An Efficient High-Order Algorithm For Scattering by Inhomogeneous Penetrable Media

An Efficient High-Order Algorithm
for Scattering by
Inhomogeneous Penetrable Media
by
AMBUJ PANDEY
Department of Mathematics and Statistics
Indian Institute of Technology, Kanpur
January 2016
An Efficient High-Order Algorithm
for Scattering by
Inhomogeneous Penetrable Media
A Thesis Submitted
in Partial Fulfilment of the Requirements
for the Degree of
DOCTOR OF PHILOSOPHY
by
AMBUJ PANDEY
to the
Department of Mathematics and Statistics
Indian Institute of Technology, Kanpur
January, 2016
Synopsis
Name of the Student : Ambuj Pandey

Roll Number : Y7208061
Degree for which
submitted : Ph.D.
Department : Mathematics and Statistics
Thesis Title : An Efficient High-Order Algorithm for
Scattering by Inhomogeneous Penetrable Media
Thesis Supervisor : Prof. Akash Anand and Prof. B.V. Rathish Kumar
Month and year of
submission : January, 2016
Efficient and accurate simulation of acoustic or electromagnetic wave scattering by

inhomogeneous penetrable media is of paramount importance because of its appearance in
a broad range of applications, such as medical imaging, underwater optics, radar, sonar
and geophysics. Because of plethora of modern applications, the demand for accurate
numerical methods, that do not compromise much in terms of computational speed, is
increasing.
Formally, the d-dimensional direct acoustic scattering problem that we consider in this
thesis for d = 2 and 3, is described as follows: given an obstacle Ω, a bounded open
subset of Rd , with a smooth boundary ∂Ω, and an incident time-harmonic acoustic wave
i
ui satisfying
∆ui (x) + κ2 ui (x) = 0, x ∈ Rd , (0.0.1)
where κ = ω/c0 is the wavenumber, ω is the angular frequency, and c0 is the constant
speed of wave outside the inhomogeneity Ω, find the total acoustic field u that satisfies
∆u(x) + κ2 n2 (x)u(x) = 0, x ∈ Rd , (0.0.2)
with the refractive index n(x) = c0 /c(x), where c is the speed of acoustic wave, which
is allowed to vary with position within Ω and the scattered field us := u − ui satisfies
Sommerfeld radiation condition
∂us

(d−1)/2
lim r − ikus = 0, (0.0.3)
r→∞ ∂r
qP
d
where r = kxk2 = j=1 x2j .
There are two major approaches to solve this equation; the first one is by direct dis-
cretization of the differential equation or its variational formulation, and the other that
discretize after reformulating the problem as the integral equation
Z
2
u(x) + κ Gκ (x, y)m(y)u(y)dy = ui (x), x ∈ Rd , (0.0.4)
Rd
known as the Lippmann–Schwinger equation, where Gκ (x, y) is the fundamental solution

of the free space Helmholtz equation (0.0.1) and m(x) = 1 − n2 (x). The major drawback
in the former approach is, in order to deal with radiation condition (0.0.3), the infinite
domain has to be truncated into a finite one, and an artificial boundary condition needs to
be imposed on the boundary of truncated domain. These procedures often require a large
number of unknowns which lead to large linear systems. In addition, accurate artificial
boundary conditions with efficient numerical implementations are quite difficult to con-
struct; the error associated with such boundary conditions typically dominates the error
in the computed solution. In the later approach, however, Sommerfeld radiation condi-
tion (0.0.3) is directly incorporated in the integral equation formulation as a consequence
of the fundamental solution satisfying Sommerfeld radiation condition. Thus, automatic
ii
satisfiability of radiation condition and absence of dispersion errors that typically accom-
panies schemes belonging to the former class makes integral equation based solvers much
attractive for the numerical solution of scattering problems.
In recent years, a lot of progress has been made toward numerical solution of the
Lippmann-Schwinger Integral equation, including both, direct and iterative solvers. Al-
though, existing fast numerical scheme converges to high-order for smooth scattering con-
figuration, they produces only linear or quadratic convergence in the presence of material
discontinuity. In 2007, Anand and Reitich, proposed an algorithm for the solution of acous-
tic volumetric scattering problem in two and three dimensions, that exhibits high-order
convergence but it is designed to be computationally efficient only for “thin” scattering
configurations. Our present work, in fact, is a non-trivial extension of ideas presented by
Anand and Reitich wherein we obtain fast high-order solver for general scattering config-
urations.
We believe that this thesis is the first, fast integral equation based solver for the so-
lution of general scattering configuration that retains high-order accuracy in presence of
material discontinuity. The content of this thesis has been divided into seven chapters.
The description of these chapters is given below:
Chapter 1 includes introduction and preliminary discussions that are relevant to the
subject matter of this thesis. This chapter starts with a brief discussion of important ap-
plication of the scattering phenomena and a derivation of the mathematical model given
in the form of the Helmholtz equation which, in fact, encompasses many of the most im-
portant mathematical and computational issues that arise in scattering theory. This is
followed by, an overview of the mathematical tools for establishing the equivalence of the
integral equation formulation, which are also used in subsequent chapters, a derivation
of Lippmann–Schwinger integral equation is also included. Subsequently, a short review
of existing numerical methods for the solution of wave scattering problems is discussed.
Finally, in the context of our integral equation approach to the aforementioned scattering
problems, a brief survey of methodologies relating to the numerical treatment of Lipp-
mann–Schwinger integral equation (0.0.4) is delineated.
iii
In Chapter 2, we have provided a detail discussion of our algorithm for the solution
of Lippmann–Schwinger integral equation (0.0.4) in two dimensions. The method works
with overlapping coordinate charts as a description of the given scatterer. In particular,
it employs “partitions of unity” to simplify the implementation of high-order quadratures
along with suitable changes of parametric variables to analytically resolve the singularities
present in the integral operator to achieve desired accuracies in approximations. To deal
with the discontinuous material interface in a high-order manner, a specialized quadrature
is used in the boundary region. The approach further utilizes an FFT based strategy that
uses equivalent source approximations to accelerate the evaluations of large number of
interactions that arise in the approximation of the volumetric integral operator and thus
achieves a reduced computational complexity of O (N logN ) for an N -point discretization.
In Chapter 3, we presented a variety of numerical results that demonstrate numerical
properties and performance of approach in two dimensions described in Chapter 2. In
particular, we exemplify the adaptability and applicability of the proposed methodology
in dealing with complex shape scatterer. We have also demonstrated that the proposed
algorithm does provide high-order accuracy for discontinuous scatters while maintaining
the desired computational efficiency.
Chapter 4 is dedicated to presenting a fast and accurate computational strategy to solve
the Lippmann–Schwinger integral equation (0.0.4) in three dimensions. While several is-
sues that arise in the evaluation of three dimensional integral in (0.0.4) are analogous to the
two dimensional counterpart, new challenges arise, nevertheless, mainly due to increased
dimensionality but also as a result of the difference in the behavior of the corresponding ker-
nel singularity. Similar to the two dimensional approach, we have adapted the framework
of overlapping parameterizations for describing general scattering geometry and employ a
partitions of unity to simplify the design of quadratures for approximation of the integral
operator. As in two dimensions, a high-order convergence is achieved through analytic
resolution of singularities via suitable change of parametric variables. Accelerated evalu-
ation of the non-singular integrals is accomplished by three dimensional equivalent source
representation on a Cartesian grids and employing three dimensional FFTs.
iv
In order to exemplify the high-order accuracy of the approach, and to demonstrate the
effect of its accelerated evaluation, we present computational results obtained from our
implementation of the three dimensional algorithm in Chapter 5.
In Chapter 6, we derive the error bounds for the interpolations and quadrature rule
described in Chapter 2 and 4. In particular, our error analysis establishes theoretically
super algebraic convergence of the proposed quadrature scheme in a case in which refractive
index n(x) is smooth within the scatterer Ω.
Final conclusions of this thesis work has been presented in Chapter 7. An outline of
the topics to be pursued in these directions in future have also been included. The relevant
references are given in the Bibliography.
v
Acknowledgment
It has been a privilege to spend these years in the Department of Mathematics and Statis-
tics, IIT Kanpur. Its members will always remain dear to me. This thesis would not
have been possible without the help and support from many people to whom I am greatly
indebted.
First and foremost, I would like to express my deepest gratitude to my thesis super-
visors, Professor Akash Anand and Professor B.V. Rathish Kumar, for their excellent
guidance, invaluable suggestions and generous help at all the stages of my research work.
I would like to thank them for introducing me to the interesting field of computational
acoustics and electromagnetics.
I would like to express my sincere thanks to Professor Pravir Dutt for his valuable
suggestions and support during my research work, and more importantly for drawing my
attention to Professor Akash Anand. I am grateful to Professor Nandini Nilakantan for
her support and encouragement at all the stages of my research work. I express my heart-
felt thanks to Professor Debasis Kundu and Professor Amit Mitra for being so kind and
concerned for my well doing in career. I take this opportunity to thank Professor Shobha
Madan who has always been encouraging and accommodating. I also extend my sincere
thanks to all the faculty members of the Department of Mathematics and Statistics, in par-
ticular, Professor M. K. Kadalbajoo, Professor Sameer Chavan, Professor Parasar Mohanty
and Professor Sudipta Dutta for their cooperation during this period.
I am indebted to all my school teachers, specially, Avsar Ali sir, Jakir Ali sir and
Ramesh Shrivastva sir, who motivated me to do well in my studies. I am also indebted to
vii
my college teachers Dr. Anjani Kumar Singh, Dr. E.K. Eklof and Dr. Ashok Pandey, not
only their teaching but also their moral support and motivation in my odd days to continue
my studies. I am grateful to my university teachers Professor P.N. Pandey, Professor R.P.
Shukla and Professor D.P. Chaudhary for many fruitful discussions and making us feel
what research is all about.
Friends are always an integrated part of our life. Without their support and help our
journey of life is impossible. My best friend late Anand Narayan Pandey was the person
who motivated me to pursue higher study in mathematics. His unfortunate death was a
big loss to me. I acknowledge the role of Manish Bhaiyaa for his concern and interest in
my academic works. He has been a great source of encouragement even in my odd times.
I owe my thanks to M.Sc. friends, who always very close to my hearts, Karamdev, Qaiser,
Ranu Singh, Uday and Sashank. Specially, the support I got from Karamdev can never
be expressed in words. My heartiest thanks to Abdullah and Binoy, who helped me in
different ways to boost my moral during my PhD program. In many ways, their company
evolve me to be a better human being.
My special thanks goes to Jagabandhu for his careful reading of the initial manuscript
and providing many helpful comments.
I am thankful to Awanish, Arun, Surjit, Lokpati, Indira and Alpesh, to be with me
during the difficult times of my life. The 4 P.M. discussions over tea at the chemical canteen
is one of the most memorable part of my life in IIT Kanpur. I am very lucky to have them
as my friends.
I thank to all my friends for their companionship and I will cherish their memories
ever. Special thanks to Arbaz, Shanker, Samir, Anurag, Anuj, Parul, Gopal, Vamsinadh,
Abhishek, Abdul Halim, Sheela.
I find no words to thank my Buaa Manorama Shukla and Phupha Ashok Shukla for
providing me a support during my initial education. Their support had changed my life a
lot. I don’t know how to express my affection to their son Prashant and daughter Gudiya.
I sincerely thank my uncle Laxmi Raman Pandey for his continuous encouragement. I
express my deep sense of gratitude to my bade papa Late Shri Kamla Shanker Pandey and
viii
my uncle Late Shri Ravi Shanker Pandey who had been constant source of inspiration and
encouragement throughout my education. My heart felt gratitude to my father Shri Vimla
Shanker Pandey. His everlasting support and motivation always gave me the strength to
continue my work without thinking about result. His tremendous faith on me always make
me confident to be able to do more hard work and accomplish my dreams honestly. I am
thankful to my mother, badi maa and chachi for their love and support in all respects. My
deep love and regards to my brothers Pradeep, Sudhir, Pankaj, Rajneesh, Ashu and Sanad
whose love and support have been an constant source of inspiration.
I am also grateful to the Council of Scientific and Industrial Research (CSIR) India for
providing me fellowship to carry out this research work.
Last but not the least, I thank the Almighty God for the uncountable blessings, the
opportunities and strength He showered on me which make my life easy.
ix
Dedicated
to
My Father Shri Vimla Shanker Pandey
and
Uncle Late Shri Kamla Shanker Pandey
Contents
Synopsis i
Acknowledgments vii
Contents xi
List of Figures xiii
List of Tables xix
1 Introduction and Preliminaries 1

1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 The Helmholtz equation . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 Scattering Formulation . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Fundamental Solution . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.4 Integral Equation Formulation . . . . . . . . . . . . . . . . . . . . . 12
1.3 Brief Review of Numerical Methods for Scattering Problems . . . . . . . . 17
1.3.1 Differential Equation Based Methods . . . . . . . . . . . . . . . . . 17
1.3.2 Integral Equation Based Methods . . . . . . . . . . . . . . . . . . . 19
1.4 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5 Organization of The Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2 An efficient high-order integral equation solver in two dimensions 27
xi
2.1 Scatterer Parametrization, Nyström Nodes And Partitions of Unity . . . . 28
2.2 Decomposition of The Integral Operator . . . . . . . . . . . . . . . . . . . 33
2.3 Fast and Accurate Evaluation of Integrals . . . . . . . . . . . . . . . . . . 38
2.3.1 Singular Integration . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.2 Singular Integration Over Boundary Patches . . . . . . . . . . . . . 46
2.4 Non-Singular Integration: Acceleration . . . . . . . . . . . . . . . . . . . . 51
3 Computational Results In Two Dimensions 57
4 An efficient high-order integral equation solver in three dimensions 87

4.1 Principal Components of The Method In Three Dimensions . . . . . . . . . 88
4.2 Singular Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3 Non-singular integration: Acceleration . . . . . . . . . . . . . . . . . . . . 105
5 Computational Results In Three Dimensions 109
6 Error Analysis 127

6.1 Interpolation Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.2 Quadrature Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.2.1 Quadrature Analysis In Two Dimensions . . . . . . . . . . . . . . . 145
6.2.2 Quadrature Analysis In Three Dimensions . . . . . . . . . . . . . . 156
7 Conclusions and Future Directions 163
Bibliography 167
xii
List of Figures
2.1 The figure on the top left depicts the discretization for the scatterer coming
from a typical volumetric computational grid. Other three figures display the
overlapping coordinate patches and computational grid on it. . . . . . . . . . . 30
2.2 A typical point in the set Tk for a patch Pk when N1 = 8, N2 = 4 . . . . . . . 31
2.3 An illustration of the function χ(s) used as a building block in the construction
of partitions of unity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Partition of unity corresponding to interior patch for disc shape scatterer. . . . 34
2.5 Partition of unity corresponding to Interior patch for disc shape scatterer. . . 35
2.6 Partition of unity on disc shape scatterer. . . . . . . . . . . . . . . . . . . . . 36
2.7 Partition of unity on bean shape scatterer. . . . . . . . . . . . . . . . . . . . . 37
2.8 Change of variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.9 An illustration of the locations xij,` of equivalent sources (points displayed in

blue color) on the parallel faces of cells Cij . The image also depicts, for a target
point x, the set N (x) that defines the adjacency. . . . . . . . . . . . . . . . . 50
3.1 Plane wave exp(iκd · x) incidence along the incidence direction d = (1, 0) with
wave number κ = 20 on a penetrable unit disc. Fields computed in whole square
region displayed in blue color. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
xiii
3.2 Visualization of real part of the fields for scattering of a plane wave exp(iκd · x)
with κ = 20, d = (1, 0) by a penetrable unit disc. A computational grid of size
2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to
obtain an error of 0.032% measured in max norm. . . . . . . . . . . . . . . . . 72
3.3 Visualization of imaginary part of the fields for scattering of a plane wave
exp(iκd · x) with κ = 20, d = (1, 0) by a penetrable unit disc. A computa-
tional grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes
quadrature is used to obtain an error of 0.032% measured in max norm. . . . . 73
3.4 visualization of absolute value of the fields for scattering of a plane wave exp(iκd·
x) with κ = 20, d = (1, 0) by a penetrable unit disc. A computational grid of
size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used
to obtain an error of 0.032% measured in max norm. . . . . . . . . . . . . . . 74
3.5 Visualization of real part of the fields for scattering of a plane wave exp(iκd · x)
√ √
with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A computational
grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature
is used to obtain an error of 0.06% measured in max norm. . . . . . . . . . . . 75
3.6 Visualization of imaginary part of the fields for scattering of a plane wave
√ √
exp(iκd · x) with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A
computational grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-
Cotes quadrature is used to obtain an error of 0.06% measured in max norm. . 76
3.7 Visualization of absolute value of the fields for scattering of a plane wave
√ √
exp(iκd · x) with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A
computational grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-
Cotes quadrature is used to obtain an error of 0.06% measured in max norm. . 77
3.8 Three-patch covering of the Bean shape scatterer depicted in Figure 3.9(a).
First and last figure corresponds to boundary patch while the middle figure
represent voluminous interior patch. . . . . . . . . . . . . . . . . . . . . . . . . 78
3.9 Plane wave incidence exp(iκd · r) with κ = 10, d = (1, 0) on a penetrable bean
shape scatterer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
xiv
3.10 Visualization of real part of fields for scattering by penetrable bean shape scat-
terer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 × 129 × 129,
Newton-Cotes five point quadrature is used for transverse integration over
boundary patches. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.11 Visualization of imaginary part of fields for scattering by penetrable bean shape
scatterer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 × 129 ×
129, Newton-Cotes five point quadrature is used for transverse integration over
3.12 Visualization of absolute value of the fields for scattering by penetrable bean
shape scatterer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 ×
129×129, Newton-Cotes five point quadrature is used for transverse integration
over boundary patches. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.13 Visualization of real part of fields for scattering by penetrable bean shape scat-
√ √
terer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2×65×129+1×129×
129, Newton-Cotes five point quadrature is used for transverse integration over
3.14 Visualization of imaginary part of fields for scattering by penetrable bean shape
√ √
scatterer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2 × 65 × 129 + 1 ×
129×129, Newton-Cotes five point quadrature is used for transverse integration
over boundary patches. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.15 Visualization of absolute value of the fields for scattering by penetrable bean
√ √
shape scatterer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2 × 65 ×
129 + 1 × 129 × 129, Newton-Cotes five point quadrature is used for transverse
integration over boundary patches. . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1 Three patch covering of full spherical shape geometry and Nyström discretiza-
tion grids. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2 Partition of unity on boundary patch of spherical shape scatterer. . . . . . . . 92
4.3 Partition of unity on interior patch of spherical shape scatterer. . . . . . . . . 93
xv
5.1 Spherical shape scatterer, boundary is covered by two overlapping coordinate
patches. Discretization on the outer surface of the scatterer coming from a
typical volumetric computational grid. . . . . . . . . . . . . . . . . . . . . . . 112
5.2 Three overlapping coordinate patches and corresponding computational grid

used for scattering calculation of the bean shape scatterer. . . . . . . . . . . . 118
5.3 Two patch covering of outer surface of Bean shape scatterer and the Nyström
grid comes from a typical volumetric computational grid. . . . . . . . . . . . . 119
5.4 Visualization of real part of the total and scattered field on the three orthogo-
nal slices from penetrable spherical shape scatterer of acoustical size κa = 20
√
and refractive index n(x) = 2, under the plane wave incidence coming form
positive z-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.5 Visualization of imaginary part of the total and scattered field on the three
orthogonal slices from penetrable spherical shape scatterer of acoustical size
√
κa = 20 and refractive index n(x) = 2, under the plane wave incidence
coming form positive z-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.6 Visualization of absolute value of the total and scattered field on the three
orthogonal slices from penetrable spherical shape scatterer of acoustical size
√
κa = 20 and refractive index n(x) = 2, under the plane wave incidence
coming form positive z-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.7 Visualization of real part of the total and scattered field on the three orthogonal
slices from penetrable bean shape scatterer of acoustical size κa = 20 and
√
refractive index n(x) = 2 under the plane wave incidence coming form positive
y-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.8 Visualization of imaginary part of the total and scattered field on the three
orthogonal slices from penetrable bean shape scatterer of acoustical size κa = 20
√
and refractive index n(x) = 2, under the plane wave incidence coming form
positive y-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
xvi
5.9 Visualization of absolute value of the total and scattered field on the three
orthogonal slices from penetrable bean shape scatterer of acoustical size κa = 20
√
and refractive index n(x) = 2 under the plane wave incidence coming form
positive y-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.1 Trigpoly interpolation at off grid points using refined data obtained by FFTs.
Top figure presents original coarse uniform grid in an interval [a, b]. Middle
figure presents refined grid where the function values is obtained by inverse
FFTs. Bottom figure presents local polynomial interpolation of fixed degree R
in a neighborhood of off grid point t (encircled and marked by star). . . . . . 133
xvii
List of Tables
3.1 Convergence study: Forward map computation for a disc of acoustic size κa = 4
√
and refractive index n = 3 when 3-point Newton-Cotes quadrature employed
for t2 -integration over boundary patches. . . . . . . . . . . . . . . . . . . . . . 60
3.2 Convergence study: plane wave scattering by a disc of acoustic size κa = 4 and
√
refractive index n = 3 when 3-point Newton-Cotes quadrature employed for
t2 -integration over boundary patches. . . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Convergence study: Forward map computation for a disc of acoustic size κa =
√
10 and refractive index n = 2 when 3-point Newton-Cotes quadrature em-
ployed for t2 -integration over boundary patches. . . . . . . . . . . . . . . . . . 61
3.4 Convergence study: plane wave scattering by a disc of acoustic size κa = 10
√
√
√
3.7 Convergence study: Forward map computation for a disc of acoustic size κa = 4
√
xix
3.8 Convergence study: plane wave scattering by a disc of acoustic size κa = 4 and
√
refractive index n = 3 when 5-point Newton-Cotes quadrature employed for
t2 -integration over boundary patches. . . . . . . . . . . . . . . . . . . . . . . . 63
√
√
√
√
3.13 Computational cost for accelerated and non-accelerated algorithm. . . . . . . . 65
3.14 Accuracy of acceleration for fixed wavenumber κ = 8π . . . . . . . . . . . . . 66
3.15 Accuracy of acceleration when wavenumber κ increases but number of point
per wavelength is fixed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.16 Convergence study: Forward map computation for a bean of acoustic size
√
κa = 10 and refractive index n = 2 when 3-point Newton-Cotes quadrature
employed for t2 -integration over boundary patches. . . . . . . . . . . . . . . . 67
√
3.17 Convergence for the bean shape scatterer with κa = 10 and n = 2, when
3-point Newton-Cotes quadrature employed for boundary patch integration in
transverse integration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.18 Convergence study: Forward map computation for a bean of acoustic size
√
xx
√
3.19 Convergence for the bean shape scatterer with κ = 10 and n = 2, when
5-point Newton-Cotes quadrature employed for boundary patch integration in
transverse integration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.20 Convergence study: Forward map computation for a bean of acoustic size κa =
4 and refractive index n = sin(πx) cos(πy) for x = (x, y) ∈ Ω, when 3-point
Newton-Cotes quadrature employed for t2 -integration over boundary patches. . 69
3.21 Convergence for the bean shape scatterer with κa = 4 and n = sin(πx) cos(πy)
for x = (x, y) ∈ Ω, when 3-point Newton-Cotes quadrature employed for bound-
ary patch integration in transverse integration. . . . . . . . . . . . . . . . . . . 69
3.22 Convergence study: Forward map computation for a bean of acoustic size κa =
4 and refractive index n = sin(πx) cos(πy) for x = (x, y) ∈ Ω, when 5-point
Newton-Cotes quadrature employed for t2 -integration over boundary patches. . 69
3.23 Convergence for the bean shape scatterer with κa = 4 and n = sin(πx) cos(πy)
for x = (x, y) ∈ Ω, when 5-point Newton-Cotes quadrature employed for
boundary patch integration in transverse integration. . . . . . . . . . . . . . . 70
5.1 Convergence study: Forward map computation for a sphere of acoustic size
√
5.2 Convergence study: Plane wave scattering by a sphere of acoustic size κa = 2

√
√

√
xxi
√
√
√
√
10π
5.9 Accuracy of acceleration for fixed wavenumber κ = 3
. . . . . . . . . . . . . 117
5.10 Accuracy of acceleration for increasing acoustical cell size. . . . . . . . . . . . 117
xxii
Chapter
1
Introduction and Preliminaries
Wave scattering is a natural phenomenon that occurs everyday in our life, for example,
scattering of light wave in the atmosphere by dust particle, sound wave scattering by packs
of fish in the ocean, scattering of light for traffic signals, scattering of ultrasound waves
by biological tissues, etc. Because of its vital importance in many field of science and
engineering the accurate simulation of wave propagation problem has become a fascinating
discipline for last few decades and has drawn the attention of engineers, computer sci-
entists, physicists, and mathematician. Indeed, in real life, most of our interaction with
external world takes place by extracting informations from the waves scattered by objects
in it. As when energy wave propagate through a medium encounters an obstacle along its
propagation path may undergo propagation effect like a reflection, refraction, transmission,
diffraction etc. This scattering phenomena serves as a source of information on the physical
world surrounding us. The object or obstacle responsible for wave scattering is known as
scatterer. Although, we don’t realize it, scattering phenomena plays a very important role
in our daily life. For example, it is well known that our eyes can only see those objects
which interact with visible light in some way, for instance through reflection or refraction,
which are two forms of scattering. In addition, scattering of light (which is known to be
1
Chapter 1. Introduction and Preliminaries
as an electromagnetic wave) explain why the sky is blue and it also explains why the sun
looks red at sunrise and sunset. Indeed, in the present world, it is virtually impossible to
find an area in technology that does not utilize the scattering phenomena.
In particular, scattering of electromagnetic and acoustic waves itself find applications
in a wide spectrum of areas in engineering and industry, including ultrasound imaging,
underwater optics, military, biomedical, communications, material science, plasma physics,
radar and remote sensing to name but just a few.
The mathematical models for such problems are based on physical conservation laws,
and lead to partial differential equations. It is desirable to solve the scattering problems
using analytical methods and obtain closed-form solutions. However, because of the com-
plexity of the governing partial differential equations, only a limited number of scattering
geometries with enough symmetry can be solved exactly using analytical techniques typi-
cally employ separation of variables to transform the underlying partial differential equation
into ordinary differential equations where solutions can then be expressed in terms of special
functions. For solution of scattering problems of real life applications such methodologies
do not apply and we must turn to numerical methods to obtain an accurate approxima-
tion. With the advances in computer hardware and numerical techniques from the past
few decades, it has become possible to solve the computationally challenging scattering
problems numerically and obtain solutions with acceptable accuracy. Nevertheless search
for accurate and efficient algorithms for numerical solution of scattering problems is far
from over and challenges remain due to its specific computational bottlenecks, particularly
for large scale problems.
1.1 Applications
Analysis and understanding of scattering phenomena is crucial to the development and
innovation of advanced technology impacts on a wide variety of applications spanning
many fields. Remote sensing techniques, such as RADAR (Radio Detection And Ranging),
SONAR (Sound Navigation And Ranging), are known to be based on the principal of
acoustic and electromagnetic wave scattering [91], [101, Chapter 5]. For example, the
2
1.1. Applications
knowledge of scattering of radar waves is vital to applications in scientific, commercial, and

military operations [16, 40]. The most common application can be seen in the navigation
and tracking of crafts at sea and in the air. RADAR sends out electromagnetic waves,
which hit the target and a part of the wave is reflected back to the radar set where they
are received, amplified, and processed electronically to yield an image showing the object’s
location. The strength of the reflected wave is also characteristic of the target and the
environment in which the systems are operating. Radar systems are capable of accurately
determining the distance to their targets by recording the length of the time for the signals
to travel to the target and to be reflected back to the receiver and by utilizing the fact that
the speed of radar waves is same as the speed of light. In fact, modern radar technology is
capable of predicting both, location of an object and direction of its movement. As they
can tell how fast and in which direction a target is moving, radars are used by military to
track the enemies missiles and submarines.
Similarly SONAR also utilizes principals of acoustic and electromagnetic scattering
and found myriad applications in wide variety of fields. It allows scientists, geologists,
and archaeologists to map topography and subsurface features on earth and on objects
within the solar system. Because SONAR signals easily penetrate water, they are ideal
for navigation and identification of mineral and energy resources under water. SONAR
waves are also used by submarines to detect other vessels. From the past few decades,
considerable mathematical and computational progress has been made in elucidating and
understanding the scattering theory involved in such a process [45, 72].
Wave scattering is also utilized by geophysicists in detection and imaging of under-
ground structure such as an oil, gas or minerals deposit [50, Chapter 16] , [27, 75]. To
obtain the information about under ground or under water structure the boundary of the
earth is perturbed by an artificial explosion or by a natural earthquake, this perturbation
produces waves that travels through geodesics and hits the oil and mineral deposits and
reflects back to the surface of the earth. Instruments either at underground or at the
surface detect the reflections from these waves, allowing researchers to extract information
from these reflected waves.
3
Wave scattering phenomena is at heart of medical imaging techniques such as X-Ray

Computed Tomography (CT) (X-Ray is high-frequency electromagnetic wave that propa-
gate almost in straight line), Magnetic Resonance Imaging (MRI), Positron Emitted To-
mography (PET), where image of the specific part of the human body is created from
minimal invasive procedure [48]. Probably, among all X-Ray, Computed Tomography is
the simplest example of a medical imaging method. The key idea behind this technique
is, when X-Ray beam of known intensity impinges on a medium such as muscle or brain
tissue, as a scattering effect, some of its energy get absorbed and some passes through it,
depending on the radio-density of the medium. By measuring the change in the intensity
of X-Ray beam one can produce specific image of the interior of the body. To recover dif-
ferent properties of the tissue, modern medical images uses different type of waves instead
of X-rays. For example, Optical Coherence Tomography (OCT) uses light waves to image
nontransparent tissues and this enabling OCT to be applied in a wide range of medical
specialties [51, 98, 99]. It is interesting to note that mathematics to realize the topographic
image were devolved in 1917 by Austrian mathematician J. Radon, 62 years later A. Cor-
mack and G. Hounsfield [63] in 1979 rediscover this formula by combining the Physics of
X-Rays with mathematics of tomography and got the Nobel prize in Medicine for their
work of X-Ray Computed Tomography. This is an example of the need for interdisciplinary
work between mathematics and science.
Microwave imaging is another important application of scattering phenomena in medical
science such as brain imaging, and cancer detection in human body [71]. It emerged as
a very powerful technique in early detection of cancer cells in the breast [53, 55, 84] and
leukemia in the bone marrow of the leg [38, 39]. In fact, because of significant success in
cancer detection microwave imaging is considered as one of the main pillars in biomedical
fields of comprehensive cancer care.
These are just few examples of the wide range of fields in which scattering theory
played an important role. Despite the great benefit that these techniques provide, much
more potential benefit remains to gained through deeper understanding of mathematics
and innovative algorithm for solving these scattering problems.
4
1.2. Mathematical Preliminaries
1.2 Mathematical Preliminaries

In this section we collect mathematical preliminaries for the proposed thesis. We start by
briefly reviewing the mathematical formulation of acoustic wave propagation which under
the assumption of time-harmonic scattering boils down to the Helmholtz equation. Al-
though the Helmholtz equation models the time-harmonic propagation of acoustic waves,
in some settings it can also be derived from the heat conduction equation, Schrödinger
equation, telegraph equation or other wave type or evolutionary equations [60]. For exam-
ple, in elastic wave scattering, it is well known that the Helmholtz decomposition enables
to decouple elastic waves into p-waves and s-waves. In particular, former are the curl-free
gradient of a scalar potential, which satisfy an acoustic wave equation. If the medium is
a fluid then only p-waves can propagate, which are easily modeled in frequency domain
by the Helmholtz equation via the scalar potential. Scattering by electromagnetic wave is
modeled by Maxwell’s equations which can be decomposed into a transverse magnetic and
a transverse electric component. With this decomposition, Maxwell’s equations reduce to
the Helmholtz equation for the z-component of either of the transverse components. In the
following subsection we briefly derive Helmholtz equation for scattering by acoustic and
electromagnetic wave in an isotropic medium.
1.2.1 The Helmholtz equation

Consider the propagation of small amplitude time harmonic sound waves in an inviscid
fluid. We denote the velocity vector field by U = U (x, t), and pressure with p = p(x, t),
density with ρ = ρ(x, t) , entropy with S = S (x, t), where x ∈ R3 . If no external forces
are acting on the fluid, then the motion is governed by (see [36, 68])
∂U 1
+ (U · ∇)U + ∇p = 0 (1.2.1)
∂t ρ
which is Euler’s equation. The equation of continuity is given by
∂ρ
+ ∇ · (ρU ) = 0, (1.2.2)
∂t
5
the adiabatic hypothesis by

∂S
+ U · ∇S = 0 (1.2.3)
∂t
and the state equation by
p = f (ρ, S ), (1.2.4)
where f is a function depending on the fluid. If we assume that U , p, ρ and S are small
perturbations of the static states U = 0, p = p0 = constant, ρ = ρ0 (x) and S = S (x)
with p0 = f (ρ0 , S0 ) then we can express
U (x, t) = U1 (x, t) + · · · (1.2.5)
p(x, t) = p0 (x) + p1 (x, t) + · · · (1.2.6)
ρ(x, t) = ρ0 (x) + ρ1 (x, t) + · · · (1.2.7)
S (x, t) = S0 (x) + S1 (x, t) + · · · (1.2.8)
where 0 < 1 and the dots denote higher order term in . If we substitute (1.2.5)-
(1.2.8) into (1.2.1)-(1.2.4) and retaining only the terms of order then we obtain linearized
equations
∂U1 1
+ ∇p1 = 0 (1.2.9)
∂t ρ0
∂ρ1
+ ∇ · (ρ0 U1 ) = 0 (1.2.10)
∂t
∂p1 2 ∂ρ1
= c (x) + U1 · ∇ρ0 (1.2.11)
∂t ∂t
where the speed of acoustic wave c is defined by
∂f
c2 (x) = f (ρ0 (x), S0 (x)) . (1.2.12)
∂ρ
From this we obtain that p1 satisfies wave equation
∂ 2 p1

2 1
= c (x)ρ0 (x)∇ · ∇p1 . (1.2.13)
∂t2 ρ0 (x)
If we neglect the term involving grad ρ0 and assume that p1 is time-harmonic which means
it can be described as a product of a spatial function and a periodic temporal function, i.e,
p1 (x, t) = Re u(x)e−iωt ,

(1.2.14)
6
1.2.2. Scattering Formulation
where ω > 0 denotes the frequency, then spatial part u(x) of the function p1 (x, t) satisfies
the reduced wave equation or Helmholtz equation
∆u(x) + κ2 n2 (x)u(x) = 0 (1.2.15)
named in the honor of German physicist Hermann Ludwig Ferdinand von Helmholtz
(1821 − 1894) a pioneer in acoustics, electromagnetism and physiology. Here, wave number
κ is given by
ω
κ=
c0
and refractive index n(x) is given by
c0
n(x) = ,
c(x)
where c0 is the speed of sound in the vacuum.

