Scattering Theory

Lecture Notes

Prof. Dr. Ruming Zhang

Institut für Mathematik
Technische Universität Berlin
Wintersemester 2023/24

February 5, 2024
2
Contents

1 Introduction 1

2 Functional analysis 3
2.1 Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Compact Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 The Adjoint Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 The Helmholtz equation 11

3.1 Time-harmonic acoustic waves . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Time-harmonic electromagnetic waves . . . . . . . . . . . . . . . . . . . . . 12
3.3 Green’s Theorem and Formula . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Farfield pattern and further properties . . . . . . . . . . . . . . . . . . . . 19

4 Boundary integral equations 25

4.1 Uniquenss of the exterior Dirichlet problem . . . . . . . . . . . . . . . . . . 25
4.2 Single- and double layer potentials . . . . . . . . . . . . . . . . . . . . . . 27
4.3 Numerical implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5 Inverse problems 41
5.1 Ill-posed problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3 Singular value decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.4 Tikhonov regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

6 Inverse scattering problems 55

6.1 Uniqueness of the inverse obstacle problems . . . . . . . . . . . . . . . . . 56
6.2 The linear sampling method . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.3 The iterative method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

7 Scattering by an Orthotropic Medium 73

7.1 Orthotropic Medium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.2 Mathematical framework for the variational methods . . . . . . . . . . . . 74
7.3 Variational method for scattering problems . . . . . . . . . . . . . . . . . . 76

i
ii CONTENTS

7.4 Scattering by an Orthotropic medium . . . . . . . . . . . . . . . . . . . . . 79

7.5 Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.6 Interior Transmission Problems . . . . . . . . . . . . . . . . . . . . . . . . 83
7.7 Uniqueness of the inverse problems . . . . . . . . . . . . . . . . . . . . . . 89

8 Scattering from inhomogeneous medium 95

8.1 Lippmann-Schwinger equation . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.2 The unique continuation principle . . . . . . . . . . . . . . . . . . . . . . . 98
8.3 Uniqueness results for the inverse problems . . . . . . . . . . . . . . . . . . 102
8.4 Stability estimates for the inverse problems . . . . . . . . . . . . . . . . . . 106
8.5 Numerical solution to the inverse medium problem . . . . . . . . . . . . . 110
Chapter 1

Introduction

The scattering theory includes the study of direct and inverse wave scattering problems.
Waves are always modelled by certain partial differential equations (PDEs) with proper
boundary conditions. Here I give an example to show what are the direct/inverse scattering
problems, and what are the interesting topics.
Given a Helmholtz equation:

∆u + k 2 nu = f ∈ B(0, R),

where k > 0 is the wave number, n > 0 is the function valued refractive index, f is the
source and B(0, R) is a ball centered at 0 with the radius R > 0.
The direct problem is to study the solution u, when everything (k, n, f ) is known.
We are interested in the well-posedness of the problem (existence, uniqueness, stability),
as well as numerical methods to solve the problems. The inverse problem is the opposite.
When some of the information is missing (for example, n), but we can measure the solution
u at some points. Then how can we reconstruct n? We are interested in the uniqueness
and stability of the problem, as well as efficient numerical schemes.
Inverse scattering theory has been a very active field in applied mathematics from the
1980’s. There are already a lot of successful applications of the scattering theory, despite
its non-complete mathematical theory. Some important examples are listed as follows:

• Medical imaging (ultrasound, magnetic resonance imaging (MRI), computed tomog-

raphy (CT)).

• Nondestructive testing.

• Underwater detection.

From mathematical point of view, the study of the inverse scattering problems depends
on a good understanding of direct scattering problems. In this lecture, we will mainly focus
on the direct scattering problems. The main topics are:

• Functional analysis

• The Helmholtz equation, Green’s theorem and radiation condition

1
2 CHAPTER 1. INTRODUCTION

• Potential theory, boundary integral equations, scattering by bounded obtacles

• Variational method, Scattering by orthotropic media

• The Lippmann-Schwinger equation, scattering by an inhomogeneous medium

Chapter 2

Functional analysis

2.1 Normed Spaces

Definition 2.1.1. Let X be a vector space over the field of complex numbers C and there
is a real-valued function x 7→ ∥x∥. If ∥ · ∥ satisfies:

1. ∥x∥ ≥ 0,

2. ∥x∥ = 0 if and only if x = 0,

3. ∥tx∥ = |t|∥x∥ for all t ∈ C,

4. ∥x + y∥ ≤ ∥x∥ + ∥y∥

for all x, y ∈ X, then it is a norm on X. A vector space X equipped with a norm is called
a normed space.

Example 2.1.2. X is the vector space of complex valued continuous functions in the
bounded interval [a, b]. Then
∥φ∥ := max |φ(x)|
x∈[a,b]

is a norm X. This normed space is denoted by C[a, b]].

Example 2.1.3. X is the vector space of Lebesgue square integrable functions in the
bounded interval [a, b]. Then
Z b
∥φ∥ : |φ(x)|2 dx .
a

is a normon X. This normed space is denoted by L2 [a, b]].

Definition 2.1.4. Let X be a normed space, we introduce a topological structure on X:

• A sequence {φn } ⊂ X converges to φ ∈ X if ∥φn − φ∥ → 0 as n → 0;

• A subspace U is closed if it contains all limits of convergent sequences of U .

• The closure U of U is the set of all limits of convergent sequences of U .

3
4 CHAPTER 2. FUNCTIONAL ANALYSIS

• A set U is dense in X if U = X.
Definition 2.1.5. A sequence {φn } ⊂ X is called a Cauchy sequence if for every ε > 0,
there is an integer N > 0 such that ∥φm − φn ∥ ≤ ε for all m, n > N .
Definition 2.1.6. A subset U of X is complete if every Cauchy sequence converges to
an element of U .
Definition 2.1.7. A complete normed space X is called a Banach space.
Example 2.1.8. 1. Let X be the space of polynomials in the bounded interval [a, b]
equipped with the norm
∥φ∥ := max |φ(x)|.
x∈[a,b]

X is not a Banach space.

2. C[a, b] is a Banach space.
3. L2 [a, b] is a Banach space.
Now we introduce an inner product and the Hilbert space.
Definition 2.1.9. Let X be a vector space over the filed of complex numbers C. A function
(·, ·) : X × X → C such that
1. (φ, φ) ≥ 0,
2. (φ, φ) = 0 if and only if φ = 0,
3. (φ, ψ) = (ψ, φ),
4. (αφ + βψ, ξ) = α(φ, ξ) + β(ψ, ξ) for all α, β ∈ C
for all φ, ψ, ξ ∈ X is called an inner product on X. The space X is called an inner product
space.
Definition 2.1.10. A complete inner product space X is called a Hilbert space.
Example 2.1.11. The inner product of the space L2 [a, b] is defined by:
Z b
(φ, ψ) = φ(x)ψ(x) dx .
a
2
The space L [a, b] equipped with the inner product (·, ·) is a Hilbert space.
Theorem 2.1.12. The inner product satisfies the Cauchy-Schwarz inequality:
|(φ, ψ)|2 ≤ (φ, φ)(ψ, ψ).
Definition 2.1.13. Two elements φ and ψ of a Hilbert space are called orthogonal if
(φ, ψ) = 0 and we write φ⊥ψ. The set
U ⊥ := {ψ ∈ X : ψ⊥U }
is called the orthogonal complement of the subset U .
Theorem 2.1.14. When U is a closed subspace,
M
X=U U ⊥.
2.2. LINEAR OPERATORS 5

2.2 Linear Operators

Definition 2.2.1. An operator between two vector spaces A : X → Y is linear if
A(αφ + βψ) = αAφ + βAψ
for all φ, ψ ∈ X and α, β ∈ C.
Definition 2.2.2. A linear operator A : X → Y is bounded if there exists a positive
constant C such that
∥Aφ∥ ≤ C∥φ∥
for every φ ∈ X. The norm of A is defined by
∥A∥ := sup ∥Aφ∥, φ ∈ X.
∥φ∥=1

Theorem 2.2.3. Let X and Y be normed spaces and A : X → Y a linear operator. Then
A is continuous if it is continuous at one point.
Proof. Suppose A is continuous at φ0 ∈ X Then for every φ ∈ Xand φn → φ we have
that
Aφn = A(φn − φ + φ0 ) + A(φ − φ0 ) → Aφ0 + A(φ − φ0 ) = Aφ
since φn − φ + φ0 → φ0 .
Theorem 2.2.4. Let X and Y be normed spaces and A : X → Y a linear operator. Then
A is continuous if and only if it is bounded.
Proof. Let A : X → Y be bounded and let {φn } be a sequence in X such that φn → 0 as
n → ∞. Then ∥Aφn ∥ ≤ C∥φn ∥ implies that Aφn → 0 as n → ∞, i.e. A is continuous at
φ = 0. By Theorem 2.2.3, A is continuous for all φ ∈ X. Conversely, let A be continuous
and assume that there is no C such that ∥Aφ∥ ≤ C∥φ∥ for all φ ∈ X. Then there exists
a sequence φn with ∥φn ∥ = 1 such that ∥Aφ∥ ≥ n. Let ψn := ∥Aφn ∥−1 φn . Then ψn → 0
as n → ∞ and hence by the continuity of A we have that Aψn → Aψ0 = 0 which is a
contradiction since ∥Aψn ∥ = 1 for every integer n. Hence A must be bounded.
Example 2.2.5. Let K(x, y) be continuous on [a, b] × [a, b] and define A : L2 [a, b] →
L2 [a, b] by
Z b
(Aφ)(x) := K(x, y)φ(y) dy .
a
Then
Z b
2
∥Aφ∥ = |(Aφ)(x)|2 dx
a
Z b Z b 2
= K(x, y)φ(y) dy dx
a a
Z b Z b  Z b 
2 2
≤ |K(x, y)| dy |φ(y)| dy dx
a a a
Z bZ b
2
= ∥φ∥ |K(x, y)| dx dy .
a a
6 CHAPTER 2. FUNCTIONAL ANALYSIS

Hence A is bounded and

Z b Z b 1/2
2
∥A∥ ≤ |K(x, y)| dx dy .
a a

Theorem 2.2.6 (Riesz Representation Theorem). Let X be a Hilbert space. Then for
each bounded linear functional F : X → C there exists a unique f ∈ X such that
F (φ) = (φ, f )
for every φ ∈ X. Furthermore, ∥f ∥ = ∥F ∥.
Definition 2.2.7. Given a vector space X over a field F , the dual space X ′ is defined as
the set of all linear functionals φ : X → F .

