independent of the diffusion coefficient D. Observe, from the

above relation, that the signs of v and x coincide. Intuitively this is consistent –
the expected location of a Brownian particle situated at x > 0 at some time
t, at an earlier time, is less than x which spatial asymmetry stems from that
in the prescribed initial condition. In other words, the particle is more likely to
have come from the origin than (positive) infinity. The initial density (an infinite
concentration at the origin) induces an essential spatial asymmetry with respect

to the backward drift (cf. eqn (3.3)).7
Remarks. On time symmetry. Under ‘time’ reversal T :t
→ −t these quantities
transform as

T [bi ] = −b̃i
(3.29)
T [b̃i ] = −bi
and hence

T [ui ] = ui
(3.30)
T [v i ] = −v i .
The reversal of v i under T is in accordance with intuition since this velocity is

identical to the Eulerian velocity field of fluid mechanics (Batchelor 1967), while
the invariance of ui under T is consistent with the osmotic equation (3.27).
Similarly for the forward and backward generators we have
T [D] = −D̃. (3.31)
In addition, we observe that the pair of (covariant) FPEs (3.9) can be obtained
from each other immediately by invoking the time-reversal operation T (since ρ,
∇, σ ij are each invariant under this map, while ∂/∂t reverses sign).
Remarks. On detailed balance. In a general context, there exist two distinct
notions of stochastic equilibrium, namely that of ordinary equilibrium and the
stronger condition of detailed balance. Ordinary ‘equilibrium’ is attained when
the probability density occurring in the FPE is time invariant. If, in addition,
the stochastic current vanishes, then the condition of ‘detailed balance’ is said
to hold. Clearly, from the above discussion, detailed balance implies equilibrium,
but not conversely.
We return to examine such kinematical properties, in the context of detailed

balance for specific scattering processes, in Sections 5.3, 8.3.2, and 9.4.1.
The interested reader may wish further to study the application of this
stochastic differential geometry to the dynamics of quantum mechanical dif-
fusion on general manifolds, as discussed in Field (2003). This paper explores
non-degenerate diffusion processes on an arbitrary manifold, the dynamics of
which arise from a principle of least action for a Lagrangian consisting of a ki-
netic term quadratic in the forward drift of the process and a local potential.
The equation governing the action emerges as a stochastic Hamilton–Jacobi con-
dition and is expressed in terms of the geometry determined by the Levi-Civita
7 Note that, in the definition of the backward drift (3.3), the differentiation is still forwards
in time, the distinction between the forward drift arising from the subtle difference in the
positioning of the time increment with respect to the conditioning.
KINEMATICS 19
connection of the diffusion tensor. It is argued that there are essentially two dy-
namical structures for the rate of change of the drift in the presence of a local
potential, consistent with the requirement of time-reversal symmetry. In both
cases a conserved energy is identified. An alternative wave function and associ-
ated operator description reveal a complex structure in the dynamical equations,
thus extending the earlier results of Edward Nelson on the stochastic treatment
of the Schrödinger equation (Nelson 1967a, 1985).
4
EXAMPLES OF STOCHASTIC DIFFERENTIAL EQUATIONS
We begin the chapter with some general remarks concerning the physical mo-
tivation for the application of stochastic differential equation (SDE) theory to
problems in the mathematical description of scattering from random media. Such
scattering problems are characterized by two essential ingredients. First, the ex-
istence of a random population, whose size fluctuates in time. A measure of the
size of this population (in the continuum limit) is what is referred to as the scat-
tering cross-section and its statistical characteristics are independent of those of
any external electromagnetic radiation that might be used to indirectly probe
the properties of the population. From an experimental point of view, the cross-
section is usually the object of primary interest and various experiments are
designed to infer its behaviour from the properties of electromagnetic radiation
that scatters from or propagates through the random medium.
Second, the electromagnetic component consists of a known transmitted wave
that interacts with the random population, and is subsequently forward- or
backscattered. The random elements that enter here are the fluctuations in mean
power as determined by the cross-section, together with microscopic phase ran-
domization effects that occur for each component of the scattered field that is
essentially an electromagnetic property. The resulting electromagnetic energy
therefore entails two degrees of randomness, inherited from the scattering popu-
lation size and its microscopic details that affect the degree of coherence amongst
the component scattered waves associated with each member of the population.
Thus, we are confronted with a coupled system that has an inherent two-fold
randomness, namely, the scattering population and the associated electromag-
netic radiation. Both of these components of the state of the system evolve
temporally as well exhibiting spatial correlation. Thus, the scattering processes
that we shall consider are fundamentally continuous time-random entities. It is
therefore appropriate, and indeed necessary, to employ a stochastic methodology
that is able to capture the dynamics of such processes in as complete a manner
as possible, beyond eliminating the inherent randomness in the system.
From this point of view, conventional statistical descriptions, though they
may be established and correct, do not represent a complete physical descrip-
tion, capturing only certain ensemble average characteristics, such as distribu-
tion, correlation, power spectra and certain higher order statistics. SDE theory
in partnership with the Ito calculus, on the other hand, provides a rigourous
mathematical framework that pertains to random states of the system that are
separated in time on an infinitesimal scale. The theory is therefore able, in prin-
ciple, to yield all higher order statistical properties of such a physical system.
20
GEOMETRIC BROWNIAN MOTION 21
The identification of the parameters in SDEs that make this possible, in this
monograph, is achieved through both experimental and independent theoretical
reasoning.
The following sections provide some common illustrative examples in SDE
theory, so that the reader can become familiarized with the basic concepts and
techniques that we shall require in the study of scattering processes.
4.1 Geometric Brownian motion

The following example is of particular interest in financial mathematics in mod-
elling the behaviour of a stock price St , and its exact (random) solution illustrates
the basic principles involved in applying Ito’s formula.
We take the increment in St in proportion to the instantaneous value to
behave according to the SDE
dSt
= µdt + σdWt . (4.1)
St
In the absence of the reciprocal term on the left-hand side, the solution entails
a straightforward integration of the right-hand side, thus ST = µT + σWT (for
S0 = 0). For geometric Brownian motion, however, it is necessary to consider
the stochastic differential of the logarithmic process which, via Ito’s formula, has
the SDE
dSt dSt2
d (log St ) = − (4.2)
St 2St2
in which the novice to the Ito calculus should carefully observe the presence of
the second-order term on the right-hand side. Then substituting (4.1) we find

1
d (log St ) = µ − σ 2 dt + σdWt (4.3)
2
which has the merit of a pure differential on the left-hand side and eliminating the
presence of St on the right-hand side. The exact (random) solution is therefore

1
ST = S0 exp µ − σ 2 T + σWT . (4.4)
2

Now, the exponent is Gaussian with mean µ − 12 σ 2 T and variance σ 2 T . Hence
ST has lognormal behaviour, which example is of central importance in the
Black–Scholes theory of option pricing (cf. Appendix D).
4.2 Bessel process

This example is a variant on the Wiener process in n-dimensions, with a natural
geometrical interpretation, and which again illustrates the utility of Ito’s formula.
22 SDE EXAMPLES
In Rn the Bessel process is defined as the Euclideanmetric distance of a n-

(i)
dimensional Wiener process from the origin. Thus, if Wt is a collection of
independent Wiener processes, we define the Bessel process as

n
(i)
Bt :=
(n)
Wt 2 (4.5)
i=1
(n)
for integer n > 1. Our aim is now to calculate the SDE satisfied by Bt , accord-
ing to the rules of Ito calculus. From Ito’s formula we have
n −1/2  n 
1 (i) 2 (j)
(n)
dBt = W  d Wt 2 
2 i=1 t j=1
3 −3/2  n 2
1
n
−
(i)
Wt 2  d Wt 2  .
(j)
(4.6)
2 i=1 j=1
2
Now let us examine the term of the form d W occurring above in isolation.
Via Ito’s formula and the identity dW 2 = dt we can express each term in the
summation as

(j) (j) (j)
d Wt 2 = 2Wt dWt + dt (4.7)
and thus
 
n n
(i)
= 2  Wt 2  dW̃t + ndt
(j) 2
d Wt (4.8)
j=1 i=1
in which we invoke a new (related) Wiener process W̃t (for n > 1).8 Combining
these results we deduce that the Bessel process satisfies the SDE
(n) 1
dBt = (n)
(n − 1)dt + dW̃t (4.9)
2Bt
(n)
for Bt = 0.9
(n)
Remarks. Concerning positivity. Directly from its definition, Bt is every-
where positive (or zero); this is reflected in its SDE by the divergence of the drift
as zero is approached, which property (in spite of the persistence of the Wiener
fluctuating term) ‘repels’ the process from the origin. (The reader should com-
pare the discussion of natural boundaries in Chapter 5.)
8 The way in which the components combine to yield a new Wiener process arises from the
stability of the Gaussian distribution, as discussed in Appendix A.

9 For the trivial case n = 1 using the same techniques we find dB (1) = sgn(W )dW which
t t t
is to be taken for W <, > 0 separately, thus maintaining positivity.
LAPLACE’S EQUATION 23
4.3 Solution to Laplace’s equation

This (rather more abstract) example illustrates the intriguing connection be-
tween stochastic process theory and (elliptic) partial differential equations. We
consider the solution to Laplace’s equation
∇2 φ = 0 (4.10)
on some domain D ⊂ Rn , subject to the boundary condition φ|∂D = f for some

prescribed function f on the boundary ∂D. This boundary value problem can
be solved expediently via SDE theory (simulation), as follows.
We consider a pure Brownian motion Xt whose stochastic differential is sim-
ply that of the Wiener process, dWt , starting at an interior point x ∈ D. Let
this trajectory evolve from the initial time until it (first) reaches the boundary
∂D10 – denote this (random) ‘hitting’ point as Xτ ∈ ∂D so that τ is the ‘hitting
time’. Then consider the expected value of the prescribed function f acquired at
the hitting point, conditioned on the starting point x, i.e.
φ(x) := Ex [f (Xτ )] . (4.11)
This is evidently a function of x and, since the generator of Brownian motion

is the Laplacian (cf. the heat equation), the function φ so constructed satisfies
Laplace’s equation. Also, the boundary condition is manifestly satisfied since for
x ∈ D, we have τ = 0 and Xτ = x (so Ex in (4.11) has no effect).
Similar constructions apply to the solution of Poisson’s equation and other
elliptic partial differential equations in relation to more general SDEs (see e.g.
Oksendal 1998).
The theory of the Ornstein–Uhlenbeck process is of paramount importance in
our treatment of scattering and we shall develop this separately in Chapter 6.
10 The existence of a ‘hitting’ time at which this occurs is ensured by the continuity of the
Brownian path.
Part II
Dynamics of the scattering process
In this part, we develop the detailed stochastic dynamics of the electromagnetic

scattering from a random fluctuating population. We focus largely on the K-
scattering model and its dynamical and statistical properties. Such a description
extends naturally to processes of weak scattering, for which the scattered com-
ponent is weak in comparison with an additional signal component that lies in
coherent superposition.
In Chapter 5, we present a classification of continuous time diffusion pro-
cesses for the local scattering cross-section and scattered intensity for an elec-
tromagnetic field scattering from a random medium, such that the asymptotic
marginal distribution for the intensity is the K-distribution. These processes are
represented as Ito stochastic differential equations (SDEs), which enable iden-
tification of their stochastic volatilities with certain free functions that serve
to calibrate the model. Later, in Chapter 12, we shall use these results in a
study of the volatility behaviour of the electromagnetic intensity scattered from
a random phase screen, which establishes the form of the volatility function to
a high accuracy. This model is then used to derive a SDE for the complex-
valued amplitude, which is then applied to radar scattering from the sea surface
for the purpose of anomaly detection. Chapter 6 develops, from first principles,
the SDE pertinent to the case of Rayleigh scattering, for which the scattering
cross-section is considered to be constant. In Chapter 7 we provide an exposition
of the dynamics of fluctuating discrete populations which, in the limit of large
population size, leads to a continuum population whose dynamics emerges in the
form of an SDE for the scattering cross-section. Using this continuum population
model, in combination with the description of the unit power Rayleigh process, in
Chapter 8 we develop a complete dynamical description of the K-scattering com-
plex amplitude process, its statistical properties and asymptotic behaviour. The
geometrical structure of the infinitesimal fluctuations in the scattered amplitude
are derived and interpreted physically. The effect of a Doppler term is consid-
ered and represented in terms of the amplitude SDE. The chapter concludes
with an iterative (numerical) scheme for integration of the Rayleigh and gamma
population constituents of the K-scattering process.
Chapter 9 develops the corresponding dynamical theory of weak scattering,
which is designed to include the effect of a coherent offset contribution in the
25
26 SCATTERING DYNAMICS
resultant amplitude with respect to which the scattered component is weak.

This is seen to arise in essentially three distinct ways, namely Rice, homodyned
K, and generalized K-scattering, depending on the manner in which the cross-
section fluctuations are incorporated into the resultant amplitude. For each of
these cases, we account for the form of the resultant amplitude in terms of a ran-
dom walk model and provide the SDE dynamics. The corresponding geometry of
the infinitesimal amplitude fluctuations and their interrelationship amongst the
various cases is derived which, in the case of generalized K-scattering, exhibits
a rich geometrical structure that we analyse in detail. The asymptotic statisti-
cal properties and detailed balance are derived in each case. The part concludes
in Chapter 10 with some recent results concerning scattering from general pop-
ulations in which we indicate how to obtain sharp statistical inference of the
scattering cross-section via high-frequency sampling as developed later, in terms
of simulated data, in Chapter 14.
It is worth clarifying some notational issues at this point. We adopt the con-
sistent notation throughout for a continuous time stochastic processes qt , with
Ito differential dqt , We shall consistently adopt the notation for the decompo-
sition of a general Ito process qt into drift and volatility terms, respectively, as
(q) (q) (q)
dqt = bt dt + σt dWt , with respect to some (fixed) probability measure on
(q,p)
the space of paths, and define the diffusion coefficients Σ(·) by dqt dpt = Σt dt
(q)
and abbreviate the squared volatility via dqt2 = Σt dt. (The notation dqt2 , means
that the (stochastic) differential ‘d’ is taken before its square.) The error surface
S q of a vector process q i is then defined by the (inverse) quadratic form relation
Σ(q)−1 (δq) = 1. We shall adopt the Einstein summation convention through-
out (as explained e.g. in Penrose and Rindler 1984), unless explicitly indicated
otherwise.
It will be beneficial for the reader to be familiar with basics of electromagnetic
scattering, diffusion theory, some basic measure theory, and the Ito stochastic
calculus (Oksendal 1998), the latter as far as we have developed in Part I.
5
DIFFUSION MODELS OF SCATTERING
Diffusion models have proved highly successful in their application to a large

number of problems in applied physics, as described e.g. in Risken (1989). In this
chapter we consider the diffusion dynamics underlying electromagnetic scattering
processes from random media, such that the asymptotic distribution for the
scattered intensity is the K-distribution (Jakeman 1980; Jakeman and Tough
1988; Tough 1987; Ward 1981).
Our analysis does not entail the usual separation of a signal from noise. On the
contrary we should think of the pure signal as an inherently stochastic process,
a feature which makes our study of particular interest. Accordingly, our purpose
is to characterize the noise signal by identifying certain parameters in stochastic
differential equations (SDEs) that are used to model the process, namely, the
drift and volatility (diffusion) coefficients. In particular, the stochastic volatility
coefficient is of special significance since it is independent of the probability mea-
sure used to define the underlying Wiener process of the SDE, and is therefore
instantaneously observable without recourse to an ensemble average. Identifi-
cation of the volatility function enables a comparison to be made between the
theoretically predicted and empirically observed volatilities, as discussed in terms
of real experimental data in Part III. In the K-distributed domain these volatil-
ities should correlate very strongly, and thus one is led to a method for anomaly
detection based on where such correlation becomes weak.
The chapter is organized as follows. In Section 5.2 we introduce the electro-
magnetic scattering process as a vector diffusion process in the cross-section and
the scattered intensity, under the assumption that the scattered intensity and
intensity scattered from single scatterer are the K-distribution and Rayleigh dis-
tribution, respectively. A (partial) classification is thus made of vector diffusion
processes such that these marginal distributions are preserved. The volatility
coefficients for the component processes emerge as free functions subject to cer-
tain natural boundary requirements which are derived.11 Given the choice of
volatility function, preservation of the joint distribution implies that the drifts
are accordingly constrained to have a precise form. We complete such classifi-
cation in Section 5.3 via a detailed investigation of the correlation properties
amongst the components of such a vector scattering diffusion model, in terms of
their relationship to the asymptotic distributional properties and the conditions
for stochastic equilibrium and detailed balance.
11 The situation as regards natural boundaries should be compared with the account of the
Bessel process given in Section 4.2.
27
28 DIFFUSION MODELS
We observe that from a physical point of view that only the volatility func-
tions are observable and explain this feature in the context of Girsanov’s theorem
which effects a change of probability measure on the space of all possible paths
of the process. We are thus led to an infinite dimensional family of stochastic
volatility models for the scattering process such that the discriminating features
depend only on the instantaneous observed volatilities and are thus drift inde-
pendent. In this connection, we remark that similar techniques involving change
of measure and stochastic volatility have been studied in the context of financial
mathematics, in particular, in the Black–Scholes theory of option pricing (see
Appendix D).
5.1 Non-Gaussian statistical models

Earlier work has shown that the distribution of the electromagnetic intensity z
for scattering from an object with a fixed (δ-function distribution) scattering
cross-section x, equal to the expected value of z, is given by the negative expo-
nential distribution (or equivalently the Rayleigh distribution for the square root
intensity)

1 −z
Rz (z) = exp (5.1)
z z
(Jakeman and Tough 1988). The usual treatment of the K-distributed intensity
process then makes an assumption that the scattering cross-section x is governed
by a birth–death–migration (BDI) population process such that it is asymp-
totically Γ-distributed and, thereby, one obtains the joint distribution P(x, z).
Integrating out the x variable leads to the K-distribution Kν (z) for the (uncon-
ditioned) intensity
√
2z ν/2 Kν (2 z)
Kν (z) = (5.2)
Γ(ν + 1)
where Kν is a modified Bessel function of the second kind (see e.g. Jeffreys and
Jeffreys 1966).
However, the physical scattering observations relate directly to the scattered
radiation, giving rise to the complex-valued amplitude process and correspond-
ing electromagnetic intensity. On the other hand, no direct measurements of the
scattering cross-section are made. Thus from a physical point of view the be-
haviour of the scattering cross-section should be inferred rather than postulated
in advance. From large time averages of measurements of the scattered radia-
tion, we can deduce that the intensity is, to a close approximation, K-distributed.
The additional assumption of a Rayleigh distributed intensity given a δ-function
distributed cross-section (population size) is a direct consequence of a uniform
distribution in the phases ϕn of the contributions to the scattered electric field
amongst a population of independent identically distributed (i.i.d.) scatterer am-
plitudes (this conclusion follows readily from the central limit theorem applied
to a component of the scattered electric field). More precisely if we decompose
STATISTICAL MODELS 29
the total scattered electric field into the sum of contributions from individual
scatterers, thus

E= an exp(iϕn ), (5.3)
n
then the form factors an are taken to be i.i.d. random variables (with any distri-
bution, not necessarily Gaussian) and the phases to be distributed uniformly (cf.
Jakeman and Tough 1988). These two features of the electromagnetic scattering,
which are physically motivated, then imply the Γ distribution for the scattering
cross-section, as seen from the following result.
Lemma 5.1 For an intensity with P(z|x) and P(z) given respectively by the
Rayleigh distribution Rx (z) and the K-distribution Kν (z), the scattering cross-
section necessarily has a gamma distribution Γν (x). Accordingly the joint distri-
bution P(x, z) is equal to Rx (z)Γν (x).
Proof Suppose,
for some unknown distribution
in the cross-section P(x), that
Kν (z) = Rx (z)P(x)dx, i.e. Kν (z) = [exp(−z/x)/x]P(x)dx for some value of
ν. Since this Kν is known to be generated by Γν , we have

−z q(x)
exp dx ≡ 0, ∀z, (5.4)
x x
where q = P − Γν , so that existence of the above integral is ensured
(observe also
that q = 0). Setting u = 1/x the above integral becomes exp(−uz)Q(u)du,
where Q(u) = −q(1/u)/u, and this integral must vanish for all values of z.
Regarding z as a Laplace transform variable, we see from the Laplace transform
inversion theorem that Q ≡ 0. Thus P = Γν . 2
Observe that the same principle applies to different distributions for the intensity,
e.g. if the modulus amplitude is governed by the Weibull distribution, which is
sometimes used in radar clutter modelling. More precisely, if the intensity has a
given (observed) distribution, and the Rayleigh assumption for the conditional
distribution P(z|x) is preserved, then a corresponding argument leads to the
conclusion that the distribution of the cross-section is uniquely determined.
In summary therefore, as a consequence of the various physical assumptions
and an observed K-distribution for the intensity, the joint distribution for x, z
is given by
xν−1 exp(−x − z/x)
P∞ (x, z) = (5.5)
Γ(ν + 1)
in which ν ≥ −1 is the so-called ‘shape parameter’ as discussed in the radar con-
text in Chapter 13 (equal to α − 1 of Chapter 7). We can now proceed to classify
(vector) diffusion processes in the variables x, z that possess the above joint dis-
tribution. To enable the classification we first observe the following elementary
result (cf. also Wong 1963 for a discussion of sufficient conditions for a diffusion
to have certain asymptotic distributions, governed by Pearson’s equation).
30 DIFFUSION MODELS
Lemma 5.2 Given the Fokker–Planck equation (FPE) ∂P/∂t = −∂x (βP) +
∂x2 (ΣP) the relations between the drift, volatility and asymptotic distribution P∞ ,
if such exists, can be summarized as
x
K β
P∞ = exp (5.6)
Σ Σ
β = Σ∂x log(ΣP) (5.7)
x
k βP
Σ= + . (5.8)
P P
Thus any two of {P, β, Σ} implies the other (modulo k in the case of Σ).12
Proof The existence of an asymptotic distribution P∞ implies that ∂x [−βP∞ +
∂x (ΣP∞ )] ≡ 0. Thus the square-bracketed expression is some function f (t),
which must vanish given the decay of P∞ (x) as x → ∞. Therefore βP∞ =
∂x (ΣP∞ ), whereupon the required expression for β is immediate, and integration
yields those for P∞ , Σ. 2
5.2 Dynamics of the vector scattering process

This section provides a description of the stochastic dynamics of the processes for
the scattering cross-section and scattered intensity, a vector scattering process
(xt , zt ), under the assumption that the intensity for a point scatterer is Rayleigh
distributed and the observed scattered intensity is K-distributed (Ward 1981;
Jakeman and Tough 1988). The point of view, at this stage, is to generate a large
class of dynamics that have appropriate statistical properties. This construction
is non-unique, but identifies a certain parametric sub-class of models – thus it is
appropriate to situations where one desires to simulate data that is known to pos-
sess certain statistical characteristics but, perhaps, where one is less concerned
with having a detailed physical understanding of the origins of these dynamics.
A more principled and constructive approach, that isolates a specific dynamics
uniquely from this class, can be taken essentially from first principles. Such an
approach, in contrast, derives statistical properties as emergent consequences of
the detailed physical model, and will be provided to enhance and complement
the current treatment later in Chapter 8.
We begin by exploiting the results of Lemma 5.2 from which we are able
to deduce, for P∞ (x, z) prescribed by (5.5), a general form of the FPE for the
time-dependent joint distribution P(x, z, t). For the moment, for simplicity, and
also to make contact with the work of previous authors (Jakeman and Tough
1988; Tough 1987), we shall neglect the correlation between x and z – later in
Section 5.3 we shall develop the classification theory in a more general setting
that includes such correlation which in turn corresponds to the construction of
K-scattering dynamics provided in Chapter 8.
12 The quantities k and K are constants of integration, equal to each other if the lower limits
of integration in the above system are chosen to coincide.

SCATTERING DYNAMICS 31
Proposition 5.3 The FPE for K-scattering processes takes the parametric form
2
∂P ∂ ∂ 1−ν z ∂φ
=A (φP) + − 2 +1 φ− P
∂t ∂x2 ∂x x x ∂x
2
∂ ∂ ψ ∂ψ
+B (ψP) + − P (5.9)
∂z 2 ∂z x ∂z
for squared volatility functions φ(x, z, t), ψ(x, z, t). The corresponding non-linearly
coupled SDEs for the vector diffusion process (xt , zt ) are

ν−1 z ∂φ (1)
dxt = A + 2 −1 φ+ dt + (2Aφ)1/2 dWt, (5.10)
x x ∂x

∂ψ ψ (2)
dzt = B − dt + (2Bψ)1/2 dWt , (5.11)
∂z x
(1) (2)
where Wt , Wt are independent Wiener processes.
Proof For equilibrium, set the left-hand side of (5.9) equal to zero. Since A, B
are independent constants we require that the expressions in square brackets on
the right-hand side vanish separately. The form of the joint distribution (5.5)
together with (5.7) imply the forms taken by the drift coefficients in (5.9), given
the squared volatility free functions φ, ψ. 2
The reciprocals of the (characteristic frequency) constants A, B represent inde-
pendent correlation timescales for the respective scattering processes xt , zt . In
tensor notation (Risken 1989) we introduce the operator L̂ = −∂i β i + 12 ∂i ∂j σ ij
in terms of which the FPE can be expressed as ∂ρ/∂t = L̂ρ. The detailed balance
condition, which is stronger than the stationarity (i.e. equilibrium) condition,
requires that each individual transition in the master equation is perfectly bal-
anced. This condition can be expressed as the operator equation L̂ρ = ρL̂† acting
on all functions f (Risken 1989), whereby the choice f = 1 implies stationar-
ity. Alternatively, in terms of the ‘current’ v i = β i − 12 ρ−1 ∂j (σ ij ρ) the FPE
can be written ∂ρ/∂t + ∂i (ρv i ) = 0 and the detailed balance condition becomes
v i = 0. Since the constants A, B in (5.9) are independent and arbitrary, the
detailed balance condition is satisfied for any choice of volatility functions φ, ψ.
In typical physical situations, such as those we shall describe in Chapter 12, zt
de-correlates much more rapidly than its companion xt , so that A B. The
frequency A is inherent to the scattering population and is independent of the
incident electromagnetic field; in contrast, on dimensional grounds B is propor-
tional to the wave-number of the illuminating radiation and thus B ∼ c|k|.13
The (reciprocals of the) constants A, B can be determined experimentally from
the scattered electromagnetic intensity time series {zt } by measuring the period
between successive peaks over the two characteristic long and short time scales,
13 The ∼ symbol here indicates an (approximate) scaling relation.
32 DIFFUSION MODELS
respectively (see e.g. §6 in Jakeman and Tough 1988); cf. also §IIC in Field and
Tough (2003b) for a detailed theoretical account of the observability of these
characteristic frequency constants.
5.2.1 Natural boundaries

The squared volatility coefficients φ(x, z, t), ψ(x, z, t) are free functions subject
to the natural boundary condition, as follows. Since the physical quantities x, z
are positive, we require the set {x = 0} ∪ {z = 0} for x, z > 0 to bound the
vector process (xt , zt ). Thus the volatilities of the component processes must
vanish at the origin and the drifts must tend to positive (possibly infinite) val-
ues. Such a requirement ensures that (xt , zt ) remains in the positive quadrant
of the (x, z) plane without imposing additional boundary conditions. Thus the
positivity property becomes an inherent feature of the stochastic dynamics and
is correspondingly easier to handle. We shall assume that both functions φ, ψ
are twice differentiable in x, z, respectively, and that these derivatives remain
bounded in some closed interval [0, x0 ], x0 > 0 (but are not necessarily con-
tinuous there). This requirement is natural given the presence of the second
derivatives in the FPE for the joint distribution (5.9). Since the natural bound-
ary condition requires φ, ψ → 0 as x, z → 0 respectively, we may write, for
sufficiently small x > 014
φ(x) = xφ(1) (0) + Rφ (x), (5.12)

φ (1)
(x) = φ (1)
(0) + Rφ (x), (5.13)
and likewise for ψ(z) with remainder term Rψ (z). The Taylor theorem
x with
remainder (see e.g. Jeffreys and Jeffreys 1966) states that Rφ (x) = 0 (x − u)
φ(2) (u)du and that this remainder satisfies the inequalities
1 2 1
x inf [φ(2) ] ≤ Rφ (x) ≤ x2 sup[φ(2) ] (5.14)
2 [0,x] 2 [0,x]
x inf [φ(2) ] ≤ Rφ (x) ≤ x sup[φ(2) ] (5.15)
[0,x] [0,x]
with corresponding inequalities for ψ(z). Thus we deduce the limiting behaviour
of the drifts b(x) , b(z) in the respective component SDEs (5.10), (5.11). We find

(x) (1) Rφ (1) 1
lim [b ] = νφ (0) + z lim + zφ (0) lim . (5.16)
x→0 x→0 x2 x→0 x
The second term on the right-hand side is bounded by virtue of (5.14). Thus
necessary and sufficient conditions for the natural boundary requirement
14 We suppress the (z, t) and (x, t) arguments in the sets of functions {φ, R }, {ψ, R }
φ ψ
respectively for notational clarity.
SCATTERING DYNAMICS 33
x
0 < limx→0 b(x) ≤ ∞ are φ(1) > 0, or else φ(1) (0) = 0 with 0 (x − u)φ(2) (u)du >
0, and sufficient that inf [0,x] [φ(2) ] > 0. Correspondingly for b(z) , we find
lim [b(z) ] = ψ (1) (0) (5.17)

x→0
so that a necessary and sufficient condition on the component drift for the natural
boundary to exist with respect to z is 0 < ψ (1) (0) ≤ ∞.
As a special case of (5.9) (Jakeman and Tough 1988; Tough 1987) when
φ(x, z, t) = x and ψ(x, z, t) = z we find
2
∂P ∂ ∂ z
=A (xP) + x − ν − P
∂t ∂x2 ∂x x
2
∂ ∂ z
+B (zP) + − 1 P , (5.18)
∂z 2 ∂z x
where P = P(x, z, t). Observe that the natural boundary condition is satisfied
for this choice of volatility functions. The associated non-linearly coupled SDEs
(5.10), (5.11) for the vector diffusion process (xt , zt ) become

ν−x+z (1)
dxt = A dt + (2Ax)1/2 dWt (5.19)
x

1−z (2)
dzt = B dt + (2Bz)1/2 dWt . (5.20)
x
This special case can be motivated by the corresponding situation for the Rayleigh
(Gaussian speckle) process as described in Section 6.1.
In respect of classification of electromagnetic scattering processes, the novelty
of the present approach rests on the following key observation.
Lemma 5.4 The volatility functions φ, ψ are instantaneously observable quan-

tities without recourse to an ensemble average.
This result, for infinitesimal dt, follows from the squares of the Ito differen-
tials (5.10) and (5.11). The implementation of the result for discrete sampled
(experimental) data is explained later in Chapter 12. According to the Ito ma-
nipulation rules dWt2 = dt, dW dt = dt2 = 0, these are
dx2t = 2Aφ dt, (5.21)

dzt2 = 2Bψ dt. (5.22)
The argument for the local observability of the volatility functions can also be
understood from the point of view of Girsanov’s theorem (Oksendal 1998 and
Appendix D) which states that, via a change of measure on the space of paths,
the process dW̃t = dWt + γt dt can be regarded as a pure Brownian motion. More
34 DIFFUSION MODELS
precisely, if P, Q are probability measures on the space of paths and if Wt is a

P-Wiener process, then W̃t is a Q-Wiener process, where the change of measure
process Mt (i.e. the Radon–Nikodym derivative dQ/dP) is given explicitly by
the stochastic integral15
t
1 t 2
Mt = exp − γs dWs − γs ds . (5.23)
0 2 0
Observe from Ito’s formula that dMt = −Mt γt dWt so that Mt is manifestly a
martingale with respect to the original measure P, i.e. EPs Mt = Ms for s ≤ t.
In respect of our electromagnetic scattering processes we may re-express (5.10)
and (5.11) in the Q(i) ‘drift-neutral’ measures as
(1)
dxt = (2Aφ)1/2 dW̃t , (5.24)
(2)
dzt = (2Bψ)1/2 dW̃t . (5.25)
Remarks. On the instantaneous observability of volatility. It is of considerable

interest how our methodology resembles the Black–Scholes theory of option pric-
ing (see e.g. Ch. 12 in Oksendal 1998) in financial mathematics – via the change
of measure argument, effected via Girsanov’s theorem, the stochastic volatili-
ties are the only observable parameters of the SDEs for an individual sample
path of the stochastic process. In accordance with this observation, in the finan-
cial context, via the introduction of a ‘risk-neutral’ measure, the Black–Scholes
no-arbitrage option price is independent of the drift in the underlying stock
price. In our physical application, the observability argument leads to a crite-
rion for instantaneous anomaly detection for which the drifts in the underlying
electromagnetic scattering processes are of no relevance.
The significance of the volatility over the drift in respect of observability
represents a fundamental shift of emphasis in the approach to the problem
of anomaly detection in the context of electromagnetic scattering. Previous
approaches have used temporal correlations for the intensity process, i.e. quan-
tities such as
ztα11 ztα22 . . . ztαnn (5.26)
which is a moment of order α1 + · · · + αn . These moments, while they may

provide discriminating measures of data properties, have the shortcoming that
they depend on the probability measure P and the corresponding diffusion drift.
In contrast, the volatility analysis is independent of P and requires no ensemble
average ·.
t
15 We require γt to satisfy Novikov’s condition, namely that exp( 12 0 γs2 ds) < ∞.
CORRELATED VECTOR SCATTERING 35
5.3 Correlation in the vector scattering process

In the foregoing treatment we have effectively neglected the possibility of correla-
(x) (z)
tion between the noise processes Wt and Wt . This assumption is represented
mathematically by the diffusion tensor diagonal (x, z) matrix representation

2Aφ 0
Σ(·,·) = . (5.27)
0 2Bψ
Now, this special case can be extended to include explicit correlation between
(x) (z)
Wt and Wt . We shall represent this mathematically by a non-zero off diagonal
term in the diffusion matrix Σ(x,z) = 2Cχ for some correlation function χ =
χ(x, z, t), where C is an (independent) constant with the dimensions of frequency
whose reciprocal represents the cross-correlation timescale.16 Thus, in the (x, z)
representation we shall have

(·,·) 2Aφ 2Cχ
Σ = . (5.28)
2Cχ 2Bψ
In accordance with the theory of the K-distribution, as developed in the context
of scattering above, we maintain the requirement that the joint x, z asymptotic
distribution be preserved, which we recall is equal to
xν−1 exp(−x − z/x)
P∞ (x, z) = . (5.29)
Γ(ν + 1)
We recall also the joint FPE
∂P 1
= −∂i (P bi ) + ∂i ∂j (P Σij ) (5.30)
∂t 2
which must have (5.29) as an asymptotic solution, obtained by setting the left-
hand side of (5.30) equal to zero.
The joint FPE can alternatively be expressed in terms of the ‘current’ v i as
∂P
+ ∂i (P v i ) = 0 (5.31)
∂t
in which v i := bi − 12 P −1 ∂j (P Σij ). This is the equation of continuity (conser-
vation of probability, or physically mass if P represents the concentration of a
cloud of massive particles).
Remarks. On the relation with fluid dynamics. The vector v i is the current
familiar from fluid mechanics. Observe that, even in the presence of stochastic be-
haviour (in the sense of a non-zero Wiener term in the SDE dynamics), the right-
hand side of the equation of continuity is zero. Thus, fluid dynamics should not
16 In fact, as we shall see in Chapter 8, from the point of view of a random walk model this
cross-correlation has a timescale inherited from the (fluctuations in the) scattering cross-section
and thus depends on the constant A.
36 DIFFUSION MODELS
be viewed as an approximation to a stochastic dynamics and associated FPE –

rather, it is an exact description for the fluid current v, as distinct from the
forward drift b, in a stochastic context.
The relationship between the equilibrium and detailed balance conditions

in one dimension can be established as follows (assuming rapid enough decay
in P (xi , t) at infinity in x). A steady state (∂P∞ /∂t = 0) implies that P v i is
equal to some function f (t) of time only, by virtue of (5.31), whereupon letting
|x| → ∞ and using the decay properties of P we must have f = 0. It follows
that the current vanishes, and thus detailed balance is equivalent to equilibrium.
The corresponding situation in higher dimensions is more subtle – we develop
this below.
The vanishing of the time derivative of the joint distribution that is a condi-
tion for equilibrium implies, via (5.31), that the vector P v i is solenoidal, i.e. it
has vanishing divergence. In accordance with a basic theorem from vector calcu-
lus, it can therefore be expressed in terms of a vector potential A as P v = ∇×A
which in our case shall be interpreted in terms of the embedding of v (an in-
herently two-dimensional vector) into Euclidean 3-space, and the potential A
necessarily exists in the three-dimensional space.
Before engaging in the detailed mathematics, we begin with some informal
remarks concerning detailed balance in respect of the (xt , zt ) vector scattering
process. Consider a visual representation of this process, in the x, z positive
quadrant, as an abstract fluid flow in case v = 0 whose trajectories (streamlines)
have abstract ‘velocity’ v which is prescribed as the current of the stochastic
process (xt , zt ) according to the ideas of Section 3.2. This constitutes a deter-
ministic mapping of points τ : Qt
→ Qt+τ where Qt has coordinates (x, z) for
some fixed initial point, and arises through (5.31) from ensemble average prop-
erties of the vector scattering process. Such a flow could occur with fluid density
constant in time, i.e. equilibrium, and yet with a non-zero fluid current. Now we
ask why, intuitively/physically in terms of this abstract fluid, should the (xt , zt )
process asymptotically satisfy detailed balance v = 0. Now, this non-vanishing
flow can be interpreted as a type of non-stationarity , in the precise sense that it
yields canonical inequivalent times as pertains to a fluid particle immersed in the
current flow (in the asymptotic time domain). This flow becomes an emergent
feature of the ensemble average properties of the vector scattering process – its
vanishing (detailed balance) is to be regarded as a notion of strict equilibrium.
In what follows, we shall investigate both the situations where detailed balance
is imposed, which we shall regard as a physical constraint, and also for complete-
ness explore what happens if we relax this assumption.
So, suppose that we appeal to our physical argument above and assume
strict equilibrium, i.e. detailed balance, motivated in part by our intuition and
also by the fact, as we shall see later, that detailed balance is a consequence of
the random walk scattering model. Then, we must solve the equation v i = 0,
i.e. P bi = 12 ∂j (P Σij ). Accordingly, we seek a general solution to the coupled
system
 (x)
Pb = 1
2 [∂x (P Σxx ) + ∂z (P Σxz )] ,
(5.32)

P b(z) = 1
2 [∂x (P Σzx ) + ∂z (P Σzz )] .
Before solving this system let us recall briefly the 1-dimensional case. Consider
that we are prescribed the asymptotic distribution P and we find the drift for
an arbitrary volatility function. Then the system above reduces simply to the
scalar relation b = 12 Σ∂x [log(ΣP )] – thus knowledge of P uniquely establishes
the relationship between the drift and volatility (diffusion) coefficients (cf. Wong
1963). In the two-dimensional case of the vector scattering process (x, z) that we
are concerned with here, the asymptotic (joint) distribution is indeed prescribed
according to (5.29) and thus we follow a similar approach to the one-dimensional
case. First, let us recall some simple identities for the partial derivatives of the
joint distribution with respect to x, z:

 ∂P z ν−1

 = P −1 + + ,
 ∂x x2 x
(5.33)



 ∂P = − P .
∂z x
From (5.32) this leads us to an expression for the x-drift as

z ν−1 −1
2b(x) = ∂x Σ(x,x) + Σ(x,x) −1 + 2 + + ∂z Σ(x,z) + Σ(x,z) .
x x x
(5.34)
and similarly for the z-drift

z ν−1 −1
2b(z)
= ∂x Σ (x,z)
+Σ(z,x)
−1 + 2 + + ∂z Σ (z,z)
+Σ(z,z)
.
x x x
(5.35)
Thus, analogous to the one-dimensional case, we have an explicit mapping from

Σ(·,·) to b(·) provided by the joint distribution P . We remark that, although our
interest here is confined to the vector scattering case (5.29), the methods we have
developed would apply equally well to any abstract joint distribution with suit-
able decay properties, and may therefore also be of pure mathematical interest
in their own right.
Given this explicit relationship between the drift and volatility functions of
the vector scattering process, which incorporates the presence of cross-correlation,
we can verify the generalized dynamics for a special K-scattering model, derived
from the physical considerations of a random walk model. For this we take the
38 DIFFUSION MODELS
diffusion tensor to have the special form17 :

ij Ax % Az
Σ =2 . (5.36)
Az Az 2 x + Bxz
Observe, as remarked above, that the off-diagonal cross-correlation term scales

with the constant A, which originates from the fluctuations in the scattering
cross-section process. Note however that the cross-correlation function χ coin-
cides with the intensity variable. Substituting the matrix expression for Σ(x,z)
into (5.34), and upon cancellation of various terms, we obtain for the x-drift
b(x) = A(ν + 1 − x). (5.37)
Similarly for the z-drift a straightforward calculation using (5.35) and arranging
terms according to A and B dependence, yields

z(ν + 1 − x)
b(z) = A + B(x − z) (5.38)
x
in accordance with the analysis of the random walk model as developed in Chap-
ter 8. This completes our discussion of the vector K-scattering process in the
case that detailed balance is imposed.
Now, we turn to consider the case of classifying K-scattering models if there
is ‘weak’ equilibrium, i.e. an asymptotic limiting distribution exists but the is
a non-zero current for the (xt , zt ) as described in terms of the abstract fluid
flow above. We consider the density-weighted immersion of the two-dimensional
flow determined by the vector field v into Euclidean 3-space, co-ordinatized by
(X, Y, Z),18 thus
v → V = P v̆ (5.39)
in which we define v̆ := (vx , vz , 0). The (weak) equilibrium condition can then
be expressed (locally) as
V =∇×A (5.40)
and the vector potential A is necessarily a three-dimensional object. From the

geometry of the immersion we may assume ∂/∂Z ≡ 0, whereas general the
vector potential (which is subject to a gauge freedom, as identified below) has
all components non-vanishing. Accordingly we have

VX = ∂Y AZ
(5.41)
VY = −∂X AZ
17 Orig. eqn (3.2) in Field and Tough (2003b); note, for comparison, set α − 1 in the latter
reference equal to ν herein.