As we have enunciated above Helmholtz equation encompass many of most computa-
tional issues arising in scattering theory. For example, equation (1.2.15) can also be derived
from Maxwell equation. Indeed, these underlying similarities in the mathematical models,
leads to similar numerical effects in computational implementations.
Although, equation (1.2.15) describes the propagation of acoustic wave in an inhomoge-
neous medium, for any meaningful solution of the scattering problem, we must specify how
the wave is initiated and what is the boundary of the region containing the fluid. Therefore,
in the next subsection we will present a precise formulation of obstacle scattering problem
by acoustic waves.
1.2.2 Scattering Formulation

Generally the obstacle scattering problem can be divided into two main classes namely
scattering by impenetrable and penetrable obstacle which we describe briefly below:
To this end we begin by expressing the total field by u as a sum of the given incident
field ui and a scattered field us
u = ui + us .
7
The incident field ui is simply the wave-function that would exist in the absence of the
scatterer and satisfy free space Helmholtz equation
∆ui + κ2 ui = 0, (1.2.16)
2π
in all of Rd , where κ = λ
denotes the wave number and λ is the wavelength of incident
wave ui . However, as we mentioned in the beginning of this chapter, the scattered field us
represents a wave produced by the reflection and diffraction of incident wave ui when it
impinges on the scatterer Ω, a bounded open subset of Rd with smooth boundary ∂Ω.
The mathematical formulation for the case of impenetrable scatterer is to find the total
field u ∈ C 1 (Rd \ Ω) ∩ C 2 (Rd \ Ω) such that
∆u(x) + κ2 u(x) = 0, x ∈ Rd \ Ω (1.2.17)
u(x) = 0, x ∈ ∂Ω. (1.2.18)
Depending on the physical characteristics of the scatterer Ω, one can impose different
boundary conditions instead of Dirichlet boundary condition (1.2.18) on the boundary
of scatterer Ω. For instance, if the normal derivative of the total field u vanishes on the
interface ∂Ω, that is, if the surface of scatterer is sound hard then the appropriate boundary
condition is the Neumann boundary condition:
∂u
(x) = 0, x ∈ ∂Ω.
∂ν
Depending upon the context, more general boundary condition may be relevant, “Robin”
or “Impedance” boundary condition of the form

∂u
u + iµ (x) = 0, x ∈ ∂Ω.
∂ν
Solution of the problem (1.2.17) together with boundary condition (1.2.18) is not unique
unless a condition at infinity is imposed [36]; here the relevant condition is the Sommerfeld
radiation condition
∂us

(d−1)/2
lim r − iκus = 0, (1.2.19)
r→∞ ∂r
8
1.2.3. Fundamental Solution
where r = kxk2 and limit in equation (1.2.19) exist uniformly in all directions x/kxk2 . In
fact, advantage of radiation condition (1.2.19) is in two fold; on one hand mathematically
it ensures the unique solvability of (1.2.17)-(1.2.18), on the other hand physically it ensures
that scattered field us is outgoing.
Another important class of scattering problems which is main objective of this thesis,
often called volumetric scattering, consists in the scattering of time-harmonic acoustic
waves by a penetrable scatterer, where the incoming waves penetrate the obstacle and the
interior structure of the obstacle strongly influences the scattering process [36, pp.211-214].
In such a case the total field u satisfy
∆u(x) + κ2 n2 (x)u(x) = 0 (1.2.20)
in Rd , d = 2, 3. Here, it is important to note that refractive index n(x) varies arbitrarily in

the interior of inhomogeneity Ω and takes value one outside the scatterer with discontinuity
across the interface ∂Ω. Thus, the problem of scattering by penetrable inhomogeneous
media results in a Helmholtz equation with a variable coefficient. In conclusion, scattering
problem under consideration is described as: for a given incident field ui , satisfying (1.2.16),
find the scattered field us ∈ C 2 (Rd ), d = 2, 3, such that [94, page 60]
u(x) = ui (x) + us (x) (1.2.21)
∆u(x) + κ2 n2 (x)u(x) = 0, (1.2.22)
∆ui (x) + κ2 ui (x) = 0, (1.2.23)

s
(d−1)/2 ∂u s
lim r − iκu = 0. (1.2.24)
r→∞ ∂r
In the next subsection, we discuss the fundamental solution of Helmholtz equation
in two and three dimensions which serve as a basis for integral equation formulation of
equations (1.2.21)- (1.2.24), on which we base our thesis.
1.2.3 Fundamental Solution

The fundamental solutions play a very decisive role in the integral equation based methods.
A comprehensive survey on the fundamental solution of elliptic equation due to Miranda
9
can be found in [90]. In general, it is not easy to obtain the fundamental solution explicitly,
however, for free space Helmholtz equation explicit formula is known.
The fundamental solutions for free space Helmholtz equation is also referred as Green’s
function or point source solution Gκ satisfying
∆Gκ (x, y) + κ2 Gκ (x, y) = −δ(x − y) for x, y ∈ Rd ,
where δ(x − y) is the Dirac delta function (distribution) defined by the properties

0, if x 6= y

δ(x − y) =
∞, otherwise,

and Z
f (x)δ(x − y)dx = f (y), for y ∈ Rd .
Rd
Here, f (x) is an arbitrary function. Note that the function Gκ (x, y) satisfy Helmholtz
equation only in Rd \ {x}. Explicit formula for fundamental solution Gκ (x, x0 ), subject to
the radiation condition (1.2.24), in two and three dimensions is given by

 i H10 (κ|x − y|), in R2

0 4
Gκ (x, x ) = (1.2.25)
 1 eiκ|x−y| ,

in R3 .
4π |x−y|
Here H10 , denotes the cylindrical Hankel function of the first kind of order zero [36, 37].
The Hankel function is also known as a cylindrical Bessel’s functions of third kind and
defined by complex linear combination of cylindrical Bessel’s functions of the first kind
Jν (x) and second kind Yν (x):
Hν1 (x) = Jν (x) + iYν (x).
We should note that in both the dimensions 2D and 3D, fundamental solution |Gκ (x, y)| →
∞ as x → y and become singular when x = y as explained in [36]. Moreover, it satis-
fies the Sommerfeld radiation condition (1.2.24) uniformly for y in compact sets, when
|x| = r → ∞.
10
1.2.3. Fundamental Solution
Now, we state few definitions and results which ensure the existence of integral of these
fundamental solutions over scattering region Ω and its boundary ∂Ω.
Definition 1.2.1. Let K : Ω × Ω → C be a function defined and continuous for all

x, y ∈ Ω ⊂ Rd , x 6= y. Then the function K(x, y) is said to be weakly singular kernel if
there exists positive constant M and α ∈ (0, d] such that
K(x, y) ≤ M x − y α−d , x, y ∈ Ω ⊂ Rd , x 6= y.

(1.2.26)
Theorem 1.2.2. [77, Theorem 2.22] Suppose K(x, y) denotes weakly singular kernel de-
fined on Ω × Ω. Then the linear operator A : C(Ω) → C(Ω), defined by
Z
(Aψ)(x) = K(x, y)ψ(y)dy x ∈ Ω,
Ω
is a compact operator on C(Ω).
Remark 1.2.1. The fundamental solution Gκ (x, y) is weakly singular kernel function in
Rd , d = 2, 3.
Definition 1.2.3. Let ϕ be a real or complex valued function defined on Ω ⊂ Rd . Then

the function ϕ is said to be uniformly Hölder continuous with Hölder exponent 0 < α ≤ 1
if there exists a constant C > 0 such that
|ϕ(x) − ϕ(y)| ≤ C|x − y|α (1.2.27)
for all x, y ∈ Ω ⊂ Rd .
By C p,α (Ω) we denote the space of all bounded functions whose p-th derivative are
uniformly Hölder continuous with exponent α. It is a Banach space equipped with the
norm:
If p = 0
k ϕ k0,α :=k ϕ kα,Ω := sup |ϕ(x)| + |ϕ|α,Ω , (1.2.28)
x∈Ω
and if p 6= 0
k ϕ kp,α =k ϕ k∞ + k ∇ϕ kp−1,α , (1.2.29)
11
where
|ϕ(x) − ϕ(y)|
|ϕ|α,Ω = sup (1.2.30)
x,y∈Ω |x − y|α
x6=y
is a Hölder semi-norm and this is the smallest constant for which (1.2.27) hold and is
named as a Hölder constant of ϕ.
Remark 1.2.2. For α < β each function ϕ ∈ C 0,β (Ω) is also belong to C 0,α (Ω).
Remark 1.2.3. If Hölder exponent α = 1, then the function is called Lipschitz continuous
and any uniformly Hölder or Lipschitz continuous function is also uniformly continuous.
1.2.4 Integral Equation Formulation

In this section, we reproduce an equivalent integral equation formulation corresponding to
partial differential equation formulation of scattering by penetrable inhomogeneous media
(1.2.21)- (1.2.24). We start by stating Green’s integral theorem which is a very basic
tool in the integral equation formulation of partial differential equation. We assume that
inhomogeneous region Ω is a bounded domain of class C 2 and ν denotes the unit normal
vector to the boundary ∂Ω directed into the exterior of Ω.
Theorem 1.2.4. Green’s first identity

If u ∈ C 1 (Ω) and w ∈ C 2 (Ω), then
Z
∂w
u∆w + ∇u · ∇wdx = u ds.
∂ν
∂Ω
Theorem 1.2.5. Green’s second identity

If u, w ∈ C 2 (Ω), then
Z Z
∂w ∂u
(u∆w − w∆u)dx = u −w ds.
∂ν ∂ν
Ω ∂Ω
The following Green’s representation theorem allows to represent smooth functions in

a smooth bounded domain by a superposition of certain potentials.
12
1.2.4. Integral Equation Formulation
Theorem 1.2.6. Green’s representation theorem

If u ∈ C 2 (Ω) ∩ C 1 (Ω) then for x ∈ Ω we have Green’s formula
Z
∂u ∂Gκ (x, y)
u(x) = (y)Gκ (x, y) − u(y) ds(y) (1.2.31)
∂ν ∂ν(y)
∂Ω
Z
− ∆u(y) + κ2 u(y) Gκ (x, y)dy, x ∈ Ω,
Ω
where the domain integral exists as an improper integral. In particular, if u is a solution

of Helmholtz equation
∆u + κ2 u = 0 in Ω,
then Z
∂u ∂Gκ (x, y)
u(x) = (y)Gκ (x, y) − u(y) ds(y), x ∈ Ω.
∂ν ∂ν(y)
∂Ω
We may change the hypothesis and assume u ∈ C 2 (Ω) ∩ C(Ω), in this case we have to
assume the existence of the normal derivative as a limit for x ∈ ∂Ω, i.e.,
∂u
(x) = lim ν(x) · ∇u(x − hν(x)), x ∈ ∂Ω.
∂ν h→+0
Green’s formula (1.2.31) provides an amazing finding that any function u ∈ C 2 (Ω) ∩
C 1 (Ω) can be expressed as a sum of three potentials namely a single layer potential
Z
u(x) = (Sϕ)(x) = ϕ(y)Gκ (x, y)ds(y), x ∈ / ∂Ω, (1.2.32)
∂Ω
a double layer potential

Z
∂Gκ (x, y)
v (x) = (Dϕ)(x) = ϕ(y) ds(y), x ∈
/ ∂Ω, (1.2.33)
∂ν(y)
∂Ω
and a volume potential

Z
w (x) = (V ϕ)(x) = ϕ(y)Gκ (x, y)dy, x ∈ Rd , (1.2.34)
Ω
with density ϕ. Notably, single and double layer potential solve the homogeneous Helmholtz
equation in Ω and in Rd \ Ω while volume potential solve the non-homogeneous Helmholtz
13
equation and they also satisfy the Sommerfeld radiation condition (1.2.24). Most inter-
estingly, any solution u(x) of Helmholtz equation in any arbitrary shaped domain can be
expressed solely as a combination of single and double layer potentials with density ϕ = u
∂u
and ϕ = ∂ν
respectively [36, page-39].
The integral equation formulation for scattering by penetrable inhomogeneous media

(1.2.21) -(1.2.24) rely heavily on the smoothness properties of volume potential (1.2.34).
Here, we present few important results which concern with the investigation of this poten-
tial.
Lemma 1.2.4. [73, Lemma 3.7] Let ϕ : Ω → C be a piecewise continuous. Then the
volume potential w ∈ C 1 (R3 ) and
Z
∂w ∂Gκ
(x) = ϕ(y) (x, y)dy, x ∈ R3 , i = 1, 2, 3.
∂xi ∂xi
Ω
For proof we refer to [73].
Theorem 1.2.7. If density ϕ ∈ C 0,α (Ω), then the volume potential w ∈ C 2 (Ω)∩C ∞ (R3 \Ω)
and 
−ϕ,

in Ω
∆w + κ2 w = (1.2.35)
3
0, in R \ Ω.

Furthermore, if we consider a domain Ω0 of class C 2 such that Ω ⊂ Ω0 and extend ϕ by

zero in Ω0 \ Ω then for x ∈ Ω and i, j ∈ {1, 2, 3}, we have
∂ 2w ∂ 2 Gκ
Z Z
∂Gκ
(x) = (x, y)[ϕ(y) − ϕ(x)]dy − ϕ(x) (x, y)νj (y)ds(y). (1.2.36)
∂xi ∂xj ∂xi ∂xj ∂xi
Ω0 ∂Ω0
For proof we refer to [73].

Now using Green’s theorem and properties of volume potential w, we show that solving
scattering problem (1.2.21)-(1.2.24) is equivalent to solving Lippmann-Schwinger integral
equation [36]
Z
2
u(x) + κ Gκ (x, y)m(y)u(y)dy = ui (x), for x ∈ Rd , (1.2.37)
Rd
14
1.2.4. Integral Equation Formulation
for total field u. Here m(x) = 1 − n2 (x). Note that the function m(x) takes the value zero
outside the scatterer Ω and hence it can be discontinuous across the boundary ∂Ω. Since,
m(x) has compact support and kernel Gκ (x, y) is weakly singular, therefore, the integral
in equation (1.2.38) exists as an improper integral. Moreover, for all x ∈ Rd
Z Z
Gκ (x, y)m(y)u(y)dy = Gκ (x, y)m(y)u(y)dy.
Rd Ω
Thus, equation (1.2.37) can be made to read as

Z
2
u(x) + κ Gκ (x, y)m(y)u(y)dy = ui (x), for x ∈ Rd . (1.2.38)
Ω
In the following theorem we reproduce equivalence between PDE and integral equation
formulation in three dimensions, one can readily adapt these ideas to obtain equivalence
for the two dimensional case.
Theorem 1.2.8. Let u ∈ C 2 (R3 ) be a solution of the scattering problem (1.2.21)-(1.2.24),

then u is a solution of (1.2.38). Conversely, let u ∈ C(R3 ) is a solution of (1.2.38) then
u ∈ C 2 (R3 ) and u is a solution of (1.2.21)-(1.2.24).
Proof. First we assume u ∈ C 2 (R3 ) be a solution of the scattering problem (1.2.21)-(1.2.24).

Consider an arbitrary point x ∈ R3 and then choose an open ball B of radius r with exterior
unit normal ν such that the support of m and x contained in B. Now applying Green’s
representation formula (1.2.31) to u, we have
Z
∂u ∂Gκ (x, y)
u(x) = (y)Gκ (x, y) − u(y) ds(y)
∂ν ∂ν(y)
∂B
Z
− ∆u(y) + κ2 u(y) Gκ (x, y)dy
B
It can be readily derive that total field u satisfies
∆u + κ2 u = mκ2 u.
Thus,
Z
∂u ∂Gκ (x, y)
u(x) = (y)Gκ (x, y) − u(y) ds(y) (1.2.39)
∂ν ∂ν(y)
∂B
15
Z
2
−κ Gκ (x, y)m(y)u(y)dy.
B
An application of Green’s second theorem (1.2.5) to ui and Gκ (x, y) gives

!
Z i
∂u ∂G κ (x, y)
ui (x) = (y)Gκ (x, y) − ui (y) ds(y). (1.2.40)
∂ν ∂ν(y)
∂B
In addition, radiation condition (1.2.24) implies that

!
∂us
Z
∂G κ (x, y)
lim (y)Gκ (x, y) − us (y) ds(y) = 0. (1.2.41)
r→∞ ∂ν ∂ν(y)
∂B
i s
Using u = u + u , and (1.2.40)-(1.2.41) in (1.2.39) we conclude that u satisfies (1.2.38).
Conversely, let u ∈ C 2 (R3 ) be a solution of Lippmann-Schwinger equation (1.2.38) and we
define Z
s 2
u (x) = −κ Gκ (x, y)m(y)u(y), x ∈ R3 . (1.2.42)
R3
s
Clearly, u satisfies Sommerfeld radiation condition (1.2.24) as the Green’s function Gκ (x, y)
satisfies Sommerfeld radiation condition (1.2.24) uniformly with respect to y on compact
set and m has compact support. Indeed, integral (1.2.42) is an acoustic volume potential
and therefore from the properties of volume potential (Theorem 1.2.7), when we apply
(∆ + κ2 ) on us we obtain
∆us + κ2 us = κ2 mu. (1.2.43)
Since, ∆ui + κ2 ui = 0, therefore,
∆u + κ2 u = (∆ui + κ2 ui ) + (∆us + κ2 us ) = κ2 mu,
and this complete the proof.
Question of existence and uniqueness for the scattering problem (1.2.21)-(1.2.24), has
been addressed by several authors. For the constant inhomogeneity case, the integral
equation (1.2.38) admits unique solution established in [37, pp.100-101];[78]. The case
when m ∈ C 1 has been discussed in [36]. For more general case when m is allowed to
be discontinuous across the the surface of scatterer, that is the case here, existence and
uniqueness results have been studied by several authors; see [85, 86, 104, 105] for references.
16
1.3. Brief Review of Numerical Methods for Scattering Problems
Remark 1.2.5. 1. Once the total field u known on the scatterer Ω, using equation
(1.2.38), it can be evaluated at any point in the space.
2. Since, the function m(x) vanishes outside the scatterer Ω therefore, integrand of vol-
ume integral in equation (1.2.38) can be extended in whole Rd without compromising
smoothness of integrand. Moreover, for all x ∈ Rd
Z Z
Gκ (x, y)m(y)u(y)dy = Gκ (x, y)m(y)u(y)dy.
Rd Ω
1.3 Brief Review of Numerical Methods for Scattering

Problems
In this section we review briefly the strengths and limitations of different numerical tech-
niques for numerical simulation of scattering problems. The quest for effective numerical
method for scattering problems, specially which find real life application, is far from over
and it still remains an area of active and ongoing research in computational and applied
mathematics. With the advancement of computational resources, many tremendous in-
novative and useful numerical methods have been proposed to solve these problems. A
plenty of methodologies have been proposed for the numerical solution of scattering prob-
lems, nevertheless, the most successful numerical methods can be broadly classified into two
categories namely differential equation based method and integral equation based method.
In the following sections, we briefly review both these methodologies and try to highlight
their strengths and limitations.
1.3.1 Differential Equation Based Methods

Differential equation based methods directly solve the partial differential equation model
of scattering problems by discretizing a region of space enclosing the scatterer. A variety of
algorithms have been proposed such as finite-difference methods, finite element methods,
method of moment [18] etc. Due to availability of simple time-stepping schemes, the finite
difference methods are very popular in time-domain simulations (see for example [81, 100]
17
and references therein). Although, finite difference schemes are probably the easiest to
implement and lead to sparse matrix but it suffer from problems in treating complex shaped
obstacle, specially if high-order convergence is desired. On the other hand, in contrast with
finite difference methods, finite element methods are more difficult to implement but it can
easily handle the complex geometry, a feature that, in fact, largely justifies its popularity
[11, 66].
Although, these methods have the advantage that, the resulting linear system is sparse
therefore, by avoiding calculations on zeros, one can significantly speed up computations
and reduce the memory requirements. Since, scattering problems are described in an
unbounded domain and of course one can not discretize the entire domain. Therefore, one
must replace their infinite physical domain with finite computational domain. To do this,
usually the unbounded domain is truncated by an artificial boundary enclosing the region
of inhomogeneity which yield a bounded computational domain. Of course, by reducing
the size of the truncated domain one can certainly reduce the computation cost, but in
order to satisfy Sommerfeld radiation condition (1.2.24), a large computational domain
is required with an appropriate boundary condition on the artificial boundary. Therefore,
this procedure gives rise to a large number of unknowns and correspondingly to large linear
systems. On the other hand, the newly introduced boundary condition on the artificial
boundary should make boundary as transparent as possible, or in other words, all scattered
waves leave the domain without any reflection back into the computational domain.
Although, several artificial boundary conditions have been proposed to achieve this kind
of property such as, absorbing [44, 47, 76], non-reflecting [56, 57, 58, 69], perfectly matching
layer boundary conditions [13, 30, 32, 35] etc, but none of these boundary conditions yield
zero reflection for all angle of incidence. Therefore, the error associated with such boundary
conditions typically dominates the error in the computed solution and still it continue as
a main source of errors in differential equation base simulations. Moreover, differential
equation based method also suffers from significant numerical dispersion/pollution errors
[12, 67], which cause a wave to have a different phase velocity on a computational grid
compared to the exact solution. More precisely, for fixed mesh length h, the numerical
18
1.3.2. Integral Equation Based Methods
accuracy deteriorates rapidly as wave number κ increases. Although, this problem can be
alleviated by enforcing the condition κ2 h < 1, but this condition entails a huge system of
linear equations which has become highly indefinite as wave number κ increases.
These problems inhibit the use of the differential equation methods for accurately com-
puting electromagnetic/acoustic scattering, specially for large scatterer.
1.3.2 Integral Equation Based Methods

As compared to differential equation based method, integral equation based methodology
rely on the simulation of the equivalent integral equation formulation of the corresponding
partial differential equation of the scattering problems. Most distinguished feature of inte-
gral equation based method is that the Sommerfeld radiation condition (1.2.24) is directly
incorporated in the integral formulation. Hence, unlike differential equation based solver,
they do not require the introduction of artificial boundaries. Indeed, in the derivation
of equivalent integral equation formulations, the radiation condition (1.2.24) is explicitly
imposed by simply choosing an appropriate outgoing Green’s function [36, 37]. Thus, the
computational domain is confined to the scattering obstacles and the number of unknowns
arising upon discretization of such equations are relatively small compared to differential
equation based methods, especially in surface scattering applications. In addition, an inte-
gral equation solver propagates the field analytically with a closed-form Green’s function,
as opposed to a differential equation based methodology that propagates the field on a nu-
merical grid. Hence, there are usually fewer grid-dispersion errors for an integral equation
solver compared with a differential equation based method.
The major drawback of this approach is that, upon desensitization, it leads dense
linear system which requires O(N 2 ) memory storage, where N is the number of unknowns.
This linear system can be solved by using either a direct solver or an iterative solver. A
direct solver, such as the LU decomposition (Gaussian elimination), entails a computational
complexity of O(N 3 ). On the other hand, in contrast with direct solver, iterative methods
produces solution only in O(Niter N 2 ) computational cost, where O(N 2 ) is the cost of matrix
vector multiplication in each iteration and O(Niter ) is the number of iterations to achieve
19
desired convergence. Normally, in most of the scattering calculations number of iterations

are very less in comparison to number of unknowns. Therefore, the cost of iterative solver
can be taken as the cost of matrix vector multiplication, that is, O(N 2 ). Although, it is
obviously advantageous to solve the dense linear system using an iterative solver but still
O(N 2 ) cost precludes their application to the analysis of scattering from large structures.
Another major difficulty associated with integral equation approach is the singularity of
Green’s function in the scattering region which pose challenges for accurate integration.
Thus, specialized quadrature formulas must be used to attain acceptable accuracy without
an undue computational burden.
Indeed, the integral equation based solvers are not competitive unless a specialized
strategy is used to reduce the complexity of matrix vector multiplications. In an effort
to remedy this difficulty, a significant part of all recent researches relating to the numer-
ical solution of integral equations in this context has focused on the fast techniques that
can reduce the computational labor of the matrix-vector multiplications in O(N p log N )
operations with p significantly less than 2.
In the past, many fast methods have been proposed to expedite the matrix-vector mul-
tiplication of the iterative solver. These techniques made Integral equation based solvers
competitive in terms of computational complexity and memory requirements with those
of the differential equation solvers. These include the Conjugate Gradient Fast Fourier
Transform Method (CG-FFT) [17, 28, 82, 109], the Fast Multipole Method [34, 61, 96], the
Wavelet Expansion Method (WEM) [103, 110], the Impedance Matrix Localization Scheme
(IML) [26], the Adaptive Integral Methods (AIM) [15, 14]. The key concept behind these
novel methodologies is to obtain a sparse matrix starting with the classical method of
moments (MoM) solution procedure. In WEM and IML, special set of basis functions are
used to represent the unknown quantity. This generate a sparse matrix, which in turn,
accelerate the matrix vector multiplications. On the other hand, however, in FFT, AIM,
FMM sparse matrices are obtained by handling the influence of the Green’s function in a
novel way. Since, wavelets are scale invariant and boundary integral equation of electro-
statics belongs to the class of Calderon Zygmund operators [4], where the integral kernel
20
1.3.2. Integral Equation Based Methods
is in infinitely smooth therefore WEM are well-suited for solving static or low frequency
(elliptic) problems. In these problems, application of wavelet basis functions sparsify the
resulting matrices to O(N log N ) elements, reducing the computational complexity from
order O(N 2 ) to O(N log N ). However, for wavelike problems, use of wavelet basis func-
tions in WEM reduces the solution time by a constant factor but not the computational
complexity [103]. This is because, the associated integral equation has an oscillatory ker-
nel or, in other words, wavelike problems are not scalar invariant. The IML technique,
on the other hand, work well for smooth geometries, but not for non-smooth geometries.
Perhaps, the Fast Multipole Method (FMM) and its variants hold most promise in provid-
ing a fast method whose applicability is not restricted by the geometry of scatterer. It is
a hierarchically based method which relies on directional approximations of the Green’s
function using addition theorems and fast translations. Although, FMM yields O(N log N )
complexity, however, the large constant factor in this asymptotic bound require rather a
large value of N before any advantage over the O(N 2 ) complexity is observed. Indeed,
this large asymptotic bound make it incompetent to other fast algorithms in certain cases.
In addition, while FMM is well suited for the solution of low-frequency problems but it
becomes unstable at higher frequencies [41, 79].
Another class of fast algorithms are based on Fast Fourier transform (FFT) which
reduces the computational complexity of the iterative solvers from O(N 2 ) to O(N log N ).
In these methods, the convolution in the integral equation is computed by means of FFT
and multiplication in the Fourier space by utilizing the fact that the Green’s function of the
free space Helmholtz equation is translation invariant. When the FFT is incorporated into
the conjugate gradient (CG) method to compute the solution then the resulting method is
called the CG-FFT method. Unfortunately, in case of singular integral kernel CG-FFT does
not perform well with respect to accuracy. In addition, the CG-FFT requires the meshing
of the scattering region on uniform grids, which in turn, preclude its applicability for
complex geometry. As a remedy to this problem, Adaptive Integral Method (AIM) based on
computation of an equivalent source distribution on Cartesian grid is introduced. Indeed,
this method retains the advantages of CG-FFT and offers excellent modeling capability
21
and flexibility for the analysis of scattering from very complex structures. It reduces the
the computational cost to O(N 3/2 log N ) and O(N log N ) for surface and volume scattering
calculations respectively. Though, AIM offer significant reduction in computational cost,
its memory requirements tend to be very large, specially for surface scattering calculation.
An excellent, efficient extensions of AIM in three dimension were realized in [23, 24] where
the cost of scattering calculations was reduces to O(N 4/3 log N ). An important feature
of this method is that, equivalent sources are distributed on the uniform Cartesian grid
located on the faces of a cube enclosing the scattering region, which in turn, allow the use
of FFT to accelerate the integral computations. The same idea used in [6, 7] for accelerated
computation of volumetric scattering from thin structures. In light of significant advantage
of fast algorithm introduced in [23, 24] over aforementioned fast strategies, we have chosen
this algorithm to be used and further enhanced in this thesis for full volumetric scattering
calculations.
For more detail description and numerical analysis of integral equation based method
we refer [9, 37]. An extensive survey and bibliography of numerical methods for wave
scattering problem can be found in [31, 92].
1.4 Previous Work

Although, several differential equation based methodologies have been proposed for the
solution of acoustic scattering by penetrable inhomogeneous media, for example [12, 52,
74, 89, 88]. All most all methodologies suffer from the problem mentioned in the previous
section. To the best of our knowledge, none of the methods report high-order convergence
in the presence of material discontinuity across the scatterer interface, which we have
achieved in the proposed thesis. Since, this thesis is devoted for the development of an
accurate and efficient solution methodology for the equivalent integral equation formulation
of this problem, therefore, in the following paragraph, we provide a succinct overview of
the recent work for integral equation based simulations, that is, method reported for the
solution of Lippmann-Schwinger integral equation (1.2.38).
Significant effort has been devoted for the simulation of Lippmann-Schwinger integral
22
1.4. Previous Work
equation (1.2.38), including both direct and iterative methods. Direct solver is particularly
well suited for multiple right hand side and for small size problem. As, we have mentioned
in the previous section use of direct solvers require large amount of memory and its compu-
tational cost grows like O(N 3 ), which would severely limit the size of problems that could
be solved in practice, specially in three dimensions. As a remedy to this problem, couple of
fast direct solver [1, 5, 29, 46, 54] have been introduced. The algorithms proposed in these
solvers are hierarchical in nature, and, produces the scattering matrix which transforms
the incident field into the scattered field. The basic idea is to divide the entire computa-
tional domain into sufficiently small non-overlapping subdomains and then compute the
scattering matrix in each subdomains independent of others. Then, a merging, followed by
subdivision of computational domain, accounts for multiple scattering interactions between
subdomains. These algorithms reduce the cost of direct solver from O(N 3 ) to O(N 3/2 ) in
two dimension and would be O(N 2 ) in three dimensions. Still these solvers require large
amount of memory as compared to iterative solver and provide low order convergence for
discontinuous scatterer.
The strategy in this wok is to employ an iterative solver for the solution of Lippmann-
Schwinger integral equation (1.2.38) in two and three dimensions.
Perhaps the most natural way to solve the Lippmann-Schwinger integral equation (1.2.38)
equation by means of an iterative solvers, is to discretize the integral equation on an uni-
form grid and use the FFT to obtain the convolution type integral in (1.2.38). Here, use
of FFT will reduce the cost of matrix vector multiplication of iterative solver such as con-
jugate gradient method, from O(N 2 ) to O(N log N ) and provide solution in fast manner
[106, 107]. Unfortunately, for discontinuous functions, this method yields only first or-
der convergence. In addition, singularities of the free space Green’s function also present
significant difficulties for this method.
As we enunciated in the previous section, Fast Multiple Method (FMM), is most widely
used algorithm for efficient solution of integral equations that describe electromagnetic
or acoustic scattering in two or three dimensions. However, for volumetric scattering
problem, which is the subject of this thesis, FMM finds very few applications due to the
23
difficulties discussed in the previous section. In [62], FMM in conjunction with FFT has
been applied for the solution of volumetric acoustic scattering problem where scattering
elements are defined on a regular grid. The regular arrangements of scattering element
coincide with the grid imposed by recursive FMM group subdivisions and at the finest
level neighboring interactions are obtained by means of FFT. This use of FFT at finest
level made possible to avoid the low frequency breakdown problem, the most substantial
issue arises in implementation of FMM. In addition, use of FFT reduces dramatically
the computational and storage overhead associated with FMM setup. In spite of several
advantage offered by this algorithm, requirement of regular scattering element, limits, its
applicability for complex geometry and due to singularity in the Green’s function, method
does not perform well from accuracy point of view.
In [102], Vainikko proposed two fast approximation methods of computational com-
plexity O(N log N ), for the solution of Lippmann-Schwinger integral equation (1.2.38) in
three dimension, where scatterer is assumed to be smooth. The first method is simple
cubature (quadratures) and notably provide second order convergence even in the presence
of jump discontinuity in the refractive index n(x) across the interface of scatterer. In the
second method, inhomogeneity is covered by the cube of size at least 2d, where d is the
typical diameter of inhomogeneity and then periodic solution is produced by cutting of
Green’s function and modifying integral equation (1.2.38). It is shown that this method
provide arbitrary high-order convergence for smooth scatterer. However, for discontinuous
scatterer, produces only first order convergence.
Liu and Gedney [83] proposed an O(N 2 ) locally corrected Nyström scheme for scattering
calculations in two dimensions. For volumetric scattering problem, this method is capable
to produce high-order convergence which is not limited by the smoothness of scatterer.
However, its quadratic computational cost made it incompetent for applications of large
scale problems.
In a series of papers [21, 22, 64, 65], stemming from the Ph.D. thesis of Hyde [64], Bruno
and Hyde proposed an iterative, fast high order method which is also applicable in the case
of discontinuous scatterers. In the two dimensions [21, 22], algorithm separate the convo-
24
1.4. Previous Work
lution integral into radial and angular integration by using change of polar coordinates and
addition theorem for two dimensional Green’s function. High-order approximation in the
radial integration is performed by Chebyshev polynomial approximation of smooth func-
tion multiplied by Hankel functions of integer order. High-order integration in the angular
variable is achieved by replacing the scattering object with its truncated Fourier series and
use of FFTs. On the other hand, in three dimensions [65], high-order approximation is
gained through the use of Trapezoidal rule and Fourier series approximation in Cartesian
coordinates. Although, in both the dimensions, methods are very efficient, capable of deal-
ing with non-smooth scattering geometries and requires only O(N log N ) computational
cost but it produce only second order convergence for discontinuous scatterers.
In the methodology introduced in [80], the whole inhomogeneity region is covered by
some bigger domain, and, then solution of Lippmann-Schwinger equation is obtained by
using scaled and shifted Gaussians centered on a uniform grid . The direct computation
of convolution with the Green’s function and Gaussians reduces computational cost of the
algorithm and provide numerical solution in fast manner. It has been shown that method
converges spectrally for smooth scatterers and proposed quadrature produces negligible
error. Since, spectral convergence of the algorithm depend on the smoothness of scatterer
and therefore method does not perform well for the case of discontinuous scatterer.
The approach in [8] is based on construction of a band-limited version of the Green’s
function and has a computational complexity O(N log N ). The interesting feature of this
approach is that, it is capable of automatically dealing with scattering objects of low
regularity and at higher level of accuracy, computational effort consumed by the algorithm
is of the same magnitude as straightforward FFT-implementations of the same grid-size.
Method suffers from accuracy problem for discontinuous scatterer and produces only second
order convergence.
Another approach for solution of acoustic volumetric scattering problem is introduced
in [6, 7], though, high-order convergent, is designed to be computationally efficient only
for “thin” scattering configurations. This scheme gains high-order convergence through a
combination of changes of parametric variables (in order to resolve the singularities present
25
in the Green’s function) and by suitably employing “partitions of unity” to yield smooth
and periodic integrand away from vicinity of target points. Our present approach, in
fact, is a non trivial extension of ideas presented in [6, 7] wherein we obtain a solver for
general scattering configurations that retains high-order accuracy even in the presence of
material discontinuity. To the best of our knowledge, this is the first fast integral equation
solver which provide high-order convergence for full volumetric scattering problem with a
discontinuous material interface.
1.5 Organization of The Thesis

Remaining thesis is organized as follows: In Chapter 2, we present detail description of
our algorithm for the solution of Lippmann-Schwinger integral equation (1.2.38) in two
dimensions. We first describe the basic component of our algorithm and then we explain
high-order integration scheme for regular (integral with smooth integrand) and singular
integrals. We conclude the chapter by discussing acceleration strategy for accelerated
computation of non-singular non-adjacent interactions. In order to demonstrate the effi-
ciency and high-order convergence of our two dimensional algorithm, variety of numerical
results are presented in Chapter 3. In Chapter 4, we extend our two dimensional algorithm
for solution of Lippmann-Schwinger integral equation (1.2.38) to three dimensions. Numer-
ical results obtained by implementation of our three dimensional algorithm for variety of
scatterers, which demonstrate the high-order convergence of our algorithm, presented in
Chapter 5. In Chapter 6, we derive the error bounds for the interpolations and quadrature
rule described in Chapters 2 and 4. In particular, our error analysis establishes theoret-
ically super algebraic convergence of the proposed quadrature scheme in a case in which
refractive index n(x) is smooth within the scatterer Ω. Finally, our conclusions and rec-
ommendations presented in Chapter 7.
26
Chapter
2
An efficient high-order integral
equation solver in two dimensions
In this chapter, we present an efficient high-order algorithm for the solution of Lippmann-
Schwinger integral equation (1.2.38) numerically in two dimensions. We discretize equation
(1.2.38) using Nyström scheme and the resulting linear system is solved by means of it-
erative generalized minimal residual method (GMRES) [97]. Since, the GMRES solver
requires evaluation of volume integral operator
Z
K[u](x) = Gκ (x, y)m(y)u(y)dy (2.0.1)
Ω
of Lippmann-Schwinger equation(1.2.38) in each iteration and the accuracy of method

depends on the accurate approximation of this integral. Therefore, it suffices to describe
an efficient integration scheme which approximates the integral (2.0.1) accurately.
The chapter is organized as follows: Section (2.1) presents basic component of our
numerical scheme where we start with representing scatter Ω in terms of overlapping co-
ordinate patches, then we discuss Partition of Unity and discretization of domain Ω. In
section (2.2), we introduce some notation and then decompose the volume integral op-
27
Chapter 2. An efficient high-order integral equation solver in two dimensions
erator (2.0.1) into local integrals over the overlapping coordinate patches. Section (2.3),
in turn, presents discussion in detail of our high-order efficient integration scheme for the
evaluation of integrals over patches. Finally, in section (2.4), we describe our acceleration
scheme for the fast evaluation of non-singular integrals.
2.1 Scatterer Parametrization, Nyström Nodes And

Partitions of Unity
The Nyström method under consideration, following the idea introduced in [6, 23], uses
overlapping coordinate patches (local charts, in the language of differential geometry) for
description of scatterer Ω. We note that, compactness of Ω (closure of Ω) guaranteed
existence of such a local charts for scatterer Ω. Thus we will start by describing the
scatterer Ω by a collection P = {Pk }K
k=1 of K overlapping coordinate patches, such that
1. Each Patch Pk is an open set in Ω for k = 1, · · · , K, and its union cover entire Ω,
that is,
K
[
Ω⊆ Pk .
k=1
Specific instance of this procedure for disc shape scatterer which we have used for
the numerical experiment in Chapter 3 is shown in figure (2.1).
2. The set Pk is homeomorphic to an open set Hk ⊆ (0, 1)2 in R2 , via smooth invertible
parametrization
ξk : Hk → Pk
(t1 , t2 ) 7→ ξk (t1 , t2 )
such that the determinant of Jacobians

h i
ξk0 (t1 , t2 ) = det ∂ξk (t1 ,t2 )
∂t1
∂ξk (t1 ,t2 )
∂t2
> 0, k = 1, · · · , K
(ξk0 : Hk → R) are smooth functions of t1 and t2 .
28
2.1. Scatterer Parametrization, Nyström Nodes And Partitions of Unity
In practice, for the sake of accuracy, existence of a significant region of overlap between
overlapping patches is desired. Hence, we prefer to use as minimum number of patches as
possible to cover Ω.
Now, we introduce a set of (N1 + 1) × (N1 + 1) Nyström nodes on the k-th patch, say
Tk , by taking the image of the equispaced grid {(i1 /N1 , i2 /N2 ) | 0 ≤ i` ≤ N` , ` = 1, 2} in
[0, 1]2 under the map ξk (see Figure (2.2)). Further, the union of these grid points, that is,
K
[
T= Tk , (2.1.1)
k=1
where
Tk = {ξk (i1 /N1 , i2 /N2 ) | 0 ≤ i` ≤ N` , ` = 1, 2} , (2.1.2)
defines the complete set of Nyström discretization points where we seek to approximate
the solution of equation (1.2.38).
Our algorithm is based on the decomposition of volume integral in (2.0.1) into sum of
K integrals over the overlapping coordinate patches Pk . To achieve this decomposition, we
utilize partitions of unity (POU) subordinated to covering P of Ω, i.e., set of non-negative
functions { ωk (x) : k = 1, · · · , K}, such that
1. ωk (x) is defined, smooth, and nonnegative in whole Ω
2. ωk (x) vanishes together with all of its derivatives in Ω \ Pk

K
P
3. ωk (x) = 1 ∀x ∈ Ω.
k=1
We announce this set of POU functions as the “fixed” partition of unity, in contrast with
certain “floating” partitions of unity that we will introduce in section (2.3.2). In fact,
floating partitions of unity play very crucial role in acceleration of integral computation
and its accurate approximation.
Indeed, partition of unity is a key component of our algorithm and construction of
such POU functions is not difficult. We build partition of unity by constructing functions
{ψk } satisfying first two conditions in the above. Then, POU functions ωk can be simply
29
(a) Computational grid on disc shape scatterer (b) Boundary Patch
(c) Boundary Patch (d) Interior Patch
Figure 2.1: The figure on the top left depicts the discretization for the scatterer coming
from a typical volumetric computational grid. Other three figures display the overlapping
coordinate patches and computational grid on it.
30
2.1. Scatterer Parametrization, Nyström Nodes And Partitions of Unity
Figure 2.2: A typical point in the set Tk for a patch Pk when N1 = 8, N2 = 4

.
obtained by defining it as
ψk
ωk = K
.
P
ψj
j=1
Hence, once we construct functions {ψk }, construction of POU functions ωk are straight-
forward. For the sake of accuracy, it is advantageous to use partitions of unity with small
derivatives, more precisely POU functions cannot have transition regions which are too
steep. However, this difficulty can be avoided by choosing an appropriate ψk and by in-
creasing the overlapping region among the patches Pk . In order to construct ψk , we utilize
another function χ(s, s0 , s1 ), depicted in figure (2.3), as a building block. As illustrated in
31
Figure 2.3: An illustration of the function χ(s) used as a building block in the construction
of partitions of unity.
Figure (2.3), function χ(s, s0 , s1 ) takes value one if s ≤ s0 for some real number s0 and
varying smoothly from 1 to 0 as s goes from s0 to s1 , for real number s1 and zero otherwise.
In our implementation, we have define χ(s, s0 , s1 ) as follows [23]

for s ≤ s0 ,



 1,


χ(s, s0 , s1 ) = exp 2e−1/r , for s0 < s < s1 , where r = |s|−s0
, (2.1.3)
 r−1 s1 −s0



0, for s ≥ s1 .

