Title

Surface integral equation method for analyzing electromagnetic

scattering in layered medium

Advisor(s)

Chew, WC; Jiang, L

Author(s)

Chen, Yongpin.; .

Citation

Issued Date

URL

Rights

2011

http://hdl.handle.net/10722/174463

The author retains all proprietary rights, (such as patent rights)

and the right to use in future works.

Surface Integral Equation Method for Analyzing

Electromagnetic Scattering in Layered Medium

by

Yongpin CHEN ()

A thesis submitted in partial fulfillment of the requirements for the degree of

Doctor of Philosophy at The University of Hong Kong

October 2011

Abstract of thesis entitled

Surface Integral Equation Method for Analyzing

Electromagnetic Scattering in Layered Medium
Submitted by

Yongpin CHEN
for the degree of Doctor of Philosophy
at The University of Hong Kong
in October 2011
Surface integral equation (SIE) method with the kernel of layered medium Green's
function (LMGF) is investigated in details from several fundamental aspects. A
novel implementation of discrete complex image method (DCIM) is developed to
accelerate the evaluation of Sommerfeld integrals and especially improve the far
field accuracy of the conventional one. To achieve a broadband simulation of thin
layered structure such as microstrip antennas, the mixed-form thin-stratified
medium fast-multipole algorithm (MF-TSM-FMA) is developed by applying
contour deformation and combining the multipole expansion and plane wave
expansion into a single multilevel tree. The low frequency breakdown of the
integral operator is further studied and remedied by using the loop-tree
decomposition and the augmented electric field integral equation (A-EFIE), both
in the context of layered medium integration kernel. All these methods are based
on the EFIE for the perfect electric conductor (PEC) and hence can be applied in
antenna and circuit applications. To model general dielectric or magnetic objects,
the layered medium Green's function based on pilot vector potential approach is
generalized for both electric and magnetic current sources. The matrix
representation is further derived and the corresponding general SIE is setup.
Finally, this SIE is accelerated with the DCIM and applied in quantum optics,
such as the calculation of spontaneous emission enhancement of a quantum
emitter embedded in a layered structure and in the presence of nano scatterers.

Declaration

I declare that the thesis and the research work thereof represents my own work,
except where due acknowledgement is made, and that it has not been previously
included in a thesis, dissertation or report submitted to this University or to any
other institution for a degree, diploma or other qualifications.

Signed
Yongpin CHEN

c 2011 Yongpin CHEN

SURFACE INTEGRAL EQUATION METHOD FOR ANALYZING

ELECTROMAGNETIC SCATTERING IN LAYERED MEDIUM

BY
YONGPIN CHEN

DISSERTATION
Submitted in partial fulfillment of the requirements
for the degree of Doctor of Philosophy in Electrical and Electronic Engineering
in the Graduate School of the
University of Hong Kong, 2011

Hong Kong

ABSTRACT

Surface integral equation (SIE) method with the kernel of layered medium
Greens function (LMGF) is investigated in details from several fundamental
aspects. A novel implementation of discrete complex image method (DCIM)
is developed to accelerate the evaluation of Sommerfeld integrals and especially improve the far field accuracy of the conventional one. To achieve a
broadband simulation of thin layered structure such as microstrip antennas,
the mixed-form thin-stratified medium fast-multipole algorithm (MF-TSMFMA) is developed by applying contour deformation and combining the multipole expansion and plane wave expansion into a single multilevel tree. The
low frequency breakdown of the integral operator is further studied and remedied by using the loop-tree decomposition and the augmented electric field
integral equation (A-EFIE), both in the context of layered medium integration kernel. All these methods are based on the EFIE for the perfect electric
conductor (PEC) and hence can be applied in antenna and circuit applications. To model general dielectric or magnetic objects, the layered medium
Greens function based on pilot vector potential approach is generalized for
both electric and magnetic current sources. The matrix representation is
further derived and the corresponding general SIE is setup. Finally, this
SIE is accelerated with the DCIM and applied in quantum optics, such as
the calculation of spontaneous emission enhancement of a quantum emitter
embedded in a layered structure and in the presence of nano scatterers.

ii

To my wife and my parents, for their love and support.

iii

ACKNOWLEDGMENTS

First and foremost, I would like to thank my mentor and thesis adviser,
Professor Weng Cho CHEW, for providing me this opportunity to study
with him. His patient and careful guidance during the last four years has
totally changed my way of thinking, re-shaped my knowledge structure, and
finally built my self-confidence. His enthusiasm, persistence, and diligent
work ethic has inspired me much and will continue to affect my career and
life.
I would also like to thank my co-adviser, Dr. Lijun JIANG, who always
encouraged me when I was frustrated, shared with me his working experience
in industry, and helped me a lot in technical details.
I feel grateful to my former supervisors Professor Zaiping NIE and Professor Jun HU at UESTC, who opened me the door of the fascinating CEM
world, encouraged and supported me to pursue study at HKU, and always
pay kind attention to my recent status.
Many thanks to my friends and colleagues with the Electromagnetics and
Optics Laboratory, to Dr. Wallace C. H. CHOY for providing me TA opportunities and inviting me to audit his course on organic devices, to Dr.
Sheng SUN for giving me suggestions on possible directions about circuit
simulation, to Dr. Wei SHA for sharing his knowledge on modern physics,
to Dr. Yang LIU for teaching me lots of mathematics, to Dr. Yat Hei LO for
sharing his expertise on computer and optics, to Dr. Shaoying HUANG, Dr.
Min TANG, Dr. Osman GONI, Dr. Bo ZHU, Phillip ATKINS, Peng YANG,
Qi DAI, Jun HUANG, Zuhui MA, Shiquan HE, Yumao WU, Yan LI, Nick
HUANG, Ping LI for their help and friendship.
I also wish to thank alumni from both UIUC and UESTC. A special thanks
to Dr. Yuan LIU for helping me a lot when I first came to HKU, though we
have never met with each other, to Dr. Liming XU for sharing his code and
teaching me the transmission line method, to Dr. Jie XIONG for sharing her
iv

code and polishing my understanding on many important concepts, to Dr.

Su YAN for working together to develop the MLFMA code, to Dr. Zhiguo
QIAN for his help on A-EFIE.
I thank my parents and my little sister for their love and support. Though
I cannot go back to my hometown every year, their encouragement through
the telephone line always warms my heart.
Finally, I thank my wife, Min MENG, for her love, her sharing of joy
and sadness, and her shouldering of so much responsibility alone when I was
pursuing my study.

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

LIST OF ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . xiii

CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . .
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Organization of the Thesis . . . . . . . . . . . . . . . . . . . .
CHAPTER 2 A NOVEL IMPLEMENTATION OF
COMPLEX IMAGE METHOD (DCIM) . . . . .
2.1 Introduction . . . . . . . . . . . . . . . . . .
2.2 Layered Medium Greens Function (LMGF)
2.3 Discrete Complex Image Method (DCIM) .
2.4 Numerical Results . . . . . . . . . . . . . . .
2.5 Summary . . . . . . . . . . . . . . . . . . .

DISCRETE
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .

.
.
.
.
.
.

1
1
3

. 4
. 4
. 6
. 9
. 14
. 15

CHAPTER 3 MIXED-FORM THIN-STRATIFIED MEDIUM FASTMULTIPOLE ALGORITHM (MF-TSM-FMA) . . . . . . . . . . .

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Review of Thin-Stratified Medium Fast-Multipole Algorithm (TSM-FMA) . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Mixed-Form Thin-Stratified Medium Fast-Multipole Algorithm (MF-TSM-FMA) . . . . . . . . . . . . . . . . . . . . . .
3.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

24
30
36

CHAPTER 4 REMEDIES FOR LOW FREQUENCY BREAKDOWN OF INTEGRAL OPERATOR IN LAYERED MEDIUM
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Loop-Tree Decomposition and Frequency Normalization . .
4.3 Basis Rearrangement via Connection Matrix . . . . . . . .
4.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . .
4.5 Augmented Electric Field Integral Equation (A-EFIE) . .

37
37
39
42
45
45

vi

.
.
.
.
.
.

.
.
.
.
.
.

19
19
21

4.6
4.7
4.8
4.9

Charge Neutrality Issue

Electrostatic Limit . . .
Numerical Results . . . .
Summary . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

51
56
59
60

CHAPTER 5 A NEW LAYERED MEDIUM GREENS FUNCTION (LMGF) FORMULATION FOR GENERAL OBJECTS .
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Dyadic Form of LMGF . . . . . . . . . . . . . . . . . . . .
5.3 Matrix Representation of LMGF . . . . . . . . . . . . . . .
5.4 Duality Principle of LMGF . . . . . . . . . . . . . . . . . .
5.5 Surface Integral Equation . . . . . . . . . . . . . . . . . .
5.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . .
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

63
63
65
73
82
84
85
89

.
.
.
.

.
.
.
.

90
90
91
95

CHAPTER 6 DCIM ACCELERATED SURFACE INTEGRAL

EQUATION (SIE) METHOD FOR NANO-OPTICAL APPLICATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 SIE with DCIM Acceleration . . . . . . . . . . . . . . . . .
6.3 Greens Function Approach in Spontaneous Emission (SE)
6.4 Surface Plasmon Polaritons (SPPs) and Localized Surface
Plasmons (LSP) . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Quantum Emitter Coupled to Surface Plasmon Resonance
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . 98
. . 101
. . 105

CHAPTER 7 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . 106

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
LIST OF PUBLICATIONS . . . . . . . . . . . . . . . . . . . . . . . . 119
AUTHORS BIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . 122

vii

LIST OF TABLES

2.1 CPU time comparison (seconds) . . . . . . . . . . . . . . . . . 15

3.1 Layout of radiation and receiving patterns . . . . . . . . . . . 27
5.1 Matrix element of KE : no line integral (103 ) . . . . . . . . . 79
5.2 Matrix element of KE : testing line integral (103 ) . . . . . . . 79
5.3 Matrix element of KE : testing & basis line integrals (104 ) . . 81
6.1 CPU time for single frequency calculation . . . . . . . . . . . 102

viii

LIST OF FIGURES

2.1 A general layered medium. . . . . . . . . . . . . . . . . . .

2.2 A typical microstrip structure (unit: mm). . . . . . . . . .
2.3 One branch-cut case: Sommerfeld integration path (SIP),
Sommerfeld branch cut (SBC) and poles in the complex k
plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Two branch-cut case: Sommerfeld integration path (SIP),
Sommerfeld branch cuts (SBC) and poles in the complex
k plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 The magnitude of gss versus k0 for the microstrip structure. The asymptotic behavior is 1/2 , which agrees with
the theoretical prediction. . . . . . . . . . . . . . . . . . .
2.6 The magnitude of g versus k0 for the microstrip structure. The far field is dominated by the pole contribution,

which has the asymptotic behavior of 1/ . . . . . . . . .

2.7 A 3-layer model (unit: mm). . . . . . . . . . . . . . . . . .
2.8 The magnitude of gss versus k0 for the 3-layer model. The
asymptotic behavior is 1/2 . . . . . . . . . . . . . . . . . .
2.9 The magnitude of g versus k0 for the 3-layer model. The
asymptotic behavior is 1/2 . . . . . . . . . . . . . . . . . .
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
3.10
3.11

. . 6
. . 11

. . 11

. . 13

. . 15

. . 16
. . 16
. . 17
. . 17

The worst interaction in applying the addition theorem. . . . .

Relative error by using plane wave expansion. . . . . . . . . .
Relative error by using multipole expansion. . . . . . . . . . .
The MF-TSM-FMA structure. . . . . . . . . . . . . . . . . . .
The geometrical structure of the 8 4 microstrip array. . . . .
The radiation patterns of the microstrip array in E plane
( = 0 ) and H plane ( = 90 ) at 9.42 GHz. . . . . . . . . . .
The bistatic RCS of the microstrip array at f = 2.5 GHz
and (i , i ) = (60 , 0 ). . . . . . . . . . . . . . . . . . . . . . .
A dipole on an EBG structure, top and side view. . . . . . . .
Radiation pattern of the dipole in E plane, due to the layered medium assumption, the ground plane is infinitely large. .
The geometrical model of a 30 30 microstrip array. . . . . .
Bistatic RCS of the 30 30 microstrip array. . . . . . . . . . .
ix

23
23
28
30
32
32
33
33
34
34
35

3.12 Memory requirement (solid line) and CPU time per iteration (dash line) versus number of unknowns. . . . . . . . . . 35
4.1
4.2
4.3
4.4
4.5

4.6

4.7

4.8

4.9

4.10

4.11

4.12

Local loop and global loop (dashed lines). . . . . . . . . . . .

RCS of a sphere. . . . . . . . . . . . . . . . . . . . . . . . . .
Geometrical model of the capacitor. . . . . . . . . . . . . . .
Negative input reactance of the capacitor, r2 = 6.0. . . . . . .
The geometrical structure of the loop inductor embedded
in a seven-layer medium, unit: mm. The central layer is
a magnetic material. A delta gap excitation is applied at
the center of the top arm. . . . . . . . . . . . . . . . . . . . .
The condition number versus frequency for the rectangular
loop. The condition number is unbounded when decreasing the frequency. Charge neutrality enforcement (CNE)
makes the condition number constant. . . . . . . . . . . . . .
The eigenvalue distribution for the rectangular loop at 1
Hz. The smallest eigenvalue is removed away from the
origin after the charge neutrality enforcement (CNE). . . . . .
The geometrical model of the half loop embedded in a fivelayer medium (including the PEC layer), unit: mm. A
delta gap excitation is applied at the center of the top arm. . .
The condition number versus frequency for the half loop.
Since it is connected to the ground plane, charge neutrality
cannot be guaranteed. The condition number is bounded
when decreasing the frequency without any special treatment.
The geometrical model of the circular parallel plate capacitor, with a dielectric layer (r = 2.65) inserted in between.
A delta gap is applied at the edge. The mesh is refined to
capture the fringing effect. . . . . . . . . . . . . . . . . . . . .
The input reactance of the rectangular loop. A-EFIE agrees
well with the loop-tree (LT) decomposition, while the traditional EFIE breaks down quickly when decreasing the
frequency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The input reactance of the half loop. A-EFIE maintains
the scale invariance very well while the traditional EFIE
breaks down quickly when decreasing the frequency. Since
the non-magnetic dielectric is transparent to the inductor,
a PEC half space model is applied to validate the results. . . .

41
43
44
44

53

54

54

55

55

57

60

61

4.13 The capacitance of the circular parallel plate capacitor. AEFIE I represents the capacitance extracted from current,
while A-EFIE Q means the capacitance extracted from
charge. The A-EFIE current suffers from an inaccuracy
problem while the A-EFIE charge is stable. The result
agrees with the static solver. Both are further validated
by the analytic solution. When the frequency is below 1
MHz, the relative error of A-EFIE Q is around 0.1%. . . . . . 61
5.1 A homogeneous dielectric object is embedded in a layered
medium. The external excitation is either a plane wave or
a Hertzian dipole. Equivalent electric and magnetic currents are induced on the boundary which then generate the
scattered field. . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 An electric or magnetic dipole is radiating in a seven-layer
medium (unit: m). The layered medium is both dielectric
and magnetic, and the layer constants are shown in the
figure. The source point is in Layer 2 and the observation
line is in Layer 5. . . . . . . . . . . . . . . . . . . . . . . .
5.3 The magnitude of electric field generated by an electric
dipole. The dipole polarization is ( = 20o , = 30o ) and
is working at f = 300 MHz. The result is validated by the
transmission line method (solid line). . . . . . . . . . . . .
5.4 The magnitude of electric field generated by a magnetic
dipole. The dipole polarization is ( = 20o , = 30o ) and
is working at f = 300 MHz. The result is validated by the
transmission line method (solid line). . . . . . . . . . . . .
5.5 Cases where testing line integral exists. The testing function is straddling the interface, all radiation from the RWG
or half-RWG basis functions (triangles) in color needs invoking testing line integral, while radiation from others
does not need this line integral. . . . . . . . . . . . . . . .
5.6 Line integral test. The testing function is at the top interface. Radiation from: basis function 1: no line integral
activated; basis function 2: testing line integral activated;
basis function 3: both testing and basis line integrals activated. (unit: m). . . . . . . . . . . . . . . . . . . . . . . .
5.7 A homogeneous capsule structure embedded in a 5-layer
medium, where h = 0.6, r = 0.15 (unit: m). . . . . . . . .
5.8 The scattered field inside the object with = 2 and = 2.
Since there is no contrast, the scattered field recovers the
incident field (solid line). . . . . . . . . . . . . . . . . . . .

xi

. . 65

. . 73

. . 74

. . 74

. . 80

. . 80
. . 86

. . 86

5.9 The scattered field outside the object with = 4 and = 1.

The tangential components of the E field are continuous at

the interface. Symbol: field at d+

5 , solid line: field at d5 . .
5.10 The scattered field outside the object with = 4 and = 1.
The normal component of the E field differs by a factor of
3 (4 /5 ) at the interface. . . . . . . . . . . . . . . . . . . .
5.11 A homogeneous cuboid embedded in a 4-layer medium
(unit: m). It is penetrating different layers. . . . . . . . . .
5.12 The scattered field of the PEC object, calculated from
EFIE and MFIE. . . . . . . . . . . . . . . . . . . . . . . .
5.13 The scattered field of the dielectric object. . . . . . . . . .
6.1 Transition of a two-level atomic system. . . . . . . . . . .
6.2 SPPs can be sustained in the interface of dielectric-metal
half space. . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 LSP can be excited by a metallic nano sphere under an EM
field excitation. . . . . . . . . . . . . . . . . . . . . . . . .
6.4 A gold nano sphere with radius 20 nm is located above a
gold slab with thickness 30 nm. A z-polarized emitter is
located at the middle of the sphere and slab. . . . . . . . .
6.5 Normalized SER for the three cases: (a) only slab (SPPs
enhanced SE); (b) only nano sphere (LSP enhanced SE);
(c) both (both effects). . . . . . . . . . . . . . . . . . . . .
6.6 A gold nano bowl is located in a layered medium with two
gold layers. A z-polarized emitter is located at the center
of the aperture. The dimensions are shown in the figure. .
6.7 Mesh of the gold nano bowl. . . . . . . . . . . . . . . . . .
6.8 Normalized SER. The first peak corresponds to the peak
from SPPs of the layered medium at 520 nm, the second
peak is from the nano bowl effect around 590 nm. . . . . .

xii

. . 87

. . 87
. . 88
. . 88
. . 89
. . 96
. . 99
. . 100

. . 103

. . 103

. . 104
. . 104

. . 105

LIST OF ABBREVIATIONS

SIE

Surface Integral Equation

LMGF

Layered Medium Greens Function

PEC

Perfect Electric Conductor

CEM

Computational Electromagnetics

FDM

Finite Difference Method

FEM

Finite Element Method

MoM

Method of Moments

GPR

Ground-Penetrating Radar

LED

Light-Emitting Diode

GMRES

Generalized Minimal Residual

DCIM

Discrete Complex Image Method

PML

Perfectly Matched Layers

RFFM

Rational Function Fitting Method

SBC

Sommerfeld Branch Cut

SIP

Sommerfeld Integration Path

TE

Transverse Electric

TM

Transverse Magnetic

MF-TSM-FMA Mixed-Form Thin-Stratified Medium Fast-Multipole Algorithm

FMM

Fast Multipole Method

MLFMA

Multilevel Fast Multipole Algorithm

xiii

AIM

Adaptive Integral Method

CG-FFT

Conjugate Gradient Fast Fourier Transform

MLMDA

Multilevel Matrix Decomposition Algorithm

ACA

Adaptive Cross Approximation

RCS

Radar Cross Section

EBG

Electromagnetic Band Gap

CPU

Central Processing Unit

PC

Personal Computer

A-EFIE

Augmented Electric Field Integral Equation

CAD

Computer-Aided Design

RWG

Rao-Wilton-Glisson

CCIE

Current and Charge Integral Equation

SPIE

Separated Potential Integral Equation

MPIE

Mixed Potential Integral Equation

CNE

Charge Neutrality Enforcement

LT

Loop-Tree

2D

Two-Dimensional

3D

Three-Dimensional

VIE

Volume Integral Equation

BOR

Body of Revolution

TL

Transmission Line

PMCHWT

Poggio-Miller-Chang-Harrington-Wu-Tsai

EFIE

Electric Field Integral Equation

MFIE

Magnetic Field Integral Equation

SE

Spontaneous Emission

SPPs

Surface Plasmon Polaritons

LSP

Localized Surface Plasmons

xiv

LDOS

Local Density of States

EM

Electromagnetics

QED

Quantum Electrodynamics

PC

Photonic Crystal

SER

Spontaneous Emission Rate

DOS

Density of States

OLED

Organic Light-Emitting Diode

MEMS

Microelectromechanical Systems

GIBC

Generalized Impedance Boundary Condition

xv

CHAPTER 1
INTRODUCTION

1.1 Motivation
With the development of computer technology and scientific computation,
numerical simulation becomes as important as theoretical study and experimental verification in understanding the nature of our world. Computational
electromagnetics (CEM) is one branch of numerical techniques to model all
electromagnetic phenomena governed by the Maxwells equations, covering
from circuit engineering, microwave and antenna engineering, to optical engineering.
The methods in CEM can be classified into different categories according to
different criteria, such as different computation domains (time or frequency),
or different equation forms (differential or integral) [1]. The time domain
methods solve the time domain EM fields directly, while the frequency domain methods solve the harmonic fields based on Fourier transform. On the
other hand, the differential equation methods solve the Maxwells equations
in differential form, such as finite difference method (FDM) [2], finite element method (FEM) [3], while the integral equation methods first extract
the point response (Greens function) of a differential equation, and solve the
integral equation formulated from equivalence principle (surface or volume),
such as the method of moments (MoM) [4].
Integral equation method in the frequency domain is one of the most powerful methods in analyzing electromagnetic radiation and scattering. The key
of this method is the Greens function, which describes the EM response of
a point source in a specific environment. It satisfies the boundary condition
automatically and hence does not require any domain truncation or absorption boundary condition in the computation. Also, if the surface equivalence
principle is applied, the surface integral equation (SIE) can be obtained,
1

where the unknowns are only associated with the boundary of the scatterers
and thus can be made minimum.
Although the SIE can always be derived mathematically, the Greens functions for arbitrary inhomogeneous medium is not trivial and can only be
obtained numerically, which may be as complex as solving the original problem. For this reason, most research on surface integral equation methods are
focused on homogeneous environment, such as aerospace applications. However, if the inhomogeneity is one dimensional piecewise, or so called layered
medium, the Greens function can be determined analytically in the spectral (Fourier) domain, and the spatial domain counterpart can be obtained
by simply inverse Fourier transforming it [5]. In fact, this scenario consists of a broad class of applications in both microwave and optical regimes,
such as microstrip antennas and microwave circuits, geophysical exploration,
ground-penetrating radar (GPR), solar cell, light-emitting diode (LED), and
lithography, etc.
This thesis is focused on the SIE method with layered medium Greens
function (LMGF), and addresses several fundamental problems of this method.
First, the spatial domain LMGF can only be obtained by numerically integrating the spectral domain counterpart, expressed as a Sommerfeld integral [6]. This integral is oscillatory and slowly convergent, which makes
the repeated evaluation in MoM extremely time expensive. An acceleration
technique at the Greens function level is necessary.
Second, the SIE leads to a full matrix, which requires an O(N 2 ) memory
and an O(N 3 ) or O(N 2 ) computational time if a direct inverse such as LU
decomposition or an iterative method such as the GMRES is applied [7] [8]. A
fast algorithm is indispensable for large scale computation, such as radiation
of an antenna array. The fast algorithm developed shall also cover a broad
frequency range to achieve a broadband simulation.
Third, when the structure under investigation is much smaller then the
wavelengthwhich is common in circuit simulationthe integral operator
may suffer from a low frequency breakdown. Remedies shall be proposed to
overcome this difficulty for the specific integration kernel of LMGF.
Fourth, though the commonly used metal is a good conductor at the radio
or microwave frequencies and the electric field integral equation (EFIE) [9]
for a perfect electric conductor (PEC) can be implemented in most applications, the ability of dielectric modeling is important in other situations such
2

as underground exploration or optical applications. For example, metal is

penetrable to the order of wavelength at the optical frequency band, and
hence becomes dielectric. A full sets of LMGFs shall be developed to obtain
a general SIE.
Finally, the power of CEM should not be restricted to classical electrodynamics [10], but shall also be highlighted in quantum electrodynamics such
as quantum optics [11]. Such attempt will be made by applying the tools
developed in this thesis.