2.3 Compact Operators

Definition 2.3.1. A subset U of a normed space X is called compact if every sequence in
U contains a subsequence that converges to an element in U . U is called relatively compact
if U is compact.
Definition 2.3.2. An operator A : X → Y is a compact operator if it maps each bounded
set in X into a relatively compact set in Y . Or equivalently, for each bounded sequence
{φn } ⊂ X, the sequence {Aφn } has a convergent subsequence in Y .
Theorem 2.3.3. Let X be a normed space and Y a Banach space. Suppose An : X → Y
is a compact operator for each integer n and there exists a linear operator A such that
∥A − An ∥ → 0 as n → ∞. Then A is a compact operator.
Proof. Let {φm } be a bounded sequence in X. We will use a diagonalization procedure to
show that {Aφm } has a convergent subsequence in Y . Since A1 is a compact operator,
{φm } has a subsequence {φ1,m } such that {A1 φ1,m } is convergent. Similarly, {φ1,m } has
a subsequence {φ2,m } such that {A2 φ2,m } is convergent. Continuing in this manner, we
see that the diagonal sequence {φm,m } is a subsequence of {φm } such that, for every fixed
positive integer n, the sequence {An φm,m } is convergent. Since {φm } is bounded, say
∥φm ∥ ≤ C for all m, ∥φm,m ∥ ≤ C for all m. We now use the fact that ∥A − An ∥ → 0 as
n → ∞ to conclude that for each ε > 0 there exists an integer n0 = n0 (ε) such that
ε
∥A − An0 ∥ <
3C
and, since {An0 φm,m } is convergent, there exists and integer N = N (ε) such that
ε
∥An0 φj,j − An0 φk,k ∥ ≤
3
for j, k > N . Hence, for j, k > N , we have that
∥Aφj,j − Aφk,k ∥ ≤ ∥Aφj,j − An0 φj,j ∥ + ∥An0 φj,j − An0 φk,k ∥ + ∥Aφk,k − An0 φk,k ∥
ε
≤ ∥A − An0 ∥∥φj,j ∥ + + ∥A − An0 ∥∥φk,k ∥ < ε.
3
Thus {Aφm,m } is a Cauchy sequence and therefore convergent in the Banach space Y .
2.4. THE ADJOINT OPERATOR 7

Example 2.3.4. Consider the operator A : L2 [a, b] → L2 [a, b] defined as in the previous
example by Z b
(Aφ)(x) := K(x, y)φ(y) dy
a

where K(x, y) is continuous on [a, b] × [a, b]. Then A is a compact operator.

Now we introduce a result from the well-known Fredholm theory.
Theorem 2.3.5 (Riesz Theorem). Let A : X → X be a compact operator on a normed
space X. Then either 1) the homogeneous equation

φ − Aφ = 0

has a nontrivial solution φ ∈ Xor 2) for each f ∈ X the equation

φ − Aφ = f

has a unique solution φ ∈ X. If I − A is injective (and hence bijective), then (I − A)−1 :

X → X is bounded.

2.4 The Adjoint Operator

Definition 2.4.1. Let X and Y be Hilbert spaces and A : X → Y be a bounded linear
operator. Then there exists a unique linear operator A∗ : Y → X such that

(Aφ, ψ) = (φ, A∗ ψ)

for every φ ∈ X and ψ ∈ Y . Then A∗ is called the adjoint of A and ∥A∥ = ∥A∗ ∥.
Theorem 2.4.2. Let X and Y be Hilbert spaces and let A : X → Y be a compact operator.
Then A∗ : Y → X is also a compact operator.
Proof. Let ∥ψn ∥ ≤ C for some positive constant C. Then, since A∗ is bounded, AA∗ :
Y → Y is a compact operator. Hence, by passing to a subsequence if necessary, we may
assume that the sequence {AA∗ ψn } converges in Y . But

∥A∗ (ψn − ψm )∥2 = (AA∗ (ψn − ψm ), ψn − ψm ) ≤ 2C∥AA∗ (ψn − ψm )∥,

i.e., {A∗ ψn } is a Cauchy sequence and hence convergent. We can now conclude that A∗ is
a compact operator.
Lemma 2.4.3. Let U be a closed subspace of a Hilbert space X. Then U ⊥⊥ = U .
L ⊥
U and X = U ⊥ U ⊥⊥ .
L
Proof. Since U is a closed subspace, we have that X = U
Hence for φ ∈ X we have that φ = φ1 + φ2 where φ1 ∈ U and φ2 ∈ U ⊥ and φ = ψ1 + ψ2
where ψ1 ∈ U ⊥⊥ and ψ2 ∈ U ⊥ . In particular, 0 = (φ1 − ψ1 ) + (φ2 − ψ2 ) and since it is
easily verified that U ⊂ U ⊥⊥ we have that φ1 − ψ1 = ψ2 − φ2 ∈ U ⊥ . But φ1 − ψ1 ∈ U ⊥⊥
hence φ1 = ψ1 . We can now conclude that U ⊥⊥ = U .
8 CHAPTER 2. FUNCTIONAL ANALYSIS

Theorem 2.4.4. Let X and Y be Hilbert spaces. Then for a bounded linear operator
A : X → Y we have that if A(X) := {y ∈ Y : y = Ax for some x ∈ X} is the range of A
then
A(X)⊥ = N (A∗ ) and N (A∗ )⊥ = A(X).
Proof. We have that g ∈ A(X)⊥ if and only if (Aφ, g) = 0 for every φ ∈ X. Since
(Aφ, g) = (φ, A∗ g) we can now conclude that A∗ g = 0, i.e. g ∈ N (A∗ ). On the other
⊥⊥ ⊥
hand, by Lemma 2.4.3, A(X) = A(X) = N (A∗ )⊥ since A(X)⊥ = A(X) = N (A∗ ).

2.5 Sobolev spaces

Suppose Ω ⊂ Rn is an open domain with sufficiently smooth boundary ∂Ω. We introduce
some basic knowledge of the Sobolev space in this chapter.
Let C k (Ω) be the set of functions on Ω whose partial derivatives up to order k exist
and are continuous. Thus C ∞ (Ω) is the set of functions on Ω with continuous any order
partial derivatives. C0∞ (Ω) is the subset of all compactly supported functions in C ∞ (Ω).
Definition 2.5.1. Let α = (α1 , . . . , αn ) be a multi-index of non-negative integers αℓ . A
locally integrable function has a weak derivative corresponding to α if there is some locally
integrable function uα such that
Z Z
|α|
α
u(D φ) dx = (−1) uα φ dx , ∀ φ ∈ C0∞ (Ω) (testfunction),
Ω Ω

where
∂ |α| φ(x)
Dα φ(x) = and |α| = |α1 | + · · · + |αn |.
∂xα1 1 · · · ∂xαnn
uα is a a weak derivative of u corresponding to α, set Dα u := uα .
Example 2.5.2. The absolute value function f (x) = |x| where x ∈ (−1, 1) is not differ-
entialble at 0. However, define

−1, −1 < x < 0;

g(x) = 0, x = 0;

1, 0 < x < 1.


Given any test function φ ∈ C0∞ (−1, 1),

Z 1 Z 0 Z 1
− g(x)φ(x) dx = φ(x) dx − φ(x) dx
−1 −1 0
Z 0 Z 1
0 ′ 1
= xφ(x) −1 − xφ (x) dx − xφ(x) 0 + xφ′ (x) dx
−1 0
Z 0 Z 1
=− xφ′ (x) dx + xφ′ (x) dx
−1 0
Z 1 Z 1
= |x|φ′ (x) dx = f (x)φ′ (x) dx
−1 −1
2.5. SOBOLEV SPACES 9

Thus Df (x) = g(x). The weak derivative exists but the strong derivative is not defined at
0.

Definition 2.5.3. The Lp -norm (where p ∈ [1, ∞)) of a meassurable function f : Rn → R

is defined by:
Z 1/p
p
∥f ∥p := |f | .
Ω

The L∞ -norm ) of f is defined by

∥f ∥∞ = ess sup|f | = inf {C ≥ 0 :, |f (x)| ≤ C for almost every x} .

Definition 2.5.4. If a measurable function f satisfies ∥f ∥p < ∞ (p ∈ [1, ∞]), then it

is p-integrable and is called an Lp -function. Lp (Ω) is the Lp -space of all Lp -functions.
Lploc (Ω), which is the space of all functions that are p-integrable in any bounded subset of
Ω, is called the local Lp -space.

Theorem 2.5.5. Let the inner product defined in Ω be defined as:

Z
(φ, ψ) = φ(x)ψ(x) dx .
Ω

1 1
The dual space of Lp (Ω) is Lq (Ω), where p, q ∈ (1, ∞) and p
+ q
= 1.

Definition 2.5.6. For any non-negative integer k and real number p ∈ [1, ∞], define the
Sobolev space W k,p (Ω) by:

W k,p (Ω) := {u ∈ Lp (Ω) : ∀|α| ≤ k, Dα u ∈ Lp (Ω)} .

The W k,p -norm (p ∈ [1, ∞)) for a function u ∈ W k,p (Ω) is defined by
 1/p  1/p
XZ X
∥u∥k,p =  |Dα u(x)|p dx  = ∥Dα u(x)∥pp  .
|α|≤k Ω |α|≤k

The W k,∞ -norm for a function u ∈ W k,∞ (Ω) is defined by

∥u∥k,∞ = max ∥Dα u∥∞ .

|α|≤k

Define the compact support Sobolev space W0k,p (Ω) by:

k,p
W0k,p (Ω) := C0∞ (Ω) .

Remark 2.5.7. The function f (x) = |x| does not belong to C 1 (−1, 1), but belongs to
W 1,p (−1, 1).
10 CHAPTER 2. FUNCTIONAL ANALYSIS

Definition 2.5.8. For any non-negative integer k and real number q ∈ (1, ∞), the space
W −k,q (Ω) is defined by
W −k,q (Ω) := {φ ∈ (C0∞ (Ω))′ : ∥φ∥W −k,q < ∞}
1 1
where p
+ q
= 1,
(φ, u)
∥φ∥W −k,q = sup ,
u∈C0∞ (Ω),u̸=0 ∥u∥W k,p (Ω)
′
and (·, ·) is defined as in Theorem 2.5.5. In particular, H −1 (Ω) = (H01 (Ω)) .
Theorem 2.5.9. We introduce some results of Sobolev embeddings in Ω ⊂ Rn :
• For any p ∈ N, when W n,p (Ω) ⊂ W m,p (Ω) where m, n ∈ N and m ≤ n.
• (Kondrachov embedding theorem) If k > ℓ and 1
p
− k
n
< 1
q
− nℓ , then W k,p (Ω) ⊂ W ℓ,q
and the embedding is compact.
• (Morrey’s inequality) Let n < p ≤ ∞ and γ = 1 − np . Then W 1,p (Ω) is continuously
embedded in C 0,γ (Ω), where C 0,γ is the Hölder continuous space with index γ.
Theorem 2.5.10. The operator T : u → u|∂Ω is continuous from W 1,p (Ω) to Lp (∂Ω). T
is called the trace operator.
In particular, denote H k (Ω) := W k,2 (Ω) and H0k (Ω) := W0k,2 (Ω). (Hilbert spaces) At
the end of this section, we introduce the Soboleve space H s [0, 2π] in the periodic domain.
n o
Definition 2.5.11. The orthonormal system √12π eimt : m ∈ Z is complete in L2 [0, 2π].
Then H s [0, 2π] is a Hilbert space with inner product
X
(φ, ψ)p := (1 + m2 )s am bm
m∈Z

where am , bm are the Fourier coefficients of φ, ψ, respectively.