32
2.2. Decomposition of The Integral Operator
As an illustration, example of POU functions on individual patches and on whole scatterer

for disc shape geometry are presented in Figures 2.4 to 2.6. In fact, our procedure for
construction of POU is independent of scatterer geometry and using this procedure one
can construct POU for any complex shape domain. For example, Figure (2.7), presents
POU functions on complex shape geometry, namely, on bean shape geometry. In these
figures one can clearly see POU functions take value one in the non-overlapping region
and decaying smoothly towards zero along the boundary of those patches Pk where they
overlap.
2.2 Decomposition of The Integral Operator

An essential step of our algorithm involves decomposition of integral operator (2.0.1) into
integrals over the overlapping coordinate patches Pk . In this regard, we utilize partitions
of unity functions {ωk (x)|k = 1, · · · , K} subordinated to covering P. In detail, with the
help of POU we can decompose the evaluation of integral operator (2.0.1) at a point x as
a sum of K integrals as follows:
K Z
X
K[u](x) = Gκ (x, y)m(y)u(y)ωk (y)dy. (2.2.1)
k=1 P
k
Thus, breaking of the integral operator (2.0.1) into integrals over overlapping coordinate
patches is an important step towards decoupling of the interactions of various patches with
the target point x. Note here that for all those patches Pk whose closure does not intersect
with the boundary of Ω, that is, ∂Ω, the corresponding ωk vanishes along with all of its
derivatives on the boundary of the patch, for example see Figure (2.5). On remaining
patches, however, where one of the edges coincide with the boundary of Ω, POU functions
ωk clearly does not vanish and, in fact, attains the value 1 (see Figure 2.4 for examples). For
the clarity of presentation of the proposed algorithm, and to distinguish between patches
based on this criterion, we partition the set of patches P into two sets, namely, set of
boundary patches
n o
P B = Pk ∈ P k ∈ IB
33
Figure 2.4: Partition of unity corresponding to interior patch for disc shape scatterer.
34
Figure 2.5: Partition of unity corresponding to Interior patch for disc shape scatterer.
35
Figure 2.6: Partition of unity on disc shape scatterer.
36
Figure 2.7: Partition of unity on bean shape scatterer.
37
and set of interior patches

n o
I
P = Pk ∈ P k ∈ II ,

where
IB = {k|Pk ∩ ∂Ω 6= Ø}
and
II = {k|Pk ∩ ∂Ω = Ø},
define an index set for boundary and interior patches respectively. Note that P B ∩ P I = Ø
and P B ∪ P I = P. Now, clearly, the integrals under the summation in (2.2.1) can be
re-expressed as a sum of integrals over boundary and interior patches
XZ XZ
K[u](x) = Gκ (x, y)m(y)u(y)ωk (y)dy+ Gκ (x, y)m(y)u(y)ωk (y)dy. (2.2.2)
k∈IB P k∈II P
k k
Indeed, the performance of our iterative solver hinges on our ability to approximate the
integrals in (2.2.2) accurately, in a computationally efficient manner. In the next section,
we provide a detailed description of a fast, high order integration scheme in two dimensions
for the evaluation of these integrals..
Remark 2.2.1. 1. In this chapter, we often refer to x in equation (2.2.2) as a target

point whereas points y therein have sometimes been called source points.
2. While discussing the approximations on boundary patches, we identify the second

coordinate variable in [0, 1]2 with the transverse parameter. We, therefore, use the
notation t = (t1 , t2 ) to denote a point in the corresponding parameter space and
assume that the boundary of Ω coincides with t2 = 0.
2.3 Fast and Accurate Evaluation of Integrals

The difficulty in accurately computing integrals in (2.2.2) is significantly more when the
target point x lies in the integration patch Pk compared to the case when it does not.
Indeed, when target point occurs in the integration patch, that is, x ∈ Pk , owing to the
38
2.3. Fast and Accurate Evaluation of Integrals
singularity of the kernel Gκ (x, y) at y = x, the integrand is unbounded within the inte-
gration domain and direct use of a standard quadratures yield inaccurate approximations.
Thus, specialized quadrature rules must be developed and used to deal with such singular
integrals. The case when target point x lies out side the integration patch, that is, x 6∈ Pk ,
in contrast, does not present this challenge. To effectively present these two contrasting
scenarios, we further refine the index sets IB and II by introducing target point dependent
index sets
MB (x) = {k ∈ IB | x ∈ Pk }
and
MI (x) = {k ∈ II | x ∈ Pk } .
Now we can rewrite (2.2.2) as

 
X X Z
K[u](x) =  +  Gκ (x, y)m(y)u(y)ωk (y) dy
k∈MB (x) k6∈MB (x) Pk
 
X X Z
+ +  Gκ (x, y)m(y)u(y)ωk (y) dy. (2.3.1)
k∈MI (x) k6∈MI (x) Pk
Clearly, each of the integrals above can be rewritten in the parametric coordinates
Z ZZ
Gκ (x, y)m(y)u(y)ωk (y) dy = Gκ (x, ξk (t1 , t2 ))ϕk [u](t1 , t2 )ξk0 (t1 , t2 ) dt1 dt2 , (2.3.2)
Pk [0,1]2
where ϕk [u](t1 , t2 ) = m(ξk (t1 , t2 ))u(ξk (t1 , t2 ))ωk (ξk (t1 , t2 )), and ξk0 denotes the Jacobian
of the transformation ξk . This transformed integral into parametric variables offers two
significant advantages over integrals in (2.3.1). First, unknown density ϕk [u](t1 , t2 ) can
be easily discretized and second integration in parametric space is easy to perform rather
than the integration on arbitrary shape patches. Indeed, this transformation made our
algorithm independent of patch geometry which is one of the most important feature of
this algorithm.
Now we begin the discussion of high-order evaluation of integrals for the cases k 6∈
MB (x) and k 6∈ MI (x) in (2.3.1). In these cases, target point x does not lie in the in-
tegration patch and hence, integral kernel Gκ (x, y) remains non-singular throughout the
39
region of integration. Though adopting a single high-order approximation scheme for all
the integrals under the above choice is possible, we actually utilize two different quadratures
to take advantage of a more favorable behavior of the integrands in the later case when
k 6∈ MI (x). In this scenario, density ϕk [u] vanish to high order at the boundary of the
integration domain since corresponding POU function ωk vanishes to high-order along the
boundary of integration interval and therefore, the whole integrand has smooth periodic
extensions to R2 . It is well known that, the trapezoidal rule exhibits super-algebraic con-
vergence for smooth and periodic integrands, which we indeed employ to obtain accurate
approximations in this case. In contrast, when k 6∈ MB (x), a straightforward use of trape-
zoidal rule does not produce high-order accuracy as the integrands do not vanish at t2 = 0.
We overcome this minor difficulty by utilizing the trapezoidal rule for the integration with
respect to t1 -variable while employing a composite Newton-Cotes quadrature in transverse
variable t2 , to achieve approximations whose rate of convergence directly depends on the
order of Newton-Cotes used and could be enhanced arbitrarily as long as the smoothness
of m within Ω allows it.
As mentioned above, integral kernel Gκ (x, y) become singular when target point x
happens to be in an integration patch, therefore, integrands in (2.3.1) corresponding to
the cases when k ∈ MB (x) and k ∈ MI (x) are singular. To achieve rapidly convergent
approximations, we rely on analytic resolution of singularities through suitable changes of
parametric variables and application of high-order quadratures to resulting smooth inte-
grands. In addition, we also localize the region where such coordinate transformations are
affected to a small neighborhood of the singular point using a suitable smooth compactly
supported cut-off function. Indeed, we see that
ZZ ZZ
Gκ (x, ξk (t)) · · · dt = Gκ (x, ξk (t)) · · · η(t; ξk−1 (x)) dt (2.3.3)
[0,1]2 [0,1]2
ZZ
+ Gκ (x, ξk (t)) · · · (1 − η(t; ξk−1 (x))) dt,
[0,1]2
where η(·; t0 ) is a C ∞ function that it is compactly supported in a neighborhood of the

point t0 ∈ [0, 1]2 while η ≡ 1 in a smaller neighborhood of t0 . Note that, similar to fixed
40
2.3. Fast and Accurate Evaluation of Integrals
POU, discussed in section (2.1), the pair (η(t; ξk−1 (x)), 1 − η(t; ξk−1 (x))) is also a partition
of unity. Since, support of (η(t; ξk−1 (x)) depends on the location of target point x and
our algorithm demands such partition of unity for each target point x, therefore, we refer
collection of all such partitions of unity as “Floating” partitions of unity.
This localization of singularity as seen in the first integral on the right hand side of
(2.3.3) brings in a two-fold benefit, namely, (a) it limits the relatively expensive treatment
of singularity which, among other computational challenges, also demands an interpolation
of grid-data to off-grid quadrature points to a small integration domain, and (b) it allows
for additional speed-up in the computation of non-singular second term on the right hand
side of (2.3.3). Note that this regular integral, of course, can be integrated to high-order
using the same numerical quadratures that we apply in the case when target point x does
not lie in the integration patch, as explained above.
It is straightforward to see that the direct application of the integration scheme that we
have described above, leads to a computational complexity of O(N 2 ), where N = |T| is the
size of the set T, the total number of quadrature points in the Nyström scheme. However,
as we outline below, one can improve the computational complexity of this methodology
to O(N 1+δ/2 log N ), where δ ∈ (0, 1], by breaking the overall computation of these inte-
grals into two parts – a relatively small but specialized calculation in the neighborhood
of singularity and remaining non-singular contributions arising from the voluminous bulk.
Toward this, we begin by bounding the inhomogeneity Ω by a square cell C of side length
A. This cell is further partitioned into L2 identical cells Cij (i, j = 1, · · · , L) of side length
H = A/L such that the bounding cell C contains L of these smaller cells along its sides (for
an example, see Figure 2.9). We assume, without loss of generality, that for the chosen side
length A, cells Cij do not admit inner acoustical resonance. This, of course, can be easily
ensured by necessary adjustment in the choice of A so that −κ2 is not a Dirichlet eigenvalue
of the Laplace operator in the cell Cij . These cells are used to break the computation of
integrals appearing in equation (2.3.1) into adjacent and non-adjacent calculations which,
as pointed out above, is primarily motivated by our desire for the method to have a more
favorable computational complexity than the quadratic cost in the number of unknowns.
41
We say a source point y is adjacent to a fixed target point x ∈ Cij , if it belongs to the
set N (x) defined by
N (x) = {y | y ∈ Ckl for some k, l satisfying |k − i| ≤ 1, |l − j| ≤ 1} . (2.3.4)
Obviously, we say y is non-adjacent to x if y 6∈ N (x). Based on this, we separate the

integrals over Pk in (2.3.1) into integrals over Pk ∩ N (x) accounting for the adjacent
contributions and the computationally large non-adjacent contributions for which a variant
of the two face FFT based acceleration strategy introduced in [23] is employed for efficient
calculations. Furthermore, if x ∈ Pk ∩ N (x) then we call adjacent interaction as singular
adjacent interaction and non-singular adjacent otherwise.
We, thus, split integrals in (2.3.1) to rewrite the expression for K[u](x) as
  
 X X Z XZ 
K[u](x) =  +  + Gκ (x, y)m(y)u(y)ωk (y) dy
 P ∩N (x) P ∩N (x)c
k∈MB (x) k6∈MB (x) k k∈IB k
  
 X X Z XZ 
+  +  + Gκ (x, y)m(y)u(y)ωk (y) dy,

k∈MI (x) k6∈MI (x) Pk ∩N (x) k∈II Pk ∩N (x)c 
(2.3.5)
where N (x)c denotes the compliment of N (x). Now, in the following, we provide a succinct
overview of the computational cost of the algorithm.
The computational cost of carrying out the non-singular adjacent calculations for all
target points x ∈ T, is
Λadj = N × O N δ/2 ,

(2.3.6)
where O(N δ/2 ) denotes the number of additional points supplied for the evaluation of
singular integral over region Pk ∩ N (x) up to the desired accuracy and 0 < δ ≤ 1.
The cost involved in computing the contributions to integrals coming from Pk ∩ N (x)c
for all target points x ∈ T, as we will explain later in section (2.4) is given by
2
1/2 1/2 3/2 3 3/2 N
Λnonadj = O(LN log(LN )) + O(N /L ) + O(N /L) + O . (2.3.7)
L2
This, of course, suggests that by choosing the parameter L = O(N 1/2 ) in (2.3.7), the com-
putational complexity of the non-adjacent calculations reduces to the desired O(N log N ).
42
2.3.1. Singular Integration
Thus, the total cost of computing K[u] at all Nyström nodes is given by
Λ = Λnonadj + Λadj ,
= O(N log N ) + O(N 1+δ/2 )
= O(N 1+δ/2 ). (2.3.8)
Note that, the number δ ∈ (0, 1] is a constant that can be chosen to control the
complexity and accuracy of the algorithm. The value of δ in Equation (2.3.8) is mostly
determined by the radius of the floating partition of unity η. By using η that shrink faster,
we can speed up the algorithm at the cost of letting the error decrease with lower rate of
convergence. On the contrary, by using η whose support shrinks slower, we can enhance
the accuracy of the approximation, but overall complexity of the algorithm will increase.
Thus, for limited regularity of the solution there is always a trade off between accuracy and
complexity. However, if material properties are infinitely differentiable, so is the solution
within the scatterer Ω, then the proposed algorithm converges to high-order for any small
values of δ arbitrarily close to zero. In fact, if we choose δ by the following formula
log(log N )
δ=2
log N
then for infinitely differentiable solution we achieve desired O(N log N ) computational
complexity while maintaining high-order accuracy.
2.3.1 Singular Integration

Recall that when the target point x belongs to the integration patch Pk , we break the
integral as follows:
Z ZZ
Gκ (x, y)m(y)u(y)ωk dy = Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)η(t; ξk−1 (x)) dt
Pk [0,1]2
ZZ
+ Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)(1 − η(t; ξk−1 (x))) dt
ξk−1 (Pk ∩N (x))
Z
+ Gκ (x, y)m(y)u(y)ωk (y)(1 − η(ξk−1 (y); ξk−1 (x))) dy.
Pk ∩N (x)c
(2.3.9)
43
As mentioned above, we rely on the FFT-accelerator for computation of the third term
in (2.3.9). This, however, entails that we compute this integral as a convolution which,
in turn, requires that the integrand remains independent of localization cut-off function
η. This, of course, can easily be achieved by ensuring that the support of η(·, ξk−1 (x)) is
contained in N (x). The second term in this expression, which we refer to as local correction
to the accelerated numerics, is computationally small. We delay a more detailed discussion
on these aspects to Section (2.4). We detail the approximation of the singular integration
ZZ
sing
Kk [u](x) = Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)η(t; ξk−1 (x)) dt, (2.3.10)
[0,1]2 ∩ supp η
the first term on the right hand side of equation (2.3.9), next.
2.3.1.1 Singular integration over interior patches
In this section, we describe high-order quadrature rule for evaluation of Kksing [u](x) when
k ∈ MI (x). In this case, we define the localization function η(t; ξk−1 (x)) as follows:
|t − ξk−1 (x)|

−1
η(t; ξk (x)) = χ , 0, 1 , (2.3.11)
r
where r > 0 is appropriately chosen real number. We extend the domain of integration in
(2.3.10) to the disc of radius r around ξk−1 (x) = (tx1 , tx2 ) for each target point x ∈ Pk ∩ T.
Obviously, because of the fact that ϕk [u] vanishes to high order in all directions on the
boundary of the integration domain, this process does not compromise on the smoothness
of the integrand. To perform the integration, we then change to polar coordinates centered
at (tx1 , tx2 ):
t1 = tx1 + ρ cos θ
t2 = tx2 + ρ sin θ.
If we let
ϕ̃k [u](tx1 + ρ cos θ, tx2 + ρ sin θ) =
ϕk [u](tx1 + ρ cos θ, tx2 + ρ sin θ)ξk0 (tx1 + ρ cos θ, tx2 + ρ sin θ)η((tx1 + ρ cos θ, tx2 + ρ sin θ); (tx1 , tx2 )),
44
2.3.1.1. Singular integration over interior patches
then the integral (2.3.10) takes the form
Kksing [u](x)
Z2π rZ1/M
=M dθ ρ2M −1 Gκ (x, ξk (tx1 + ρM cos θ, tx2 + ρM sin θ))ϕ̃k [u](tx1 + ρM cos θ, tx2 + ρM sin θ) dρ
0 0
(2.3.12)
for M > 1. The appearance of additional factor ρ2M −1 (the Jacobian of the polar change of
variables) in the integrand of integral (2.3.12) cancels the kernel singularity as ρ2M −1 Gκ (x, ξk (tx1
+ρM cos θ, tx2 +ρM sin θ)), clearly, is a (2M −1)-times continuously differentiable function of
ρ and θ. Additionally, the cut-off function η vanishes to high order at the boundary of inte-
gration interval in ρ variable. Thus, use of Trapezoidal rule yields high-order convergence
for the ρ-integration. Moreover, the θ-integral can also be approximated to high order
by employing Trapezoidal rule quadrature as the corresponding integrand varies smoothly
and periodically with respect to θ.
While the change to polar coordinates provides a way to resolve the singularity analyt-
ically, the proposed application of Trapezoidal rule demands that we provide the density
values ϕ̃k [u] at points outside of the computational grid T. This necessitates employ-
ing an efficient and accurate interpolation strategy for evaluation of ϕ̃k [u] at these newly
transformed grid points. Toward this, we adapt the following Fourier refined polynomial
interpolation introduced in [23] to retain high order accuracy while maintaining computa-
tional efficiency:
1. Based on the density values ϕ̃k [u] on the original grid points, obtain the Fourier
coefficient for each one of the grid line {(t1 , t2 ) : t1 , t2 ∈ [0, 1]} by means of the fast
Fourier transforms(FFT).
2. Evaluate the Fourier series on much refined uniform grid. This evaluation can be
performed accurately and efficiently by means of FFTs.
3. Construct an interpolating polynomial of fixed degree, say, R on each interval of the

refined grid using the values obtained by step 2.
45
Error incurred by this interpolation procedure is much smaller compared to the other
source of error introduced in the overall algorithm, thus one can safely ignore the error
produced by this interpolation procedure. Note that, the actual choice of the refinement
factor, in second step, is always a trade-off between accuracy and computational time.
2.3.2 Singular Integration Over Boundary Patches

When target point x lies in one of the boundary integration patches, say Pk with k ∈
MB (x), we adopt a different approximation strategy to evaluate Kksing [u](x). The first
deviation comes in the form of the choice of the cut-off function η where a circular support
used for interior patches becomes unsuitable for use near the physical boundary of the
domain, that is, near t2 = 0. In this case, therefore, following the ideas introduced in [6],
we take the localization function to have a rectangular support around ξk−1 (x) = (tx1 , tx2 ),
given by
|t1 − tx1 | |t2 − tx2 |

η((t1 , t2 ); (tx1 , tx2 )) =χ , 0, 1 χ , 0, 1 . (2.3.13)
r r
We represent Kksing [u](x) as
Z1
Kksing [u](x) = Jk (t2 ; x)χ(|t2 − tx2 |/r, 0, 1)dt2 (2.3.14)
0
where
x +r
tZ
1
Jk (t2 ; x) = Gκ (x, ξk (t1 , t2 ))ϕk [u](t1 , t2 )ξk0 (t1 , t2 )χ(|t1 − tx1 |/r, 0, 1)dt1 . (2.3.15)
tx
1 −r
The evaluation of Jk (t2 ; x) poses difficulties owing to singularity present in the integrand
at t2 = tx2 and the near singularity in the vicinity of singular surface t2 = tx2 . Although, in
the near singular integrals integral kernel Gκ (x, y) is not singular and therefore integrand
is not singular. Nevertheless, when target point x approaches to singular surface t2 = tx2
integral kernel Gκ (x, y) becomes nearly singular and oscillatory. This singular nature of
the integrand can not be appropriately resolved by the polar change of variables (used
for the singularity resolution in the singular integral over interior patches) [2, page 64].
46
2.3.2. Singular Integration Over Boundary Patches
Thus, in order to circumvent these difficulties, we use different type of change of variable
t1 = tx1 + %(τ ), as one depicted in Figure (2.8), where the smooth invertible odd function
%(τ ) satisfies
dm %(τ )
= 0 for m = 0, ..., M, (2.3.16)
dτ m τ =0

(for example, %(τ ) = τ M +1 for even, non-negative integer M ). We now change the integra-
tion variable in (2.3.15) to obtain
−1 (r)
%Z
Jk (t2 ; x) = Gκ (x, ξk (tx1 +%(τ ), t2 ))ϕk [u](tx1 +%(τ ), t2 )ξk0 (tx1 +%(τ ), t2 )χ(%(τ )/r, 0, 1)%0 (τ ) dτ.
−%−1 (r)
(2.3.17)
This change of variable renders the integrand M times differentiable in τ that are also
uniformly bounded. In addition, in view of the factor χ(%(τ )/r, 0, 1), the integrand in
equation (2.3.17) vanishes on the boundary of the integration interval i.e. τ = ±%−1 (r),
together with all of its derivatives. Thus the integrand in integral (2.3.17) is a smooth
and periodic function which can be integrated to high-order by means of Trapezoidal
rule. Again, this change in a variable, produces a set of quadrature point in τ that do not
coincide with the computational grid on the integration patch Pk requiring an interpolation
strategy for which we again utilize the FFT refined polynomial interpolation, discussed in
subsection (2.3.1.1).
The final step in high-order approximation of Kksing [u](x) relates to the the computa-
tion of t2 -integral in (2.3.14). The main difficulty, here, is encountered in the form of a
jump discontinuity in the t2 -derivative of Jk (t2 ; x) at t2 = tx2 [6]. One can see this as a
consequence of jump relations in the single-layer potential (see Theorem 3.1 in [36] ). To
work around this, we split the integral in (2.3.14) as
x
Zt2
0
Z1
+ Jk (t2 ; x)χ(|t2 − tx2 |/r, 0, 1)dt2 ,
tx
2
47
Figure 2.8: Change of variables.
where both integrands, in principle, can be approximated to high-order by means of Q-

point Newton-Cotes quadrature. This, however, presents a practical difficulty in the form
of requiring at least Q-equidistant grid points in [0, tx2 ], and [tx2 , 1], that, of course, is not
available when tx2 is close to either 0 or 1. The implementation of a Q-point quadrature,
in such cases, requires values of Jk (t2 ; x) at points other than the original grid points.
The direct interpolation of Jk (t2 ; x), however, is neither accurate, (on account of non-
smoothness of Jk (t2 ; x) at t2 = tx2 ), nor efficient (because of its dependence on tx2 ). To
avoid this expensive computation, we further split (2.3.18) as follows:
L1 Z tx I1 Z tx
X 2 −(l1 −1)(Q−1)h2 X 2 −L1 (Q−1)h2 −(i1 −1)h2
Kksing [u](x) = Fk (t2 ; x) dt2 + Fk (t2 ; x) dt2
l1 =1 tx
2 −l1 (Q−1)h2 i1 =1 tx
2 −L1 (Q−1)h2 −i1 h2
L2 Z tx I2 Z tx
X 2 +l2 (Q−1)h2 X 2 +L2 (Q−1)h2 +i2 h2
+ Fk (t2 ; x) dt2 + Fk (t2 ; x) dt2
l2 =1 tx
2 +(l2 +1)(Q−1)h2 i2 =1 tx
2 +L2 (Q−1)h2 +(i2 −1)h2
48
2.3.2. Singular Integration Over Boundary Patches
Z tx
2 −L1 (Q−1)h2 −I1 h2
Z 1
+ Fk (t2 ; x) dt2 + Fk (t2 ; x) dt2 , (2.3.19)
0 tx
2 +L2 (Q−1)h2 +I2 h2
where Fk (t2 ; x) = Jk (t2 ; x)χ(|t2 − tx2 |/r, 0, 1), h2 = 1/N2 denotes the mesh size in t2
direction and numbers L1 , L2 , I1 , I2 are obtained as
tx2 1 − tx2

L1 = , L2 = ,
(Q − 1)h2 (Q − 1)h2
x
(1 − tx2 ) − L2 (Q − 1)h2

t2 − L1 (Q − 1)h2
I1 = , I2 = ,
h2 h2
with brc denoting the largest integer less than or equal to the real number r. Note that
when tx2 lies on one of the parallel grid lines, that is, tx2 = jh2 for some j, then last two
integrals in (2.3.19) vanish. Moreover, the computation of the second and the fourth set
of integrals therein, requires at the most 2 × (Q − 1) × (Q − 1) additional grid points, first
(Q − 1) × (Q − 1) in the vicinity of 0 and the remaining in the vicinity of 1. The values,
Jk (t2 ; x) at these additional grid points are obtained by, first interpolating the smooth
density ϕk [u] at these extra grid lines followed by the integration in (2.3.17).
It is clear that the strategy described above is valid when the target point x ∈ T happens
to lie on one of the parallel grid lines. Obviously, the overall computational scheme require
us to compute Kksing [u](x) for target point x that do not coincide with any of the grid
lines. In particular, this happens when the integration on a boundary patch corresponds
to a target point coming from an interior patch. To deal with such interactions between
boundary and interior patches, we exploit precomputed values of
X
KB [u](x) = Kk [u](x). (2.3.20)
k∈IB
at the Nyström nodes coming from the discretization of boundary patches, i.e., at x ∈
∪k∈IB Tk . Details of this strategy are described as follows:
Clearly, KB [u](x) accounts for the interaction of target point x with the set of boundary
patches P B . Notice that, Kk [u](x) can be viewed as volume potential defined over Ω whose
density ϕk [u] ∈ C m,α (Ω), by theorem (1.2.7) Kk [u] ∈ C m+2,α (Ω). Thus KB [u] ∈ C m+2,α (Ω).
Now, rewriting Equation (2.3.20) as
X
KB [u](x) = ω̃k (x)KB [u](x), (2.3.21)
k∈IB
49
where {ω̃k (x)|k ∈ IB } denotes partitions of unity subordinated to set of boundary patches
P B . Since, partition of unity functions ω̃k (x) vanishes for all k 6∈ MB (x), equation (2.3.21)
can be rewritten as
X
KB [u](x) = ω̃k (x)KB [u](x). (2.3.22)
k∈MB (x)
Here, ω̃k (x)KB [u](x) = 0 if x ∈ P B \Pk . Now, on each boundary patch we use precomputed
data {ω̃k (x)KB [u](x)|x ∈ Tk } to set up an FFT-refined polynomial interpolation schemes,
discussed in section (2.3.1.1), making possible accurate evaluations of ω̃k (x)KB [u](x) at
off-grid lines. Finally, we approximate integral over boundary patches, that is, KB [u](x),
efficiently and accurately by using formula (2.3.22). Due to efficiency of Fast Fourier
Transform, computational complexity of the whole procedure is bounded by O(NB log NB ),
where NB is the size of set {∪k∈IB Tk }.
Figure 2.9: An illustration of the locations xij,` of equivalent sources (points displayed in
blue color) on the parallel faces of cells Cij . The image also depicts, for a target point x,
the set N (x) that defines the adjacency.
50
2.4. Non-Singular Integration: Acceleration
2.4 Non-Singular Integration: Acceleration

In this section, we describe our acceleration strategy, for the evaluation of nonsingular
integrals
!Z
X X
reg
K [u](x) = + Gκ (x, y)m(y)u(y)ωk (y) dy, (2.4.1)
k∈IB k∈II Pk ∩N (x)c
that appear in (2.3.5). We rely on an equivalent source approximation strategy [23, 24] to
improve the overall computational efficiency. While the equivalent source methodology is
given for the three dimensional Helmholtz kernel in [23], we can readily adapt those ideas
to obtain a variant for the two dimensional case, as discussed below.
Indeed, we can accurately approximate Kreg [u](xq ) at each grid point xq ∈ T at an
O(N 2 ) cost by employing a high order quadrature, say, given by
X
Kareg [u](xq ) = w` Gκ (xq , y` )m(y` )u(y` )ωk (y` ). (2.4.2)
y` ∈T\N (xq )
Thus, for large size problem, owing to quadratic computational complexity, the appli-
cability of this algorithm becomes prohibitively expensive and it becomes hard to solve
with existing computational resources. Thus, for a large scale problem, looking for an
acceleration strategy is becomes imperative.
Indeed, the basic idea of our present approach is a modification of two of the most
advanced methods, namely, Adaptive Integral Method (AIM) introduced by Bleszynski
et. al. [14, 15], and precorrected-FFT methods [93], wherein a couple of new ideas are
introduced leading to significant memory savings with fast numerics which leads to super
algebraic convergence as the problem size increases. Noticing that (2.4.2) is a discrete
convolution whose contributing sources y` are located somewhat irregularly in Ω, prevented
the direct use of FFT to compute this convolution. Our acceleration strategy seeks to
replace sources y` by a certain set of “equivalent sources” placed on a Cartesian grid, that
produce an accurate approximation for the convolution, to facilitate the use of FFT for
computing the discrete convolution at an O(N log N ) cost.
51
To this end, we begin by noting that each one of the discretization points xq ∈ T, which
we refer as “true sources” in this discussion, lies in one of the cell Cij contained with in the
volume Ω. The key concept of the algorithm is to substitute the true sources contained
with in cell Cij by certain “equivalent sources” placed on the two opposite and parallel
faces of cell Cij such that the field produced by these equivalent sources approximate the
field produced by true sources within cell Cij to high order at all point in the space that
are nonadjacent to cell Cij . Since, these equivalent sources are placed on uniform Cartesian
grids, therefore its interactions can be computed efficiently on account of FFTs and then
easily transformed back into the scatterer Ω. For the sake of clarity we divide the complete
technique in three main steps, namely:
1. Two face equivalent source representation on Cartesian grids. For each cell
Cij , we seek to approximate the field produced by the scattering region contained
with in cell Cij , i.e., Z
Gκ (x, y)m(y)u(y)dy (2.4.3)
Ω∩cij
with the field produced by two independent sets of equivalent sources Fij , comprising
(m) (m)
of both acoustical monopoles σij,` Gκ (x, xij,` ) of intensity σij,` and acoustical dipoles
(d) ∂Gκ (x,xij,` ) (d)
σij,` ∂x`
of intensity σij,` , placed on an equipoised Cartesian grids xij,` , l =
1, · · · , N eq on the two parallel and opposite faces of cell cij . More precisely, for each
reg
cell Cij , if Ka,ij [u](x) denotes the quantity
reg
X
Ka,ij [u](x) = w` Gκ (x, y` )m(y` )u(y` )ωk (y` ),
y` ∈Cij
(m) (d)
then we seek constants σij,` and σij,` defining the approximating quantity
N eq
(d) ∂Gκ (x, xij,` )
reg,eq
X (m)
Ka,ij [u](x) = σij,` Gκ (x, xij,` ) + σij,` (2.4.4)
`=1
∂ν(xij,` )
such that the quantity

N coll
X reg reg,eq
2
Ka,ij [u](xq ) − Ka,ij [u](xq )
q=1
52
is minimized for a fixed N coll number of evaluation points xq on the boundary of the
set N (x) corresponding to the cell Cij . The locations xij,` of equivalent sources in
(2.4.4) are equidistant points placed on the two opposite and parallel sides of the cell
as shown in Figure 2.9 and ν(xij,` ) denotes the unit normal to adjacent set N (x) at
reg reg,eq
xij . Note that Ka,ij [u](x) as well as Ka,ij [u](x) which is a combination of discretize
single and double layer potentials, are a radiating solution of the Helmholtz equation
∆v + κv = 0 in R2 \ cij and closeness of approximation of former by the later follows
from well-posedness of the underlying exterior Dirichlet problem in N (x)c . It is
known that one can achieve this approximation up to the prescribed accuracy O(),
provided the number N eq is chosen as [6]

eq κA log()
N = max , √ .
L 2 log(3/ 2)
(m) (d)
The unknown constants σij,` and σij,` are obtained as solution to the overdetermined
linear system  
(m),ij
σ
A  = b, (2.4.5)
(d),ij
σ
eq reg,eq
where A is ncoll×N matrix corresponds to the evaluation of Ka,ij [u](x) and b is
reg
column vector of size ncoll × 1 corresponds to the evaluation of Ka,ij [u](x) at ncoll
points located at the boundary of set N (x). It is important to note that, as the
geometry is identical for each cell Cij , the QR factorization [59] of the above matrix
A need only be computed once and saved for repeated use. For numerical stability
of the least-squares solver, we choose N coll = 4N eq . A straightforward counting
shows that this process of equivalent source computation for all cells Cij requires
O(N 3/2 /L3 ) + +O(N 3/2 /L) operations in total, where O(N 3/2 /L3 ) corresponds for
QR factorization of the matrix A and O(N 3/2 /L) for evaluation of right hand side
vector b in (2.4.5) for all cells Cij .
2. Evaluation of non-adjacent field at Cartesian grids and FFTs. Once the in-
(m) (d)
tensities σij,` and σij,` have been computed on the Cartesian grids xij,` , ` = 1, · · · , N eq ,
53
for each cell Cij evaluation of the non-adjacent field values at the equivalent source
location as an approximation of true field Kareg [u](x) can be obtained as
L X
L j+1
i+1 X
X reg,eq
X reg,eq
Kareg,eq [u](x) = Ka,kl [u](x) − Ka,kl [u](x). (2.4.6)
k=1 l=1 k=i−1 l=j−1
Since, both summations in (2.4.6) are exact convolution on the Cartesian grids, they
can be approximated efficiently and accurately by means of FFTs with a compu-

tational cost of O LN 1/2 log(LN 1/2 ) . The values obtained in this manner pro-
vide accurate approximations for non-adjacent non-singular interactions at points
xij,` , ` = 1, · · · , N eq on the boundary of cells Cij .
3. Evaluation of non-adjacent field at true source points. Finally, to obtain

the approximation of nonadjacent nonsingular interactions (2.3.5) at a true source
location, say for an x ∈ Cij , we solve the free space Helmholtz equation within Cij ,
with Dirichlet boundary data coming from Kareq,eq [u](xij,` ).
To efficiently obtain solutions to these well-posed Dirichlet boundary value problems,

we utilize a discretized plane wave expansion of the form [3, 49]
Nw
X
Kareq,eq [u](x) ≈ γ` exp(iκd` .x), (2.4.7)
`=1
where the unit vectors d` sample the surface of unit disc with sufficient degree of
uniformity. Use of this approach is motivated by the spectral convergence of the
above expansion with respect to the number of unit vectors d` used and the fact that
the wave expansion coefficients γ` can be obtained as solutions to overdetermined
linear systems of the form Bγ = βij where the matrix B remains unchanged for
each cell, again, owing to the identical geometry of cells Cij s.
Computation of βij for all cells and QR factorization of matrix B requires O(N 3/2 /L)
and O(N 3/2 /L3 ) operations respectively. Thus, Kareg,eq [u](x) can be evaluated at all
true source locations x ∈ Cij at a computational cost of O(N 3/2 /L3 ) + O(N 3/2 /L)
and the overall computational cost for the evaluation of non-adjacent non-singular
54
interactions is given by
Λnon-adj = O LN 1/2 log(LN 1/2 ) + O(N 3/2 /L3 ) + O(N 3/2 /L).