18 Upper case is used here to distinguish the Euclidean coordinates from the vector scattering
process (x, z).

and since VZ vanishes
∂X AY = ∂Y AX . (5.42)
Thus, without loss of generality, we may set A = (0, 0, A) for some scalar po-
tential function A ≡ A(X, Y ) and hence V = (∂Y , −∂X , 0)A. Observe here that
A has the status of an arbitrary scalar function (with suitable decay at ‘spatial’
infinity). This enables us to construct a general expression for the forward drift
of the process as
P b = P bD.B. + (∂Y , −∂X , 0)A (5.43)
consisting on the right-hand side of a detailed balance part and current contri-
bution respectively, wherein the former is obtained by setting A = 0. In respect
of the latter, we identify the current in terms of a gauge freedom in choice of this
scalar function. Observe that the relation P bD.B. = 12 ∂· (Σ· · P ) is unaffected by
the presence of a gauge term involving A. We therefore summarize the scheme for
classifying the dynamics of asymptotic equilibrium as follows: we begin with a
prescribed asymptotic joint distribution (e.g. that appropriate to K-scattering),
choose the diffusion coefficients Σ(·,·) and gauge potential A as free (indepen-
dent) functions, and then construct the drift according to (5.43) – this provides
a complete classification of the diffusion dynamics, incorporating both detailed
balance and current contributions. Functionally, we can summarize this situation
by the statement
b = bD.B. [P, Σ] + P −1 (∂Y , −∂X , 0)A. (5.44)
In Chapter 8, we shall then see how specific choice of the functions φ, ψ, χ occur-
ring in the (x, z) representation of Σ(·,·) may be obtained, from first principles,
via considerations of a random (phasor) walk model. It emerges that detailed
balance is satisfied for such a model and so we may assume A = 0 in (5.44).
Thus, a physical – as opposed to parametric data driven – model emerges as
a special case of the general structure above. The latter is constructed merely
to preserve the joint (x, z) asymptotic distribution, while the former is a spe-
cific instance of this with the essential property that, in addition, it specifies all
correlation information and higher order statistics.
6
RAYLEIGH SCATTERING
In this chapter we develop the stochastic dynamical theory of Rayleigh scattering,

from the point of view of a random walk model. Physically this is appropriate to
forward and backward scattering experiments where the scattering population is
effectively constant over the observation period. From a theoretical point of view,
the purpose of the chapter is to present a self-contained derivation that follows
essentially from first principles. We shall see how the development, according to
a random walk model pertaining to the components of the scattered field, leads
to an elegant mathematical derivation of the Ornstein–Uhlenbeck process for the
resultant amplitude. This derivation of the scattering dynamics, motivated by
a physical model (the coherent superposition of a number of randomly phased
scattered wave components) is of inherent pure mathematical interest, as well as
being satisfying from a theoretical physics point of view.
The chapter is organized as follows. We begin in Section 6.1 with a basic ac-
count of the dynamics of the quadrature components of the Rayleigh scattering
process, from the point of view of constructing these dynamics from its key statis-
tical properties, such as certain correlation functions and higher order statistics.
Given such (limited) statistical characteristics one can consistently posit a set
of Ornstein–Uhlenbeck SDEs for the quadrature amplitude components, such
that the appropriate statistics emerge. This approach is made physically and
mathematically complete, in Section 6.2 by considering the inverse construction,
namely that of deriving statistical characteristics from first principled dynamical
ones. As opposed to merely constructing a (non-unique) dynamics from desired
statistics, we posit a random walk (phase diffusion) model for the scattered am-
plitude. It is shown, from this point of view, in the Rayleigh case of a fixed step
number, that the amplitude necessarily obeys a complex Ornstein–Uhlenbeck
(COU) equation. A corresponding SDE in the K-scattering case, incorporating
step number fluctuations, shall be derived later in Chapter 8.
6.1 Quadrature components

It is familiar that the complex-valued amplitude or ‘received envelope’ can be
expressed as the sum of its quadrature components according to
Ψt = It + iQt (6.1)
so that I, Q denote the ‘in-phase’ and ‘quadrature-phase’ components of the

coherent amplitude (e.g. radar signal), respectively (e.g. in Helstrom 1960). The
40
QUADRATURE COMPONENTS 41
intensity zt can then be expressed in terms of these quantities according to

zt = |Ψt |2 = It2 + Q2t . (6.2)
Now, in the Rayleigh case, it is consistent from a statistical point of view to
describe It , Qt as a pair of Ornstein–Uhlenbeck processes (see e.g. Oksendal
1998) with SDEs
√ (I)
dIt = −αIt dt + 2dWt (6.3)
√ (Q)
dQt = −αQt dt + 2dWt . (6.4)
The Fokker–Planck equation (FPE) associated with either of these separately is
∂P ∂2P ∂(IP)
= 2
+α (6.5)
∂t ∂I ∂I
whose stationary solution is
α 1/2
−αI 2
P(I) = exp . (6.6)
2π 2
This shows that It is an asymptotically Gaussian-distributed random variable
with mean zero and variance I 2 = 1/α and thus, asymptotically, we have
z = 2/α.
6.1.1 Rayleigh intensity
To calculate dzt we require Ito’s formula (see e.g. Oksendal 1998 and Chapter 2)
which states that, for an Ito process Xt satisfying dXt = µt dt + σt dWt (where
µt , σt are general stochastic processes, not necessarily
Ito processes), the2 process

Ft = f (Xt , t) has the stochastic differential dFt = ∂f ∂t + µt
∂f
∂Xt + 1 2 ∂ f
σ
2 t ∂X 2 dt +
t
∂f
σt ∂X t
dWt . Hence the SDE satisfied by zt is derived according to
dzt = 2It dIt + 2Qt dQt + dIt2 + dQ2t

√ (I)
√ (Q)
= [4 − 2α(It2 + Q2t )]dt + 2 2It dWt + 2 2Qt dWt . (6.7)
Since the sum of two independent Gaussian random variables is itself a Gaussian
random variable (see Appendix A), whose variance is the sum of the variances
of its constituent parts, we write
√ (I)
√ (Q) √ (z)
2 2It dWt + 2 2Qt dWt = 2 2zt dWt . (6.8)
The SDE for zt therefore becomes
√ (z)
dzt = (4 − 2αzt )dt + 2 2zt dWt (6.9)
whose associated FPE is
∂P ∂ 2 (4zP) ∂((4 − 2αz)P)
= − . (6.10)
∂t ∂z 2 ∂z
42 RAYLEIGH SCATTERING
6.1.2 Statistical properties

The stationary solution of this equation is

α −αz
P(z) = exp (6.11)
2 2
and thus the mean value of z is what we expect from (6.6). In this way the
analysis of the component I, Q processes using the Ornstein–Uhlenbeck process
gives insight into the origin of the square root volatility for the intensity zt . It is
interesting to note that the FPE describing the Rayleigh process also encodes the
factorization properties of the intensity distribution, which result from its being
the sum of squares of two Gaussian processes, each with its own characteristic
factorization properties. If we change variables to z = αz/2, t = 2αt, the FPE
(6.10) becomes
∂P ∂ 2 (zP) ∂((z − 1)P)

= + . (6.12)
∂t ∂z 2 ∂z
As discussed in detail in Jakeman and Tough (1988), the Green’s function of this
FPE can be expanded in terms of Laguerre polynomials which, in turn, encode
the combinatorial factors implicit in the factorization properties of the Gaussian
random variables It , Qt . Thus, for example, we find the exponential decay of the
two-time correlation function according to

min(m,n)
n!m!
z(t)n z(0)m = n!m! exp(−rt). (6.13)
r=0
(n − r)!(m − r)!(r!)2
In much the same way, it can be shown that the expansion of the Green’s func-
tion of the FPE (6.5) in terms of Hermite polynomials captures the factorization
properties of the constituent Ornstein–Uhlenbeck processes. In each case the lin-
earity of the deterministic part of the underlying SDE establishes the exponential
decay in the correlation function characteristic of a Markov process.
The constant volatility in the Ornstein–Uhlenbeck case does not impose a
natural boundary at the origin, so that the process can take positive and nega-
tive values. Conversely, the square root volatility emerging in the Rayleigh case
establishes a natural boundary that maintains the positivity of zt . Observe that
the forms taken by the SDEs (6.3) and (6.4) inherently capture these fundamen-
tal properties of the physical processes they describe. The special cases (5.18),
(5.19) and (5.20) will be significant in the derivation of K-scattering dynamics
from a random walk model in Chapter 8 and in the experimental analysis of
Chapter 12.
Remarks. On random motion in a potential. The situation of stochastic volatil-
ity discussed here should be compared with the more familiar case of constant
RANDOM WALK 43
diffusion coefficient and the description of a particle at location xt in a potential

V (x) subject to a random force. The motion is governed by the SDE

∂V
dxt = −α dt + (2D)1/2 dWt (6.14)
∂x
and the particle location has the equilibrium Boltzmann distribution

−αV
P∞ (x) ∝ exp . (6.15)
D
6.2 Random walk model

We develop the random walk model with step number fluctuations due to Jake-
man (orig. Jakeman 1980; cf. Jakeman and Tough 1988) as a continuous time
diffusion process.
In the Rayleigh case consider the random walk model for the scattered electric
field (cf. Jakeman 1980; Tough 1987; Jakeman and Tough 1988)
N
& '
(N ) (j)
Et = exp iϕt (6.16)
j=1
for constant population size N . Since Maxwell’s equations for the electromag-
netic field possess U (1) gauge invariance with respect to duality rotations, i.e.
multiplication by exp(iΛ) for (constant
) Λ (cf. Penrose and Rindler 1984), the as-
sumption of independence of ϕ(j) implies that these
&phases'are uniformly dis-
(j)
tributed. Accordingly, in (6.16) the phase factors exp iϕt are independent
and uniformly distributed on the unitcircle in the complex plane C. Our (phase)
(j)
diffusion model therefore takes ϕt as a collection of (displaced) Wiener pro-
cesses evolving on a suitable timescale,
(j) 1 (j)
ϕt = ∆(j) + B 2 Wt , (6.17)
( )
with the random initializations ∆(j) chosen as a set of independent random
variables uniformly distributed on the interval [0, 2π). The effect
& of' these ini-
(j)
tializations is to render each component phasor process exp iϕt , and thus
the resultant amplitude obtained from their coherent addition, stationary (cf.
Remarks On stationarity below). The component phase dynamics is simply
(j) 1 (j)
dϕt = B 2 dWt , (6.18)
(j) 2
with squared volatility dϕt = Bdt.
44 RAYLEIGH SCATTERING
6.2.1 Ornstein–Uhlenbeck process

The resultant amplitude dynamics can now be computed explicitly, using the
mathematics of the Ito calculus, as developed in Chapter 2. We shall deduce
presently that, from the basic assumptions of the random walk model, the dy-
namics of Rayleigh scattering is necessarily that of the COU process.
From Ito’s formula (e.g. Oksendal 1998; Karatzas and Shreve 1988) the Ito
differential of (6.16) is
N
& '
(N ) (j) 1 (j) 2 (j)
dEt = idϕt − dϕt exp iϕt . (6.19)
j=1
2
N & '
(j) (j)
The first term j=1 idϕt exp iϕt on the right-hand side of (6.19) consists
of a sum of independent randomly phased Wiener processes, with variance equal
to BN dt, while the second term is independent of the scatterer label j. Thus
from (6.19) we can write
(N ) 1 (N ) 1
dEt = − BEt dt + (BN ) 2 dξt , (6.20)
2
* (j)|dξt | = dt,
2
where ξt is a complex Wiener process satisfying dξt2 = 0. The
process ξt is adapted to the filtration F (ϕ)
= j F , where F is the filtration
(j)
(j)
appropriate to the component scatterer
& phase
' ϕt . The amplitude process Ψt is
(N ) 1
then defined by Ψt = limN →∞ Et /N̄ 2 and satisfies the SDE
1 1
dΨt = − BΨt dt + (Bx) 2 dξt , (6.21)
2
where the continuous valued random variable x, the average scattering
+ , ‘power’,
arises from an asymptotically large population via x = limN →∞ N/N̄ . Observe
that the process exhibits mean reversion (towards the origin), is of constant
volatility, and has an asymptotic stable Gaussian distribution.
Remarks. On geometry of phase wrapping. The effect of the exponentiation
in (6.16) is to ‘phase wrap’ the process onto the unit circle in the complex
plane. Thus, a coordinate discontinuity in the phase process at the boundary of
a coordinate interval of length 2π is mapped to a continuous behaviour on the
unit circle. The manifold of the phasor process is thus a cylinder, with symmetry
axis corresponding to time, on which the sample paths of the phasor process are
continuous (with probability 1).
Remarks. On stationarity. With regard to stationarity, observe primarily that,
given the random uniform initializations (6.17) the distribution of phase remains
uniform over the unit circle at all times – the initial distribution is uniform due
to the nature of ∆(j) and this coincides with the asymptotic distribution. The
RANDOM WALK 45
asymptotic uniform property is most readily apparent by considering the Green’s

function (1.8) phase wrapped onto the unit circle according to
∞

f wrap (θ) = f (θ − 2nπ), (6.22)
n=−∞
where θ is restricted to a coordinate interval [0, 2π) and f is the distribution

appropriate to the ordinary Wiener process on R.19 For large times the phase
wrapped distribution, given the asymptotic behaviour of (1.8), tends to a uniform
distribution on the unit circle. Now, stationarity of the phasor process follows
from the fact that its propagator (a phase wrapped version of that of the Wiener
process) is time translation invariant, combined with the Markov property that
facilitates factorization of the multi-time joint distribution of the process into
two-point propagators. Each component phasor process is therefore (strict sense)
stationary. It follows that the resultant amplitude, according to (6.16), also has
this property.20
Remarks. On the complex nature of the derivation. It is quite intriguing, from
a mathematical perspective and from the point of view of being driven by consid-
erations of a fundamentally physical nature (namely wavelike interference), that
the COU process21 can be derived from such a random phasor diffusion model,
in contrast to which is its application on a comparatively ad hoc basis in other
contexts in the literature, such as financial modelling.
19 Equation (6.22) ensures that f wrap is normalized.

20 These remarks should be contrasted with the case of the pure Wiener process, which is
trivially non-stationary (in both the strict and wide senses) due to the time dependence of its
single point distribution (1.8) and autocorrelation function.
21 A corresponding derivation for real-valued Ornstein–Uhlenbeck processes (without involv-
ing complexification) appears to be unavailable.

7
POPULATION DYNAMICS
The chapter is organized as follows. We begin in Section 7.1 with a discussion of

the dynamics of the scattering cross-section and characterize this process in terms
of a higher order transition process for a discrete scattering population. This is
specialized to the case of the birth–death–immigration (BDI) model in Section
7.2. As the size of the population becomes asymptotically large, in Section 7.3,
we obtain a corresponding limiting description in terms of a normalized contin-
uous process, whose probability density obeys the Kramers–Moyal equation. In
doing so we elucidate under what circumstances such a continuous process is a
diffusion. As a special case we recover the diffusion model applied to scatter-
ing by Jakeman (1980) that corresponds to the BDI model, i.e. for the case of
first-order transitions of birth, death and immigration.
7.1 Master equations and the Kramers–Moyal expansion

Consider a first-order master equation (Van Kampen 1961; Risken 1989) for the
time evolution of the probability PN (t) of the event Xt = N for a continuous time
integer-valued population process (cf. Fig. 7.1). Denoting the (time dependent)
generation and recombination rates by GN (t), RN (t) respectively, we find – from
conservation of probability – that
dPN
= GN −1 PN −1 − (GN + RN )PN + RN +1 PN +1 . (7.1)
dt
This is the discrete analogue of the continuous master equation

∂P(x, t)
= [w(x → x)P(x , t) − w(x → x )P(x, t)]dx . (7.2)
∂t
Now, from the identity for the Taylor expansion f (x ± l) ≡ exp(±l∂/∂x)f (x) we
deduce that, for the continuous valued counterpart xt of the discrete Nt with
step size l, the above rate equation can be re-expressed as

∂P(x, t) ∂ ∂
= exp −l − 1 (G(x, t)P(x, t)) + exp l − 1 (R(x, t)P(x, t)).
∂t ∂x ∂x
(7.3)
This expression is to be compared with the Kramers–Moyal expansion for the
evolution of the probability density
∞ n
∂P ∂
= − [D(n) P] (7.4)
∂t n=1
∂x
46
BDI PROCESSES 47
where the Kramers–Moyal coefficients are defined as the moments

- .
1 (Xt+δt − Xt )n
Dn (x, t) = lim . (7.5)
n! δt→0 δt
Observe in relation to (7.4) the Pawula theorem (see e.g. Risken 1989), which
states that for the probability density P to remain positive at all times, the
expansion (7.4) must either truncate at second order, i.e. reduce to the Fokker–
Planck equation (FPE), or be of infinite order. Combining (7.3) and (7.4) we
see that for a first-order transition process, the Kramers–Moyal coefficients are
determined by the generation and recombination rates according to
ln
D(n) = [G(x, t) + (−1)n R(x, t)] . (7.6)
n!
7.2 Birth–death–immigration processes
In the special case of the BDI model, that is of particular interest to modelling
random fluctuations in scattering populations, the generation and recombination
rates appearing in (7.1) take the forms
G = λN + ν (7.7)
R = µN (7.8)
for positive constants λ, µ, ν, the birth, death and immigration rates, respec-
tively. The asymptotic stationary solution to (7.1) for λ < µ is then the negative
binomial distribution

N +α−1 (N̄ /α)N
PN (∞) = (7.9)
N (1 + N̄ /α)N +α

where ab is a binomial coefficient, α = ν/λ. We deduce that the asymptotic
mean, for λ < µ, is given by
ν
N̄ = . (7.10)
(µ − λ)
The reader should consult Appendix E for a more detailed discussion of the
properties of the BDI model.
The fluctuations in this process are non-Gaussian in nature, as evidenced
by the super-Poissonian normalized variance of the population size N . More
precisely, we define the Fano factor (Fano 1947) according to
VarN
F= . (7.11)
N
In the Gaussian case this is equal to unity, whereas for the negative binomial
case (7.9) we find
N
F= +1 (7.12)
α
so that F > 1, which is the property that leads to ‘bunching’ or ‘clustering’ in
the population (Jakeman and Tough 1988).
48 POPULATION DYNAMICS
7.3 Continuum diffusion limit

In this section we discuss the population dynamics of the cross-section appro-
priate to the modelling of scattering of electromagnetic radiation from certain
types of random media that commonly arise. We extend the analysis and present
a generalization of the BDI model for a first-order transition process (Bartlett
1966) in the context of the Kramers–Moyal expansion (Kramers 1940; Moyal
1949). In doing so, we shall elucidate the features of the transition rates in the
underlying discrete population process that determine whether the associated
continuous (normalized) process is a diffusion.
We begin by writing xnor = x/N̄ so that x̄nor = 1. Then, as N̄ → ∞ (for an
asymptotically large scattering population), xnor becomes a continuous valued
random variable. The probability density for the re-scaled population process
satisfies
P(x, t) ≡ PN (t) = PN̄ x (t). (7.13)
Setting the series expansion parameter l = 1, the evolution of this density is

determined by
∞
n
∂P 1 ∂
= n
− [D(n) P] (7.14)
∂t n=1
N̄ ∂x nor
or in terms of a re-scaled time parameter t = t/N̄

∞ n
∂P 1 ∂
= − [D(n) P]. (7.15)
∂t n=1
N̄ n−1 ∂x nor
In the BDI case we evaluate (7.15) in this limit, as follows. For notational
clarity we shall drop the prime and ‘nor’ suffix on t, x, respectively. From (7.6)
the summation of (7.15) contains

∂
− [ν(1 − x)P] (7.16)
∂x
for n = 1, which equality is exact for finite N̄ ; n = 2 yields

2
1 ∂
{[x(λ + µ) + ν/N̄ ]P} (7.17)
2 ∂x
for finite N̄ (ν remains bounded as N̄ → ∞) which, as N̄ → ∞, becomes

2
∂
(λxP). (7.18)
∂x
For n ≥ 3 the sum contains derivatives of D(n) /N̄ n−1 which, for first-order
transitions, is equal to
CONTINUUM LIMIT 49
1 (G(x, t) + (−1)n R(x, t))

(7.19)
n! N̄ n−1
for step size l = 1. These terms vanish in the limit N̄ → ∞ since G, R and,
consequently, Dn , are linear in N̄ . Thus, the normalized process xnor is a diffusion
with probability density governed by the FPE
∂P ∂ ∂2
=− [ν(1 − x)P] + (λxP) . (7.20)
∂t ∂x ∂x2
In this derivation we have used the expression for the asymptotic mean in terms
of the population parameters (7.10). The SDE associated with the above FPE
can now be read off as
dxt = ν(1 − x)dt + (2λx)1/2 dWt . (7.21)
The asymptotic distribution for this process is then P∞ (x) ∝ xα−1 exp(−αx)
(α = ν/λ), so the re-scaled variate αx has the gamma distribution
xα−1 exp(−x)
Γα (x) = , (7.22)
Γ(α)
where Γ(α) is the gamma function (see e.g. Jeffreys and Jeffreys 1966). The
situation in respect of the limiting N̄ behaviour and the population parameters
can therefore be summarized as follows. If we let ν → ∞, keeping N̄ fixed, then
the discrete Poisson distribution emerges from (7.9). If one now lets N̄ → ∞, the
associated continuous distribution can be identified as a delta function. The delta
function also occurs if the limits are taken in the reverse order, thus N̄ → ∞,
via µ → λ yields a gamma distribution, and ν → ∞, with fixed mean, reduces
the gamma distribution to a delta function.
Remark. The case of the pure Poisson process is obtained when the birth and
death rates both vanish and there is a constant immigration rate.
For a more general first-order master equation, we expect the transition coef-
ficients G, R to have higher order polynomial behaviour. In this respect we have
the following result.
Proposition 7.1 Given a first-order discrete valued transition process Nt , a
necessary and sufficient condition for the associated continuous valued normal-
ized random variable xnor to be of diffusion type (with respect to the re-scaled
time t ) is that G, R are O(N α ) for α < 3 and G − R ∼ O(N β ) for β < 2.
Proof Sufficiency is immediate from the N̄ → ∞ limiting behaviour of (G(x, t)+
(−1)n R(x, t))/N̄ n−1 . Conversely, the zero contribution for n = 3 implies the
required behaviour for G − R and thus trivially that G − R ∼ O(N β ) for
β < 3. Combining this with the relation G + R ∼ O(N β ), β < 3 which fol-
lows from the zero contribution for n = 4, yields the required behaviour for G, R
separately.
50 POPULATION DYNAMICS
N+2
(2) (2)
GN RN+2
N+1
(1) (1)
GN RN+1
N
(1) (1) (2)
GN–1 RN (2)
GN–2 RN
N–1
N–2
Fig. 7.1. Higher order transition process.
It is of some interest that, even for a first-order master equation, the expan-
sion (7.15) is of infinite order generically, and accordingly the fluctuations in the
process can not be driven by the usual Wiener process (cf. Jakeman et al. 2003).
For higher order transitions the situation is depicted in Fig. 7.1. We can express
the evolution of the probability density in a similar fashion as
∞
(i)
dPN (i) (i) (i)
= {GN −i PN −i − (GN + RN )PN + RN +i PN +i }, (7.23)
dt i=1
(i) (i)
where GN −i = RN = 0 for i > N , thus maintaining N ≥ 0. Correspondingly
the Kramers–Moyal coefficients are given by
∞
ln n (i) n (i)
D =
(n)
i (G + (−1) R ) . (7.24)
n! i=1
The order of the Kramers–Moyal expansion for higher order transition processes
can be summarized by the following result.
Proposition 7.2 The Kramers–Moyal expansion (7.15) is finite order as N̄ → ∞
if and only if for all integer m > 2
∞

im [G(i) + (−1)m R(i) ] = o(N m−1 ) (7.25)
i=1
CONTINUUM LIMIT 51
as N → ∞. Correspondingly, the continuous normalized stochastic process xnor

is a diffusion with respect to the re-scaled time t = t/N̄ .
The alternative possibility is supplied by the Pawula theorem.

We refer the reader to Appendix E for details of the solution to the rate
equation for the BDI model via the partition function. A complete treatment of
the special limiting cases that arise in that context is also provided therein.
8
DYNAMICS OF K-SCATTERING
We derive the stochastic dynamics of the complex-valued amplitude resulting

from coherent scattering from a random population of scatterers when this be-
comes asymptotically large. Considerations of a random walk model, introduced
by Jakeman, are used to derive stochastic differential equations (SDEs) for the
amplitude and corresponding intensity and phase stochastic processes. An anal-
ysis of the correlation structure in the fluctuations is provided and interpreted
geometrically in terms of the gauge invariant properties of the field and the
Markov property. A Fokker–Planck description for the evolution of the proba-
bility density is given and the equilibrium and detailed balance conditions are
shown to hold. Expressions for the intensity autocorrelation function and power
spectral density are provided in closed form. The practical implications of the
stochastic theory are discussed.
Recent developments in the diffusion-based analysis of scattering from ran-
dom media, reported in Field and Tough (2003a), have led to significant results
that enable the identification of K-distributed noise processes in electromag-
netic scattering. The results comprise various SDEs for the scattered amplitude,
intensity, phase, and scattering cross-section motivated by a combination of the-
oretical considerations and analysis of empirical data.
The purpose of the present chapter is to formulate the stochastic dynamics
of the electromagnetic field, scattered from a random medium that consists of a
collection of independent component scatterers, whose population size has fluctu-
ations in accordance with the birth–death–immigration (BDI) model (Bartlett
1966). This is achieved from first principles via considerations of the complex
random walk model introduced by Jakeman (Jakeman 1980). Our results thus
provide the theoretical foundation of the anomaly detection technique reported
earlier (Field and Tough 2003a), and extend the model to include a detailed
description of the intensity autocorrelation function and power spectral density.
The chapter is organized as follows. Section 8.1 recalls the complex amplitude
SDE in the case of a constant scattering cross-section, i.e. Rayleigh scattering,
that follows from considerations of a complex random walk model, and extends
this to the K-distributed case via the insertion of step number fluctuations.
The resulting SDE for the amplitude process is used to derive corresponding
SDEs for the intensity and phase processes, and expressions for the squared
volatilities of each of these processes are provided. This framework is used to
explain the geometrical structure of the correlations in the fluctuations of the
complex amplitude in Section 8.2.
52
K-AMPLITUDE 53
In Section 8.3, we provide a Fokker–Planck description for the joint probabil-

ity density of the scattering cross-section and intensity processes and study the
asymptotic behaviour. It is verified that the model possesses the joint probability
appropriate to the K-distribution, and that the condition for detailed balance is
also satisfied.
Section 8.4 provides a detailed analysis of the finite-time correlation properties
of the stochastic model. The Green’s functions arising from the Fokker–Planck
description are computed and expressions for the autocorrelation functions and
power spectral density are provided in closed form, and the spectral proper-
ties discussed. The chapter concludes with an account of the interpretation and
experimental implications of the theoretical framework proposed.
8.1 Stochastic dynamics of K-amplitude process

8.1.0.1 K-distributed noise In the case of step number fluctuations in the ran-
dom walk model (6.16), we define the amplitude Ψt in a similar manner to the
Rayleigh case above, with+ the modification
, that we employ a time dependent Nt
such that xt = limNt →∞ Nt /N̄ . Thus
 
 1 Nt & '
(j)
Ψt = lim exp iϕt (8.1)
Nt →∞  N̄ 2 
1
j=1
 
 N 12 1 Nt & '
t (j)
= lim exp iϕt (8.2)
Nt →∞  N̄ 
1
2
N t j=1
1
2
= xt γt , (8.3)
& 1
'
(N )
where γt = limN →∞ Et t /Nt2 .
Remark. On the failure of the central limit theorem. It is important to observe
the failure of central limit theorem to apply in this case, because of the nature
of the square root normalization and the fact that the number of steps in the
random walk model fluctuates in time.
According to the arguments given in the Rayleigh case above, γt is a complex
Ornstein–Uhlenbeck (COU) process which obeys the SDE
1 1
dγt = − Bγt dt + B 2 dξt . (8.4)
2
Observe from (6.21), therefore, that γt is a unit power Rayleigh process. The
above
+ equation , for γt can be solved by considering the stochastic differential
d exp( 12 Bt)γt , which leads to the solution
t
1 1 1
γt = exp − Bt γ0 + B 2 exp Bs dξs . (8.5)
2 0 2
54 K-SCATTERING
We deduce the expectation formulae

1
E [γt ] = exp − Bt γ0 , (8.6)
2
+ 2,
E |γt | = 1 + exp(−Bt)(|γ0 |2 − 1). (8.7)
+ 2,
From (8.7) it follows that limt→∞ E |γt | = 1 and so from (8.3) we find the
intensity process, defined by zt = |Ψt |2 , satisfies
(x)
E[zt |Ft ] = xt (8.8)
where the expectation is conditional on a sharp value of the cross-section and
thus averages of the phases ϕ(j) in (6.16) only. The SDE for Ψt , as determined
by (8.3), can now be derived, via the Ito product formula d(Xt Yt ) ≡ Xt dYt +
Yt dXt + dXt dYt . For this purpose, it is convenient to introduce the square-
1/2
root cross-section rt = xt and thus, from (8.3), we find dΨt = rt dγt + γt drt .
Observe that the cross term drt dγt does not feature in this relation owing to
(x) (ϕ)
the independence of Wt and ξt , which originate from the intrinsic scattering
population and the scattered electromagnetic field, respectively.
The BDI model (Bartlett 1966) posits a first-order master equation for the
population Nt , with respective generation and recombination rates G = λN + ν,
R = µN . For an asymptotically large population, N → ∞, we deduce (Tough
1987; Jakeman and Tough 1988; Field and Tough 2003a) that the re-scaled pop-
ulation variate x
→ αx satisfies the SDE
1 (x)
dxt = A(α − xt )dt + (2Axt ) 2 dWt (8.9)
(x)
for an independent Wiener process Wt , where α = ν/λ. This leads to the
asymptotic Γ-distribution for xt ,
xα−1 exp(−x)
Γα (x) = (8.10)
Γ(α)
1/2
so that Var[x] = x = α. From Ito’s formula applied to rt = xt we find
1/2 3/2
drt = dxt /2xt − dx2t /8xt , and thus from (8.9)
12
2(α − xt ) − 1 A (x)
drt = A dt + dWt . (8.11)
4rt 2
This leads to the following result.
Theorem 8.1 (Field equations.) In the K-distributed case the scattered ampli-
tude is governed by the SDE
1 1
dΨt 1 B2 2(α − xt ) − 1 A 2 (x)
= − Bdt + dξt + A dt + dWt . (8.12)
Ψt 2 γt 4xt 2xt
This evolution is invariant under the U (1) gauge transformation Ψt
→ exp(iΛ)Ψt ,
for constant Λ.
K-AMPLITUDE 55
Remark. In many respects this result, which is derived from very fundamental
physical and mathematical principles, represents the central theoretical advance
within the entire body of the monograph.
In the above expression, A and B are independent constants with the di-
mension of frequency, and they may take arbitrary values. In most situations
of interest, however, such as those reported in Field and Tough (2003a) and
developed in Chapter 12, the wavelength of the illuminating radiation is such
that the two corresponding reciprocal correlation timescales satisfy A B. The
description of Rayleigh scattering (i.e. constant scattering cross-section) is then
recovered when A = 0. Theorem 8.1 implies the following result.
Corollary 8.2 The squared volatility of the amplitude process Ψt is given by

Azt
|dΨt |2 = Bxt + dt. (8.13)
2xt
It is the linearity of the right-hand side above in zt that, in part, enables the
anomaly detection mechanism described in Section 12.2 (orig. Field and Tough
(2003a)).
8.1.1 Intensity
The stochastic differential of the intensity process zt can be expressed in terms
of the amplitude via the identity
dzt = Ψ∗t dΨt + Ψt dΨ∗t + |dΨt |2 (8.14)
which follows from zt = |Ψt |2 . From (8.4) we find
Ψ∗t dΨt + Ψt dΨ∗t = xt (γt∗ dγt + γt dγt∗ ) + 2|γt |2 rt drt

1 2zt
= −Bzt dt + B 2 xt (γt∗ dξt + γt dξt∗ ) + drt . (8.15)
rt
The terms involving dξt above can be combined in terms of a real-valued Wiener
(ϕ)
process Wt according to
12
2zt
γt∗ dξt γt dξt∗
(ϕ)
+ ≡ dWt . (8.16)
xt
We deduce from (8.13), (8.14), (8.15) that

1 (ϕ) 2zt Azt
dzt = −Bzt dt + (2Bzt xt ) 2 dWt + drt + Bxt + dt. (8.17)
rt 2xt
In combination with (8.11) this leads to the following result.

56 K-SCATTERING
Proposition 8.3 The intensity SDE is given by
1
Azt (α − xt ) 2Azt2 2 (z)
dzt = B(xt − zt ) + dt + 2Bxt zt + dWt (8.18)
xt xt
(z) (x)
in which Wt is correlated with Wt of (8.9), and satisfies
12 12
Azt2 (z) 1 (ϕ) A (x)
Bxt zt + dWt = (Bxt zt ) 2 dWt + zt dWt . (8.19)
xt xt
(ϕ) (j)
The filtration of Wt arises from the constituent phases ϕt in the random
(x)
walk according to (8.16), while that of Wt stems solely from the fluctuations
in the endogenously specified population model – such (scattering) model is thus
sometimes referred to as the endogenous model. Observe that, if B += 0, ,(8.19)
implies W (z) = W (x) , and from (8.4) γt is constant, so |γt |2 = E |γt |2 = 1.
Accordingly, zt = xt is a solution of (8.18), as required by (8.3). From Proposition
8.3 we obtain the following result.
Corollary 8.4 The squared intensity volatility is determined by

2Azt2
dzt2 = 2Bxt zt + dt. (8.20)
xt
Alternatively, in terms of the amplitude process Ψt the squared volatility in the

intensity zt can be expressed as
dzt2 ≡ Ψ2t dΨ∗t 2 + Ψ∗t 2 dΨ2t + 2zt |dΨt |2 (8.21)
which, from (8.12), leads to the above expression for dzt2 . Observe that for A B
the dominant contribution to the squared intensity volatility is proportional to
the instantaneous value of the intensity. Thus for a Rayleigh timescale B −1 , over
which xt remains approximately constant, the time series for dzt2 and zt should
exhibit strong correlation. This feature has been experimentally verified in a case
of optical scattering, which is described in Section 12.1 (orig. §4(a) of Field and
Tough (2003a)).
√
In terms of the square-root intensity Rt = zt , an application of Ito’s formula
to (8.20) yields the following result.
Corollary 8.5 The squared volatility in the modulus amplitude is determined by

1 Azt
dRt2 = Bxt + dt. (8.22)
2 xt
K-AMPLITUDE 57
8.1.2 Phase
The complex amplitude process can be expressed in polar form Ψt = Rt exp(iθt )
and thus, writing iθt = log(Ψt /Rt ), we deduce from Ito’s formula that
2 2
dΨt 1 dΨt dRt 1 dRt
idθt = − − + . (8.23)
Ψt 2 Ψt Rt 2 Rt
Since the left-hand side is purely imaginary we can express dθt in terms of Ψt
alone as
2 ∗ 2
1 dΨt 1 dΨt dΨt 1 dΨ∗t

dθt = − − − . (8.24)
2i Ψt 2 Ψt Ψ∗t 2 Ψ∗t
Accordingly, the squared phase volatility is determined by the identity
|dΨt |2 dΨ2t dΨ∗t 2

dθt2 ≡ − − . (8.25)
2zt 2
4Ψt 4Ψ∗t 2
From (8.12) we have

2 1 1
dΨt 1 dΨt A(α − xt − 1) 1 A 2 (x) B2
− = − B dt + dWt + dξt .
Ψt 2 Ψt 2xt 2 2xt γt
(8.26)
Hence, from (8.24), θt obeys the SDE

1
B2
dθt = (γ ∗ dξt − γt dξt∗ ). (8.27)
2i|γt |2 t
As in the derivation of the SDE for the intensity, we can express the terms
(θ)
involving ξt in (8.27) as a distinct real-valued Wiener process Wt according to
12
1 ∗ zt
(γ dξt − γt dξt∗ ) ≡
(θ)
dWt . (8.28)
2i t 2xt
Thus we obtain the following result.

Proposition 8.6 The resultant phase θt of the complex K-amplitude process Ψt
obeys the SDE
12
Bxt (θ)
dθt = dWt (8.29)
2zt
which has vanishing drift.

