Chew, 2009

MC1483_Chew_Lecture.
qxd 8/18/08 12:29 PM Page i
Integral Equation Methods for

Electromagnetic and Elastic
Waves
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page ii
Copyright © 2009 by Morgan & Claypool
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means – electronic, mechanical, photocopy, recording, or any other except for brief quotations
in printed reviews, without the prior permission of the publisher.
Integral Equation Methods for Electromagnetic and Elastic Waves

Weng Cho Chew, Mei Song Tong and Bin Hu
www.morganclaypool.com
ISBN-10: 1598291483 paperback

ISBN-13: 9781598291483 paperback
ISBN-10: 1598291491 ebook

ISBN-13: 9781598291490 ebook
DOI10.2200/S00102ED1V01Y200807CEM012
A Publication in the Morgan & Claypool Publishers series

LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #12
Lecture #12
Series Editor: Constantine A. Balanis, Arizona State University
Series ISSN: 1932-1252 print

Series ISSN: 1932-1716 electronic
First Edition
10 9 8 7 6 5 4 3 2 1
Printed in the United States of America

MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page iii
Integral Equation Methods for

Electromagnetic and Elastic
Waves
Weng Cho Chew

University of Hong Kong and
University of Illinois at Urbana-Champaign
Mei Song Tong

University of Illinois at Urbana-Champaign
Bin Hu
Intel Research
SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #12
M
&C Mor gan &Cl aypool Publishers
iv
ABSTRACT
Integral Equation Methods for Electromagnetic and Elastic Waves is an outgrowth of several years
of work. There have been no recent books on integral equation methods. There are books written
on integral equations, but either they have been around for a while, or they were written by math-
ematicians. Much of the knowledge in integral equation methods still resides in journal papers.
With this book, important relevant knowledge for integral equations are consolidated in one place
and researchers need only read the pertinent chapters in this book to gain important knowledge
needed for integral equation research. Also, learning the fundamentals of linear elastic wave theo-
ry does not require a quantum leap for electromagnetic practitioners.
Integral equation methods have been around for several decades, and their introduction to elec-
tromagnetics has been due to the seminal works of Richmond and Harrington in the 1960s. There
was a surge in the interest in this topic in the 1980s (notably the work of Wilton and his cowork-
ers) due to increased computing power. The interest in this area was on the wane when it was
demonstrated that differential equation methods, with their sparse matrices, can solve many prob-
lems more efficiently than integral equation methods. Recently, due to the advent of fast algo-
rithms, there has been a revival in integral equation methods in electromagnetics. Much of our
work in recent years has been in fast algorithms for integral equations, which prompted our inter-
est in integral equation methods. While previously, only tens of thousands of unknowns could be
solved by integral equation methods, now, tens of millions of unknowns can be solved with fast
algorithms. This has prompted new enthusiasm in integral equation methods.
KEYWORDS
Integral Equations, Computational Electromagnetics, Electromagnetic Waves, Linear Vector
Spaces, Energy Conservation Theorem, Low-Frequency Problems, Dyadic Green's Function,
Fast Inhomogeneous Plane Wave Algorithm, Elastic Waves
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page v
Dedication
We dedicate this book

to our predecessors
in this field.
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page vi
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page vii
vii
Contents
Preface v
Acknowledgements vii
1 Introduction to Computational Electromagnetics 1

1.1 Mathematical Modeling—A Historical Perspective . . . . . . . . . . . . . . . 1
1.2 Some Things Do Not Happen in CEM Frequently—Nonlinearity . . . . . . . 3
1.3 The Morphing of Electromagnetic Physics . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Comparison of Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Low-Frequency Regime . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 Mid Frequency Regime . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.4 High-Frequency Regime . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.5 Quantum Regime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Matched Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Why CEM? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Time Domain versus Frequency Domain . . . . . . . . . . . . . . . . . . . . . 8
1.7 Differential Equation versus Integral Equation . . . . . . . . . . . . . . . . . . 9
1.8 Nondissipative Nature of Electromagnetic Field . . . . . . . . . . . . . . . . . 11
1.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Linear Vector Space, Reciprocity, and Energy Conservation 21

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Linear Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Inner Products for Electromagnetics . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 Comparison with Mathematics . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Comparison with Quantum Mechanics . . . . . . . . . . . . . . . . . . 25
2.4 Transpose and Adjoint of an Operator . . . . . . . . . . . . . . . . . . . . . . 26
2.5 Matrix Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.6 Compact versus Noncompact Operators . . . . . . . . . . . . . . . . . . . . . 28
2.7 Extension of Bra and Ket Notations . . . . . . . . . . . . . . . . . . . . . . . 29
2.8 Orthogonal Basis versus Nonorthogonal Basis . . . . . . . . . . . . . . . . . . 30
2.9 Integration by Parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.10 Reciprocity Theorem—A New Look . . . . . . . . . . . . . . . . . . . . . . . 33
i
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page viii
viii INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES
2.10.1 Lorentz Reciprocity Theorem . . . . . . . . . . . . . . . . . . . . . . . 35

2.11 Energy Conservation Theorem—A New Look . . . . . . . . . . . . . . . . . . 36
2.12 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 Introduction to Integral Equations 43

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 The Dyadic Green’s Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Equivalence Principle and Extinction Theorem . . . . . . . . . . . . . . . . . 44
3.4 Electric Field Integral Equation—A Simple Physical Description . . . . . . . 48
3.4.1 EFIE—A Formal Derivation . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5 Understanding the Method of Moments—A Simple Example . . . . . . . . . . 51
3.6 Choice of Expansion Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.7 Closed Surface versus Open Surface . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.1 EFIE for Open Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.2 MFIE and More . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.8 Internal Resonance and Combined Field Integral Equation . . . . . . . . . . . 56
3.9 Other Boundary Conditions—Impedance Boundary Condition, Thin Dielectric
Sheet, and R-Card . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.10 Matrix Solvers–A Pedestrian Introduction . . . . . . . . . . . . . . . . . . . . 59
3.10.1 Iterative Solvers and Krylov Subspace Methods . . . . . . . . . . . . . 59
3.10.2 Effect of the Right-Hand Side—A Heuristic Understanding . . . . . . 62
3.11 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4 Integral Equations for Penetrable Objects 71

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Scattering by a Penetrable Object Using SIE . . . . . . . . . . . . . . . . . . 72
4.3 Gedanken Experiments for Internal Resonance Problems . . . . . . . . . . . . 75
4.3.1 Impenetrable Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.2 Penetrable Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3.3 A Remedy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.4 Connection to Cavity Resonance . . . . . . . . . . . . . . . . . . . . . 82
4.3.5 Other remedies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Volume Integral Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4.1 Alternative Forms of VIE . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4.2 Matrix Representation of VIE . . . . . . . . . . . . . . . . . . . . . . . 87
4.5 Curl Conforming versus Divergence Conforming Expansion Functions . . . . 89
4.6 Thin Dielectric Sheet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.6.1 A New TDS Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.7 Impedance Boundary Condition . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.7.1 Generalized Impedance Boundary Condition . . . . . . . . . . . . . . 95
4.7.2 Approximate Impedance Boundary Condition . . . . . . . . . . . . . . 96
4.7.3 Reflection by a Flat Lossy Ground Plane . . . . . . . . . . . . . . . . 96
4.7.4 Reflection by a Lossy Cylinder . . . . . . . . . . . . . . . . . . . . . . 98
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page ix
CONTENTS ix
5 Low-Frequency Problems in Integral Equations 107

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 Low-Frequency Breakdown of Electric Field Integral Equation . . . . . . . . . 107
5.3 Remedy—Loop-Tree Decomposition and Frequency Normalization . . . . . . 109
5.3.1 Loop-Tree Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3.2 The Electrostatic Problem . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.3.3 Basis Rearrangement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
−1
5.3.4 Computation of K · Q in O(N ) Operations . . . . . . . . . . . . . . 118
5.3.5 Motivation for Inverting K Matrix in O(N ) Operations . . . . . . . . 119
5.3.6 Reason for Ill-Convergence Without Basis Rearrangement . . . . . . . 122
5.4 Testing of the Incident Field with the Loop Function . . . . . . . . . . . . . . 124
5.5 The Multi-Dielectric-Region Problem . . . . . . . . . . . . . . . . . . . . . . . 124
5.5.1 Numerical Error with Basis Rearrangement . . . . . . . . . . . . . . . 127
5.6 Multiscale Problems in Electromagnetics . . . . . . . . . . . . . . . . . . . . . 128
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6 Dyadic Green’s Function for Layered Media and Integral Equations 135
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.2 Dyadic Green’s Function for Layered Media . . . . . . . . . . . . . . . . . . . 136
6.3 Matrix Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.4 The ∇ × Ge Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.5 The L and K Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.6 The Ez -Hz Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.6.1 Example 1: Half Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.6.2 Example 2: Three-Layer Medium . . . . . . . . . . . . . . . . . . . . . 150
6.7 Validation and Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7 Fast Inhomogeneous Plane Wave Algorithm for Layered Media 159

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.2 Integral Equations for Layered Medium . . . . . . . . . . . . . . . . . . . . . 160
7.3 FIPWA for Free Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.4 FIPWA for Layered Medium . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.4.1 Groups not aligned in ẑ axis . . . . . . . . . . . . . . . . . . . . . . . 166
7.4.2 Groups aligned in ẑ axis . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.4.3 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8 Electromagnetic Wave versus Elastic Wave 193

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2 Derivation of the Elastic Wave Equation . . . . . . . . . . . . . . . . . . . . . 193
8.3 Solution of the Elastic Wave Equation—A Succinct Derivation . . . . . . . . 197
8.3.1 Time-Domain Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.4 Alternative Solution of the Elastic Wave Equation via Fourier-Laplace Transform200
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page x
x INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES
8.5 Boundary Conditions for Elastic Wave Equation . . . . . . . . . . . . . . . . 202

8.6 Decomposition of Elastic Wave into SH, SV and P Waves for Layered Media 203
8.7 Elastic Wave Equation for Planar Layered Media . . . . . . . . . . . . . . . . 205
8.7.1 Three-Layer Medium Case . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.8 Finite Difference Scheme for the Elastic Wave Equation . . . . . . . . . . . . 210
8.9 Integral Equation for Elastic Wave Scattering . . . . . . . . . . . . . . . . . 211
8.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Glossary of Acronyms 223
About the Authors 227

MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xi
xi
Preface
This monograph is an outgrowth of several years of work. It began with a suggestion by Joel
Claypool from Morgan & Claypool Publishers and Constantine Balanis of Arizona State
University on writing a monograph on our recent contributions to computational electromagnet-
ics (CEM). Unfortunately, we had just written a book on CEM a few years ago, and it was hard
to write a new book without repeating many of the materials written in Fast and Efficient
Algorithms in Computational Electromagnetics (FEACEM).
In the midst of contemplating on a topic, it dawned upon me that there was no recent books
on integral equation methods. There had been books written on integral equations, but either they
have been around for a while, such as books edited by Mittra, and by Miller, Medgyesi-Mitschang
and Newman, or they were written by mathematicians, such as the book by Colton and Kress.
Much of the knowledge in integral equation methods still resides in journal papers.
Whenever I have to bring researchers to speed in integral equation methods, I will refer them to
study some scientific papers but not a book. So my thinking was that if we could consolidate all
important knowledge in this field in one book, then researchers just need to read the pertinent
chapters in this book to gain important knowledge needed for integral equation research.
Integral equation methods have been around for several decades, and their introduction to
electromagnetics has been due to the seminal works of Richmond and Harrington. After the ini-
tial works of Richmond and Harrington in the 1960s, there was a surge in the interest in this topic
in the 1980s (notably the work of Wilton and his coworkers) due to the increased power of com-
puters. The interest in this area was on the wane when it was demonstrated that differential equa-
tion methods, with its sparse matrices, can solve many problems more efficiently than integral
equations. However, in recent decades (1990s), due to the advent of fast algorithms, there was a
revival in integral equation methods in electromagnetics.
Much of our work in recent years has been in fast algorithms for integral equations, which
prompted our interest in integral equation methods. While previously, only tens of thousands of
unknowns can be solved by integral equation methods, now, tens of millions of unknowns can be
solved with fast algorithms. This has prompted new enthusiasm in integral equation methods. This
enthusiasm had also prompted our writing the book FEACEM, which assumed that readers
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xii
xii INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES
would have the basic knowledge of integral equation methods. This book would help to fill in that
knowledge.
With this book, important relevant knowledge for integral equations are consolidated in one
place. Much of the knowledge in this book exists in the literature. By rewriting this knowledge, we
hope to reincarnate it. As is in most human knowledge, when it was first discovered, only few could
understand it. But as the community of scholars came together to digest this new knowledge,
regurgitate it, it could often be articulated in a more lucid and succinct form. We hope to have
achieved a certain level of this in this book.
Only one chapter on this book is on fast algorithm, namely, Chapter 7, because the topic of
this chapter has not been reported in FEACEM. This chapter is also the outgrowth of BH’s PhD
dissertation on fast inhomogeneous plane wave algorithm (FIPWA) for free space and layered
media.
Since many researchers in electromagnetics can easily learn the knowledge related to elastic
waves, as many of our researchers do, we decide to include Chapter 8 in this book. It is written pri-
marily for practitioners of electromagnetics—Learning linear elastic wave theory does not require
a quantum leap for electromagnetic practitioners. Because we have recently applied some of our
expertise in solving the elastic wave problems, including the application of fast algorithm, we
decide to include this chapter.
In this book, Chapters 1 to 6 were primarily written by WCC, while Chapter 7 was written
by BH, and MST wrote Chapter 8, the last Chapter. We mutually helped with the proofreading
of the chapters, and in particular MST helped extensively with the proofreading, editing, as well
as the indexing of the book.
As for the rest of the book, Chapter 1 introduces some concepts in CEM, and tries to con-
trast them with concepts in computational fluid dynamics (CFD), which itself is a vibrant field that
has been studied for decades.
Chapter 2 is on the fundamentals of electromagnetics. We relate some fundamental concepts
to concepts in linear vector space theory. It is de rigueur that students of CEM be aware of these
relationships. After all, a linear integral equation eventually becomes a matrix equation by project-
ing it into a subspace in the linear vector space.
Chapter 3 gives a preliminary introduction to integral equation methods without being
bogged down by details. It emphasizes mainly on a heuristic and physical understanding of inte-
gral equation methods. For many practitioners, such a level of understanding is sufficient.
Chapter 4 is not for the faint hearted as it delves into more details than Chapter 3. It dis-
cusses integral equations for impenetrable bodies as well as penetrable bodies. It deals with a very
important issue of uniqueness of many of these integral equation methods. It introduces Gedanken
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xiii
PREFACE xiii
experiments to prove the nonuniqueness of these solution methods, and discusses remedies for
them.
Chapter 5 is on low-frequency methods in integral equations. Integral equation methods, as
well as differential equation methods, are plagued with instabilities when the frequency is extreme-
ly low, or the wavelength is very long compared to the structure being probed. This is because we
are switching from the regime of wave physics to the regime of circuit physics (discussed in
Chapter 1). This transition presage a change in solution method when one goes into the low-fre-
quency regime, which I have nicknamed “twilight zone!”
Chapter 6 reports on how one formulates the integral equations in the layered media. The
derivation of the dyadic Green’s function for layered media is often a complex subject. It involves
complicated Sommerfeld integrals, which often are poorly convergent. We propose a new formu-
lation, which harks back to the technique discussed in Waves and Fields in Inhomogeneous Media, as
oppose to the Michalski-Zheng formulation, which is quite popular in the literature.
Integral equation methods are in general quite complex compared to differential equation
methods. To convert integral equations to matrix equations, one has to project the Green’s opera-
tor onto a subspace by numerical quadrature. This numerical quadrature involves singular integrals,
and can be quite challenging to perform accurately. We regret not being able to discuss these
numerical quadrature techniques in this book. These often belong to the realm of applied mathe-
matics on numerical quadrature of singular integrals. Much has been written about it, and the
monograph does not allow enough space to discuss this topic.
However, we hope that we have discussed enough topics in this short monograph that can
pique your interest in this very interesting and challenging field of integral equation methods.
Hopefully, many of you will pursue further research in this area, and help contribute to new knowl-
edge to this field!
Weng Cho CHEW

June 2008, Hong Kong
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xiv
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xv
xv
Acknowledgements
Many in our research group gave feedback and helped proofread the manuscript at various
stages. Thanks are due Mao-Kun LI, Yuan LIU, Andrew Hesford, Lin SUN, Zhiguo QIAN, Jie
XIONG, Clayton Davis, William Tucker, and Bo HE. During my stay as a visiting professor
at Nanyang Technological University, I received useful feedback from Eng Leong TAN on
some of the chapters. Peter Monk gave useful feedback on part of Chapter 2.
We are grateful to our research sponsors for their supports in the last decade, notably,
AFOSR, SAIC-DEMACO, Army-CERL, Raytheon-UTD, GM, SRC, Intel, IBM, Mentor-
Graphics, and Schlumberger. WCC also thanks supports from the Founder Professorship and
Y.T. Lo Endowed Chair Professorship, as well as supports from the IBM Visiting Professorship
at Brown University, the Distinguished Visitor Program at the Center for Computational
Sciences, AFRL, Dayton, the Cheng Tsang Man Visiting Professorship at NTU, Singapore,
and the MIT Visiting Scientist Program. Supports from Arje Nachman and Mike White of the
government agencies are gratefully acknowledged. We appreciate interactions with partners
and colleagues in the industry, notably, Steve Cotten, Ben Dolgin, Val Badrov, HP Hsu,
Jae Song, Hennning Braunisch, Kemal Aygun, Alaeddin Aydiner, Kaladhar Radhakrishnan,
Alina Deutsch, Barry Rubin, Jason Morsey, Li-Jun Jiang, Roberto Suaya, Tarek Habashy,
and John Bruning. Also, discussions with university partners on various projects such as Ben
Steinberg, Steve Dvorak, Eric Michielsen, and John Volakis have been useful. Last but not
least, constant synergism with members of the EM Laboratory at Illinois, Shun-Lien Chuang,
Andreas Cangellaris, Jianming Jin, Jose Schutt-Aine, and Jennifer Bernhard are enlightening.
This book is also dedicated to the memory of Y.T. Lo, Don Dudley, and Jin Au Kong, who
have been WCC’s mentors and friends.
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page xvi
MC1483_Chew_Lecture.qxd 8/18/08 12:29 PM Page 1
CHAPTER 1
Introduction to Computational
Electromagnetics
1.1 Mathematical Modeling—A Historical Perspective
In the beginning, mathematics played an important role in human science and technology
advancements by providing arithmetic, geometry, and algebra to enable scientists and tech-
nologists to solve complex problems. For example, precalculus was used by Kepler to arrive
at laws for predicting celestial events in the 1600s. In the late 1600s and early 1700s, Leibniz
and Newton developed early calculus for a better description of physical laws of mechanics [1].
Later, the development of differential equations enabled the better description of many physi-
cal phenomena via the use of mathematics. The most notable of these were the Navier-Stokes
equations for understanding the physics of fluids in the 1800s [2–4].
Computational fluid dynamics (CFD) [5] has a longer history than computational electro-
magnetics (CEM) (Richardson 1910 [6], Courant, Friedrichs, and Lewy 1928 [7], Lax 1954 [8],
Lax and Wendroff 1960 [9], MacCormack 1969 [10]). The early works are mainly those of
mathematicians looking for numerical means to solve partial differential equation (PDE), with
the knowledge that many such equations have no closed form solution. Hence, the reason for
the early development of CFD is simple: many fluids problems are analytically intractable.
This stems from the nonlinearity inherent in the Navier-Stokes equations. Early fluids cal-
culations were done numerically, albeit laboriously, with the help of human “computers”. In
fact, the design of the first electronic computers was driven by the need to perform laborious
fluids calculations (Eniac 1945, Illiac I 1952 [11]).
In contrast, most electromagnetic media are linear. Hence, the governing equations for
electromagnetics are, for the most part, linear. The linearity of electromagnetic equations
allows for the application of many analytic techniques for solving these equations.
1
2 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES
The electromagnetic equations are:

∂B
∇×E=− (1.1)
∂t
∂D
∇×H= +J (1.2)
∂t
∇·D=ρ (1.3)
∇·B=0 (1.4)
These equations are attributable to the works of Faraday (Eq. (1.1), Faraday’s law 1843 [12]),
Ampere (Eq. (1.2), Ampere’s law 1823 [13]), Coulomb (Eq. (1.3), Coulomb’s law 1785 [14]; it
was enunciated in a divergence form via Gauss’ law), and Gauss (Eq. (1.4), Gauss’ law 1841
[15]). In 1864, Maxwell [16] completed the electromagnetic theory by adding the concept of
displacement current (∂D/∂t) in Eq. (1.2). Hence, this equation is also known as generalized
Ampere’s law. The set of equations above is also known as Maxwell’s equations. The original
Maxwell’s equations were expressed in some 20 equations. It was Heaviside [17] who distilled
Maxwell’s equations into the above four equations. Hence, they are sometimes known as the
Maxwell-Heaviside equations. These equations have been proven to be valid from subatomic
length scale to intergalactic length scale.
Because of their linearity, one useful technique for solving electromagnetic equations is the
principle of linear superposition. The principle of linear superposition further allows for the
use of the Fourier transform technique. Hence, one can Fourier transform the above equations
in the time variable and arrive at
∇ × E = iωB (1.5)
∇ × H = −iωD + J (1.6)
∇·D=ρ (1.7)
∇·B=0 (1.8)
The preceding equations are the frequency-domain version of the electromagnetic equations.
When the region is source free, that is, ρ = 0 and J = 0, as in vacuum, B = µ0 H and
D = ²0 E, the simplest homogeneous solution to the above equations is a plane wave of the
√
form a exp(ik · r) where a is orthogonal to k and |k| = ω µ0 ²0 . Hence, Maxwell’s work
bridged the gap between optics and electromagnetics because scientists once thought that the
electromagnetic phenomenon was different from the optical phenomenon.
Also, for electrodynamic problems, the last two of the above equations can be derived
from the first two with appropriate initial conditions. Hence, the last two equations can be
ignored for dynamic problems [18].
The use of Fourier transform technique brings about the concept of convolution. The
concept of convolution leads to the development of the Green’s function technique [19]. From
the concept of Green’s functions, integral equation can be derived and solved [18].
As a result, electromagneticists have a whole gamut of analytic tools to play with. Many
engineering problems and theoretical modeling can be solved by the application of these tools,
offering physical insight and shedding light on many design problems. The use of intensive
computations to arrive at critical data can be postponed until recently. Hence, CEM concepts
INTRODUCTION TO COMPUTATIONAL ELECTROMAGNETICS 3
did not develop until 1960s [20, 21]. Although laborious calculations were done in the early
days, they were not key to the development of many technologies.
1.2 Some Things Do Not Happen in CEM Frequently—

Nonlinearity
Nonlinearity is not completely absent in electromagnetics. Though most nonlinear phenomena
are weak, they are usually amplified by short wavelengths, or high frequencies. When the
wavelength is short, any anomalies will be amplified and manifested as a cumulative phase
phenomenon. For instance, the Kerr effect, which is a 10−4 effect, is responsible for the
formation of solitons at optical frequencies [22].
Nonlinear effects do not occur frequently in electromagnetics. A nonlinear phenomenon
has the ability to generate high frequency fields from low frequency ones (Just imagine f (x) =
x2 . If x(t) = sin ωt, f [x(t)] = sin2 ωt has a frequency component twice as high as that of
x(t)). Nonlinearity can give rise to shock front and chaos. These phenomena do not occur in
linear electromagnetics.
Nonlinearity in electromagnetics can be a blessing as well as a curse. The successful
propagation of a soliton in an optical fiber was heralded as a breakthrough in optics for
optical communications whereby a pulse can propagate distortion-free in an optical fiber
for miles needing no rejuvenation. However, the nonlinear phenomenon in an optical fiber
is besieged with other woes that make the propagation of a pure soliton difficult. After
two decades of research, no commercially viable communication system using solitons has
been developed. On the contrary, optical fibers are engineered to work in the linear regime
where dispersion management is used to control pulse shapes for long distance propagation.
Dispersion management uses alternating sections of optical fibers with positive and negative
dispersion where pulse expansion is followed by pulse compression to maintain the integrity
of the signal [23].
1.3 The Morphing of Electromagnetic Physics

Electromagnetic field is like a chameleon—its physics changes as the frequency changes from
static to optical frequencies [24]. Electromagnetics is described by fields—a field is a function
of 3D space and time. These fields satisfy their respective PDEs that govern their physics.
Hence, electromagnetic physics changes depending on the regime. The morphing of the
physics in electromagnetics can be easily understood from a simple concept: the comparison
of derivatives. PDEs that govern electromagnetics have many derivative operators in them—
depending on the regime of the physics, different parts of the partial derivative operators
become important.
Another important concept is that of the boundary condition. The boundary condition
dictates the manner in which a field behaves around an object. In other words, a field has to
contort around an object to satisfy the boundary condition. Hence, the variation of a field
around an object is primarily determined by the geometrical shape of the object.
1.3.1 Comparison of Derivatives

Much physics of electromagnetics can be understood by the boundary condition and the
comparison of derivatives.1 Because the boundary condition shapes the field, the derivative
of the field is inversely proportional to the size of the object. If the typical size of the object
is of length L, then
∂ 1 ∂2 1
∼ , ∼ 2 (1.9)
∂x L ∂x2 L
From the above, we can see that around small structures, the space derivative of the field is
large.
Because the Maxwell-Heaviside equations contain primarily first derivatives in space and
time, it is the comparison of the importance of these derivatives that delineates the different
regimes in electromagnetics. For a time-harmonic field, with angular frequency ω, we can
define an associated plane wave with wavelength λ such that c/λ = ω/(2π).
∂ c ∂ 1
∼ω∼ ↔ ∼ (1.10)
∂t λ ∂x L
The comparison of time derivative (which is proportional to λ−1 ) to the space derivative
(which is proportional to L−1 ) delineates different regimes in electromagnetics.
Consequently, electromagnetic physics can be divided into the following regimes:
• Low frequency, L ¿ λ.
• Mid frequency, L ≈ λ.
• High frequency, L À λ.
When the object size is much less than the wavelength, the physics of the field around
it is dominated by the fact that the time derivative is unimportant compared to the space
derivative. The field physics is close to that of static field.
In the mid-frequency regime, the space derivative is equally important compared to the
time derivative, but the size of the object, and hence the boundary condition, still controls
the nature of the solution.
In the high-frequency regime, the boundary condition plays a lesser role compared to
the equations in shaping the solution. Hence, the solution resembles the homogeneous so-
lutions of electromagnetic equations, which are plane waves. Plane waves are mathematical
representation of rays, and hence, ray physics becomes more important in this regime.
In the above category, we are assuming that the point of observation is of the same order
as the size of the object L. If one were to observe the field at different distances d from the
object, the physics of the field will change according to the above argument. Hence, we have:
• Near field, d ¿ λ.
• Mid field, d ≈ λ.
• Far field, d À λ.
1 This technique is often used in fluids to arrive at numbers such as the Reynolds number [2].
In the near-field regime, electrostatic physics is important. In the mid-field regime, wave
physics is important, whereas in the far-field regime, ray physics is important.2 We will
elaborate on the above regimes in more detail in the following discussions.
1.3.2 Low-Frequency Regime

When the frequency is low or the wavelength is long, the governing equation for the elec-
tromagnetic field is the static equation, which is Laplace’s equation or Poisson’s equation.
Laplace’s equation is an elliptic equation where the solution is smooth without the existence
of singularities (or shocks) [25]. Furthermore, solution of Laplace’s equation is nonoscillatory,
implying that it cannot be used to encode spatial information needed for long distance com-
munication. Because the solution is smooth, multigrid [26–28] and wavelets [29, 30] methods
can be used to accelerate the solution in this regime with great success.
Moreover, at static, the magnetic field and the electric field are decoupled from each other.
Hence in this regime, electromagnetic equations reduce to the equation of electrostatics and
equation of magnetostatics that are completely decoupled from each other [31–33]. The equa-
tion of electrostatics also governs the working principle of capacitors, whereas the equation of
magnetostatics governs the working principle of inductors. Furthermore, when the frequency
is very low, the current decomposes itself into a divergence-free component and a curl-free
component. This is known as the Helmholtz decomposition, and it can be gleaned from the
following equations, which are reduced from electromagnetic equations when the frequency is
vanishingly small or ω → 0.
¾
∇×H=J
magnetostatic (1.11)
∇·B=0
¾
∇×E=0
electrostatic (1.12)
∇ · D = ρ = lim ∇·J
iω
The top part of Eq. (1.11) is Ampere’s law. It says that a current J produces a magnetic
field H. However, since ∇ · (∇ × H) = 0, it follows that this current satisfies
∇·J=0 (1.13)
or is divergence-free.
The last equation in (1.12) follows from the continuity equation
∂ρ
∇·J+ = 0 or ∇ · J = iωρ (1.14)
∂t
It implies that the charge ρ that produces the electric field E in (1.12) comes from the
nondivergence free part of J. Hence, in the limit when ω → 0,
J = Jirr + Jsol (1.15)
namely, the current J decomposes itself into an irrotational component Jirr (which is curl-
free but not divergence-free) and solenoidal component Jsol (which is divergence-free but not
2 It is to be noted that in optics, the ray physics regime can be further refined into the Fresnel zone and
the Fraunhoffer zone [34].

curl-free). The irrotational current models the electroquasistatic fields, whereas the solein-
odal current models the magnetoquasistatic fields. Again, as noted earlier, this is known as
Helmholtz decomposition. In order for the last equation in (1.5) to produce a finite charge ρ,
it is seen that Jirr ∼ O(ω) when ω → 0.
Many CEM methods, which are developed for when ω is nonzero, break down as ω → 0.
This is because those methods, although capable of capturing the physics of the electromag-
netic field when ω is not small, fail to capture the physics correctly when ω → 0.
1.3.3 Mid Frequency Regime

The mid-frequency regime is where the electric field and magnetic field are tightly coupled
together. It is also the regime where electromagnetic fields become electromagnetic waves
because of this tight coupling. The field, instead of being smooth as in solutions to Laplace’s
equation, becomes oscillatory as is typical of a wave phenomenon. The oscillatory nature of
the field can be used to encode information, and this information can travel uncorrupted over
vast distances such as across galaxies. Hence, electromagnetic waves are extremely important
for long-distance communications.
In the mid-frequency regime, electromagnetic fields can be described by hyperbolic-type
PDEs. If a wave contains broadband signals all the way from static to high frequencies, it can
propagate singularities such as a step discontinuity. This is the characteristic of hyperbolic
PDEs.
The oscillatory nature of the field in this regime also precludes methods such as multigrid
and wavelets from working well in this regime because the solution is not infinitely smooth
[24, 30].
1.3.4 High-Frequency Regime

The high-frequency regime is where the wavelength becomes extremely short compared to the
dimension of interest. In this regime, the wave solution becomes plane-wave like. Plane waves
behave like rays. Even though electromagnetic field is still described by the wave equation,
a hyperbolic equation, asymptotic methods using the short wavelength as a small parameter
can be used to extract the salient feature of the solution. Alternatively, we can think of rays
as particle-like. This is what happens to light rays.
Consequently, light rays can be described by the transport equation and the eikonal equa-
tion. These equations can be derived from the wave equation under the high-frequency and
short-wavelength approximation [35].
The high-frequency regime also allows for the beam-forming of electromagnetic waves.
For instance, electromagnetic field can be focused as in optics [36–45]. Hence, antennas with
high directivity such as a reflector antenna can be designed at microwave frequencies. At
lower frequencies, such an antenna will have to be unusually large to increase the directivity
of electromagnetic wave.
1.3.5 Quantum Regime

In addition to the above, one has to be mindful of another regime when the frequency is
high. A quantum of electromagnetic energy is proportional of hf , and if hf is not small
(h = 6.624 × 10−34 Js is Planck’s constant and f is the frequency in Hertz, we then have the
quantum regime. In this regime, electromagnetics has to be described by quantum physics.
The electromagnetic field, which is an observable, has a quantum operator analogue. This is
a property not derivable from the wave equation, but is a result of the quantum nature of
light. Hence, in the optical regime when light wave interacts with matter, it is important to
consider quantum mechanics to fully model its physics [46].
Because of the minute value of Planck’s constant, the quantization of electromagnetic
energy is unimportant unless the frequency is very high, as in optical frequencies. Because
of the quantization of electromagnetic energy, light waves interact more richly with matter,
giving rise to a larger variety of physical phenomena such as nonlinear phenomena.
1.4 Matched Asymptotics

Matched asymptotics was first used in fluids to solve problems perturbatively by dividing
into regions near the boundary and away from the boundary, because the physics of fluids
is different near a boundary surface and away from a boundary surface. One way to apply
matched asymptotics in electromagnetics is to use the near-field/far-field dichotomy. The
yardstick for near and far field is then the wavelength. Another way is to exploit the difference
in the solution when it is close to a singularity and far away from a singularity.
In electromagnetics, matched asymptotics can be used to piece together solutions in dif-
ferent regions in space [47–49]. For instance, the scattering solution of a small sphere is
first obtained by Rayleigh using matched asymptotics: the near field is used to match the
boundary condition on sphere, whereas the far field is obtained by using radiation from a
dipole field. Kirchhoff obtained the approximate solution of the capacitance of two circular
disks [50] by breaking the solution into region close to the edge of the disk (where the field
is singular) and away from the edge where the solution is smooth. The work was followed by
other workers [51–54]. Approximate formulas for the resonant frequencies of microstrip disks
and approximate guidance condition for microstrip lines can be similarly obtained [55, 56].
1.5 Why CEM?

Because of the linearity of Maxwell-Heaviside equations, analytic tools have helped carried
the development of electromagnetic technology a long way. Theoretical modeling has often
played an important role in the development of science and technology. For a long time,
theoretical modeling in electromagnetics can be performed with the abundance of analytic
tools available. The analytic solutions often correspond to realistic problem in the real world.
Hence, early theoretical modeling in electromagnetics is replete with examples of math-
ematical virtuosity of the practitioners. Examples of these are the Rayleigh solution to the
waveguide problem [57], solution of Rayleigh scattering by small particles [34,58], Sommerfeld
solutions to the half-plane problem [59] and the half-space problem [60], Mie series solution
to the scattering by spheres [61,62] etc. Because electromagnetics was predated by the theory
of sound and fluids [63], much of the mathematics developed for sound and fluids can be used
in electromagnetics with some embellishment to account for the vector nature of the field.
Many of the scattering solutions by canonical shapes is documented in Bowman et al. [64].
Later, when science and technology called for solutions to more complex problems, the-
oretical modeling was performed with the help of perturbation and asymptotic methods. In
asymptotic methods, the wavelength is usually used as a small parameter for a perturba-
tion expansion. This is the field of physical optics, geometrical optics, geometrical theory of
diffraction, etc [36–45]. The solution is usually ansatz based. For example, when the fre-
quency is high, or the wavelength short, the wavenumber k = 2π/λ is large, the scattering
solution from a complex object may be assumed to be dominated by a few salient features
as a consequence of reflection and diffraction. Hence, the approximate solution may be a
superposition of solutions of the form
· ¸
eikr 1 1 1
ψ sca (r) = α a0 + a1 + a2 + a 3 + · · · (1.16)
r kr (kr)2 (kr)3
The constants α and ai ’s are obtained by matching solution with canonical solutions, such as
the Sommerfeld half-plane solution, and the Mie series solution after Watson’s transformation
[65].
However, the leaps and bounds progress of computer technology opens up new possibilities
for solving the electromagnetic equations more accurately for theoretical modeling in many
applications. This is especially so when the structure is complex. Instead of being driven
by qualitative engineering where physical intuition is needed, a quantitative number can be
arrived at using CEM that can be used as a design guide.
Moreover, the use of CEM can reduce our dependency on experiments, which are usually
by cut-and-try method. Experiments are usually expensive, and if design decisions can be
made by virtual experiment or virtual prototyping using simulation and modeling techniques
based on CEM, design cost can be substantially reduced. CEM preserves a large part of the
physics of the electromagnetic equations when they are solved, and hence, it can represent a
high-fidelity virtual experiment quite close to the real world.
Asymptotic and approximate methods serve the design needs at high frequencies. How-
ever, they become less reliable when the frequency is low or when the wavelength is long. For
this regime, CEM methods are more reliable than approximate methods. One good example
is the scattering from an aircraft by mega-Hertz electromagnetic waves, where the wavelength
is not short enough to allow for short-wavelength approximation [66].
1.6 Time Domain versus Frequency Domain

In many physical problems, such as fluids, the equations are inherently nonlinear. Hence,
they do not lend themselves to Fourier transform methods. For many applications, however,
the electromagnetic equations are linear. Hence one can choose to solve the electromag-
netic equations in the frequency domain other than the time domain, as noted before. The
frequency-domain equations are obtained by Fourier transforming the electromagnetic equa-
tions in time.
Alternatively, one can achieve the simplicity of frequency domain by assuming a pure
time-harmonic solution that is completely sinusoidal where the time dependence at any given
point in space is of the form A cos(ωt + φ). Since one can write
1 ¡ −iφ −iωt ¢
A cos(ωt + φ) = Ae e + C.C. (1.17)
2
where C.C. stands for “complex conjugate”, it is simpler to work with complex time-harmonic
signals of the form e−iωt with complex amplitudes. Such signals are called phasors [34]. To
recover the real signal, one just adds the complex conjugate part to the complex signal, or
simply, takes the real part of the complex signal.
For linear equations, a pure time-harmonic solution can exist as an independent solution,
whereas for nonlinear equations, the incipient of a time-harmonic signal will generate higher
harmonics, causing the mixing of harmonics (see Section 1.2).
The advantages of solving the electromagnetic equations in the frequency domain is the re-
moval of time dependence of the electromagnetic equations, and hence the resultant equations
are simpler (all ∂/∂t can be replaced with −iω). Another advantage is that each frequency
component solution can be solved for independently of the other frequency component—hence,
they can be solved in an embarrassingly parallel fashion with no communication overhead at
all. This multitude of time-harmonic solutions for different frequencies can then be superposed
to obtain the time domain solution [67].
On the other hand, the time-domain solution has to be sought by considering the inter-
dependence of time steps. Although in parallel computing, interprocessor communication can
be reduced via the domain-decomposition method, some communications are still required
[68, 69].
1.7 Differential Equation versus Integral Equation

For a nonlinear problem, it is more expedient to solve the equations in their differential forms.
However, when a problem is linear, the principle of linear superposition allows us to first derive
the point source response of the differential equation. This is the Green’s function [19], or the
fundamental solution of the differential equation. With the Green’s function, one can derive
an integral equation whose solution solves the original problem. We shall illustrate this in
the frequency domain for a simple scalar wave equation [18].
Given the wave equation in the time domain
1 ∂2
∇2 φ(r, t) − φ(r, t) = s(r, t) (1.18)
v 2 ∂t2
By assuming a complex time-harmonic signal with e−iωt dependence, the wave equation
becomes the Helmholtz equation as follows with an arbitrary source on the right hand side:
∇2 φ(r, ω) + k 2 φ(r, ω) = s(r, ω), k = ω/v (1.19)
The Green’s function is the solution when the right-hand side is replaced by a point source,
∇2 g(r, r0 ) + k 2 g(r, r0 ) = −δ(r − r0 ) (1.20)

where the dependence of g(r, r0 ) on ω is implied. By the principle of linear superposition, the
solution to the original Helmholtz equation can be expressed as
Z
φ(r, ω) = − dr0 g(r, r0 )s(r0 , ω) (1.21)
V
For an impenetrable scatterer, the Green’s function can be used to derive an integral
equation of scattering Z
−φinc (r) = dS 0 g(r, r0 )s(r0 ), r ∈ S (1.22)
S
0
where φinc (r) is the incident field, g(r, r ) is the free-space Green’s function given by
0
0 eik|r−r |
g(r, r ) = (1.23)
4π|r − r0 |
and s(r0 ) is some unknown surface source residing on the surface of the scatterer. Physically,
the above equation says that the surface source s(r0 ) generates a field that cancels the incident
field within the scatterer. To solve the integral equation implies finding the unknown s(r0 ).
The advantage of solving an integral equation is that often times, the unknown s(r0 ) resides
on a 2D manifold or surface [70].3 Whereas if a differential equation is solved, the unknown
is the field φ(r) that permeates the whole of a 3D space. Consequently, it often requires more
unknowns to solve a differential equation compared to an integral equation. More precisely,
the unknown count for s(r0 ) on a 2D surface scales as L2 where L is the typical length of the
scatterer, whereas the unknown count for the field φ(r) scales volumetrically as L3 . Hence,
when the problem size is very large, the cruelty of scaling law will prevail, implying that
differential equations usually require more unknowns to solve compared to integral equation.
Another advantage of solving integral equation is that the Green’s function is an exact
propagator that propagates a field from point A to point B. Hence, there is no grid dispersion
error of the kind that exists in numerical differential equation solvers where the field is prop-
agated from point A to point B via a numerical grid. For wave problems, the grid dispersion
error causes an error in the phase velocity of the wave, causing cumulative phase error that
becomes intolerable. Hence, the grid dispersion error becomes a more severe problem with
increased problem size. The grid discretization density has to increase with increased problem
size, prompting the unknown scaling to increase with more than L3 [68, 71].
On the other hand, numerical differential equation solvers are easier to implement com-
pared to integral equation solvers. In integral equation solvers, a proper way to evaluate
integrals with singular integrand is needed. For an N unknown problem, the differential
equation solver also solves an equivalent sparse matrix system with O(N ) storage greatly
reducing the storage requirements, but the matrix system associated with integral equation
is usually a dense matrix system requiring O(N 2 ) storage and more than O(N 2 ) central pro-
cessing unit (CPU) time to solve. However, the recent advent of fast integral equation solver
has removed these bottlenecks [66, 70, 72–75]. Previously, only dense matrix systems with
tens of thousands of unknowns can be solved, but now, dense matrix systems with tens of
millions of unknowns can be solved [76, 77].
3 The pros and cons of differential equation and integral equation solvers were discussed in greater detail
in this reference. We are repeating them here for convenience.

1.8 Nondissipative Nature of Electromagnetic Field

Electromagnetic waves in vacuum and many media are nondissipative. That explains the
intergalactic distance over which electromagnetic waves can travel. The popular Yee scheme
[20] in CEM is nondissipative albeit dispersive. The nondissipative nature of a numerical
scheme preserves a fundamental property of the electromagnetic field. However, many nu-
merical schemes in CFD have been borrowed to solve electromagnetic problems [5, 9, 78–80].
These numerical schemes are dissipative, and hence, one has to proceed with caution when
applying these numerical schemes to solve CEM problems.
The nondissipative nature of the electromagnetic field also implies that two electromag-
netic systems can affect each other even though they may be very far apart. Even short
length scale phenomenon can propagate over long distances such as light rays.
1.9 Conclusions
Differential equation solvers are used almost exclusively in many fields because of the nonlin-
earity of the relevant equations. On the other hand, because of the linearity of the Maxwell-
Heaviside equations, analytic methods, such as the time-domain method and the integral
equation method, can be used to arrive at alternative methods to solve the electromagnetic
equations such as integral equation solvers. Traditional integral equation solvers are ineffi-
cient, but with the advent of fast solvers, they are more efficient than before, and have become
even more efficient than differential equation solvers in many applications.
There has been an interest in applying differential equation techniques to solve electromag-
netic equations. When applied to an aircraft scattering problem, these solvers use resources
that scale more pathological compared to fast integral equation solvers. One way to curb the
scaling law in resource usage is to combine differential equation solvers with integral equation
solvers where the computational domain is truncated by a fast integral equation solver ( [70]
Chapter 13, [81],). If memory and CPU resource are not an issue, differential equation tech-
niques can be used to obtain broadband simulation data in electromagnetics. If memory alone
is not an issue, fast time-domain integral equation solver can be used to obtain broadband
data in electromagnetics ( [70] Chapter 18, [82]).
The morphing of electromagnetic physics from static to long wavelengths, to mid wave-
lengths, and then to short wavelengths implies that different techniques need to be used in
these different regimes. Many complex problems remain in electromagnetics that call for more
sophisticated computational algorithms. One of these problems is the simulation of complex
structures inside a computer at the level of motherboard, package, and chip, as shown in Fig-
ure 1.1. It entails multiscale features that cannot be easily tackled when using conventional
methods.
A computer at the finest level consists of computer chips that contain tens of millions
of transistors. Here, X and Y lines in a multilayer fashion route electrical signals among
these transistors (Figure 1.1(c)). A chip forms a data processing unit that has thousands
of electrical input and output wires. A chip communicates outside itself including receiving
power supply via the package (Figure 1.1(b)). Different processing units together form a
system, and different systems, including a power supply system, form a computer as seen at
(a) Board level complexity. (b) Package level complexity. (c) Chip level complexity.
Figure 1.1: A complex multiscale structure coming from the real world as found inside a
computer at the level of (a) motherboard, (b) package, and (c) chip. (Courtesy of Intel.)
the motherboard level (Figure 1.1(a)) just inside the computer chassis.
Appendix: Complexity of an Algorithm

In almost all modern numerical methods of solving electromagnetic problems, an object (or
region, or domain) is first decomposed into many small subobjects (or subregions, or subdo-
main) which are much smaller than a wavelength. Then a matrix equation is derived that
accounts for the interaction of these subobjects (or subregions, or subdomains). Each sub-
oject is usually given one or more unknowns. The number of subojects (or subregions, or
subdomains) represents the degree of freedom of an object, and we often say that an ob-
ject has been discretized in a numerical method. The number of unknowns to be solved for
represents the degree of freedom of the system.
The two most important measures of the efficiency of a numerical method are: (1) the
memory requirements, and (2) the computational time. Ideally, both of them should be
small. For example, the storage of the unknowns in the system, and the pertinent matrix
elements requires the usage of the random access memory (RAM) of a computer for efficiency.
The solving of the matrix equation requires computational resource such as CPU time. The
memory usage may scale as O(N α ) and the CPU time may scale as O(N β ) where α and β
are numbers larger than one. These scaling properties of the memory and computation time
with respect to the unknown number N are respectively known as the memory complexity
and the computational complexity.
At this point, it is prudent to define the meaning of O(N β ) (read as “order N β ”) and
similar expressions. When
R ∼ O(N β ), N → ∞, (A-1)
one implies that there exists a constant C such that
R → CN β , N → ∞, (A-2)
where R is the resource usage: it can be memory or CPU time. In this case, we say that
the asymptotic complexity of the algorithm is O(N β ) or just N β . For large-scale computing,
where N is usually very large, the term “complexity” refers to “asymptotic complexity”. The
complexity of an algorithm also determines the scaling property of the resource usage by the
algorithm. For instance, if β = 2, the resource usage quadruples every time the number of
unknowns, N , doubles.
When R refers to CPU time, the constant C in Equation (A-2) depends on the speed of
the hardware, and the detail implementation of the algorithm. However, the complexity of
an algorithm does not depend on the constant C.
Depending on the algorithm used, α can be as large as two and β can be as large as three.
The intent of an efficient numerical algorithm is to make these numbers as close to one as
possible.
When α and β are large, the resource needed can easily swamp the largest supercomputers
we have. For instance, Gaussian elimination for inverting a matrix is of complexity N 3 . The
CPU time can be easily of the order of years when N is large. This is known as the “cruelty
of computational complexity” in computer science parlance.
Recently, the advent of fast algorithms in reducing α and β when solving these equations,
as well as the leaps and bounds progress in computer technology, which helps to reduce the
constant C in Equation (A-2), make many previously complex unsolvable problems more
accessible to scientists and engineers using a moderately sized computer.
Alternatively, some workers have designed algorithms to solve problems using the hard
drive as well as the RAM as storage space. These are usually known as out-of-core solvers.
Even then, if the memory and computational complexity of the algorithm is bad, the hard-
drive memory can easily become insufficient.
Parallel computers can be used to speed up solutions to large problems, but they do
not reduce the computational complexity of an algorithm. If a problem takes CPU time
proportional to N β , namely,
T → CN β , N → ∞, (A-3)
an ideal parallel algorithm to solve the same problem with P processors has CPU time that
scales as
1
T → CN β , N → ∞. (A-4)
P
Such an ideal algorithm is said to have a perfect speed up, and 100% parallel efficiency.
However, because of the communication cost needed to pass data between processors, the
latency in computer memory access as problem size becomes large, 100% parallel efficiency
is generally not attained. In this case, the CPU time scales as
e
T → CN β , N → ∞, (A-5)
100P
where e is the parallel efficiency in percentage. As can be seen from above, the use of parallel
computers does not reduce the computational complexity of the algorithm, but it helps to
reduce the proportionality constant in the resource formula.
Bibliography
[1] D. E. Smith, History of Mathematics, New York: Dover Publication, 1923.
[2] A. R. Choudhuri, The Physics of Fluids and Plasma, Cambridge UK: Cambridge Univer-
sity Press, 1998.
[3] C. L. M. H. Navier, “Mémoire sur les lois du mouvement des fluides,” Mém. Acad. Sci.
Inst. France, vol. 6, pp. 389–440, 1822.
[4] G. G. Stokes, “On the theories of the internal friction of fluids in motion, and of the
equilibrium and motion of elastic solids,” Trans. Cambridge Philos. Soc., vol. 8, pp. 287–
319, 1845.
[5] J. C. Tannehill, D. A. Anderson, and R. H. Pletcher, Computational Fluid Mechanics

and Heat Transfer, First edition, New York: Hemisphere Publ., 1984. Second edition,
Philadelphia: Taylor & Francis, 1997.
[6] L. F. Richardson, “The approximate arithmetical solution by finite differences of physical

problems involving differential equations, with an application to th stresses in a masonry
dam,” Philos. Trans. R. Soc. London, Ser. A, vol. 210, pp. 307–357, 1901.
[7] R. Courant, K. O. Friedrichs, and H. Lewy, “Über die partiellen differenzengleichungen

der mathematischen physik,” Math. Ann., vol. 100, pp. 32–74, 1928. (Translated to: “On
the partial difference equations of mathematical physics,” IBM J. Res. Dev., vol. 11, pp.
215–234, 1967.
[8] P. D. Lax, “Weak solutions of nonlinear hyperbolic equations and their numerical com-
putation,” Commun. Pure Appl. Math., vol. 7, pp. 159–193, 1954.
[9] P. D. Lax and B. Wendroff, “Systems of conservation laws,” Commun. Pure Appl. Math.,
vol. 13, pp. 217–237, 1960.
[10] R. W. MacCormack, “The effect of viscosity in hypervelocity impact cratering,” AIAA

Paper 69-354, Cincinnati, OH, 1969.
[11] P. E. Ceruzzi, A History of Modern Computing, Cambridge, MA: MIT Press, 1998.
1
[12] M. Faraday, “On static electrical inductive action,” Philos. Mag., 1843. M. Faraday,
Experimental Researches in Electricity and Magnetism. vol. 1, London: Taylor & Francis,
1839.; vol. 2, London: Richard & John E. Taylor, 1844; vol. 3, London: Taylor & Francis,
1855. Reprinted by Dover in 1965. Also see M. Faraday, “Remarks on Static Induction,”
Proc. R. Inst., Feb. 12, 1858.
[13] A. M. Ampère, “Mémoire sur la théorie des phénomènes électrodynamiques,” Mem.

Acad. R. Sci. Inst. Fr., vol. 6, pp. 228–232, 1823.
[14] C. S. Gillmore, Charles Augustin Coulomb: Physics and Engineering in Eighteenth Cen-
tury Frrance, Princeton, NJ: Princeton University Press, 1971.
[15] C. F. Gauss, “General theory of terrestial magnetism,” Scientific Memoirs, vol. 2, ed.
London: R & J.E. Taylor, pp. 184–251, 1841.
[16] J. C. Maxwell, A Treatise of Electricity and Magnetism, 2 vols, Clarendon Press, Oxford,
1873. Also, see P. M. Harman (ed.), The Scientific Letters and Papers of James Clerk
Maxwell, Vol. II, 1862–1873, Cambridge, U.K.: Cambridge University Press, 1995.
[17] O. Heaviside, “On electromagnetic waves, especially in relation to the vorticity of the
impressed forces, and the forced vibration of electromagnetic systems,” Philos. Mag., 25,
pp. 130–156, 1888. Also, see P. J. Nahin, “Oliver Heaviside,” Sci. Am., pp. 122–129, Jun.
1990.
[18] W. C. Chew, Waves and Fields in Inhomogeneous Media, New York: Van Nostrand
Reinhold, 1990, reprinted by Piscataway, New Jersey: IEEE Press, 1995.
[19] G. Green, An Essay on the Application of Mathematical Analysis to the Theories of

Electricity and Magnetism, T. Wheelhouse, Nottingham, 1828. Also, see L. Challis and
F. Sheard, “The green of Green functions,” Phys. Today, pp. 41–46, Dec. 2003.
[20] K. S. Yee, “Numerical solution of initial boundary value problems involving Maxwell’s
equation in isotropic media,” IEEE Trans. Antennas Propag., vol. 14, pp. 302–307, 1966.
[21] R. F. Harrington, Field Computation by Moment Method, Malabar, FL: Krieger Publ.,
1982. (Original publication, 1968.)
[22] Y. R. Shen, The Principles of Nonlinear Optics, New York: John Wiley & Sons, 1991.
[23] G. P. Agrawal, Fiber-Optic Communication Systems, New York: John Wiley & Sons,
1997.
[24] W. C. Chew, “Computational Electromagnetics—the Physics of Smooth versus Oscilla-

tory Fields,” Philos. Trans. R. Soc. London, Ser. A, Math., Phys. Eng. Sci. Theme Issue
Short Wave Scattering, vol. 362, no. 1816, pp. 579–602, Mar. 15, 2004.
[25] R. J. LeVeque, Numerical Methods for Conservation Laws, ETH Lectures in Mathematics
Series, Basel: Birkhuser Verlag, 1990.
[26] R. P. Federonko, “A relaxation method for solving elliptic equation,” USSR Comput.
Math. Phys., vol. 1, pp. 1092–1096, 1962.
[27] A. Brandt, “Multilevel adaptive solutions to boundary value problems,” Math. Comput.,
vol. 31, pp. 333–390, 1977.
[28] O. Axelsson and V. A. Barker, Finite Element Solution of Boundary Value Problems:
Theory and Computation, New York: Academic Press, 1984.
[29] S. Jaffard, “Wavelet methods for fast resolution of elliptic problems,” SIAM J. Nu-
mer. Anal., vol. 29, no. 4, pp. 965–986, 1992.
[30] R. L. Wagner and W. C. Chew, “A study of wavelets for the solution of electromagnetic
integral equations,” IEEE Trans. Antennas Propag., vol. 43, no. 8, pp. 802–810, 1995.
[31] D. R. Wilton and A. W. Glisson, “On improving the electric field integral equation at
low frequencies,” URSI Radio Sci. Meet. Dig., pp. 24, Los Angeles, CA, Jun. 1981.
[32] J. S. Zhao and W. C. Chew, “Integral Equation Solution of Maxwell’s Equations from
Zero Frequency to Microwave Frequencies,” IEEE Trans. Antennas Propag., James R.
Wait Memorial Special Issue, vol. 48. no. 10, pp. 1635–1645, Oct. 2000.
[33] W. C. Chew, B. Hu, Y. C. Pan and L. J. Jiang, “Fast Algorithm for Layered Medium,”
C. R. Phys., vol. 6, pp. 604–617, 2005.
[34] J. A. Kong, Electromagnetic Wave Theory, New York: John Wiley & Sons, 1986.
[35] M. Born and E. Wolf, Principles of Optics, New York: Pergammon Press, 1980.
[36] B. D. Seckler and J. B. Keller, “Geometrical theory of diffraction in inhomogeneous

media,” J. Acoust. Soc. Am., vol. 31, pp. 192–205, 1959.
[37] J. Boersma, “Ray-optical analysis of reflection in an open-ended parallel-plane waveguide:

II-TE case,” Proc. IEEE, vol. 62, pp. 1475–1481, 1974.
[38] H. Bremmer, “The WKB approximation as the first term of a geometric-optical series,”
Commun. Pure Appl. Math., vol. 4, pp. 105, 1951.
[39] D. S. Jones and M. Kline, “Asymptotic expansion of multiple integrals and the method
of stationary phase,” J. Math. Phys., vol. 37, pp. 1–28, 1958.
[40] V. A. Fock, Electromagnetic Diffraction and Propagation Problems, New York: Pergamon
Press, 1965.
[41] R. C. Hansen, (ed.), Geometric Theory of Diffraction, Piscataway, New Jersey: IEEE
Press, 1981.
[42] R. G. Kouyoumjian and P. H. Pathak, “A uniform geometrical theory of diffraction for

an edge in a perfectly conducting surface,” Proc. IEEE, vol. 62, pp. 1448–1461, 1974.
[43] S. W. Lee and G. A. Deschamps, “A uniform asymptotic theory of electromagnetic

diffraction by a curved wedge,” IEEE Trans. Ant. Progat., vol. 24, pp. 25–35, 1976.
[44] P. H. Pathak, “An asymptotic analysis of the scattering of plane waves by a smooth
convex cylinder,” Radio Sci., vol. 14, pp. 419, 1979.
[45] S. W. Lee, H. Ling, and R. C. Chou, “Ray tube integration in shooting and bouncing
ray method,” Microwave Opt. Tech. Lett., vol. 1, pp. 285–289, Oct. 1988.
[46] H. Haken, Light: Waves, Photons, Atoms, North Holland, Amsterdam, 1981.
[47] M. D. Van Dyke, Perturbation Methods in Fluid Mechanics, New York: Academic Press,
1975.
[48] J. Kevorkian, J. D. Cole, Perturbation Methods in Applied Mathematics, New York:

Spring Verlag, 1985.
[49] A. H. Nayfeh, Perturbation Methods, New York: Wiley-Interscience, 1973.
[50] G. Kirchhoff, “Zur Theorie des Kondensators,” Monatsb. Deutsch. Akad. Wiss., Berlin,
pp. 144, 1877.
[51] S. Shaw, “Circular-disk viscometer and related electrostatic problems,” Phys. Fluids, vol.
13, no. 8, pp. 1935, 1970.
[52] W. C. Chew and J. A. Kong, “Asymptotic formula for the capacitance of two oppositely
charged discs,” Math. Proc. Cambridge Philos. Soc., vol. 89, pp. 373–384, 1981.
[53] W. C. Chew and J. A. Kong, “Microstrip capacitance for a circular disk through matched
asymptotic expansions,” SIAM J. Appl. Math., vol. 42, no. 2, pp. 302–317, Apr. 1982.
[54] S. Y. Poh, W. C. Chew, and J. A. Kong, “Approximate formulas for line capacitance
and characteristic impedance of microstrip line,” IEEE Trans. Microwave Theory Tech.,
vol. MTT-29, no. 2, pp. 135–142, Feb. 1981.
[55] W. C. Chew and J. A. Kong, “Asymptotic eigenequations and analytic formulas for the
dispersion characteristics of open, wide microstrip lines,” IEEE Trans. Microwave Theory
Tech., vol. MTT-29, no. 9, pp. 933–941, Sept. 1981.
[56] W. C. Chew and J. A. Kong, “Asymptotic formula for the resonant frequencies of a
circular microstrip antenna,” J. Appl. Phys., vol. 52, no. 8, pp. 5365–5369, Aug. 1981.
[57] J. W. Strutt Rayleigh (Lord Rayleigh), “On the passage of electric waves through tubes,
or the vibra cylinder,” Phi. Mag., vol. 43, pp. 125–132, 1897.
[58] E. A. Ash and E. G. S. Paige (eds.), Rayleigh Wave Theory and Application, Berlin:
Spring-Verlag, 1985.
[59] A. Sommerfeld, “Mathematische Theorie der Diffraction,” Math. Ann., vol. 47, no. s319,
pp. 317–374, 1896.
[60] A. Sommerfeld, “Uber die Ausbreitung der Wellen in der drahtlosen Telegraphie,”
Ann. Phys., vol. 28, pp. 665–737, 1909.
[61] G. Mie, “Beiträge zur optik trüber medien speziel kolloidaler metallösungen,”
Ann. Phys. (Leipzig), vol. 25, pp. 377, 1908.
[62] P. Debye, “Der Lichtdruck auf Kugeln von beliebigem Material,” Ann. Phys., vol. 4,
no. 30, pp. 57, 1909.
[63] J. W. Strutt Rayleigh (Lord Rayleigh), Theory of Sound, New York: Dover Publ.,
1976. (Originally published 1877.)
[64] J. J. Bowman, T. B. A. Senior, and P. L. E. Uslenghi, Electromagnetic and Acoustic
Scattering by Simple Shapes, Amsterdam: North-Holland, 1969.
[65] G. N. Watson, “The Diffraction of Electric Waves by the Earth,” Proc. R. Soc. London.
Ser. A, vol. 95, 1918, pp. 83–99.
[66] J. M. Song, C. C. Lu, W. C. Chew, and S. W. Lee, “Fast Illinois solver code (FISC),”
IEEE Antennas Propag. Mag., vol. 40, no. 3, pp. 27–33, 1998.
[67] J. M. Song and W. C. Chew, “Broadband time-domain calculations using FISC,” IEEE
Antennas Propag. Soc. Int. Symp. Proc., San Antonio, Texas, vol.3, pp. 552–555, Jun.
2002.
[68] A. Taflove, Computational Electrodynamics: The Finite Difference Time Domain
Method, Norwood, MA: Artech House, 1995.
[69] W. Hoefer, V. Tripathi, and T. Cwik, “Parallel and Distributed Computing Techniques
for Electromagnetic Modeling,” in IEEE APS/URSI Joint Symposia, University of Wash-
ington, 1994.
[70] W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song, eds., Fast and Efficient Algo-
rithms in Computational Electromagnetics, Berlin: Artech House, 2001.
[71] R. Lee and A. C. Cangellaris, “A study of discretization error in the finite element
approximation of wave solution,” IEEE Trans. Antennas Propag., vol. 40, no. 5, pp. 542-
549, 1992.
[72] V. Rokhlin, “Rapid solution of integral equations of scattering theory in two dimensions,”
J. Comput. Phys., vol. 36, no. 2, pp. 414–439, 1990.
[73] C. C. Lu and W. C. Chew, “A multilevel algorithm for solving boundaryvalue scattering,”
Microwave Opt. Tech. Lett., vol. 7, no. 10, pp. 466-470, Jul. 1994.
[74] J. M. Song and W. C. Chew, “Multilevel fastmultipole algorithm for solving combined
field integral equations of electromagnetic scattering,” Microwave Opt. Tech. Lett., vol.
10, no. 1, pp. 14-19, Sept. 1995.
[75] B. Dembart and E. Yip, “A 3D fast multipole method for electromagnetics with multiple
levels,” Eleventh Ann. Rev. Prog. Appl. Comput. Electromag., vol. 1, pp. 621-628, 1995.
[76] S. Velamparambil, W. C. Chew, and J. M. Song, “10 million unknowns, is it that large,”
IEEE Antennas Propag. Mag., vol.45, no.2, pp. 43–58, Apr. 2003.
[77] M. L. Hastriter, “A Study of MLFMA for Large-Scale Scattering Problems,” Ph.D.