Theorem 2.5.12. Some properties of the Sobolev space H s [0, 2π] is listed as follows.
• (Rellich’s Theorem) If q > p then H q [0, 2π] is dense in H p [0, 2π], and the imbedding
operator is compact.
• (Sobolev Imbedding Theorem) When p > 1/2, then any function in H p [0, 2π] equals
to a function in C[0, 2π] almost everywhere.
Next we define the Sobolev space H p (∂Ω) where Ω is a simply connected bounded
domain in R2 . Moreover, ∂Ω is C k . This means that ∂Ω has a 2π-periodic parameterization
∂Ω = {(x1 (t), x2 (t)) : t ∈ [0, 2π), x ∈ C k [0, 2π]}.
For a function φ defined on ∂Ω, it belongs to H p (∂Ω) if φ(x1 (t), x2 (t)) as a 2π-periodic
function of t, belongs to H p [0, 2π].
Theorem 2.5.13. Let Ω ⊂ Rn be a bounded Lipschitz domain. Then there is a constant
C > 0 such that
φ ∂Ω H 1/2 (∂Ω) ≤ C∥φ∥H 1 (Ω) .
Chapter 3

The Helmholtz equation

In this chapter, we will get the Helmholtz equation from both the wave equation and the
Maxwell’s equation. The Green’s theorem and radiation condition will also be introduced.

3.1 Time-harmonic acoustic waves

For detailed mathematical modelling from the physical laws we refer to Section 2.1 in
Inverse Acoustic and Electromagnetic Scattering Theory (D. Colton and R. Kress). The
velocity potential U = U (x, t) satisfies the wave equation:

1 ∂ 2U
= ∆U
c2 ∂t2
where c > 0 is the speed of sound. Notation:

∂ 2 U (x, t) ∂ 2 U (x, t) ∂ 2 U (x, t)

∆U = + +
∂x21 ∂x22 ∂x23

in R3 . For a time-harmonic wave, U has the following form:

U (x, t) = Re u(x)e−iωt .


where ω > 0 is the frequence. Let u(x) = u1 (x) + iu2 (x) where both u1 and u2 are real,
then
U (x, t) = u1 (x) cos(ωt) + u2 (x) sin(ωt).
Use the representation in the wave equation, we have:

ω2 ω2
− u1 (x) cos(ωt) − u2 (x) sin(ωt) = ∆u1 (x) cos(ωt) + ∆u2 (x) sin(ωt).
c2 c2
Let k := ω/c > 0 be the wavenumber, then

[∆u1 (x) + k 2 u1 (x)] cos(ωt) + [∆u2 (x) + k 2 u2 (x)] sin(ωt) = 0.

11
12 CHAPTER 3. THE HELMHOLTZ EQUATION

Since the above equation holds for any t > 0, it is straightforward to get the following
PDEs:

∆u1 (x) + k 2 u1 (x) = 0;

∆u2 (x) + k 2 u2 (x) = 0.

From the definition of u1 and u2 , finally we get the Helmholtz equation for u:

∆u(x) + k 2 u(x) = 0.

3.2 Time-harmonic electromagnetic waves

We are now interested in electromagnetic scattering problems. Electric field E, electric
displacement field D, magnetic H-field H and B-field B. Current intensity J , charge
density ϱ. Linear Maxwell’s equations:


 ∂t B + ∇ × E = 0 (Faraday’s Law)

∇ × H = J + ∂ D (Ampere’s Law)
t


 ∇ · (D) = ϱ (Gauss’ Electric Law)
∇ · (B) = 0

(Gauss’ Magnetic Law)

Here
  i j k
∂f ∂f ∂f ∂f1 ∂f2 ∂f3 ∂ ∂ ∂
∇f = , , , ∇·f = + + , ∇×f = ∂x1 ∂x2 ∂x3
∂x1 ∂x2 ∂x3 ∂x1 ∂x2 ∂x3
f1 f2 f3

Relationships between the fiels:

D = ε(x)E and B = µH − M,

where ε ∈ C is the permitivity, µ ∈ C is the magnetic permeability, M is magnetization

of the material.

Reduction to the Helmholtz equation. When there is no current (J = 0), no charge

(ϱ = 0) and no magnetization (M = 0), the system is reduced to two fields E and H:


 ∇ × E + µ∂t H = 0

∇ × H − ε∂ E = 0
t


 ∇·E =0
∇·H=0


For time-harmonic waves,

1 1
H(x, t) = √ H(x)e−iωt , E(x, t) = √ E(x)e−iωt .
µ ε
3.2. TIME-HARMONIC ELECTROMAGNETIC WAVES 13

Then we get the system with E and H:



 ∇ × E − ikH = 0

∇ × H + ikE = 0


 ∇·E =0
∇·H =0


√
where the wave number k = ω εµ. Both E and H are Divergence free. Take ∇× to the
first equation on both side, we get:
∇ × ∇ × E − k 2 E = 0.
Exercise. When ∇ · u = 0, we have ∇ × ∇ × u = −∆u. Finally, we show that the electric
field E satisfies the Helmholtz equation:
∆E + k 2 E = 0.
Example 3.2.1. The Helmholtz equation in one-dimensional space (ODE):
u′′ + k 2 u = 0.
There are two solutions to this second order ODE:
u1 (x) = eikx u2 (x) = e−ikx .
Related velocity potential:
U1 (x) = cos(kx − ωt) U2 (x) = cos(−kx − ωt).
Where do the two waves propagate? Try to answer by yourself !
Direct scattering problem. The lecture will be focused on the following scattering
problem. The obstacle D has a C 2 smooth boundary ∂D. It is not penetrable, and the
incident wave ui satisfies the Helmholtz equation:
∆ui + k 2 ui = 0 ∈ R2 . (3.2.1)
The incident field ui is scattered and produces the scattered field us , which satisfies
∆us + k 2 us = 0 ∈ R2 \ D. (3.2.2)
The total field u := ui + us satisfies certain boundary conditions:
• Sound soft: u = 0;
• Sound hard: ∂u
∂ν
= 0;
• Impedance: ∂u
∂ν
+ iλu = 0.
To guarantee the solution is physical, it is necessary to introduce the Sommerfeld radiation
condition:
√ ∂us
 
s
lim r − iku = 0. (3.2.3)
r→∞ ∂r
14 CHAPTER 3. THE HELMHOLTZ EQUATION

3.3 Green’s Theorem and Formula

In this section, consider the scattering problems in R3 . The Sommerfeld radiation condition
in three dimensional domain is written as:
 s 
∂u s
lim r − iku = 0. (3.3.1)
r→∞ ∂r

Lemma 3.3.1 (Divergence Theorem). Suppose D ⊂ Rn is a bounded domain and ∂D is

piecewise smooth. The vector field F ∈ Rn is continuous differentiable and satisfies
Z I
(∇ · F)dx = F · νdS,
D ∂D

where ν the unit normal vector to the boundary ∂D directed to the exterior of D.

Theorem 3.3.2. Let D be a C 1 -continuous bounded domain and ν is the normal vector
directed to the exterior of D. Suppose u ∈ C 1 (D) and v ∈ C 2 (D), then we have Green’s
first theorem: Z Z
∂v
(u∆v + ∇u · ∇v) dx = u ds. (3.3.2)
D ∂D ∂ν

If u ∈ C 2 (D), then we have Green’s second theorem:

Z Z  
∂v ∂u
(u∆v − v∆u) dx = u −v ds. (3.3.3)
D ∂D ∂ν ∂ν

Proof. i) Let F = u∇v, then

     
∂ ∂v ∂ ∂v ∂ ∂v
∇ · F = ∇ · (u∇v) = u + u + u
∂x1 ∂x1 ∂x2 ∂x2 ∂x3 ∂x3
∂u ∂v ∂ 2v
= +u 2
∂x1 ∂x1 ∂x1
∂u ∂v ∂ 2v
+ +u 2
∂x2 ∂x2 ∂x2
∂u ∂v ∂ 2v
+ +u 2
∂x3 ∂x3 ∂x3
= ∇u · ∇v + u∆v.

Then apply the divergence theorem,

Z Z I I I
∂v
(u∆v + ∇u · ∇v) dx = (∇ · F)dx = F · νdS = u∇v · νdS = u dS.
D D ∂D ∂D ∂D ∂ν

ii) Exchange the rolls of u and v, we have:

Z Z
∂u
(v∆u + ∇v · ∇u) dx = v ds.
D ∂D ∂ν
3.3. GREEN’S THEOREM AND FORMULA 15

The difference between the two equations results in:

Z Z  
∂v ∂u
(u∆v − v∆u) dx = u −v ds.
D ∂D ∂ν ∂ν

Now we aim to formulate boundary integral representations for solutions of the Helmholz
equations. First we need to introduce the fundamental solution

1 eik|x−y|
Φ(x, y) = , x ̸= y.
4π |x − y|

The fundamental solution satisfies the following equation:

∆x Φ(x, y) + k 2 Φ(x, y) = δ(x − y).

From direct computation (Exercise), it is easy to check that the fundamental solution
satisfies the Helmholtz equation when x ̸= y. When x approaches y, it becomes singular.
The delta function δ(x) is a generalied function which satisfies
Z
(δ, φ) = δ(x)φ(x)dx = φ(0).
R3

Remark 3.3.3. The fundamental solution of the Helmholtz equation in two dimensional
space is:
i (1)
Φ(x, y) = H0 (k|x − y|),
4
(1)
where H0 (r) is the Hankel function of the first kind. When |x − y| << 1,
1 1
Φ(x, y) = log + O(1).
2π |x − y|
R
Example 3.3.4. A function u(x) = R3 Φ(x, y)f (y)dy, x ∈ R3 is a solution to the
Helmholtz equation ∆u + k 2 u = f in R3 .

In the following, we will develop (boundary) integral representations for interior/exterior

Helmholtz problems. Let’s begin with the interior problems.