The result has the following consequence.
58 K-SCATTERING
Corollary 8.7 The squared resultant phase volatility is given by

Bxt
dθt2 = dt. (8.30)
2zt
Alternatively, (8.30) can be derived by applying (8.25) to (8.12). This result
accords with the general scaling and symmetry arguments for the behaviour of
the squared phase volatility put forward in §4 of Field and Tough (2003a). The
situation should be contrasted with the differentiable model for the process Ψt
(Jakeman et al. 2001) for which the intensity-weighted phase-derivative, instead
of its square, has minimal variance.
These various relations allow the K-amplitude dynamics to be recast in terms
of W (x) , W (θ) , W (r) as follows.
Corollary 8.8 The K-amplitude satisfies the SDE
1 1
dψt 2(α − xt ) − 1 1 Azt + Bx2t 2 (z) Bxt 2 (θ)
= A − B dt + dWt + i dWt
ψt 4xt 2 2xt zt 2zt
(8.31)
in which, alternatively, the Wiener terms can be expressed as
1 1
A 2 (x) Bxt 2 (r) (θ)

dWt + dWt + idWt . (8.32)
2xt 2zt
(x)
The following result, implied by (8.12) and the identities dξt2 = dξt dWt = 0,
is relevant in connection with the geometry of fluctuations for weak scattering
processes discussed in Section 9.3.
Corollary 8.9 The product cross-section/K-amplitude stochastic differentials
satisfy
dx2t = 2Axt dt, (8.33)
dxt dψt = Aψt dt, (8.34)

2 Aψt2
dψt = dt, (8.35)
2xt

Azt
|dψt | =
2
+ Bxt dt. (8.36)
2xt
The results above lead to expressions for the frequency constants A and B,
(j)
as follows. With respect to an average over the phase fluctuations dϕt there
exists a residual constant term in the squared phase volatility, i.e.
& ' 1
E dθt2 /dt|F (ϕ) = B. (8.37)
2
In principle, this enables the Rayleigh constant B to be deduced from scattering
data (alternatively an estimate of the Rayleigh correlation timescale B −1 can
GEOMETRY OF FLUCTUATIONS 59
be found from measuring the time difference between successive peaks in the
intensity time series {zt }). Expression (8.30) implies that xt = 2B−1 zt dθt2 /dt
and so the instantaneous values of the cross-section xt , and therefore rt , are
observable through the squared phase fluctuations. Consequently, the constant
A can be deduced from the square of (8.11)
drt2 1
= A (8.38)
dt 2
(cf. §3 in Field and Tough (2003a) for an account of the observability of the
squared volatilities for discretely sampled time series data).
8.2 Geometry of K-amplitude fluctuations

In relation to the discussion of Chapter 3 and Section 3.1 therein, we apply the
concept of diffusion tensor arising in stochastic differential geometry to the de-
tailed amplitude dynamics appropriate to K-scattering. These stochastic dynam-
ics, as presented in Theorem 8.1, enable one to deduce the correlation structure
in the fluctuations of the complex amplitude process Ψt .
A geometric insight into these properties can be gained from the symmetry
properties of the process Ψt . Theorem 8.1 shows that the SDE for Ψt is invariant
under the U (1) gauge transformation Ψt
→ exp(iΛ)Ψt . Since the drift in (8.26)
is real-valued, the identity (8.24) implies that θt has vanishing drift, as seen
explicitly from Proposition 8.6. The resulting asymptotic probability distribution
is therefore U (1) symmetric.
In respect of time evolution, the SDE (8.12) has the Markov property (see Ap-
pendix C), that the evolution it determines depends on the instantaneous value
of Ψt and is independent of the history of the process {Ψt |t < t}. This feature
yields a preferred symmetry, namely the instantaneous radial direction deter-
mined by Ψt and its orthogonal θ direction. The diffusion tensor σ ij , determined
by dΨi dΨj = σ ij dt is real and symmetric and therefore can be diagonalized over
C. In the non-degenerate case its eigenvectors constitute a unique orthogonal
pair corresponding to the directions in which the component Wiener increments
are independent. On grounds of the above symmetry, we anticipate that the
space of eigen-directions contains the instantaneous radial and attendant θ di-
rection. This geometrical property can be verified explicitly from (8.16), (8.19),
and (8.28), which imply the following.
(z) (θ)
Proposition 8.10 The cross-correlation between Wt , Wt vanishes identi-
(z) (θ)
cally, i.e. dWt dWt = 0. Correspondingly dRt dθt = 0, i.e. the fluctuations in
Rt , θt are statistically independent.
In terms of the I, Q component representation the coordinate transformations
I = R cos θ, Q = R sin θ and the property dRt dθt = 0 imply the geometric
relation
dIt dQt = cos θt sin θt (dRt2 − Rt2 dθt2 ). (8.39)
60 K-SCATTERING

Proposition 8.11 The It , Qt components of Ψt are independent if and only
if σ(R) = Rσ(θ) , i.e. σ(z) = 2zσ(θ) . A departure from this relation induces a
correlation between the Wiener increments in It , Qt .
Alternatively, this result can be derived using the contravariance of the diffu-
sion tensor σ ij , which enables one to translate between its I, Q and R, θ compo-
nents via the above coordinate transformation (e.g. Risken 1989). In the general
case, we find from (8.20), (8.30) that
2
σ(z) 4Az 3
= 4z 2 + . (8.40)
2
σ(θ) Bx2
This relation can be used to characterize the geometry of the fluctuations as

follows.
Proposition 8.12 In the K-distributed case, A = 0, the diffusion tensor is non-
degenerate, and the fluctuations δIt , δQt are correlated. The (comoving) error
surface S of δΨt , defined by the quadratic form
σ II δIt2 + 2σ IQ δIt δQt + σ QQ δQ2t = 1, (8.41)
is an ellipse whose major axis lies in the instantaneous radial direction defined
by Ψt . Degeneracy occurs only in the Rayleigh case, A = 0, for which S is a
circle, i.e. the fluctuations in Ψt are isotropic.
We remark, in general, that the random variables It , Qt possess a joint prob-
ability distribution that is U (1) symmetric, i.e. given by a surface of revolution
about the perpendicular axis to the origin in the I, Q-plane. Nevertheless It ,
Qt are correlated in general, and become independent only in the Rayleigh case,
A = 0, for which the surface of revolution is Gaussian. In this case the component
It , Qt processes can be described by the pair of (uncoupled) Ornstein–Uhlenbeck
processes determined as the real and imaginary parts of (6.21).
8.3 Asymptotic behaviour

In this section we shall investigate the equilibrium and detailed balance prop-
erties of K-scattering. The mathematical tools involved are the FPE and the
concept of stochastic current as introduced in Section 3.2.
8.3.1 Equilibrium distribution

We recall the covariant form of the FPE for the asymptotic joint distribution
P(x, z, t) (e.g. Risken 1989)
∂P 1
=− ∂i (bi P) + ∂i ∂j (σ ij P). (8.42)
∂t i
2 i,j
ASYMPTOTICS 61
From (8.9), (8.18), (8.19) the components of the diffusion tensor in the x, z
coordinate representation are given by

ij Ax Az
σ =2 (8.43)
Az Bxz + Az 2 /x
while the drift vector has components

A(α − x)
βi = 2 . (8.44)
B(x − z) + Az(α − x)/x
From (8.42), therefore, we deduce the following.

Proposition 8.13 The FPE for the joint distribution of the cross-section xt
and intensity zt is
∂P ( )
= B −∂z [(x − z)P] + x∂z2 [zP]
∂t
z(α − x)P
+ A −∂z − ∂x ((α − x)P)
x
2
z P
+ A ∂z2 + 2∂x ∂z (zP) + ∂x2 (xP) . (8.45)
x
This admits the asymptotic joint distribution
xα−2 exp(−x − z/x)

P= . (8.46)
Γ(α)
Proof The derivation of (8.45) follows immediately from (8.42), (8.43), and
(8.44), while the following identities for the derivatives of the joint distribution
P
∂z P = −
x z
∂z (zP) = 1 − P
x
2 z
∂z2 (zP) = − + 2 P
x x

α−1 z
∂x P = + 2 −1 P
x x
2 2 3
α−1 z α − 1 2z
∂x P =
2
+ 2 −1 − − 3 P (8.47)
x x x2 x
enable one to verify that (8.46) is an asymptotic solution of (8.45).

Observe from (8.9) and (8.45) that the model behaviour of the scattering
cross-section is endogenously specified, i.e. the parameters involved arise from
62 K-SCATTERING
the population alone, independently of the electromagnetic field. Thus (8.9) is

independent of zt . Nevertheless, there exists a non-linear coupling between the xt ,
(x)
zt variables, owing to the presence of xt in (8.18) and the correlation of Wt ,
(z)
Wt according to (8.19). Therefore xt , zt are statistically dependent random
variables, which relation is symmetric. The situation in regard to the endoge-
nous specification of the evolution of the cross-section through (8.9) should be
contrasted with the previous discussion given in Tough (1987), and its general-
izations in Field and Tough (2003a) in which it is necessary that the SDE for
(x) (z)
xt has an explicit zt dependence, and in which the Wiener processes Wt , Wt
are considered to be independent. Although these analyses preserve the joint
distribution appropriate to K-distributed noise, these models are not so natural
from a physical point of view.
8.3.2 Detailed balance

The covariant FPE (8.42) can be re-expressed as the equation of continuity
∂P
+ ∂i (Pv i ) = 0 (8.48)
∂t i
where the current v i is defined by
1
v i = bi − P −1 ∂j (σ ij P). (8.49)
2 j
In addition to the equilibrium condition ∂P/∂t = 0, the condition for detailed

balance states that v i = 0. Explicit calculation using (8.47) shows that v i van-
ishes asymptotically. Alternatively, a more intuitive argument for this property
is as follows. From (8.3), we have the factorization zt = xt ut , where ut = |γt |2 ,
in which the factors xt , ut are independent random variables. The coordinate

transformation xi
→ xi : (x, z)
→ (x, u) recasts the joint distribution (8.46) and
SDEs (8.9), (8.18) such that

Aβ 1 (x)
βi = , (8.50)
Bβ 2 (u)

i j AΣ1 1 (x) 0
σ = 2 2 (8.51)
0 BΣ (u)
where the functions β · (·), Σ· (·) are determined from (8.4), (8.9) and are indepen-
dent of A, B. The equilibrium condition in the (x, u) representation, obtained by
setting the left-hand side of (8.42) equal to zero, implies detailed balance, since
this condition holds for arbitrary values of the constants A, B. Consequently,
v i in the (x, z) representation also vanishes, since v i transforms homogeneously
(i.e. tensorially) under coordinate transformations (see e.g. Risken 1989).
CORRELATION AND SPECTRA 63
8.4 Correlation and spectra

For simplicity, we adopt a timescale such that the constant B of (8.4) is equal to
unity. The independent constant A will then satisfy A 1 in most practical sit-
uations of interest (e.g. scattering at radar wavelength), although this condition
is not necessary for the validity of the expressions that follow in this section.
8.4.1 Intensity autocorrelation

It is convenient to write the intensity process in the product representation zt =
ut xt . From (8.4) the process ut = |γt |2 satisfies the SDE
√ (u)
dut = (1 − ut )dt + 2ut dWt , (8.52)
√
where γt dξt∗ + γt∗ dξt = 2ut dWt . The propagator (i.e. Green’s function for the
(u)
corresponding FPE) for the process ut is given by
√
1 u + u0 exp(−t) 2 exp(−t/2) uu0
P (u, t|u0 ) = exp − I0
1 − exp(−t) 1 − exp(−t) 1 − exp(−t)
(8.53)
where Iα denotes the modified Bessel function (e.g. Jeffreys and Jeffreys 1966).
In a similar manner, the propagator for (8.9) is given by
(α−1)/2
1 x exp(At) (x + x0 exp(−At))
P (x, t|x0 ) = exp −
1 − exp(−At) x0 1 − exp(−At)
√
2 exp(−At/2) xx0
× Iα−1 . (8.54)
1 − exp(−At)
This can be re-expressed as a series expansion
∞
n!
P (x, t|x0 ) = xα−1 exp(−x) exp(−Ant)Lα−1
n (x)Lα−1
n (x0 ), (8.55)
n=0
Γ(n + α)
where the Laguerre polynomials Lα

n are defined by
n
x−α exp(x) d
Lα
n (x) = (xα+n exp(−x)) (8.56)
n! dx
(cf. Wong 1963 for corresponding derivations). Combining (8.53), (8.55) leads to
the following result.
64 K-SCATTERING
Proposition 8.14 The propagator for (8.45) is given by
P (z, x, t|z0 , x0 )
(α−1)/2
1 x exp(At)
=
x(1 − exp(−t))(1 − exp(−At)) x0

z/x + z0 exp(−t)/x0 x + x0 exp(−At)
× exp − exp −
1 − exp(−t) 1 − exp(−At)
4 √
2 exp(−t/2) zz0 2 exp(−At/2) xx0
× I0 Iα−1 . (8.57)
1 − exp(−t) xx0 1 − exp(−At)
Thus a general two-point correlation function can be expressed as the integral
∞
F1 (xt , zt )F2 (x0 , z0 ) = dx dz dx0 dz0 F1 (x, z)F2 (x0 , z0 )P (x, z, t|x0 , z0 )
0
xα−2 exp(−z0 /x0 − x0 )
× 0
. (8.58)
Γ(α)
In particular, we deduce the following important consequence.
Corollary 8.15 The intensity autocorrelation function is determined by the fol-
lowing expression:
zt z0 = ut u0 xt x0
= α(α + exp(−At))(1 + exp(−t)). (8.59)
8.4.2 Power spectral density

In the additional presence of a Doppler frequency shift ω0 22 , the process γt of
Section 6.2 is modified to obey the SDE

1
dγt = − + iω0 γt dt + dξt . (8.60)
2
The amplitude process Ψt determined by (8.12) is stationary, since there is no
explicit time dependence in (8.12), the phase distribution is uniform, and the
modulus amplitude Rt has a stationary distribution in accordance with the sta-
tionary K-distribution for the intensity (it is assumed that the distributions of
the initial values of Ψ and its associated processes are given by their asymptotic
stationary distributions). Therefore, we apply the Wiener–Khintchine theorem
which asserts that the power spectral density S(ω) is equal to the Fourier trans-
form (denoted tilde) of the autocorrelation function, i.e.
Ψ̃(ω)Ψ̃(ω ) = πδ(ω − ω )S(ω), (8.61)
where S(ω) = Ψ5 ∗
t Ψ0 .
22 The presence of ω0 = 0 is important in radar applications; see e.g. Helstrom 1960.

CORRELATION AND SPECTRA 65
Remarks. Concerning power spectra and stationarity. In the case of a non-

stationary process the power spectral density is defined as in the left-hand side
of (8.61), i.e. the autocorrelation function in the frequency domain – thus, in
general, it is a function of a pair of frequencies. The stationarity property leads
to the delta function dependence on the right-hand side of (8.61) so that, in
this case, the power spectral density becomes essentially a function of a single
frequency only.
The amplitude autocorrelation function satisfies
√
Ψt Ψ∗0 = xt x0 exp(−|t|/2 − iω0 t). (8.62)
√
Using the propagator expansion (8.55), the evaluation of the factor xt x0
proceeds according to
∞ ∞
√ xα−1 exp(−x0 ) √
xt x0 = dx dx0 0 P (x, t|x0 ) xx0
0 0 Γ(α)
∞ ∞ 2
1 n! α− 12 α−1
= exp(−Ant) x exp(−x)Ln (x)dx
Γ(α) n=0 Γ(n + α) 0
∞ 2
1 n! Γ(α + 12 )Γ(n − 12 )
= exp(−Ant) √
Γ(α) n=0 Γ(n + α) n!2 π

Γ(α + 12 )2 1 1
= 2F1 − , − , α, exp(−At) . (8.63)
Γ(α)2 2 2
Here the hypergeometric function√ 2F1 is identified from its series expansion.
When t → ∞, (8.63) approaches x2 , as anticipated from the decorrelation of
xt , x0 over large times. As t → 0 we find, from the identity due to Gauss,
Γ(c)Γ(c − a − b)
2F1 (a, b, c, 1) = , (8.64)
Γ(c − a)Γ(c − b)
that (8.63) reduces to the anticipated form
√ Γ(α + 1)
lim xt x0 = = α = x. (8.65)
t→0 Γ(α)
Expressions (8.62), (8.63) lead to the following result.

Proposition 8.16 The autocorrelation function of the complex amplitude pro-
cess Ψt is given by

∗ Γ(α + 12 )2 1 1 |t|
Ψt Ψ0 = 2F1 − , − , α, exp(−At) exp − − iω0 t . (8.66)
Γ(α)2 2 2 2
According to the Wiener–Khintchine theorem, a Fourier transform of this result
has the following consequence.
66 K-SCATTERING
Corollary 8.17 The power spectral density of the K-distributed noise process
characterized by (8.12) is given by

Γ(α + 12 )2 ∞ 1 1 t
S(ω) = 2 2F1 − , − , α, exp(−At) exp − cos((ω − ω0 )t).
Γ(α)2 0 2 2 2
(8.67)
Expanding the hypergeometric function 2F1 as a series and integrating term by

term, the resulting series is recognized as a generalized hypergeometric function
of unit argument. Thus
S(ω)

3F2 (−1/2, −1/2, (1/2 + i(ω − ω0 ))/A; α, 1 + (1/2 + i(ω − ω0 ))/A; 1)
= 2
1/2 + i(ω − ω0 )
Γ(α + 12 )2
× . (8.68)
Γ(α)2
These calculations illustrate how the compound representation of the ampli-
tude (8.3) facilitates the analysis of the associated FPE (8.45). In terms of the
constituent spectra for the two component factors in (8.3) observe that, since
these component processes are independent, the autocorrelation of the resul-
tant amplitude factorizes into that of the components. Therefore, according to
the Wiener–Khintchine and convolution theorems, the power spectrum of the
resultant amplitude is equal to the convolution of the component spectra.
The explicit dependence on the electromagnetic frequency scale B in the
above expressions may be restored most simply on dimensional grounds.23 The
frequency scales A and B, which satisfy A B (for typical carrier frequencies),
may be found from an experimentally observed autocorrelation by fitting these
parameters according to the theoretical expression (8.66). From the series expan-
sion of the hypergeometric function, the radar cross-section (RCS) component of
the resultant amplitude autocorrelation (8.66) may be written as a sum of terms
proportional to exp(−nAt) (Fayard 2008), whose spectra are therefore Cauchy
or ‘Lorentzian’, while the Rayleigh spectral component is also Cauchy.24 Since
the Cauchy distribution is stable (see Appendix A), via the Fourier convolution
S (ψ) = S (r) ∗ S (γ) and taking the leading (n = 1) term in the hypergeomet-
ric expansion, it follows that the spectrum of the resultant amplitude ψ is also
(approximately) Cauchy, with FWHM equal to 2A + B. (The DC part of the
RCS spectrum, equal to αδ(ω) and reflecting merely the fact that the RCS has
a constant non-zero mean value of α, has been removed for clarity.)
Observe with regard to the above analysis the special case that the cross-
section is constant (A = 0). Then the amplitude autocorrelation consists solely
for example, a term exp(− 12 |t|) becomes exp(− 12 B|t|) in general units of time, etc..
23 Thus,
24 Anautocorrelation function R(τ ) = exp(− 12 kt) has associated power spectrum S(ω) =
2k/π(k2 + 4ω 2 ) with full-width-half-maximum (FWHM) therefore equal to k.
INTERPRETATION AND IMPLICATIONS 67
of an exponential decay term and thus the spectrum is exactly Lorentzian, i.e. a
Cauchy distribution (see Appendix A).
8.5 Interpretation and implications

The study provides the first theoretical account of K-distributed noise processes
in which the continuous time dynamical features of the electromagnetic scat-
tering process are fully captured. This has been achieved via the formulation
of SDEs for the scattered amplitude, using the primitive assumptions of the
complex random walk model.
The results substantiate an earlier proposal for anomaly detection in the
context of such processes (Field and Tough 2003a) based on the concept of ob-
servability in the fluctuations of the complex amplitude process over a sample
path. In this respect, Corollary 8.2 leads to a correlation between the observed
|dΨt |2 and its predicted value of the form c(|dΨt |2 , zt ), which should approxi-
mate unity within the domain of validity of the model. This feature enables an
anomaly detection mechanism for K-distributed noise processes, which has been
successfully tested on experimental data, and is described later in Section 12.2
herein (cf. §4(b) of Field and Tough (2003a) where the result first appears). It is
of considerable importance that (8.13) can be derived from theoretical consider-
ations alone, as described in Section 8.1.
The formulation of the continuous time dynamics is more fundamental than
knowledge of certain statistical properties of a model and, moreover, implies the
form of all correlation functions and higher order statistics. In this respect we
have provided closed form expressions for the intensity autocorrelation function
and the power spectral density, which should be applicable in situations of radar
and laser physics. The tractability of such expressions is facilitated by the use of
computational tools such as Mathematica (Wolfram 1999).
The methodology we describe admits the generalization of the SDE for the
scattering cross-section (8.9) to more general endogenous models of population
processes, such as described in orig. §2 of Field and Tough (2003a) and Section
7.3 herein. This could include corresponding descriptions of the electromagnetic
scattering processes that lead e.g. to the Weibull distribution (applied to scat-
ter from land clutter), the intensity compound K-distribution for various radar
parameters (e.g. applied to synthetic aperture radar), and other examples (cf.
Jakeman and Tough 1988).
Corollary 8.7 implies that the instantaneous value of the scattering cross-
section is observable from the scattered amplitude. The cross-section is of pri-
mary significance in anomaly detection within a random scattering medium, e.g.
the sea-surface, heterogeneous media, which had previously been regarded as
a hidden physical variable whose instantaneous values were not deducible from
the scattering data. Expression (8.22) for the squared volatility in the modulus
amplitude should also find application to problems in incoherent radar detection
where the total scattered phase information is not available.
68 K-SCATTERING
We have recalled in this chapter that the random walk model with step num-
ber fluctuations, due to Jakeman (see Jakeman 1980; Jakeman and Tough 1988),
accounts for certain statistical properties of K-scattering. In addition, we have
provided the extension to a complete dynamical description, in terms of contin-
uous time diffusion processes. This dynamical extension is generalized further
in Chapter 9, where we demonstrate how to include the effect of weak scatter-
ing in superposition with a coherent offset signal, in a corresponding stochastic
dynamical framework.
9
MODELS OF WEAK SCATTERING
In this chapter, we extend the K-scattering model of Chapter 8, in which the

continuous time dynamics of the K-scattering process were derived, to include
the effect of the presence of a coherent offset or ‘signal’ in the scattering ampli-
tude. The weak scattering amplitudes are characterized in terms of continuous
time biased random walk models, and the corresponding stochastic dynamics
derived. The stochastic differential geometry of the resultant amplitude fluctua-
tions is derived in relation to that of pure K-scattering. Asymptotic distributions
of amplitude, intensity and phase are provided, and the condition for detailed
balance shown to hold.
Significant progress has recently been made in our understanding of the
dynamics of models of electromagnetic scattering in the context of diffusion
processes. Deviations from Rayleigh (Gaussian) scattering have been success-
fully formulated in the context of K-distributed scattering processes (Field and
Tough 2003b) and have formed the basis of an anomaly detection technique that
has been successfully applied to maritime radar scattering and laser propagation
experiments (Field and Tough 2003a).
The models considered previously assume a uniform (asymptotic) distribu-
tion of phase. In this chapter, we consider how anisotropic phase distributions
can be accommodated within the framework provided by stochastic differen-
tial equations (SDEs) that has proved to be useful in K-scattering. We have
seen how a simple random walk model provides a physically motivated descrip-
tion of the scattering process (Jakeman and Tough 1988) that at the same time
makes useful contact with the SDE formulation of the problem (Field and Tough
2003a,b). In earlier work, un-biased random walk models have provided useful
insight into the Gaussian and non-Gaussian statistics of radiation scattered suffi-
ciently strongly for its phase to be effectively randomized, and to take a uniform
asymptotic distribution. A biased random walk model of weak scattering has
been discussed in detail in Jakeman and Tough (1987). Their analysis led to the
so-called generalized K-scattering model. The present chapter re-addresses this
problem, replacing the static, characteristic function, approach with one in which
the dynamics is captured by a set of coupled SDEs. A fairly complete analysis
is possible that also makes contact with the Rice and homodyned K descrip-
tions of weak scattering (Jakeman 1980). This allows for a detailed description
of the geometry of the resultant amplitude fluctuations, which is shown to be
different in some significant respects from that encountered in the K-distributed
case (Field and Tough 2003b). In addition to developing this SDE description,
we study the phase distributions implicit in these models in more detail than
has been reported previously.
69
70 WEAK SCATTERING
It will be necessary for the reader to review the results of Chapter 8 (orig.
Field and Tough 2003b), which are essential in the present context for the treat-
ment of weak scattering.
9.1 Weak scattering amplitudes

In situations of strong back-scattering, such as that occur e.g. in radar appli-
cations, the phases of the back-scattered components are taken to be uniformly
randomized and correspondingly the dynamics and asymptotic distributions of
the resultant amplitude process Ψt are invariant under Ψt
→ eiΛ Ψt . This is no
longer the case for ‘weak’ scattering however, i.e. situations where the Rayleigh
component of the scatter is weak in comparison to some coherent offset contribu-
tion. In these cases the mean amplitude is offset from zero, and the asymptotic
resultant phase distribution is anisotropic. We have seen in the K-distributed
1/2
case that ψt = xt γt where γ describes the (unit-power) Rayleigh process (Field
and Tough 2003b) according to (8.4). When this process lies in superposition with
a coherent offset amplitude t , the resultant amplitude process Ψt depends on
the relative scalings of the offset and (modulated) Rayleigh components with
respect to population size. There are essentially three cases to consider, each of
which can be understood in terms of the random walk model (6.16) by imposing
a bias on each step s(j) , whose physical origin is the coherent offset contribu-
tion. We shall describe these cases in the order of Rice, homodyned, and general-
ized K-scattering, thus introducing physical features (noise, K-noise, fluctuating
coherently scattered beam) in a natural order that is mirrored in the increasing
complexity of the calculations.
9.1.0.1 Rice We assume that the number of scatterers is constant in time,

with a constant offset contribution t = a. Thus modifying the random walk
model (6.16) we write
 
s(j)
N
 7 89 : 
(N )
Et = a + exp[iϕ(j) 
t ] . (9.1)

j=1
1
Scaling by 1/N , 1/N 2 for the respective terms under the summation, in the
xt -continuum limit (N → ∞) this becomes
ΨR
t = a + γt . (9.2)
9.1.0.2 Homodyned K The situation here is the same as for K-scattering with
the superposition of a constant offset t = a that does not fluctuate with Nt .
In the continuum limit this amounts to adding a constant to the K-amplitude,
thus
ΨHK
t = a + ψt . (9.3)
STOCHASTIC DYNAMICS 71
9.1.0.3 Generalized K In a similar fashion (9.1) is modified to become

 
s(j)
Nt
7 89 :
(N )
Et t = a + exp[iϕ(j) 
t ] . (9.4)

j=1
in which the t-dependence of the limit of summation is to be observed, i.e. the

step number of the biased random walk has fluctuations. Scaling by the reciprocal
mean and root mean populations respectively, the offset becomes t = axt and
we have
ΨGK
t = axt + ψt (9.5)
in the continuum limit. Observe with respect to scaling in the continuum pop-
ulation limit that, in each case, we have divided by the (unique) length scale
factors, appropriate to the relevant terms in s(j) separately, which yield finite
non-zero resultant amplitudes.
9.2 Stochastic dynamics

The stochastic dynamics of the weak scattering amplitudes described above can
be calculated from the underlying K-scattering dynamics presented in Chapter 8.
We shall make use of the identities for the (resultant) intensity and phase stochas-
tic differentials in terms of the (resultant) amplitude,
dZt ≡ Ψ∗t dΨt + Ψt dΨ∗t + |dΨt |2 ,

2 ∗ 2
1 dΨt 1 dΨt dΨt 1 dΨ∗t (9.6)

dΘt ≡ − − − ,
2i Ψt 2 Ψt Ψ∗t 2 Ψ∗t
and their products

(Ψ∗ ) (Ψ,Ψ∗ )
+ Ψ∗2
(Z) (Ψ)
Σt ≡ Ψ2t Σt t Σt + 2Zt Σt ,
∗
(Z,Θ) Ψt (Ψ)
Σt ≡ Σt ,
Ψt (9.7)

(Ψ,Ψ∗ ) (Ψ) (Ψ∗ )

(Θ) 1 2Σt Σt Σt
Σt ≡ − − .
4 Zt Ψ2t Ψ∗2
t
In combination with (9.2), (9.3), (9.5), (8.9) and the results of Corollary 8.9
these identities enable us to derive
the SDEs satisfied by Zt , Θt in terms of
(r) (θ) (x)
the component Wiener processes Wt , Wt , Wt encountered in Chapter 8.
The dynamics are simplest for Rice scattering owing to the differential of (9.2).
(In the context of radar applications, the Rice scattering model is referred to
as a ‘Swerling zero target in Rayleigh clutter’, where the ‘target’ strength is
72 WEAK SCATTERING
represented by the signal (assumed constant over the timescale of interest) and
the Rayleigh process γt represents background ‘clutter’.) More care is required
in the calculations for the homodyned and generalized K-scattering processes
owing to certain cross-terms that arise. Nevertheless the strategy is the same
for each case, and we are led to the dynamical characterizations of the vector
scattering process St = (xt , Zt , Θt )tr according to the scheme
dSti = βti dt + σti dWti (9.8)
(no summation over i) for a collection of Wiener processes {Wti | ∀i} (not nec-
essarily independent) with respective drift and diffusion coefficients β i , Σij de-
termined by
E[dSti ]
βti = ,
dt (9.9)
dSti dStj = Σij
t dt.
The corresponding Fokker–Planck equation (FPE) (e.g. Risken 1989) for the
joint probability density ρt (x, Z, Θ) is then
∂t ρ + ∂i (ρV i ) = 0 (9.10)
where the vector scattering current V i is defined by
1
V i = β i − ρ−1 ∂j (Σij ρ). (9.11)
2
9.2.0.4 Rice The amplitude dynamics of the Rice process is identical to that
of the Rayleigh process and the cross-section is constant and equal to unity, as
evident from (9.2). We deduce from the identities above that, in terms of the
geometry of the underlying Rayleigh process, the resultant intensity satisfies the
SDE
& 1
' 1
& 1 '
(r) (θ)
dZt = B 1 − ut − aut2 cos θt dt + (2B) 2 (ut2 + a cos θt )dWt − a sin θt dWt .
(9.12)
Likewise the resultant phase satisfies
1/2 B 1/2 & (r) 1/2 (θ)

'
− 12 Baut sin θt dt + 2 a sin θt dWt
+ (ut + a cos θt )dWt
dΘt = .
1/2
a2 + ut + 2aut cos θt
(9.13)

STOCHASTIC DYNAMICS 73
Proposition 9.1 The Rice vector scattering process St has drift

& 1
'
i B 1 − Z + aZ 2 cos Θ
β = 1
, (9.14)
−(Ba sin Θ)/2Z 2
and diffusion tensor

2BZ 0
Σij = . (9.15)
0 B/2Z
9.2.0.5 Homodyned K From (9.3) the amplitude dynamics is identical to that
of the K-process. Thus for the intensity, in terms of the underlying K-scattering
geometry, we find

Azt (α − xt ) 1 α − xt − 12
dZt = + B(xt − zt ) + azt cos θt A
2
− B dt
xt xt
1
& 1
' 1
(r) (θ)
+ (2Bxt ) 2 a cos θt + zt2 dWt − (2Bxt ) 2 [a sin θt ]dWt
1
2A 2 & 12 '
(x)
+ azt cos θt + zt dWt . (9.16)
xt
Likewise for the resultant phase we find
2
3
zt2 12
1 1
azt2 sin θt 2(α − xt ) − 1 1
dΘt = A − z + a cos θt − B dt
Zt 4xt 2xt Zt t 2
1 1 1
Bxt 2 a sin θt (r) Bxt 2 zt2 + a cos θt (θ)
+ dWt + dWt
2 Zt 2 Zt
1
Azt 2 a sin θt (x)
+ dWt . (9.17)
2xt Zt
Thus, in terms of the resultants (Z, Θ), we deduce that the vector homodyned
K-scattering process (xt , Zt , Θt ) has the following structure.
Proposition 9.2 The drift vector is given by
 
& A(α − x) '
A 1 1 
βi =  (α − x)Z + a x − α − 12 Z 2 cos Θ + 12 a2 + B(x − Z + aZ 2 cos Θ).
x & 1
' 
a sin Θ
1 A 2(α−x)−1
4x − 1
1 (Z 2 − a cos Θ) −
1
2 B
Z2 2xZ 2
(9.18)
The (symmetric) diffusion tensor is
 Aa sin Θ 
2Ax 2A(Z − aZ 1/2 cos Θ) Z 1/2
 & 1/2
' 
cos Θ)2
Σij =  ... 2Z A(Z −a + Bx (Aa sin Θ)(Z 1/2 − a cos Θ)/x  .
 x & 2 2 ' 
Aa sin Θ
... ... 1
2Z x + Bx
(9.19)
74 WEAK SCATTERING
9.2.0.6 Generalized K The differential of the amplitude (9.5) contains both a

K-scattering component and an explicit fluctuating part from the cross-section,
i.e. dΨt = adxt + dψt . This leads, in terms of the K-scattering geometry, to the
intensity SDE

zt (α − xt ) 1 1
dZt =A + 2a xt (α − xt + 1) + 3azt α − xt +
2 2
cos θt dt
xt 2
1
1
1

(r)
+ B xt − zt − axt zt2 cos θt dt + (2Bxt ) 2 zt2 + axt cos θt dWt

1 (θ) 1 zt 1
(x)
− (2Bxt ) 2 axt sin θt dWt + (2Axt ) 2 + 2a2 xt + 3azt2 cos θt dWt .
xt
(9.20)
Likewise for the phase we find

1 1
azt2 sin θt zt + 2a2 x2t + 3axt zt2 cos θt 1

dΘt = A + xt − α − − Bxt dt
2Zt Zt 2
1 1 1
Bxt 2 axt sin θt (r) Bxt 2 zt2 + axt cos θt (θ)
+ dWt + dWt
2 Zt 2 Zt
1
Axt zt 2 a sin θt (x)
− dWt . (9.21)
2 Zt
Combining these results we obtain the following.

Proposition 9.3 The vector generalized K-scattering process has drift
 
A(α − x)
 & ' 
β i = A Z(α/x − 1) + 12 a2 x + aZ 2 (α − x + 32 ) cos Θ + B(x − Z + aZ 2 x cos Θ)
1 1
( + , )
a sin Θ A (Z 1/2 + ax cos Θ)/Z 1/2 − (α − x + 12 ) − Bx /2Z 1/2
(9.22)
and (symmetric) diffusion tensor

 
2Ax 2A(Z + axZ 1/2 cos Θ) −Aax sin Θ/Z 1/2
Σij =  ... 2A(Z + axZ 1/2 cos Θ)2 /x + 2BxZ −Aa(Z 1/2 + ax cos Θ) sin Θ  .
... ... (B + Aa2 sin2 Θ)x/2Z
(9.23)
9.3 Geometry of amplitude fluctuations

In a similar fashion to Section 8.2, we make contact here between the stochastic
differential geometry of Chapter 3 and the detailed (amplitude) dynamics of
weak scattering processes.
AMPLITUDE FLUCTUATIONS 75
We begin with some purely geometrical results concerning the correlation

structure in the amplitude fluctuations. Combining drift terms as quantities of
1
o(dt 2 ), we write the amplitude stochastic differential as
1
dΨt = iRt exp(iΘt )dΘt + exp(iΘt )dRt + o(dt 2 )
1
= αt exp[i(Θt + φt )] + iβt exp[i(Θt + φt )] + o(dt 2 ), (9.24)
where αt , βt are real-valued Ito differentials and φt is chosen so that their Ito
product αt βt vanishes, i.e. the Wiener components of αt , βt are statistically inde-
pendent (see e.g. Karatzas and Shreve 1988). Comparing the two decompositions
1
of dΨt above, it follows that (neglecting terms of o(dt 2 ))
αt cos φt − βt sin φt = dRt

(9.25)
αt sin φt + βt cos φt = Rt dΘt .
Therefore
(αt2 − βt2 ) sin 2φt = 2Rt dRt dΘt
(9.26)
(αt2 − βt2 ) cos 2φt = dRt2 − Rt2 dΘ2t
up to o(dt). From (9.6),

& ∗ '
(α) (β) Ψt (Ψ) (Ψ∗ )
Σt − Σt 1
sin 2φt = 2i Ψt Σ t − Ψt
Ψt∗ Σ t
& ∗ ' (9.27)
(α) (β) Ψt (Ψ) (Ψ∗ )
Σ t − Σt cos 2φt = 12 Ψt Σ t + Ψt
Ψ∗ Σt .
t
Also, from (9.25) we find αt2 + βt2 = dRt2 + Rt2 dΘ2t so that
(α) (β) (Ψ,Ψ∗ )
Σt + Σt = Σt . (9.28)
We deduce from (9.26) that

<
(α) (β) (Ψ) (Ψ∗ )
Σt − Σt = ± Σ t Σt (9.29)
which, combining with (9.28), yields the following result.
Lemma 9.4
<
(α) (Ψ,Ψ∗ ) (Ψ) (Ψ∗ )
Σt = 1
2 Σt ± Σ t Σt
< (9.30)
(β) (Ψ,Ψ∗ ) (Ψ) (Ψ∗ )
Σt = 1
2 Σt ∓ Σ t Σt .
with ± corresponding to the major/minor axes of the error surface of the resul-
tant amplitude, respectively.
76 WEAK SCATTERING
Observe that (9.29), (9.30) have the appropriate symmetry under interchange
α ←→ β. The angle φt represents a rotation in the geometry of the resultant
amplitude fluctuations relative to the case of pure K-scattering, for which φ = 0.
From (9.27) this angle is determined as follows.
Lemma 9.5 The phase rotation φt , that yields an orthogonal dyad (see Fig.
9.1) associated with independent Wiener increments in the resultant amplitude
process Ψt , satisfies the geometrical identity
(Z,Θ)
4Zt Σt
tan 2φt = (Z) (Θ)
. (9.31)
Σt − 4Zt2 Σt
Equivalently, in terms of the resultant complex amplitude process, we have the
geometrical identity
+ ,
Ψ2t dΨ∗2
t
tan 2φt = − , (9.32)
[Ψ2t dΨ∗2
t ]
where , denote the real and imaginary parts, respectively.
Before applying this geometry to the weak scattering processes described
earlier, as a preliminary we give a result which provides the relationship between
the structure of the diffusion tensor that arises in the cases of homodyned and
generalized K-scattering.
Proposition 9.6 The transformation a
→ −axt maps the homodyned to the
generalized K-scattering diffusion tensor of the vector scattering process
(xt , Zt , Θt ).
Proof Choose an arbitrary instant of time, labelled t = 0. Define
(GK)
Ψt = axt + ψt
(9.33)
= −ax0 + ψt
(HK)
Ψt
for all t ≥ 0, coincident at t = 0. Thus ψ0 = 2ax0 + ψ0 and otherwise ψt ,
ψt are considered independent K-scattering processes. The
result is equivalent
(·)
to the corresponding (complex-valued) vector processes xt , Ψt having the
same diffusion tensor, at the chosen instant. The amplitude components are best
computed using the complex polarization, i.e.
2
iī dΨt dΨt dΨ∗t
Σ dt = . (9.34)
... dΨ∗2t
The results of Corollary 8.9 and the above relation between ψt , ψt at t = 0 imply
(·)2 (·)
that dΨ0 are identical. Likewise the expressions for |dΨ0 |2 coincide, by virtue

of the cosine rule applied to P RR of Fig. 9.1. The same method shows that
(·)
dxt dΨt are identical at the chosen instant. 2
The image point R has the physical interpretation of a fluctuating canceling
beam, π out of phase with the original t (R). A result corresponding to Proposi-
tion 9.6 does not hold for the vector scattering drift, as evident from comparing
Propositions 9.2 and 9.3.
AMPLITUDE FLUCTUATIONS 77
C f
C
c
Q u
R'
I
– 0 R
D D
Fig. 9.1. Geometry of fluctuations for weak scattering processes depicting

orthogonal dyad with respect to which resultant amplitude fluctuations decor-
relate.
9.3.0.7 Rice The situation here is straightforward since the resultant ampli-
tude dynamics is identical to that of Rayleigh scattering. Thus, as we have
(Z) (Θ)
seen in Proposition 9.1, the cross term Σ(Z,Θ) vanishes, so that Wt , Wt
are independent. The error surface S at P of Fig. 9.1 is circular, i.e. the ampli-
tude fluctuations are isotropic and φt can take any value (both the numerator
and denominator in (9.31) are identically zero).
9.3.0.8 Homodyned K Using Lemma 9.5 and (9.19) we find

a sin Θt
tan φt = 1/2
(9.35)
Zt − a cos Θt
and minus its reciprocal for the perpendicular φ
→ φ+π/2. Thus, in terms of the
geometry of Fig. 9.1 (e.g. by drawing a perpendicular from R to the line OP ) we
see that φt = θt − Θt (OP̂ R), i.e. the (major) axis of S HK coincides with that of
the underlying K-scattering process, as anticipated from the differential of (9.3).
The radial and angular components of the resultant amplitude fluctuations decor-
relate (the diffusion coefficient Σ(Z,Θ) of (9.19) vanishes) if the major/minor axis
of the error ellipse of the K-amplitude fluctuations is aligned with the resultant
amplitude (respectively the first/second factor in Σ(Z,Θ) vanishes). In the lat-
ter case P lies on the boundary ∂D and OP , P R are perpendicular. Inside the
domain D the error surface S rotates (anticlockwise for P shown in the upper
78 WEAK SCATTERING
half plane in Fig. 9.1) and Σ(Z,Θ) <, > 0 according as P lies in the upper/lower
half plane, while the opposite situation holds for the complement D̄.
9.3.0.9 Generalized K Perhaps the most interesting geometrical features eme-

rge for generalized K-scattering. In this case, the coherent offset (e.g. a fluctuat-
ing beam) t = axt has intrinsic fluctuations, arising from those in the scattering
population (cf. in Fig. 9.1 the boundaries ∂D and ∂D fluctuate in time). Using
Lemma 9.5, Proposition 9.6, and the homodyned result (9.35) it is immediate
that, for generalized K-scattering,
axt sin Θt
tan φt = − 1 (9.36)
2
Zt + axt cos Θt
(and minus the reciprocal). The above tangent corresponds to an axis of S GK

along R P (as seen e.g. by drawing a perpendicular from R to the continuation
in Fig. 9.1 of OP ). In contrast to homodyned K-scattering, the symmetry axes of
the error surface S GK of the resultant amplitude are no longer aligned to those of
the underlying K-amplitude. For alignment of the axes of S GK and S K to occur,
1
we require the above tangent to coincide with tan φ0 = ax sin Θ/(Z 2 − ax cos Θ),
1
or minus the reciprocal, which occurs if Z = 0 or Z 2 = ax. In the latter case
P lies on the circle C shown, consistently on which P R, P R are perpendicular.
The major axis of S GK at a general P can be identified by writing
dΨt = (At + iBt )ẑt + o(dt1/2 ), (9.37)
where At , Bt are independent real-valued (Wiener components of) Ito differ-

entials, and ẑt = (Ψt + axt )/|Ψt + axt | corresponding to a unit vector in the
axial direction R P . Then we have the squared relation dΨ2t = (A2t − Bt2 )ẑt2 .
Comparing with the expression for dΨ2t derived from Corollary 8.9 and (9.5)
we find
(A) (B) A|Ψt + axt |2

Σt − Σt = , (9.38)
2xt
so Σt ≥ Σt with equality if and only if A = 0 or Ψt + axt = 0, i.e. P = R .