Thesis, Dept. ECE, U. Illinois, Jun. 2003.
[78] V. Shankar, “A gigaflop performance algorithm for solving Maxwell’s Equations of elec-
tromagnetics,” AIAA Paper 911578, Honolulu, Jun. 1991.
[79] J. S. Shang, and D. Gaitonde, “Scattered Electromagnetic Field of a Reentry Vehicle,”

AIAA Paper 940231, Reno, Nevada, Jan. 1994.
[80] J. S. Shang, “Computational Electromagnetics,” ACM Comput. Surv., vol. 28, no. 1, pp.
97–99, 1996.
[81] J. M. Jin and V. V. Liepa, “Application of hybrid finite element method to electro-
magnetic scattering from coated cylinders,” IEEE Trans. Antennas Propag., vol. 36, pp.
50–54, Jan. 1988.
[82] A. A. Ergin, B. Shanker, and E. Michielssen, “Fast evaluation of transient wave fields
using diagonal translation operators,” J. Comput. Phys., vol. 146, pp. 157–180, 1998.
21
CHAPTER 2
Linear Vector Space,

Reciprocity, and
Energy Conservation
2.1 Introduction
Linearity allows us to use the Fourier transform technique to transform and simplify the
electromagnetic equations. We express all field and current functions of time using Fourier
transform. For example, Z ∞
1
F(r, t) = F(r, ω)e−iωt dω (2.1)
2π −∞
Substituting the above into the electromagnetic equations for the fields and currents, the
equations become independent of time in the Fourier space, but the fields and currents are
now functions of frequency and they are complex valued. Furthermore, ∂F(r, t)/∂t Fourier
transforms to −iωF(r, ω). Consequently, the electromagnetic equations in the Fourier space
become
∇ × E(r, ω) = iωB(r, ω) − M(r, ω) (2.2)

∇ × H(r, ω) = −iωD(r, ω) + J(r, ω) (2.3)
∇ · D(r, ω) = ρe (r, ω) (2.4)
∇ · B(r, ω) = ρm (r, ω) (2.5)
Each of these fields and currents can then be considered a representation of a time-harmonic
signal. These complex fields and currents in the Fourier space are generally referred to as
being in the frequency domain. The magnetic current M(r, ω), and the magnetic charge
ρm (r, ω) are regarded as fictitious, and are added to symmetrize the above equations.
In the electromagnetic equations, the second two equations are derivable from the first two
equations for time varying fields and current. Hence, there are actually only two equations
2
with four unknown quantities E, H, D, and B, assuming that the driving sources of the
system J and M are known. The constitutive relations, which describe the material media,
have to be invoked to render the above equations solvable [1]. For free space, the constitutive
relations are B = µ0 H and D = ²0 E, thus reducing the number of unknowns from four to
two.
If the material media is nonlinear for which there is a nonlinear relation between D and E
for instance, then a purely time-harmonic E will produce a D with more than one harmonics
because of the second and higher order harmonic generation. Then a pure time-harmonic
(monochromatic) signal cannot exist in such a system.
As mentioned in the Chapter 1, linearity permits the definition of the Green’s function
and the expression of solutions by invoking the principle of linear superposition. Conse-
quently, the linearity of electromagnetic equations makes their solutions amenable to a large
number of mathematical techniques, even though many of these techniques are restricted to
linear problems only, such as linear vector space techniques. Because there are a number of
mathematical concepts that are important for understanding computational electromangetics
(CEM) and integral equation solvers, these concepts will be reviewed here.
This chapter is necessary also because a large part of electromagnetic theory can be un-
derstood from the concepts of linear vector spaces. The introduction of linear vector space
concepts to students of computational electromagnetics is mandatory (de rigueur) because
such concepts are very useful in understanding wave physics and computational electromag-
netics (CEM). But many students of electromagnetics are not introduced to such concepts.
In this chapter, additionally, we will look at the reciprocity theorem and the energy
conservation theorem from the viewpoint of a linear vector space [2]. It can be shown that
these properties are intimately related to the symmetry and Hermitian properties of the linear
operators in electromagnetics.
Some of the concepts expressed here are buried in, or may be derived from, the works of
Reed and Simon [3], Felsen and Marcuvitz [4], Hanson and Yakovlev [5], or Collin [6], but
the points expressed here have not been explicitly stated before.
2.2 Linear Vector Spaces

For linear electromagnetics, the computation is done in a linear vector space with linear
operators. Examples of linear vector spaces are found in linear algebra, where one usually
deals with vectors in N dimensions. Solving linear systems with N unknowns forms the
underlying gist of CEM. How then do we arrive at such systems from the electromagnetic
equations?
The electromagnetic equations are defined in a continuum space. The continuum space
entails four physical space-time dimensions described by the x, y, z, and t variables. Because
these space-time variables are infinitely divisible, electromagnetic fields and sources are de-
scribed by infinite degree of freedom. One can think of them as an equivalent matrix system
with uncountably infinite number of variables. An example of such infinite dimensional space
with uncountably infinite degree of freedom is the Hilbert space.
However, it is cumbersome if not impossible to perform computations in this infinite di-
mensional space. Hence, a systematic way is needed to approximate this space with a discrete,
LINEAR VECTOR SPACE, RECIPROCITY, AND ENERGY CONSERVATION 23
finite dimensional space. One way to arrive at a discrete, finite system is with the subspace
projection method. This is the underlying principle behind the finite element method (FEM)
and the method of moments (MOM). Another method is to use finite-difference approxima-
tion of space time in a finite domain of space, or to use Nyström method to approximate an
integral equation. In Nyström method, integrals are approximated with quadrature rules, so
that a continuum integral is replaced with a discrete summation.
Most of the linear vector space concepts are similar to those in vector space in linear
algebra, except that the vectors can be infinite dimensional. These vectors can be countably
infinite, or uncountably infinite. A finite summation in linear algebra will be replaced by an
infinite summation, or an integral, which may or may not converge.
A linear vector space V consists of a set of infinitely many vectors that can be added and
scaled. The added vectors are members of the space V , and so are the scaled vectors. A set
of vectors, {fn , n = 1, . . . , N }, are linearly independent if
N
X
an fn = 0 (2.6)
n=1
implies that an = 0, n = 1, . . . , N . This set of vectors spans the vector space V if every
element f in the vector space can be written as
N
X
f= an fn (2.7)
n=1
The above concept can be extended to the case when N is infinite, or when the vectors
are uncountably infinite. In the case of uncountably infinite set, an integral replaces the
summation above.
For this linear vector space, we can define linear operators (linear transformations) that
transform a vector from linear vector space V to another vector in linear vector space W . For
example we can write
f = Lv (2.8)
where f is a new vector generated when the operator L operates on v. Here, f can be in the
same space as v lives in or it can be outside the space that v is in. The space of functions
that L can operate on is known as the domain of L or D(L). The space spanned by vectors
that Lv can generate for all v ∈ D(L) is known as the range of L or R(L).
The domain or range of L can change depending on the form of L. For instance, if L
contains derivatives, they may render the action of L on some functions undefined, eliminating
these functions from the domain and range of L in the classical sense. However, one can
expand the domain and range to contain distributions.
In electromagnetics, L can be a differential operator, an integral operator or a mixture
thereof. We can think of (2.8) as the elevation of linear algebra concept where L is analogous
to a matrix operator and v, and f are simple vectors in matrix algebra. Instead of being
indexed by discrete indices in matrix algebra, (2.8) can represent functions whose “indices”
are a continuum, and hence of infinite dimension. For example, an explicit form of (2.8) in
one dimension is Z b
f (x) = dx0 L(x, x0 )v(x0 ), a<x<b (2.9)
a
This is the analogy of

f =L·v (2.10)
in matrix algebra.
Also, in electromagnetics, one works with fields in the Fourier space or the frequency
domain, which allows for complex numbers. Hence, in frequency-domain electromagnetics, a
vector space embodies complex-valued vectors as elements. As such, the inner products in
electromagnetics have to be defined so that they can account for the possibilities of different
physical interpretations. We will first outline the inner products in electromagnetics and
compare them with the inner products commonly used in mathematics and quantum physics.
2.3 Inner Products for Electromagnetics

To use any subspace projection method, we need to define an inner product in the linear
vector space. Before defining an inner product, one should be cognizant that a function f (x)
is actually the analogy of a component of a vector fj in linear algebra, where the former is
“indexed” by an indenumerable variable x, whereas the latter is indexed by a denumerable
integer. This analogy can be extended to multidimensional functions in space. By drawing
upon this analogy, the inner product in an infinite dimensional space can be defined.
In electromagnetics, the inner product between two functions is defined as [7–10]
Z
hf, gi = f (r)g(r)dr, (2.11)
where the integral is assumed to converge. This inner product involves elements f (r) and
g(r), which either live in one, two, or three dimensions.
The above is analagous to
N
X
ft · g = fi gi (2.12)
i=1
in matrix algebra, and hence, we call this the inner product between two functions or infinite
dimensional vectors. The difference between an inner product of a finite dimensional vector
and an infinite dimensional vector is that the latter inner product may not exist because the
integral may diverge.
The inner product in Eq. (2.11) is physically related to the reaction between two field
quantities [11]. We shall call the above the reaction inner product. In the literature, it has
been called the pseudo inner product [5] or the bilinear form [12]. In subsequent sections,
unless otherwise stated, we will use h, i to imply an electromagnetic or reaction inner product.
The electromagnetic reaction inner product has its usefulness: it can be shown that for a
large class of media, called reciprocal media, the electromagnetic equations are analogous to
a complex symmetric system in linear algebra.
In CEM, when the reaction inner product is used, it is to find the matrix representation of
electromagnetic operators. If the electromagnetic operator is symmetric, the matrix system
obtained is complex symmetric, resulting in storage reduction. Because the reaction inner
product is given the physical meaning of a reaction, such a physical concept can be used as
a sanity check in many experimental measurements and engineering designs [11].
When we need to define an inner product for the energy, we will explicitly denote one of
the vectors with a complex conjugation, namely,
Z
hf ∗ , gi = f ∗ (r)g(r)dr, (2.13)
where f ∗ (r) means the complex conjugation of f (r). Because the above has the physical
meaning of complex energy or power [13], we shall call the above the energy inner product.
In this paper, f and g can be n-vectors (column vectors with n elements). In this case, the
products in the integrands of the right-hand sides of (2.11) and (2.13) are replaced by dot
products.
2.3.1 Comparison with Mathematics

The above inner product is different from that defined in the mathematical literature [12,14],
where the inner product is invariably defined as
Z
hf, giH = f ∗ (r)g(r)dr. (2.14)
where the subscript H emphasizes that the conjugation of f is implied. The above inner
product has its elegance because it can be used to define the norm of a function as
Z
||f || = f ∗ (r)f (r)dr.
2
(2.15)
With this definition, the norm of a function ||f ||2 > 0 if f 6= 0. A norm can be used as a
metric to measure the “distance” between two vectors.
A linear vector space with a defined norm is known as a normed vector space. Different
norms can be used to distinguish different normed linear vector spaces. For example, the set
of all functions f such that the integral in (2.15) exists and is of finite value is known as the
L2 space. More complicated norms can be defined to allow for different normed spaces. For
instance, a norm can be defined such that
Z Z
2
||f ||A = f ∗ (r)A(r, r0 )f (r0 )drdr0 , (2.16)
where f and A are chosen so that the above is always a bounded positive number. When
||f ||2A is positive for all candidate f , then A is termed a positive operator. When ||f ||2A is
bounded, for all f such that ||f || = 1, then A is termed bounded.
2.3.2 Comparison with Quantum Mechanics

The electromagnetic inner products are different from the inner product of quantum mechanics
[15, 16]. In quantum mechanics, one studies quantum systems, which have (experimentally)
measurable quantities called observables and state vectors describing the state of a quantum
system. The observables are measurable quantities such as momentum and position, and
they are represented by operators in a quantum system. The state vector is represented by a
function f (r). The inner product between two state vectors is defined in Dirac’s bra and ket
notation as Z
hf |gi = f ∗ (r)g(r)dr. (2.17)
This is similar to the energy inner product previously defined. If f = g, then the integrand
on the right-hand side represents a probability function, and this definition of inner product
ensures the positivity of this integrand. Probability functions, which are
R normalized to one,
cannot be destroyed. In other words, the net “energy” of a state vector, |f (r)|2 dr, is always
constant. Hence, a quantum system is “lossless” from that perspective. This is reflected by
the fact that the Hamiltonian operator that governs the time evolution of the state vector is
Hermitian, with real eigenvalues.
To convert an observable operator A of a quantum system into something that can be
measured in the laboratory, one computes the quantity
Z
hf |A|f i = f ∗ (r)A(r, r0 )f (r0 )drdr0 , (2.18)
which is called the expectation value of the operator A. Because this value represents a
measurable quantity, it must always be real. Hence, the operator A has to be self-adjoint or
Hermitian to ensure its reality. In other words, the definition of quantum-mechanics inner
product ensures the reality of the expectation value.
2.4 Transpose and Adjoint of an Operator

Many mathematics books define the adjoint of an operator, but not the transpose. Such
definitions may be sufficient for quantum mechanics, but are not suitable for electromagnetics
[10]. We define the transpose of an operator G, denoted by Gt , to be
hf, Ggi = hg, Gt f i, ∀g ∈ D(G), ∀f ∈ R(G), (2.19)
where D(G) and R(G) stand for the domain and the range of G, respectively. The definition
of the transpose of a matrix is simple because we just need to swap the indices. But in an
infinite-dimensional vector space, it is more convenient to define the transpose using inner
products given by (2.19). Notice that the range and the domain of Gt are swapped compared
to those of G. An operator is symmetric if Gt = G.1
The adjoint of an operator G, denoted by Ga , is defined by
hf ∗ , Ggi = hg ∗ , Ga f i∗ , ∀g ∈ D(G), ∀f ∈ R(G). (2.20)
This concept is similar to the conjugate transpose of a matrix. An operator is self-adjoint if

Ga = G. It is similar to the concept of a Hermitian matrix. Hence, we shall use G† to denote
the adjoint of an operator as well, which is a common notation used to denote the conjugate
transpose of a matrix in linear algebra.
1 In many textbooks, this is called the “adjoint” instead of the “transpose” of an operator [6]. We prefer
to call this the “transpose” because of its analogy to the “transpose” concept in linear algebra.
2.5 Matrix Representation

An operator in an infinite-dimensional space is difficult to compute. It is customary to reduce
such an operator to an approximate, finite-dimensional matrix operator by finding its matrix
representation.
Given an operator equation
Lf = g, (2.21)
where f and g are functions, we can let
N
X
f (r) ≈ fn bn (r), (2.22)
n=1
where bn (r) are linearly independent vectors or functions, fn are constants, DN = span{bn (r), n =
1, . . . , N } is a space that approximates the original domain space of L, or D(L). DN may be
a subspace of D(L) because DN is finite-dimensional but D(L) is infinite-dimensional. If L is
a differential operator, then f in (2.21) is nonunique unless boundary conditions are specified.
Hence, L has to be augmented with boundary conditions to be invertible.
We can use (2.22) in (2.21), multiply the resultant equation with
tm (r), m = 1, . . . , N
and integrate to obtain

N
X
fn htm , Lbn i = htm , gi, m = 1, . . . , N. (2.23)
n=1
In order for (2.21) to have a solution, g must be in the range space R(L) of L. Therefore,
if RN = span{tm , m = 1, . . . , N } is a good approximation to R(L), then (2.23) is a good
replacement equation to solve instead of (2.21). Moreover, (2.23) can be written in a matrix
form
L · f = g, (2.24)
where
[L]mn = Lmn = htm , Lbn i, (2.25)
[f ]n = fn , (2.26)
[g]n = htn , gi = gn . (2.27)
In the above, L is the matrix representation of the operator L, and f and g are the vector
representations of the functions f and g in the subspaces spanned by {fn , n = 1, . . . , N } and
{gn , n = 1, . . . , N }, respectively.
The functions bn (r), n = 1, . . . , N are known as the expansion (or basis) functions, and
together, they form a basis that spans the subspace that approximates the domain of the
operator. The functions tn (r), n = 1, . . . , N are known as the testing (or weighting) functions,
and together, they form a basis that spans the subspace that approximates the range of the
operator. The above method is a subspace projection method, where the linear operator has
been approximated by subspaces that approximate its domain and its range spaces.
The above method of finding the matrix representation of an operator is known as the
method of moments (MOM) or the boundary element method (BEM) when the operator is
an integral operator. It is known as the finite element method (FEM) when the operator is a
differential operator; the name Galerkin’s method is also used when the testing function is the
same as the expansion function. Mathematically, MOM and FEM are similar methods, where
an operator equation in an infinite-dimensional space is projected onto a subspace spanned
by a finite set of expansion functions.
When the operator is symmetric, the range space is the same as the domain space. In
this case, the matrix representation of the original operator is a symmetric matrix when the
testing functions are the same as the expansion functions, and memory saving in matrix
storage ensues.
Because we have approximated an infinite dimensional space with a subspace of finite
dimension, the matrix representation will not be an exact representation. But it can be made
increasingly accurate by increasing N if the set of expansion functions is complete when N
tends to infinity. In other words, they can be used to approximate any function to arbitrary
accuracy when N tends to infinity.
However, letting N tend to infinity is impractical for most computations. The trick is
on how to obtain a good solution with finite computational resource, that is, with N being
a finite number. Invocation of heuristics is necessary in this case. To make (2.24) a high
fidelity representation of the original equation (2.21), the vector representations f and g must
be good representations of the original functions f and g. For instance, if g is singular, one
may invoke heuristics to infer that f is also singular, and hence, appropriately distribute the
expansion functions to capture the essence of their singularities. Otherwise, an exorbitantly
large number of expansion functions may be needed to approximate these singular functions
well.
As previously noted, an operator L may operate on some function f that may produce a
field g that is singular. But in CEM, we are interested in Lmn = htm , Lbn i being a computable
quantity. There could be choice of expansion and testing functions that may render the matrix
element Lmn infinite or undefined. This is highly undesirable in a numerical computation.
Hence, we restrict our expansion and testing functions with computable inner products when
the matrix representation of the operator is sought. This is the same as requiring physical
quantities to be finite, because the inner product in electromagnetics can be interpreted as
either a reaction or an energy.
Oftentimes, seeking the matrix representation of an operator can also be related to the
minimization of a quadratic functional by the Rayleigh-Ritz method ( [10, Chapter 5]). The
minimization of the quadratic functional can be shown to be equivalent to solving the original
operator equation.
2.6 Compact versus Noncompact Operators

A smooth function, which is bandlimited, can be approximated well by a finite number of
expansion functions, such as Fourier series or a set of polynomials, to within exponential
accuracy. A compact operator can be thought of as a smoothing operator. It converts a

function with high spectral components to a smooth function. In other words, the range
space of a compact operator can be approximated by a finite number of expansion function
accurately.
In CEM, the matrix representations of the operators are often obtained by using subdo-
main expansion and testing functions. These are functions whose domains have finite supports
in the 2D or 3D space over which they are defined. An example of them is a pulse expansion
function in one dimension. We use h to describe the typical size of the subdomain. When
h is small, these functions are capable of modeling fields and sources that vary rapidly as a
function of space: in other words, they can model functions with high spectral frequencies.
Moreover, when h is small, many expansion functions are needed to model a function, and
hence the dimension N of the matrix representation increases.
The nature of a compact operator can be understood from its matrix representation when
h → 0 or N → ∞. When h → 0, the degrees of freedom of the function increase, and more
high-frequency eigenvectors can be formed. However, these higher-frequency eigenvectors will
have smaller eigenvalues because of the smoothing property of the compact operator. Hence,
the eigenvalue of the matrix representation, which are more numerous, become increasingly
smaller and cluster around 0. In other words, a compact operator annihilates high-frequency
eigenvectors. On the other hand, when its eigenvalues become increasingly larger when h → 0,
the operator is noncompact. A noncompact operator enhances the high frequency components
of a function it acts on.
Many integral operators eliminates the high-frequency components, and hence, they are
compact. Differential operators, on the other hand, amplify the high-frequency components,
and hence, they are noncompact.
2.7 Extension of Bra and Ket Notations

In Hilbert space, the use of bra and ket notation is insufficient for electromagnetics as two
kinds of inner products are used in electromagnetics, whereas in quantum mechanics, or
mathematics, only one kind of inner product is implied. However, the notations in linear
algebra are quite complete and unambiguous. We will adopt notations similar to the bra and
ket, such that they do not give rise to ambiguity when used in electromagnetics.
Because the inner product
hf, gi (2.28)
is analogous to f t · g, we will liken gi to be analogous to a column vector in linear algebra,
while hf to be analogous to a row vector (the transpose of a column vector) in linear algebra.
The comma between hf and gi denoting an inner product, is analogous to the dot in linear
algebra notation. We can think of hf as the transpose of f i. For a conjugate transpose, we
will use the notation hf ∗ . For the lack of a better name, we will call hf a braem, and gi a
ketem for use in electromagnetics.
Furthermore, we will use
L, gi (2.29)
to denote the action of the operator L on the vector gi. This is analogous to L · g in linear
algebra, and hence, the comma here is still analogous to a dot product. Hence, we can think
of L, gi in this extended notation here to be analogous to Lg in many mathematics literature.
Furthermore, for a projection of the operator onto two vectors, we denote it by
hf, L, gi. (2.30)
This is analogous to f t ·L·g in linear algebra where each comma is analogous to a dot product.
For the outer product between two vectors in Hilbert space, we will use the notation f ihg,
and this is analogous to fgt in linear algebra. It is to be noted that f ihg is not analogous to
f (x)g(x) but f (x)g(x0 ).
Dirac has used a vertical bar to denote an inner product. In his notation, a vector is
denoted by |f i, whereas hf | invariably is analogous to conjugate transpose in linear algebra.
He has used L|gi for (2.29), and hf |L|gi for (2.30) [15, 16]. Because of the implicit conjugate
transpose in the notations, they are not convenient for use in electromagnetics.
2.8 Orthogonal Basis versus Nonorthogonal Basis

When a basis set is orthogonal, they can be normalized such that they are orthonormal, so
that
hfn , fm i = δnm (2.31)
where δnm is the Kronecker delta function. If fn and fm are members of an orthonormal
basis set, then we can define an identity kernel in the subspace spanned by fn such that
N
X
I(x, x0 ) = fn (x)fn (x0 ) (2.32)
n=1
We define the subspace as Sn = span{fn (x), n = 1, . . . , N, a < x < b}. If g(x) ∈ Sn , then the
action (or operation) of I(x, x0 ) on g(x) is given by
Z b N
X
dx0 I(x, x0 )g(x0 ) = fn (x)hf, gi = g(x) (2.33)
a n=1
P
We can also think of n fn (x)fn (x0 ) as the outer product between two vectors indexed by x
and x0 . Hence, it can also be expressed as
N
X
I= fn ihfn (2.34)
n=1
We see that I behaves like an integral operator in this example.

When fn (x) are nonorthogonal, but complete in a subspace, we can expand a function
g(x) in that subspace by
XN
g(x) = an fn (x) (2.35)
n=1
To solve for an , we test the above by fm (x), m = 1, . . . , N to get

N
X
hfm , gi = an hfm , fn i , m = 1, . . . , N (2.36)
n=1
The above can be expressed as a matrix equation
g =F·a (2.37)
where
[g]n = hfn , gi, [F]mn = hfm , fn i, [a]n = an (2.38)
The matrix F is also known as the Grammian [17]. Hence, we can solve (2.37) to get
−1
a=F ·g (2.39)
we can write (2.35) as

−1
g(x) = f t (x) · a = f t (x) · F ·g (2.40)
where [f (x)]n = fn (x). Hence, we can express g = hf , gi in (2.38) and (2.40) becomes
−1
g(x) = f t (x) · F · hf , gi (2.41)
or
−1
gi = f t i · F · hf , gi (2.42)
We can identify an identity operator
−1
I = f ti · F · hf (2.43)
−1
When fn ’s are orthonormal, F = I, and the above becomes
I = f t i · hf (2.44)
Eq. (2.44) is an alternative way of writing (2.34).
2.9 Integration by Parts

One of the most frequently used mathematical tricks in electromagnetics is integration by
parts. It is based on the product rule of derivatives. Such a rule appears for scalar and vector
fields in different form [18].
∇ · (φA) = ∇φ · A + φ∇ · A (2.45)
∇ · (A × B) = B · ∇ × A − A · ∇ × B (2.46)
∇(φ1 φ2 ) = φ2 ∇φ1 + φ1 ∇φ2 (2.47)
∇ × (φA) = φ∇ × A − A × ∇φ (2.48)
It is from this rule, together with Gauss’ divergence theorem, that many integral identities
follow. One can integrate (2.45) over a volume V to obtain, after invoking Gauss’ theorem,
I Z Z
n̂ · (φA)dS = dV (∇φ · A) + dV (φ∇ · A) (2.49)
S V V
where S is the surface of V . If either φ or A, or both are of finite support, we can pick V to
be larger than the support of φ or A. Then the left-hand side of (2.49) is zero because φ or
A, or both are zero on S, and we have
Z Z
dV ∇φ · A = − dV (φ∇ · A) (2.50)
V V
Similarly, from (2.46), we have

I Z Z
n̂ · (A × B)dS = dV B · ∇ × A − dV A · ∇ × B (2.51)
S V V
If either A or B, or both are of finite support, then, we can pick V to be larger than the
support of A or B, and the left-hand side of (2.51) is zero, because either A or B, or both
are zero on S, and we have
Z Z
dV B · ∇ × A = dV A · ∇ × B (2.52)
V V
Eq. (2.47) is similar to Eq. (2.45) if we replace φ = φ1 and A = aφ2 where a is a constant
vector. Therefore, from (2.47),
I Z Z
n̂(φ1 φ2 )dS = dV φ1 ∇φ2 + dV φ2 ∇φ1 (2.53)
S V V
The above becomes Z Z

dV φ1 ∇φ2 = − dV φ2 ∇φ1 (2.54)
V V
if either φ1 or φ2 is of finite support.
To derive a formula from (2.48), we use the fact that
Z Z
dV ∇ × F = dS(n̂ × F) (2.55)
V S
This can be derived by applying Gauss’ divergence theorem to the function C = a × F

where a is an arbitrary constant, and noting that ∇ · C = ∇ · (a × F) = −a · ∇ × F and
n̂ · a × F = −a · n̂ × F. Using the above, we have
I Z Z
dS n̂ × φA = dV φ∇ × A − dV A × ∇φ (2.56)
V V
or Z Z
dV A × ∇φ = dV φ∇ × A (2.57)
V V
if either φ or A is of finite support.
The above are some of the most commonly used integration-by-parts formulas in electro-
magnetics.
2.10 Reciprocity Theorem—A New Look

The reciprocity theorem is attributed to Lorentz [19], but variations of it have been derived. A
simpler form for the Poisson’s equation has been stated by George Green [20]. The reciprocity
theorem can be restated from the perspective of a linear vector space: the linear operator
associated with the electromagnetic equations is a symmetric operator. For the statement
of the reciprocity theorem, the electromagnetic equations are augmented with a fictitious
magnetic current M to make them symmetrical. Namely, the first two equations are
∇ × E(r, ω) = iωB(r, ω) − M(r, ω), (2.58)

∇ × H(r, ω) = −iωD(r, ω) + J(r, ω). (2.59)
For anisotropic media where B = µ·H, and D = ²·E, we can write electromagnetic equations
as · ¸ · ¸ · ¸ · ¸ · ¸
0 ∇×I E ² 0 E J
· + iω · = . (2.60)
∇×I 0 H 0 −µ H −M
where ∇ × I · E = ∇ × E, and overbar over a boldface character implies that it is a second
rank tensor. The above can be more compactly rewritten as
· ¸ · ¸
E J
L = , (2.61)
H −M
where · ¸ · ¸
0 ∇×I ² 0
L= + iω . (2.62)
∇×I 0 0 −µ
We shall prove that the above operator is symmetric when ² and µ are symmetric tensors.
Many authors associate a negative sign with one of the ∇× operators, but that will make the
L operator nonsymmetric [4, 21, 22].
For the fields produced by two different groups of sources, we have
· ¸ · ¸
E1 J1
L = , (2.63)
H1 −M1
· ¸ · ¸
E2 J2
L = . (2.64)
H2 −M2
Multiplying (2.63) by [E2 , H2 ] and (2.64) by [E1 , H1 ], we have
¿ · ¸À ¿ · ¸À
E1 J1
[E2 , H2 ] , L = [E2 , H2 ] , = hE2 , J1 i − hH2 , M1 i, (2.65)
H1 −M1
¿ · ¸À ¿ · ¸À
E2 J2
[E1 , H1 ] , L = [E1 , H1 ] , = hE1 , J2 i − hH1 , M2 i, (2.66)
H2 −M2
where the inner product above is defined by integrating over a volume V bounded by a surface
S. Writing out the left-hand sides of (2.65) and (2.66) explicitly, we have
LHS(2.65) = hE2 , ∇ × H1 i + hH2 , ∇ × E1 i + iωhE2 , ² · E1 i − iωhH2 , µ · H1 i, (2.67)

LHS(2.66) = hE1 , ∇ × H2 i + hH1 , ∇ × E2 i + iωhE1 , ² · E2 i − iωhH1 , µ · H2 i. (2.68)

If ² and µ are symmetric tensors, then the last two terms of (2.67) and (2.68) are equal to
each other. We can apply integration by parts to (2.67) to get
hE2 , ∇ × H1 i + hH2 , ∇ × E1 i = hH1 , ∇ × E2 i + hE1 , ∇ × H2 i
I I
+ dS n̂ · (E2 × H1 ) − dS n̂ · (E1 × H2 ), (2.69)
S S
where S is the surface bounding V . Now using (2.69) in (2.67), and then subtracting (2.67)
and (2.68), we have
I I
LHS(2.65) − LHS(2.66) = dS n̂ · (E2 × H1 ) − dS n̂ · (E1 × H2 ), (2.70)
S S
The surface integral can be made to vanish if the surface S is a perfect electric conductor
(PEC), a perfect magnetic conductor (PMC), or a surface with a certain impedance boundary
condition (IBC), implying that the right-hand side of the above is zero.
Alternatively, if the sources J1 , M1 , J2 , M2 are of finite support, we can make the volume
infinitely large so as to take S to infinity. Using the fact that E and H satisfy the radiation
condition and degenerate to plane waves, the surface integrals in (2.69) cancel and hence
vanish.
Hence, we conclude that LHS(2.65) = LHS(2.66).2 This also implies that the L operator
defined in (2.61) is a symmetric operator when J and M are of finite support and the fields
they generate satisfy the radiation condition. Alternatively,
Hence, it follows from (2.65) and (2.66) that
hE2 , J1 i − hH2 , M1 i = hE1 , J2 i − hH1 , M2 i. (2.71)
The above is known as the reciprocity theorem. As can be seen, it is a consequence of the
symmetry (in a mathematical sense) of the electromagnetic equations, under the assumption
that µ and ² are symmetric tensors (not Hermitian!).
We can physically interpret hEi , Jj i as the electric field due to sources Ji and Mi “mea-
sured” by source Jj , and −hHi , Mj i as the magnetic field due to sources Ji and Mi “mea-
sured” by source Mj . This “measurement” is symmetric according to the reciprocity theorem
if the medium is also symmetric or reciprocal.
The proof can be generalized easily to bianisotropic media of finite extent. For this case,
the constitutive relation is
· ¸ · ¸ · ¸
D ² ξ E
= · . (2.72)
B ζ µ H
The corresponding modification to the L operator is
· ¸ · ¸
0 ∇×I ² ξ
L= + iω . (2.73)
∇×I 0 −ζ −µ
To retain the symmetry of the L, in addition to requiring symmetry in ² and µ, we require
t
ξ = −ζ . These are also the requirements for reciprocity in bianisotropic media [23].
2 We are associating [E1 , H1 ]t with g and [E2 , H2 ] with f in (2.19) to arrive at this.
2.10.1 Lorentz Reciprocity Theorem

The reciprocity theorem is sometimes written with the surface integrals from (2.69) aug-
menting the terms in (2.71). In this case, it is commonly known as the Lorentz reciprocity
theorem:
hE2 , J1 i − hH2 , M1 i = hE1 , J2 i − hH1 , M2 i

I I
+ dS n̂ · (E2 × H1 ) − dS n̂ · (E1 × H2 ). (2.74)
S S
The implication of the above is that the L operator is not a symmetric operator when the
integration of the inner product is taken over a finite volume, but one can augment the L
operator to make it symmetric when the inner product is defined over a finite volume.
The nonsymmetric part for finite volumes arises from the differential operator, which can
be symmetrized. To this end, we define the curl differential operator to be
· ¸
0 ∇×I
C= . (2.75)
∇×I 0
The above operator is symmetric if the inner product is defined over an infinite volume, and
if the field satisfies the radiation condition at infinity. To make the above operator symmetric
even when the inner product is defined over a finite volume, we augment C with an additional
term, defining a new curl differential operator to be
Ĉ = C + S, (2.76)
where S is an operator with the property that a surface integral will be induced when it is
sandwiched in the inner product definition. Namely,
¿ · ¸À I
E1 1
[E2 , H2 ] , S =− dS n̂ · (E2 × H1 − E1 × H2 ). (2.77)
H1 2 S
The S operator has the sifting property of the Dirac delta function, but in a three-dimensional
space. Hence, it converts a volume integral on the left-hand side to a surface integral. With
this definition,
¿ · ¸À
E1
[E2 , H2 ] , Ĉ = hE2 , ∇ × H1 i + hH2 , ∇ × E1 i
H1
I
1
− dS n̂ · (E2 × H1 − E1 × H2 ), (2.78)
2 S
¿ · ¸À
E2
[E1 , H1 ] , Ĉ = hE1 , ∇ × H2 i + hH1 , ∇ × E2 i
H2
I
1
− dS n̂ · (E1 × H2 − E2 × H1 ). (2.79)
2 S
Using (2.69), it can be shown that (2.78) and (2.79) are the same. Hence, Ĉ is a symmetric
operator even when the integral is taken over a finite volume.
The term S[E, H]t can be added to both sides of (2.61) to symmetrize the operator on the
left-hand side, and the Lorentz reciprocity theorem indicated in (2.74) can be derived.
The S operator can be thought of as an operator that will produce an equivalence surface
current on S when it operates on a field. When these sources are tested with another field, a
surface integral is produced. Away from the surface S, this operator has no effect. Physically,
one can think that the S operator places an equivalence current on the surface S so as to
generate a field external to S that cancels the original external field there. By the extinction
theorem [10], this equivalence surface current does not generate a field inside S. Consequently,
one needs only integrate over the finite volume V .
2.11 Energy Conservation Theorem—A New Look

The electromagnetic energy conservation theorem in the time domain is expressed using the
Poynting theorem, which was derived by Poynting [24] in 1884 and Heaviside in 1885 [25].
The energy conservation theorem in the Fourier space, making use of complex fields, was
probably first derived by Stratton [13].
We can look at the energy conservation theorem from a different perspective using our
understanding of vector spaces, operators, and their eigenvalues. We know that Hermitian
operators have real eigenvalues, and they are intimately related to the lossless nature of the
systems they describe, as is the case with a quantum system. We rewrite electromagnetic
equations into a form that produces a Hermitian system when the medium is lossless. To this
end, we rewrite (2.60) as
· ¸ · ¸ · ¸ · ¸ · ¸
0 −i∇ × I E ² 0 E J
· +ω · = −i . (2.80)
i∇ × I 0 H 0 µ H M
The above can be compactly rewritten as

· ¸ · ¸
E J
L = −i , (2.81)
H M
where · ¸ · ¸
0 −i∇ × I ² 0
L= +ω . (2.82)
i∇ × I 0 0 µ
When ² and µ are Hermitian, it can be easily shown, using the definition for self-adjointness,
that L is Hermitian under appropriate boundary conditions. The appropriate boundary
condition is obtained if the above equation is restricted to a resonant modes enclosed by a
surface S and the perfect-electric or perfect-magnetic boundary condition is imposed on the
wall of the resonant modes. In other words, the resonant modes must have no wall loss.
If the resonant modes is source free, (2.80) reduces to
· ¸ · ¸ · ¸ · ¸
0 −i∇ × I E ² 0 E
· = −ω · . (2.83)
i∇ × I 0 H 0 µ H
The above can be thought of as a generalized eigenvalue problem where −ω is the eigenvalue.
If ² and µ are Hermitian, and because the differential operator above is also Hermitian for
a resonant modes without wall loss, the above equation can only have real eigenvalues. Real
eigenvalues correspond to modes whose fields do not decay with time. This implies that the
resonant modes of the resonant modes do in fact correspond to those of a lossless cavity. We
shall see below that the Hermitian nature of ² and µ is in fact necessary for a lossless medium.
Multiplying (2.80) by [E∗ , H∗ ], and integrating over a volume V , we have
¿ · ¸À ¿ · ¸À
E J
[E∗ , H∗ ] , L = [E∗ , H∗ ] , −i = −i (hE∗ , Ji + hH∗ , Mi) . (2.84)
H M
The above inner product is defined by integrating over a volume V . Writing out the left-hand
sides of (2.84) explicitly, we have
LHS(2.84) = −ihE∗ , ∇ × Hi + ihH∗ , ∇ × Ei + ωhE∗ , ² · Ei + ωhH∗ , µ · Hi. (2.85)
Using integration by parts, it can be shown that

I
−ihE∗ , ∇ × Hi = −ih∇ × E∗ , Hi + i dS n̂ · (E∗ × H)
S
I
= −ω hµ · H , Hi + i hM , Hi + i dS n̂ · (E∗ × H),
∗ ∗ ∗
S
(2.86)
where the normal n̂ points outward from the surface S. Similarly, it can be shown that
I
ihH , ∇ × Ei = −ω h² · E , Ei + i hJ , Ei + i dS n̂ · (E × H∗ ).
∗ ∗ ∗ ∗
(2.87)
S
Consequently, using (2.86) and (2.87) in (2.85), we have

·I ¸
LHS(2.84) = 2i<e dS n̂ · (E∗ × H)
S
+2iω=m{hE∗ , ² · Ei + hH∗ , µ · Hi}
+ihM∗ , Hi + ihJ∗ , Ei. (2.88)
Using Eq. (2.88), Eq. (2.84) can be rewritten as

·I ¸
<e dS n̂ · (E × H∗ ) = −ω=m{hE∗ , ² · Ei + hH∗ , µ · Hi}
S
−<e{hE, J∗ i + hH, M∗ i}. (2.89)
The above is reminiscent of the complex Poynting theorem, except that it makes a statement
only on the real part of the complex power E × H∗ . It makes no statement on the imaginary
part of the complex power, which is expressed in complex Poynting theorem. This is because
the differential operator in (2.80) is Hermitian, and it will produce a real quantity under the
inner product we have used here.
In the above, the real part of a complex power corresponds to time-average power. If the
region enclosed by S is source-free, so that J and M are zero, and if ² and µ are Hermitian,
so that hE∗ , ² · Ei +hH∗ , µ · Hi is pure real, then the right-hand side of (2.89) is zero, and then
there is no time-average power flowing into S, implying that the region is lossless. Hence, the
Hermitian nature of ² and µ is necessary for a lossless medium.
For a bianisotropic medium, Eq. (2.82) becomes
· ¸ · ¸
0 −i∇ × I ² ξ
L= +ω . (2.90)
i∇ × I 0 ζ µ
†
The conditions for a lossless medium are that ² and µ are Hermitian and ξ = ζ .
To derive a statement on the imaginary part of the complex power, we replace the differ-
ential operator in (2.80) with a skew Hermitian one (we define this to be A† = −A). In this
case, the electromagnetic equations are written as
· ¸ · ¸ · ¸ · ¸ · ¸
0 i∇ × I E ² 0 E J
· −ω · =i . (2.91)
i∇ × I 0 H 0 −µ H −M
Upon multiplying the above by [E∗ , H∗ ], and integrating over a volume V , we have a new
equation that becomes
ihE∗ , ∇ × Hi + ihH∗ , ∇ × Ei − ωhE∗ , ² · Ei + ωhH∗ , µ · Hi = ihE∗ , Ji − ihH∗ , Mi. (2.92)
Upon simplifying the above using (2.86) and (2.87), we have

·I ¸
=m dS n̂ · (E × H∗ ) = ω<e{hH∗ , µ · Hi − hE∗ , ² · Ei}
S
+=m{hH, M∗ i − hE, J∗ i}. (2.93)
The above is the energy conservation theorem regarding the imaginary part of the complex
power, which is the reactive power. The reactive power is proportional to the difference
between the stored energy in the magnetic field and the electric field, which is indicated by
the first term on the right-hand side. The second term on the right-hand side corresponds to
the reactive power absorbed or exuded by the sources M and J. The above derivation can
also be generalized to bianisotropic media.
In a time-harmonic sinusoidal system, reactive power is the power that one pumps into a
system in one cycle, and then withdraws from it in the next cycle. This happens, for instance,
when a capacitor is hooked to a sinusoidal generator source, where in one cycle the capacitor
is charged (and hence taking power from the generator), while in another cycle the capacitor
is discharged (hence returning power back to the generator).
It is interesting to note that the energy conservation theorem can be derived using our
understanding of Hermitian operators. Hermitian operators have the physical meaning of
“energy” conserving operators. To derive energy conservation statements, we are also mo-
tivated to construct an operator equation whose operator becomes Hermitian in the lossless
case.
It is to be noted that in wave physics, a lossless medium does not necessarily imply a
Hermitian operator. It is only in the case of a closed cavity with lossless media as well as no
wall loss that the resulting system is Hermitian. Hence, in (2.83), an appropriate boundary
condition corresponding to no wall loss needs to be imposed before the operator is Hermitian.
For instance, if the radiation condition is imposed on the wave at infinity, the resultant
operator is non-Hermitian. Hence, when the free-space Green’s function, exp(ik|r − r0 |)/|r −
r0 |), forms the kernel of an integral operator, the operator is symmetric, but not Hermitian.
A radiating system, even though it is embedded in a lossless medium of infinite extent, is not
lossless, because it leaks energy to infinity.
2.12 Conclusions
Although most works do not relate the reciprocity theorem and energy conservation theorem
to the properties of linear operators in linear vector spaces, we show that, in fact, they are
intimately related. It will be better to express these theorems from a modern perspective
using linear vector space concepts, which are used frequently in computational sciences. The
reciprocity theorem is related to the symmetry of the electromagnetic operators, and the loss-
less nature of a medium is related to the Hermitian nature of the associated electromagnetic
operators. Moreover, one can define a skew-Hermitian electromagnetic operator from which
the energy conservation of the reactive component of complex power can be derived.
Bibliography
[1] J. A. Kong, Theory of Electromagnetic Waves, New York: Wiley-Interscience, 1975.
[2] W. C. Chew, “A new look at reciprocity and energy conservation theorems in electro-
magnetics,” Research Report No.: CCEML 11-06, Oct. 19, 2006. also in IEEE Trans.
Antennas Propag., vol. 56, no. 4, pp. 970–975, April 2008.
[3] M. C. Reed and B. Simon, Methods of Modern Mathematical Physics, Vol. III: Scattering
Theory, New York: Academic Press, 1979. Section XI.10.
[4] L. B. Felsen and N. Marcuvitz, Radiation and Scattering of Waves, Englewood Cliffs, NJ:
Prentice Hall, 1973. Sections 1.4c and 1.5b.
[5] G. Hanson and A. Yakovlev, Operator Theory for Electromagnetics: An Introduction, New
York: Springer-Verlag, 2002.
[6] R. E. Collin, Field Theory of Guided Waves, Second edition, New York: Wiley-IEEE
Press, 1990.
[7] L. Cairo and T. Kahan, Variational Techniques in Electromagnetism, New York: Gordon
and Breach, 1965.
[8] R. Harrington, Field Computation by Moment Method, Malabar, FL: Krieger, 1983. (First
printing 1968.)
[9] C. H. Chen and C. D. Lien, “The variational formulation for non-self-adjoint electromag-
netic problems.” IEEE Trans. Microwave Theory Tech., vol. 28, pp. 878–886, 1980.
[10] W. C. Chew Waves and Fields in Inhomogeneous Media, New York: Van Nostrand
Reinhold, 1990. Reprinted by Piscataway, NJ: IEEE Press, 1995.
[11] V. H. Rumsey, “Reaction concept in electromagnetic theory,” Phys. Rev., vol. 94, pp.
1483–1491 (1954), Errata, vol. 95, pp. 1705, 1954.
[12] M. Artin, Algebra, Englewood Cliffs, NJ: Prentice Hall, 1991.
[13] J. A. Stratton Electromagnetic Theory, New York: McGraw-Hill Book Company Inc.,
1941.
[14] I. Stakgold, Boundary Value Problems of Mathematical Physics, vol. I, London: Macmil-
lan Co., 1967.
[15] E. Merzbacher, Quantum Mechanics, New York: John Wiley & Sons, 1970.
[16] J. J. Sakurai, Modern Quantum Mechanics, New York: Wiley-Interscience, 1985.
[17] T. K. Moon and W. C. Stirling, Mathematical Methods and Algorithms for Signal Pro-
cessing, Englewood Cliffs, NJ: Prentice Hall, 2000.
[18] C. T. Tai, Dyadic Green’s Functions in Electromagnetic Theory, second edition, Piscat-
away, NJ: IEEE Press, 1993.
[19] H. A. Lorentz, “The theorem of Poynting concerning the energy in the electromagnetic
field and two general propositions concerning the propagation of light,” Amsterdammer
Akad. Weten., vol. 4, pp. 176, 1896.
[20] N. M. Ferrers (Editor), Mathematical Papers of George Green, New York: Chelsea Publ.,
1997.
[21] C. Altman and K. Suchy, Reciprocity, Spatial Mapping, and Time Reversal in Electro-
magnetics, Boston: Kluwer Academic Pub., 1991.
[22] I. V. Lindell, A. H. Sihvola, and K. Suchy, “Six-vector formalism in electromagnetics

of bi-anisotropic media,” J. Electromagn. Waves and Appl., vol. 9, no. 7/8, pp. 887–903,
1995.
[23] J. A. Kong, “Theorems of bianisotropic media,” Proc. IEEE, vol. 60, pp. 1036–1046,
1972.
[24] J. H. Poynting, “On the transfer of energy in the electromagnetic field,” Philos. Trans.
R. Soc. London, vol. 175, Part II, pp. 343–361, 1884.
[25] O. Heaviside, “Electromagnetic induction and its propagation,” The Electrician, 1885.
43
CHAPTER 3
Introduction to
Integral Equations
3.1 Introduction
This chapter serves to give a pedestrian introduction of integral equations to a newcomer in
computational electromagnetics (CEM) who will learn most of surface integral equations from
this chapter. The more detailed subjects will be dealt with in later chapters. The derivation
of various theorems, principles, and operators will be done in the frequency domain.
In the previous chapters, we consider frequency domain solutions to be the solutions of
the electromagnetic equations in the Fourier space. One can also consider frequency domain
solutions when the fields or signals are all purely time harmonic. All time-harmonic vector
fields can be regarded as
h i
F(r, t) =<e F̃(r, ω)e−iωt = F̃r (r, ω) cos(ωt) + F̃i (r, ω) sin(ωt) (3.1)
where F̃(r, ω) = F̃r (r, ω) + iF̃i (r, ω) is a complex field which is independent of time. The
complex function F̃(r, ω) is also known as the phasor representation of the time harmonic
field F(r, t). It is related to the Fourier transform of the time harmonic field. By substituting
the above into the electromagnetic equations, they can be simplified just as Fourier transform
is used to simplify these equations as was done in the previous chapters.
3.2 The Dyadic Green’s Function

Before we can derive integral equations for vector electromagnetic field, it is most convenient
to express these integral equations in terms of the dyadic Green’s function [1–4]. The dyadic
Green’s function is the point source response to the vector wave equation,
∇ × ∇ × E(r) − k 2 E(r) = iωµJ(r). (3.2)
4
derivable from the electromagnetic equations for a homogeneous medium where k 2 = ω 2 µ² is

a constant.
By letting Z
E(r) = iωµ dr0 J(r0 ) · G(r0 , r), (3.3)
V
0
where G(r , r) is the dyadic Green’s function, and using the above in (3.2), assuming that
the current J(r0 ) is arbitrary, one can show that the homogeneous medium dyadic Green’s
function is the solution to the equation
∇ × ∇ × G(r, r0 ) − k 2 G(r, r0 ) = I δ(r − r0 ). (3.4)
Its solution is given by

µ ¶ ik|r−r0 |
∇∇ e
G(r, r0 ) = I+ 2 . (3.5)
k 4π |r − r0 |
Physically, a dyadic Green’s function, when operating on an electric current, yields the electric
field produced by the current. By the principle of linear superposition, it can be used to
calculate the field due to a distributed current source via (3.3) As a consequence of the
reciprocity theorem, one can show that [7]
t
G (r0 , r) = G(r, r0 ). (3.6)
Then, taking its transpose, (3.3) becomes

Z
E(r) = iωµ dr0 G(r, r0 ) · J(r0 ). (3.7)
V
In the above, r is referred to as the field point (or observation point), while r0 is the source
point.
Because this dyadic Green’s function generates electric field from electric current, it is also
known as the electric dyadic Green’s function. Because of the double derivative on the scalar
Green’s function, the dyadic Green’s function is very singular or hypersingular, and hence, its
numerical evaluation via integration often poses a problem when the field point is close to the
source point. However, in CEM, the practice has been to avoid the evaluation of the dyadic
Green’s function in its above form. The singularity is reduced by various mathematical tricks
such as integration by parts (see Chapter 2).
3.3 Equivalence Principle and Extinction Theorem

The building blocks of surface integral equations, which appear in different forms, are the
equivalence principle and the extinction theorem. The equivalence principle is similar to
Huygens’ principle, but its mathematical form is not attributable Huygens. The mathe-
matical form for scalar wave was derived by Kirchhoff and Green [5]. The modern vector
electromagnetic form was contributed by Stratton and Chu [6], and also Franz [8]. To derive
INTRODUCTION TO INTEGRAL EQUATIONS 45
S inf
J2
V2
V1 S
J1
Figure 3.1: Derivation of equivalence principle and extinction theorem.
the equivalence principle and extinction theorem, we start with the vector wave equation
(3.2) where k2 is assumed to be constant everywhere. Figure 3.1 illustrates the setup for
deriving the equivalence principle and the extinction theorem. Here, (3.2), J(r) represents
the currents J1 (r) and J2 (r), where J1 (r) is in V1 and J2 (r) is in V2 . Here, V1 is bounded by
surface S and Sinf while V2 is bounded by S.
Premultiplying (3.4) by E(r) and postmultiplying (3.2) by G(r, r0 ), and then subtracting
the resultant equations, we have
E(r) · ∇ × ∇ × G(r, r0 ) − ∇ × ∇ × E(r) · G(r, r0 ) =δ(r − r0 )E(r)
− iωµJ(r) · G(r, r0 )
(3.8)
and integrating the difference over a volume V1 , we have
E(r0 )−E1 (r0 )

Z
£ ¤
= dV E(r) · ∇ × ∇ × G(r, r0 ) − ∇ × ∇ × E(r) · G(r, r0 ) .
V1
(3.9)
where Z
E(r0 ) = dr δ(r − r0 )E(r), (3.10)
V1

Z
0
E1 (r ) = iωµ dr J1 (r) · G(r, r0 ) (3.11)
V1
The above is field produced by J1 in V1 . Because J2 is not in V1 , it does not contribute to

the above volume integral.
Using the identity that
£ ¤
−∇ · E(r) × ∇ × G(r, r0 ) + ∇ × E(r) × G(r, r0 )
= E(r) · ∇ × ∇ × G(r, r0 ) − ∇ × ∇ × E(r) · G(r, r0 ), (3.12)
Eq. (3.9), with the help of Gauss’ divergence theorem, can be rewritten as
E(r0 )−E1 (r0 )

I
£ ¤
= dS n̂ · E(r) × ∇ × G(r, r0 ) + ∇ × E(r) × G(r, r0 )
S+Sinf
I
£ ¤
= dS n̂ × E(r) · ∇ × G(r, r0 ) + iωµ n̂ × H(r) · G(r, r0 ) .
S+Sinf
(3.13)
where the normal n̂ is a normal on the surface S that points outward. The term E(r0 ) is
zero if r0 ∈
/ V1 because of the sifting property of delta function (one can appreciate this only
by going through the detail of the derivation that arrives at the above). Hence, the above
equation can be written as
¾ I
r0 ∈ V1 , E(r0 ) 0
£
0 =E 1 (r ) + dS n̂ × E(r) · ∇ × G(r, r0 )
r ∈/ V1 , 0
S+Sinf
¤
+ iωµ n̂ × H(r) · G(r, r0 ) . (3.14)
By making use of the far field or radiation condition, the integral over Sinf is shown to vanish.
It can be shown that a finite source will generate a field that decays as 1/r as r → ∞,
where r is the source point to field point separation. Hence, E, H, G, and ∇ × G individually
will decay as 1/r. Therefore, each of the terms on the right-hand side of (3.14) decays as
1/r2 . But this reason alone is not sufficient to ignore the contribution from the integral over
Sinf , as the surface area of Sinf grows as r2 when r → ∞. However, it can be shown that
the two terms on the right-hand side of (3.14) cancel each other to leading order, causing
the total integrand to decay as 1/r3 . This is the correct reason that the integral over Sinf
vanishes when r → ∞.
Then by swapping r and r0 , the above finally becomes
¾ I
r ∈ V1 , E(r) £
=E1 (r) + dS 0 n̂0 × E(r0 ) · ∇0 × G(r0 , r)
r ∈ V2 , 0
S
¤
+ iωµ n̂0 × H(r0 ) · G(r0 , r) . (3.15)
Taking the transpose of the above equation, and making use of the following (which can be
proven by using the reciprocity theorem [3, 7])
£ ¤t £ ¤t
∇ × G(r, r0 ) = ∇0 × G(r0 , r) , G(r, r0 ) = G(r0 , r) (3.16)
we have
¾ I
r ∈ V1 , E(r) £
=E1 (r) + dS 0 ∇ × G(r, r0 ) · n̂0 × E(r0 )
r ∈ V2 , 0
S
¤
+ iωµ G(r, r0 ) · n̂0 × H(r0 ) . (3.17)
In the above, we can let Ms (r0 ) = −n̂0 × E(r0 ) and Js (r0 ) = n̂0 × H(r0 ), which can be thought
of as equivalence surface magnetic and electric currents respectively impressed on the surface
S, and the above then becomes
¾ I
r ∈ V1 , E(r) £
=E1 (r) − dS 0 ∇ × G(r, r0 ) · Ms (r0 )
r ∈ V2 , 0
S
¤
− iωµ G(r, r0 ) · Js (r0 ) . (3.18)
The physical meaning of the above is as follows: The total field E in V1 is generated by the
source J1 inside V1 as well as source outside V1 such as J2 in our example, and it can be
decomposed into the field E1 on the right-hand side which is generated by J1 , plus field gener-
ated by equivalence surface currents Ms and Js impressed on the surface S. The equivalence
surface currents generate the same field produced by J2 or any source in V2 enclosed by S.
Moreover, the total field is zero if the field point is placed in V2 : this is reflected by the
lower part of the left-hand side of (3.18). This implies that the equivalence surface currents
produce a field that exactly cancels E1 in V2 . This is known as the extinction theorem.
It may be instructive to go through different cases to see how the extinction theorem
manifests itself.
Case (a): J1 = 0
In this case, E1 = 0 and E is entirely due to J2 , and hence, E = E2 . Consequently, (3.18)
becomes
¾ I
r ∈ V1 , E2 (r) £
= − dS 0 ∇ × G(r, r0 ) · M2s (r0 )
r ∈ V2 , 0
S
¤
− iωµ G(r, r0 ) · J2s (r0 ) . (3.19)
From the above, one notices that the equivalence surface currents produce the correct field
outside the volume V2 , but produces zero field inside V2 . In other words, if we think of the
equivalence surface currents as propagating information away from the source J2 , it correctly
does so for field points outside V2 , but extincts the field for points inside V2 that contains the
source J2 .
Case (b): J2 = 0
In this case, E = E1 , and (3.18) after some rearrangement becomes
¾ I
r ∈ V1 , 0 £
= dS 0 ∇ × G(r, r0 ) · M1s (r0 )
r ∈ V2 , E1 (r)
S
¤
− iωµ G(r, r0 ) · J1s (r0 ) . (3.20)
From the above, we see that the equivalence currents generate no field inside V1 , the region
that contains the source J1 , whereas it generates the correct field (save for a minus sign) in
region V2 that does not contain the source. The minus sign can be absorbed by redefining
the surface normals. From the above, one notes that the physical interpretation of (3.20) is
actually no different from that of (3.19).
As a final note, in deriving Eq. (3.9), if we integrate over V2 instead, a similar equation
for the above for V2 can be derived. By invoking duality, similar equations for the magnetic
field can be derived. The above is the mathematical statement of equivalence principles in
electromagnetics as espoused by Harrington [9].
3.4 Electric Field Integral Equation—A Simple Physical

Description
To simply see how an integral equation can arise physically, let us consider a scattering prob-
lem involving a metallic scatterer [perfect electric conductor (PEC)]. Consider an incident
electric field from an electromagnetic wave that impinges on the scatterer. The first phe-
nomenon that happens is that electric current will be induced on the scatterer. The current
produces an electric field that exactly cancels the incident electric field inside the scatterer,
as no electric field should exist inside.
In addition to canceling the incident electric field, the current also generates an electro-
magnetic field outside the scatterer yielding a scattered field. The currents have to cooperate
with each other to produce a field that cancels the incident field inside the scatterer. To coop-
erate, the currents communicate with each other via the Green’s function. Their cooperation
with each other can be mathematically described via the integral equation
Z
−Einc (r) =iωµ G(r, r0 ) · Js (r0 )dr0 , r ∈ S − (3.21)
S
where G(r, r0 ) is the dyadic Green’s function defined in (3.5) that produces an electric field
due to a point electric current source, Js (r0 ) is the electric surface current on the surface
of the scatterer, Einc (r) is the incident electric field, and S − represents the surface of the
scatterer that is just inside the surface S. In the above, the unknown is the surface current
Js (r0 ) on the surface of the scatterer.
As n̂ × E is zero on the surface of the PEC, Ms = 0. By assuming that E1 = Einc ,
we arrive at the above integral equation. The above equation has three vector components,
but Js , a surface current, has only two vector components. Hence, a vector equation with
two components is sufficient to solve for Js . Consequently, we can reduce (3.21) to have two
vector components by taking its tangential components, namely,
Z
−n̂ × Einc (r) =iωµn̂ × G(r, r0 ) · Js (r0 )dr0 , r ∈ S (3.22)
S
As tangential E is continuous across a current sheet, the above integral equation need only
be applied on S. This integral equation is also known as the electric field integral equation
(EFIE) [10–14].
3.4.1 EFIE—A Formal Derivation
Esca Jinc
S S+ Einc
PEC
V2
V1
Figure 3.2: Scattering by a PEC (perfect electric conductor) object.
The above integral equation can also be derived more rigorously by invoking the extinction
theorem in the lower part of Eq. (3.15). Figure 3.2 shows the scattering by a PEC. We assume
that Jinc is a current that generates an incident field Einc that impinges on a PEC object.
Electric current will be induced on the PEC object, which in turn radiates to generate the
scattered field. We can imagine a surface S + that is just large enough to contain S, and apply
the equivalence principle and extinction theorem to it.
When equivalence currents are placed on S + , they produce the total field outside S + and
zero field inside S − . Hence, the PEC object can be removed without disturbing the field
outside. The equivalence currents are given by
Jeq = n̂ × H, Meq = E × n̂ on S+ (3.23)
where E and H represent total field. But as E × n̂ = 0 on S + , only Jeq is needed to represent
the field outside. Hence, the equivalence principle and extinction theorem becomes
¾ Z
r ∈ V1 , E(r)
= Einc (r) + iωµ dS 0 G(r, r0 ) · Jeq (r0 ) (3.24)
r ∈ V2 , 0 S+
When we apply the extinction theorem part of the above equation to S, we obtain the integral
equation of scattering, viz.,
Z
r ∈ S, 0 = Einc (r) + iωµ dS 0 G(r, r0 ) · Jeq (r0 ) (3.25)
S+
The magnetic field version of (3.24) can be derived by taking the curl of (3.24), namely,
¾ Z
r ∈ V1 , H(r)
= Hinc (r) + ∇ × dS 0 G(r, r0 ) · Jeq (r0 ) (3.26)
r ∈ V2 , 0 S+
The corresponding integral equation of scattering is

Z
r ∈ S, 0 = Hinc (r) + ∇ × dS 0 G(r, r0 ) · Jeq (r0 ) (3.27)
s+
In the above, Jeq is practically the same as Js because S + and S are infinitesimally close
together, and the tangential H on both these surfaces are practically equal to each other.
For a resonant modes whose surface is S, the requirement that n̂ × E = 0 on the cavity
wall will ensure null solution inside the resonant modes except at the cavity resonance. Hence,
away from cavity resonances, it is sufficient to include the tangential component of (3.25) and
(3.27) to get Z
r ∈ S, 0 = n̂ × Einc (r) + iωµn̂ × dS 0 G(r, r0 ) · Jeq (r0 ) (3.28)
S
Z
r ∈ S, 0 = n̂ × Hinc (r) + n̂ × ∇ × dS 0 G(r, r0 ) · Jeq (r0 ) (3.29)
S+
Eq. (3.28) is known as the electric field integral equation (EFIE) for a PEC, whereas Eq.
(3.29) is known as the magnetic field integral equation (MFIE) for a PEC. They break down
at the resonant frequency of the resonant modes enclosed by the PEC surface S, because at
the resonance frequency, the extinction of the tangential components of the E and H fields
does not guarantee the extinction of the field inside the scatterer. This point will be further
elaborated in Section 3.8.
A distinction between EFIE and MFIE is that the current on the surface of the PEC gives
rise to a discontinuous H field across the surface, but a continuous E field across the same
surface. Hence, the EFIE (3.28) holds true irrespective of whether we impose r to be on S,
S + , or S − . However, for the MFIE, (3.29) holds true because of the extinction theorem, and
it is important that we impose r on S such that S is contained in S + .
In the literature, it is a common practice to define the L operator such that
Z
E(r) = L(r, r ) · J(r ) = iωµ G(r, r0 ) · J(r0 )dr0
0 0
(3.30)
and K operator such that

Z
H(r) = K(r, r0 ) · J(r0 ) = ∇ × G(r, r0 ) · J(r0 )dr0 (3.31)
Z
= ∇g(r, r0 ) × J(r0 )dr0
In the above, one uses ∇ × ∇g(r, r0 ) = 0 to simplify the K operator, and integration is implied
over repeated variable r0 .
3.5 Understanding the Method of Moments—A Simple