1.2 Organization of the Thesis

The organization of the thesis is as follows: Chapter 2 first reviews one popular acceleration technique, the discrete complex image method (DCIM), for
fast evaluation of LMGF. The cause of far field inaccuracy of the conventional DCIM is explained and a novel implementation based on Sommerfeld
branch cut is proposed. In Chapter 3, a fast algorithm suitable for microstrip structuresthe quasi-3D thin-stratified medium fast-multipole algorithm (TSM-FMA)is first introduced, followed by the discussion of low frequency breakdown of it. A mixed-form TSM-FMA (MF-TSM-FMA) combing the multipole expansion and plane wave expansion is proposed to achieve
a fast and broadband simulation. Next, in Chapter 4, the low frequency
breakdown of the integral operator is further investigated. The loop-tree
decomposition based on quasi-Helmholtz decomposition and the augmented
electric field integral equation (A-EFIE) based on the augmentation technique in the context of layered medium are both studied. After that, in
Chapter 5, a new LMGF formulation is developed to obtain a general SIE
for dielectric modeling. The matrix representation is derived to reduce the
singularity of LMGF and the line integral issue associated with it is discussed in details. On top of the SIE solver, in Chapter 6, the DCIM is
incorporated and is applied to study the matter-field interaction in a complex environmentspontaneous emission enhancement of a quantum emitter
embedded in a layered medium and in the presence of nano scatterers. Finally, in Chapter 7, the conclusion is made and several possible research
topics are discussed.

CHAPTER 2
A NOVEL IMPLEMENTATION OF
DISCRETE COMPLEX IMAGE METHOD
(DCIM)
A novel implementation of discrete complex image method (DCIM) based
on the Sommerfeld branch cut is proposed to accurately capture the far-field
behavior of the layered medium Greens function (LMGF) as a complement
to the traditional DCIM. By contour deformation, the Greens function can
be naturally decomposed into branch-cut integration (radiation modes) and
pole contributions (guided modes). For branch-cut integration, matrix pencil
method is applied, and the alternative Sommerfeld identity in terms of kz
integration is utilized to get a closed-form solution. The guided modes are
accounted for with a pole-searching algorithm. Both one-branch-cut and
two-branch-cut cases are studied. Several numerical results are presented to
validate this method. For the sake of completeness, the LMGF formulation
will also be briefly summarized in this chapter.

2.1 Introduction
Sommerfeld first investigated a dipole radiating above a half space using
Hertzian potential [6]. This problem was further studied and extended to
general multilayered medium by various researchers [12][15], [5]. There are
several approaches to derive the layered medium Greens function (LMGF),
such as the popular transmission line analog [16], [17], Hertzian potential
approach [6], [18], Ez Hz formulation [19], [20], and pilot vector potential
approach [5], [21], etc. No matter which method is applied, however, the
LMGF can only be expressed as an infinite, oscillatory and slowly-convergent
integral, Sommerfeld integral, making the numerical evaluation process very
inefficient. Several methods have been developed to expedite the evaluation of the LMGF, such as the function approximation approach [22], [23],
the path deformation technique [5], [24], the tabulation and interpolation
4

method [25], [26], the perfectly matched layers (PML) method [27], [28], the
singularity subtraction method [29], and the fast all-modes combined with
numerical modified steepest descent path method [30], [31] etc.
Function approximation in the spectral domain is one of the most popular
methods. In this method, the integration kernel is first approximated by
certain simple functions, and the integral is then evaluated in a closed
form by applying relevant integration identities. Though lots of function
approximation techniques are available from a numerical analysis point of
view, those candidates with closed-form identities of the infinite integrals
in our context can finally be utilized. This leads to the following methods:
the complex discrete complex image method (DCIM) [22], [32] (based on the
complex exponential functions), the rational function fitting method (RFFM)
[23], [33] (based on the rational fraction functions), or their combination [34].
The popular DCIM has unpredictable errors when the interaction is in
the far field region ( 0). The original sampling path cannot effectively

capture certain singularities, which correspond to the guided modes or surface

waves and lateral waves physically. Several efforts have been made to remedy
this problem. A two-level approximation [35] was proposed to separate the
sharp-transition region from the smooth-varying region, with higher sampling
rate in the former part to capture the singularities. In [36], the surfacewave poles are extracted explicitly for a general multilayer medium before
applying the DCIM, which is also suggested in [22]. Other attempts are made
to deform the sampling path to carry more pole singularities. For instance,
in [37], a direct DCIM was developed to push the sampling path closer to
the poles and rely on the matrix pencil method [38] itself to take care of the
singularities. Meanwhile, spatial error control is another big issue in DCIM
and a recent progress can be found in [39].
Though the pole singularities can be captured successfully by the above
methods, the branch-point singularities are still not considered. These singularities contribute to the lateral wave when the source and observation points
are at the interface of the layers [5]. Recently, a three-level DCIM [40] was
proposed to bring the sampling path closer to the branch point to capture
more information of this singularity. In the RFFM, the continuous spectrum
in the far field is also considered based on a vertical path and asymptotic
analysis [41]. In this chapter, we propose an alternative implementation
of the DCIM to capture these singularities. Other than introducing extra
5

Figure 2.1: A general layered medium.

segments of the sampling path to approach the singularities, we simply deform the sampling path to the Sommerfeld branch cut (SBC) [5] when is
relatively large. The matrix pencil method is applied to approximate the
function along the SBC, which can be mapped into the real axis in the kz
plane [42]. The pole contributions are accounted for by applying a robust
pole-searching algorithm [43].
For the sake of completeness, the basic formulation of the LMGF such
as the propagation factor and the matrix-friendly formulation will be briefly
reviewed. It can also be found in Chapter 3, Chapter 4 and will be generalized
in Chapter 5.

2.2 Layered Medium Greens Function (LMGF)

The LMGF is the impulse response of a dipole in this medium. The configuration is shown in Figure 2.1, where each layer is characterized by the
permittivity and permeability (losses are included). The dipole is assumed to be at r in layer m and the observation point is at r in layer n.
The LMGF (in dyadic form) can be expressed as [21],
r) = ( z)( z)g TE (r, r )
G(r,
+

1
2
knm

( z)( z)g

(2.1)
TM

(r, r )

2
where knm
= 2n m . The g TE/TM (r, r ) is expressed as a Sommerfeld inte-

gral,

dk
(1)
H0 (k )F (k, z, z )
(2.2)
kmz k
p 2
(1)
where F (k, z, z ) is the propagation factor [5], kmz = km
k2 , and H0 (k )
is the first kind Hankel function of order 0 (the time convention is assumed
to be eit ).
i
g(r, r ) =
8

2.2.1 Propagation Factor

The expression of propagation factor depends on the relative positions of the
source point and the observation point. In the following expressions, the
argument k is neglected for simplicity.
A. n = m
The source point and the observation point are in the same layer. There
are primary field as well as secondary field. Then

F (z, z ) = eikmz |zz |

mR
m,m1 eikmz (2dm +z +z)
+ M
mR
m,m+1 eikmz (2dm+1 z z)
+ M

(2.3)

mR
m,m+1 R
m,m1 eikmz (2dm+1 2dm +z z)
+ M
mR
m,m1 R
m,m+1 eikmz (2dm+1 2dm z +z)
+ M
where the generalized reflection coefficients and Fresnel reflection coefficients
[5] are
iki+1,z (2di+2 2di+1 )

i,i+1 = Ri,i+1 + Ri+1,i+2 e

R
(2.4)
i+1,i+2 eiki+1,z (2di+2 2di+1 )
1 + Ri,i+1 R
iki1,z (2di 2di1 )

i,i1 = Ri,i1 + Ri1,i2 e

R
i1,i2 eiki1,z (2di 2di1 )
1 + Ri,i1 R

Rij =
p=

(2.5)

pj kiz pi kjz
pj kiz + pi kjz

(2.6)

,
,

(2.7)

TE
TM

and

B. n > m

h
i1
m = 1 R
m,m1 R
m,m+1 e2ikmz (dm+1 dm )
M

(2.8)

The observation layer is above the source layer. There is only secondary
field. Then
h
i

ikmz (dm+1 z )
ikmz (dm+1 2dm +z )

F (z, z ) = e
+ Rm,m1 e
(2.9)
h
i
iknz (zdn )
iknz (2dn+1 dn z)

e
+ Rn,n+1 e
Mm Tmn
where

n1
Y

Tmn =

j=m+1

 ik (d d )

e jz j+1 j Sj1,j Sn1,n

Tj1,j
.

1 Rj,j1Rj,j+1eikjz (2dj+1 2dj )

Sj1,j =

(2.10)

(2.11)

C. n < m
The observation layer is below the source layer. There is also only secondary field. Then
h
i

m,m+1 eikmz (2dm+1 dm z )

eikmz (z dm ) + R
h
i
iknz (dn+1 z)
iknz (z+dn+1 2dn )

m Tmn
e
+ Rn,n1e
M

F (z, z ) =

where
Tmn =

m1
Y

j=n+1

Sj+1,j =

 ik (d d )

e jz j+1 j Sj+1,j Sn+1,n

Tj+1,j

1 Rj,j+1Rj,j1eikjz (2dj+1 2dj )

(2.12)

(2.13)

(2.14)

2.2.2 Matrix-Friendly Formulation

The matrix-friendly formulation can be derived in the implementation of
method of moments (MoM), where the spatial derivatives can be moved from
the Greens function to either testing or basis functions. In this manner,
the singularity of the LMGF can be reduced and the Sommerfeld integral
converges more rapidly. Here we only summarize the basic expressions and

the detailed derivation can be found in [21]. In the electric field integral
equation (EFIE) formulation, there are five scalar Greens functions in the
matrix-friendly formulation.
gss (r, r ) = k2 g TE (r, r)

(2.15)

2
gzz (r, r ) = kmn
g TM (r, r ) z z g TE (r, r)
n
gz1 (r, r) =
z g TM (r, r ) + z g TE (r, r)
m
m
gz2 (r, r ) =
z g TM (r, r ) + z g TE (r, r)
n

(2.16)

g (r, r) =

z z TM
g (r, r ) g TE (r, r)
2
knm

(2.17)
(2.18)
(2.19)

where the partial derivative with respect to z and z can be easily implemented in the spectral domain, namely, z = iknz and z = ikmz where
the signs are determined by the relative positions of the source and observation points.

2.3 Discrete Complex Image Method (DCIM)

2.3.1 Traditional DCIM
The Greens functions in (2.15)(2.19) can be expressed as an infinite integral
of the following type
i
g() =
8

dk

k (1)
H (k )
g (k ).
kz 0

(2.20)

If the integration kernel can be approximated by a series of complex exponentials,

g(k ) =

M
X

ai eikz bi

(2.21)

i=1

by applying the Sommerfeld identity [5],

i
eikr
=
r
2

dk

p
k (1)
H0 (k )eikz z , r = 2 + z 2
kz
9

(2.22)

the infinite integral can be calculated in a closed form,

g() =

M
X
i=1

ai

eikri
, ri =
4ri

2 + b2i .

(2.23)

The complex exponential series can be obtained by, for example, the matrix
pencil method [38], which approximates a function with real argument by
y(t) =

M
X

Ri eSi t .

(2.24)

i=1

The mapping of the the real variable t to kz is given by [35],



t
kz = k it + 1
, 0 t T0
T0


(2.25)

where ai and bi in (2.23) can be obtained by

bi =

iSi T0
, ai = Ri eikbi
k(1 iT0 )

(2.26)

The far field prediction of the traditional DCIM is poor, and various remedies
have been proposed [35][37], [40].

2.3.2 DCIM Based on the Sommerfeld Branch Cut

When the transverse distance is large, we deform the integration path to the
Sommerfeld branch cut.

One Branch-Cut Case

If the layered medium is backed by a PEC ground plane, there is only one
branch cut associated with the top layer [5]. A typical microstrip structure
falls into this case, as is shown in Figure 2.2. Due to the Cauchys theorem
and Jordans lemma, the integral defined along the Sommerfeld integration
path (SIP) is equivalent to the path integral along the Sommerfeld branch
cut (SBC) and some pole contributions, as shown in Figure 2.3.
Based on the deformed path, the Greens function can be expressed as a

10

Figure 2.2: A typical microstrip structure (unit: mm).

Figure 2.3: One branch-cut case: Sommerfeld integration path (SIP),

Sommerfeld branch cut (SBC) and poles in the complex k plane.

11

superposition of the following two terms,

g = gbranch + gpole
where

Z
k (1)
i
gbranch =
dk
H (k )
g (k )
8 SBC
kN z 0
1 X k,q (1)
gpole =
H (k,q )Res [
g (k,q )]
4 q kN z,q 0

(2.27)

(2.28)
(2.29)

where q is the number of poles and Res [

g (k,q )] is the residue of the kernel.
The locations of the poles and relevant residues can be obtained by a robust pole-searching algorithm [43]. Pole contributions are represented by the
Hankel function, which has the following asymptotic behavior,
(1)

H0 (k )

2 i(k 4 )
e
(k )
k

(2.30)

If k is real, we have
gpole

p
1/

(2.31)

The gbranch represents the radiation modes from the branch cut integration,
which can be obtained in closed form by the new DCIM. By transforming
the variable from k to kN z , (2.28) becomes
gbranch

i
=
8

(1)

dkN z H0 (k )
g (k )

with
dkN z =

k
dk
kN z

(2.32)

(2.33)

In order to use (2.24) to approximate the kernel of (2.32), we can let

kN z = t

T0
, 0 t T0 .
2

(2.34)

Then, ai and bi can be obtained similarly from Si and Ri ,

bi =

Si
, ai = Ri eibi T0 /2
i

(2.35)

Once it is approximated by complex exponentials, we can apply the alterna12

Figure 2.4: Two branch-cut case: Sommerfeld integration path (SIP),

Sommerfeld branch cuts (SBC) and poles in the complex k plane.
tive Sommerfeld identity in terms of kz integration [5] to get a closed-form
solution of gbranch ,
i
eikr
=
r
2

(1)

dkz H0 (k )eikz z , r =

p
2 + z 2

(2.36)

The radiation modes include spatial wave and the lateral wave, which have
the following asymptotic behavior respectively
gspatial 1/

(2.37)

glateral 1/2

(2.38)

From (2.31), (2.37), and (2.38), we can see that usually the surface wave
dominates in the far field. However, at the interface, the spatial waves of the
primary term and the secondary term cancel each other and the lateral wave
can be observed, if there are no pole contribution for certain cases. This
cancelation in DCIM was first analyzed in details in [40].

Two Branch-Cut Case

For a general layered medium, there are two branch cuts associated with the
top and bottom layers, where radiation modes can be supported. In this
case, the path and possible poles are shown in Figure 2.4.
13

In this case, (2.20) becomes

g = gbranch,N + gbranch,1 + gpole

(2.39)

where gbranch,N and gpole are similar to those in (2.32) and (2.29), while
gbranch,1 has the form of
gbranch,1

i
=
8

(1)

dk1z H0 (k )

k1z
g(k ).
kN z

(2.40)

2.4 Numerical Results

Several numerical results are presented in this section. Only gss and g are
used as illustrative examples here. The microstrip shown in Figure 2.2 is
first studied. The working frequency is f = 3 GHz, and the source point
and observation point are at the interface between the air and dielectric
substrate. We apply the traditional DCIM for small , and switch it to the
new implementation when is large. The transition region can be set in
100 < k0 < 101 . In the following examples, we set the transition point at
k0 = 100.5 . For this microstrip problem, only one real TM pole is found. The
gss and g are calculated in Figure 2.5 and Figure 2.6; both agree well with
those from numerical integration. In Figure 2.6, since there is no TE pole,
we can observe that the asymptotic behavior of gss is 1/2 , which represents
the lateral wave. It agrees with the results by the three-level DCIM in [40]
except for a constant due to the definition of the Greens function. It is also
reported in [40] that the popular two-level DCIM cannot correctly capture
the branch point contribution in this case. In Figure 2.6, since g contains
both TE and TM waves, the pole contribution dominates in the far field,

which is of 1/ .
The acceleration of DCIM lies in the fact that multiple sourceobservation
points can share the same image coefficients. For the numerical examples
shown in Figure 2.5 and Figure 2.6, we compare the computational time
for calculating 5,000 k0 samples. The central processing unit (CPU) time
is listed in Table 2.1. We can observe that the computation can be much
accelerated by using DCIM.
To validate the two-branch cut case, a 3-layer model with lossy material
14

4
Numerical
DCIM

log10|gss|

0
2
4
6
8
3

0
1
log10(k0)

Figure 2.5: The magnitude of gss versus k0 for the microstrip structure.
The asymptotic behavior is 1/2 , which agrees with the theoretical
prediction.
Table 2.1: CPU time comparison (seconds)
LMGF

Numerical

DCIM

gss

56.82

0.59

74.69

2.44

shown in Figure 2.7 is studied. The working frequency is f = 1.5 GHz.

Both gss and g are calculated and compared with the numerical integration
results, as are shown in Figure 2.8 and Figure 2.9. Again, good agreement
can be observed. In this case, the poles are general complex numbers away
from the real axis, so their contributions decay quickly when is large. The
lateral wave dominates in the far field with the asymptotic behavior of 1/2 ,
as are shown in Figure 2.8 and Figure 2.9.

2.5 Summary
A novel implementation of the DCIM based on Sommerfeld branch cut is
proposed to improve the far field prediction of the LMGF. By contour deformation, the Greens function can be naturally decomposed into the radiation
modes and guided modes. The guided modes can be obtained by a robust
pole-searching algorithm and the radiation modes can be calculated in a
15

0
Numerical
DCIM

log10|g|

2
3
4
5
6
3

0
1
log10(k0)

Figure 2.6: The magnitude of g versus k0 for the microstrip structure.

The far field is dominated by the pole contribution, which has the

asymptotic behavior of 1/ .

Figure 2.7: A 3-layer model (unit: mm).

16

4
Numerical
DCIM

log10|gss|

0
2
4
6
8
3

0
1
log10(k0)

Figure 2.8: The magnitude of gss versus k0 for the 3-layer model. The
asymptotic behavior is 1/2 .

2
0

Numerical
DCIM

log10|g|

2
4
6
8
10
12
3

0
1
log10(k0)

Figure 2.9: The magnitude of g versus k0 for the 3-layer model. The
asymptotic behavior is 1/2 .

17

closed form, so that the evaluation can be made efficient compared to the
direct numerical integration. For small interaction, we simply switch back
to the traditional DCIM to capture the near field. One should note that
in this new implementation, when becomes small, the length of sampling
path in spectral domain increases, and it becomes harder to approximate the
kernel. Efforts can be made to improve this DCIM in the near field, such
as extracting the asymptotic behavior analytically. At the same time, for
cases when poles are very close to the branch cut, the accuracy of function
approximation may be affected and more careful treatment of the poles is
necessary. Such efforts shall also be carried out in the future to improve this
DCIM.

18

CHAPTER 3
MIXED-FORM THIN-STRATIFIED
MEDIUM FAST-MULTIPOLE
ALGORITHM (MF-TSM-FMA)
A mixed-form thin-stratified medium fast-multipole algorithm (MF-TSMFMA) is proposed for fast simulation of general microstrip structures at both
low and mid-frequencies. The newly developed matrix-friendly formulation
of layered medium Greens function (LMGF) is applied in this algorithm.
For well-separated interactions, the contour deformation technique is implemented to achieve a smoother and exponentially convergent integral. The
two-dimensional addition theorem is then incorporated into the integrand to
expedite the matrix-vector product. In our approach, multipole expansion
(low-frequency fast-multipole algorithm) as well as the plane wave expansion
(mid-frequency fast-multipole algorithm) of the translational addition theorem are combined into a single multilevel tree to capture quasi-static physics
and wave physics simultaneously. The outgoing wave is represented first in
terms of multipole expansion at leafy levels, and then switched to plane wave
expansion automatically at higher levels. This seamless connection makes the
algorithm applicable in simulations, where subwavelength interaction (circuit
physics) and wave physics both exist.

3.1 Introduction
The simulation of microstrip structures has attracted intensive study for
many years. Since this kind of structure usually involves a thin layered
medium, the LMGF can be applied to reduce the number of unknowns. All
the acceleration techniques introduced at the beginning in Chapter 2 can be
implemented to expedite the process during the matrix filling stage in the integral equation method. However, since the integral equation method leads
to a full matrix, the memory consumption is large and the matrix-vector
product is still time-consuming in iterative solvers. Fast algorithms are in19

dispensable for large-scale computation. In the last two decades, various

fast algorithms were proposed, first based on the free space Greens function,
and then extended to other kernels. The incomplete list includes the fast
multipole method (FMM) [44], [45] and the multilevel fast multipole algorithm (MLFMA) [46][48], the adaptive integral method (AIM) [49], [50],
the conjugate gradient fast Fourier transform (CG-FFT) [51], [52] and the
precorrected FFT method [53][55], the multilevel Greens function interpolation method [56], the kernel-independent or black-box FMM [57], [58],
the multilevel matrix decomposition algorithm (MLMDA) [59], the adaptive
cross approximation (ACA) algorithm [60], [61], the IES3 [62], and the IEQR algorithm [63] etc. These algorithms can roughly be categorized into two
classes, the algebraic algorithms, which are kernel independent and are thus
more flexible, and the physics-based algorithms, which utilize the physical
property of the concrete kernel and are thus usually more efficient, especially
for dynamic field. In this chapter, we only focus on one of the physical algorithms and the applications in microstrip structures. A detailed survey of
the algebraic algorithms can be found in [60].
For the physical algorithm, both MLFMA- and FFT-based methods have
been extended to layered medium problems, for instance, in the parasitic
extraction and simulation of microstrip antennas and arrays. The former one
factorizes the Greens function by using the translational addition theorem
in a multilevel manner [47], [48], [64][67] and the latter one takes advantage
of the transverse translational invariance property of the interactions and
applies FFT to accelerate the computation in the Fourier domain [50], [52],
[55], [68], [84]. Both these methods can achieve an O(N log N) computational
complexity in the dynamic regime.
MLFMA is more efficient than FFT if the objects are sparse, however,
MLFMA usually suffers from a low frequency breakdown problem [70], which
means that the plane wave expansion of the addition theorem, attractive at
mid-frequencies, is error uncontrollable due to the finite precision of the computer. The multipole expansion is usually implemented instead in the low
frequency regime. This expansion is suitable for scale-invariant interaction,
but becomes inefficient rapidly when entering the electrodynamic regime. For
many broadband large-scale simulation, both electrodynamic interaction and
sub-wavelength interaction exist in one problem. This requires a broadband
fast algorithm capable of capturing the wave physics and quasi-static physics
20

simultaneously. Motivated by the idea of mixed-form fast multipole algorithm in free space [70], a mixed-form thin-stratified medium fast-multipole
algorithm (MF-TSM-FMA), based on the newly-formulated LMGF [21], is
developed in this chapter. In this MF-TSM-FMA, the multipole expansion
and the plane wave expansion are combined into one multilevel tree, where
different scales of interaction can be separated by the multilevel nature of
the MLFMA [71]. A transition equation can be derived to connect these two
expansions.

3.2 Review of Thin-Stratified Medium Fast-Multipole

Algorithm (TSM-FMA)
3.2.1 TSM-FMA Based on the Plane Wave Expansion
We show the LMGF again here [21],
r) = ( z)( z)g TE (r, r )
G(r,
+

1
2
knm

( z)( z)g

(3.1)
TM

(r, r )

where
g

TE/TM

i
(r, r ) =
8

dk
(1)
H0 (k )F TE/TM (k , z, z )
kmz k

(3.2)

To obtain a TSM-FMA, the path deformation technique is necessary to improve the convergence of the Sommerfeld integral.