(2.4.8)
55
Chapter
3
Computational Results In Two
Dimensions
In this chapter, we provide several numerical experiments to demonstrate the high-order

accuracy and capability of our two dimensional algorithm to solve the acoustic scattering
problem by penetrable inhomogeneous media. All the computations have been carried out
using in-house code devolved during the course of this Ph.D. in the object oriented frame
work of C++ programming language. modules have been developed for the visualization
of computational results. In all the cases ( evaluation of Fourier series for the construc-
tion of interpolation and evaluation of discrete convolution in acceleration strategy) FFT
computation has been carried out using fftw3 package. We have also used some relevant
subroutines of LAPACK package for the solution of linear least square system arising in
accelerated computation. In all the experiments, near fields are computed under the plane
wave incidence, ui (x) = eiκx.d , where d denotes the direction of propagation of incom-
ing plane wave. Solution of the dense linear system, arising from the discretization of
Lippmann-Schwinger equation (1.2.38) is obtained by means of the iterative solver GM-
RES [97]. We have used iterative solver GMRES in fully complex arithmetic, since, in
57
Chapter 3. Computational Results In Two Dimensions
comparison to double-dimension it requires a lower number of iterations for convergence

[65]. In all the experiment, we have computed relative max error ∞ and relative root mean
square error 2 at Nyström nodes T using following formulae:
kuexact − uapprox k∞
∞ =
kuexact k∞
N  21
P exact
 i=1 |u (xi ) − uapprox (xi )|2 
2 = 
 N
 ,

P
|uexact (xi )|2
i=1
where uapprox denotes the solution computed by our algorithm and uexact is exact solution.
The order of convergence of our algorithm which relates to the rate at which the error
in the computed solution decreases as the discretization scale decreases is given by the
following formula:

EN
Order = log2 , (3.0.1)
E2N
where EN denote the relative error corresponding to N number of discretization points. In
all the convergence tables, we write “Order” to denote the numerical order of convergence of
the approximation and “Iter” stands for number of GMRES iterations required to produce
the desired accuracy.
As we have explained in chapter 2, high-order convergence of the proposed algorithm
rely on the accuracy of the quadrature scheme used for the evaluation of integral opera-
tor (2.0.1), that is, accuracy in forward map computation. Therefore, to demonstrate the
high-order convergence of our algorithm we provide convergence table for solution along
with corresponding convergence table for forward map computation. Since, GMRES iter-
ation process has to be terminated after achieving desired error in residual we provide an
error tolerance to stop GMRES iteration process. Knowledge of accuracy in forward map
computation provide great help in this regard. In fact, in all of our scattering calculations,
presented in this chapter, we have chosen GMRES error tolerance, according to accuracy in
the corresponding forward map computation. The exact analytical solution for scattering
problems are seldom available or hard to obtain [19]. Therefore, in all the convergence
58
tables, computed solutions are compared either against the exact solution or approximate
solution which is obtained by application of our own numerical method with a very fine
discretization.
In all the convergence tables, presented in this chapter, the format of computational
grid specification p1 × n1 × n2 + q1 × m1 × m2 implies that a p1 number of boundary
patches, each with n1 × n2 discretization points, and a q2 number of interior patches with
n1 × n2 points are employed in the computation and they constitute the so called Nyström
grid T . In all the numerical experiments, it is assumed that the incident plane wave is
propagating along the positive x-axis, i.e., d = (1, 0). As the approximation scheme for the
integral operator is based on the use of spectrally accurate Trapezoidal rule and Q-point
Newton-Cotes quadrature, the expected convergence rates (computed by using formula
(3.0.1)) are essentially that of Newton-Cotes, that is, either Q or Q + 1 depending on
whether Q is even or odd, provided other parameters, such as the degree of the polynomial
in the interpolators, the number M for the change of variable %, etc. are chosen favorably.
We demonstrate the increase in the convergence rate with respect to increasing Q through
different sets of experiments, one using a 3-points while other using 5-points Newton-Cotes
quadrature. For the construction of interpolation we have used 16-fold grid refinement,
i.e., the length of the refined grid is 16 times smaller than that of original grid points.
For the pictorial visualization, we present a figures created by plotting the numerical
solution obtained by our algorithm. For the sake of visualization of fields on and off the
scatterer, we cover the inhomogeneity by square and then solution is computed in whole
of the square region by means of precomputed solution on the scatterer. Note that, in all
examples, refractive index n(x) is assumed to be one outside the inhomogeneity, that is,
material properties are allowed to be discontinuous across the surface of scatterer.
Example 3.0.1. (Convergence study for a simple scatterer)
As a first exercise, we compute the acoustic scattering by a penetrable disc for which
analytical solution can be computed, therefore, a comparison with exact solution is possible.
In order to perform simulation for this scatterer we decompose it into three overlapping
59
coordinate patches; two boundary and one interior. For this scattering configuration, we
present a convergence study, in Tables 3.1 to 3.6, corresponding to
√
1. κa = 4, n = 3
√
2. κa = 10, n = 2
√
3. κa = 20, n = 2,
where a denotes the diameter of the inhomogeneity of the scatterer which is taken to
be 2 in this example. The numerical solutions in this study are computed for several
discretization levels while employing a 3-point Newton-Cotes quadrature for t2 -integration
in boundary patches. In this case, we clearly see the methodology achieve the expected rate
of convergence. To show an enhanced order of convergence, we repeat the same experiment
but use a 5-point Newton-Cotes quadrature in the computation of the integral operator.
The corresponding results are presented in Tables 3.7 to 3.12, where, as expected, we see
an increased rate of convergence. In all the tables N stands for number of unknowns.
Grid Size N L2 L∞
ε2 Order ε∞ Order
2 × 3 × 9 + 1 × 9×9 135 1.38e-01 - 1.92e-01 -
2 × 5 × 17 + 1 × 17×17 459 6.37e-02 1.12e+00 6.21e-02 1.63e+00
2 × 9 × 33 + 1 × 33×33 1683 1.03e-02 2.62e+00 1.10e-02 2.50e+00
2 × 17 × 65 + 1 × 65×65 6435 6.07e-04 4.09e+00 8.93e-04 3.62e+00
2 × 33 × 129 + 1 × 129×129 25155 3.51e-05 4.11e+00 4.97e-05 4.17e+00
Table 3.1: Convergence study:

√ Forward map computation for a disc of acoustic size κa = 4
and refractive index n = 3 when 3-point Newton-Cotes quadrature employed for t2 -
integration over boundary patches.
For Pictorial visualization, we display, in Figures (3.1)-(3.7), the near fields obtained
by using our methodology corresponding to the following two experiments:
√
1. κa = 20, n = 2, d = (1, 0)
√
2. κa = 20, n = 2, d = (1, 0).
60
Grid Size N Iter L2 L∞

2 × 3 × 9 + 1 × 9×9 135 4 3.22e-01 - 3.28e-01 -
2 × 5 × 17 + 1 × 17×17 459 6 7.68e-02 2.07e+00 7.39e-02 2.15e+00
2 × 9 × 33 + 1 × 33×33 1683 7 1.32e-02 2.54e+00 1.38e-02 2.42e+00
2 × 17 × 65 + 1 × 65×65 6435 10 1.19e-03 3.47e+00 1.38e-03 3.32e+00
2 × 33 × 129 + 1 × 129×129 25155 12 4.41e-05 4.76e+00 5.90e-05 4.55e+00
Table 3.2: Convergence

√ study: plane wave scattering by a disc of acoustic size κa = 4 and
refractive index n = 3 when 3-point Newton-Cotes quadrature employed for t2 -integration
over boundary patches.
Grid Size N L2 L∞
2×3×9+ 1×9×9 135 3.14e-01 - 2.28e-01 -
2×5×17+ 1×17×17 459 8.00e-02 1.97e+00 7.35e-02 1.63e+00
2×9×33+ 1×33×33 1683 1.52e-02 2.40e+00 1.37e-02 2.43e+00
2×17×65+ 1×65×65 6435 1.16e-03 3.71e+00 1.10e-03 3.63e+00
2×33×129+ 1×129×129 25155 5.12e-05 4.50e+00 6.71e-05 4.04e+00


2 × 3 × 9 + 1 × 9×9 135 10 7.04e-01 - 6.21e-01 -
2 × 5 × 17 + 1 × 17×17 459 12 2.68e-01 1.39e+00 2.73e-01 1.18e+00
2 × 9 × 33 + 1 × 33×33 1683 16 3.92e-02 2.78e+00 4.77e-02 2.52e+00
2 × 17 × 65 + 1 × 65×65 6435 22 3.38e-03 3.54e+00 4.14e-03 3.52e+00
2 × 33 × 129 + 1 × 129×129 25155 26 1.47e-04 4.52e+00 2.37e-04 4.13e+00

√ plane wave scattering by a disc of acoustic size κa = 10
Numerical results, corresponding to first experiment, are reported in the last row of Table
(3.12). For the second experiment where the incidence wave impinges on the obstacle at
45 degree angle with positive x-axis, require 71 GMRES iterations to produce a relative
61
Grid Size N L2 L∞
2 × 3 × 9 + 1 × 9×9 135 9.63e-01 - 1.16e+00 -
2 × 5 × 17 + 1 × 17×17 459 2.30e-01 2.07e+00 3.35e-01 1.79e+00
2 × 9 × 33 + 1 × 33×33 1683 2.42e-02 3.25e+00 2.98e-02 3.49e+00
2 × 17 × 65 + 1 × 65×65 6435 2.23e-03 3.44e+00 3.43e-03 3.12e+00
2 × 33 × 129 + 1 × 129×129 25155 1.57e-04 3.83e+00 2.82e-04 3.61e+00


2 × 3 × 9 + 1 × 9×9 135 4 1.17e+00 - 2.16e+00 -
2 × 5 × 17 + 1 × 17×17 459 14 8.41e-01 4.78e-01 1.01e+00 1.12e+00
2 × 9 × 33 + 1 × 33×33 1683 33 2.63e-01 1.68e+00 4.33e-01 1.22e+00
2 × 17 × 65 + 1 × 65×65 6435 48 2.47e-02 3.41e+00 3.10e-02 3.81e+00
2 × 33 × 129 + 1 × 129×129 25155 61 1.76e-03 3.81e+00 1.76e-03 4.14e+00

Grid Size N L2 L∞
2×5×9+1×9×9 171 6.17e-02 1.39e-01 -
2 × 9 × 17 + 1 × 17 × 17 595 1.34e-02 2.20e+00 2.70e-02 2.36e+00
2 × 17 × 33 + 1 × 33 × 33 2211 1.23e-03 3.45e+00 3.12e-03 3.12e+00
2 × 33 × 65 + 1 × 65 × 65 8515 4.63e-05 4.73e+00 6.83e-05 5.51e+00
2 × 65 × 129 + 1 × 129 × 129 33411 1.56e-06 4.89e+00 1.81e-06 5.24e+00

error of 0.03% in the max norm ∞ .
Example 3.0.2. (Computational Efficiency)
62
GridSize N Iter L2 L∞
2×5×9+1×9×9 171 4 3.04e-01 3.09e-01 -
2 × 9 × 17 + 1 × 17 × 17 595 6 6.54e-02 2.22e+00 7.50e-02 2.04e+00
2 × 17 × 33 + 1 × 33 × 33 2211 7 4.53e-03 3.85e+00 6.48e-03 3.53e+00
2 × 33 × 65 + 1 × 65 × 65 8515 9 9.78e-05 5.53e+00 1.26e-04 5.68e+00
2 × 65 × 129 + 1 × 128 × 129 33411 12 1.75e-06 5.80e+00 2.29e-06 5.78e+00
Table 3.8: Convergence

√ study: plane wave scattering by a disc of acoustic size κa = 4 and
refractive index n = 3 when 5-point Newton-Cotes quadrature employed for t2 -integration
over boundary patches.
Grid Size N L2 L∞
2×5×9+1× 9×9 171 2.24e-01 - 2.25e-01 -
2×9×17+1×17 ×17 595 3.31e-02 2.76e+00 4.32e-02 2.38e+00
2×17×33+1×33×33 2211 3.03e-03 3.45e+00 4.41e-03 3.29e+00
2×33×65+1×65×65 8515 1.23 e-04 4.62e+00 1.26 e-04 5.12e+00
2×65×129+1×129×129 33411 2.77e-06 5.47e+00 2.03e-06 5.96e+00


2×5×9+1×9×9 171 10 7.09e-01 - 5.96e-01
2 × 9 × 17 + 1 × 17 × 17 595 11 1.22e-01 2.54e+00 1.27e-01 2.24e+00
2 × 17 × 33 + 1 × 33 × 33 2211 15 8.78e-03 3.80e+00 1.03e-02 3.62e+00
2 × 33 × 65 + 1 × 65 × 65 8515 19 2.84e-04 4.95e+00 3.35e-04 4.94e+00
2 × 65 × 129 + 1 × 129 × 129 33411 23 7.76e-06 5.19e+00 8.53e-06 5.30e+00

We next demonstrate a tempered growth in computational cost of the proposed ac-

celerated algorithm while maintaining a fixed computational error by comparing our ap-
proximate integral operator against the continuous operator (2.0.1). Toward this, with
63
Grid Size N L2 L∞
2 × 5 × 9 + 1 × 9×9 171 6.32e-01 - 8.28e-01 -
2 × 9 × 17 + 1 × 17×17 595 9.73e-02 2.70e+00 1.48e-01 2.48e+00
2 × 17 × 33 + 1 × 33×33 2211 1.42e-02 2.77e+00 1.80e-02 3.04e+00
2 × 33 × 65 + 1 × 65×65 8515 8.09e-04 4.14e+00 8.33e-04 4.43e+00
2 × 65 × 129 + 1 × 129×129 33411 1.33e-05 5.92e+00 1.74e-05 5.58e+00
Table 3.11: Convergence study: Forward

√ map computation for a disc of acoustic size
κa = 20 and refractive index n = 2 when 5-point Newton-Cotes quadrature employed
for t2 -integration over boundary patches.

2 × 5 × 9 + 1 × 9×9 171 29 1.43e+00 - 1.37e+00 -
2 × 9 × 17 + 1 × 17×17 595 47 9.07e-01 6.59e-01 1.13e+00 2.86e-01
2 × 17 × 33 + 1 × 33×33 2211 53 3.76e-01 1.27e+00 4.51e-01 1.32e+00
2 × 33 × 65 + 1 × 65×65 8515 53 1.37e-02 4.78e+00 1.88e-02 4.59e+00
2 × 65 × 129 + 1 × 129×129 33411 65 2.72e-04 5.66e+00 3.21e-04 5.87e+00

every doubling of the computational grid size, we double the wave number κ to keep the
number of points per wavelength unchanged, thus, fixing the accuracy level. For this set
of experiments, we again use a disc scatterer with a constant refractive index n = 2 so
that the solution of Lippmann-Schwinger equation (1.2.38) is infinitely smooth within the
scatterer Ω. The results in Table 3.13 show that, while for small values of N , the cost
of non-accelerated algorithm is comparable to that of the accelerated version, as problem
size increase, there is a substantial gain in terms of computational cost. For instance, for
the scatterer of size κa = 32, accelerated computations are 88 times faster than its non-
accelerated counterpart. Since, solution of Equation (1.2.38) is infinitely smooth within the
scatterer Ω, therefore, exponent δ in the Equation (2.3.8) can be chosen very close to zero
without compromising in the accuracy of the approximation. In particular, for sufficiently
64
small value of δ the time for accelerated computation in Table 3.13, exhibits the growth
according to the computational complexity of O(N log N ).
κa Grid Size ε2 ε∞ Time(s)/Iteration

accelerated Non-Accelerated
4 2 × 16 × 32 + 1 × 32 × 32 1.18e-03 3.15e-03 2.57 3.93
8 2 × 32 × 64 + 1 × 64 × 64 3.90e-04 5.67e-04 8.53 39.56
16 2 × 64 × 128 + 1 × 128 × 128 7.90e-04 7.85e-04 32.8 626.47
32 2 × 128 × 256 + 1 × 256 × 256 8.64e-05 9.05e-05 122.31 10785.9
Table 3.13: Computational cost for accelerated and non-accelerated algorithm.
Example 3.0.3. (Accuracy of non-singular non-adjacent interactions)
The previous examples demonstrated the high-order convergence and efficient computa-
tional time for the proposed methodology. We, in particular, noted significant gains for the
accelerated scheme in terms of computational cost. On the other hand, though Table 3.13
show that the accuracy levels of the numerical method are maintained even when frequency
of wave oscillations increase, provided we ensure a fixed number of discretization points per
wavelength, an interesting picture emerges when we look approximations coming from the
reg,eq reg
accelerator alone. To see this, we compare the values of Ka,ij [u] against that of Ka,ij [u] at
increasing levels of frequency while the number of equivalent sources remained unchanged.
As seen in Table 3.15, approximations converge rapidly with increasing wavenumber. This
observation, in turn, is utilized in a more effective load balancing between adjacent and
non-adjacent calculations to achieve favorable computational cost.
We also include, in Table 3.14, a similar study, this time keeping the wavenumber
constant (κ = 8π) while increasing the number of equivalent sources, N eq , where we again
see rapid improvement in accuracy levels.
Example 3.0.4. (Convergence study: complex scattering media)
65
Equivalent Source/cell ε2 ε∞
4×4 1.60e-02 4.66e-02
6×4 4.14e-05 1.68e-04
8×4 3.53e-08 1.85e-07
10 × 4 1.60e-10 1.48e-09
12 × 4 2.86e-12 2.82e-11
Table 3.14: Accuracy of acceleration for fixed wavenumber κ = 8π
κH Equivalent Source/cell ε2 ε∞
2 4×4 5.10e-05 1.30e-04
4 8×4 3.05e-09 1.94e-08
8 13 × 4 3.00e-12 6.70e-12
16 30× 4 1.74e-12 1.14e-12
Table 3.15: Accuracy of acceleration when wavenumber κ increases but number of point
per wavelength is fixed.
The methodology presented in this thesis can, of course, be applied to acoustic scat-
tering calculations from penetrable media with complicated geometries as well as variable
material properties.
We begin with an example that demonstrates adaptability and applicability of our
algorithm in dealing with scatterers that have relatively complex geometrical description.
Toward this, we consider scattering by a penetrable bean shaped scatterer whose boundary
is described by following parametric equation
r(t) = cos t + 0.65 cos 2t − 0.65, 0 ≤ t ≤ 2π. (3.0.2)
In order to perform simulation for this scatterer we decompose it into three overlapping
coordinate patches; two boundary and one interior ( see Figure (3.8)). Since for this scat-
tering configuration analytical solution is not available, therefore, we use numerical solution
obtained by our algorithm at a finer discretization for convergence study comparisons. All
the computation performed under plane wave incidence.
In our fist example for this scattering configuration, acoustical size of the scatterer is
√
taken to be κa = 10 and refractive index of the medium is n = 2. The first sets of
66
computations, for which the results are presented in Tables 3.16 and 3.17, employ 3-point
Newton-Cotes quadrature for t2 -integration over boundary patches. As expected, we see
a convergence rate of order 4 in the solution. We repeat the experiment with 5-points
Newton-Cotes quadrature and report the results in Tables Tables 3.18 and 3.19, where we
again see an enhanced order of convergence.
Grid Size N L2 L∞
2 × 3 × 9 + 1 × 9×9 135 4.46e-01 - 3.38e-01 -
2 × 5 × 17 + 1 × 17×17 459 1.27e-01 1.82e+00 9.87e-02 1.77e+00
2 × 9 × 33 + 1 × 33×33 1683 1.79e-02 2.82e+00 1.61e-02 2.61e+00
2 × 17 × 65 + 1 × 65×65 6435 1.75e-03 3.36e+00 1.71e-03 3.24e+00
2 × 33 × 129 + 1 × 129×129 25155 7.25e-05 4.59e+00 1.45e-04 3.56e+00

√ map computation for a bean of acoustic size

ε2 Order ε∞
Order
2 × 3 × 9 + 1 × 9×9 135 4 1.95e+00 - 1.82e+00
-
2 × 5 × 17 + 1 × 17×17 459 11 3.89e-01 2.33e+00 3.84e-01
2.25e+00
2 × 9 × 33 + 1 × 33×33 1683 15 3.88e-02 3.32e+00 4.07e-02
3.24e+00
2 × 17 × 65 + 1 × 65×65 6435 20 4.10e-03 3.24e+00 4.83e-03
3.08e+00
2 × 33 × 129 + 1 × 129×129 25155 25 2.28e-04 4.17e+00 4.55e-04
3.41e+00
√
Table 3.17: Convergence for the bean shape scatterer with κa = 10 and n = 2, when
3-point Newton-Cotes quadrature employed for boundary patch integration in transverse
integration.
Of course, our algorithm is not restricted to consideration of constant or piecewise

constant material properties. In order to demonstrate this fact, in our final experiment, we
carried out scattering computations for a scatterer of acoustical size κa = 4 with variable
refractive index n, given by
n(x) = sin(πx1 ) cos(πx2 ) for x = (x1 , x2 ) ∈ Ω.
67
Grid Size N L2 L∞
2×5×9+1×9×9 171 4.12e-01 4.21e-01 -
2 × 9 × 17 + 1 × 17 × 17 595 4.05e-02 3.35e+00 5.79e-02 2.86e+00
2 × 17 × 33 + 1 × 33 × 33 2211 3.60e-03 3.49e+00 5.51e-03 3.39e+00
2 × 33 × 65 + 1 × 65 × 65 8515 1.47e-04 4.61e+00 1.31e-04 5.39e+00
2 × 65 × 129 + 1 × 129 × 129 33411 3.40e-06 5.43e+00 3.09e-06 5.41e+00

√ map computation for a bean of acoustic size

ε2 Order ε∞
Order
2×5×9+1×9×9 171 12 1.85e+00 1.31e+00
-
2 × 9 × 17 + 1 × 17 × 17 595 16 1.56e-01
3.57e+00 3.35e-01
1.96e+00
2 × 17 × 33 + 1 × 33 × 33 2211 18 1.90e-02
3.04e+00 2.94e-02
3.51e+00
2 × 33 × 65 + 1 × 65 × 65 8515 24 6.13e-04
4.95e+00 6.36e-04
5.53e+00
2 × 65 × 129 + 1 × 129 × 129 33411 30 1.17e-05
5.71e+00 1.30e-05
5.61e+00
√
Table 3.19: Convergence for the bean shape scatterer with κ = 10 and n = 2, when
5-point Newton-Cotes quadrature employed for boundary patch integration in transverse
integration.
Numerical results corresponding to three and five point Newton-Cotes quadrature are pre-
sented in Tables 3.20 to 3.23, respectively. Convergence study of these tables clearly
support that the high-order nature of our algorithm, as expected, continues to hold for
scattering media with variable material properties and this is one of the most powerful
features of this method.
For the sake of pictorial visualization, results of scattering computation under plane
wave incidence corresponding to
√
1. κa = 20, n = 2, d = (1, 0)
√ √ √
2. κa = 20, n = 2, d = (1/ 2, 1/ 2)
are shown in Figures figs. 3.9 to 3.15. The plots depicts the solution obtained when residual
68
Grid Size N L2 L∞
2 × 3 × 9 + 1 × 9×9 135 1.79e-01 - 1.75e-01 -
2 × 5 × 17 + 1 × 17×17 459 7.24e-02 1.31e+00 7.00e-02 1.33e+00
2 × 9 × 33 + 1 × 33×33 1683 7.41e-03 3.29e+00 7.02e-03 3.32e+00
2 × 17 × 65 + 1 × 65×65 6435 6.98e-04 3.41e+00 7.23e-04 3.28e+00
2 × 33 × 129 + 1 × 129×129 25155 1.75e-05 5.32e+00 5.40e-05 3.74e+00
Table 3.20: Convergence study: Forward map computation for a bean of acoustic size
κa = 4 and refractive index n = sin(πx) cos(πy) for x = (x, y) ∈ Ω, when 3-point
Newton-Cotes quadrature employed for t2 -integration over boundary patches.

2 × 3 × 9 + 1 × 9×9 135 3 1.01e+00 - 1.25e+00 -
2 × 5 × 17 + 1 × 17×17 459 4 7.60e-02 3.74e+00 9.11e-02 3.77e+00
2 × 9 × 33 + 1 × 33×33 1683 5 7.71e-03 3.30e+00 8.36e-03 3.45e+00
2 × 17 × 65 + 1 × 65×65 6435 6 6.56e-04 3.55e+00 6.36e-04 3.72e+00
2 × 33 × 129 + 1 × 129×129 25155 7 2.02e-05 5.02e+00 5.79e-05 3.46e+00
Table 3.21: Convergence for the bean shape scatterer with κa = 4 and n = sin(πx) cos(πy)
for x = (x, y) ∈ Ω, when 3-point Newton-Cotes quadrature employed for boundary patch
integration in transverse integration.
Grid Size N L2 L∞
2×5×9+1×9×9 171 8.78e-02 3.12e-01 -
2 × 9 × 17 + 1 × 17 × 17 595 1.65e-02 2.41e+00 6.22e-02 2.32e+00
2 × 17 × 33 + 1 × 33 × 33 2211 1.30e-03 3.67e+00 5.64e-03 3.46e+00
2 × 33 × 65 + 1 × 65 × 65 8515 7.54e-05 4.11e+00 1.40e-04 5.33e+00
2 × 65 × 129 + 1 × 129 × 129 33411 1.12e-06 6.07e+00 2.08e-06 6.07e+00
Table 3.22: Convergence study: Forward map computation for a bean of acoustic size
κa = 4 and refractive index n = sin(πx) cos(πy) for x = (x, y) ∈ Ω, when 5-point Newton-
Cotes quadrature employed for t2 -integration over boundary patches.
in GMRES iteration process had reached 10−4 .
69

2×5×9+1×9×9 171 3 5.25e-02 1.69e-01 -
2 × 9 × 17 + 1 × 17 × 17 595 4 9.12e-03 2.53e+00 3.34e-02 2.34e+00
2 × 17 × 33 + 1 × 33 × 33 2211 5 7.99e-04 3.51e+00 3.14e-03 3.41e+00
2 × 33 × 65 + 1 × 65 × 65 8515 7 4.51e-05 4.15e+00 8.14e-05 5.27e+00
2 × 65 × 129 + 1 × 129 × 129 33411 8 6.14e-07 6.20e+00 1.23e-06 6.04e+00
Table 3.23: Convergence for the bean shape scatterer with κa = 4 and n = sin(πx) cos(πy)
for x = (x, y) ∈ Ω, when 5-point Newton-Cotes quadrature employed for boundary patch
integration in transverse integration.
70
(a) Disc Shape Scatterer
(b) Plane Wave Incidence, Re(ui )
Figure 3.1: Plane wave exp(iκd · x) incidence along the incidence direction d = (1, 0) with
wave number κ = 20 on a penetrable unit disc. Fields computed in whole square region
displayed in blue color.
71
(a) Real part of the scattered field, Re(us )
(b) Real part of the total field, Re(u)
Figure 3.2: Visualization of real part of the fields for scattering of a plane wave exp(iκd · x)
with κ = 20, d = (1, 0) by a penetrable unit disc. A computational grid of size 2 × 65 ×
129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to obtain an error of
0.032% measured in max norm.
72
(a) Imaginary part of the scattered field, Im(us )
(b) Imaginary part of the total field, Im(u)
Figure 3.3: Visualization of imaginary part of the fields for scattering of a plane wave
exp(iκd · x) with κ = 20, d = (1, 0) by a penetrable unit disc. A computational grid of
size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to obtain
an error of 0.032% measured in max norm.
73
(a) Absolute part of the scattered field, |us |
(b) Absolute part of the total field, |u|
Figure 3.4: visualization of absolute value of the fields for scattering of a plane wave
exp(iκd · x) with κ = 20, d = (1, 0) by a penetrable unit disc. A computational grid of
size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to obtain
an error of 0.032% measured in max norm.
74
(a) Real part of the scattered field,Re(us )
(b) Real part of the total field,Re(u)
Figure 3.5: Visualization √ part of the fields for scattering of a plane wave exp(iκd · x)
√ of real
with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A computational grid of size
2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to obtain an
error of 0.06% measured in max norm.
75
(a) Imaginary part of the scattered field,Im(us )
(b) Imaginary part of the total field,Im(u)
Figure 3.6: Visualization of imaginary

√ part
√ of the fields for scattering of a plane wave
exp(iκd · x) with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A computational
grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to
obtain an error of 0.06% measured in max norm.
76
(a) Absolute value of the scattered field,|us |
(b) Absolute value of the total field,|u|
Figure 3.7: Visualization of absolute

√ value
√ of the fields for scattering of a plane wave
exp(iκd · x) with κ = 10, d = (1/ 2, 1/ 2) by a penetrable unit disc. A computational
grid of size 2 × 65 × 129 + 1 × 129 × 129 and 5-point Newton-Cotes quadrature is used to
obtain an error of 0.06% measured in max norm.
77
Figure 3.8: Three-patch covering of the Bean shape scatterer depicted in Figure 3.9(a).
First and last figure corresponds to boundary patch while the middle figure represent
voluminous interior patch.
78
(a) Bean shape scatterer
(b) Plane wave incidence
Figure 3.9: Plane wave incidence exp(iκd · r) with κ = 10, d = (1, 0) on a penetrable bean
shape scatterer.
79
Figure 3.10: Visualization of real part of fields for scattering by penetrable bean shape
scatterer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 × 129 × 129, Newton-
Cotes five point quadrature is used for transverse integration over boundary patches.
80
Figure 3.11: Visualization of imaginary part of fields for scattering by penetrable bean
shape scatterer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 × 129 ×
129, Newton-Cotes five point quadrature is used for transverse integration over boundary
patches.
81
(a) Absolute part of the scattered field, |us |
(b) Absolute part of the total field, |u|
Figure 3.12: Visualization of absolute value of the fields for scattering by penetrable bean
shape scatterer, At frequency κ = 10, d = (1, 0), Grid Size 2 × 65 × 129 + 1 × 129 ×
patches.
82
Figure 3.13: Visualization of real part of√fields√for scattering by penetrable bean shape
scatterer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2 × 65 × 129 + 1 × 129 ×
patches.
83
Figure 3.14: Visualization of imaginary part√of fields

√ for scattering by penetrable bean
shape scatterer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2 × 65 × 129 + 1 × 129 ×
patches.
84
(a) Absolute value of the scattered field, |us |
(b) Absolute value of the total field, |u|
Figure 3.15: Visualization of absolute value of√the fields

√ for scattering by penetrable bean
shape scatterer, At frequency κ = 10, d = (1/ 2, 1/ 2), Grid Size 2 × 65 × 129 + 1 × 129 ×
patches.
85
Chapter
4
An efficient high-order integral
equation solver in three dimensions
An efficient high-order Nyström scheme, presented in Chapter (3), in the context of solu-
tion of Lippmann-Schwinger integral equation (1.2.38) in two dimensions, can be readily
generalized to three dimensions. As in the two dimensional approach, the goal is to design
an integration scheme that evaluate the volume integral operator
Z
K[u](x) = Gκ (x, y)m(y)u(y)dy (4.0.1)
Ω
efficiently in three dimension without compromising in accuracy. Then resulting linear

system is solved by means of GMRES iterative linear solver and total field u is obtained.
Notice that, most of the issues that arise in the evaluation of integral (4.0.1) are largely
analogous to those presented in two dimensional counterpart. Nevertheless, due to in-
creased dimensionality and substantially different behavior of the corresponding kernel
singularity, new challenges appear in treating fully three-dimensional structures. More
precisely, in contrast with the two-dimensional case, integral kernel (free space Green’s
exp(iκ|x−y|) 1
function) Gκ (x, y) = 4π|x−y|
in three dimensions have a singularity of order r
rather
87
Chapter 4. An efficient high-order integral equation solver in three dimensions
than logarithmic singularity, when target point x is either close to source point y or is equal
1
to y. In addition, singularity of the type r
in the three dimensions can not be resolved
appropriately by the techniques used in two dimensional setting such as polar change of
variable. Therefore, it must be treated adequately. In the following sections we see that
these issues can be successfully resolved by using appropriate change of variables.
In this chapter, we present a generalization of the integration scheme presented in

Chapter 2 for two dimensions. Indeed, like two dimensional counterpart, our three dimen-
sional integration scheme is also based on combination of two main basic elements: (1)
high-order evaluation of singular integrals by means of analytical resolution of kernel sin-
gularity together with partitions of unity, and (2) evaluation of non-singular non-adjacent
interactions by means of 3-D Fast Fourier Transform.
This chapter is organized as follows: in Section (4.1), we present main algorithmic

components of our numerical scheme in three dimensions, where, like in two dimensional
case, we decompose volume integral operator over domain Ω into local integrals over the
overlapping coordinate patches.We will devote section (4.2) for the detailed description of
our approximation strategy for the high-order evaluation of singular integrals. Finally, in
section (4.3), a detailed account of acceleration strategy for efficient and accurate evaluation
of non-singular non-adjacent interactions is presented.
4.1 Principal Components of The Method In Three

Dimensions
As mentioned above, we base our numerical strategy on solving the linear system, arising
out of discretization of the Lippmann-Schwinger integral equation which is extension of
ideas presented in Chapter 2. We focus our presentation on the computational technique for
accurate and efficient evaluation of the volume integral in Equation (4.0.1). We represent
the scatterer Ω ⊂ R3 as a collection P = {Pk }K
k=1 of K overlapping coordinate patches
where the k-th patch is homeomorphic to an open set Hk ⊂ [0, 1]3 via a smooth invertible
88
4.1. Principal Components of The Method In Three Dimensions
parametrization
ξk = ξk (t1 , t2 , t3 ), for t = (t1 , t2 , t3 ) ∈ [0, 1]3 , k = 1, · · · , K.
As an example, decomposition of spherical shape scatterer into three overlapping coordi-

nate patches is shown in figure (4.1). These overlapping patches, in conjunction with a par-
titions of unity (POU) subordinated to covering P, that is, functions ωk (x) : k = 1, . . . , K,
satisfying
K
X
ωk (x) = 1 for all x ∈ Ω,
k=1
∞
where for each k, ωk ∈ C (Ω) with its support contained in Pk , reduce the evaluation of
(4.0.1) to computation of integrals over these K patches. Such partitions of unity can be
easily constructed by adopting the procedure discussed in Chapter (2). Note that, for all
those patches Pk whose closure does not intersect with the boundary of Ω, that is, ∂Ω, the
corresponding ωk vanishes to high order along with all of its derivatives on the boundary
of the patch. On remaining patches, however, where one of the edges coincide with the
boundary of Ω, ωk clearly does not vanish and, in fact, attains the value 1 (see figures (4.2)
and (4.3) for examples).
For the clarity of presentation, and to distinguish between patches based on this cri-
terion, we follow the convention set in the two dimensional case and partition the set of
patches P into two sets, namely, set of boundary patches
n o
P B = Pk ∈ P k ∈ IB
and set of interior patches

n o
I
P = Pk ∈ P k ∈ II ,

where
IB = {k|Pk ∩ ∂Ω 6= Ø}
and
II = {k|Pk ∩ ∂Ω = Ø}
89
(a) Spherical Shape Scatterer (b) Boundary Patch
(c) Interior Patch (d) Boundary Patch
Figure 4.1: Three patch covering of full spherical shape geometry and Nyström discretiza-
tion grids.
90
define an index set for boundary and interior patches respectively. Note that these two set
of patches have empty intersection and its union cover entire scattering region Ω. Now,
similar to the two dimensional case, the integral (4.0.1) can be rewritten as a sum of
integrals over boundary and interior patches
XZ XZ
K[u](x) = Gκ (x, y)m(y)u(y)ωk (y)dy + Gκ (x, y)m(y)u(y)ωk (y)dy.
k∈IB Pk k∈II Pk
(4.1.1)
In view of this, we introduce a set of quadrature points on the k-th patch, say Tk , by taking
the image of the uniform Cartesian grid {(i1 /N1 , i2 /N2 , i3 /N3 ) | 0 ≤ i` ≤ N` , ` = 1, 2, 3} in
[0, 1]3 under the smooth invertible map ξk . Further, the union of these grid points, that is,
K
[
T= Tk , (4.1.2)
k=1
where
Tk = {ξk (i1 /N1 , i2 /N2 , i3 /N3 ) | 0 ≤ i` ≤ N` , ` = 1, 2, 3} (4.1.3)
define the complete set of Nyström discretization points where we seek to approximate the
solution of Lippmann-Schwinger integral equation (1.2.38) in the three dimensions.
Since the proposed algorithm is based on the Nyström methodology, therefore, the goal
is to find the values of total field u at each of the Nyström grid points x ∈ T. Indeed, the
performance of our iterative solver hinges on our ability to approximate the integrals in
(4.1.1) accurately, in a computationally efficient manner at each iteration.
Remark 4.1.1. 1. In the following sections, we often refer to x in equation (4.1.1) as

a target point whereas points y therein have sometimes been called source points.
2. While discussing the approximation on boundary patches, we identify the 3rd coordi-
nate variable in [0, 1]3 with the transverse parameter. We, therefore, use the notation
t = (t1 , t2 , t3 ) to represent a point in the corresponding parameter space and assume
that the boundary of scatterer Ω coincides with t3 = 0.
91
Figure 4.2: Partition of unity on boundary patch of spherical shape scatterer.
92
Figure 4.3: Partition of unity on interior patch of spherical shape scatterer.
93
As in two dimension, the difficulty in high-order approximation of integrals in (4.1.1) is

significantly more, when the target point x lies in the integration patch Pk compared to the
case when it does not. Indeed, when x ∈ Pk , owing to the singularity of the kernel Gκ (x, y)
and its subsequent derivative at y = x, the integrand is unbounded within the integration
domain and direct application of a standard quadratures yield inaccurate approximations.
Thus, specialized quadrature rules must be developed and used to deal with such singular
integrals. The case when x 6∈ Pk , in contrast, does not present this challenge because all
the factors in the integrand of integrals in (4.1.1) are smooth. To effectively present these
two contrasting scenarios, similar to the two dimensional case, we refine the index sets IB
and II by introducing target point dependent index sets
MB (x) = {k ∈ IB | x ∈ Pk }
and
MI (x) = {k ∈ II | x ∈ Pk }
to rewrite (4.1.1) as
 
X X Z
K[u](x) =  +  Gκ (x, y)m(y)u(y)ωk (y) dy
k∈MB (x) k6∈MB (x) Pk
 
X X Z
+ +  Gκ (x, y)m(y)u(y)ωk (y) dy. (4.1.4)
k∈MI (x) k6∈MI (x) Pk
Clearly, each of the above integrals in the parametric coordinates can be made to read as
Z
Gκ (x, y)m(y)u(y)ωk (y) dy
Pk
ZZZ
= Gκ (x, ξk (t1 , t2 , t3 ))ϕk [u](t1 , t2 , t3 )ξk0 (t1 , t2 .t3 ) dt1 dt2 dt3 , (4.1.5)
[0,1]3
where
ϕk [u](t1 , t2 , t3 ) = m(ξk (t1 , t2 , t3 ))u(ξk (t1 , t2 , t3 ))ωk (ξk (t1 , t2 , t3 )),
and ξk0 denotes the Jacobian of the transformation ξk .
94
To this end, we begin with the case, when target point x does not lie in the integration
patch Pk , in other words, when k 6∈ MB (x) and k 6∈ MI (x) in (4.1.4), then source
point y is sufficiently away from the target point x and hence integral kernel Gκ (x, y)
remains non-singular throughout the region of integration. Though, adopting a single high-
order approximation scheme for both scenarios is possible, we actually utilize two different
quadratures to take advantage of a more favorable behavior of the integrands in the later
case when k 6∈ MI (x), where density ϕk [u] vanishes to high order at the boundary of the
integration domain on account of partitions of unity function ωk . Thus we can extend the
domain of integration in whole of R3 without affecting the value of integral (4.1.4). As it is
well known that, the trapezoidal rule exhibits super-algebraic convergence for smooth and
periodic integrands [10, 77], which we indeed employ to obtain accurate approximations
in this case. In contrast, when k 6∈ MB (x), a straightforward use of trapezoidal rule
does not produce high-order accuracy as although integrand is smooth but does not vanish
towards the discontinuity of inhomogeneity, that is, at t3 = 0. This is because, partition
of unity does not vanishes in this direction. We remedy this minor difficulty by utilizing
the trapezoidal rule for the integration with respect to t1 and t2 variables while employing
a composite Newton-Cotes quadrature in transverse variable t3 . The rate of convergence
of this approximation scheme directly depends on the order of Newton-Cotes quadrature
used in transverse variable t3 and could be enhanced arbitrarily as long as the smoothness
of m(x) = 1 − n2 (x) within Ω allows it.
As alluded above, integrands in (4.1.4) corresponding to the cases when k ∈ MB (x)
and k ∈ MI (x) are singular–a difficulty that does not exist in the previous case. This
kernel singularity possess major difficulty for numerical integrations and has to be removed
for high-order approximation. Similar to the two dimensions, to achieve rapidly convergent
approximations, we rely on analytic resolution of singularities through suitable changes of
parametric variables and application of high-order quadratures resulting in smooth inte-
grands. We localize the region where such coordinate transformations are affected to an
adequately small neighborhood of the singular point using a suitable smooth compactly
supported cut-off function. Indeed, we split integral in the right hand side of (4.1.5) into
95
two parts as follows:

ZZZ ZZZ
Gκ (x, ξk (t)) · · · dt = Gκ (x, ξk (t)) · · · η(t; ξk−1 (x)) dt
[0,1]3 [0,1]3
ZZZ
+ Gκ (x, ξk (t)) · · · (1 − η(t; ξk−1 (x))) dt, (4.1.6)
[0,1]3
where η(·; t0 ) is a C ∞ function such that it is compactly supported in a neighborhood of

the point ξk−1 (x) = t0 ∈ [0, 1]3 while η ≡ 1 in a smaller neighborhood of t0 . As in two
dimensions, this localization of singularity as seen in the first integral on the right hand
side of (4.1.6), brings in a two-fold benefit, namely, (a) it limits the relatively expensive
treatment of singularity since the region of integration reduces into the supp of η, which,
among other computational challenges, also demands an interpolation of grid-data to off-
grid quadrature points to a small integration domain, and (b) it allows for additional
speed-up in the computation of non-singular second term on the right hand side of (4.1.6).
Note that this regular integral, of course, can be integrated to high-order using the same
numerical quadratures that we apply in the case of x 6∈ Pk , as explained above.
It is straightforward to see that the direct application of the aforementioned integration
scheme leads to a computational complexity of O(N 2 ), where N = |T| is the size of the
set T, the total number of quadrature points in the Nyström scheme. As alluded in the
introduction, O(N 2 ) computational cost is a significant bottleneck for scattering simulation
of large scale problem. Therefor, to make our algorithm applicable for more practical
problems it is essential to compute these integrals efficiently.
Similar to the two dimensions, one can accelerate integral computations and attains
a reduced complexity O(N 1+δ/3 ), where δ ∈ (0, 1], by breaking the overall computation
of these integrals into two parts – a relatively small but specialized calculation in the
neighborhood of singularity and remaining non-singular contributions arising from the
voluminous bulk. Toward this, we begin by enclosing the inhomogeneity Ω by a cube
cell C of side length A. This cell is further partitioned into L3 identical cubical cells Cijk
(i, j, k = 1, · · · , L) of side length H = A/L such that the bounding cell C contains L
of these smaller cells along its sides. We assume, without loss of generality, that for the
96
chosen side length A and parameter L, cells Cijk do not admit inner acoustical resonance
[54, 87], more precisely the real number −κ2 is not a Dirichlet eigenvalue of the Laplace
operator in the domain Cijk . This, of course, can be easily ensured by necessary adjustment
in the choice of parameters A and L. We utilize these cells to break the computation of
integrals appearing in equation (4.1.4) into adjacent and non-adjacent calculations, which,
as pointed out above, is primarily motivated by our desire for the method to have a more
favorable computational complexity than the cumbersome O(N 2 ) cost.
Analogous to the two dimension, we call a source point y is adjacent to a fixed target
point x ∈ Cijk , if it belongs to the set N (x) defined by
N (x) = {y | y ∈ Clmn for some l, m, n satisfying |l − i| ≤ 1, |m − j| ≤ 1, |n − k| ≤ 1} .