Example
Method of moments (MOM) [10] is a way of solving an integral equation by converting it
into a matrix equation. In other words, it is a way of finding the matrix representation of an
integral operator (see Chapter 2), which can be quite complex. In the previous chapter, we
have discussed this in an abstract manner, but in this section, we will give a concrete example
as an illustration. We will illustrate MOM with a simple PEC scatterer.
MOM finds the matrix representation of an integral operator by first expanding and
approximating the unknown current in terms of a finite set of expansion functions:
N
X
J(r) = In Jn (r) (3.32)
n=1
In the above, Jn (r) is a known as expansion function or a basis function, while In ’s are
unknowns yet to be sought. So the search for the unknown current Js has been transferred
to the search for the finite set of unknown numbers denoted by In ’s.
Substituting the above into the integral equation, we have
XN Z
−n̂ × Einc (r) =iωµ In n̂ × G(r, r0 ) · Jn (r0 )dr0 , r ∈ S (3.33)
n=1 s
The above is not quite a matrix equation yet. It is an equation that depends on r. To remove
the r dependence, we dot multiply the above by −n̂ × Jm (r), m = 1, · · · , N , and integrate
the resultant dot product over the surface of the scatterer. By so doing, we have, after using
that n̂ × Jm (r) · n̂ × Einc (r) = −Jm (r) · Einc (r)
Z X N Z Z
− Jm (r) · Einc (r)dr =iωµ In drJm (r) · G(r, r0 ) · Jn (r0 )dr0 ,
s n=1 s s
m = 1, ..., N (3.34)
Now the left-hand side is only dependent on m and the double integral on the right-hand side
is only dependent on m, n. The above procedure is known as testing the equation with testing
functions. Thus the above now is a set of linear algebraic equation that can be written more
concisely as
N
X
Vm = Zmn In , m = 1, ..., N (3.35)
n=1
or more compactly in vector notation

V =Z · I (3.36)
where V is a column vector that contains Vn as its elements, Z is a matrix that contains Zmn
as its elements, and I is a column vector that contains In . More explicitly,
Zmn =iωµhJm , G, Jn i (3.37)
Vm = − hJm , Einc i (3.38)
where we have used the short hand

Z Z
hJm , G, Jn i = drJm (r) · G(r, r0 ) · Jn (r0 )dr0 (3.39)
s s
Z
hJm , Einc i = Jm (r) · Einc (r)dr (3.40)
s
The above inner product is the reaction inner product discussed in Chapter 2.
The above method for converting an integral equation into a matrix equation is alter-
natively known as the Galerkin’s method. In the above derivation, the testing functions
need not be from the same basis set as the expansion functions. However, if the operator is
symmetric, which is true of the above, choosing the testing functions to be the same as the
expansion functions, assuming the reaction inner product, gives rise to a symmetric matrix
system, reducing the storage requirements of the system.
3.6 Choice of Expansion Function

In general, the testing function in MOM can be different from the expansion function. Because
in the above, we have chosen them to be the same, we will talk about the choice of expansion
function here. One popular choice of expansion function is the Rao-Wilton-Glisson (RWG)
function [15], which is the workhorse of most MOM computations. To use this expansion
function, the surface of a scatterer has to be discretized into small elements each of which is
a triangle.
Figure 3.3: A geometrical discretization of a parabola with triangular patches.
Figure 3.3 shows a geometrical discretization (also called tessellation, meshing, gridding)
of an object with triangular patches. Triangular patches are known as simplices in a two
dimensional manifold, and all two-dimensional manifolds can be tessellated with triangles [16].
Hence, they are very versatile in describing surfaces of geometrical objects. For metallic
scatterers, only surface currents are induced on the scatterer, and hence, we need to only
mesh the surface of the scatterer.
After the object is meshed, the RWG function can be described. The RWG function
straddles two adjacent triangles, and hence, its description requires two contiguous triangles
as shown in the Figure 3.4. The expression for the expansion function describing the current
on the triangle is (
`n
s 2A±
ρ±
n (r), r ∈ Tn±
Λn (r) = n (3.41)
0, otherwise
where ± is used to denote the respective triangles, `n is the edge length of the contiguous
edge between the two triangles, A± ±
n is the area of the respective triangles, and ±ρn (r) is the
vector from the point r to the apex of the respective triangles, and Tn± is the support of the
respective triangles.
Figure 3.4: The geometry description and the current flowing on a Rao-Wilton-Glisson (RWG)
function.
As the current function Jn (r) is a vector function, the RWG function is also a vector
function. A current that flows from one triangle to another triangle via the contiguous edge
represents the RWG function. Hence, there is one RWG function for every common edge
of two contiguous triangles. Consequently, only inner edges of a surface require an RWG
function.
One thing is clear—for the RWG function to reside on the surface of the object, the object
must be meshed with good triangles. The triangle mesh shown in Figure 3.5 is considered a
bad mesh, because more than two triangles are sharing an edge, and hence, an RWG function
cannot be described on such a mesh.
The RWG function is vastly popular for a simple reason: an arbitrary surface can be
approximated fairly well with a triangulated surface. This is because a triangle is a simplex
for a 2D manifold (while a line segment is a simplex for a line, and a tetrahedron is a simplex
for a volume). If a surface is not approximated well, it can be further improved by mesh
refinement. This is not the case for a surface that is described by a quad patch, where
frustration can occur. Another reason for its popularity is its simplicity: it is of the lowest
order, is divergence conforming function1 and is defined on a pair of triangles. Its divergence
conforming property makes its representation of the current physical. One may argue that a
1A divergence conforming function is one whose divergence is finite—more of it in the next Chapter.
Figure 3.5: Defective meshes happen when an RWG function cannot be defined on such
meshes.
rooftop function defined over two quad patches has similar properties, but it does not have
the versatility of triangular patches.
If one views the center of a triangular patch as a node in a graph, then the RWG functions
allow the flow of current from one node to another, and each node is connected to three other
nodes in the graph (see Figure 3.6). This graph forms a circuit network on which current
can flow. Hence, RWG functions replace a current on a surface approximately by a circuit
network where each node is connected to every other nodes (except for the boundary nodes).
Hence, a smooth-flowing current is replaced by a somewhat tortuous current flow which is
not in a smooth line. Because of the nature of electromagnetic physics, when the scale of this
“circuitous” flow is happening at a lengthscale much smaller than wavelength, this flow can
still capture the physics of the current flow. The graininess of this current flow is due to the
low-order nature of the RWG basis function. One minimizes the effect of this graininess by
making the patch size smaller, or alternatively, one can go to higher-order basis, albeit with
more work.
Figure 3.6: The current in a mesh flows from triangle center to triangle center (left). The
centers of the triangles can be connected to form a network on which the current flows (right).
3.7 Closed Surface versus Open Surface

When an object is described by a surface, there are two classes of surfaces: closed surfaces, or
open surfaces. A closed surface is a surface that encloses a nonzero volume (see Figure 3.7).
A spherical surface is an example of a closed surface, so is a doughnut surface, or a doughnut
with two holes. All topological equivalence of a sphere is also a closed surface.
An open surface is like a thin disk: it is a surface that encloses no volume. Hence, sharp
or zero thickness edges occur in open surfaces (see Figure 3.3). An annular disk, or a disk
with two holes is still an open surface. All topological equivalence of these thin surfaces are
also open surfaces.
A thin sheet object may be thought of as a limiting case of an open surface object whose
volume has collapsed to zero. In this case, the thin sheet object has two original surfaces that
have collapsed to one, and these two surfaces coincide in space.
Figure 3.7: Example of closed surfaces: (left) An ogive. (right) A flying saucer.
3.7.1 EFIE for Open Surfaces

The EFIE introduced in Eq. (3.22) is valid for both open and closed surface. But when it
is applied to an open surface, one has to treat both sides of an open surface as one surface,
and the surface current is the sum of the surface currents on both sides of the surface. This
is because tangential E field is continuous across a current-carrying open surface, and hence,
Eq. (3.22) is valid on both sides of an open surface.
If one collapses a closed surface into a open surface, for example, by reducing a flying
saucer into one with zero thickness, in the case of EFIE, the equation before the collapse
has been imposed on both the upper and lower surfaces of the flying saucer, but after the
collapse, it is only imposed on the collapsed surface which is one surface. Hence, the number
of independent equations (one can imagine the EFIE has been reduced to a set of linear
algebraic equations by using MOM) has seemingly decreased as now we have more unknowns
than equations. However, one can sum the currents on the top and bottom surfaces into one
current, making the number of equations equal to the number of unknowns.
3.7.2 MFIE and More

The MFIE is obtained by the invocation of the extinction theorem. One can rewrite the
MFIE equation (3.29) with the simplification in (3.31)
Z
−n̂ × Hinc (r) =n̂ × ∇g(r, r0 ) × Js (r0 )dr0 , r ∈ S − (3.42)
S
±
since S and S are infinitesimally close to each other. MFIE can only be written for closed
surfaces because it is based on the extinction theorem which is valid only for closed surfaces.
Hence, the field points in the above equation cannot be taken outside the closed surface S.
When a closed surface collapses to a thin open surface as in the case of a flying saucer, the
field points become redundant, and the number of equations is smaller than the number of
unknowns. Alternatively, one can argue that the extinction theorem cannot be invoked for
an open surface. Hence, the resultant equation becomes ill-conditioned, and not solvable.
The gradient on the Green’s function makes it more singular, and the principal value
integral method can be used to evaluate this singular integral [17]. After applying such
technique, the form of MFIE becomes
Z
1
n̂ × Hinc (r) = Js (r) − n̂ × P.V. ∇g(r, r0 ) × Js (r0 )dr0 , r ∈ S
2 s
(3.43)
where the first term on the right-hand side is the residue contribution of the singular integral,
and the second term, a principal value integral. The second term converges to the same value
irrespective of if the field point is on S, S + , or S − . This equation can be solved by using
MOM as for EFIE. In the above, since the unknown current appears both inside and outside
the integral, it is known as the second-kind integral equation. (In contrast, EFIE is known
as the first-kind integral equation.)
3.8 Internal Resonance and Combined Field Integral Eq-

uation
For closed surfaces, EFIE is only imposed on the surface of the scatterer. Hence, even if we
are solving for the scattering solution from a solid PEC object, EFIE does not distinguish that
from the solution of the thin-shell PEC object. Though simple that it may seem, this model
has a problem. A thin-shell PEC object is also a resonance cavity. Since MOM approximates
an actual physical object with facets and approximates its solution, electromagnetic energy
will leak into the thin-shell PEC object, exciting the internal resonance modes of the cavity.
These excited internal resonances contribute to the extraneous surface current on the surface
S which is quite different from the actual physical surface current that one wants to solve
for, that is, that of a solid PEC scatterer. Moreover, when the operating frequency is near
the resonance of the interior cavity, the interior resonance mode will also leak energy to the
exterior, giving rise to erroneous scattered field.
One may argue that the internal resonance surface current of a resonant modes is a non-
radiating current, and hence, the scattered field should not be overly affected. But in an
approximate model of facets and approximate expansion functions, this is not a perfect non-
radiating current. It is a weakly radiating current.
MFIE also suffers from the internal resonance problem as the EFIE, but the internal
resonance is of a different nature. MFIE is derived from the invocation of the extinction
theorem, such that the total field is extinct at S − . The nature of internal resonance for
MFIE is quite different from that of EFIE, and we will discuss this in greater detail in the
next chapter.
The internal resonance problem is related to the existence of a null-space solution for the
EFIE. Even if we can solve the EFIE exactly, the current from the internal resonance mode be-
longs to the null space of the EFIE integral operator. For a PEC cavity with lossless medium,
this internal resonance frequency is pure real, and can coincide (collide) with the operating
frequency of one’s calculation. Hence, the current is nonunique when the operating frequency
is exactly at the internal resonance frequency. At frequencies close to the internal resonance
frequency, the equation is ill-conditioned because the pertinent matrix representation of the
EFIE operator is also ill-conditioned: small eigenvalue exists because the resonance condition
is almost satisfied. The internal resonance mode of the EFIE operator is characterized by the
boundary condition n̂ × E = 0 on S.
The ill-conditioned matrix system due to internal resonances is deleterious to iterative
solvers as the iteration count needed to solve such problems becomes exceedingly large (see
last section of this chapter). For noniterative solvers, roundoff error can exist in the solution
because of ill conditioning. The internal resonance gives rise to spurious currents flowing on
the surface of the scatterer, and hence, the current obtained can be very wrong when this
occurs. Luckily, there is a way to remove this problem. The remedy is to formulate the
scattering problem so that the solution external to the scatterer is the same, but the interior
solution corresponds to a lossy wall cavity resonance [18]. Hence, the interior resonance
solution can only occur as a damped sinusoidal signal. In other words, the internal resonances
of the cavity correspond to complex resonances. The operating frequency of the calculation,
which is usually real, does not coincide (or collide) with the complex resonance frequencies
(which correspond to poles in the complex plane below the real axis). Consequently, when
the scatterer is excited by a sinusoidal time-harmonic signal, the interior resonance solution
does not grow to a large value to corrupt the scattered field in the exterior.
Alternatively, the remedy is to use a combined field integral equation (CFIE) [11], which
is a linear superposition of EFIE and MFIE. Usually, the integral equations are superposed
as
CFIE =αEFIE + η(1 − α)MFIE (3.44)
p
with α chosen to be about 0.5 and η = µ/² is the intrinsic impedance of space. It is
important that the above equations be combined in a manner so that an effective lossy
boundary condition ensues. As a consequence, the null-space solution of CFIE can occur
only at complex resonance frequencies. (We omit the proof here as it is similar to that for
scalar wave equation [19]. It will be discussed in more detail in the next chapter.) Hence, the
equation is free from internal resonance problem. A wrongly combined CFIE can correspond
to reactive boundary condition on the wall rather than dissipative wall losses. Reactive
boundary condition is lossless. In this case, there will still be real resonance modes rather
than complex resonance modes, plaguing calculations with real frequencies.
3.9 Other Boundary Conditions—Impedance Boundary

Condition, Thin Dielectric Sheet, and R-Card
So far, we have described the scattering solution to a scatterer where the surface scatterer
is modeled by a PEC. This is an idealization, and most likely, some surfaces are not PEC
surfaces. In this case, we can model a lossy surface or a reactive surface with an impedance
boundary condition (IBC) [20]. More of this will be discussed in Chapter 4.
For an IBC, instead of imposing the boundary condition that the tangential electric field
is zero on the surface of the scatterer, one imposes the condition that the tangential electric
field is [21]
Etan =ηs Js (3.45)
where ηs is the surface impedance in ohms. When the surface impedance is zero, we recover
the PEC case. For a thinly coated PEC surface, we can use the following equation to estimate
the surface impedance:
r µr ¶
µ ²µ
ηs = − i tan k0 d (3.46)
² ²0 µ0
where d is the thickness, ² and µ are the complex permittivity and permeability of the
material coating, ²0 and µ0 are free-space permittivity and permeability, and k0 is the free-
space wavenumber. The above is obtained using a simple transmission-line model or a layered-
medium model, and should be quite accurate when the surface is almost flat, and the thickness
is thin.
For lossy, bulky, large dielectric objects one may use the formula
r
µ
ηs = (3.47)
²
which is only an approximate concept. A thin dielectric sheet (TDS) can be modeled as a zero
thickness surface with a special impedance condition [12, 13]. Such an impedance condition
is
i
Z= (3.48)
ωd∆²
where ω is the operating angular frequency, ∆² = ² − ²0 and d is its thickness. Notice that
the above impedance is purely reactive when ² is pure real. When the permittivity is made
to be lossy, the TDS can be used to model a resistive sheet or an R-card. The difference
between an IBC and a TDS is that IBC is best used for modeling coating of an impenetrable
scatterer. On the other hand, a TDS can be used to model a semitransparent dielectric sheet
that allows the transmission of electromagnetic field through it. For instance, one may want
to model the windscreen of a car or a cockpit with a TDS. A more elaborate version of the
TDS has recently been developed [22].
3.10 Matrix Solvers–A Pedestrian Introduction

A pedestrian introduction to matrix solver is given here so that one can distinguish the
difference between solving an equation iteratively and noniteratively. This section does not
do justice to the vast amount of literature that have been written on matrix solvers. It is
hoped that the reader will gain at least a pedestrian knowledge of matrix solvers here.
Once the matrix equation is obtained, a whole sleuth of methods can solve the matrix
equation. The more recent advances have been on fast solvers in solving the matrix system
iteratively. But before describing the iterative solvers for the matrix equation, we will describe
some traditional solvers for the matrix system.
The simplest method of solving a matrix system is to use Gaussian elimination or LUD
(lower-upper triangular decomposition) [23]. These solvers are also known as direct inversion
solvers or noniterative solvers. One reason for their convenience is that there have been
many canned routines for such methods. However, their complexities are bad: If the matrix
system has N unknowns, the CPU time grows as N 3 . Hence, it uses a humongous amount
of CPU resource when N is large. One advantage of such a method is that when the LU
factorization is performed, the CPU cost of solving for a new right-hand side is proportional
to N 2 . However, the matrix system has to be available, and hence, all the matrix elements
have to be stored in the computer memory, requiring storage proportional to N 2 .
3.10.1 Iterative Solvers and Krylov Subspace Methods

Another way of solving a matrix system is to use an iterative method [24]. This method allows
one to solve a matrix system by performing a small number of matrix-vector multiplies at each
iteration. As one can produce the matrix-vector product without generating or storing the
matrix, the method can be made matrix free, greatly reducing the storage requirements. In
this case, only the unknowns need to be stored, and the memory requirement is proportional
to N instead of N 2 . For large N , this can be a tremendous saving in memory. For sparse
matrices, and fast algorithms for dense matrices,
¡ such
¢ matrix-vector product can be effected in
O(N ) or O(N log N ) operations, instead of O N 2 operations. So if the number of iterations
can be kept small in these methods, one can greatly improve the solution time of the equation
when N is large.
There are umpteen ways to solve a matrix equation iteratively. But we will illustrate the
concept with a very simple iterative solver. Consider a matrix system
A · x =b (3.49)
where A = D + T, the matrix D is the diagonal part of A while T is its off-diagonal part.
The above equation can be rewritten as
A =D + T
D · x =b − T · x
−1 ¡ ¢
x =D · b − T · x (3.50)
Since D is diagonal, finding its inverse is trivial. We can use the last equation to seek the
solution iteratively by writing it as
−1 ¡ ¢
xn =D · b − T · xn−1 (3.51)
In the equation above, we can start with an initial guess for x, substitute it into the right-
hand side and then solve for its new estimate on the left-hand side. The new value can then
be substituted into the right-hand side to obtain a newer estimate of x. This iteration goes
on until the value of x ceases to change. This method is called a fixed-point iteration also
known as Jacobi iteration.
In the above equation, the major cost of each iteration is a matrix-vector product T · x.
For a dense matrix with N unknowns, the cost of the matrix-vector product involves N 2
operations. So if the iterative process converges in Ni iterations, then the cost of solving the
above equation involves N 2 Ni operations. If Ni ¿ N , this method can be much faster than
the direct solver, which involves N 3 operations.
The above is only a simple illustration of an iteration solver, which is based on fixed
point iteration. There are a whole host of different iterative solvers making an alphabet soup
such as CG (conjugate gradient), BCG or BiCG (biconjugate gradient), BCGSTAB (BiCG
stabilized), GMRES (generalized minimal residual), QMR (quasi-minimal residual), TFQMR
(transpose free QMR) etc. Unfortunately, there is no one iterative solver that is the panacea
for all problems. One pretty much has to try different ones to see which one works the best
for our problems.
Many iterative solvers are based on the Krylov subspace method. One can understand
a lot about the properties of these iterative solvers through the Krylov subspace concept
without going through their full machineries [25–28]. In the Krylov subspace method, one
estimates the best solution to a matrix equation (3.49) by performing a number of matrix-
vector multiplies. One defines the initial estimate of the solution as x0 , and the initial residual
error as r0 = b − A · x0 . After K matrix-vector multiplies, K independent ¡ vectors,
¢ each of
length N , are generated that span a space called the Krylov subspace KK A, r0 , or
¡ ¢ n 2 K−1
o
KK A, r0 = span r0 , A · r0 , A · r0 , . . . , A · r0 (3.52)
An iterative solver finds the optimal solution to (3.49) by making use of these K independent
vectors to minimize the residual error. It finds
¡ the¢ optimal solution at the K-th iteration by
choosing xK = x0 + zK such that zK ∈ KK A, r0 . In this manner, the residual at the K-th
iteration is
rK = b − A · xK
n 2 K
o
= r0 − A · zK ∈ span r0 , A · r0 , A · r0 , . . . , A · r0
¡ ¢
∈ KK+1 A, r0 (3.53)
In other words, the best estimate solution has the residual error in the K-th iteration as
2 K
rK = r0 + a1 A · r0 + a2 A · r0 + · · · + a2 A · r0
K
X k 0
¡ ¢
= ak A · r0 = PK A · r0 (3.54)
k=0
where a0 = 1, and the residual error for exact solution is zero, whereas that of the optimal
0 0
one is as small ¡as possible.
¢ In the above, PK (x) is a K-th order polynomial with PK (0) = 1,
0
and hence, PK A is a polynomial of the matrix A.
Different iterative solvers use different strategies in finding the ak ’s. For instance, in
the CG method (applicable to a positive definite Hermitian system), at every new iteration,
when a new independent vector is added to the Krylov subspace by performing a matrix-
vector product, it constructs a new residual error vector that is orthogonal to all the previous
ones. Because in an N dimensional space, there could be only N independent orthogonal
vectors of length N , the exact solution is found after N iterations.
In practice, however, the arithmetics in a computer is not exact, and numerical roundoff
error often precludes the convergence of the CG method in N steps. Numerical roundoff causes
the loss of orthogonality in these vectors. As a result, different methods are constructed to
remedy this convergence issue. Moreover, methods are developed for non-Hermitian systems.
To understand the convergence issue, one can expand r0 in terms of the eigenvectors of
A, namely,
XN
¡ ¢
r0 = vn wnt · r0 (3.55)
n=1
assuming that the eigenvalues and eigenvectors are distinct and that a reciprocal set of vectors
wn exists such that wn · vn = δij . Using the above in the k-th term of (3.54), one gets
N
X
k ¡ ¢
A · r0 = λkn vn wnt · r0 (3.56)
n=1
From the above, one sees that when k → ∞, the eigenvectors with small eigenvalues are
deemphasized in the k-th term, whereas the large eigenvalue terms are emphasized swamping
the small eigenvalue terms. Hence, the new independent vector that can be generated by
this process is limited by numerical roundoff. In the extreme case when one eigenvector has
a very large eigenvalue, while all the other eigenvalues are very small or close to zero, each
matrix-vector product does not add a new independent eigenvector to the Krylov subspace.
Therefore, when the dynamic range of eigenvalues is very large, the iterative method is not
effective in generating new independent vectors after a while because of numerical roundoff.
The condition number of a matrix is the ratio of the largest singular value of the matrix
to the smallest singular value (For positive definite Hermitian matrix, the eigenvalues are also
the singular values). For matrices with large condition numbers, the eigenvector constituents
in the system become imbalanced after a number of iterations. Therefore, the number of
iterations needed to solve the matrix equation is related to the condition number of the
matrix system.
Iterative ¡method
¢ based on Krylov subspace method seeks the minimum of krK k subject
to zK ∈ KK A, r0 , or
° 0 ¡ ¢ ° ° 0 ¡ ¢°
min krK k = min °PK A · r0 ° ≤ min °PK A ° kr0 k (3.57)
0 0
zK ∈KK (A,r0 ) PK (x) PK (x)
√
The vector norm krk = |r| = r† · r or the magnitude of the vector, while the matrix norm
° ° ©° ° ª
°A° = max °A · x° / kxk
x6=0
−1
If the matrix A is diagonalizable such that A = U · Λ · U , where Λ is a diagonal matrix
0
¡ ¢ 0
¡ ¢ −1
that contains the eigenvalues of A on its diagonal, then PK A = U · PK Λ · U . As a
result,
° 0 ¡ ¢° ° ° ° 0 ¡ ¢° ° °
°PK A ° ≤ °U° °PK Λ ° ° °U °
−1 °
(3.58)
° ° ° °° °
where the inequality °A · B° ≤ °A° °B° has been used. Furthermore
° 0 ¡ ¢°
°PK Λ ° = max |PK 0
(λi )| (3.59)
λi ∈Λ(A)
¡ ¢
where Λ A means the set of all eigenvalues of A. Finally, one gets
° °°° −1 °
°
min krK k ≤ °U° °U ° kr0 k min
0 (x)
max |PK0
(λi )| (3.60)
zK ∈KK (A,r0 ) PK λi ∈Λ(A)
0
Therefore, to minimize the residual error, one picks the polynomial PK (x) such that it has
min-max value at the point when x is evaluated at the eigenvalues of A. It is seen that when
0
the eigenvalues are close to the origin, because PK (0) = 1, it is hard to find such a min-max
polynomial. Also, if the eigenvalues are scattered widely over the complex plane, finding such
a polynomial is difficult. However, if the eigenvalues are clustered around points away from
the origin, finding such a °polynomial is easier.
° °° ° −1 °
The factor °U° °U ° is the condition number of the matrix U. For ill-conditioned U,
this number can be fairly large, and (3.60) is not a very tight bound. For a normal matrix2
for which Hermitian matrix is a special case, the matrix A can be diagonalized such that U
−1 †
is unitary and U = U . The norm of a unitary matrix is 1, and the above bound is tight.
−1 t
For a diagonalizable complex symmetric system, we can show that U = U implying that
U is complex orthogonal, we expect the bound to be tight since complex orthogonal matrices
generally have good condition numbers.
3.10.2 Effect of the Right-Hand Side—A Heuristic Understanding

The convergence of an iterative solver often depends on the right-hand side of the equation.
In scattering, this corresponds to the nature of the exciting source that generates the incident
2A † †
normal matrix is one for which A · A = A · A .
field. We can gain a heuristic understanding of the effect of the right-hand side by performing
an eigenvalue analysis of the pertinent equation (3.49). We let
N
X
b= vn (wnt · b) (3.61)
n=1
and let
N
X
x= cn vn (3.62)
n=1
Using the above in (3.49), and using similar technique to (3.55), we deduce that cn = wnt ·b/λn
or
XN
x= vn wnt · b/λn (3.63)
n=1
From the above, we observe that the importance of the eigenvector in a solution depends
very much on the right-hand side and the eigenvalue. The right-hand side will excite different
eigenvectors of the system depending on its nature or the inner product wnt · b. When
eigenvectors with small eigenvalues are excited, poor convergence of the iterative solution
will ensue because of the swamping of small eigenvalues by large eigenvalues in the Krylov
subspace (see discussion after Eq. (3.55)).
A way to reduce the condition number of a matrix is to precondition it by multiplying the
matrix equation by its approximate inverse. It is hoped that the resultant matrix equation,
with a smaller condition number, requires a smaller number of iterations to solve. The
construction of a preconditioner remains an art rather than a science. The matrix is often
constructed with the aid of heuristics or physical insight.
The disadvantage of an iterative solver is that the number of iterations needed is un-
predictable, and very much problem dependent. Also, whenever the right-hand side of the
equation changes, the process has to begin all over again.
3.11 Conclusions
In this chapter, we introduce the concept of integral equations and the method used to solve
them. A very popular method is MOM which has been used with great success in solving a
number of practical problems. Coupled with fast solvers and parallel computers, it can solve
up to tens of millions of unknowns making electrically large problems solvable by numerical
methods, that were previously unsolvable [19, 29–31].
Figure 3.8 shows the current on an 83 Ford Camaro when it is illuminated by an elec-
tromagnetic field generated by a Hertzian dipole at 1 GHz. It is solved with the Fast Illi-
nois Solver Code (FISC) [32], a MOM code where the multilevel fast multipole algorithm
(MLFMA) is applied to accelerate the solution. The body of the car is modeled as a PEC,
the windscreens are modeled as TDS, and the tires are modeled with IBC.
Figure 3.9 shows the current distribution on a 02 Cadillac Deville when an XM antenna
is mounted on its roof [33]. Because of the disparate mesh sizes involved, the matrix system
is extremely ill-conditioned. A specially designed preconditioner, the self-box inclusion (SBI)
Figure 3.8: The current on a 83 Ford Camaro solved with MOM and the MLFMA accelerated
fast solver called the Fast Illinois Solver Code (FISC) [32].
preconditioner [34], has to be used in order to arrive at a converged solution. The windows
of the car are modeled with TDS.
The XM antenna consists of a probe-fed square patch with edges cleaved so that a circu-
larly polarized wave can be generated by the antenna. The antenna has a dielectric substrate
which is modeled as a volume integral equation (VIE). It has a radome that is modeled as a
TDS.
(a) Current distribution.
(b) XM antenna. (c) Triangulation of the car geometry.
Figure 3.9: (a) The current distribution on an 02 Cadillac Deville when (b) an XM antenna
is mounted on its roof, solved with MOM and the MLFMA accelerated fast solver called the
FastAnt [33]. (c) shows the triangulated geometry description of the car.
Bibliography
[1] P. M. Morse and H. Feshbach, Methods of Theoretical Physics, New York: McGraw-Hill,
1953.
[2] C. T. Tai, Dyadic Green’s Functions in Electromagnetic Theory, Piscataway, NJ: IEEE
Press, 2nd ed., 1994.
[3] J. A. Kong, Theory of Electromagnetic Waves, New York: Wiley-Interscience, 1975.
[4] A. D. Yaghjian, “Electric dyadic Green’s functions in the source region,” Proc. IEEE, vol.
68, pp. 248–263, 1980.
[5] G. Green, An Essay on the Application of Mathematical Analysis to the Theories of Elec-
tricity and Magnetism, T. Wheelhouse, Nottingham, 1828. Also, see L. Challis and F.
Sheard, “The green of Green functions,” Phys. Today, vol. 56, no. 2, pp. 41-46, Dec. 2003.
[6] J. A. Stratton and L. J. Chu, “Diffraction theory of electromagnetic waves,” Phys. Rev.,
vol. 56, pp. 99-107, Jul. 1939.
[7] W. C. Chew Waves and Fields in Inhomogeneous Media, New York: Van Nostrand Rein-
hold, 1990. Reprinted by Piscataway, NJ: IEEE Press, 1995.
[8] W. Franz, “Zur formulierung des huygenschen prinzips,” Z. Naturforsch., vol. 3a, pp.
500-506, 1948.
[9] R. F. Harrington, Time-Harmonic Electromagnetic Fields, New York: McGraw-Hill, 1961.
[10] R. F. Harrington, Field Computation by Moment Methods, Malabar, FL: Krieger, 1968.
[11] R. F. Harrington and J. R. Mautz, “E-field, H-field, and combined field solutions for
conducting bodies of revolution,” A.E.Ü., vol. 32, no. 4, pp. 159–164, Apr. 1978.
[12] R. F. Harrington and J. R. Mautz, “An impedance sheet approximation for thin dielectric
shells,” IEEE Trans. Antennas Propag., vol. AP-23, pp. 531534, Jul. 1975.
[13] E. Bleszynski, M. Bleszynski, and T. Jaroszewicz, “Surface integral equations for electro-
magnetic scattering from impenetrable and penetrable sheets,” IEEE Antennas Propag.
Mag., vol. 35, no. 6, pp. 14–26, Dec. 1993.
6
[14] R. Mittra, editor, Computer Techniques for Electromagnetics, New York: Hemisphere
Publisher, 1978.
[15] S. M. Rao, G. R. Wilton, and A. W. Glisson, “Electromagnetic scattering by surfaces of
arbitrary shape,” IEEE Trans. Antennas Propag., vol. 30, no. 3, pp. 409–418, May 1982.
[16] P. P. Silvester and R. L. Ferrari, Finite Elements for Electrical Engineers, Cambridge,
MA: Cambridge University Press, 1986.
[17] D. R. Wilton, S. M. Rao, A. W. Glisson, D. Schaubert, O. Al-Bundak, C. Butler, “Poten-
tial Integrals for Uniform and Linear Source Distributions on Polygonal and Polyhedral
Domains,” IEEE Trans. Antennas Propag., AP-32, No. 3, pp. 276–281, Mar. 1984.
[18] D. R. Wilton, “Review of current status and trends in the use of integral equations in
computational electromagnetics,” Electromagn., vol. 12, pp. 287-341, 1992.
[19] W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song, eds., Fast and Efficient Algo-
rithms in Computational Electromagnetics, Boston, MA: Artech House, 2001.
[20] K. M. Mitzner, “Effective boundary conditions for reflection and transmission by an
absorbing shell of arbitrary shape,” IEEE Trans. Antennas Propag., vol. AP-16, pp. 706–
712, 1968.
[21] J. A. Stratton, Electromagnetic Theory, New York: McGraw-Hill, 1941.
[22] I. T. Chiang and W. C. Chew, “Thin dielectric sheet simulation by surface integral
equation using modified RWG and pulse bases,” IEEE Trans. Antennas Propag., vol. 54,
no. 7, pp. 1927-1934, Jul. 2006.
[23] G. H. Golub and C. F. Van Loan, Matrix Computations, Baltimore, MD: Johns Hopkins
University Press, 1983.
[24] R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkout, R.
Pozo, C. Romine, and H. van der Vorst, Templates for the Solution of Linear Systems:
Building Blocks for Iterative Methods, Second Edition, Philadelphia, PA: SIAM, 1994.
[25] Y. Saad and M. H. Schultz, “GMRES: a generalized minimal residual algorithm for
solving nonsymmetric linear systems,” SIAM J. Sci. Stat. Comput., vol. 7, pp. 856–869,
1986.
[26] T. A. Driscoll, K. C. Toh, L. N. TREFETHEN, “From potential theory to matrix itera-
tions in six steps,” SIAM Rev., vol. 40, no. 3, pp. 547–578, 1998.
[27] J. Liesen and P. Tichy, “Convergence analysis of Krylov subspace methods,” GAMM-
Mitteilungen, vol. 27, no. 2, pp. 153–173, 2004.
[28] C. Siefert and E. de Sturler, “Preconditioners for generalized saddle-point problems,”
SIAM Journal on Numerical Analysis, vol. 44, no. 3, pp. 1275-1296, 2006.
[29] S. Velamparambil, W. C. Chew, and J. M. Song, “10 million unknowns, is it that large,”
IEEE Antennas Propag. Mag., vol. 45, no. 2, pp. 43–58, Apr. 2003.
[30] M. L. Hastriter, “A study of MLFMA for large scale scattering problems,” Ph.D. Thesis,
Dept. ECE, U. Illinois, Urbana-Champaign, Jun. 2003.
[31] L. Gurel and O. Ergul, “Fast and accurate solutions of extremely large integral-equation
problems discretized with tens of millions of unknowns,” Electron. Lett., accepted for
publication.
[32] J. M. Song, C. C. Lu, W. C. Chew, and S. W. Lee, “Fast Illinois solver code (FISC),”
IEEE Antennas Propag. Mag., vol. 40, no. 3, pp. 27–33, 1998.
[33] H. Y. Chao, K. Pirapaharan, V. Bodrov, T. J. Cui, H. P. Hsu, G. Huff, X. J. Zhang,

J. S. Zhao, J. Bernhard, and W. C. Chew, “Simulation of vehicle antennas by the multilevel
fast multipole algorithm,” Antennas Appl. Symp., Allerton Park, Illinois, Sept. 18–20,
2002.
[34] H. Y. Chao, “A multilevel fast multipole algorithm for analyzing radiation and scattering
from wire antennas in a complex environment,” Ph.D. Thesis, Dept. ECE, U. Illinois,
Urbana-Champaign, Jan. 2003.
71
CHAPTER 4
Integral Equations
for Penetrable Objects
4.1 Introduction
The previous chapter illustrates the derivation of integral equations for a perfect electric
conductor (PEC) object. PEC objects are some of the simplest scattering objects we can
think of. They are also termed impenetrable objects as the electromagnetic energy does
not penetrate into them. Another case of an impenetrable object is the perfect magnetic
conductor (PMC) object. It is the dual of the PEC. In these objects, either the tangential
electric field or tangential magnetic field is zero on the entire surface, so that there could be
no net power flow (real or complex) into them.
However, many objects are penetrable such as dielectric objects, magnetic objects, or
lossy objects. It is possible to have net power flow into or through these objects. Hence, the
tangential electric field or the tangential magnetic field cannot be zero on the entire surface
of the object.
The subject of scattering by penetrable scatterers has been fervently studied by many
workers throughout the years. Before numerical methods were in vogue, approximate analytic
methods were popular such as the high frequency methods [1]. At one time, the extended
boundary condition (EBC) method [2] was popular, because of its simplicity. Other methods
that have been used are generalized multipole method [3], and more recently, differential
equation solvers such as finite element (FEM) and finite difference methods (FDM) [4–8].
There are two main types of integral equations for penetrable scatterers: surface integral
equation (SIE) when the inhomogeneity is piecewise constant, and volume integral equation
(VIE) when the inhomogeneity is highly heterogeneous. Early works were done in two di-
mensions [9–11] in the 1960s. Three dimensional integral equation work based on the method
of moments (MOM) has only become popular in the 1970s to the 1990s (see an account by
Poggio and Miller [12] and Miller et al. [13]). The first comprehensive paper on SIE method
for 3D electromagnetic scatterer was reported by Midgyesi-Mitschang et al. [14]. The readers
are referred to these works plus the references there in for a historical account.
7
We will first derive the SIEs for penetrable bodies. We shall assume that the permittivity
² and the permeability µ are different from their free-space values, but are constant or ho-
mogeneous inside these bodies. The case of lossy bodies is easily included by making the ²
or (and) the µ complex. Then later, we will derive the volume integral equations (VIEs) for
inhomogeneous penetrable scatterers.
4.2 Scattering by a Penetrable Object Using SIE

We assume that Jinc is a current that generates an incident field Einc that impinges on a
penetrable object (see Figure 4.1). Polarization currents will be induced on the penetrable
object, which in turn radiates to generate the scattered field. We can imagine a surface
S + that is just large enough to enclose S, the surface of the penetrable object and apply
the equivalence principle and extinction theorem to it [2, 15]. The equivalence principle and
extinction theorem have been stated in words by Harrington [15]. But we will show their
mathematical forms here (see derivation in the previous chapter).
When equivalence surface currents are placed on S + , they produce the total field outside
S and zero field inside S − (which is a surface just small enough to be enclosed in S). Hence,
+
the scattering object can be removed without disturbing the field outside. The equivalence
surface currents are given by
Js = n̂ × H, Ms = E × n̂ on S+ (4.1)
+
where E and H represent total fields and n̂ is a outward normal on S . Consequently, the
equivalence principle and the extinction theorem become [16]
¾ Z Z
r ∈ V1 , E(r)
= Einc (r)+iωµ dS 0 G(r, r0 )·Js (r0 )−∇× dS 0 G(r, r0 )·Ms (r0 ) (4.2)
r ∈ V2 , 0 S + S +
When we apply the extinction theorem part of the above equation to S − , we obtain the
integral equation of scattering, viz.,
Z Z
− 0 0 0
r ∈ S , 0 = Einc (r) + iωµ dS G(r, r ) · Js (r ) − ∇ × dS 0 G(r, r0 ) · Ms (r0 ) (4.3)
S+ S+
The above is also known as the electric field integral equation (EFIE) for a penetrable ob-
ject.1 Understanding the equivalence principle and the extinction theorem above is key to
the understanding of the internal resonance problem as shall be shown in the next section.
Using operator notations, we can rewrite (4.3) more compactly as
−Einc (r) = L1E (r, r0 ) · Js (r0 ) + K1E (r, r0 ) · Ms (r0 ), r ∈ S−, r0 ∈ S + (4.4)
where integrations are implied over repeated variables (extension of the index notation to
continuous variables), and L1E and K1E assume the property of medium 1. More explicitly,
they are µ ¶
∇∇
L1E (r, r0 ) = iωµ1 G1 (r, r0 ) = iωµ1 I + 2 g1 (r, r0 ) (4.5)
k1
1 When only the tangential components of the above equation is taken, the equation is still known as EFIE
in the literature.
INTEGRAL EQUATIONS FOR PENETRABLE OBJECTS 73
K1E (r, r0 ) = −∇ × G1 (r, r0 ) = −∇g1 (r, r0 ) × I (4.6)

where 0
eik1 |r−r |
g1 (r, r0 ) = (4.7)
4π|r − r0 |
The subscripts E of the above operators imply that they are generating E field from either
the electric current Js or from the magnetic current Ms respectively.
The above constitutes one equation with two unknowns, and the solution for Js and Ms
cannot be obtained by solving only one equation alone—more equations are needed. To derive
more equations, we apply the extinction theorem to region 2 in an inside-out manner. As a
result, we have
0 = L2E (r, r0 ) · Js (r0 ) + K2E (r, r0 ) · Ms (r0 ), r ∈ S+, r0 ∈ S − . (4.8)
In the above, we are assuming that Js and Ms are residing on S − , namely, r0 ∈ S − , but they
are the same as those for (4.4) since Js = n̂ × H, and Ms = E × n̂, and tangential E and H
are continuous across the interface S from S − to S + . We can impose (4.4) for r ∈ S − , and
(4.8) for r ∈ S + and obtain two integral equations for two unknowns Js and Ms . Since Js
and Ms are surface currents on a 2D manifold (or surface), only the tangential components
of the Eqs. (4.4) and (4.8) are imposed to solve for the current unknowns.
Even though we have enough equations to solve for the number of unknown functions, the
solution to the penetrable scatterer problem is fraught with nonuniqueness problem at the
internal resonance of the resonant modes formed by the surface S filled with exterior material.
This is because we impose the extinction theorem on a surface S, for the electric field only or
for the magnetic field only, which does not guarantee the extinction of the field in the entire
volume surrounded by S. We shall show this point later.
To derive equations that are free of internal resonances, we need the aid of the magnetic
field equivalence of the above equations. The magnetic field version of (4.2) can be derived
by taking the curl of (4.2) or by invoking the duality principle.
¾ Z Z
r ∈ V1 , H(r)
= Hinc (r)+iω² dS 0 G(r, r0 )·Ms (r0 )+∇× dS 0 G(r, r0 )·Js (r0 ) (4.9)
r ∈ V2 , 0 S+ S+
The corresponding integral equation of scattering is

Z Z
− 0 0 0
r ∈ S , 0 = Hinc (r) + iω² dS G(r, r ) · Ms (r ) + ∇ × dS 0 G(r, r0 ) · Js (r0 ) (4.10)
S+ S+
The above is the magnetic field integral equation (MFIE) for a penetrable object.
In a similar manner, MFIE can be compactly written using operator notation
−Hinc (r) = L1H (r, r0 ) · Ms (r0 ) + K1H (r, r0 ) · Js (r0 ), r ∈ S−, r0 ∈ S + (4.11)
where integral over repeated variables is implied, and more explicitly, the operators are
µ ¶
∇∇ 1
L1H (r, r0 ) = iω²1 G1 (r, r0 ) = iω²1 I + 2 g1 (r, r0 ) = 2 L1E (r, r0 ) (4.12)
k1 η1
Figure 4.1: Scattering by a penetrable object inside which the electric field and the magnetic
field are nonzero.
K1H (r, r0 ) = ∇ × G1 (r, r0 ) = ∇g1 (r, r0 ) × I = −K1E (r, r0 ) (4.13)

Similar to the EFIE case, we apply extinction theorem to region 2 in an inside-out manner
to obtain
0 = L2H (r, r0 ) · Ms (r0 ) + K2H (r, r0 ) · Js (r0 ), r ∈ S + , r0 ∈ S − (4.14)
It is to be noted that the MFIE is not independent from the EFIE since they are derivable
from one another by the action of the curl operator. However, we now have four equations
and two unknowns to solve for.
To prevent the internal resonance problem, a myriad of remedies have been proposed. One
is to use the combined field integral equation (CFIE) as has been used for the PEC scatterer
case. This will reduce the four equations we have derived to two. Alternatively, a weighted
sum of the external and internal equations is obtained both for the EFIE and the MFIE to
reduce the number of equations to two. This is known as either the PMCHWT [14, 17–19]
method or the Müller method [20] depending on how the weights are chosen. It can be shown
that these weighted-sum methods are free of internal resonance problems.
In the weighted-sum method, we add (4.4) and (4.8) with a weighting factor to obtain
−α1 Einc (r) = [α1 L1E + α2 L2E ] · Js (r0 ) + [α1 K1E + α2 K2E ] · Ms (r0 ), r ∈ S− (4.15)
Again, we can add (4.11) and (4.14) with a weighting factor to get
−η1 β1 Hinc (r) = η1 [β1 K1H + β2 K2H ] · Js (r0 ) + η1 [β1 L1H + β2 L2H ] · Ms (r0 ), r ∈ S + (4.16)
where η1 is used to ensure that (4.15) and (4.16) are of the same dimension. Again, only
the tangential components of the above equations need to be imposed to generate enough
equations to solve for the surface currents Js and Ms , since they live on a 2D manifold.
When α1 = α2 = β1 = β2 = 1, the above formulation is known as the PMCHWT (Poggio,
Miller, Chang, Harrington, Wu, Tsai) formulation. When α1 = β1 = 1, α2 = − ²²12 , and
β2 = µµ12 , it is known as the Müller formulation. When the tangential component of the above
equations are extracted, the corresponding K operators have the principal value terms (see
previous chapter). Because of the difference in the direction of the normals, the principal value
terms cancel in the the PMCHWT formulation. They do not cancel in the Müller formulation.
The Müller formulation is good for low contrasts, whereas the PMCHWT formulation is good
for high contrasts.
4.3 Gedanken Experiments for Internal Resonance Prob-

lems
The internal resonance problem of integral equations is a problem that has plagued solutions
of scalar integral equations and vector integral equations [18–29]. Although the existence of
these internal resonances can be easily proved for some cases, their existence for other cases
is less well understood.
For example in electromagnetics, the existence of internal resonance for the EFIE for
perfect electric conducting (PEC) scatterer can be easily shown. But the existence of this
problem for MFIE for PEC scatterer is less obvious. Their existence for penetrable scatterers
is even less obvious. The internal resonance problem in EFIE for PEC scatterers is related to
the internal cavity resonance and nonradiating current of a resonance cavity. The physical
character of the internal resonance can be easily understood. However, the physical character
of the internal resonance problem for MFIE for PEC scatterers is less well understood. For
penetrable scatterers, this is even less well understood.
In this section, we will prove, via Gedanken experiments (thought experiment), the ex-
istence of the internal resonance problems for MFIE for PEC, and for penetrable scatterers
when EFIE or MFIE alone is applied [30]. These proofs reveal with physical clarity the
reasons of the internal resonance problem, as well as the physical character of the internal
resonance problem.
The internal resonance problem exists both for impenetrable as well as penetrable scat-
terers when certain integral equations are used to solve for their scattering solutions. For
simplicity, we shall discuss the case of impenetrable scatterers first.
4.3.1 Impenetrable Objects

Some integral equations for impenetrable objects exhibit internal resonance problems. When
the solution of the integral equation is sought at the frequency that corresponds to the internal
resonance of the resonant modes formed by the object, the solution to the integral equation is
nonunique. Such is the case with the EFIE and the MFIE. However, this internal resonance
problem can be remedied.
There are two major kinds of impenetrable objects in electromagnetics, the PEC and the
PMC. Since the PMC is the dual of the PEC, we will consider the PEC here only. Other
impenetrable scatterers are those with reactive boundary conditions such that all energy is
reflected from the scatterer.
EFIE-PEC Case
The internal resonance problem for an EFIE when solving a PEC scattering problem will be
reviewed here. The EFIE for a PEC scattering problem is given as
n̂ × n̂ × LE (r, r0 ) · Jeq (r0 ) = −n̂ × n̂ × Einc (r), r∈S (4.17)
where Z
LE (r, r0 ) · Jeq (r0 ) = iωµ dS 0 G(r, r0 ) · Jeq (r0 ) (4.18)
S
In the above, the −n̂ × n̂× extracts the tangential components of the field. It can be replaced
by the dyad It = I − n̂n̂ where I is a unit dyad, and the above can be rewritten as
It · LE (r, r0 ) · Jeq (r0 ) = −It · Einc (r), r∈S (4.19)
When the EFIE is solved by imposing tangential n̂ × E = 0 on the surface of the scatterer,
the field is not extinct inside the scatterer if the operating frequency of the time-harmonic
incident wave is at one of the resonance frequencies of the PEC cavity formed by the surface
S. We can also imagine that the scattering solution from a thin PEC shell is being solved, for
which nontrivial interior solution to this PEC shell exists at the resonance frequencies of this
resonant modes. Consequently, a solution to the EFIE exists even when no exciting incident
field is present. In other words, the EFIE in the following has a nontrivial solution at internal
resonance, indicating the presence of the null-space solution to (4.19):
It · LE (r, r0 ) · Jeq (r0 ) = 0, r∈S (4.20)
The existence of a null-space solution makes the solution to the equation nonunique. Hence,
when the matrix representation of the EFIE operator is sought, the matrix equation is ill
conditioned.
To elaborate further, Eq. (4.20) is the equation for the internal resonance current of a
PEC cavity at its resonance frequencies. This current produces nontrivial electromagnetic
field inside the cavity, but it produces no field outside the cavity. This can also be understood
from the equivalence principle viewpoint. Hence, the resonance current is nonradiating.
Because the null-space current is a nonradiating current, we may be led to think that
this resonance current will not contribute to the scattered field, and hence, may still obtain
the correct scattered field even though the current is wrong and nonunique. However, the
scattered field is still fraught with small error. The reason is that the numerical solution is an
approximate solution, and this approximate null-space current is not a perfect nonradiating
source. At the resonance frequency of the resonant modes, this nonradiating current will still
leak some energy to the outside of the resonant modes contributing to error in the scattered
field [31].
Because of the nonuniqueness of the integral equation at resonance, the matrix system
representing the integral equation becomes ill conditioned. Hence, when iterative solvers are
used to solve for the solution, the number of iterations required is large, or the iterative
solution is nonconvergent. However, a noniterative solver such as lower-upper triangular
decomposition (LUD) can be used to solve the matrix system near resonance. The solution
usually exhibits larger error in a narrow band near the resonance frequencies of the resonant
modes.
Scattering from a 2D PEC circular cylinder is used as an example here [30], with pulse
expansion functions and point-matching. The diagonal elements of the MOM matrix are
calculated by small argument approximation of the Hankel function and the off-diagonal
elements are evaluated using the one-point quadrature rule [17, p. 41].
The first resonance of TM (transverse magnetic) modes in a circular cylinder is ka =
2.40482555769577 for TM01 mode, where k is the wave number and a is the radius of the
circle. Figure 4.2(a) shows the current distribution calculated by solving EFIE using MOM
and by Mie series at ka = 2.40482555769577. Large difference (relative RMS error: 0.436)
between the MOM results and Mie series is observed. The difference is the resonant electric
current induced on the PEC circular cylinder surface due to the incident plane wave. Because
the resonant electric current does not radiate, very accurate scattered fields as shown in
Figure 4.2(b) are obtained with a relative error as small as 0.001 in comparison with the Mie
series.
6 7
Bistatic Scattering Width (λ)

5 Mie Series 6 Mie Series
Surface Current (mA/m)
MOM result MOM result

5
4
4
3
3
2
2
1 1
0 0
0 30 60 90 120 150 180 0 30 60 90 120 150 180
φ (degrees) φ (degrees)
(a) Current distribution (b) Scattering width

Figure 4.2: (a) Current distribution and (b) scattering width at ka = 2.40482555769577 for
a circular cylinder due to a TM plane wave incidence.
MFIE-PEC Case
The MFIE is defined by
n̂ × KH (r, r0 ) · Jeq (r0 ) = −n̂ × Hinc (r), r ∈ S− (4.21)
where Z
KH (r, r0 ) · Jeq (r0 ) = ∇ × dS 0 G(r, r0 ) · Jeq (r0 ) (4.22)
S
The internal resonance problem of an MFIE for a PEC scattering problem is less well under-
stood. At the internal resonance of the MFIE, the scattered field solution has larger error
compared to that of EFIE. We shall provide a lucid physical explanation for this phenomenon.
To understand this, we perform a Gedanken experiment (see Figure 4.3) to construct a
null-space solution to the MFIE. We assume that the incident field on the scatterer, instead
M
S
inc
inc
µ ,ε1 1 µ ,ε
1 1
V
σ→∞ S
Figure 4.3: Presence of null-space solution for some integral equations for PEC scatterers can
be proved by performing a Gedanken experiment.
of coming from far away, is generated by equivalence currents impressed on a surface Sinc ,
where Sinc is larger than the surface of the PEC scatterer so that the PEC scatterer is
completely enclosed inside Sinc . In general, by the equivalence principle, two impressed
currents Jinc = n̂ × Hinc and Minc = Einc × n̂ are required on the surface Sinc .
First we pick the operating frequency of the Gedanken experiment to be at the resonance
frequency of the PMC formed by Sinc . In this manner, a nontrivial resonance solution exists
inside this PMC. Next we pick the equivalence currents to be the resonance current on the
surface of this PMC. Then only magnetic equivalence current, Minc = Einc × n̂, is needed,
because the electric equivalence current, Jinc = n̂ × Hinc = 0 on Sinc . Also, just inside
Sinc , n̂ × Hinc = 0 on Sinc since this is the PMC solution. However, the incident field (both
electric and magnetic) inside Sinc is not zero. Therefore, when a scatterer is placed inside
Sinc , scattered field is generated and exists outside the scatterer.
Now, with this experimental picture in mind, we can gradually shrink the size of Sinc such
that it becomes S + , where S + is a surface just infinitesimally larger than S to enclose it. As
the size of Sinc is gradually shrunk, we also alter the operating frequency of the impressed
current so that it becomes the resonance frequency of the new PMC after shrinkage. Because
of this, n̂ × Hinc = 0 always on the surface of the cavity formed by S + , but n̂ × Einc 6= 0,
and only the equivalence magnetic current is needed. Also, there is always an incident field
inside Sinc even when it becomes S + . So when we put a PEC inside S + , electric current will
be induced on the PEC, and a scattered field is generated outside S + .
By the above Gedanken experiment, we have constructed a nontrivial solution to the
MFIE. The driving term of the equation, n̂ × Hinc is zero, because it corresponds to the
tangential magnetic field on the surface of the PMC cavity formed by S + at resonance.
Tangential magnetic field is continuous from S + to S − ; therefore this field is also zero on S − .
Hence, the right-hand side of Eq. (4.21) is zero. Therefore, this nontrivial solution is also the
null-space solution to the MFIE, which is the solution to the following equation:
n̂ × KH (r, r0 ) · Jeq (r0 ) = 0, r∈S (4.23)

Two facts become clear in the above Gedanken experiment:
• The MFIE has a null-space solution when the operating frequency is at the resonance
frequency of the PMC formed by S. This frequency is exactly the same as the resonance
frequency of a similar PEC cavity by duality principle. Therefore, MFIE has the same
breakdown frequencies as EFIE.
• Unlike the null-space solution or current of the EFIE, this null-space current is quite
different physically. It is not the resonance current of the PEC cavity formed by S. But
it is an induced current on the surface S due to an incident field with n̂ × Hinc = 0 and
n̂×Einc 6= 0 on S + . Hence, this current is a good radiator, generating a strong scattered
field outside the scatterer. This is actually observed by MFIE numerical calculation near
the internal resonance frequency of the resonant modes.
The MFIE for a TE (transverse electric) wave scattering from a 2D PEC circular cylinder
is studied in the same way [30]. The diagonal elements are 0.5 and the off-diagonal elements
are evaluated using one-point rule. Figure 4.4 shows the current distribution and scattering
width at ka = 2.40482555769577. The difference is the electric current induced on the surface
due to the incident wave, which radiates and leads to large error in the scattered fields.
By the duality principle, the above can also be applied to prove the existence of internal
resonance problem when EFIE or MFIE is used to solve for the scattering solution of a PMC
scatterer.
5 2.5
Bistatic Scattering Width (λ)
Mie Series
Surface Current (mA/m)
4 2 MOM result
3 1.5
2 Mie Series 1
MOM result
1 0.5
0 0
0 30 60 90 120 150 180 0 30 60 90 120 150 180
φ (degrees) φ (degrees)
(a) Current distribution (b) Scattering width

Figure 4.4: Similar to Figure 4.2, solution of the magnetic field integral equation for a TE
plane wave.
4.3.2 Penetrable Objects

When a penetrable object such as a dielectric scatterer is solved by MFIE or EFIE alone, in-
ternal resonance problem occurs. On first thought, this is not obvious as no PEC cavity seems
to be formed here. Moreover, the resonance frequency of a dielectric scatterer must be com-
plex because of radiation damping. Hence, this resonance problem is coming from somewhere
else. We again shall provide a lucid physical explanation as to why this is happening.
J
S
inc
inc
µ ,ε
1 1
V
µ ,ε S
2
2 2
Figure 4.5: The presence of null-space solution for some integral equations for penetrable
scatterers can be proved by performing a Gedanken experiment.
Since the surface currents in (4.4) and (4.8) live in a 2D manifold, to maintain the same
number of equations as unknowns in the discretized equation, only the tangential components
of (4.4) and (4.8) are imposed when solving them. These are known as the EFIE form of
the equation for penetrable scatterers. We can show that It ·(4.4) and It ·(4.8) have internal
resonance problem by a similar Gedanken experiment.
To prove the existence of the null-space solution for the EFIE equation when applied to
the penetrable scatterer, we consider an incident field generated by an equivalence current
Jinc impressed on a surface Sinc (Figure 4.5). Again, we pick the frequency such that it is
the resonance frequency of the PEC cavity formed by Sinc in the absence of the scatterer.
Then we make Jinc the resonance current of the PEC cavity formed by Sinc . Only electric
current will be needed by the equivalence principle, since n̂ × Einc = 0 on the surface of Sinc .
But Hinc 6= 0, and Einc 6= 0 inside Sinc . There will be a scattered field Hsca and Esca in
this problem when a penetrable scatterer such as a dielectric scatterer is placed inside Sinc .
Now, we adjust and shrink Sinc so that it just becomes S + , which is infinitesimally larger
than S so as to enclose it, and also we alter the frequency continuously so that the operating
frequency is always equal to the resonance frequency of the resonant modes formed by Sinc .
Then, there exists a trivial n̂ × Einc in (4.4) such that the scattering solution is nontrivial.
Namely Js and Ms are nontrivial, even when n̂ × Einc = 0. This proves the existence of
null-space solutions to the tangential components of (4.4) and (4.8).
For the case when MFIE alone is used to solve for the scattering solution of a penetrable
scatterer, we can repeat the above Gedanken experiment with a PMC resonance cavity that
generates an equivalence magnetic current. We use this resonance current as the equivalence
current to generate an incident field and a scattered field that have associated currents on the
surface of the scatterer. These currents are in the null space of the tangential components of
the equations (4.11) and (4.14).
A few things become clear in this Gedanken experiment.
• The scattering solutions of a penetrable scatterer have internal resonance problem when
only EFIE or MFIE alone is used.
• The solution is nonunique at the resonance frequencies of the resonant modes formed
by the surface S. These are exactly the same frequencies that one encounters in solving
the PEC or PMC scatterer case when EFIE or MFIE is used to solve the problem.
• The incident field acts like the field from a resonant electric current of a PEC cavity
(for EFIE alone) or a resonant magnetic current of a PMC cavity (for MFIE alone) on
the penetrable scatterer. They will appear as zero on the left-hand sides of (4.4) and
(4.11) when tangential components of the equations are taken. However, these incident
fields are capable of inducing a scattered field from a scatterer. Therefore, there will be
nonzero equivalent surface currents Js and Ms that satisfy these equations, which form
the null-space solutions of the equations. These null-space currents are good radiators
that produce erroneous scattered fields at these frequencies, hence contributing to the
scattered field errors.
4.3.3 A Remedy
With the above insight, the source of the internal resonance problem becomes clear: The
extinction of tangential electric field or tangential magnetic field on a surface S does not
guarantee the extinction of the total field inside the surface S. This is because internal
resonance solutions exist inside the cavity formed by S at its resonance frequencies. These
resonance fields, depending on whether the resonant modes is a PMC or a PEC, have either
n̂ × H = 0 or n̂ × E = 0 on S.
The remedy for this is to arrive at equations such that when the appropriate field com-
ponents are extinct on the surface S, it guarantees the extinction of the total field inside S.
The CFIE is such an equation. It is a linear superposition of the EFIE and MFIE.
αIt · LE (r, r0 ) · Jeq (r0 ) + (1 − α)ηn̂ × KH (r, r0 ) · Jeq (r0 )

= −αIt · Einc (r) − (1 − α)ηn̂ × Hinc (r), r ∈ S− (4.24)
The null-space solution of the above has to satisfy the equation
αIt · LE (r, r0 ) · Jeq (r0 ) + (1 − α)ηn̂ × KH (r, r0 ) · Jeq (r0 ) = 0, r ∈ S− (4.25)
The above is the equation for a resonant modes solution with boundary condition that
αEt (r) + (1 − α)ηn̂ × H(r) = 0, r ∈ S− (4.26)
The above is equivalent to that the tangential electric field and the tangential magnetic
field satisfy an impedance boundary condition (IBC) together on the surface S. If the IBC
is made lossy, then the resonance solutions inside the surface S have complex resonance
frequencies [32]. These complex resonances correspond to damped sinusoidal signals inside
the resonant modes formed by the surface S. Hence, when the solution is sought at a real
time harmonic frequency, this complex resonance solution is not possible and the integral
equation does not have a null space.
4.3.4 Connection to Cavity Resonance

The proof of the uniqueness of CFIE has been given [18]. However, no proof of its connection
to the cavity resonance problem has been reported, even though assertion has been made
in the literature [17]. The partial proof is shown by considering a cavity resonance problem
where the resonant modes is filled with material of region 1. We can formulate the integral
equation for cavity resonance by using (4.8) but with material of region 1.
0 = L1E (r, r0 ) · Js (r0 ) + K1E (r, r0 ) · Ms (r0 ), r ∈ S+ (4.27)
Now the boundary condition on a lossy cavity wall is
Es = Zs Js , or Ms = −Zs n̂ × Js (4.28)
Using the above in (4.27), then
0 = L1E (r, r0 ) · Js (r0 ) − Zs K1E (r, r0 ) · n̂0 × Js (r0 ), r ∈ S+ (4.29)
The above is the integral equation for the cavity resonance problem. Since it is a source-free
problem, it does not have a nontrivial solution except when it is solved at the resonance
frequencies of the cavity. In other words, the integral operator in (4.29) does not have a null
space except at the resonance frequencies of the cavity. When the resonant modes is lossy,
its resonance frequency is complex. Hence, at real ω, the integral operator in (4.29) does not
have a null space.
We can show that (4.29) and (4.25) are related. To this end, we test the above with a
tangential current JT which is tangential to S and is supported at r ∈ S + .
0 = hJT (r), L1E (r, r0 ), Js (r0 )i − Zs hJT (r), K1E (r, r0 ) · n̂0 ×, Js (r0 )i,
r0 ∈ S, r ∈ S+ (4.30)
From the above, the transpose of the integral operator shall be derived. Because L1E (r, r0 ) is
a symmetric operator, its transpose is just itself. Looking more carefully at the second part
of the operator,
hJT (r), K1E (r, r0 ) · n̂0 ×,Js (r0 )i