Deformation of the Integration Path

For well separated interactions, exponential convergence can be achieved by
deforming the Sommerfeld integration path into the vertical branch cut for
a thin layered medium. In this structure, the vertical dimension is tightly
confined compared to the horizontal dimension and the source-observation
dependent steepest descent path for a thick layered medium evolves to be
source-observation independent [5], letting us to implement the fast algorithm. Possible pole singularity should be included in such deformation ac21

cording to Cauchys theorem [47]. The detailed implementation about the

path integration, and the choice of Riemann sheets, can be found in [67] and
will not be repeated here.

Two Dimensional MLFMA

The transverse interaction of the LMGF is the Hankel function, so 2D MLFMA
can be implemented to accelerate the computation. The core equations of
the MLFMA are listed below [72],
(1)
H0 (k ji )

1
=
2

JI () =

P
X

djJ ()
JI ()Ii ()

(3.3)

Hp(1) (k JI )eip(J I 2 )

(3.4)

p=P

jJ () = eik jJ

Ii () = eik Ii

(3.5)

where J and I are box centers for the observation point j and the source
point i, respectively. Here k is in general complex. In implementing TSMFMA, the z-dependent propagation factor F (z, z ) can be easily factorized
and embedded into the radiation and receiving patterns [67].

3.2.2 Low Frequency Breakdown of the TSM-FMA

When the operation frequency decreases, the plane wave expansion becomes
unstable due to the numerical overflow of the outgoing-to-outgoing translator

. It has been discussed for free space MLFMA (spherical harmonics) by

[70]. We will show that this is also the case for layered medium problems
(cylindrical harmonics) for general complex argument k . The asymptotic
expression of the Hankel function is
(1)

H0 (k ) Ceik0 eu

(3.6)
2

where k = k0 + iu2 , eik0 is the propagating wave while eu is the decaying

factor. For a fixed quadrature point u, we can show that numerical instability of the plane wave expansion exists in the low frequency regime. In
implementing the addition theorem, the worst interaction situation is shown
22

Figure 3.1: The worst interaction in applying the addition theorem.

Figure 3.2: Relative error by using plane wave expansion.

in Figure 3.1. For a given quadrature u, if we fix the frequency, then the
imaginary part of the argument may introduce different decays for different
, making the Hankel function become negligible rapidly for large . To
make a fair comparison, we may fix the transverse distance and change the
frequency. Let the box size to be b = 5 mm and u = 4. The relative errors of
the plane wave expansion with the expansion order P for different box sizes
are shown in Figure 3.2. It is shown that the expansion is unstable when
the box size is much smaller than the wavelength. Since in near region, the
evanescent wave dominates and the field is scale invariant, the plane wave
expansion based on wave physics is no longer suitable.

23

3.3 Mixed-Form Thin-Stratified Medium

Fast-Multipole Algorithm (MF-TSM-FMA)
3.3.1 TSM-FMA Based on the Multipole Expansion
The multipole expansion, which uses the undiagonalized addition theorem,
is free of low-frequency breakdown since normalization can be made and
are well balanced. However, since the
different translators (
and )
operator cannot be calculated analytically in the multipole expansion, we
will apply the matrix-friendly formulation of the LMGF.

Matrix-Friendly LMGF
The matrix element in the MoM procedure can be expressed in the matrixfriendly LMGF as [21]

z2
z1
ss
zz
+ Zji
+ Zji
Zji = Zji
+ Zji
+ Zji

(3.7)

ss
Zji
= im hfs (rj ), gss (rj , ri ), fs (ri )i

(3.8)

zz
Zji
= im h
z f(rj ), gzz (rj , ri ), z f(ri )i

(3.9)

where

z1
Zji
= im h
z f(rj ), gz1 (rj , ri ), f(ri )i

(3.10)

z2
Zji
= im h f(rj ), gz2 (rj , ri ), z f(ri )i

(3.11)

Zji
= im h f(rj ), g (rj , ri ), f(ri )i

(3.12)

gss (rj , ri ) = k2 g T E (rj , ri )

(3.13)

2
gzz (rj , ri ) = kmn
g T M (rj , ri ) z z g T E (rj , ri )
n
gz1 (rj , ri ) =
z g T M (rj , ri ) + z g T E (rj , ri )
m
m
gz2 (rj , ri ) =
z g T M (rj , ri ) + z g T E (rj , ri )
n

(3.14)

g (rj , ri ) =

z z T M
g (rj , ri ) g T E (rj , ri )
2
knm

24

(3.15)
(3.16)
(3.17)

Compared with (3.1), the matrix-friendly Greens function has lower-order

singularity, since several spatial derivatives have been moved to the basis
function.

Factorization of the Propagation Factor

The factorization of the propagation factor is critical in implementing a generalized TSM-FMA, because the 2D MLFMA can only factorize the transverse interaction, and the vertical interaction should be factorized manually
to bind to the radiation and receiving patterns. The propagation factor takes
different forms if the positions of source and observation points are different.
According to the expressions shown in Chapter 2, the propagation factor can
be easily factorized as followings,
A. n = m
F (z, z ) = F1 (z, z ) + fv2 (z)I2 (m)fr2 (z )

(3.18)

+fv3 (z)I3 (m)fr3 (z )

where

F1 (z, z ) = eikmz |zz |

(3.19)

fv2 (z) = eikmz (dm+1 z)

(3.20)

m,m1 eikmz (dm+1 2dm +z ) ]

fr2 (z ) = [eikmz (dm+1 z ) + R

(3.21)

mR
m,m+1
I2 (m) = M

(3.22)

fv3 (z) = eikmz (zdm )

(3.23)

m,m+1 eikmz (2dm+1 dm z ) ]

fr3 (z ) = [eikmz (z dm ) + R

(3.24)

mR
m,m1
I3 (m) = M

(3.25)

F (z, z ) = fv (z)I(m, n)fr (z )

(3.26)

n,n+1 eiknz (2dn+1 dn z) ]

fv (z) = [eiknz (zdn ) + R

(3.27)

m,m1 eikmz (dm+1 2dm +z ) ]

fr (z ) = [eikmz (dm+1 z ) + R

(3.28)

m Tmn
I(m, n) = M

(3.29)

B. n > m

where

25

C. n < m
F (z, z ) = fv (z)I(m, n)fr (z )

(3.30)

n,n1 eiknz (z+dn+1 2dn ) ]

fv (z) = [eiknz (dn+1 z) + R

(3.31)

m,m+1 eikmz (2dm+1 dm z ) ]

fr (z ) = [eikmz (z dm ) + R

(3.32)

m Tmn
I(m, n) = M

(3.33)

where

Here I(m, n) only depends on layer parameters. When n = m, the modulus

sign in the direct term F1 (z, z ) = eikmz |zz | is a problem, however, we can
show that this modulus sign can be removed in the context of vertical branch
cut integration [67]. After implementing factorization of the propagation
factor, the addition theorem can be applied to accelerate the computation.
However, there are totally ten scalar components that need to be factorized
separately, hence optimization of the patterns are necessary.

Layout of Radiation and Receiving Patterns

For a certain source-observation-layer pair, the partial derivative z or z in
the matrix-friendly formulation (3.13)(3.17) is acting directly on the propagation factor. It can be calculated easily according to the concrete expression
of F (z, z ). For example, if n > m,
z z F (z, z )
= z f (z)z f (z )I(m, n)
h
i
ikmz (dm+1 z )
ikmz (dm+1 2dm +z )

= ikmz e
+ Rm,m1 e
h
i
iknz (zdn )
iknz (2dn+1 dn z)

m
iknz e
Rn,n+1 e
Tmn M

(3.34)

By investigating the matrix-friendly formulas (3.7)(3.17), we know that

there are eight scalar components and one 2D vector component requiring
factorization separately. The memory requirement is huge if we store all
the radiation and receiving patterns directly. In our implementation, the
radiation and receiving patterns are filled on the fly in the process of matrixvector product. In order to reduce the number of MLFMA processes, we can
set two receiving patterns for one radiation pattern. For example, The TE
26

Table 3.1: Layout of radiation and receiving patterns

Radiation

Receiving A

Receiving B

g T E (r )
x f(r )

x f(r)g T E (r)

TE

TE

(r )
y f(r )

(r ) f(r )

z g T E (r )
z f(r )
g T M (r )
z f(r )

z g T M (r ) f(r )

y f(r)g

TE

f(r)g

(r)

TE

0
z f(r)z g T E (r)

(r)

f(r)g T E (r)
z f(r)g T M (r)
z f(r)g T M (r)

z f(r)z g T E (r)

f(r)z g T M (r)
f(r)z g T M (r)

z1
part of Zji
and Zji
can be accelerated at the same time. If we denote

g (r, r) = g (r)g (r )

(3.35)

then we can set one radiation pattern and two receiving patterns to be
Ir = f(r )g T E (r )

(3.36)

Iva = f(r)g T E (r)

(3.37)

Ivb = z f(r)z g T E (r)

(3.38)

By doing this, two interactions can share the same radiation patterns,
outgoing waves and incoming waves. However, minor modification is needed
in the last stage, where incoming waves at the leafy level are disaggregated
into observation points by using different receiving patterns. The pattern
layout is listed in Table 3.1. By categorizing the patterns, one may only
need to calculate six kinds of interactions instead of ten.

Two Dimensional Low Frequency MLFMA

In the low frequency regime, multipole expansion is applied to factorize the
Hankel function. The general core equations are listed in the following [72]
mn (ji ) =

XX
M

mM (jJ )M N (JI )N n (Ii )

(3.39)

(1)

mn (ji ) = Hmn (k ji )ei(mn)ji

27

(3.40)

Figure 3.3: Relative error by using multipole expansion.

mn (ji ) = Jmn (k ji )ei(mn)ji

(3.41)

For Hankel function, we have

(1)

H0 (k ji ) =

XX
M

0M (jJ )M N (JI )N 0 (Ii )

(3.42)

Though this multipole expansion is equivalent to the plane wave expansion mathematically, the former is stable in the low frequency regime since
normalization can be easily implemented to balance the and , to make
each numerical step error controllable. For the same setup as in Figure 3.1,
the relative error of the multipole expansion is shown in Figure 3.3. It is
obvious that the expansion is stable at low-frequencies. However, when the
frequency increases, the multipole expansion becomes very inefficient since
more and more expansion terms are required to capture the wave physics.
So it should be switched back to plane wave representation to describe the
dynamic field.

28

3.3.2 Mixed-Form Expansion

For many applications such as broadband microstrip antennas or arrays, the
interaction may involve both quasi-static field and dynamic field. Neither of
the two expansions can be used independently to capture the two different
fields. Following the idea of [70], one can combine the two expansions into
a multilevel tree and switch between them dynamically. Based on addition
theorem and integral representation of the Bessel function, we can derive the
following mixed-form core equation
1
mn (ji ) =
2

JI ()Ii ()ei(+ 2 )n
dei(+ 2 )m jJ ()

(3.43)

This equation combines the multiple expansion and plane wave expansion.
Substituting it into (3.39), we can construct a mixed-form TSM-FMA. At
leafy levels, the box size is much smaller compared to the wavelength, where
interaction is in the sub-wavelength regime, the outgoing wave is calculated
by multipole expansion. With the process of aggregation, the box size becomes large enough, where the interaction may enter the dynamic regime, the
outgoing wave is switched into the plane wave representation to capture the
wave physics. Similar idea can be applied in the process of disaggregation.
In matrix notation, the MF-TSM-FMA for a certain k is expressed as,
H0 (k ji ) = [jJ1 ]1P1 [J1 J2 ]P1 P2 [T ]P2 K3 [J2 J3 ]K3 K3
(1)

[I]TK3 K4 [J3 J4 ]K4 K4 [

J4 I4 ]K4 K4 [I4 I3 ]K4 K4
[I]K4 K3 [I3 I2 ]K3 K3 [T ]K3 P2 [I2 I1 ]P2 P1 [I1 i ]P1 1
(3.44)
where [I] is the interpolation matrix in the dynamic regime, and [T ] transforms the multipole outgoing wave into plane wave outgoing wave.



T = ei(+ 2 )n KP

(3.45)

The implementation scheme is shown in Figure 3.4. We denote the level

where I2 in the figure resides as the switch level. From our experience, the
switch level can be chosen with box size to be around Re [k ] b = 0.2
0.4, namely 0.10 0.20 for Re [k ] = k0 , where 0 is the wavelength in
free space. If the switch box size is smaller than the box size of the leafy
29

Figure 3.4: The MF-TSM-FMA structure.

level, all the interaction are calculated in plane wave expansion and the MFTSM-FMA goes back to the mid-frequency TSM-FMA.
For dynamic field, the computational complexity is O(N log N), while for
quasi-static field, the computational complexity is O(N). The complexity
of MF-TSM-FMA should stay in between depending on the location of the
switch level. However, the real computational time may depend on other
factors. For example, the translation matrix in multipole expansion is full,
which may take longer time even though the complexity is lower for a given
problem.

3.4 Numerical Results

In this section, several numerical results are demonstrated to validate our
algorithm. We have first simulated an 8 4 corporate-fed microstrip array
available in the literature [48], [50], [52], [54]. The geometrical structure of
this array is shown in Figure 3.5. The detailed parameters are l = 10.08 mm,
= 11.79 mm, d1 = 1.3 mm, d2 = 3.93 mm, l1 = 12.32 mm, l2 = 18.48
mm, D1 = 23.58 mm, D2 = 22.40 mm. The thickness of the substrate is
30

d = 1.59 mm, and the relative permittivity and permeability are r = 2.2
and r = 1. The radiation patterns are calculated by feeding at the input
port at the working frequency of f = 9.42 GHz. The patterns in both Eplane ( = 0o ) and H-plane ( = 90o ) are shown in Figure 3.6. The results
agree well with the results extracted from [36]. Here, the box size at the
leafy level is 0.30 , and a five-level FMA is set up. Since the switch level is
below the leafy level, the MF-TSM-FMA applies the plane-wave expansion
for all levels. In order to activate the low-frequency FMA, we decrease the
frequency to f = 2.5 GHz, and calculate the radar cross section (RCS) of
this microstrip array. The results are shown in Figure 3.7. The leafy level
box size is 0.080 , and the switch level is 4. The number of unknown is 9,394,
and the memory consumption of the MF-TSM-FMA is 40 MB. Next, to show
the capability of modeling vertical structures, a dipole antenna mounted on
a 4 4 mushroom-type electromagnetic band gap (EBG) is simulated, with
its geometrical structure shown in Figure 3.8. The geometrical parameters
are: w = 29 mm, g = 1 mm, d = 1 mm, h = 2 mm, L = 63 mm, W = 1
mm, H = 2 mm. The FR4 is used as the substrate with = 4.4 and
tan = 0.02. The antenna is working at f = 2.2 GHz. The radiation pattern
in the E plane is calculated and shown in Figure 3.9 (in unit of dB). Due
to the layered medium assumption, there is no back lobe and side lobes in
the radiation pattern. A 30 30 microstrip array [47] shown in Figure 3.10
is then simulated at frequency of 2.2 GHz, 1.1 GHz, and 550 MHz, where
a = b = 6 cm, L = 3.66 cm, W = 2.6 cm, the thickness of the substrate is

d = 0.158 cm, r = 2.17. The number of unknown is 117,000. A seven-level

MF-TSM-FMA is constructed. At the high frequency end (f = 2.2 GHz),
the RCS is compared with the one sampled from [47] and good agreement is
achieved. At the low frequency end (f = 550 MHz), the leafy level box size is
0.03 0 , and the switch level is 5, which corresponds to a box size of 0.12 0 .
The memory requirement is 250 MB, the central processing unit (CPU) time
for each iteration is 2.2 min. Finally, we list the memory consumption and
CPU time per iteration versus the number of unknowns in Figure 3.12, where
the proportions (the number of levels) of multipole expansion and plane wave
expansion are set to be similar. All the examples run on a personal computer
(PC) with Intel 2.66 GHz processor. It is obvious that this algorithm can
be used in the simulation of large scale microstrip structures with acceptable
computational cost.
31

Figure 3.5: The geometrical structure of the 8 4 microstrip array.

Radiation pattern (dB)

10
20
30
40

E at = 90
E at = 0
MFTSMFMA
Ref.

50
60

80

60

40

20

0
( )

20

40

60

80

Figure 3.6: The radiation patterns of the microstrip array in E plane

( = 0 ) and H plane ( = 90 ) at 9.42 GHz.

32

35
MoM
MFTSMFMA

40

Bistatic RCS (dBsm)

45
50
55
60
65
70
75
80

50

100

150

200
()

250

300

350

Figure 3.7: The bistatic RCS of the microstrip array at f = 2.5 GHz and
(i , i ) = (60 , 0 ).

Figure 3.8: A dipole on an EBG structure, top and side view.

33

Figure 3.9: Radiation pattern of the dipole in E plane, due to the layered
medium assumption, the ground plane is infinitely large.

Figure 3.10: The geometrical model of a 30 30 microstrip array.

34

40

Bistatic RCS (dBsm)

60

80

100

120
f=2.2 GHz
Ref.
f=1.1 GHz
f=0.55 GHz

140

160

10

20

30

40

50

60

70

80

90

()

300

200

100

10

Time per iteration (min)

Memory requirement (MB)

Figure 3.11: Bistatic RCS of the 30 30 microstrip array.

0
12
4

x 10

Figure 3.12: Memory requirement (solid line) and CPU time per iteration
(dash line) versus number of unknowns.

35

3.5 Summary
A mixed-form thin-stratified medium fast multipole algorithm is developed in
this chapter. By path deformation, the LMGF can be expressed by severalterm summation of Hankel functions weighted by z-dependent propagation
factors. For each quadrature point, the Hankel function can be accelerated
by 2D MLFMA. The matrix-friendly formulation of the LMGF is applied,
and interaction is categorized into six components if we introduce two receiving patterns for one radiation pattern. Due to the numerical instability
of the TSM-FMA in the low frequency regime, a mixed-form TSM-FMA is
proposed for broadband simulation. By utilizing the multilevel nature of
the MLFMA, interactions with different scales are accelerated by different
FMA expansions, namely multipole expansion and plane wave expansion. A
transform equation is derived to embed the two expansions into one MLFMA
tree. Numerical results show the accuracy and efficiency of this algorithm. It
should be noted that in the very low frequency regime, the integral equation
in layered medium may suffer from another kind of low frequency breakdown,
where the existence of the null space of the operator plagues the numerical
procedure. Before modeling large problems in extremely low frequency, one
must develop an efficient algorithm accounting for the two breakdowns at
the same time.

36

CHAPTER 4
REMEDIES FOR LOW FREQUENCY
BREAKDOWN OF INTEGRAL
OPERATOR IN LAYERED MEDIUM
When the working frequency further decreases, the low frequency breakdown
of the integral operatorthe electric field integral equation (EFIE) occurs.
Two remediesthe loop-tree decomposition and the augmented electric field
integral equation (A-EFIE)which were first developed in free space, are
extended to layered medium in this chapter. In the loop-tree decomposition,
the current is decomposed into divergence-free part and non-divergence-free
part according to quasi-Helmholtz decomposition when frequency tends to
zero, in order to capture both capacitance and inductance physics. Frequency normalization and basis rearrangement are applied to stabilize the
matrix system. In the A-EFIE, the traditional EFIE can be cast into a
generalized saddle-point system, by separating charge as extra unknown list
and enforcing the current continuity equation. Frequency scaling for the
matrix-friendly layered medium Greens function (LMGF) is analyzed when
frequency tends to zero. Rank deficiency and the charge neutrality enforcement of the A-EFIE for LMGF is discussed in detail. The electrostatic limit
of the A-EFIE is also analyzed. Without any topological loop-searching algorithm, electrically small conducting structures embedded in a general layered
medium can be simulated by using this new A-EFIE formulation.

4.1 Introduction
Computational electromagnetics becomes indispensable as a computer-aided
design (CAD) methodology in various electrical engineering applications,
such as in integrated circuit and wireless communication device. The operating frequency of the electrical systems keeps on increasing to several
gigahertz, meanwhile fabrication process has achieved nanoscale. Hence, a
broadband simulation tool is badly needed for capturing circuit physics of
37

the tiny structures as well as wave physics for the whole package. Unfortunately, however, in the sub-wavelength regime, or so-called low frequency
regime, the electric field and magnetic field decouple, and the total current
in Maxwells equations decomposes into a divergence-free part and a curlfree part, with different frequency scaling properties. In this situation, the
commonly used EFIE method solved by the method of moments (MoM)
[4] with the Rao-Wilton-Glisson (RWG) basis function [73] suffers from a
low frequency breakdown problem, where vector potential gradually looses
its significance compared with the scalar potential part when the frequency
decreases, and the EFIE operator becomes singular [1]. Various approaches
have been proposed to overcome this problem in the last few years. One of the
most popular remedies is the loop-tree or loop-star decomposition [74], [75],
where the solenoidal and irrotational components of the unknown current can
be separated due to the quasi-Helmholtz decomposition (also known as Hodge
decomposition), to capture inductance physics and capacitance physics when
the frequency tends to zero. However, even after frequency normalization,
the matrix is still ill-conditioned. Preconditioning is necessary to improve
the convergence when iterative solvers are applied. Several effective preconditioners have been proposed, either based on the basis-rearrangement, where
the favorable property of electrostatic problems is utilized [76], or based on
the near-field interactions, where the incomplete factorization with a heuristic drop strategy is applied [77]. By using the Calderon identity and the dual
basis or Buffa-Christiansen basis function [78], [79], a more effective preconditioner has been constructed [80][82]. The loop-tree or loop-star method
has also been implemented with the LMGF, which is more versatile in the
simulation of printed antenna and planarly integrated circuit [83][85].
However, one big issue associated with the loop-tree or loop-star method
is the loop-search process. It is a bottleneck for complicated interconnecting geometries with increasing number of unknowns, where many entangled
global loops may exist. Situation becomes even worse when layered medium
with conducting ground plane is involved, where extra implicit global loops
are introduced because of the vias. To avoid the loop-search process, the
idea of separating current and charge to construct a stable formulation has
been studied recently. The current and charge integral equation (CCIE)
method [86] puts charges into the extra unknown list and manipulates the
equation system to be of the second kind. While in the separated potential
38

integral equation (SPIE) method [87], the scalar potential is included as the
unknowns, where resistive loss and dielectric loss are introduced to flatten
the condition number when the frequency is low. In the recently developed
augmented electric field integral equation (A-EFIE) method [88][90], the
similar idea of separating current and charge as independent unknowns is
applied. By enforcing the current continuity equation explicitly and implementing a proper frequency scaling, the EFIE can be cast into a generalized
saddle point system [91]. With the help of a constraint preconditioner for
the saddle point system and the mixed-form fast multipole algorithm, a real
package problem with more than one million unknowns has been successfully
solved on a personal computer [89].
In this chapter, we discuss both loop-tree decomposition and the A-EFIE
involving the LMGF.

4.2 Loop-Tree Decomposition and Frequency

Normalization
If traditional RWG function [73] is used to expand the current, we have
J(r) =

N
X
i=1

where
fi (r) =

fi (r)Ii = f(r) I

+
i
,
2A+
i

2Ai ,
i

0,

r Ti+

r Ti

(4.1)

(4.2)

otherwise

The RWG basis function is defined on an adjacent triangular patch pair, and

A
i is the area of the two triangles associated with the i-th basis, and the i
is the vector pointing to a point r from the two vertices. Here, the function
is normalized by its edge length for convenience.