(4.1.7)
Obviously, we say y is non-adjacent to x if y 6∈ N (x). Clearly, for each target point x
corresponding N (x) contains at most 36 neighboring cells. Based on this, we separate
the integrals over Pk in (4.1.4) into integrals over Pk ∩ N (x) accounting for the adjacent
contributions and the computationally large non-adjacent contributions, that is, integrals
over Pk ∩ N (x)c , where N (x)c denotes the compliment of N (x), for which a variant of
the two face FFT based acceleration strategy introduced in [23] is employed for efficient
calculations.
We thus, split integrals in (4.1.4) into adjacent and non-adjacent interactions by rewrit-
ing the expression for K[u](x) as
  
 X X Z XZ 
K[u](x) =  +  + Gκ (x, y)m(y)u(y)ωk (y) dy
 P ∩N (x) P ∩N (x)c
k∈MB (x) k6∈MB (x) k k∈IB k
  
 X X Z XZ 
+  +  + Gκ (x, y)m(y)u(y)ωk (y) dy.

k∈MI (x) k6∈MI (x)Pk ∩N (x) k∈II Pk ∩N (x)c 
(4.1.8)
We note here that the computational cost of carrying out the adjacent calculations for
all target points x ∈ T, where we propose to use O(N δ/3 ) grid points locally for accurate
97
evaluation of singular integration, is given by
Λadj = O(N 1+δ/3 ), (4.1.9)
where δ ∈ (0, 1] is mostly decided by the size of support of floating partition of unity η.
The computational cost needed in the evaluation of contributions to integrals coming from
Pk ∩ N (x)c for all target points x ∈ T, which we will explain later in section (4.3) is given
by !
5
N2

2 2 N3
Λnon-adj = O(LN log(LN )) + O
3 3 +O . (4.1.10)
L6 L2
This, of course, suggests that appropriate choice of parameter L reduces the over all com-
plexity for this calculation dramatically. For instance if we choose L = O(N 1/3 ), then
the computational complexity of the algorithm for non-adjacent calculations reduces to
the desired O(N log N ). Thus, the total computational cost for the evaluation of volume
integral operator (4.0.1) for all Nyström discretization points in T is given by
Λ = Λnon-adj + Λadj
= O (N log N ) + O(N 1+δ/3 )
= O(N 1+δ/3 ).
Note that, when material properties are infinitely differentiable within the scatterer Ω, our
algorithm exhibits O(N log N ) computational complexity for
log(log N )
δ=3
log N
while maintaining high-order accuracy of the approximation.
Now, in the following sections we discuss detail description of our integration scheme for
accurate evaluation of singular integration and acceleration strategy for efficient calculation
of non-singular non-adjacent interactions.
4.2 Singular Integration

This section will address the issue of high-order integration problem that arises when the
target point x belongs to the integration patch Pk . Recall that, in this case integral kernel
98
4.2. Singular Integration
Gκ (x, y) becomes unbounded when |x−y| goes to zero. This singular behavior of integrand
presents a significant difficulty in numerical integration and has to be resolved for accurate
approximations. In this section we have resolved this singular behavior of integrand by
introducing an appropriate change of parametric variable.
To this end, we start by breaking the integral (4.1.5) as follows:
Z ZZZ
Gκ (x, y)m(y)u(y)ωk (y) dy = Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)η(t; ξk−1 (x)) dt
Pk [0,1]3
ZZZ
+ Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)(1 − η(t; ξk−1 (x))) dt
ξk−1 (Pk ∩N (x))
Z
+ Gκ (x, y)m(y)u(y)ωk (y)(1 − η(ξk−1 (y); ξk−1 (x))) dy.
Pk ∩N (x)c
(4.2.1)
As enunciated above, we rely on the FFT-accelerator for computation of the third term
in (4.2.1) which accounts for the major part of integration. This, however, entails that
we compute this integral as a three dimensional convolution which, in turn, requires that
the integrand remains independent of localization cut-off function η. This, of course, as
in two dimension, can easily be achieved by ensuring that the support of η(·, ξk−1 (x)) is
contained in N (x). The second term in this expression, which we refer to as local correction
in the context of accelerated numerics, is computationally small and can be integrated to
high-order by means of trapezoidal rule in all parametric variable. We will give the detail
discussion of these aspects in section (4.3). The first term
ZZZ
Kksing [u](x) = Gκ (x, ξk (t))ϕk [u](t)ξk0 (t)η(t; ξk−1 (x)) dt, (4.2.2)
[0,1]3 ∩ supp η
on the right hand side of equation (4.2.1) is singular integral and detailed description of
high-order approximation of this integral is the subject of our following subsections.
99
4.2.0.1 Singular integration over interior patches
This subsection describes the high-order computation of Kksing [u](x) when k ∈ II . In our
implementation, we have defined localization function η as follows:
|t − ξk−1 (x)|

−1
η(t; ξk (x)) = χ , 0, 1 , (4.2.3)
r
where r > 0 is appropriately chosen real number. Note that, since support of η(t; ξk−1 (x))
contained with in ball of radius r centered at ξk−1 (x) = (tx1 , tx2 , tx3 ), therefore, we can extend
the domain of integration in (4.2.2) to the ball of radius r around ξk−1 (x) = (tx1 , tx2 , tx3 ).
Since ϕk [u] in eq. (4.2.2) vanishes to high order along the boundary of the integration
domain, this process does not compromise on the smoothness of the integrand. Now in
order to resolve the kernel singularity in (4.2.2), we change to spherical coordinates centered
at (tx1 , tx2 , , tx3 ):
t1 = tx1 + ρ cos θ sin φ
t2 = tx2 + ρ sin θ sin φ
t3 = tx3 + ρ cos φ.
If we take
ξ˜kx (ρ, θ, φ) = ξk (tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ),
0
ξ˜kx (ρ, θ, φ) = ξk0 (tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ),
η̃ x (ρ, θ, φ) = η ((tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ); (tx1 , tx2 , tx3 ))
and
0
ϕ̃xk [u](ρ, θ, φ) = ϕk [u](tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ)ξ˜kx (ρ, θ, φ)η̃ x (ρ, θ, φ)
then the integral (4.2.2) takes the form
Zπ Z2π Zr
Kksing [u](x) = 2
ρ Gκ ˜
x, ξk (ρ, θ, φ) ϕ̃xk [u](ρ, θ, φ) sin φ dρ dθ dφ.
x
(4.2.4)
0 0 0
100
4.2.0.1. Singular integration over interior patches
The appearance of additional factor ρ2 sin φ (the Jacobian of the spherical change of vari-
ables in parametric variables) in the integrand of integral (4.2.4) cancels the kernel sin-

gularity as ρ2 Gκ x, ξ˜kx (ρ, θ, φ) , clearly, is a smooth function of ρ. Additionally, the
smooth cut-off function η vanishes to high order at the boundary of integration interval in
ρ variable and therefore application of trapezoidal rule yields high-order convergence for
the ρ-integration. The θ-integral can also be approximated to high order by employing
trapezoidal rule quadrature as the corresponding integrand is smooth and periodic. The
integrand in φ-variable is smooth but not periodic. Thus, trapezoidal rule is not appro-
priate quadrature for approximation of this integral as it yields low order convergence.
However, this integral can certainly be approximated to high-order by using classical high-
order quadrature rule, for instance use of Clenshaw-Curtis quadrature [33] gives high-order
convergence and thus our high-order integration problem for (4.2.2) is solved.
While the change to spherical coordinates provides a way to resolve the singularity
analytically, the proposed quadratures demands that we provide the values ϕ̃xk [u] at points
outside of the computational grid T. This necessitates employing an efficient and accurate
interpolation strategy for evaluation of ϕ̃xk [u] at these newly transformed grid points in new
coordinates (ρ, θ, φ). Toward this, we adapt the following Fourier refined polynomial inter-
polation, which is similar to the one introduced in [20], with a suitable choice of polynomial
degree, to retain high order accuracy while maintaining computational efficiency:
1. Compute Fourier coefficients of density ϕk [u](ξk−1 (x)) on each of the original grid
point {ξk−1 (x) = (t1 , t2 , t3 ) ∈ [0, 1]3 | x ∈ Tk } by using its values on these grid points.
Since, density ϕk [u](ξk−1 (x)) is smooth and periodic, therefore, this computation can
be performed efficiently by means of three-dimensional FFTs.
2. Evaluate the Fourier series on much refined uniform grid. This is accomplished by
taking three-dimensional inverse FFTs of original coarse data padded with zeros.
3. To obtain the value of density ϕk [u](ξk−1 (x)) at off grid points, we collect the values of
ϕk [u](ξk−1 (x)) at the (R+1)×(R+1)×(R+1), refined grid points in the neighborhood
101
of ξk−1 (x). Now using these values construct an interpolating polynomial of degree
R and finally obtain the value of ϕk [u] at off grid point ξk−1 (x).
This interpolation procedure is highly accurate and requires only O(NI log NI ), floating
P
point operations, where NI = k∈II |Tk | denote the total number of Nyström nodes coming
from the discretization of interior patches. Moreover, error incurred by this interpolation
procedure is negligible compared to the other sources of error introduced in the overall
algorithm. Thus one can safely ignore the error produced by this interpolation procedure.
4.2.0.2 Singular integration over boundary patches
This subsection addresses the high-order integration problem for the case when target point
x lies in one of the boundary integration patches, say Pk with k ∈ MB (x). In contrast
with interior patches, we adopt a different approximation strategy to evaluate Kksing [u](x).
The first difference comes in the form of the choice of the cut-off function η where a circular
support used for interior patches becomes unsuitable for use near the physical boundary
of the domain, that is, near t3 = 0. In this case, following the ideas introduced in [7], we
take the localization function supported in cube around ξk−1 (x) = (tx1 , tx2 , tx3 ), given by
|(t1 , t2 ) − (tx1 , tx2 )| |t3 − tx3 |

x x x
η((t1 , t2 , t3 ); (t1 , t2 , t3 )) = χ , 0, 1 χ , 0, 1 . (4.2.5)
r r
We represent Kksing [u](x) as

Z1
0
where
Jk (t; x) =
|(t1 , t2 ) − (tx1 , tx2 )|
ZZ
0
Gκ (x, ξk (t1 , t2 , t))ϕk [u](t1 , t2 , t)ξk (t1 , t2 , t)χ , 0, 1 dt1 dt2 .
r
B ((tx
1 ,t2 ),r )
x
(4.2.7)
The evaluation of Jk (t; x) poses difficulties owing to singularity present in the integrand
at the surface t = tx3 . In addition, high-order integration of Jk (t; x) is exacerbated by
102
4.2.0.2. Singular integration over boundary patches
near singularity and oscillatory behavior of integrand in the vicinity of singular surface
t = tx3 , because, when t approaches to tx3 , integral kernel Gκ (x, ξk (t1 , t2 , t)) becomes nearly
singular and oscillatory. To alleviate these difficulties, the following two fold strategy is
used:
a. First, use a change of polar coordinate system (ρ, θ) centered at the point (tx1 , tx2 )
t1 = tx1 + ρ cos θ
t2 = tx2 + ρ sin θ,
which transform the integral in (4.2.7) to
Z2π Zr
Jk (t; x) = dθ Gκ (x, ξk (tx1 + ρ cos θ, tx2 + ρ sin θ, t)) ρϕ̃k [u](tx1 +ρ cos θ, tx2 +ρ sin θ, t) dρ,
0 0
(4.2.8)
where ϕ̃k [u](tx1 + ρ cos θ, tx2 + ρ sin θ, t) =

0 |ρ|
ϕk [u](tx1 + ρ cos θ, tx2 + ρ sin θ, t)ξk (tx1 + ρ cos θ, tx2 + ρ sin θ, t)χ , 0, 1 .
r
It can be readily verified that, owing to appearance of additional factor ρ, integrand in the
radial integral of (4.2.8) is uniformly integrable with respect to variable θ.
b. Using an additional change of variables ρ = %(τ ), where the smooth invertible odd
function %(τ ) satisfies
dm %(τ )
= 0 for m = 0, ..., M, (4.2.9)
dτ m τ =0

(for example, %(τ ) = τ M +1 for even, non-negative integer M ), equation (4.2.8) can be
written as
−1 (r)
%Z
Z2π
0
Jk (t; x) = dθ Gκ (x, ξk (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t)) %(τ )% (τ )f˜k (τ, θ, t) dτ,
0 0
(4.2.10)
where
f˜k (τ, θ, t) = ϕ̃k [u](tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t).
103
This change of variable, renders the integrand in (4.2.10) M times differentiable in

τ that are also uniformly bounded. In addition, in view of the localization function χ,
the integrand with respect to τ -variable in equation (4.2.10) vanishes on the boundary of
the integration interval, i.e., τ = 0 and τ = %−1 (r), together with all of its derivatives.
Thus, the integrand with respect to τ -variable in integral (4.2.10) is a smooth and periodic
function which can be integrated to high-order by means of trapezoidal rule. Further, the
application of trapezoidal rule in θ-variable also provide high-order convergence since the
corresponding integrand is smooth periodic function of θ.
Note that, as in two dimensions, this change in a parametric variables produces a set
of quadrature points in (τ, θ) that do not lie on the Cartesian grid associated with the
given integration patch Pk requiring an interpolation strategy for which we utilize the
FFT refined polynomial interpolation [7] which is known to be accurate while maintaining
computational efficiency. The final step in high-order approximation of Kksing [u](x) relates
to the the computation of t3 -integral in (4.2.6). The main difficulty, here, is encountered
in the form of a jump discontinuity in the t-derivative of Jk (t; x) at t = tx3 . In order to
overcome this difficulty, we split the integral in (4.2.6) as
x
Zt3 Z1
Kksing [u](x) = Jk (t; x)χ(|t − tx3 |/r, 0, 1)dt + Jk (t; x)χ(|t − tx3 |/r, 0, 1)dt, (4.2.11)
0 tx
3
where both integrands are smooth in its domain of integration and in principle can be
approximated to high-order by means of Q-point Newton-Cotes quadrature. This, however,
as in two dimensions, present a practical difficulty in the form of requiring at least Q-
equidistant grid points in [0, tx3 ], and [tx3 , 1], that, of course, is not available when tx3 is
close to either 0 or 1. The direct interpolation of Jk (t; x), however, is neither accurate, (on
account of non-smoothness of Jk (t; x) at t = tx3 ), nor efficient (because of its dependence
on tx3 ). However, this expensive computation can be easily avoided by splitting (4.2.11)
exactly in the same way as in (2.3.19).
Remark 4.2.1. It is clear that the strategy described above is valid when the target point
x ∈ T happens to lie on one of the parallel surface {(t1 , t2 , t3j )|(t1 , t2 ) ∈ [0, 1]2 }. Obviously,
104
4.3. Non-singular integration: Acceleration
the overall computational scheme require us to compute Kksing [u](x) for target point x that
do not coincide with any of the grid lines. In particular, this happens when the integration
on a boundary patch corresponds to a target point coming from an interior patch. To deal
with such interactions between boundary and interior patches, we follow the same strategy
as in two dimensions. We first compute
X
KB [u](x) = Kk [u](x)
k∈IB
at all boundary discretization points x ∈ ∪k∈IB Tk and then set up an FFT-refined poly-
nomial interpolation schemes for all boundary patch Pk k ∈ IB making possible accurate
evaluations of KB [u](x) at off-grid points.
4.3 Non-singular integration: Acceleration

As mentioned in the section (4.1), for the fast computation of non-singular integrals
!
X X Z
Kreg [u](x) = + Gκ (x, y)m(y)u(y)ωk (y) dy, (4.3.1)
k∈IB k∈II Pk ∩N (x)c
that appear in (4.1.8), we rely on an equivalent sources approximation strategy [23] to

improve the overall computational efficiency.
Indeed, we can accurately approximate Kreg [u](xq ) at each grid point xq ∈ T at an
O(N 2 ) cost by employing a high order quadrature, say, given by
X
Kareg [u](xq ) = w` Gκ (xq , y` )m(y` )u(y` )ωk (y` ). (4.3.2)
y` ∈T\N (xq )
Noticing that (4.3.2) is a discrete convolution whose contributing sources y` are located
somewhat irregularly in Ω. As in two dimensions, the acceleration strategy seeks to replace
them by a certain set of “equivalent sources” placed on a Cartesian grid, that produces an
accurate approximation for the convolution, to facilitate the use of three dimensional FFT
for computing the discrete convolution at an O(N log N ) cost.
reg
More precisely, for each cell Cijk , if Ka,ijk [u](x) denotes the quantity
reg
X
Ka,ijk [u](x) = w` Gκ (x, y` )m(y` )u(y` )ωk (y` ),
y` ∈Cijk
105
(m) (d)
then we seek constants σijk,` and σijk,` defining the approximating quantity
eq
N
(d) ∂Gκ (x, xijk,` )
reg,eq
X (m)
Ka,ijk [u](x) = σijk,` Gκ (x, xijk,` ) + σijk,` . (4.3.3)
`=1
∂ν(xijk,` )
such that
N coll
X reg reg,eq
2
Ka,ijk [u](xq ) − Ka,ijk [u](xq )
q=1
is minimized for a fixed N coll number of evaluation points xq on the boundary of the set
N (x). The locations xijk,` of equivalent sources in (4.3.3) are equidistant points placed on
the two opposite and parallel faces of the cells cijk .
It is known that one can achieve this approximation up to the prescribed accuracy
O(), provided the number N eq is chosen as [23]
κ2 A2 2

eq
N = max 4 2 , 4c () ,
L
where
log()
c() = − .
log(3)
(m) (d)
The unknown quantity σijk,` and σijk,` in eq. (4.3.3) are obtained as solution to the
overdetermined linear system Aσ ijk = b. It is important to note that, as the geometry
is identical for each cell Cijk , the QR factorization of the above matrix A need only be
computed once and saved for repeated use. For numerical stability of the least-squares
solver, we choose N coll = 4N eq . A straightforward counting shows that this process of
equivalent source computation for cells requires O(N 5/3 /L2 ) + O(N 2 /L6 ) operations in
total.
Clearly, for x ∈ Cijk , the computation of
L X
L X
L i+1 j+1 k+1
X reg,eq
X X X reg,eq
Kareg,eq [u](x) = Ka,ijk [u](x) − Ka,lmn [u](x)
i=1 j=1 k=1 l=i−1 m=j−1 n=k−1
as an approximation to Kareg [u](x) at all points on the grid can now be obtained by means of

three dimensional FFTs with a computational cost of O LN 2/3 log(LN 2/3 ) . The values
106
4.3. Non-singular integration: Acceleration
obtained in this manner provide accurate approximations for non-adjacent non-singular

interactions at points xijk,` on the faces of cells Cijk .
Finally, to obtain these values at a true source location, say for x ∈ Cijk , we solve
the free space Helmholtz equation within Cijk , with Dirichlet boundary data coming from
Kareq,eq [u](xijk,` ). To efficiently obtain solutions to these well-posed Dirichlet boundary
value problems, we utilize a discretized plane wave expansion of the form [49]
N w
X
Kareq,eq [u](x) ≈ γ` exp(iκd` .x), (4.3.4)
`=1
where the unit vectors d` sample the surface of unit sphere with sufficient degree of unifor-
mity. Use of this approach is motivated by the spectral convergence of the above expansion
with respect to the number of unit vectors d` used and the fact that the wave expansion
coefficients γ` can be obtained as solutions to overdetermined linear systems of the form
Bγ = βijk where the matrix B remains unchanged for each cell, again, owing to the
identical geometry of cells Cijk s.
Thus, Kareg,eq [u](x) can be evaluated at all true source locations x ∈ Cijk at a computa-
tional cost of O(N 5/3 /L2 ) + O(N 2 /L6 ) and the overall cost of evaluation of non-adjacent
non-singular interactions, therefore, stands at
5
!
N2

2 2 N3
Λnon-adj = O(LN log(LN )) + O
3 3 +O . (4.3.5)
L6 L2
107
Chapter
5
Computational Results In Three
Dimensions
This chapter presents several numerical results obtained by implementing our three dimen-
sional algorithm, discussed in chapter (4). The goals of this section are to demonstrate
the high-order convergence and capability of our algorithm in dealing with complex shape
scatterer through numerical examples in three dimensions. As in two dimensions, all the
algorithmic components of the proposed algorithm are implemented in C++ apart from
the figure visualization, which uses Matlab. In all the cases ( evaluation of Fourier series
for the construction of interpolation and evaluation of discrete convolution in accelera-
tion strategy) FFT computations are carried out by of means fftw3 package. We have
also used some relevant subroutines of LAPACK package for the solution of linear least
square system, arises in three dimensional accelerated computation. In all the experiments,
near fields are computed under the plane wave incidence, ui (x) = eiκx.d , where d denotes
the direction of propagation of incoming plane wave. Solution of the dense linear system
arising from the discretization of three dimensional Lippmann-Schwinger equation Equa-
tion (1.2.38) is obtained by means of the iterative solver GMRES [97]. Following formulae
109
Chapter 5. Computational Results In Three Dimensions
have been used for the computation of relative max error ∞ and relative root mean square
error 2 at Nyström nodes T:
kuexact − uapprox k∞
∞ =
kuexact k∞
N  21
P exact approx 2
 i=1 |u (xi ) − u (xi )| 
2 = 

N
 ,

P
|uexact (xi )|2
i=1
where uapprox denotes the solution computed by our algorithm and uexact is exact solution.
The order of convergence of our algorithm which relates to the rate at which the error
in the computed solution decreases as the discretization scale decreases is given by the
following formula:

EN
Order = log2 , (5.0.1)
E2N
where EN denote the relative error corresponding to N number of discretization points. In
all the convergence tables, we write “Order” to denote the numerical order of convergence of
the approximation and “Iter” stands for number of GMRES iteration required to produce
the desired accuracy.
Although, we have implemented accelerated algorithm for the fast evaluation of nonad-
jacent interactions but still some part of the code has not been optimized yet. Similar to
the two dimensional case, in all the convergence tables, presented in this chapter, we use
notation p1 × n1 × n2 × n3 + q1 × m1 × m2 × m3 to denote the underlying computational
grid. Implying that a p1 number of boundary patches, each with n1 × n2 × n3 discretization
points, and a q2 number of interior patches with n1 ×n2 ×n3 grid points on each. These two
sets of points constitute the Nyström grid T. Our approximation scheme for the integral
operator, in contrast to two dimensional counter part, is based on the use of three different
quadrature namely, spectrally accurate trapezoidal rule, Clenshaw-Curtis quadrature and
Q-point Newton-Cotes quadrature. Therefore, the expected convergence rates (computed
by using formula (5.0.1)) are essentially that of Newton-Cotes, that is, either Q or Q + 1
depending on whether Q is even or odd, provided other parameters, such as the degree
110
of the polynomial in the interpolators, the number M for the change of variable % , etc
are chosen favorably. We demonstrate the increase in the convergence rate with respect to
increasing Q through different sets of experiments, one using a 3-points while other using
5-points Newton-Cotes quadrature.
As we have discussed in chapter 4, our interpolation strategy demands evaluation of
Fourier series on finer grid which is saved and used further for polynomial interpolation.
Although, use of FFT evaluation of Fourier series on finer grid is very fast but at higher
discretization, a large amount of memory is required to save these values. Therefore, in
contrast to the two dimensional case, we have used 8-fold refinement instead of 16. In other
words, the length of the refined grid is 8 times smaller than that of original grid points.
For the pictorial visualization, we will display couple of figures created by plotting the
numerical solution obtained by implementation of our algorithm. For the sake of visualiza-
tion of fields on and off the scatterer, we have computed the fields on the orthogonal slices
which intersect the scatterer and also cover some planner region outside the inhomogeneity
by means of precomputed solution on the Nyström nodes.
Example 5.0.1. (Convergence study for a simple scatterer)
The first example that we study to verify the high-order convergence of our three dimen-
sional algorithm is a penetrable sphere. For this example, problem is explicitly solvable, so
that the comparison with the exact solution is possible. The incident plane wave impinges
on the scatterer from the positive z-axis which is described as eiκx.(0.0.1) = eiκz in the Carte-
sian coordinates. The whole scatterer is covered by three overlapping coordinate patches,
namely, two boundary patches and one interior patch(see figure (4.1) in chapter (4)). Since
the boundary ∂Ω of scatterer Ω can not be represented by a single parametrization, there-
fore, more than one patches have been used to describe the boundary ∂Ω of scatterer
by overlapping coordinate patches. In practice, we noticed, it is advantageous to utilize
as large patches as the geometry of scatterer permits. Indeed, this approach provides a
compromise between a desirable global discretization and the impossibility of representing
the boundary of complicated scatterer by a single invertible parametrization. As we have
111
Figure 5.1: Spherical shape scatterer, boundary is covered by two overlapping coordinate
patches. Discretization on the outer surface of the scatterer coming from a typical volu-
metric computational grid.
described in Chapter 4, for the sake of efficiency, we allow substantial overlap between the
patches (see Figure 5.1), so that the derivatives of the partition of unity remain small.
First in order to demonstrate the high-order convergence of the algorithm we present
a convergence study, in Tables 5.1 to 5.8, corresponding to following experiments:
√
1. κa = 2, n(x) = 3 for x ∈ Ω
√
2. κa = 10, n(x) = 2 for x ∈ Ω.
112
Here, a denotes the diameter of inhomogeneity of the scatterer which is taken to be 2 in

this example.
The numerical solutions in this study are computed for several discretization levels while
employing a 3-point Newton-Cotes quadrature for the approximation of t3 -integration in
boundary patches. Numerical results corresponding to these experiments clearly illustrate
that the methodology achieves the expected rate of convergence. Further, to show an
enhanced order of convergence, we repeat the same experiment but use a 5-point Newton-
Cotes quadrature instead of 3-point in the computation of the integral operator Equa-
tion (4.0.1). The corresponding numerical results are presented in Tables 5.5 to 5.8, where,
as expected, we see an increased rate of convergence.
Grid Size L2 L∞
2×3×5×5+1×5×5×5 1.26e-01 - 2.30e-01 -
2×5×9×9+1×9×9×9 5.10e-02 1.30e+00 1.94e-01 2.46e-01
2 × 9 × 17 × 17 + 1 × 17 × 17 × 17 1.06e-02 2.27e+00 2.87e-02 2.76e+00
2 × 17 × 33 × 33 + 1 × 33 × 33 × 33 6.48e-04 4.02e+00 1.83e-03 3.97e+00
2 × 33 × 65 × 65 + 1 × 65 × 65 × 65 2.93e-05 4.47e+00 1.15e-04 4.00e+00
Table 5.1: Convergence study: √ Forward map computation for a sphere of acoustic size
κa = 2 and refractive index n = 3 when 3-point Newton-Cotes quadrature employed for
t3 -integration over boundary patches.
Grid Size Iter L2 L∞

2×3×5×5+1×5×5×5 1 9.63e-02 - 1.47e-01 -
2×5×9×9+1×9×9×9 1 8.73e-02 1.42e-01 1.41e-01 6.04e-02
2 × 9 × 17 × 17 + 1 × 17 × 17 × 17 2 7.64e-03 3.51e+00 2.20e-02 2.68e+00
2 × 17 × 33 × 33 + 1 × 33 × 33 × 33 3 3.77e-04 4.34e+00 1.15e-03 4.27e+00
2 × 33 × 65 × 65 + 1 × 65 × 65 × 65 4 1.48e-05 4.67e+00 6.74e-05 4.09e+00

√ Plane wave scattering by a sphere of acoustic size κa = 2
113
Grid Size L2 L∞
2×3×5×5+1×5×5×5 6.88e-01 - 5.30e-01 -
2×5×9×9+1×9×9×9 1.24e-01 2.47e+00 1.50e-01 1.82e+00
2 × 9 × 17 × 17 + 1 × 17 × 17 × 17 2.46e-02 2.33e+00 4.06e-02 1.88e+00
2 × 17 × 33 × 33 + 1 × 33 × 33 × 33 1.62e-03 3.92e+00 2.43e-03 4.06e+00
2 × 33 × 65 × 65 + 1 × 65 × 65 × 65 7.53e-05 4.43e+00 1.45e-04 4.06e+00

√ map computation for a sphere of acoustic size

2×3×5×5+1×5×5×5 4 1.38e+00 - 1.00e+00 -
2×5×9×9+1×9×9×9 9 3.82e-01 1.85e+00 4.17e-01 1.27e+00
2 × 9 × 17 × 17 + 1 × 17 × 17 × 17 13 8.31e-02 2.20e+00 7.14e-02 2.54e+00
2 × 17 × 33 × 33 + 1 × 33 × 33 × 33 17 4.74e-03 4.13e+00 4.00e-03 4.16e+00
2 × 33 × 65 × 65 + 1 × 65 × 65 × 65 19 1.99e-04 4.57e+00 2.12e-04 4.24e+00

Grid Size L2 L∞
2×5×5×5+1×5×5×5 1.35e-01 - 2.87e-01 -
2×9×9×9+1×9×9×9 3.58e-02 1.92e+00 1.61e-01 8.32e-01
2 × 17 × 17 × 17 + 1 × 17 × 17 × 17 1.61e-02 1.15e+00 3.56e-02 2.18e+00
2 × 33 × 33 × 33 + 1 × 33 × 33 × 33 4.20e-04 5.26e+00 1.17e-03 4.93e+00
2 × 65 × 65 × 65 + 1 × 65 × 65 × 65 5.69e-06 6.21e+00 3.51e-05 5.06e+00
Table 5.5: Convergence study: √ Forward map computation for a sphere of acoustic size
κa = 2 and refractive index n = 3 when 5-point Newton-Cotes quadrature employed for
t3 -integration over boundary patches.
For the pictorial visualization of the computed total and scattered field, next set of
examples correspond to simulations for significantly large scatterer κa = 20. The incident
plane wave again propagate along the positive z-axis and refractive index of the material
114

2×5×5×5+1×5×5×5 1 8.93e-02 - 1.71e-01 -
2×9×9×9+1×9×9×9 1 8.19e-02 1.25e-01 1.42e-01 2.61e-01
2 × 17 × 17 × 17 + 1 × 17 × 17 × 17 2 5.16e-03 3.99e+00 1.95e-02 2.86e+00
2 × 33 × 33 × 33 + 1 × 33 × 33 × 33 4 1.11e-04 5.54e+00 5.48e-04 5.15e+00
2 × 65 × 65 × 65 + 1 × 65 × 65 × 65 5 2.00e-06 5.79e+00 1.15e-05 5.57e+00

Grid Size L2 L∞
2×5×5×5+1×5×5×5 6.29e-01 - 5.88e-01 -
2×9×9×9+1×9×9×9 1.11e-01 2.51e+00 1.45e-01 2.02e+00
2 × 17 × 17 × 17 + 1 × 17 × 17 × 17 1.13e-02 3.29e+00 3.85e-02 1.91e+00
2 × 33 × 33 × 33 + 1 × 33 × 33 × 33 5.15e-04 4.45e+00 9.53e-04 5.33e+00
2 × 65 × 65 × 65 + 1 × 65 × 65 × 65 1.05e-05 5.62e+00 3.16e-05 4.91e+00

√ map computation for a sphere of acoustic size

2×5×5×5+1×5×5×5 4 2.31e+00 - 2.47e+00 -
2×9×9×9+1×9×9×9 9 3.80e-01 2.61e+00 4.37e-01 2.50e+00
2 × 17 × 17 × 17 + 1 × 17 × 17 × 17 13 3.46e-02 3.45e+00 5.37e-02 3.02e+00
2 × 33 × 33 × 33 + 1 × 33 × 33 × 33 20 5.19e-04 6.06e+00 8.96e-04 5.91e+00
2 × 65 × 65 × 65 + 1 × 65 × 65 × 65 22 9.73e-06 5.74e+00 2.05e-05 5.45e+00