(A) (B)
Strict inequality implies R P is the major axis of S GK with (circular) degeneracy

(Z,Θ)
otherwise. From the expression for Σt in (9.23), decorrelation of the radial
and angular components of Ψt occurs if P lies on the I axis or boundary ∂D .
In the latter case, R P is the major axis of S GK and OP , P R are perpendicular
(P = R ). Inside D in the upper/lower half plane, Σ(Z,Θ) >, < 0 and S rotates
with a corresponding orientation.
A measure of the total uncertainty t in the
& resultant
' amplitude Ψt is provided
(α) (β) ij
by the eigenvalue product Σt Σt = det Σt . For homodyned K-scattering,
Corollary 8.9 and (9.30) imply
ASYMPTOTIC BEHAVIOUR 79
  zt 

 7 89 :

 1 A 1/2 
(Z − 2aZt cos Θt + a2 ) +Bxt 
= 2  xt t
(α)
Σt (9.39)



1
2 Bxt
so that t = 14 B(Azt +Bx2t ). Setting A = 0, xt = 1 for Rice and using Proposition

9.6 and (9.39) for generalized K-scattering we deduce the hierarchy of increasing
(in the sense of the proliferation of terms that arise) uncertainties
1 2
 4 B (Rice)

t = 4 B(Azt + Bx2t ) (HK)
1
(9.40)

1 1/2
4 B(Azt + Bxt ) + ABZt axt cos Θt
2
(GK).
These geometrical properties of the amplitude fluctuations should provide

various means for anomaly detection, through the observability of the squared
volatilities (cf. Field and Tough 2003a and the discussion in Chapter 12) and
their departure, for a = 0, from the pure K-scattering values.
9.4 Asymptotic behaviour

The effect of the offset in the mean amplitude Ψ = 0 for a = 0 is that the
resulting (asymptotic) phase distributions are non-uniform. Expressions for these
distributions can be calculated for the various processes we have described. We
begin by deriving the joint asymptotic probability distribution functions (p.d.f.)
for the cross-section, modulus amplitude, and phase, and from these deduce the
marginal p.d.f.s of these quantities by integration.
9.4.0.10 Rice Noting that ψt is a complex Gaussian process, we see that the
familiar Rice process (Rice 1954) emerges as the model for weak scattering. If
we write the amplitude and phase of the scattered field as (E, Θ) their joint
distribution takes the form

E exp − E 2 + a2 − 2Ea cos Θ
P (E, Θ) = . (9.41)
π
From this we can derive the familiar result for the marginal p.d.f. of the field
amplitude, the Rice distribution,

P (E) = 2E exp − E 2 + a2 I0 (2Ea) , (9.42)
where I0 is the modified Bessel function of the first kind. The phase distribution
associated with the Rice scattering model can be obtained from (9.41) by inte-
gration over E. The result can be expressed in a reasonable closed form in terms
of the error function,
80 WEAK SCATTERING
PRice (冟a)
1.4
1.2
0.8
0.6
0.4
0.2

–3 –2 –1 1 2 3
Fig. 9.2. Phase p.d.f. for the Rice scattering model, a = 0, 12 , 1, 32 , 2, 52 .
∞ ∞
exp −a2 sin2 Θ 2
P (Θ|a) = P (E, Θ) dE = E exp − (E − a cos Θ) dE
π
o 0
∞
exp −a2 sin2 Θ
= (E + a cos Θ) exp −E 2 dE
π
−a cos Θ
1
a cos Θ
= exp −a2 + √ (1 + erf (a cos Θ)) exp −a2 sin2 Θ . (9.43)
2π 2 π
Figure 9.2 shows the behaviour of this function, for differing values of a, whose
square can be interpreted as a signal to noise power ratio.
9.4.0.11 Homodyned K In the case of the homodyned K-scattering process,

which can be represented through (9.3), the joint p.d.f. of the cross-section, field
amplitude and phase is
Ebα α−2 %
P (x, E, Θ) = x exp(−bx) exp − E 2 + a2 x exp(2Ea cos Θ/x) .
πΓ (α)
(9.44)
The field amplitude p.d.f. associated with the homodyned K-scattering model
cannot be rendered in a simple closed form for general values of the modified25
shape parameter α. Its compound representation takes the form
25 Modified in the sense that ‘shape parameter’ often refers to the quantity ν = α − 1, in a
radar context.
∞
2EbαE %
P (E) = xα−2 exp(−bx) exp − E 2 + a2 x I0 (2Ea/x) dx. (9.45)
Γ (α)
0
The asymptotic phase distribution for the homodyned K-scattering model can-
not be evaluated in closed form. The compound representation of the process
indicates that the phase p.d.f. can be written as
∞
bα
P (Θ|a, b, α) = P (Θ|a, x) exp(−bx) xα−1 dx, (9.46)
Γ (α)
0
where we define
2 2 2
1 −a a cos Θ a cos Θ −a sin Θ
P (Θ|a, x) = exp + √ 1 + erf √ exp .
2π x 2 πx x x
(9.47)
This can be recast in the form
P (Θ|a, b, α)
2 α/2 √ ab α2 − 14 cos Θ a2 sin2 Θ 2 + 4
α 1
=
a b
= Kα 2 ba +2 √ Kα−1/2 2 ba2 sin2 Θ
πΓ (α) πΓ (α)
α+1 α+1 1 <
a b 2 2
cos Θ
+2 Kα−1 2 ba2 sin2 Θ + t2 cos2 Θ
πΓ (α)
0
2 α−1
× sin Θ + t2 cos2 Θ 2 dt
(9.48)
by using the integral representation of the error function
√ 1
√ 2a x cos Θ
erf a x cos Θ = √ exp −s2 a2 x cos2 Θ ds. (9.49)
π
0
The representation (9.48) while a little arcane appears, when implemented

in Mathematica (Wolfram 1999), to be more stable and efficient than a direct
numerical integration of (9.46). The corresponding plots of the phase p.d.f. for
the homodyned K-scattering model are shown in Figs. 9.3 and 9.4.
9.4.0.12 Generalized K To generalize the weak scattering model to the non-
Gaussian regime we allow the number of steps in the biased random walk to
fluctuate according to (9.4). The joint distribution of the cross-section, field
amplitude, and phase now takes the form

Ebα α−2 E 2 + a2 x2
P (x, E, Θ) = x exp(−bx) exp − exp(2Ea cos Θ) .
πΓ (α) x
(9.50)
82 WEAK SCATTERING
PHK (冟1,a, a)
1.4
1.2 a = 0.1
0.8 a = 1.0
0.6
a = 10.0
0.4
0.2

–3 –2 –1 1 2 3
Fig. 9.3. Phase p.d.f.s derived from the homodyned K-scattering model,
α = 0.1, 1.0, 10.0.
PHK (冟a,1,1)
2
a = 0.5
1.5
0.5
a = 2.0

–3 –2 –1 1 2 3
Fig. 9.4. Phase p.d.f.s derived from the homodyned K-scattering model,
a = 0.5, 1.0, 1.5, 2.0.
(Here we have introduced the scale parameter b to relax the condition that
the power in the complex Ornstein–Uhlenbeck process is taken as unity.) This
provides us with the ‘compound’ representation of the generalized K-scattering
process in accordance with (8.3). This is to be contrasted with the corresponding
result for the homodyned K-scattering process above. Thus by integration we
have the field amplitude p.d.f. given by
4Ebα =
P (E) = I (2Ea) K 2E a2+b (9.51)
(α−1)/2 0 α−1
Γ (α) (a2 + b)
which is essentially the result obtained in Jakeman and Tough (1987) using the
method of characteristic functions. The calculation of the asymptotic phase dis-
tribution for Rice scattering can be extended straightforwardly to the generalized
and homodyned K-scattering models, essentially by exploiting the compound
representation (8.3). Thus using (9.47) we construct
∞
bα
P (Θ|a, b, α) = P (Θ|ax, x) exp(−bx) xα−1 dx. (9.52)
Γ (α)
0
This consists of three terms; two are straightforward while the third can be
expressed in terms of a hypergeometric function. To this end we have
∞ α
bα 1 1 b
xα−1 exp − b + a2 x dx = ,
Γ (α) 2π 2π b + a2
0
∞
bα 1 1
√ a cos Θ xα− 2 exp − b + a2 sin2 Θ x dx (9.53)
Γ (α) 2 π
0
a cos Θ Γ (α + 1/2) 1
= < 2Γ (α) √π 1 + a2 sin2 Θ%bα .
2
b + a2 sin Θ
The third term can be evaluated by substituting (9.49) and integrating over x,
thus
∞
√ √
exp − a2 sin2 Θ + b x xα−1 a x cos Θerf a x cos Θ dx
0
1
2Γ (α + 1) 2 −(α+1) (9.54)
= √ a cos2 Θ b + a2 sin2 Θ + a2 s2 cos2 Θ ds
π
0
2Γ (α + 1) a2 cos2 Θ 1 3 −a2 cos2 Θ
= √ α+1 2 F1 , α + 1; ; 2 2 .
π a2 sin2 Θ + b 2 2 a sin Θ + b
On bringing these results together, we obtain

P (Θ|a, b,
α) α
1 b a cos Θ Γ (α + 1/2) 1
= 2
+ < √ 2
% α
2π b + a b + a2 sin2 Θ 2Γ (α) π 1 + a sin Θ b
2

α a2 cos2 Θ 1 1 3 −a2 cos2 Θ
+ 2 2 % 2 F1 , α + 1; ; 2 2 .
π a sin Θ + b 1 + a2 sin2 Θ b α 2 2 (a sin Θ + b)
(9.55)
It is interesting to compare this result with that derived in an analysis of the

performance of interferometric synthetic aperture radar [eqn (53) in Tough 1991],
84 WEAK SCATTERING
PGK (冟1, a, a)
a = 10
0.5
0.4
0.3
0.2
a = 0.1
0.1

–3 –2 –1 1 2 3
Fig. 9.5. Phase p.d.f.s derived from the generalized K-scattering model,
α = 0.1, 1.0, 10.0.
to which the above result reduces when the (modified) shape parameter α takes
integer values. Figure 9.5 shows the phase p.d.f. derived from the generalized K-
scattering model. We have chosen a = 1, x = 1 and α = 0.1, 1, 10. Noise with
larger ‘spikes’, associated with lower values of α, results in a broader distribution
of phase. In Fig. 9.6 we show the variation in the phase distribution with the
parameter a, keeping the mean noise power x = 1 and α = 1. The phase
distribution becomes narrower as the parameter a increases. Comparison with
Fig. 9.2 shows that, while the mean noise power is the same in each, the more
appreciable spikes in the character of the noise is manifest in a broader phase
distribution.
The most marked difference between the phase p.d.f.s derived from the ho-
modyned and generalized K-scattering models is evident at small values of α
(i.e. less than unity), where a singular behaviour is observed at the origin. This
can be seen quite clearly in Fig. 9.4. When α takes larger values, a behaviour
more reminiscent of that seen in Fig. 9.2 emerges, as the noise becomes more
Gaussian in character. In the case where α = 1, the phase p.d.f. displays a
cusp at the origin, irrespective of the value of a; this can be seen in Fig. 9.4.
The differences between the phase p.d.f.s derived from the homodyned and gen-
eralized models can be understood qualitatively in terms of the ‘signal’ fluc-
tuating with xt in the latter, but remaining constant in the former. Jakeman
and Tough (1987) discuss the implications of this difference between the mod-
els in some detail, without making explicit reference to the asymptotic phase
p.d.f.s.
PGK (冟a,1,1)
1
a = 2.0
0.8
0.6
0.4
0.2 a = 0.5

–3 –2 –1 1 2 3
Fig. 9.6. Phase p.d.f.s derived from the generalized K-scattering model,
a = 0.5, 1.0, 1.5, 2.0.
9.4.1 Detailed balance

The detailed balance condition holds (asymptotically) for each of the weak scat-
tering processes we have described. This result follows essentially from the con-
dition being satisfied in the case of pure K-scattering, and the ways in which
the weak scattering processes can be represented as functions of an underly-
ing K-scattering process. To complete the analysis we shall need the following
result.
Lemma 9.7 If a (complex-valued/n-dimensional) Ito diffusion process xit satis-
fies detailed balance (at time t) then so does the transformed process xît = f (xit ),
i.e. (in contrast to the drift) if xit has vanishing current then so does xît .
Proof Applying Ito’s formula to the components xî we find
1
bî = Piî bi + Pijî Σij , Σîĵ = Piî Pjĵ Σij , (9.56)
2
where Piî denotes the transition matrix of partial derivatives ∂xî /∂xi , with a
corresponding notation for second derivatives. Attention should be paid to the
non-tensorial nature of the second term in the drift transformation, which is
characteristic of the Ito calculus. The probability density transforms as ρ̂J = ρ
where J is the Jacobian of P , i.e. εî1 î2 ...în J = εi1 i2 ...in Piî11 Piî22 ...Piînn . Using the
ĵ
identity ∂j log J ≡ Pĵk Pjk and the relation (9.11) we deduce the vector current
transformation V î = Piî V i , i.e. the current transforms as a tensor. The equivalent
holomorphic/anti-holomorphic representation follows, via a complex change of
coordinates. (The reader should compare orig. Field (2003) and Section 3.2 for an
86 WEAK SCATTERING
account of the geometrical structure of the transformations in drift and current

involved here.) 2
It is known that the K-scattering amplitude ψt (asymptotically) satisfies

detailed balance, as evident from the compound representation (8.3) and the
equilibrium condition (see Section 8.3.2 and also Field and Tough 2003b). As a
consequence of Lemma 9.7, the relations (9.2), (9.3), (9.5), and the independence
of the processes xt , γt we obtain the following result.
Corollary 9.8 The vector scattering processes (xt , Zt , Θt ), in each case of weak
scattering described here, namely Rice, homodyned K, and generalized K, satisfy
the condition of detailed balance, asymptotically.
Alternatively, and in rather more pedestrian fashion, detailed balance can be
shown by explicit calculation using (9.11) and the expressions for the asymptotic
distributions given above. In the presence of Doppler (Field and Tough 2003b)
detailed balance is no longer satisfied, and the current V amounts to a (rigid)
rotation of the Argand Ψ-plane, at a corresponding angular frequency.
We conclude the chapter with some general remarks. The K-distribution pro-
vides a useful model of the non-Gaussian statistics of strongly scattered radiation
with a uniform distribution of phase. In recent work (Field and Tough 2003a,b) a
description of the K-scattering process in terms of SDEs has been developed that
makes direct contact with a simple underlying random walk model of scattering.
In this chapter we have extended this analysis to models of weak scattering, in
which the distribution of phase is non-uniform.
The K-scattering process can be derived from an isotropic random walk
with a fluctuating number of steps. To incorporate a non-uniform distribution of
phase, we consider a random walk on which a preferred direction or bias has been
imposed. In the case where the random walk has a large, but fixed, number of
steps, the Rice process emerges as a model for weak scattering. We have analysed
the phase distribution associated with this model, and established the connection
between its random walk formulation and a description in terms of SDEs.
The extension of the Rice scattering model to the non-Gaussian regime is
effected when we allow the number of steps in the biased random walk to fluc-
tuate. We have shown how this leads to the generalized K-scattering process
discussed in Jakeman and Tough (1987) and have made contact between this
model and the homodyned K-scattering process. In each case we have char-
acterized the associated distribution of phase in detail, and have developed a
description in terms of SDEs and their equivalent FPEs. This complements the
earlier dynamical description of K-scattering (Field and Tough 2003a,b).
The results of this chapter have implications for detection schemes where the
signal behaviour (represented by the coherent offset in the resultant amplitude)
can, to a reasonable extent, be modelled in the context of ambient K-distributed
noise (cf. the results reported in §4 of Field and Tough 2003a). The results of
Section 9.3 indicate a method for anomaly detection based on departures in
the geometry of the resultant amplitude fluctuations from that expected in the
pure K-scattering case. The results should find application in adaptive imaging
problems, in the de-noising of optical images (signal separation from noise, i.e.
extraction of t from Ψt ) and anomaly detection in radar back-scatter where a
(coherent) reflection contribution is involved (cf. Jakeman and Tough 1987).
10
SCATTERING FROM GENERAL POPULATIONS
In this chapter we derive the stochastic dynamics for scattering of a wavelike field
from a large population of scatterers whose dynamics is arbitrary. This leads to
a result concerning the observability of the scattering cross-section in terms of
the resultant phase fluctuations that is independent of the population dynamics.
An emergent concept is a certain notion of an ideal filter. The diffusion-based
model of K-scattering arises, encountered in Chapter 8, as a special case. The
experimental implications of the results in a variety of contexts are discussed
later in Part III.
Motivated by the possible application of the recent results on electromagnetic
scattering from random media to more general situations (e.g. medical imaging,
wireless communications) than those encompassed by the K-distribution model
(Field and Tough 2003a,b), the current chapter focuses on a significant new re-
sult in connection with inference of the scattering cross-section or population,
in local time. A special case of the result was reported previously in Field and
Tough (2003b) [see paragraph above eqn (2.35) therein], stating that the instan-
taneous values of the cross-section are deducible through the phase fluctuations
in the scattered field. This was demonstrated theoretically in the context of K-
scattering. Intriguingly, the same result holds for an arbitrary population. More
precisely, given the structure of a random walk model, component phase, and
step number fluctuations, the result holds for an arbitrary specification of popu-
lation dynamics. In this sense, the result is a geometrical feature of (the dynam-
ical extension of) Jakeman’s random walk model with step number fluctuations
(Jakeman 1980), and as such should apply to a large number of experimental
situations involving interference effects of wavelike fields arising from random
populations. From a filtering point of view, the result represents an improve-
ment on Kalman/particle filtering methods, since an exact expression for the
‘hidden’ state (the population level) in terms of the additional phase degrees of
freedom can be derived.
10.1 Extended random walk model

It is well known that Rayleigh scattering can be described by a random walk
model for the scattered field amplitude (cf. Jakeman 1980; Tough 1987; Jakeman
and Tough 1988) with a fixed number of steps. The pertinent expression for the
88
EXTENDED RANDOM WALK 89
resultant field amplitude according to this model is
s(j)
N 7 89
& ':
(N )
(j)
Et = aj exp iϕt (10.1)
j=1
with (constant) population size N , random phasor step s(j) . Notice here, for the
purpose of generality, the inclusion of the component amplitude ‘form factors’ aj ,
i.e. the strength of the individual scattering components may vary throughout
the medium. Also in this chapter we shall allow for the possibility that the
component phases ϕ(j) are correlated for short times, as may occur by virtue of
some special initial condition. As such, we refer to this random walk construction
as the ‘extended random walk model’ by comparison with the simpler model of
the previous chapter.
In the aforementioned references, all components of this expression are in
effect considered at a given instant of time, thus not addressing the question
of continuous time evolution properties or ‘dynamics’. This extra structure is
supplied by a (phase) diffusion model (Field and Tough 2003b) which takes
(j)
the component phases {ϕt } to be a collection of (displaced) Wiener processes
(j) 1 (j)
evolving on a suitable timescale. Thus ϕt = ∆(j) + B 2 Wt , with component
(j)
random initialization ∆ uniformly distributed on the interval [0, 2π). In situa-
tions where the component phasors s(j) are aligned initially (e.g. for the received
‘T 2’ signal arising in magnetic resonance imaging, described e.g. in Ernst et al.
1987 and discussed in Chapter 14) (cf. also Field and Bain 2008), the initial-
izations ∆(j) are identical for all j, whereas in some cases (e.g. the statistical
description of radar scattering from the sea surface, cf. §4(b) in Field and Tough
2003a) it is more appropriate to draw ∆(j) independently. In any case, these
primitive assumptions enable us to derive the dynamics of Rayleigh scattering,
essentially from first principles. The stochastic differential of (10.1), according
to Ito’s formula (e.g. Oksendal 1998; Karatzas and Shreve 1988), is given by
N
& '
(N ) (j) 1 (j) 2 (j)
dEt = aj idϕt − dϕt exp iϕt . (10.2)
j=1
2
If we write dζt for the first term on the right-hand side above, then for t ≥ T ,
(j)
where T is the ‘phase decoherence’ time such that {ϕt | t ≥ T } have negligible
1/2
correlation, we have |dζt |2 = j aj Bdt, and therefore dζt =
2 2
j aj B 1/2 dξt
where ξt is a complex-valued Wiener process (satisfying |dξt |2&= dt, dξt2' = 0).
(N ) 1
Defining the (normalized) Rayleigh amplitude by γt = limN →∞ Et /N 2 leads
to the resultant dynamics (cf. Field and Tough 2003b).
Proposition 10.1 For sufficiently large times t ≥ T the dynamics of Rayleigh
scattering is given by the complex Ornstein–Uhlenbeck equation
90 GENERAL POPULATIONS
1 1 1
dγt = − Bγt dt + B 2 a2 2 dξt . (10.3)
2
If {∆(j) } are assumed independent then the result holds for arbitrarily small
times.
Remarks. On the asymptotic distribution of phase. If the phases are initialized
at some value, say the real direction in the complex plane, then the distribution
of the resultant Rayleigh amplitude is Gaussian, of non-zero mean, by virtue of
the central limit theorem applied to the i.i.d. collection of random phasors (with
a square root scaling). Observe that, for small times, the major axis of its error
surface S – an ellipse – is oriented in the direction of the imaginary axis, since
the magnitude of the fluctuations (for each component phasor) in that direction
is dominant. As relaxation occurs, i.e. the asymptotic equilibrium distribution is
approached, the surface S stretches in the real direction, and tends to a circular
geometry in the limit t → ∞. At the same time, the mean value – geometrically
the centre of gravity of the ellipse – tends to the origin of the complex plane.
Accordingly, the resultant asymptotic distribution of phase is uniform.
If we re-scale the (Rayleigh) amplitude according to γt
→ a2 −1/2 γt , then
the re-scaled field satisfies (10.3) with the form factors equal to unity. In what
follows we shall therefore assume the field to be scaled in this way, i.e. a2 =
1. In the case of a fluctuating number of steps N
→ N + t in (10.1),
, we define
the (continuous-valued) cross-section as xt = limNt →∞ Nt /N̄ . The resultant
+ ,
(normalized) amplitude ψt = limN →∞ E (Nt ) /N̄ therefore has the compound
representation
1
ψt = xt2 γt , (10.4)
& 1
'
(N )
where γt = limN →∞ Et t /Nt2 , and in which xt and γt are independent pro-
cesses. The intensity zt has the compound representation zt = xt ut , where
ut = |γt |2 is the instantaneous intensity of the component (unit power) Rayleigh
process (cf. the analysis of asymptotic behaviour and propagators in §s III, IV
of Field and Tough 2003b).
It is worth clarifying at this point the precise meaning and definitions of
the various amplitudes that have occurred in the exposition of the random walk
model. We begin with E, as the superposition of N random phasors, where N is
fixed. For a large population, N → ∞, and the root mean square (r.m.s.) of E
tends to infinity. Thus, to obtain a finite resultant in the limit of an asymptot-
ically large number of scatterers, we define a ‘normalized’ Rayleigh amplitude
1
γ, by dividing through by the r.m.s. value N 2 . (Equivalently, we could absorb
this normalization into the form factors aj .) The term ‘Rayleigh’ refers to the
fact that the number of scatterers is fixed. In the general case that the scatter-
ing population fluctuates in time, we define the ‘normalized’ amplitude ψ as in
1
the case of Rayleigh scattering, dividing E by the r.m.s. value N̄ 2 , where now
GENERALIZED DYNAMICS 91
the number of terms Nt in the random walk fluctuates. Re-arranging the result-
ing expression produces the compound representation of the resultant amplitude
(10.4), for a general scattering process. The Rayleigh amplitude is then recovered
if the cross-section is unity.
A corresponding dynamical situation for ‘weak’ scattering processes, i.e. where
the field ψt lies in (weak) superposition with a coherent offset signal t , should
be possible for a general type population as a generalization of the results de-
scribed in Chapter 9 (orig. Field and Tough 2005), although this extension is
not addressed in the current monograph.
10.2 Generalized dynamics

In this section we propose a general scheme for describing scattering/interference
of wavelike fields from random media, for which the statistical characteristics of
the (scattering) population are taken to be arbitrary. With regard to the cross-
section and intensity variables, in ‘Bayesian’ terms one may write P(x|z) ∝
P(z|x)P(x) and interpret P(z|x) as the ‘likelihood’ function L, P(x) as the
‘prior’ P and P(x|z) as the ‘posterior’ distribution. In K-scattering, the con-
stant (with respect to x) of proportionality is the reciprocal K-distribution for
the intensity z. Our development shall entail the following, that we first preserve
the likelihood L as the Rayleigh distribution, cf. the universality (under appro-
priate conditions) of arguments of central limit theorem type, second modify the
prior P , i.e. consider general (endogenous) population dynamics appropriate to
more complex population processes, under the assumption that xt remains an
Ito process (cf. Field and Tough 2003a), and third preserve the mathematical
structure of the random walk model (10.1) describing the resultant amplitude
process.
Accordingly, we specify that the underlying ‘signal’ xt is an Ito process that
satisfies the generalized (in the sense of its relationship with the K-scattering
model) stochastic differential equation (SDE)
1 (x)
dxt = Abt dt + (2AΣt ) 2 dWt (10.5)
in which the drift and diffusion parameters bt , Σt are, respectively, (real-valued)
(x)
stochastic processes, not necessarily Ito processes, adapted to the filtration Ft
(x)
corresponding to the Wiener process Wt . In other words, the continuous pop-
ulation dynamics is taken to lie within the general category of Ito processes. The
special case of a diffusive population behaviour arises when the SDE parameters
are functions of state, i.e. bt = b(t, xt ) and Σt = Σ(t, xt ) for given functions
b(·, ·) and Σ(·, ·), in which case a corresponding Fokker–Planck description for
the time evolution of the probability density is possible (e.g. Risken 1989). The
case of K-scattering (a special type of diffusion model) is obtained by setting
b(t, x) = (α − x), Σ(t, x) = x, and arises as the continuous-valued (large N )
limit of the birth–death–immigration (BDI) model (see Bartlett 1966; Field and
Tough 2003b). We shall not require that xt be a diffusion in what follows, how-
ever. The generalized dynamics of the resultant amplitude process can now be
derived according to the scheme outlined at the end of the previous section. For
arbitrary γt , xt an application of Ito’s formula to (10.4) yields
dψt dγt dxt dx2

= + − 2t . (10.6)
ψt γt 2xt 8xt
This enables the resultant amplitude dynamics to be calculated under the as-
sumption that γt is a unit power Rayleigh process according to (10.3), with unit
form factors.
Proposition 10.2 The generalized resultant amplitude dynamics is given by
1 1
dψt bt Σt 1 AΣt 2 (x) B2
= A − 2 − B dt + dWt + dξt . (10.7)
ψt 2xt 4xt 2 2x2t γt
Observe that ∂/∂B acting on the drift/volatility parameters in (10.7) yields ex-
pressions that are independent of bt , Σt , as expected from the endogenous spec-
ification of population dynamics (10.5). Using the vanishing of the Ito products
(x)
dξt2 , dξt dWt , and the property |dξt |2 = dt, the above result yields the squared
amplitude fluctuations as follows.
Corollary 10.3
2
dψt AΣt
= dt, (10.8)
ψt 2x2t

AΣt zt
|dψt | =
2
+ Bxt dt. (10.9)
2x2t
The generalized intensity dynamics can be computed from Proposition 10.2 and
the identity dzt ≡ ψt dψt∗ + ψt∗ dψt + dψt dψt∗ .
Proposition 10.4 The generalized intensity SDE is given by

bt zt 1 zt (x) 1 (r)
dzt = A + B(xt − zt ) dt + (2AΣt ) 2 dWt + (2Bxt zt ) 2 dWt
xt xt
(10.10)
where
12
2zt
(γt∗ dξt + γt dξt∗ ) ≡
(r)
dWt . (10.11)
xt
The intensity squared volatility is

AΣt zt
dzt2 = 2zt + Bxt dt. (10.12)
x2t
The propositions above reduce to the appropriate expressions in K-scattering

(cf. Field and Tough 2003a,b) for appropriate choice of b, Σ. In respect of the
generalized resultant phase dynamics, recall (Field and Tough 2003b) the identity
for the phase differential in terms of the amplitude
2
dψt 1 dψt
dθt ≡ − , (10.13)
ψt 2 ψt
where denotes the imaginary part. Since the right-hand side of (10.8) is real-
valued, only the first term on the right-hand side of (10.13) contributes to dθt ,
in respect of which
1
dψt B 2 dξt dξt∗
= − ∗ . (10.14)
ψt 2i γt γt
Thus we can deduce the phase behaviour for a general population.

Proposition 10.5 The generalized resultant phase dynamics is given by the
SDE
12
Bxt (θ)
dθt = dWt (10.15)
2zt
(θ)
where Wt satisfies
12
1 ∗ 2zt
(θ)
dWt . (10.16)
i t xt
Observe that, in contrast to the situation for the resultant amplitude and inten-
sity SDEs (10.7) and (10.10), this is functionally identical to the corresponding
result in K-scattering (i.e. independent of the population parameters bt , Σt ),
the essential difference lying in the evolutionary structure of the processes xt , zt .
Observe from (10.11), (10.16) that the radial and angular fluctuations in the re-
sultant amplitude are statistically independent in the general case. The squared
phase volatility obtained from (10.15) leads to the central result of the chapter.
Theorem 10.6 The instantaneous values of the scattering cross-section are ob-
servable through the intensity-weighted squared phase fluctuations according to
2 dθt2
xt = zt (10.17)
B dt
if xt is an Ito process, not necessarily a diffusion, and throughout space and time.
The (experimental) significance of this result is that the relation (10.17) is
exact and moreover independent of the dynamics of xt . The result resembles (but
is distinct from) the minimal variance of the intensity-weighted phase derivative
Fig. 10.1. Geometry of random walk for generalized scattering process – gener-
ically (a.e.) each component phasor point lies on two circles.
discussed in Jakeman et al. (2001), for differentiable processes. In the present sit-
uation however, the processes considered are not differentiable, and instead the
squared phase differential arises. Since the elements of the random walk model
(10.1), preserved by the generalized framework given in the exposition of this
section for dealing with general populations, are the only essential ingredients
involved here, the result is geometrical in nature (cf. Fig. 10.1). With regard to
this geometry, the derivations above show that the dynamics of the resultant
field are not affected if the radii for each component phasor are drawn indepen-
dently from an arbitrary probability distribution. (The result of the theorem was
anticipated from a physical point of view previously in Field and Tough 2003a;
see discussion following Prop. 4.1. therein.)
A slight complication is posed in the computation of dθt2 from experimental
data, owing to the discontinuous-valued behaviour of θt at coordinate intervals
of 2π. This is resolved by instead using the (continuous-valued) ‘phase-wrapped’
process wt = exp(iθt ), whose stochastic differential is dwt = exp(iθt )[idθt − 12 dθt2 ],
which enables the squared phase fluctuations to be computed from the single-
valued process wt via |dwt |2 = dθt2 . In respect of discrete-time implementation,
we remark that if Wt is a Wiener process, then δWt = Wt+h − Wt is normally
distributed as N (0, h), so that its square is a ‘chi-squared’ χ2 (1) variable. The
sum of n such variables is therefore distributed as χ2 (n), from which an esti-
mate of dqt2 from δqt can be obtained (via the weak law of large numbers) by
considering the interval from t to t + δt divided into n ‘pulse’ intervals each of

length h and letting n → ∞ before taking the limit δt → dt (Howison 2004; see
also Higham 2004 and the exposition to §4 in Field and Tough 2003a). In this
respect we observe the following significant result.
Proposition 10.7 In order to achieve an improved signal to noise ratio (SNR),
it is sufficient merely to increase the pulse-rate n, without necessarily requiring
a high amplitude signal xt .
The implications of this idea are discussed in Chapter 14 in an experimen-
tal context, and illustrated with some synthetically generated data. The reader
should compare the remarks in Section 14.1 for issues concerning discrete sam-
pling and bounded pulse rate, and the optimal choice of a smoothing parameter
that effects the result of the theorem above. The structure of (10.5), (10.10)
could be interpreted as an instance of the ‘generalized’ Kalman filter (see Ch.
6 in Oksendal 1998) in which the unknown state xt is to be estimated from
observations of zt . It is significant that in this situation the dynamics of the fil-
ter stem from first principles and that the resulting statistics are non-Gaussian
(notwithstanding the Gaussian nature of the Wiener process). The noise origi-
(x)
nates through two components, namely the ‘intrinsic’ system noise Wt which
derives from fluctuations in the (endogenously specified) population model, and
the ‘measurement’ noise ξt arising from the particulars of the wavelike interfer-
ence effects. The latter should be viewed as an exogenous ‘device’ whose purpose
is to probe the true underlying state of the system that is of primary interest,
in this case the ‘signal’ xt .
Our development has demonstrated (an instance of) how, instead of attempt-
ing to filter the received signal to eliminate the noise (e.g. via a Kalman or par-
ticle filter), one can exploit the statistical fluctuation properties of the noise to
infer the exact values of the underlying signal. This notion might appropriately
be termed an ideal filter and constitutes a shift of viewpoint from the various
conventional approaches to enhancement of signal to noise. We discuss the ex-
perimental implications of this framework in greater detail in Chapter 14.
Finally, we remark in proof that stochastic models of the kind discussed in this
part have recently been applied successfully to modelling of the wireless channel
in communications, in which the latter can be viewed as a multi-path propagation
problem consisting of a combination of phase fluctuations and Doppler effects
for each component path. The theoretical results predict modifications to the
channel spectra that are observed experimentally, and a corresponding state-
space model can be constructed (Feng et al. 2007; Feng and Field 2008).
Part III
Simulation and experiment
In this part we develop simulation techniques for scattering processes and de-
scribe some experimental tests that verify our theoretical development. This de-
velops the interrelationships between theoretical prediction, real, and simulated
data, and is illustrated in a specific case of scattering from a general population.
Chapter 11 provides a detailed account of the simulation of K-scattering
processes, including the influence of Doppler, target returns, and improved ac-
curacy second-order algorithms for numerical integration of the gamma process.
In Chapter 12 we provide two independent experimental tests of the proposed
stochastic theory, and provide a scheme for anomaly detection. The first in Sec-
tion 12.1 is a study of data gathered from electromagnetic scattering from a
random phase screen, in the absence of anomalies. A study is made of the (for-
ward) scattered intensity which is observed to be very closely K-distributed. The
volatility function of the intensity is measured and observed to correlate very
strongly with the instantaneous square root intensity, thus providing a calibra-
tion of the stochastic model. This calibration setting is applied to an independent
set of (complex-valued) radar return data in Section 12.2. The radar scattering
is from a region of the ocean surface in which a tethered anomalous reflecting
body is placed and whose location is known. It is previously known that such data
typically contains a large number of spurious ‘clutter spike’ anomalies and that
the phase behaviour is an important indicator of a genuine anomaly (Luttrell
2001). Accordingly, we derive a complex-valued stochastic differential equation
(SDE) for the amplitude of the scattered radiation, under the assumption that
the phase behaviour is uniform. We exhibit a strong correlation between the ob-
served versus the theoretical prediction of the volatility for the complex-valued
process in the clutter domain. At the location of the anomaly this correlation be-
comes weak, which provides a means of detection, isolating the anomaly to a high
degree of accuracy. With the theory firmly established from these experiments,
Chapter 13 provides a discussion of the non-linear character of the stochastic
dynamics of radar sea clutter, from independent experimental and theoretical
points of view.
The part concludes in Chapter 14 with some simulations that illustrate the
concepts of scattering from general populations as developed earlier in Chapter
10. The implications of these results for tracking populations in various physical
contexts are discussed.
97
11
SIMULATION OF K-SCATTERING
We have seen in Chapter 8 that the stochastic differential equations (SDEs)

arising in the description of scattering generically contain Wiener fluctuating
terms whose noise power is a function of the state of the system, i.e. the system
is said to have ‘multiplicative noise’ or, in other words, it falls under the category
of a ‘stochastic volatility’ model. In this chapter we develop the tools necessary
for the numerical integration/simulation of such systems in a fairly complete
general context. From this analysis the case of additive noise, for which the noise
power is independent of the state of the system, will emerge as a special case.
This framework is applied to Rayleigh and gamma processes that are of special
interest in K-scattering. We examine in detail the effect of Doppler for discrete
time-series data and describe the simulation of a target in a clutter environment.
The chapter ends with an investigation of the refinement in terms of second-order
algorithms for the gamma process.26
11.1 Iterative schemes

Consider a differential equation expressed in the Langevin form
dx
= F (x) + σ (x) f (t) (11.1)
dt
in which f (t) is a white noise process, whose integral is the Wiener process. The
interpretation of this equation has, in the statistical physics literature, been the
source of considerable controversy and dispute. It is no longer possible to as-
sociate a Langevin equation of this form unambiguously with a Fokker–Planck
equation (FPE), without additional specification of how the fluctuating terms
undergo integration.27 Pusey and Tough (1982) have discussed the iterative
solution of equations of this type in some detail, in such a way that the ambigui-
ties are avoided. Indeed in the mathematical and engineering literature, the util-
ity of such equations has long been recognized, and considerable effort has been
expended in accommodating them within a consistent and rigourous calculus.
The analysis given by Pusey and Tough ibid. produces results that are in
accordance with this Ito, as opposed to Stratonovich, calculus, which is the
convention followed throughout this book (cf. Appendix B). The resulting theory
26 The author acknowledges the contribution of Dr. Robert Tough with regard to the majority
of material in this chapter.