Z Z
= dS dS 0 JT (r) · ∇g(r, r0 ) × (n̂0 × Js (r0 ))
+
ZS ZS
= dS dS 0 JT (r) × ∇g(r, r0 ) · (n̂0 × Js (r0 ))
S+ S
Z Z
0
= dS dS (n̂0 × Js (r0 )) · ∇0 g(r, r0 ) × JT (r)
S S+
Z Z
= − dS 0 dSJs (r0 ) · n̂0 × (∇0 g(r, r0 ) × JT (r))
S S+
= −hJs (r ), n̂ × K1H (r0 , r), JT (r)i
0 0
(4.31)
Notice that in the above g(r, r0 ) = g(r0 , r). Moreover, r and r0 are dummy variables. Hence,
it can be seen that the transpose of the operator K1E (r, r0 ) · n̂0 × is −n̂0 × K1H (r0 , r) [16].
Hence, if Zs = (1 − α)η/α is substituted in (4.29) and the resultant equation multiplied by

α, then the integral operators in (4.29) and (4.25) are related to each other by a transpose
relation.
When the matrix representations of these operators are sought with the proper expansion
functions for the range and the domain spaces, then the matrices are also the transpose of
each other. It is well known in matrix theory that if a matrix has a right null space, it will
also have a left null space, and vice versa. However, this concept is not easily extendible to
linear operators in infinite dimensional space. But we shall assume this to be true for this
class of problems.
Therefore, we shall assume that if the transpose of an operator has no null space, the
operator itself has no null space. Hence, if (4.29) has no null space when ω is real by physical
argument, then (4.25) has no null space. However, if Zs is chosen wrongly to give rise to a
lossless, reactive IBC, internal cavity resonances at real ω still exist, and hence (4.29) will
still have a null space, and so would the transpose operator. Hence, the proper choice of the
combination parameter in CFIE is important so that the right physics occurs in the solution.
It is to be noted that the combined source integral equation (CSIE) produces an operator
that is the transpose of the CFIE operator [33]. Hence, CSIE is free of the internal resonance
problem because it is directly connected to cavity resonance.
4.3.5 Other remedies

The above use of CFIE is one remedy to the internal resonance problem. Other remedies are:
1. Extinction of both the tangential and the normal components of the electric or the
magnetic field to guarantee the removal of the resonance solution.
2. Extinction of the tangential components of both the electric and the magnetic field to
guarantee the removal of the resonance solution.
3. Extinction of the tangential component of the electric or the magnetic field at two
surfaces close to each other to remove the internal resonance solution.
4. Insertion of shorts in the resonance cavity to push the resonance frequency of the reso-
nant modes higher (see [29] and references therein).
5. Insertion of a lossy object inside the resonance cavity to cause the resonance frequency
to become complex [34].
However, the above remedies require more equations than unknowns for an impenetrable
scatterer. Hence the CFIE for impenetrable scatterers offers a solution where the equation
number is not increased compared to the number of unknowns.
For a penetrable scatterer, remedy 2 above can be used effectively. The tangential compo-
nents of (4.4), (4.8), (4.11), and (4.14) constitute four equations and two surface unknowns Js
and Ms . The integral equations can be added to the exterior equations to reduce the number
of equations to two, but at the same time, both the tangential components of the electric and
magnetic field are extinct at the surface S − . This will guarantee the extinction of the total
field inside the resonant modes even at internal resonances, and hence, remove the null-space
solutions that we have constructed by our Gedanken experiment. Numerical calculations with
the weighted-sum methods of PMCHWT (Poggio, Miller, Chang, Harrington, Wu, Tsai) and
Müller do not indicate internal resonance problem for this reason.
4.4 Volume Integral Equations

In this section, the VIE is derived similar to the derivation given in [16]. The vector wave
equation for a general inhomogeneous, anisotropic medium is given by
∇ × µ−1 2
r · ∇ × E(r) − ω ²r (r) · µ0 ²0 E(r) = iωµ0 J(r) (4.32)
where µr (r) = µ(r) ²(r) 2

µ0 , and ²r (r) = ²0 . After subtracting ∇ × ∇ × E(r) − ω µ0 ²0 E(r) from
the sides of the equation, the above can be rewritten as
£ ¤
∇ × ∇ × E(r) − ω 2 µ0 ²0 E(r) = iωµ0 J(r) + ∇ × I − µ−1 r (r) · ∇ × E(r)
£ ¤
− ω 2 µ0 ²0 I − ²r (r) · E(r) (4.33)
The dyadic Green’s function corresponding to the differential operator on the left hand side
can be derived, viz.,
∇ × ∇ × G(r, r0 ) − k02 G(r, r0 ) = Iδ(r − r0 ) (4.34)
where k02 = ω 2 µ0 ²0 , and µ ¶
∇∇
G(r, r ) = I + 2 g(r, r0 )
0
(4.35)
k0
G(r, r0 ) · a can be thought of as the field due to a point source aδ(r − r0 ) located at r = r0 .
By the principle of linear superposition, we treat the right hand side of (4.33) as equivalence
volume sources, and write down the solution E(r) as
Z
E(r) = iωµ0 G(r, r0 ) · J(r)dr0
V+
Z
£ ¤
+ G(r, r0 ) · ∇0 × I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
V+
Z
£ ¤
− k02 G(r, r0 ) · I − ²r (r0 ) · E(r0 )dr0 (4.36)
V+
where V + is a volume that is slightly larger than the support of the scatterer V defined by the
volume where µr or ²r departs from I. The first term can be considered the field generated
by the current source J in the absence of the scatterer; hence, we call it Einc (r), a known
field. Consequently, (4.36) becomes an integral equation for E(r), or
Z
£ ¤
inc
E(r) = E (r) + G(r, r0 ) · ∇0 × I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
V+
Z
£ ¤
− k02 G(r, r0 ) · I − ²r (r0 ) · E(r0 )dr0 (4.37)
V+
The second term on the right-hand side represents the scattered field due to induced magnetic
polarization current from the inhomogeneous permeability, whereas the third term represents
that due to the induced electric polarization current from the inhomogeneous permittivity. A
similar derivation has been presented in [16] for isotropic media. This derivation clearly sepa-
rates the scattered field from the inhomogeneous permeability from that of the inhomogeneous
permittivity.
4.4.1 Alternative Forms of VIE

The above is not suitable for computational electromagnetics. However, the equation can
be made computationally friendly through mathematical manipulations. The dyadic Green’s
function in the second term on the right-hand side of (4.37) has a term that involves the
double ∇ operator. To reduce numerical quadrature error, it is a general wisdom to move the
del operator from more singular terms to less singular terms using integration by parts [35].
But it can be shown that
Z Z
∇∇ g(r, r0 ) · ∇0 × F(r0 )dr0 = −∇ ∇0 g(r, r0 ) · ∇0 × F(r0 )dr0
V + +
ZV
= −∇ ∇0 × ∇0 g(r, r0 ) · F(r0 )dr0 = 0 (4.38)
V+
Consequently, (4.37) becomes

Z
£ ¤
E(r) = Einc (r) + g(r, r0 )∇0 × I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
+
Z V
2
£ ¤
−k0 G(r, r0 ) · I − ²r (r0 ) · E(r0 )dr0
V+
Z
inc
£ ¤
= E (r) + ∇ × g(r, r0 ) I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
V +
Z
£ ¤
−k02 G(r, r0 ) · I − ²r (r0 ) · E(r0 )dr0 (4.39)
V+
where integration by parts and that ∇g(r, r0 ) = −∇0 g(r, r0 ) have been used. The meaning
of G(r, r0 ) in (4.39) still has to be properly defined because the dyadic Green’s function is
plagued with singularities whose evaluation has to be taken with great care. However, we can
mitigate
£ the¤ effect of the singularity if we define the action of the dyadic Green’s operator on
I − ²r (r0 ) · E(r0 ) to mean
Z
£ ¤
dr0 G(r, r0 ) · I − ²r (r0 ) · E(r0 )
V
µ ¶ Z
∇∇ £ ¤
= I+ 2 · g(r, r0 ) I − ²r (r0 ) · E(r0 )dr0
k0 V
Z Z
£ ¤ ∇ £ ¤
= dr g(r, r ) I − ²r (r0 ) · E(r0 ) + 2
0 0
g(r, r0 )∇0 · I − ²r (r0 ) · E(r0 )dr0 (4.40)
V k0 V
In other words, we never move both the ∇ operators completely inside the integrand to render
it ill defined. Also, integration by parts has been used to arrive at the form of the last integral.
The integrals in (4.40) are always well defined.
We can cast the VIE yet into another form which is more suitable for curl-conforming
basis. To this end, we make use of the identity for the dyadic Green’s function using (4.34)
as
1 £ ¤
G(r, r0 ) = 2 ∇ × ∇ × G(r, r0 ) − Iδ(r − r0 )
k0
1 £ ¤
= 2 ∇ × ∇ × Ig(r, r0 ) − Iδ(r − r0 ) (4.41)
k0
Using (4.41) in (4.39), we have
Z
inc
£ ¤
²r (r) · E(r) = E (r) + ∇ × g(r, r0 ) I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
+
ZV
£ ¤
−∇× ∇g(r, r0 ) × I − ²r (r0 ) · E(r0 )dr0 (4.42)
V+
where we have moved the delta function contribution from (4.41) in (4.42) to the left-hand
side. By using integration by parts, one arrives at
Z
inc
£ ¤
²r (r) · E(r) = E (r) + ∇ × g(r, r0 ) I − µ−1 0 0 0
r (r ) · ∇ × E(r )dr
0
V +
Z
©£ ¤ ª
−∇× g(r, r0 )∇0 × I − ²r (r0 ) · E(r0 ) dr0 (4.43)
V+
The above formulations, with the curl operators outside the integrals and some curl operators
inside the integrals, are more suitable for curl-conforming expansion and testing functions.
When the medium is highly anisotropic, the form in (4.42) is preferable, whereas if the medium
is isotropic and piecewise constant, the form in (4.43) is preferable.
We can also derive the magnetic field equivalence of the above equations by duality prin-
ciple. From (4.39), we have
Z
£ ¤
H(r) = Hinc (r) + ∇ × g(r, r0 ) I − ²−1 0 0 0
r (r ) · ∇ × H(r )dr
0
V +
Z
2
£ ¤
−k0 G(r, r0 ) · I − µr (r0 ) · H(r0 )dr0 (4.44)
V+
From (4.42), we have

Z
£ ¤
µr (r) · H(r) = Hinc (r) + ∇ × g(r, r0 ) I − ²−1 0 0 0
r (r ) · ∇ × H(r )dr
0
Z
£ ¤
−∇× ∇g(r, r0 ) × I − µr (r0 ) · H(r0 )dr0 (4.45)
The above equations can also be obtained by taking the curl of the EFIEs. They should be
used if the magnetic field is the dominant field inside the volumetric scatterer, or when the
dominant inhomogeneity is coming from the permeability inside the scatterer.
4.4.2 Matrix Representation of VIE

In this section, we shall derive the matrix representation of one of the integral equations we
have derived. The matrix representations of the others should be similar.
When µr = I, the second term on the right-hand side of (4.39) vanishes, and it is cus-
tomary to write (4.39) as
Z
inc 2
E(r) = E (r) − k0 G(r, r0 ) · ξ(r0 ) · D(r0 )dr0 (4.46)
V+
where
1 £ −1 0 ¤
ξ(r0 ) = ²r (r ) − I (4.47)
²0
The normal component of D(r) is continuous across the interface of a volume element, and
it is expedient to expand D(r) in terms of divergence conforming expansion functions. Work
on this has been reported in the past [36].
However, when µr 6= I, the existence of ∇ × E(r) in (4.39) calls for the use of curl-
conforming functions such as the edge elements. Such expansion functions have been used
frequently in the finite element literature [6, 7, 37, 38]. When linear edge elements are used,
the matrix representation of (4.39) assumes a simple form. Hence, it is prudent to discuss
finding the matrix representation of the above when curl-conforming expansion function is
used for expanding the electric field E. Assuming that V + is infinitesimally larger than V , the
support of the scatterer, we can break V + into a union of nonoverlapping Vn ’s. For example,
each of the Vn ’s can represent a tetrahedron if the volume V is tesselated into a union of N
tetrahedrons. Consequently, we let
N X
X 6
E(r) = Ini Eni (r), r∈V (4.48)
n=1 i=1
where the Eni (r) is the electric field in the n-th tetrahedron corresponding to the i-th edge,
which is represented by the tetrahedral edge element. The expansion function Eni (r) has
several properties that make it useful for representing E(r).
1. Within each element, the electric field in any direction can be expressed by the six
expansion functions defined on the six edges.
2. Eni (r) has a constant tangential component along edge i, and has no tangential com-
ponent along the other five edges.
3. The component of Eni (r) tangential to the i-th edge is constant and continuous, this
ensures the continuity of the component of electric field tangent to the edges.
4. Ii represents the amplitude of the tangential component of E(r) along the i-th edge.
5. The divergence of Eni is zero inside the tetrahedron for linear edge elements, but it
produces delta function singularities on the faces of the tetrahedron.
S We assume that the volume of integration is divided £into a union 0

¤ of V 0
n such that V =
n Vn , where Vn ’s are nonoverlapping. Within each Vn , I − ²r (r ) · E(r ) is differentiable,
but
£ it is discontinuous
¤ from region to region . The action of the divergence operator on
I − ²r (r0 ) · E(r0 ) will generate delta function singularities which are equivalent to surface
charges. These singularities will introduce a surface integral term. Consequently, using (4.48)
in (4.40), we have the scattered field due to the inhomogeneous permittivity part of the
scatterer to be
X Z
² 2
£ ¤
E (r) = −k0 Ini dr0 g(r, r0 ) I − ²r (r0 ) · Eni (r0 )
n,i Vn
X Z
£ ¤
− Ini ∇ g(r, r0 )∇0 · I − ²r (r0 ) · Eni (r0 )dr0
n,i Vn−
X Z
£ ¤
− Ini ∇ n̂0 · g(r, r0 ) I − ²r (r0 ) · Eni (r0 )dS0 (4.49)
n,i Sn
The surface integral term above can also be derived using careful integration by parts argu-
ment, but the derivation is more laborious.
When linear edge elements are used for Eni such that ∇ · Eni = 0 inside Vn , and we
assume that ²r is piecewise constant inside each Vn , the second term in the above vanishes,
leaving only the first and the third term. On testing the above with Emj (r), m = 1, . . . , N ,
j = 1, . . . , 6, and integrating over Vm+ which is infinitesimally larger than Vm , we have after
using integration by parts that
X £ ¤ ®
hEmj , E² i = −k02 Ini Emj (r), g(r, r0 ) I − ²r (r0 ) , Eni (r0 )
n,i
X Z Z
£ ¤
+ Ini dr∇ · Emj (r) n̂0 · g(r, r0 ) I − ²r (r0 ) · Eni (r0 )dr0 (4.50)
+
n,i Vm Sn
Again, if Emj is a linear edge element, its divergence is zero inside Vm , but has delta function
singularities on the faces of Vm , resulting in a surface integral. Consequently, the above
becomes
X £ ¤ ®
hEmj , E² i = −k02 Ini Emj (r), g(r, r0 ) I − ²r (r0 ) , Eni (r0 )
n,i
X Z Z
£ ¤
+ Ini dS n̂ · Emj (r) n̂0 · g(r, r0 ) I − ²r (r0 ) · Eni (r0 )dr0 (4.51)
n,i Sm Sn
The above surface integral can also be derived using integration by parts if the integral in
(4.50) is over Vm instead of over Vm+ .
By applying the same procedure to the part of the scattered field due to the permeability
in (4.39), we have
X £ ¤ 0
hEmj , Eµ i = Ini h∇ × Emj (r), g(r, r0 ) I − µ−1 0 0
r (r ) , ∇ × Eni (r )i (4.52)
n,i
If the edge element tetrahedral expansion functions are used in the above, their curls give
rise to a regular field (constant in terms of first order element) inside the tetrahedrons, and
surface singularities on the surfaces of the tetrahedrons. However, continuity of the tangential
components of the field will cause these surface charges or singularities that come from ∇0 ×
Eni (r0 ) to cancel each other. Moreover, in testing functions associated with ∇×Emj (r) in the
above, the surface singularities will cancel each other except on the surface of the scatterer
described by µ−1 r (r).
Consequently, using (4.51) and (4.52) together with the matrix representation of the rest
of (4.39), one arrives at the matrix representation of the integral equation (4.39),
N X
X 6
einc
mj = Amj,ni Ini , m = 1, . . . , N, j = 1, . . . , 6 (4.53)
n=1 i=1
where
einc
mj = hEmj , Einc i (4.54)
Amj,ni = hEmj , Eni i − A²mj,ni− Aµmj,ni (4.55)
£ ¤ ®
Amj,ni = −k0 Emj (r), g(r, r ) I − ²r (r0 ) , Eni (r0 )
² 2 0
Z Z
£ ¤
+ dS n̂ · Emj (r) n̂0 · g(r, r0 ) I − ²r (r0 ) · Eni (r0 )dr0 (4.56)
Sm S
µ £n ¤ 0
Amj,ni = h∇ × Emj (r), g(r, r ) I − µ−1
0 0
r (r ) , ∇ × Eni (r )i
0
(4.57)
In the above, the tangential continuity of the electric field has not been used. Using the
fact that tangential E is continuous from element to element, the redundancy of Ini can be
removed, as the Ini ’s from contiguous elements should be the same. This is very much like
the assembly process in the FEM with edge elements where redundancies in unknowns are
removed. Some numerical results using method presented here are given in [39]. Figure 4.6
shows the comparison of the Mie series solution (exact) with the proposed method. It is
seen that as the number of unknowns increases, the agreement with the Mie series solution
improves.
4.5 Curl Conforming versus Divergence Conforming Ex-

pansion Functions
When we find the matrix representation of an operator such as computing
Aij = hTi , LFj i (4.58)
a prerequisite is that such an element exists so that Aij is finite. For instance, if
L = ∇∇ (4.59)
Then, we can show, after integration by parts that,
Aij = −h∇ · Ti , ∇ · Fj i (4.60)

−10
RCS(dB/λ )
2
−30
Mie
VIE with 2212 meshes
VIE with 784 meshes
−50
0 60 120 180
θ(Degrees)
Figure 4.6: Radar cross section for a homogeneous sphere with composite material at moderately
low frequency of Mie series (solid line) and the proposed method. The radius of the sphere a = 0.15λ
with the parameters ²r = 1.5, µr = 2.2 with unknowns N = 784 and N = 2, 212 respectively.
If Fj (r) and Ti (r) are discontinuous functions, then the ∇·Fj (r) and ∇·Ti (r) can have delta-
function-like singularities and the inner product (4.60) is undefined
R∞ when these singularities
collide in their evaluation. For instance, in one dimension, −∞ dxδ(x)δ(x) = ∞ due to delta
function collision. Divergence conforming functions are functions whose divergence is finite
and hence Aij exists [40].
Similarly, if
L=∇×∇×I (4.61)
then (4.58), after using integration by parts, becomes
Aij = h∇ × Ti , ∇ × Fj i (4.62)
If Ti (r) and Fj (r) are discontinuous functions such that ∇ × Ti (r) and ∇ × Fj (r) can have
delta-function-like singularities, (4.62) may be infinitely large or undefined. Another possible
scenario such that (4.58) is undefined is when
Z
LFj = ∇ g(r, r0 )∇0 · Fj (r0 )dr0 (4.63)
V+
Then the equivalence of (4.60) for this operator is

Aij = −h∇ · Ti , g, ∇0 · Fj i (4.64)
If g(r, r0 ) has a singularity, and if both ∇ · Ti and ∇0 · Fj have delta-function-like singularities,
they will render (4.64) undefined.
4.6 Thin Dielectric Sheet

The thin dielectric sheet (TDS) model can be obtained as a limiting case of a VIE when it
is applied to a thin dielectric structure [41–43]. It has also been fervently studied by many
workers in recent years. A literature survey is given in [44].
Consider the VIE involving only dielectric inhomogeneity,

Z
−Jv (r)
Ei (r) = − iωµ0 g(r, r0 )Jv (r0 )dV 0
iω²0 [²r (r) − 1] V
Z
∇
+ g(r, r )∇ · Jv (r0 )dV 0
0 0
(4.65)
iω²0 V
where 0
eik0 |r−r |
g(r, r0 ) = (4.66)
4π|r − r0 |
The VIE can be simplified if we assume that the volume current Jv (r) inside the TDS is only
tangential in nature. This is true when the contrast of the TDS is high, so that the normal
component of the field is small in it, and only tangential component dominates. Furthermore,
we can assume that the tangential component of Jv is approximately a constant inside the
TDS and that its thickness is much smaller than the wavelength. In this case, we can replace
the volume integral with a surface integral and that the volume current can be replaced by a
surface current Js (r) such that
Js (r) = τ Jv (r) (4.67)
where τ is the effective thickness of the TDS. Consequently, (4.65) becomes
Z
i ∼
E (r) = Zs Js (r) − iωµ0 g(r, r0 )Js (r0 )dS 0
S
Z
∇
+ g(r, r0 )∇0 · Js (r0 )dS 0 (4.68)
iω²0 S
and
i
Zs (r) = (4.69)
ω²0 [²r (r) − 1]τ
Eq. (4.68) bears a striking resemblance to EFIE for a surface PEC scatterer. Hence, a
computer code for EFIE can be easily modified to accept the TDS model. Moreover, if we
let ²(r) ∼ iωµσ, when σ → ∞, then Zs → 0, and we must recover the scattering by a PEC
object. If the TDS is a lossy dielectric sheet, one can replace the Zs with a resistive sheet to
model the loss inside.
4.6.1 A New TDS Formulation

The aforementioned formulation for the conventional TDS is not valid for low contrast di-
electric sheets. Hence a new formulation is required for the low contrast case which has been
recently reported [44, 45]. To this end, we write the VIE as
Z
i D(r) k02
E = + g(r, r0 )χ(r0 )D(r0 )dV 0
²0 ²(r) ²0 V
Z
∇
+ g(r, r0 )∇0 · [χ(r0 )D(r0 )] dV 0 (4.70)
²0 V
where
χ(r) = ²−1
r (r) − 1 (4.71)
Moreover, one can show that
∇ · (χD) = (∇χ) · D + χ∇ · D = (∇χ) · D (4.72)
since ∇ · D = 0 in a source-free region. Consequently, the last integral in (4.70) becomes

Z Z
0 0 0 0 0
I= g(r, r )∇ · [χ(r )D(r )] dV = g(r, r0 )(∇0 χ(r0 )) · D(r0 )dV 0 (4.73)
V V
0
In the above, χ(r ) = 0 outside the TDS volume, and it is a constant inside the TDS volume.
It can be shown that
Z Z
∇0 χ(r0 ) · F(r0 )dV 0 = dS 0 n̂0 · [χ(r0 )F(r0 )] (4.74)
V S
0 0
if ∇ · F(r ) is nonsingular. Consequently, we have
Z
I = χ dS 0 n̂0 · D(r0 )g(r, r0 ) (4.75)
S
when χ is assumed constant within TDS. The surface integral can be broken into contribution
from the side surface plus the contribution from the top and bottom surfaces. Hence, we can
write (4.75) as
Z Z
I=χ dS 0 n̂0t · Dt (r0 )g(r, r0 ) + χ dS 0 n̂0 · D+ 0 0
n (r )g(r, r )dS
0
+
St Sn
Z
+χ dS 0 n̂0 · D− 0 0
n (r )g(r, r )dS
0
(4.76)
−
Sn
where St represents the side surface and Dt represents the component of D tangential to
the TDS, Sn± and D± n represent the top and bottom surfaces and their respective normal
components.
We can use the ∇ · D = 0 condition to relate Dt , and D±
n , viz.,
∂
∇ · D(r) = 0 = ∇t · Dt (r) + n̂+ · Dn (r),
∂n
D+ − D−
≈ ∇t · Dt (r) + n̂t · n n
(4.77)
τ
Z Z
I∼=χ dS 0 n̂0t Dt (r0 )g(r, r0 ) + χ dS 0 n̂0+ · D+ 0
n (r )g(r, r )
0
+
St Sn
Z Z
− χτ dS ∇t · Dt (r )g(r, r0 ) − χ
0 0 0
dS 0 n̂0+ · D+ 0 0
n (r )g(r, r ) (4.78)
− +
Sn Sn
10 10
Mie series Mie series
TDS with only tan. current TDS with only tan. current
5 TDS with full current 5 TDS with full current
0 0
−5 −5
σ ( θ , φ = 90o ) / λ20 ( d B )
σ ( θ , φ = 0o ) / λ20 ( d B )
−10 −10
−15 −15
−20 −20
−25 −25
−30 −30
−35 −35
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
θ (deg) θ (deg)
(a) (b)
Figure 4.7: Scattering by a dielectric shell with radius 1.0 m, thickness 0.05 m, εr = 2.6, and
a vertically polarized incident wave from (θ, φ) = (1800 , 00 ) at 0.2 GHz. (a) RCS at φ = 00 .
(b) RCS at φ = 900 .
Note that in the above, we can approximate
∂g
g(r, r0 )|Sn+ − g(r, r0 )|Sn− ' τ (r, r0 )|Sn+ + O(τ 3 ) (4.79)
∂n0

Z Z
I∼
= χτ dl0 n̂0t · Dt (r0 )g(r, r0 ) − χτ dS 0 ∇0t · Dt (r0 )g(r, r0 )
−
C Sn
Z
∂g(r, r0 )
+ χτ dS 0 n̂0t · D+
n (r0 ) (4.80)
Sn + ∂n0
Consequently, the integral equation for TDS can be approximated as

Z
D(r)
i k02 τ
E (r) ' − χ(r)D(r0 )g(r, r0 )dS 0
²0 ²r (r) ²0 S
½ Z Z
∇ 0 0 0 0
+ χτ dl n̂t · Dt (r )g(r, r ) − χτ dS 0 ∇0t · Dt (r0 )g(r, r0 )
²0 C −
Sn
Z 0
¾
0 0 + 0 ∂g(r, r )
+χτ dS n̂ · Dn (r ) (4.81)
+
Sn ∂n0
The above integral equation can be solved by expanding Dt (r0 ) in terms of divergence con-
forming expansion function such as RWG function, and Dn (r0 ) can be expanded in terms
of functions as simple as a pulse function. This integral equation requires less number of
unknowns to solve compared to a full-fledge VIE modeling of a TDS (see Figures 4.7, 4.8).
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
1.5 1.5
−1 1 −1 1
1 0.5 1 0.5
0.5 0 0.5 0
0 −0.5 0 −0.5
−0.5 −1 −0.5 −1
−1 −1.5 −1 −1.5
0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) (b)
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
1.5
−1 1
1 0.5
0.5 0
0 −0.5
−0.5 −1
−1 −1.5
0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1
(c)
Figure 4.8: Current distribution of the case in the previous figure. (a) Tangential current.
(b) Normal current. (c) Full current.
4.7 Impedance Boundary Condition

There have been extensive work done on impedance boundary condition (IBC) [43, 46–50].
The advantage of IBC is that it can greatly simplify the numerical solution of a penetrable
object if the penetrable object can be approximated by IBC.
IBCs are used to model a lossy scatterer. In theory, the scattering solution of a lossy
scatterer can be formulated in terms of SIEs for penetrable objects. When the conductivities
of such objects are high, the integral equations can be approximated using perturbation
arguments.
4.7.1 Generalized Impedance Boundary Condition

For a penetrable object, the integral equations for the exterior region can be written using
either the electric field or the magnetic field, and similarly for the interior region. Because
only current unknowns living on a 2D manifold are solved for, only the tangential components
of these equations are needed. For the exterior region, by taking the tangential component
of (4.4), we have for EFIE
It · [L1E (r, r0 )Js (r0 ) + K1E (r, r0 )Ms (r0 )] = −It · Einc (r), r ∈ S− (4.82)
where It = I − n̂n̂ extracts the tangential component of a 3D vector at r. Similarly, by taking
the tangential component of (4.11), the MFIE is
It · [L1H (r, r0 )Ms (r0 ) + K1H (r, r0 )Js (r0 )] = −It · Hinc (r), r ∈ S− (4.83)
For the interior region, from (4.8), we have for the EFIE
It · [L2E (r, r0 )Js (r0 ) + K2E (r, r0 )Ms (r0 )] = 0, r ∈ S+ (4.84)
From (4.14), we have for the MFIE
It · [L2H (r, r0 )Ms (r0 ) + K2H (r, r0 )Js (r0 )] = 0, r ∈ S+ (4.85)
0 0
In principle, one can find Ms (r ) in terms of Js (r ) from (4.84) or (4.85). Using (4.85),
£ ¤−1
Ms (r0 ) = − It · L2H (r0 , r00 ) It · K2H (r00 , r000 )Js (r000 )
= I(r0 , r000 )Js (r000 ) (4.86)
The above is the formal solution to the integral equation. The actual numerical solution will
be sought by the subspace projection method such as MOM (see previous chapter). The above
equation is preferred because in many subspace projection methods, the matrix representation
of It · L2H is better conditioned than the matrix representation of It · K2H , and hence, its
inversion is more easily sought. Also, when medium 2 is highly conductive, It · L2H becomes
sparse and quasi diagonal, and its inversion is inexpensive.
Eq. (4.86) represents a nonlocal impedance map. It is a generalized impedance boundary
condition (GIBC). It also represents an exact way to solve (4.85). Using the above in (4.82),
we can solve for Js , namely,
h ¡ ¢−1 i−1
Js = − It · L1E − It · K1E It · L2H It · K2H It · Einc (4.87)
The inversion of the above can also be sought by iterative solvers if needed.
4.7.2 Approximate Impedance Boundary Condition

For highly conductive objects, simplifying approximation to (4.86) can be obtained. For
instance, in the case of quasi-planar half-space, in place of (4.86), one may use a simple local
impedance map as an approximation, that is,
n̂ × Ms ≈ Zs Js (4.88)
p
where Zs = η2 ≈ ωµ2 /(iσ2 ) is the surface impedance [51] which is the same as the intrinsic
impedance of medium 2. In fact, the above can be derived from the GIBC. In this manner,
one can solve (4.82) or (4.83) with (4.88) to obtain Js . Again, (4.82) or (4.83) alone may be
fraught with internal resonance problem. In this case, one can combine them and solves the
CFIE to avoid internal resonances.
The above is an approximate method to solve the interior equations. We can also simplify
the exterior equations. As an example, we rewrite (4.82) as
It · Einc (r) + It · L1E (r, r0 )Js (r0 ) = It · K1E (r, r0 )Ms (r0 ), r ∈ S− (4.89)
and we approximate the right-hand side with a local approximation. For example, for a
quasi-planar surface, we may approximate the right-hand side as
1
It · K1E (r, r0 )Ms (r0 ) ≈ − n̂ × Ms (r), r ∈ S− (4.90)
2
This second local approximation, together with our first local approximation, (4.88), yields
the approximation that
1
It · K1E (r, r0 )Ms (r0 ) ≈ − Zs Js , r ∈ S− (4.91)
2
When Zs is very small, and when the above is used in (4.89), the right-hand side of (4.89) is
much smaller than the individual terms on the left hand side. Hence, we may also approximate
(4.89) as
It · Einc (r) + It · L1E (r, r0 )Js (r0 ) ≈ 0, r ∈ S − (4.92)
and solve for Js directly. Then we can use (4.88) to obtain an approximate value for Ms .
One can also easily derive the MFIE equivalence of the above equations.
We shall investigate how good these local approximations are by studying reflection from
a flat surface and reflection from a cylindrical surface.
4.7.3 Reflection by a Flat Lossy Ground Plane

Let us consider first a general reflection of a plane wave by a penetrable half-space (Figure 4.9).
This problem has closed form solution, and the electric field in region 1 can be written as
¡ ¢
E = x̂E1 e−ik1 z + R12 eik1 z (4.93)
E1 ¡ −ik1 z ¢
H = −ŷ e − R12 eik1 z (4.94)
η1
Figure 4.9: Reflection by a conductive half space can be used to derive an impedance boundary
condition.
The above is valid even for a general half space. In general, the tangential E field is not
zero at z = 0. Also, the extinction theorem for z < 0 can be applied exactly for this problem
even if the half space is not conductive.
To apply the extinction theorem, we need Ms 6= 0 and Js 6= 0 in the following equation
Einc + L1E (Js ) + K1E (Ms ) = 0, z<0 (4.95)
We ignore the It · operation because for this 1D problem, the field is naturally tangential. In
the above, the equivalence surface current Js equals the discontinuity of the magnetic field.
Hence, from (4.94)
E1
Js = x̂ (1 − R12 ) (4.96)
η1
Furthermore, Ms equals the discontinuity in the total tangential E field, which in this case,
from (4.93), is
Ms = −ŷE1 (1 + R12 ) (4.97)
In (4.95), L1E (Js ) is the E field produced by an infinite electric current sheet (4.96) extended
over the xy plane, and for z < 0, the electric field it produces is
1
L1E (Js ) = x̂E1 (1 − R12 )e−ik1 z (4.98)
2
Remember that in invoking the extinction theorem, we remove medium 2, and replace it with
medium 1; hence, we use the wavenumber k1 , in (4.98) for z < 0. In (4.95), K1E (Ms ) is the
electric field produced by an infinite magnetic current sheet (4.97), and for z < 0
1
K1E (Ms ) = x̂E1 (1 + R12 )e−ik1 z (4.99)
2
The incident field is
Einc = x̂E1 e−ik1 z (4.100)
Using (4.98), (4.99), and (4.100) in (4.95), we see the extinction theorem is exactly satisfied
for z < 0 for this general penetrable half space problem.
Next, we want to see if the equivalent currents Ms and Js are related in a simple fashion
when the half-space is metallic or when its conductivity is high. To this end, we make a high
conductivity approximation to R12 where
µ2 k1
µ2 k1 − µ1 k2 1− µ1 k2 1 − η2 η1−1
R12 = =− µ2 =− ' −1 + 2η2 η1−1 (4.101)
µ2 k1 + µ1 k2 1+ µ1
k1
k2
1 + η2 η1−1
q q
µi ωµ2
where ηi = ²i . If region 2 is highly conductive, η2 ' iσ2 → 0, σ2 → ∞. Using this in
(4.96) and (4.97), we have
E1
Js ∼
= x̂2 , Ms = −ŷ2E1 η2 η1−1 (4.102)
η1
q
ωµ2
It is clear that n̂ × Ms = Zs Js where Zs = η2 = iσ2 . Hence the surface impedance
approximation, the first local approximation indicated by (4.88) is at least good for this
problem with the appropriate value of Zs .
Furthermore, since
1
K1E (Ms ) = ± n̂ × Ms + P.V.K1E (Ms ) (4.103)
2
We can ignore the P.V.K term: for a flat surface, it is exactly zero, and for a quasi-planar
surface, it should be small. Hence, the second local approximation indicated by (4.90) is also
reasonable. Consequently, (4.95) becomes
1
Einc + L1E (Js ) − n̂ × Ms = 0, z<0 (4.104)
2
If we let n̂ × Ms = η2 Js , then (4.104) can be readily solved for Js as this is a simple 1D
problem. We can also ignore the Ms term in (4.104) and solve for Js and show that it is
correct to leading order in η2 when we assume that η2 is a small parameter.
4.7.4 Reflection by a Lossy Cylinder

Many lossy structures are wire-like or cylinder-like. Hence, it is useful to see how the IBC is
affected by the presence of a curvature as in an infinitely long cylinder.
For simplicity of analysis, we consider a single cylindrical harmonic wave with ∂/∂z = 0
incident on the conducting cylinder. The general penetrable circular cylinder scattering can
be solved exactly (Figure 4.10). We have
h i
E1 = ẑE1 Jn (k1 ρ) + R12 Hn(1) (k1 ρ) einφ (4.105)
E2 = ẑE1 T12 Jn (k2 ρ)einφ (4.106)

i h 0
i
H1φ = E1 Jn0 (k1 ρ) + R12 Hn(1) (k1 ρ) einφ (4.107)
η1
Figure 4.10: Reflection by a lossy cylindrical surface can be used to derive an IBC for a
circular wire.
i
H2φ = E1 T12 Jn0 (k2 ρ)einφ (4.108)
η2
For n 6= 0, the magnetic field may have a nonzero Hρ component, but only the Hφ component
is needed for the ensuing discussions.
To show the extinction theorem in action for ρ < a, we need the surface equivalence
electric and magnetic currents. The surface equivalence electric current Js is
iE1 h 0 0
i
Js = ẑ Jn (k1 a) + R12 Hn(1) (k1 a) einφ = ẑJ0 einφ (4.109)
η1
The surface equivalence magnetic current Ms is

h i
Ms = φ̂E1 Jn (k1 a) + R12 Hn(1) (k1 a) einφ = φ̂M0 einφ (4.110)
To invoke the extinction theorem, we remove medium 2, and investigate the fields produced
by a current sheet Js . If we have Js = ẑJ0 einφ , it can be shown that such an electric current
sheet produces an electric field which is continuous across the current sheet, and a magnetic
field which is discontinuous across the same current sheet. The electric field is given by
(
(1)
η1 J0 πk1 a inφ Jn (k1 a)Hn (k1 ρ), ρ > a
L1E (Js ) = EJ = −ẑ e (1) (4.111)
2 Hn (k1 a)Jn (k1 ρ), ρ < a
The corresponding HJφ is given by

(
(1)0
iJ0 πk1 a inφ Jn (k1 a)Hn (k1 ρ), ρ > a
HJφ =− e (4.112)
2 H (1) (k1 a)Jn0 (k1 ρ), ρ < a
On the other hand, the magnetic current sheet Ms = φ̂M0 einφ will give rise to a discontinuous
electric field and a continuous magnetic field. Namely, we have
( (1)
M0 πk1 a inφ Jn0 (k1 a)Hn (k1 ρ), ρ > a
K1E (Ms ) = EM = −ẑ e (1) 0 (4.113)
2i Hn (k1 a)Jn (k1 ρ), ρ < a
(
(1) 0
M0 πk1 a inφ Jn0 (k1 a)Hn (k1 ρ), ρ > a
HM φ =− e (1) 0
(4.114)
2η Hn (k1 a)Jn0 (k1 ρ), ρ < a
The incident field in this problem is given by the first part of (4.105), viz.,
Einc = ẑE1 Jn (k1 ρ)einφ (4.115)
To show the validity of the extinction theorem, we need to take terms from (4.111) and (4.113)
for ρ < a together with (4.115), and show that
η1 J0 πk1 a (1) M0 πk1 a (1)0
E1 − Hn (k1 a) − Hn (k1 a) = 0 , (4.116)
2 2i
with J0 and M0 given by (4.109) and (4.110). To evaluate them, the reflection coefficient R12
is needed, and it can be derived to be [16]
η1 Jn0 (k2 a)Jn (k1 a) − η2 Jn (k2 a)Jn0 (k1 a)
R12 = − (1) (1)0
(4.117)
η1 Jn0 (k2 a)Hn (k1 a) − η2 Jn (k2 a)Hn (k1 a)
As a result, J0 is given by
iE1 h 0 0
i
J0 = Jn (k1 a) + R12 Hn(1) (k1 a)
η1
−2i
iE1 Jn0 (k2 a) πk 1a
= (1) (1)0
(4.118)
where the Wronskian of Bessel functions has been used, that is,
0 2i
Hn(1) (x)Jn0 (x) − Jn (x)Hn(1) (x) = − .
πx
Similarly, M0 is
h i
M0 = E1 Jn (k1 a) + R12 Hn(1) (k1 a)
−2i
E1 η2 Jn (k2 a) πk 1a
= (1) (1)0
(4.119)
Consequently, we can show by substituting (4.118) and (4.119) that (4.116) is true for a
general penetrable circular cylinder.
From the above, by comparing (4.118) with (4.119), we deduce that
Jn (k2 a)
M0 = −iη2 J0 (4.120)
Jn0 (k2 a)
The above is equivalent to

Jn (k2 a)
n̂ × Ms = −iη2 Js (4.121)
Jn0 (k2 a)
The above is dependent on the the harmonic order n, and hence it is not convenient for use.
It should be applied to the dominant harmonic around a cylinder. It can be shown that when
the radius of the cylinder is small compared to wavelength, the dominant harmonic is the
n = 0 harmonic.
In the limit when k2 corresponds to a highly conductive medium, using the large argument
approximations to the Bessel functions in (4.120) or (4.121), the above becomes M0 = η2 J0
or n̂ × Ms = η2 Js . This result is valid for arbitrary n that is finite. Hence, the equivalence
of (4.82) for a cylinder can be solved by using the above IBCs.
Next we would like to know if the approximation
1
K1E (Ms ) ' − n̂ × Ms , r ∈ S−
2
is still good for a cylinder. From (4.113), by setting ρ = a− , it can be seen that when
k1 a → ∞ with k1 having a small amount of loss, and the large argument approximations are
used for the Hankel and Bessel functions, or the radius of curvature of the cylinder is large
compared to wavelength, this approximation is a good one. This confirms our assertion that
this approximation is good for quasi-planar surfaces. The small amount of loss is needed so
that the Bessel wave, which is a standing wave, can be approximated by an incoming wave
around the cylinder.
Finally, we want to study how the K1E (Ms ) term can be approximated around a circular
wire. Because the monopole or n = 0 component is dominant when k1 a → 0, we may
approximate (4.113) as
K1E (Ms ) ≈ n̂ × Ms C0 (4.122)
where
πk1 a (1) 0
C0 = H0 (k1 a)J0 (k1 a) (4.123)
2i
When k1 a → 0, we can further have the approximation that
C0 ∼ 1, k1 a → 0 (4.124)
Therefore, we see that the local approximation given by (4.122) for a thin wire is in fact quite
different compared to the local approximation of a quasi-planar surface given by (4.90).
To test the idea espoused here, we calculate the input admittance of a copper loop excited
by a delta gap source. The copper loop is 1984 µm by 122 µm, and it has a cross section of 4
µm by 4 µm. For copper, the conductivity is chosen as 5.8e7 s/m. The skin depth decreased
0
10 0.02
GIBC
IBC−WI
0
IBC−MPW
−1
10 FastHenry
−0.02
−0.04
−2
10
Re (Yin) (A)
Im (Yin) (A)
−0.06
−3
10 −0.08
GIBC
−0.1 IBC−WI
−4
IBC−MPW
10 FastHenry
−0.12
−0.14
−5
10
0.01 0.1 1 10 50 0.01 0.1 1 10 50
frequency (GHz) frequency (GHz)
Figure 4.11: Real and imaginary parts of Yin for the copper rectangular loop. In the figure, GIBC is
compared with three other results. IBC-WI refers to IBC based on wave impedance of the conductive
medium. IBC-MPW refers to IBC based on the modified plane wave model introduced in [52].
FastHenry is the code from J. White’s group at MIT [53].
from 20.9 µm at 0.01 GHz to 0.3 µm at 50 GHz. Via the GIBC, the input admittance could be
compared to the one from FastHenry [53], IBC-WI based on wave impedance and IBC-MPW
based on modified plane wave model [52]. Since empirically FastHenry works for electrical
size less than 0.1 wavelength, it will provide accurate results up to about 10 GHz. From the
numerical results in Figure 4.11, the GIBC shows accurate solution for conductor analysis
over a broadband of frequencies. More of this work including numerical examples can be
found in [54].
4.8 Conclusions
In this chapter, we have discussed the formulations of integral equations for penetrable objects.
They can be solved with SIEs. We have shown by Gedanken experiments how some of
the formulations avoid the internal resonance problem commonly encountered in SIEs. We
have also discussed the use of VIEs to solve the inhomogeneous medium penetrable object
case. Then we presented various formulations for TDS. Finally, we have discussed the use of
impedance boundary condition (IBC) and generalized impedance boundary condition (GIBC)
to approximately model lossy penetrable objects.
Bibliography
[1] R. C. Hansen, ed., Geometric Theory of Diffraction, New York: IEEE Press, 1981.
[2] P. C. Waterman, “Scattering by dielectric obstacles,” Alta Freq., vol. 38, pp. 348–352,
1969.
[3] A. C. Ludwig, “The generalized multipole method,” Comput. Phys. Commun., vol. 68,
pp. 306–314, 1991.
[4] M. A. Morgan, ed., Finite Element and Finite Difference Methods in Electromagnetic
Scattering, Amsterdam: Elsevier, 1990.
[5] A. Taflove and K. R. Umashankar, “Review of FDTD numerical modeling of electromag-

netic wave scattering and radar cross-section,” Proc. IEEE, vol. 77, pp. 682–699, 1989.
[6] P. P. Silvester and R. L. Ferrari, Finite Elements for Electrical Engineers, Cambridge,
MA: Cambridge University Press, 1986.
[7] J. M. Jin, The Finite Element Method in Electromagnetics, New York: Wiley, 1993.
[8] J. L. Volakis, A. Chatterjee and L. C. Kempel, Finite Element Method for Electromagnet-
ics: Antennas, Microwave Circuits, and Scattering Applications, New York: IEEE Press
and Oxford University Press, 1998.
[9] R. F. Harrington, Field Computation by Moment Methods, Malabar, FL: Krieger, 1968.
[10] J. Richmond, “Scattering by a dielectric cylinder of arbitrary cross section shape,” IEEE
Trans. Antennas Propag., vol. 13, no. 3, pp. 334-341, May 1965.
[11] J. Richmond, “TE-wave scattering by a dielectric cylinder of arbitrary crosssection

shape,” IEEE Trans. Antennas Propag., vol. 14, no. 4, pp. 460-464, Jul. 1966.
[12] A. J. Poggio and E. K. Miller, “Integral equation solutions of three-dimensional scat-

tering problems,” in Computer Techniques for Electromagnetics, R. Mittra, Ed., 2nd ed.,
Hemisphere, New York, 1987, pp. 159–264.
[13] E. K. Miller, L. Medgyesi-Mitschang, and E. H. Newman, Ed.,Computational Electro-

magnetics: Frequency-Domain Method of Moments, New York: IEEE Press, 1992.
1
[14] L. N. Medgyesi-Mitschang, J. M. Putnam, and M. B. Gedera, “Generalized method of

moments for three-dimensional penetrable scatterers,” J. Opt. Soc. Am. A, vol. 11, pp.
1383–1398, Apr. 1994.
[15] R. F. Harrington, Time-Harmonic Electromagnetic Fields, New York: McGraw-Hill,

1961.
[17] W. C. Chew, J.-M. Jin, E. Michielssen, and J. M. Song, Fast and Efficient Algorithms
in Computational Electromagnetics, Boston: Artech House, 2001.
[18] J. R. Mautz and R. F. Harrington, “H-field, E-field, and combined-field solutions for
conducting body of revolution,” AEÜ, vol. 32, pp. 157–164, Apr. 1978.
[19] R. F. Harrington, “Boundary integral formulations for homogenous material bodies,” J.

Electromagn. Waves and Appl., vol. 3, no. 1, pp. 1-15, 1989.
[20] C. Müller, Foundation of the Mathematical Theory of Electromagnetic Waves, Berlin:

Springer-Verlag, 1969.
[21] P. Werner “On the exrerior boundary value problem of the perfect reflection for the
stationary electromagnetic wave fields.” J. Math. Anal. Appl., vol. 7, pp. 348–396, 1963.
[22] H. A. Schenck, “Improved integral formulation for acoustic radiation problems,” J.

Acoust. Soc. Am., vol. 44, pp. 41–58, 1968.
[23] K. M. Mitzner, “Numerical solution of the exterior scattering problem at the eigenfre-
quencies of the interior problem.” URSI Meeting Digest, pp. 75, Boston, 1968.
[24] A. J. Burton and G. F. Miller, “The application of integral equation methods to the
numerical solution of some exterior boundary value problems.” Proc. R. Soc., Lond. A323,
pp. 201–210, 1971.
[25] J. C. Bolomey and W. Tabbara, “Numerical aspects on coupling between complementary

boundary value problems.” IEEE Trans. Antennas Propag., vol. 21, pp. 356–63, 1973.
[26] R. Mittra and C. A. Klein, “Stability and convergence of moment method solutions,”
In Numerical and Asymptotic Techniques in Electromagnetics, Ed. R. Mittra. New York:
Springer-Verlag, 1975.
[27] N. Morita, “Resonant solutions involved in the integral equation approach to scattering
from conducting and dielectric cylinders.” IEEE Trans. Antennas Propag.27, pp. 869–871,
1979.
[28] A. D. Yaghjian, “Augmented electric- and magnetic-field integral equations,” Radio Sci.,
vo1. 16, pp. 987–1001, 1981.
[29] A. F. Peterson, “The interior resonance problem associated with surface integral equa-
tions of electromagnetics: numerical consequences and a survey of remedies,” Electro-
magn., vol. 10, pp. 293–312, 1990.
[30] W. C. Chew and J. M. Song, “Gedanken experiments to understand the internal reso-
nance problems of electromagnetic scattering,” Electromagn., vol. 27, no. 8, pp. 457–472,
Nov.-Dec. 2007.
[31] J. M. Song, W. W. Shu, and W. C. Chew, “Numerical resonances in method of moments,”

IEEE APS Int. Symp. Dig., Honolulu, Hawaii, pp. 4861-4864, Jun. 2007.
[32] D. R. Wilton, “Review of current status and trends in the use of integral equations in
computational electromagnetics,” Electromagn., vol. 12, pp. 287-341, 1992.
[33] J. R. Mautz and R. F. Harrington, “A combined-source solution for radiation and scat-
tering from a perfectly conducting body,” IEEE Trans. Antennas Propag., AP-27, no. 4,
pp. 445–454, Jul. 1979.
[34] J. C. Monzon and N. J. Damaskos, “A scheme for eliminating internal resonances: the
parasitic body technique,” IEEE Trans. Antennas Propag., vol. 42, no. 8, pp. 1089–1091,
1994.
[35] K. F. Warnick and W. C. Chew, “Error analysis of surface integral equation methods,”
in Fast and Efficient Algorithms in Computational Electromagnetics, eds. W. Chew, J.
Jin, E. Michielssen, J. Song, Boston: Artech House, 2001.
[36] H. Gan and W. C. Chew, “A discrete BCG-FFT algorithm for solving 3D inhomogeneous
scatterer problems,” J. Electromagn. Waves and Appl., vol. 9, no. 10, pp. 1339-1357, 1995.
[37] J. C. Nedelec, “Mixed finite elements in R3 ,” Numer. Math., vol. 35, pp. 315–341, 1980.
[38] A. Bossavit: “On Finite Elements for the Electricity Equation,” in The Mathematics of
Finite Elements (J.R. Whiteman, ed.), London: Academic Press, pp. 85–92, 1982.
[39] W. C. Chew and L. E. Sun, “A novel formulation of the volume integral equation for
electromagnetic scattering,” Tenty-Third Ann. Rev. Prog. Appl. Comput. Electromag.,
Verona, Italy, Mar. 19–23, 2007.
[40] R. D. Graglia, D. R. Wilton, and A. F. Peterson, “Higher order interpolatory vector

bases for computational electromagnetics,” IEEE Trans. Antennas Propag., vol. 45, no.
3, pp. 329–342, Mar. 1997.
[41] R. F. Harrington and J. R. Mautz, “An impedance sheet approximation for thin dielectric
shells,” IEEE Trans. Antennas Propag., vol. AP-23, pp. 531-534, Jul. 1975.
[42] E. H. Newman and M. R. Schrote, “An open surface integral formulation for electro-
magnetic scattering by material plates,” IEEE Trans. Antennas Propag., vol. AP-32, pp.
672-678, Jul. 1984.
[43] E. Bleszynski, M. Bleszynski, and T. Jaroszewicz, “Surface-integral equations for electro-

magnetic scattering from impenetrable and penetrable sheets,” IEEE Antennas Propag.
Mag., vol. 35, pp. 14-25, Dec. 1993.
[44] I. T. Chiang and W. C. Chew, “Thin dielectric sheet simulation by surface integral
equation using modified RWG and pulse bases,” IEEE Trans. Antennas Propag., vol. 54,
no. 7, pp. 1927–1934, Jul. 2006.
[45] I. T. Chiang and W. C. Chew, “A Coupled PEC-TDS Surface Integral Equation Ap-
proach for Electromagnetic Scattering and Radiation from Composite Metallic and Thin
Dielectric Objects,” IEEE Trans. Antennas Propag., vol. 54, no. 11, pp. 3511–3516, Nov.
2006.
[46] S. W. Lee and W. Gee, “How good is the impedance boundary condition?” IEEE Trans.
Antennas Propag., vol. AP-35, pp. 1313–1315, Nov. 1987.
[47] T. B. A. Senior and J. L. Volakis, “Derivation and application of a class of generalized
impedance boundary conditions,” IEEE Trans. Antennas Propag., vol. AP-37, pp. 1566–
1572, Dec. 1989.
[48] D. J. Hoppe and Y. Rahmat-Samii, “Scattering by superquadric dielectric coated cylin-
ders using high order impedance boundary conditions,” IEEE Trans. Antennas Propag.,
vol. 40, pp. 1513–1523, Dec. 1992.
[49] O. Marceaux and B. Stupfel, “High-order impedance boundary condition for multilayer
coated 3-D objects,” IEEE Trans. Antennas Propag., vol. 48, pp. 429–436, Mar. 2000.
[50] B. Stupfel and Y. Pion, “Impedance boundary conditions for finite planar and curved
frequency selective surfaces,” IEEE Trans. Antennas Propag., vol. 53, pp. 1415–1425 Apr.
2005.
[51] J. A. Stratton, Electromagnetic Theory, New York: McGraw-Hill, 1941.
[52] B.-T. Lee, “Efficient Series Impedance Extraction Using Effective Internal Impedance,”
Ph.D. Thesis, University of Texas at Austin, Aug. 1996.
[53] M. Kamon, M. J. Tsuk, and J. White, “FastHenry: A multipole-accelerated 3-D in-

ductance extraction program,” IEEE Trans. Microwave Theory Tech., vol. 42, no. 9, pp.
1750-58, Sept. 1994.
[54] Z. G. Qian, W. C. Chew, and R. Suaya, “Generalized impedance boundary condition for
conductor modeling in surface integral equation,” IEEE Trans. Microwave Theory Tech.,
vol. 55, no. 11, pp. 2354–2364, Nov. 2007.
107
CHAPTER 5
Low-Frequency Problems in
Integral Equations
5.1 Introduction
The development of a number of technologies has called for the solutions of complex structures
when the structures are small compared to wavelength, and yet the complexity of these
structures cannot be ignored. These are the area of high-speed circuit in computer chip
and package designs, the area of small antennas where the antennas can be a fraction of
a wavelength, but its complex design is important, as well as the area of nanotechnology
and metamaterial. These complex subwavelength structures are richly endowed with physics,
especially when they are coupled with nonlinear devices such as transistors and diodes. Their
microscopic designs manifest themselves macroscopically in a vastly different manner. An
example of such is in radio frequency identification (RFID) tag [1].
One can relate the richness of the physics that is buried in microcircuits as due to the
singularity of the dyadic Green’s function ( [2, Chap. 7]). Such a singularity is not found for
instance in the elastodynamic dyadic Green’s function [3], marking a contrast between micro
devices in the mechanical world and those in the electrical world.
It is important to solve the electromagnetic equations in multiscales, and yet preserve
the microscale physics, as it can be manifested on a macroscopic scale. This is unlike some
area where the microscale physics can be homogenized for their connection to the macroscale
physics. In this chapter, we discuss the solutions of integral equations when the wavelength
is much larger than the dimensions of the structures.
5.2 Low-Frequency Breakdown of Electric Field Integral

Equation
The electric field integral equation (EFIE) suffers from a low-frequency breakdown prob-
lem [4–11]. This problem, however, is not specific to integral equations. It is also germane to
1
differential equation solutions of electromagnetics. In other words, a numerical electromag-

netic solver, when written for mid frequency, may not work well for low frequency.
We can foresee this problem by looking at the electromagnetic equations when ω → 0 [11].
Right at zero frequency, they become
∇ × E = 0, (5.1)
∇ × H = J, (5.2)
∇·J
∇ · ²E = ρ = limω→0 ıω , (5.3)
∇ · µH = 0. (5.4)
Notice that from (5.2), the current J that produces the magnetic field must be divergence
free. However, the current J that produces the charge ρ in (5.3) cannot be divergence free,
but in order to keep the right-hand side of (5.3) bounded, this J ∼ O(ω), when ω → 0.
Hence we see the total J naturally decomposes into a solenoidal (divergence free) part and a
irrotational (curl-free) part. Namely
J = Jsol + Jirr (5.5)
This is also known as the Helmholtz decomposition [12]. Moreover, the curl-free part, Jirr ∼
O(ω) when ω → 0 as alluded to above, and hence,
|Jirr | ¿ |Jsol | , ω→0 (5.6)
If the total J is the unknown current to be solved for, the above poses a severe numerical
problem.1 One may think that perhaps Jirr is not important when ω → 0, since it is so much
smaller than Jsol . However, this is not the case. The electromagnetic equations separate into
the magnetoquasistatic equations and the electroquasistatic equations. In circuit theory, the
magnetoquasistatic equations are important for describing the world of inductors, whereas
the electroquasistatic equations are important for the world of capacitors. Both are equally
important in capturing circuit physics. Hence, the accuracy of one cannot be compromised
over the other. The solenoidal current Jsol represents eddy currents that produce primarily
the magnetic field at low frequency, whereas the irrotational current Jirr represents charge
currents that produce primarily the electric field at low frequency.
This problem can also be seen when we look at the EFIE operator (see the previous
chapter),
Z Z
1
LE Js = iωµ dS 0 g(r, r0 )Js (r0 ) + ∇ dS 0 g(r, r0 )∇0 · Js (r0 ) (5.7)
S iω² S
where Js is the surface current on the surface S. The first term is due to the vector potential
or it corresponds to the electric field generated by a time-varying magnetic field. We shall
call it the induction term. The second term is due to the scalar potential, or it corresponds
to the electric field produced by the charge in the system. We shall call it the charge term.
1 The disparity between the magnitudes of J
irr and Jsol remains true if the above argument is repeated
for a conductive medium.
LOW-FREQUENCY PROBLEMS IN INTEGRAL EQUATIONS 109
When ω → 0, the second term dominates over the first term, and we have
Z
1
LE Js ' ∇ dS 0 g(r, r0 )∇0 · Js (r0 ) (5.8)
iω² S
Since there exists Js (r0 ) such that ∇0 · Js (r0 ) = 0, that is, when Js = Jsol , the above integral
operator has a null space. The eigenvalues of the operator LE in (5.7) flip-flop between
very large values for Js = Jirr when the charge term dominates, and very small eigenvalues
for Js = Jsol when the induction term dominates. Hence, the matrix representation of the
above operator becomes extremely ill-conditioned when ω → 0. This is the reason for the
low-frequency breakdown of the EFIE operator.
Another point of view of the low-frequency breakdown is to look at the differential equa-
tion. One can convert the electromagnetic equations to a vector wave equation for a lossless
inhomogeneous medium as
∇ × µ−1 ∇ × E − ω 2 ²E = iωJ (5.9)
When ω → 0, the above becomes
∇ × µ−1 ∇ × E ' iωJ (5.10)
Eq. (5.10) has no unique solution because ∇ × A = b does not guarantee a unique A because
A0 = A + ∇φ will equally satisfy ∇ × A0 = b. In other words, the curl operator has a
null space. Hence, the matrix representation of the differential operator in (5.10) yields a
badly conditioned system. The breakdown of the electromagnetic equations when ω → 0 is
also because that ω = 0 is mathematically a singular perturbation point for solutions to the
electromagnetic equations [13, p. 159].
The fact that ω = 0 is a singular perturbation point is both a curse as well as a blessing.
A curse because it makes the development of robust numerical solvers in this regime (which
is referred to as the twilight zone [14, 15]) difficult. It is a blessing because it is this difficulty
that enriches the world of circuit physics. In other words, it is not simple to transition from
the world of wave physics to the world of circuit physics. If one transitions smoothly from the
world of wave physics to the world of the circuit physics, however, the richness in the world
of circuit physics will be revealed. This is almost like the Alice in Wonderland story!
5.3 Remedy—Loop-Tree Decomposition and Frequency