Theorem 3.3.5. Suppose D is bounded and C 2 , ν is the unit normal vector to the exterior
of D. Let u ∈ C 2 (D). Then
Z   Z
∂u(y) ∂Φ(x, y)
∆u(y) + k 2 u(y) Φ(x, y)dy, x ∈ D.


u(x) = Φ(x, y) − u(y) dS−
∂D ∂ν(y) ∂ν(y) D

In particular, if u satisfies the Helmholtz equation ∆u + k 2 u = 0, then

Z  
∂u(y) ∂Φ(x, y)
u(x) = Φ(x, y) − u(y) dS, x ∈ D.
∂D ∂ν(y) ∂ν(y)
16 CHAPTER 3. THE HELMHOLTZ EQUATION

Proof. Fix any point x ∈ D. Let S(x, ϱ) = ∂B(x, ϱ) where B(x, ϱ) ⊂ D is a small ball
centered at x with radius ϱ. Let Dϱ := D \ B(x, ϱ) be the domain between ∂D and S(x, ϱ),
then Φ(x, y) is smooth with respect to y ∈ Dϱ . Apply Green’s second theorem in Dϱ :
Z   Z
∂u(y) ∂Φ(x, y)
Φ(x, y) − u(y) dS = (∆u(y)Φy (x, y) − u(y)∆y Φ(x, y)) dx.
∂Dϱ ∂ν(y) ∂ν(y) Dϱ

With the fact that ∆y Φ(x, y) + k 2 Φ(x, y) = 0 in Dϱ , above equation is written as:
Z  
∂u(y) ∂Φ(x, y)
Φ(x, y) − u(y) (:= I − II)
S(x,ϱ) ∂ν(y) ∂ν(y)
Z   Z
∂u(y) ∂Φ(x, y)
∆u(y) + k 2 u(y) Φ(x, y)dx


= Φ(x, y) − u(y) dS −
∂D ∂ν(y) ∂ν(y) Dϱ
Z   Z
∂u(y) ∂Φ(x, y)
∆u(y) + k 2 u(y) Φ(x, y)dx


= Φ(x, y) − u(y) dS −
∂D ∂ν(y) ∂ν(y) D
Z
2


+ ∆u(y) + k u(y) Φ(x, y)dx(:= III).
B(x,ϱ)

First consider the limit of (III) when ϱ → ∞. Since u is C 2 -smooth, there is a constant
C > 0 such that
∆u(y) + k 2 u(y) ≤ C

holds uniformly for y ∈ B(y, ϱ). From the definition of Φ(x, y),

1
|Φ(x, y)| = .
4π|x − y|

Then Z Z
2 C 1
|(III)| ≤ ∆u(y) + k u(y) |Φ(x, y)| dy = dy.
B(x,ϱ) 4π B(x,ϱ) |x − y|
Use spherical coordinate:

y1 = x1 + r sin θ cos φ, y2 = x2 + r sin θ sin φ, y3 = x3 + r cos θ

where θ ∈ [0, π] and φ ∈ [0, 2π). Then

π 2π ϱ
Cϱ2
Z Z Z
C 1 2
|(III)| ≤ r sin θdrdφdθ = ,
4π 0 0 0 r 2

which tends to 0 as ϱ → 0+ . Thus limϱ→0+ (III) = 0.

Similar to (III), it is easy to prove that limϱ→0+ (I) = 0 (Exercise). Then we focus on
(II). From direct computation, on S(x, ϱ),

eik|x−y| (y − x)
 
1
∇y Φ(x, y) = ik − .
|x − y| 4π|x − y| |x − y|
3.3. GREEN’S THEOREM AND FORMULA 17

(y−x)
Since ν(y) = |x−y|
,

1 eikϱ
 
∂Φ(x, y)
= ∇y Φ(x, y) · ν(y) = ik − .
∂ν(y) ϱ 4πϱ

Thus with spherical coordinate again and the mean value theorem, (II) reads

1 eikϱ
Z  
(II) = u(y) ik − dS
S(x,ϱ) ϱ 4πϱ
Z π Z 2π
1 eikϱ 2
 
= u(y) ik − ϱ sin θdφdθ
0 0 ϱ 4πϱ
Z π Z 2π
1 eikϱ 2
 
= u(x) ik − ϱ sin θdφdθ
0 0 ϱ 4πϱ
Z π Z 2π
1 eikϱ 2
 
+ (u(y) − u(x)) ik − ϱ sin θdφdθ
0 0 ϱ 4πϱ
Z π Z 2π
1 eikϱ 2
 
ikϱ
= u(x)(4ikπϱ − e ) + (u(y) − u(x)) ik − ϱ sin θdφdθ.
0 0 ϱ 4πϱ

When ϱ → 0+ , the first term tends to −u(x) and the second term tends to 0 due to the
continuity of u at x. Thus limϱ→0+ (II) = −u(x). Thus finally we get the result of the
Theorem from the limits of (I), (II) and (III).

In scattering theory, we are always interested in the wave field outside a bounded
domain. Suppose u satisfies
∆u + k 2 u = 0 in R3 \ D
and satisfies the Sommefeld radiation condition (radiating waves).
Theorem 3.3.6. Let u ∈ C 2 (R3 \ D) be the radiating solution of the Helmholtz equation
outside D, then
Z  
∂Φ(x, y) ∂u(y)
u(x) = u(y) − Φ(x, y) dS, x ∈ R3 \ D.
∂D ∂ν(y) ∂ν(y)

Proof. Let B(0, R) be the ball centered at 0 with radius R > 0 and its sphere is S(0, R).
Moreover, B(0, R) is sufficiently large such that D ⊂ B(0, R). Let DR := B(0, R) \ D.
Apply the boundary integral formula for the interior of DR , then
Z   Z  
∂u(y) ∂Φ(x, y) ∂u(y) ∂Φ(x, y)
u(x) = Φ(x, y) − u(y) dS− Φ(x, y) − u(y) dS, x ∈ R3 \D.
S(0,R) ∂ν(y) ∂ν(y) ∂D ∂ν(y) ∂ν(y)

Denote the first term by (I) and estimate (I) in the following.
We already known from the Sommerfeld radiation condition:
 
∂u 1
− iku = o , R → ∞.
∂ν R
18 CHAPTER 3. THE HELMHOLTZ EQUATION

Also, for the fundamental solution Φ (when R = |x| >> y),

 
1
|Φ(x, y)| = O , R → ∞.
R

From the proof of Theorem 3.3.4,

 
∂Φ(x, y) 1
− ikΦ(x, y) = O , R → ∞.
∂ν(y) R2

We only need to show Z

|u|2 dS = O(1), R → ∞.
S(0,R)

From the Sommerfeld radiation condition and the direct calculation,

Z 2  ! Z 2
∂u 2 2 ∂u ∂u
+ k |u| + 2kIm u dS = − iku dS = o(1), R→∞
S(0,R) ∂ν ∂ν S(0,R) ∂ν

Apply the Green’s first theorem in DR (v = u), we have:

Z Z Z Z
∂u ∂u
|∇u|2 − k 2 |u|2 dx.


u dS − u dS = (u∆u + ∇u · ∇u)dx =
S(0,R) ∂ν ∂D ∂ν DR DR

Take the imaginary part,

Z  Z 
∂u ∂u
Im u dS = Im u dS ,
S(0,R) ∂ν ∂D ∂ν

where is independent of R. Thus the integral on S(0, R) is uniformly bounded w.r.t. R.

Thus Z
|u|2 dS = O(1).
S(0,R)

Now we are prepared to estimate the term (I).

(I) = (II) + (III),

where
Z  
∂u(y)
(II) = − iku(y) Φ(x, y)dS;
S(0,R) ∂ν(y)
Z  
∂Φ(x, y)
(III) = u(y) ikΦ(x, y) − dS.
S(0,R) ∂ν(y)

From previous results,

Z    
∂u(y) 2 1 1
|(II)| ≤ − iku(y) |Φ(x, y)|dS = 4πR o O → 0, R → ∞+ .
S(0,R) ∂ν(y) R R
3.4. FARFIELD PATTERN AND FURTHER PROPERTIES 19

Use Cauchy-Schwarz inequality,

!
Z  Z 2
∂Φ(x, y)
|(III)|2 ≤ |u(y)|2 dS ikΦ(x, y) − dS
S(0,R) S(0,R) ∂ν(y)
   
2 1 1
= O(1)4πR O 4
=O → 0, R → ∞+ .
R R2

Note that all the above results hold for the two dimensional spaces

3.4 Farfield pattern and further properties

Get back to the scattering problem. Given the incidnet field ui = eikx·d (plane wave). Here
d = (d1 , d2 , d3 ) ∈ R3 is a unit vector. The total field u = ui + us , where us is the scattered
field which satisfies the Sommerfeld radiation condition. Suppose we are considering the
following problem:

∆us + k 2 us = 0 in R3 \ D;
us = −ui on ∂D.

From Theorem 3.3.6, it is easy to get the integral representation of us in R3 \ D:

∂Φ(x, y) ∂us (y)

Z  
s s
u (x) = u (y) − Φ(x, y) dS, x ∈ R3 \ D.
∂D ∂ν(y) ∂ν(y)

On the other hand, since ui satisfies the Helmholtz equation globally in R3 thus obvious
in D; for fixed x ∈ R3 \ D, Φ(x, y) also satisfies the Helmholtz equation in D. Thus we
can apply equation (3.3.3) to ui and Φ(x, ·) in D:

∂Φ(x, y) ∂ui (y)

Z Z  
i i i


0= u (y)∆y Φ(x, y) − ∆u (y)Φ(x, y) dx = u (y) − Φ(x, y) dS
D ∂D ∂ν(y) ∂ν(y)

Combine the above formulas we have:

Z  
s ∂Φ(x, y) ∂u(y)
u (x) = u(y) − Φ(x, y) dS, x ∈ R3 \ D
∂D ∂ν(y) ∂ν(y)

which finally arrives the following formula with u = 0 on ∂D:

Z
s ∂u(y)
u (x) = − Φ(x, y)dS, x ∈ R3 \ D
∂D ∂ν(y)

and Z
i ∂u(y)
u(x) = u (x) − Φ(x, y)dS, x ∈ R3 \ D.
∂D ∂ν(y)
20 CHAPTER 3. THE HELMHOLTZ EQUATION

The above formula is known as Huygens’ principle (secondary sources on the boundary).

The Far Field pattern concerns the asymptotic behaviour of the solution when |x| → ∞.
From the formula, we need to consider the asymptotic behaviour of Φ(x, y) for very large
|x|. Let x = x̂|x|, then
s
p x̂ · y |y|2
|x − y| = |x|2 + |y|2 − 2|x|x̂ · y = |x| 1 − 2 + 2.
|x| |x|
√
Use the Taylor’s expansion of 1 + z at 0:
√ 1
1 + z = 1 + z + O |z|2 ,


2
 
|y|2
and let z := −2 x̂·y
|x|
+ |x| 2 , then |z| = O
1
|x|
and

|y|2
  
x̂ · y 1
|x − y| = |x| 1 − − +O
|x| 2|x|2 |x|2
    
x̂ · y 1 1
= |x| 1 − +O 2
= |x| − x̂ · y + O .
|x| |x| |x|
Use the similar argument, we can easily get when |x| >> 1:
1
eik|x−y| eik(|x|−x̂·y+O( |x| )) eik|x| −ikx̂·y
 
1
Φ(x, y) = =    = e +O .
4π|x − y| 4π |x| − x̂ · y + O |x| 1 4π|x| |x|2

Plug into the formula for us , we have:

Z
s ∂u(y)
u (x) = − Φ(x, y)dS(y)
∂D ∂ν(y)
eik|x|
Z  
∂u(y) −ikx̂·y 1
=− e dS(y) + O .
4π|x| ∂D ∂ν(y) |x|2
For a general solution us without particular boundary condition on ∂D, we have the
following estimation:

eik|x|
Z    
s ∂ −ikx̂·y ∂u(y) −ikx̂·y 1
u (x) = u(y) e − e dS(y) + O .
4π|x| ∂D ∂ν(y) ∂ν(y) |x|2
Define the Far Field pattern as:
Z  
∂ −ikx̂·y ∂u(y) −ikx̂·y
u∞ (x̂) := u(y) e − e dS(y),
∂D ∂ν(y) ∂ν(y)
then
eik|x|
 
s 1
u (x) = u∞ (x̂) + O .
4π|x| |x|2
3.4. FARFIELD PATTERN AND FURTHER PROPERTIES 21

At the end of this charpter, we will also introduce some further properties of the
scattered field and the far field pattern. Let’s begin with the two dimensional case, where
(1)
Φ(x, y) = 4i H0 (k|x − y|). It is known (F. Cakoni and D. Colton) that
(1)
X
H0 (k|x − y|) = Hn(1) (k|x|)Jn (k|y|)einθ
n∈Z

when |x| > |y| and θ = θ − θy , Jn is the Bessel function. Recall Theorem 3.3.6,
Z  
∂Φ(x, y) ∂u(y)
u(x) = u(y) − Φ(x, y) dS
∂D ∂ν(y) ∂ν(y)
Z !
i ∂ X (1) in(θ−θy ) ∂u(y) X (1) in(θ−θy )
= u(y) H (k|x|)Jn (k|y|)e − H (k|x|)Jn (k|y|)e dS
4 ∂D ∂ν(y) n∈Z n ∂ν(y) n∈Z n
Z  
i X (1) inθ ∂ −inθy ∂u(y) −inθy
= H (k|x|)e u(y) Jn (k|y|)e − Jn (k|y|)e dS
4 n∈Z n ∂D ∂ν(y) ∂ν(y)
X
:= an Hn(1) (k|x|)einθ .
n∈Z

This is the explicit structure of the outgoing wave in 2D.