115
√
is 2. We discretize the scatterer with 2 × 33 × 65 × 65 + 1 × 65 × 65 grid points and
the GMRES iteration is terminated when the residual is reduced by a factor of 10−4 . By
using the resulting solution on the Nyström grid points T, we computed the solution on
the three perpendicular orthogonal slices and results are presented in Figures 5.4 to 5.6.
In Figure 5.4, we present the real part of the scattered and total field while in Figures 5.5
and 5.6, we visualize imaginary and absolute value of scattered and total field.
Example 5.0.2. (Accuracy of non-singular non-adjacent interactions)
In chapter 4, we have delineated the acceleration strategy for the expedited evaluation
of non-adjacent interactions in three dimensions. This strategy is based on substitution
of true sources within each cell by certain equivalent sources on the two parallel faces
of individual cells. The non-adjacent field generated by these equivalent sources, i.e.,
reg,eq
Ka,ijk [u], is equal, outside the adjacent set N (x), to the field generated by true source
reg
Ka,ijk [u], inside that cell in the least square sense. In this example, we present the numerical
results that corroborate the feasibility of such equivalent source identification.
For fixed wavenumber κ and fixed cell size H, field generated by true sources, i.e., true
field, compared against its approximated value computed by equivalent source approxima-
tions. The relative error of this approximation in max norm ∞ and in root mean square
norm 2 are reported in Table (5.9). As we anticipated, approximate field converges to
true field with spectral accuracy as number of equivalent source point on the parallel faces
of cells cijk increases. The relative error incurred by approximation for different cell size,
shown in Table (5.10). These numerical results clearly demonstrate the spectral accuracy
of approximation for small as well as large acoustical cells. Another important observation
is the accuracy of our approximation does not deteriorate when we double the acoustical
size of cell κH keeping the number of point per wavelength constant.
These numerical results confirm the spectral accuracy of acceleration strategy used for
the approximation of discrete counterpart of non-singular non-adjacent interactions and
therefore the overall error of the algorithm is dominated by quadrature error incurred in
the approximation of integrals over patches.
116
Equivalent Source/cell ε2 ε∞
6×7×7×7 3.72e-02 3.18e-01
6 × 10 × 10 × 10 2.17e-06 1.24e-05
6 × 14 × 14 × 14 4.24e-09 1.23e-08
6 × 18 × 18 × 18 7.72e-10 8.34e-10
10π
Table 5.9: Accuracy of acceleration for fixed wavenumber κ = 3
κH Equivalent Source/cell ε2 ε∞
2 6×6×6×6 6.12e-06 4.80e-05
4 6 × 12 × 12 × 12 1.47e-08 2.50e-08
8 6 × 24 × 24 × 24 5.92e-10 6.38e-10
16 6 × 36 × 36 × 36 1.52e-09 2.53e-09
Table 5.10: Accuracy of acceleration for increasing acoustical cell size.
Example 5.0.3. (Scattering by complex shape scatterer)
In order to show the feasibility of our algorithm to deal with complex shape scatterer
we next provide simulation results for bean shape obstacle. The boundary of this obstacle
is defined by the equation
x2 (β1 R cos πz
R
+ y)2 z 2 2
2 πz + 2 πz + 2 = R ,
a1 (1 − β3 cos R ) a2 (1 − β2 cos R ) a3
where a1 = 0.8, a2 = 0.8, a3 = 1, β1 = 0.3, β2 = 0.4 and β3 = 0.1. This scatterer is
non-symmetric, non-convex, as depicted in Figure (5.3). The largest dimension of this
geometry is along z-coordinate and is equal to 2R. In order to apply our algorithm, we
have decomposed this geometry into three overlapping coordinate patches namely, two
boundary patches and one bulky interior patch (see Figure (5.2)). As in the spherical
geometry shape, we allow substantial overlap of the patches (see Figure 5.3), so that the
derivatives of the partition of unity remain small. Note that, [7, 23, 25], to describe the
surface of bean shape scatterer more than two patches have been used however, we use only
two patches to decompose boundary as a overlapping coordinate patches (see Figure 5.3).
This is an important observation, because, as we explain in Example 1, use of minimum
number of patches saves significant amount of time.
117
(a) Bean shape scatterer (b) Boundary patch
(c) Boundary patch, discretization grid com- (d) Bulk Interior patch, discretization
ing from a typical volumetric computational coming from a typical volumetric compu-
grid. tational grid.
Figure 5.2: Three overlapping coordinate patches and corresponding computational grid
used for scattering calculation of the bean shape scatterer.
118
Figure 5.3: Two patch covering of outer surface of Bean shape scatterer and the Nyström
grid comes from a typical volumetric computational grid.
We computed plane wave scattering propagating along positive y-direction, that is,
d = (0, 1, 0) by an inhomogeneous bean shape obstacle of size κa = 20 with refractive
√
index n = 2. We discretized the set of boundary patches with 2 × 33 × 65 × 65 and set
of interior patches with 1 × 33 × 65 × 65 grid points. The GMRES iteration is terminated
after 63 iteration by then the residual is reduced by a factor of 10−3 . By using the resulting
solution we have computed the fields on the the three orthogonal planes parallel to the
coordinate axis. Simulation results for this experiments are presented in Figures 5.7 to 5.9.
119
In Figure (5.7), the real part of the computed scattered field together with the total field
are given. In Figures 5.7 to 5.9 imaginary and absolute part of computed solutions are
presented.
120
Figure 5.4: Visualization of real part of the total and scattered field on the three orthogonal
slices from penetrable
√ spherical shape scatterer of acoustical size κa = 20 and refractive
index n(x) = 2, under the plane wave incidence coming form positive z-axis.
121
Figure 5.5: Visualization of imaginary part of the total and scattered field on the three
orthogonal slices from penetrable
√ spherical shape scatterer of acoustical size κa = 20 and
refractive index n(x) = 2, under the plane wave incidence coming form positive z-axis.
122
(a) Absolute value of the scattered field, Abs(us )
(b) Absolute value of the total field, Abs(u)
Figure 5.6: Visualization of absolute value of the total and scattered field on the three
√ spherical shape scatterer of acoustical size κa = 20 and
refractive index n(x) = 2, under the plane wave incidence coming form positive z-axis.
123
Figure 5.7: Visualization of real part of the total and scattered field on the three orthogonal
slices from
√ penetrable bean shape scatterer of acoustical size κa = 20 and refractive index
n(x) = 2 under the plane wave incidence coming form positive y-axis.
124
Figure 5.8: Visualization of imaginary part of the total and scattered field on the three
√ bean shape scatterer of acoustical size κa = 20 and
refractive index n(x) = 2, under the plane wave incidence coming form positive y-axis.
125
(a) Absolute part of the scattered field, Abs(us )
(b) Absolute part of the total field, abs(u)
Figure 5.9: Visualization of absolute value of the total and scattered field on the three
√ bean shape scatterer of acoustical size κa = 20 and
refractive index n(x) = 2 under the plane wave incidence coming form positive y-axis.
126
Chapter
6
Error Analysis
The Nyström method proposed in this thesis is based on approximation of volume integral
operator (1.2.38) whose kernel is weakly singular in its domain of integration. It is quite
well known that, one of the main hindrance in the design of integral equation based solver
concern development of accurate quadrature rules for approximation of singular integrals.
In Chapter 2 and Chapter 4 we have described an efficient high-order quadrature scheme
for the approximation of volume integral operator (1.2.38). In this chapter, we provide
theoretical error estimates for the quadrature described in Chapters 2 and 4. In particular
we establish high-order convergence of the proposed quadrature in a case in which the
refractive index n(x) of the material is smooth within the scatterer.
As, we have enunciated in Chapters 2 and 4, our high-order integration scheme splits
the kernel into two parts; regular and singular. Then, integrals with both type of kernel
treated separately. The necessary function values required for the approximation of singu-
lar integrals are obtained by highly accurate interpolators. Thus, error produced by our
integration scheme comprises two main errors: error in the approximation of integrals by
quadratures and error incurred by interpolators.
In Section 6.1 we establish error estimates for our interpolators in two and three dimen-
127
Chapter 6. Error Analysis
sions. Subsequently, in Section 6.2, we derive error estimate for the quadrature discussed
in Chapters 2 and 4.
Remark 6.0.1. 1. Throughout this chapter we use C to denote a generic constant;

different instance may not denote the same constant.
p,α
2. Throughout in this chapter, by, Cper we denote the space of all periodic functions
whose p-th derivative is Hölder continuous.
6.1 Interpolation Errors

In this section we provide error bounds on the interpolation operators which we have
discussed in chapter 2 and chapter 4. Recall that, our interpolation strategy is consist of
two steps. First, we obtain the value of unknown periodic density ϕk [u] on an uniform
refined grid by evaluating its discrete Fourier transform on these grids. Subsequently,
polynomial interpolation of fixed degree R is used to obtain the density values ϕk [u] at
points not on the grid. It is important to note that, use of polynomial interpolation of
fixed degree R, plays an essential role in both, the performance of the overall interpolation
scheme (see chapter 2 and chapter 4 for details) and its theoretical analysis. With regard
to the latter, use of fixed degree polynomial interpolations provide us bound on its operator
norm, since for a fixed degree polynomial interpolation Lebesgue number is constant and
does not grow with reduction in mesh length.
To this end, we first derive error estimates for one dimensional interpolations and
subsequently use these results to derive error estimate in two and three dimensions.
p,α
Suppose ϕ ∈ Cper ([a, b]), where a < b ∈ R. Then one dimensional trigonometric
interpolation on N number of even equidistant mesh points
4[a,b],N = {ti |ti = a + ih, 0 ≤ i < N, h = (b − a)/N } , (6.1.1)
is given by
N
X −1
P1,N
T
[ϕ](t) = ϕ(ti )LTi (t), for all t ∈ [a, b]. (6.1.2)
i=0
128
6.1. Interpolation Errors
Here, LTi (t) denotes the Lagrange basis for trigonometric interpolation, given by [77, page
182] ( )
n−1
1 X
LTi (t) = 1+2 cos kL(t − ti ) + cos nL(t − ti ) , (6.1.3)
N k=1
where n = N/2 and L = 2π/(b − a). Note that
LTi (tj ) = δij , 1 ≤ i, j < N,
where δij is Kronecker’s delta 

1 if i = j

δij =
0 if i 6= j.

Theorem 6.1.1. [77, Theorem 11.4] For N ≥ 2, the trigonometric interpolation operator
with N equidistant interpolation points has norm
P1,N < C ln(N ),

T
∞
where C > 0 is a positive constant independent of N .
Following theorem provide a uniform error estimate and implies uniform convergence
for the one dimensional trigonometric interpolation of Hölder continuous functions.
Theorem 6.1.2. [77, Theorem 11.6] Let p ∈ N ∪ {0} and 0 < α ≤ 1. Then for the one
dimensional trigonometric interpolation we have
p+α
P1,N [ϕ] − ϕ ≤ C ln(N )
T b−a
∞
kϕkp,α
N
p,α
for all ϕ ∈ Cper ([a, b]) and some constant C depending on p and α.
If h = (b − a)/N denotes the mesh length then the above result can be rewritten as
P1,N [ϕ] − ϕ ≤ Chp+α ln(1/h) kϕk .

T
∞ p,α (6.1.4)
For ϕ ∈ C p,α ([a, b]), one dimensional Lagrange polynomial interpolation of degree R
with respect to uniform mesh points 4[a,b],R = {ti |ti = a + ih, 0 ≤ i ≤ R, h = (b − a)/R}
is given by
R
X
P1,R,I
P
[ϕ](t) = ϕ(ti )LPi (t), for all t ∈ [a, b], (6.1.5)
i=0
129
where I in the subscript denote the length of interval [a, b]. Here LPi (t) denotes the La-
grange basis for polynomial interpolation, given by
w(t)
LPi (t) = , (6.1.6)
(t − ti )w0 (ti )
where
w(t) = (t − t0 )(t − t1 ) · · · (t − tR ) (6.1.7)
and w0 (t) denotes the derivative of w(t) with respect to t. Note that LPi (tj ) = δij for
0 ≤ i, j ≤ R.
Theorem 6.1.3. Let p ∈ N ∪ {0} and 0 < α ≤ 1. Then for the error in one dimensional
Lagrange polynomial interpolation of ϕ ∈ C p,α ([a, b]) of degree R > p with respect to
uniform mesh points 4[a,b],R is given by
p+α
P1,R,I [ϕ] − ϕ ≤ C(1 + ΛR )
P b−a
∞
, (6.1.8)
R
where
R
X
ΛR = P1,R,I
P P
= max Li (t)
∞ a≤t≤b
i=0
is the Lebesgue constant and C > 0 is another constant depending on ϕ.
Proof. For ϕ ∈ C p,α ([a, b]), let ϕPR best approximation of ϕ with respect to the algebraic
polynomial of degree less than or equal to R in maximum norm. Then, by Jackson’s
theorem (see [70, Theorem-3.7.2]) there exist a constant C such that for R > p
p+α
ϕR − ϕ ≤ C b − a
P
∞
. (6.1.9)
R
Now, writing
P1,R,I
P
[ϕ] − ϕ = P1,R,I
P
ϕ − ϕPR − ϕ − ϕPR

and then by invoking triangle inequality, we have
P1,R,I [ϕ] − ϕ ≤ P1,R,I

P P P

ϕ − ϕPR

∞
ϕ − ϕR ∞
+ ∞
P
≤ 1 + P1,R,I ∞ ϕ − ϕPR ∞ .

130
Now by using inequality (6.1.9), we get desired estimate

p+α
P1,R,I [ϕ] − ϕ ≤ C(1 + ΛR )
P b − a
∞
.
R
Note that estimate (6.1.8) does not implies the convergence of P1,R,I
P
[ϕ] to ϕ with the
increase in degree of interpolating polynomial R. Because, as we increase the degree of in-
terpolating polynomial at uniform mesh points Lebesgue number ΛR grows exponentially.
However, use of piecewise polynomial Lagrange interpolation of fixed degree certainly guar-
anteed the convergence.
In the following theorem we recall the well known error formula for Lagrange polynomial
interpolation which can be found, for example, in [70].
Theorem 6.1.4. Let p ∈ N ∪ {0} and 0 < α ≤ 1. Then for the error in one dimensional
Lagrange polynomial interpolation of ϕ ∈ C p,α ([a, b]) of degree R < p with respect to
uniform mesh points 4[a,b],R is given by
w(t)
P1,R,I
P
[ϕ](t) − ϕ(t) = ϕR+1 (ξ), (6.1.10)
(R + 1)!
where w(t) is given by eq. (6.1.7) and ξ ∈ (a, b).
Next, we define trigpoly interpolation that is utilized for the evaluation of density values
at off grid points.
As, we have explained in Chapters 2 and 4 in order to perform interpolation accurately
we refine the original grid by some refinement factor. In rest of this section, we denote
this refinement factor by symbol F . For given t ∈ [a, b], by 4t[a,b],R , we denote set of R + 1
equidistant mesh points of 4[a,b],N F , in the neighborhood of point t (see Figure 6.1). To
be precise, for t ∈ [a, b], let
t−a
j= N
b−a
and
R
k = min max j− , 0 , NF − R (6.1.11)
2
131
then
4t[a,b],R = {ti |ti = a + ihF , k ≤ i ≤ k + R, hF = (b − a)/N F } ,
= 4[a+khF ,a+(k+R)hF ],R .
p,α
For ϕ ∈ Cper ([a, b]), p ≥ 0, we define one dimensional trigpoly interpolation, with
respect to the interpolation points 4[a,b],N , as
P1,R,N,F
TP
[ϕ](t) = P1,R,Rh
P
P1,N [ϕ] (t)
T
F ,t
(6.1.12)
for all t ∈ [a, b]. Here, P1,R,Rh

P
F ,t
[ϕ](t) is one dimensional Lagrange polynomial interpola-
tion of ϕ of fixed degree R with respect to uniform mesh points 4[a+khF ,a+(k+R)hF ],R , where
k is given by Equation (6.1.11).
Our next theorem provides an error estimate for the one dimensional trigpoly interpo-
lation defined in Equation (6.1.12).
Theorem 6.1.5. Suppose p ∈ N ∪ {0} and 0 < α ≤ 1. Then for fixed R, F ≥ 1 error in
p,α
the one dimensional trigpoly interpolation of ϕ ∈ Cper ([a, b]) is given by
 n
 b−a p+α
o
b−a R+1
max ln(N ) N

 , NF if R < p


P1,R,N,F [ϕ] − ϕ ≤ C
TP
b−a R+α (6.1.13)

∞  N F
if R = p


ln(N ) b−a p+α

if R > p

N
where C is some constant depend on ϕ, R, F .
Proof. For t ∈ [a, b], writing
P1,R,N,F
TP
[ϕ](t) − ϕ(t) = P1,R,Rh
P
P (t) − P1,R,Rh [ϕ](t) + P1,R,Rh
T P P
F ,t 1,N [ϕ] F ,t F ,t
[ϕ](t) − ϕ(t).
By invoking triangle inequality in the above equation, we have
P1,R,N,F [ϕ] − ϕ ≤ P1,R,Rh P P P

TP P T P
P
∞ F ,t 1,N [ϕ] − 1,R,RhF ,t [ϕ]
∞
+ 1,R,RhF ,t − ϕ
∞
≤ P1,R,RhF ,t ∞ P1,N [ϕ] − ϕ ∞ + P1,R,RhF ,t [ϕ] − ϕ ∞ .

P T P
(6.1.14)
132
Figure 6.1: Trigpoly interpolation at off grid points using refined data obtained by FFTs.
Top figure presents original coarse uniform grid in an interval [a, b]. Middle figure presents
refined grid where the function values is obtained by inverse FFTs. Bottom figure presents
local polynomial interpolation of fixed degree R in a neighborhood of off grid point t
(encircled and marked by star).
As, P1,R,Rh
P
F ,t
[ϕ] is the polynomial Lagrange interpolation of ϕ with respect to uniform
mesh points 4[a+khF ,a+(k+R)hF ],R , where k is given by Equation (6.1.11). Therefore, for
R < p, by employing theorem 6.1.4 in an interval [a + khF , a + (k + R)hF ], we have
R+1
P1,R,Rh ,t [ϕ] − ϕ ≤ C (Rh/F )
P
∞
. (6.1.15)
F
(R)R+1
By Theorem 6.1.2, we have

p+α
P1,N [ϕ] − ϕ ≤ C ln(N )
T b−a
∞
. (6.1.16)
N
133
Now, by using Equations (6.1.15) and (6.1.16) in Equation (6.1.14), we have

p+α R+1
P1,R,N,F [ϕ] − ϕ ≤ C P1,R,Rh
TP P b−a b−a
∞

F ,t ∞
ln(N ) +C
N NF
( p+α R+1 )
b−a b−a
≤ C max ln(N ) , .
N NF
Thus for a fixed R < p, we have

( p+α R+1 )
P1,R,N,F [ϕ] − ϕ ≤ C max ln(N )
TP b − a b − a
∞
, .
N NF
Estimate for the case R = p follows by using the fact that pth derivative of ϕ is Hölder
continuous. However, following the same argument, the estimate for R > p follows from
theorem 6.1.3 and theorem 6.1.2.
Now we provide error estimates for two dimensional interpolators. We first introduce
some notation which considerable simplify our presentation.
Let for a1 < b1 , a2 < b2 ∈ R
4[a1 ,b1 ]×[a2 ,b2 ],N = 4[a1 ,b1 ],N1 × 4[a2 ,b2 ],N2 , (6.1.17)
be an equidistant subdivision of rectangle [a1 , b1 ] × [a2 , b2 ] with an even number of grid

points in each direction. Note that total number of grid points is given by N = N1 N2 and
O(N1 ) = O(N2 ).
p,α
Suppose ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ]) , p ≥ 0 and 0 < α ≤ 1. Then for each fixed
t1 ∈ [a1 , b1 ], let P1,t
T
2 ,N2
[ϕ] be the trigonometric interpolate to ϕ(t1 , t2 ) in the t2 variable
interpolating ϕ(t1 , t2 ) with respect to uniform mesh points 4[a2 ,b2 ],N2 . Specifically,
N
X 2 −1
P1,t
T
2 ,N2
[ϕ](t1 , t2 ) = ϕ(t1 , t2,i )LT2,i (t2 ),
i=0
where LT2,i (t2 ) is given by (6.1.3).

For each fixed t2 ∈ [a2 , b2 ], let P1,t
T
1 ,N1
[ϕ] be the trigonometric interpolate to ϕ(t1 , t2 ) in
the t1 variable interpolating ϕ(t1 , t2 ) with respect to uniform mesh points 4[a1 ,b1 ],N1 . Then
N
X 1 −1
P1,t
T
1 ,N1
[ϕ](t1 , t2 ) = ϕ(t1,j , t2 )LT1,j (t1 ).
j=0
134
p,α
Now, we define two dimensional trigonometric interpolation for ϕ(t1 , t2 ) ∈ Cper ([a1 , b1 ] × [a2 , b2 ])
with respect to uniform mesh points 4[a1 ,b1 ]×[a2 ,b2 ],N as
P2,N
T
[ϕ](t1 , t2 ) = P1,t
T
P1,t2 ,N2 [ϕ](t1 , t2 ) ∀(t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ].
T
1 ,N1
(6.1.18)
It can be readily verified that

N
X1 −1 N
X 2 −1
P2,N
T
[ϕ](t1 , t2 ) = ϕ(t1,i , t2,j )LT1,i (t1 )LT2,j (t2 ) (6.1.19)
i=0 j=0
and
P2,N
T
[ϕ](t1 , t2 ) = P1,t
T
P1,t2 ,N2 [ϕ](t1 , t2 ) = P1,t P1,t1 ,N1 [ϕ](t1 , t2 ) .
T T
T
1 ,N1 2 ,N2
Let I1 = b1 − a1 and I2 = b2 − a2 , then for for ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ]) we define two
dimensional Lagrange polynomial interpolation of degree R with respect to uniform mesh
points 4[a1 ,b1 ]×[a2 ,b2 ],R2 as
P2,R,I
P
[ϕ](t1 , t2 ) = P1,t
P
P1,t2 ,R,I2 [ϕ](t1 , t2 ) ∀(t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ]. (6.1.20)
P
1 ,I2 1 ,R,I1
Here, for fixed t2 ∈ [a1 , b1 ], P1,t

P
1 ,R,I1
[ϕ](t1 , t2 ) denotes the Lagrange interpolate of degree R
to ϕ(t1 , t2 ) in t1 variable with respect to uniform mesh points 4[a1 ,b1 ],R . Similarly, for fixed
t1 ∈ [a2 , b2 ], P1,t
P
2 ,R,I2
[ϕ](t1 , t2 ) denotes the Lagrange interpolate of degree R to ϕ(t1 , t2 ) in
t2 variable with respect to uniform mesh points 4[a2 ,b2 ],R . It is not difficult to verify that
R X
X R
P2,R,I
P
1 ,I2
[ϕ](t1 , t2 ) = ϕ(t1,i , t2,j )LP1,i (t1 )LP2,j (t2 )
i=0 j=0
and
P2,N,I
P
P P
P P P
P P
P
1 ,I2
[ϕ](t1 , t2 ) = 1,t1 ,R,I1 1,t2 ,R,I2 [ϕ](t1 , t2 ) = 1,t2 ,R,I2 1,t1 ,R,I1 [ϕ](t1 , t2 ) .
Moreover, the operator norm of P2,R

P
is given by
R X
X R
P2,R,I ,I =
P P
max L1,i (t1 ) LP2,j (t2 ) .
1 2 ∞
(t1 ,t2 )∈[a1 ,b1 ]×[a2 ,b2 ]
i=0 j=0
Note that for fixed R, P2,R,I

P
≤ C, for some constant C.
,I
1 2 ∞
135
Theorem 6.1.6. Suppose ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ]), p ≥ 0 and 0 < α ≤ 1. Then for
fixed R we have
 n o
b1 −a1 p+α b2 −a2 p+α



max R
, R
if R > p

 n o
P2,R,I ,I [ϕ] − ϕ
P
≤ C max b1 −a1 R+α b2 −a2 R+α (6.1.21)

1 2 ∞  R
, R
if R = p

 n o
b1 −a1 R+1 b2 −a2 R+1

max

, if R < p
R R
where C is constant depending on degree R, F and ϕ.
Proof.
P2,R,I
P
P P
P − P1,t [ϕ] + P1,t
P P P
,I
1 2
[ϕ] − ϕ = 1,t 1 ,R,I 1 1,t2 ,R,I 2
[ϕ] 1 ,R,I1 1 ,R,I1
[ϕ] − ϕ.
By invoking triangle inequality we have
P2,R,I ,I [ϕ] − ϕ ≤ P1,t P P + P1,t

P P P P
P
1 2 ∞ 1 ,R,I 1 1,t 2 ,R,I 2
[ϕ] − 1,t 1 ,R,I 1
[ϕ] ∞ 1 ,R,I 1
[ϕ] − ϕ
∞
≤ P1,t1 ,R,I1 ∞ P1,t2 ,R,I2 [ϕ] − ϕ ∞ + P1,t1 ,R,I1 [ϕ] − ϕ ∞ .

P P P
By theorem (6.1.3)
P2,R,I ,I [ϕ] − ϕ ≤
P
1 2 ∞
p+α p+α
P1,t ,R,I P1,t b 2 − a2 b 1 − a1
+ P1,t
P P P
1 ∞

2 ,R,I2 ∞

1 ,R,I1 ∞
.
1
R R
For fixed degree R > p,

( p+α p+α )
P2,R,I ,I [ϕ] − ϕ ≤ C max
P b 1 − a1 b 2 − a2
∞
, .
1 2
R R
The estimate for R ≤ p follows from the theorem 6.1.4.
p,α
Theorem 6.1.7. Suppose ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ]), p ≥ 0 and 0 < α ≤ 1. Then
( p+α p+α )
P2,N [ϕ] − ϕ ≤ C ln(N1 ) max
T b 2 − a2 b 1 − a1
∞
ln N2 , ln N1 , (6.1.22)
N2 N1
where C is constant depending on ϕ.
136
Proof. Writing
P2,N
T
[ϕ] − ϕ = P1,t
T
P − P1,t [ϕ] + P1,t
T T T
1 ,N1 1,t2 ,N2 [ϕ] 1 ,N1 1 ,N1
[ϕ] − ϕ.
P2,N [ϕ] − ϕ ≤ P1,t P P + P1,t

T T T T
T
∞ 1 ,N 1 1,t2 ,N 2
[ϕ] − 1,t 1 ,N1
[ϕ] ∞ 1 ,N1
[ϕ] − ϕ
∞
≤ P1,t1 ,N1 ∞ P1,t2 ,N2 [ϕ] − ϕ ∞ + P1,t1 ,N1 [ϕ] − ϕ ∞ .

T T T
(6.1.23)
By theorem (6.1.1)
P1,t ,N = O(ln N1 )
T
1 1 ∞
(6.1.24)
and by theorem 6.1.2

p+α
P1,t ,N [ϕ] − ϕ ≤ C ln(N2 )
T b2 − a2
∞
(6.1.25)
2 2
N2
and p+α
P1,t ,N [ϕ] − ϕ C ln(N1 )
T b1 − a1
∞
. (6.1.26)
1 1
N1
Now, by using eqs. (6.1.24) to (6.1.26) in eq. (6.1.23), we have desired estimate
( p+α p+α )
P2,N [ϕ] − ϕ ≤ C ln(N1 ) max
T b 2 − a 2 b 1 − a 1
∞
ln N2 , ln N1 .
N2 N1
In the following paragraph, by 4t[a1 ,b1 ]×[a2 ,b2 ],R , we denote (R + 1) × (R + 1) rectangular
uniform mesh points of 4[a1 ,b1 ]×[a2 ,b2 ],N in the neighborhood of point t = (t1 , t2 ) and by
P2,R,Rh
P
1 /F,Rh2 /F,t
[ϕ], we denote two dimensional Lagrange polynomial interpolation with
respect to uniform mesh points 4t[a1 ,b1 ]×[a2 ,b2 ],R , as defined in (6.1.20).
p,α
Suppose ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ]) , p ≥ 0, 0 < α ≤ 1 and refinement factor F ≥ 1.
Then, we define two dimensional trigpoly interpolation with respect to uniform mesh points
4[a1 ,b1 ]×[a2 ,b2 ],N as
P2,R,N,F
TP
[ϕ](t) = P2,R,Rh
P
1 /F,Rh2 /F,t
P2,N
T
[ϕ](t) (6.1.27)
for all t = (t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ]. Note that degree R of the two dimensional Lagrange
polynomial interpolation is fixed.
137
Following theorem provide error estimate for two dimensional trigpoly interpolation
defined in (6.1.27).
Theorem 6.1.8. Suppose ϕ ∈ C p,α and p ≥ 0. Then for fixed R and F ≥ 1, we have
 p+α p+α
b2 −a2 b1 −a1
ln(N1 ) max ln N2 , N1 ln N1 if R > p



 N2


 R+α R+α
P2,N,F [ϕ] − ϕ
TP b2 −a2 b1 −a1
∞
≤ C max N2 F
, N1 F
if R = p



 R+1 R+1
 b2 −a2 b1 −a1
max

N2 F
, N1 F if R < p
(6.1.28)
where C is constant depend on R, F and ϕ.
Proof. Writing
P2,N,F
TP
[ϕ] − ϕ
= P2,R,Rh
P
P − P2,R,Rh [ϕ] + P2,R,Rh
T P P
1 /F,Rh2 /F,t 2,N [ϕ] 1 /F,Rh2 /F,t 1 /F,Rh2 /F,t
[ϕ] − ϕ.
P2,N [ϕ] − ϕ ≤ P2,R,Rh P P

TP P T P

∞ 1 /F,Rh2 /F,t 2,N [ϕ] − 2,R,Rh 1 /F,Rh2 /F,t [ϕ] +
∞
P2,R,Rh /F,Rh /F,t [ϕ] − ϕ

P
1 2 ∞
(6.1.29)
≤ P2,R,Rh P2,N [ϕ] − ϕ + P2,R,Rh

P T P
1 /F,Rh2 /F,t ∞ ∞ 1 /F,Rh2 /F,t [ϕ] − ϕ .
∞
(6.1.30)

Since, for fixed R, P2,R,Rh1 /F,Rh2 /F,t
P
is bounded and by Theorem 6.1.6 for R > p

∞
( p+α p+α )
P2,R,Rh /F,Rh /F,t [ϕ] − ϕ ≤ C max
P b 2 − a2 b1 − a1
∞
, (6.1.31)
1 2
N2 N1
and by Theorem 6.1.7

( p+α p+α )
P2,N,F [ϕ] − ϕ ≤ C ln(N1 ) max
T b 2 − a2 b 1 − a1
∞
ln N2 , ln N1 .
N2 N1
(6.1.32)
138
Now, by using eqs. (6.1.30) to (6.1.31) in the inequality (6.1.30) we get the desired estimate
for R > p
( p+α p+α )
P2,N,F [ϕ] − ϕ ≤ C ln(N1 ) max
TP b 2 − a2 b 1 − a1
∞
ln N2 , ln N1 .
N2 N1
Estimate for the case R ≤ p is identical to the R > p, and is omitted.
Thus, the above result shows that if O(N1 ) = O(N2 ) then for sufficiently smooth
function ϕ, P2,N,F
TP
[ϕ] − ϕ ∞ → 0 as N → ∞ with high-order.
In the following paragraphs we provide an error estimate for three dimensional inter-
polators. Again, to simplify our presentation we first introduce following notations.
For a1 < b1 , a2 < b2 , a3 < b3 ∈ R, let
4[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ],N = 4[a1 ,b1 ],N1 × 4[a2 ,b2 ],N2 × 4[a3 ,b3 ],N3 , (6.1.33)
be an equidistant subdivision of set [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ] with an even number of grid
points in each direction. Note that the total number of grid points is given by N = N1 N2 N3
and O(N1 ) = O(N2 ) = O(N3 ).
Similar to the two dimensional approach, let for each fixed (t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ],
P1,t
T
1 ,N1
[ϕ] denotes the trigonometric interpolate of ϕ(t1 , t2 , t3 ) in t1 variable with respect to
uniform mesh points 4[a1 ,b1 ],N1 . In a similar fashion, for each fixed (t1 , t3 ) ∈ [a1 , b1 ]×[a3 , b3 ],
P1,t
T
2 ,N1
[ϕ] denotes the trigonometric interpolate of ϕ(t1 , t2 , t3 ) in t2 variable with respect
to uniform mesh points 4[a2 ,b2 ],N2 and for each fixed (t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ], P1,t
T
3 ,N3
[ϕ]
denotes the trigonometric interpolate of ϕ(t1 , t2 , t3 ) in t3 variable with respect to uniform
mesh points 4[a3 ,b3 ],N3 .
p,α
We define three dimensional trigonometric interpolation of ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ] ×
[a3 , b3 ]), p ≥ 0 with respect to uniform mesh points 4[a1 ,b1 ]×[a2 ,b2 ]×[a2 ,b2 ],N , as
P3,N
T
[ϕ](t1 , t2 , t3 ) = P1,t
T
P1,t2 ,N2 P1,t
T T
3 ,N3 1 ,N1
[ϕ](t1 , t2 , t3 ) (6.1.34)
for all (t1 , t2 , t3 ) ∈ [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. Moreover, it can be readily verified that
N
X 1 −1 N
X2 −1 N
X3 −1
P3,N
T
[ϕ](t1 , t2 , t3 ) = ϕ(t1,i , t2,j , t3,k )LT1,i (t1 )LT2,j (t2 )LT3,k (t3 ),
i=0 j=0 k=0
139
where LT1,i , LT2,j , LT3,k are Lagrange basis for trigonometric interpolation of the form (6.1.3).
Let I1 = b1 −a1 , I2 = b2 −a2 , I3 = b3 −a3 . Then for ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]),
we define three dimensional Lagrange polynomial interpolation of degree R with respect
to uniform mesh points 4[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ],R3 as
P3,R,I
P
[ϕ](t1 , t2 , t3 ) = P1,t
P
P1,t2 ,R,I2 P1,t
P P
1 ,I2 ,I3 3 ,R,I3 1 ,R,I1
[ϕ](t1 , t2 , t3 ) (6.1.35)
for all (t1 , t2 , t3 ) ∈ [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. Here, for fixed (t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ],
P1,t
P
1 ,R,I1
[ϕ](t1 , t2 , t3 ) denotes the Lagrange polynomial interpolate of degree R to ϕ(t1 , t2 , t3 )
in t1 variable with respect to uniform mesh points 4[a1 ,b1 ],R . Similarly, for each fixed
(t1 , t3 ) ∈ [a1 , b1 ] × [a3 , b3 ], P1,t
P
2 ,R,I2
[ϕ] denotes the Lagrange polynomial interpolate of
ϕ(t1 , t2 , t3 ) in t2 variable with respect to uniform mesh points 4[a2 ,b2 ],R and for each fixed
(t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ], P1,t
P
3 ,R,I3
[ϕ] denotes the Lagrange polynomial interpolate of
ϕ(t1 , t2 , t3 ) in t3 variable with respect to uniform mesh points 4[a3 ,b3 ],R .
It is not difficult to verify that
R X
X R X
R
P3,R,,I
P
1 ,I2 ,I3
[ϕ](t1 , t2 , t3 ) = ϕ(t1,i , t2,j , t3,k )LP1,i (t1 )LP2,j (t2 )LP3,k (t3 ).
i=0 j=0 k=0
The operator norm of P3,R,,I

P
1 ,I2 ,I3
is given by
R X
X R X
R
P3,R,,I ,I ,I =
P P
max L1,i (t1 ) LP2,j (t2 ) LP3,k (t3 ) .
1 2 3 ∞
(t1 ,t2 ,t3 )∈[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ]
i=0 j=0 k=0
Note that for fixed R, P3,R,,I

P
≤ C, for some constant C.
1 ,I2 ,I3 ∞
By, 4t[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ],R , we denote (R + 1) × (R + 1) × (R + 1) cubical uniform
mesh points of 4[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ],N in the neighborhood of point t = (t1 , t2 , t3 ) and by
P3,R,,I
P
1 ,I2 ,I3 ,t
[ϕ], we denote three dimensional Lagrange polynomial interpolation with re-
spect to uniform mesh points 4t[a1 ,b1 ]×[a2 ,b2 ]×[a3 ,b3 ],R .
p,α
Suppose ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]) , p ≥ 0 and 0 < α ≤ 1. Then, we define
three dimensional trigpoly interpolation with respect to uniform mesh points 4[a1 ,b1 ]×[a2 ,b2 ]×[a2 ,b2 ],N
as
P3,R,N,F
TP
[ϕ](t) = P3,R,,I
P
1 ,I2 ,I3 ,t
P3,N
T
[ϕ](t) (6.1.36)
140
for all t = (t1 , t2 , t3 ) ∈ [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. Note that degree R of three dimensional
Lagrange polynomial interpolation is fixed.
Now, In the Theorems 6.1.9 to 6.1.11 we provide error estimate on the above dis-
cussed three kind of interpolations whose proofs are almost identical to those that of
Theorems 6.1.6 to 6.1.8 respectively, and thus stated below without proof.
p,α
Theorem 6.1.9. Suppose ϕ ∈ Cper ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]), p ≥ 0 and 0 < α ≤ 1.
Then
( p+α p+α p+α )
P3,N [ϕ] − ϕ ≤ C ln(N1 ) ln(N2 ) ln(N3 )
T b 1 − a1 b 2 − a2 b3 − a3
∞
, , ,
N1 N2 N3
(6.1.37)
where C is constant depending on ϕ.
Theorem 6.1.10. Suppose ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]), p ≥ 0 and 0 < α ≤ 1.
Then for fixed polynomial degree R, we have
P3,R,I ,I ,I [ϕ] − ϕ
P
1 2 3 ∞
 p+α p+α p+α
b1 −a1 b2 −a2 b3 −a3
max , N2 , N3 if R > p



 N1


 R+α R+α R+α
b1 −a1 b2 −a2 b3 −a3
≤ C max N1
, N2 , N3 if R = p (6.1.38)



 R+1 R+1 R+1
 b1 −a1 b2 −a2 b3 −a3
max

N1
, N2 , N3 if R < p
where C is constant depending on degree R, F and function ϕ.
Theorem 6.1.11. Suppose ϕ ∈ C p,α , and p ≥ 0. Then for fixed polynomial degree R and
F ≥ 1, we have
P3,N,F [ϕ] − ϕ
TP
∞
 p+α p+α p+α
b1 −a1 b2 −a2 b3 −a3
ln(N1 ) ln(N2 ) ln(N3 ) max , N2 , N3 if R > p



 N1


 R+α R+α R+α
b1 −a1 b2 −a2 b3 −a3
≤ C max N1 F
, N2 F
, N3 F
if R = p



 R+1 R+1 R+1
 b1 −a1 b2 −a2 b3 −a3
max

N1 F
, N2 F , N3 F if R < p
(6.1.39)
141
where the constant C depend on R and ϕ.
All the error estimates proved in this section are under the assumption that ϕ ∈ C p,α
for 0 ≤ p and 0 < α ≤ 1. In the case where ϕ ∈ C ∞ , by using Jackson theorems for
C ∞ functions and following the same arguments, it is straightforward to show that error
incurred by these interpolators converges to zero faster than any power of N .
6.2 Quadrature Analysis

In this section we provide an error estimate of the proposed high-order quadrature scheme
for the approximation of volume integral operator of Lippmann-Schwinger integral equa-
tion (1.2.38) in the two and three dimensions. As elucidated in chapter 2 and chapter 4,
proposed high-order quadrature scheme comprises of three basic interpolatory quadrature
rules, namely, trapezoidal rule, Q-points composite Newton-Cotes and Clenshaw-Curtis
quadrature rule. First two quadrature rules have been used in the two dimensional al-
gorithm while all the three have been used in the three dimensional setting. Therefore,
before going into the details of quadrature errors in the proposed integration scheme, we
first provide an error estimate for all three kinds of quadrature that will be used in further
discussion of error estimates of the proposed quadrature scheme.
Theorem 6.2.1. Let L[a,b],N

N C,Q
[ϕ] denote the closed Q-point composite Newton-Cotes quadra-
ture for a function ϕ, tabulated at (N + 1) uniformly distributed points of mesh length
h = (b − a)/N in an interval [a, b]. Then for ϕ ∈ C p,α ([a, b]), 0 < α ≤ 1 and Q < p we
have

ChQ+α+1 if Q is odd
Z b 
ϕ(t) dt − L[a,b],N [ϕ] ≤
N C,Q

(6.2.1)
a ChQ+α if Q is even,

where C is a constant depending on Q and ϕ.
The following Lemma provides an error estimate for the approximation to the integral
of a function having corner type of discontinuity within region of integration by Q-point
composite Newton-Cotes quadrature rule.
142
6.2. Quadrature Analysis
Lemma 6.2.2. Suppose ϕ ∈ C p,α ([a, t0 ) ∪ (t0 , b]), 0 < α ≤ 1, Q < p such that ϕ(t) has
corner type singularity at t = t0 and
L1
X
L˜[a,b],N
N C,Q
[ϕ] = L[tN0 −l
C,Q
1 (Q−1)h,t0 −(l1 −1)(Q−1)h],N
[ϕ]
l1 =1
I1
X
+ L[tN0 −L
C,Q
1 (Q−1)h−i1 h,t0 −L1 (Q−1)h−(i1 −1)h],N
[ϕ]
i1 =1
L2
X
+ L[tN0 +(l
C,Q
2 +1)(Q−1)h,t0 +l2 (Q−1)h],N
[ϕ]
l2 =1
I2
X
+ L[tN0 +L
C,Q
2 (Q−1)h+(i2 −1)h,t0 +L2 (Q−1)h+i2 h],N
[ϕ]
i2 =1
+ L[a,t
N C,Q
0 −L1 (Q−1)h−I1 h],N
[ϕ]
+ L[tN0 +L
C,Q
2 (Q−1)h+I2 h,b],N
[ϕ],
where h = (b − a)/N denotes the mesh size and numbers L1 , L2 , I1 , I2 are obtained as

t0 − a b − t0
L1 = , L2 = ,
(Q − 1)h (Q − 1)h

t0 − a − L1 (Q − 1)h (b − t0 ) − L2 (Q − 1)h
I1 = , I2 = ,
h h
with btc denoting the largest integer less than or equal to the real number t. Then

ChQ+α+1 if Q is odd
Z b 
LÑ C,Q [ϕ] ≤

ϕ(t) dt − [a,b],N 
a ChQ+α if Q is even,
where C is a constant depending on Q and ϕ.
Rb
Proof. We break the integral a
ϕ(t) dt as
Zb L1 Z
X t0 −(l1 −1)(Q−1)h I1 Z
X t0 −L1 (Q−1)h−(i1 −1)h
ϕ(t) dt = ϕ(t) dt + ϕ(t) dt
l1 =1 t0 −l1 (Q−1)h i1 =1 t0 −L1 (Q−1)h−i1 h
a
L2 Z t0 +l2 (Q−1)h
X I2 Z t0 +L2 (Q−1)h+i2 h
X
+ ϕ(t) dt + ϕ(t) dt
l2 =1 t0 +(l2 +1)(Q−1)h i2 =1 t0 +L2 (Q−1)h+(i2 −1)h
143
Z t0 −L1 (Q−1)h−I1 h Z b
+ ϕ(t) dt + ϕ(t) dt. (6.2.2)
a t0 +L2 (Q−1)h+I2 h
Now application of theorem (6.2.1) under the each integral of the above summation provide
the desired result.
The following theorem provides an error estimate for the approximation to the integral
of a smooth periodic function by composite trapezoidal rule quadrature.