Normalization
The remedy to this low-frequency breakdown is to solve the problem in such a way so that
the world of inductors is separated from the world of the capacitors. That is, we have to
forcefully decompose the current into two parts by performing a Helmholtz decomposition.
In the preceding discussion, it is seen that the charge term is responsible for producing
instability of the EFIE operator depending on if it is acting on a divergence-free current or
nondivergence-free current. To stabilize it, it is necessary to decompose the current into a
divergence-free part and the nondivergence-free part [4, 5]. This is not a full Helmholtz de-
composition, but a quasi-Helmholtz decomposition current. The Rao-Wilton-Glisson (RWG)
basis2 is amenable to the aforementioned quasi-Helmholtz decomposition, but not a full
2 We use the word basis here to mean a set of functions that spans a space.
Helmholtz decomposition.3
Instead of using RWG functions as a basis to span the approximate linear vector space,
we decompose the RWG basis into the loop basis whose members have zero divergence, and
the tree basis (or star basis) whose members have nonzero divergence. This is known as the
loop-tree or loop-star decomposition, which is a quasi-Helmholtz decomposition because the
tree or the star expansion functions are not curl-free. We shall call the space spanned by
RWG basis the RWG space, the subspace spanned by the tree (or star) basis the tree (or
star) space, and that spanned by loop basis the loop space.4
Consequently, we expand [11]
NL
X NC
X
Js (r0 ) = ILn JLn (r0 ) + ICn JCn (r0 ) (5.11)
n=1 n=1
where JLn (r0 ) is a loop expansion function such that ∇ · JLn (r0 ) = 0, and JCn (r0 ) is a tree
expansion (or a star expansion) function such that ∇ · JCn (r0 ) 6= 0 and it is used to model
the charge in the system. We write the above compactly as
t t
Js (r0 ) = JL (r0 ) · IL + JC (r0 ) · IC (5.12)
where £ ¤
JL (r0 ) n = JLn (r0 ), [IL ]n = ILn (5.13a)
£ ¤
JC (r0 ) n = JCn (r0 ), [IC ]n = ICn (5.13b)
When the above is substituted into the EFIE,
LE Js = −Einc (5.14)
and testing the above with the same set of functions as in Galerkin’s method, we have a
matrix equation · ¸ · ¸ · ¸
ZLL ZLC IL VL
· = (5.15)
ZCL ZCC IC VC
where
VL = −hJL (r), Einc (r)i (5.16a)
VC = −hJC (r), Einc (r)i (5.16b)
t
ZLL = iωµhJL (r), g(r, r0 ), JL (r0 )i (5.16c)
t
ZLC = iωµhJL (r), g(r, r0 ), JC (r0 )i (5.16d)
3 It is interesting to note that the stabilizing of the differential equation (5.9) is done via a quasi-Helmholtz
decomposition where the field is decomposed into a curl-free and the noncurl-free part [4].
4 This is also known as the tree-cotree decomposition in some work [16].
t t
ZCL = iωµhJC (r), g(r, r0 ), JL (r0 )i = ZLC (5.16e)
t
ZCC = iωµhJC (r), g(r, r0 ), JC (r0 )i
i
− ω² h∇ · JC (r), g(r, r0 ), ∇0 · JC (r0 )i (5.16f)
In this case, the integral is over the supports of the expansion and testing functions.
When ω → 0, the frequency-scaling behavior of the matrix system is as follows:
· ¸ · ¸ · ¸
ZLL (O(ω)) ZLC (O(ω)) IL VL
· = (5.17)
ZCL (O(ω)) ZCC (O(ω −1 )) IC VC
where O(ω α ) in the parentheses represents the scaling properties of the matrix elements with
respect to frequency ω when ω → 0. Hence, even after the loop-tree decomposition, the
matrix system is still ill conditioned because of the imbalance of the matrix elements; some
blocks have small values, whereas the lower right block is inordinately large when ω → 0.
One still needs to frequency-normalize the matrix system in order to obtain a stable system.
However, one should note that this determinate frequency scaling is because of the quasi-
Helmholtz decomposition, which isolates the frequency-scaling behavior into different blocks.
This makes the block matrices amenable to frequency normalization.
To this end, we multiply the first row of the equation by ω −1 , and rearrange the matrix
system such that
· −1 ¸· ¸ · −1 ¸
ω ZLL (O(1)) ZLC (O(ω)) IL ω VL
= (5.18)
ZCL (O(ω)) ωZCC (O(1)) ω −1 IC VC
As can be seen, now the diagonal blocks are of O(1) and the off diagonal blocks are of O(ω)
when ω → 0. Now the system is diagonally dominant and hence is better conditioned.
Frequency normalization can be regarded as a kind of post- and prediagonal precondition-
ing such that · −1 ¸ · ¸ · ¸ · −1 ¸
ω 0 1 0 1 0 ω 0
·Z· · ·I= ·V (5.19)
0 1 0 ω 0 ω −1 0 1
or
−1
P P R · Z · PP O · PP O · I = PP R · V (5.20)
where PP R and PP O , the pre- and the postpreconditioning matrices, can be identified by
comparing (5.19) with (5.20).
The above frequency normalization boosts the importance of the vector potential term,
or the ZLL matrix elements, and suppresses the divergence of the ZCC matrix elements. It
also boosts the importance of the IC and VL . These two terms tend to be vanishingly small
from our discussion in Eq. (5.6) when ω → 0.
In the original EFIE, as ω → 0, it becomes an electroquasistatic system where the electric
field dominates over the magnetic field. Frequency normalization makes the magnetic field
equally important compared to the electric field, and hence, the magnetoquasistatic field
is solved to the same accuracy as the electroquasistatic field. Consequently, the physics of
inductors is captured with the same accuracy as the physics of capacitors.
To elucidate the physics of the magnetoquasistatic part in the above testing procedure,
since ∇ · JLi = 0 for the loop expansion function, we can express JLı (r) = ∇ × n̂ψi (r). Then
an element of the ZLL matrix in (5.18) becomes
£ ¤
ZLL ij = iωµh∇ × n̂ψi (r), g(r, r0 ), JLj (r0 )i
= iωµhn̂ψi (r), ∇ × Ig(r, r0 ), JLj (r0 )i
= iωµhn̂ψi (r), Hj (r)i (5.21)
where integration is over the r variable, and Hj (r) is the magnetic field produced by JLj (r).
Hence, testing by the loop function is equivalent to testing the normal component of the
magnetic field. In solving a magnetoquasistatic problem, it is sufficient to impose the bound-
ary condition that the normal component of the magnetic field be zero on a metallic surface.
Consequently, the upper left block reduces to a magnetoquasistatic problem for low frequency.
Similarly, from (5.16f), it is seen that the lower right block reduces to an electroquasi-
static problem when ω → 0. Moreover, the electrostatic problem is decoupled from the
magnetostatic problem at ω = 0. This point is underscored in the Eqs. (5.1), (5.2), (5.3),
and (5.4) in the beginning of this chapter.
It is to be noted that right when ω = 0, the problem decomposes into two uncoupled
electrostatic and magnetostatic problems. These problems are important for describing the
worlds of capacitors and inductors respectively, and hence are important for circuit physics.
When the frequency increases, these two worlds start to couple to each other, and it often
is important to include the physics of these two worlds concurrently. We shall call this the
regime of full-field physics, since both the E and H fields are involved, but the regime of
full-wave physics has not fully set in yet. It is in fact found that the loop-tree decomposition
is deleterious in modeling full-wave physics [17]. When the regime of full-wave physics sets
in, the RWG functions are better suited to model this regime.
5.3.1 Loop-Tree Decomposition

The loop-tree decomposition of a subspace spanned by RWG expansion functions or rooftop
expansion functions can be found quite easily [18–21]. We will discuss the case whose surface
is triangulated and that the approximate subspace for the current is spanned by the RWG
basis.
There are two cases to consider, the open surface and the closed surface. In the open
surface case, we first introduce cuts on the triangulated surface (see Figure 5.1). The cuts
should be such that the triangulated surface remains simply connected. Namely, if one thinks
of the cuts as being introduced on a sheet of paper, the paper should not be broken into two or
more pieces. Also, the number of cuts should be taken such that no loops can be formed from
the remaining RWG expansion functions on the surface. An RWG function is then removed
from the edges that have been cut. Then associated with each cut, one introduces a loop
expansion function. Hence, for every RWG function removed, a loop function is introduced.
For a simple open surface, the number of cuts is equal to the number of interior nodes of the
surface. Hence, the number of loops is equal to the number of interior nodes NN int .
Figure 5.1: Loop-tree decomposition for an open surface. The serrated cuts represent RWG
expansion functions removed from the edges, and the loops represent loop expansion function
around a node.
For a simple closed surface, every node is an interior node. However, there are only
Nnode −1 loops where Nnode is number of nodes on the triangulated geometry (see Figure 5.2).
If one adds a loop to every node on a closed surface, then the loop added to the last (final)
node is dependent on the previous loops added. One can see this for a tetrahedron, which is
the simplest triangulated closed surface (see Figure 5.2). A loop is dependent on other loops
if it can be formed by a linear superposition of these other loops.
In addition to simple loops described above, there are global loops. Global loops are
distinguished by their irreducibility to a loop around a single node. For instance, if we have
a ribbon, there will be a loop around a ribbon, which cannot be reduced to a loop around
a single node. A ribbon is topologically the same as an open surface with a hole in it (see
Figure 5.3). For a surface with two holes, the number of global loops is two. The number of
global loops is proportional to the number of holes in an open surface.
A closed surface with a hole is equivalent to a doughnut (see Figure 5.4). For a doughnut-
shape object, the number of global loops is two, as can be seen from the figure. If we add one
more hole to the doughnut, the number of global loops increases to four, and so on. When
more than one global loop exist, they may be entangled and special care has to be taken to
account for them [22].
5.3.2 The Electrostatic Problem

Even though the lower right block reduces to the electrostatic problem when ω → 0, the matrix
is unlike the capacitance matrix from solving an electrostatic problem. When iterative solvers
are applied to the above equation, convergence is still slow. To gain further insight, we will
review the electrostatic problem here.
Figure 5.2: Loop-tree decomposition for a closed surface. A tetrahedron represents the sim-
plest triangulated closed surface. The number of loops is one less than the number of nodes.
Figure 5.3: When an open surface has a hole, a global loop is formed.
An electrostatic boundary value problem can be cast as [23]
∇2 φ(r) = s(r) (5.22)

φ(r) = 0, r ∈ S
If s(r) is not charge neutral, we may assume that there are charges at infinity that neutralize
the net charge of s(r). The integral equation for the above can be written as
Z
φ(r) = φ0 (r) + dr0 g0 (r, r0 )σ(r0 ) = 0, r∈S (5.23)
S
1
where σ(r0 ) is the surface charge on S, g0 (r, r0 ) = 4π|r−r 0 | is the static Green’s function, and
φ0 (r) is the potential generated by the source s(r). The above can be converted into a matrix
equation by letting
X N
0
σ(r ) = Qn σn (r0 ) (5.24)
n=1
Figure 5.4: When a closed surface has a hole in it forming a doughnut, two global loops are
formed.
where σn (r0 ) is an expansion function for expanding the charge and N is the number of
expansion functions used. Using (5.24) in (5.23), and testing the result by σm (r), m =
1, . . . , N, and integrating over S, and finally, imposing that the total φ = 0 on S gives
N
X
−hσm , φ0 i = Qn hσm , g0 , σn i (5.25)
n=1
The above is a matrix equation

A·Q=V (5.26)
−1
Since in a capacitive system, Q = C·V, the matrix A = C is the inverse of the capacitance
matrix, and is given by £ ¤
A mn = hσm , g0 , σn i (5.27)
[Q]n = Qn , [V]m = −hσm , φ0 i (5.28)
The above matrix equation is easily proven to be Hermitian positive definite, with good
convergence property when solved with iterative solvers.
When σm (r) is chosen to be a pulse expansion function, we expect Eq. (5.26) to be related
to the frequency normalized ZCC . However, we find that
£ ¤ i
lim ωZCC mn = − h∇ · Jm (r), g0 (r, r0 ), ∇ · Jn (r0 )i (5.29)
ω→0 ²
When Jm (r) is chosen to be an RWG basis, ∇ · Jm (r) does not become a pulse function,
but a pulse pair of opposite signs (see Figure 5.5). We shall call this the pulse-pair function.
Hence, the matrix in (5.29) is the matrix representation of the integral operator in the sub-
space spanned by pulse-pair functions. Moreover, each of those pulse-pair functions is charge
neutral, namely, the total integral sum of the charge is zero.
Expanding σ(r) in terms of pulse expansion functions, we have
Np
X
σ(r) = Qn Pn (r) (5.30)
n=1
Figure 5.5: A pulse-pair function resulting from taking the divergence of the RWG function.
where Np is the number of pulse functions used. We can connect this with the charge that is
expanded in terms of the pulse-pair functions, namely,
NC
X
σ(r) = In ∇ · Λn (r) (5.31)
n=1
where Λn (r) represents an RWG function, In is the unknown coefficient, and NC is the
number of the subset of the RWG functions that span the tree space, R the subspace of non-
divergence-free functions
R complement to the loop space. In (5.31), S
σ(r)dS = 0, or it is
charge neutral because S ∇ · Λn (r)dS = 0. Hence, if we impose charge neutrality on (5.30),
we can make the connection between Qn and In , because ∇ · Λn (r) is the superposition of
two pulse functions. This connection is necessary to improve on the convergence property of
the frequency normalized equation (5.18).
5.3.3 Basis Rearrangement

The above method so far gives rise to a matrix system with no null space when ω → 0,
but the matrix system is still not well-conditioned enough for iterative solvers. The basis
rearrangement method [11] has to be invoked to make the solution of the matrix system more
amenable to iterative solvers. The basis rearrangement relates the electroquasistatic part of
(5.18) to (5.26), the matrix equation of an electrostatic
R problem.
If we assume that Pn (r) is normalized such that S dSPn (r) = 1, then the charge neutrality
condition for (5.30) is
Np
X
Qn = 0 (5.32)
n=1
or
Np −1
X
QNp = − Qn (5.33)
n=1
Hence, using (5.33), (5.30) can be written as

Np −1
X
σ(r) = Qn [Pn (r) − PNF (r)] (5.34)
n=1

R
where now σ(r) is charge neutral, that is, S σ(r)dS = 0. Charge neutrality is true, because
the pulse functions are the consequence
R of splitting an RWG function into two, but each RWG
function is charge neutral, namely, S ∇ · Λn (r)dS = 0. We can express the above as
σ(r) = Nt (r) · Q (5.35)
where Nt (r) and Q are vectors of length Np − 1.
£ t ¤
N (r) n = Pn (r) − PNF (r), n = 1, . . . , Np − 1 (5.36)
[Q(r)]n = Qn , n = 1, . . . , Np − 1 (5.37)
From the continuity equation, we note that
∇ · Js = iωσ(r) (5.38)
We can apply this to the surface current and surface charge density on the surface of a perfect
electric conduction (PEC).
At low frequency, the solenoidal part of Js does not play a role in generating charges, and
hence, only the tree current plays a role in (5.38). The tree or charge current is related to
the NC tree RWG function as
t
JC = JC (r) · IC (5.39)
t
where JC (r) is a 3 × NC matrix and IC is an NC × 1 vector. Using (5.39) in (5.38), we have
t
∇ · JC (r) · IC = iωNt (r) · Q (5.40)
t
where ∇ · JC (r) is a 1 × NC vector. Taking the inner product of (5.40) with P(r), where
[P(r)]n = Pn (r), n = 1, . . . , Np − 1, we have
t
hP(r), ∇ · JC (r)i · IC = iωhP(r), Nt (r)i · Q (5.41)
In this manner, hP(r), Nt (r)i is the identity matrix because Pn (r) is normalized. Conse-
quently, (5.41) becomes
K · IC = iωQ (5.42)
where K is a square matrix given by
t
K = hP(r), ∇ · JC (r)i (5.43)
and more explicitly, £ ¤
K mn = hPm (r), ∇ · Λn (r)i (5.44)
K is an (NC − 1) × (NC − 1) sparse matrix dense to O(NC ) and can be inverted in O(NC )
operations. Eq. (5.42) implies that IC ∼ O(ω), when ω → 0, as is required in our earlier
discussion.
Consequently, the original method of moments (MOM) matrix (5.18) can be written in a
form more suggestive of circuit theory, namely,
" #· ¸ " 1 #
LLL iωLLC IL VL
= ³ t ´−1
iω
−1 (5.45)
iωLCL CCC Q K · VC
where
1
LLL = ZLL , (5.46)
iω
1 −1
LLC = ZLC · K , (5.47)
iω
1 t t
LCL = (K )−1 · ZCL = LLC , (5.48)
iω
−1
³ t ´−1 −1
CCC = iω K · ZCC · K (5.49)
and LLL is the self-inductance matrix represented in the loop basis, LLC has a unit of induc-
tance, and iωLLC can be regarded as the mutual impedance matrix between the inductance
and the capacitance, and CCC is the self-capacitance matrix represented in the tree or charge
basis.
Because inductance is proportional to µ and capacitance is proportional to ², and in MKS
units, µ and ² are vastly different in numerical values, the corresponding diagonal blocks in
(5.45) are vastly different. A remedy is to express (5.45) in dimensionless form, or multiply
the top row in (5.45) by µ−1 , and the bottom row by ² to further normalize the system.
For acceleration purposes, the final equation may be written in terms of a sequence of
matrix-vector products. For instance, we can rewrite (5.45) as
" # " # · ¸ " 1 #
I 0 I 0 IL VL
³ t ´−1 · Ft · Z ³ ´
iω
RWG · F · −1 · = t −1 (5.50)
0 K K Q K · VC
When an iterative solver is used to solve the matrix equation, the matrix-vector product
can be decomposed into a sequence of matrix-vector products as above. A fast matrix-vector
product can be written for ZRWG ·a, and be used to accelerate the total matrix-vector product.
In the above, ZRWG is a conventional impedance matrix constructed from using RWG
functions as a basis, but it can represent a matrix system obtained by using other basis set.
F is a transformation matrix that changes from the conventional basis to the loop-tree basis.
It is a sparse matrix with O(N ) elements and hence, its associated matrix-vector product
can be performed in O(N ) operations. The frequency and matrix normalization can be
incorporated in F. The construction of the F matrix requires the knowledge of the loop-tree
decomposition of a structure. This can be formed by loop-tree search algorithms [18–21].
The matrix-vector product associated with ZRWG · a can be accelerated by a low-frequency
fast multipole algorithm (LF-FMA) [24] requiring O(N ) operations (for simplicity, we use
−1
the acronym LF-FMA to replace LF-MLFMA). Hence, if K · Q can be effected in O(N )
operations, the whole matrix vector product can be effected in O(N ) operations.
−1
5.3.4 Computation of K · Q in O(N ) Operations
−1
As discussed above, it is imperative to effect K · Q in O(N ) operations. Otherwise it will
−1
be the bottleneck in the entire matrix-vector product. Applying K · Q in O(N ) operations
is equivalent to solving
K · I = iωQ (5.51)
in O(N ) operations. We regroup IC , the nondivergence-free current into a part that involves
junctions, IC1 , and the rest that does not involve junctions, IC2 . Hence, (5.51) can be written
as · ¸ · ¸ · ¸ · ¸
Q1 IC1 A1 B1 IC1
iω =K· = · (5.52)
Q2 IC2 A2 B2 IC2
Then by the matrix partitioning method,
" −1 −1 −1 −1 −1 −1
#
−1 A1 + A1 · B1 · D2 · A2 · A1 , −A1 · B1 · D1
K = −1 −1 −1 (5.53)
−D2 · A2 · A1 , D2
In the above, the dimension of A1 is much smaller than that of B2 , and D2 is a small matrix
given by
−1 −1
D2 = B2 − A2 · A1 · B2 (5.54)
−1 −1
Hence, the main cost of K · Q is the action of A1 on a vector y on finding
−1
x = A1 · y (5.55)
which is similar to solving

y = A1 · x (5.56)
³ t ´−1
Similarly, when we need to calculate the action of K on a vector, we need to calculate
the action of ³ t ´−1
A1 ·x=y (5.57)
5.3.5 Motivation for Inverting K Matrix in O(N ) Operations

The connection matrix K that connects charge due to current basis such as RWG to the
pulse basis for a triangle is a sparse matrix system. Many sparse matrices, such as those in
finite element method (FEM), cannot be inverted in O(N ) operation. The cost of inverting
2
sparse matrices arising from FEM is O(N Bw ) where Bw is the bandwidth of the matrix.
Surprisingly, the K matrix here can be inverted in O(N ) operations.
We can motivate such an inversion because for one dimensional structure, it can be done
in O(N ) operations. For the K matrix involving a tree basis, it can be thought of as a string
of one dimensional problems hooked up like branches of a tree.
A Simple 1D Problem
Before discussing the general problem, it is pedagogical to study the one dimensional (1D)
problem, as it can be more lucidly described. In a simple 1D problem regarding a strip, the
strip current can be expanded in a simple rooftop function
 x−xn

(1 − ln+1 ), xn < x < xn+1 ,
Jn (x) = x̂Jn (x) = x̂ (1 − xnl −x ), xn−1 < x < xn , (5.58)

 n
0, otherwise.
an interior
edge
(a)
xn
ln-1 ln+1
Hn(x)
(b)
xn-1 xn xn+1
x
Pn(x)
(c)
xn-1 x
xn
Figure 5.6: Illustration of the 1D tree problem.
The divergence of this current expansion function yields a charge that is proportional to the
simple Haar function [25] shown in Figure 5.6. The surface current can be expanded in terms
of the simple rooftop function, viz.,
N
X
Js (x) = In Jn (x) (5.59)
n=1
Taking the divergence, we have the surface charge

N
X
iωσ(x) = In Hn (x) (5.60)
n=1
where the Haar function is defined as


−1

−ln+1 , xn < x < xn+1 ,
−1
Hn (x) = ln , (5.61)
xn−1 < x < xn ,


0, otherwise.
But we can also expand the surface charge in terms of the pulse basis
n+1
X
σ(x) = Qn Pn (x), (5.62)
n=1
where (
−1/2
ln−1 , xn−1 < x < xn ,
Pn (x) = (5.63)
0, otherwise.
There is one more pulse (patch) function than rooftop functions as can be seen from the
figure. However, we can assign one of the patches to be a charge-neutralizing patch so that
the number of unknowns in (5.60) and (5.62) is the same. Charge neutrality is necessary
because the charge in Eq. (5.60) is naturally neutral, and also the Haar function in (5.61)
has zero net area. For simplicity, we choose the last patch in the figure to be the neutralizing
function.
Multiplying (5.62) by iω, equating the right hand sides of (5.62) and (5.60), and testing
with a pulse function Pm (x), m = 1, . . . , N, we have
N
X
iωQm = hPm (x), Hn (x)iIn
n=1
N
X
= Kmn In , m = 1, . . . , N (5.64)
n=1
or
iωQ = K · I (5.65)
It is quite clear that when the pulse function is used to test the left-most patch of the structure,
only one unknown from the right hand side of (5.64) is involved. In other words, the first row
of K has only one element, and hence, one can solve for I1 in terms of Q1 . The subsequent row
involves two elements, but one of the two unknowns is already solved for from the previous
row. Therefore, one can recursively solve for all the In ’s in terms of the Qn ’s, n = 1, . . . , N .
Here, QN +1 is picked for charge neutrality.
A General Tree Problem

For a general surface whose spanning tree (which is the set of RWG basis complementary to
the loop basis) has been found, the tree graph will be shown in Figure 5.7. Each node in
the tree represents the center of a triangular patch, which supports half of the spanning tree
RWG function, or the quad patch which supports half of a general rooftop. Each segment
between two nodes can be associated with the current that flows between two patches. If we
go through the same process of constructing the connection matrix K as in the 1D problem,
then the patch unknown that is associated with a tip of the dendritic branch is only related
to one current unknown. We can recursively solve for the unknowns starting from the branch
tips until a junction node is reached. For RWG basis, a junction node cannot have more than
three segments connected to it, even though some junction nodes may be associated with
Figure 5.7: Schematic of a general tree. Each node represents the center of an RWG expansion
function, and the line indicates the possibility of current flowing from one center to another
center.
more than three current unknowns, for example, when a quad patch is used. Eventually, for
the RWG basis case, after the unknowns on two of the open branches are solved for, the third
current unknown can be solved for. Finally all current unknowns, In ’s, can be solved for in
terms of Qn ’s, in O(N ) operation, because the CPU cost of recursion is proportional to the
number of nodes or N .
Figure 5.8 shows the loop-tree decomposition of a conical antenna [26]. The conical
antenna consists of two circular, annular surfaces. It is driven from the inner edges which are
connected by a circular cylinder. The outer edges are connected by four wires which can be
shorted or inductively loaded.
The current on the triangulated conical antenna is expanded in terms of RWG functions.
The center of each triangle is assigned a node, and the centers are connected to form loops
(see Figure 5.8(b)). Hence, an RWG function, which is associated with an inner edge on a
triangulated surface, lives between two nodes. The RWG basis is then decomposed into the
loop basis and the tree basis. In the case of Figure 5.8(b), when a group of RWG functions
form closed loops, they form a loop function that belongs to the loop basis. In the case of
Figure 5.8(c), the RWG functions do not form closed loops. Each RWG function living on a
segment of the tree belongs to the tree basis.
5.3.6 Reason for Ill-Convergence Without Basis Rearrangement

Without basis rearrangement, the charge part of the integral operator is acting on an unknown
current through the divergence operator, viz.,
Z
∇
LC Js = dr0 g(r, r0 )∇0 · Js (r0 ) (5.66)
iω² S
If we restrict Js (r0 ) to nondivergence-free basis, such as the tree basis, it turns out that there
are currents in the tree basis where ∇0 · Js (r0 ) ' 0 implying the existence of small eigenvalues
even when the above is applied to the tree space [27].
A way to avoid the divergence operator is via basis rearrangement so that the fundamental
unknown is the charge, which is related to the current via ∇ · Js (r) = iωσ(r), and hence Eq.
(a) Triangulated conical antenna.
(b) Loop functions of the conical antenna. (c) Tree functions of the conical antenna.
Figure 5.8: The decomposition of (a) a conical antenna into (b) a set of loop functions plus
(c) a set of tree functions.
(5.66) becomes Z
∇
LC Js = dr0 g(r, r0 )σ(r0 ) = L̂C σ (5.67)
²
In the above, K is the matrix representation of the divergence operator in the tree space [see
−1
(5.42)], while K is a preconditioner that removes the effect of the divergence operator.
5.4 Testing of the Incident Field with the Loop Function

When the frequency is very low, we need to test the right-hand side of the EFIE, which
consists of the incident fields, with the loop functions. This is computed as
VnL = −hJLn (r), Einc (r)i (5.68)
When the frequency is low, the above integral in the inner product can yield numerical
inaccuracies [28]. To remedy this problem, we use the same trick as in Eq. (5.21), and
express JLn (r) = ∇ × n̂ψn (r). By using integration by parts, we convert (5.68) to
VnL = iωµhψn (r), n̂ · Hinc (r)i (5.69)
so that VnL drops linearly with ω explicitly when ω → 0, whereas this explicit linear depen-
dence on ω is not in (5.68).
The use of low-frequency MOM (LF-MOM) with loop-tree decomposition to solve small
problems is demonstrated in the modeling of a Hertzian dipole in Figure 5.9(a) [11]. The input
impedance is calculated with LF-MOM where the equation is solved using conjugate gradient
(CG) or low-frequency multilevel fast multipole algorithm (LF-MLFMA). The capacitive
model where the input impedance is modeled solely by the capacitance between the two
spheres is also shown.
It is seen that at low frequencies, the dipole is a capacitive structure, in agreement with the
capacitive model. At higher frequencies, the capacitive model starts to depart from the full-
field model using LF-MOM where the magnetic field is solved concurrently with the electric
field. The reason is that the input feed wires of the Hertzian dipole start to become inductive.
At even higher frequencies as shown in Figure 5.9(b), where a nonloop-tree model is used,
the structure starts to resonate like an LC tank circuit.
Figure 5.10 shows the use of this technique to calculate the input impedance of the conical
antenna shown in Figure 5.8 [26]. With this technique, the input impedance can be calculated
to very low frequencies without low-frequency breakdown, but without this technique as in the
finite element radiation method (FERM) code developed at Lincoln Lab [29], the calculation
at low frequency is impossible.
5.5 The Multi-Dielectric-Region Problem

The problem involving a single dielectric region can be solved in terms of the PMCHWT
(Poggio-Miller-Chang-Harrington-Wu-Tsai) formulation (see previous chapter). Even the
10
10 2500
Plain CG Resistance
LF−MLFMA 2000 Reactance
10
8
Capacitor approximation
1500 1m 1m
1m 1m
Impedance (ohm)
Reactance (ohm)
6
10 Z in
1000
4 500
10
1m 1m
1m 1m 0
10
2 Z in
−500
0
10 −10 −8 −6 −4 −2 0
−1000
10 10 10 10 10 10 0 0.02 0.04 0.06 0.08 0.1
Frequency (GHz)
Frequency (GHz)
(a) (b)
Figure 5.9: Input impedance of a Hertzian dipole from (a) very low frequency to (b) microwave
frequency.
(a) (b)
Figure 5.10: Input impedance of a conical antenna at very low frequencies indicating (a) the
conductance, and (b) the susceptance. Data are taken from [26].
ε0,µ0
ε1,µ1 ε2,µ2
Figure 5.11: The multiple dielectric region problem.
Figure 5.12: The multidielectric region problem with contacting regions can be solved as the
limiting case of the noncontact region problem.
multidielectric-region problem can be solved in terms of PMCHWT formulation, or the com-

posite region problem as shown in Figure 5.11 [30]. The PMCHWT formulation involves
the use of the EFIE operator which becomes unstable or ill-conditioned when ω → 0. The
MFIE operator or the K operator does not exhibit low-frequency breakdown as in the EFIE
operator, but because of the disparate amplitude of the eddy current with respect to the
charge current, numerical inaccuracy still persists [31]. Hence, it is still necessary to use a
loop-tree decomposition method to stabilize the EFIE operator [28, 32–34]. However, per-
forming a loop-tree decomposition for a composite surface shown in Figure 5.11 is a difficult
task. Instead, we break the composite object into two nontouching objects with closed sur-
faces as shown in Figure 5.12. The loop-tree decomposition can then be preformed for the
individual closed surfaces, and the corresponding integral operators can be stabilized as such.
As a result, the solution for Figure 5.11 can be obtained by letting t → 0. We call this the
contact-region modeling method [33]. The disadvantage of such a method is that the number
of unknowns needed to solve the problem is increased, but the loop-tree decomposition on
the surface currents can be performed.
5.5.1 Numerical Error with Basis Rearrangement

A simple way to rearrange the basis so that the charges from the RWG currents are connected
to the pulse basis yields numerical instability. This is because of the numerical cancelation
that happens. So when a large problem is solved, the error becomes unacceptable, and
iterative solution method does not converge [28].
One can appreciate this problem as follows. Originally, when we take the divergence of
the RWG basis to obtain the charge, a pair of pulse bases of opposite signs are formed. We
shall call this the short patch pair (SPP) function. This is the expansion function for the
charge before basis rearrangement.
(a) (b)
+ +
_ +
_ _
A short patch-pair function A long patch-pair function
Figure 5.13: A short patch-pair function and a long patch-pair function.
After basis rearrangement, the fundamental expansion function is a patch pair which is
shown in Figure 5.13(b). A patch pair can be formed by linearly superposing a number of
SPP functions, where there are charge cancelations between the patch pair. Before basis
rearrangement, these SPPs are producing electric field of opposite signs, where supposedly
after basis rearrangement, they are to cancel exactly, but in a numerical setting, they do
not cancel exactly, giving rise to numerical errors. Hence, when the total field is tested at
a field point where exact cancelation should have occurred, numerical error persists. This is
especially true when large problems are solved where the separation between the patch pair
can be very large, and many SPPs are involved before they cancel to form the long patch pair
(LPP).
As a remedy to this problem, we first form a matrix representation of the Green’s operator
using a single-patch expansion function, that is to find
Anm = hσn (r), g(r, r0 ), σm (r0 )i (5.70)
directly where σn (r) is a single-patch expansion function. Then a linear transformation

converts the single-patch basis to the patch-pair basis, because a patch pair function is just
a linear sum of two single-patch expansion function. Since the transformation is the change
of basis, a corresponding similarity transform can be performed on the matrix in Eq. (5.70)
to obtain the matrix representation based on patch-pair basis. The transformation is not full
rank because for each object where charge neutrality is maintained, the rank is reduced by
one. For Nb independent noncontacting objects, the rank is reduced by Nb .
In this alternative implementation, the LF-FMA [24] and other low-frequency algorithms
[35–38] can still be used to accelerate the matrix-vector product associated with Amn . The
1
10
PEC
10 mm
2 mm
0
Input Reactance ( Ω )
10
ε = 3.0
r
−1
10
4 mm Consistent mesh, 1980 unknowns

Inconsistent mesh, 1986 unknowns
Inconsistent mesh, 32172 unknowns
20 mm
6 7 8
10 10 10
Frequency (Hz)
(a) Geometry (b) Input reactance vs. frequency

Figure 5.14: The input reactance of an inductance structure using contact region modeling.
transformation matrix involved with the similarity transform is sparse with O(N ) elements.
Hence, a matrix-vector product can be performed in O(N ) operations.
Figures 5.14(a) and 5.14(b) show the use of contact-region modeling to model an inductive
ring on top of a dielectric substrate [34]. It is shown that the meshes on the two contacting
surfaces need not be the same, and yet the result can be reasonable. The inductor-like PEC-
dielectric composite structure as shown in Figure 5.14(a), where the top part is a square PEC
ring located on the dielectric substrate. For convenience of display, the top part is shifted
upwards in Figure 5.14(a). A delta-gap source is placed peripherally in the middle of one arm
of the PEC part. The input reactance is shown in Figure 5.14(b), which demonstrates the
inductor-like character of the structure. Again the input resistance is negligible compared to
the reactance. Good agreement among different meshes is observed.
Figures 5.15(a), 5.15(b), 5.15(c) show the use of LF-MOM for the solution of some large
structures using LF-MLFMA acceleration with the stabilization of the patch pair described
above [28]. It is only with stabilization that such a large problem can be solved. This example
involves two layers of wires and a ground plane. The width of each wire is 5 µm and the
thickness is 1 µm. The distance between two adjacent parallel wires is 4 µm. The ground
plane is 210 × 210 µm2 . The first layer of wires is 4 µm above the ground plane and the
second layer is 8 µm above. Two ports are defined as in the figure. The conductivity values
of the wires and the ground plane is 3.7 × 107 S/m. The entire structure is assumed to be in
a silicon dioxide background with ²r = 3.8.
5.6 Multiscale Problems in Electromagnetics

The solution to the multiscale problems in electromagnetics is important, especially so in
the area of computer design. Inside a computer chassis, electromagnetic problems of different
length-scales occur. Moreover, unlike some problems, homogenization method cannot be used
to simplify the problems of small length-scale so that they can be interfaced with problems
of larger length-scales.
Recently, the mixed-form fast multipole algorithm (MF-FMA) [17] and broadband fast
algorithm [39, 40] have been developed so that integral equation solvers can be accelerated
7
σ =3.7× 10 s/m
f = 1 GHz 3
210 µ m
Port 2
−1
Port 1
−2
(a) (b)
7 5
σ =3.7× 10 s/m
f = 1 GHz
4
−1
−2
−3
(c)
Figure 5.15: (a) Two layers of wires with a ground plane. (b) Charge magnitude on the
three-layer structure. (c) Current magnitude on the three-layer structure.
all the way from very low frequency for objects with very tiny length-scale compared to
wavelength to mid length-scale where the objects are of the order of wavelength or larger.
These methods will be suitable for accelerating integral equation solvers for problems pertinent
to multiscale problems.
However, because of the disparate nature of the physics of small length-scale and physics
of mid-length-scale, it is necessary to decouple these two physical regimes when solutions are
sought. The equivalence principle algorithm (EPA) [41,42] will be a good method to decouple
the solutions of these two regimes. In this manner, solvers can be used separately to solve
for solutions in the regime of small length-scales, and the regime of mid length-scales. This
is still an active area of research.
5.7 Conclusions
This chapter introduces the low-frequency problems and solutions in solving integral equa-
tions. It suggests using the loop-tree decomposition to overcome the low-frequency breakdown
problem. The basis rearrangement method is introduced to arrive at a better conditioned
matrix system, and the physical reason is also given. The low-frequency technique is very im-
portant in solving problems associated with computer technology where wave physics meets
circuit physics. It is also important in modeling nanostructures where the wavelength is much
longer than the size of the structures. However, most structures, as those in computer tech-
nology (see end of Chapter 1), are multiscale. The multiscale problems presage new solution
techniques. We are currently investigating a stable low-frequency solution without the need
for loop-tree decomposition [43].
Bibliography
[1] J. Blau, “Supermarket’s futuristic outlet,” IEEE Spectrum, vol. 41, no. 4, pp. 21–22, 25,
Apr. 2004.
[2] W. C. Chew Waves and Fields in Inhomogeneous Media, New York: Van Nostrand Rein-
hold, 1990. Reprinted by Piscataway, NJ: IEEE Press, 1995.
[3] M. S. Tong and W. C. Chew, “Unified boundary integral equation for the scattering of
elastic wave and acoustic wave: solution by the method of moments,” Waves Random
Complex Media, vol. 18, pp. 303–324, May 2008.
[4] D. R. Wilton and A. W. Glisson, “On improving the electric field integral equation at low
frequencies,” URSI Radio Sci. Meet. Dig., pp. 24, Los Angeles, CA, Jun. 1981.
[5] J. R. Mautz and R. F. Harrington, “An E-field solution for a conducting surface small
or comparable to the wavelength,” IEEE Trans. Antennas Propag., vol. 32, pp. 330–339,
Apr. 1984.
[6] E. Arvas, R. F. Harrington, and J. R. Mautz, “Radiation and scattering from electrically
small conducting bodies of arbitrary shape,” IEEE Trans. Antennas Propag., vol. 34, pp.
66–77, Jan. 1986.
[7] W. Wu, A. W. Glisson, and D. Kajfez, “A comparison of two low-frequency formulations

for the electric field integral equation,” Tenth Ann. Rev. Prog. Appl. Comput. Electromag.,
vol. 2, pp. 484–491, Monterey, California, Mar. 1994.
[8] M. Burton and S. Kashyap, “A study of a recent, moment-method algorithm that is

accurate to very low frequencies,” Appl. Comput. Electromagn. Soc. J., vol. 10, no. 3, pp.
58–68, Nov. 1995.
[9] F. Yuan, “Analysis of power/ground noise and decoupling capacitors in printed circuit
board systems,” IEEE Electromagn. Compat. Symp. Proc., pp. 425–430, Austin, TX,
Aug. 1997.
[10] G. Vecchi, “Loop-star decomposition of basis functions in the discretization of EFIE,”

IEEE Trans. Antennas Propag., vol. 47, pp. 339–346, Feb. 1999.
1
[11] J. S. Zhao and W. C. Chew, “Integral equation solution of maxwell’s equations from
zero frequency to microwave frequencies,” IEEE Trans. Antennas Propag., James R. Wait
Memorial Special Issue, vol. 48. no. 10, pp. 1635–1645, Oct. 2000.
[12] R. E. Collins, Foundations for Microwave Engineering, New York: McGraw-Hill, 1966.
[13] D. Colton and R. Kress, Integral Equation Methods in Scattering Theory, Melbourne,
FL: Krieger Publ., 1983.
[14] W. C. Chew, B. Hu, Y. C. Pan and L. J. Jiang, “Fast algorithm for layered medium,”
C. R. Phys., vol. 6, pp. 604–617, 2005.
[15] W. C. Chew, L. J. Jiang, Y. H. Chu, G. L. Wang, I. T. Chiang, Y. C. Pan, and J. S. Zhao,
“Toward a more robust and accurate cem fast integral equation solver for IC applications,”
IEEE Trans. Adv. Packag., vol. 28, no. 3, pp. 449–464, Aug. 2005.
[16] R. Albanese and G. Rubinacci, “Integral formulation for 3D eddy-current computation
using edge elements,” IEE Proc., vol. 135, pt. A, no. 7, pp. 457–462, Sept. 1988.
[17] L. J. Jiang and W. C. Chew, “A mixed-form fast multipole algorithm,” IEEE Trans.
Antennas Propag., vol. AP-53, no. 12, pp. 4145–4156, Dec. 2005.
[18] L. Bai, “An efficient algorithm for finding the minimal loop basis of a graph and its
applications in computational electromagnetics,” M.S. Thesis, Dept. ECE, U. Illinois,
Urbana-Champaign, Sept. 2000.
[19] V. I. Okhmatovski, “An efficient algorithm for generation of loop-tree basis in 2.5D
interconnect models,” Elect. Perf. Electron. Packag., Austin, TX, 2005.
[20] S. Chakraborty, D. Gope, G. Ouyang, and V. Jandhyala, “A three-stage preconditioner
for geometries with multiple holes and handles in integral equation-based electromagnetic
simulation of integrated packages,” Elect. Perf. Electron. Packag., Austin, TX, 2005.
[21] H. Y. Chao and W. C. Chew, “A linear-time algorithm for extracting tree and loop
bases in computational electromagnetics,” IEEE Antennas Propag. Soc. Int. Symp. Proc.,
Albuquerque, NM, pp. 4115–4118, 2006.
[22] Y. Liu, Y. Chu and W. C. Chew, “LFFMA and induction well-logging modeling,” IEEE
Antennas Propag. Soc. Int. Symp. Proc., Washington, DC, vol. 3A, pp. 224–227, Jul. 2005.
[23] J. D. Jackson, Classical Electrodynamics, 3rd ed., New York: Wiley, 1999.
[24] J. S. Zhao and W. C. Chew, “Three dimensional multilevel fast multipole algorithm from
static to electrodynamic,” Microwave Opt. Tech. Lett., vol. 26, no. 1, pp. 43–48, Jul. 2000.
[25] I. Daubechies, “Orthonormal bases of compactly supported wavelets,” Commun. Pure
Appl. Math., vol XLI, pp. 909–996, 1988.
[26] J. S. Zhao, W. C. Chew, and P. E. Mayes, “Accurate analysis of electrically small conical
antennas by using the low-frequency method,” Antennas Appl. Symp., Monticello, IL, pp.
135–151, Sept. 19–21, 2001.
[27] K. F. Warnick and W. C. Chew, “Error analysis of surface integral equation methods,”
in Fast and Efficient Algorithms in Computational Electromagnetics, eds. W. Chew, J.
Jin, E. Michielssen, J. Song, Boston: Artech House, 2001.
[28] Y. H. Chu and W. C. Chew, “Large-scale computation for electrically small structures
using surface integral equation method,” Microwave Opt. Tech. Lett., vol. 47, no. 6, pp.
525–530, Dec. 2005.
[29] D. A. Shnidman and S. Lee, “The finite element radiation model (FERM) program,”
Third Ann. Rev. Prog. Appl. Comput. Electromag., Monterey, CA, Mar. 24–26, 1987.
[30] L. N. Medgyesi-Mitschang, J. M. Putnam, and M. B. Gedera, “Generalized method of
moments for three-dimensional penetrable scatterers,” J. Opt. Soc. Am. A, vol. 11, pp.
1383–1398, Apr. 1994.
[31] Y. Zhang, T. J. Cui, W. C. Chew and J. S. Zhao, “Magnetic field integral equation at
very low frequencies,” IEEE Trans. Antennas Propag., vol. 51, no. 8, pp. 1864–1871, Aug.
2003.
[32] S. Y. Chen, W. C. Chew, J. M. Song, and J. S. Zhao, “Analysis of low frequency scattering
from penetrable scatterers,” IEEE Geosci. Remote Sens., vol. 39, no. 4, pp. 726–735, Apr.
2001.
[33] Y. Chu, W. C. Chew, S. Chen, and J. Zhao, “A surface integral equation method for
low-frequency scattering from a composite object,” IEEE Trans. Antennas Propag., vol.
51, no. 10, pp. 2837-2844, Oct. 2003.
[34] Y. Chu and W. C. Chew, “A multi-level fast multipole algorithm for electrically small
composite structures,” Microwave Opt. Tech. Lett., vol. 43, no. 3, pp. 202–207, Nov. 2004.
[35] L. Greengard, J. F. Huang, V. Rokhlin, and S. Wandzura, “Accelerating fast multipole
methods for the Helmholtz equation at low frequencies,” IEEE Comput. Sci. Eng., vol. 5,
no. 3, pp. 32–38, 1998.
[36] L. Xuan and R. J. Adams, “Computing the translation operator by sampling the Green
function at low frequencies,” URSI Radio Sci. Meet. Dig., Columbus, OH, pp. 683, Jun.
2003.
[37] L. J. Jiang, and W. C. Chew, “Low frequency inhomogeneous plane wave algorithm–LF-
FIPWA,” Microwave Opt. Tech. Lett., vol. 40, no. 2, pp. 117–122, Jan. 2004.
[38] E. Darve and P. Havè, “Efficient fast multipole method for low-frequency scattering,” J.
Comput. Phys., vol. 197, no. 1 pp. 341–363, 2004.
[39] H. Wallén and J. Sarvas, “Translation procedures for broadband MLFMA,” Prog. Elec-
tromagn. Res., PIER 55, pp. 47-78, 2005.
[40] H. Cheng, W. Y. Crutchfield, Z. Gimbutasa, L. F. Greengard, J. F. Ethridgea, J. Huang,
V. Rokhlin, N. Yarvin, and J. Zhao, “A wideband fast multipole method for the Helmholtz
equation in three dimensions,” J. Comput. Phys., vol. 216, no. 1, pp. 300–325, 2006.
[41] M. K. Li and W. C. Chew, “A domain decomposition scheme based on equivalence

theorem,” Microwave Opt. Tech. Lett., vol. 48, no. 9, pp. 1853–1857, Sept. 2006.
[42] M. K. Li and W. C. Chew, “Wave-Field Interaction With Complex Structures Using

Equivalence Principle Algorithm,” IEEE Trans. Antennas Propag., vol. 55, no. 1, pp.
130–138, Jan. 2007.
[43] Z. G. Qian and W. C. Chew, “An augmented EFIE for high speed interconnect analysis,”
Microwave Opt. Tech. Lett., scheduled for October 2008.
135
CHAPTER 6
Dyadic Green’s Function for Layered

Media and Integral Equations
6.1 Introduction
In computational electromagnetics (CEM), modeling with layered medium is quite common.
This problem can be formulated in terms of integral equations using the layered medium
Green’s function. As such, the dyadic Green’s function needs to be derived. When the
layered medium Green’s function is used, unknowns are only associated with the embedded
scatterers or objects, but not with the interfaces of the layers. Hence, it greatly reduces the
number of unknowns required to model the problem accurately. The integral equation thus
formulated is then converted into a matrix equation by the method of moments (MOM) [1]
described in the previous chapters.
In this chapter, we introduce an elegant way of deriving the matrix representation of
the dyadic Green’s function for layered media (DGLM) [2]. The derivation is based on the
pilot vector potential approach where two vector potentials in terms of TE and TM fields
are derived. The derivation is succinct and the resulting Sommerfeld integrals only involve
zeroth order Bessel function and two scalar potentials.
The problem of a dipole over a half space was first solved by Sommerfeld [3] in 1909 using
Hertz vector potential. Kong in 1972 suggested using the Ez and Hz formulation to solve
such problems, extending them to layered media [4]. The layered medium dyadic Green’s
function can only be written in terms of Fourier integrals which are expressed in terms of
Sommerfeld integrals [5, 6]. The evaluation of these integrals is time consuming, especially if
they are associated with high order coordinate-space singularities. Because these integrals are
spectral integrals, the coordinate-space singularities translate into higher spectral components
in the Fourier space, causing their slow convergence.
The filling of these matrix elements involves evaluation of Sommerfeld integrals for singular
functions in the coordinate space. These singular functions are due to the spatial derivatives
1
of the Green’s function. To evaluate these matrix elements, it is important that the matrix
representations be manipulated into a form whose coordinate-space singularities are as weak
as possible. We shall call such matrix elements the matrix-friendly forms. By this, we mean
that their associated Sommerfeld integrals are more rapidly convergent.
6.2 Dyadic Green’s Function for Layered Media

The DGLM has been derived in various forms [5–10]. Michalski and Zheng [7] used the mixed
potential formulation to derive a dyadic Green’s function. The Michalski-Zheng formulation
is vastly popular and has been adopted by many workers1 [11–18]. We will start with a dyadic
Green’s function that is succinctly derived in [6], and then manipulate it into a form that is
matrix-friendly for evaluating elements of its matrix representation.
To derive the dyadic Green’s function for a layered medium, first we derive that for a
homogeneous medium. The dyadic Green’s function for a homogeneous medium (or free
space) exists in many different forms. However, we will use the form that is expressed in
terms of a linear superposition of plane waves. This is with the hindsight that the interaction
of plane waves with layered medium can be easily solved for.
Using the z-directed pilot vector potential approach [19], the free space dyadic Green’s
function can be found using the vector wave function expansion approach. These are the M,
N and L functions where M, N are divergence free but not curl free (solenoidal) whereas L
is curl free (irrotational) but not divergence free. In Cartesian coordinates, when ẑ is chosen
as the pilot vector, they are defined in terms of plane waves as
M(k, r) = ∇ × ẑeik·r = ik × ẑeik·r (6.1)

1 1
N(k, r) = ∇ × M(k, r) = − k × k × ẑeik·r (6.2)
k k
L(k, r) = ∇eik·r = ikeik·r (6.3)
Here, M and N are transverse waves because they are orthogonal to the direction of the plane
wave propagation or the k vector, whereas L is a longitudinal wave because it points in the
direction of the plane wave propagation or the k vector.
The dyadic Green’s function, when expressed in terms of the vector wave functions, is
Z Z∞Z ·
0 1 M(k, r) M(−k, r0 )
G(r, r ) = dk
(2π)3 (k 2 − km
2 )k 2
s
−∞
¸
N(k, r) N(−k, r0 ) L(k, r) L(−k, r0 )
+ − . (6.4)
(k 2 − km
2 )k 2
s
2 k2
km
√
where ks2 = kx2 + ky2 , k 2 = kx2 + ky2 + kz2 , km = ω µm ²m is the wavenumber of the homogeneous
medium, and dk = dkx dky dkz is a three-dimensional Fourier integral. Hence, k is not a
constant at this point. A continuum of Fourier wave number is needed because the above
corresponds to a Fourier expansion in an infinitely large box in the x, y, and z directions.
1 This list is by no means complete.
DYADIC GREEN’S FUNCTION FOR LAYERED MEDIA AND INTEGRAL EQUATIONS 137
The above choice of pilot vector ẑ makes the vector M function transverse to ẑ. If the
above represents the electric field dyadic Green’s function, M corresponds to a TE to z wave,
whereas N resembles the electric field of a TM to z wave. The L function is needed for
completeness. This follows from Helmholtz theorem that a vector field can be decomposed
into the sum of a solenoidal (divergence-free) field and an irrotational (curl-free) field.
A contour integration technique can be applied to the above to evaluate the dkz integration
in closed form. Consequently, the dyadic Green’s function can be decomposed into TEz and
TMz components. Hence, the dyadic Green’s function for a homogeneous medium can be
derived in terms of vector wave functions as [6, p. 411]
ZZ ∞
i dks
G(r, r0 ) = 2 2
[M(ks , r)M(−ks , r0 )
8π −∞ kmz k s
ẑ ẑ
+N(ks , r)N(−ks , r0 )] − 2 δ(r − r0 ), (6.5)
km
where
M(ks , r)M(−ks , r0 ) = (ks × ẑ)(ks × ẑ)eiks ·(rs −rs ) eikmz |z−z |

0 0
(6.6)
1
N(ks , r)N(−ks , r0 ) = 2 (km± × ks × ẑ)(km∓ × ks × ẑ) · eiks ·(rs −rs ) eikmz |z−z |
0 0
(6.7)
km
0
km± = ks ± ẑkmz , ks = x̂kx + ŷky , and the
p upper sign is2chosen when z > z and the lower
0 2 2
sign is chosen when z < z . Here, kmz = km − ks and ks = ks · ks .
In (6.4), the dyadic Green’s function represents the electric field produced by a point
current source at r0 . Hence, this field must be divergence-free when r 6= r0 . The L function is
generally nondivergence free, and seemingly, it cannot be nonzero outside the source region.
But because the ∇ · L = −k 2 eik·r which is divergence free at k = 0, the L function can be
nonzero outside the source point for k = 0. In fact, k = 0q with k 6= 0 is possible, where
the three components of the k vector are related by kz = ± kx2 + ky2 , so that L defined in
(6.3) can be nonzero. Consequently, looking at (6.4), for the L term, a pole exists at k = 0
giving rise to a pole contribution for the L term when the contour integration technique is
used to evaluate the dkz integration. This field is nonzero outside the source point, but it is
Laplacian (static) in nature, and hence, is unphysical for a time-harmonic field. Fortunately,
because N ∼ k1 , k → 0, a pole contribution from the N term produces a static field that
exactly cancels this unphysical field. The last term in (6.5) is the remnant of this cancelation.
The above integrals are ill-convergent when the source point and the field point are copla-
nar, viz., z = z 0 . This is because this plane contains the source point singularity whose
Fourier transform is ill-convergent (see [13,14,18] for details). However, numerical evaluation
of them can be avoided because they have closed form in coordinate space for the homogeneous
medium case, that is,
µ ¶ ikm |r−r0 |
0 1 e
G(r, r ) = I + 2 ∇∇ (6.8)
km 4π|r − r0 |
In the above, the M function corresponds to TEz waves whereas the N function corre-
sponds to TMz waves. For a layered medium, we can replace the above by [6]
M(ks , r)M(−ks , r0 ) = (∇ × ẑ)(∇0 × ẑ)eiks ·(rs −rs ) F TE (ks , z, z 0 ),

0
(6.9)
µ ¶µ 0 0
¶
∇ × ∇ × ẑ ∇ × ∇ × ẑ
· eiks ·(rs −rs ) F TM (ks , z, z 0 ), (6.10)
0
N(ks , r)N(−ks , r0 ) = −
iω²n iωµm
where we have assumed that the source point is in region m whereas the field point is in
region n. The function F α (ks , z, z 0 ) above, where α is either TE or TM, satisfies the following
ordinary differential equation [6, 20]
· ¸
d 1 d 1 2 2i
+ k (z) F α (ks , z, z 0 ) = kz (z)δ(z, z 0 ) (6.11)
dz p(z) dz p(z) z p(z)
p
where p = µ for TE waves, and p = ² for TM waves. Moreover, kz = k 2 (z) − ks2 . When
k(z) is piecewise constant denoted by km for region m, F α can be found in closed form using a
recursive procedure [6, p. 76]. Furthermore, F α describes the propagation of Bz or Dz fields
for TE and TM fields respectively in an inhomogeneous layer. For a homogeneous medium,
F α (ks , z, z 0 ) = exp(ikmz |z − z 0 |).
After using Eqs. (6.9) and (6.10) in Eq. (6.5), and exchanging the order of derivatives
and integrals, we can rewrite the dyadic Green’s function for layered media as
Ge (r, r0 ) =(∇ × ẑ)(∇0 × ẑ)g TE (r, r0 )

1
+ 2 (∇ × ∇ × ẑ)(∇0 × ∇0 × ẑ)g TM (r, r0 ) , |z − z 0 | > 0, (6.12)
knm
2
where knm = ω 2 ²n µm . The subscript e is used to denote that this is an electric dyadic Green’s
function that generates electric field when acting on an electric current source. The above
representation is strictly correct only when |z −z 0 | > 0, because when the derivative operators
act on exp(ikmz |z − z 0 |), a singularity is produced at z = z 0 which is not the same as that in
Eq. (6.5). Here, g TE (r, r0 ) and g TM (r, r0 ) consist of the primary (direct) field terms and the
secondary (reflected) field terms. In a layer that does not contain the source point, the field
is entirely secondary. The primary field term contains the source singularity, and hence, its
representation in the Fourier space is inefficient. Because it is the same as the homogeneous
medium case for which closed form exists, it is more expedient to express it in closed form in
coordinate space (see Eq. (6.8)).
Another point to be noted is that in (6.12), the derivative operators are taken out of the
spectral integrals. Hence, g TE (r, r0 ) and g TM (r, r0 ) in their coordinate-space representations,
are less singular than the dyadic Green’s function itself. As such, their spectral integrals are
more rapidly convergent compared to the spectral integrals in (6.5).
The explicit expressions for g TE (r, r0 ) and g TM (r, r0 ) are
ZZ ∞
i dks iks ·(rs −r0s ) TE
g TE (r, r0 ) = 2 2
e F (ks , z, z 0 ) (6.13)
8π −∞ k mz ks
ZZ ∞
i dks iks ·(rs −r0s ) TM
g TM (r, r0 ) = 2 2
e F (ks , z, z 0 ) (6.14)
8π −∞ k mz ks
The above Fourier integrals can be expressed in terms of Sommerfeld integrals involving a
single integral and Bessel function (where kρ = ks ) using the techniques described in [6],
namely
Z ∞
TE 0 i dkρ
g (r, r ) = J0 (kρ |rs − r0s |)F TE (kρ , z, z 0 ), (6.15)
4π 0 kmz kρ
Z ∞
i dkρ
g TM (r, r0 ) = J0 (kρ |rs − r0s |)F TM (kρ , z, z 0 ). (6.16)
4π 0 kmz kρ
To arrive at the above, we have made use of the identity that
ZZ ∞ Z 2π Z ∞
dks = dα kρ dkρ (6.17)
−∞ 0 0
Z 2π
1 0
J0 (kρ |rs − r0s |) = eikρ ·(rs −rs ) dα (6.18)
2π 0
where kρ = x̂kρ cos α+ ŷkρ sin α. Because of the 1/ks2 or the 1/kρ term in the integrands, these
spectral integrals are more rapidly convergent compared to those in (6.5). When F TM and
F TE are defined in (6.15)–(6.16), the generalized reflection coefficients R̃TE and R̃TM [6, Chap.
2] embody polarization effects of the layered medium.
Notice that in (6.15) and (6.16), poles exist at kρ = 0, but these poles are unphysical
causing the integrals to diverge. This dilemma is nonexistent because g TE and g TM are not
the final physical quantities, but Ge in (6.12). In fact, the poles disappear because they
cancel each other when (6.15) and (6.16) are substituted in (6.12).
6.3 Matrix Representation

The primary field term is best expressed in coordinate space in closed form as in Eq. (6.8).
Hence, its matrix representation is straightforward and can be done following conventional
methods. The secondary field terms do not contain singularities, and are regular, except when
both the field point and source point are at the interface—a near singularity exists when the
source point and field points are close to the interface.
In the following, the manipulation to obtain the matrix representation of the Green’s
function operator is intended for the secondary fields where the image source point and the
field point do not coincide. Focussing on the TM wave term for the dyadic Green’s function,
we have
TM,S 1
Ge (r, r0 ) =
2
(∇ × ∇ × ẑ)(∇0 × ∇0 × ẑ)g TM,S (r, r0 )
knm
1
= 2 (∇∇ · ẑ + kn2 ẑ)(∇0 ∇0 · ẑ + km
2
ẑ)g TM,S (r, r0 )
knm
1
= 2 (∇∇0 ∂z ∂z0 + kn2 ∇0z ∇0 + km 2
∇∇z + kn2 km 2
ẑ ẑ)g TM,S (r, r0 ) , (6.19)
knm
where the superscript S indicates that these are secondary field terms, and ∇z = ẑ∂z . In
the above, we have assumed that (∇2 + kn2 )g TM,S (r, r0 ) = 0 and (∇02 + km
2
)g TM,S (r, r0 ) = 0,
because we assume that the field point and the image source point do not coincide. Likewise,
the TE wave term for the DGLM is
TE,S
Ge (r, r0 ) = (∇ × ẑ)(∇0 × ẑ)g TE,S (r, r0 ) . (6.20)
The electric field due to an electric current can be expressed with the dyadic Green’s
function for an inhomogeneous medium as [6, p. 411]
Z
E(r) = iω dr0 Ge (r, r0 )µ(r0 ) · J(r0 ). (6.21)
V
Next, we evaluate the matrix element for the impedance matrix as
Zij = iωµm hJT i (r), Ge (r, r0 ), Jj (r0 )i

Z Z
= iωµm drJT i (r) · dr0 Ge (r, r0 ) · Jj (r0 ), (6.22)
where we have assumed that the source expansion function Jj (r0 ) is entirely in the m-th region.
When a source expansion function straddles two or more regions, the above expression has
to be modified accordingly.
In the above, the matrix representation of the primary field term is obtained in the con-
ventional method. For the secondary field, the term associated with TM waves is
TM,S 1 £
hJT i (r), Ge (r, r0 ), Jj (r0 )i = 2
h∇ · JT i (r), ∂z ∂z0 g TM,S (r, r0 ) , ∇0 · Jj (r0 )i
knm
− kn2 hJT i (r) · ẑ, ∂z0 g TM,S (r, r0 ) , ∇0 · Jj (r0 )i
2
− km h∇ · JT i (r), ∂z g TM,S (r, r0 ) , ẑ · Jj (r0 )i
¤
+kn2 km
2
hẑ · JT i (r), g TM,S (r, r0 ) , ẑ · Jj (r0 )i . (6.23)
In deriving the above, we have made use of integration by parts as much as possible to move
the ∇ operators away from g TM,S . Notice that ∂z ∂z0 g TM,S is the most singular function since
derivatives enhance the order of the singularity. If the z derivative operators are brought into
the spectral integral of g TM,S , the resultant integral will be slowly convergent. The other
integrals are more benign and are rapidly convergent when numerically integrated.
For the term associated with TE waves, using (6.20), we can express it as
TE,S £ ¤
hJT i (r), Ge (r, r0 ), Jj (r0 )i = hJT i (r), (∇s × ẑ)(∇0s × ẑ)g TE,S (r, r0 ) , Jj (r0 )i, (6.24)
where ∇s = x̂∂x + ŷ∂y . If JT i and Jj are curl-conforming volume expansion function, integra-
tion by parts can be used to move the curl operators to the expansion and testing functions.
We manipulate the expression by first showing that
(ẑ × ∇s )(ẑ × ∇0s ) = ks2 Is − ∇s ∇0s

= Is ks2 − (∇∇0 − ∇z ∇0 − ∇∇0z + ∇z ∇0z ). (6.25)
(The first equality above can be easily proved in the spectral domain.) Then we can rewrite
(6.24) as
TE,S
hJT i (r), Ge (r, r0 ), Jj (r0 )i =hJT Si (r), gsTE,S (r, r0 ), JSj (r0 )i
− h∇ · JT i (r), g TE,S (r, r0 ), ∇0 · Jj (r0 )i
− hJT i (r) · ẑ, ∂z g TE,S (r, r0 ), ∇0 · Jj (r0 )i
− h∇ · JT i (r), ∂z0 g TE,S (r, r0 ), ẑ · Jj (r0 )i
− hJT i (r) · ẑ, ∂z ∂z0 g TE,S (r, r0 ), ẑ · Jj (r0 )i.
(6.26)
In the above,
Z ∞
i dkρ kρ
gsTE,S (r, r0 ) = J0 (kρ |rs − r0s |)F TE,S (kρ , z, z 0 ), (6.27)
4π 0 kmz
where the primary field term has been removed in F TE,S , and JT Si is the part of JT i that is
transverse to ẑ, and similarly for JSj .
We can combine the integrals in (6.26) with those in (6.23) to arrive at five basic terms.
By combining the TM and the TE Green’s functions together, and collecting like terms, for
the secondary field, we get
S
hJT i (r), Ge (r, r0 ), Jj (r0 )i
¿ À
∂z ∂z0 TM,S 0 TE,S 0 0 0
= ∇ · JT i (r), 2 g (r, r ) − g (r, r ), ∇ · Jj (r )
knm
µn
− hJT i (r) · ẑ, ∂z0 g TM,S (r, r0 ) + ∂z g TE,S (r, r0 ), ∇0 · Jj (r0 )i
µm
²m
− h∇ · JT i (r), ∂z g TM,S (r, r0 ) + ∂z0 g TE,S (r, r0 ), ẑ · Jj (r0 )i
²n
2
+ hẑ · JT i (r), kmn g TM,S (r, r0 ) − ∂z ∂z0 g TE,S (r, r0 ), ẑ · Jj (r0 )i
+ hJT Si (r), gsTE,S (r, r0 ) , JSj (r0 )i. (6.28)
Evaluating (6.28) involves the computation of two basic Green’s functions g TM,S , g TE,S and
their z derivatives. The gsTE,S Green’s function can be obtained from g TE,S by taking Rs
derivatives, where Rs = |rs −r0s |. It can be shown by using the property of the Bessel equation
that
1 ∂ ∂ TE,S
Rs g + gsTE,S = 0 (6.29)
Rs ∂Rs ∂Rs
these Green’s functions, g TM,S , g TE,S , can be computed and tabulated via Sommerfeld inte-
grals, and their derivatives can be approximated by finite difference on tabulated values to
evaluate the above five terms.
Alternatively, for higher accuracy, one may want to evaluate the z derivatives exactly by
bringing them inside the spectral integrals to avoid the need for finite difference. To this
end, the Green’s functions sandwiched between the source expansion function and the test
function can be combined into one integral, and tabulated for efficiency. Also, in the special
case when the source point and the field point are in the uppermost layer or the lowermost
layer, the second and the third terms in (6.28) can be further combined, leaving only four
integrals.
We consider the matrix representation when Rao-Wilton-Glisson (RWG) functions are
used for expansion and testing. Using the vector notation shown in Figure 6.1, the matrix
elements for the impedance matrix can be explicitly expressed as
2
X Z Z n
l j lk pq pq
Zjk = iωµm p q dr dr0 Zss,jk + Zzz,jk + Zzpq1 ,jk + Zzpq2 ,jk
p,q=1
4S j Sk Sjp Skq
o
pq
+ZΦ,jk (6.30)
where
pq ¡ ¢
Zss,jk = ẑ × ẑ × ξpj · (ẑ × ẑ × ξ qk ) gsTE (r, r0 ) (6.31)
pq ¡ p¢ q £ 2 TM TE
¤
Zzz,jk = ẑ · ξj (ẑ · ξk ) kmn g − ∂z ∂z 0 g (6.32)
· ¸
¡ ¢ µn
Zzpq1 ,jk = − 2²q ξpj · ẑ ∂z0 g TM + ∂z g TE (6.33)
µm
· ¸
²m
Zzpq2 ,jk = − 2²p (ξ qk · ẑ) ∂z g TM + ∂z0 g TE (6.34)
²n
· ¸
pq ∂z ∂z0 TM
ZΦ,jk =4²pq 2
g − g TE (6.35)
knm
½ ½
1, p = q 1, s = 1
where ²p,q = , ²s = , s = p, q. The subscripts ss, z1 , z2 , zz, Φ
−1, p 6= q −1, s = 2
pq
of Z·,jk follow [7] for ease of comparison.
q =1
k
q =2
k
p=2
j
p =1
j
Figure 6.1: The notation for the definition of the RWG [21] functions in this Chapter. The
function is also used as a testing function.
6.4 The ∇ × Ge Operator

In addition to calculating the electric field due to an electric current source, we need to find
the magnetic field due to an electric current source. Then the other operators can be found
by invoking duality.
By taking the curl of the Ge operator, the primary field term from (6.4), is given by
0
P eikm |r−r |
∇× Ge (r, r0 ) =∇×I (6.36)
4π|r − r0 |
By taking the curl of the secondary field term, from (6.19), we have
TM,S µn
∇ × Ge (r, r0 ) = (∇ × ẑ)(∇0 × ∇0 × ẑ)g TM,S (r, r0 ) (6.37)
µm
From (6.20), we have
TE,S
∇ × Ge (r, r0 ) = (∇ × ∇ × ẑ)(∇0 × ẑ)g TE,S (r, r0 ) (6.38)
As Z
H(r) = ∇ × Ge (r, r0 ) · J(r0 )dr0 (6.39)
V
Equations (6.37) and (6.38) will generate the magnetic field when operating on the electric
current. In fact (6.37) and (6.38) are related by duality.
When we need to find the matrix representation of the above operators, we rewrite
TM,S µn
∇ × Ge (r, r0 ) = (∇ × ẑ)(∇0 ∇0 · ẑ + km
2
ẑ)g TM,S (r, r0 ) (6.40)
µm
Consequently, we have
TM,S
hJT i (r), ∇ × Ge (r, r0 ), Jj (r0 )i
µn £
= −hJT i (r), ∇ × ẑ∂z0 g TM,S (r, r0 ) , ∇0 · Jj (r0 )i
µm
2
¤
+km hJT i (r), ∇ × ẑg TM,S (r, r0 ) , ẑ · Jj (r0 )i (6.41)
TE,S
Similarly, for ∇ × Ge (r, r0 ), we have
TE,S
∇ × Ge (r, r0 ) = (∇∇ · ẑ + km
2
ẑ)(∇0 × ẑ)g TE,S (r, r0 ) (6.42)
TE,S
hJT i (r), ∇ × Ge (r, r0 ), Jj (r0 )i = −h∇ · JT i (r), ∇0 × ẑ∂z g TE,S (r, r0 ) , Jj (r0 )i
2
+km hJT i (r) · ẑ, ∇0 × ẑg TE,S (r, r0 ) , Jj (r0 )i (6.43)
When JT i and Jj are curl-conforming volume expansion functions, an additional inte-

gration by parts can be used to move the ∇× operators onto them. However, for surface
currents, which are distributed over a thin sheet in space, their curls are singular, and hence,
it is prudent to leave the curl operators the way they are in the above.
6.5 The L and K Operators

The Ge and ∇ × Ge operators above allow us to formulate the EFIE and MFIE for impen-
etrable objects embedded in layered media. To formulate integral equations for penetrable
objects in a layered medium, it is necessary to derive the full suite of L and K operators as
have been done for the homogeneous medium case, which are
Z
E(r) = iω Ge (r, r0 )µ(r0 ) · J(r0 )
= Le (r, r0 ) · J(r0 ) (6.44)
where
Le (r, r0 ) = iωGe (r, r0 )µ(r0 ), (6.45)
and
Z
H(r) = µ−1 (r) ∇ × Ge (r, r0 )µ(r0 ) · J(r0 )
= Km (r, r0 ) · J(r0 ) (6.46)
where
Km (r, r0 ) = µ−1 (r)∇ × Ge (r, r0 )µ(r0 ) (6.47)
and in the above, integration is implied over repeated variables.