In 3D, it is more complex and we will not give the details here. Following similar
approaches with the 2D case,
∞ X
X n
u(x) = am (1) m
n hn (k|x|)Yn (x̂),
n=0 m=−n

x (1,2)
where x̂ = ∥x∥ , Ynm is the spherical harmonic, hn is the spherical Hankel functions of
the first and second kind (D. Colton and R. Kress).
Remark 3.4.1. For general solutions of the exterior solution of the Helmholtz equation
in 2D, it has the following form:
X
an Hn(1) (k|x|) + bn Hn(2) (k|x|) einθ .


u(x) =
n∈Z

For outgoing waves, bn = 0 for all n ∈ Z.

Theorem 3.4.2. The solution u of the Helmholtz equation is real analytic w.r.t. its
variables.
With the expansion, we arrive at the following Rellich’s lemma.
Theorem 3.4.3 (Rellich’s Lemma). Let u ∈ C 2 (R2 \ D) be a solution of the Helmholtz
equation satisfying Z
lim |u|2 dS(x) = 0,
R→∞ |x|=R

then u = 0 in R2 \ D.
22 CHAPTER 3. THE HELMHOLTZ EQUATION

Proof. From the expansion of the solution, when |x| ≥ R, we have:

X
an Hn(1) (k|x|) + bn Hn(2) (k|x|) einθ := cn (R)einθ .


u(x) =
n∈Z

From direct computation,

Z X
lim |u|2 dS(x) = 2πR |cn (R)|2 → 0, R → ∞,
R→∞ |x|=R n∈Z

thus for any n ∈ Z,

R|cn (R)|2 → 0, R → ∞.
Recall the asymptotic behaviour of the Hankel function:
2  nπ π 
Hn(1) (kR) = + O r−3/2


exp i R − −
πR 2 4
(2) (1)
and Hn (kR) = Hn (kR), the above implies

|an |2 + |bn |2 + 2Re ((−1)n−1 an bn e2iR ) → 0, R → 0.

For any sufficiently small ε > 0, there is a Rε > 0 such that

0 ≤ |an |2 + |bn |2 + 2Re ((−1)n−1 an bn e2iR ) ≤ ε

holds uniformly for R ≥ Rε . Let R1 = nπ and R2 = (n + 1/2)π for a sufficiently large

integer n such that R1 , R2 > Rε , then

0 ≤ |an |2 + |bn |2 + 2Re ((−1)n−1 an bn e2inπ ) ≤ ε;

0 ≤ |an |2 + |bn |2 + 2Re ((−1)n−1 an bn e2inπ+iπ ) ≤ ε.

Since e2inπ = 1 and e2inπ+iπ = −1, sum up above inequalities, we have

0 ≤ 2|an |2 + 2|bn |2 ≤ 2ε.

From abitrary choice of ε > 0, we finally have an = bn = 0. Thus u = 0 for |x| ≥ R.

Since u is analytic in R2 \ D, u = 0 for |x| ≥ R implies that u = 0 in R2 \ D.

Corollary 3.4.4. Let us ∈ C 2 (R2 \ D) be the radiating solution of the Helmholtz equation.
If
∂us
Z
Im us dS ≥ 0,
∂D ∂ν
then us = 0 in R2 \ D.
3.4. FARFIELD PATTERN AND FURTHER PROPERTIES 23

Proof. Recall the proof of Theorem 3.3.6, from the Sommerfeld radiation condition:
2 !
∂us
Z  s
∂u
+ k 2 |us |2 + 2kIm us dS = o(1).
S(0,R) ∂ν ∂ν

Use the condition, we have: Z

|us |2 = 0.
S(0,R)

From Rellich’s lemma, us = 0 in R2 \ D.

24 CHAPTER 3. THE HELMHOLTZ EQUATION
Chapter 4

Boundary integral equations

In this chapter, we will establish the integral equation for the following exterior Dirichlet
boundary value problem. Given an incident field ui which is continuous on the boundary
of D, we look for the solution u ∈ C 2 (R2 \ D) ∩ C(R2 \ D) such that:
∆u + k 2 u = 0 in R2 \ D; (4.0.1)
u = ui + us ; (4.0.2)
√ ∂u
 s 
lim r − ikus = 0; (4.0.3)
r→∞ ∂r
u = 0 on ∂D. (4.0.4)
Remark 4.0.1. Since u is only continuous up to the boundary, its gradient ∇ on the
boundary ∂D is not well defined.

4.1 Uniquenss of the exterior Dirichlet problem

To study the uniquenss of the solution, we need to introduce the following lemma.
Lemma 4.1.1. Let u ∈ C 2 (R2 \ D) ∩ C(R2 \ D) be a solution of the Helmholtz equation
in R2 \ D with u = 0 on ∂D. Let R be sufficiently large, DR := {y ∈ R2 \ D : |y| < R}
and SR := {y ∈ R2 : |y| = R}. Then ∇u ∈ L2 (DR ) and
Z Z
2 2 2

∂u
|∇u| − k |u| dx = u dS(x).
DR SR ∂ν

Proof. First suppose u is real valued. Let ψ ∈ C 1 (R) be an odd function such that

0, −1 ≤ t ≤ 1;

ψ(t) = t, t > 2 or t < −2;

 1
C -continuous, otherwise.

Let un := ψ(nu)
n
for n ∈ N. It implies that un = 0 when |un | ≤ n1 , un = u when |un | ≥ n2 .
Thus ∥u − un ∥∞ → 0, when n → ∞. From the continuity of u in R2 \ D and u = 0 on
∂D, un ∈ C 1 (R2 \ D).

25
26 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

Choose a small neighbourhood of D, denoted by Dn such that un = 0 on Dn \ D. Then

u ∈ C 2 (R2 \ Dn ) and we apply the Green’s theorem in Dn :
Z Z Z
2 ∂u
∇un · ∇udx = k un udx + un dS.
DR \Dn DR \Dn SR ∂ν

Let n → ∞, since u is continuous up to the boundary, the right hand side tends to
Z Z
2 2 ∂u
k u dx + u dS.
DR SR ∂ν

Thus the left hand side is uniformly bounded. Since

∇un = ψ ′ (nu)∇u → ∇u, n → ∞.

Thus the left hand side equals to

Z
ψ ′ (nu)|∇u|2 dx
DR

and is uniformly bounded w.r.t. n. Let n → ∞, we have

Z Z Z
2 2 2 ∂u
|∇u| dx = k u dx + u dS.
DR DR SR ∂ν

The lemma is proved for real valued u.

If u is complex valued, i.e., u = v + iw, then define
ψ(nv) ψ(nw)
un = +i
n n
then similar approaches can prove the final result.

Thus we are prepared to study the uniqueness of the solution.

Theorem 4.1.2. The exterior Dirichlet problem has at most one solution.
Proof. Suppose u1 and u2 are two different solutions for the exterior Dirichlet problem.
Let u := u1 − u2 satisfies the exterior Dirichlet problem with u = 0 on ∂D. From above
lemma, Z Z
2 2 2

∂u
|∇u| − k |u| dx = u dS(x).
DR SR ∂ν

Take the imaginary part on both sides, we have:

Z
∂u
Im u dS(x) = 0.
SR ∂ν

Since u satisfies the Sommerfeld radation condition, from the corollary of Rellich’s lemma,
we have u − 0 in R2 \ D. Thus u1 = u2 .
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 27

4.2 Single- and double layer potentials

The existence of solutions will be studied via the boundary integral equations. At the very
beginning, we need to introduce the single- and double- layer potentials and operators.
For an integrable function φ, define the single- and double-layer potentials:
Z
u(x) = φ(y)Φ(x, y)dS(y), x ∈ R2 \ ∂D;
Z ∂D
∂Φ(x, y)
v(x) = φ(y) dS(y), x ∈ R2 \ ∂D.
∂D ∂ν(y)

We can also define the single- and double-layer operators:

Z
(Sφ)(x) = 2 φ(y)Φ(x, y)dS(y), x ∈ ∂D;
∂D
Z
∂Φ(x, y)
(Kφ)(x) = 2 φ(y) dS(y), x ∈ ∂D.
∂D ∂ν(y)

The normal derivatives of S and K are also defined, denoted by K ′ and T :

Z
′ ∂Φ(x, y)
(K φ)(x) = 2 φ(y)dS(y), x ∈ ∂D;
∂D ∂ν(x)
∂ 2 Φ(x, y)
Z
(T φ)(x) = 2 φ(y)dS(y), x ∈ ∂D.
∂D ∂ν(x)∂ν(y)

In this section, we review some important properties of the single- and double-layer
potentials, and the single- and layer-operators. First we have two facts:

• Both u and v satisfy the Helmholtz equation in R2 \ ∂D;

• both u and v satisfy the Sommerfeld radiation condition at the infinity.

We first recall the jump relations of the single- and double-layer potentials.

Theorem 4.2.1. Let ∂D be C 2 and φ ∈ C(∂D). Then the single-layer potential u is

continuous in R2 and
∥u∥∞,R2 ≤ C∥φ∥∞,∂D
where C only depends on ∂D. On the boundary, we have:
Z
u(x) = φ(y)Φ(x, y)dS(y), x ∈ ∂D;
∂D
Z
∂u± ∂Φ(x, y) 1
(x) = φ(y) dS(y) ∓ φ(x), x ∈ ∂D;
∂ν ∂D ∂ν(x) 2

where
∂u±
(x) := lim+ ν(x) · ∇u(x ± hν(x))
∂ν h→0
28 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

is to be understood in the sense of uniform convergence on ∂D and the integral exists as

improper integrals.