PT
[ϕ] denotes the closed composite trapezoidal rule quadrature
p,α
for a smooth periodic function ϕ ∈ Cper ([a, b]), 0 < α ≤ 1, tabulated at (N + 1) uniformly
distributed points of mesh length h = (b − a)/N in an interval [a, b]. Then for p ≥ 0 we
have
Z b
ϕ(t) dt − L[a,b],N [ϕ] ≤ Chp+α
P T

a
where C is a constant depending on ϕ.
Our next theorem provides an error estimate for the approximation to the integral of
a smooth function by Clenshaw-Curtis quadrature rule. Proof of the following result can
be found in [43].

CC
[ϕ] denotes the Clenshaw-Curtis quadrature rule for a smooth
function ϕ ∈ C p,α ([a, b]), 0 < α ≤ 1, tabulated at
1 1
tj = (b − a)µj + (a + b), j = 0, 1, · · · , N,
2 2
where

N −j
µj = cos π, , j = 0, 1, · · · , N,
N
contained in an interval [a, b]. Then for p ≥ 0 we have
Z b
1
ϕ(t) dt − L[a,b],N [ϕ] ≤ Ch
CC
p+α
ln ,

a h
where h = max{hj = tj+1 − tj : j = 0, 1, · · · , N − 1.}.
144
6.2.1. Quadrature Analysis In Two Dimensions
6.2.1 Quadrature Analysis In Two Dimensions

In this subsection we provide an error estimate for the approximation of integral operator
(2.0.1) in the two dimensions by the integration scheme discussed in chapter 2. More
precisely, we will prove that, for given smooth refractive index n ∈ C p,α (Ω), 0 < α ≤ 1,
Q,R,F
and hence for smooth u (total field), as N → ∞ our approximate operator KN [u],
defined in eq. (6.2.8), converges to continuous operator K[u] with high-order accuracy.
Note that total number of unknown N is given by N = N1 N2 and O(N1 ) = O(N2 ).
As we have discussed in chapter 2, in the two dimensions integral operator (2.0.1) can
be re-expressed as
X X
B B
K[u](x) = Kk,sing [u](x) + Kk,reg [u](x) (6.2.3)
k∈MB (x) k6∈MB (x)
X X
I I
+ Kk,sing [u](x) + Kk,reg [u](x),
k∈MI (x) k6∈MI (x)
where
ZZ
B
Kk,reg [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 ) dt1 dt2 , (6.2.4)
[0,1]2
ZZ
B
Kk,sing [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 )) dt1 dt2 (6.2.5)
[0,1]2
−1 (r)
%Z
Z1
+ G̃κ,B (x, ξk (τ, t2 ))ϕ̃xk,B [u](τ, t2 )ηBx (τ, t2 ) dτ dt2 ,
0 %−1 (−r)
ZZ
I
Kk,reg [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 ) dt1 dt2 , (6.2.6)
[0,1]2
ZZ
I
Kk,sing [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 )) dt1 dt2 (6.2.7)
[0,1]2
−1 (r)
Z2π %Z
+ G̃κ,I (x, ξk (τ, θ))ϕ̃xk,I [u](τ, θ)ηIx (τ, θ) dτ dθ
0 0
and
ϕ̃k [u](t1 , t2 ) = ϕk [u](t1 , t2 )ξk0 (t1 , t2 ),
145
G̃κ,B (x, ξk (τ, t2 )) = %0 (τ )Gκ (x, ξk (tx1 + %(τ ), t2 )),
ϕ̃xk,B [u](%(τ ), t2 ) = ϕk [u](tx1 + %(τ ), t2 )ξk0 (tx1 + %(τ ), t2 ),
ηBx (τ, t2 ) = η((tx1 + %(τ ), t2 ); (tx1 , tx2 )),
G̃κ,I (x, ξk (τ, θ)) = %(τ )%0 (τ )Gκ (x, ξk (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ)),
ϕ̃xk,I [u](%(τ ), t2 ) = ϕk [u](tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ)ξk0 (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ),
ηIx (τ, θ) = η((tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ); (tx1 , tx2 )).
Note that all of the above integrals have a smooth integrand in its region of integration.
Now, we define approximate operator for the high-order approximation of these integrals
of smooth functions as follows:
Q,R,F
X B,Q,R,F
X B,Q
KN [u](x) = Kk,N,sing [u](x) + Kk,N,reg [u](x) (6.2.8)
X I,R,F
X
I
+ Kk,N,sing [u](x) + Kk,N,reg [u](x),
where
B,Q
[u](x) = L[0,1],N
N C,Q
L[0,1],N1 ,t1 [Gκ (x, ξk (., .))ϕ̃k [u]] ,
PT
Kk,N,reg 2 ,t2
(6.2.9)
B,Q,R,F
[u](x) = L[0,1],N
N C,Q
L[0,1],N1 ,t1 [Gκ (x, ξk (., .))ϕ̃k [u](1 − η)]
PT
Kk,N,sing 2 ,t2
h h x ii
+ L˜[0,1],N
N C,Q
L PT
P
TP x
2 ,t2 [%−1 (−r),%−1 (r)],N δ ,τ G̃κ,B (x, ξ k (., .)) 1,R,N,F ϕ̃k,B [u] ηB ,
1
(6.2.10)
I
[u](x) = L[0,1],N
PT
L
PT
Kk,N,reg 2 ,t2 [0,1],N1 ,t1 [Gκ (x, ξ k (., .))ϕ̃k [u]] , (6.2.11)
I,R,F
[u](x) = L[0,1],N
PT
L[0,1],N1 ,t1 [Gκ (x, ξk (., .))ϕ̃k [u](1 − η)]
PT
Kk,N,sing 2 ,t2
h h ii
+ L[0,2π],N
PT
δ ,θ L[0,%−1 (r)],N δ ,τ
PT
G̃κ,I (x, ξk (., .)) P2,R,N,F
TP x
ϕ̃k,I [u] ηIx .
2 1
(6.2.12)
Subscript N1δ , N2δ in eqs. (6.2.10) and (6.2.12) denote the number of points used for accurate
evaluation of singular integrals, where exponent δ ∈ (0, 1/2].
As we have explained in Chapter 2, in the two dimensional setting, non-singular bulky
integral has been computed with complexity O(N log N ) by choosing acceleration param-
eter L = O(N 1/2 ). Acceleration strategy demands that region of singular integration must
146
be contained within the adjacent set N (x). This can be easily ensured by restricting the
support of floating partitions of unity within N (x). This implies radius of η should be less
than or equal to cell size H = O(h1/2 ). Thus, radius of η should be chosen as O(hδ ), where
δ ∈ (0, 1/2].
Remark 6.2.5. 1. In the above equations additional subscript in the quadrature and
interpolation operators denotes the application of the operators in that variable while
keeping other are fixed. For example L[0,1],N
N C,Q
2 ,t2
[ϕ](t1 , t2 ) denotes the application of
Q-points composite Newton-Cotes quadrature in t2 variable while keeping t1 fixed.
2. In the interpolation operators P1,R,N,F

TP
, P2,R,N,F
TP
, P3,R,N,F
TP
, subscript F denotes the
level of refinement in the original grids.
3. As we have explained in the Chapter 2, region of integration in the singular integration

is significantly small compare to the regular integration. Therefore, we can evaluate
singular integral accurately by using O(N δ ) points instead of N , where δ ∈ (0, 1/2].
The value of δ is decided by the choice of radius of η which can be chosen to control
the complexity and accuracy of the proposed quadrature.
In the following we prove couple of results which serve as a basis in deriving error
estimate in the approximation of continuous operator (6.2.3) by the operators defined in
eqs. (6.2.9) to (6.2.12).
Lemma 6.2.6. Let

h h ii
Lk,N
B
[u] = L˜[0,1],N
N C,Q
2 ,t2
L PT
δ
[%−1 (−r),%−1 (r)],N ,τ G̃κ,B (x, .) ϕ̃x
k,B [u]ηB
x
,
1
and
h h x ii
Lk,N,Interp
B
[u] = L˜[0,1],N
N C,Q
L PT
P
TP x
2 ,t2
−1 −1 δ
[% (−r),% (r)],N ,τ G̃κ,B (x, .) 1,R,N,F ϕ̃k,B [u] ηB .
1
where N = N1 N2 and O(N1 ) = O(N2 ). Then for refractive index n ∈ C p,α (Ω) 0 < α ≤ 1
and F ≥ 1, R < p we have
h R+1
Lk,N [u] − Lk,N,Interp [u] ≤ C L˜
[0,1],N2 ,t2 L[%−1 (−r),%−1 (r)],N1δ ,τ
N C,Q
B B
P T
.

∞ ∞ ∞ F
147

where h = 1/N1 . Moreover, since L[%P−1 T
≤ (%−1 (r) − %−1 (−r)) and for

(−r),%−1 (r)],N1δ ,τ ∞

fixed Q, L˜[0,1],N
N C,Q
2 ,t2
is bounded, therefore,
∞
R+1
Lk,N [u] − Lk,N,Interp
B B
h
[u] ∞ ≤ C .
F
Proof. First note that

sup G̃ (x, y) ≤ C,

κ,B
(x,y)∈Ω×Ω
and |ηBx (.)| ≤ 1. Then it is easy to see that

B B

[u] ≤

Ñ C,Q P T
C L[0,1],N2 ,t2 L[%−1 (−r),%−1 (r)],N δ ,τ ϕ̃xk,B [u] − P1,R,N,F
TP
x
ϕ̃k,B [u] . (6.2.13)

∞ 1 ∞
By theorem (6.1.5)
R+1
ϕ̃k,B [u] − P1,R,N,F
x TP
x h
ϕ̃k,B [u] ∞ ≤ C .
F
Using this inequality in eq. (6.2.13), we can conclude the desired result
h R+1
Lk,N [u] − Lk,N,Interp [u] ≤ C L˜
[0,1],N2 ,t2 L[%−1 (−r),%−1 (r)],N1δ ,τ
N C,Q
B B
P T
.

∞ ∞ ∞ F
Lemma 6.2.7. Let

h h ii
Lk,N
I
[u] = L[0,2π],N
PT
δ ,θ L PT
[0,%−1 (r)],N δ ,τ G̃κ,I (x, .) ϕ̃x
k,I [u]η x
I
2 1
and
h h x ii
Lk,N,Interp
I
[u] = L[0,2π],N
PT
L PT
P
TP x
δ ,θ [0,%−1 (r)],N δ ,τ G̃κ,I (x, .) ϕ̃
2,R,N,F k,I [u] ηI ,
2 1
where N = N1 N2 and O(N1 ) = O(N2 ). Then for refractive index n ∈ C p,α (Ω) 0 < α ≤ 1
and F ≥ 1, R < p we have

I I

[u] ∞
148
( R+1 R+1 )
1 1
≤ C L[0,2π],N L[0,%−1 (r)],N1δ ,τ max
PT PT
, .

δ
2 ,θ ∞ ∞ N1 F N2 F

Moreover, since L[0,2π],N L = |%−1 (r)| ≤ C, therefore,
PT PT
= 2π and

δ ,θ [0,%−1 (r)],N δ ,τ
2 ∞ 1 ∞
( R+1 R+1 )
I I
1 1
[u] ∞ ≤ C max , .
N1 F N2 F
Proof. Proof of this lemma is similar to that of lemma(6.2.6) and desired estimate can be
obtained by noting

sup G̃κ,I (x, y) ≤ C,

(x,y)∈Ω×Ω
|ηIx (.)| ≤ 1 and by invoking theorem (6.1.8).

0 0 0
Theorem 6.2.8. Let ϕ ∈ C p,α [a, b] × [c, t2 ) ∪ [a, b] × (t2 , d] , c ≤ t2 ≤ d, and 0 < α ≤ 1,
Q < p such that
0
1. ϕ(t1 , t2 ) has a corner type singularity at t2 = t2 .
∂ l ϕ(t1 ,t2 ) ∂ l ϕ(.,t2 )

2. For all l ≤ p, ∂tl1
is uniformly bounded and ∂tl1
is periodic on [a, b] for all
t2 ∈ [c, d]. Then
d b  n o
Z Z C max hp+α
1 , hQ+1+α
2 if Q is odd,
LÑ C,Q L P T

ϕ(t) dt − [c,d],N2 ,t2 [a,b],N ,t
1 1
[ϕ]≤
 n o
C max hp+α , hQ+α if Q is even,
c a 1 2
(6.2.14)
where C is constant depending on ϕ and h1 = (b−a)/N1 , h2 = (d−c)/N2 , t = (t1 , t2 ).
Proof. Let
Zd Zb
EN [ϕ] = ϕ(t1 , t2 ) dt1 dt2 − L˜[c,d],N
N C,Q
2 ,t2
L[a,b],N
PT
1 ,t1
[ϕ]
c a
 
Zd Zb
=  ϕ(t1 , .) dt1 − L[a,b],N
PT
1 ,t1
[ϕ] dt2
c a
Zd
+ L[a,b],N
PT
1 ,t1
[ϕ] dt2 − L˜[c,d],N
N C,Q
2 ,t2
L[a,b],N
PT
1 ,t1
[ϕ].
c
149
Since in the above, integral and operator L[a,b],N

PT
1 ,t1
commutes and
L˜[c,d],N
N C,Q
2 ,t2
L[a,b],N
PT
1 ,t1
[ϕ] = L[a,b],N
PT
1 ,t1
L˜[c,d],N
N C,Q
2 ,t2
[ϕ].
Therefore

Zd Zb
|EN [ϕ]| ≤ ϕ(t1 , .) dt1 − L[a,b],N1 ,t1 [ϕ] dt2
P T

c a
d
Z
+ L[a,b],N1 ,t1 ∞ ϕ(., t2 ) dt2 − L˜[c,d],N2 ,t2 [ϕ] .
N C,Q
PT
(6.2.15)

c
As, ϕ(., t2 ) is periodic in [a, b] for all t2 ∈ [c, d] and its derivative up to order p with respect
to t1 is uniformly bounded. Therefore, by lemma (6.2.2) there exist a constant C depending
on ϕ such that
d 
Q+1+α
Ch2 if Q is odd
Z 
ϕ(., t2 ) dt2 − L˜ N C,Q

[c,d],N2 ,t2 [ϕ] ≤ (6.2.16)

 ChQ+α if Q is even.
c 2
By Theorem 6.2.3, we have

b
Z
ϕ(t1 , .) dt1 − L[a,b],N [ϕ] ≤ Ch1p+α .
PT

1 ,t1
(6.2.17)

a
Now by using eqs. (6.2.16) and (6.2.17) in eq. (6.2.15), we can conclude the desired estimate

(6.2.14) by noting that L[a,b],N
PT
= b − a.

1 ,t1
∞
Theorem 6.2.9. Let ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ]), 0 < α ≤ 1, 0 ≤ p and for all l = (l1 , l2 )
∂ |l| ϕ(t) ∂ |l| ϕ(t)
with |l| = |l1 | + |l2 | ≤ p, ∂tl
is uniformly bounded and ∂tl
is periodic for all t =
(t1 , t2 ) ∈ [a1 , b1 ] × [a2 , b2 ]. Then
b b
Z 2 Z 1
ϕ(t) dt − L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] ≤ C max hp+α , hp+α
PT PT

1 2 , (6.2.18)

a2 a1
where C is constant depending on ϕ and h1 = (b1 − a1 )/N1 , h2 = (b2 − a2 )/N2 .
150
Proof. Let
Zb2 Zb1
EN [ϕ] = ϕ(t1 , t2 ) dt1 dt2 − L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ]
a2 a1
 
Zb2 Zb1
=  ϕ(t1 , .) dt1 − L[aP T,b 1 1 ],N1 ,t1
[ϕ] dt2
a2 a1
Zb2
+ L[aP1T,b1 ],N1 ,t1 [ϕ] dt2 − L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ].
a2

Zb2 Zb1
|EN [ϕ]| ≤ ϕ(t1 , .) dt1 − L[a1 ,b1 ],N1 ,t1 [ϕ] dt2
PT

a2 a1
b
Z 2
+ L[a1 ,b1 ],N1 ,t1 [ϕ] dt2 − L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] .
P T P T P T

a2
Note that
L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ] = L[aP1T,b1 ],N1 ,t1 L[aP2T,b2 ],N2 ,t2 [ϕ].
Now, by theorem (6.2.3) we have

|EN [ϕ]| ≤ C h1p+α + L[aP1T,b1 ],N1 ,t1 ∞ h2p+α .

We can conclude the desired estimate (6.2.18) by noting that L[a1 ,b1 ],N1 ,t1
PT
= b1 −a1 .

∞
In the following theorems we provide error estimate for the approximation of integral
operators given by Equations (6.2.4) to (6.2.7) by means of approximate operators defined
in Equations (6.2.9) to (6.2.12). In rest of this section, we denote h1 = 1/N1 , h2 = 1/N2
and N = N1 N2 .
Theorem 6.2.10. If refractive index n ∈ C p,α (Ω), p > Q and 0 < α ≤ 1. Then there exist
a constant C such that
 n o
 C max h1p+α , hQ+1+α
2 if Q is odd,
B B,Q
sup Kk,reg [u](x) − Kk,N,reg [u](x) ≤ (6.2.19)

n o
x∈Ω\Pk C max hp+α , hQ+α

if Q is even,.
1 2
151
Proof. For each fix x ∈ Ω \ Pk and fix k ∈ IB define
ϕx [u] (ξk (t1 , t1 )) = Gκ (x, ξk (t1 , t2 ))ϕk [u](t1 , t2 )ξk0 (t1 , t2 )
for all (t1 , t2 ) ∈ [0, 1]2 . Since x ∈

/ Pk , therefore Gκ (x, ξk (t1 , t2 )) is smooth in whole integra-
tion region. Because n ∈ C p,α (Ω), therefore ϕk ∈ C p,α ([0, 1]2 ). Moreover, for all l = (l1 , l2 )
∂ |l| ϕk (t) ∂ |l| ϕk (.,t2 )
with |l| = |l1 | + |l2 | ≤ p, ∂tl
is periodic in [0, 1]
for all t2 ∈ [0, 1]. Thus, derivative of function ϕx up to order p is uniformly bounded
∂ |l| ϕx (.,t2 )
and ∂tl
are vanishes at the endpoints of integration interval along t1 variable for all
t2 ∈ [0, 1]. Now the desired estimate (6.2.19) can be obtained by invoking Theorem 6.2.8
for function ϕx .
Theorem 6.2.11. Suppose refractive index n ∈ C p,α (Ω), 0 ≤ p and 0 < α ≤ 1. Then
there exist a constant C such that
[u](x) ≤ C max hp+α , hp+α

I I

sup Kk,reg [u](x) − Kk,N,reg 1 2 . (6.2.20)
x∈Ω\Pk
Proof. Similar to the proof of the theorem (6.2.10), for each fix x ∈ Ω \ Pk and fix k ∈ II
define
ϕx [u] (ξk (t1 , t1 )) = Gκ (x, ξk (t1 , t2 ))ϕk [u](t1 , t2 )ξk0 (t1 , t2 )
for all (t1 , t2 ) ∈ [0, 1]2 . Since, Gκ (x, ξk (t1 , t2 )) is smooth in Pk and ϕk [u] is smooth and
periodic, therefore, ϕx [u] is smooth and periodic function of t1 and t2 in [0, 1]2 . Thus,
the desired estimate (6.2.20) can be directly obtained by employing theorem (6.2.9) to the
smooth and periodic function ϕx [u].
Theorem 6.2.12. If refractive index n ∈ C p,α (Ω), 0 < α ≤ 1 and Q, R < p, F ≥ 1. Then
there exist a constant C such that

B B,Q,R,F
sup Kk,sing [u](x) − Kk,N,sing [u](x)

x∈Ω
 n o
h1 R+1
C max h(1−δ)p+α

1
δp+α
, h1 , h2
(Q+1)(1−δ)+α
, F if Q is odd,
≤ n o (6.2.21)
C max h(1−δ)p+α , hδp+α , hQ(1−δ)+α , h1 R+1

if Q is even,
1 1 2 F
where δ ∈ (0, 1/2].
152
Proof. Error in the approximation of singular integral is comprises of two error: quadrature
error and interpolation error incurred in the approximation of function values at quadrature
points. In view of this, we consider
ZZ
B B,Q,R,F
Kk,sing [u](x) − Kk,N,sing [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 )) dt1 dt2
[0,1]2
− L[0,1],N
N C,Q
L
PT
,t
2 2 [0,1],N 1 ,t1
[Gκ (x, ξ k (., .))ϕ̃k [u](1 − η)]
−1 (r)
%Z
Z1
+ G̃κ,B (x, ξk (τ, t2 ))ϕ̃xk,B [u](τ, t2 )ηBx (τ, t2 ) dτ dt2
0 %−1 (−r)
− Lk,N,Interp
B
[u] + Lk,N,Interp
B
[u] − Lk,N
B
[u]. (6.2.22)
Note that for all t = (t1 , t2 ) ∈ [0, 1]2 and for all l = (l1 , l2 ) with |l| = |l1 | + |l2 | ≤ p
|l|
∂ (Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 ))) C1

≤ (6.2.23)
∂tl r|l|

∂ |l| G̃ (x, ξ (τ, t ))ϕ̃x [u](τ, t )η x (τ, t ) C
κ,B k 2 k,B 2 B 2 2
≤ |l| , (6.2.24)

l l2 r
∂τ ∂t21
where r is the diameter of the support of η. Now for p < R, by invoking triangle inequality
and Theorem 6.2.8, lemma 6.2.6, and eqs. (6.2.23) and (6.2.24) in eq. (6.2.22), we have
R+1 !

B,Q,R,F
hp+α
1 hQ1 +α
δp+α h1
+ 2 Q1 + h1
B
Kk,sing [u](x) − Kk,N,sing [u](x) ≤ C + ,

r p r F
where Q1 is either equal to Q + 1 + α or Q + α depending upon Q is even or odd. As,
r = O(hδ1 ) = O(hδ2 ), therefore,

B B,Q,R,F

x∈Ω
R+1 !
(1−δ)p+α Q (1−δ)+α h1
≤ C max h1 , hδp+α
1 , h2 1 , .
F
Theorem 6.2.13. Suppose refractive index n ∈ C p,α (Ω), 0 ≤ p, 0 < α ≤ 1 and R < p,
F ≥ 1. Then there exist a constant C such that

I I,R,F

x∈Ω
153
n o
(1−δ)p+α (1−δ)p+α
≤ C max hδp+α
1 , h1 , hδp+α
2 , h2 , H (h1 , h2 , F, R) . (6.2.25)
where δ ∈ (0, 1/2] is the radius of support of floating partitions of unity, F ≥ 1, and
( R+1 )
R+1
h1 h2
H (h1 , h2 , F, R) = max , .
F F
Proof. Similar to the, approximation of singular integral over boundary patch, error in the
approximation of singular integral over interior patches composed of two parts: quadrature
error and interpolation error incurred in the approximation of function values at quadrature
points. In view of this, we consider
ZZ
I I,R,F
Kk,sing [u](x) − Kk,N,sing [u](x) = Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 )) dt1 dt2
[0,1]2
− L[0,1],N
PT
L
PT
,t
2 2 [0,1],N ,t
1 1
[Gκ (x, ξ k (., .))ϕ̃k [u](1 − η)]
−1 (r)
Z2π %Z
+ G̃κ,I (x, ξk (τ, θ))ϕ̃xk,I [u](τ, θ)ηIx (τ, θ) dτ dθ
0 0
− Lk,N,Interp
I
[u] + Lk,N,Interp
I
[u] − Lk,N
I
[u]. (6.2.26)
Note that for all t = (t1 , t2 ) ∈ [0, 1]2 and for all l = (l1 , l2 ) with |l| = |l1 | + |l2 | ≤ p
|l|
∂ (Gκ (x, ξk (t1 , t2 ))ϕ̃k [u](t1 , t2 )(1 − η((t1 , t2 ); (tx1 , tx2 ))) C1

≤ (6.2.27)
∂tl r|l|

∂ |l| G̃ (x, ξ (τ, θ))ϕ̃x [u](τ, θ)η x (τ, θ) C
κ,I k k,I I 2
≤ |l| , (6.2.28)

l
∂τ ∂θ l r
1 2

where r is the radius of the support of η. Since R < p, therefore, by invoking triangle
inequality and Theorem 6.2.9, lemma 6.2.7, and eqs. (6.2.27) and (6.2.28) in eq. (6.2.26),
we have

I I,R,F
Kk,sing [u](x) − Kk,N δ ,sing [u](x) ≤

p+α
hp+α

h1 2 δp+α δp+α
C + p + h1 + h2 + H (h1 , h2 , F, R) .
rp r
Since r = O(hδ1 ) = O(hδ2 ), therefore

I I,R,F
sup Kk,sing [u](x) − Kk,N [u](x)

β ,sing
x∈Ω
154
n o
(1−δ)p+α (1−δ)p+α
≤ C max hδp+α
1 , h1 , hδp+α
2 , h2 , H (h1 , h2 , F, R) .
Finally, in the following theorem we provide our final error estimate for the approxi-
mation of continuous operator (6.2.3) by the discrete operator defined in (6.2.8).
Theorem 6.2.14. Suppose refractive index n ∈ C p,α (Ω), 0 < α ≤ 1 and Q, R < p, 1 ≤ F .
Then we have

Q,R,F
sup K[u](x) − KN [u](x)

x∈Ω
n o
min{δp+α,(1−δ)p+α} min{δp+α,(1−δ)p+α,Q1 (1−δ)+α}
≤ C max h1 , h2 , H (h1 , h2 , R, F )
where C is constant depending on R, Q, F and δ ∈ (0, 1/2],

( R+1 )
R+1
h1 h2
H (h1 , h2 , R, F ) = max ,
F F
and Q1 is equal to Q if Q is even and Q + 1 otherwise.
Proof.
X
Q,R,F B B,Q,R,F
K[u](x) − KN [u](x) ≤ Kk,sing [u](x) − Kk,N,sing [u](x)

k∈MB (x)
X
B B,Q
+ Kk,reg [u](x) − Kk,N,reg [u](x)

k6∈MB (x)
X I,R,F

I
+ Kk,sing [u](x) − Kk,N,sing [u](x)

k∈MI (x)
X
I
Kk,reg I
+ [u](x) − Kk,N,reg [u](x) .
k6∈MI (x)
Now by using theorems 6.2.10 to 6.2.13, for Q, R < p and F ≥ 1, we have

Q,R,F
sup K[u](x) − KN [u](x)

x∈Ω
n o
min{δp+α,(1−δ)p+α} min{δp+α,(1−δ)p+α,Q1 (1−δ)+α}
≤ C max h1 , h2 , H (h1 , h2 , R, F ) .

Q,R,F
By the above theorem we conclude that if n ∈ C ∞ (Ω), then K[u](x) − KN [u](x)

∞
→ 0 as max{h1 , h2 } → 0 with high-order accuracy in the two dimensions.
155
6.2.2 Quadrature Analysis In Three Dimensions

In this subsection, we provide an error estimate for the approximation of volume integral
(4.0.1) by the high-order integration scheme described in Chapter 4. Most of the results
of this subsection are simple extension of the results in the preceding subsection.
Recall integral operator (4.0.1) can be expressed as
X X
B B
K[u](x) = Kk,sing [u](x) + Kk,reg [u](x) (6.2.29)
X X
I I
+ Kk,sing [u](x) + Kk,reg [u](x),
where
ZZZ
B
Kk,reg [u](x) = Gκ (x, ξk (t1 , t2 , t3 ))ϕ̃k [u](t1 , t2 , t3 ) dt1 dt2 dt3 (6.2.30)
[0,1]3
ZZZ
B
Kk,sing [u](x) = Gκ (x, ξk (t1 , t2 , t3 ))ϕ̃k [u](t1 , t2 , t3 )(1 − η((t1 , t2 , t3 ); (tx1 , tx2 , tx3 )) dt1 dt2 dt3
[0,1]3
−1 (r)
Z1 Z2π %Z
+ G̃κ,B (x, ξk (τ, θ, t3 ))ϕ̃xk,B [u](τ, θ, t3 )ηBx (τ, θ, t3 ) dτ dθ dt2 (6.2.31)
0
Z Z0Z 0
I
Kk,reg [u](x) = Gκ (x, ξk (t1 , t2 , t3 ))ϕ̃k [u](t1 , t2 , t3 ) dt1 dt2 dt3 (6.2.32)
[0,1]3
ZZZ
I
Kk,sing [u](x) = Gκ (x, ξk (t1 , t2 , t3 ))ϕ̃k [u](t1 , t2 , t3 )(1 − η((t1 , t2 , t3 ); (tx1 , tx2 , tx3 )) dt1 dt2 dt3
[0,1]3
Zπ Z2π Zr
+ G̃κ,I (x, ξk (ρ, θ, φ))ϕ̃xk,I [u](ρ, θ, φ)ηIx (ρ, θ, φ) dρ dθ dφ (6.2.33)
0 0 0
and
ϕ̃k [u](t1 , t2 , t3 ) = ϕk [u](t1 , t2 , t3 )ξk0 (t1 , t2 , t3 ),
G̃κ,B (x, ξk (τ, θ, t3 )) = %0 (τ )%(τ )Gκ (x, ξk (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t3 )),
ϕ̃xk,B [u](τ, θ, t3 ) = ϕk [u](ξk (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t3 ))
ξk0 (tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t3 ),
156
6.2.2. Quadrature Analysis In Three Dimensions
ηBx (τ, θ, t3 ) = η ((tx1 + %(τ ) cos θ, tx2 + %(τ ) sin θ, t3 ); (tx1 , tx2 , tx3 )) ,
G̃κ,I (x, ξk (τ, θ)) = ρ2 Gκ (x, ξk (tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ)) ,
ϕ̃xk,I [u](%(τ ), t2 ) = ϕ̃k [u](tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ),
ηIx (τ, θ) = η((tx1 + ρ cos θ sin φ, tx2 + ρ sin θ sin φ, tx3 + ρ cos φ), (tx1 , tx2 , tx3 )).
We define our approximate operator as
Q,R,F
X B,Q,R,F
X B,Q
KN [u](x) = Kk,N,sing [u](x) + Kk,N,reg [u](x) (6.2.34)
X I,R,F
X
I
+ Kk,N,sing [u](x) + Kk,N,reg [u](x),
where
B,Q
[u](x) = L[0,1],N
N C,Q
L[0,1],N2 ,t2 L[0,1],N
PT PT
Kk,N,reg 3 ,t3 1 ,t1
[Gκ (x, ξk (., ., .))ϕ̃k [u]] , (6.2.35)
B,Q,R,F
[u](x) = L[0,1],N
N C,Q
L[0,1],N2 ,t2 L[0,1],N
PT PT
Kk,N,sing 3 ,t3 1 ,t1
[Gκ (x, ξk (., ., .))ϕ̃k [u](1 − η)]
h h h x iii
+ L˜[0,1],N
N C,Q
L PT
L PT
P
TP x
3 ,t3 [0,2π],N δ ,θ [0,%−1 (r)],N δ ,τ G̃κ,B (x, ξ k (., ., .)) ϕ̃
2,R,N,F k,B [u] ηB ,
2 1
(6.2.36)
I
[u](x) = L[0,1],N
PT
L L
PT PT
Kk,N,reg 3 ,t3 [0,1],N2 ,t2 [0,1],N1 ,t1 [Gκ (x, ξ k (., ., .))ϕ̃ k [u]] , (6.2.37)
I,R,F
[u](x) = L[0,1],N
PT
L[0,1],N2 ,t2 L[0,1],N
PT PT
Kk,N,sing 3 ,t3 1 ,t1
[Gκ (x, ξk (., ., .))ϕ̃k [u](1 − η)]
h h h x iii
+ L[0,π],N
CC
δ ,φ L[0,2π],N δ ,θ L[0,r],N δ ,ρ G̃κ,I (x, ξk (., ., .)) P3,R,N,F ϕ̃k,I [u] ηI
PT PT
TP x
.
3 2 1
(6.2.38)
Subscript N1δ , N2δ , N3δ in eqs. (6.2.36) and (6.2.38) denote the number of points used for
accurate evaluation of singular integrals where exponent δ ∈ (0, 1/3].
In the following theorem we provide an error estimate for the approximation of inte-
gration of smooth function by means of combination of trapezoidal and Clenshaw-Curtis
quadrature rule.
Theorem 6.2.15. Let ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]), 0 < α ≤ 1, and 0 < p such
∂ |l| ϕ(t)
that for all l = (l1 , l2 , l3 ) with |l| = |l1 | + |l2 | + |l3 | ≤ p, ∂tl
is uniformly bounded and
157
∂ |l| ϕ(.,.,t3 )
∂tl
is periodic in [a1 , b1 ] × [a2 , b2 ] for all t3 ∈ [a3 , b3 ]. Then there exist a constant C
depending on ϕ such that
b b b
Z 3 Z 2 Z 1
1
ϕ(t) dt − L[a3 ,b3 ],N3 ,t3 L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] ≤ C max h1 , h2 , h3 ln
p+α p+α p+α
CC PT PT