The EFIE has the form of

r)(r ), J(r)i = t i
t ihG(r,

r )(r) J(r ) = t Ei (r) (4.3)

dr G(r,

r ) is the LMGF, and Ei (r) is the

where t is the unit tangential vector, G(r,
39

excitation. After applying the Galerkin procedure [4] to the EFIE, a matrix
equation can then be set up
I =V
Z
(4.4)
where the matrix elements in the matrix-friendly formulation [21] are shown
in (3.7)(3.17).
When 0, the matrix elements have different frequency scaling behaviors. The detailed analysis can be found in Section 4.5 and here we only list
the scaling for the lossless case.
Z ss O(), Z zz O(), Z z1 O(), Z z2 O(), Z O( 1)

(4.5)

where O( ) represents the scaling properties with respect to frequency

when 0. It is obvious that Z will be much larger than the remaining
four terms in the low frequency regime. If they are summed up directly,
some information will be lost because of the limited numerical precision. As
a remedy, we decompose the RWG basis into loop basis which is divergencefree and tree basis which is non-divergence-free. Then the term of Z with
O( 1) can be easily extracted from the other terms with O().
J(r) =

NL
X

fLi (r)ILi +

NT
X

fT i (r)IT i

(4.6)

i=1

i=1

The loop functions can be obtained from the RWG functions [92], and the
tree functions have the same form as the RWG functions. The loop-tree basis
spans the same linear space as the RWG basis. For a general structure with
a hole, loop basis contains both local loop and global loop, as shown in
Figure 4.1.
Applying the loop-tree function as both expansion function and testing
function, the matrix equation then takes the form of
"

LL Z
LT
Z
TL Z
TT
Z

# "

IL
IT

"

VL
VT

(4.7)

where
LL = Z
ss + Z
zz
Z
LL
LL

(4.8)

LT = Z
ss + Z
zz + Z
z1
Z
LT
LT
LT

(4.9)

40

Figure 4.1: Local loop and global loop (dashed lines).

TL = Z
ss + Z
zz + Z
z2
Z
TL
TL
TL

(4.10)

TT = Z
ss + Z
zz + Z
z1 + Z
z2 + Z

Z
TT
TT
TT
TT
TT

(4.11)

When 0, the frequency scaling behavior of the matrix system is

"

LL (O()) Z
LT (O())
Z
T L (O()) Z
T T (O( 1))
Z

# "

IL
IT

"

VL
VT

(4.12)

We observe that after the loop-tree decomposition, components with different

frequency scaling behaviors are isolated into different blocks, enabling the
use of frequency normalization to balance the matrix system. Considering
the imbalance between im and (in )1 in MKS system, ik0 is chosen to
normalize the linear system, instead of i, to further balance the diagonal
elements so that the two blocks are almost of the same order.
"
# "
# "
#
LL (O(1)) Z
LT (O())
(ik0 )1 Z
IL
(ik0 )1 VL

=
T L (O())
T T (O(1))
Z
ik0 Z
(ik0 )1 IT
VT
(4.13)
The matrix equation behaves much better after frequency normalization
since contribution from ZLL is elevated while the contribution from ZT T is
suppressed and the diagonal elements are dominant.

41

4.3 Basis Rearrangement via Connection Matrix

As has been reported in [76], when applying iterative method in solving
the matrix equation, the performance is still not good because of the electroquasi-static block (right lower block). The situation is similar here, and basis
rearrangement is needed to make the matrix equation amenable to iterative
solvers. After such rearrangement, the electro-quasi-static part in terms of
operator with spatial derivatives, can be expressed in terms of a smoother
operator as that in classical electrostatic problems, where no convergence
problem occurs.
In electrostatic problems, the electric charge density can be expanded by
pulse basis. For a charge-neutral system, a particular patch is chosen as a
neutralizing patch, where a negative pulse is defined. Usually we choose the
last patch as the neutralizing patch, and the charge density is
(r) =

NX
P 1
i=1

[Pi (r) PNp (r)]Qi =

NX
P 1

Ni (r)Qi

(4.14)

i=1

R
where Pi (r) is the normalized pulse basis satisfying s dsPi (r) = 1. Here Ni (r)
is the new expansion function to describe charge density in a charge-neutral
system. It can be written in a more compact form,
(r) = Nt (r) Q

(4.15)

Since the loop current does not produce charges, we can connect the tree
current to charge via the continuity equation
t (r) IT = iNt (r) Q
J
T

(4.16)

Taking the inner product of the above equation with P(r), we get


t (r) IT = K
IT = i Q
P(r), J
T

(4.17)

connects the current coefficient with the charge coefThe sparse matrix K
ficient. Substituting (4.17) into (4.13), a system with good convergence can

42

0
LT
RWG

RCS (dBsm)

50

100

150

200

250
0
10

10

10
10
Frequency (KHz)

10

10

Figure 4.2: RCS of a sphere.

be finally established as follows
"

LL
LT K
1
(ik0 )1 Z
Z
t )1 Z
T L ik0 (K
t )1 Z
TT K
1
(K

# "

IL
Q

"

(ik0 )1 VL
t )1 VT
(K

(4.18)

For the electro-quasi-static part, the matrix in terms of the function N is


N N = N, g, Nt
Z

(4.19)


T T = f , g, f t
Z

(4.20)

t N = f
K

(4.21)

NN = (K
t )1 Z
TT K
1
Z

(4.22)

while the matrix in terms of the tree function or the RWG function f is

Since

we can get

It can be easily found that the element in the electro-quasi-static part of

(4.18) is indeed expressed in terms of basis N, as expected.

43

Figure 4.3: Geometrical model of the capacitor.

10

LT + Free Space
LT + Layered Medium
6

Negative input reactance (Ohm)

10

10

10

10

10

10
1
10

10

10
10
Frequency (KHz)

10

10

Figure 4.4: Negative input reactance of the capacitor, r2 = 6.0.

44

4.4 Numerical Results

Some numerical results are demonstrated in this section. Figure 4.2 shows
the RCS of a unit sphere above a three-layered medium with its center 1.2
m above the first interface, the thickness of layer 2 is 0.3 m. The relative
permittivities are r3 = 1.0, r2 = 2.56 and r1 = 6.5. The plane wave is
incident from i = 60 , i = 0 , and the RCS is evaluated at the observation angle of s = 60 , s = 90 . Both loop-tree and RWG functions are
applied for comparison. It can be shown that the two results agree very well
at midfrequencies, but with the decrease of operating frequencies, the RWG
result becomes unstable, while the loop-tree result maintains the scale invariance property well. A rectangular parallel capacitor embedded in a layered
medium is shown in Figure 4.3. Both layer 1 and layer 3 are air, while layer 2
is dielectric. The negative input reactance is plotted in Figure 4.4, compared
with the one from free space. It is demonstrated again that loop-tree basis
maintains the low-frequency scale invariance property very well.

4.5 Augmented Electric Field Integral Equation

(A-EFIE)
As has been mentioned, the loop-searching process is a big challenge for
complex interconnecting geometries. The global loop shown in Figure 4.1
is especially difficult to find. This section extends the recently developed
A-EFIE [89], which is absent of loop searching, to layered medium.

4.5.1 Review of A-EFIE for Free Space Greens Function

Compared with the LMGF, the free space Greens function has the closedform expression,



r ) = I +
G(r,
g(r, r)
(4.23)
2
k0
where k0 is the wave number in free space and g(r, r) is the solution to the
scalar Helmholtz wave equation with a point source,

g(r, r) =

eik0 |rr |
4|r r |

45

(4.24)

Again we first use the RWG basis function [73] shown in (4.2) to expand the
induced current. In free space, the mixed potential integral equation (MPIE)
is favorable since the scalar vector potential and scalar potential are well
defined under the Lorentz gauge, and the operator in the Greens function
can be easily moved to the RWG basis function. The surface divergence of
the RWG basis is analytical.

s fi (r) =

1
A+
i
A1
i

0,

r Ti+

r Ti

(4.25)

otherwise

After applying the Galerkin procedure [4], a matrix system shown in (4.4)
can be obtained and the matrix can be expressed as
= ik0 0 A
+ 0 S

Z
ik0

(4.26)

where 0 is the free space wave impedance. The A and S correspond to the
magnetic vector potential and electric scalar potential
ji = hfj (r), g(r, r), fi (r )i
[A]

(4.27)

ji = h fj (r), g(r, r), fi (r )i

[S]

(4.28)

From (4.26)(4.28), we can see that the vector potential block and the scalar
potential one are imbalanced when the frequency is low, namely k0 0,
since they are in different frequency order. The EFIE operator becomes
singular because any divergence-free current is a solution to the EFIE in the
quasi-static limit,
I0
S

(4.29)

To balance the system, the charge can be separated and added into the
unknown list to make the system stable in an augmented fashion [88], [89].
We define the normalized pulse basis function on each triangular patch as
pi (r) =

1
Ai

0,

46

r Ti

otherwise

(4.30)

If further defining a patch-to-patch scalar potential matrix as,

ji = hpj (r), g(r, r), pi (r )i
[P]

(4.31)

we can obtain the relationship between the patch-pair based (in terms of
divergence of RWG basis) scalar potential matrix and the patch-based one
=D
T P
D

(4.32)

relates the domain of the RWG basis and the

where the incidence matrix D
patch basis,

1, Patch j is the positive part of RWG i

[D]ji =
1, Patch j is the negative part of RWG i

0,
otherwise

(4.33)

Due to the current continuity equation, we have

J = ik0 c0
D

(4.34)

where c0 is the light speed in vacuum and is the charge unknowns. Substituting the above equations into the EFIE matrix equation, and enforcing the
current continuity equation explicitly, we can arrive at the following A-EFIE
system
"
# "
# "
#
D
T P

A
ik0 J
01 V

=
(4.35)

D
k02I
c0
0
This equation is the generalized saddle point system with the lower right
block nearly equals to zero and various methods can be applied to solve this
problem efficiently [91].

4.5.2 A-EFIE for Layered Medium Greens Function

Again the matrix-friendly formulation of LMGF is applied. Note that the
notation of the matrices is slightly different from (3.7)(3.12) but the scalar
Greens functions are the same as (3.13)(3.17). Here, im is extracted out

47

for convenience in constructing the A-EFIE.

= im {Z
ss + Z
zz + Z
z1 + Z
z2 + Z
}
Z

(4.36)

ss ]ji = hfjs (r), gss(r, r ), fis (r )i

[Z

(4.37)

zz ]ji = h
[Z
z fj (r), gzz (r, r ), z fi (r )i

(4.38)

z1 ]ji = h
[Z
z fj (r), gz1 (r, r), fi (r )i

(4.39)

z2 ]ji = h fj (r), gz2 (r, r), z fi (r )i

[Z

(4.40)

]ji = h fj (r), g (r, r ), fi (r )i

[Z

(4.41)

where

Frequency Scaling
Since the primary (direct) term can be analyzed in the similar way as in free
space, only secondary (reflected or transmitted) terms are considered in this
section. We first assume that the layered medium is lossless. For general
case, namely i 6= j and i 6= j , when 0, kiz ik , the frequency
scaling of the Fresnel reflection coefficient is
Ri,j =

pj pi
pj kiz pi kjz
=
O( 0)
pj kiz + pi kjz
pj + pi

(4.42)

where p = for TE wave and p = for TM wave. Then we can get the
frequency scaling for other quantities:
i,j O( 0), M
m O( 0), Tmn O( 0)
R

(4.43)

Subsequently, the frequency scaling for the propagation factor is

F (k , z, z ) O( 0)

(4.44)

Finally the frequency scaling for the matrix element in (4.37)(4.41) is

ss O( 0), Z
zz O( 0), Z
z1 O( 0), Z
z2 O( 0), Z
O( 2)
Z
(4.45)

48

We can separate the matrix into two parts according to the frequency scaling
 ss

= mr Z
+Z
zz + Z
z1 + Z
z2
A

(4.46)

Hence, we can have

= knm Z

S
nr

(4.47)

= ik0 0 A
+ 0 S

Z
ik0

(4.48)

O( 0), S
O( 0)
A

(4.49)

where

Equation (4.48) has the same form as Equation (4.26) in free space, which
allows us to augment the EFIE in a similar fashion as in (4.35).
Since most material is non-magnetic, namely i = j , we discuss this
situation separately. The frequency scaling of the Fresnel reflection coefficient
for TE wave is a high order term of frequency,
TE
Ri,j
=

kiz kjz
j kiz i kjz
=
O( 2)
j kiz + i kjz
kiz + kjz

(4.50)

Then we have
T E O( 2)
R
i,j

(4.51)

T E O( 0), T T E O( 0)
M
m
mn

(4.52)

However, we still have

ss , other four terms in (4.38)(4.41)

Notice the fact that except for the Z
consist of TE as well as TM wave, and the TM part is still on O( 0). This
means that the frequency scaling for these four terms in (4.45) are still valid.
ss , the leading
By careful dimensional analysis, we can show that even for Z
order term is still O( 0). In a word, for non-magnetic material, (4.45)(4.49)
are also valid.
Dielectric loss and conductor loss can be introduced to alleviate the the
low frequency breakdown in free space [87]. For a structure embedded in
a layered medium, if we introduce dielectric loss to each layer, since the
equivalent permittivity is
r = r +
49

i
0

(4.53)

the frequency scaling of the Z becomes

O( 1)
Z

(4.54)

so the scalar potential matrix becomes

O( 1)
S
The A-EFIE can then take the alternative form
"
# "
# "
#
(ik0 )1 D
T P

A
J
(ik0 0 )1 V

D
ik0I
c0
0

(4.55)

(4.56)

Consistency Validation
We discuss two extreme cases to analyze the consistency of the A-EFIE for
layered medium Greens function. One with homogeneous layers (free space)
and the other with perfect electrical conductor (PEC) layer (half space), both
of which have closed form Greens functions.
In free space, the EFIE can be separated into two parts, the vector potential part and the scalar potential part, both of which are scalar problems
with scalar Greens function, because of the homogeneity of the medium,
shown in (4.26)-(4.28). However, in the layered medium, the response of a
dipole is polarization dependent. A vertical electric dipole can only generate a TM wave, while a horizontal electric dipole generates TE as well as
TM waves. The polarization dependence leads to the difficulty in defining
ss
a uniform scalar potential. In the matrix friendly formula, we can see Z
zz manifest the different response of a horizontal and a vertical dipole
and Z
in a layered medium. If we asymptotically make the inhomogeneity disappear, namely, making mr 1 and mr 1 for each layer, and applying the
Sommerfeld identity
i
g(r, r ) =
4

dk
0

J0 (k )eikz |zz |
kz

(4.57)

ss and Z
zz recover the polarization independent vector potential in
the Z
, it goes back to the scalar potential
(4.27). Similar situation holds for the Z
part in (4.28) when removing the inhomogeneity. The physical meanings of
50

 z1 and Z
z2 are ambiguous due to the lack of exact
the remaining two terms Z
definition of scalar potential. There is no correspondence in free space, since
the TE and TM waves cancel each other when the layered medium degrades
into a free space. By appearance, we can interpret it as the cross interaction
between charge and the vertical current. According to their same frequency
ss and Z
zz , we can group them together to obtain the layered
scaling with Z
medium A-EFIE, as is done in (4.46).
For a half space with a PEC layer, the image method can be applied and
the dyadic Greens function can be expressed in a closed form [12]


1

r ) = I [g(r, r) gi (r, r )] + 2
G(r,
z zgi (r, r )
k02

(4.58)

where g(r, r) and gi (r, r ) is the free space scalar Greens function with real
and S
defined as
source point r and image source point r , with A
ji = hfj (r), g(r, r) gi (r, r ), fi (r )i + 2h
[A]
z fj (r), gi (r, r), z fi (r )i (4.59)
ji = h fj (r), g(r, r) gi (r, r), fi (r )i
[S]

(4.60)

where we assume the interface is at z = 0, so r = (x , y , z ), and r =

(x , y , z ). Notice the fact that RT E = 1 and RT M = 1 at the interface
in the propagation factor F (k , z, z ), we can reproduce (4.59) and (4.60) by
our general A-EFIE formula, with the help of Sommerfeld identity.
z1 and Z
z2 cannot be validated by these two cases, we
Although the terms Z
can show their significance by numerical examples, where vertical structure
exists in a dielectric layered medium, and the two terms are always there
with nonzero value.

4.6 Charge Neutrality Issue

Charge neutrality enforcement is very important in the A-EFIE for lowfrequency problems, as stated in [89]. The motivation of enforcing the charge
neutrality is because of the rank deficiency of the A-EFIE. For the A-EFIE
shown in (4.35), the upper block is exactly the same as the traditional EFIE
except that the scalar potential part is expressed in terms of patch basis
and charge is separated as a set of independent unknowns. The lower block
51

constraints the current continuity condition to make the augmented system

solvable. However, rank deficiency exists in the A-EFIE matrix, due to the
Here, k 2 is a eigenvalue of the A-EFIE
definition of the incidence matrix D.
0
and it tends to zero finally when the frequency goes to zero [89]. Usually, the
deflation method [93] can be applied to remove the smallest eigenvalue, for
example, in the CCIE Formula [86]. Motivated by the basis rearrangement
preconditioner in the loop-tree decomposition [76], as has also been discussed
in Section 4.3, we can also apply the charge neutrality enforcement to remedy
this problem [89]. This is driven by the physical observation of the problem,
and can be easily extended to different layered medium problems.
We discuss this issue in the context of layered medium. If it is backed by
a conducting ground plane, which is a common situation in the circuit problems, it acts as a charge bath and absorbs the extra charge of the structure,
so we should distinguish situations whether there is a via connected to the
ground in some parts of the structure. As will be shown in the following,
the condition number is always bounded when the frequency goes to zero, by
properly enforcing the charge neutrality condition.

4.6.1 Structures Not Connected to a Ground Plane

If the structure is not connected to the charge bath, the total charge is always neutral. This condition shall be enforced when in low frequency domain
is
due to the rank deficiency. By definition of the RWG basis, the matrix D
linearly dependent or singular. It is evident that the lower block sub-matrix
of the A-EFIE is rank deficient when the frequency is low (k0 0). It is
necessary to enforce the charge neutrality condition to make the lower block
full rank.
Two transform matrices can be introduced to fulfill the charge neutrality
enforcement [89], and the final A-EFIE system becomes:
"

T P
B

A
D
D

F
k02Ir

# "

ik0 J
c0 r

"

01 V
0

(4.61)

with the forward and backward transform as

,
r = F

r
=B
52

(4.62)

Figure 4.5: The geometrical structure of the loop inductor embedded in a

seven-layer medium, unit: mm. The central layer is a magnetic material. A
delta gap excitation is applied at the center of the top arm.
where r is the reduced charge unknowns and the Ir is the reduced identity
matrix.
A rectangular loop embedded in a seven-layer medium is shown in Figure
4.5, with its layer parameters specified in the figure. The condition numbers
versus frequency are demonstrated in Figure 4.6. We can see that without charge neutrality enforcement, the condition number grows unboundedly
when decreasing the frequency. It increases in the order of 1/k02 because of
the right lower block. The eigenvalue distribution of the A-EFIE matrix at
f = 1 Hz is shown in Figure 4.7. After the charge neutrality enforcement,
the smallest eigenvalue has been removed away from the origin. We can also
observe that when frequency increases, however, the lower-right block is a
identity matrix scaled by k02 . Thus the lower block is no longer singular and
such enforcement is no longer necessary.

4.6.2 Structures Connected to a Ground Plane

In this case, the charge neutrality condition cannot be guaranteed since the
is no longer
charge bath absorbs the extra charge. The incidence matrix D
singular. In this situation, no special treatment is needed since the A-EFIE
system is full rank. For a half rectangular loop connected to the ground
plane shown in Figure 4.8, the condition number versus frequency is shown
in Figure 4.9. Since the A-FEIE matrix is no longer singular because of the
ground plane, the condition number remains constant when decreasing the
frequency without any special treatment.
Here, the forward and backward transform matrices become the identity
53

18

10

With CNE
Without CNE

16

10

14

Condition number

10

12

10

10

10

10

10

10

10

10

10
Frequency (Hz)

10

10

Figure 4.6: The condition number versus frequency for the rectangular loop.
The condition number is unbounded when decreasing the frequency. Charge
neutrality enforcement (CNE) makes the condition number constant.

x 10

With CNE
Without CNE

Imaginary part of eigenvalues

3
2
1
0
1
2
3
0

3
4
5
Real part of eigenvalues

9
3

x 10

Figure 4.7: The eigenvalue distribution for the rectangular loop at 1 Hz.
The smallest eigenvalue is removed away from the origin after the charge
neutrality enforcement (CNE).

54

Figure 4.8: The geometrical model of the half loop embedded in a five-layer
medium (including the PEC layer), unit: mm. A delta gap excitation is
applied at the center of the top arm.
5

3.04

x 10

Condition number

3.038

3.036

3.034

3.032
Without CNE
3.03
0
10

10

10
Frequency (Hz)

10

10

Figure 4.9: The condition number versus frequency for the half loop. Since
it is connected to the ground plane, charge neutrality cannot be
guaranteed. The condition number is bounded when decreasing the
frequency without any special treatment.
matrix in this situation.
=B
= I
F

(4.63)

This observation can be used as a guideline when dealing with complex

structures. For a structure with s independent surfaces, each with pk triangular patches, ik inner edges and gk ground edges, k = 1, 2...s. If there are m
surfaces connected to the ground plane, then the total number of unknowns
is
s
X
N=
(ik + gk + pk ) m
(4.64)
k=1

One should note that though the number of unknowns in A-EFIE increases

much compared to the loop-tree decomposition, where the number of un55

knowns is the same as the number of RWG basis, the memory requirement
F
and B

increases marginally since all transformation matrices such as D,

are sparse and consume marginal memory when iterative solver is applied.

4.7 Electrostatic Limit

In the electrostatic limit, the electric potential is expressed by the following
boundary value problem [1],
2 = s(r)

(4.65)

(r) = 0 r S

(4.66)

and the integral equation becomes

Z

dr Gs (r, r)(r ) = 0 0 (r) r S

(4.67)

where is the surface charge density, Gs (r, r ) is the static layered medium
Greens function [94] and 0 is the potential generated by the source s(r). In
this limit, a typical problem is the parallel plate capacitor, shown in Figure
4.10. Here we connect the two plates with a narrow strip so that we can
apply the electrodynamic analysis.
In this limit, the A-EIFE suffers from an inaccuracy problem, even though
the matrix is nonsingular, because the current is a higher order term in
frequency. To capture the current accurately to arbitrary order, the perturbation method should be applied [95]. We will show that the charge is always
stable and describes the electrostatic physics. In DC, the A-EFIE becomes
"

s D
T P
s B

A
D

F
0

# "

ik0 J
c0 r

"

01 V
0

(4.68)

s and P
s are matrix evaluated at k0 = 0. For this problem, the
where A
current J disappears, while the charge remains constant,
J 0, c

(4.69)

Since the matrix is still full rank, we have unique solution. However, the
56

z (mm)

50
0
50

1000
500
0
500
x (mm)1000
1000

500

500

1000

y (mm)

Figure 4.10: The geometrical model of the circular parallel plate capacitor,
with a dielectric layer (r = 2.65) inserted in between. A delta gap is
applied at the edge. The mesh is refined to capture the fringing effect.
current is no longer correct due to the finite numerical precision. If we
discard the ik0 J term manually, we have
T P
s B
r = 0 V
D

(4.70)

It is shown that the electrostatic information is included in the A-EFIE. In

free space, it is straightforward to check that the static Greens function in
(4.67) and (4.70) is the same one
Gs (r, r) =

1
4|r r |

(4.71)

In a layered medium, we will show the static form of the general matrix s in the following subsection, which agrees with
friendly Greens function in P
that of [94] applied in the electrostatic analysis except for minor differences
such as the layer index, constant 1/40 and the sign definition in the Fresnel
reflection coefficient.