The double-layer potential v is continuously extended from D to D and from R2 \ D to

2
R \ D with Z
∂Φ(x, y) 1
v± (x) = φ(y) dS(y) ± φ(x), x ∈ ∂D,
∂D ∂ν(y) 2
where
v± (x) := lim+ v(x ± hν(x))
h→0

and the integral exists as improper integrals. Moreover,

∥v∥∞,D ≤ C∥φ∥∞,∂D , ∥v∥∞,R2 \D ≤ C∥φ∥∞,∂D

and  
∂v ∂v
lim (x + hν(x)) − (x − hν(x)) = 0, x ∈ ∂D
h→0+ ∂ν ∂ν
holds uniformly on ∂D

Now we go on with the properties of the operator s S, K.

Theorem 4.2.2. Let ∂D be of class C 2 . The operators S, K are bounded and compact
from C(∂D) to C(∂D).

Now we are prepared to apply the layer approach to consider the existence of solutions
to the exterior Dirichlet boundary problem.

Example 4.2.3. We are looking for the solution u to the exterior problem in the form of
Z  
∂Φ(x, y)
u(x) = − iηΦ(x, y) φ(y)dS(y), x ∈ R2 \ ∂D,
∂D ∂ν(y)

where φ ∈ C(∂D) is the density function and the parameter η ∈ R \ {0} is fixed. Then
from the jump relations, with u = f , we have:

(I + K − iηS)φ = 2f.

Since S and K are both compact, I + (K − iηS) is a Fredholm type opertors, which implies

• either I + (K − iηS) is bounded invertible, i.e., (I + (K − iηS))−1 is a bounded linear

operator, or

• the equation (I + (K − iηS))φ = 0 has a nontrival solution.

It implies that, to show the invertibility of the operator, it is sufficient to prove that the
uniqueness of the solutions. Thus we need to show that if f = 0, then the solution φ = 0.
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 29

Suppose f = 0 on ∂D, since u is the radiating soluton of the Helmholtz equation with
zero Dirichlet boundary condition, it implies that u = 0 in R2 \ D. Thus

∂u
u|+ = 0, =0 on ∂D.
∂ν +

From the jump relatio again,

∂u ∂u
u|+ − u|− = φ, − = iηφ.
∂ν + ∂ν −

Since u also satisfies the Helmholtz equation in D, apply the Green’s first theorem to u and
u, Z Z
2 ∂u
[|∇u| + ∆uu]dx = udS(x),
D ∂D ∂ν

which results in Z Z
2 2 2
[|∇u| − k |u| ]dx = iη |φ|dS(x).
D ∂D

Take the imaginary part of the above equation on both sides, we have φ = 0 on ∂D. The
uniqueness of the solution is proved, thus the operator I + (K − iηS) is bounded invertible.
It means that for any f ∈ C(∂D), there is a unique φ ∈ C(∂D) such that

(I + (K − iηS))φ = 2f.

Then the solution u is given by the combination of single and double layer potentials with
the density function φ.

Further properties for the singule- and double-layer potentials are listed as follows.
First we need the definition of Hölder contuity. A function φ : G → C is called α-Hölder
continuous (0 < α < 1), if there is a constant C > 0 such that

|φ(x) − φ(y)| ≤ C|x − y|

holds uniformly for x, y ∈ G. Define the norm of the Hölder continuous space C 0,α (G) as:

|φ(x) − φ(y)|
∥φ∥0,α = sup |φ(x)| + sup .
x∈G x,y∈Gx̸=y |x − y|α

Also define C 1,α (G) by the norm

∥φ∥1,α = ∥φ∥∞ + ∥∇φ∥0,α .

Theorem 4.2.4. Suppose 0 < α < β ≤ 1 and G is compact. Then the identity operators

I β : C 0,β (G) → C(G), I α,β : C 0,β (G) → C 0,α (G)

are compact.
30 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

As the extension of Theorem 4.2.1, we have the jump relations in the Höler continuous
spaces:

Theorem 4.2.5. Let ∂D be C 2 and φ ∈ C(∂D). Then the single-layer potential u is

uniformly Hölder continuous in R2 and

∥u∥α,R2 ≤ Cα ∥φ∥∞,∂D

where C only depends on ∂D. When φ ∈ C 0,α (∂D), ∇u is uniformly Hölder continuously
extended from D to D, and from R2 \ D to R2 \ D. On the boundary:
Z
1
∇u|± = φ(y)∇x Φ(x, y)dS(y) ∓ φ(x)ν(x), x ∈ ∂D;
∂D 2

where
∇u|± = lim+ ν(x)∇u(x ± hν(x))
h→0

and
∥∇u∥α,D ≤ Cα ∥φ∥α,∂D , ∥∇u∥α,R2 \D ≤ Cα ∥φ∥α,∂D .
The double-layer potential v with density φ ∈ C 0,α (∂D) is uniformly Hölder continu-
ously extended from D to D and from R2 \ D to R2 \ D with

∥v∥α,D ≤ Cα ∥φ∥α,∂D , ∥v∥α,R2 \D ≤ Cα ∥φ∥α,∂D .

When the density φ ∈ C 1,α (∂D), the first order derivative is uniformly Hölder continuously
extended from D to D and from R2 \ D to R2 \ D with

∥∇v∥α,D ≤ Cα ∥φ∥1,α,∂D , ∥v∥α,R2 \D ≤ Cα ∥φ∥1,α,∂D .

The properties of the opertors S, K, K ′ , T are also given in the Hölder continuous
spaces.

Theorem 4.2.6. Let ∂D is C 2 . Then S, K, K ′ are bounded from C(∂D) to C 0,α (∂D), and
S, K are bounded from C 0,α (∂D) to C 1,α (∂D). The operator T is bounded from C 1,α (∂D)
to C 0,α (∂D).

The properties of the operators are also extended to more general spaces.

Theorem 4.2.7. Suppose ∂D is C 2 . The operator S is bounded from L2 (∂D) to H 1 (∂D).

If ∂D is C 2,α , then K and K ′ are bounded from L2 (∂D) to H 1 (∂D), and T is bounded
from H 1 (∂D) to L2 (∂D).

Corollary 4.2.8. Suppose ∂D is C 2 . Then S is bounded from H −1/2 (∂D) to H 1/2 (∂D). If
∂D is C 2,α . Then K and K ′ are bounded from H −1/2 (∂D) to H 1/2 (∂D), and T is bounded
from H 1/2 (∂D) to H −1/2 (∂D).
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 31

Note that S and K are also compact from C 1,α (∂D) to C 1,α (∂D). We can still prove
that I + K − iηS is bounded invertible in C 1,α . when f ∈ C 1,α (∂D), the density function
φ ∈ C 1,α (∂D) exists and is unique. Thus u, which is defined by the singl- and double-
layer potentials, as well as its first order derivative, is well defined in R2 \ ∂D and Hölder
continuous up to the boundary from the interior and exterior of D. Thus we use the jump
relation again,
∂u 1
= (iηI − iηK ′ + T ) φ := Af,
∂ν + 2
1,α 0,α
where A : C (∂D) → C (∂D) is defined by

A = (iηI − iηK ′ + T ) (I + K − iηS)−1 .

It is called the Dirichlet-to-Neumann map.

Follow-up for Example 4.2.3. According to the boundary integral formulation for u
(see Section 3.4):
Z
s ∂u(y)
u (x) = − Φ(x, y)dS(y), x ∈ R2 \ D.
∂D ∂ν(y)
Why don’t we seek for the solution us in the following forms?
Z
s
u (x) = φ(y)Φ(x, y)dS(y),
∂D

or Z
s ∂Φ(x, y)
u (x) = φ(y) dS(y),
∂D ∂ν(y)
Let’s take the first one as anRexample. Suppose the solution u in Example 4.2.3 we seek
for is in the form of u(x) = 2 ∂D φ(y)Φ(x, y)dS(y), then from the boundary condition, we
have:
(I + S)φ = f.
We need to prove that I + S is bounded invertible. When f ∈ C(∂D), S is a compact
operator from C(∂D) to itself thus we will prove the invertibility by proving the uniqueness.
Let φ ∈ C(∂D) be a solution of (I + S)φ = 0. Since f = 0, we know that u = 0 in R2 \ D.
Then u|+ = ∂u∂ν
= 0 on ∂D. We extend the solution to D with the same formula, then
+
from the jump relations,
∂u ∂u
u|− = u|+ = 0, = + 2φ = 2φ.
∂ν − ∂ν +

Thus u is the solution to the following equation:

∂u
∆u + k 2 u = 0 D, u = 0, = 2φ on ∂D.
∂ν
If u ̸= 0 solves ∆u + k 2 u = 0 in D with u = 0 on ∂D, then it is a Dirichlet eigenfunction,
with λ := k 2 a Dirichlet eigenvalue.
32 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

For a fixed bounded domain D, the Dirichlet eigenvalues are real, positive, and have no
limit point. Thus they can be arranged in increasing order:

0 < λ1 ≤ λ2 ≤ λ3 ≤ · · · , λn → ∞.

When k 2 happends to be a Dirichlet eigenvalue λn , then let u ̸= 0 be the eigenfunction

related to λn , the function
1 ∂u
φ :=
2 ∂ν
solves the equation (I + S)φ = 0. Now we need to show that φ ̸= 0. If φ = 0, then u is
the solution to the Helmholtz equation with u = ∂u ∂ν
= 0. Thus u is a global solution of
the Helmholtz equation in R . Since u = 0 outside D and it is real analytic, u = 0 in R2 .
2

This contradicts with u ̸= 0 in D. Thus φ ̸= 0 on ∂D. Now we find a nontrivial solution

φ to the equation (I + S)φ = 0, so I + S is not invertible.
Example 4.2.9. When D is a disk B(0, R), we can even find out the eigenvalues explicity.
We are looking for the eigenfunctions in the form of

un (r, θ) = yn (kr)einθ , n ∈ Z.

Then y = yn (r) is a solution of the Bessel’s equatino:

n2
 
′′ 1 ′
y + y + 1 − 2 y = 0.
r r
The solutions are Bessel functions Jn (r) (Section 3.2, Cakoni and Colton 2006). Thus the
solutions have the form of un (r, θ) = Jn (kr)einθ . Thus for any n, there are a sequence

Figure 4.1: Bessel functions of the first kind. (mathworks.com)

(m) (m)
of zeros rn (m = 1, 2, 3, . . . and n ∈ Z). When kR = rn , un (r, θ) = Jn (kr)einθ is a
Dirichlet eigenfunction.
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 33

In the following, we will apply the boundary integral equations to other problems.

Example 4.2.10. Consider the exterior Neumann problem. We study the unique solvabil-
ity of the radiating solution in C 2 (R2 \ D) ∩ C(R2 \ D) of

∂u
∆u + k 2 u = 0 in R2 \ D, = g on ∂D,
∂ν
where g ∈ C(∂D) is the Neumann data.