By invoking duality, it follows that
Z
H(r) = iω Gm (r, r0 )²(r0 ) · M(r0 )
= Lm (r, r0 ) · M(r0 ) (6.48)
Z
−1
E(r) = −² (r) ∇ × Gm (r, r0 )²(r0 ) · M(r0 )
= Ke (r, r0 ) · J(r0 ) (6.49)
where
Lm (r, r0 ) = iωGm (r, r0 )²(r0 ) (6.50)
Ke (r, r0 ) = −²−1 (r)∇ × Gm (r, r0 )²(r0 ) (6.51)
In the above,
Gm (r, r0 ) = (∇ × ẑ)(∇0 × ẑ)g TM (r, r0 )

1
+ 2 (∇ × ∇ × ẑ)(∇0 × ∇0 × ẑ)g TE (r, r0 ) , |z − z 0 | > 0, (6.52)
kmn
6.6 The Ez -Hz Formulation

The Ez -Hz formulation was used in some of our earlier works [8–10,22–25]. For completeness,
we will describe it here. In this formulation, the Ez and Hz components of the electromagnetic
fields are first derived in the layered medium due to a point source, and other components
can then be derived from them [4]. But in this section, we will derive the fields so that there
will be less derivative operators operating on these spectral integrals. This will enable the
easier evaluations of the spectral integrals.
Consider an arbitrary Hertzian dipole point source J(r) = α̂0 I0 `δ(r−r0 ), carrying a current
I0 with effective length `, and located in region m in a layered medium. For simplicity, we
can assume that I0 ` = 1. The electric field it produces can be derived, and from these
expressions, we can glean off the dyadic Green’s function. By breaking an arbitrary point
source into vertical and horizontal components, the solution due to this point source can be
written in the spectral domain as
µ ¶
1 ∂2 1 ∂
Ẽmz = ẑ · α̂0 1 + 2 TM
g̃m − 2 ∇s · α̂0 g̃m
TM
, (6.53)
km ∂z 2 km ∂z 0
where the first term is due to the vertical component of the source whereas the second term
is due to its horizontal component. In general,
TM T M,P T M,R
g̃m = g̃m + g̃m (6.54)
where 0 0
T M,P −ωµm eiks ·(rs −rs )+ikmz |z−z |
g̃m = , (6.55)
8π 2 kmz
0
T M,R −ωµm eiks ·(rs −rs ) £ T M −ikmz z T M ikmz z
¤
g̃m = 2
Bm e + Dm e . (6.56)
8π kmz
TM TM T M,P
Bm and Dm can be found as in [6]. Here, g̃m is the primary field contribution whereas
T M,R T M,P
g̃m is due to reflections at boundaries at z = −dm−1 and z = −dm . Because g̃m is
T M,R
the same as the homogeneous medium case, we can focus on g̃m . Moreover, Ẽmz due to
T M,R ∂2 2
g̃m can be further simplified as using ∂z 2 → −kmz to obtain
µ 2
¶
R 0 ks 1 ∂ 0 T M,R
Ẽmz = ẑ · α̂ 2 − 2 ∇s · α̂ g̃m . (6.57)
km km ∂z 0
Using [6, Eq. (2.3.17)] and from the above, we can derive the other components of the TM
electric field to obtain
µ ¶
1 ∂ R 1 ∂ k2 1 ∂
ẼTms
M,R
= 2 ∇s Ẽmz = 2 ∇s ẑ · α̂0 2s − 2 ∇ s · α̂ 0 T M,R
g̃m . (6.58)
ks ∂z ks ∂z km km ∂z 0
Adding the z component to the above, we have
µ
1 1 ∂ ∂
ẼTmM,R = 2
∇s ∇z · α̂0 − 2 2 ∇s ∇s · α̂0
km km ks ∂z ∂z 0
2
¶
0 ks 1 0 0 T M,R
+ẑ ẑ · α̂ 2 − 2 ∇z ∇s · α̂ g̃m . (6.59)
km km
We can write
1 1 1 kz2 1 ks2
2
∇ s ∇ z = 2
(∇∇ z − ∇ z ∇ z ) = 2
∇∇ z + ẑ ẑ 2
= 2
∇∇ z + ẑ ẑ − ẑ ẑ 2
. (6.60)
km km km km km km
µ ¶
1 1 ∂ ∂ 1 0
ẼTmM,R = 2
∇∇ z − 2 k 2 ∂z ∂z 0
∇ s ∇ s + ẑ ẑ − 2
∇ z ∇ s · α̂0 g̃m
T M,R
. (6.61)
km km s km
T M,R
We can always define a new function gmh such that
∂ T M,R ∂ T M,R
0
g̃m = − g̃mh . (6.62)
∂z ∂z
T M,R
It can be shown that such a gmh can always be found. In this case, (6.61) can be written
as
µ ¶ µ ¶
T M,R 1 0 T M,R 1 1 ∂2 T M,R
Ẽm = 2
∇∇z + ẑ ẑ · α̂ g̃m + 2 ∇s ∇s + ∇z ∇s · α̂0 g̃mh . (6.63)
km km ks2 ∂z 2
∂2 2 2
Using ∂z 2 = −kmz = −km + ks2 , the above can be rewritten as
µ ¶ µ ¶
T M,R 1 0 T M,R 1 1 T M,R
Ẽm = 2
∇∇z + ẑ ẑ · α̂ g̃m + 2
∇∇s − 2 ∇s ∇s · α̂0 g̃mh . (6.64)
km km ks
For the TE wave, we can focus again on the reflected wave to obtain the spectral domain
solution as
R 1 1
H̃mz = ẑ · ∇s × α̂0 T E,R
g̃m = α̂0 · ẑ × ∇s g̃ T E,R , (6.65)
iωµm iωµm m
where 0
ωµm eiks ·(rs −rs ) £ T E −ikmz z
T E,R T E ikmz z
¤
=− 2
g̃m Bm e + Dm e . (6.66)
8π kmz
Then, the TE electric field in region m is
1
α̂ · ẼTms
E,R
= α̂ · ẼTmE,R = − (α̂ · ẑ × ∇s ) (α̂0 · ẑ × ∇s ) g̃m
T E,R
. (6.67)
ks2
Using the fact that (an identity easily derived in the spectral domain),
1 1
(α̂ · ẑ × ∇s ) (α̂0 · ẑ × ∇s ) = −α̂ · ẑ ẑ · α̂0 − 2 α̂ · ∇s ∇s · α̂0 − α̂ · α̂0
ks2 ks
1
= −α̂ · Īs · α̂0 − 2 α̂ · ∇s ∇s · α̂0 , (6.68)
ks
where Īs = x̂x̂ + ŷ ŷ, we can rewrite (6.67) as

µ ¶
1
T E,R
Ẽm = Īs + 2 ∇s ∇s · α̂0 g̃m
T E,R
. (6.69)
ks
Combining (6.64) and (6.69), we have the total reflected electric field as
µ ¶ µ ¶
1 1 1 T M,R
ẼR
m = 2
∇∇ z + ẑ ẑ · α̂ 0 T M,R
g̃m + 2
∇∇ s − ∇ s ∇ s · α̂0 g̃mh
km km ks2
µ ¶
1
+ Īs + 2 ∇s ∇s · α̂0 g̃m T E,R
. (6.70)
ks
Because the electric field due to a point current source J(r) = α̂δ(r − r0 ) is [see (6.21)]
ẼR ˜R 0 0 0
m (r) = iω G (r, r )µ(r ) · α̂ (6.71)
one can extract an expression for the dyadic Green’s function in the spectral domain for the
reflected wave term from the above.
The field in region n < m due to a source in region m is of the form
µ ¶
k2 1 ∂
Ẽnz = ẑ · α̂0 2s − 2 ∇ s · α̂ 0
g̃nT M , (6.72)
km km ∂z 0
where
ωµm eiks ·(rs −rs ) ²m T M h iknz z i

0
g̃nT M = − 2
An e TM
+ R̃n,n−1 e−2iknz dn−1 −iknz z . (6.73)
8π kmz ²n
ATn E can be found using the technique described in [6, pp. 53, 78]. The transverse components
of the field can be found to be
µ ¶
TM 1 ∂ 1 ∂ 1 ∂2
Ẽns = 2 ∇s Ẽnz = 2 ∇s ẑ − 2 ∇s ∇s · α̂0 g̃nT M . (6.74)
ks ∂z km ∂z ks ∂z∂z 0
Adding the z component to the above, we have
µ ¶
TM 1 1 ∂2 ks2 1 0
Ẽn = 2
∇s ∇z − 2 2 ∇s ∇s + ẑ ẑ 2 − 2 ∇z ∇s · α̂0 g̃nT M . (6.75)
km km ks ∂z∂z 0 km km
By using the same trick as before, the first term and third term can be combined, viz,
2
1 1 1 1 knz 1 kn2 ks2
2
∇ s ∇ z = 2
∇∇ z − 2
∇ z ∇ z = 2
∇∇ z + 2
ẑ ẑ = 2
∇∇ z + 2
ẑ ẑ − 2
ẑ ẑ. (6.76)
km km km km km km km km
Similar to before, we can define
∂ TM ∂ TM
0
g̃n = − g̃nh , (6.77)
∂z ∂z
to combine the second and the fourth terms, viz,
1 ∂2 1 ∂2 k2
− 2
∇s ∇s − ∇0z ∇s = 2 2 ∇s ∇s + ∇z ∇s = − nz ∇s ∇s + ∇z ∇s
ks ∂z∂z 0 ks ∂z ks2
k2
= − n2 ∇s ∇s + ∇∇s . (6.78)
ks
Consequently, we have
µ ¶
1 ¡ ¢ 1 kn2
ẼTn M = 2
∇∇z + kn2 ẑ ẑ · α̂0 g̃nT M − 2 ∇s ∇s − ∇∇s · α̂0 g̃nh
TM
, (6.79)
km km ks2
The TE field in region n < m in the spectral domain is

1 TE
H̃nz = (α̂0 · ẑ × ∇s ) g̃ ,
iωµn n
where
ωµm eiks ·(rs −rs ) µm T E h iknz z i
0
g̃nT E = − A e + R̃ TE
e−2iknz dn−1 −iknz z
(6.80)
8π 2 kmz µn n n,n−1
The TE field can be shown to be

µ ¶
∇ s ∇s
ẼTn E = Īs + · α̂0 g̃nT E . (6.81)
ks2
The total electric field is then

µ ¶
1 ¡ ¢ 1 kn2
Ẽn = 2
∇∇z + kn2 ẑ ẑ · α̂0 g̃nT M − 2 ∇ ∇
s s − ∇∇ 0 TM
s · α̂ g̃nh
km km ks2
µ ¶
∇ s ∇s
+ Īs + · α̂0 g̃nT E . (6.82)
ks2
Similar to (6.70), one can extract an expression for the dyadic Green’s function in the spectral
domain by noticing that
˜
Ẽn (r) = iω G(r, r0 )µ(r0 ) · α̂0 (6.83)
When the above is used to compute the MOM matrices, they need to be integrated to
yield the spatial domain fields. It is to be noted that when the spectral domain integral is
performed, the derivative operators are taken out of the spectral integrals. Integration by
parts are then used to move the derivative operators to the expansion and testing functions.
6.6.1 Example 1: Half Space

Let us assume that m = N = 2, so that it corresponds to two half spaces (see Figure 6.2). In
TM
this case, Dm = 0 in (6.56) and
0
TM
Bm = B2T E = R12
T M −ik2z (d1 +z )−ik2z d1
e . (6.84)
0
ωµ2 eiks ·(rs −rs ) T M −ik2z (d1 +z0 )−ik2z d1 −ik2z z
g̃2T M,R = − R12 e . (6.85)
8π 2 k2z
T M,R
Also, g̃2h defined in (6.62) is
T M,R
g̃2h = −g̃2T M,R . (6.86)
Region 1
− d1
Region 2
z = z'
Figure 6.2: A point source in a lower half space.
g̃2T E,R is similar to (6.85) with R12 TE

replacing R12 TM
. Consequently, (6.70) for this example
becomes
µ ¶
1 1
ẼR
2 = ∇∇ z + ẑ ẑ · α̂g̃2T M,R − 2 ∇∇s · α̂0 g̃2T M,R + Īs · α̂0 g̃2T E,R
ks2 k2
1
+ 2 ∇s ∇s · α̂0 (g̃2T E,R + g̃2T M,R )
ks
1 ˆ
= − 2 ∇∇ · α̂0 g̃2T M,R + ẑ ẑ · α̂0 g̃2T M,R + Īs · α̂0 g̃2T E,R
k2
1
+ 2 ∇s ∇s · α̂0 (g̃2T E,R + g̃2T M,R ), (6.87)
ks
ˆ = ∇s − ∇ z .
where ∇
In region 1, g̃1T M of (6.73) becomes
0
ωµ2 eiks ·(rs −rs ) ²2 T M ik1z z
g̃1T M = − · A1 e , (6.88)
8π 2 k2z ²1
where
0
AT1 M = e−ik2z (d1 +z ) T21
T M ik1z d1
e . (6.89)
TM
In this case g̃1h defined in (6.77) becomes,
TM k2z T M
g̃1h = g̃ , (6.90)
k1z 1
and (6.82) becomes

µ ¶
1 ¡ ¢ 1 k22 k2z T M
Ẽ1 = ∇∇z + k12 ẑ ẑ · α̂0 g̃1T M − 2 ∇ s ∇ s − ∇∇ s · α̂0 g̃
2
k1 k2 k 2 k1z 1
µs ¶
∇s ∇s
+ Īs + · α̂0 g̃1T E , (6.91)
ks2
where g̃1T E is similar to g̃1T M in (6.88), but with ( ²²12 )AT1 M replaced by ( µµ12 )AT1 E , and AT1 E is
similar to AT1 M in (6.89) but with T21 TM
replaced by T21 TE
, or
0
ωµ2 eiks ·(rs −rs ) µ2 −ik2z (d1 +z0 ) T E ik1z d1 +ik1z z
g̃1T E = − e T21 e . (6.92)
8π 2 k2z µ1
6.6.2 Example 2: Three-Layer Medium
Region 1
Region 2 − d1
z = z'
− d2
Region 3
Figure 6.3: A point source embedded in a three-layer medium.
For field in region 2, B2T M and D2T M of (6.56) can be obtained from [6, Eq. (2.4.5)].
h 0
i
T M ik2z (d2 +z 0 )
B2T M = e−ik2z d1 R21
TM
e−ik2z (d1 +z ) + eik2z (d2 −d1 ) R23 e M̃2 , (6.93)
h 0
i
T M −ik2z (d1 +z 0 )
D2T M = eik2z d2 R23
TM
eik2z (d2 +z ) + eik2z (d2 −d1 ) R21 e M̃2 , (6.94)
where h i−1
T M T M 2ik2z (d2 −d1 )
M̃2 = 1 − R23 R21 e , (6.95)
and 0
ωµ2 eiks ·(rs −rs ) £ T M −ik2z z ¤
g̃2T M,R =− 2 B2 e + D2T M eik2z z . (6.96)
8π k2z
T M,R
Moreover, g̃2h as defined in (6.62) is
ωµ2 eiks ·(rs −rs ) k1z h i

0
T M,R T M −ik2z z T M ik2z z
g̃2h = −B̂2 e + D̂ 2 e . (6.97)
8π 2 k2z k2z
where
h 0
i
T M ik2z (d2 +z 0 )
B̂2T M = e−ik2z d1 R21
TM
−e−ik2z (d1 +z ) + e−ik2z (d2 −d1 ) R23 e M̃2 , (6.98)
h 0
i
T M −ik2z (d1 +z 0 )
D̂2T M = eik2z d2 R23
TM
eik2z (d2 +z ) − eik2z (d2 −d1 ) R21 e M̃2 . (6.99)

µ ¶ µ ¶
R 1 0 T M,R 1 1 T M,R
Ẽ2 = ∇∇z + ẑ ẑ · α̂ g̃2 + ∇∇s − 2 ∇s ∇s · α̂0 g̃2h
k22 k22 ks
µ ¶
∇s ∇s
+ Īs + · α̂0 g̃2T E,R . (6.100)
ks2
In region 1, g̃1T M of (6.73) is similar to (6.88), but with

h 0 0
i
AT1 M = T21
T M −ik2z d1
e e−ik2z z + eik2z (z +2d1 ) R23
TM
M̃2 eik1z d1 (6.101)
TM
which can be gathered from (2.4.8) and (2.4.9) of [6]. In this case, g̃1h defined in (6.77)
becomes
0
TM ωµ2 eiks ·(rs −rs ) k2z ²2 T M ik1z z
g̃1h =− 2 Â e , (6.102)
8π k2z k1z ²1 1
where h i
0 0
ÂT1 M = T21
T M −ik2z d1
e −e−ik2z z + eik2z (z +2d1 ) R23
TM
M̃2 eik1z d1 . (6.103)

µ ¶
1 ¡ ¢ 1 k12
Ẽ1 = 2 ∇∇z + k12 ẑ ẑ · α̂0 g̃1T M − 2 ∇s ∇s − ∇∇s · α̂0 g̃1h TM
k2 k2 ks2
µ ¶
∇ s ∇s
+ Īs + · α̂0 g̃1T E , (6.104)
ks2
µ2 T E ²2 T M
where g̃1T E is similar to g̃1T M defined in (6.88), but with µ1 A1 replacing ²1 A 1 , and also
AT1 E is similar to (6.101) but with TE replacing TM.
6.7 Validation and Result

From these integrals, this new approach gives rise to matrix elements which agree with the
formulation derived in [10] term by term. It also verifies its validity, because the results
in [10] have been derived differently using Ez and Hz formulation. We have also checked the
expressions in (6.28) against those derived by the Ez -Hz formulation for a general layered
medium, and they are proved to be equivalent to each other [26].
Another validation is by comparing the radar cross section (RCS) of a sphere above a
three-layered medium when calculated by the approach of Michalski-Zheng formulation [7]
and this one. The incident wave is at an elevation angle of 30◦ and an azimuth angle of 0◦ ,
with the operating frequency at 300 MHz. The layer parameters are µ0 , ²1 = 1.0, ²2 = 2.56,
²3 = 6.5+0.6i. The first layer is the upper half-space of infinite size, second one with thickness
of 0.3 m, and the third one is the lower half-space of infinite size. The PEC sphere has a
radius of 1m and is placed in layer 1, 10 m above the first interface. Figure 6.4 demonstrates
the excellent agreement of the two formulations.

q
φ−plane at f = 0.3 GHz , El = 30 deg. File:spherepec.rcs

30
20
10
σdBsm
0
−10
VV, [6]
−20 HH, [6]
HH, Eqn. (22)
VV, Eqn. (22)
−30
−200 −150 −100 −50 0 50 100 150 200
Azimuth [deg]
Figure 6.4: The radar cross section of a PEC sphere above a three-layered medium.
6.8 Conclusions
We have arrived at an alternative and elegant way of deriving the matrix representation of
the dyadic Green’s operator in the electric field integral equation (EFIE). The derivation is
based on the pilot vector potential approach where the two vector potentials in terms of TE
and TM fields are derived.
Starting with the dyadic Green’s function formulas in [6], we manipulate them into a
form that is suitable for matrix-element evaluation for CEM, for example, in the MOM.
All integrals involve only Bessel function of zeroth order, and they can be reduced to two
basic integrals for the Green’s function g TM and g TE . When these two Green’s functions are
tabulated, the other integrals needed can be obtained by finite difference on the tabulated
values. This derivation is more succinct than previous derivations.
Appendix: Singularity Subtraction for Scalar Potentials

The dyadic Green’s function introduced in Section 6.2 requires the evaluation of two canon-
ical scalar potentials given by (6.15) and (6.16). But the integrands of these integrals have
pole singularities at the origin, which disappear (or cancel each other) when they are sub-
stituted into the equation for the dyadic Green’s function. Hence, these singularities do not
contributed to the final value of the dyadic Green’s function, and they can be extracted and
ignored. This shall be discussed here.
A typical integral for the scalar potential is a divergent integral of the form
Z ∞
F (kρ , z, z 0 , ρ0 )
I= dkρ (A-1)
0 kρ
for a layered medium Green’s function. The above can be rewritten as
Z ∞ · ¸
1 0 0 0 0 a2
I= dkρ F (kρ , z, z , ρ ) − F (0, z, z , ρ ) 2
0 kρ kρ + a2
Z ∞
1
+ F (0, z, z 0 , ρ0 )a2 dkρ ¡ 2 ¢
0 kρ kρ + a2
= IR + IS (A-2)
where
Z ∞ · ¸
1 a2
IR = dkρ F (kρ , z, z 0 , ρ0 ) − F (0, z, z 0 , ρ0 ) 2 (A-3)
0 kρ k ρ + a2
Z ∞
1
IS = F (0, z, z 0 , ρ0 )a2 dkρ ¡ 2 2
¢ (A-4)
0 k ρ k ρ +a
IR is now a convergent integral which may be tabulated. IS is singular, but can be integrated
in closed form. By partial fraction
Z ∞ · ¸
0 0 1 kρ
IS = lim F (0, z, z , ρ ) dkρ − 2
²→0 ² kρ k ρ + a2
· ¸
1 ¡ ¢ ∞
= F (0, z, z 0 , ρ0 ) lim ln kρ − ln kρ2 + a2 (A-5)
²→0 2 ²
or
· ¸
1 ¡ ¢
IS = F (0, z, z 0 , ρ0 ) lim − ln ² + ln ²2 + a2
²→0 2
= −F (0, z, z 0 , ρ0 ) lim ln ² + F (0, z, z 0 , ρ0 ) ln a (A-6)
²→0
The singular term in the above will cancel each other in the full dyadic Green’s function or
else it will contribute to a nonphysical field. Hence, it can be ignored a priori and
IS = F (0, z, z 0 , ρ0 ) ln a (A-7)
The above singularity subtraction technique is nonunique, and other ways of subtraction that
yield better convergence for the integral are possible.
Bibliography
[1] R. F. Harrington, Field Computation by Moment Methods, Melbourne, FL: Krieger Publ.,
1982.
[2] W. C. Chew, J. L. Xiong, and M. A. Saville, “A matrix-friendly formulation of layered

medium Green’s function,” Antennas Wirel. Propag. Lett., vol. 5, no. 1, pp. 490–494, Dec.
2006.
[3] A. Sommerfeld, “Uber die Ausbreitung der Wellen in der drahtlosen Telegraphie.” Ann.
Phys., vol. 28, pp. 665–737, 1909.
[4] J. A. Kong, “Electromagnetic field due to dipole antennas over stratified anisotropic me-
dia.” Geophys., vol. 38, pp. 985–996, 1972.
[5] L. Tsang, J. A. Kong, and R. T. Shin, Theory of Microwave Remote Sensing, Melbourne,
FL: Wiley Interscience, 1985.
[7] K. A. Michalski and D. Zheng, “Electromagnetic scattering and radiation by surfaces of

arbitrary shape in layered media, Part I: Theory,” IEEE Trans. Antennas Propag., vol.
38, pp. 335–344, 1990.
[8] J. S. Zhao, W. C. Chew, C. C. Lu, E. Michielssen, and J. M. Song, “Thin-stratified medium

fast-multipole algorithm for solving microstrip structures,” IEEE Trans. Microwave The-
ory Tech., vol. 46, no. 4, pp. 395–403, Apr. 1998.
[9] T. J. Cui and W. C. Chew, “Fast evaluation of sommerfield integrals for em scattering
and radiation by three-dimensional buried objects,” IEEE Geosci. Remote Sens., vol. 37,
no. 2, pp. 887–900, Mar. 1999.
[10] W. C. Chew, J. S. Zhao, and T. J. Cui, “The layered medium Green’s function—A new
look,” Microwave Opt. Tech. Lett., vol. 31, no. 4, pp. 252–255, 2001.
[11] G. Dural and M. I. Aksun, “Closed-form Green’s functions for general sources and strat-
ified media,” IEEE Trans. Microwave Theory Tech., vol. 43, pp. 1545–1552, Jul. 1995.
1
[12] Y. L. Chow, N. Hojjat, S. Safavi-Naeini, “Spectral Green’s functions for multilayer media
in a convenient computational form,” IEE Proc.-H., vol. 145, pp. 85–91, Feb. 1998.
[13] K. A. Michalski, “Extrapolation methods for Sommerfeld integral tails (Invited Review
Paper),” IEEE Trans. Antennas Propag., vol. 46, pp. 1405–1418, Oct. 1998.
[14] W. Cai and T. Yu, “Fast calculation of dyadic Green’s functions for electromagnetic
scattering in a multi-layered medium,” J. Comput. Phys., vol. 165, pp. 1–21, 2000.
[15] F. Ling and J. M. Jin, “Discrete complex image method for Green’s Functions of general
multilayer media,” IEEE Microwave Guided Wave Lett., vol. 10, no. 10, pp. 400–402, Oct.
2000.
[16] J. Y. Chen, A. A. Kishk, and A. W. Glisson, “Application of new MPIE formulation
to the analysis of a dielectric resonator embedded in a multilayered medium coupled to
a microstrip circuit,” IEEE Trans. Microwave Theory Tech., vol. 49, pp. 263–279, Feb.
2001.
[17] T. M. Grzegorczyk and J. R. Mosig, “Full-wave analysis of antennas containing hori-
zontal and vertical metallizations embedded in planar multilayered media,” IEEE Trans.
Antennas Propag., vol. 51, no. 11, pp. 3047–3054, Nov. 2003.
[18] E. Simsek, Q. H. Liu, and B. Wei, “Singularity subtraction for evaluation of Green’s
functions for multilayer media,” IEEE Trans. Microwave Theory Tech., vol. 54, no. 1, pp.
216–225, Jan. 2006.
[19] W. W. Hansen, “Transformations useful in certain antenna calculations,” J. Appl. Phys.,
vol. 8, pp. 282–286, Apr. 1937.
[20] W. C. Chew and S. Y. Chen, “Response of a point source embedded in a layered medium,”
Antennas Wirel. Propag. Lett., vol. 2, no. 14, pp. 254–258, 2003.
[21] S. M. Rao, D. R. Wilton, and A. W. Glisson, “Electromagnetic scattering by surfaces of
arbitrary shape,” IEEE Trans. Antennas Propag., vol. 30, no. 3, pp. 409-417, May 1982.
[22] B. Hu, W. C. Chew, and S. Velamparambil, “Fast inhomogeneous plane wave algorithm
for the analysis of electromagnetic scattering,” Radio Sci., vol. 36, no. 6, pp. 1327–1340,
2001
[23] B. Hu, W. C. Chew, E. Michielssen and J. S. Zhao, “Fast inhomogeneous plane wave
algorithm for the fast analysis of two-dimensional scattering problems,” Radio Sci., vol.
34, pp. 759–772, Jul.–Aug. 1999.
[24] B. Hu and W. C. Chew, “Fast inhomogeneous plane wave algorithm for electromagnetic
solutions in layered medium structures - 2D case,” Radio Sci., vol. 35, pp. 31–43, Jan.–Feb.
2000.
[25] B. Hu and W. C. Chew, “Fast inhomogeneous plane wave algorithm for scattering from
objects above the multi-layered medium,” IEEE Geosci. Remote Sens., vol. 39, no. 5, pp.
1028–1038, May 2001.
[26] W. C. Chew, “Field due to a point source in a layered medium—symmetrized formula-

tion,” Private Notes, Mar. 25, 2003.
159
CHAPTER 7
Fast Inhomogeneous Plane Wave

Algorithm for Layered Media
7.1 Introduction
Although new developments in fast algorithm result in the ability to solve large-scale electro-
magnetic simulation problem with presently available computational resources, all of them are
targeted to problems with homogeneous background, also known as free-space (homogeneous
space) problems. However, there are numerous applications of practical interests, where the
free-space approximation is not sufficient. For example, correct modeling of electromagnetic
phenomena in areas such as remote sensing, interconnect characterization, microstrip anten-
nas and monolithic microwave integrated circuits (MMIC), calls for the capability to analyze
structures in the vicinity of stratified medium containing an arbitrary number of dielectric
layers. The casting of this problem into integral based form was systematically addressed by
Michalski and Zheng [1], where several versions of dyadic Green’s function for layered medium
problem were derived. Later, Chew et al. developed an alternative formulation [2–4] and suc-
cessfully applied to the development of fast method for layered medium [5,6]. As shown later,
this formulation is more easily adapted to fast methods developed here and results in similar
efficiency as that for free-space problems.
Although theoretical development for solving layered medium problem using integral equa-
tion based solvers was considered complete, two obstacles exist for solving such problems effi-
ciently. The first difficulty comes from the necessity to fill the impedance matrix when using
the traditional solution method. The core part is to evaluate the spatial Green’s functions,
which are expressed in the Sommerfeld-type integral forms. Because no general analytic closed
forms are available, the numerical integration along the Sommerfeld integration path (SIP)
is necessary and it is quite time consuming. Many techniques have been proposed to reduce
this computation time. The majority of those techniques can be categorized into two main
approaches. The first approach is represented by well known discrete complex image method
(DCIM) [7–10]. The main idea is to fit the spectrum of dyadic Green’s function encountered
in Sommerfeld integrals using a set of special functions, for which closed-form Sommerfeld
1
integration results exist. This approach garners significant attention in recent years because
of its simplicity and ease of implementation. The second approach is to perform Sommer-
feld integration through a cleverly designed integration path so that fast convergence can be
achieved in numerical integration. For example, Cui and Chew [6] have demonstrated the
efficiency to perform the numerical integration along the fast converging steepest descent path
(SDP). By using these techniques, the impedance matrix filling time can be reduced to be in
the same order of magnitude as its free-space counterpart. However, the total computational
complexity to fill the impedance matrix remains O(N 2 ).
Although the matrix-filling issue for layered medium problems can be mitigated using
aforementioned techniques, the second difficulty, that is, solving the resultant matrix equation
within a reasonable amount of time, poses a major hurdle for solving the practical problems.
Similar to free-space problems, the impedance matrix resulted from MOM discretization for
layer medium problems is a dense, complex one in general. Because the straightforward fast
multipole method (FMM) implementation for layered medium problem is not possible, there
are multiple efforts to develop efficient methods for large-scale layered medium problems. For
example, the adaptive integral method (AIM) [11] coupled with DCIM [12], thin-stratified
medium fast-multipole algorithm (TSM-FMA) developed by Zhao et al. to solve microstrip
type of problems [5], steepest descent fast multipole method (SDFMM) developed by Jand-
hyala et al. [13], MLFMA-based implementatation using the asymptotic image term developed
by Geng et. al [14], and FMM-PML-MPIE developed by Vande Ginste et al. [15, 16].
In this chapter, we will review the recent progress on fast inhomogeneous plane wave
algorithm (FIPWA) and its application to layered medium problems.
7.2 Integral Equations for Layered Medium

In this section, without loss of generality of describing the algorithm involved, the electromag-
netic scattering from a three-dimensional perfect electric conducting (PEC) surface scatterer
S, in the presence of a layered medium, is studied. The stratified medium is characterized by
N layers with different dielectric properties as shown in Figure 7.1. The air-dielectric interface
is assumed to be at z = 0, and the scatterer is above all planar dielectric layers. To compute
the scattering from the target, the electric field integral equation (EFIE) for inhomogeneous
medium is formulated as
Z
iωn̂ × Ge (r, r0 )µ(r0 ) · J(r0 )dr0 = −n̂ × Einc (r), ∀r ∈ S, (7.1)
S
and the magnetic field integral equation (MFIE) for inhomogeneous medium is
Z
1
−1
µ(r) n̂ × ∇ × Ge (r, r0 )µ(r0 ) · J(r0 )dS 0 − J(r) = −n̂ × Hinc (r), (7.2)
S 2
where n̂ is the normal outward vector at point r, Ge (r, r0 ) is the dyadic Green’s function, J(r0 )
is the surface current distribution, and Einc (r) and Hinc (r) are the fields of the impinging
plane wave in the presence of the layered medium.
To reduce the effect of the internal resonances (see Chapters 3 and 4), combined field
integral equation (CFIE) for closed conducting objects is utilized and it is simply a linear
FAST INHOMOGENEOUS PLANE WAVE ALGORITHM FOR LAYERED MEDIA 161
1 air-dielectric interface
2
3
N-1
. . .
N
Figure 7.1: The PEC scatterer S above a layered medium. The outermost layers are Layer 1
and Layer N .
combination of EFIE and MFIE, that is,
CFIE = αEFIE + η(1 − α)MFIE (7.3)
with α ∈ [0, 1] and normally α = 0.2 ∼ 0.5 is chosen.

Before discretizing Eqs. (7.1)–(7.3), the symmetric dyadic Green’s function can be derived
as follows [3] (see also Chapter 6):
iω α̂ · G(r, r0 )µ(r0 ) · α̂0 = (α̂s · α̂s0 )(g0 − g T E ) + αz αz0 (g0 + g T M )

1 1
+ 2 α̂ · ∇∇ · α̂0 g0 + 2 α̂ · ∇∇ · α̂00 g T M
k k
+α̂ · ∇s ∇s · α̂00 g EM , (7.4)
where 0
iωµ eik1 |r−r |
g0 (r, r0 ) = , (7.5)
4π |r − r0 |
Z
g β (r, r0 ) = dkρ g̃ β (kρ , r, r0 ), (7.6)
SIP
β = (T E), (T M ), EM, (7.7)

α̂ = α̂s + α̂z , (7.8)
0
α̂ = α̂s0 + α̂z0 , (7.9)
Im{k ρ }
Re{k ρ}
Figure 7.2: Illustration of the Sommerfeld integration path (SIP). The shift of the path is to
avoid the singularities at kρ = 0 and the surface wave poles.
α̂00 = −α̂s0 + α̂z0 , (7.10)

ωµ kρ T M (1) 0
g̃ T M (kρ , r, r0 ) = − R̃ (kρ )H0 (kρ |ρ − ρ0 |)eik1z (z+z ) , (7.11)
8π k1z
ωµ kρ T E (1) 0
g̃ T E (kρ , r, r0 ) = R̃ (kρ )H0 (kρ |ρ − ρ0 |)eik1z (z+z ) , (7.12)
8π k1z
1 ¡ ¢
g̃ EM (kρ , r, r0 ) = 2 g̃ T E (kρ , r, r0 ) − g̃ T M (kρ , r, r0 ) , (7.13)
kρ
q
and k1 is the wave number for the residing layer of the scatterer, k1z = k12 − kρ2 , and R̃T E
and R̃T M represent the generalized reflection coefficients for the layered medium, as defined
in [17]. The spectral domain Green’s function in terms of cylindrical waves can be derived
from those of plane waves in Chapter 6 by using the relationship that
ZZ ∞ Z ∞
1 0
dks F (kρ )eiks ·(rs −rs ) = dkρ kρ J0 (kρ ρ)F (kρ )
2π −∞ 0
Z
1 ∞ (1)
= dkρ kρ H0 (kρ ρ)F (kρ ) (7.14)
2 −∞
It should be noted that all information related to the layered medium, such as the loca-
tion and the dielectric property of each layer, is embedded within the generalized reflection
coefficients. Therefore, the Green’s function presented in Eq. (7.4) is applicable to arbitrary
layered structures. In the above equations, g0 represents the free-space (or homogeneous
medium) Green’s function and the g T E and g T M denote the reflected wave (secondary field)
parts, which are introduced to account for the effect of the layered medium. The integration
in Eq. (7.6) is performed along the SIP in the kρ plane, which is shown in Figure 7.2.
Each component of the dyadic Green’s function can be cast in the following generic form:
Z
kρ (1)
g0 (rj , ri ) = dkρ H0 (kρ |ρj − ρi |)eik1z (zj −zi ) , (7.15a)
SIP k1z
Z
(1)
g R (rj , ri ) = dkρ W (kρ )H0 (kρ |ρj − ρi |)eik1z (zj +zi ) , (7.15b)
SIP
Im{k ρ }
Re{k ρ}
Figure 7.3: The folded SIP, which is used for the integral with the Bessel function kernel.
where W (kρ ) denote the weight function because it can be considered as the weights applied
to the inhomogeneous plane waves pointing to different directions. For the sake of clarity,
Eqs. (7.15a) and (7.15b) are used to describe the algorithm.
When |ρj − ρi | → 0, it is better to work with the following formula:
Z
kρ
g0 (rj , ri ) = 2 dkρ J0 (kρ |ρj − ρi |)eik1z (zj −zi ) , (7.16a)
F SIP k1z
Z
g R (rj , ri ) = 2 dkρ W (kρ )J0 (kρ |ρj − ρi |)eik1z (zj +zi ) , (7.16b)
F SIP
to avoid the divergence of the Hankel function. Here, FSIP stands for the folded SIP, as
shown in Figure 7.3.
Observation 1. It is easy to observe the similarity between g0 and g R in their spectral
integral representations. Because g0 represents the free-space Green’s function and FIPWA
has been developed to accelerate the matrix-vector multiplication for this kind of kernel, it is
suggested that the same techniques be applied to g R .
7.3 FIPWA for Free Space

FIPWA was originally developed as an alternative to FMM to accelerate the matrix vector
multiplication encountered in the iterative solution of the electromagnetic scattering problems
in free-space. It is briefly reviewed here for the sake of completeness. Readers are referred
to [18] for more details.
To simplify the discussion, the kernel of the free-space Green’s function, that is, eikr /r,
is used to illustrate the algorithm. By using the Weyl identity [17] and using the symmetric
property of the free-space Green’s function, its integral representation can be formulated as
Z 2π Z
eikrji ik
= dφ dθ sin θeik(θ,φ)·rji , (7.17)
rji 2π 0 F SIP
where k(θ, φ) = k k̂(θ, φ) = k(x̂ sin θ cos φ + ŷ sin θ sin φ + ẑ cos θ) and rji = rj − ri .
observation group
11
00
00rjC
C
11
00
11
00
11
j
θ0
rCC’
00
11
i
rC’i00
11 C’
source group
Figure 7.4: Illustration of the source and observation groups. The source group is centered
at rC 0 and the observation group is centered at rC .
Denoting rC 0 and rC as the centers of the groups containing the elements i and j, respec-
tively, and using rji = rjC + rCC 0 + rC 0 i , as shown in Figure 7.4, the above integral can be
written as
Z Z
eikrji ik 2π
= dφ dθ sin θeik(θ,φ)·rjC · eik(θ,φ)·rCC 0 · eik(θ,φ)·rC 0 i . (7.18)
rji 2π 0 F SIP
To evaluate the integral efficiently, the SDP can be defined. When the source and obser-
vation groups are aligned along the z axis, that is, rCC 0 = ẑzCC 0 , the SDP on the θ plane is
chosen and shown in Figure 7.5. The integration on φ is within [0, 2π] and can be performed
fairly easily.
We rewrite Eq. (7.18) as
Z 2π Z
eikrji
= dφ dθf (θ)B(θ, φ)eikzCC 0 cos θ , (7.19)
rji 0 Γ
where
ik
f (θ) = sin θ, B(θ, φ) = BjC · BC 0 i = eik(θ,φ)·rjC · eik(θ,φ)·rC 0 i , (7.20)
2π
and B(θ, φ) is the combination of the radiation and receiving pattern function for the source
and observation groups, respectively, and f (θ) is considered to be the weight function. De-
noting Ωq = (θq1 , φq2 ), Wq = (wq1 , wq2 ), and q = (q1 , q2 ) as the appropriate abscissas and
θI
SIP
SDP
θ0 π/2 θR
Path I
Path II
Figure 7.5: Illustration of the steepest descent path (SDP) in the θ plane for the free-space
problems with the SIP in the background. The θ0 is defined in Figure 7.4.
weights for the corresponding numerical quadrature methods for θ and φ along the SDP, Eq.
(7.19) can be evaluated as
eikrji X
= f (Ωq )B(Ωq )eikzCC 0 cos θq1 . (7.21)
rji q
To diagonalize the translation operator, the value of the pattern function along the SDP,
that is, B(Ωq ), can be obtained using the interpolation and extrapolation techniques as fol-
lows: X
B(Ωq ) = B(Ωs )I(Ωq , Ωs ), (7.22)
s
where I(Ωq , Ωs ) is the interpolation and extrapolation kernel and Ωs denotes the samples
on the real θ and φ axis. After substituting Eq. (7.22) into (7.21), it can be derived as
eikrji X
= f (Ωq )B(Ωq )eikzCC 0 cos θq1
rji q
" #
X X
= f (Ωq ) B(Ωs )I(Ωq , Ωs ) eikzCC 0 cos θq1
q s
X X
= B(Ωs ) f (Ωq )I(Ωq , Ωs )eikzCC 0 cos θq1
s q
X
= B(Ωs ) · TCC 0 (Ωs ). (7.23)
s
Here, TCC 0 (Ωs ) is the diagonal translator and Eq. (7.23) can be used to construct an FMM-
like fast algorithm to accelerate the matrix-vector multiplication. As detailed in [18], when
source and observation groups are not aligned in the z axis, the diagonal translation operator
can be obtained by performing the coordinate system rotation with the help of interpolation.
It has been observed that FIPWA achieves the diagonal translator using a different method
from that used by FMM. Except for the different diagonalization procedures, these two al-
gorithms are quite similar. Therefore, the computational complexity can be analyzed in a
similar fashion, which has been discussed by many researchers, such as Song et al. [19]. In
summary, the memory requirement for the multilevel implementation is O(N ) and the CPU
time per iteration scales as O(N log N ).
7.4 FIPWA for Layered Medium

In this section, FIPWA is extended to accelerate the matrix-vector multiplication for the
problems involving a layered medium. As mentioned before, the Green’s function for the
layered medium can be separated into a direct (primary) part, which is the same as the free-
space Green’s function, and a reflected-wave part. In the previous section, FIPWA has been
developed to handle the free-space problems. Therefore, in this section, effort focuses on the
reflected-wave part.
7.4.1 Groups not aligned in ẑ axis

Before describing the algorithm, the coordinate transform is performed by using
kρ = k sin θ, kz = k cos θ (7.24)
and Eq. (7.15b) becomes

Z
(1)
g R (rj , ri ) = dθW 0 (θ)H0 (k sin θ|ρj − ρi |)eik cos θ(zj +zi ) , (7.25)
SIP
where W 0 (θ) = k cos θW (θ), and ri and rj denote the source and observation points, respec-
tively. As discussed in the previous section, the source points and observation points are
grouped, as shown in Figure 7.6, where the source and observation groups are centered at
rC 0 and rC , respectively. Comparing Eq. (7.25) with its free-space counterpart, the following
observations can be made.
Observation 2. The g R (rj , ri ) can be interpreted as the field originated from the source
at ri , reflected by the dielectric layers and received by observation points located at rj .
Equivalently, this effect can also be viewed as the field radiated from the mirror image of
ri with respect to the air-dielectric interface, denoted as riI . Thus, the mirror images of
the source points can also be combined to form a group, whose center, denoted as rC 0I , is
also the mirror image of the center for source group located at rC 0 . This grouping scheme
is illustrated in Figure 7.6. Without introducing ambiguity, the mirror images of the source
observation group
source group 1C
0
i 1j
0
1
0
1
0
C’
air-dielectric interface
C’ I 0
1
1
0
iI
image of the source group
Figure 7.6: The grouping of the source and observation points. The mirror image of the
source points can also be combined to form a mirror image group, both with respect to the
air-dielectric interface.
point and source group are referred to as the source point and source group in the following
discussion. It is to be cautioned that this is not the same as invoking the image theorem.
Observation 3. The basic expression of g R (rj , ri ) is similar to the free-space one. The
only difference lies in the weight function W 0 (θ). Therefore, all techniques developed for the
FIPWA in free-space can be utilized for this kernel.
In the following, several important issues of the proposed algorithm, such as the proper
choice of the SDP, the inclusion of the pole and branch point contributions, and the diago-
nalization of the translation matrices, are discussed.
Steepest descent path

By using the asymptotic expansion of the Hankel function, that is,
r
(1) 2 i(x− π )
H0 (x) ∼ e 4 , x → ∞, (7.26)
xπ
the integrand in Eq. (7.25) behaves asymptotically as eikρ sin θ+ikz cos θ . Thus, an SDP, along
which the integrand decays exponentially away from its saddle point, can be defined in the θ
plane. As the proper choice of the SDP has been discussed extensively in [20] and [21], the
whole procedure is not repeated here. One possible choice of the SDP consists of three parts.
They are:
Path 1 : θ = θlef t − θR + iθI ,
cos θR cosh θI = 1,
θR > 0, θI > 0;
Path 2 : θlef t ≤ θR ≤ θright ;
Path 3 : θ = θright + θR − iθI ;
cos θR cosh θI = 1;
θR > 0, θI > 0; (7.27)
where θlef t and θright represent the smallest and largest angles that the vectors, directed from
the source points to the observation points located in the respective groups, make with the
z axis, and they are shown in Figure 7.7. The SDP path is illustrated with the SIP in the
background in Figure 7.8.
Two-dimensional fast multipole method

Before proceeding to discuss the diagonalization, the Hankel function kernel should be ex-
panded using the plane wave. Here, the 2D FMM is chosen to perform this task.
From [22], the Hankel function can be expanded using the plane waves as follows:
N
(1) 1 X
H0 (k|ρj − ρi |) = βjC (k, φq2 ) · τ (k, φq2 , ρCC 0I ) · βC 0I iI (k, φq2 ), (7.28)
N q =1
2
where τ is the diagonal translation coefficient of 2D FMM and is calculated using
N
X π
τ (k, φq2 , ρCC 0I ) = Hn(1) (kρCC 0I )e−in[φCC 0I −φq2 − 2 ] , (7.29)
q=−N
and
βjC (k, φq2 ) = eik(xjC cos φq2 +yjC sin φq2 ) , (7.30)
βC 0I iI (k, φq2 ) = eik(xC 0I iI cos φq2 +yC 0I iI sin φq2 ) , (7.31)
where xjC = xj − xC , and so on. The βC 0I iI and βjC are the 2D radiation and receiving
patterns for the source and observation groups, respectively. Because k is a complex number,
they are both inhomogeneous plane waves.
In principle, the 2D FIPWA [20] can also be used to obtain the plane wave expansion of the
Hankel function, as long as the proper SDP for complex wave numbers is defined (remember
the integration is now performed along the SDP and k sin θ is a complex number in general).
Integration along the steepest descent path

The property of the integrand along the SDP has been studied in detail in [20] and [21]. On
Paths 1 and 3, the integrand decays exponentially, whereas fast oscillation is observed on
Path 2. Therefore, proper quadrature rules, such as the Gauss-Laguerre and Gauss-Legendre
method, are chosen. For simplicity, (θq1 , wq1 ) denote the proper quadrature sample and weight
observation group
C
11
00
θright
θleft
11
00
C’ I
image of the source group
Figure 7.7: Illustration of observation group and the image of the source group. The image
of the source group is centered at rC 0I .
Path 1 SIP
SDP
θleft θright π/2 θR

−π/2
Path 2 Pole Path 3
*
*
Branch Point
ΓB
Figure 7.8: The SDP is shown, with the SIP in the background. One contributing pole and
the branch point with the constant phase path, that is, ΓB , passing through it are also shown.
along the SDP in the θ plane and (φq2 , wq2 ) is for the φ integration. Thus, the numerical
integration is evaluated as
X (1)
g SDP (rj , ri ) = wq1 W 0 (θq1 )H0 (k sin θq1 |ρj − ρi |)eik cos θq1 (zj +zi )
q1
1 X
= wq1 W 0 (θq1 )eik cos θq1 (zj +zi )
N q
1
N
X
× βjC (k sin θq1 , φq2 ) · τ (k sin θq1 , φq2 , ρCC 0I ) · βC 0I iI (k sin θq1 , φq2 )
q2 =1
N
XX
= BjC (Ωq ) · T SDP (Ωq , rCC 0I ) · BC 0I iI (Ωq ), (7.32)
q1 q2 =1
where
Ωq = (θq1 , φq2 ), q = (q1 , q2 ), (7.33)
BjC (Ωq ) = eik(Ωq )·rjC , (7.34)
ik(Ωq )·rC 0I iI
BC 0I iI (Ωq ) = e , (7.35)
k(Ωq ) = k (x̂ sin θq1 cos φq2 + ŷ sin θq1 sin φq2 + ẑ cos θq1 ) , (7.36)
wq1 0
T SDP (Ωq , rCC 0I ) = W (θq1 )τ (k sin θq1 , φq2 , ρCC 0I )eik cos θq1 zCC 0I , (7.37)
N
and rji = rj − ri .
In Eq. (7.32), the numerical integration is performed through the translation of the
inhomogeneous plane waves from the source group to the observation group using the diagonal
translation operator T SDP . The diagonal translator in this context means that there are no
mutual interactions among the inhomogeneous plane waves pointing in different directions.
Although the T SDP is diagonal, this translation scheme is not desirable for two reasons.
First, the number of quadrature samples, especially on Path 2, is directly proportional to the
distance between two groups. Therefore, a large number of Ωq is expected for distant groups.
Second, as the integration path, that is, the SDP, is dependent on the relative position of
the source group with respect to the observation groups, different sets of quadrature samples
are required if this relative position changes. As is well known, the efficiency of the FMM-
like algorithm is based on the precomputing and storing of the far-field pattern samples and
the translation operator, that is, BC 0I iI , BjC and T SDP , before the actual matrix-vector
multiplication. To account for all possible source-observation group pairs, Eq. (7.32) requires
one to precompute and store the samples for all possible SDPs. Obviously, it is inefficient
in view of both the computation time and memory storage required. Previously developed
algorithms [5], [13] store the samples along the integration path and thus are only suitable
for the quasi-planar objects.
To solve this problem, the interpolation and extrapolation techniques are used. First, for
the sake of clarity, Eq. (7.32) is rewritten as
X
g SDP (rj , ri ) = SDP
BjC (Ωq ) · TCC 0I (Ωq ) · BC 0I iI (Ωq ). (7.38)
q
Denoting Ωs = (θs1 , φs2 ) and s = (s1 , s2 ) as the samples for the homogeneous plane waves,
which means that θs1 and φs2 are real, and I(Ωq , Ωs ) as the interpolation/extrapolation
kernel, the samples of the far-field patterns along the SDP can be obtained by
X
BjC (Ωq ) · BC 0I iI (Ωq ) = BjC (Ωs ) · BC 0I iI (Ωs )I(Ωq , Ωs ). (7.39)
s
Then, Eq. (7.38) can be manipulated in the same fashion as in Eq. (7.23), that is,
X
g SDP (rj , ri ) = SDP
BjC (Ωq ) · TCC 0I (Ωq ) · BC 0I iI (Ωq )
q
X X
SDP
= TCC 0I (Ωq ) · BjC (Ωs ) · BC 0I iI (Ωs ) · I(Ωq , Ωs )
q s
X
SDP
= BjC (Ωs ) · T̃CC 0I (Ωs ) · BC 0I iI (Ωs ), (7.40)
s
where
X
SDP SDP
T̃CC 0I (Ωs ) = TCC 0I (Ωq ) · I(Ωq , Ωs ), (7.41)
q
SDP
and T̃CC 0I is the desired diagonal translation coefficients. Consequently, the integration along
the SDP is performed by translating the homogeneous plane waves from the source group to
SDP
the observation group, with the help of T̃CC 0I . The advantage of this manipulation is that
the far-field patterns are sampled on the real axis, that is, Ωs in Eq. (7.40), and the sampling
criteria are only dependent on the requirement of accurate reconstruction in Eq. (7.39). In
other words, the way of the sampling, such as the number of the samples, is only determined
by the group properties, such as the size of the groups, and it is independent of the source-
observation group pairs involved in the numerical integration. Thus, only one set of samples
of the far-field patterns is necessary for each group.
Observation 4. The radiation pattern function from the image source group, that is,
BC 0I iI (Ωs ), can be constructed from the source group, that is, BC 0 i (Ωs ), in Eq. (7.20) after
exploiting the symmetric properties between them. Therefore, it is not necessary to store
BC 0I iI (Ωs ) when implementing the algorithm.
Pole contribution
In contrast to the free-space version, the weight function W 0 (θ), or essentially the generalized
reflection coefficients embedded within it, has singularities, manifested as the pole and branch
point in the complex θ plane. As a result, the deformation of the integration path from the
SIP to SDP inevitably encompasses some of them, and the effects of these singularities to the
final integration result must be included. As an illustration, a contributing pole is shown in
Figure 7.8.
Using the same technique proposed in [21], all the poles in the complex θ plane can be
located by using a recursive method. Assuming that θp is one of the contributing poles, the
effect, denoted as g P , can be computed using

(1)
g P (rj , ri ) = <es{θp }H0 (k sin θp |ρj − ρi |)eik cos θp (zj +zi )
1
= <es{θp }eik cos θp (zj +zi )
N
XN
× βjC (k sin θp , φq2 ) · τ (k sin θp , φq2 , ρCC 0I ) · βC 0I iI (k sin θp , φq2 )
q2 =1
N
X
= BjC (θp , φq2 ) · T P (θp , φq2 , rCC 0I ) · BC 0I iI (θp , φq2 ), (7.42)
q2 =1
where <es{θp } is the residue of the weight function W 0 (θ) at θ = θp , and T P is the translation
operator for the pole contribution and is calculated using
<es{θp }
T P (θp , φq2 , rCC 0I ) = τ (k sin θp , φq2 , ρCC 0I )eik cos θp zCC 0I . (7.43)
N
Observation 5. As the locations of the poles are fixed in the θ plane, the pole contribution
can be accounted for by using a 2D FMM, as manifested in Eq. (7.42).
Pole determination
Although it is quite straightforward to include the pole contribution in FIPWA, a major
problem is to find the accurate locations of all poles in a specific region. A handful of root-
searching routines for nonlinear equations are available, but almost all of them can only
locate one root in the given region. As the locations and the number of poles are not known
beforehand, existing root solvers miss the locations of some poles. As a remedy, a new
recursive method is introduced here to accurately determine arbitrary number of poles for a
given region.
By using the Cauchy theorem in complex variables, which states that the contour integral
of an analytic function is zero, this problem is solved successfully. The basic procedure
is as follows. Given a specific region, normally a rectangular box, the contour integral is
performed around it. A zero result implies that no pole exists inside. If the integration result
is nontrivial, the box is further divided into smaller boxes around which the contour integral
is performed. Obviously, only boxes with nontrivial contour integration results are further
subdivided. In this manner, the approximate locations of the poles can be determined. Then,
some root-searching routines can be used to refine them.
This approach has two advantages. First, this method is systematic and accurate. The
method is mathematically sound. By subdividing the box in a multilevel manner, very small
cells can be achieved in only a few levels, implying that the poles can be located with very
high accuracy. Secondly, it is very efficient, because only the reflection coefficient, which has
a closed form, possesses the poles, it is easy to perform the contour integral. As mentioned
before, in each level, only several contour integrations are actually performed. Therefore,
little time is used to locate the poles. A by-product of this method is that the residue is
obtained directly from the result of contour integral. This is useful when the structure under
Table 7.1: Locations of Poles
Number real part (φR ) imag part (φI ) |R̃T M | at this point
1 2.04328364451110 −0.164791069566742 1.1157E+12
2 2.04822034059814 −1.26312918284433 2.0315E+10
3 2.40029093077478 −1.11651541355518 3.1057E+14
4 2.91654903716033 −1.49117966962597 6.2335E+13
0
Pole 1 6
−0.5 3
2
φI
Pole 3
0
−1
Pole 2
−1
Pole 4 −2
−3
−1.5
0 0.5 1 1.5 2 2.5 3
φR
Figure 7.9: Contour plot of R̃T M in the φ plane. The region plotted is φR ∈ (0, π) and
φI ∈ (− π2 , 0).
study is complex and the derivation of an analytic expression for the residue is tedious, if not
impossible.
As an illustration of the effectiveness of the proposed pole-locating algorithm, the following
example is shown. The working frequency is f = 3 GHz. The permittivity of the substrate
is εr = 2.56 + i2.56, the thickness is t = 5 cm, and it is backed with the PEC ground plane.
This structure is not typical in microstrip analysis and is chosen so that multiple poles exist.
Figure 7.9 is the contour plot of the reflection coefficient R̃T M in the φ plane. It is shown
that four poles are located on this plane. The poles found by this routine are listed in Table
7.1. Comparing this table with Figure 7.9, it is obvious that the program finds all four poles.
For all other root-searching routines tested so far, only one pole (No. 1 in the table) can be
located.
Because the poles are fixed in the φ plane, the computation time in the order of O(N ) is
required to evaluate Eq. (7.42) in a multilevel implementation.
Branch point contribution

For layered medium problems, two branch points exist in the kρ plane, which are located at
kρ = k1 and kρ = kN , and k1 and kN are the wave numbers of the outermost layers. As
discussed in [21], the branch point at k1 is eliminated under the coordinate transform. Thus,
the only branch point in the θ plane corresponds to kρ = kN , which is shown in Figure 7.8.
A constant phase path passing through the branch point, denoted as ΓB , can be designed
so that the integrand decays exponentially away from the branch point. The branch point
contribution, denoted as g B , can be evaluated by performing the numerical integration along
the constant phase path as follows:
Z
£ ¤ (1)
g B (rj , ri ) = dθ W 0+ (θ) − W 0− (θ) H0 (k sin θ|ρj − ρi |)eik cos θ(zj +zi )
Γ
XB £ ¤ (1)
= wn W 0+ (θn ) − W 0− (θn ) H0 (k sin θn |ρj − ρi |)eik cos θn (zj +zi )
n
1 X
= wn W(θn )eik cos θn (zj +zi )
N n
N
X
× βjC (k sin θn , φq2 ) · τ (k sin θp , φq2 , ρCC 0I ) · βC 0I iI (k sin θp , φq2 )
q2 =1
N
XX
= BjC (θn , φq2 ) · T B (θn , φq2 , rCC 0I ) · BC 0I iI (θn , φq2 ), (7.44)
n q2 =1
where (θn , wn ) is the numerical quadrature coefficient and the diagonal translator T B is
expressed as
wn
T B (θn , φq2 , rCC 0I ) = W(θn )τ (k sin θn , φq2 , ρCC 0I )eik cos θn zCC 0I , (7.45)
N
with W(θn ) = W 0+ (θn ) − W 0− (θn ). Here, W 0+ (θn ) is the value of the weight function where
kρ lies in the upper Riemann sheet, that is, <e{kN z } > 0, whereas W 0− (θn ) corresponds to
the lower Riemann sheet, that is, <e{kN z } < 0.
7.4.2 Groups aligned in ẑ axis