57

4.7.1 Electrostatic Layered Medium Greens Function

In the electrostatic limit, the LMGF shown in [94] can be derived from our
general matrix-friendly formulation
Z
1 X +
1
TM

dJ0 ()eZi i (4.72)

Gs (r, r ) = z z g (r, r ) =
nr
4nr i 0

where the Greens function is described by several images with weight i and
distance Zi
A. m = n
There are one primary term and four image terms.
0 = 1, Z0 = |z z |

(4.73)

m,m1 M
m , Z1 = 2dm + z + z
1 = R

(4.74)

m,m+1 M
m , Z2 = 2dm+1 z z
2 = R

(4.75)

m,m+1 R
m,m1 M
m , Z3 = 2hm + z z
3 = R

(4.76)

m,m+1 R
m,m1 M
m , Z4 = 2hm z + z
4 = R

(4.77)

B. m < n
There are four image terms.
+
1 = Tmn
Mm , Z1 = z + z

(4.78)

n,n+1 T + M

2 = R
mn m , Z2 = 2dn+1 z z

(4.79)

m,m1 T + M

3 = R
mn m , Z3 = 2dm + z + z

(4.80)

+
m,m1 R
n,n+1 Tmn
4 = R
Mm , Z4 = 2dm + 2dn+1 + z z

(4.81)

C. m > n
There are also four image terms.

1 = Tmn
Mm , Z1 = z z

(4.82)

n,n1 T M

2 = R
mn m , Z2 = 2dn + z + z

(4.83)

m,m+1 T M

3 = R
mn m , Z3 = 2dm+1 z z

(4.84)

58


m,m+1 R
n,n1 T M

4 = R
mn m , Z4 = 2dm+1 2dn z + z

where,

(4.85)

i+1,i+2 e2hi+1
Ri,i+1 + R

Ri,i+1 =
i+1,i+2 e2hi+1
1 + Ri,i+1 R

(4.86)

2hi1

i,i1 = Ri,i1 + Ri1,i2 e

R
i1,i2 e2hi1
1 + Ri,i1 R

(4.87)

m =
M
+
Tmn
=

1
m,m1 R
m,m+1 e2hm
1R

n1
Y

Tj,j+1
j+1,j+2 e2hj+1
1 Rj+1,j R
j=m

Tmn
=

m1
Y
j=n

Tj+1,j
j,j1e2hj
1 Rj,j+1R

hi = di+1 di

(4.88)

(4.89)

(4.90)
(4.91)

4.8 Numerical Results

Several numerical results are presented in this section. The input reactance
of the rectangular loop shown in Figure 4.5 is calculated, where the results
are compared with the loop-tree (LT) decomposition and traditional EFIE in
Figure 4.11. The EFIE breaks down quickly when the frequency decreases,
while the A-EFIE is very stable and agrees very well with the loop-tree
decomposition. We also calculate the input reactance of the half rectangular
loop mentioned above. Because it is connected to a conducting ground plane,
the current can flow along this half loop. The input reactance is shown in
Figure 4.12 and compared with traditional EFIE. Similar phenomenon can be
observed. Since non-magnetic dielectric is transparent to the loop inductor,
the PEC half space model can be used to validate the result, which is also
shown in Figure 4.12. Good agreement with the A-EFIE for general layered
medium can be observed. Finally, a circular parallel plate capacitor with
radius of unit length (a = 1 m) is shown in Figure 4.10 with a dielectric layer
inserted in between (r = 2.65). The distance is set to be d/a = 0.1. The
capacitance extracted from current and charge and static solver are shown in
Figure 4.13. The current suffers from an inaccuracy problem, as mentioned in
59

10

Input reactance (Ohm)

10

10

10

AEFIE
LT
EFIE

10

10

10

10

10
Frequency (Hz)

10

10

Figure 4.11: The input reactance of the rectangular loop. A-EFIE agrees
well with the loop-tree (LT) decomposition, while the traditional EFIE
breaks down quickly when decreasing the frequency.
last section, while the charge is accurate and agrees with the static analysis.
The analytic result from asymptotic expansion [96] is also shown to validate
the numerical results. In this example, when the frequency is below 1 MHz,
the relative error of the A-EFIE with charge information is around 0.1%. If
the frequency is increased, wave physics begins to play a role and the parallel
plate is no longer a pure capacitor.

4.9 Summary
Two remedies for low frequency breakdown of the EFIE operator involving the LMGF are discussed in this chapter. By decomposing the current
into the divergence-free part and the non-divergence-free part according to
the property of the current in the low frequency regime, we can capture
both capacitance and inductance physics, which are of the same importance
in circuit physics. Mathematically, loop-tree decomposition combined with
frequency normalization makes the matrix stable. The connection matrix
further makes the final system well conditioned, as is done in free space. Due
to the complexity of pole searching, the A-EFIE for LMGF is developed in
this chapter. The frequency scaling is analyzed for both lossless and lossy
media. The rank deficiency of the A-EFIE in layered media depends on if the

60

10

Input reactance (Ohm)

10

10

10

10

AEFIE
EFIE
PEC halfspace

10

10

10

10

10
Frequency (Hz)

10

10

Figure 4.12: The input reactance of the half loop. A-EFIE maintains the
scale invariance very well while the traditional EFIE breaks down quickly
when decreasing the frequency. Since the non-magnetic dielectric is
transparent to the inductor, a PEC half space model is applied to validate
the results.

Capacitance (nF)

0.9

0.8

0.7
AEFIE I
AEFIE Q
Static
Analytic

0.6

0.5
0
10

10

10

10
10
Frequency (Hz)

10

10

Figure 4.13: The capacitance of the circular parallel plate capacitor.

A-EFIE I represents the capacitance extracted from current, while A-EFIE
Q means the capacitance extracted from charge. The A-EFIE current
suffers from an inaccuracy problem while the A-EFIE charge is stable. The
result agrees with the static solver. Both are further validated by the
analytic solution. When the frequency is below 1 MHz, the relative error of
A-EFIE Q is around 0.1%.

61

charge neutrality condition is satisfied. For independent structures, the enforcement is necessary in the low frequency regime, while at mid-frequencies,
such implementation is no longer necessary. For structures connected to the
ground, the A-EFIE matrix is full rank, and no special treatment is needed.
The electrostatic limit is analyzed and compared with the static formulation.
One should note that in the electrostatic limit, though the matrix is well conditioned, the A-EFIE has an inaccuracy problem due to the extremely small
current, which has a high-order frequency dependence. Remedies such as the
perturbation method shall be implement to capture the high-order term of
the current in the future.

62

CHAPTER 5
A NEW LAYERED MEDIUM GREENS
FUNCTION (LMGF) FORMULATION FOR
GENERAL OBJECTS
In this chapter, a new Greens function formulation is developed systematically for modeling general homogeneous (dielectric or magnetic) objects in
a layered medium. The dyadic form of the Greens function is first derived
based on the pilot vector potential approach. The matrix representation in
the moment method implementation is then derived by applying integration
by parts and vector identities. The line integral issue in the matrix representation is investigated, based on the continuity property of the propagation
factor and the consistency of the primary term and the secondary term. The
duality principle of the LMGF is also discussed to make the formulation succinct. After that, the extinction theorem is revisited in the inhomogeneous
background and a surface integral equation for general homogeneous objects
is set up. Different from the popular mixed potential integral equation formulation, this method avoids the artificial definition of scalar potential. The
singularity of the matrix representation of the Greens function can be made
as weak as possible. Several numerical results are demonstrated to validate
the formulation developed at the end of this chapter.

5.1 Introduction
During the last several decades, intensive study has been carried out to model
objects embedded in a layered medium by using integral equation methods
and the LMGF. The two-dimensional (2D) analysis was developed in [97],
where scattering from a conducting cylinder partially buried in a half space
is analyzed. For more practical applications, the three-dimensional (3D)
analysis was carried out and applied to many problems on radiation and
scattering from perfect electric conducting (PEC) objects in the background
of layered medium [98], [16], [99], [52], [100], [101]. Among them, the LMGF,
63

based on transmission line analog and the corresponding mixed potential

integral equation (MPIE) developed in [16], becomes one of most popular
formulations. A correction vector is introduced to make the definition of a
scalar potential possible, which is due to the fact that a scalar potential of
a point source in layered medium does not in general exist [102]. Different
choice of the vector potential and correction term leads to three different
formulations: Formulations A, B, and C.
For analysis of general scatterers in layered medium, the volume integral
equation (VIE) method is usually applied [103], [104]. Though VIE can handle inhomogeneous objects, the number of unknowns is typically large and the
equation should be reformulated if there are contrasts in both permittivity
and permeability. On the other hand, the SIE is favorable for homogeneous
objects due to its smaller unknown number and elegant form. Different from
the PEC case, both surface electric current and magnetic current are required
and more types of Greens functions are needed in the SIE formulation for
general homogeneous (dielectric or magnetic) objects [17], [18], [1]. A dielectric body of revolution (BOR) buried in soil covered with a layer of snow
is solved by SIE in [105], where Formulation C in [16] is applied. SIE is
further extended for general dielectric target in a multilayered medium in
[106], with application to radar-based sensing of plastic land mines. In [107],
SIE is applied to analyze a dielectric resonator in layered medium coupled to
a microstrip circuit. Though MPIE is also applied, the Greens function is
different from previous one. The vector potential is kept in the simple form
as is developed by Sommerfeld [6], while the scalar potential is manipulated
in a dyadic way. Other attempts are also made in solving this problem, for
example, a spectral integral method is developed for 2D PEC and dielectric
objects with a closed boundary in layered medium [108].
To avoid the artificial definition of the scalar potential in layered medium,
a new formulation of LMGF is developed in EFIE for modeling PEC objects
[21]. By applying integration by parts, the singularity can be reduced as
weak as possible. It is shown in [109] that this formulation is as convenient
as the popular MPIE method developed in [16].
In this chapter, we formulate the Greens function in a systematic way and
extend it to model general homogeneous objects [110]. The configuration is
shown in Figure 5.1. To make the formulation succinct, only the electric-type
e and G
e ) is discussed here, and the magnetic-type
Greens function (G
64

Figure 5.1: A homogeneous dielectric object is embedded in a layered

medium. The external excitation is either a plane wave or a Hertzian
dipole. Equivalent electric and magnetic currents are induced on the
boundary which then generate the scattered field.
m and G
m ) can be obtained by applying the duality
Greens function (G
principle of the LMGF, which will be shown later on.

5.2 Dyadic Form of LMGF

The dyadic form of the layered medium Greens function describes the electric or magnetic field generated by an electric or magnetic dipole in a layered
medium. In other existing literature, one usually defines four dyadic Greens
EJ (r, r ), G
HJ (r, r ), G
EM (r, r ) and G
HM (r, r). In this chapfunctions, G
e (r, r )
ter, however, we will apply the notation slightly differently, namely G
m (r, r) and their curls [1]. It is convenient to introduce the integral
and G
operators with kernels of different Greens functions to express the electromagnetic fields generated by arbitrary electric and magnetic current sources.
So we will first introduce different sets of integral operators and then derive
the expressions of their kernels.

5.2.1 The Integral Operators

There are generally four integral operators [1].
E(r) = LE (r, r) J(r ) + KE (r, r ) M(r )
65

(5.1)

H(r) = LH (r, r) M(r ) + KH (r, r) J(r )

(5.2)

where J denotes the electric current and M is the magnetic current, r is the
observation point and r is the source point. These expressions are valid for
any inhomogeneous medium; however, we will focus on the planarly layered
medium here. The operator LE is defined as [1]
Z

LE (r, r ) = i

e (r, r)(r )
dr G

(5.3)

e is the electric-type dyadic Greens function. According to Maxwells

where G
equations, the operator KH can then be defined as

KH (r, r ) = (r)

e (r, r )(r )
dr G

(5.4)

Due to the duality principle of the Greens function shown in Section 5.4, we
can further define the other two operators as,

LH (r, r ) = i

KE (r, r ) = (r)

m (r, r )(r )
dr G

(5.5)

m (r, r)(r )
dr G

(5.6)

It should be noted that in homogeneous space, usually two operators are

involved (LE and KE , or simply L and K), since LH and KH can be easily
obtained by LE and KE explicitly, either by a factor of wave impedance (1/ 2 )
or a negative sign [1], due to the fact that for homogeneous medium Greens
e = G
m (see Section 5.4). However, in layered medium,
function, we have G
the relation can only be determined by the duality principle implicitly, so we
prefer to stay with the four operators in this chapter.
It should also be noted that the dyadic Greens function can typically be
expressed in a uniform manner, where the direct term is absorbed in the
total expression in the spectral domain, as is done for example in [17]. However, from implementation point of view, we will always separate the Greens
function into the primary term (direct interaction in a homogeneous environment) and the secondary term (due to the reflection and transmission of the
layered medium) in the following, since the primary term can be determined
in closed form and be calculated separately. In the following, we only focus

66

 e and G
e , since G
m and G
m can be easily obtained by the
on G
duality principle of the Greens function shown in Section 5.4.

5.2.2 Primary Term of the Greens Function

The primary term is the same as the homogeneous medium Greens function.
It is nonzero only when the observation point and the source point are in
the same layer. For the sake of completeness, we will also briefly list the
expressions of the primary term.

e
Expression of G
e of the primary term in (5.3) is
The G



Ge (r, r ) = I + 2
g(r, r)
km

(5.7)

with the scalar Greens function

g(r, r) =

eikm R
eikm |rr |
=
4|r r |
4R

(5.8)

where the subscript m is the layer index of the source point (and n for
observation layer, here m = n), and km is the wave number in layer m. In
Cartesian coordinates, the dyadic Greens function can be further written as,

 
 ik R 1 
m
e (r, r ) = I R
R
+ I 3R
R

G
g(r, r)
2 R2
km

(5.9)

where I is the identity dyad and

= R = r r = 1 [(x x )
R
x + (y y )
y + (z z )
z]
R
|r r |
R

(5.10)

e
Expression of G
e of the primary term in (5.4), we have
For G
e (r, r ) = Ig(r, r)
G
67

(5.11)

It can also be obtained in closed form as

e (r, r) =
G
z
y

z
0
x

zz
(y y )

R1 ikm

1
R

x g(r, r)
0
(z z )
y y
0
x x

g(r, r)

(5.12)

(x x )
0

5.2.3 Secondary Term of the Greens Function

The secondary term of the Greens function is from the reflection and transmission of the layered medium; it can be expressed as [21]
e (r, r ) = G
TE (r, r ) + 1 G
TM (r, r)
G
e
e
2
knm

(5.13)

2
where knm
= 2 n m , and

TE (r, r) = ( z)( z)g TE (r, r)

G
e

(5.14)

TM (r, r ) = ( z) ( z) g TM (r, r )
G
e

(5.15)

where
g (r, r) =
=

i
8 2
i
4

RR +

dks
eiks (rs rs ) F (ks , z, z )
kmz ks2

R +
0

(5.16)

dk
J (k )F (k , z, z )
kmz k 0

here represents TE or TM wave, F (k , z, z ) is the propagation factor [5],

p 2
[71], and kmz = km
k2 . In Cartesian coordinates, we have ks = kx x +ky y,
rs = x
x + y y, while in cylindrical coordinates, we further have kx = k cos ,
ky = k sin , and x x = cos , y y = sin . If we denote a 2D inverse
Fourier transform and a Sommerfeld integral as [16]
F 1 {f(kx , ky )} =

1
(2)2

RR +

dkx dky f(kx , ky )

ikx (xx )+iky (yy )

68

(5.17)

1
Sn {f(k )} =
2

dkf(k )Jn (k )kn+1

(5.18)

we have
i
g (r, r ) = F 1
2

F (ks , z, z )
kmz ks2

i
= S0
2

F (k , z, z )
kmz k2

(5.19)

The spatial derivatives in Cartesian coordinates can be transformed to be

higher order Sommerfeld integrals in the cylindrical system by applying the
integral representation of the Bessel function [5]
1
Jn (k ) =
2

deik cos +inin 2 .

(5.20)

The expressions have been derived in [16] (Appendix I) and will not be repeated here. Notice that the second-order Sommerfeld integrals are further
simplified to lower-order counterparts by using the recursive property of the
Bessel function [111], [18]
J2 (k ) =

2
J1 (k ) J0 (k ).
k

(5.21)

All the relations will be utilized in deriving the following Greens function
components (the time convention here is eit and is different from that of
[16]).

e
Expression of G
We first consider the TE wave, since
z = y x x y
we have

y y

TE (r, r) =
G
x y
e
0

where
GTE
e,xx

i
cos 2S1
=
2

F TE
kmz k2

(5.22)

y x 0
x x
0

0 g TE (r, r )
0

i
+ (1 cos 2) S0
4
69

F TE
kmz

(5.23)

(5.24)

GTE
e,xy

GTE
e,yy

i
=
sin 2S1
2

F TE
kmz k2

i
sin 2S0
4

F TE
kmz

TE
GTE
e,yx = Ge,xy
 TE 
 TE 
i
F
i
F
= cos 2S1
+ (1 + cos 2) S0
2
2
kmz k
4
kmz

(5.25)
(5.26)
(5.27)

Similarly, for TM wave, we have

z = x z x + y z y + ks2 z

(5.28)

then

x x z z x y z z x z k2

TM

TM (r, r) =
G
y x z z y y z z y z k2 g (r, r )
e
x z k2
y z k2
k4
where
i
GTM
e,xx = 2 cos 2S1

+ 4i

z z F TM
kmz k2

z z F TM
kmz

(1 + cos 2) S0
n
o
z z F TM
i
GTM
=

sin
2S
1
e,xy
2
kmz k2
n
o
F TM
+ 4i sin 2S0 z kzmz


i
z F TM
TM
Ge,xz = cos S1
2
kmz

TM
GTM
e,yx = Ge,xy
n
o
z z F TM
i
GTM
=
cos
2S
1
e,yy
2
kmz k2
n
o
F TM
+ 4i (1 cos 2) S0 z kzmz


i
z F TM
TM
Ge,yz = sin S1
2
kmz


i
z F TM
TM
Ge,zx = cos S1
2
kmz


i
z F TM
TM
Ge,zy = sin S1
2
kmz

70

(5.29)

(5.30)

(5.31)

(5.32)
(5.33)

(5.34)

(5.35)
(5.36)
(5.37)

GTM
zz

i
= S0
2

k2 F TM
kmz

(5.38)

e
Expression of G
e , from (5.13)(5.15), we simply have
For G
e (r, r ) = G
TE (r, r) + 1 G
TM (r, r )
G
e
e
2
knm

(5.39)

TE (r, r) = ( z)( z)g TE (r, r )

G
e

(5.40)

TM (r, r ) = k 2 ( z) ( z) g TM (r, r).

G
e
n

(5.41)

where

Here (5.41) is derived due to the fact that for the secondary field, we have


2 + kn2 g (r, r) = 0.

(5.42)

Again we first consider the TE wave, from (5.22), (5.28), and (5.40), we have

x z y x z x 0
TE

TE (r, r) =
G
y z y y z x 0 g (r, r )
e
k2 y
k2 x 0

(5.43)

where,
n TE o


i
z F
TE
=

G
sin
2S
1
e
2
kmz k2
xx
n TE o
+ 4i sin 2S0 zkFmz
n TE o


z F
i
TE
G
=
cos
2S
1
e
2
kmz k2
xy
n TE o
4i (1 + cos 2) S0 zkFmz
n TE o


i
z F
TE
cos
2S
G
=
1
e
2
kmz k2
yx
n TE o
i
+ 4 (1 cos 2) S0 zkFmz




TE
TE
G
= G
e

yy

71

xx

(5.44)

(5.45)

(5.46)

(5.47)


F TE
kmz
 TE 
F
i
= cos S1
.
2
kmz



i
TE
G
= sin S1
e
zx
2


TE
G
e
zy

(5.48)
(5.49)

Similarly, for the TM wave, we have

y x z
y y z
y k2
2 TM

TM (r, r ) =
G
x x z x y z x k2 kn g (r, r )
e
0
0
0
where

n
o
z F TM
i
=

sin
2S
kn2
1
2
kmz k2
xx
n
o
F TM
+ 4i sin 2S0 zk mz
kn2
n
o


z F TM
i
TM

Ge xy = 2 cos 2S1 kmz k2 kn2

n
o
F TM
+ 4i (1 cos 2) S0 zk mz
kn2
 TM 


i
F
TM

kn2
Ge xz = sin S1
2
kmz
n
o


z F TM
i
TM

Ge yx = 2 cos 2S1 kmz k2 kn2

n
o
F TM
4i (1 + cos 2) S0 zk mz
kn2




TM
TM
G
= G
e
e
yy
xx
 TM 


i
F
TM
G
=
cos
S
kn2
1
e
yz
2
kmz


TM
G
e

(5.50)

(5.51)

(5.52)

(5.53)

(5.54)

(5.55)
(5.56)

To validate the dyadic form of the Greens function derived in this section,
an electric dipole radiating in a seven-layer medium is investigated. The
layered medium is shown in Figure 5.2, and the working frequency is f = 300
MHz. The layered medium is both dielectric and magnetic. The source is
at (x = 0, y = 0, z = 1.4) m and the observation line is at (3 x
3, y = 1, z = 0.3) m. The polarization of the dipole is ( = 20o , = 30o ).
e is involved.
The electric field of an electric dipole is first evaluated, where G
The result is compared with that from the transmission line (TL) method
[17] as shown in Figure 5.3. Good agreement is observed. To further test the
magnetic-type Greens function, the electric field of a magnetic dipole is also
72

Figure 5.2: An electric or magnetic dipole is radiating in a seven-layer

medium (unit: m). The layered medium is both dielectric and magnetic,
and the layer constants are shown in the figure. The source point is in
Layer 2 and the observation line is in Layer 5.
m is involved. The result is shown in Figure 5.4 and
calculated, where G
again agrees well with the TL method.

5.3 Matrix Representation of LMGF

The dyadic form of the Greens function in the last section can map the
current source to the field directly. However, the singularity of the Greens
function is high in spatial domain due to the spatial derivatives, or it has high
frequency component in the spectral domain, which leads to slow convergence
of the Sommerfeld integral. Usually, in moment method implementation [4],
MPIE is applied, where the Greens function has lower singularity. As has
already been mentioned in the introduction, the scalar potential in layered
medium is ambiguous; so careful manipulation should be made to obtain
a proper MPIE. Due to this fact, a matrix-friendly formulation for PEC
objects without defining any artificial potential is developed in [21]. In this
formulation, integration by parts is applied wherever possible to transfer the
partial derivative from the Greens function to the basis functions, where
the Rao-Wilton-Glisson (RWG) basis [73] is applied as both expansion and
testing functions. Hence, the singularity of the Greens function in the spatial
domain can be made as weak as possible. In other words, the Sommerfeld
integral can be made rapidly convergent.
In this section, we will extend the matrix representation for general homo73

300
Ex
Ey

250
Electric field (V/m)

Ez
200
150
100
50
0
3

1
0
1
Observation point x (m)

Figure 5.3: The magnitude of electric field generated by an electric dipole.

The dipole polarization is ( = 20o , = 30o ) and is working at f = 300
MHz. The result is validated by the transmission line method (solid line).

1.4
Ex

Electric field (V/m)

1.2

Ey
Ez

1
0.8
0.6
0.4
0.2
0
3

1
0
1
Observation point x (m)

Figure 5.4: The magnitude of electric field generated by a magnetic dipole.

The dipole polarization is ( = 20o , = 30o ) and is working at f = 300
MHz. The result is validated by the transmission line method (solid line).

74

geneous (dielectric or magnetic) objects. Again we only discuss the LE and

e and G
e are involved since the LH and KE can be easily
KH where G
obtained by the duality principle of the Greens function.

5.3.1 Matrix Representation of Primary Term

The primary term is from the direct interaction in a homogeneous environment, hence the conventional MPIE form can be applied since the auxiliary
potentials can be easily determined if the Lorentz gauge is utilized [73]. The
matrix representation is
hfj (r), LE (r, r ), fi (r )i
e (r, r), fi (r )i
= im hfj (r), G

(5.57)

= im hfj (r), g(r, r ), fi (r )i

+ i1m h fj (r), g(r, r), fi (r )i
hfj (r), KH (r, r ), fi (r )i
=

m
hf (r),
n j

e (r, r), fi (r )i
G

(5.58)

I, fi (r )i
R
where the inner product is defined as hf(r), X(r)i = s drf(r) X(r) for real
=

m
hf (r), g(r, r)
n j

function f(r). Also fj (r) is the j-th testing function and fi (r ) is the i-th basis
function, g(r, r) is the scalar homogeneous medium Greens function shown
in (5.8).