Proof. The uniqueness comes from Corollary 3.4.4. Note that the solution is only con-
tinuious up to the boundary, limiting processes are needed (exercise).
We seed for the solution in the form of
Z  
2 ∂Φ(x, y)
u(x) = φ(y)Φ(x, y) + iη(S0 φ)(y) dS(y), x ∈ R2 \ D,
∂D ∂ν(y)

where φ ∈ C(∂D). Here S0 is defined by the fundamental solution Φ0 (x, y) with wavenum-
ber k = 0: (
1 1
− 2π ln |x−y| , x, y ∈ R2 ;
Φ0 (x, y) = 1
4π|x−y|
, x, y ∈ R3 .
Thus they are real valued functions.
Why we don’t use the form from Example 4.2.3 directly? When the density function
φ ∈ C(∂D), the normal direvative on ∂D of the double layer potential is not well defined.
Thus we replace φ by the term S02 φ. When φ ∈ C(∂D), S0 φ ∈ C 0,α (∂D) and S02 φ ∈
C 1,α (∂D). According to Theorem 4.2.5, the normal derivative of the double layer potential
with density function S02 φ ∈ C 1,α (∂D) is well defined and Hölder continuous up to the
boundary.
With the jump relations and the boundary condition, we have:

(I − K ′ − iηT S02 )φ = −2g.

Recall that K ′ is bounded from C(∂D) to C 0,α (∂D) thus is compact from C(∂D) to
C(∂D). Since S02 is bounded from C(∂D) to C 0,α (∂D) and T is bounded from C 1,α (∂D)
to C 0,α (∂D), T S02 is bounded from C(∂D) to C 0,α (∂D) thus compact from C(∂D) to
C(∂D). Then K ′ + iηT S02 is comact from C(∂D) to C(∂D). Thus I − K ′ − iηT S02 is a
Fredholm operator, we need to prove its invertibiliy by proving the unique solutions.
Let g = 0, from the uniqueness of the exterior Neumann problem, u = 0 in R2 \ 0.
Extend u to D. Then we apply the jump relations:

∂u ∂u
u|+ − u|− = iηS02 φ, − = −φ,
∂ν + ∂ν −

thus u is a solution to the Helmholtz equation in D with boundary conditions

∂u
u|− = −iηS02 φ, = φ.
∂ν −
34 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

Apply the first Green’s theorem to the interiour problem,

Z Z Z Z
2 2 2 ∂u
(∇u · ∇u + u∆u)dx = (|∇u| − k |u| )dx = u dS(y) = iη φS02 φdS(y).
D D ∂ ∂ν ∂D

Here we need to use some properties of S0 . From the definition,

Z Z
2
(S0 φ)(x) = 4 Φ0 (x, y) Φ0 (y, z)φ(z)dS(z)dS(y).
∂D ∂D

Since Φ0 (x, y) is real valued, S02 φ = S02 φ. Thus

Z Z Z Z
2
φS0 φdS(y) = φ(x) Φ0 (x, y) Φ0 (y, z)φ(z)dS(z)dS(y)dS(x)
∂D Z∂D ∂D ∂D

= |(S0 φ)(y)|2 dS(y).

∂D

This implies that

Z Z
2 2 2
(|∇u| − k |u| )dx = iη |(S0 φ)(y)|2 dS(y).
D ∂D

Still we take the imaginary part, it results in S0 φ = 0 on ∂D. Let

Z
v(x) = 2 Φ0 (x, y)φ(y)dS(y),
∂D

then v ∈ C(R2 ) and satisfies

∆v = 0 in R2 \ ∂D, v = 0 on ∂D.

Note here v is a harmonic function and tends to 0 at the infinity. From the maximum-
minimum principle for harmonic functions, a harmonic function reaches the maximum and
minimum values on the boundary, thus v = 0 in R3 . From the jump relations φ = 0. Thus
the uniqueness is proved, which results in the invertibility of I − K ′ − T S02 . The proof is
finished.
When g ∈ C 0,α (∂D), we can get φ ∈ C 0,α (∂D) from the same argument of the Dirichlet
problem. Thus the solution u belongs to C 1,α (R2 \ D) and on the boundary,
1
iηS02 + iηKS02 + s φ.


u|+ =
2
Define the operator B : C 0,α (∂D) → C 1,α (∂D) by

−1
B := iηS02 + iηKS02 + s K ′ − I + iηT S02


,

we have u = Bg and it is called a Neumann to Dirichlet operator. Thus A is the inverse

operator of B.
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 35

Example 4.2.11. Consider the exterior impedance problem. We study the unique solv-
ability of the radiating solution in C 2 (R2 \ D) ∩ C 1 (R2 \ D) of

∂u
∆u + k 2 u = 0 in R2 \ D, + iλu = h on ∂D,
∂ν
where h ∈ C(∂D) is the boundary data, λ ∈ C(∂D) and it is positive.

Proof. Uniqueness is easily proved (exercise). We are looking for the solution in the form
of the single-layer potential:
Z
u(x) = φ(y)Φ(x, y)dS(y), x ∈ R2 \ D.
∂D

From the boundary condition on ∂D,

(I − K ′ − iλS)φ = 2h.

Thus we want to prove the existence of the above equation. Recall that for φ ∈ C(∂D),
S, K ′ are compact thus it is a Fredholm operator. Again we only need to prove that h = 0
implies φ = 0.
When h = 0, from the uniqueness, u = 0 in R2 \ D. Thus

∂u
u|+ = =0 on ∂D.
∂ν +

Extend the definition of u into the domain D,

∂u ∂u
u|− = u|+ = 0, = +φ=φ on ∂D.
∂ν − ∂ν +

Thus u|D is the soluton of the Helmholtz equation in D with u = 0. If k 2 is a Dirichlet

eigenvalue, and let u be the corresponding nontrival eigenfunction. Then φ = ∂u
∂ν
̸= 0.
−
Recall u is the single-layer potential with the density function φ, it is a solution to the
Helmholtz eqution with u = 0 on ∂D. Thus u = 0 in R2 \ D due to the uniqueness
of the exterior Dirichlet boundary condition. Thus on the boundary, u|+ = ∂u ∂ν
= 0
+
which implies that it satisfies the homogeneous impedance boundary condition. Thus φ
is a nontrival solution to the equation thus we cannot prove the existence through the
uniqueness.
To overcome this difficulty, we modify the single-layer potential by defining a new
kernel:
Γ(x, y) = Φ(x, y) + X (x, y)
where
iX
X (x, y) := an Hn(1) (kr)Hn(1) (kry )ein(θ−θy )
4 n∈Z
36 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

with the polar coorinates (r, θ) for x and (ry , θy ) for y. The coefficients an are chosen such
that the series converges for |x|, |y| > R for some R > 0 such that B(0, R) ⊂ D. Moreover,
the function X (x, y) satisfies
n o
∆x X (x, y) + k 2 X (x, y) = 0 in R2 \ B(0, R) ∪ {y} .

Then we seek for the solution in a modifed single-layer potential:

Z n o
u(x) := φ(y)Γ(x, y)dS(y), x ∈ R2 B(0, R) ∪ ∂D
∂D

with the same jump relations. Reason: the function X (x, y) is analytic w.r.t. x and y when
x, y ∈ R2 \ B(0, R).
R Thus when y ∈ ∂D, the function is smooth w.r.t. x ∈ R2 \ B(0, R)
thus the integral ∂D Γ(x, y)φ(y)dS(y) is C ∞ continuous across the boundary. Then we
get the equation  
′
I − K − iλS φ = 2h
e e

and consider its unique solvability.

From h = 0, we have still u = 0 in R2 \ D. From the jump relations, we have:

∂u
u|− = 0, =φ on ∂D.
∂ν −

Extend the definition of u into D \ B(0, R), and we known u = 0 on ∂D. Find two positive
numbers R1 , R2 such that R < R1 < R2 such that B(0, R2 ) ⊂ D. Recall that
(1)
X
H0 (k|x − y|) = Jn (kr)Hn(1) (kyr )ein(θ−θy ) , |x| < |y|,
n∈Z

the function u has the following expansion in B(0, R2 ) \ B(0, R):

Z !
iX iX
u(x) = an Hn(1) (kr)Hn(1) (kry )ein(θ−θy ) + Jn (kr)Hn(1) (kyr )ein(θ−θy ) φ(y)dS(y).
∂D 4 n∈Z 4 n∈Z

Finally we get the simplified form:

X
bn Jn (kr) + an Hn(1) (kr) einθ .


u(x) =
n∈Z

Apply Green’s second theorem to u and u,

Z
0= (u∆u − u∆u)dx
D\B(0,R)
Z   Z  
∂u ∂u ∂u ∂u
= u|− − u|− dS − u −u dS,
∂D ∂ν − ∂ν − ∂B(0,R1 ) ∂ν ∂ν
4.2. SINGLE- AND DOUBLE LAYER POTENTIALS 37

thus
Z  
∂u ∂u
0= u −u dS
∂B(0,R1 ) ∂ν ∂ν

X
2

d  (1)

= 2π |bn | Jn (kR) + an Hn(1) (kR) Jn (kR) + an Hn (kR)
n∈Z
dr
 d 
(1) (1)


− Jn (kR) + an Hn (kR) Jn (kR) + an Hn (kR)
dr
We take any n and compute the term directly:

Jn (kR) + an Hn(1) (kR) Jn′ (kR) + an (Hn(2) )′ (kR)

− Jn (kR) + an Hn(2) (kR) Jn′ (kR) + an (Hn(1) )′ (kR)

=Jn (kR)Jn′ (kR) + an Jn (kR)(Hn(2) )′ (kR) + an Hn(1) (kR)Jn′ (kR) + |an |2 Hn(1) (kR)(Hn(2) )′ (kR)
−Jn (kR)Jn′ (kR) − an Jn′ (kR)Hn(2) (kR) − an (Hn(1) )′ (kR)Jn (kR) − |an |2 Hn(2) (kR)(Hn(1) )′ (kR)
 
(1) (1)
=an Jn (kR)(Hn )′ (kR) − Jn′ (kR)Hn (kR)
+an Hn(1) (kR)Jn′ (kR) − (Hn(1) )′ (kR)Jn (kR)

+|an |2 Hn(1) (kR)(Hn(2) )′ (kR) − Hn(2) (kR)(Hn(1) )′ (kR) .