27 Indeed some workers, such as Van Kampen (1981), have characterized such equations as
being devoid of meaning.
99
100 K-SCATTERING SIMULATION
of SDEs has found widespread application in signal processing, control theory

and mathematical finance (e.g. Oksendal 1998).
dxt = F (xt ) dt + σ (xt ) dWt (11.2)
which is to be understood in the Ito interpretation, as developed in Chapter 2.

We now distinguish between integration with respect to time and with respect
to the Brownian measure and write the associated integral equation in the form
t t
∆xt = xt − x0 = F (xu ) du + σ (xu ) dWu . (11.3)
0 0
It is natural to develop the drift F and volatility σ in Taylor series, which yields
F (n) σ (n)
t t
n n
∆xt = (∆xu ) du + (∆xu ) dWu . (11.4)
n
n! n
n!
0 0
A solution to this is now developed via iteration. We calculate terms up to O(t2 );

Pusey and Tough ibid. carried through their analysis to this level in the multi-
plicative noise case. Taking only the n = 0 contributions gives us the following
terms
∆xt = F t + σWt . (11.5)
This simple stochastic analogue of the Euler integration algorithm forms the
basis of many widely used Brownian dynamics algorithms that incorporate hy-
drodynamic interactions (Ermak and McCammon 1978). Taking the first four
terms in the Taylor expansions in (11.4) gives us
t t t
(1) 1 2
∆xt = F t + σWt + F (1)
∆xu du+σ ∆xu dWu + F (2) (∆xu ) du
2
0 0 0
t t t
1 2 1 3 1 3
+ σ (2) (∆xu ) dWu + F (3) (∆xu ) du+ σ (3) (∆xu ) dWu + · · · .
2 6 6
0 0 0
(11.6)
We consider the orders of the various terms generated

√ by this process, and make
use of the properties F t = O(t), Wt = O( t). Integration with respect to u
raises the order by one while integration with respect to Wu raises it by one half.
To generate terms of order two, we must therefore keep terms up to order t
in the integrand of
t
F (1) ∆xu du (11.7)
0
ITERATIVE SCHEMES 101
while terms up to O(t3/2 ) must be retained in the integrand of
t
(1)
σ ∆xu dWu . (11.8)
0
Not all these are provided by the lowest order solution (11.5) when it is intro-
duced into (11.1). Nonetheless, it does generate terms of orders t, t3/2 , t2 . These
in turn should be re-inserted into (11.1) to produce extra terms in the iteration.
Thus, from (11.5) alone, we have
t2
∆xt = F t + σWt + F F (1)
2
t t
3/2 (1)
(1)
+F σ Wu du O t + Fσ udWu O t3/2
0 0
t
+σ (1) σ Wu dWu (O (t))
(11.9)
0
t t
1 1
+ F (2) σ Wu2 du O t2 + σ (2) σ 2 Wu2 dWu O t3/2
2 2
0 0
t t
σ (3) σ 3
+σ (2)
σF uWu dWu O t2 + Wu3 dWu O t2 .
6
0 0
Here the order of each term in t is included in parentheses, where it is not

obvious. Terms that are of higher order in t than the second have been omitted.
By inspection we see that the term
t
(1)
σ σ Wu dWu (11.10)
0
should be incorporated into the integrand in (11.7) and in that of
t
1 (2) 2
σ (∆xu ) dWu . (11.11)
2
0
There they give rise to terms of the form
t u
F (1) σ (1) σ du Ws dWs (11.12)
0 0
and
t u
σ (2) σ (1) σ 2 dWu Wu dWs Ws . (11.13)
0 0
2
Each of these is of O(t ) and so need be iterated no more.
We now turn our attention to (11.8); including all terms up to O(t3/2 ) ob-
tained thus far we generate the following new terms
t u t u
(1) 2 (1) (1)
σ σ dWu Ws dWs + σ F σ dWu Ws ds
0 0 0 0
t u t u
1
+ σ (1) F σ dWu s dWs + σ (1) σ (2) σ 2 dWu Ws2 dWs . (11.14)
2
0 0 0 0
The last three of these terms are of order t ; the first, however, is O(t3/2 ) and 2
so should be iterated once more. This leads us to a contribution of the form

3 t u s
(1)
σ σ dWu dWs dWv Wv . (11.15)
0 0 0
This process of successive re-substitution enables us to generate all the terms of

a required order in a systematic fashion; as a result we obtain a solution correct
to O(t2 ) of the form:
2 t t t
(1) t (1) (1) (1)
∆xt = F t + σWt + F F +F σ Wu du + F σ udWu + σ σ Wu dWu
2
0 0 0
t t t
1 1
+ F (2) σ 2 Wu2 du + σ (2) σ 2 Wu2 dWu +σ (2)
Fσ uWu dWu
2 2
0 0 0
t u t u
(1) (1) (2) (1) 2
+F σ σ du Ws dWs + σ σ σ dWu Wu dWs Ws
0 0 0 0
t 2 t u
t u
σ (3) σ 3 3 (1) (1) (1)
+ Wu dWu + σ σ dWu Ws dWs + σ F σ dWu Ws ds
6
0 0 0 0 0
t u t u
2 1
+ σ (1) F dWu sdWs + σ (1) σ (2) σ 2 dWu Ws2 dWs
2
0 0 0 0
t u s
3
+ σ (1) σ dWu dWs dWv Wv . (11.16)
0 0 0
ITERATIVE SCHEMES 103
This process could be carried on to higher order; the number of terms rapidly
becomes very large. Perhaps the most striking qualitative difference between this
and the corresponding result obtained in the additive noise case is the presence
of a term with a linear time dependence in the stochastic part of the increment.
Now that we have developed an iterative solution we calculate the expecta-
tion value of the increment in x occurring over the time t. All terms in (11.16)
containing odd powers of W average to zero. Among those containing even pow-
ers of W the majority of the expectation values go to zero as a consequence of
the Ito rule (2.6)
 t 
u
E  dWu dWs g (u, s) = 0. (11.17)
0 0
For example, we consider the simple case where g is unity:
t u t
Wt2 − t
dWu dWs = dWu Wu = ; E Wt2 = t,
2
0 0 0 t  (11.18)
u
E  dWu dWs  = 0.
0 0
Here we observe the occurrence of an extra term arising in the integration by

parts within the Ito calculus; it is this term that ensures that the expectation
value of the iterated integral is indeed zero. Examination of each of the terms
occurring in (11.16) shows that
t
(1) t 12
E (∆xt ) = F t + F F + F (2) σ E Wu2 du
2 2
0
t2 1
= F t + F F (1) + F (2) σt2 . (11.19)
2 4
This is identical with the corresponding result obtained in the additive noise
case, to second order in time.
We now wish to evaluate the mean square of the stochastic part of the incre-
ment, i.e.
t2
∆x̂t = ∆xt − F t − F F (1) (11.20)
2
again retaining only terms up to second order in t. (We have introduced the caret
notation because superscript notation has already been used in this case.) Thus
we need only consider the terms in (11.16) up to and including those that are
O(t3/2 ),
t t
(1) (1)
∆x̂t =σWt + F σ Wu du + F σ u dWu
0 0
t t
(1) 1
+σ σ Wu dWu + σ (2) σ Wu2 dWu + O t2 . (11.21)
2
0 0
From this we see that

 2 
t
2  
E  Wu dWu  
2
E (∆x̂t ) = σ 2 E Wt + σ (1) σ 2
0
 
t
+ 2F (1) σ 2 E Wt Wu du
0
   
t t

+ 2F σ (1) σE Wt u dWu  + σ (2) σ 2 E Wt Wu2 dWu  + O t2 .
0 0
(11.22)
The various expectation values can be evaluated, using the Ito calculus. We recall
the familiar relation E[Wt2 ] = t. The Ito isometry (see Section 2.1.1) allows us
to evaluate the next contribution:
 2 
t t
  t2
E Wu dWu   = E Wu2 du = . (11.23)
2
0 0
The remaining averages can be obtained as follows

 
t t t
t2
E Wt Wu du = E (Wt Wu ) du = u du = , (11.24)
2
0 0 0
    
t t
t2 t2
EWt u dWu  = EWt tWt − Wu du = t2 − = , (11.25)
2 2
0 0
     
t t
W3 t2
EWt Wu2 dWu  = EWt  t − Wu du = . (11.26)
3 2
0 0
RAYLEIGH, GAMMA PROCESSES 105
In the final derivation we have exploited the Gaussian factorization property of

> ? > ?2
the Brownian process Wt4 = 3 Wt2 , and thus re-expressed the integral over
the Wiener measure.28 Altogether we find that
2 2
2 t
E (∆x̂t ) = σ 2 t + σ (1) σ + 2F (1) σ 2 + 2F σσ (1) + σ (2) σ 3 . (11.27)
2
We treat the corresponding analysis for a vector process in Appendix G.
11.2 Rayleigh and gamma processes

Having generated an iterative solution to a generic SDE with multiplicative noise,
we now apply these results to SDEs describing processes of particular interest in
the study of non-Gaussian scattering processes.
First, we recall the Rayleigh process, described by the scalar SDE
√
dzt = (1 − z) dt + 2zt dWt . (11.28)
The relationship between this SDE and other descriptions of the intensity of a
Gaussian speckle process is discussed in Tough (1987). To make contact with the
general results above, we make the identifications


 F (z0 ) = (1 − z0 ) ,

F (1) (z0 ) = −1,


 (2)
F (z0 ) = 0
 √ (11.29)

 σ (z0 ) = 2z< 0,
 (1)
σ (z0 ) = 2 z20 ,
1
 <

 σ (2) (z0 ) = − 1 23 .
4 z 0
When these are substituted into (11.19) we find that
t2 1
E (∆z) = F t + F F (1) + F (2) σt2 + O t3
2 4

(11.30)
t
= (1 − z0 ) t 1 − + O t3 .
2
The mean square value of the stochastic part of the increment in z can be
evaluated as follows
28 This can be derived by calculating the stochastic differential d(W 4 ), integrating, and taking
t
the expected value.
2 2
2 t
E (∆ẑ) = σ 2 t + σ (1) σ + 2F (1) σ 2 + 2F σσ (1) + σ (2) σ 3 + O t3
2
3
= 2z0 t + (1 − 3z0 ) t + O t .
2
(11.31)
In the special case of the Rayleigh process described by (11.28) further
progress can be made, albeit by eschewing the iterative approach. Instead, we
identify the FPE that is stochastically equivalent to this SDE and solve explicitly
for its propagator. Using this we can evaluate expectations of the form
+ ,
E [(zt − zo )] , E (zt − zo )2 (11.32)
for arbitrary time intervals t. By making a suitable small time expansion we
can check the validity of our more general analysis in this case. The statistics
of this increment are of interest in the context of improving the precision in
the stochastic volatility analyses of Chapter 12; the results below that hold for
arbitrary times identify precision corrections to the O(t) treatment that we have
in practice applied to the experimental data analysis, which serves to illustrate
the theory sufficiently well. However, in situations that present themselves where
the sample time is not as small as one would like, the higher order correction
analysis we provide here should be useful.
We recall that the FPE equivalent to (11.28) is
∂2 ∂ ∂
(zP (z, t|z0 )) + ((z − 1) P (z, t|z0 )) = P (z, t|z0 ) (11.33)
∂z 2 ∂z ∂t
whose fundamental solution or propagator satisfies the initial condition
P (z, 0|z0 ) = δ (z − z0 ) . (11.34)
(Further background information on this topic can be found in Tough 1987;
our notation follows that of this reference quite closely.) The propagator can be
expanded in terms of the Laguerre polynomials as
∞

P (z, t|z0 ) = exp(−z) Ln (z) Ln (z0 ) exp(−nt) (11.35)
n=0
n
1 d
Ln (z) = exp(z) (exp(−z) z n ) .
n! dz
Using this expansion, and some basic properties of these polynomials, we find
that
∞
E (zt − z0 ) = (z − z0 )P (z, t|z0 ) dz
0
= (1 − z0 ) (1 − exp (−t))

t2 t3
= (1 − z0 ) t − + + O t4 (11.36)
2 6
RAYLEIGH, GAMMA PROCESSES 107
∞
2 2
E (zt − z0 ) = (z − z0 ) P (z, t|z0 ) dz
0
2 2
= (1 − exp(−t)) + z0 (4 exp(−t) − 2) (1 − exp(−t)) + 2 (1 − exp(−t))
z02

13
= 2z0 t + 2 − 5z0 + z0 t − 2 − z0 + z0 t3 + O t4 .
2 2 2
(11.37)
3
Higher order moments can, in principle, be determined from the characteristic
function, which is most conveniently derived using an alternative form for the
propagator
√
1 (z + z0 θ) 2 θzz0
P (z, t|z0 ) = exp − I0 ,
1−θ 1−θ 1−θ
θ = exp(−t) . (11.38)
Here I0 is a zeroth-order modified Bessel function of the first kind. Using this
we find that
∞ (−s)n E ((z − z )n )
t 0
E(exp(−s (zt − z0 ))) =
n=0 n!
∞ √
1 θ 1 2 θzz0
= exp z0 s + exp −z s + I0 dz
1−θ 1−θ 1−θ 1−θ
0

sz0 (1 + s) (1 − θ)
exp
1 + s (1 − θ)
= . (11.39)
1 + s (1 − θ)
Expanding this characteristic function in a power series and equating coefficients
of powers of s verifies the results (11.36) and (11.37). These in turn allow us to
verify the results obtained to order t2 by iteration. In particular, the mean square
of the stochastic part of the increment can be calculated as

2 2
E ∆ẑ 2 = E (zt − z0 ) − (1 − z0 ) t2 + O t3

= 2z0 t + (1 − 3z0 ) t2 + O t3 . (11.40)
The results we have presented in this section demonstrate the application of

the general formulae derived earlier, and provide us with some further insights
into the SDE description of the K-scattering process. In the special case of the
Rayleigh process, we have been able to calculate equivalent quantities in more
detail by making recourse to the closed form solution of the corresponding FPE
that is available in this case. Limiting forms of these results provide verification
of our more general iterative analysis. In Section 11.3, we discuss briefly how
the results we have derived for the K-scattering process can be exploited in a
numerical simulation algorithm.
So far in this chapter we have examined the iterative solutions of some SDEs
that occur in the description of clutter processes. We have considered the case
where the SDEs incorporates multiplicative noise, and so has been interpreted
within the Ito convention. A corresponding simpler treatment holds for the case
of iterative solutions of ordinary differential equations and Langevin equations
incorporating additive noise. The greater complexity of the multiplicative noise
calculations has limited us to results valid up to second order only, whereas the
corresponding Langevin case would naturally yields results valid up to third order
in time. This fact highlights some of the problems encountered in the analysis of
systems with multiplicative, as opposed to additive, noise.
The general results obtained have been specialized to the relatively simple
SDEs that describe the Rayleigh and K-scattering processes. As these incorpo-
rate multiplicative noise, we have generated results valid up to second order in
time. Recent experimental work (Field and Tough 2003a) has focussed attention
of the behaviour of increments in the intensity of the Rayleigh and K processes
that occur over time intervals that are very short in comparison to the correlation
timescale of the process concerned. Accordingly, this analysis can be carried out
using a first-order expansion in time. The results we present here are useful in
assessing the validity of this approximation in cases where the relevant sample
time interval, for example as dictated by experimental design, is a significant
proportion of the correlation time. In the special case of the Rayleigh process
we are able to make significantly more progress, by adopting the Fokker–Planck
formulation of the problem. The results we derive specialize to provide a verifi-
cation of the analysis for more general processes, and should be of relevance to
analyses of empirical data with shorter correlation timescales and constraints on
the available sample frequency ranges.
In addition to providing some insight into the temporal development of
the system described by SDEs incorporating multiplicative noise, the results
of this iterative analysis also allow us to simulate their behaviour numerically, in
effect providing numerical solution to the SDEs concerned. One way to achieve
this, which has been applied to Langevin type equations (Greenside and Helfand
1981) is very much in the spirit of the Runge–Kutta algorithm, frequently ap-
plied to the solution of ordinary differential equations. These workers develop
a method of stepping forward in time, combining several contiguous values of
the dependent variables in such a way that agreement with a Taylor series de-
velopment of the solution is maintained to some stated order in the time step
(Whittaker and Watson 1969). When applying this method to SDEs, both the
deterministic and stochastic parts of the solution have to be controlled. Given the
difficulties encountered in characterizing the stochastic part of the solution, it is
reasonable to require agreement to second order in the time increment; typically
Runge–Kutta algorithms applied to deterministic equations seek agreement to
fourth or even higher order. We shall apply these methods to the solution of the
SDEs discussed in Section 11.3, and so provide a more accurate method for the
simulation of K-scattering processes.
COMPOUND K 109
11.3 Compound K-model

In Sections 11.1 and 11.2, we described how stochastic processes described by
SDEs with multiplicative noise can be simulated by matching the mean and
mean square increments generated in a Runge–Kutta-like algorithm with those
obtained from a systematic iterative solution of the SDE.The computational

2
complexity restricts the accuracy of these algorithms to O (∆t) . Thus, in a
situation where two coupled processes evolve with widely differing de-correlation
for A B in Theorem 8.1) the algorithm will be accurate only
times (e.g.
2 3
to O (∆t) , even though A3 (∆t) and higher order terms may be negligibly
small. In the case of the K-scattering model of Chapter 8, in a radar context the
so-called ‘compound K model’, it is possible to overcome this limitation.
As the compound form represents the K-scattering process as a product of
two independent processes, described as in Section 8.1, one might suppose that
the characteristic time of the Rayleigh process (i.e. the reciprocal of the constant
B) would define the appropriate integration time step. However, it proves to be
possible to propagate the Rayleigh process along over arbitrary time steps, using
the following artifice. We have seen that the Rayleigh process of unit power is
obtained by summing the squares of two component Ornstein–Uhlenbeck pro-
cesses. Each of these is described by a linear SDE and so can be integrated
explicitly over any time step. Thus it is only the slower gamma modulation
process
that
limits the step size, as the overall algorithm has an accuracy of
2
O (A∆t) .
The coupled SDEs we solve to simulate the complex Ornstein–Uhlenbeck
process are
dp = − 12 pdt − ωqdt + √1 dWp (t) ,
2
(11.41)
dq = − 12 qdt + ωpdt + √1 dWq (t) .
2
We can write these as a linear matrix equation

d(exp(At) · u) = √1 exp(At) dB (t) ,
2
p
u= ,
q

1/2 ω (11.42)
A= = 12 1 + ωP,
−ω 1/2

Wp (t)
B (t) = .
Wq (t)
This we can integrate to yield
t
u (t) = exp(−At) · u (0) + exp(−A (t − t )) · dB (t ) . (11.43)
0
The mean of this process is given by

u (t) = exp(−At) · u (0) . (11.44)
The exponential matrix can be evaluated, exploiting the trivial commutation of
1 and P and P’s identification as the matrix square root of the identity matrix,
P · P = −1. Thus we find that

cos ωt sin ωt
exp(At) = exp 2t ,
−sin ωt cos ωt
cos ωt − sin ωt (11.45)
(exp(At)) = exp 2t
T
.
sin ωt cos ωt
We can now evaluate the covariance matrix of the process as follows:
@ A
T
(u (t) − u (t)) (u (t) − u (t))
- .
1 t t T
= (exp(−A (t − t1 ) · dB (t1 ))) (exp(−A (t − t2 ) · dB (t2 )))
2 0 0
1 t T
= exp(−A (t − t1 )) · (exp(−A (t − t1 ))) dt1
20
1 t 1
= 1 exp(−t1 ) dt1 = (1 − exp(−t)) 1. (11.46)
2 0 2
Here we have exploited the Ito isometry (see Section 2.1.1) to evaluate the expec-
tation. Using these results, we can step the complex Ornstein–Uhlenbeck (COU)
process through arbitrary time intervals via

p (tn ) − (tn − tn−1 ) cos (ω (tn − tn−1 )) − sin (ω (tn − tn−1 ))
= exp
q (tn ) 2 sin (ω (tn − tn−1 )) cos (ω (tn − tn−1 ))
<
p (tn−1 ) gp
× + (1−exp(−(t2n −tn−1 ))) . (11.47)
q (tn−1 ) gq
in which gp , gq are independent Gaussian random variables drawn from a dis-
tribution with zero mean and a unit variance. This representation of the COU
process can be used to generate a unit mean Rayleigh intensity process through
2 2
u (tn ) = p (tn ) + q (tn ) . (11.48)
The Doppler frequency can be suppressed if we wish only to generate the inten-
sity; in these circumstances we have
<
p (tn ) = exp −(tn −t
2
n−1 )
p (tn−1 ) + (1−exp(−(t n −tn−1 )))
2gp ,
< (11.49)
q (tn ) = exp −(tn −t
2
n−1 )
q (tn−1 ) + (1−exp(−(tn −tn−1 )))
2gq .
This unit power Rayleigh process can be simulated over arbitrary time steps. To
yield a K-scattering process it should be multiplied by the correlated gamma
INFLUENCE OF DOPPLER 111
process,
generated
by the solution of (8.9) which will typically have accuracy
2
to O (A∆t) characteristic of the entire compound process. Observe that the
procedure we have outlined here provides us with a relatively efficient method
for the simulation of the complex amplitude of the scattered field, arising as the
product according to (8.3).
In this section we have seen that the compound representation of the K-

scattering process also facilitates its (computer) simulation. In particular, it is
possible to construct an algorithm whose accuracy depends only on the ratio
of the time step of numerical integration to the long time characteristic of the
decay of the correlation with respect to the constituent gamma process.
11.4 Influence of Doppler on volatilities

The anomaly detection method described in Chapter 12 is based on the determi-
nation of the volatility of the complex field, based on the measurement change in
the returned complex field over short periods of time. In the radar data analysed
thus far this time step is determined by the pulse repetition frequency, and is
of the order of 1 ms. The returned field develops through diffusive and Doppler
effects; it is the former that is characterized by the volatility. For the determina-
tion of the volatility to be accurate, Doppler effects should be negligible over the
timescale of the measurement. We will now consider whether this is the case.
To compare the effects of Doppler and diffusion we consider the simple model
of a COU process, as discussed in Section 8.1. The underlying SDE is
4
1 β (p)
dp = − βpdt − ωqdt + dWt ,
2 2
4 (11.50)
1 β (q)
dq = − βqdt + ωpdt + dWt .
2 2
The SDE associated with the intensity z = p2 + q 2 is obtained using the usual
rules of Ito calculus as
= (z)
dz = β (1 − z) dt + 2βzdWt . (11.51)
This is stochastically equivalent to the FPE
∂P (z, t) ∂ ∂2
=β ((z − 1) P (z, t)) + β 2 (zP (z, t)) (11.52)
∂t ∂z ∂z
which has the stationary solution P (z) = exp (−z). The correlation function of
the intensity implicit in this model exhibits a simple exponential decay, falling
off as exp(−βt). Because the SDEs for the COU process are linear they can be
integrated over an arbitrary time step, as explained earlier. Thus we have
p0 = p (0), q0 = q (0) ;
∆p = p (t) − p (0); ∆q = q (t) − q (0)

∆p cos ωt − sin ωt 10 p0
= exp(−βt/2) −
∆q sin ωt cos ωt 01 q0
4
(1 − exp (−βt)) gp
+ . (11.53)
2 gq
The g-functions are drawn independently from a zero mean, unit variance Gaus-
sian distribution. These results form the basis of a simple simulation of coherent
clutter, which is discussed in more detail in the next section.
To establish the behaviour of the square modulus @ of the increment
A in the
2 2
complex field Ψ = p + iq, we make an expansion of (∆p) + (∆q) ; here the
angular bracket denotes an averaging over the gs. In this way, we find that
@ A
2 2
(∆p) + (∆q) = 1 − exp(−βt) + p20 + q02 (1 + exp(−βt)
−2 exp(−βt/2) cos ωt) . (11.54)
We can expand this up to second order in t as
@ A
2 2 1 z0 ω 2 z0
(∆p) + (∆q) ∼ βt + β t − +
2 2
+ 2 ,
2 4 β (11.55)
2 2
z0 = p0 + q0 .
To ascertain the effects of the Doppler evolution on the measurement of the
volatility in the complex field (which we can see from (11.55) to be equal to β),
we must introduce estimates of the various terms in this expression.
An independent analysis of the data presented in Section 12.2 (Ward 2002),
for which the time step t is known to be 1 ms, indicates that the intensity
correlation function decays in the order of 10 ms; thus we approximate β ≈ 100.
The Doppler frequency is identified as 100 Hz, i.e. an angular frequency ω =
200π. When these values are introduced into (11.55) we find that
|∆Ψ|2
βt = 1 + 0.1 − 12 + 4π 2 + 14 z0 (11.56)
≈ 1 + 4z0 .
From this we see that, for the data analysed in Field and Tough (2003a) and
Section 12.2, Doppler induced effects are comparable with the diffusive evolution
of the complex field and so can influence the determination of the volatility. It
is particularly striking that the intensity dependence of the Doppler term is the
same as that interpreted as a contribution to the volatility of the type previously
accounted for by the endogenous model. The interaction of the diffusive, Doppler,
and power modulation terms and their effects on the performance of the anomaly
detection can best be assessed by simulation studies. Methods with which these
may be implemented will be discussed in the following.
COHERENT CLUTTER 113
11.5 Coherent clutter

In the previous section, we analysed the COU process to gain some insight into
the procedure adopted in the measurement of volatilities. We will now show
how the COU process provides us with a useful simulation tool that allows us
to reproduce many of the features observed in the data analysis underpinning
the volatility-based anomaly detector. When combined multiplicatively with the
square root of a gamma process, whose simulation is described in the follow-
ing section, the COU and related processes should provide us with a valuable
simulation tool for the investigation of the performance of the volatility based
anomaly detector described in Section 12.2.
The simulation procedure implicit in (11.53) is trivial to implement and rep-
resents the extension to the complex domain of a method described, for example,
by Pike and co-workers (Hughes et al. 1973) some thirty or more years ago.29 The
effect of a Doppler term on the evolution of the returned field, represented as a
complex number in the Argand plane, can be illustrated as follows. Figures 11.1–
11.3 show this evolution for zero, small and significant Doppler respectively; in
the last of these the relative contributions of the diffusive and Doppler effects
are much the same as in the data analysis discussed in the previous section. The
qualitative differences between these plots are quite evident.
We now assess the extent to which the complex field volatility, measured
over a single time step, tracks the intensity of the field. In Fig. 11.4, we display
traces of the intensity (solid) and measured field volatility (dotted) derived from
simulated COU data for which β = 1.0, ω = 6.0, ∆t = 0.1. The plots have
been scaled so that the range of each is the same; a strong correlation between
the intensity and measured field volatility is evident in this plot. In Fig. 11.5,
however, we see that this correlation is greatly reduced when the COU process
has no Doppler component. This simple simulation, for typical radar parameter
values, highlights the care that is needed in the interpretation of measured field
amplitude volatilities, in that the terms of O(δt2 ) may have a significant con-
tribution. Thus the results of Section 12.2 depend on the presence of Doppler
for such high degrees of correlation in the clutter domain. Measurements of the
intensity volatility can be analysed similarly. Figure 11.6 shows the correlation
between the intensity (solid) and measured intensity volatility (dotted) derived
from simulated COU data for which there is no Doppler component. Once again
we have scaled the plots to facilitate their comparison. Thus, unlike the case of
amplitude volatilities above, the situation as regards correlation for the intensity
volatility is unaffected by Doppler, as anticipated since the Doppler transforma-
tion, consisting of rigid rotations of the Argand plane representing the amplitude,
leaves the intensity process invariant.30
29 A Mathematica notebook that implements this and other simulation procedures is available
from the author.

30 This fact can be readily established from computing the intensity SDE from the amplitude
SDE containing the Doppler term as provided by (8.60).

1.5
0.5
–2 –1.5 –0.5 0.5 1 1.5
–1.5
Fig. 11.1. Realization of a COU process; β = 1.0, ω = 0, ∆t = 0.1.
11.5.1 Target returns in clutter

A simple model of a ‘target plus clutter’ (radar) return can be constructed by
adding a coherent signal of constant amplitude and prescribed Doppler frequency
directly to the COU process. It is interesting to note that a simple FFT displays
even a fairly weak target signal quite prominently, as long as there is a sig-
nificant difference between the Doppler frequencies characteristic of the clutter
and target. The introduction of fluctuations into the target return amplitude,
perhaps modelled by a gamma or similar process, might yield a more realistic
model. Figure 11.7 shows an (I, Q) plot of this simple target plus clutter model;
the Fourier transform of this data is shown in Fig. 11.8, and demonstrates the
efficacy of simple Doppler processing in this case.
In this section we have assembled the components of a simulation capability
that should allow us to assess the performance of the volatility-based anomaly
detector described in Section 12.2 [orig. Field and Tough (2003a)]. An analysis
of the impact of the Doppler characteristics of the returned signal on the mea-
surement of the field volatility indicates that some care will be necessary in the
separation and identification of the processes involved. This also highlights the
need for further controlled simulation and real-data-based studies of the anomaly
detector, to establish its robustness and general applicability.
Some simple preliminary simulations based on the COU process confirm our
findings on the influence of Doppler on measured volatilities. The incorporation
SECOND-ORDER ALGORITHMS 115
–2 –0.5 0.5 1.5
–2
Fig. 11.2. Realization of a COU process; β = 1.0, ω = 1.0, ∆t = 0.1.
of a coherent target return into this model is straightforward, and allows us to

demonstrate the effectiveness of simple FFT processing in isolating the target
when its Doppler characteristics differ significantly from those of the clutter.
11.6 Second-order algorithms

The endogenous model of the clutter process represents the returned field as
the product of a COU process with a stochastically varying amplitude, i.e. the
square root of a gamma process. Thus, to complete our simulation capability,
1.5
0.5
1.5 1 1 1.5
–1.5
–2
Fig. 11.3. Realization of a COU process; β = 1.0, ω = 6.0, ∆t = 0.1.
1.5
0.5
20 40 60 80 100
Fig. 11.4. Comparison of the intensity (solid) and measured field volatility (dot-
ted) of a COU process with β = 1.0, ω = 6.0, ∆t = 0.1.
20 40 60 80 100
Fig. 11.5. Comparison of the intensity (solid) and measured field volatility (dot-
ted) of a COU process with β = 1.0, ω = 0, ∆t = 0.1.
1.5
0.5
20 40 60 80 100
Fig. 11.6. Comparison of the intensity (solid) and measured intensity volatility
(dotted) of a COU process with β = 1.0, ω = 0, ∆t = 0.1.
we must devise an effective method for the simulation of such a process. As we

have discussed previously in Section 8.1, the temporal evolution of a gamma
process xt is described by the SDE (8.9). We observe that the volatility term
depends on the current value of xt , i.e., it is an instantaneous stochastic volatil-
ity.31 Closed form solutions of SDEs of this type are not generally available; the
initial temporal evolution of the process can, nonetheless, be characterized by
a systematic iteration of the SDE, re-expressed in an integral form. A heuristic
development of this solution was given by Pusey and Tough (1982), in a dis-
cussion presented in the context of the Brownian dynamics of hydrodynamically
interacting particles. A more rigourous and detailed analysis of the problem was
31 The reader should compare the discussion of models in mathematical finance in
Appendix D.
–2 –1.5 –1 –0.5 0.5 1 1.5
–1
–2
Fig. 11.7. An IQ plot of a target plus clutter return; the target power is one
quarter that of the clutter, while the target Doppler is twice that of the
clutter.
given by Klauder and Petersen (1985), whose results have been re-derived re-
cently by Quiang and Habib (2000). In essence, these workers determined the
mean and variance of the increment in a process described by a SDE, to second
order in the time step and expressed in terms of the drift and volatility functions
and their derivatives.
A simple simulation of the gamma process proceeds as follows. Integrating
(8.9) we obtain the simple recursion
=
x (n + 1) = x (n) + A (ν − x (n)) ∆t + 2Ax (n) ∆tg (n) . (11.57)
Here ∆t is the time step over which the process evolves and the g(n) are indepen-
dent Gaussian distributed random variables, with zero mean and unit variance.
This simple procedure reproduces the mean and variance of the increment in the
process to linear order in ∆t, and is frequently referred to as an Euler algorithm.
50 100 150 200 250
Fig. 11.8. Amplitude of Fourier transform of the preceding data, demonstrating

the isolation of the coherent target from the clutter.
(Much of the greatest part of the solution of SDEs by numerical simulation is

based on this method, as in Quiang and Habib.) Increments in x whose mean
and variance are correct to quadratic order in ∆t could be constructed using the
algorithms presented in the latter two references. Klauder and Petersen display
considerable ingenuity to obviate the need for evaluation of derivatives of the
drift and volatility functions, basing their algorithm on the classic Runge–Kutta
algorithm for the solution of ordinary differential equations (e.g. Fox and Mayers
1968). However, in the case of the gamma process these drift and volatility func-
tions are particularly simple; thus it is possible to characterize the increments
in the process directly, rather than by iteration. In this special case, we can
therefore produce an algorithm, accurate to second order in ∆t, that is almost
as simple to implement as the Euler algorithm (11.57).
Recall from Chapter 3 that the FPE stochastically equivalent to (8.9) is
∂ ∂ ∂2
P (x, t|x0 ) = ((x − ν) P (x, t|x0 )) + 2 (xP (x, t|x0 )) . (11.58)
A∂t ∂x ∂x
The propagator of this can be constructed in a reasonable closed form32 as
(ν−1)/2
1 x exp (At) (x + x0 exp (−At))
P (x, t|x0 ) = exp −
1 − exp (−At) x0 1 − exp (−At)
√
2 exp (−At/2) xx0
× Iν−1 . (11.59)
1 − exp (−At)
In the absence of the stochastic driving term, (8.9) reduces to
32 Here I
ν−1 is a modified Bessel function of the first kind (see Section 9.6 in Abramowitz
and Stegun 1970).
dxD
= A (ν − xD ) (11.60)
dt
which can be solved to yield
xD = ν − (ν − x0 ) η,
(11.61)
xD ≡ xD (t) , x0 ≡ x (0) , η = exp(−At) .
Here, the subscript D denotes the deterministic part of x. We can characterize
the stochastic part of the development of x by evaluating the generating function
∞
C (s) = exp(−s (x − xD )) P (x, t|x0 ) dx. (11.62)
0
A development of C(s) in powers of s then allows us to identify the required

expectation values
∞ n
(−1) n n
C (s) = s (x − xD ) . (11.63)
n=0
n!
Recalling that η = exp(−At), and on substituting (11.59) into (11.62), we find

that
exp ((1 − η) s (ν + ηsx0 /(1 + (1 − η) s)))
C (s) = ν . (11.64)
(1 + (1 − η) s)
Equation (11.64) can now be expanded as a power series in s to give
2
C (s) = 1 + 1
2 (1 − η) ((1 − η) ν + 2ηx0 ) s2 − 1
3 (1 − η) ((1 − η) ν + 3ηx0 ) s3 + · · · .
(11.65)
We note that the term linear in s vanishes; the mean value of the stochastic
part of x vanishes identically, as expected. We can now read off the following:
@ (x − xD )A = 0,
2
(x − xD ) = (1 − η) [(1 − η) ν + 2ηx0 ] , (11.66)
@ A
3 2
(x − xD ) = 2 (1 − η) [(1 − η) ν + 3ηx0 ] .
Finally, we introduce the above expression for η into these results, and expand
up to cubic order in the time step, which leads us to
(x − xD ) = 0,
@ A
2 2 7x0 3
(x − xD ) = 2x0 At + (ν − 3x0 ) (At) + − ν (At) + · · · ,
3
@ A
3 2 3
(x − xD ) = 6x0 (At) + 2 (6x0 − ν) (At) + · · · . (11.67)
It is interesting to note that the mean cube of the stochastic part of x has a
quadratic time contribution; in earlier work (Klauder and Petersen 1985; Quiang
and Habib 2000; Pusey and Tough 1982) this term has not been analysed explic-
itly for SDEs with multiplicative noise, and has been assumed to be third order
in time.
If we now expand xD to second order in time as
xD = x0 + (ν − x0 ) At (1 − At/2) (11.68)
we see that we can construct an algorithm similar in structure to (11.57) that
nonetheless generates increments whose mean and variance are maintained to
quadratic order in the time step:

A∆t
x (n + 1) = x (n) + A (ν − x (n)) ∆t 1 −
2 (11.69)
<
2
+ 2Ax (n) ∆t + (ν − 3x (n)) (A∆t) g (n) .
One of the problems faced when generating a gamma Markov process by
simulation of the solution of the SDE (8.9) is that presented by a particularly
large random term causing x to take a negative value precluded by the natural
barrier inherent in the SDE itself. This potential inconsistency can be removed
by working with the logarithm of x, which can legitimately take negative values.
It is straightforward to construct the SDE satisfied by y = log x; application of
Ito’s formula yields

dy 1 d2 y 2
dy = dx + (dx)
dx 2 dx2
1 √ 1 1
= (ν − x) dt + 2x dW − · 2x dt
x 4 2 x2

ν−1 2
= − 1 dt + dW
x x

√ −y
= ((ν − 1) exp (−y) − 1) dt + 2 exp dW. (11.70)
2
While this identification of the appropriate SDE is sufficient for us to construct an
Euler type algorithm, essentially by inspection, we might wish to include terms
of higher order in the time step, much as in the direct simulation of the gamma
variate itself. To calculate the required higher order statistics of the increment
in the log gamma process we again make use of the analytic form (11.59) of the
propagator of the gamma Markov process. The generating function approach
adopted earlier has to be modified slightly. An appropriate moment generating
function can be constructed as follows:
s
x
exp(s (log (x) − log (x0 ))) =
- s . x 0
x 2
@ A
= 1 + s (log (x) − log (x0 )) + s2 (log (x) − log (x0 )) · · · .
2
x0
(11.71)
We evaluate this function (essentially the Mellin transform of the propagator

(11.59)) as
- s . ∞
x s
= (x/x0 ) P (x, t|x0 ) dx
x0
0
∞
1 1 x0 η 1
= exp − xs+(ν−1)/2
1 − η (x0 η)(ν−1)/2 1 − η xs0
0
√
x 2 xx 0 η
× exp − Iν−1 dx. (11.72)
1−η 1−η
The integration over the modified Bessel function can be effected analytically
(eqn 11.4.28 in Abramowitz and Stegun 1970) to give us
- s . s
x (1 − η) ηx0 Γ (ν + s) x0 η
= exp − 1 F1 ν + s; ν; . (11.73)
x0 xs0 1−η Γ (ν) 1−η
Here we have encountered the confluent hypergeometric function 1 F1 (Chapter

13 ibid.). In the construction of the simulation algorithm we focus attention on
the short time behaviour of this function, where η → 1. Thus the standard series
representation of the confluent hypergeometric function (eqn 13.1.2 ibid.) is not
particularly useful in this context. Fortunately, (11.73) can be recast using the
asymptotic representation
Γ(b) exp(z) −1

1 F1 (a : b : z) ∼ b−a 2 F0 b − a, 1 − a; z ; z → ∞; (z) > 0
Γ(a)z
(11.74)
developed by Copson (1970). The term 2 F0 is the generalized hypergeometric

function, whose formal series expansion is
∞
1 Γ (a + n) Γ (b + n) n
2 F0 (a, b; q) = q . (11.75)
Γ (a) Γ (b) n=0 n!
Using this we find that