75

5.3.2 Matrix Representation of Secondary Term

For the secondary term, the matrix representation of the LE operator is [21]
hfj (r), LE (r, r), fi (r )i
= im hfjs (r), ge,ss(r, r ), fis (r )i
+ im h
z fj (r), ge,zz (r, r), z fi (r )i

(5.59)

+ im h
z fj (r), ge,zd(r, r ), fi (r )i
+ im h fj (r), ge,dz (r, r ), z fi (r )i
+ im h fj (r), ge,dd(r, r ), fi (r )i
where fs =
z z f is the horizontal projection of the basis function and
ge,ss

i
= S0
2

F TE
kmz




1
i
TE
2
TM
ge,zz = S0 z z F + kmn F
2
kmz k2



i
n
1
TE
TM
ge,zd = S0
z F
z F
2
m
kmz k2



i
m
1
TE
TM
z F z F
ge,dz = S0
2
n
kmz k2



i
z z TM
1
TE
ge,dd = S0
F + 2 F
.
2
knm
kmz k2

(5.60)

(5.61)
(5.62)
(5.63)
(5.64)

One should note that for RWG basis function straddling the interface of two
adjacent layers, the coefficient m should be modified accordingly. Since in
real implementation the matrix element is accounted for by half RWG-half
RWG (triangle-triangle) interactions, the expression in (5.59) is convenient
and shall not cause confusion. The derivation details can be found in [21]
and will not be repeated here.
We pay more attention to the matrix representation of the KH operator
e here. For the TE wave, from (5.40), (5.42), we
with the kernel of G
have
TE (r, r) = ( z + k 2 z)( z)g TE (r, r )
G
(5.65)
e
n
76

It can be divided into two parts and the matrix representation can be obtained by applying integration by parts,
TE (r, r), fi (r )i
hfj (r), G
e
= h fj (r), z zg TE (r, r), fi (r )i
+

(5.66)

dlfj (r) n
hz zg TE (r, r ), fi (r )i
c

+ h
z fj (r), kn2 zg TE (r, r ), fi (r )i
Similarly from (5.41), we have


TM (r, r) = k 2 ( z) z + k 2 z g TM (r, r )
G
e
n
m

(5.67)

then the matrix representation is

TM (r, r ), fi (r )i
hfj (r), G
e
= hfj (r), kn2 zz g TM (r, r ), fi (r )i
+ hfj (r),

dl kn2
c

zz g

TM

(5.68)

(r, r )fi (r ) n
i

2
+ hfj (r), kn2 km
zg TM (r, r ), z fi (r )i

The curl operator can be approximated by the finite difference method in

spatial domain. For accurate evaluation, the curl can also be implemented
in the spectral domain, as is done in Section 5.2. In summary, from (5.4),

77

(5.22), (5.66), and (5.68), the matrix representation of the KH operator is

hfj (r), KH (r, r), fi (r )i

where

m
h
n

m
n

m
h
z fj (r), gce,zs(r, r), fis (r )i
n

m
hf (r), gce,sd(r, r ),
n js

m
hf (r), c
n js

m
hf (r), gce,sz (r, r), z fi (r )i
n js

fj (r), gce,ds (r, r), fis (r )i

dlfj (r) n
hgce,ds (r, r), fis (r )i

fi (r )i

dl gce,sd(r, r )fi (r ) n
i

#


sin i
z F TE
S1
gce,ds (r, r ) =
2
kmz k2
cos
"
#
 TE 
sin

i
F

2
S1
gce,zs (r, r ) = kn
kmz k2
cos 2
"
#


sin

i
z F TM
n

S1
gce,sd(r, r ) =
m cos 2
kmz k2
"
#
 TM 

sin

i
F

2
gce,sz (r, r ) = kmn
S1
2
kmz k2
cos

(5.69)

"

(5.70)

(5.71)

(5.72)

(5.73)

where gce is a 2D vector (x and y components).

5.3.3 Line Integral Issue

Although the line integral is absent in MPIE in homogeneous environment
(namely only the matrix representation of the primary term is involved), it
is always an issue in layered medium when the object is penetrating more
than one layer and the RWG basis defined on two adjacent triangles has
to straddle the interface of two layers. However, if the kernel is continuous
across the interface, such line integral from positive and negative triangles of
a RWG basis function can be canceled.
In fact, this is usually the case. The continuity of the kernel can be verified
78

Table 5.1: Matrix element of KE : no line integral (103 )

Z++

Z+

Z+

Matrix (real)

9.662

1.080

Matrix (imag)

0.623

5.147

Dyadic (real)
Dyadic (imag)

0.189

4.348

1.027 0.756 0.939

1.118 0.558 0.577

9.133 3.538
0.967

2.568

0.816 3.538
0.315 2.568

Table 5.2: Matrix element of KE : testing line integral (103 )

Z++

Z+

Z+

Matrix (real)

2.870

2.103

3.868

1.436

2.015

Matrix (imag)

1.521

Dyadic (real)

0.911

0.437

0.547

0.119

2.015

Dyadic (imag)

4.274 0.712

0.054 0.353 0.512

0.185 1.303
0.385 1.303

by applying the continuity of the propagation factor F (z, z ) in (5.16). When

the observation point r comes across the interface, we have [5]
F1 (z+ , z ) = F2 (z , z )

(5.74)

p1
1 (z+ )z F1 (z+ , z ) = p2 (z )z F2 (z , z )

(5.75)

where p = for TE wave and p = for TM wave. To analyze the source

point r , the symmetry relation of the F (z, z ) can be applied [112]
p(z )
p(z)
F (z, z ) =
F (z , z).

kz (z )
kz (z)

(5.76)

Hence we have the following relations

p1 (z+
)
p2 (z
)

F
(z,
z
)
=
F (z, z
)
1
+

k1z (z+ )
k2z (z ) 2

(5.77)

k1z
(z+ )z F1 (z, z+
) = k2z
(z )z F2 (z, z
)

(5.78)

The continuity of the kernel of the LE in (5.59) is shown in [109]. The one of
KH in (5.69) can be verified similarly since the horizontal partial derivatives
(( ) z) in (5.66) and (5.68) do not affect the vertical continuity due to the
phase matching condition.
79

Figure 5.5: Cases where testing line integral exists. The testing function is
straddling the interface, all radiation from the RWG or half-RWG basis
functions (triangles) in color needs invoking testing line integral, while
radiation from others does not need this line integral.

Figure 5.6: Line integral test. The testing function is at the top interface.
Radiation from: basis function 1: no line integral activated; basis function
2: testing line integral activated; basis function 3: both testing and basis
line integrals activated. (unit: m).

80

Table 5.3: Matrix element of KE : testing & basis line integrals (104 )
Matrix (real)
Matrix (imag)
Dyadic (real)
Dyadic (imag)

Z++

Z+

1.406

0.063

0.680

1.406
0.680

2.765
0.063

2.765

Z+

0.697 1.123 0.917

3.379

0.131

1.426

0.697 1.123 0.917

3.379

0.131

1.426

5.3.4 Consistency Analysis

However, the propagation factor shown in (5.74)(5.78) contains all fields
information (including the primary field if necessary), the continuity is valid
only when all the fields are expressed as the Sommerfeld integral of the
propagation factor. Since we have separated the primary term in closed-form
to avoid ill-convergent self-term Sommerfeld integral, and also to gain the
convenience of singularity treatment, we have to make sure that the matrix
representation of the two terms are consistent before we can surely remove
the line integral. By setting the basis function and the testing function in
different layers, and then making the layers homogeneous, we can easily test
the consistency. The secondary term is consistent with the primary term in
the matrix representation of LE [90], hence there is always no line integral

in (5.59). However, for the KH , we find that the secondary term is not
consistent with the primary term. In the primary term, the operator
acts directly on the Greens function in (5.58), while in the secondary term,
it is transferred to either testing or basis function in (5.69). This operator
transfer reduces the singularity of the Greens function; however, it also
destroys the continuity of the propagation factor and leads to non-canceled
line integral for interactions where the primary term is involved. For the
testing line integral (the first line integral in (5.69)), it needs to be calculated
only when the testing RWG function straddles the interface and the basis
RWG functions or half-RWGs (either positive or negative triangles) are in
the same two layers shown in Figure 5.5. The necessity of basis line integral
(the secondary line integral in (5.69)) can be analyzed similarly. Since we are
filling the matrix elements by triangle-triangle pairs other than RWG-RWG
pairs, the line integral can be easily judged and added if necessary; and also
the number of Sommerfeld integrals can be reduced to minimum.

81

To validate the matrix representation of KE , we consider the interaction

between one testing function and several basis functions shown in Figure
5.6, the representative basis functions are picked from a mesh of a sphere
embedded in a four-layer medium. The working frequency is f = 300 MHz.
The results are validated by applying the dyadic form Greens function. The
testing function straddles the interface, and three different basis functions are
considered. Table 5.1 shows the radiation from basis 1, where no primary
field is involved. Hence, no line integral is activated. We can observe that
though the interactions between triangle-triangle pairs differ due to the line
integral, it can be canceled and the final matrix element is the same. Table
5.2 shows the radiation from basis 2, in this case, the testing line integral
should be activated, and again the final matrix element agrees. Finally,
the radiation from basis 3 is calculated and shown in Table 5.3. Now both
testing line integral and basis line integral need to be activated and the matrix
representation exactly recovers the dyadic form.

5.4 Duality Principle of LMGF

m and G
m in (5.5) and (5.6) can
The magnetic type Greens function G

be obtained from the duality principle. One form of the duality principle
reads [5],
E H, H E, ,
(5.79)
M J, J M, m , m

5.4.1 Primary Term of the Greens Function

Since the scalar homogeneous medium Greens function in (5.8) satisfies the
scalar Helmholtz equation
2
(2 + km
)g(r, r) = (r, r )

(5.80)

gm (r, r) = ge (r, r) = g(r, r)

(5.81)

we can easily show that

82

Hence




Gm (r, r ) = Ge (r, r ) = I + 2
g(r, r)
km

(5.82)

Note that the m is the layer index in km while it is magnetic-type Greens

m.
function in gm or G

5.4.2 Secondary Term of the Greens Function

The propagation factor F (z, z ) of the secondary term satisfies the following
differential equation [1]



d 1 d
1 2
+
k (z) F (z, z ) = 0
dz p(z) dz p(z) z

(5.83)

It is obvious that
F TE (z, z ) F TM (z, z ), F TM (z, z ) F TE (z, z )

(5.84)

g TE (r, r) g TM (r, r), g TM (r, r ) g TE (r, r)

(5.85)

hence,

Finally, we have
TE (r, r )
m (r, r ) = G
TM (r, r) + 1 G
G
m
2
kmn m

(5.86)

TM (r, r) = ( z)( z)g TM (r, r )

G
m

(5.87)

TE (r, r ) = ( z) ( z) g TE (r, r)
G
m

(5.88)

e (r, r) G
m (r, r), G
m (r, r) G
e (r, r )
G

(5.89)

where

Hence,
e and G
m.
and similar duality holds between G
With these properties, we can now easily verify the correctness of definition
of the LH and KE in (5.5) and (5.6).

83

5.5 Surface Integral Equation

As the LMGF is formulated properly, we can set up the surface integral
equation for general homogeneous objects by applying the extinction theorem
[5]. For the object shown in Figure 5.1, if we put r S and r S + , we

have

o
Eoinc (r) = LoE (r, r ) J(r) + KE
(r, r ) M(r ).

(5.90)

Here we use superscript o to represent that the Greens function is evaluated in the outside region (i will be used for inside region). Similarly, if
r S + , r S ,
i
0 = LiE (r, r) [J(r )] + KE
(r, r ) [M(r )]

(5.91)

where the negative signs of the currents are due to the unique definition of
the unit normal vector in one problem (J = n
H, M = E n
). One should
note that for Li and Ki , the choice of the Greens function has some freedom.

As long as the material and the boundary condition is maintained, the field
can be uniquely determined. The Greens function may even not satisfy the
radiation boundary condition (non-physical). However, for simplicity, we
usually choose the homogeneous medium Greens function.
In practice, we may have internal resonance problems if we only use the
above two E-type equations. To avoid it, the H-type equations are usually
applied too. The H-field equations read
o
Hoinc (r) = LoH (r, r) M(r ) + KH
(r, r) J(r )

(5.92)

i
0 = LiH (r, r) [M(r )] + KH
(r, r) [J(r )]

(5.93)

The popular PMCHWT (Poggio, Miller, Chang, Harrington, Wu, Tsai)

formulation [113] is applied here though other alternatives are available in
homogeneous medium [114]. Finally, the surface integral equation in matrix
notation reads
"
#
"
# "
#
o
i
Eoinc 
(LoE + LiE ) (KE
+ KE
)
J 

 =

o
i
Hoinc tan
(KH
+ KH
) (LoH + LiH )
M tan

84

(5.94)

5.6 Numerical Results

Several numerical results are presented in this section. Figure 5.7 shows a
capsule structure embedded in a 5-layer medium. The parameters of the
layered medium are shown in the figure. The dimension of the object is
h = 0.6 m, r = 0.15 m. A Hertzian dipole is at (x = 0.1, y = 0.1, z = 0.4) m
with polarization of ( = 20o , = 30o ). The working frequency is f = 300
MHz. The object is first set to be transparent for sanity check, namely there
is no contrast between the object and the surrounding layer ( = 2, = 2).
The scattered field at the center line of the object (z = 0) is calculated by the
SIE. In this case, the equivalent electric and magnetic current should generate
exactly the incident field according to the extinction theorem. The result is
shown in Figure 5.8 and agrees well with the theoretical prediction. In fact,
this sanity-check model activates almost all ingredients of the layered medium
Greens function and SIE developed in this chapter. For real non-transparent
objects, only the Greens function inside the object needs to be modified,
which is trivial since it is the homogeneous medium Greens function. Next,
the material of the object is set to be non-transparent ( = 4 and = 1), and
the scattered field due to the same Hertzian dipole is calculated and shown in
Figure 5.9 and Figure 5.10. The observation line is set to be at the interface
(z = d5 = 0.55 m). We can observe that the tangential components of the
E field are continuous while the normal component differs by a factor of 3
(4 /5 ), which satisfy the boundary condition. Finally, a cuboid penetrating
different layers are shown in Figure 5.11. The cuboid is tilted by = 45o
and the dimension is a = 0.6 m, b = c = 0.2 m. It is illuminated by a
TE plane wave with normal incidence (f = 300 MHz). To further validate
the Greens functions and operators for the straddling case, the cuboid is
e and LE involved) and
first filled with PEC material. Hence both EFIE (G
e and KH involved) can be implemented. The scattered coMFIE ( G

polarization fields are evaluated at z = 0.5. The results are shown in Figure
5.12 and good agreement can be observed. Finally, the cuboid is filled with
dielectric material ( = 5 and = 1). The scattered field is again calculated
and shown in Figure 5.13, which is weaker than that of the PEC case.

85

Figure 5.7: A homogeneous capsule structure embedded in a 5-layer

medium, where h = 0.6, r = 0.15 (unit: m).

500
Ex
Ey

Electric field (V/m)

400

Ez
300

200

100

0.3

0.2

0.1
0
0.1
Observation point x (m)

0.2

0.3

Figure 5.8: The scattered field inside the object with = 2 and = 2.
Since there is no contrast, the scattered field recovers the incident field
(solid line).

86

40
Ex

Electric field (V/m)

35

Ey

30
25
20
15
10
5
0
3

1
0
1
Observation point x (m)

Figure 5.9: The scattered field outside the object with = 4 and = 1.
The tangential components of the E field are continuous at the interface.

Symbol: field at d+
5 , solid line: field at d5 .

50
Ez at d+5
Ez at d5

Electric field (V/m)

40

30

20

10

0
3

1
0
1
Observation point x (m)

Figure 5.10: The scattered field outside the object with = 4 and = 1.
The normal component of the E field differs by a factor of 3 (4 /5 ) at the
interface.

87

Figure 5.11: A homogeneous cuboid embedded in a 4-layer medium (unit:

m). It is penetrating different layers.

0.25
EFIE
MFIE

Electric field (V/m)

0.2

0.15

0.1

0.05

0
3

1
0
1
Observation point x (m)

Figure 5.12: The scattered field of the PEC object, calculated from EFIE
and MFIE.

88

0.1
Dielectric object

Electric field (V/m)

0.08

0.06

0.04

0.02

0
3

1
0
1
Observation point x (m)

Figure 5.13: The scattered field of the dielectric object.

5.7 Summary
A new Greens function formulation for modeling homogeneous objects in
layered medium is systematically developed in this chapter. The dyadic
form of the Greens function is developed by applying the pilot vector potential approach, and all components are derived in terms of order 0 and
order 1 Sommerfeld integrals. The matrix representation is further derived
by applying integration by parts and vector identities. The line integral issue for interactions between straddling basis functions is discussed in details
for different situations, based on the continuity property of the propagation
factor and the consistency analysis of the primary term and the secondary
term. The duality principle of the LMGF is further discussed to make the
formulation succinct. Finally, the extinction theorem is revisited in inhomogeneous environment and the relevant integral equations are set up. Several
numerical results are presented to validate this formulation.

89

CHAPTER 6
DCIM ACCELERATED SURFACE
INTEGRAL EQUATION (SIE) METHOD
FOR NANO-OPTICAL APPLICATIONS
The SIE involving LMGF developed in Chapter 5 is accelerated by the DCIM
introduced in Chapter 2 and is applied to nano-optical applications. We
only focus on the situation where the nano scatterers are embedded in a
single layer, which is common in various optical applications. Some issues in
incorporating DCIM to SIE in this case will be discussed. The basic theory
of spontaneous emission (SE) of a quantum emitter and the surface plasmon
resonance will also be briefly reviewed. SE is then studied numerically for an
emitter embedded in a layered medium and in the presence of nano scatterers,
where the emitter is coupled to the surface plasmon polaritons (SPPs) and
the localized surface plasmons (LSP).

6.1 Introduction
When atoms are prepared in the excited states, they spontaneously emit light
even in the vacuum. This interesting phenomenon attracted much attention
after the development of quantum mechanics in 1920s [116]. Purcell in 1946
first demonstrated that the SE of a quantum emitter is not an intrinsic
property of the emitter but can be modified when it is located in a cavity,
now known as the Purcell effect [117]. SE can be explained by the atom-field
interaction, after the quantization of the electromagnetic (EM) field. The
vacuum fluctuation of the EM field perturbs the atom system and causes the
spontaneous emission of photons. If the local electromagnetic environment
is changed, which leads to the change of local density of states (LDOS), the
emission can be modified. Purcells work stimulated relevant studies, led to
the development of laser (firstly in microwave spectrum, known as maser),
and finally started the subject of cavity quantum electrodynamics (Cavity
QED) [118]. In the so-called weak coupling regime, the SE enhancement and
90

inhibition of an atom can be controlled by local EM environment such as a

cavity, which now plays an important role in a variety of novel optoelectronic
designs such as light-emitting diode (LED), laser, solar cell, etc. Recently,
the photonic crystal (PC) [119] or surface plasmon polaritons (SPPs) [120]
other than the cavity mode are applied to control SE [121], [122]. In the
strong coupling regime, the entanglement can be manipulated, which can be
used for quantum information.
We will focus on SE in the weak coupling regime. The spontaneous emission rate (SER) can be calculated by the Fermis golden rule [11], [116], [123],
it can be further related to the LDOS after some mathematical derivation.
The LDOS counts the number of EM modes where photons can be emitted at
the specific location of the emitter. By using the Greens function approach,
the LDOS can be easily obtained. Hence, this quantum electrodynamic problem can finally be cast into a classical EM scattering problem.
In this chapter, the SIE developed previously will be applied to analyze
SE enhancement of an emitter embedded in a layered medium with metal
layers and in the presence of metallic nano scatterers, where the emitter can
be coupled to the SPPs formed by the dielectric-metal interface and the LSP
formed by the nano scatterers. It is observed that SE can further be enhanced
by the coupling of the two resonances. For the sake of completeness, we will
discuss some special issues in applying DCIM to SIE, introduce the Greens
function approach in calculating SE and briefly review the surface plasmon
resonance.

6.2 SIE with DCIM Acceleration

The broadband simulation in visible spectrum (400 nm800 nm) requires
efficient computation of single frequency response. The direct evaluation of
Sommerfeld integrals in SIE makes the matrix filling extremely slow. DCIM
introduced in Chapter 2 will be applied to accelerate the calculation. The
near field information is important in the simulation since the nano scatterers
are usually in the sub-wavelength scale. Hence, the traditional DCIM [22]
will be applied here since it preserves good accuracy in the near field region.
In the following, we will only focus on the special features in combining DCIM
to SIE. The main idea of DCIM can be found in Chapter 2 and will not be
91

repeated here. We will also focus our attention on the secondary field since
the primary field has closed-form solution.

6.2.1 Position Independent DCIM

Since the nano scatterers are confined in a single layer, the DCIM can be
made position (z, z ) independent [115]. From (2.3), the propagation factor
of the secondary field can be written as follows when n = m,
F (z, z ) = F1 (z, z ) + F2 (z, z ) + F3 (z, z ) + F4 (z, z )

(6.1)

mR
m,m1 eikmz z1
F1 (z, z ) = M

(6.2)

mR
m,m+1 eikmz z2
F2 (z, z ) = M

(6.3)

mR
m,m+1 R
m,m1 eikmz z3
F3 (z, z ) = M

(6.4)

mR
m,m1 R
m,m+1 eikmz z4
F4 (z, z ) = M

(6.5)

z1 = 2dm + z + z

(6.6)

z2 = 2dm+1 z z

(6.7)

z3 = 2dm+1 2dm + z z

(6.8)

z4 = 2dm+1 2dm z + z

(6.9)

where

and

The position dependent factor is already in the form of exponentials and can
be extracted during the function fitting in DCIM process, if we separate the
propagation factor into four parts, as are shown in (6.2)(6.5). Once the
DCIM coefficients ai and bi are obtained, the position-dependent factor can
be easily added back to the Greens function; hence, (2.23) becomes
g() =

Mj
4 X
X
j=1 i=1

aji

eikm rji
, rji =
4rji

92

2 + (bi + zj )2

(6.10)

6.2.2 DCIM via Order 1 Sommerfeld Identity

In the LE operator, the ordinary (order 0) Sommerfeld identity [5] can be
applied, we show it again here for clarity.
eikm r
=i
r

dk

k
J0 (k )eikmz |z|
kmz

(6.11)

However, due to the horizontal partial derivatives (x , y , etc), order 1

Sommerfeld integrals occur in the KH operator when they are calculated in
the spectral domain. We will apply the order 1 Sommerfeld identity [115] in
this case.

eikm r
(1 ikm r) = i
r3

k2
dk
J1 (k )eikmz |z|
kmz

(6.12)

Alternatively, we can also use the order 0 identity in this operator and calculate the horizontal partial derivatives in the spatial domain. For example,
x

eikm r
eikm r
= 3 (ikm r 1)x
r
r

(6.13)

The two can be shown to be equivalent to each other.