If two groups are adjacent to each other in the ρ direction, that is, ρ → 0 for the source and
observation points within the groups, the algorithm described in the previous section is no
longer valid because of the divergence of the Hankel function. Therefore, another approach
is proposed, which uses Eq. (7.16b) with Bessel function as the kernel.
Using the identity of the Bessel function,
Z 2π
1
J0 (kρ ρ) = dφekρ (φ)·ρ , (7.46)
2π 0
and performing the same coordinate transform in Eq. (7.24), Eq. (7.16b) can be rewritten as
Z 2π Z
g R (rj , ri ) = dφ dθW 0 (θ)ek(θ,φ)·(rj −ri ) , (7.47)
0 F SIP
0 1
where W (θ) = sin θW (θ). The FSIP is shown in Figure 7.3. As mentioned before, the
π
only difference between this Green’s function and the free-space one, that is, Eq. (7.17), is
the weight function. Therefore, the choice of the proper SDP is the same as the free-space
case, which is shown in Figure 7.4. The diagonalization of the translation matrix is similar
to the free-space one, as detailed in [18]. The pole and branch point contributions have to
be included in the final integration result, which can be treated similarly as in the previous
sections.
7.4.3 Observations
In the previous section, the application of the FIPWA to the layered medium problem is
discussed in detail. After carefully going through the derivation, several observations are in
order:
Observation 6. As mentioned at the start of this chapter, this approach treats the reflected
part of the Green’s function in the same way as the free-space one. The Sommerfeld-type
integral presented in the reflected part is evaluated rigorously using the method of steepest
descent. In addition, the numerical integration is performed by translating the homogeneous
plane waves between groups, with the help of a diagonal translator. Furthermore, the ra-
diation pattern and the receiving pattern for the free-space problem can be reused for the
reflected-wave part. It is necessary to store additional far-field pattern samples for the re-
flected part.
Observation 7. Because the layered medium information is embedded in the generalized
reflection coefficients, that is, R̃T M and R̃T E , this approach, when applied to solve the prob-
lems with the objects above the layered medium, can handle arbitrary number of layers, with
arbitrary parameters such as the dielectric constant and the thickness for each layer. By using
this scheme, only one image term is actually required for arbitrary layered medium structure.
Observation 8. There are several methods previously designed for the layered medium
problems using the SDP idea, such as the steepest descent fast multipole algorithm (SDFMA)
[13] and the TSM-FMA [5]. Both methods make some assumptions, either about the structure
under study or about the layered medium, and both limit their abilities to handle general
problems. This approach can be considered a generalized one in which those restrictions are
lifted.
Observation 9. The analysis of the computational complexity of this approach is easy. In
this algorithm, the reflected wave part is treated similarly to the free-space part. As mentioned
previously, the radiation and receiving patterns for the direct radiation term can be reused
for the reflected-wave part. Therefore, the added computational cost for the layered medium
problems is limited to the translation from the image source group to the observation group.
The pole and branch point contributions do not affect the final computational complexity
because these singularities are fixed in the θ plane. Following the similar procedure discussed
in [18], the CPU time per iteration for the multilevel implementation is O(N log N ), where
N is the number of unknowns.
1m
3m
0.2m
0.3m
Figure 7.10: Illustration of the circular cylinder above a two-layer medium. The permittivity
of the first dielectric layer is εr,2 = 2.56 and the bottom one is εr,3 = 6.5 + i0.6.
Observation 10. It is well known that the translation coefficients for the free-space problems
are only the function of the relative position of the source and observation groups. On the
contrary, in the layered medium problems, the translation coefficients for the reflected-wave
part are also dependent on the location of the source group with respect to the air-dielectric
interface. Therefore, additional storage of the translation coefficients is required. From the
implementation point of view, the major memory consumption of the algorithm is the storage
for the samples of the radiation patterns. Therefore, this additional storage does not affect
the total memory requirement much, and the scaling property of the memory usage with
respect to the number of unknowns is unchanged.
7.5 Numerical Results

In this section, some numerical results are presented to show the validity of this approach. The
multilevel version of this algorithm is implemented on top of the ScaleME (scaleable multipole
engine), a portable implementation of the dynamic MLFMA code [23]. The scattering from
two objects, including a sphere, a circular cylinder, and a realistic tank, is calculated. The
plane wave excitation is always assumed and all scattered fields are computed in the far-field
zone. All examples are solved using a 600 MHz DEC Alpha workstation with 2 GB RAM,
except the ones involving more than one million unknowns.
The first example is a circular cylinder above a two-layer medium, with permittivities
of the two layers as εr,2 = 2.56 and εr,3 = 6.5 + i0.6. The thickness of the first dielectric
layer is t = 0.3 m. The cylinder is 3 m long and has a diameter of 1 m. The distance
between the lower end of the cylinder and the interface is 0.2 m, as shown in Figure 7.10.
The object is discretized by 6,472 small triangle patches at f = 600 MHz and the number
25
20
15
bistatic RCS (dB)

10
0
VV−pol FIPWA
VV−pol full matrix CG
HH−pol FIPWA
HH−pol full matrix CG
−5
−10
−180 −120 −60 0 60 120 180
φsca
Figure 7.11: The bistatic radar cross section for a circular cylinder above a two-layer medium.
The incident plane wave is at (θinc , φinc ) = (60o , 0o ). The scattered field is calculated at
θsca = 60o and φsca ∈ [−180o , 180o ]. The cylinder is discretized with 6,472 small triangle
patches and the number of unknowns is N = 9,708. The results are compared with those
computed using the traditional CG program with full matrix-vector multiplication.
of unknowns is N = 9,708. The incident plane wave direction is θinc = 60o and φinc = 0o
and the bistatic RCS is computed at θscat = 60o and φscat ∈ [−180o , 180o ]. A five-level
ML-FIPWA (multilevel FIPWA) program is used in the example. The results are shown in
Figure 7.11 and are compared to those calculated using a conventional MOM code with the
iterative solver using full matrix-vector multiplication. Very close agreement between the two
codes is observed.
——————————-
The next example is a tank model sitting on the ground, with permittivity of the ground
material as εr = 6.5 + i0.6. The dimensions of the tank model are 8.5 × 3.74 × 2.19 m. A
patch model of the tank is shown in Figure 7.12. The incident plane wave direction is fixed at
θinc = 60o and φinc = 0o . The bistatic RCS is computed at θscat = 60o and φscat ∈ [0o , 360o ].
In Figure 7.13, the bistatic RCS at f = 100 MHz is computed using a five-level ML-FIPWA
code and compared with the one calculated using the conventional MOM code. The tank is
discretized using 2,816 small triangle patches and the number of unknowns is N = 8,334.
The scaling properties of the memory requirement and the CPU time per matrix-vector
multiplication are shown in Figure 7.14. They are compared with the ones from the free-
space codes, including a free-space ML-FIPWA code and the Fast Illinois Solver Code (FISC)
[19, 24]. To make a fair comparison of the computational complexity between this algorithm
and the free-space ones, the parameters for the fast algorithms, such as the number of samples,
the order of interpolation scheme, and the number of levels, are chosen to be the same for
both the free-space and the layered medium codes. To generate this figure, three frequency
z
y
x
Figure 7.12: Illustration of the tank model sitting on the ground. The dimensions of the tank
are 8.5 × 3.74 × 2.19 m.
30
VV−pol FIPWA
25 HH−pol FIPWA
20
15
10
bistatic RCS (dB)
−5
−10
−15
−20
0 60 120 180 240 300 360
φsca
Figure 7.13: The bistatic radar cross section for the tank model at f = 100 MHz. The incident
plane wave is at (θinc , φinc ) = (60o , 0o ). The scattered field is calculated at θsca = 60o and
φsca ∈ [0o , 360o ]. The tank is discretized with 5,556 small triangle patches and the number
of unknowns is N = 8,334.
3
10
2
10
1
10
0
10 FISC − Memory
MLFIPWA Free Space − Memory
MLFIPWA Layer − Memory
FISC − CPU
MLFIPWA Free Space − CPU
MLFIPWA Layer − CPU
−1
10
3 4 5 6 7
10 10 10 10 10
Number of Unknowns
Figure 7.14: The scaling properties of the memory requirement and the CPU time for one
matrix-vector multiplication with respect to the number of unknowns.
50
V−pol
H−pol
40
30
20
bistatic RCS (dB)
10
−10
−20
−30
0 60 120 180 240 300 360
φsca
Figure 7.15: The bistatic RCS of the tank at f = 1,200 MHz. The solid line is the VV
response and the dashed line is the HH response.
Figure 7.16: The current distribution of the tank residing in free-space. The incident plane
wave direction is (θinc , φinc ) = (60o , 0o ) and it is H-polarized. The working frequency is
f = 1, 200 MHz and the number of unknowns involved is N = 1, 210, 458.
points are chosen. They are 100 MHz, 200 MHz, and 400 MHz, corresponding to the number
of unknowns N = 8,334, N = 33,921, and N = 133,578, respectively. From this figure, it is
shown that the memory increases by merely 5% and the CPU time per iteration increases by
about 25% compared with the free-space ML-FIPWA code. Even compared with FISC, the
increase is about 13% and 30%, respectively. This is a good performance as the MLFMA im-
plemented in FISC is tightly coupled to the MOM code and several specialized optimizations
for efficiency are possible. This result shows that the layered medium problems can be solved
as efficiently as the free-space ones. The additional cost to account for the layered medium
is fairly small.
To show the capability of this algorithm, the scattering of this tank at f = 1,200 MHz is
solved with the number of unknowns as N = 1,210,458. An eight-level program is utilized and
requires about 6.6 GB and about 65 hours of CPU time to achieve the residual error at 0.001.
The problem was solved on one R10000/195-MHz processor of the SGI Origin 2000 array at
the National Center for Supercomputing Applications (NCSA). In Figure 7.15, the bistatic
RCS of the VV (vertical-vertical) and HH (horizontal-horizontal) responses is shown. If the
conventional solver is used to solve the same size problem, the estimated memory requirement
shall exceed 13 TB and will take about 30 years of CPU time to reach the convergence.
The comparison of the current distribution on the tank, when residing in free-space and
sitting on the ground, is also shown in Figures 7.16 and 7.17. The H-polarization (horizontal
polarization) results are presented and the effect of the layered medium is clearly shown.
However, there is an open issue on solving the ill-conditioned problems using the iterative
solver. For example, the tank model sits on top of a two-layer medium, which is the same
medium as in Figure 7.10. The working frequency is f = 100 MHz. The scattering results of
Figure 7.17: The current distribution of the tank sitting on the ground. The incident plane
wave direction is (θinc , φinc ) = (60o , 0o ) and it is H-polarized. The working frequency is
f = 1, 200 MHz and the number of unknowns involved is N = 1, 210, 458.
this problem are compared in Figure 7.18. It is obvious that the results obtained using FIPWA
is a little distorted compared to the ones obtained using the full matrix solver, especially in
the back-scattering region. It turns out the resultant impedance matrix is an ill-conditioned
one and the number of iteration used is insufficient; hence a preconditioning procedure is
required.
The same implementation can be readily applied to the buried object problem, which is
a contra-posed problem to the one with objects above layered medium. All of the examples
shown here are the PEC targets buried in the ground, with lossy and lossless ground materials
presented. The plane wave excitation is always assumed and all scattered fields are computed
in the far-field zone. In the following, the scattering from three types of targets is computed.
They include a buried land mine, buried unexploded ordnance, and a buried bunker. Most
of them are of interest to the remote sensing community.
The first example is a buried land mine, whose geometry is shown in Figure 7.19. The
land mine is constructed with two circular cylinders. The diameter for the top one is 17.3
cm and 27 cm for the bottom one. The heights of the two cylinders are 1.62 cm and 7 cm,
respectively. The working frequency is 600 MHz and the object is discretized using 1,764
small triangle patches. The mine is buried within a ground material with εr = (3.3, 0.3) and
the buried depth is 85 cm. The incident plane wave direction is fixed at θinc = 60o and
φinc = −90o . In Figure 7.20, the bistatic RCS, including the HV (horizontal-vertical) and
HH (horizontal-horizontal) response, is computed using the three-level ML-FIPWA code and
the conventional iterative solver with a direct matrix-vector multiplication. The figure shows
close agreement between the ML-FIPWA program and the conventional solver.
The second example deals with an unexploded ordnance model, as shown in Figure 7.21.
30
VV−pol FIPWA
25 HH−pol FIPWA
20
15
bistatic RCS (dB)

10
−5
−10
−15
50 100 150 200 250 300 350
φsca
Figure 7.18: The bistatic RCS of the tank sitting on a two-layer medium at f = 100 MHz.
The layered medium is described in Figure 7.9. The solid line is the VV response and the
dashed line is the HH response. There is some distortion in the backscattering region, because
of the ill-condition of the impedance matrix.
Figure 7.19: The geometry of the land mine. It has been discretized by using 1,764 small
triangle patches, and the number of unknowns is N = 2,646.
−30
−35
−40
−45
−50
bistatic RCS (dB)
−55
−60
−65
−70
HH−pol FIPWA
−75 HV−pol FIPWA
HV−pol full matrix CG
−80
0 60 120 180 240 300 360
φsca
Figure 7.20: The bistatic RCS for the buried land mine problem.
Figure 7.21: The geometry model of the unexploded ordnance. It has been discretized by
using 5,688 small triangle patches, and the number of unknowns is N = 8,532.
−20
−25
−30
−35
bistatic RCS (dB)

−40
−45
−50
VV−pol FIPWA
VH−pol FIPWA
−55 VH−pol full matrix CG
−60
0 60 120 180 240 300 360
φsca
Figure 7.22: The bistatic RCS for the buried unexploded ordnance problem. The VV response
and VH response are shown here.
The geometry is extracted from [14]. It is a cylinder with a hemispherical endcap with a
diameter of 40.6 cm and a length of 153 cm. The object has a symmetry axis and it makes
an angle of 30o relative to the z-axis. The working frequency is 600 MHz and the object is
discretized using 5,688 small triangle patches. The object is buried within a ground material
with εr = 3.5 and the buried depth is 85 cm. The plane wave is incident from θinc = 60o
and φinc = −90o . Again, the bistatic RCS is computed using a five-level ML-FIPWA code
and the conventional iterative solver, which is used as a benchmark. The results are shown
in Figure 7.22 and the close agreement between two solvers is shown.
As a more practical large-scale buried object problem, the scattering from a underground
bunker model is solved. As shown in Figure 7.23, the dimensions of the bunker are 5 m by 5
m with the height of 2 m. The object is buried within a ground material with εr = (3.3, 0.3)
and the buried depth is 2.5 m. The scaling properties of the CPU time per iteration and
memory requirement are shown in Figures 7.24 and 7.25, respectively. They are compared
with the free-space ML-FIPWA code for the same structure. To have a fair comparison,
the background material and the parameters for FIPWA are chosen to be the same for both
programs. From these figures, it is shown that compared to the free-space code, the CPU time
increases by merely 10% and the memory required to solve the same size problem increases
by about 13%. These timing results are similar to what have been observed for objects above
layered medium, that is, the buried object problems can be solved almost as efficiently as
the free-space problem. This excellent scaling property is again mainly due to the plane
wave expansion of the dyadic Green’s function for the buried object problems. To generate
these figures, four frequencies are chosen. They are 75 MHz, 150 MHz, 300 MHz, and 600
MHz, corresponding to the number of unknowns as N =7,674, N =30,090, N =120,180, and
N =480,696.
Figure 7.23: The geometry of the underground bunker. The one shown here is discretized by
using 5, 116 small triangle patches.
CPU Time for One Matrix Vector Multiplication

2
10
FIPWA Free Space
FIPWA Buried
1
10
CPU time (s)
0
10
−1
10
4 5 6
10 10 10
Number of Unknowns
Figure 7.24: The scaling property of the CPU time per iteration for the different sizes of the
underground bunker problem. It is compared with the free-space ML-FIPWA code, and CPU
time required for the buried object problem increases by less than 10%.
Kernel Memory
4
10
FIPWA Free Space
FIPWA Buried
3
10
Memory (MB)
2
10
1
10
4 5 6
10 10 10
Number of Unknowns
Figure 7.25: The scaling property of the required memory for different sizes of the under-
ground bunker problems. It is compared with the free-space ML-FIPWA code, and the
memory required for the buried object problem increases by about than 13%.
−30
−40
−50
bistatic RCS (dB)
−60
−70
−80
−90
−100
0 50 100 150 200 250 300 350
φsca
Figure 7.26: The bistatic RCS (VV-pol) from the underground bunker at f = 900 MHz. The
number of unknowns is N =1,074,588 and it is solved by using one SGI R10000/195-MHz
processor in the Origin 2000 array at NCSA. It uses about 3.94 GB memory and takes about
18 h of CPU time to achieve the residual error of 0.001.
To show the capability of this algorithm, the bunker problem has been solved at f = 900
MHz. The number of unknowns for this frequency is N =1,074,588 and an eight-level ML-
FIPWA program is used. To solve this problem requires 3.94 GB memory and about 18 h
of CPU time on one R10000/195-MHz processor of the SGI Origin 2000 array at NCSA, and
the residual error is set at 0.001. In Figure 7.26, the bistatic RCS of the VV response is
shown. If a conventional solver is used to solve the same size problem, the estimated memory
requirement will exceed 9.3 TB and takes about 11 years of CPU time, which is unrealistic.
This also underscores the importance of the fast method in solving large-scale problems.
7.6 Conclusions
In this chapter, the FIPWA has been successfully applied to solve the layered-medium prob-
lems, by exploiting the similarity of the Green’s function encountered in a layered medium
with its free-space counterpart. This approach has several merits, such as being efficient,
accurate, and simple. Compared to the previously developed methods for solving the prob-
lems in this category, this method is more general and versatile. The multilevel algorithm is
implemented and the CPU time per iteration is shown to be O(N log N ), and the memory
requirement scales as O(N ). It is interesting to note that the computational labor of this
algorithm increases marginally over its free-space counterpart; typically the increase is less
than 20%. As a result, we conclude that the problems in the presence of the layered medium
can be solved as efficiently as the free-space ones. A large-scale problem with more than one
million unknowns has been solved using this multilevel program.
Bibliography
[1] K. A. Michalski, and D. Zheng, “Electromagnetic scattering and radiation by surfaces of

arbitrary shape in layered media, Parts I and II,” IEEE Trans. Antennas Propag., vol.
38, pp. 335–352, Mar 1990.
[2] W. C. Chew, “Field due to a point source in a layered medium - symmetrized formulation,”
Private Notes, March 25, 2003.
[3] W. C. Chew, J. S. Zhao, and T. J. Cui, “The layered medium Green’s function - A new
look,” Microwave Opt. Tech. Lett., vol. 31, no. 4, pp. 252–255, 2001.
[4] W. C. Chew, J. L. Xiong, and M. A. Saville, “A Matrix-Friendly Formulation of Layered

Medium Green’s Function,” Antennas Wirel. Propag. Lett., vol. 5, no. 1, pp. 490–494, Dec
2006.
[5] J. S. Zhao, W. C. Chew, C. C. Lu, E. Michielssen, and J. M. Song, “Thin-stratified medium

fast-multipole algorithm for solving microstrip structures,” IEEE Trans. Microwave The-
ory Tech., vol. 46, no. 4, pp. 395–403, Apr 1998.
[6] T. J. Cui, and W. C. Chew, “Fast evaluation of Sommerfeld integrals for EM scattering
and radiation by three-dimensional buried objects,” IEEE Geosci. Remote Sens., vol. 37,
no. 2, pp.887–900, Mar 1999.
[7] D. G. Fang, J. J. Yang, and G. Y. Delisle, “Discrete image theory for horizontal electric
dipole in a multilayer medium,” IEE Proc.-H, vol. 135, pp. 295–303, Oct 1988.
[8] Y. L. Chow, J. J. Yang, D. G. Fang, and G. E. Howard, “A closed-form spatial Green’s

function for the thick microstrip substrate,” IEEE Trans. Microwave Theory Tech.,vol.
39, pp. 588–592, Mar 1991.
[9] M. I. Aksun, “A robust approach for the derivation of closed-form Green’s functions,”
IEEE Trans. Microwave Theory Tech., vol. 44, pp. 651–658, May 1996.
[10] V. I. Okhmatovski, and A. C. Cangellaris, “Evaluation of layered media Green’s function

via rational function fitting,” IEEE Microwave Compon. Lett., vol. 14, pp. 22–24, Jan
2004.
[11] E. Bleszynski, M. Bleszynski, and T. Jaroszewicz, “AIM: Adaptive integral method

compression algorithm for solving large scale electromagnetic scattering and radiation
problems,” Radio Sci., vol 31, pp. 1225–1251, Sep–Oct 1996.
[12] F. Ling, and J. M. Jin, “Discrete complex image method for Green’s functions of general
multilayer media,” IEEE Microwave Guided Wave Lett., vol. 10, no. 10, pp. 400–402, Oct
2000.
[13] V. Jandhyala, E. Michielssen, B. Shanker, and W. C. Chew, “A combined steepest
descent fast multipole algorithm for the fast analysis of three-dimensional scattering by
rough surfaces,” IEEE Geosci. Remote Sens., vol. 36, pp. 738–748, May 1998.
[14] N. Geng, A. Sullivan, and L. Carin, “Multilevel fast-multipole algorithm for scattering
from conducting targets above or embedded in a lossy half space,” IEEE Geosci. Remote
Sens., vol. 38, pp. 1561–1573, Jul 2000.
[15] D. Vande Ginste, H. Rogier, D. De Zutter, and F. Olyslager, “A fast multipole method
for layered media based on the application of perfectly matched layers - the 2-D case,”
IEEE Trans. Antennas Propag., vol. 52, no. 10, pp. 2631–2640, Oct 2004.
[16] D. Vande Ginste, E. Michielssen, F. Olyslager, D. De Zutter, “An efficient perfectly
matched layer based multilevel fast multipole algorithm for large planar microwave struc-
tures,” IEEE Trans. Antennas Propag., vol. 54, no. 5, pp. 1538–1548, May 2006.
[17] W. C. Chew, Waves and Fields in Inhomogeneous Media. New York: Van Nostrand
[18] B. Hu, W. C. Chew, and S. Velamparambil, “Fast inhomogeneous plane wave algorithm
for the analysis of electromagnetic scattering,” Radio Sci.., Volume 36, Issue 6, pp. 1327–
1340, 2001
[19] J. M. Song, C. C. Lu, and W. C. Chew, “Multilevel fast multipole algorithm for electro-
magnetic scattering by large complex objects,” IEEE Trans. Antennas Propag., vol 45,
pp. 1488–1493, Oct. 1997.
[20] B. Hu, W. C. Chew, E. Michielssen and J. S. Zhao, “Fast inhomogeneous plane wave
algorithm for the fast analysis of two-dimensional scattering problems,” Radio Sci., vol.
34, pp. 759–772, Jul–Aug 1999.
[21] B. Hu and W. C. Chew, “Fast inhomogeneous plane wave algorithm for electromagnetic
solutions in layered medium structures - 2D case,” Radio Sci., vol. 35, pp. 31–43, Jan–Feb
2000.
[22] C.C. Lu and W.C. Chew, “A fast algorithm for solving hybrid integral equation,” IEE
Proc.-H, vol. 140, no. 6, pp.44455-460, Dec. 1993.
[23] S. Velamparambil, J. M. Song, and W. C. Chew, “ScaleME: A portable, distributed
memory multilevel fast multipole kernel for electromagnetic and acoustic integral equa-
tion solvers,” Ctr. Comput. Electromag., Dept. Elect. Comp. Eng., U. Illinois, Urbana-
Champaign, Urbana, IL, Tech. Rep. CCEM-23-99, 1999.
[24] J. M. Song and W. C. Chew, “Large scale computations using FISC,” in IEEE APS Int.
Symp. Dig., 2000, pp. 1856–1859.
193
CHAPTER 8
Electromagnetic Wave versus

Elastic Wave
8.1 Introduction
Students of electromagnetics eventually become expert in wave physics. However, many
disciplines in engineering and science involve wave physics. For instance, elastic wave is used
in nondestructive testing, ultrasound imaging, seismic exploration, well logging, etc. The
subject of elastic wave is within the grasp of students of electromagnetics easily. Even though
this book is mainly about electromagnetic wave, we will introduce the elastic wave concepts
here so that students can easily learn about this subject when they need this knowledge in
the real world. Other areas that can benefit from the understanding of wave physics are the
areas of quantum physics, ocean acoustics, and marine engineering.
This chapter is written for advanced students of electromagnetics who have had some
advanced level courses in the area. Some mathematical sophistication is assumed in the
reader. For some preliminary background in electromagnetics, one can consult [1, Chapter
1]. This chapter is also developed from a set of lecture notes written for advanced students
in electromagnetics [2].
8.2 Derivation of the Elastic Wave Equation

The field of elastodynamics is broad, but in this chapter, we will consider small perturbations
to the equations of elastodynamics. In this manner, the equations can be greatly simplified
by linearization, giving rise to the elastic wave equation. By so doing, many wave phenomena
in elastic solids can be analyzed using the linearized equation. In elasticity, where strain and
stress analysis is involved, tensors of higher rank are involved. Hence, it is more expedient to
use indicial notation (or index notation) to describe the equation sometimes.
The elastic wave equation governs the propagation of waves in solids. We shall illustrate
its derivation as follows: the wave in a solid causes small perturbation of the particles in the
solid. The particles are displaced from their equilibrium positions. The elasticity of the solid
1
will provide the restoring forces for the displaced particles. Hence, the study of the balance
of these forces will lead to the elastic wave equation [3–5].
The displacement of the particles in a solid from their equilibrium position causes a small
displacement field u(x, t) where u is the displacement of the particle at position x at time
t. Here, x is a position vector in three dimensions. (Usually, in electromagnetics and many
physics literature, we use r for position vector, but we use x here so that indicial notation can
be used conveniently. In indicial notation, x1 , x2 , and x3 refer to x, y, and z, respectively.)
The displacement field u(x, t) will stretch and compress distances between particles. For
instance, particles at x and x + δx are δx apart at equilibrium. But under a perturbation by
u(x, t), the change in their separation is given by
δu(x, t) = u(x + δx, t) − u(x, t) (8.1)
By using Taylor’s series expansion, the above becomes
δu(x, t) ' δx · ∇u(x, t) + O(δx2 ) (8.2)
or in indicial notation,
δui ' ∂j ui δxj (8.3)
∂ui
where ∂j ui = ∂xj .
This change in separation δu can be decomposed into a symmetric and
an antisymmetric parts as follows:
symmetric antisymmetric
1z }| { 1z }| {
δui = (∂j ui + ∂i uj ) δxj + (∂j ui − ∂i uj ) δxj (8.4)
2 2
Using indicial notation, it can be shown that
[(∇ × u) × δx]i = ²ijk (∇ × u)j δxk

= ²ijk ²jlm ∂l um δxk (8.5)
where ²ijk is a Levi-Civita alternating tensor. From the identity that [1, Appendix B]
²ijk ²jlm = −²jik ²jlm = δim δkl − δil δkm (8.6)
we deduce that
[(∇ × u) × δx]i = (∂k ui − ∂i uk )δxk (8.7)
Hence,
stretch rotation
z }| { 1 z }| { 1
δui = eij δxj + [(∇ × u) × δx]i = (e · δx)i + [∇ × u) × δx]i (8.8)
2 2
where we have used an overline over a boldface to denote a second rank tensor, and we have
defined the second rank tensor e as
1
eij = (∂j ui + ∂i uj ) (8.9)
2
The first term in (8.8) results in a change in distance between the particles, whereas the
second term, which corresponds to a rotation, has a higher order effect. This can be shown
ELECTROMAGNETIC WAVE VERSUS ELASTIC WAVE 195
easily as follows: The perturbed distance between the particles at x and x+δx is now δx+δu.
The length square of this distance is (using (8.8))
(δx + δu)2 ∼
= δx · δx + 2δx · δu + O(δu2 )
= |δx|2 + 2δx · e · δx + O(δu2 ) (8.10)
assuming that δu ¿ δx, because u is small, and δu is even smaller. Hence, the last term in
(8.10) can be ignored. The second term in (8.8) vanishes in (8.10) because it is orthogonal to
δx.
The above analysis shows that the stretch in the distance between the particles is deter-
mined to first order by the first term in (8.8). The tensor e describes how the particles in a
solid are stretched in the presence of a displacement field: it is called the strain tensor. This
strain produced by the displacement field will produce stresses in the solid.
Stress in a solid is described by a stress tensor T . Given a surface 4S in the body of
solid with a unit normal n̂, the stress in the solid will exert a force on this surface 4S. This
force acting on a surface, known as traction, is given by (see Figure 8.1)
T = n̂ · T 4S (8.11)
n̂
T
∆S S
Figure 8.1: The balance of forces on an elastic body.
Hence, if we know the traction on the entire closed surface S of a volume V , the total
force acting on the body is given by
I Z
n̂ · T dS = ∇ · T dV (8.12)
S V
where the second equality follows from Gauss’ theorem, assuming that T is defined as a
continuous function of space. This force caused by stresses in the solid, must be balanced by
other forces acting on the body, for example, the inertial force and body forces. Hence, by
Newton’s law, Z Z Z
∂2u
ρ 2 dV = ∇ · T dV + f dV (8.13)
V ∂t V V
where ρ is the mass density and f is a force density, for example, due to some externally
applied sources in the body. The left-hand side is the inertial force (mass times acceleration)
while the right-hand side is the total applied force on the body. Because (8.13) holds true for
an arbitrary volume V , we conclude that
∂2u
ρ =∇·T +f (8.14)
∂t2
as our equation of motion.
Because the stress force, the first term on the right hand side of (8.14), is caused by strains
in the solid, T should be a function of e. Under the assumption of small perturbation,
T should be linearly dependent on ē (generalized Hooke’s law). The most general linear
relationship between the second rank tensors is [3]
Tij = Cijkl ekl (8.15)
The above is the constitutive relation for a solid, where Cijkl is a fourth rank tensor.
For an isotropic medium, Cijkl should be independent of any coordinate rotation. The
most general form for a fourth rank tensor that has such independence is
Cijkl = λδij δkl + µ1 δik δjl + µ2 δil δjk (8.16)
Furthermore, because ekl is symmetric, Cijkl = Cijlk . Therefore, the simplified form of Cijkl
is
Cijkl = λδij δkl + µ(δjk δil + δil δik ) (8.17)
Consequently, in an isotropic medium, the constitutive relation is characterized by two con-
stants λ and µ known as Lamé constant. Note that as a consequence of (8.17), Tij = Tji in
(8.15). Hence, both the strain and the stress tensors are symmetric tensors in an isotropic
medium. The above arguments hold true for inhomogeneous media as well. Reciprocity the-
orem similar to that in electromagnetics exists in linear elastodynamics according to Chapter
2.
Using (8.17) in (8.15), after using (8.9), we have that
Tij = λδij ell + µ(eij + eji )
= λδij ∂l ul + µ(∂j ui + ∂i uj ), (8.18)
which is true for inhomogeneous and isotropic media. Then
(∇ · T )j = ∂i Tij = ∂j (λ∂l ul ) + ∂i (µ∂j ui ) + ∂i (µ∂i uj )
= λ∂j (∂l ul ) + (∂l ul )∂j λ + µ∂j (∂i ui ) + (∂j ui )∂i µ + ∂i µ∂i uj
= (λ + µ)[∇∇ · u]j + (∇ · µ∇u)j + (∇ · u)(∇λ)j + [(∇u) · ∇µ]j
(8.19)
If µ and λ are constants of positions, that is, for a homogeneous medium, the above becomes
∇ · T = (λ + µ)∇∇ · u + µ∇2 u (8.20)
∂2u
ρ = (λ + µ)∇∇ · u + µ∇2 u + f (8.21)
∂t2
which is the elastic wave equation for homogeneous and isotropic media. By using the identity
that ∇2 u = ∇∇u − ∇ × ∇u, the above equation can be rewritten as
(λ + 2µ)∇∇ · u − µ∇ × (∇ × u) − ρü = −f , (8.22)
where the double over dots implies second derivative with respect to time.
8.3 Solution of the Elastic Wave Equation—A Succinct

Derivation
By Fourier transform, Z ∞
1
u(x, t) = dωe−iωt u(x, ω) (8.23)
2π −∞
the elastic wave equation (8.22) becomes
(λ + 2µ)∇∇ · u − µ∇ × (∇ × u) + ω 2 ρu = −f (8.24)
where u = u(x, ω), and f = f (x, ω) now.

The above equation supports transverse waves as well as longitudinal waves. The trans-
verse wave means that if the displacement u is a plane wave, it represents motion that is
transverse to the k vector of the plane wave. For a longitudinal wave, the u is in the direction
of the k vector if u is a plane wave. Hence, the displacement u can be decomposed into the
transverse wave part plus the longitudinal wave part. The transverse waves are known as
shear waves (S waves), whereas the longitudinal waves are known as compressional waves. In
acoustics, the longitudinal waves are the same as the pressure waves (P waves). Hence, S and
P waves are also used to denote these two types of waves.
To see this decomposition, one takes ∇× of the above equation, and defines
Ω = ∇ × u, (8.25)
the rotation of u, we have
µ∇ × (∇ × Ω) − ω 2 ρΩ = ∇ × f (8.26)
The Ω represents the field of the S wave. Because ∇ × (∇ × Ω) = ∇∇ · Ω − ∇2 Ω, and that

∇ · Ω = ∇ · ∇ × u = 0, the above is just
1
∇2 Ω + ks2 Ω = − ∇ × f (8.27)
µ
q
where ks2 = ω 2 ρ/µ = ω 2 /c2s and cs = µρ . Hence, this wave propagates with the velocity cs
which is the shear wave velocity.
The solution to the above equation is
Z
1
Ω(x, ω) = dx0 gs (x − x0 )∇0 × f (x0 , ω) (8.28)
µ

0
where gs (x − x0 ) = eiks |x−x | /4π|x − x0 | is the scalar Green’s function for the shear wave.
Taking the divergence of (8.24), and defining
θ = ∇ · u, (8.29)
the dilatational part of u, we have
(λ + 2µ)∇2 θ + ω 2 ρθ = −∇ · f . (8.30)
In the above, θ is related to the P wave. The solution to the above is

Z
1
θ(x, ω) = dx0 gc (x − x0 )∇0 · f (x0 , ω) (8.31)
λ + 2µ
p 0
where kc2 = ω 2 ρ/(λ + 2µ) = ω 2 /c2c , cc = (λ + 2µ)/ρ, and gc (x − x0 ) = eikc |x−x | /4π|x − x0 |
is the scalar Green’s function for the compressional wave. Notice that the P wave or the
compressional wave travels with a velocity of cc .
From (8.24), after using (8.25) and (8.29), we deduce that
f 1 1
u(x, ω) = − + 2 ∇ × Ω − 2 ∇θ (8.32)
µks2 ks kc
In a source-free region where f = 0, the above is the Helmholtz decomposition of the dis-
placement field u into a divergence-free part, the S wave, and a curl-free part, the P wave.
Notice that when µ = 0, from (8.26), Ω is nonzero only in the source region, and is zero
outside the source. Hence, only compressional wave exists when µ = 0, and this can be used
to model a fluid medium which is an acoustic medium. An acoustic medium is a special case
of an elastic medium with µ = 0 where shear waves cannot exist.
To further manipulate the above, using (8.28) and (8.31), we have
Z
f 1
u(x, ω) = − 2 + 2 ∇ × dx0 gs (x − x0 )∇0 × f (x0 , ω)
µks µks
Z
1
− ∇ dx0 gc (x − x0 )∇0 · f (x0 , ω) (8.33)
(λ + 2µ)kc2
Using integration by parts, and the fact that ∇0 g(x − x0 ) = −∇g(x − x0 ), we obtain
Z
f 1
u(x, ω) = − 2 + 2 ∇ × ∇ × dx0 gs (x − x0 )f (x0 , ω)
µks µks
Z
1
− ∇∇ dx0 gc (x − x0 )f (x0 , ω) (8.34)
(λ + 2µ)kc2
Using the usual identity to simplify ∇ × ∇ × A, and the fact that
∇2 gs (x − x0 ) + ks2 gs (x − x0 ) = −δ(x − x0 ),
we arrive at
µ ¶ Z
∇∇ 1
u(x, ω) = I+ 2 · dx0 gs (x − x0 )f (x0 , ω)
ks µ
Z
∇∇ 1
− 2 · dx0 gc (x − x0 )f (x0 , ω) (8.35)
kc λ + 2µ
We can write the above more concisely as

Z
u(x) = dx0 G(x, x0 ) · f (x0 ) (8.36)
with the ω dependence suppressed, and the dyadic Green’s function G(x, x0 ) is defined as
µ ¶
∇∇ 1 ∇∇ 1
G(x, x0 ) = I + 2 gs (x − x0 ) − 2 gc (x − x0 ). (8.37)
ks µ kc λ + 2µ
The first term of the dyadic Green’s function produces the transverse wave, and is similar
to the electromagnetic dyadic Green’s function [1, Chapter 1]. The second term produces
the longitudinal wave, and does not exist in a simple electromagnetic medium. It is to be
noted that the scalar Green’s functions have 1/|x − x0 | singularity, and the double ∇ operator
will accentuate this singularity in both terms above. This singularity is also known as the
hypersingularity in the computational electromagnetics parlance. However, it can be shown
that these hypersingularities from both terms above cancel each other when x → x0 . Since
the electromagnetic dyadic Green’s function only has the first term (transverse wave), the
hypersingularity persists.
The hypersingularity, often studied in electromagnetics as the singularity of the dyadic
Green’s function [1, Chapter 7], is both a curse and a blessing! It is a curse because it makes
the numerical evaluation of many integrals involving the dyadic Green’s function difficult. It
is a blessing because it implies that there is rich physics in this hypersingularity. It is within
this hypersingularity that dwells the world of circuit physics that we use to create a plethora
of new technologies. Because within the proximity of this singularity, where dimensions of
structures are much smaller than the wavelength, lie the world of magnetoquasistatics (the
world of inductors), and the world of the electro-quasistatics (the world of the capacitors)
(see Chapter 5).
8.3.1 Time-Domain Solution

If f (x, t) = x̂j δ(x)δ(t), then f (x, ω) = x̂j δ(x) and
µ ¶ iks r
∇∇ e ∇∇ eikc r
u(x, ω) = I + 2 · x̂j − 2 · x̂j (8.38)
ks 4πµr kc 4π(λ + 2µ)r
where r = |x|, or in indicial notation

µ ¶
eiks r 1 eiks r eikc r
ui (x, ω) = δij + ∂ i ∂j − (8.39)
4πµr 4πr µks2 (λ + 2µ)kc2

p p
Since ks = ω/cs = ω ρ/µ , kc = ω/cc = ω ρ/(λ + 2µ), the above can be inverse Fourier
transformed to yield
δ(t − r/cs )
ui (x, t) = δij
4πµr
·µ ¶ µ ¶ µ ¶ µ ¶¸
1 r r r r
− ∂ i ∂j t− u t− − t− u t− (8.40)
4πrρ cs cs cc cc
2
∂
Writing ∂i ∂j = (∂i r)(∂j r) ∂r 2 , the above becomes
δ(t − r/cs )
ui (x, t) = [δij − (∂i r)(∂j r)]
4πµr
· µ ¶ µ ¶¸
1 r r
− (∂i r)(∂j r) tu t − − tu t −
2πρr3 cs cc
δ(t − r/cc )
+ (∂i r)(∂j r) . (8.41)
4π(λ + 2µ)r
Consequently, if f (x, t) is a general body force, by convolutional theorem,

r
fj (x, t − cs )
ui (x, t) = [δij − (∂i r)(∂j r)]
4πµr
Z r/cs
1
+(∂i r)(∂j r) τ fj (x, t − τ )dτ
2πρr3 r/cc
fj (x, t − crc )
+(∂i r)(∂j r) (8.42)
4π(λ + 2µ)r
8.4 Alternative Solution of the Elastic Wave Equation

via Fourier-Laplace Transform
Alternatively, the elastic wave equation
(λ + µ)∇∇ · u + µ∇2 u − ρü = −f (8.43)
can be solved by Fourier-Laplace transform. We let

Z ∞ Z ∞
1 −iωt
u(x, t) = dωe dkeik·x u(k, ω), (8.44)
(2π)4 −∞ −∞
then (8.43) becomes

−(λ + µ)kk · u − µk 2 u + ω 2 ρu = −f (8.45)
2
where u = u(k, ω), f = f (k, ω) now, and k = k · k. The above can be formally solved to
yield
£ ¤−1
u(k, ω) = (λ + µ)kk + µk2 I − ω 2 ρI · f. (8.46)
The inverse of the above tensor commutes with itself, so that the inverse must be of the form
αI + βkk, that is, £ ¤ £ ¤
(λ + µ)kk + µk 2 I − ω 2 ρI · αI + βkk = I (8.47)
Solving the above yields that
1
α= (8.48)
(µk 2 − ω 2 ρ)
−(λ + µ)
β =
[µk 2 − ρω 2 ] [(λ + 2µ)k 2 − ρω 2 ]
µ (λ + 2µ)
= − 2 +
ρω [µk 2 − ρω 2 ] ρω 2 [(λ + 2µ)k 2 − ρω 2 ]
1 1
= − 2 + 2 (8.49)
2 2
ks µ [k − ks ] kc (λ + 2µ) [k 2 − kc2 ]
where ks2 = ω 2 ρ/µ, kc2 = ω 2 ρ/(λ + 2µ). Consequently,
· ¸
kk f (k, ω) kk · f (k, ω)
u(k, ω) = I − 2 · + (8.50)
ks µ(k 2 − ks2 ) kc2 (λ + 2µ)(k 2 − kc2 )
To obtain u(x, ω), we use the relation that
Z ∞
1
u(x, ω) = dkeik·x u(k, ω). (8.51)
(2π)3 −∞
First, we can perform the integral assuming that we have a point source for f . It can be
shown by contour integration technique [1, Chapters 2 and 7] that
Z
1 1 eik0 r
3
dkeik·x 2 2 = (8.52)
(2π) (k − k0 ) 4πr
where r = |x|. Hence, it can be shown that
Z µ ¶ µ ¶
1 ik·x kk 1 ∇∇ eiks r
dke I− 2 = I+ 2 (8.53)
(2π)3 ks µ(k 2 − ks2 ) ks 4πµr
p
where ks = ω/cs , and cs = µ/ρ is the shear wave velocity. Similarly,
Z
1 ik·x kk ∇∇ eikc r
dke = − (8.54)
(2π)3 kc2 (λ + 2µ)(k2 − kc2 ) kc2 4π(λ + 2µ)r
p
where kc = ω/cc , cc = (λ + 2µ)/ρ is the compressional wave velocity.
By the convolutional theorem,
µ ¶ Z 0
∇∇ eiks |x−x |
u(x, ω) = I + 2 · dx0 f (x0 , ω)
ks 4πµ|x − x0 |
Z 0
∇∇ 0 eikc |x−x |
− 2 · dx f (x0 , ω) (8.55)
kc 4π(λ + 2µ)|x − x0 |
The above is exactly the same as (8.35), the result obtained in the previous Section 8.3.
8.5 Boundary Conditions for Elastic Wave Equation
δ
n̂
u( ) , τ ( )
1 1
µ1 , λ1
u( ) , τ (
2 2)
µ2 , λ2
Figure 8.2: Boundary condition at the interface of two elastic media.
The equation of motion for elastic waves is
∇ · T + f = −ω 2 ρu (8.56)
By integrating this over a pillbox whose thickness is infinitesimally small at the interface
between two regions (see Figure 8.2), and assuming that f is not singular at the interface, it
can be shown that
(1) (2)
n̂ · T = n̂ · T (8.57)
The above corresponds to three equations for the boundary conditions at an interface.
If (8.57) is written in terms of a local Cartesian coordinates with ẑ being the unit normal
n̂, then (8.57) is equivalent to the continuity of Tzz , Tzx , and Tzy across an interface. More
explicitly, one can express
Tzz = λ∇ · u + 2µ∂z uz = λ(∂x ux + ∂y uy ) + (λ + 2µ)∂z uz (8.58)
In the above λ and µ are functions of z in the local coordinates, since the medium parameters
are assumed to vary across the interface. The z derivative is assumed to be taken across
the interface as well. Furthermore, Tzz cannot be singular (does not contain Dirac delta
functions), then sum of the terms on the right-hand side involving ∂x ux , ∂y uy , and ∂z uz must
be regular. Since ∂x and ∂y are tangential derivatives, and ux and uy are smooth functions
of x and y, ∂x ux and ∂y uy are regular. Therefore ∂z uz must be regular. Requiring ∂z uz to
be regular implies that at the interface
u(1) (2)
z = uz (8.59)
Furthermore,
Tzx = µ∂x uz + µ∂z ux , (8.60)
by the same token, the continuity of Tzx implies that ∂z ux must be regular. This induces the
boundary condition that
u(1)
x = ux
(2)
(8.61)
By the same argument from Tzy , we have
u(1) (2)
y = uy (8.62)
Hence, in addition to (8.57), we have boundary conditions (8.59), (8.61), and (8.62) which
form a total of six boundary conditions at a solid-solid interface.
At a fluid-solid interface, µ1 = 0 in the fluid region assuming that it is region 1, then from
(8.60), Tzx and Tzy are zero at the interface, in order for them to be continuous across an
interface. Furthermore, ux and uy need not be continuous anymore. The boundary conditions
are (8.57) and (8.59), a total of four boundary conditions.
At a fluid-fluid interface, only Tzz is nonzero. Its continuity implies λ∇·u = p is continuous
or the pressure is continuous. Furthermore, it induces the boundary condition (8.59). Hence
there are only two boundary conditions.
8.6 Decomposition of Elastic Wave into SH, SV and P

Waves for Layered Media
In a planar layered medium where the interfaces are horizontal, the transverse waves, or the
shear waves, can be classified as shear horizontal (SH) where the particle motion is parallel
to the interface, and shear vertical (SV) where the particle motion is in a plane orthogonal to
the horizontal interface. The SH and SV waves are analogous to the TEz and TMz waves in
electromagnetics [1, Chapter 2]. The P wave is the longitudinal wave, and in the case when
the shear wave disappears, as in a purely acoustic medium, it is the pressure wave.
From the boundary conditions, it can be seen that the SH waves propagate through a
planar layered medium independent of the SV and P waves. Moreover, the SV and P waves
are coupled together at the planar interfaces by the boundary condition. Hence, given an
arbitrary source, we can use the Weyl or Sommerfeld identity to expand the waves into plane
waves. If these plane waves can be further decomposed into SH, SV and P waves, then the
transmission and reflection of these waves through a planar layered medium can be easily
found using the method of Chapter 2 in reference [1].
It has been shown previously in Section 8.3 that an arbitrary source produces a displace-
ment field in a homogeneous isotropic medium given by
f 1 1
u(x) = − + 2 ∇ × Ω − 2 ∇θ (8.63)
µks2 ks kc
| {z } | {z }
us up
Outside the source region the first term is zero. The second term corresponds to S waves
while the third term corresponds to P waves. Hence, in a source-free region, the S component
of (8.63) is
1
us (x) = 2 ∇ × Ω (8.64)
ks
From the definition of Ω, we have
Ω = ∇ × us (x). (8.65)
In the above, Ω was previously derived in the Section 8.3 to be
Z iks |x−x0 |
1 0 e
Ω(x) = ∇ × dr f (x0 ) (8.66)
µ 4π|x − x0 |
Using the Sommerfeld-Weyl identities, (8.66) can be expressed as a linear superposition of

plane waves. Assuming that Ω and us are plane waves in (8.64) and (8.65), replacing ∇
by iks , when we assume the vertical axis to be the z axis, we note that only the SV waves
have usz 6= 0, and only the SH waves have Ωz 6= 0. Hence, we can use usz , namely, the z
component of us , to characterize SV waves; and we can use Ωz , namely, the z component of
Ω, to characterize SH wave.
Assuming that f (x) = âAδ(x), that is, a point excitation polarized in the â direction, then
(8.66) becomes
A eiks r
Ω(x) = (∇ × â) (8.67)
µ 4πr
where r = |x|.
In the electromagnetic case, the z components of the H and E fields are used to charac-
terize TEz and TMz waves, respectively. We can adopt the same method here. Consequently,
using the fact that ∇2 gs = −ks2 gs outside the point source (where gs is the scalar Green’s
function for the shear wave), extracting the z component of the S-wave field, usz = ẑus follows
from (8.64) to be
A eiks r
usz (x) = 2
(ẑ ẑks2 + ∇z ∇) · â (8.68)
µks 4πr
The z component of (8.67) characterizes an SH wave while (8.68) characterizes an SV wave.
For a student of electromagnetics, one can think that Ω is analogous to the H field, while u
is analogous to the E field.
The P wave can be characterized by θ which has been previously derived to be (see (8.31))
Z 0
∇· eikc |x−x |
θ = dr0 f (x0 )
λ + 2µ 4π|x − x0 |
A∇ · â eikc r
= (8.69)
λ + 2µ 4πr
for this particular point source. Alternatively, P waves can be characterized by upz = −(1/kc2 )∂z θ
Using the Sommerfeld identity [1, Chapter 2],
Z ∞
eikr kρ ikz |z|
=i dkρ e J0 (kρ ρ), (8.70)
r 0 kz
p p
where kρ2 + kz2 = k 2 , ρ = x2 + y 2 , r = ρ2 + z 2 , the above can be expressed as
Z ∞
iAẑ · (∇ × â) kρ iksz |z|
Ωz (r) = dkρ e J0 (kρ ρ), SH, (8.71)
4πµ 0 ksz
Z
iA(ẑ · âks2 + ∂z∇ · â) ∞ kρ iksz |z|
usz (r) = dkρ e J0 (kρ ρ), SV, (8.72)
4πµks2 0 k sz
Z ∞
p ±A∇ · â
uz (r) = dkρ kρ eikcz |z| J0 (kρ ρ), P (8.73)
4π(λ + 2µ)kc2 0
where kρ2 + ksz

2
= ks2 , kρ2 + kcz
2
= kc2 . The above have expressed the field from a point source
in terms of a linear superposition of plane waves. We have used upz to characterize P waves
so that it has the same dimension as usz .
When the point source is placed above a planar layered medium, the SH wave characterized
by Ωz will propagate through the layered medium independently of the other waves. Hence,
the SH wave for a point source on top of a layered medium can be expressed as
Z h i
iAẑ · (∇ × â) ∞ kρ
Ωz (r) = dkρ J0 (kρ ρ) eiksz |z| + RHH eiksz (z+2d1 ) (8.74)
4πµ 0 ksz
Since the SV waves and the P waves are always coupled together in a planar layered
medium, we need to write them as a couplet:
· s ¸ Z ∞ · s ¸
uz ikz |z| ez±
u z>0
φ= = dk k
ρ ρ e p (8.75)
upz 0 uez± z<0
where
iA(ẑ · âks2 ± iksz ∇s± · â) ±A∇p± · â
esz± =
u J0 (kρ ρ), epz± =
u J0 (kρ ρ) (8.76)
4πµks2 ksz 4π(λ + 2µ)kc2
· ¸
ksz 0
kz = , ∇s± = ∇s ± ẑiksz , ∇p± = ∇s ± ẑikcz . (8.77)
0 kcz
In the above ∇s = x̂∂x + ŷ∂y . When the point excitation is placed on top of a layered medium,
we have Z ∞ h i
φ= dkρ kρ eikz |z| · u
e ± + eikz (z+d1 ) · R · eikz d1 · u
e − J0 (kρ ρ) (8.78)
0
£ s ¤
where e t±
u = u epz± and R is the appropriate reflection matrix describing the reflection
ez± , u
and cross-coupling between the SV and P waves.
The above derivation could be repeated with the Weyl identity if we so wish. The reflection
coefficient RHH and the reflection matrix R can be found by matching boundary conditions.
Then the field due to an elastic point source in or on top of a layered medium can be derived
similar to the formalism described in Chapter 2 in [1].
8.7 Elastic Wave Equation for Planar Layered Media

The elastic wave equation for isotropic inhomogeneous media is
∂j (λ∂l ul ) + ∂i (µ∂i uj ) + ∂i (µ∂j ui ) + ω 2 ρuj = 0 (8.79)
In vector notation, this may be written as

←
−
∇(λ∇ · u) + ∇ · (µ∇u) + (µ∇u) · ∇ + ω 2 ρu = 0 (8.80)
←
−
where ∇ operates on terms to its left.
∂
If λ and µ are functions of z only, and ∂x = 0 and u = x̂ux , that is, we have an SH wave
only, then extracting the x̂ component of (8.80) leads to
∇s · µ∇s ux + ω 2 ρux = 0 (8.81)
∂ ∂
where for this example, ∇s = ŷ ∂y + ẑ ∂z . Hence for this problem, a displacement field
SH SH
µ1 , λ1
µ2 , λ2 x y
SH
Figure 8.3: Reflected and transmitted SH plane wave at the interface of two-layer media
polarized in x with no variation in x is a pure shear wave with no coupling to the P wave.
Even when µ and λ are discontinuous in z, only SH waves will be reflected and transmitted,
as shown in Figure 8.3.
However, if the incident plane wave is an SV (shear vertical) wave, the displacement of the
particles at an interface will induce both P (compressional) and SV reflected and transmitted
waves. To see this, we let us = ŷuy + ẑuz . Then (8.80) becomes
←
−
∇s (λ∇s · us ) + ∇s · (µ∇s us ) + (µ∇s us ) · ∇ s + ω 2 ρus = 0. (8.82)
The above is the equation that governs the shear and compressional waves in a one-
∂
dimensional inhomogeneity where ∂x = 0.
∂
Since λ, µ and us are smooth functions of y, ∂y is smooth. Since µ and λ functions which
are discontinuous with a step jump in the z direction, the z derivative may be discontinuous
unless boundary conditions are satisfied by the field. To this end, we extract the z component
of the terms in (8.82) that contain z derivative. They are
µ ¶
∂ ∂
λ∇s · us + 2µ uz (8.83)
∂z ∂z
∂
In order for ∂z to be nonsingular in (8.82), we require
∂
λ∇s · us + 2µ uz (8.84)
∂z
to be continuous. This is the same as requiring Tzz to be continuous.