- s .
x s 1−η
∼ η 2 F0 −s, 1 − ν − s; , η → 1. (11.76)
x0 x0 η
Finally, we introduce the exponentially decaying correlation function
η = exp(−A∆t) . (11.77)
We now make an expansion in both s and ∆t from which we can extract the
mean and mean square values of the increment in the logarithm. Thus we have
- s .
x ν−1 3ν − ν 2 − 2 (ν − 1) 2
=1+s − 1 A∆t + + (A∆t) · · ·
x0 x0 x20 x0
2
A∆t ν − 5ν + 5 3 − 2ν 1 2
+s2
+ + + (A∆t) · · · + · · · .
x0 2x20 2x0 2
(11.78)
From this we can read off

ν−1 3ν − ν 2 − 2 (ν − 1) 2
(log (x) − log (x0 )) = − 1 A∆t + + (A∆t)
x0 x20 x0
@ A
2 A∆t ν 2 − 5ν + 5 3 − 2ν 2
(log (x) − log (x0 )) =2 + + + 1 (A∆t)
x0 x20 x0
(11.79)
and calculate the variance of the logarithmic increment as
@ A A∆t 4 − 3ν + x0
2 2 2
(log (x) − log (x0 )) − (log (x) − log (x0 )) = 2 + (A∆t) .
x0 x20
(11.80)
The linear terms in time in the mean and variance of the increment are those that
one would expect from (11.70) and might incorporate into an Euler algorithm.
The quadratic correction terms can be accounted for just as was done in (11.69)
for the case of the gamma, as opposed to the log-gamma, process. In a final
step we regenerate the gamma process by exponentiation. Thus, the log-based
algorithm can be summarized as follows:



 − 1) + ((ν
y (n) = y (n −1) exp(−y (n − 1))− 1) A∆t

 3ν − ν 2
− 2 exp(−2y (n − 1))

 + (A∆t)
2

 + (ν − 1) exp(−y (n − 1))
 B
(4 − 3ν) exp(−2y (n−1)) 2

 + 2A∆t exp(−y (n−1)) + (A∆t) g (n)

 + exp(−y (n−1))






x (n) = exp(y (n)) .
(11.81)
The final component of the clutter simulation within the compound model is
the generation of a gamma process, described by a SDE. Simple Euler algorithms
are almost invariably used to effect such simulations. Algorithms of higher order
accuracy have been discussed in the literature; their generality unfortunately
renders them rather forbidding. However, in the case of the gamma process,
significant analytic progress has been made and simple algorithms, of an accuracy
equivalent to that of these more general methods, have been constructed for the
generation of the gamma variate and its logarithm.
12
EXPERIMENTAL TESTS
In this chapter we study two independent examples of data gathered from elec-
tromagnetic scattering from random media. The first example involves scattering
from a uniform randomizing optical phase screen, which is applied as a calibra-
tion tool for the intensity volatility in our diffusive model of K-distributed noise.
The second example consists of ocean surface radar scattering data that contains
K-distributed sea clutter, a multitude of ‘clutter spike’ anomalies, and a genuine
target for which our methodology is used to isolate the anomaly with respect to
range.
In both cases we shall study the electromagnetic scattered amplitude stochas-
tic processes and compare the empirically observed volatilities with the theoreti-
cal predictions from our stochastic model. In the absence of anomalies we expect
a strong correlation between these two quantities, via the standard statistical
correlation measure
(X − X̄)(Y − Ȳ )
c(X, Y ) = . (12.1)
(VarX VarY )1/2
Observe that this standard measure, which satisfies −1 ≤ c(X, Y ) ≤ 1, is zero for
independent variables, unity for ‘perfectly correlated’ variables, and is invariant
under (positive) affine transformations of the random variables X
→ λX + µ,
λ > 0.
Since the data were supplied in discrete time, an important issue arises con-
cerning the instantaneous observability of the stochastic volatility explained in
Lemma 5.4. Although the relations (5.24), (5.25) are exact within the framework
of the Ito calculus, these differential relations are idealizations of the incremen-
tal properties that can be physically measured. In practice, physical observations
occur at a discrete set of times and, in accordance with the fact that the physi-
cal processes exist in continuous time, the significance of the Ito calculus is that
sample times remain unspecified. For discrete time sampling of an Ito process
qt , satisfying the stochastic differential equation (SDE) dqt = µt dt + σt dWt , we
may write
δqt = µt δt + σt nt δt1/2 + o(δt), (12.2)
where nt is drawn from a normal distribution with zero mean and unit variance
N (0, 1), so n2t = 1. It follows that δqt2 = σt2 n2t δt + o(δt). Given the Ito inter-
pretation of the (discrete) SDE (12.2), the independent increments property of
Wt implies that σt and δWt = nt δt1/2 are independent (this independence also
124
OPTICAL WAVELENGTH 125
holds if σt is non-Markov). Therefore, taking expectations ·, at fixed t with

respect to the random variable nt , we find33
σt2 = δqt2 /δt + o(1). (12.3)
If qt is a diffusion then σt = σ(t, qt ), so that δqt2 = σ(t, qt )2 n2t δt + o(δt). Strictly,

the expectation · above should be interpreted as an ensemble average, for a
given t, qt . Nevertheless, since the sample paths of qt are continuous with unit
probability, a local time approximation to δqt2 is obtained by smoothing the raw
δqt2 values by averaging over a time window [t−∆, t+∆]. This approximation may
be improved upon (O’Loghlen 2001) by ordering the qt values before smoothing
the time series for δqt2 . More precisely, we can permute the discrete sample times
{ti }
→ {p(ti )} such that {qp(ti ) } is an ordered set, then evaluate the expectation
δqp(t
2
i)
by smoothing, and apply the inverse permutation to the result. This
method provides an improved empirical estimate of dqt2 /dt, to be compared with
the theoretical prediction σ(t, qt ), the latter of which is instantaneously deducible
from the instantaneous observed values {qt }.
12.1 Scattering at optical wavelength

The dynamical framework of Chapter 8 applies equally well to forward-and back-
scattered radiation. Moreover, it is an exact description for all parts of the elec-
tromagnetic spectrum, the carrier frequency being represented by the constant
B. In this section we shall consider an optical forward scattering experiment.
The same type of analysis would apply to many situations of optical scattering,
e.g. propagation of laser light through turbulent atmospheric conditions, and in
optical instrumentation in theoretical astronomy.
12.1.1 Random phase screen

The data consisted of a complex-valued time series (sampled at 10 kHz) for the
I, Q components of the electromagnetic field (forward) scattered from a ran-
domizing phase screen. The details of the experimental situation are described
in Ridley et al. (2002) and the resulting intensity data were observed to be very
closely K-distributed, as shown in Fig. 12.1. The squared volatility function ψ
of (5.11) was assumed to have power law behaviour with respect to the intensity
(cf. the discussion following (5.19), (5.20) above) and to deduce the appropriate
power a comparison was made of |δzt |p and zt using the smoothing technique
above to calculate the expectation ·. The results are shown in Fig. 12.2 and
indicate a strong correlation between these quantities according to the measure
(12.1), with a maximum occurring at p ≈ 2. A comparison of the time series for
this value of p is shown in Fig. 12.3. The analysis demonstrates that the squared
33 We assume that the data are sampled at a sufficiently high frequency for the o(1) term in
(12.3) to be negligible so that δt1/2 δt and the δWt contribution to (12.2) dominates the
drift over a sample period, as evidenced, for example, in Fig. 12.3.
126 EXPERIMENTAL TESTS
–2
–4
Logarithmic probability distribution
–6
–8
–10
–12
–14
–16
0 5 10 15
Normalized intensity
Fig. 12.1. Random phase screen data, the logarithmic probability distribution
for the normalized forward-scattered intensity, showing comparison of the
nearest fit theoretical K-distribution (solid line) and the experimentally ob-
served histogram. (The discrepancy in the tail of the K-distribution is well
known and a more refined model would be necessary to model such extreme
intensities.)
volatility function ψ is to a close approximation equal to the intensity, which

accords with the theoretical discussion of Section 8.1.
Remarks. On the suppression of fluctuating terms. Observe that the time

series behaviour for this stochastic dynamics is radically different from what
would be obtained from suppressing the fluctuating term in the underlying in-
tensity SDE. Since the drift and volatility coefficients are functions of the state
(the intensity itself), an abscissa of constant intensity will intersect such (deter-
ministic) path at most once; however, this geometrical construct yields multiple
crossings for the fluctuating time series, as readily apparent from Fig. 12.3.
12.2 Scattering at radar wavelength

In this section we provide a detailed experimental study of scattering at radar
wavelength for a case of sea surface scattering. We examine the nature of the
amplitude fluctuations in light of our theoretical results, and demonstrate a close
agreement between theory and experiment. Our results also enable construction
RADAR WAVELENGTH 127
0.95
0.9
Correlation function
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
50 3
45 2.5
40 2
35 1.5
30 1
Smoothing parameter/0.1 ms Volatility exponent
Fig. 12.2. Correlation of zt and |dzt |p for random phase screen data. The values
p ≈ 2, smoothing window parameter ∆ ≈ 40 give an optimal correlation
of approximately 95%. For visualization purposes, the correlation function
shown is M − (M − c)1/2 where M = sup(c). The surface c is approximately
flat near its maximum (i.e. the absolute value of its derivatives are close to
zero). This is a desirable property from the point of view of the sensitivity of
the technique to the precise values chosen for the volatility exponent p and
smoothing parameter ∆, which are inevitably subject to empirical errors.
of an anomaly detection mechanism, which is applied to the sea surface scattering

data to isolate a body in space to a high degree of accuracy.
12.2.1 Coherent sea clutter
The data consists of K-distributed radar returns from a region of the ocean sur-
face, whose I, Q components (e.g. Helstrom 1960) constitute a complex-valued
two-dimensional space-time array Ψ(t, x) = I(t, x) + iQ(t, x), with pulse rep-
etition frequency (PRF) 1kHz, over a range of 300 m at a resolution 0.3 m,
and within which a tethered target is situated whose location was known. The
intensity data are shown in Fig. 12.4. The data were monochromatic, so that per-
sistent temporal correlation due to the illuminated physical objects can occur
over appreciable timescales, and there is no decorrelation arising from compo-
nent frequencies of the illuminating beam. In accordance with the findings from
the anomaly-free K-distributed data of Section 12.1, we adopt the behaviour
2
σ(z) ∼ z, i.e. p ≈ 2 for the stochastic volatility of the intensity process. The
view is therefore that the appropriate volatility exponent is a consequence of the
physics of the scattering process, which is common to both sets of K-distributed
data. A validation of this choice of calibration shall be a predominance of high
correlation between the various predicted and observed volatilities under this
Normalised intensity (solid line)/power of volatility (dotted line)
3.5
2.5
1.5
0.5
0
5500 6000 6500 7000 7500
Sample number at 10 kHz
Fig. 12.3. Random phase screen data – comparison between raw intensity and
|dz|p (both normalized by their respective means over the time shown). The
figure shows a strong correlation for p ≈ 2.
assumption, as evident in Fig. 12.6. The methodology of Section 12.1 involving

only the intensity process, when applied to the current set of data revealed a pre-
ponderance of ‘clutter spike’ anomalies with respect to range. As a result of this,
our analysis was enhanced by studying the stochastic volatility behaviour of the
complex-valued amplitude Ψ directly. It is the inclusion of the phase informa-
tion contained in Ψ that is the important feature in the avoidance of false alarms
caused by ‘clutter spike’ anomalies (Luttrell 2001), and which contains impor-
tant correlation information between the I and Q components. Accordingly, the
analysis of Ψ exploits all the physical degrees of freedom in the electromagnetic
field, which is desirable from the point of view of discrimination. (From a com-
putational point of view the analysis involving Ψt is only of the order a factor of
two more intensive compared to the corresponding analysis of the phase screen
data which involved only the intensity zt .)
For a given range we express the complex-valued amplitude in polar form
√
Ψt = Rt exp(iθt ), so that Rt = zt . In order to calculate dΨt , we apply Ito’s
1/2 3/2
formula to the modulus amplitude Rt , which yields dRt = dzt /2zt −dzt2 /8zt .
(R)
Writing dzt = Bb(z) dt + (2B)1/2 σ(z) dWt so that dzt2 = 2Bσ(z)
2
dt, we deduce
that
100
200
Range cell at resolution 0.3 m
300
400
500
600
700
800
900
1000
500 1000 1500 2000 2500 3000
Sample number at PRF 1 kHz
Fig. 12.4. Sea clutter space-time intensity.

1/2
B 2 B (R)
dRt = 2R b (z) − σ 2
(z) dt − σ(z) dWt . (12.4)
4R3 2R2
Ito’s formula yields the identity d exp(iθt ) = exp(iθt )[idθt − 12 dθt2 ] and thus by
the Ito product rule d(Xt Yt ) ≡ Xt dYt +Yt dXt +dXt dYt (see e.g. Oksendal 1998)
we deduce

1
dΨt = exp(iθt ) iRt dθt + dRt + idRt dθt − Rt dθt . 2
(12.5)
2
This identity yields the following geometric expressions for the squared volatili-
ties of the amplitude process Ψt ,
+ ,
dΨ2t = exp(2iθt ) dRt2 − Rt2 dθt2 + 2iRt dRt dθt , (12.6)
|dΨt |2 = dRt2 + Rt2 dθt2 . (12.7)
The phase θt is taken to be a stochastic quantity driven by an independent
(θ) (θ)
Wiener process Wt and which obeys the SDE dθt = ωdt + σ(θ) dWt . The
(R) (θ)
independence of Wt , Wt renders the dRt dθt term in (12.6) zero (see e.g.
Karatzas and Shreve 1988). An explicit derivation of this independence in the
Wiener increments in R, θ from a random walk model is provided in Field and
Tough (2003b). Thus we obtain the following result.34
34 Observe that the absolute value and square operations on dΨ do not commute with one
t
another, i.e. |dΨ2t | = |dΨt |2 .
Proposition 12.1 The squared volatilities of the amplitude process Ψt are given
in terms of the intensity and phase squared volatilities by

Bσ(z)
2
dΨ2t = exp(2iθt ) − 2
zt σ(θ) dt, (12.8)
2zt

Bσ(z)
2
|dΨt |2 = 2
+ zt σ(θ) dt. (12.9)
2zt
Observe that Proposition 12.1 does not involve drift quantities and only the
instantaneous volatilities in the intensity and phase feature explicitly in (12.8),
(12.9).
In the pure Rayleigh case, the I, Q processes can be described by the pair
of (uncoupled) Ornstein–Uhlenbeck processes (6.3), (6.4) from which we deduce
2
(e.g. via Ito’s formula applied to arctan(I/Q)) that σ(θ) = 2/z thus exhibiting
singular behaviour at zeroes of the intensity. This suggests that such singular be-
haviour should occur also in the K-distributed case which constitutes a Rayleigh
process with a modulated cross-section. The phase volatility can be calculated
from the normalized amplitude process via the identity |d exp(iθt )|2 = dθt2 , which
can be applied directly to the data and avoids difficulties arising from the discon-
tinuity in θt at 2π. A comparison of the intensity and (squared) phase volatility
is shown in Fig. 12.5, which reveals singular behaviour in σ(θ) near zeroes of
the intensity, characteristic of the Ornstein–Uhlenbeck (Rayleigh) process, and
approximate constant behaviour when the intensity lies above some threshold.
In respect of the singular behaviour at zeroes of the intensity, our results
are consistent with studies of the phase derivative, for a differentiable amplitude
process, reported in Jakeman et al. (2001). In contrast, however, in the differen-
tiable case it is the intensity, rather than amplitude-weighted phase derivative
that has minimal variance. This distinction between the differentiable and dif-
fusion models for the amplitude Ψt should enable one to select which type of
model is appropriate in different physical situations (see orig. §IIC in Field and
Tough 2003b for a detailed theoretical account of the phase behaviour in the
non-Gaussian case).
2
From a physical point of view the behaviour of σ(θ) is anticipated as follows.
The amplitude Ψ = I + iQ is generated from (some component of) the complex-
valued electric field E. One therefore expects a contribution to σ(θ) that varies
inversely with R, since |δE(θ) | ∼ Rδθ and |δE(θ) | contains an ‘intensive’ part
that is independent of the total field intensity Etot
2
. This observation is consistent
with the Ornstein–Uhlenbeck description of the Rayleigh process given above. In
N
the random walk model (Jakeman and Tough 1988) with Etot ∼ n=1 exp(iϕn ),
there is a fluctuation term that relates only to the random phases ϕn and which
is reflected in W (2) of the intensity SDE (5.20). Simultaneously, in the case of a
time varying scattering cross-section, there exist fluctuations in θt induced by an
7
phase volatility
6
Intensity/(squared) phase volatility
intensity
5
0
0 50 100 150 200 250 300 350 400
Fig. 12.5. Comparison of intensity and squared phase volatility.
‘extensive’ contribution to |δE(θ) |2 that scales with the scattering cross-section

|Etot |2 , i.e. the number of steps N in the random walk model of Jakeman
and Tough (1988). This feature is consistent with the appearance of a square
root volatility in the SDE for the scattering cross-section (5.19). For a Rayleigh
(Gaussian speckle) process the cross-section xt is constant, and so the ‘extensive’
contribution to |δE(θ) | does not arise.
(R) (θ)
Assuming that dW (R,θ) are independent, i.e. dWt dWt = 0, it follows
(I) (Q)
from the coordinate transformations I = R cos θ, Q = R sin θ that dWt dWt
2
vanishes if and only if σ(R) = Rσ(θ) , or equivalently σ(z) = 4z 2 σ(θ)
2
. This special
relation holds in the Rayleigh case for which the Wiener increments in I, Q are in-
dependent. In more general circumstances, however, we expect a departure from
this relation, so that in the K-distributed case the fluctuations in the scattering
cross-section induce a correlation between the increments dIt , dQt .
Explicit derivations of these various properties of the amplitude process in
the K-distributed case from the random walk model are described in Chapter 8.
2
These calculations reveal that σ(z) contains a term proportional to the constant
A that is quadratic in z, and a linear term proportional to B. This explains
the observed behaviour, for A B, of the volatility exponent p ≈ 2 in the
analysis of the intensity data, and yields a linear term in |dΨt |2 according to
2
(12.9). The analysis of the phase shows that σ(θ) is proportional to the constant
B and varies inversely with z and linearly with x in accordance with the general
scaling arguments adduced above. Consequently for A = 0, i.e. the K-distributed
case, the Wiener increments in the I, Q components of the field are correlated.
The special relation above for independence of the quadrature components is
recovered in the Rayleigh case for which A = 0.
12.2.2 Anomaly detection
As a consequence of Proposition 12.1 and the invariance of c(X, Y ) under linear
transformations, therefore, we obtain the following result.
Corollary 12.2 For a K-distributed35 process described by (5.19), (5.20) the in-
stantaneous correlation between the observed and theoretical |dΨt | squared volatil-
ities reduces to
cΨ = c(|dΨt |2 , zt ). (12.10)
The results of applying this correlation measure at each range cell are shown
in Fig. 12.6, which reveals high correlation at all ranges except the target loca-
tion. The persistence of this local spatial de-correlation over time is shown in
Fig. 12.7. Using these correlation results, plots were generated of the time series
for the theoretical and observed |dΨ|2 squared volatilities at both the maximum
and minimum correlation ranges as discerned from Fig. 12.6. This was carried
out in both real and permuted time, as explained above. The results shown in
Figs. 12.8–12.11 illustrate the instantaneous nature of the correlation, and how
this is destroyed at the target location.
The plots of the inverse predictive functions shown in Figs. 12.9 and 12.11
in fact provide the approximate cumulative distribution functions (c.d.f.) for the
respective random variables, when the vertical axis of the inverse function is
re-scaled to the interval [0, 1]. This feature would be useful in obtaining a mea-
sure of discrimination between the corresponding distributions, since the Fisher
information geodesic distance between two distributions pi can be approximated
directly in terms of the discrete c.d.f.s ci as cos θij = α [δci (α)δcj (α)]1/2 . In
the present approach, however, we assume that, in the clutter domain, the un-
derlying distributions are identical and apply the statistical correlation measure
(12.1) to the distinct random variables. It may however be possible to exploit
the Fisher information discrimination measure through a corresponding SDE
description for the target (cf. Wootters 1981).
The approach to anomaly detection described in the present section should
be compared with Haykin (1999) in which real empirical data are used to build
the predictive sea clutter model.
It is of some (theoretical) interest at this point to consider the possibility of
an anomaly detection scheme based on the same type of principle of comparison
35 This 2
result would not hold for the pure Rayleigh process because in that case σ(z) ∝ z,
2 ∝ 1/z and the right-hand side of (12.9) is then constant.
σ(θ)
1
0.9
0.8
0.7
0.6
Correlation
0.5
0.4
0.3
0.2
0.1
0
0 200 400 600 800 1000 1200
Fig. 12.6. Correlation of |dΨ|2 and σΨ 2

according to Corollary 12.2 versus range.
Mean correlation = 98.11%. Maximum correlation 99.94% occurring at range
cell 725. Minimum correlation 9.48% occurring at range cell 839. A threshold
of 80% isolates the anomaly to within 4 range cells, i.e. approximately 1 m.
(The signal is sampled for 1 s at PRF 1kHz over a range 360 m. The smoothing
parameter ∆ is 50 samples.)
between observed and theoretical squared volatilities, but that does not involve
the scattering cross-section xt explicitly. Indeed, such a scheme can be readily
constructed, as follows (albeit up to second order in the time differential). First
recall the squared differential relations for the amplitude and its modulus:

|dΨt |2 = (Bxt + Azt /2xt )dt,
(12.11)
dRt2 = 12 (Bxt + Azt /xt )dt.
Taking linear combinations of these yields

 1
 |dΨt |2 − dRt2 = Bxt dt
2 (12.12)
 2dR2 − |dΨ |2 = Azt dt
 t t
2xt
whose product gives the quadratic time differential relation
1
|dΨt |2 − dRt2 2dRt2 − |dΨt |2 = ABzt dt2 . (12.13)
4
In principle, this general relation facilitates an alternative stochastic volatility
based anomaly detection, without concern for the properties of xt which is no-
tably absent in (12.13).36
36 The author is grateful to Dr. Tao (Stephen) Feng for pointing out the existence of such a
scheme.
1.6
1.4
1.2
1-correlation
1
0.8
0.6
25
0.4
20
0.2
15
0 10
0
200 400 5 Time window translation/0.1s
600 800 1000 0
1200
Fig. 12.7. 3-dimensional plot showing range and time horizontally and
1 − c(|dΨt |2 , zt ) vertically; each point along the time axis corresponds to
a windowed sample of 1 s, thus covering a total time duration of 3 s of data.
The figure illustrates the persistence of the anomaly detection mechanism
over time – the object of interest disappears out of view for a short period
while it is submerged due to bulk wave motion, and then re-surfaces.
We close this chapter with some concluding remarks on the experimental

aspects of our scattering models. Our stochastic treatment of the electromagnetic
scattering from random media reveals a new technique for anomaly detection,
which is based on the instantaneous observability of the stochastic volatility of
the complex-valued amplitude process. The emphasis on the stochastic volatility
constitutes a fundamental change in viewpoint as to the choice of significant
parameters, and lies in contrast to previous thinking which has focussed on the
dynamics contained in the drift. In this respect, there is a similarity between
our approach to the scattering problem and the mathematical theory of option
pricing for which the absence of an ensemble average leads to an arbitrage-free
pricing mechanism that is drift independent (cf. Appendix D).
In the analysis of the radar data of Section 12.2, an essential ingredient has
been the behaviour of the complex-valued amplitude process, and how the form
of its volatility function induces correlation properties between the I, Q compo-
nents of the amplitude. In the K-distributed case, these correlations arise from
fluctuations in the scattering cross-section. Such correlations are correspondingly
absent in the pure Rayleigh (Gaussian speckle) process, for which there exists
a description for I, Q in terms of a pair of uncoupled Ornstein–Uhlenbeck pro-
cesses.
Predicted (solid line)/observed (dotted line) squared volatility
0
200 300 400 500 600 700
Fig. 12.8. Comparison of predicted and measured (smoothed) squared volatility

for the process |dΨt |. The range cell was chosen to give maximum correlation.
5
4.5
Predicted (solid line)/observed
(dotted line) squared volatility
4
3.5
3
2.5
2
1.5
1
0.5
00 100 200 300 400 500 600 700 800 900

Permuted sample number at 1 kHz
Fig. 12.9. Permuted-time representation of predicted and observed (smoothed)

squared volatility for the process |dΨt |. The predicted values are permuted in
ascending order; the required permutation is then applied also to the observed
values, which are subsequently smoothed. The range cell was chosen as in the
previous figure.
3.5
predicted (solid line)/observed
2.5
1.5
0.5
0
0 100 200 300 400 500 600 700 800 900 1000
Sample number at 1kHz
Fig. 12.10. Comparison of predicted and observed volatilities; as in Fig. 12.8

except with range cell chosen at the location of the anomaly where the cor-
relation is minimized.
3.5
Predicted (solid line)/observed
2.5
1.5
0.5
0
0 100 200 300 400 500 600 700 800 900
Permuted sample number at 1 kHz
Fig. 12.11. Comparison of predicted and observed volatilities; as in Fig. 12.9

except with range cell chosen at the location of the anomaly where the cor-
relation is minimized.
In furtherance of this study one hopes to exploit the freedom in the volatility
functions in the classification of K-distributed processes provided in Chapter 5,
thus exploring different calibrations for σ(z) and the underlying phase volatility
σ(θ) from a simulation or data-driven point of view. Our analysis could be ex-
tended to include non-Markovian models for the behaviour of xt , zt , Ψt via the
introduction of extra variables into a larger number of coupled SDEs (cf. Tough
1987), and in addition, to explore other models for the scattering cross-section,
which do not fall into the general category of Ito processes. The analysis of the
discrete sampled data in this chapter, which integrates the underlying Ito SDEs
to first order only, could be enhanced by higher order numerical integration
schemes (Klauder and Petersen 1985) as described in Chapter 11. The aim of
pursuing these various avenues would be to optimize the correlation between the
predicted and observed volatilities in the predominant K-distributed domain.
The techniques we have described could also likely be adapted to situations that
occur in the modelling of financial time series with a view to detecting anoma-
lous behaviour, and to other situations of propagation of acoustic waves through
random media.
13
NON-LINEAR DYNAMICS OF SEA CLUTTER
We review experimental evidence for the non-linearity of sea clutter and the role
of the z-parameter or Mann–Whitney rank-sum statistic in quantifying this non-
linear behaviour in the context of a hybrid amplitude modulation (AM)/frequency
modulation (FM) model for sea clutter, viewed as a cyclostationary process. An
independent theoretical derivation of the stochastic dynamics of radar scattering
in a sea clutter environment, in terms of a pair of coupled stochastic differential
equations (SDEs) for the received envelope and radar cross-section (RCS), en-
ables the identification of non-linearity in terms of the shape parameter for the
RCS. We are led to conclude that, from both experimental and theoretical points
of view, the dynamics of sea clutter are non-linear with a consistent degree of
non-linearity that is determined by the sea state.37
Haykin et al. (2002) advocated a state-space formalism for the processing
of radar signals in the presence of sea clutter (i.e. radar backscatter from an
ocean surface). Such a model not only accounts for the temporal dimension of
sea clutter in an explicit manner but also its statistical characterization. Basic
to this formalism is whether the underlying dynamics of sea clutter are linear,
or non-linear.
In the detailed experimental study reported in Haykin et al. (2002), it was also
demonstrated that sea clutter is a non-linear dynamic process, with the degree of
non-linearity increasing as the ‘sea state’ becomes higher. The conclusion reached
on the non-linearity of sea clutter was based on two premises, using real-life data
collected with an instrument-quality coherent radar system:
1. The characterization of sea clutter embodies two forms of continuous-wave
modulation:
(i) amplitude modulation, which is linear, and
(ii) frequency modulation, which is non-linear
The latter phenomenon is responsible for the non-linearity of sea clutter.
2. The z-parameter, denoting the Mann–Whitney rank-sum statistic, is less
than the special value −3, which is a strong indicator of non-linearity.
With regard to point 1, it is also noteworthy that in another study that
focussed on the spectral characterization of sea clutter using the Loève trans-
form (Haykin and Thomson 1998), it was discovered for the first time that sea
37 The author acknowledges the input and collaboration of Prof. Simon Haykin with regard
to the material on the z-parameter and the hybrid AM/FM model in this chapter.
138
NON-LINEAR DYNAMICS 139
clutter is a cyclostationary process. Cyclostationarity is ordinarily associated

with modulation. But knowing that sea clutter is cyclostationary, it does not tell
us the type of modulation involved in the characterization of its waveform.
In this section, we expand on the characterization of sea clutter as a non-
linear dynamic process, using a principled theoretical approach. In particular, the
approach is rooted in SDE theory. The issue of the dynamics of radar scattering
in a sea clutter environment has been addressed in the literature independently
from both theoretical and experimental points of view. Perhaps most notable in
the former case, Field and Tough (2003a,b) develop a theoretical basis for the dy-
namics which is demonstrated to agree with experimental data to a remarkable
degree of accuracy. In the latter case, Haykin et al. (2002) study experimental
data to motivate a line of argument leading to the conclusion that sea clutter
is inherently non-linear (and, indeed, possibly chaotic). Here we bring these two
independent lines of development together in a consistent way in order to estab-
lish the non-linear nature of sea clutter from both physical and mathematical
viewpoints. More precisely, the scattering dynamics can be derived from first
principles in terms of a pair of SDEs for the received envelope and the radar
cross-section (RCS) that feature a non-linear coupling and encode the statistical
character of the sea state in terms of a certain ‘shape parameter’. Examination
of the differentiable parts in this system of SDEs reveals a corresponding ‘noise-
free skeleton’ that is a non-linear vector process, with a degree of non-linearity
dependent on the shape parameter in a manner consistent with that shown ex-
perimentally by Haykin and co-workers. This significant development affirms the
case for the non-linear character of radar sea clutter.
The chapter is organized as follows. Section 13.1 provides a summary of the

experimental study that led to the formulation of a hybrid AM/FM model, and
the conclusion that sea clutter is a non-linear dynamic process. Section 13.2
summarizes the essential ingredients of SDE theory necessary for the basic inter-
pretation of the SDE dynamics of radar sea clutter. We then apply this formalism
to establish the non-linear character of the stochastic dynamics of the vector pro-
cess consisting of the RCS and resultant back-scattered amplitude or ‘received
envelope’. This is achieved from first principles via an extended random walk
model, as developed in Chapter 8. The extent of the non-linearity in the result-
ing SDE description is quantified in terms of a certain ‘shape parameter’ (the
relative variance in the RCS, minus one) that encodes the sea state. We conclude
with a discussion of the interplay between the two independent lines of enquiry
that lead to the common conclusions concerning the non-linear character of radar
sea clutter. We also indicate how our results may suggest which types of experi-
ments to perform to further substantiate and enhance the theoretical framework,
and discuss future prospects for the investigation of chaotic dynamics.
We refer the reader also to the recent book by Haykin (2006 ed.), where
the experimental results are mentioned in the broader context of adaptive radar
signal processing.
140 NON-LINEAR DYNAMICS
13.1 Hybrid AM/FM model of sea clutter

In an independent study reported in Gini and Greco (2001), sea clutter was
viewed as a fast ‘speckle’ process multiplied by a ‘texture’ component that rep-
resents the slowly varying power level of the sea clutter signal; such a model
is perceptually satisfying. The slow variation of the sea clutter power level was
attributed to the large ocean waves passing through the observed ocean patch.
The speckle was modeled as a stationary compound complex Gaussian process,
and the texture was modelled as a harmonic process.
Inspired by the Gini–Greco model of sea clutter, Haykin and coworkers car-
ried out an extensive physical study of sea clutter collected by the instrument
quality coherent IPIX radar, where the radar data were recorded on the East
Coast of Canada (Haykin et al. 2002). In that paper, it was demonstrated that
AM and FM play important roles in the waveform description of sea clutter. The
hybrid AM/FM model of sea clutter has been substantiated further in Greco and
Gini (2007).
To explain the physical presence of modulation in sea clutter, we observe
that when a large wave passes across a patch of the ocean surface, it will first
accelerate and then decelerate the water’s motion on the ocean surface. The
continuous tilting of the ocean surface by the waves gives rise to amplitude
modulation.
Moreover, the ocean wave will cause a cyclic motion of the instantaneous
velocity of scatterers on the ocean surface, thereby giving rise to frequency mod-
ulation as another characteristic of the sea clutter waveform. When the mean
velocity of the scatterers is high at a given instant of time, then the spectral
spread (i.e., the bandwidth occupied by the frequency modulation) around that
mean is correspondingly high, which is in perfect accordance with modulation
theory.
It is well known that, unlike AM, FM is a non-linear process (Haykin 2001).
Therefore, the presence of FM in the physical behaviour of the sea clutter wave-
form leads us to hypothesize that sea clutter is a non-linear dynamic process.
To validate this hypothesis Haykin et al. use 78 different coherent radar data
sets to compute the z-parameter, which denotes the Mann–Whitney rank-sum
statistic (Siegel 1956). The results of this test are reproduced in Fig. 13.1, where
the z-parameter is plotted against the spectral width modulation. A value of z
less than −3 is considered to be a strong reason for rejecting the null hypothesis
that the sea clutter data under test can be described by linearly correlated noise.
In Fig. 13.1, we clearly see that the large majority of the experimental points lie
below z = −3. Those points were in actual fact representative of high sea states.
On the basis of these experimental results, Haykin et al. concluded that sea
clutter is indeed a non-linear dynamic process, with the degree of non-linearity
increasing with increasing sea state.
–2
–4
–6
Z[–]
–8
–10
–12
–14
–16
20 40 60 80
NMAD(df/dt ) [Hz]
Fig. 13.1. z-parameter value versus Normalized Median Absolute Deviation

NMAD (φ̇), computed for 78 data sets measured by the IPIX radar at various
experimental conditions.
13.2 Non-linear dynamics from SDE theory

SDE theory has significant implications for statistical signal processing. It has
recently proven successful in this context in the application to radar sea clutter
(Field 2002; Field and Tough 2003a; Field and Haykin 2008). In a more general
physical context, including optical propagation, the stochastic calculus has led
to substantial new theoretical developments in the subject of electromagnetic
scattering from random media (Field and Tough 2003b; Field 2005).
More recently, SDE techniques have been applied to wireless channel mod-
elling (Feng et al. 2007) to include the effects of phase fluctuations in multi-path
reception. The fact that similar techniques are applicable to both the radar
backscattering and wireless propagation problems stems from the fact that each
is multi-path in nature, with the only essential difference being that for radar
the receiver and transmitter are co-located. This latter feature, however, does
not affect the structure of the mathematical model used to describe the resulting
amplitude signal.
Here we shall consider the RCS and received envelope processes to evolve
according to the dynamics governed by a SDE. In the context of the RCS such
dynamics arise from taking the continuum limit of a generic population dynamic
model for the (discrete) number of component scatterers. For the scattered ra-
diation the origin of the SDE dynamics lies in the behaviour of the component
phases which are taken to evolve in time according to a Wiener process on a
suitable (Rayleigh) timescale.
Thus we are able to represent the essential ingredients of the radar back-
scatter temporally, in the form of a continuous time stochastic process, say qt ,
which evolves in time according to
dqt = bt dt + σt dWt . (13.1)
Herein, bt is a random process, and represents the ordinary time derivative of

the process qt in the case that σ vanishes. The quantity σt , on the other hand, is
the amplitude of the noise or fluctuating part of qt , in general a random process,
and referred to as the ‘stochastic volatility’ of qt . In the cases we study, it will
become apparent that bt = b(t, qt ) and σt = σ(t, qt ) for some specific functions
b, σ, and accordingly the process qt is called a ‘diffusion’.
In contrast to the part of dqt containing bt , the σt term contributes an es-
sential part to qt that is not differentiable, in the ordinary sense that d/dt
is well-defined. Nevertheless, the (Ito) stochastic differential of qt can be well
defined.
In the engineering physics literature, one is perhaps more familiar with the
Langevin equation for the time derivative
dq
= bt + σt Γt (13.2)
dt
in which Γt is the familiar white noise process and Γt has the autocorrelation
property Γt Γt = δ(t − t ). For our purposes it will be sufficient to understand
and interpret from the dynamical equations for the RCS and the received radar
amplitude that, in a discrete-time setting,
δqti = bti δt + σti nti δt1/2 , (13.3)
where {ti } is a discrete set of observation times, δt = ti+1 − ti , and {nti } are a
collection of independent N (0, 1) random variables. Then the above properties of
q and its time derivative are evident (see Oksendal 1998 for a detailed rigourous
account).
The essence of the approach taken is therefore to postulate the exact dynam-
ics in continuous time, and then sample at a discrete set of times corresponding
to the physical measurements. This procedure is inevitably more precise than
an attempt at a model that is fundamentally discrete time in nature, since the
physical observables are not quantized in time.
We shall assume the (dynamical extension of the) random walk model for the
resultant back-scattered amplitude or ‘received envelope’
s(j)
N 7
89
& ':
(N ) (j)
Et = aj exp iϕt (13.4)
j=1
with (fluctuating) population size N , random phasor step s(j) , and ‘form factors’
aj . The key result of relevance to our discussion is obtained by taking the (Ito)
stochastic differential of (13.4). This provides the following coupled stochastic
dynamics of the RCS and scattered amplitude/received envelope Ψt .38
Proposition 13.1 The dynamics of the RCS and received envelope for radar
sea clutter, with shape parameter ν = α − 1, are given by the following set of
non-linearly coupled SDEs:
(x)
dxt = A(α − xt )dt + (2Axt )1/2 dWt (13.5)
1/2
dΨt 2(α − xt ) − 1 1 A (x) B 1/2
= A − B dt + dWt + dξt (13.6)
Ψt 4xt 2 2xt γt
in which γt is a unit power Rayleigh process, whose dynamics are obtained by
setting xt equal to a constant of unity and A = 0 in the above system.
(The original proof of this result appears as Proposition 2.1 in Field and
Tough 2003b, and we refer the reader to Chapter 8 herein for a detailed math-
ematical derivation.) Thus, the non-linear SDE for Ψt is derived theoretically
from first principles beginning with the random walk model for the scattered
electric field under the assumption of a uniform phase distribution.39 An imme-
diate consequence of this dynamical equation is the ‘noise-free skeleton’, obtained
by setting the volatility coefficients of the fluctuating Wiener terms equal to zero.
Accordingly, the randomness of the process is eliminated and the residual dy-
namics are deterministic and differentiable. Physically, this corresponds to an
evolution conditioned on the current state of the system and then averaged over
an ensemble.40 The concept of the residual noise-free part is explored further
below.
The constants A, B in (13.5), (13.6) have the physical dimension of frequency,
so that their reciprocals represent correlation timescales for the RCS modulation
and Rayleigh scattered components, respectively. The constant B is electromag-
netic in origin with a value B ∼ c|k| where k is the wave vector of the carrier. In
radar situations, the illuminating radiation is such that A B, with the value
of A being determined as an intrinsic property of the statistics of the scattering
38 We
√
can express Ψt = It +jQt (j = −1), the familiar sum of its ‘in-phase’ and ‘quadrature-
phase’ components.
39 The assumption of a uniform phase distribution can be relaxed and a corresponding de-
tailed dynamical description in terms of SDEs has been given in Field and Tough (2005).
40 In other words, for an Ito process q with SDE dq = b dt + σ dW , the ensemble average
t t t t t
evolution is determined by E[dqt ] = bt dt.
surface, independent of the electromagnetic wave. Accordingly in radar the cor-

relation time for the RCS is much longer than that of the Rayleigh speckle. (The
reader should compare also the discussion of amplitude and frequency modula-
tion given at the end of this section.)
This set of coupled stochastic dynamical equations is manifestly non-linear
by virtue of the reciprocal term in xt appearing in the amplitude equation, and
only reduces to linear dynamics in the special case that A vanishes, i.e. the
scattering cross-section is constant (Rayleigh scattering).
13.3 Radar parameters

In the present context, it is worth remarking on some of the key salient features
of the SDE theory, in relation to the sensitivity analysis of sea clutter to certain
radar parameters. Most notably, this kind of description is illuminating in respect
of the following issues.
13.3.1 Superposition
In light of the SDE theory, we may argue that the SDE of sea clutter is inde-
pendent of the amplitude profile of a transmitted pulse, provided the transmit
energy is maintained constant. This property, which is derived explicitly in Field
(2005), is related to the fact that the form factors (i.e. the amplitude weight-
ings) in (13.4) may be taken as unity for an asymptotically large population
(cf. also Jakeman and Tough 1988 where the emergent statistical properties are
independent of the choice of form factors).
For a radar pulse of constant amplitude, suppose that the two halves of the
pulse have transmit frequencies ω1 and ω2 . Then we may consider the correlation
between the SDE of sea clutter for the two portions as frequency ω1 increases
relative to ω2 . The transmit frequencies are proportional to the Rayleigh con-
stant B appearing in (13.6) (B ∼ c|k|, c is the speed of light, k is the carrier
wave vector), and the relationship between the two SDEs, for the two different
transmit frequencies, is through (13.6): the two terms involving the constant A
are the same for both SDEs; on the other hand, the terms involving the Rayleigh
constant have different B values corresponding to the two transmit frequencies.
Nevertheless, on physical grounds, the two complex Wiener processes ξt for each
transmit frequency should be considered perfectly correlated. The reason for this
correlation is that the physical origin of the component phase fluctuations φ(j)
is (microscopic) Doppler – the Doppler frequency ratio ω1 /ω2 is a function of
the radial velocity of the j-th member of the population, so the micro-Doppler
phase shift scales with the transmit frequency; the ξt process is the same for
any transmit frequency (assuming these are transmitted simultaneously) as this
depends only on the behaviour of the component scatterer.
In a similar fashion, consider the simultaneous transmission of two pulses of
constant amplitude, with two different frequencies as above, and the resulting
SDE of sea clutter received by a common antenna. Since Maxwell’s equations
of electromagnetism are linear, the resulting Ψ is a linear superposition Ψ =
RADAR PARAMETERS 145
k1 Ψ(1) + k2 Ψ(2) where k1 , k2 are the relative intensities of the two transmit
waveforms, normalized so that k1 + k2 = 1, and Ψ(i) are the constituent complex
amplitude processes, both satisfying the SDE (13.6), with different Rayleigh
constants B corresponding to the two transmit frequencies. Since the beams are
simultaneous, the ξ processes are perfectly correlated, with the remaining parts
of (13.6) involving the constant A being the same for both transmit frequencies.
Thus, the non-linear dynamics do not infringe the principle of superposition
inherent in Maxwell’s equations.41
13.3.2 Sea state and polarization

Next, consider the two different like-polarizations ‘HH’ and ‘VV’. The SDE the-
ory encapsulates the spikes in the RCS of sea clutter due to ‘HH’ versus the
noise-like character due to ‘VV’ as follows. The cross-section SDE (13.5) con-
tains the parameter α, where α = ν/λ and λ, ν are the birth and immigra-
tion parameters of the underlying (BDI) population model, respectively. Now
Var[x] = E[x] = α, so the absolute magnitude of fluctuations in the RCS, that
give rise to the K-distribution for the intensity (as opposed to the Rayleigh
‘noise-like’ distribution), become more appreciable as α increases. However, the
appropriate theoretical measure of ‘spikiness’ is the relative variance R given by
Var[x]
R= (13.7)
(E[x])2
which is the physical parameter of interest since it is dimensionless and invari-
ant under re-scaling of the RCS. In the case of K-scattering that we consider
here, R is equal to 1/α, and therefore the horizontal ‘HH’ like-polarization has
small α, with larger α for vertical like-polarization ‘VV’. The SDE theory ex-
plains that if the ratio α of the immigration to birth rates is small, then the
sea clutter possesses spikes. It is therefore a natural mathematical, as opposed
to a detailed phenomenological, way of encoding this physical property of the
sea surface.42 Correspondingly, there are two different K-distributions for the
intensity, indexed by different values of the shape parameter ν = α − 1, for the
respective polarizations.
The situation as regards the extent of the temporal fluctuations in the RCS
for low/moderate/high values of the shape parameter is well illustrated in Fig.
13.2. The figure demonstrates the extreme departures from the mean value for
large R, which represents, in physical terms, sea spikes or glints in the scattering
surface. The dynamical nature of these extreme departures can be understood
further, to some extent, by calculating the SDE of the normalized RCS x = x/α,
which is seen to be
41 It is necessary to assume here that the scattering populations N (1) and N (2) pertaining
to the different transmit frequencies are equal.