6.2.3 Poles in Matrix Representation of KH Operator

In the matrix representation of the LE operator, except for ge,ss, there seems
to be poles associated with the TE and TM waves in the integrand in
(5.61)(5.64) at k = 0. Take ge,zz for example,
o


2
z z F TE + kmn
F TM kmz1 k2



2 F TM
R +
(z z F TE +kmn
)
k
i
= 4 0 dk kmz J0 (k )
k2

ge,zz =

i
S
2 0

(6.14)

the function in the spectral domain that needs to be approximated is

2
z z F TE + kmn
F TM
g(k ) =
k2

(6.15)

Because the TE wave is always accompanied by the TM wave, the poles

can be canceled and this function is regular at k = 0.
In the matrix representation of the KH operator, however, TE wave and
93

TM wave work separately. Take the gce,ds for example,

gce,ds (r, r ) =

"

sin
cos

"

sin
cos

i
S
2 1

i
4

z F TE
kmz k2

R +
0

(6.16)

k2
dk kmz
J1 (k)

F TE
k2

the function to be approximated by DCIM is

g(k ) =

z F TE
k2

(6.17)

It can be shown that though the integrand in (6.16) is regular at k = 0 (there

is also a k2 in the numerator that can cancel the one in the denominator),
the function in (6.17) in DCIM is unfortunately singular at k = 0.
Note that the matrix representation is derived to reduce the singularity of
the LMGF, which means that the Sommerfeld integral converges more rapidly
compared with the dyadic LMGF in direct numerical evaluation. In DCIM,
however, the slow convergence of the dyadic Greens function can also be
remedied by techniques such as the two-level approximation [35], where the
sampling path can be made longer along the real axis with sparser sampling
rate. By investigating the expressions of the dyadic LMGF in (5.44)(5.49)
and (5.51)(5.56) carefully, we found that only two Sommerfeld integrals have
possible pole problems in DCIM
S1

z F TE
kmz k2

S1

z F TM
kmz k2

(6.18)

Fortunately, the two always appear in pairs, hence the poles can be canceled in a similar way as in the matrix representation of the LE operator. In
a word, when using DCIM in KH , we choose the dyadic LMGF other than
the matrix representation.

94

6.3 Greens Function Approach in Spontaneous

Emission (SE)
6.3.1 Spontaneous Emission Rate (SER)
The emitter is assumed to be a two-level system, located at r = r0 . The
transition process is shown in Figure 6.1. The initial state of the atomfield system is |ii = |e, 0i, where the atom is at excited state, and the EM
field is at the 0-photon state; the final state is |f i = |g, 1kii after transition,
where the atom jumps to the ground state and the EM field is now at the
1-photon state (ki mean different degenerate EM modes with frequency )
The SER of this system is determined by the Fermis golden rule [123],
which can be derived from the perturbation method [116], [11]
=

2 X
I |ii|2(i f )
|hf |H
2
h
f

(6.19)

The interaction Hamiltonian reads

I =

H
pE

(6.20)

is the dipole moment operator, which can be expressed as

where p
= er = p (|gihe| + |eihg|)
p

(6.21)

p = hg|
p|ei = he|
p|gi

(6.22)

and
has the form of
The electric field operator E
=
E

i
Xh
+

Ek a
k (t) + Ek a
k (t)

(6.23)

where
a
k (t) = a
k (0)eik t

(6.24)

a
k (t) = ak (0)eik t

(6.25)

95

Figure 6.1: Transition of a two-level atomic system.

Here, ak (0) and ak (0) are the annihilation and creation operators. After
substituting all these expressions into (6.19), we have
=

0 X
[p (uk uk ) p] (k 0 )
h
0 k

(6.26)

where the normalized EM eigen modes are

E+
k

h
k
uk ,
20

E
k

h
k
u
20 k

(6.27)

The SER can be further expressed by the partial LDOS p

=

0 2
|p| p (r0 , 0 )
3h0

(6.28)

with
p (r0 , 0 ) = 3

X
k

where p = pnp .

[np (uk uk ) np ] (k 0 )

(6.29)

6.3.2 Greens Function Approach

The eigen modes satisfy the Helmholtz equation.
uk (r, k )
96

k2
uk (r, k ) = 0
c2

(6.30)

The electric-type Greens function defined in (5.1), (5.3), satisfies the same
equation with excitation of delta function.
2
r , ) G(r,
r , ) = I(r r )
G(r,
c2

(6.31)

According to the eigen-modes expansion, we have

r , ) =
G(r,

ck (r , )uk (r, k )

(6.32)

The expansion coefficients can be easily determined by the orthonormal property of the eigen modes and hence we have
r, ) =
G(r,

c2

uk (r , k )uk (r, k )
k2 2

(6.33)

After some mathematical derivation, we can further obtain the following

relationship.
2 X


r, ) = c
Im G(r,
uk (r, k )uk (r, k )( k )
2 k

(6.34)

Finally, the partial LDOS in (6.29) can be expressed by the the imaginary
part of the Greens function as
p (r0 , 0) =



60
0 , r0 , ) np }
{n

Im
G(r
p
c2

(6.35)

The LDOS counts the number of electromagnetic modes at specific position

by definition, based on which the density of states (DOS) can be obtained
by integrating over space.
(r0 , 0 ) =

X
k

|uk |2 (k 0 )

(6.36)

If the emitter has no fixed dipole axis, the emission rate is obtained by
averaging over all possible directions. In this case, the partial LDOS becomes
the same as the LDOS.
(r0 , 0) =



20
0 , r0 , 0 ) }
Im{Tr G(r
2
c
97

(6.37)

where Tr means trace. Alternatively, the SER can also be derived from
the fluctuation-dissipation theorem [124], [125], which will not be discussed
here.
2
(r, )E
(r , )i = h
hE
Im{G (r, r , )}
(6.38)
c2 0
Specifically, in vacuum (free space), the imaginary part of the trace of the
dyadic Greens function shown in (5.7) is


1
0 (r0 , r0 , 0 ) } = k0 = 0 .
Im{Tr G
3
6
6c

(6.39)

Hence, the SER in free space is

0 =

03 |p|2
30h
c3

(6.40)

which is identical with the one obtained by Einstein from Boltzmann statistics and Plancks formula in the black body radiation [116].
To quantify the SE enhancement in an inhomogeneous environment, it is
convenient to define the normalized SER or Purcell factor as


0 , r0 , 0 ) }
Im{Tr G(r

(r0 , 0 )


=
=
(6.41)
0 (r0 , r0 , 0 ) }
0
0 (r0 , 0)
Im{Tr G
In summary, the local environment changes the eigen modes of the EM

field, and thus changes the partial LDOS or LDOS. By calculating the Greens
function, and making the source point and the observation point identical
(r = r = r0 ), one can obtained the partial LDOS and finally the SER, as
are shown in (6.28) and (6.35).

6.4 Surface Plasmon Polaritons (SPPs) and Localized

Surface Plasmons (LSP)
6.4.1 SPPs at Dielectric-Metal Interface
As has been mentioned, noble metals (Au, Ag, Cu, etc) are good conductors
at microwave or far-infrared spectrum; however, as the frequency goes to
the infrared or visible region, the skin depth increases considerably. Strong
98

Figure 6.2: SPPs can be sustained in the interface of dielectric-metal half

space.
absorption occurs due to transitions between electronic bands. The dielectric
constant (electric permittivity) of metals can be described by either Drude
model, which is based on the free electron gas, or Lorentz model, which
also adds a term of restoring force. The dielectric constant of noble metals
has a negative real part in optics, which leads to a variety of interesting
phenomena such as the surface plasmon polaritons (SPPs) and the localized
surface plasmons (LSP).
The SPPs arises from the coupling of EM fields to the oscillations of the
metals electron plasma. The EM field propagates at the interface of a half
space (2-layer medium) formed by a normal non-absorbing dielectric with 2
and a metal with Re [1 ()] < 0. The field is evanescently confined in the
vertical direction, shown in Figure 6.2. Notice that only the TM wave can
support the SPPs since the TE wave cannot satisfy the proper boundary
condition. If is the propagation constant along the interface (x direction),
we have the following dispersion relationship [120]
= k0

1 2
1 + 2

(6.42)

This leads to the dispersion curve of the SPPs lying to the right of the light
line. When the frequency tends to the surface plasmon frequency,
sp =

p
1 + 2

(6.43)

and vg 0 in lossless case, where p is the plasma frequency of the

99

Figure 6.3: LSP can be excited by a metallic nano sphere under an EM

field excitation.
bulk metal. In this limit, we have
1 () + 2 = 0

(6.44)

6.4.2 Localized Surface Plasmons (LSP) of Metallic Nano

Particle
LSP comes from the coupling of EM field and oscillation of conduction electrons of metallic nano particles. It leads to strong field amplification both
inside and in the near-field region outside the particle, and hence are now
applied in various optical devices. Since the particle is typically much smaller
than the wavelength of light, LSP can be analyzed simply from the quasistatic approximation [120]. For a nano sphere with dielectric constant 1 ()
and radius a located in the surrounding medium with 2 shown in Figure 6.3,
it responds to the external EM field as a dipole and the polarizability can be
obtained as
= 4a3

1 () 2
1 () + 22

(6.45)

The resonance of happens when

1 () + 22 = 0
which is different from the resonance condition of SPPs in (6.44).

100

(6.46)

6.5 Quantum Emitter Coupled to Surface Plasmon

Resonance
Surface plasmon resonance is recently applied to enhance SE in optoelectronic
devices, for example in the quantum well lasers [126], [127]. The improved
photoluminescence and low threshold current can be achieved due to the
strong optical confinement in surface plasmon resonance. However, most
theoretical works employed approximate methods to model this problem and
convincing results can hardly be found in the literature. In this chapter, we
will apply the full-wave 3D electromagnetic surface-integral-equation solver
based on the layered medium Greens function, which is developed previously,
to study this problem.
As has been mentioned, the computation of SER of a quantum emitter in
a plasmonic system can be cast into a classical EM scattering problem. More
specifically, we put a Hertzian dipole to excite the structure, and calculate
the scattered field at the same location. Once the total field is obtained
numerically, the Greens function (of the whole hybrid structure) can be
deduced accordingly (actually the imaginary part of the Greens function
since the real part is singluar).
From numerical point of view, there are various methods available in computational electromagnetics (CEM) in solving the aforementioned classical
scattering problem. For example, the finite element method (FEM) can be
applied to obtain the Greens function. However, since the whole structure
has to be discretized by tetrahedrons, the number of unknowns is thus huge,
especially when considering that the substrates are typically large. For localized plasmons, very fine mesh is required around the nano scatterers to
capture the highly concentrated and localized field, multiscale discretization
is hence required in FEM and the condition of the matrix is bad. Also, if
the number of layers increases in certain optimizations, the mesh has to be
regenerated and the number of unknowns increases accordingly. However, for
the scheme developed here, such problems no longer existonly the surface
of the nano-scatterers needs to be discretized, and the increase of the number
of layers does not add extra modeling complexity.
The first example is shown in Figure 6.4, where a gold nano sphere with
radius 20 nm is located above a gold slab with thickness 30 nm. The zpolarized emitter stays in between and the distances to the slab and the
101

Table 6.1: CPU time for single frequency calculation

Numerical

DCIM

Matrix filling

1.5 days

9.2 minutes

Excitation

2.1 minutes

1 second

Scattered field

2.1 minutes

1 second

sphere are both 10 nm. We scan the visible spectrum from 400 nm to 800 nm,
with the step of 10 nm. The dielectric constant of the gold is obtained from
a more accurate Brendel-Bormann model [128], which satisfies the KramersKronig relation. We consider the following three cases: (a) SE of an emitter
in the presence of the slab; (b) SE of an emitter in the presence of the
nano sphere; (c) SE of an emitter in the presence of both structures. The
normalized SER are calculated and shown in Figure 6.5. For case a, the SER
of the emitter is enhanced by the SPPs of the slab, while for case b, the
SER is enhanced by the LSP of the nano sphere. In vacuum, the resonant
frequency of SPPs and LSP are similar and around 510 nm. For case c, the
SER is strongly enhanced by the coupling of SPPs and LSP at the same
resonant frequency.
For the nano sphere here, a reasonably fine surface mesh can be obtained
by using 1620 triangular elements. The number of unknowns (electric and
magnetic current) is 4860, hence it only requires around 180 MB to store the
matrix if single precision is used. The CPU times for matrix filling, RHS
vector (excitation) filling and scattered field calculation at a single frequency
(wavelength) point are also listed in Table 6.1 for further comparison. This
simulation is run on a laptop computer with Intel 2.00 GHz processor. It is
observed that the computation can be much accelerated by using DCIM.
Next, we consider a nano bowl structure embedded in a layered medium
with two gold layers shown in Figure 6.6. The mesh of the nano bowl is
shown in Figure 6.7. The normalized SER is calculated and shown in Figure
6.8. Two sub-figures are also demonstrated to show the respective effects
from the layered medium and the nano bowl.

102

Figure 6.4: A gold nano sphere with radius 20 nm is located above a gold
slab with thickness 30 nm. A z-polarized emitter is located at the middle of
the sphere and slab.

350

Normalized SER

300

Slab
Sphere
Both

250
200
150
100
50
0
400

450

500

550
600
650
Wavelength (nm)

700

750

800

Figure 6.5: Normalized SER for the three cases: (a) only slab (SPPs
enhanced SE); (b) only nano sphere (LSP enhanced SE); (c) both (both
effects).

103

Figure 6.6: A gold nano bowl is located in a layered medium with two gold
layers. A z-polarized emitter is located at the center of the aperture. The
dimensions are shown in the figure.

80
z (nm)

60
40
20
0
100
50
x (nm)0

50

50
50

100

0 y (nm)

Figure 6.7: Mesh of the gold nano bowl.

104

Figure 6.8: Normalized SER. The first peak corresponds to the peak from
SPPs of the layered medium at 520 nm, the second peak is from the nano
bowl effect around 590 nm.

6.6 Summary
DCIM is incorporated into the SIE formulation and the developed 3D full
wave EM solver is applied to study SE enhancement of a quantum emitter
in plasmonic structures. The special features of combination of DCIM and
SIE are discussed in details for the case when the scatterers are confined in a
single layer. The fundamental principle and computational technique of SE is
reviewed, which is based on the Fermis golden rule and the Greens function
approach. SPPs of a air-metal interface and LSP of a metallic nano particle
are also discussed for completeness. Finally, several numerical results are
demonstrated to show the SE enhancement of a quantum emitter coupled to
SPPs and LSP. It is observed that the SE can further be enhanced from the
coupling of the two plasmonic resonances.

105

CHAPTER 7
CONCLUSION

Surface integral equation (SIE) method involving the layered medium Greens
function (LMGF) is important in analyzing electromagnetic radiation and
scattering in planarly layered structures, such as the microstrip antenna,
integrated circuit, organic light-emitting diode (OLED), and optical forces
in microelectromechanical systems (MEMS) etc.
In this thesis, the SIE with LMGF is investigated systematically. The acceleration technique for evaluation of Greens function is first studied. The
popular discrete complex image method (DCIM) is reviewed and a novel implementation scheme based on Sommerfeld branch cut is proposed. Next,
a broadband fast integral equation solver, the mixed-form thin-stratified
medium fast-multipole algorithm (MF-TSM-FMA), is developed to expedite the matrix-vector product in the iterative solvers. After that, the lowfrequency breakdown of the integral operator is discussed and the loop-tree
decomposition and the augmented electric field integral equation (A-EFIE)
are studied. In order to model general homogeneous objects in layered
medium, a new formulation of the LMGF based on the pilot vector potential
approach is developed and the corresponding SIE is then obtained. Finally,
attempt is made to apply the developed full wave three-dimensional (3D)
electromagnetic solver to nano-optics, and the spontaneous emission (SE) of
a quantum emitter located in a plasmonic system is studied.
Future work can be done to incorporate the generalized impedance boundary condition (GIBC) to model the finite conductance in circuit simulation.
The finite element method (FEM) can also be combined with the developed
SIE to model inhomogeneous objects in under-ground detection. For objects
with arbitrary penetration such as in optical lithography, DCIM can be further combined with the interpolation technique to expedite matrix filling and
fast integral equation solvers such as the kernel-independent algorithm can
be further developed for large-scale problems.
106

REFERENCES

[1] W. C. Chew, M. S. Tong, and B. Hu, Integral Equations for Electromagnetic and Elastic Waves, Morgan & Claypool Publishers, 2009.
[2] A. Taflove and S. C. Hagness, Computational Electrodynamics : the
Finite-Difference Time-Domain Method, 2nd ed. Boston, MA: Artech
House, 2000.
[3] J. Jin, The Finite Element Method in Electromagnetics, 2nd ed. New
York: Wiley, 2002.
[4] R. F. Harrington, Field Computation by Moment Methods, New York:
IEEE Press, 1993.
[5] W. C. Chew, Waves and Fields in Inhomogeneous Media, Van Nostrand
Reinhold, 1990, Reprinted by IEEE Press, 1995.

[6] A. Sommerfeld, Uber

die ausbreitung der wellen in der drahtlosen
telegraphie, Ann. Phys., vol. 4, no. 28, pp. 665-736, Jan. 1909.
[7] G. H. Golub, and C. F. Van Loan, Matrix Computations, 3rd ed. Baltimore: Johns Hopkins University Press, 1996.
[8] Y. Saad and M. H. Schultz, GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems, SIAM J. Sci.
Statist. Comput., vol. 7, no. 3, pp. 856-869, July 1986.
[9] A. F. Peterson, S. L. Ray and R. Mittra. Computational Methods for
Electromagnetics, New York: IEEE Press, 1998.
[10] J. D. Jackson, Classical Electrodynamics, 2nd ed. New York: Wiley,
1975.
[11] C. Gerry, and P. Knight, Introductory Quantum Optics, Cambridge
University Press, 2005.
[12] C. T. Tai, Dyadic Greens Functions in Electromagnetic Theory. Scranton, PA: International Textbook, 1971.

107

[13] L. B. Felsen and N. Marcuvitz, Radiation and Scattering of Waves.

Englewood Cliffs, NJ: Prentice-Hall, 1973.
[14] J. A. Kong, Theory of Electromagnetic Waves, New York: Wiley Intersci., 1975.
[15] L. Tsang, J. A. Kong, and R. T. Shin, Theory of Microwave Remote
Sensing. New York: Wiley Intersci., 1985.
[16] K. A. Michalski and D. Zheng, Electromagnetic scattering and radiation by surfaces of arbitrary shape in layered media, part I: Theory,
IEEE Trans. Antennas Propagat., vol. 38, no. 3, pp. 335-344, Mar.
1990.
[17] K. A. Michalski, and J. R. Mosig, Multilayered media Greens functions in integral equation formulations, IEEE Trans. Antennas Propagat., vol. 45, no. 3, pp. 508-519, Mar. 1997.
[18] P. Yla-Oijala, M. Taskinen, and J. Sarvas, Multilayered media Greens
functions for MPIE with general electric and magnetic sources by the
Hertz potential approach, Progress in Electromagn. Res. PIER 33, pp.
141-165, 2001.
[19] J. A. Kong, Electromagnetic field due to dipole antennas over stratified anisotropic media, Geophysics, vol. 38, pp. 985-996, 1972.
[20] W. C. Chew, J. S. Zhao and T. J. Cui, The layered medium Greens
function a new look, Microw. Opt. Tech. Lett., vol. 31, no. 4, pp.
252-255, 2001.
[21] W. C. Chew, J. L. Xiong and M. A. Saville, A matrix-friendly formulation of layered medium Greens function, IEEE Antennas Wireless
Propagat. Lett., vol. 5, pp. 490-494, 2006.
[22] D. G. Fang, J. J. Yang, and G. Y. Delisle, Discrete image theory for
horizontal electric dipoles in a multilayered medium, IEE PROCEEDINGS, vol. 135, Pt. H, no. 5, pp. 297-303, Oct. 1988.
[23] A. C. Cangellaris, and V. I. Okhmatovski, Novel closed-form Greens
function in shielded planar layered media, IEEE Trans. Microw. Theory Tech., vol. 48, no. 12, pp. 2225-2232, Dec. 2000.
[24] T. J. Cui and W. C. Chew, Fast evaluation of Sommerfeld integrals
for EM scattering and radiation by three-dimentional buried objects,
IEEE Trans. Geoscience Remote Sensing, vol. 37, no. 2, pp. 887-900,
1999.

108

[25] G. J. Burke and E. K. Miller, Modeling antennas near to and penetrating a lossy interface, IEEE Trans. Antennas Propagat., vol. 32,
pp. 1040-1049, 1984.
[26] J. Y. Chen, A. A. Kishk and A. W. Glisson, A 3D Interpolation model
for the calculation of the Sommerfeld integrals to analyze Dielectric
resonators in a multilayered medium, Electromagnetics, vol. 20, pp.
1-15, 2002.
[27] H. Derudder, F. Olyslager, and D. D. Zutter, An efficient series expansion for the 2D Greens function of a microstrip substrate using
perfectly matched layers, IEEE Microw. Guided Wave Lett., vol. 9,
no. 12, pp. 505-507, Dec. 1999.
[28] F. Olyslager and H. Derudder, Series representation of Green dyadics
for layered media using PMLs, IEEE Trans. Antennas Propag., vol.
51, no. 9, pp. 2319-2326, Sept. 2003.
[29] E. Simsek, Q. H. Liu, and B. Wei, Singularity subtraction for evaluation of Greens function for multilayer media, IEEE Trans. Microw.
Theory Tech., vol. 54, no. 1, pp. 216-225, Jan. 2006.
[30] L. Tsang, and B. Wu, Electromagnetic fields of Hertzian dipoles in
layered media of moderate thickness including the effects of all modes,
IEEE Antennas Wireless Propagat. Lett., vol. 6, pp. 316-319, 2007.
[31] B. Wu, and L. Tsang, Fast computation of layered medium Greens
functions of multilayers and lossy media using fast all-modes method
and numerical modified steepest descent path method, IEEE Trans.
Microwave Theory Tech., vol. 56, no. 6, pp. 1446-1454, Jun. 2008.
[32] Y. L. Chow, J. J. Yang, D. G. Fang and G. E. Howard,A closedform spatial Greens function for the thick microstripsubstrate, IEEE
Trans. Microwave Theory Tech., vol. 39, no. 3, pp. 588-592, 1991.
[33] V. I. Okhmatovski and A. C. Cangellaris, Evaluation of layered media
Greens functions via rational function fitting, IEEE Microw. Wireless
Components Lett., vol. 14, no. 1, pp. 22-24, Jan. 2004.
[34] A. G. Polimeridis, T. V. Yioultsis, and T. D. Tsiboukis, A robust
method for the computation of Greens functions in stratified media,
IEEE Trans. Antennas Propagat., vol. 55, no. 7, pp. 1963-1969, Jul.
2007.
[35] M. I. Aksun, A robust approach for the derivation of closed-forrn
Greens functions, IEEE Trans. Microw. Theory Tech., vol. 44, no. 5,
pp. 651-658, May 1996.
109

[36] F. Ling, and J.-M. Jin, Discrete complex image method for Greens
functions of general multilayer media, IEEE Microw. Guided Wave
Lett., vol. 10, no. 10, pp. 400-402, Oct. 2000.
[37] M. Yuan, T. K. Sarkar, and M. Salazar-Palma, A direct discrete complex image method from the closed-form Greens functions in multilayered media, IEEE Trans. Microw. Theory Tech., vol. 54, no. 3, pp.
1025-1032, Mar. 2006.
[38] T. K. Sarkar and O. Pereira, Using the matrix pencil method to estimate the parameters of a sum of complex exponentials, lEEE Antennas Propagat. Magazine, vol. 37, no. 1, pp. 48-55, Feb. 1995.
[39] E. P. Karabulut, A. T. Erdogan, and M. I. Aksun, Discrete complex
image method with spatial error criterion, IEEE Trans. Microw. Theory Tech., vol. 59, no. 4, pp. 793-802, Apr. 2011.
[40] A. Alparslan, M. I. Aksun, and K. A. Michalski, Closed-form Greens
functions in planar layered media for all ranges and materials, IEEE
Trans. Microw. Theory Tech., vol. 58, no. 3, pp. 602-613, Mar. 2010.
[41] F. Mesa, R. R. Boix, and F. Medina, Closed-form expressions of multilayered planar Greens functions that account for the continuous spectrum in the far field, IEEE Trans. Microw. Theory Tech., vol. 56, no.
7, pp. 1601-1614, Jul. 2008.
[42] Y. P. Chen, W. C. Chew, and L. Jiang, A novel implementation of
discrete complex image method for layered medium Greens function,
IEEE Antennas Wireless Propagat. Lett., vol. 10, pp. 419-422, 2011.
[43] B. Hu, and W. C. Chew, Fast inhomogeneous plane wave algorithm for electromagnetic solutions in layered medium structures: Twodimensional case, Radio Sci., vol. 35, no. 1, pp. 31-43, Jan.-Feb. 2000.
[44] V. Rokhlin, Rapid solution of integral equations of classic potential
theory, J. Comput. Phys., vol. 60, no. 2, pp. 187-207, Sept. 1985.
[45] R. Coifman, V. Rokhlin, and S. Wandzura, The fast multipole method
for the wave equation: A pedestrian prescription, IEEE Antennas
Propagat. Mag., vol. 35, pp. 7-12, June 1993.
[46] J. M. Song, C. C. Lu, and W. C. Chew, Multilevel fast multipole
algorithm for electromagnetic scattering by large complex objects,
IEEE Trans. Antennas Propagat., vol. 45, no. 10. pp. 1488-1493, Oct.
1997.