,

h3
a3 a2 a1
(6.2.39)
where h1 = (b1 − a1 )/N1 , h2 = (b2 − a2 )/N2 , h2 = (b3 − a3 )/N3 and t = (t1 , t2 , t3 ).
Proof. Let
Zb3 Zb2 Zb1

EN [ϕ] = ϕ(t) dt − L[aCC
3 ,b3 ],N3 ,t3
L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ].
a3 a2 a1

b b b 
Z 3 Z 2 Z 1
 ϕ(t1 , ., .) dt1 − L[a ,b ],N ,t [ϕ] dt2 dt3
P T

|EN | ≤ 1 1 1 1

a3 a2 a1
b b 
Z 3 Z 2
L[a1 ,b1 ],N1 ,t1 [ϕ] dt2 − L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] dt3
PT PT PT

+  

a3 a2
b
Z 3
+ L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] dt3 − L[a3 ,b3 ],N3 ,t3 L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] .
P T P T CC P T P T

a3
It is easy to check that
L[aP1T,b1 ],N1 ,t1 L[aP2T,b2 ],N2 ,t2 [ϕ] = L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ]
L[aCC
3 ,b3 ],N3 ,t3
L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 [ϕ] = L[aP2T,b2 ],N2 ,t2 L[aP1T,b1 ],N1 ,t1 L[aCC
3 ,b3 ],N3 ,t3
[ϕ],
therefore,
b b b 
Z 3 Z 2 Z 1
ϕ(t1 , ., .) dt1 − L[a1 ,b1 ],N1 ,t1 [ϕ] dt2 dt3
PT

|EN | ≤
 

a3 a2 a1

Zb3 Zb2
+ L[a1 ,b1 ],N1 ,t1 ∞ ϕ(., t2 , .) dt2 − L[a2 ,b2 ],N2 ,t2 [ϕ] dt3
PT P T

a3 a2
158
b
Z 3
+ L[a2 ,b2 ],N2 ,t2 ∞ L[a1 ,b1 ],N1 ,t1 ∞ ϕ(., ., t3 ) dt3 − L[a3 ,b3 ],N3 ,t3 [ϕ] .
PT PT CC

a3
Note that kL[aP1T,b1 ],N1 ,t1 k∞ = b1 − a1 and kL[aP2T,b2 ],N2 ,t2 k∞ = b2 − a2 . Now by invoking
Theorems 6.2.3 and 6.2.4 in the above equation we have the desired estimate

p+α p+α p+α 1
|EN | ≤ C max h1 , h2 , h3 ln .
h3
The proofs of the Theorems 6.2.16 and 6.2.17 are almost identical to the those of the
Theorems 6.2.8 to 6.2.13 respectively, and thus stated below without proof.
0 0
Theorem 6.2.16. Let ϕ ∈ C p,α [a1 , b1 ] × [a2 , b2 ] × [a3 , t3 ) ∪ [a1 , b1 ] × [a2 , b2 ] × (t3 , b3 ] ,
0
a3 ≤ t3 ≤ b3 , and 0 < α ≤ 1, Q < p such that
0
1. ϕ(t1 , t2 , t3 ) has a corner type singularity at t3 = t3 .
∂ |l| ϕ(t) ∂ |l| ϕ(.,.,t3 )

2. For all l = (l1 , l2 ) with |l| = |l1 |+|l2 | ≤ p, ∂ l1 t1 ∂ l2 t2
is uniformly bounded and ∂ l1 t1 ∂ l2 t2
is periodic on [a1 , b1 ] × [a2 , b2 ] for all t3 ∈ [a3 , b3 ]. Then

b b b
Z 3 Z 2 Z 1
L˜ N C,Q
L L
P T P T

ϕ(t) dt − [c,d],N3 ,t3 [a ,b
2 2 ],N ,t
2 2 [a ,b
1 1 ],N ,t
1 1
[ϕ]

a3 a2 a1
 n o
C max h1p+α , hp+α

2 , h Q+1+α
3 if Q is odd,
≤ n o
C max hp+α , hp+α , hQ+α

if Q is even,
1 2 3
where C is some positive constant and h1 = (b1 − a1 )/N1 , h2 = (b2 − a2 )/N2 , h3 =

(b3 − a3 )/N3 and t = (t1 , t2 , t3 ).
Theorem 6.2.17. Let ϕ ∈ C p,α ([a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]), 0 < α ≤ 1, 0 ≤ p and for
∂ |l| ϕ(t) ∂ |l| ϕ(t)
all l = (l1 , l2 , l3 ) with |l| = |l1 | + |l2 | + |l3 | ≤ p, ∂tl
is
periodic for all t = (t1 , t2 , t3 ) ∈ [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. Then
b b b
Z 3 Z 2 Z 1
ϕ(t) dt − L[a3 ,b3 ],N3 ,t3 L[a2 ,b2 ],N2 ,t2 L[a1 ,b1 ],N1 ,t1 [ϕ] ≤ C max hp+α , hp+α , hp+α
P T P T P T

1 2 2 ,

a3 a2 a1
(6.2.40)
159
where C is constant depending on ϕ and h1 = (b1 − a1 )/N1 , h2 = (b2 − a2 )/N2 , h3 =

(b3 − a3 )/N3 .
In the following theorem we provide error estimates for the approximation of continuous
operators defined in Equations (6.2.30) to (6.2.33) by the approximate operators defined
in Equations (6.2.35) to (6.2.38).
Theorem 6.2.18. If refractive index n ∈ C p,α (Ω), 0 < α ≤ 1 and Q, R < p, F ≥ 1. Then

 Chmin{δp+α,(1−δ)p+α,(Q+1)(1−δ)+α} if Q is odd,
B B,Q
sup Kk,reg [u](x) − Kk,N,reg [u](x) ≤

x∈Ω\Pk Chmin{δp+α,(1−δ)p+α,Q(1−δ)+α} if Q is even,

(6.2.41)
I I

sup Kk,reg [u](x) − Kk,N,reg [u](x) ≤ Chmin{δp+α,(1−δ)p+α} , (6.2.42)
x∈Ω\Pk

 CHb (h1 , h2 , h3 , F, p, Q + 1, R, α, δ) if Q is odd,
B B,Q,R,F
sup Kk,sing [u](x) − Kk,N δ ,sing [u](x) ≤

x∈Ω CHb (h1 , h2 , h3 , F, p, Q, Rα, δ) if Q is even,

(6.2.43)

I I,R,F
sup Kk,sing [u](x) − Kk,N β ,sing [u](x) ≤ Hi (h1 , h2 , h3 , F, p, R, α, δ) (6.2.44)

x∈Ω
where
h = max {h1 , h2 , h3 } ,
Hb (h1 , h2 , h3 , F, p, Q, R, α, δ) =
( ( R+1 R+1 ))
h1 h2
max hmin{δp+α,(1−δ)p+α,Q(1−δ)+α} , max , , (6.2.45)
F F
( R+1 )
h
Hi (h1 , h2 , h3 , F, p, R, α, δ) = max hmin{δp+α,(1−δ)p+α} , , (6.2.46)
F
and δ ∈ (0, 1/3].
160
Proof. Proof follows from the Theorems 6.1.6, 6.1.10 and 6.2.15 to 6.2.17. Since, proof of
the estimates (6.2.41) to (6.2.44) are almost identical to the those of the Theorems 6.2.8
to 6.2.13 respectively, and are omitted.
In the following theorem we provide our final error estimate for the approximation of
integral operator (6.2.29) by approximate operator defined in (6.2.34).
Theorem 6.2.19. Suppose refractive index n ∈ C p,α (Ω), 0 < α ≤ 1, Q, R < p and F ≥ 1.
Then we have

Q,R,F
sup K[u](x) − KN [u](x) ≤ C max {Hb (h1 , h2 , h3 , F, p, Q1 , R, α, δ), Hi (h1 , h2 , h3 , F, p, R, α, δ)} ,

x∈Ω
where C is constant depending on p, Q, R, α, F , δ ∈ (0, 1/3] , Hi (h1 , h2 , h3 , F, p, α, δ) and

Hb (h1 , h2 , h3 , F, p, Q1 , R, α, δ) is given by eqs. (6.2.45) and (6.2.46) respectively and Q1 is
equal to Q if Q is even and Q + 1 otherwise.
Proof. By Equation (6.2.29) and (6.2.34) and invoking triangle inequality we have
X
Q,R,F B B,Q,R,F
K[u](x) − KN [u](x) ≤ Kk,sing [u](x) − Kk,N,sing [u](x)

k∈MB (x)
X
B B,Q
+ Kk,reg [u](x) − Kk,N,reg [u](x)

k6∈MB (x)
X I,R,F

I
+ K [u](x) − K [u](x)

k,sing k,N,sing
k∈MI (x)
X
I
Kk,reg I
+ [u](x) − Kk,N,reg [u](x) .
k6∈MI (x)
Now proof is follows simply by invoking Theorem 6.2.18 in the above equation.

Q,R,F
By the above theorem we conclude that if n ∈ C ∞ (Ω) then K[u](x) − KN [u](x)

∞
→ 0 as h = max {h1 , h2 , h3 } → 0 with high-order accuracy in the three dimensions.
161
Chapter
7
Conclusions and Future Directions
In this thesis, we proposed a fast high-order method for scattering of acoustic waves by
penetrable inhomogeneous media which requires O(N 1+δ/d ) computational cost for each
iteration of GMRES in Rd , for d = 2, 3. The constant δ ∈ (0, 1] can be chosen to achieve
favorable computational complexity and accuracy of the algorithm. In a case, when mate-
rial properties are infinitely differentiable within the scatterer Ω, the proposed algorithm
converges to high-order with computational complexity of order O(N log N ). Our rapid
convergent approximations are obtained through a use of specialized quadrature rule that
combine a number of elements. Starting from parametric representations of the scatterer,
a careful treatment of corner type singularities of the volume potentials in the trans-
verse direction, “near-singularities” in the tangential direction, and efficient interpolation
strategies; the evaluations are further expedited by utilizing a strategy based on equiv-
alent sources approximation on the Cartesian grids. We presented a series of numerical
experiments to demonstrate its performance in terms of computational efficiency as well
as numerical accuracy. We emphasize that this algorithm is designed to retain high-order
convergence even when material properties jump across the scattering interface and can be
employed for simulating scattering of acoustic waves by geometrically complex structures.
163
Chapter 7. Conclusions and Future Directions
For instance, in our numerical experiments where function m(x) = 1 − n2 (x) is discon-
tinuous across the material interface which, of course, limits the global smoothness of the
total field u, (in fact u ∈ C 1 (R3 )), we still obtain rapidly convergent approximations. This,
of course, become possible as a result of carefully avoiding integrating across the material
interface. To the best of our knowledge, this is the first solver which provide high-order
convergence with discontinuous material interface, where computational cost grows linearly
with problem size.
The work reported in this thesis, suggest a number of interesting and important re-
search directions for the solution of integral equations arising from scattering by acoustic
and electromagnetic waves. For example, in this thesis we have made assumption that
material density is constant through out the space. If we relax this condition and allow
the density vary as a function of position with in the inhomogeneity, then the solution of
acoustic scattering problem is obtained by solving Bergmann’s equation, subject to the
certain transmission condition across the interface between the inhomogeneity and sur-
rounding material [86]. An equivalent integral equation formulation corresponding to this
scattering problem presented in [86, Theorem 3.1], where integral kernel is weakly singular.
We believe that, our method can be extended to obtain an efficient high-order solver for
aforementioned problem. In the same paper [86], author have discussed integral equation
formulation for electromagnetic scattering problem from bounded inhomogeneous media.
This problem can also be solved by the strategy proposed in this thesis, in fact, currently, we
are working on these problems. Lippmann-Schwinger type integral equation also arises in
the context of acoustic/electromagnetic scattering by orthotropic and anisotropic medium
[42, 95]. Although, integral equations corresponding to these scattering problems contain
more complicated singularities than the one considered in this thesis, nevertheless, we be-
lieve that, a careful treatment of singularities in conjunction with the strategy developed
in this thesis will provide an efficient high-order solver for these problems too.
As we have seen in chapter 2, and 4, the number of iterations increases significantly
with the increase in wavenumber of incoming wave. This is because, Lippmann–Schwinger
integral equation is numerically ill-conditioned. Indeed, for scattering object of large size
164
Chapter 7. Conclusions and Future Directions
(κa >> 1), or of large contrast (m(x) >> 1), the iterative solver converges so slowly
that it becomes infeasible to obtain the desired degree of accuracy in residuals. This
difficulty, results in a challenging computational problem especially in three dimensions.
A possible remedy of this problem is proposed by introducing a preconditioning matrix
for the linear system, obtained by discretization of integral equation. To the best of
our knowledge, only limited attempt has been made for developing good preconditioners
related to the Lippmann–Schwinger equation and it remain a copious subject matter for the
future research [21, 29, 108]. In future we intend to develop an efficient preconditioning
technique for the solution of scattering by acoustic/electromagnetic wave by penetrable
inhomogeneous media in two and three dimensions.
Finally, in this thesis we have only considered scattering object with smooth boundaries.
However, inhomogeneities with sharp feature such as corner and cusps play a significant
role in many fields of science and engineering applications. In future, we intend to extend
our algorithm for such scattering object, which is by no means trivial if we want to retain
the high-order convergence rate.
165
Bibliography
[1] Aguilar, J. and Chen, Y., A high-order, fast algorithm for scattering calculation in
two dimensions, Computers & Mathematics with Applications, 47 (2004), no. 1,
1–11.
[2] Ainsworth, M., Davies, P., Duncan, D. B., Martin, P. A. and Rynne, B., Topics in
Computational Wave Propagation: Direct and Inverse Problems, vol. 31, Springer
Science & Business Media (2012).
[3] Aki, K. and Richards, P. G., Quantitative seismology, vol. 1 (2002).
[4] Alpert, B., Beylkin, G., Coifman, R. and Rokhlin, V., Wavelet-like bases for the fast
solution of second-kind integral equations, SIAM Journal on Scientific Computing,
14 (1993), no. 1, 159–184.
[5] Ambikasaran, S., Borges, C., Imbert-Gerard, L.-M. and Greengard, L., Fast, adap-
tive, high order accurate discretization of the lippmann-schwinger equation in two
dimension, arXiv preprint arXiv:1505.07157, (2015).
[6] Anand, A., An efficient high-order algorithm for scattering from penetrable thin struc-
tures, University of Minnesota (2006).
[7] Anand, A. and Reitich, F., An efficient high-order algorithm for acoustic scattering
from penetrable thin structures in three dimensions, The Journal of the Acoustical
Society of America, 121 (2007), no. 5, 2503–2514.
167
Bibliography
[8] Andersson, F. and Holst, A., A fast, bandlimited solver for scattering problems in
inhomogeneous media, Journal of Fourier Analysis and Applications, 11 (2005), no. 4,
471–487.
[9] Atkinson, K. E., The numerical solution of integral equations of the second kind, 4,
Cambridge university press (1997).
[10] Atkinson, K. E., An introduction to numerical analysis, John Wiley & Sons (2008).
[11] Babuška, I. and Suri, M., The p and h-p versions of the finite element method, basic
principles and properties, SIAM review, 36 (1994), no. 4, 578–632.
[12] Bayliss, A., Goldstein, C. I. and Turkel, E., On accuracy conditions for the numerical
computation of waves, Journal of Computational Physics, 59 (1985), no. 3, 396–404.
[13] Berenger, J.-P., A perfectly matched layer for the absorption of electromagnetic waves,
Journal of computational physics, 114 (1994), no. 2, 185–200.
[14] Bleszynski, E., Bleszynski, M. and Jaroszewicz, T., A fast integral-equation solver
for electromagnetic scattering problems, in Antennas and Propagation Society Inter-
national Symposium, 1994. AP-S. Digest, vol. 1, pp. 416–419, IEEE (1994).
[15] Bleszynski, E., Bleszynski, M. and Jaroszewicz, T., Aim: Adaptive integral method for
solving large-scale electromagnetic scattering and radiation problems, Radio Science,
31 (1996), no. 5, 1225–1251.
[16] Borden, B., Mathematical problems in radar inverse scattering, Inverse Problems, 18
(2002), no. 1, R1.
[17] Borup, D. T. and Gandhi, O. P., Fast-fourier-transform method for calculation of

sar distributions in finely discretized inhomogeneous models of biological bodies, Mi-
crowave Theory and Techniques, IEEE Transactions on, 32 (1984), no. 4, 355–360.
[18] Bourlier, C., Pinel, N. and Kubické, G., Method of moments for 2D scattering prob-
lems: basic concepts and applications, John Wiley & Sons (2013).
168
Bibliography
[19] Bowman, J. J., Senior, T. B. and Uslenghi, P. L., Electromagnetic and acoustic
scattering by simple shapes, Tech. rep., DTIC Document (1970).
[20] Bruno, O. P. and Geuzaine, C. A., An o (1) integration scheme for three-dimensional
surface scattering problems, Journal of Computational and Applied Mathematics,
204 (2007), no. 2, 463–476.
[21] Bruno, O. P. and Hyde, E. M., An efficient, preconditioned, high-order solver for scat-
tering by two-dimensional inhomogeneous media, Journal of Computational Physics,
200 (2004), no. 2, 670–694.
[22] Bruno, O. P. and Hyde, E. M., Higher-order fourier approximation in scattering by

two-dimensional, inhomogeneous media, SIAM Journal on Numerical Analysis, 42
(2005), no. 6, 2298–2319.
[23] Bruno, O. P. and Kunyansky, L. A., A fast, high-order algorithm for the solution of
surface scattering problems: basic implementation, tests, and applications, Journal
of Computational Physics, 169 (2001), no. 1, 80–110.
[24] Bruno, O. P. and Kunyansky, L. A., Surface scattering in three dimensions: an accel-
erated high–order solver, in Proceedings of the Royal Society of London A: Mathemat-
ical, Physical and Engineering Sciences, vol. 457, pp. 2921–2934, The Royal Society
(2001).
[25] Bu, F., Lin, J. and Reitich, F., A fast and high-order method for the three-dimensional
elastic wave scattering problem, Journal of Computational Physics, 258 (2014), 856–
870.
[26] Canning, F. X., Improved impedance matrix localization method [em problems], An-
tennas and Propagation, IEEE Transactions on, 41 (1993), no. 5, 659–667.
[27] Carrion, P., Inverse problems and tomography in acoustic and seismology, penn publ,
Co., Atlanta, GA, (1987).
169
Bibliography
[28] CÁtedra, M. F. and Gago, E., Spectral domain analysis of conducting patches of
arbitrary geometry in multilayer media using the cg-fft method, Antennas and Prop-
agation, IEEE Transactions on, 38 (1990), no. 10, 1530–1536.
[29] Chen, Y., A fast, direct algorithm for the lippmann–schwinger integral equation in
two dimensions, Advances in Computational Mathematics, 16 (2002), no. 2-3, 175–
190.
[30] Chew, W. and Liu, Q., Perfectly matched layers for elastodynamics: a new absorbing
boundary condition, Journal of Computational Acoustics, 4 (1996), no. 04, 341–359.
[31] Chew, W. C., Jin, J.-M., Lu, C.-C., Michielssen, E. and Song, J. M., Fast solution
methods in electromagnetics, Antennas and Propagation, IEEE Transactions on, 45
(1997), no. 3, 533–543.
[32] Chew, W. C. and Weedon, W. H., A 3d perfectly matched medium from modified
maxwell’s equations with stretched coordinates, Microwave and optical technology
letters, 7 (1994), no. 13, 599–604.
[33] Clenshaw, C. W. and Curtis, A. R., A method for numerical integration on an auto-
matic computer, Numerische Mathematik, 2 (1960), no. 1, 197–205.
[34] Coifman, R., Rokhlin, V. and Wandzura, S., The fast multipole method for the wave
equation: A pedestrian prescription, Antennas and Propagation Magazine, IEEE, 35
(1993), no. 3, 7–12.
[35] Collino, F. and Monk, P., The perfectly matched layer in curvilinear coordinates,
SIAM Journal on Scientific Computing, 19 (1998), no. 6, 2061–2090.
[36] Colton, D. and Kress, R., Inverse acoustic and electromagnetic scattering theory,
vol. 93, Springer (2012).
[37] Colton, D. and Kress, R., Integral equation methods in scattering theory, vol. 72,
SIAM (2013).
170
Bibliography
[38] Colton, D. and Monk, P., A linear sampling method for the detection of leukemia
using microwaves, SIAM Journal on Applied Mathematics, 58 (1998), no. 3, 926–
941.
[39] Colton, D. and Monk, P., A linear sampling method for the detection of leukemia
using microwaves ii, SIAM Journal on Applied Mathematics, 60 (1999), no. 1, 241–
255.
[40] Conyers, L. B. and Goodman, D., Ground-penetrating radar, AltaMira Press An

Introduction for Archaeologist (1997).
[41] Darve, E., The fast multipole method: numerical implementation, Journal of Com-
putational Physics, 160 (2000), no. 1, 195–240.
[42] Dassios, G. and Karadima, K., Time harmonic acoustic scattering in anisotropic
media, Mathematical methods in the applied sciences, 28 (2005), no. 12, 1383–1401.
[43] de De Villiers, J., Mathematics of approximation, vol. 1, Springer Science & Business
Media (2012).
[44] Djellouli, R., Farhat, C., Macedo, A. and Tezaur, R., Finite element solution of two-
dimensional acoustic scattering problems using arbitrarily shaped convex artificial
boundaries, Journal of Computational Acoustics, 8 (2000), no. 01, 81–99.
[45] Donohue, D., Ku, H.-C. and Thompson, D. R., Application of iterative moment-
method solutions to ocean surface radar scattering, Antennas and Propagation, IEEE
Transactions on, 46 (1998), no. 1, 121–132.
[46] Duan, R. and Rokhlin, V., High-order quadratures for the solution of scattering prob-
lems in two dimensions, Journal of Computational Physics, 228 (2009), no. 6, 2152–
2174.
171
Bibliography
[47] Engquist, B. and Majda, A., Absorbing boundary conditions for numerical simulation
of waves, Proceedings of the National Academy of Sciences, 74 (1977), no. 5, 1765–
1766.
[48] Epstein, C. L., Introduction to the mathematics of medical imaging, Siam (2008).
[49] Felsen, L. B. and Marcuvitz, N., Radiation and scattering of waves, vol. 31, John
Wiley & Sons (1994).
[50] Friedman, A., Mathematics in industrial problems, vol. 67, Springer Science & Busi-
ness Media (2012).
[51] Fujimoto, J. G., Pitris, C., Boppart, S. A. and Brezinski, M. E., Optical coherence
tomography: an emerging technology for biomedical imaging and optical biopsy, Neo-
plasia, 2 (2000), no. 1, 9–25.
[52] Gan, H., Levin, P. and Ludwig, R., Finite element formulation of acoustic scattering
phenomena with absorbing boundary condition in the frequency domain, The Journal
of the Acoustical Society of America, 94 (1993), no. 3, 1651–1662.
[53] Gharsalli, L., Ayasso, H., Duchêne, B. and Mohammad-Djafari, A., Inverse scat-
tering in a bayesian framework: application to microwave imaging for breast cancer
detection, Inverse Problems, 30 (2014), no. 11, 114011.
[54] Gillman, A., Barnett, A. H. and Martinsson, P.-G., A spectrally accurate direct so-
lution technique for frequency-domain scattering problems with variable media, BIT
Numerical Mathematics, 55 (2014), no. 1, 141–170.
[55] Giorgi, G., Brignone, M., Aramini, R. and Piana, M., Application of the inhomoge-
neous lippmann–schwinger equation to inverse scattering problems, SIAM Journal on
Applied Mathematics, 73 (2013), no. 1, 212–231.
[56] Givoli, D., Non-reflecting boundary conditions, Journal of Computational Physics,

94 (1991), no. 1, 1–29.
172
Bibliography
[57] Givoli, D., High-order local non-reflecting boundary conditions: a review, Wave mo-
tion, 39 (2004), no. 4, 319–326.
[58] Givoli, D. and Keller, J. B., Non-reflecting boundary conditions for elastic waves,
Wave motion, 12 (1990), no. 3, 261–279.
[59] Golub, G. H. and Van Loan, C. F., Matrix computations. 1989.
[60] Gumerov, N. A. and Duraiswami, R., Fast multipole methods for the Helmholtz equa-
tion in three dimensions, Elsevier (2005).
[61] Hamilton, L., Macdonald, P., Stalzer, M., Turley, R., Visher, J. and Wandzura, S.,
3d method of moments scattering computations using the fast multipole method, in
Antennas and Propagation Society International Symposium, 1994. AP-S. Digest,
vol. 1, pp. 435–438, IEEE (1994).
[62] Hesford, A. J. and Waag, R. C., The fast multipole method and fourier convolu-
tion for the solution of acoustic scattering on regular volumetric grids, Journal of
computational physics, 229 (2010), no. 21, 8199–8210.
[63] Hounsfield, G. N., Computerized transverse axial scanning (tomography): Part 1.

description of system, The British journal of radiology, 46 (1973), no. 552, 1016–
1022.
[64] Hyde, E. M., Fast, high-order methods for scattering by inhomogeneous media, Ph.D.
thesis, California Institute of Technology (2002).
[65] Hyde, E. M. and Bruno, O. P., A fast, higher-order solver for scattering by penetrable
bodies in three dimensions, Journal of Computational Physics, 202 (2005), no. 1,
236–261.
[66] Ihlenburg, F., Finite element analysis of acoustic scattering, vol. 132, Springer Sci-
ence & Business Media (2006).
173
Bibliography
[67] Jameson, L., High order schemes for resolving waves: number of points per wave-
length, Journal of scientific computing, 15 (2000), no. 4, 417–439.
[68] Jones, D. S., Acoustic and electromagnetic waves, Oxford/New York, Clarendon
Press/Oxford University Press, 1986, 764 p., 1 (1986).
[69] Keller, J. B. and Givoli, D., Exact non-reflecting boundary conditions, Journal of
computational physics, 82 (1989), no. 1, 172–192.
[70] Kendall, A. and Weimin, H., Theoretical numerical analysis–a functional analysis
framework, Texts in Applied Mathematics, 39 (2001).
[71] Khalil, M. H. and Jiadong, X., Inverse scattering in microwave imaging for detection
of malignant tumor inside the human body., International Journal of Signal Process-
ing, Image Processing & Pattern Recognition, 7 (2014), no. 1.
[72] Khankhoje, U. K., van Zyl, J. J. and Cwik, T. A., Computation of radar scatter-
ing from heterogeneous rough soil using the finite-element method, Geoscience and
Remote Sensing, IEEE Transactions on, 51 (2013), no. 6, 3461–3469.
[73] Kirsch, A. and Hettlich, F., The mathematical theory of time-harmonic maxwell’s
equations (2015).
[74] Kirsch, A. and Monk, P., An analysis of the coupling of finite-element and nyström
methods in acoustic scattering, IMA Journal of numerical analysis, 14 (1994), no. 4,
523–544.
[75] Kooij, B. and Lambert, M., Nonlinear inversion of a buried object in te-scattering,
in Antennas and Propagation Society International Symposium, 1997. IEEE., 1997
Digest, vol. 4, pp. 2617–2620, IEEE (1997).
[76] Kosloff, R. and Kosloff, D., Absorbing boundaries for wave propagation problems,
Journal of Computational Physics, 63 (1986), no. 2, 363–376.
[77] Kress, R., Linear integral equations, vol. 82, Springer.
174
Bibliography
[78] Kress, R. and Roach, G., Transmission problems for the helmholtz equation, Journal
of Mathematical Physics, 19 (1978), no. 6, 1433–1437.
[79] Labreuche, C., A convergence theorem for the fast multipole method for 2 dimensional
scattering problems, Mathematics of Computation of the American Mathematical
Society, 67 (1998), no. 222, 553–591.
[80] Lanzara, F., Maz’ya, V. and Schmidt, G., Numerical solution of the lippmann–
schwinger equation by approximate approximations, Journal of Fourier Analysis and
Applications, 10 (2004), no. 6, 645–660.
[81] Lee, C. F., Shin, R. and Kong, J., Finite difference method for electromagnetic scat-
tering problems, Ph.D. thesis, Massachusetts Institute of Technology, Dept. of Elec-
trical Engineering and Computer Science (1990).
[82] Lin, J. and Chew, W., Bicg-fft t-matrix method for solving for the scattering solution
from inhomogeneous bodies, Microwave Theory and Techniques, IEEE Transactions
on, 44 (1996), no. 7, 1150–1155.
[83] Liu, G. and Gedney, S., High-order nyström solution of the volume efie for tm-wave
scattering, Microwave and Optical Technology Letters, 25 (2000), no. 1, 8–11.
[84] Lui, H.-S., Fhager, A. and Persson, M., On the forward scattering of microwave breast
imaging, Journal of Biomedical Imaging, 2012 (2012), 6.
[85] Martin, P., Acoustic scattering by inhomogeneous spheres, J. Acoust. Soc. Am, 111
(2002), no. 5, 2013–2018.
[86] Martin, P., Acoustic scattering by inhomogeneous obstacles, SIAM Journal on Applied
Mathematics, 64 (2003), no. 1, 297–308.
[87] McLean, W. C. H., Strongly elliptic systems and boundary integral equations, Cam-
bridge university press (2000).
175
Bibliography
[88] Meddahi, S., Márquez, A. and Selgas, V., Computing acoustic waves in an inho-
mogeneous medium of the plane by a coupling of spectral and finite elements, SIAM
Journal on Numerical Analysis, 41 (2003), no. 5, 1729–1750.
[89] Medvinsky, M., Tsynkov, S. and Turkel, E., High order numerical simulation of
the transmission and scattering of waves using the method of difference potentials,
Journal of Computational Physics, 243 (2013), 305–322.
[90] Miranda, C., Partial differential equations of elliptic type, vol. 2, Springer-Verlag
(2013).
[91] Otremba, Z. and Piskozub, J., Modelling of the optical contrast of an oil film on a
sea surface, Optics Express, 9 (2001), no. 8, 411–416.
[92] Peterson, A. F., Ray, S. L., Mittra, R., of Electrical, I. and Engineers, E., Computa-
tional methods for electromagnetics, vol. 2, IEEE press New York (1998).
[93] Phillips, J. R. and White, J. K., A precorrected-fft method for electrostatic analysis
of complicated 3-d structures, Computer-Aided Design of Integrated Circuits and
Systems, IEEE Transactions on, 16 (1997), no. 10, 1059–1072.
[94] Potthast, R., Point sources and multipoles in inverse scattering theory, CRC Press
(2010).
[95] Potthast, R. et al., Electromagnetic scattering from an orthotropic medium, J. Inte-

gral Equations Appl, 11 (1999), no. 2, 197–215.
[96] Rokhlin, V., Rapid solution of integral equations of scattering theory in two dimen-
sions, Journal of Computational Physics, 86 (1990), no. 2, 414–439.
[97] Saad, Y. and Schultz, M. H., Gmres: A generalized minimal residual algorithm for
solving nonsymmetric linear systems, SIAM Journal on scientific and statistical com-
puting, 7 (1986), no. 3, 856–869.
176
Bibliography
[98] Schmitt, J., Yadlowsky, M. and Bonner, R., Subsurface imaging of living skin with
optical coherence microscopy, Dermatology, 191 (1995), no. 2, 93–98.
[99] Schmitt, J. M., Knuttel, A., Yadlowsky, M. and Eckhaus, M., Optical-coherence
tomography of a dense tissue: statistics of attenuation and backscattering, Physics in
Medicine and Biology, 39 (1994), no. 10, 1705.
[100] Taflove, A. and Hagness, S., Computational electrodynamics: The finite-difference

time-domain method (artech house, boston), Tech. rep., ISBN 978-1-58053-832-9
(2005).
[101] Tsang, L., Kong, J. A. and Ding, K.-H., Scattering of Electromagnetic Waves, The-
ories and Applications, vol. 27, John Wiley & Sons (2004).
[102] Vainikko, G., Fast solvers of the lippmann-schwinger equation, in Direct and inverse
problems of mathematical physics, pp. 423–440, Springer (2000).
[103] Wagner, R. L. and Chew, W. C., A study of wavelets for the solution of electro-
magnetic integral equations, Antennas and Propagation, IEEE Transactions on, 43
(1995), no. 8, 802–810.
[104] Werner, P., Zur mathematischen theorie akustischer wellenfelder, Archive for rational
mechanics and analysis, 6 (1960), no. 1, 231–260.
[105] Werner, P., Beugungsprobleme der mathematischen akustik, Archive for rational me-
chanics and analysis, 12 (1963), no. 1, 155–184.
[106] Xu, X. M. and Liu, Q. H., Fast spectral-domain method for acoustic scattering prob-
lems, Ultrasonics, Ferroelectrics, and Frequency Control, IEEE Transactions on, 48
(2001), no. 2, 522–529.
[107] Xu, X. M. and Liu, Q. H., The bcgs-fft method for electromagnetic scattering from
inhomogeneous objects in a planarly layered medium, Antennas and Wireless Propa-
gation Letters, IEEE, 1 (2002), no. 1, 77–80.
177
Bibliography
[108] Ying, L., Sparsifying preconditioner for the lippmann-schwinger equation, arXiv
preprint arXiv:1408.4495, (2014).
[109] Zhang, Z. and Liu, Q., Three-dimensional weak-form conjugate-and biconjugate-

gradient fft methods for volume integral equations, Microwave and Optical Technology
Letters, 29 (2001), no. 5, 350–356.
[110] Zion, B. and Leviatan, Y., On the use of wavelet expansions in the method of mo-
ments, IEEE Transaction on antennas and propagation, 41 (1993), no. 5.
178

An Efficient High-Order Algorithm For Scattering by Inhomogeneous Penetrable Media

Uploaded by

Copyright:

Available Formats

An Efficient High-Order Algorithm For Scattering by Inhomogeneous Penetrable Media

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Efficient High-Order Algorithm For Scattering by Inhomogeneous Penetrable Media

Uploaded by

Copyright:

Available Formats

An Efficient High-Order Algorithm

Department of Mathematics and Statistics

Indian Institute of Technology, Kanpur

in Partial Fulfilment of the Requirements

for the Degree of

Name of the Student : Ambuj Pandey

Efficient and accurate simulation of acoustic or electromagnetic wave scattering by

∆u(x) + κ2 n2 (x)u(x) = 0, x ∈ Rd , (0.0.2)

known as the Lippmann–Schwinger equation, where Gκ (x, y) is the fundamental solution

List of Figures xiii

List of Tables xix

1 Introduction and Preliminaries 1

2 An efficient high-order integral equation solver in two dimensions 27

3 Computational Results In Two Dimensions 57

4 An efficient high-order integral equation solver in three dimensions 87

5 Computational Results In Three Dimensions 109

6 Error Analysis 127

7 Conclusions and Future Directions 163

2.2 A typical point in the set Tk for a patch Pk when N1 = 8, N2 = 4 . . . . . . . 31

2.6 Partition of unity on disc shape scatterer. . . . . . . . . . . . . . . . . . . . . 36

2.7 Partition of unity on bean shape scatterer. . . . . . . . . . . . . . . . . . . . . 37

2.8 Change of variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.9 An illustration of the locations xij,` of equivalent sources (points displayed in

5.2 Three overlapping coordinate patches and corresponding computational grid

5.2 Convergence study: Plane wave scattering by a sphere of acoustic size κa = 2

5.4 Convergence study: Plane wave scattering by a sphere of acoustic size κa = 10

knowledge of scattering of radar waves is vital to applications in scientific, commercial, and

Wave scattering phenomena is at heart of medical imaging techniques such as X-Ray

1.2 Mathematical Preliminaries

1.2.1 The Helmholtz equation

which is Euler’s equation. The equation of continuity is given by

the adiabatic hypothesis by

U (x, t) = U1 (x, t) + · · · (1.2.5)

p(x, t) = p0 (x) + p1 (x, t) + · · · (1.2.6)

ρ(x, t) = ρ0 (x) + ρ1 (x, t) + · · · (1.2.7)

S (x, t) = S0 (x) + S1 (x, t) + · · · (1.2.8)

∆u(x) + κ2 n2 (x)u(x) = 0 (1.2.15)

where c0 is the speed of sound in the vacuum.

1.2.2 Scattering Formulation

∆u(x) + κ2 u(x) = 0, x ∈ Rd \ Ω (1.2.17)

u(x) = 0, x ∈ ∂Ω. (1.2.18)

∆u(x) + κ2 n2 (x)u(x) = 0 (1.2.20)

in Rd , d = 2, 3. Here, it is important to note that refractive index n(x) varies arbitrarily in

u(x) = ui (x) + us (x) (1.2.21)

∆u(x) + κ2 n2 (x)u(x) = 0, (1.2.22)

∆ui (x) + κ2 ui (x) = 0, (1.2.23)

1.2.3 Fundamental Solution

∆Gκ (x, y) + κ2 Gκ (x, y) = −δ(x − y) for x, y ∈ Rd ,

Hν1 (x) = Jν (x) + iYν (x).

Definition 1.2.1. Let K : Ω × Ω → C be a function defined and continuous for all

is a compact operator on C(Ω).

Definition 1.2.3. Let ϕ be a real or complex valued function defined on Ω ⊂ Rd . Then

|ϕ(x) − ϕ(y)| ≤ C|x − y|α (1.2.27)

1.2.4 Integral Equation Formulation

Theorem 1.2.4. Green’s first identity

Theorem 1.2.5. Green’s second identity

The following Green’s representation theorem allows to represent smooth functions in

Theorem 1.2.6. Green’s representation theorem

where the domain integral exists as an improper integral. In particular, if u is a solution

U (x, t) = U1 (x, t) + · · · (1.2.5)

p(x, t) = p0 (x) + p1 (x, t) + · · · (1.2.6)

ρ(x, t) = ρ0 (x) + ρ1 (x, t) + · · · (1.2.7)

S (x, t) = S0 (x) + S1 (x, t) + · · · (1.2.8)

error of 0.03% in the max norm ∞ .