42 However, the SDE theory does not explain why for ‘HH’ polarization one should expect
the population to behave this way, the phenomenological reasons for which we do not describe
here.
3.5
small
3
Radar cross-section
2.5
2 large
moderate
1.5
0.5
0
0 2000 4000 6000 8000 10000 12000
Time
Fig. 13.2. Normalized RCS time-series for low/moderate/high values of the

shape parameter; simulated data with Aδt = 0.001, α = 1, 10, 100.
12
2Axt
dxt xt )dt
(x)
= A(1 − + dWt . (13.8)
α
A natural question that arises for the domain of low α (high sea state) is whether,
following a burst, the RCS returns to equilibrium and undergoes a period of
relative quiescence before a successive burst, or whether the high relative variance
is achieved through perpetual bursts of varying magnitude, with the maximum
timescale for their separation (pertaining to the largest deviations) being given by
the reciprocal of the constant A. Inspection of (13.8) reveals the latter situation
to be the case, as setting x ≈ 1 therein removes the drift term but nevertheless
produces a highly volatile behaviour for small values of α. Observe also, in this
regard, the essential difference between the roles of A and α, which determine
the correlation timescale and magnitude of excursions respectively – thus, for
instance, the single point distribution of the RCS is unaffected by A, whereas it
is α that parameterizes this distribution according to the discussion of the shape
parameter above. These features are indeed evident from the simulated data
shown in Fig. 13.2 (for α = 1). Observe also, embedded within the largest peak
to peak timescale O(A−1 ), the presence of sub-scale peaks separated by times
of smaller order, including scales less than the characteristic Rayleigh timescale
O(B−1 ).
Now, in contrast, as the sea state settles down to a low value, the (normalized)
RCS has small fluctuations away from its mean (unity), so that there is no
significant modulation of the Rayleigh scattering time series – in other words,
the scattering is approximately of constant local power.
As the sea state diminishes, correspondingly in terms of the SDE dynamics

the parameter α → ∞ and the relative variance in the RCS tends to zero. Thus
in Fig. 13.2 the non-linear term becomes less pronounced. Accordingly, as we
have seen, so does the degree of non-linearity as measured by the z-parameter
as developed in Section 13.1, which confirms the experimental findings reported
in Haykin et al. (2002).
We close this chapter with some general discussion concerning the main as-
pects of the non-linear dynamics of sea clutter. We have described a detailed
analysis of radar sea clutter data, whose primary purpose is to address the pres-
ence of non-linearity, from real experimental data. A natural quantifier for this
non-linearity is the z-parameter or Mann–Whitney rank-sum statistic, which
has been successfully applied in the context of a hybrid AM/FM model for sea
clutter. The SDE dynamical model of radar sea clutter has also been verified pre-
viously to a remarkable degree of accuracy, in terms of real experimental data
(orig. section 4(b) in Field and Tough 2003a). Moreover, an independent theo-
retical account for such a model was provided in Field and Tough (2003b), and
has served as the basis for other significant developments (Field and Tough 2005;
Field 2005). As we have seen in Section 13.2, this stochastic dynamic behaviour
is inherently non-linear, due to the broader timescale fluctuations in the RCS.
The extent of non-linearity arises naturally in the SDE description through the
relative variance or shape-parameter, which encodes the sea state. Thus, from
an SDE dynamical perspective, the non-linear character of radar sea clutter is
firmly established, both theoretically and experimentally.
Calculation of the z-parameter is from real data containing noise, the latter
being akin to the stochastic fluctuating terms present in (13.5), (13.6). However,
z has the stochastic element removed, i.e., it is not a random variable. Accord-
ingly, some ensemble averaging takes place in the calculation of z, and for this
purpose the statistical properties of ergodicity and stationarity are assumed, le-
gitimate over realistic short timescales. In terms of the parameter A of (13.5),
such timescales are short enough that the assumption of constant A is valid. Nev-
ertheless, they should be long enough (of the order of A−1 ) for the fluctuations
in the RCS (or equivalently, as we elucidate below, the frequency modulation
effect) to be appreciable so that non-linearity can indeed be detected.
From an engineering physics perspective, the dynamics of sea clutter are
perhaps more naturally viewed in terms of amplitude (AM) and frequency mod-
ulation (FM). Studies have indicated that the degree of non-linearity is governed
by the extent of FM which, in turn, is more noticeable for higher sea states
(i.e. the shape parameter ν is large). To relate this further to the SDE descrip-
tion of Section 13.2, it is convenient to view the resultant amplitude process
Ψt in the product representation Ψ = x1/2 γ, in which x is the RCS and γ is
a unit power Rayleigh process. Then, the AM consists of the fluctuations of
γt (Rayleigh ‘speckle’) which is ‘frequency’ modulated by the RCS process xt
over a much broader timescale. The FM/AM contributions therefore have char-
acteristic frequencies determined by A, B respectively. With zero FM, i.e. xt

constant, the dynamics of the resultant amplitude are rendered linear, according
to Proposition 13.1.
Observe that, whereas the FM/AM characteristics of the received envelope
do not map to unique stochastic dynamics, conversely the SDE description al-
lows for explicit extraction of both the FM/AM constituents (see Field 2005),
and therefore the SDE description is more fundamental than the spectral one.
Indeed, given the SDE dynamics, we are able to extract all higher order statisti-
cal information through the propagators obtained as solutions of the associated
Fokker–Planck equations (Risken 1989; Field and Tough 2003b). In this way, the
SDE description of sea clutter should be viewed as the most complete dynamical
description, which preserves the inherent randomness in the physical processes
involved.
We have seen that independent lines of enquiry, from theoretical and exper-
imental perspectives, lead to the common conclusion that radar sea clutter is
non-linear over appreciable timescales such that the temporal variation in the
RCS is significant. The degree of non-linearity is determined by the sea state,
which is represented by a certain ‘shape parameter’ ν that features in the SDE
for the RCS.43 Consistently, the non-linearity is also determined by the extent
of frequency modulation which, in terms of real experimental data, has been
quantified in terms of a certain z-parameter representing the Mann–Whitney
rank-sum statistic.
From a theoretical point of view the deterministic part of the stochastic
dynamics (13.5), (13.6) is non-linear, and is augmented with the addition of
fluctuating Wiener terms in the description of real experimental data, which is
inherently noisy. (The reader should compare the discussion of chaos surrounding
figure 2 in Sugihara 1994, and also Stone 1992, Sugihara and May 1990.) We
recommend further studies be made on the noise-free skeleton of the coupled
system (13.5), (13.6), which is manifestly non-linear, to establish the existence
or otherwise of an underlying deterministic chaotic behaviour. If chaos is present,
then this system of non-linearly coupled SDEs is an instance of ‘stochastic chaos’.
We remark in this respect that the presence of the Wiener fluctuating terms in the
system has the effect of stabilizing the system, so that any chaotic behaviour may
no longer be observable experimentally. These issues will be pursued elsewhere
in the journal literature.
It is worth emphasizing again that the SDE theory of sea clutter is experimen-
tally valid, in its own terms (orig. section 4(b) in Field and Tough 2003a), and
has also succeeded in practical applications, such as radar anomaly detection, to
a remarkable degree of accuracy. The theory also provides a way of generating
synthetic data, over which we have direct control, in terms of its dependence on
the sea state. Thus, in principle, we could measure the z-parameter for a data
43 More precisely, ν = α − 1 where α is the parameter in the SDE for the RCS, and α ≥ 0
arises from the parameters in the scattering population model as described in Chapter 7.
set simulated using SDEs, for which the shape parameter is known, and thereby
develop the precise relationship between the z-parameter and shape parameter
quantifiers of non-linearity. It may, indeed, also be possible to relate the two pa-
rameters in purely theoretical terms. We can also generate data with and without
noise, which forms the basis for further experiment. We suggest that these two
lines of enquiry could form the basis of future developments in the investigation
of the non-linear properties of radar scattering dynamics.
14
OBSERVABILITY OF SCATTERING CROSS-SECTION
A study of some synthetically generated data provides a clear illustration of the

consequences of Theorem 10.6 and Proposition 10.7 in a variety of experimental
contexts. The theoretical results are conveniently substantiated using synthetic
data, as such enables direct verification of the accuracy of the state estimation
from the observations, the precise value of the ‘hidden’ state being known. One
can then compare the values of the state inferred from the observations alone,
with the exact values of the underlying hidden state recorded in the simulation.
14.1 Simulated data

The cross-section was chosen to satisfy the stochastic differential equation (SDE)
(10.5) with b(t, x) = α−x, Σ(t, x) = x so that xt is a gamma variate and the pro-
cess thus generated is appropriate to the types of scattering data found in radar
applications (cf. Field and Tough 2003a). Resultant amplitude data was simu-
lated via integration of (10.7), which is achieved most effectively by a separate
numerical integration of the (independent) component SDEs (10.3) and (10.5).
We emphasize however, that the same types of numerical results as demonstrated
below should hold for an arbitrary population, as shown theoretically in The-
orem 10.6. For the purposes of the simulation, α was chosen to be large, to
avoid numerical difficulties that can arise due to the singular behaviour in the
phase fluctuations at zeros of the intensity, implied by (10.15). The results of the
simulation are provided in Fig. 14.1, which shows time series for the observed
intensity, the exact cross-section generated in the simulation (i.e. the unknown
state one is trying to estimate), and the values of this state inferred from the
observations of the scattered amplitude alone. The estimate of the state follows
from Theorem 10.6 which, for discretely sampled data, implies
zi δθi2 ∝ xi n2i (14.1)
where i is a discrete time index and {ni } are an independent collection of N (0, 1)
distributed random variables. Applying a smoothing average ·∆ to the left-hand
side (the ‘observations’) of (14.1) with window ∆ = [t0 − ∆, t0 + ∆] yields an
approximation to xt0 , with an error that tends to zero as the number of pulses
inside ∆ tends to infinity and ∆ → 0 (see discussion of χ2 statistic following
Theorem 10.6).44
44 We assume here that the sample paths of x are continuous a.s., which is a consequence of
t
(10.5).
150
EXPERIMENT 151
3
Exact (solid) and inferred (dotted) cross-section
2.5
intensity
time series; intensity as indicated
2
inferred
cross-section
1.5 exact cross-section
0.5
0
0 0.5 1 1.5 2 2.5 3
Time x 104
Fig. 14.1. Estimation of the scattering cross-section/population after relax-

ation time T through the effect of phase decoherence. (For parameter values
α = 10, ∆ = 1200 pulses, A = 10−5 , B = 10−4 the figure shows a statis-
tical correlation between the exact and inferred cross-section time series of
0.9954.)
Remarks. On finite pulse rate. In the case that the pulse rate is bounded, as in
a real experiment, one should consider the optimization of the smoothing window
– the window must be neither so large that the structure of the temporal variation
in the cross-section is lost, nor so small that an average over the normal random
seeds is no longer effected. Thus a compromise optimal value should exist, and
can indeed be derived (at least for K-scattering). We refer the interested reader
to Fayard and Field (2008) for a detailed account.
On theother hand, smoothing the zt time series, for any choice of param-
eter ∆, does not yield the desired close correlation with xt ; indeed any such
attempt to ‘decorrelate the speckle pattern’ merely produces an intensity profile
with the same general shape as the original zt , with oscillations on a timescale
approximately equal to ∆.
14.2 Experimental applications

The application of related earlier ideas in SDE theory to optical propagation and
radar scattering has been reported in §4 of Field and Tough (2003a). The current
results should be of additional value in these types of experiments, through the
ability to estimate the underlying scattering cross-section for general scattering
populations, in real time. When applied to experimental data, the results herein
provide means for studying the behaviour of random media based on statistical
152 OBSERVABILITY
analysis of the inferred cross-section alone, which has hitherto been regarded as
the ‘hidden’ state of the system. The electromagnetic scattering process should
then appropriately be viewed as a secondary exogenous device, whose purpose
is merely to extract the real time behaviour of the underlying scattering cross-
section, where the latter is the object of primary interest.
Our results also suggest the application to the physics of magnetic resonance
(MR) imaging and spectroscopy and experimental NMR (Field and Bain 2008).
The random medium (e.g. brain tissue) lies inside a background magnetic field
B0 , with which the constituent (proton) spin vectors are aligned, in their mini-
mum energy configuration. An applied RF pulse causes resonant absorption to
occur, so that the spins re-align, typically at a ‘pulse flip’ angle of 90◦ to B0 .
Radiation of this absorbed energy gives rise to the received MR signal (the ‘free
induction decay’ or FID), which is detected through the generation of electro-
motive force in a coil apparatus, due to the time varying local magnetic field.45
The MR signal has the usual in-phase (I) and quadrature-phase (Q) compo-
nents familiar from radio theory, and thus corresponds to the amplitude process
ψ = I + iQ for each point in space. For a perfectly homogeneous (total) mag-
netic field throughout the medium, each spin vector precesses at the Larmor
frequency ω0 about the longitudinal axis, where ω0 is given by the Larmor equa-
tion ω0 = γB0 and γ is the (local) gyromagnetic ratio (e.g. Ernst et al. 1987). (In
the radar scattering situation described above, ω0 corresponds to the Doppler
frequency, arising from bulk wave motion in the scattering surface.) However,
the local inhomogeneities in the net magnetic field, due to the local magnetic
properties of the medium, give rise to a process known as ‘spin-spin’ or ‘T 2 re-
laxation’ constituting the (random) exchange of energy between neighbouring
spins. These local perturbations in the total magnetic field can reasonably be
considered as independent for each component spin, so that the dynamics of
each spin vector can be modelled as a phase diffusion process with (transverse)
resultant as in (6.16) and phase initializations {∆(j) } equal. After sufficient re-
laxation time t ≥ T has elapsed, phase decoherence occurs. In principle, this
enables the spin population to be tracked according to Theorem 10.6, and thus
real-time MR images to be generated. This application is explored in greater
depth, and from a less detailed mathematical perspective, in Field (2006). In
the context of experimental NMR these theoretical ideas have been tested and
verified, as reported in Field and Bain (2008).
45 This effect is the result of Faraday’s law, i.e. Maxwell’s curl equation for a time dependent
magnetic field ∇ × E = −∂B/∂t integrated around a loop.

APPENDIX A
STABILITY AND INFINITE DIVISIBILITY
First, let us remark on the basic properties of sums of independent random vari-
ables. Suppose {Xi } is a collection of independent continuous
random variables
with density functions pXi . Let Z be their sum, Z = Xi . Then the distribution
of Z is given by the n-fold convolution, namely
pZ = pX1 ∗ pX2 ∗ · · · ∗ pXn (A.1)
in which the convolution of two functions f , g is defined as usual according to

f ∗ g(x) = f (u)g(x − u)du.+ Correspondingly,
, in terms of the characteristic
function, defined by ΦX := E eiωX , we have the n-fold product
n
C
ΦZ (ω) = ΦXi (ω). (A.2)
1
(The same formulae, of course, do not hold if Xi are statistically dependent.)

Consistently with the nomenclature, the property of stability concerns su-
perposition while infinite divisibility concerns decomposition of random variables
(distributions), and the component random variables that arise in the definitions
of each property are considered independent and identically distributed. Strictly
speaking, the two properties should correctly be considered as pertaining to dis-
tributions (not random variables).
A.0.0.1 Infinite divisibility A distribution F is said to be infinitely divisible

if, for each positive integer n, it arises as the n-fold convolution of a (distinct)
distribution F (n) with itself, i.e.
n terms
7 89 :
F = F (n) ∗ · · · ∗ F (n) . (A.3)
A.0.0.2 Stability The distribution F is said to be stable if, given {Xi } inde-
pendent and {Xi , X} identically distributed according to F , then for all choices
of constants a, b there exist constants c, d such that
.
aX1 + bX2 = cX + d (A.4)
.
with equality = in distribution.
By expressing these conditions for infinite divisibility and stability separately
in terms of characteristic functions, as explained above, one can deduce that
153
154 APPENDIX A
stability implies infinite divisibility. The converse does not hold, however. Some
pertinent cases are the following:
Gaussian. The Gaussian distribution is stable. This follows from the distribu-
tion properties:
X ∼ N (µ, σ 2 ),
1
pX (x) = √ exp(−(x − µ)2 /2σ 2 ),
2πσ 2
+ ,
ΦX (ω) = exp iµω − σ 2 ω 2 /2 .
Gamma. The gamma distribution, with basic properties
X ∼ G(α, β),
xα−1 e−x/β
pX (x) = ; x ≥ 0; α, β > 0,
Γ(α)β α
Γ(z + 1) = zΓ(z),
ΦX (ω) = (1 − iωβ)−α
is infinitely divisible. However, consistently with the above general observations,
it is not stable (as apparent from the characteristic functional form).
Cauchy
X ∼ G(α, β),
1 a
pX (x) = ,
π a2 + (x − b)2
ΦX (ω) = exp(ibω − a|ω|) .
This distribution (also referred to as the ‘Lorentzian’ distribution, of importance
in resonance theory) is stable (and therefore, like the Gaussian, also infinitely
divisible). Observe also that (for b = 0) the distribution of the sample mean µN =
N
i=1 Xi /N where Xi are drawn independently from a Cauchy distribution, has
the same distribution as Xi , so the ordinary central limit theorem drastically
fails.
On the basis of the same type of reasoning the K-distribution encountered
in the description of K-scattering is also infinitely divisible (but not stable). Of
mathematical interest, distributions for discrete multi-valued random variables
are neither infinitely divisible nor stable (as can be seen by considering the
possible sample values that occur under linear combinations as compared to
those drawn from the original distribution).
APPENDIX B
ITO VERSUS STRATONOVICH STOCHASTIC INTEGRALS
Let us compare the decompositions of the differential of a stochastic process Xt

in Ito versus Stratonovich terms:
(I)
b dt + σdWt
dXt = (B.1)
b(S) dt + σ ◦ dWt
in which b(I) , b(S) denote the drifts in the Ito and Stratonovich interpretations
respectively and ‘◦’ is a shorthand that indicates the Stratonovich prescription for
taking the stochastic integral, i.e. that the volatility is evaluated at the midpoint
of each subinterval. Now we can study the Stratonovich stochastic integral term
above and translate it into the Ito interpretation:

1
σ ◦ dWt = σ + (σ∂x σ)dWt dWt + o(dt1/2 ) (B.2)
2
in which the origin of the factor 12 on the right-hand side is that the midpoint of
the subinterval is being applied, in the sense explained above.46 Thus, neglecting
terms of o(dt1/2 ) we deduce that the Stratonovich and Ito stochastic integrands
are related by
1
σ ◦ dW = σdW + σ∂x σdt (B.3)
2
which follows from the relation dWt2 = dt. Hence we find that the two notions
of drift are related by
1
b(I) = b(S) + σ∂x σ. (B.4)
2
It is important to appreciate that the stochastic process Xt can be consistently

represented either way, provided one is clear which prescription for the stochastic
integral is being applied. Translation between the two representations occurs with
respect to the drift (only), and is according to (B.4).
46 A corresponding construction holds if the volatility is evaluated at t + αδt in the interval

i
[ti , ti + δt] thus generating a 1-parameter family of stochastic integrals – the cases of Ito and
Stratonovich then correspond to α equal to zero and 12 respectively.
155
156 APPENDIX B
It is of mathematical interest that the Stratonovich interpretation of the

stochastic integral leads to a calculus for which the ordinary rules of differenti-
ation apply. More precisely, for example, the ‘chain rule’ is valid so that, in the
Stratonovich interpretation, we have the classical differential formula
∂f
dFt = ◦ dXt . (B.5)
∂X
This result follows from the same type of argument as in (B.2) applied to f ◦
(σdW ), integrating the f factor over a semi-interval, which yields the second-
order derivative term 12 f σ 2 dt that is characteristic of the Ito calculus.
We introduce the martingale property for a stochastic process Mt
E[Mt |Fs ] = Ms , s ≤ t. (B.6)
Observe that the expectation of an Ito stochastic integral vanishes, i.e.

b
E φs dWs ≡ 0. (B.7)
a
Thus, Ito stochastic integrals are martingales, unlike their Stratonovich counter-
parts.
Finally, comparing with the development of stochastic differential geometry
in Chapter 3, first recall the relation between the Ito and Kolmogorov drifts on
(M, σ ij )
1
bi = β i + Γi (B.8)
2
where Γi = σ jk Γijk , the trace of the Christoffel symbol. Then in one-dimension
Γ1 = −σ∂x σ. (B.9)
Accordingly, we should compare (B.4) with the one-dimensional instance of

(B.8), namely
1
β = b + σ∂x σ. (B.10)
2
To see the consistency of these two points of view, observe that the Stratonovich
drift can be obtained (at a point of M) in a normal coordinated system (with
respect to σij ) and that the Stratonovich drift transforms homogeneously under
coordinate transformations xi
→ xî .
The interested mathematical reader should also compare Wong and Zakai
(1965) for a related discussion in the context of the relationship between ordi-
nary and stochastic differential equations (SDEs). We can consider an approxi-
(n)
mate solution to the SDE (B.1) generated by a sequence xt of solutions to the
APPENDIX B 157
(n)
ordinary differential equation dx(n) /dt = b + σΓ(n) in which Γt is a sequence
(n) t (n)
of processes whose integral Wt = Γs ds converges to the Wiener process
as n → ∞. Then the solution x(n) converges to the Stratonovich solution of the
SDE.
Remarks. On numerical integration. Sometimes it may be appropriate from
a numerical stability point of view to use a numerical integration (discrete time
step) scheme based on using the midpoint of each subinterval, the results of which
therefore converge to the solution of the Stratonovich SDE. When confronted
with an Ito SDE (B.1) with a volatility that is state dependent (and possibly non-
linear) one can, according to the above results, proceed to an accurate numerical
solution via a midpoint method applied to the following (Stratonovich) SDE,
which contains a modified drift according to the transformation (B.4),

1 ∂σ
dXt = b(I) − dt + σ ◦ dWt . (B.11)
2 ∂x
The solution thus generated is the desired approximation to the original Ito SDE.
Throughout the monograph the Ito interpretation is consistently adopted.

APPENDIX C
FILTRATIONS, CONDITIONAL PROBABILITY, AND
MARKOV PROPERTY
We let E[·|G] denote the expectation functional conditional on information rep-

resented by (the σ-algebra) G. Then if G ⊂ H, i.e. the information represented
by G is contained in that of H, then47 we have the iterated expectation rule
E [E[·|Ft ]|Fs ] ≡ E[·|Fs ]. (C.1)
In terms of time conditioned iterated expectations for a stochastic process

Xu we introduce the abstract notion of a ‘filtration’ Fs which (is a σ-algebra
that) represents the information contained in the history of the process up to
and including time s. Now, if we set G, H to correspond to Fs , Ft respectively,
for s ≤ t, then we obtain the tower law
Es ◦ Et ≡ Es , s≤t (C.2)
in which Es [·] ≡ E[·|Fs ].

If Es is defined as above, so that it is conditional on the history of the process,
then the tower law holds for a general process. Observe, on the other hand, if we
define Ěs as the expectation conditioned only at the instant s, then we require
the Markov property for the corresponding tower law to hold for Ě, i.e. that
Ěs Xu ≡ E[Xu |Fs ]. (C.3)
47 This follows essentially from the probability rule P[A|B]P[B] ≡ P[A ∩ B].
158
APPENDIX D
GIRSANOV’S THEOREM
Consider a process Xt that satisfies the stochastic differential equation (SDE)
dXt = bt dt + dWt , (D.1)
where Wt is a Wiener process, with respect to the probability measure P.

Consider now the process Yt := Mt Xt where Mt is the exponential change of
measure defined as
t
1 t 2
Mt := exp − bs dWs − bs ds . (D.2)
0 2 0
We deduce, via Ito’s formula, the stochastic differential of Mt
dMt = −bt Mt dWt (D.3)
and thus Mt is itself a martingale with respect to the measure P (it has vanishing
drift). It follows that the stochastic differential of Yt is given, according to the
Ito product rule, by
dYt = Mt (1 − bt Xt )dWt (D.4)
and hence Yt is also a P-martingale.

This line of argument leads to Girsanov’s theorem, which states that48 if we
make the Girsanov transformation of measure from P to Q on the space of paths
according to
dQ = MT dP, (D.5)
i.e. with Radon–Nikodym derivative dQ/dP = MT , then Xt is a Wiener process

with respect to the transformed measure Q defined on FT .
Intuitively, the theorem says that if we re-assign the probabilities of paths
according to Q, then with respect to this (ensemble average) measure the drift of
the process (the tendency for X to increase or decrease, on average) is eliminated.
This enables us to consider a single path of a stochastic process as an instance of
its drift-free counterpart. In other words, from the point of view of instantaneous
observable features of the process, the drift has no significance. It is only when
48 subject
to existence
of the change of measure, which is ensured by the Novikov condition,
E exp 12 0T b2s ds < ∞.
159
160 APPENDIX D
ensemble averages are applied that the drift concept enters into the description
and, in real time physics applications such as we have considered, it may be that
no such ensemble average is practicable or indeed available.
In contrast to the drift, the volatility is invariant under changes of measure
on path space. This is the mathematical basis of the stochastic volatility anal-
ysis applied to scattering, and implies that no ensemble average is required to
determine the volatility from a single realization (path) of a stochastic process.
Remarks. On coin tossing. As a useful simple analogy, we could consider toss-
ing a weighted coin – generate a (discrete-time/valued) process by starting at
the origin and moving up/down according as heads/tails is obtained. Suppose
now we observe a single path of this process, and that it tends to move upwards
much more than downwards. Then consider the question of whether, given the
observation of the single path, the coin is biased toward heads (positive drift) or
is a fair coin (no drift). The answer to this question is, of course, indeterminate
– the coin could either be fair and the observed path is an unlikely one, or it
could be a ‘typical’ path for a head-weighted coin. Indeed, one can see this intu-
ition reflected in the explicit expression for the change of measure (D.5) which
suppresses probabilities of paths with large drift.
D.0.1 Relation to mathematical finance

We remark briefly that exactly the same principle applies to financial data, which
is observed and hedged in real time.49 Accordingly the Black–Scholes option
price is drift independent, obtained in the ‘risk-neutral’ measure. The use of
this measure, as opposed to the original ‘real world’ measure, amounts to the
distinction between arbitrage versus expectation pricing. It is the former strategy
for which the Black–Scholes theory provides an explicit option price, and which
enables an individual in a market to eliminate risk in a random environment, by
hedging the observed stock based on their instantaneous values/fluctuations.
We refer the interested mathematical reader to Oksendal (1998) for a more
thorough and rigourous account of Girsanov’s theorem.
49 These remarks may be omitted without interrupting the main flow of the development.
They are mentioned for the purpose of highlighting the intimate connections that exist be-
tween mathematical physics and financial mathematics, and hopefully therefore to stimulate
interaction between the two scientific communities.
APPENDIX E
PARTITION FUNCTION SOLUTION TO BDI MODEL
The ‘partition function’ for the birth–death–immigration (BDI) process satisfies

the following differential equation (Bartlett 1966):
∂Πt (z) ∂Πt (z)

= (z − 1)(λz − µ) − ν(1 − z)Πt (z) (E.1)
∂t ∂z
in which the partition function is defined by Πt (z) = z Nt .50 This equation has
the finite-time closed form solution
(λ − µ)ν/λ [µ(T − 1) − (µT − λ)z]n
Πt (z) = (E.2)
[λT − µ − λ(T − 1)z]n+ν/λ
in which we abbreviate T = exp ((λ − µ) t) and n is the specified number in the

population initially. The condition for the existence of an asymptotic equilibrium
solution (distribution) is µ > λ which ensures that the function T (t) decays to
zero as t → ∞. In this limiting case, recall that the negative binomial variate N
has probability mass function

α+N −1
pN = pα q N (E.3)
N
for integer N ≥ 0 with partition function

α
p
Π∞ (z) = (E.4)
1 − qz
to be compared with the special case of the Poisson variate for which
Π∞ (z) = exp[−λ(1 − z)]. (E.5)
Thus in terms of the BDI parameters, we identify p = 1−λ/µ, q = λ/µ (p+q = 1)

and the partition function in the general case is given by
ν/
µ−λ λ
Π∞ (z) = (E.6)
µ − λz
so α = ν/λ.
√
50 If
we replace the variable z by e −1ω then the characteristic function Φt (ω) is immediately
obtained.
161
162 APPENDIX E
In case µ < λ then T (t) grows progressively larger with time. In all cases,
Πt (1) = 1 so that probability is conserved. The expressions for the mean and
variance of the population remain positive and increase with time, the normalized
variance tending to a limiting value, which we read off as
2
λ − µ2 n + λν
2 . (E.7)
((λ − µ) n + ν)
Observe that this depends on the initial population n. However, the mean value of
the population increases without bound; any value might reasonably be chosen as
n at an arbitrarily chosen ‘initial’ time. The master equation does not establish an
equilibrium population, and accordingly, the population is prevented from being
asymptotically stationary. Therefore, these values of the underlying population
rate parameters can be disregarded as a description of a physical system in
equilibrium, such as we have encountered in practice.
Remarks. On pure birth-death processes. An intriguing feature emerges in the

case of a pure birth-death process when these rates coincide, with respect to
the first two moments. From the partition function finite-time exact solution we
deduce the following expressions for the mean and variance of the population,
at time t,
mt = exp(λ − µ)t, (E.8)
λ+µ + ,
λ−µ e
(λ−µ)t
e(λ−µ)t − 1 , if λ = µ,
var(t) = (E.9)
2λt if λ = µ.
Observe, therefore, that if λ = µ then as time progresses the mean remains at

a constant value of unity and the variance grows without bound. However, it is
evident from the exact partition function solution in this case
1 − (λt − 1)(z − 1)
Πt (z) = (E.10)
1 − λt(z − 1)
that the extinction probability, conveniently obtained by setting z = 0 in the

time-dependent partition function,
p0 (t) ≡ Πt (0) (E.11)
tends to unity as t → ∞. The population is therefore said to become asymptot-

ically extinct.
This fact would seem to be irreconcilable with the properties of the first two
moments noted above, namely that the mean population size is a constant unity
and that its variance tends to infinity. The resolution of this apparent paradox
lies in considering the detailed behaviour of the occupation probabilities pN (t).
APPENDIX E 163
Although these probabilities decay to zero with time for all N ≥ 1, the rate at
which they decay decreases with time, in such a way that the noted behaviour of
the first and second moments is admissible. To see this explicitly, we may extract
the pN (t) from the partition function by writing (E.10) in the form
−1
λt 1 − λt λt
Πt (z) = + z 1− z . (E.12)
1 + λt 1 + λt 1 + λt
The probabilities pN (t) can then be obtained as the coefficients of z N in a bino-
mial series expansion of the above expression for Πt (z), whereby we obtain
N
1 λt
pN (t) = . (E.13)
λt(1 + λt) 1 + λt
Thus,
with respect to N , the population has a negative exponential decay rate
λt
log 1+λt that tends to zero as t → ∞.51
The extinction property is also interesting in terms of a sample path of the
population process Nt . Given a death rate perfectly counterbalanced by an equal
birth rate, one might perhaps intuitively expect the population level to remain
stable, in some sense, fluctuating around some positive value. The apparent para-
dox is resolved by realizing that, at some finite time and with probability 1, the
population level will hit Nt = 0 and cannot regain any positive value subse-
quently, because of the absence of immigration; i.e. a situation of population
extinction. In other words, although the transition rates are precisely symmet-
ric with respect to increasing and decreasing Nt , the property of extinction is
asymmetric in this sense, being represented by the boundary at N = 0.
Special cases of the BDI model where a limiting asymptotic distribution exists
can also be calculated for (limiting) values of the population parameters. This
question is best addressed via analysis of the characteristic or partition function.
We explain the behaviour for various values of the parameter α = ν/λ.52
E.0.1.1 α = 1: In this case we obtain the geometric distribution, and accord-
ing to the results of Section 7.3, this yields the exponential distribution in the
continuum limit.
E.0.1.2 α → 0: (i) if ν → 0, with λ, µ fixed, then Π → 1 so that the distribu-
tion is concentrated at N = 0, i.e. extinction . (ii) if λ → ∞, with ν fixed (so
µ → ∞ to maintain N > 0); say λ/µ → k ≤ 1, then for all values of k the
partition function tends to unity, so we have extinction as in (i). A case of inter-
est arises if µ = λ + 1 so that N = ν; nevertheless, Π → 1,53 i.e. asymptotic
extinction.
51 property mt = 1 can be verified from (E.13) using the series identity
∞TheNconstant mean
1 Nζ ≡ ζ/(1 − ζ)2 .
52 Recall that, for a limiting population to exist, N̄ = ν/(µ − λ) for values µ > λ.
53 This can be verified using the fact lim u
u→0 u = 1.
164 APPENDIX E
E.0.1.3 α → ∞: (i) if ν → ∞ with λ fixed, the Fano factor F > 1 and the
average population size N → ∞; then the partition function Π → 0 (for z = 1)
so that the distribution is negative binomial for all values of ν with pN → 0, ∀N .
(ii) if λ → 0 with ν fixed, then F → 1 and the limiting distribution is Poisson
with mean ν/µ.54
54 By virtue of the identity lims→0 (1 + µs)1/s = exp µ.