110

[47] 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 Theory Tech., vol. 46, no. 4, pp.
395-403, 1998.
[48] D. V. Ginste, E. Michielssen, F. Olyslager and D. D. Zutter, An efficient perfectly matched layer based multilevel fast multipole algorithm
for large planar microwave structures, IEEE Trans. Antennas Propagat., vol. 54, no. 5, pp. 1538-1548, 2006.
[49] E. Bleszynski, M. Bleszynski, and T. Jaroszewicz, AIM: Adaptive
integral method for sloving large-scale electromagnetic scattering and
radiation problems, Radio Sci., vol. 31, no. 5, pp. 1225-1251, Sept.Oct. 1996.
[50] F. Ling, C. F. Wang and J. M. Jin, An efficient algorithm for analyzing large-scale microstrip structures using adaptive integral euqation
method combined with discrete complex image method, IEEE Trans.
Microwave Theory Tech., vol. 48, no. 5, pp. 832-839, May 2000.
[51] T. K. Sarkar, E. Arvas, and S. M. Rao, Application of FFT and the
conjugate gradient method for the solution of electromagnetic radiation from electrically large and small conducting bodies, IEEE Trans.
Antennas Propagat., vol. 34, no. 5, pp. 635-640, May 1986.
[52] C. F. Wang, F. Ling and J. M. Jin, A fast full-wave analysis of scattering and radiation from large finite arrays of microstrip antennas,
IEEE Trans. Antennas Propagat., vol. 46, no. 10, pp. 1467-1474, Oct.
1998.
[53] J. R. Phillips and J. K. White, A precorrected-FFT method for electrostatic analysis of complicated 3-D structures, IEEE Trans. Comput.
Aided Des. Integr. Circuits Syst., vol. 16, no. 10, pp. 1059-1072, Oct.
1997.
[54] N. Yuan, T. S. Yeo, X.-C. Nie, and L. W. Li, A fast analysis of scattering and radiation of large microstrip antenna arrays, IEEE Trans.
Antennas Propagat., vol. 51, no. 9, pp. 2218-2226, Sept. 2003.
[55] V. Okhmatovski, M. Yuan, I. Jeffrey, and R. Phelps, A threedimensional precorrected FFT algorithm for fast method of moments
solutions of the mixed-potential integral equation in layered media,
IEEE Trans. Microwave Theory Tech., vol. 57, no. 12, pp. 3505-3517,
Dec. 2009.

111

[56] H. G. Wang, C. H. Chan, and L. Tsang, A new multilevel Greens

function interpolation method for large-scale low-frequency EM simulations, IEEE Trans. Comput. Aided Des. Integr. Circuits Syst., vol.
24, no. 9, pp. 1427-1443, Sept. 2005.
[57] L. Ying, G. Biros, D. Zorin, A kernel-independent adaptive fast multipole algorithm in two and three dimensions, J. Comp. Phys., vol.
196, 591-626, 2004.
[58] W. Fong, E. Darve, The black-box fast multipole method, J. Comp.
Phys., vol. 228, pp. 8712-8725, 2009.
[59] J. Parron, J. M. Rius, and J. R. Mosig, Application of the multilevel
matrix decomposition algorithm for the frequency analysis of large microstrip antenna array, IEEE Trans. Magn., vol. 38, no. 2, pp. 721-724,
Mar. 2002.
[60] K. Zhao, M. N. Vouvakis, and J.-F. Lee, The adaptive cross approximation algorithm for accelerated method of moments computations of
EMC problems, IEEE Trans. Electromagnetic Compatibility, vol. 47,
no. 4, pp. 763-773, Nov. 2005.
[61] R. Maaskant, R. Mittra, and A. Tijhuis, Fast analysis of large antenna
arrays using the characteristic basis function method and the adaptive
cross approximation algorithm, IEEE. Trans. Antennas Propag., vol.
56, no. 11, pp. 3440-3451, Nov. 2008.
[62] S. Kapur and D. E. Long, IES3 : A fast integral equation solver for
efficient 3-dimensional extraction, in Proc. ICCAD, pp. 448-455, 1997.
[63] K. Zhao and J.-F. Lee, A single-level dual rank IE-QR algorithm
tomodel large microstrip antenna arrays, IEEE. Trans. Antennas
Propag., vol. 52, no. 10, pp. 2580-2585, Oct. 2004.
[64] W. C. Chew and C. C. Lu, A fast algorithm to compute the wavescattering solution of a large strip, J. Comput. Phys., vol. 107, no. 2,
pp. 378-387, 1993.
[65] L. Gurel and M. I, Aksun, Electromagnetic scattering solution of conducting strips in layeredmedia using the fast multipole method, IEEE
Microwave and Guided Wave Letters, [See also IEEE Microwave and
Wireless Components Letters], vol. 6, no. 8, 1996.
[66] D. V. Ginste, E. Michielssen, F. Olyslager, and D. D. Zutter, A highperformance upgrade of the perfectly matched layer multilevel fast multipole algorithm for large planar microwave structures, IEEE Trans.
Antennas Propagat., vol. 57, no. 6, pp. 1728-1739, Jun. 2009.
112

[67] J. L. Xiong, Y. P. Chen and W. C. Chew, A quasi-3D thin-stratified

medium fast-multipole algorithm for microstrip structures, IEEE
Trans. Antennas Propagat., vol. 59, no. 7, pp. 2578-2587, Jul. 2011.
[68] Y. Zhuang, K. L. Wu, C. Wu and J. Litva, A combined full-wave CGFFT method for rigorous analysis of large microstrip antenna arrays,
IEEE Trans. Antennas Propagat., vol. 44, no. 1, pp. 102-109, 1996.
[69] F. Ling, V. I. Okhmatovski, W. Harris, S. McCracken, and A. Dengi,
Large-scale broad-band parasitic extraction for fast layout verification
of 3-D RF and mixed-signal on-chip structures, IEEE Trans. Microw.
Theory Tech., vol. 53, no. 1, pp. 264-273, Jan. 2005.
[70] L. J. Jiang and W. C. Chew, A mixed-form fast multipole algorithm,
IEEE Trans. Antennas Propagat., vol. 53, no. 12, pp. 4145-4156, 2005.
[71] Y. P. Chen, J. L. Xiong, and W. C. Chew, A mixed-form thinstratified medium fast-multipole algorithm for both low and midfrequency problems, IEEE Trans. Antennas Propagat., vol. 59, no.
6, pp. 2341-2349, Jun. 2011.
[72] J. S. Zhao and W. C. Chew, MLFMA for solving integral euqations of
2-D electromagnetic problems from static to electrodynamic, Microw.
Opt. Tech. Lett., vol. 20, no. 5, pp. 306-311, 1999.
[73] S. M. Rao, D. R. Wilton and A. W. Glisson, Electromagnetic scattering by surface of arbitrary shape, IEEE Trans. Antennas Propagat.,
vol. 30, no. 3, pp. 409-418, May 1982.
[74] D. R. Wilton and A. W. Glisson, On improving the electic field integral equation at low frequencies, in Proc. URSI Radio Science Meeting
Dig., Los Angeles, CA, Jun. 1981, pp. 24-24.
[75] G. Vecchi, Loop-star decomposition of basis functions in the discretizaiton of the EFIE, IEEE Trans. Antennas Propagat., vol. 47,
no. 2, pp. 339-346, Feb. 1999.
[76] J. S. Zhao and W. C. Chew, Integral equation solution of Maxwells
equations from zero frequency to microwave frequencies, IEEE Trans.
Antennas Propagat., vol. 48, no. 10, pp. 1635-1645, Oct. 2000.
[77] J. F. Lee, R. Lee and R. J. Burkholder, Loop-star basis functions and
a robust preconditioiner for EFIE scattering problems, IEEE Trans.
Antennas Propagat., vol. 51, no. 8, pp. 1855-1863, Aug. 2003.
[78] Q. Chen and D. R. Wilton, Electromagnetic scattering by threedimensional arbitrary complex material/conducting bodies, in Proc.
Antennas Propag. Soc. Int. Symp., 1990, pp. 590-593.
113

[79] A. Buffa and S. Christiansen, A dual finite element complex on the

barycentric refinement, Math. Comput., vol. 76, pp. 1743-1769, 2007.
[80] F. P. Andriulli, K. Cools, H. Bagei, F. Olyslager, A. Buffa, S. Christiansen, and E. Michielssen, A multiplicative Calderon preconditioner
for the electric field integral equation, IEEE Trans. Antennas Propag.,
vol. 56, no. 8, pp. 2398-2412, Aug. 2008.
[81] M. B. Stephanson and J. F. Lee, Preconditioned electric field integral
equation using Calderon identities and dual loop/star basis functions,
IEEE Trans. Antennas Propagat., vol. 57, no. 4, pp. 1274-1279, Apr.
2009.
[82] S. Yan, J. M. Jin and Z. P .Nie, EFIE Analysis of low-freuqnecy problems with loop-star decomposition and Calderon multiplicative preconditioner, IEEE Trans. Antennas Propagat., vol. 58, no. 3, pp. 857-867,
Mar. 2010.
[83] S. Uckun, T. K. Sarkar, S. M. Rao and M. Salazar-Palma, A novel
technique for analysis of electromagnetic scattering from microstrip
antennas of arbitrary shape, IEEE Trans. Microwave Theory Tech.,
vol. 45, no. 4, pp. 485-491, Apr. 1997.
[84] F. Ling, V. I. Okhmatovski, W. Harris, S. McCracken, and A. Dengi,
Large-scale broad-band parasitic extraction for fast layout verification
of 3-D RF and mixed-signal on-chip structures, IEEE Trans. Microw.
Theory Tech., vol. 53, no. 1, pp. 264-273, Jan. 2005.
[85] Y. P. Chen, J. L. Xiong, W. C. Chew and Z. P. Nie, Numerical analysis of electricallly small structures embedded in a layered medium,
Microw. Opt. Tech. Lett., vol. 51, no. 5, pp. 1304-1308, 2009.
[86] Matti Taskinen and Pasi Yla-Oijala, Current and charge integral equation formulation, IEEE Trans. Antennas Propagat., vol. 54, no. 1, pp.
58-67, Jan. 2006.
[87] D. Gope, A. Ruehli, and V. Jandhyala, Solving low-frequency EMCKT problems using the PEEC method, IEEE Trans. Adv. Packag.,
vol. 30, no. 2, pp. 313-320, May 2007.
[88] Z. G. Qian and W. C. Chew, An augmented EFIE for high speed
interconnect analysis, Microw. Opt. Tech. Lett., vol. 50, no. 10, pp.
2658-2662, Oct. 2008.
[89] Z. G. Qian and W. C. Chew, Fast full-wave surface integral equation solver for multiscale structure modeling, IEEE Trans. Antennas
Propagat., vol. 57, no. 11, pp. 3594-3601, Nov. 2009.
114

[90] Y. P. Chen, L. Jiang, Z.-G. Qian, and W. C. Chew, An augmented

electric field integral equation for layered medium Greens function,
IEEE Trans. Antennas Propagat., vol. 59, no. 3, pp. 960-968, Mar.
2011.
[91] M. Benzi, G. H. Golub, and J. Liesen, Numerical solution of saddle
point problems, Acta Numer., vol. 14, pp. 1-137, Apr. 2005.
[92] S. Y. Chen, W. C. Chew, J. M. Song and J. S. Zhao, Analysis of low
frequency scattering from penetrable scatterers, IEEE Trans. Geosci.
Remote Sensing, vol. 39, no. 4, pp. 726-735, Apr. 2001.
[93] A. Barnard, I. Duck, M. Lynn, and W. Timlake, The application of
electromagnetic theory to electrocardiography II: numerical solution of
the integral equations, Biophys. J., vol. 7, pp. 463-490, Mar. 1967.
[94] Y. C. Pan, Development of the fast multipole mehtod for stratified
medium, Ph.D., University of Illinois, Urbana-Champaign, Illinois,
United States, 2002.
[95] Z. G. Qian and W. C. Chew, Enhanced A-EFIE with perturbation
method, IEEE Trans. Antennas Propagat., vol. 58, no. 10, pp. 32563264, Oct. 2010.
[96] 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.
[97] X.-B. Xu and C. M. Butler, Scattering of TM excitation by coupled
and partially buried cylinders at the interface between two media,
IEEE Trans. Antennas Propagat., vol. 35, no. 5, pp. 529-538, May
1987.
[98] J. R. Mosig, Arbitrarily shaped microstrip structures and their analysis with a mixed potential integral equation, IEEE Trans. Microwave
Theory Tech., vol. 36, no. 2, pp. 314-323, Feb. 1988.
[99] N. Kinayman, and M. I. Aksun, Efficient use of closed-form Greens
functions for the analysis of planar geometries with vertical connections, IEEE Trans. Microwave Theory Tech., vol. 45, no. 5, pp. 593603, May, 1997.
[100] 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 Trans. Geosci. Remote Sensing, vol. 38, no. 4.
pp. 1561-1573, Jul. 2000.

115

[101] B. Hu and W. C. Chew, Fast inhomogeneous plane wave algorithm for

scattering from objects above the multilayered medium, IEEE Trans.
Geosci. Remote Sensing, vol. 39, no. 5, pp. 1028-1038, May 2001.
[102] K. A. Michalski, On the scalar potential of a point charge associated with a time-harmonic dipole in a layered medium, IEEE Trans.
Antennas Propagat., vol. 35, no. 11, pp. 1299-1301, Nov. 1987.
[103] T. J. Cui, and W. C. Chew, Fast algorithm for electromagnetic scattering by buried 3-D dielectric objects of large size, IEEE Trans.
Geosci. Remote Sensing, vol. 37, no. 5, pp. 2597-2608, Sept. 1999.
[104] X. Millard, and Q. H. Liu, Simulation of near-surface detection of
objects in layered media by the BCGS-FFT method, IEEE Trans.
Geosci. Remote Sensing, vol. 42, no. 2, pp. 327-334, Feb. 2004.
[105] N. Geng and L. Carin, Wide-band electromagnetic scattering from
a dielectric BOR buried in a layered lossy dispersive medium, IEEE
Trans. Antennas Propagat., vol. 47, no. 4, pp. 610-619, Apr. 1999.
[106] J. He, T. Yu, N. Geng, and L. Carin, Method of moments analysis of
electromagnetic scattering from a general three-dimensional dielectric
target embedded in a multilayered medium, Radio Science, vol. 35,
no. 2, 305-313, Mar.-Apr. 2000.
[107] J. Chen, A. A. Kishk, and A. W. Glisson, Application of a 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, no. 2, pp. 263-279, Feb. 2001.
[108] E. Simsek, J. Liu, and Q. H. Liu, A spectral integral method (SIM)
for layered media, IEEE Trans. Antennas Propagat., vol. 54, no. 6,
pp. 1742-1749, Jun. 2006.
[109] J. L. Xiong and W. C. Chew, A newly developed formulation suitable
for matrix manipulation of layered medium Greens functions, IEEE
Trans. Antennas Propagat., vol. 58, no. 3, pp. 868-875, Mar. 2010.
[110] Y. P. Chen, W. C. Chew, and L. Jiang, A new Greens function formulation for modeling homogeneous objects in layered medium, submitted to IEEE Trans. Antennas Propagat..
[111] Abramowitz, M. and I. Stegun, Handbook of Mathematical Functions,
New York: Dover Publications, 1970.
[112] W. C. Chew, and S.-Y. Chen, Response of a point source embedded
in a layered medium, IEEE Antennas Wireless Propagat. Lett., vol. 2,
pp. 254-258, 2003.
116

[113] R. F. Harrington, Boundary integral formulations for homogenous

material bodies, J. Electromagn. Waves and Appl., vol. 3, no. 1, pp.
1-15, 1989.
[114] P. Y. Oijala and M. Taskinen, Application of combined field integral
equation for electromagnetic scattering by dielectric and composite objects, IEEE Trans. Antennas Propagat., vol. 53, no. 3, pp. 1168-1173,
Mar. 2005.
[115] P. Y. Oijala and M. Taskinen, Efficient formulation of closed-form
Greens functions for general electric and magnetic sources in multilayered media, IEEE Trans. Antennas Propagat., vol. 51, no. 8, pp.
2106-2115, Aug. 2003.
[116] D. J. Griffiths, Introduction to Quantum Mechanics, Upper Saddle
River, NJ:Prentice Hall, 1995.
[117] E. Purcell, Phys. Rev., 69, 681, 1946.
[118] S. M. Dutra, Cavity Quantum ElectrodynamicsThe Strange Theory
of Light in a Box, NJ:John Wiley & Sons, Inc., 2005.
[119] J. D. Joannopoulos, S. G. Johnson, J. N. Winn, and R. D. Meade,
Photonic CrystalsMolding the Flow of Light, 2nd ed. NJ:Princeton
University Press, 2008.
[120] S. A. Maier, Plasmonics: Fundamentals and Applications, New York:
Springer Science+Business Media LLC, 2007.
[121] P.-F. Qiao, W. E. I. Sha, W. C. H. Choy and W. C. Chew, Systematic
study of spontaneous emission in a two-dimensional arbitrary inhomogeneous environment, Phys. Rev. A, 83, 043824, 2011.
[122] J. Barthes, G. Colas des Francs, A. Bouhelier, J.-C. Weeber, and
A. Dereux, Purcell factor for a point-like dipolar emitter coupled to
a two-dimensional plasmonic waveguide, Phys. Rev. B, 84, 073403,
2011.
[123] L. Novotny, and B. Hecht, Principles of Nano-optics, Cambridge University Press, 2006.
[124] L. D. Landau, E. M. Lifshitz, and L. P. Pitaevskii, Statistical physics,
Part 2. Oxford, UK: Pergamon Press, 1980.
[125] S. M. Barnett, B. Huttner, and R. Loudon, Spontaneous emission
in absorbing dielectric media, Phys. Rev. Lett., vol. 68, no. 25, pp.
3698-3701, Jun. 1992.

117

[126] K. Okamoto, I. Niki, A. Shvartser, Y. Narukawa, T. Mukai, and

A. Scherer, Surface-plasmon-enhanced light emitters based on InGaN
quantum wells, Nature Materials, 3, pp. 601-605, Aug. 2004.
[127] C. Walther, G. Scalari, M. I. Amanti, M. Beck, and J. Faist, Microcavity laser oscillating in a circuit-based resonator, Science, vol. 327,
pp. 1495-1497, Mar. 2010.
[128] A. D. Rakic, A. B. Djurisic, J. M. Elazar, and M. L. Majewski, Optical
properties of metallic films for verticalcavity optoelectronic devices,
Appl. Opt. vol. 37, no. 22, pp. 5271-5283, Aug. 1998.

118

LIST OF PUBLICATIONS

Journal Papers
Y. P. Chen, J. L. Xiong, W. C. Chew, and Z. P. Nie, Numerical
analysis of electrically small structures embedded in a layered medium,
Microw. Opt. Tech. Lett., vol. 51, no. 5, pp. 1304-1308, May 2009.
Y. P. Chen, L. Jiang, Z.-G. Qian, and W. C. Chew, An augmented

electric field integral equation for layered medium Greens function,

IEEE Trans. Antennas Propagat., vol. 59, no. 3, pp. 960-968, Mar.
2011.

Y. P. Chen, W. C. Chew, and L. Jiang, A novel implementation of

discrete complex image method for layered medium Greens function,
IEEE Antennas Wireless Propagat. Lett., vol. 10, pp. 419-422, 2011.
Y. P. Chen, J. L. Xiong, and W. C. Chew, A mixed-form thinstratified medium fast-multipole algorithm for both low and mid-frequency
problems, IEEE Trans. Antennas Propagat., vol. 59, no. 6, pp. 23412349, Jun. 2011.
J. L. Xiong, Y. P. Chen, and W. C. Chew, A quasi-3D thin-stratified
medium fast-multipole algorithm for microstrip structures, IEEE Trans.
Antennas Propagat., vol. 59, no. 7, pp. 2578-2587, Jul. 2011.
W. E. I. Sha, W. C. H. Choy, Y. P. Chen, and W. C. Chew, Optical design of organic solar cell with hybrid plasmonic system, Optics
Express, vol. 19, no. 17, pp. 15908-15918, Aug. 2011.
Y. P. Chen, W. C. Chew, and L. Jiang, A new Greens function
formulation for modeling homogeneous objects in layered medium,
119

submitted to IEEE Trans. Antennas Propagat..

Y. P. Chen, W. E. I. Sha, W. C. H.Choy, W. C. Chew, and L. Jiang,
Study on spontaneous emission in complex coupled plasmonic system
via a rigorous surface integral equation method, in preparation.

Conference Papers
Y. P. Chen, J. L. Xiong, W. C. Chew, and Z. P. Nie, Simulation
of structures situated in a layered medium at low frequencies, Asia
Pacific Microwave Conference, Hong Kong SAR, China, Dec. 16-20,
2008.
Y. P. Chen, W. C. Chew, and J. L. Xiong, A mixed-form thinstratified medium fast-multipole algorithm for broadband simulation,
Workshop on Computational Electromagnetics and Its Applications,
Chengdu, China, Aug. 2009.
Y. P. Chen, J. L. Xiong, and W. C. Chew, Fast and broadband

simulation of large-scale microstrip structures, Progress in Electromagnetics Research Symposium, Xian, China, Mar. 22-26, 2010.

Y. P. Chen, L. Jiang, W. C. Chew, J. L. Xiong and Z.-G. Qian,

Study on low frequency integral equation method for layered medium

Greens function, Workshop on Simulation and Modeling of Emerging
Electronics, Hong Kong SAR, China, P-4, Dec. 6-10, 2010.

Y. P. Chen, W. C. Chew, and L. Jiang, A new closed-form evaluation

of layered medium Greens function, IEEE International Symposium
on Antennas and Propagation, Spokane, USA, pp. 3211-3213, Jul. 3-8,
2011.
Y. P. Chen, L. Jiang, Z.-G. Qian, and W. C. Chew, Modeling
electrically small structures in layered medium with augmented EFIE
method, IEEE International Symposium on Antennas and Propagation, Spokane, USA, pp. 3218-3221, Jul. 3-8, 2011.

120

 Y. P. Chen, W. C. Chew, and L. Jiang, A matrix representation

of dyadic Greens function for modeling general dielectric objects embedded in a layered medium, Progress in Electromagnetics Research
Symposium, Suzhou, China, pp. 277-277, Sept. 12-16, 2011.
P.-F. Qiao, W. E. I. Sha, Y. P. Chen, W. C. H. Choy, and W. C. Chew,

Spontaneous emission in 2D arbitrary inhomogeneous environment,

Progress in Electromagnetics Research Symposium, Suzhou, China, pp.
577-577, Sept. 12-16, 2011.

Y. H. Lo, L. Jiang, Y. P. Chen and W. C. Chew, Finite-width excitation and impedance models, Progress in Electromagnetics Research
Symposium, Suzhou, China, pp. 273-273, Sept. 12-16, 2011.

121

AUTHORS BIOGRAPHY

Yongpin CHEN was born in Zhejiang, China, in 1981. He received the B.S.
and the M.S. degrees in microwave engineering from the University of Electronic Science and Technology of China (UESTC), Chengdu, in 2003 and
2006, respectively. He was a Ph.D. student at UESTC from 2006 to 2007,
and then a Research Assistant at the University of Hong Kong (HKU) in
2008. He transferred to HKU as a Ph.D. student from Jan. 1, 2009 and is
currently pursuing his Ph.D. degree in electrical and electronic engineering
at HKU.

122