Similarly, taking the y component of (8.82) that contains z derivative, we require
∂ ∂
µ uy + µ uz (8.85)
∂z ∂y
to be continuous. This is the same as requiring Tzy to be continuous. In order for (8.84) and
(8.85) to be regular, uz and uy have to be continuous functions of z. Hence, the boundary
conditions at a solid-solid interface are
u(1) (2)
z = uz (8.86)
u(1) (2)
y = uy (8.87)
∂ (1) ∂ (2)
λ1 ∇s · u(1)
s + 2µ1 uz = λ2 ∇s · u(2)
s + 2µ2 u (8.88)
∂z ∂z z
µ ¶ µ ¶
∂ (1) ∂ (1) ∂ (2) ∂ (2)
µ1 uy + uz = µ2 uy + uz (8.89)
∂z ∂y ∂z ∂y
The reflection of SH waves by a plane interface is purely a scalar problem. However, the
reflection of a P wave or an SV wave by a plane boundary is a vector problem. In this case,
us can always be decomposed into two components us = v̂uv + p̂up where p̂ is a unit vector
in the direction of wave propagation, and v̂ is a unit vector in the yz-plane orthogonal to p̂.
For z > 0, the incident wave can be written as
· inc ¸ · ¸
inc uv v0 e−ik1vz z
us = = eiky y
uinc
p p0 e−ik1pz z
· −ik z ¸· ¸
e 1vz 0 v0
= eiky y
0 e−ik1pz z p0
= e−ik1z z · u0 eiky y (8.90)
where · ¸ · ¸
v0 k1vz 0
u0 = , k1z = (8.91)
p0 0 k1pz
In the presence of a boundary, the reflected wave can be written as
· ref ¸
ref uv
us = = eik1z z · ur eiky y (8.92)
uref
p
The most general relation between ur and u0 is that
ur = R · u0 (8.93)
where · ¸
Rvv Rvp
R= (8.94)
Rpv Rpp
By the same token, the transmitted wave is
utra
s = e−ik2z z · ut eiky y (8.95)
where
ut = T · u 0 (8.96)
and · ¸
Tvv Tvp
T= (8.97)
Tpv Tpp
There are four unknowns in ur and ut which can be found from four equations as a conse-
quence of the boundary conditions (8.86) to (8.89).
8.7.1 Three-Layer Medium Case
region 1
y
0
region 2
z = −h
region 3
Figure 8.4: Reflected and transmitted waves in three-layer media
The three-layer problem can be solved using the constraint boundary conditions introduced
in Chapter 2 of [1]. Then a recursive formula can be derived from which the reflection matrix
for an N -layer problem can be easily found.
When three layers are present as shown in Figure 8.4, the field in Region 1 can be written
as
u1 = e−ik1z z · a1 + eik1z z · b1 ,
h i
e
e−ik1z z + eik1z z · R
= 12 · a1 (8.98)
e · a . The eiky y dependence is dropped assuming that it is

where we have defined b1 = R12 1
implicit.
In Region 2, we have
u2 = e−ik2z z · a2 + eik2z z · b2
h i
= e−ik2z z + eik2z (z+h) · R23 · eik2z h · a2 (8.99)
where we have defined

e−ik2z h · b2 = R23 · eik2z h · a2 (8.100)
and R23 is just the one-interface reflection coefficient previously defined.
The amplitude a2 is determined by the transmission of the amplitude of the downgoing
wave in Region 1 (which is a1 ) plus the reflection of the upgoing wave in Region 2.
As a result, we have at z = 0,
a2 = T12 · a1 + R21 · eik2z h · R23 · eik2z h · a2 (8.101)
Solving the above yields

h i−1
a2 = I − R21 · eik2z h · R23 · eik2z h · T12 · a1 (8.102)
The amplitude b1 of the upgoing wave in Region 1 is the consequence of the reflection
of the downgoing wave in Region 1 plus the transmission of the upgoing wave in Region 2.
Hence, at z = 0,
e · a = R · a + T · eik2z h · R · eik2z h · a
b1 = R (8.103)
12 1 12 1 21 23 2
e , yielding
Using (8.101), the above can be solved for R12
e =R
R + T12 · eik2z h · R23 · eik2z h
12 12
h i−1
· I − R21 · eik2z h · R23 · eik2z h · T12 , (8.104)
where R e is the generalized reflection operator for a layered medium. If a region is added
12
beyond Region 3, we need only to change R23 to R e in the above to account for subsurface
23
reflection.
The above is a recursive relation which in general, can be written as
e
R = Ri,i+1 + Ti+1,i · eiki+1,z hi+1 · Re iki+1,z h
i,i+1 i+1,i+2 · e ·
h i−1
e
I − Ri+1,i · eiki+1,z hi+1 · R iki+1,z hi+1
· Ti,i+1
i+1,i+2 · e
(8.105)
where hi+1 is the thickness of the (i + 1)-th layer. Eq. (8.102) is then
h i−1
e
ai+1 = I − Ri+1,i · eiki+1,z h · R iki+1,z h
· Ti,i+1 · ai
i+1,i+2 · e (8.106)
8.8 Finite Difference Scheme for the Elastic Wave Equa-

tion
The equation of motion for elastic waves is given by
∂ 2 ux ∂Txx ∂Txz
ρ = + (8.107)
∂t2 ∂x ∂z
∂ 2 uz ∂Txz ∂Tzz
ρ = + (8.108)
∂t2 ∂x ∂z
∂ux ∂uz
Txx = (λ + 2µ) +λ (8.109)
∂x ∂z
∂uz ∂ux
Tzz = (λ + 2µ) +λ (8.110)
∂z ∂x
µ ¶
∂ux ∂uz
Txz = Tzx =µ + (8.111)
∂z ∂x
Defining υi = ∂ui /∂t, the above can be transformed into a first-order system, that is,
µ ¶
∂υx −1 ∂Txx ∂Txz
=ρ + (8.112)
∂t ∂x ∂z
µ ¶
∂uz ∂Txz ∂Tzz
= ρ−1 + (8.113)
∂t ∂x ∂z
∂Txx ∂υx ∂υz
= (λ + 2µ) +λ (8.114)
∂t ∂x ∂z
∂Tzz ∂υz ∂υx
= (λ + 2µ) +λ (8.115)
∂t ∂z ∂x
µ ¶
∂Txz ∂υx ∂υz
=µ + (8.116)
∂t ∂z ∂x
Using a central difference scheme, the above can be written as
k+ 1 k− 1 ∆t h k i
υx,i,j2 − υx,i,j2 = ρ−1
ij
k
Txx,i+ 1 ,j − Txx,i− 1
∆x 2 2 ,j
∆t h i
+ρ−1
ij T k
1 − T k
1 (8.117)
∆z xz,i,j+ 2 xz,i,j− 2
k+ 1 k− 1 ∆t h k i
υz,i+2 1 ,j+ 1 − υz,i+2 1 ,j+ 1 = ρ−1
i+ 1 ,j+ 1
k
Txz,i+1,j+ 1 − Txz,i,j+ 1
2 2 2 2 2 2 ∆x 2 2
∆t h i
+ρ−1 1 1
i+ 2 ,j+ 2 ∆z
T k
1
zz,i+ 2 ,j+1 − T k
1
zz,i+ 2 ,j (8.118)
k+1 k ∆t h k+ 21 k+ 1
i
Txx,i+ 1
,j
− Txx,i+ 1
,j = (λ + 2µ)i+ 21 ,j υx,i+1,j − υx,i,j2
2 2 ∆x
∆t h k+ 21 k+ 1
i
+λi+ 12 ,j υz,i+ 1 ,j+ 1 − υz,i+2 1 ,j− 1 (8.119)
∆z 2 2 2 2
k+1 k ∆t h k+ 21 k+ 1
i
Tzz,i+ 1
,j
− Tzz,i+ 1
,j = (λ + 2µ)i+ 2 ,j
1 · υz,i,j+1 − υz,i,j2
2 2 ∆z
∆t h k+ 21 k+ 1
i
+λi+ 21 ,j υx,i+ 1 ,j − υx,i,j2 (8.120)
∆x 2
k+1 k ∆t h k+ 12 k+ 21
i
Txz,i,j+ 1 − Txz,i,j+ 1 = µi,j+ 1
2 ∆z
υ x,i,j+1 − υ x,i,j
2 2
∆t h k+ 21 k+ 1
i
+µi,j+ 12 υz,i+ 1 ,j+ 1 − υz,i−2 1 ,j+ 1 (8.121)
∆x 2 2 2 2
For a homogenous medium, the stability criterion is

s
1 1
υρ ∆t + <1 (8.122)
(∆x)2 (∆z)2
The above finite difference formulas are easy to implement and has been used to solve for
various practical problems [6–9].
8.9 Integral Equation for Elastic Wave Scattering

The integral equations for elastic wave scattering can be derived from its partial differential
equation (PDE) counterpart by applying the equivalence principle and extinction theorem as
described in Chapter 8 in reference [1]
The PDE of elastic wave has been previously derived in Section 8.2, and reproduced here
under a slightly different notation as
γ∇∇ · u − µ∇ × ∇ × u + ω 2 ρu = −f (8.123)
where γ = λ + 2µ.
The dyadic Green’s function for an elastic medium has been derived in Section 8.3 in Eq.
(8.37), which is repeated here as
µ ¶
0 1 ∇∇ 1 ∇∇
G(x, x ) = I+ 2 gs (x − x0 ) − gc (x − x0 ) (8.124)
µ ks γ kc2
It satisfies
γ∇∇ · G(x, x0 ) − µ∇ × ∇ × G(x, x0 ) + ω 2 ρG(x, x0 ) = Iδ(x − x0 ) (8.125)

V
ρ, λ, µ
V0 nˆ′ uI
ρ 0 , λ0 , µ0 x1
O
x3
S x2
Figure 8.5: Elastic wave scattering by a homogeneous scatterer V0 embedded in an infinite

isotropic medium V . uI is an incident wave.
Right multiplying (8.123) by G(x, x0 ) and left multiplying (8.125) by u(x), and subtracting
the result, we have
£ ¤
γ ∇(∇ · u(x)) · G(x, x0 ) − u(x) · ∇∇ · G(x, x0 )
£ ¤
−µ ∇ × ∇ × u(x) · G(x, x0 ) − u(x) · ∇ × ∇ × G(x, x0 )
= −f (x) · G(x, x0 ) − δ(x − x0 )u(x) (8.126)
Using the vector identities
∇ · [(∇ · u)B − u∇ · B] = ∇(∇ · u) · B − u · (∇∇ · B) (8.127)

∇ · [(∇ × u) × B + u × ∇ × B] = (∇ × ∇ × u) · B − u · ∇ × ∇ × B
(8.128)
we can integrate (8.126) over a volume V bounded by S, as shown in Figure 8.5, to result in
I
£ ¤
γ dS n̂ · (∇ · u(x))G(x, x0 ) − u(x)∇ · G(x, x0 )
S
I
£ ¤
−µ dS n̂ · (∇ × u(x)) × G(x, x0 ) + u(x) × ∇ × G(x, x0 )
Z S ½
0 −u(x0 ), x0 ∈ V
+ f (x) · G(x, x )dx = (8.129)
V 0, x0 ∈
/V
The above can be rewritten as

I
£ ¤
γ dS 0 ∇0 · u(x0 )n̂0 · G(x0 , x) − n̂0 · u(x0 )∇0 · G(x0 , x)
I
£ ¤
−µ dS 0 n̂0 × (∇0 × u(x0 )) · G(x0 , x) + (n̂0 × u(x0 )) · ∇0 × G(x0 , x)
½
u(x), x ∈ V
+uI (x) = (8.130)
0, x∈/V
The extinction theorem described in Chapter 8 of reference [1] is also embedded in the above
equation. We can use the above to formulate integral equations for elastic wave scattering.
In the above, one can show that
· ¸
0 1 0 ∇ 2 1∇ 2
∇ · G(x, x ) = ∇gs (x − x ) + 2 ∇ gs − ∇ gc
µ ks γ kc2
· ¸
1 0 1 0 2
= ∇gs (x − x ) + 2 ∇(−δ(x − x ) − ks gs )
µ ks
1 1
− ∇(−δ(x − x0 ) − kc2 gc )
γ kc2
1∇ 1 1 1
= − 2
δ(x − x0 ) + ∇gc + ∇δ(x − x0 )
µ ks γ γ kc2
1
= ∇gc (x − x0 ) (8.131)
γ
Using G T (x0 , x) = G(x, x0 ), [∇0 × G(x0 , x)]T = ∇ × G(x, x0 ), where the superscript T
denotes a transpose, we have
I ½ ¾
1
γ dS 0 [G(x, x0 ) · n̂0 ][∇0 · u(x0 )] − [∇0 gc (x − x0 )][n̂0 · u(x0 )]
γ
I
© ª
−µ dS 0 G(x, x0 ) · [n̂0 × ∇0 × u(x0 )] + [∇0 × G(x, x0 )] · [n̂0 × u(x0 )]
½
I u(x), x∈V
+u (x) = (8.132)
0, x∈/V
The above formula was first given by Morse and Feshbach [10, (13.1.10)]. If we add a null
quantity [11]
Z Z
−2ρc2s n̂0 · [∇0 × (u × G)]dS 0 = −2ρc2s ∇0 · [∇0 × (u × G)]dx0
S V
= 0 (8.133)
to (8.132), then the formula becomes
Z
{t(x0 ) · G(x, x0 ) − u(x0 ) · [n̂0 · Σ(x, x0 )]} dS 0 + uI (x)
S
½
u(x), x ∈ V
= (8.134)
0, x∈/V
where t is the traction vector which can be related to u by Hooke’s law, and Σ(x, x0 ) =
λI∇ · G + µ(∇G + G∇) is a third-rank Green’s tensor introduced by Pao and Varatharajulu
[11]. Eq. (8.134) is also known as the Somigliana’s identity [12]
Z
T T T
C (x)u(x) = [U (x, x0 )t(x0 ) − T (x, x0 )u(x0 )] dS 0 + uI (x) (8.135)
S
by identifying U = G and T = n̂0 · Σ, where U and T are the Stokes’ displacement and
traction tensors, respectively. In Eq. (8.135), C(x) is a tensor which takes the identity dyad
for x in V , 0 for x in V0 , and a real function of the geometry of S in the vicinity of x for
x on S. If the geometry is smooth at x, then C(x) = I/2. If we incorporate the boundary
conditions at the surface of a scatterer, which are the continuity of t and u for a solid-solid
interface as shown in Section 8.5, we have the integral equations
Z
1 T T
u(x) + [T (x, x0 ) · u(x0 ) − G (x, x0 ) · t(x0 )] dS 0
2 S
= uI (x), x ∈ S
Z
1 T T
u(x) + [G0 (x, x0 ) · t(x0 ) − T0 (x, x0 ) · u(x0 )] dS 0
2 S
= 0, x ∈ S (8.136)
The first equation above is obtained by letting the field point x approach the surface S from
the exterior of the scatterer and the second equation is achieved when x approaches S from
the interior of the scatterer. The kernels G and T are related to the medium in V while G0
and T0 are related to the medium in V0 . We have assumed that the geometry is smooth at the
field point x on S so that the coefficient 1/2 appears in the equations. From these boundary
integral equations (BIE’s)1 , the total displacement vector u and total traction vector t at
the scatterer surface can be solved. If the scatterer is a traction-free cavity, then the total
traction on the surface vanishes and the above equations reduce to
Z
1 T
u(x) + T (x, x0 ) · u(x0 ) dS 0 = uI (x), x ∈ S. (8.137)
2 S
On the other hand, if the scatterer is a fixed rigid inclusion, then the total displacement on
the surface vanishes and the above equations reduce to
Z
T
G (x, x0 ) · t(x0 ) dS 0 = −uI (x), x ∈ S. (8.138)
S
The above BIE’s can be solved using different methods such as boundary element method
(BEM) [12]– [20], method of fundamental solutions (MFS) [21], Nyström method [22], or
method of moments (MOM) [23]. In the MOM implementation as described in Chapters 2–4,
when applied to the elastic wave case, we need to decompose the unknown vectors into tan-
gential and normal components. This is because the unknown vectors are three-dimensional
1 In electromagnetic literature, these are also known as surface integral equations.
at the scatterer surface, which are different from the induced currents in electromagnetics.
Taking Eq. (8.138) as an example, we decompose the unknown traction vector as
t(x0 ) = tt (x0 ) + tn (x0 ) (8.139)
where tt (x0 ) is the tangential component and tn (x0 ) is the normal component. We then
expand the two components using Rao-Wilton-Glisson (RWG) basis and pulse basis, respec-
tively
Nt
X
tt (x0 ) = αn f n (x0 )
n=1
Nn
X
tn (x0 ) = βn n̂n (x0 ). (8.140)
n=1
In the above, f n (x0 ) is the RWG basis as shown in Figure 8.6, n̂n (x0 ) is the unit normal
vector of the nth triangle patch used as a pulse basis, and αn and βn represent the unknown
expansion coefficients to be solved. The RWG basis is redefined here as [24]
 + 0
`n 0 +

 2Sn+ Λn (x ), x ∈ Sn
0 `n − 0
f n (x ) = − Λn (x ), x ∈ Sn−
0 (8.141)

 2Sn
0, otherwise.
where `n is the length of the common edge of two neighboring triangles, Sn+ and Sn− are the
areas of the two triangles, and Λ+ 0 − 0
n (x ) and Λn (x ) are the distance vectors as indicated in
Figure 8.6. We have in total Nt nonboundary edges connecting two neighboring triangles
in which the RWG bases are defined and Nn triangles in which the pulse bases are defined.
After using the above expansion, the BIE can be written as
Nt
X Z
αn G(x, x0 ) · f n (x0 ) dS 0 +
n=1 Sn
Nn
X Z
βn G(x, x0 ) · n̂n (x0 ) dS 0 = −uI (x). (8.142)
n=1 Sn
where we omit the transpose T on G(x, x0 ) because G(x, x0 ) is symmetrical [12]. The next
step is the testing using the basis functions as the weighting functions. By doing so, the
following matrix equations are formed
Nt
X Nn
X
αn Amn + βn Bmn = Emn , m = 1, · · · , Nt
n=1 n=1
Nt
X Nn
X
αn Cmn + βn Dmn = Fmn , m = 1, · · · , Nn .
n=1 n=1
(8.143)
−
(x′)
+
x′ x′ n
x 0− n
n (x′)
x 0+ n
S n+ S n−
Figure 8.6: RWG basis f n defined in two neighboring triangles Sn+ and Sn− . These two
triangles share the nth nonboundary edge whose length is `n . Here Sn+ and Sn− also denote
the corresponding areas of the two triangles.
where
Amn = hf m (x), G(x, x0 ), f n (x0 )i
`m `n + 0 + 0
= + + hΛm (x), G(x, x ), Λn (x )i +
4Sm Sn
`m `n + 0 − 0
+ − hΛm (x), G(x, x ), Λn (x )i +
4Sm Sn
`m `n − 0 + 0
− + hΛm (x), G(x, x ), Λn (x )i +
4Sm Sn
`m `n − 0 − 0
− − hΛm (x), G(x, x ), Λn (x )i
4Sm Sn
Bmn = hf m (x), G(x, x0 ), n̂n (x0 )i
`m + 0 0
= + hΛm (x), G(x, x ), n̂n (x )i +
2Sm
`m − 0 0
− hΛm (x), G(x, x ), n̂n (x )i
2Sm
Cmn = hn̂m (x), G(x, x0 ), f n (x0 )i
`n
= hn̂m (x), G(x, x0 ), Λ+ 0
n (x )i +
2Sn+
`n
hn̂m (x), G(x, x0 ), Λ− 0
n (x )i
2Sn−
Dmn = hn̂m (x), G(x, x0 ), n̂n (x0 )i
Emn = hf m (x), uI (x)i
`m `m
= − + hΛ+ I
m (x), u (x)i −
− I
− hΛm (x), u (x)i
2Sm 2Sm
Fmn = −hn̂m (x), uI (x)i (8.144)
with
Λ+ 0
n (x ) = x0 − x+
0n
− 0 −
Λn (x ) = x0n − x0 . (8.145)
The involved inner products above are defined exemplarily as
hf m (x), G(x, x0 ), f n (x0 )i =

Z Z
dSf m (x) · G(x, x0 ) · f n (x0 ) dS 0
Sm Sn
Z
I
hf m (x), u (x)i = f m (x) · uI (x) dS. (8.146)
Sm
To illustrate the MOM solutions for the above BIE’s, we first consider the scattering by
a fixed rigid sphere with a radius of a = 1.0. The host medium has Poisson’s ratio ν = 0.25
and mass density ρ = 1.0. The incident wave is a time-harmonic dilatational plane wave
propagating along −x3 direction with a unit circular frequency (ω = 1.0) and normalized
wave number of κc a = π. Figure 8.7 shows the radial and tangential (elevated) components
of total traction along the principal cut (φ = 0◦ and θ = 0◦ ∼ 180◦ ) at the sphere surface. It
can be seen that the solutions agree with the analytical solutions very well. The analytical
solutions for general elastic wave scattering by a sphere can be found in [25].
We then consider the scattering in an elastic medium by a traction-free spherical cavity.
The cavity has a radius of a = 1.0 and the host medium is characterized by Poisson’s ratio
ν = 1/3, Young’s modulus E = 2/3 and mass density ρ = 1.0. The incident wave is the
same as before but with the normalized wave number of κc a = 0.913. Figure 8.8 plots the
radial and tangential (elevated) components of total displacement along the principal cut at
the surface. The solutions are also very close to the analytical solutions.
For the generalized case with both the host medium and scatterer being elastic, we select
λ = 0.53486, µ = 0.23077 and ρ = 1.0 for the host medium, and λ0 = 0.23716, µ0 = 0.52641
and ρ0 = 1.9852 for the elastic spherical inclusion with a unit radius. The incident wave is
the same as the one for scattering by the spherical cavity. Figure 8.9 depicts the radial and
tangential (elevated) components of total displacement at the surface along the principal cut.
These results are also in excellent agreement with the analytical solutions.
8.10 Conclusions
In this chapter, we start with the motion of particles under a small perturbation in a homo-
geneous elastic medium and derive the elastic wave equation in a PDE form. The equation
is then solved by using different techniques, that is, Helmholtz decomposition, time-domain
approach, Fourier-Laplace transformation, and finite difference scheme. For inhomogeneous
media, we derive the boundary conditions for different interfaces, that is, solid-solid, solid-
fluid and fluid-fluid interfaces, and address the elastic wave physics in planar layered media.
Moreover, we derive the corresponding integral equations from the PDE counterpart for elas-
tic wave scattering by using the equivalence principle and extinction theorem. The integral
equations are solved by using MOM which is first implemented in elastodynamics. Since the
0.7
|t | Exact
rr
Radial and Tangential Components of Traction

|t | Exact
0.6 rθ
|trr| MOM
|trθ| MOM
0.5
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140 160 180
θ (Degrees)
Figure 8.7: Radial and tangential (elevated) components of total traction along the principal
cut at the surface of a rigid sphere, kc a = π.
2.5
|ur| Exact
Radial and Tangential Components of Displacement
|uθ| Exact
2 |ur| MOM
|uθ| MOM
1.5
0.5
0
0 20 40 60 80 100 120 140 160 180
θ (Degrees)
Figure 8.8: Radial and tangential (elevated) components of total displacement along the
principal cut at the surface of a spherical cavity, kc a = 0.913.
1.4
Radial and Tangential Components of Displacement

|u | Exact
r
1.2 |uθ| Exact
|ur| MOM
1 |uθ| MOM
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120 140 160 180
θ (Degrees)
Figure 8.9: Radial and tangential (elevated) components of total displacement along the
principal cut at the surface of an elastic sphere, kc a = 0.913.
unknown functions are three-dimensional over a surface in the BIE’s, the implementation
of MOM is different from that in electromagnetics. Several numerical examples are used to
demonstrate the effectiveness of the solution technique.
Bibliography
[2] W. C. Chew, Elastic Wave Class Notes, U. Illinois, Urbana-Champaign, Fall, 1991.
[3] K. Aki and P. G. Richards, Quantitative Seismology: theory and methods, Chapter 1, San
Francisco: Freeman, 1980.
[4] J. A. Hudson, The Excitation and Propagation of Elastic Waves, pp. 15–24, Cambridge,
MA: Combridge University Press, 1980.
[5] J. D. Archenbach, Wave Propagation in Elastic Solids, Chapter 1, Amsterdam: North

Holland, 1973.
[6] Y.-H. Chen, W. C. Chew, and Q. H. Liu, “A three-dimensional finite difference code for
the modeling of sonic logging tools,” J. Acoust. Soc. Am., vol. 93, pp. 702–712, 1998.
[7] W. C. Chew and Q. H. Liu, “Perfectly matched layers for elastodynamics: A new absorbing
boundary condition,” J. Comput. Acoust., vol. 4, pp. 341–359, 1996.
[8] C. J. Randall, “Absorbing boundary condition for the elastic wave equation: velocity-
stress formulation,” Geophys., vol. 54, pp. 1141–1152, 1989.
[9] Q.-H. Liu, E. Schoen, F. Daube, C. Randall, H.-L. Liu, and P. Lee, “A three-dimensional
finite difference simulation of sonic logging,” J. Acoust. Soc. Am., vol. 100, pp. 72–79,
1996.
[10] P. M. Morse and H. Feshbach, Methods of Theoretical Physics, pp. 1770, New York:
McGraw-Hill, 1953.
[11] Y.-H. Pao and V. Varatharajulu, “Huygens’ principle, radiation conditions, and integral
formulas for the scattering of elastic waves,” J. Acoust. Soc. Am., vol. 59, pp. 1361–1371,
1976.
[12] F. J. Rizzo, D. J. Shippy, and M. Rezayat, “A boundary integral equation method for
radiation and scattering of elastic waves in three dimensions,” Int. J. Numer. Methods.
Eng., vol. 21, pp. 115–129, 1985.
2
[13] G. D. Manolis and D. E. Beskos, Boundary Element Methods in Elastodynamics, London:

Unwin Hyman, 1988.
[14] I. R. Gonsalves, D. J. Shippy, and F. J. Rizzo, “Direct boundary integral equations for
elastodynamics in three-dimensional half-spaces,” Comput. Mech., vol. 6, pp. 279–292,
1990.
[15] F. J. Rizzo and D. J. Shippy, “An advanced boundary integral equation method for
three-dimensional thermoelasticity,” Int. J. Numer. Methods. Eng., vol. 11, pp. 1753–
1768, 1977.
[16] T. A. Cruse, “An improved boundary-integral equation method for three dimensional
elastic stress analysis,” Comput. Struct., vol. 4, pp. 741–754, 1974.
[17] Y. Liu and F. J. Rizzo, “Hypersingular boundary integral equations for radiation and
scattering of elastic waves in three dimensions,” Comput. Methods Appl. Mech. Eng., vol.
107, pp. 131–144, 1993.
[18] W. S. Hall, The Boundary Element Method, London: Kluwer Academic Publ., 1994.
[19] P. J. Schafbuch, R. B. Thompson, and F. J. Rizzo, “Application of the boundary element

method to the elastic wave scattering by irregular defects,” J. Nondestr. Eval., vol. 2, pp.
113–127, 1990.
[20] P. S. Kondapalli, “Frequency response of elastic bodies of revolution by the boundary
element method,” J. Acoust. Soc. Am., vol. 98, pp. 1558–1564, 1995.
[21] P. S. Kondapalli and D. J. Shippy, “The method of fundamental solutions for transmission
and scattering of elastic waves,” Comput. Methods Appl. Mech. Eng., vol. 96, pp. 255–269,
1992.
[22] M. S. Tong and W. C. Chew, “Nyström method for elastic wave scattering by three-
dimensional obstacles,” J. Comput. Phys., vol. 226, pp. 1845–1858, 2007.
[23] M. S. Tong and W. C. Chew, “Unified boundary integral equation for the scattering of
elastic wave and acoustic wave: solution by the method of moments,” Waves Random
Complex Media, vol. 18, pp. 303–324, 2008.
[24] S. M. Rao, D. R. Wilton, and A. W. Glisson, “Electromagnetic scattering by surfaces of
arbitrary shape,” IEEE Trans. Antennas Propag., vol. 30, pp. 409–418, 1982.
[25] Y.-H. Pao and C. C. Mow, “Scattering of plane compressional waves by a spherical
obstacle,” J. Appl. Phys., vol. 34, pp. 493–499, 1963.
223
Glossary of Acronymns
AIM adaptive integral method;

BCG, BiCG biconjugate gradient;
BCGSTAB BCG stabilized;
BEM boundary element method;
BIE boundary integral equation;
CEM computational electromagnetics;
CFD computational fluid dynamics;
CFIE combined field integral equation;
CG conjugate gradient;
CPU central processing unit time;
CSIE combined source integral equation;
DCIM discrete complex image method;
DGLM dyadic Green’s function for layered media;
EFIE electric field integral equation;
EPA equivalence principle algorithm;
FDM finite difference method;
FEM finite element method;
FERM finite element radiation method;
FIPWA fast inhomogeneous plane wave algorithm;
FISC Fast Illinois Solver Code;
FMM fast multipole method;
2
FMM-PML-MPIE FMM, perfectly matched layer, mixed potential integral equation;
FSIP folded Sommerfeld integration path;
GIBC generalized impedance boundary condition;
GMRES generalized minimal residual;
H polarization horizontal polarization;
HH response horizontal-horizontal response;
HV response horizontal-vertical response;
IBC impedance boundary condition;
IBC-MPW IBC modified plane wave;
IBC-WI IBC wave impedance;
LF-FMA low-frequency fast multipole algorithm;
LF-MLFMA low-frequency multilevel fast multipole algorithm;
LF-MOM low-frequency method of moments;
LPP long patch pair;
LUD lower-upper triangular decomposition;
MF-FMA mixed-form fast multipole algorithm;
MFIE magnetic field integral equation;
MFS method of fundamental solutions;
ML-FIPWA multilevel fast inhomogeneous plane wave algorithm;
MLFMA multilevel fast multipole algorithm;
MMIC monolithic microwave integrated circuit;
MOM method of moments;
NCSA National Center for Supercomputing Applications;
P wave pressure wave, compressional wave;
PDE partial differential equation;
PEC perfect electric conducting;
PMC perfect magnetic conductor;

GLOSSARY OF ACRONYMNS 225
PMCHWT formulation Poggio-Miller-Chang-Harrington-Wu-Tsai formulation;
QMR quasi-minimal residual;

RAM random access memory;
RCS radar cross section;
RFID radio frequency identification tag;

RWG Rao-Wilton-Glisson;
S wave shear wave;

SBI self-box inclusion;
ScaleME scaleable multipole engine;
SDFMA steepest descent fast multipole algorithm;
SDFMM steepest descent fast multipole method;
SDP steepest descent path;

SH wave shear horizontal wave;
SIE surface integral equation;
SIP Sommerfeld integration path;

SPP short patch pair;
SV wave shear vertical wave;
TDS thin dielectric sheet;
TE transverse electric;
TFQMR transpose free quasi-minimal residual;
TM transverse magnetic;
TSM-FMA thin-stratified medium fast-multipole algorithm;

VIE volume integral equation;
VV response vertical-vertical response;
227
About the Authors
Weng Cho CHEW received the B.S. (1976), M.S. (1978), Engineer’s (1978), and Ph.D.
(1980) degrees in EE from MIT. He is serving as the Dean of Engineering at The University of
Hong Kong on leave of absence from the University of Illinois at Urbana-Champaign (UIUC).
Previously, he was a professor and the director of the Electromagnetics Laboratory at the
University of Illinois. Before that, he was a manager and program leader at Schlumberger-Doll
Research. He was on IEEE Adcom for AP-S and GRSS, and was active with various journals
and societies. He researches in wave physics for various applications and fast algorithms for
solving wave problems, and originates several fast algorithms for solving scattering and inverse
problems. A research group that he led, for the first time, solved dense matrix systems with
tens of millions of unknowns for integral equations of scattering. He wrote the book Waves
and Fields in Inhomogeneous Media, coauthored the book Fast and Efficient Methods in
Computational Electromagnetics, as well as numerous journal, conference publications, and
book chapters. He is an IEEE, OSA, and IOP Fellow, and was an NSF PYI (USA). He
received the Schelkunoff Best Paper Award for TAP, the IEEE Graduate Teaching Award,
UIUC Campus Wide Teaching Award, and IBM Faculty Awards. He was a Founder Professor
and is currently, a Y.T. Lo Endowed Chair Professor at UIUC. Recently, he served as an IEEE
Distinguished Lecturer, and Cheng Tsang Man Visiting Professor at Nanyang Technological
University in Singapore. ISI Citation elected him to the category of Most-Highly Cited
Authors (top 0.5%). This year, he has been chosen to receive the IEEE AP-S Distinguished
Educator Award.
Mei Song TONG is a visiting research scientist at the Center for Computational Electro-
magnetics and Electromagnetics Laboratory (CCEML), Department of Electrical and Com-
puter Engineering (ECE), University of Illinois at Urbana-Champaign (UIUC). He received
the B.S. and M.S. degrees from Huazhong University of Science and Technology (China)
in 1985 and 1988, respectively, and the Ph.D. degree from Arizona State University (ASU)
in 2004, all in electrical engineering. His research interests include numerical methods in
electromagnetics, acoustics and elastodynamics, simulation and design for antennas and
RF/microwave circuits, and electronic packaging for digital devices. He is a senior mem-
ber of the Institute of Electrical and Electronics Engineers (IEEE).
Bin HU received the B.S. degree in Electrical Engineering from Tsinghua University in 1995,
the M.S. degree from Syracuse University in 1997, and Ph.D. degree from University of Illinois
2
at Urbana-Champaign in 2001. Since 2001, he has been with Process Technology Modeling
division at the Intel Corporation, currently as a Senior Staff Research Scientist. Dr. Hu leads
the development of advanced lithography modeling techniques, a key enabling component of
resolution enhancement solutions for Intel’s next generation process technology. His research
interests include all aspects of lithography modeling techniques and fast algorithms in com-
putational electromagnetics. Dr. Hu is the co-author of 9 journal articles and book chapters,
15 conference papers, and 4 (pending) patents. At Intel, Dr. Hu received numerous awards
for his work on lithography modeling, including the prestigious Intel Achievement Award. He
was the recipient of Y.T. Lo Outstanding Research Award in 2000.
229
Index
Ez -Hz formulation, 135, 145, 151 lossless, PEC, PMC, 36, 39

L2 space, 25 lossy cavity wall, 82
L and K operators, 50, 126, 144 lossy, IBC, 57
L, M,and N functions, 137 physics, 3, 4
resonant mode for CFIE, 81
acoustic medium, 198, 203 solid-fluid interface, 217
adaptive integral method (AIM), see AIM solid-solid interface, 203, 207
adjoint operator, 26 surface of a scatterer, 214
AIM, 160 boundary element method (BEM), see BEM
Ampere’s law, 2, 5 boundary integral equation (BIE), see BIE
analytical solution, elastic wave, sphere, 217 bra and ket notation, 29
anisotropic medium, 33 braem, 29
VIE, 86 branch point contribution, 174
antenna, 6, 65, 122, 124, 159 breakdown frequency, MFIE, EFIE, 79
asymptotic method, 6, 8 broadband fast algorithm, 128
broadband signals, data, 6, 11
bandwidth of matrix, 119 buried bunker, land mine, unexploded ord-
basis function, 51, 215 nance, 181, 184, 187
MOM, 27
basis rearrangement method, 116, 122, 127, capacitance matrix of electrostatics, 113
130 Cauchy theorem, 172
beam-forming, high frequency, 6 CEM, 1, 2, 6, 8, 152
BEM, 28, 214 hypersingularity, 199
Bessel function, 101, 135, 139, 141, 152, 174 integral equation, 43
bianisotropic medium, 34, 38 layered medium, 135
biconjugate gradient linear vector space, 22
(BiCG), see iterative method nondissipative Yee scheme, 11
BIE, 214, 215, 217, 219 reaction inner product, 24
bilinear form, pseudo inner product, 24 wave physics, 22
body force, elastic body, 195, 200 central processing unit (CPU) time, see CPU
bound, convergence of ill-conditioned matrix, time
62 CFD, 1, 11
boundary condition, 27, 36, 57, 112, 202, 203 CFIE, 57, 74, 81–83, 160
constraint, layered media, 208 CG, 124, 177
coupling of SV and P waves, 203 CG, BCG, BiCG, BCGSTAB, 60
for reflection matrix, 205 charge neutrality, 114, 116, 117, 121, 127
2
pulse-pair function, 115 per matrix-vector multiplication, 177

charge term of EFIE operator, 108, 122 solve, 180, 187
circuit physics, 108, 109, 112, 130, 199 CSIE, 83
circular cylinder, 77, 176, 181 curl-conforming function, 86, 87
PEC, 2D, 79 curl-free component, 5
closed surface, 55, 56, 112, 113, 126, 160 elastic wave, 198
combined field integral equation (CFIE), see current sheet, 49, 63, 65
CFIE current, curl and divergence free parts, 108
combined source integral equation (CSIE), see cut-and-try method, 8
CSIE cylindrical surface, reflection, 96, 98
compact operator, smoothing operator, 29
comparison of derivatives, 3, 4 DCIM, 159, 160
complex permittivity, permeability, 58 defective meshes, 54
complex plane, 57, 62 degree of freedom, 13
complex power, 25, 37–39 delta function, 86–88, 202
complex resonance, 81, 82 collision, 90
complex symmetric system, 24 delta-function-like singularity, 90
composite object, surface, 126 delta-gap source, 128
compressional wave, 197, 198, 201, see also P dense matrix, 10, 59, 60
wave DGLM, 135, 136, 140
computational complexity, 13, 14, 160, 166, diagonal blocks, frequency normalized, 118
175, 177 diagonal translator, 166, 168, 170, 171, 174,
computational electromagnetics (CEM), see 175
CEM diagonalization of translator, 166–168, 175
computational fluid dynamics (CFD), see CFD dielectric scatterer, 58, 71, 79, 80
computational resource, 13, 14, 59 differential equation solution, 9–11, 71, 108,
condition number, 61–63 110
conjugate gradient (CG), see iterative method, differential equation, ordinary, layered media,
see CG 138
connection matrix, inversion, 119, 121 differential operator, 23, 28, 35, 38, 84
constant phase path, 174 Hermitian, 36, 37
constitutive relation, 22, 34, 196 noncompact, 29
contact-region modeling method, 126, 128 nonuniqueness, 27
continuity equation, 5, 117 Dirac’s bra and ket notation, 26
contour integration, 137, 172, 201 direct field, see primary field
convergence of iterative solver, 113 direct solver, 59, 60
dependence on RHS, 62 discrete complex image method (DCIM), 159
relationship to small eigenvalues, 63 displacement field, 194, 203, 206, 214, 217
understanding, roundoff error, 61 Helmholtz decomposition, 198
coordinate space, 135, 137–139 of particles, 193, 194, 206
coordinate transform, 166, 174, 175 strain, stress, stretched particles, 195
Coulomb’s law, 2 divergence conforming function, 87, 93
CPU time, 10, 13, 14, 59, 184 divergence-free component, 5, 136, 137, 198
per iteration, 166, 175, 180, 184 domain space, 23, 27, 28
O(N log N ), 187 domain-decomposition method, 9
INDEX 231
doughnut-shape object, 55, 113 weighted sum with MFIE, 74

duality principle, 48, 73, 79, 86, 143, 144 eigenvalue, 26, 36, 37, 57, 109
dyadic Green’s function, 44, 84, 85, 138, 152, compact or noncompact operators, 29
159 dynamic range, 61
canonical scalar potential, 153 energy conservation, 36
elastic medium, 107, 199, 211 small, 109, 122
electric field, 44, 48, 137, 140 eigenvalue analysis, 61–63
electromagnetic vs elastodynamic, 199 elastic medium, 193, 217
field decomposition, 137, 139, 145 Green’s function, 211
in deriving integral equation, 43 homogeneous medium scatterer, 217
in MFIE, 160 elastic wave, 193
layered medium, 135–137, 152, 159 equation of motion, 202, 210, 217
singularity cancelation, 153 integral equation, 211, 213
mixed potential formulation, 136 MOM implementation, 214
plane wave expansions, 184 PDE, 211
reflected part, 138 elastic wave equation, 196
singularity, 44, 85, 107, 199 balance of forces, 194
spectral domain, 147, 148, 162 Fourier-Laplace transform, 197, 200
symmetric, 161 homogeneous isotropic medium, 197, 205
vector wave functions, 136 elastodynamics, see elastic wave
VIE, 86 electric dyadic Green’s function, see dyadic
dyadic Green’s function for layered Green’s function, electric field
media (DGLM), see DGLM electric field cancelation inside a scatterer, 48
electric field coupling with magnetic field, 5,
eddy current, 108, 126 6
edge element, VIE, 89 electric field integral equation
EFIE, 49, 50, 56, 75–77, 79, 81 (EFIE), see EFIE
as ω → 0, 111 electric polarization current, 85
dyadic Green’s operator, 152 electromagnetic energy, 56, 71
extinction theorem, 74, 76 electromagnetic equation, 1, 4, 8, 107
first-kind integral equation, 56 differential, integral equation method, 11
Galerkin’s method, 110 frequency domain, 2, 8, 9
GIBC, 95 Maxwell’s equations, 2
impenetrable object, 144 electromagnetic physics, 3, 4
in CFIE, 81, 161 electromagnetics, relation to elastic wave, 193
inhomogeneous layered medium, 160 electroquasistatics, 108, 112, 113, 116, 199
internal resonance, 57, 75, 76, 79, 80 electrostatics, 5, 6
low-frequency breakdown, 107 elliptic equation, 5
MFIE, error comparison, 77 embedded scatterer, layered medium, 135
MOM, 56 energy conservation theorem, 22, 36, 38, 39
open, closed surface, 55, 56 energy inner product, 25, 26
operator, 108, 109, 126 EPA, 130
PEC, 50 equivalence principle, 49, 72, 76, 78, 80
penetrable scatterer, 72, 80 derivation from vector wave equation, 45
TDS, 91 elastic wave, 211, 217
equivalence surface current, 36, 47, 49, finite element method (FEM), see FEM
72, 97, 99 finite element radiation method (FERM), see
Huygens’ principle, 44 FERM
mathematical form, penetrable object, 48 FIPWA, 160, 163, 172, 177, 187
equivalence principle algorithm (EPA), see EPA 2D, 168
equivalence volume source, 84 cf. full matrix solver, 181
expansion function, 27, 29, 87, 148 diagonal translator, 166
approximate, 57 direct part, primary part, 166
completeness, 28, 51 free-space, 163, 166, 167
curl-conforming, 87, 140 layered medium, 166, 175
for range and domain space, 83 multilevel, 184
for range space of compact operator, 29 first-kind integral equation, 56
for singular and smooth functions, 28 FISC, 63, 64, 177, 180
Galerkin’s method, 28, 52 fixed rigid inclusion, 214, 217
layered media, 140, 142, 143 fluid medium, 198
patch pair, 111, 127 fluid-solid, fluid-fluid interface, 203
RWG function, 52, 53 FMM, 160
expectation value, quantum mechanics, 26 2D, 168, 172
extinction theorem, 36, 100 alternative to, 163, 166
derivation of integral equation, 49, 72, FMM-like fast algorithm, 166, 170
213 FMM-PML-MPIE, 160
elastic wave, 211, 217 folded SIP (FSIP), see FSIP
equivalence principle, 44, 45, 47, 49, 72 Fourier integral
general penetrable half space, 97, 99 layered media, 135, 139
inside-out manner, 73, 74 three-dimensional, 136
MFIE, 50, 56, 57 Fourier space, 21, 24, 36, 43
extrapolation in FIPWA, 165, 170, 171 Fourier transform, 2, 8, 21
Fourier-Laplace transform, 200, 217
far-field pattern, 170, 171, 175 free-space permittivity, permeability, 58
far-field regime, zone, 5, 7, 176, 181 frequency domain, 9, 21, 24, 43
Faraday’s law, 2 frequency normalization, 111, 116, 118
fast algorithm, 11, 13, 59, 63, 65 frequency-normalized matrix blocks, 111
Fast Illinois Solver Code (FISC), see FISC FSIP, 163, 175
fast inhomogeneous plane wave algorithm full-wave physics, 112
(FIPWA), see FIPWA
fast multipole method (FMM), see FMM Galerkin’s method, 28, 52, 110
FDM, 23, 71 Gauss’ divergence theorem, 32
FEM, 23, 28, 71, 89, 119 Gauss’ law, 2
FERM, 124 Gauss-Laguerre, Gauss-Legendre methods, 168
field point, 138–140 Gaussian elimination, 13, 59
at, near interface, 139 Gedanken experiment, 77–80, 84
coplanar with source point, 137 internal resonance, 75, 80, 102
exterior, interior to the scatterer, 214 generalized eigenvalue problem, 36
finite difference method, 141, 152, 211, 217 generalized impedance boundary condition
finite difference method (FDM), see FDM (GIBC), see GIBC
INDEX 233
generalized minimal residual (GMRES), see Hermitian tensor, 34, 36–38

iterative method Hertz vector potential, 135
generalized reflection coefficient, 139, 162, 171, Hertzian dipole, 63, 145
175 modeling, 124
generalized reflection operator, 209 Hertzian dipole, modeling, 124
geometrical optics, theory of diffraction, 8 HH (horizontal-horizontal) response, 180, 181
GIBC, 95, 96, 102 high frequency method, 71
Grammian, 31 high frequency regime, 3, 4, 6, 8, 29
Green’s function, 2, 9, 114, 135 high-speed circuit, 107
elastic wave, 198, 204 highly conductive medium, 96, 101
exact propagator, free space, 10, 39 Hilbert space, 22, 29, 30
hypersingular, 44, 56 homogeneous isotropic medium, elastic, 197,
integral equation, 9, 10 203
layered medium, 135, 153, 159, 162, 166, homogeneous medium, 44, 136–138, 145, 196
175, 187 EFIE and MFIE operators, 144
matrix representation, 127 stability criterion, 211
reflected part, 139, 175 homogeneous solution, 4
scalar, 198 homogenization method, 128
tabulated, 152 Hooke’s law, 196, 214
TM, TE decomposition, 141 Huygens’ principle, see equivalence principle
Green’s function, dyadic, see dyadic Green’s HV (horizontal-vertical) response, 181
function hyperbolic PDEs, 6
grid dispersion error, discretization density, hypersingularity, 199
10
IBC, 34, 58, 102
H-polarization, horizontal polarization, 180 CFIE, 81
Haar function, 120, 121 cylinder, 98, 101
half-plane and half-space problem, 7 lossy scatterer, advantage, 95
Hamiltonian operator, 26 IBC-MPW, IBC-WI, 102
Hankel function, 77, 163, 167, 168, 174 ill-conditioned matrix, 56, 126, 180, 181
Hankel function, Bessel function, 101 disparate mesh size, 63
hard-drive memory, 14 internal resonance, 57
harmonics, higher, mixing, 9 iterative solution, 62
Helmholtz decomposition, 5, 6, 108–110, 198, low frequency, 109
217 ill-convergent integral, layered media, 137
Helmholtz equation, 9, 10 image theorem, 167
Helmholtz theorem, 137 impedance boundary condition (IBC), see IBC
also, see Helmholtz decomposition impedance matrix, 140, 142, 159, 160, 181
Hermitian matrix, 61, 62 conventional, 118
positive definite, 61, 115 filling time, 160
Hermitian operator, system, 26, 36–39 impenetrable object, 10, 71, 75, 83
and boundary condition, 36 indicial notation, 72, 193, 194, 199
energy conservation, 22 induction term of EFIE operator, 108, 109
lossless, 26, 38, 39 inertial force, elastic body, 195
self-adjoint, 26 infinite dimensional space, 22, 24, 28
inhomogeneous isotropic medium, elastic, 205 layered medium, 140, 143

inhomogeneous layered medium, 160, 206 low frequency, 124
inhomogeneous medium, 91, 140, 196, 217 intergalactic length scale, 2
anisotropic medium, 84 interior nodes, number of, 112
permeability, permittivity, 85, 88 internal resonance, 56, 57, 83, 96, 160
piecewise constant, 71 EFIE, 57, 75, 76
VIE, 71, 72 in Gedanken experiment, 75
inhomogeneous plane wave, 168 key to understanding of, 72
inner product, 24, 28, 35, 63, 90 MFIE, 57
definition, 24, 33 PMCHWT formulation, 74
electromagnetics, 24, 28, 29 remedy, see remedy for internal resonance
energy, 26 interpolation technique, 165, 170, 171, 177
for defining transpose operator, 26 irrotational component, current, 5, 6
mathematics, 25 low frequency, 108
pseudo, reaction, 24 irrotational vector wave function L, 136
quantum mechanics, 24, 25, 29 isotropic medium, symmetric tensor, 196
input impedance, conical antenna, 124 iterative method, 59, 116, 118, 127
integral equation, 10, 11, 48–50, 75, 95 dependence on right-hand side, 62
CFIE, 57 disadvantage, 63
combined source (CSIE), see CSIE effect due to internal resonance, 57
compared to differential equation, 10 error, roundoff, 61
conversion to matrix equation, 10, 51, 52 fixed-point iteration, Jacobi iteration, 60
derived from extinction theorem, 72 Hermitian positive definite matrix, 115
elastic wave scattering, 211, 213, 217 inversion, 95
Green’s function, 2, 9, 10 Krylov subspace, 60, 62
impenetrable object, 75 nonconvergent, 76
internal resonance problem, 75, 76, 81, optimal solution, 60
82 saturation, numerical roundoff, 61
layered medium, 135, 159 slow convergence, 113
matrix representation, 89 tank model, 180
MFIE, 73 CG, BCG, BiCG, BCGSTAB, 60
penetrable object, 71, 75, 102, 144 GMRES, QMR, TFQMR, 60
pulse-pair function, 115 iterative solver , see iterative method
solid-solid interface, 214
solver, 10, 11, 22, 23, 128 Kerr effect, 3
stabilized, 126 ketem, 29
TDS, 93 Kronecker delta function, 30
integral operator, 23, 30, 39, 51
compact, 29 Lamé constant, 196
MOM, BEM, 28 Laplace’s equation, 5, 6
integration by parts, 31, 34, 37, 44 large length-scale, 128
derivative operator, 85, 86, 148 large-scale computing, 13
deriving surface integral, 88 layered medium, 58, 136, 162, 166, 175
elastic wave, 198 Ez -Hz formulation, 135, 145, 151
in VIE, 85, 86, 88–90 branch points, 174
INDEX 235
CEM, 135 low-frequency MOM (LF-MOM), see LF-MOM

CFIE, 160 low-frequency multilevel fast multipole algo-
compared to free-space code, 177 rithm (LF-MLFMA), see LF-MLFMA
efficiency, 180, 187 low-frequency physics, circuits, devices, 107
elastic wave, 203, 205, 217 lower-upper triangular decomposition (LUD),
fast methods, 159, 160 see LUD
FIPWA, 166, 175, 176, 181, 187 LPP, 127
generalized reflection operator, 209 LUD, 59, 76
Green’s function, 135, 136, 153, 159, 187
image term, 175 Müller formulation, 75
integral equation, 159, 160 magnetic field coupled to electric field, 6
large-scale problem, 160, 184, 187 magnetic field integral equation
numerical result, 180 (MFIE), see MFIE
objects above, 184 magnetic polarization current, 85
penetrable and impenetrable object, 144 magnetoquasistatics, 108, 111, 112, 199
polarization effect, 139 magnetostatics, 5
SDFMA, TSMFMA, 175 magnetostatics, inductor, 6
TEz , TMz waves, 138 mass density, 195, 217
LF-FMA, 118, 127 matched asymptotics, approximate guidance
LF-MLFMA, 118, 124, 128 condition, 7
LF-MOM, 124, 128 matrix equation, 10, 13, 59, 114, 118
linear operator, 22, 23, 28, 39 basis rearrangement, 116
linear vector space, 22, 25, 39, 110 condition number, 61
linearized equation, 193 elastic wave, 215
long patch pair (LPP), see LPP electrostatic problem, 116
long wavelength physics, 11 frequency normalization, 111
long-distance communication, 6 from integral equation, 51
longitudinal wave, 136, 197, 199, 203 Galerkin’s method, 52, 110
loop basis, 110, 112, 116, 118, 121, 122, 124 Hermitian positive definite, 115
loop, global, entangled, irreducible, 113 iterative solution, 59–61
loop-tree decomposition, 110–112, 118, 124, MOM, 135
126 symmetric, 52
picture of, 122 well conditioned, 130
remedying low-frequency breakdown, 130 matrix filling issue, 160
search algorithm, 118 matrix free, 59
Lorentz reciprocity theorem, 35, 36 matrix norm, 62
lossless cavity, medium, 26, 36–39 matrix operator, equation, 23, 24, 27, 31
lossy surface, cavity, 57, 58 matrix partitioning method, 119
lossy, lossless material, 181 matrix solution, 181
low frequency, 3, 5, 108, 117, 130 matrix-friendly form, 136
capacitive nature of dipole, 124 matrix-vector multiplication, 118, 163, 166,
low-frequency breakdown, 107, 109, 124, 126, 170, 177, 181
130 fast algorithm acceleration, 118, 127
low-frequency fast multipole algorithm (LF- Maxwell’s equations, 2
FMA), see LF-FMA Maxwell-Heaviside equations, 2, 4, 7, 11
measurable quantity, 25, 26 EFIE, 55

memory and CPU requirement, 11, 13, 52 elastic wave case, 214, 217
memory requirement, 10, 59, 166, 176, 177, FISC, 63
184, 187 implementation of, 219
meshing, gridding, tessalation, 52 matrix equation, 117, 135, 148, 152
metamaterial, nanotechnology, 107 matrix representation, 28
method of fundamental solutions (MFS), see solving integral equation, 51
MFS momentum, position, observable, 25
method of moments (MOM), see MOM monolithic microwave integrated circuit
method of steepest descent, 175 (MMIC), see MMIC
metric, norm, 25 morphing of electromagnetic physics, 3, 11
MF-FMA, 128 Morse and Feshbach formula, 213
MFIE, 50, 56, 57, 95, 96 multidielectric-region problem, 126
CFIE, 81 multigrid, 5, 6
closed surface, 56 multilevel fast multipole algorithm (MLFMA),
FIPWA, 160, 161 see MLFMA
GIBC, 95 multilevel FIPWA (ML-FIPWA), 177, 187
impenetrable object, 144 multilevel implementation, 166, 174, 175
in Gedanken experiment, 77–79 multiscale problem, 107, 128, 130
internal resonance, 57, 75, 77, 79–81 mutual impedance matrix, 118
operator, 126
PEC case, 77, 79 nanotechnology, 107, 130
penetrable scatterer, 73, 80 National Center for Supercomputing Applica-
weighted sum with EFIE, 74 tions (NCSA), see NCSA
MFS, 214 Navier-Stokes equations, 1
Michalski-Zheng formulation, 136, 151 NCSA, 180, 187
mid frequency, 108 near field, regime, 5, 7
mid length-scale, 130 net power flow, object, 71
mid wavelength, physics, 11 Newton’s law, 195
mid-field regime, 5 nonradiating current, 56, 57, 75, 76
mid-frequency regime, 4, 6 noncompact operator, 29
Mie series solution, 7, 8, 77, 89 nondivergence free L function, 137
mirror image, 166 nondivergence free part of current, 5
mixed potential formulation, 136 nonlinear device, 107
mixed-form fast multipole algorithm nonlinear phenomenon, 1, 3, 7, 9, 22
(MF-FMA), see MF-FMA nonoscillatory, Laplace’s solution, 5
MKS unit, 118 nonsymmetric operator, 33
ML-FIPWA, 177, 180, 181, 184, 187 nonunique inverse of operator, 27
MLFMA, 63, 65, 176, 180 norm, normed vector space, 25
MMIC, 159 normal matrix, 62
MOM, 23, 63, 71, 77, 95, 217 number of iterations, 59, 61, 63, 76
approximation of solution, 56 number of unknowns, 22, 55, 56, 176, 184
code, 180 CFIE, impenetrable scatterer, 83
conventional, 177 circular cylinder and tank models, 180
discretization, 160 degree of freedom, 13
INDEX 237
FIPWA example, 177, 180, 187 penetrable object, 71–73, 75, 80, 81, 95
integral equation, 93 EFIE, 72
multilevel implementation, 175 IBC, GIBC, 95, 102
VIE, 89 in layered media, 144
numerical quadrature, 165, 174 inhomogeneous medium, 102
numerical roundoff error, 57, 61 internal resonance problem, 79
Nyström method, 23, 214 penetrable scatterer, see penetrable object
perfect electric conducting (PEC), see PEC
observable quantity, operator, 25, 26 perfect magnetic conductor (PMC), see PMC
observation group, 166, 168, 174–176 perturbation expansion, method, 8
source group, 164, 166, 170, 171 phase velocity, grid dispersion error, 10
open surface, 55, 56, 112 physical optics, 8
operating frequency, 76, 78–80, 151 physics of fluids, 1, 7
operator, invertible, 27 physics of inductors, capacitors, 112
optical fiber, communication, dispersion man- physics regime, 3
agement, soliton, 3 piecewise constant medium, 71, 86, 88
optical frequency, 3, 7 pilot vector potential approach, 135–137, 152
ordinary differential equation, layered media,
Planck’s constant, 7
138
plane wave, radiation condition, 34
orthogonal, orthonormal basis, 30, 31
PMC, 34, 71, 75, 79, 81
oscillatory nature, mid-frequency regime, 6
cavity, 78, 81
out-of-core solver, 14
Gedanken experiment, 78, 80
P wave, 197, 198, 203–207 PMCHWT (Poggio-Miller-Chang-Harrington-
parallel computer, 14, 63 Wu-Tsai) formulation, see PMCHWT
partial differential equation (PDE), see PDE formulation
patch-pair basis, 127 PMCHWT formulation, 74, 75, 83, 124, 126
PDE, 1, 217 point source field, 145
PEC, 58, 63, 71, 160 elastic wave, 205
and dielectric composite structure, 128 electric field, 137, 147
cavity, 57, 76, 79–81 Hertzian dipole, 145
CFIE, 56, 74 lower half space, 149
circular cylinder, 77 three-layered media, 150
delta-gap source, 128 Poisson’s equation, 5, 33
EFIE, MFIE, 50, 75, 76 Poisson’s ratio, 217
Gedanken experiment, 78 polarization current, penetrable object, 72
ground plane, 173 polarization effect, 139
impenetrable object, 75 pole contribution, 137, 171, 172
integral equation, 48 and branch point contribution, 171
internal resonance, 50, 56, 75, 77, 81 and branch point contributions, 167, 175
MOM, 51 unphysical, 139
scatterer, 48, 49, 181 pole locating with contour integral, 171–173
sphere, 151 pole, complex resonance, 57
surface charge density, 117 Poynting theorem, 36, 37
penetrable half-space, cylinder, 96–98, 101 preconditioner, 63, 64, 111, 124, 132, 181
pressure (P) wave, see P wave, see also com- Rao-Wilton-Glisson (RWG)
pressional wave function, see RWG function
primary field contribution, 138, 141, 145 ray physics, 4–6, 11
coordinate space representation, 139 transport equation, eikonal equation, 6
matrix representation, 140 Rayleigh scattering, solution, 7
principal value integral method, 56 Rayleigh-Ritz method, 28
probability function, quantum mechanics, 26 RCS, 151
projection of operator, 30 FIPWA, 177, 180, 181, 184, 187
pseudo inner product, 24 reaction inner product, 24, 52
pulse basis, 115, 121, 127 reactive boundary condition, 75
charge neutrality, 117 CFIE, 57, 58
connection matrix, 119 IBC, 83
relation to pulse pair function, 115 reactive power, 38
pulse basis, normal component, 215 receiving pattern function, 164, 175
pulse function, 77, 93 reciprocity theorem, 22, 34, 39, 44, 47
pulse shape, compression, expansion, 3 a new look, 33
pulse-pair function, 115, 116 complex symmetric system, 24, 34
elastodynamics, 196
quad patch in spanning tree, 121 Lorentz, 35
quadrature sample, 170 symmetric measurement, 34
qualitative engineering, 8 reflected field, see secondary field
quantization of electromagnetic energy, 7 reflected-wave part, 162
quantum mechanics, 24–26, 29, 193 elastic wave, 207
quantum physics, see quantum mechanics FIPWA, 166, 175, 176
quantum system, 36 reflection coefficient, 172, 173
quantum-mechanics inner product, 26 reflection matrix, 205, 208, 209
quasi diagonal operator, conductive media, 95 remedy for internal resonance
quasi-Helmholtz decomposition, 109–111 CFIE, 74
quasi-minimal residual (QMR), see iterative insertion of lossy object, 83
method insertion of shorts, 83
quasi-planar object, 170 residual error, 60–62, 172, 173, 180, 187
resistive sheet, R-card, 58, 91
radar cross section (RCS), see RCS resonance cavity, condition, 56, 57
radiating current, 57 resonance solution, 75–81, 83
radiation condition, 34, 35, 39, 46 resonant mode, 36, 37, 73, 75–77, 79–83
radiation damping, 79 frequency, 50
radiation pattern function, 164, 171, 175, 176 leakage of energy, 76
radio frequency identification (RFID) tag, 107 source free case, energy conservation, 36
random access memory (RAM), 13 resource usage, scaling law, 11, 13, 14
range space, 23, 27–29 Riemann sheets, 174
Rao-Wilton-Glisson (RWG) rooftop function, 54, 112, 119–121
basis, see RWG function root solver, 172, 173
Rao-Wilton-Glisson (RWG) RWG basis, 53, 54, 112, 122, 215
basis function, see RWG basis func- RWG function, 115, 116, 142
tion connection matrix, 119, 127
INDEX 239
full-wave physics, 112 short-wavelength physics, 3, 6, 11

impedance matrix, 118 SIE, 43, 44, 71, 72, 95, 102
loop-tree decomposition, 110, 112, 122 signal integrity, 3
tree space, 116, 117, 121, 122 simplices, 52
VIE, 93 singular function, 135, 140
singular integral, 10, 56
S wave, 197, 203, 206 singular perturbation point at ω = 0, 109
Green’s function, 198, 204 singular value, 61
layered medium, 203 singularity, 175
saddle point, 167 at z = z 0 , 138
SBI, 63 dyadic Green’s function, 199
scaleable multipole engine elastic wave, 199
(ScaleME), see ScaleME enhancement by derivatives, 140
ScaleME, 176 ill-convergent Fourier integral, 137
scaling law, resource usage, 10, 11 in generalized reflection coefficient, 171
scattered field, 48, 49, 56, 57 meaning in Fourier space, 135
in Gedanken experiment, 78, 80, 81 pole at origin, 153
induced magnetic polarization current, 85 singularity subtraction technique
internal resonance problem, 56, 76, 81 convergence improvement, 153
penetrable object, 72 singularity, existence, propagation, 5–7
VIE, 85, 88 SIP, 159, 160, 162, 163, 168, 171
scattering by PEC, 49 deformation of, 171
SDFMA, SDFMM, 160, 175 skew Hermitian electromagnetic operator, 38,
SDP, 160 39
θ plane, 164 small length-scale physics, 128
choice of, 167, 168, 175 smoothing operator, compact operator, 29
dependence on source and field groups, solenoidal current, 5, 6
170 low-frequency, 108, 117
illustration of, 168 solenoidal vector wave function, 136, 137
numerical quadrature, 165, 168, 170, 171 solid-solid interface, 203, 207, 214
other methods using, 175 soliton, optical fiber, 3
singularities encountered, 171 Somigliana’s identity, 214
second-kind integral equation, 56 Sommerfeld half-plane solution, 7, 8
secondary field, 138–141, 143 Sommerfeld identity, 203, 204
self-adjoint operator, 26, 36 Sommerfeld integral, 135, 139, 159, 160, 175
self-box inclusion (SBI), see SBI matrix-friendly form, 136
SH wave, 203–207 tabulation, 141
shear (S) wave, see S wave Sommerfeld integration path (SIP), see SIP
shear horizontal (SH) wave, see SH wave source group, 167, 170, 171, 176
shear vertical (SV) wave, see SV wave image, 166, 167, 171, 175
shear wave velocity, 197 translation to observation group, 171
shocks, existence, 5 source point, 137, 166–168
short length scale phenomenon, 11 at, near interface, 139
short patch pair (SPP), see SPP coplanar with field point, 137
short-wavelength approximation, 8 image, 139, 140
in region m, 138 tabulated value, 141, 152

source-free region, 203 tank model, 176, 177, 180
source-observation group pair, 170, 171 Taylor’s series expansion, 194
space derivative, 4 TDS, 58, 63, 64, 90–93
spanning tree, 121 impedance condition, 58
sparse matrix, 59 semitransparent dielectric sheet, 58
sparse matrix system, 10, 119 structure, 90
dense to O(NC ), 117 TE field, 138, 140, 146, 148, 152
with O(N ) elements, 118 TEz wave, TMz wave, 203, 204
spatial domain, 148 tensor, 193, 194, 196
spectral domain, 145, 146, 148 fourth rank, general form, 196
deriving identity, 141, 146 inverse of, 201
Green’s function, 147, 148 Levi-Civita alternating, 194
spectral integral, 138, 140, 145 testing function, 29, 51, 52, 86, 89, 111, 140,
rapidly convergent, 138, 139 142, 148
Sommerfeld integral, 135 expansion function, symmetric matrix, 28
spherical cavity, elastic wave, 217 range of operator, 27
SPP, 127 tetrahedral edge element, 87, 89
stability criterion, elastic wave, 211 tetrahedron, 53
star basis, 110 thin dielectric sheet (TDS), see TDS
state vector, 25, 26 thin sheet, 55
static field, unphysical field, 137 thin wire, 101
static physics, 3, 5, 11, see also circuit physics thin-stratified medium fast-multipole algorithm
steepest descent fast multipole algorithm (TSM-FMA), see TSM-FMA
(SDFMA), see SDFMA time-domain energy conservation theorem, 36
steepest descent fast multipole method time-domain method, 8–11, 219
(SDFMM), see SDFMM TM field, 138, 140, 152
steepest descent path (SDP), see SDP topological equivalence, 55
Stokes’ displacement and traction tensor, 214 traction, 195, 214, 215, 217
strain, stress tensor, 193, 195, 196 traction-free cavity, 214
subatomic length scale, 2 transistors, diodes, nonlinear devices, 107
subdomain expansion, 29 translation matrix, 167, 175, 176
subspace projection method, 23, 24, 28, 95 translation operator, 165, 166, 170, 172
subsurface reflection, 209 transmission-line model, 58
surface charge, 114, 117, 120, 121 transpose free QMR (TFQMR), see iterative
surface current, 89, 99, 127 method
surface impedance, 58, 96, 98 transpose operator, 83
surface integral equation (SIE), see SIE transpose operator, definition, 26
SV wave, 203–207 transverse wave, 136, 147, 197, 199, 203
symmetric operator, 26, 28, 33–35, 52, 82 tree basis, 110, 116–119, 122, 124
matrix representation, 24 tree current, 117, 126
non-Hermitian, 39 tree-cotree decomposition, 110
range and domain, 28 triangular patch, 52, 54, 176, 177, 181, 184
reciprocity, 33 triangulated geometry, 65
symmetric tensor, 33, 34 triangulated surface, 112, 113, 122
INDEX 241
TSM-FMA, 160, 175

twilight zone, singular perturbation point, 109
vector wave equation, 43, 45, 84

vector wave function, 136, 137
VIE, 64, 71, 72, 84, 102
TDS, new TDS, 90, 91, 93
volumetric scatterer, 86
virtual experiment, prototyping, 8
volume integral equation (VIE), see VIE
VV (vertical-vertical) response, 180, 187
Watson’s transformation, 8
wave equation, 9, 57
wave physics, 5, 109, 130, 193, 217
wavelet, 5, 6
weight function
FIPWA, 164, 167, 172, 174, 175
inhomogeneous plane waves, 163
weighted-sum, PMCHWT, Müller method, 74,
84
weighting function, 27, 215
Weyl identity, 163, 203
also, see Sommerfeld identity
world of inductors, capacitors, 108, 109, 112
Wronskian, Bessel function, 100
Yee scheme, 11
Young’s modulus, 217
zero thickness surface, 58

Chew, 2009

Uploaded by

Copyright:

Available Formats

Chew, 2009

Uploaded by

Document Information

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

Chew, 2009

Uploaded by

Copyright:

Available Formats

What are integral equation methods?

What are integral equation methods?

What are some applications of integral equation methods?

What are some applications of integral equation methods?

MC1483_Chew_Lecture.

qxd 8/18/08 12:29 PM Page i

Integral Equation Methods for

Copyright © 2009 by Morgan & Claypool

Integral Equation Methods for Electromagnetic and Elastic Waves

ISBN-10: 1598291483 paperback

ISBN-10: 1598291491 ebook

A Publication in the Morgan & Claypool Publishers series

Series Editor: Constantine A. Balanis, Arizona State University

Series ISSN: 1932-1252 print

Printed in the United States of America

Integral Equation Methods for

Weng Cho Chew

Mei Song Tong

SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #12

We dedicate this book

1 Introduction to Computational Electromagnetics 1

2 Linear Vector Space, Reciprocity, and Energy Conservation 21

viii INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES

2.10.1 Lorentz Reciprocity Theorem . . . . . . . . . . . . . . . . . . . . . . . 35

3 Introduction to Integral Equations 43

4 Integral Equations for Penetrable Objects 71

5 Low-Frequency Problems in Integral Equations 107

7 Fast Inhomogeneous Plane Wave Algorithm for Layered Media 159

8 Electromagnetic Wave versus Elastic Wave 193

x INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES

8.5 Boundary Conditions for Elastic Wave Equation . . . . . . . . . . . . . . . . 202

Glossary of Acronyms 223

About the Authors 227

xii INTEGRAL EQUATIONS FOR ELECTROMAGNETIC AND ELASTIC WAVES

Weng Cho CHEW

1.1 Mathematical Modeling—A Historical Perspective

2 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES

The electromagnetic equations are:

INTRODUCTION TO COMPUTATIONAL ELECTROMAGNETICS 3

1.2 Some Things Do Not Happen in CEM Frequently—

1.3 The Morphing of Electromagnetic Physics

4 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES

1.3.1 Comparison of Derivatives

INTRODUCTION TO COMPUTATIONAL ELECTROMAGNETICS 5

1.3.2 Low-Frequency Regime

the Fraunhoffer zone [34].

6 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES

1.3.3 Mid Frequency Regime

1.3.4 High-Frequency Regime

INTRODUCTION TO COMPUTATIONAL ELECTROMAGNETICS 7

1.3.5 Quantum Regime

1.4 Matched Asymptotics

1.5 Why CEM?

8 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES

1.6 Time Domain versus Frequency Domain

INTRODUCTION TO COMPUTATIONAL ELECTROMAGNETICS 9

1.7 Differential Equation versus Integral Equation

∇2 φ(r, ω) + k 2 φ(r, ω) = s(r, ω), k = ω/v (1.19)

∇2 g(r, r0 ) + k 2 g(r, r0 ) = −δ(r − r0 ) (1.20)

10 INTEGRAL EQUATION METHODS FOR ELECTROMAGNETIC AND ELASTIC WAVES