APPENDIX F
SUMMARY OF K-SCATTERING
It is shown in the Rayleigh case of a fixed step number that the amplitude
obeys a complex Ornstein–Uhlenbeck equation, and a corresponding stochastic
differential equation (SDE) in the K-distributed case is derived.
F.1 Rayleigh scattering

In the Rayleigh case, consider the random walk model for the scattered electric
field (cf. Jakeman 1980; Tough 1987; Jakeman and Tough 1988) with step s(j)
s(j)
N 7 &89 ':
(N )
(j)
Et = exp iϕt (F.1)
j=1
for constant population size N . Since Maxwell’s equations for the electromag-
netic field possess U (1) gauge invariance with respect to duality rotations, i.e.
multiplication by exp(iΛ) for constant
( ) Λ (cf. Penrose and Rindler 1984), the
assumption of independence of ϕ(j) implies that these
&phases
'are uniformly
(j)
distributed. Accordingly, in (F.1) the phase factors exp iϕt are indepen-
dent and uniformly distributed
on the unit circle in C. Our (phase) diffusion
(j)
model therefore takes ϕt as a collection of (displaced) Wiener processes
(j) 1 (j)
on a suitable
( ) timescale, ϕt = ∆(j) + B 2 Wt , with the random initializations
(j)
∆ a set of independent random variables uniformly distributed on the in-
(j) 1 (j) (j)2
terval [0, 2π), and thus dϕt = B 2 dWt , dϕt = Bdt. From Ito’s formula (e.g.
Oksendal 1998; Karatzas and Shreve 1988) the Ito differential of (F.1) is
N & '
(N ) (j) 1 (j) (j)
dEt = idϕt − dϕt 2 exp iϕt . (F.2)
j=1
2
N & '
(j) (j)
The first term j=1 idϕt exp iϕt on the right-hand side of (F.2) consists of
a sum of independent randomly phased Wiener processes, with variance equal to
BN dt, while the second term is independent of the scatterer label j. Thus from
(F.2) we can write
(N ) 1 (N ) 1
dEt = − BEt dt + (BN ) 2 dξt (F.3)
2
* (j)|dξt | = dt,
2
where ξt is a complex Wiener process satisfying dξt2 = 0. The
process ξt is adapted to the filtration F (ϕ)
= j F , where F is the filtration
(j)
165
166 APPENDIX F
(j)
appropriate to the component scatterer phase
& ϕt . The' (normalized) amplitude
(N ) 1
process Et is then defined by Et = limN →∞ Et /N̄ 2 and satisfies the SDE
1 1
dEt = − BEt dt + (Bx) 2 dξt (F.4)
2
where the continuous-valued random variable x, the average scattering
+ , power,
arises from an asymptotically large population via x = limN →∞ N/N̄ .
F.2 K-distributed noise

In the case of step number fluctuations in the random walk model (F.1), we
define the K-amplitude ψt as a modification to the Rayleigh+amplitude,
, such
that we employ a time-dependent Nt such that xt = limNt →∞ Nt /N̄ . Thus
 
 1 Nt & '
(j)
ψt = lim exp iϕt (F.5)
Nt →∞  N̄ 1/2 
j=1
 
 N 1/2 1 Nt & '
t (j)
= lim exp iϕt (F.6)
Nt →∞  N̄ N
1/2 
t j=1
1
2
= xt γt , (F.7)
& 1
'
where γt = limN →∞ E (Nt ) /Nt2 . Thus the intensity has the compound repre-
sentation zt = xt ut where ut = |γt |2 is the (instantaneous) intensity of the
component Rayleigh process. According to the arguments given in the Rayleigh
case above, γt is a complex Ornstein–Uhlenbeck process which obeys the SDE
1 1
dγt = − Bγt dt + B 2 dξt . (F.8)
2
Observe from (F.4), therefore, that γt is a unit power Rayleigh process. The
above
+ , for γt can be solved by considering the stochastic differential
equation
d exp 12 Bt γt , which leads to the solution
t
1 1 1
γt = exp − Bt γ0 + B 2 exp Bs dξs . (F.9)
2 0 2
We deduce the expectation formulae

1
E[γt ] = exp − Bt γ0 , (F.10)
2
+ 2,
E |γt | = 1 + exp(−Bt)(|γ0 |2 − 1). (F.11)
+ 2,
It follows that limt→∞ E |γt | = 1 and so from (F.7) we find the intensity
process, defined by zt = |ψt |2 , satisfies the conditional expectation property
K-DISTRIBUTION 167
E|x [zt ] = xt . The SDE for ψt can then be derived by applying the Ito prod-
uct formula to (F.7). This requires the SDE for the scattering cross-section
to be specified. In accordance with the birth–death–immigration (BDI) model
(Bartlett 1966), we shall take the re-scaled population variate x
→ αx to satisfy
the SDE
1 (x)
dxt = A(α − xt )dt + (2Axt ) 2 dWt (F.12)
(x)
for an independent Wiener process Wt (Field and Tough 2003a). (In terms
of the underlying population parameters of the BDI model, α = ν/λ, the ratio
of the immigration to birth rate, the birth and death rates coinciding for an
infinite-sized population.) Accordingly, xt has an asymptotic Γ-distribution
xα−1 exp(−x)
Γα (x) = (F.13)
Γ(α)
with first two moments Var[x] = x = α. These preliminaries enable us to

provide the dynamics of the amplitude, intensity, and phase of the K-scattering
process as follows (cf. Field and Tough 2003b for a detailed treatment).
F.2.1 Amplitude
Theorem F.1 The K-amplitude is governed by the SDE
1 12
dψt 1 B2 2(α − xt ) − 1 A (x)
= Bdt + dξt + A dt + dWt . (F.14)
ψt 2 γt 4xt 2xt
This evolution is invariant under the U (1) gauge transformation ψt

→ exp(iΛ)ψt ,
for constant Λ.
F.2.2 Intensity
Proposition F.2 The K-intensity SDE is given by
1
Azt (α − xt ) 2Azt2 2 (z)
dzt = B(xt − zt ) + dt + 2Bxt zt + dWt (F.15)
xt xt
(z) (x)
in which Wt is correlated with Wt of (F.12), and satisfies
12 12
2Azt2 (z) 1 (r) 2A (x)
2Bxt zt + dWt = (2Bxt zt ) 2 dWt + zt dWt (F.16)
xt xt
(r)
and Wt is a real-valued Wiener process defined by
12
2zt
γt∗ dξt γt dξt∗
(r)
+ ≡ dWt . (F.17)
xt
168 APPENDIX F
F.2.3 Phase
Proposition F.3 The resultant phase θt of the K-amplitude process satisfies
the SDE
12
Bxt (θ)
dθt = dWt (F.18)
2zt
(θ)
where the distinct (real-valued) Wiener process Wt is defined according to
12
1 ∗ 2zt
(θ)
dWt . (F.19)
i t xt
These various relations allow the K-amplitude dynamics to be recast in terms

of W (x) , W (θ) , W (r) as follows.
Corollary F.4 The K-amplitude satisfies the SDE
1 1
dψt 2(α − xt ) − 1 1 Azt + Bx2t 2 (z) Bxt 2 (θ)
= A − B dt + dWt + i dWt
ψt 4xt 2 2xt zt 2zt
(F.20)
in which, alternatively, the Wiener terms can be expressed as

12 12
A (x) Bxt (r) (θ)
dWt + dWt + idWt . (F.21)
2xt 2zt
(x)
The following result, implied by (F.14) and the identities dξt2 = dξt dWt = 0,
is relevant in connection with the geometry of fluctuations for weak scattering
processes discussed in Section 9.3.
Corollary F.5 The product cross-section/K-amplitude stochastic differentials
satisfy
dx2t = 2Axt dt, (F.22)

dxt dψt = Aψt dt, (F.23)

Aψt2
dψt2 = dt, (F.24)
2xt

Azt
|dψt |2 = + Bxt dt. (F.25)
2xt
F.2.4 Geometry of fluctuations

(z) (θ)
We observe from Propositions F.2 and F.3 that dWt dWt = 0, so the fluctua-
tions in Rt , θt are statistically independent. The relative magnitude of the radial
and orthogonal (phase) fluctuations is determined by
K-DISTRIBUTION 169
(z)
Σt 4Azt3
= 4zt2 + (F.26)
(θ)
Σt Bx2t
which exceeds the quotient obtained in the Rayleigh case, A = 0. These rela-
tions can be used to characterize the geometry of the K-scattering amplitude
fluctuations as follows. The real and imaginary parts of the resultant amplitude
I, Q are the usual and quadrature-phase components, respectively.
Proposition F.6 In the K-distributed case, A = 0, the amplitude diffusion
tensor is non-degenerate, and the fluctuations in the and quadrature phase com-
ponents δIt , δQt are correlated. The (comoving) error surface S of δψt is an
ellipse whose major axis lies in the instantaneous radial direction defined by ψt .
Degeneracy occurs in the Rayleigh case A = 0, for which S is a circle, i.e. the
fluctuations in ψt are isotropic.
A complete account of the dynamical properties of K-scattering is given in

Chapter 8 (orig. Field and Tough 2003b).
APPENDIX G
ITERATIVE SOLUTION FOR VECTOR PROCESSES
The generalization to a vector process is straightforward. We consider the vector

x with components labeled with Greek superscripts, i.e. xα ; the drift F is also
a vector. The Brownian process is a vector also and we label its components
with Greek superscripts, while the time variable appears as a Latin subscript,
thus Wtα . Finally, the volatility is now a matrix represented by σβα . Derivatives
with respect to the components of x are denoted in standard tensor notation as
follows
α ∂F α α
∂σβα
F,β = ; σβ,γ = . (G.1)
∂xβ ∂xγ
We then have
dxα = F α (x (t)) dt + σβα (x (t)) dWtβ (G.2)
with summation on repeated Greek indices, according to the Einstein summation

convention introduced in Chapter 3. The corresponding integral equation is
t t
α α
∆x = F (x (u))du + σβα (x (u)) dWuβ . (G.3)
0 0
The iterative solution to this integral equation can be developed exactly as it

was in the scalar case. To second order in time we find this to be, having taken
care to place terms in the order corresponding to the scalar solution (11.16),
t
1
α
∆x = tF + α
σβα Wtβ + t2 F γ F,γ
α
+ στµ F,µ
α
Wsτ ds
2
0
t t t
1
+ F µ σβ,µ
α
sdWsβ + στµ σβ,µ
α
dWsβ Wsτ + στµ σσν F,µν
α
Wsτ Wsσ ds
2
0 0 0
t s t
1
+ στµ σσν σγ,µν
λ α
σβ,λ dWsβ dWuν Wuτ Wuσ + στµ F ν σβ,µν
α
sWsτ dWsβ
2
0 0 0
170
APPENDIX G 171
t s t s
+ στµ σσ,µ
ν
F,να ds dWuσ Wuτ + στµ σσν σγ,ν
λ α
σβ,µλ dWsβ Wsτ dWuγ Wuσ
0 0 0 0
t t s
1
+ στµ σσν σγλ σβ,µνλ
α
dWsβ Wsσ Wsτ Wsγ + στµ σσ,µ
ν α
σβ,ν dWsβ dWuσ Wuτ
6
0 0 0
t s t s
+ στµ F,µ
ν α
σβ,ν dWsβ Wuτ du + F µ στ,µ
ν α
σβ,λ dWsβ udWuτ
0 0 0 0
t
1
+ στµ σγν σβ,µν
α
dWsβ Wsτ Wsγ
2
0
t s u
+ στµ σσ,µ
ν λ
σγ,ν α
σβ,λ dWsβ dWuγ dWvσ Wvτ . (G.4)
0 0 0
The calculation of the expectation values for the increments and their products
proceeds in much the same way as the scalar case; some care is needed in the
ordering of the various indices. Thus we write the stochastic part of the vector
increment as
1
∆x̂α = ∆xα − tF α − t2 F γ F,γ
α
. (G.5)
2
In a similar way to the scalar case, we have
t2 µ ν α
E (∆x̂α ) = σ σ F (G.6)
4 τ τ ,µν
while the covariance matrix is given by
t2 λ β α t2 λ β α
E ∆x̂α ∆x̂λ = tσβα σβλ + σγ σγ F,β + σγα σγβ F,β λ
+ σγ F σγ,β + σγα F β σγ,β
λ
2 2
t2 t2
+ σγλ σβσ σβµ σγ,σµ
α
+ σγα σβσ σβµ σγ,σµ
λ
+ σγβ σγσ σν,β
α λ
σν,σ .
4 2
(G.7)
Extension of these results beyond second order in time in the general multi-
plicative noise case, although proceeding along similar lines, is rather computa-
tionally heavy and is not considered here.
APPENDIX H
OPEN PROBLEMS
1. Construction of scattering dynamics for populations arising from higher order

transitions and/or non-linearity in the generation and recombination coefficients.
Investigation of the continuum limit therein (in relation to the Pawula theorem).
2. Extension of the local population model to a spatially correlated one, induced
by the effect of inter-site migration, on a discrete lattice of sites and the contin-
uum limit thereof, in relation to path integrals. The projection of such spatial
model to single site and the relationship with the BDI model.
3. Optimal inference of the scattering cross-section for given sample rate through
adjustment of smoothing parameters, for K-scattering processes, and the exten-
sion to more general population dynamics.
4. Filtering of additive noise in superposition with the scattered amplitude pro-
cess for K-scattering and weak scattering processes, and the extension to more
general population models.
5. Coherent signal extraction for weak scattering processes, de-noising, and the
systematic analysis of performance and error bounds for such procedures.
6. Estimation of the scattering cross-section for weak scattering models through
analysis of fluctuations, as already achieved for the K-scattering model, to in-
clude general populations.
7. Formulation of a stochastic volatility based anomaly detection scheme for
weak scattering processes, as an extension of the corresponding results achieved
for K-scattering.
8. Incorporation of the stochastic dynamical model for radar scattering into a
Bayesian detection/tracking algorithm for sea clutter environments and perfor-
mance comparison with the stochastic volatility based anomaly detection scheme
(e.g. via computation of the associated received operator characteristic curves).
9. To investigate the possibility of chaotic behaviour, in the deterministic non-
linear evolution of the ‘noise free skeleton’ of the coupled amplitude/cross-section
dynamics.
10. To explore the fundamental physical limits that exist with respect to the
observability of the scattering cross-section from the intensity and phase fluctu-
ations, in a variety of contexts.
172
APPENDIX H 173
11. To establish the relationship between the z-parameter quantifier of non-

linearity and the shape parameter of the SDE radar scattering model, from
a theoretical point of view; to exhibit such relationship by evaluating the z-
parameter on a data set synthetically generated from the SDE model.
APPENDIX I
SUGGESTED FURTHER READING
[A] A. Ishimaru, Wave Propagation and Scattering in Random Media, Volume

I: Single scattering and transport theory; Volume II: Multiple scattering, tur-
bulence, rough surfaces and remote sensing, Academic Press, New York, 1978.
(Translated into Russian, 1981 and Chinese, 1986.) IEEE Press-Oxford Univer-
sity Press Classic Re-issue, 1997.
[B] A. Ishimaru, Electromagnetic Wave Propagation, Radiation, and Scattering,
Prentice Hall, New Jersey, 1991.
[C] Mauro Fabrizio, Angelo Morra, and Angelo Morro, Electromagnetism of
Continuous Media: Mathematical Modelling and Applications, Oxford Univer-
sity Press, July 2003, ISBN: 0198527004.
[D] Eric Jakeman and Kevin D. Ridley Modeling Fluctuations in Scattered Waves,
Taylor and Francis, June 19, 2006. ISBN: 0750310057.
[E] Keith D. Ward, Robert J.A. Tough, and Simon Watts Sea Clutter: Scattering,
the K-distribution and Radar Performance, IET Press, June 16, 2006, ISBN: 0-
86341-503-2, 978-086341-503-6.
Items [A, B] are important texts by a leading international expert in the field,
on the subject of electromagnetic scattering from random media. However, there
is no coverage of the approach using stochastic calculus, or of the K-scattering
process and its extensions to weak scattering models.
Item [C] deals with wave propagation in continuous electromagnetic media
including effects of non-locality and ‘memory’ – no coverage of scattering from
random media. Complements the proposed monograph well. An alternative re-
cent addition to the literature on the subject, item [D], should be an important
text by a pioneer in the field of non-Gaussian models of scattered radiation.
Mathematically, it is less sophisticated than the present monograph, covering
basic statistical models, Gaussian, Gamma, K-lognormal etc., random walks,
phase screen scattering, propagation through extended media, multiple scat-
tering, vector wave scattering, K-distributed noise, experimental artifacts, and
numerical simulation. The book is aimed at the applied physics/systems engi-
neering user and thus complements the present more theoretical treatment well.
Indeed one chapter describes K-distributed noise and includes mention of some
of the recent work of the current author in the journal literature, in abbreviated
and simplified form.
The most recent addition to the literature in book form, item [E], is devoted
to the study of sea clutter: scattering, the K-distribution and radar performance.
174
APPENDIX I 175
The text gives an authoritative account of our current understanding of radar sea
clutter from an engineering perspective. The authors pay particular attention to
the compound K-distribution model, which they have helped develop over the
past two decades. Evidence supporting this model, including a detailed review
of the calculation of EM scattering by the sea surface, its statistical formulation,
and practical application to the specification, design and evaluation of radar sys-
tems are included. The calculation of the performance of practical radar systems
is emphasized. This book provides a less mathematically sophisticated treatment
in the radar context which will serve as an excellent complement to the current
monograph for specialist radar engineers, and also be of interest to the wider
applied physics and mathematics academic communities.
REFERENCES
Abramowitz, M. and Stegun, I. A. (Eds.) (1970). Handbook of Mathematical

Functions. Dover, New York.
Bachelier, L. (1900). Théorie de la spéculation, Annales Scientifiques de l’École
Normale Supérieure 3 (17), 21–86.
Bartlett, M. S. (1966). An Introduction to Stochastic Processes, ch. 3, Cam-
bridge University Press.
Batchelor, G. K. (1967). An Introduction to Fluid Dynamics, Cambridge Uni-
versity Press.
Copson, E. T. (1970). An Introduction to the Theory of Functions of a Complex
Variable, Section 10.63. Clarendon Press, Oxford.
Einstein, A. (1905). On the movement of small particles suspended in a sta-
tionary liquid demanded by the molecular-kinetic theory of heat. Re-printed
in Investigations on the Theory of the Brownian Movement, A. Einstein (Au-
thor), R. Fürth (Editor), A. D. Cowper (Translator) Publisher: Dover (1956).
ASIN: B000F62YS0.
Ermak, D. L. and McCammon, J. A. (1978). J. Chem. Phys. 59, 1352.
Ernst, R. R., Bodenhausen, G. and Wokaun, A. (1987). Principles of Nuclear
Magnetic Resonance in One and Two Dimensions, Oxford: Clarendon Press.
Fano, U. (1947). Phys. Rev. 72, 26–29.
Fayard, P. (2008). Private communication.
Fayard, P. and Field, T. R. (2008). Optimal inference of the scattering cross-
section through the phase decoherence, Waves in Random and Complex Media,
in press.
Feng T., Field, T. R. and Haykin, S. (2007). Stochastic differential equation
theory applied to wireless channels, IEEE Transactions on Communications,
55 (8), 1478–1483.
Feng, T. and Field, T. R. (2008). Statistical analysis of mobile radio reception –
an extension of Clarke’s model, IEEE Transactions on Communications, in
press.
Field, T. R. (2002). Stochastic differential equations and their application to
the characterization of sea clutter , Invited Lectures, October 7–8, Adaptive
Systems Laboratory, McMaster University.
Field, T. R. (2003). Quantum diffusion on manifolds. J. Geom. Phys. 47, 484–
496.
Field, T. R. (2005). Observability of the scattering cross-section through phase
decoherence. J. Math. Phys. 46, 063305.
Field, T. R. (2006). Spin diffusion – a new perspective in magnetic resonance
imaging. Ch. 5 in New Directions in Statistical Signal Processing: From Sys-
tems to Brain, MIT Press.
177
178 REFERENCES
Field, T. R. and Bain, A. D. (2008). Origins of stochastic spin noise dynamics

and implications for experimental NMR, Phys. Rev. E, submitted; 49th ENC,
March 9–14, 2008, Asilomar Conference Grounds, Pacific Grove, California.
Field, T. R. and Haykin, S. (2008). Non-linear dynamics of sea clutter. Inter-
national Journal on Navigation and Observation, Special Issue on Modelling
and Processing of Radar Signals for Earth Observation; M. Greco, S. Watts
(eds.), forthcoming.
Field, T. R. and Tough, R. J. A. (2003a). Diffusion processes in electromagnetic
scattering generating K-distributed noise, Proc. R. Soc. Lond. A 459, 2169–
2193.
Field, T. R. and Tough, R. J. A. (2003b). Stochastic dynamics of the scatter-
ing amplitude generating K-distributed noise, J. Math. Phys. 44 (11), 5212–
5223.
Field, T. R. and Tough, R. J. A. (2005). Dynamical models of weak scattering.
J. Math. Phys. 46 (1), 13302–13320.
Fox, L. and Mayers, D. F. (1968). Computing methods for scientists and engi-
neers. Chapter 10, Clarendon Press, Oxford.
Gini, F. and Greco, M. (2001). Texture modeling and validation using recorded
high resolution sea clutter data. Proc. 2001 IEEE Radar Conference, Atlanta,
GA, May 1–3 2001, pp. 387–392.
Greco, M. and Gini, F. (2007). Sea clutter nonstationarity: the influence of long
waves. In S. Haykin (ed.), Adaptive Radar Signal Processing, Wiley; Chapter
5, pp. 159–191.
Greenside, H. S. and Helfand, E. (1981). Numerical integration of stochastic
differential equations – II, Bell Syst. Tech. J., 60, 1927–1940.
Haykin, S. (1999). Radar clutter attractor: implications for physics, signal pro-
cessing, and control. IEE Proceedings on Radar, Sonar and Navigation, Vol.
146, No. 4, pp. 177 , August, Invited.
Haykin, S. (2001). Communication Systems, 4th edn, Wiley (New York).
Haykin, S. (2006) (ed.) Adaptive Radar Signal Processing, Wiley-Interscience
ISBN-10: 0471735825; Chapter 4.
Haykin, S., Bakker, R. and Currie B. W. (2002). Uncovering nonlinear
dynamics – The case study of sea clutter. Proc IEEE 90 (5), 860–881.
Haykin, S. and Thomson, D. J. (1998). Signal detection in a nonstationary
environment reformulated as an adaptive pattern classification problem. Proc.
IEEE, Vol. 86, pp. 2325–2344.
Helstrom, C. W. (1960). Statistical Theory of Signal Detection, Pergamon,
Oxford.
Higham, D. J. (2004). An Introduction to Financial Option Valuation: Mathe-
matics, Stochastics and Computation, Cambridge University Press.
Howison, S. (2004). Private communication.
Hughes, A. J., Jakeman, E., Oliver C. J. and Pike, E. R. (1973). Photon cor-
relation spectroscopy; dependence of linewidth error on normalisation, clip
REFERENCES 179
level detector area, sample time and count rate. J. Phys. A, Math. and Gen.
6, 1327.
Hughston, L. P. (1996). Geometry of stochastic state vector reduction Proc. R.
Soc. London, Ser. A 452, 953.
Jakeman, E. (1980). On the statistics of K-distributed noise. J. Phys. A 13,
31-48.
Jakeman, E., Hopcraft, K. I. and Matthews, J. O. (2003). Distinguishing popu-
lation processes by external monitoring. Proc. R. Soc. Lond. A, 459, 623–639.
Jakeman, E. and Tough, R. J. A. (1987). The generalised K-distribution: a
statistical model for weak scattering. J. Opt. Soc. Am., A4, 1764.
Jakeman, E. and Tough, R. J. A. (1988). Non-Gaussian models for the statistics
of scattered waves. Adv. Phys. 37, 471.
Jakeman, E., Watson, S. M. and Ridley, K. D. (2001). Intensity-weighted phase-
derivative statistics. J. Opt. Soc. Am. A 18 (9), 2121–2131.
Jeffreys, H. and Jeffreys, B. S. (1966). Methods of Mathematical Physics, 3rd
Edn., Cambridge University Press.
Karatzas, I. and Shreve, S. E. (1988). Brownian Motion and Stochastic Calcu-
lus. Berlin: Springer.
Klauder, J. R. and Petersen, W. P. (1985). Numerical integration of
multiplicative-noise stochastic differential equations. SIAM J. Numerical Anal-
ysis 22, 1153–1166.
Kramers, H. A. (1940). Physica 7, 284.
Luttrell, S. P. (2001). Private communication.
Moyal, J. E. (1949). J. Roy. Statist. Soc. 11, 1358.
Nelson, E. (1967a). Dynamical Theories of Brownian Motion, Princeton Uni-
versity Press.
Nelson, E. (1967b). Tensor Analysis, Princeton University Press.
Nelson, E. (1985). Quantum Fluctuations, Princeton University Press.
Oksendal, B. (1998). Stochastic Differential Equations – An Introduction with
Applications, 5th Edition, Springer.
O’Loghlen, J. W. (2001). Private communication.
Penrose, R. and Rindler, W. (1984). Spinors and Space-time, Vol. 1, Cambridge
University Press.
Pusey, P. N. and Tough, R. J. A. (1982). Langevin approach to the dynamics
of interacting Brownian particles. J. Phys. A, 15, 1291–1308.
Quiang, J. and Habib, S. (2000). A second order stochastic leapfrog algorithm
for multiplicative noise Brownian motion. arXiv:physics/9912055v2 , 4th Jan.
2000.
Rice, S. O. (1954). Mathematical analysis of random noise. Bell. Syst. Tech.
J., 23, 24, reprinted in Selected Papers on Noise and Stochastic Processes, N.
Wax, Ed., Dover, New York.
Ridley, K. D., Watson, S. M., Jakeman, E. and Harris, M. (2002). Heterodyne
measurements of laser light scattering by a turbulent phase screen. Applied
Optics 41, No. 3, 532–542.
180 REFERENCES
Risken, H. (1989). The Fokker–Planck Equation, 2nd edn., Springer.

Siegel, A. (1956). Non-parametric Statistics for the Behavioral Sciences,
McGraw–Hill (New York).
Stone, L. (1992). Coloured noise or low-dimensional chaos? Proc. R. Soc. Lond.
B 250, 77–81.
Sugihara, G. (1994). Nonlinear forecasting for the classification of natural time
series. Phil. Trans. R. Soc. Lond. A 348, 477–495.
Sugihara, G. and May, R. (1990). Nonlinear forecasting as a way of distin-
guishing chaos from measurement error in a data series. Nature, Lond. 344,
734–741.
Tough, R. J. A. (1987). A Fokker–Planck description of K-distributed noise. J.
Phys. A 20, 551–567.
Tough, R. J. A. (1991). Interferometric detection of sea surface features. R.S.R.E.
Memorandum 4446 .
Van Kampen, N. G. (1961). Power series expansion of the master equation,
Can. J. Phys. 39, 551–567.
Van Kampen, N. G. (1981). Stochastic Processes in Physics and Chemistry.
North Holland, Amsterdam.
Ward, K. D. (1981). Compound representation of high resolution sea clutter.
Electronics Letters 17, 561.
Ward, K. D. (2002). Private communication.
Whittaker, E. T. and Watson, G. N. (1969). A Course of Modern Analysis,
Chapter XI, Cambridge University Press.
Wolfram, S. (1999). The Mathematica Book , Cambridge University Press, 4th
edition.
Wong, E. (1963). Proc. Am. Math. Soc. Symp. Appl. Math. 16, 264.
Wong, E. and Zakai, M. (1965). On the relation between ordinary and stochastic
differential equations. Int. J. Engng Sci. Vol. 3, pp. 213–229, Pergamon Press.
Wootters, W. K. (1981). Statistical distance and Hilbert space, Phys. Rev. D
23, pp. 357–362.
INDEX
amplitude, 16 clutter, 112, 113, 127

K-, 53 radar, 127, 138, 140
weak scattering, 70 scattering, 52
amplitude fluctuations, 77, 126, 169 superposition, 25, 40, 68, 70
K-scattering, 59, 74 compound model, 90, 91
weak scattering, 74 for K-scattering, 66, 67, 109
geometry of, 26, 59, 74 for weak scattering, 80–83
anomaly detection, 25, 27, 34, 52, 55, 67, contravariant
69, 79, 86, 87, 97, 111, 112, diffusion tensor, 12, 13, 60
127, 132–134, 148 correlation, 20
asymptotic, 25 function, 35, 38, 40, 42, 64, 67, 111,
distribution, 25, 27, 29, 30, 35, 37, 39, 112, 122, 127
44, 47, 49, 54, 59–61, 64, 69, standard statistical measure of, 124
70, 79, 81, 83, 84, 86, 90, 161, timescales, 31, 35, 55, 58, 108, 143, 146
163, 167 covariant, 14
stationarity, 162 derivative, 14, 16
asymptotic Fokker–Planck equation, 14, 18, 60, 62
distribution, 41 cross-section
asymptotic behaviour
dynamics, 30
weak scattering, 79
as limit of BDI process, 48, 49
autocorrelation, 5, 52, 53, 63–67, 142
generalized, 91
of Wiener process, 5
estimation of, 150, 151
instantaneous observability of, 150
Bachelier radar, 66, 138, 139, 141
Louis, 3 scattering, 20, 25, 26, 28–30, 35, 38, 46,
Bessel 52, 53, 55, 61, 67, 88, 93, 130,
function, 28, 63, 79, 107, 119, 122 131, 133, 134, 137, 144,
process, 21, 22, 27 150–152, 167
birth-death-immigration (BDI), 46, 47 current
continuum diffusion limit of, 48 stochastic, 5, 12, 17, 18, 31, 35, 36, 38,
Black–Scholes 39, 60, 62, 72, 85, 86
theory of option pricing, 21, 28, 34, 160
Brownian motion, 3, 5
and Wiener process, 5, 7, 9 detailed balance, 12, 18
geometric, 21 in K-scattering models, 62
in weak scattering models, 85
Cauchy distribution, 67, 154 deterministic dynamics, 42, 108, 120, 126,
stability of, 66 143, 148
central limit theorem, 6, 28, 53, 90, 91, diffusion
154 coefficient, 4, 17, 26, 39, 43, 72, 77
chain rule kinematics of, 16
for Stratonovich calculus, 156 on manifold, 12
Christoffel symbols, 14, 15, 156 process, 12, 13, 17, 18, 25, 27, 29, 31,
clutter, 29, 113, 114, 127 33, 43, 68, 69, 85, 152
target returns in, 114 quantum mechanical, 18
coherent, 25, 43, 69, 70, 78, 86, 87, 114, requirements of, 48–51
115, 119 tensor, 11–16, 19, 35, 38, 59–61, 73, 74,
amplitude, 40 76, 169
181
182 INDEX
Dirac covariant, 14, 18, 60, 62

delta function, 5, 9 relation to Kramers–Moyal
distribution expansion, 47, 50
K-, 25, 27–30, 35, 52–54, 60, 62, 64, 66,
67, 69, 70, 86, 88, 91, 97, gamma
124–127, 130–132, 134, 137, distribution, 29, 49, 154
145, 154, 165, 166, 169 function, 49
asymptotic, 25–30, 35, 37–39, 41, 44, variate, 121, 123, 150
47, 49, 54, 59–61, 64, 69, 70, generalized K-scattering, 26, 69–72, 74,
79, 81, 83, 84, 86, 90, 161, 163, 76, 78, 79, 81, 82, 84–86
167 Girsanov theorem, 28, 33, 34, 159, 160
Cauchy, 66, 67, 154 relevance to intensity scattering, 28, 33
gamma, 29, 49, 154 role in mathematical finance, 21, 28,
infinitely divisible, 153 34, 160
lognormal, 21 Green’s function, 42, 45, 53, 63
Rayleigh, 27–30, 91 solution to heat equation, 5, 17
Rice, 70, 72, 77, 79
stable, 44, 66, 153, 154 Haykin
Doppler, 25, 64, 86, 95, 97, 99, 110–115, Simon, 138
118, 144, 152 heat equation, 1, 3–5, 17, 23
effect on volatility analysis, 111 homodyned
drift K-scattering, 26, 69, 70, 73, 77, 78,
backward, 12–14, 18 80–83, 86
forward, 12–14, 17, 18, 36, 39 hypergeometric function, 65, 66, 83, 122
Einstein in-phase component, 40, 143, 152, 169

Albert, 3 increments, 59, 60, 76, 108, 109, 119, 121,
Annus Mirabilis, 5 129, 131, 132, 171
derivation of heat equation, 3, 4, 6, 17 independence of, 5–7, 9, 124
summation convention, 12, 16, 26, 170 infinite divisibility, 153, 154
endogenous model, 67, 112, 115 relation to stability, 153, 154
defined, 56 intensity, 41, 55
ensemble analysis of fluctuations in, 125, 126, 128
average, 3, 20, 27, 33, 34, 36, 125, 134, autocorrelation, 63
143, 147, 159, 160 iterative solution, 99, 103, 108, 170
equation of continuity, 14, 35, 62 for vector processes, 170
expectation of SDEs, 105, 108, 109, 170
conditional, 158, 166 Ito
experimental tests, 124 differential, 9
extinction, 162, 163 formula, 10
isometry, 9
Fano factor, 47, 164 product rule, 10
Faraday’s law, 152 stochastic integral, 7
Feng Ito calculus, 7
Tao (Stephen), 133
Field K-distribution, 25, 27–30, 35, 52–54, 60,
equations, 54 62, 64, 66, 67, 69, 70, 86, 88,
filtration, 44, 56, 91, 158, 165 91, 97, 124–127, 130–132, 134,
Fisher information 137, 145, 154, 165, 166
discrimination, 132 infinite divisibility of, 154
geodesic distance, 132 K-scattering, 25, 26, 30, 31, 37–40, 42,
Fokker–Planck 52, 59, 60, 68–70, 76, 79,
equation (FPE), 12–14, 18, 30–32, 35, 85–88, 91, 93, 97, 99, 107–111,
36, 41, 42, 47, 49, 60, 61, 63, 145, 154, 165, 167, 169
66, 72, 86, 99, 106, 107, 111, compound model, 66, 67, 109
119, 148 iterative schemes for, 99
INDEX 183
Kolmogoroff partition function, 51, 161

forward drift, 13, 15, 156 Pawula theorem, 47, 51, 172
Kramers–Moyal phase
coefficients, 47, 50 decoherence, 89, 151, 152
expansion, 46, 48, 50 dynamics, 43, 93
wrapping, 44, 45, 94
Lagrangian, 18 phasor, 43–45, 89, 94
Langevin random, 10, 39, 45, 89, 90, 94, 143
equation, 99, 108, 142 Poisson
Laplace distribution, 49
transform variable, 29 equation, 23
Laplace’s equation, 23 variate, 161
stochastic solution to, 23 Poisson process, 49
Levi-Civita population
connection, 12–14, 19 birth-death-immigration (BDI) process,
connection, 16 47, 161
Lorentzian spectrum, 66, 67, 154 dynamics, 46, 48, 88, 91, 92
power, 118
manifold spectral density, viii, 52, 53, 64–67
diffusion on, 12 for non-stationary process, 65
Mann–Whitney power (scattering), 20, 44, 80, 82, 84, 99,
rank-sum statistic, 138, 140, 147, 148 112, 140, 146, 166
Markov, 42, 52, 59, 121, 158 propagator, viii, 45, 63–65, 106, 107, 119,
non-, 125, 137 121, 122, 148
martingale, 34, 156, 159
property of Ito stochastic integrals, 156 quadratic variation, 4, 9, 10
master equation, 31, 46, 49, 50, 54, 162 quadrature components, 40, 132
master equation (transitions) quadrature-phase component, 40, 143,
higher order, 50 152, 169
mathematical finance, 100, 117, 160
Maxwell’s equations, 43, 144, 145, 165 radar
for time dependent magnetic field, 152 cross-section (RCS), 66, 138, 139,
metric 141–148
Euclidean, 13, 14, 16, 22 parameters, 144
role in construction of diffusion, 12, 14 sea clutter, 126, 127
molecules Radon–Nikodym derivative, 34, 159
random collisions of, 3 random phase screen, 125
random walk
Nelson discrete, 6
Edward, 19 extended model, 88, 89, 139
non-linearity, 138–140, 147, 149 model of scattering, 26, 35, 37, 38, 40,
of sea clutter, 138, 147, 148 42–44, 52, 53, 67–70, 86,
shape parameter and, 138, 139, 143, 88–91, 94, 129–131, 139, 142,
145–149 143, 165, 166
numerical solution, 108, 157 step number fluctuations, 40, 43, 52,
53, 68, 88, 166
optical scattering, 125 Rice
volatility analysis, 125 distribution, 79
Ornstein–Uhlenbeck process, 23, 40–42, scattering, 71, 79, 80, 83, 86
44, 45, 53, 60, 82, 89, 109–111, Runge–Kutta
130, 165, 166 algorithm, 108, 109, 119
osmotic current / velocity, 5, 17
osmotic equation, 17, 18 scattering
K-, 52
partial differential equations, 23 optical wavelength, 125
solution via diffusion process, 23 radar wavelength, 126
184 INDEX
scattering (Cont.) 147–151, 156, 157, 159,

Rayleigh, 40 165–168
weak, 69 stochastic differential equation SDE, 139
sea clutter, 127 stochastic equilibrium, 18, 27
z-parameter and, 140 Stratonovich integral, 8, 155
sea state, 138–140, 145–148 chain rule for, 156
z-parameter and, 140, 147, 149 superposition, 68, 70, 90, 144, 153
coherent, 25
second order algorithms, 97, 99, 115
principle of, 145
shape parameter, 29, 80, 84
weak, 91
simulation, 99, 150
Swerling
spectral density, viii, 52, 53, 64–67 target in Rayleigh clutter, 71
stability, 22, 44, 66, 153
relation to infinite divisibility, 153, 154 target returns, 97, 114, 115
state-space time reversal symmetry, 18, 19
formalism, 138 Tough
model, 95 Robert, 99
stationarity, 3, 31
non-, 36 vector scattering, 12, 27, 30, 36–38, 72,
strict sense, 45 73, 76, 86
wide sense, 5, 45 correlation in, 35
of SDE, 64 volatility, 7, 8
step number, 40, 165 instantaneous observability of, 27, 28,
fluctuations in, 40, 43, 52, 53, 68, 71, 33, 34, 124, 134, 159
88, 166 tensor transformation of, 15
stochastic differential equation (SDE), 1,
5, 13, 20–23, 25–27, 31–35, Weibull distribution, 29, 67
40–44, 49, 52–54, 56–59, 62–64, Wiener process, 7, 9, 157, 159
67, 69, 71, 72, 74, 86, 91–93, wireless propagation, 88, 95, 141
97, 99, 100, 105–109, 111, 113, SDE models for, 95, 141
117–119, 121, 123, 124, 126,
129–132, 137–139, 141–145, z-parameter, 138, 140, 141, 147–149

Electromagnetic Scattering From Random Media

Uploaded by

Copyright:

Available Formats

Electromagnetic Scattering From Random Media

Uploaded by

Document Information

Original Description:

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

Electromagnetic Scattering From Random Media

Uploaded by

Copyright:

Available Formats

INTERNATIONAL SERIES OF MONOGRAPHS

J. BIRMAN CITY UNIVERSITY OF NEW YORK

The topic covered by the monograph is the scattering of electromagnetic waves

by making, to some extent, the various chapters self-contained so that readers

II DYNAMICS OF THE SCATTERING PROCESS

III SIMULATION AND EXPERIMENT

13.3.1 Superposition 144

1.1 Einstein’s 1905 derivation

at location x at time t + τ is logically equivalent to the event that it is location

1.1.1 Green’s function solution

which has the variance property

Vart [x] = 2Dt. (1.9)

This corresponds to an individual particle motion following a stochastic process

in which Wt is the familiar Wiener process.

1.2 Brownian motion

E [(Wv − Wu )(Wt − Ws )] = 0, (1.11)

2.1 Ito stochastic integral

E [(Wv − Wu )(Wt − Ws )] = 0, (2.1)

Wt − Ws ∼ N (0, |t − s|); (2.2)

third, the process is deﬁned subject to the initialization W0 = 0.

Now, in light of the above expository remarks, a stochastic integral is deﬁned

σs dWs = lim σa+jδs (Wa+(j+1)δs − Wa+jδs ), (2.5)

In comparison with the above remarks, an important related property of such

2.1.1 Ito isometry

Thus, in terms of distribution valued functions the expected product stochas-

which does not vanish.

2.2 Ito diﬀerential

dXti = bi dt + dξti (2.9)

The chapter further develops the theory of (non-degenerate) diﬀusion processes,

3.1 Diﬀusions on manifolds

a diﬀusion point of view, in that no extra metrical ingredient is required in the

with the Christoﬀel symbols deﬁned in the standard way as

and satisﬁes the covariant FPEs

This relationship can be re-expressed in√terms of the Christoﬀel symbol Γ as

Γi := σ jk Γijk = −∂j σ ij − σ ij Γkkj (3.13)

and thus (3.11) can be written as

covariant derivative ∇b ξ a of a vector ξ a in the tangent space to T M can be

This enables us to introduce the abstract Ito diﬀerential according to

and hence (3.1) becomes the abstract geometrical equation6

3.2 Kinematics of diﬀusion

We further deﬁne the ‘current’ and ‘osmotic’ velocities v i , ui respectively accord-

To calculate these kinematical quantities for speciﬁc dynamical processes we

which is notably independent of the diﬀusion coeﬃcient D. Observe, from the

concentration at the origin) induces an essential spatial asymmetry with respect

The reversal of v i under T is in accordance with intuition since this velocity is

T [D] = −D̃. (3.31)

We return to examine such kinematical properties, in the context of detailed

4.1 Geometric Brownian motion

4.2 Bessel process

In Rn the Bessel process is deﬁned as the Euclideanmetric  distance of a n-

stability of the Gaussian distribution, as discussed in Appendix A.

4.3 Solution to Laplace’s equation

on some domain D ⊂ Rn , subject to the boundary condition φ|∂D = f for some

φ(x) := Ex [f (Xτ )] . (4.11)

This is evidently a function of x and, since the generator of Brownian motion

In this part, we develop the detailed stochastic dynamics of the electromagnetic

resultant amplitude with respect to which the scattered component is weak.

Diﬀusion models have proved highly successful in their application to a large

Bessel process given in Section 4.2.

In Rn the Bessel process is deﬁned as the Euclideanmetric distance of a n-

ztα11 ztα22 . . . ztαnn (5.26)

or in terms of a re-scaled time parameter t = t/N̄