Recall the Wronskian relations. The Wronskian is defined by:

y1 y2
W (y1 , y2 ) := .
y1′ y2′

And we know
2i 2i 4i
W (Jn (r), Hn(1) (r)) = , W (Hn(1) (r), Hn(2) (r)) = =− ,
πr πr πr
then the above equation equals to
(1)
an W (Jn (r), Hn (r)) − an W (Jn (r), Hn(1) (r)) + |an |2 W (Hn(1) (r), Hn(2) (r))
2i
= − (an + an + 2|an |2 )
πr
Ignoring the coefficients, we fianlly get
X X
0= |bn |2 (2an + 2an + 4|an |2 ) = |bn |2 (|2an + 1|2 − 1).
n∈Z n∈Z

Note that the choice of an is only related to the convergence of the series to define X (x, y).
By choosing positive an , we have bn = 0 thus u = 0 when R1 < |x| < R2 . Since u is real
analytic in D \ B(0, R), it is also 0 in this domain. Then φ = ∂u∂ν
= 0, which proves the
−
uniquenss. Thus I − Ke ′ − iλSe is bounded invertible and the existence of the solution φ
for any given h ∈ C(∂D) is proved. This finishes the proof.
38 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

4.3 Numerical implementation

Take the Dirichlet problem as an example, we are looking for the numerical solution for:

(I + K − iηS)φ = 2f,

where
Z Z
∂
(Sφ)(x) = 2 Φ(x, y)φ(y)dS(y), (Kφ)(x) = 2 Φ(x, y)φ(y)dS(y).
∂D ∂D ∂ν(y)

Parameterization of the curve ∂D:

x(t) = (x1 (t), x2 (t))

in counterclockwise orientation satisfying |x′ (t)| > 0, and x1 , x2 are both 2π-periodic ana-
lytic functions. Replace x by x(t), y by x(τ ),

i 2π (1)
Z p
(Sφ)(x(t)) = H0 (k|x(t) − x(τ )|2 )φ(x(τ )) (x′1 (τ ))2 + (x′2 (τ ))2 dτ.
2 0
Simplify the representation, we get
Z 2π
(Sφ)(t) = M (t, τ )ψ(τ )dτ
0

and the single layer operator with a similiar technique:

Z 2π
(Kφ)(t) = − L(t, τ )ψ(τ )dτ
0

so finally we get
Z 2π Z 2π
ψ(t) − L(t, τ )ψ(τ )dτ − iη M (t, τ )ψ(τ )dτ = g(t).
0 0

From the singularities of the Hankel functions, we get directly:

 
2 t−τ
L(t, τ ) = L1 (t, τ ) ln 4 sin + L2 (t, τ ),
2
 
2 t−τ
M (t, τ ) = M1 (t, τ ) ln 4 sin + M2 (t, τ )
2

where L1 , L2 , M1 and M2 are analytic functions. For details of the formulas we refer to
Chapter 3.5 in Colton and Kress 2013.
Finally, we simplify the equation as
Z 2π   Z 2π
2 t−τ
ψ(t) − ln 4 sin K1 (t, τ )ψ(τ )dτ − K2 (t, τ )ψ(τ )dτ = g(t).
0 2 0
4.3. NUMERICAL IMPLEMENTATION 39

The discretization of the second integral is particularly difficult due to the singularity. To
deal with this, we will adopt the Nyström method. We start with
Z 2π Z 2π  
2 t−τ
If := f (τ )dτ and SIf := ln 4 sin f (τ )dτ.
0 0 2

The main idea is to approximate the integrand f by a finite Fourier series and take the
advantage of the fast convergence of the Fourier series for analytic functions. Let
N
X −1
fN (τ ) := fˆn einτ
n=−N

be the finite Fourier series. Discretize the domain [0, 2π] uniformlly, then

(n − 1)(π)
0 = t1 < t2 < · · · = t2N = 2π(1 − 1/2N ), tn = .
N
To carry out the trignometric interpolation, first define
N −1
(m) 1 X in(t−tm )
φ2N (t) := e , m = 1, 2, . . . , 2N,
2N n=−N

(m) R 2π (m) π
then φ2N (tℓ ) = δℓ,m and 0
φ2N (t)dt = N
. With this definition, the finite Fourier series
has the alternative form:
2N
X
fN (τ ) = f (tm )φm
2N (τ ).
m=1

Replace the integrand f by fN :

Z 2N
2π X 2N
π X
If ≈ IN f = f (tm )φm
2N (τ )dτ = f (tm ),
0 m=1
N m=1

and
2N
2π X   2N
2 t−τ
Z X
m (m)
ISf ≈ ISN f = ln 4 sin f (tm )φ2N (τ )dτ = R2N (t)f (tm )
0 m=1
2 m=1

where
2N −1
(m) π X 1 π
R2N (t) = − cos m(t − tm ) − cos 2N (t − tm ), j = 1, 2, . . . , 2N.
N m=1 m 4N 2

With this formula,

Z 2π 2N
π X
K2 (t, τ )ψ(τ )dτ ≈ K2 (t, tm )f (tm )
0 N m=1
40 CHAPTER 4. BOUNDARY INTEGRAL EQUATIONS

and
2π   2N
2 t−τ
Z X (m)
ln 4 sin K1 (t, τ )ψ(τ )dτ ≈ K1 (t, tm )R2N (t)f (tm ).
0 2 m=1

With fixed t, the equation becomes

2N 2N
π X X (m)
ψ(t) − K2 (t, tm )f (tm ) − K1 (t, tm )R2N (t)f (tm ) = g(t).
N m=1 m=1

By taking t = tℓ (ℓ = 1, 2, . . . , 2N ), we get a full matrix where the l-th row is:

2N 2N
π X X (m)
ψ(tℓ ) − K2 (tℓ , tm )f (tm ) − K1 (tℓ , tm )R2N (tℓ )f (tm ) = g(tℓ ).
N m=1 m=1

By solving this linear equation, we finally solves the Fredholm equation numerically.
Chapter 5

Inverse problems

5.1 Ill-posed problems

The introduction begins with Hadamard’s concept of well-posedness.

Definition 5.1.1. Let A : U ⊂ X → V ⊂ Y be an operator where U is a subset of a

normed space X, V is a subset of a normed space Y . The eqution

Aφ = f

is called well-posed if A : U → V is bijective and the inverse A−1 : V → U is continuous.

Otherwise the equation is called ill-posed.

Thus for an ill-posed problem, the at least one of the following three statements is
satisfied.

• When A is not surjective, it means for some f ∈ V , there is no φ ∈ U such that

Aφ = f . Thus the existence of the solution is not satisfied. (nonexistence)

• When A is not injective, it means there is one f ∈ V corresponds to at least two

φ1 ̸= φ2 such that Aφ1 = Aφ2 = f . Thus the solution is not unique. (nonuniqueness)

• When A−1 is not continuous, it means the solution φ does not depend continuously
on f , this is called (instable).

The following will be an easy example for ill-posed problems (matrices).

Example 5.1.2. Consider the equation Ax = b where A is an n × n-matrix and b is an

n × 1-vector. Known A and x to compute b is the direct/forward problem, which is easy.
The inverse problems have different settings, the straightforward one is, known A and b,
how to get x? The problem is ill-posed when

• When det(A) = 0, then A as an operator is neither surjective nor injective. Then

for some b ∈ Rn , Ax = b does not have a solution (nonexistence); for some other
b ∈ Rn , Ax = b has multiple solutions (nonuniqueness). In this case, the problem is
ill-posed.

41
42 CHAPTER 5. INVERSE PROBLEMS

• When matrix A is invertible but ill conditioned, which means the condition number

κ(A) = ∥A∥∥A−1 ∥ >> 1.

Then the problem is ill-posed.

Example 5.1.3. The matrix A is defined as follows:
   
1 1 −1 10001 −10000
A= with A = .
1 1.0001 −10000 10000

The condition number

κ(A) = ∥A∥1 ∥A−1 ∥1 = 2.0001 × 20001 = 40004.0001

which is large. Thus A is an ill-conditioned matrix. Let

   
2 2
b= results in x = .
2 0

When we make a small change to b:

   
2 1
bδ = results in xδ = .
2.0001 1

Thus a small perturbation in b results in a large difference in x, which is called unstable.

Example 5.1.4. Consider the initial-boundary value problem
∂u 2
∂t
= ∂∂xu2 in [0, π] × [0, T ];
u(0, t) = u(π, t) = 0, 0 ≤ t ≤ T ;
u(x, 0) = φ(x), 0 ≤ x ≤ π,

where φ ∈ C[0, π] is known. With the separation of variables,

∞
2
X
u(x, t) = an e−n t sin(nx)
n=1

where the coefficients an are given by:

Z π
2
an = φ(y) sin(ny)dy.
π 0

The existence, uniqueness and stability of solutions are clear.

Now consider the inverse problems. Suppose φ is no longer known but we know f (x) :=
u(x, T ) for all x ∈ [0, π] and some large enough T . Then the aim is to find the unknown
φ from the known f . Since f (0) = f (π) = 0, f is expanded by the sin functions:
∞
2 π
X Z
f (x) = bn sin(nx), where bn = f (y) sin(ny)dy.
n=1
π 0
5.2. REGULARIZATION 43

From the form of the solution,

∞ ∞
−n2 (t−T ) 2T
X X
u(x, t) = bn e sin(nx) thus φ(x) = bn e n sin(nx).
n=1 n=1

Check the L2 -norm of φ:

∞
2X 2 2
∥φ∥ = |bn |2 e2n T ,
π n=1
which blows up when bn does not decay sufficiently fast. A small perturbation in f (i.e.,
bn ) will result in a siginificant difference in φ. The inverse problem is ill-posed.
We can even reformulate the problem in terms of the integral equaiton. Let
∞
2 X −n2 T
K(x, y) = e sin(nx) sin(ny), 0 ≤ x, y ≤ π.
π n=1

Then the problem is written as:

Z π
K(x, y)φ(y)dy = f (x).
0

Define Z π
(Aφ)(x) := K(x, y)φ(y)dy,
0
then it is a bounded linear operator in L2 [0, π]. Moreover, it is compact.
Theorem 5.1.5. Let X and Y be normed spaces and A : X → Y be a compact operator.
Then Aφ = f is ill-posed if X is not of finite dimension.
Proof. Assume A−1 exists and is continuous. Then I := A−1 A :, X → X is compact,
which is impossible in infinite dimensional spaces.

5.2 Regularization
To solve an ill-posed problem, the regularization methods are deveoped. We want to solve
the equation Aφ = f but the right hand side f is perturbed with a known error level
∥f − f δ ∥ ≤ δ.
The aim of the regularization is to construct a reasonable approximation φδ to φ. First
we need the defintion for the regularization.
Definition 5.2.1. Let X and Y be normed spaces and A : X → Y be an injective bounded
linear operator. Then a family of bounded linear operators Rα : Y → X, α > 0, such that
lim Rα Aφ = φ
α→0

for every φ ∈ X is called a regularization scheme for A. The parameter α is called the
regularization parameter.
44 CHAPTER 5. INVERSE PROBLEMS

From the definition, Rα f → A−1 f as α → 0 for each f ∈ A(X). However, a uniform

convergence is not possible.

Theorem 5.2.2. Let X and Y be normed spaces, A : X → Y is an injective compact

operator and x has infinite dimension. Then Rα connot be uniformly bounded with respect
to α as α → 0 and Rα A can not be norm convergent as α → 0.

Proof. First assume ∥Rα ∥ ≤ C as α → 0. Then since Rα f → A−1 f for each f ∈ A(X),
∥A−1 ∥ ≤ C as an operator from A(X) to X. Thus I = A−1 A is bounded and compact on
X, which contradicts with the fact that X is inifinite dimensional.
Now assume that Rα A is norm convergent, which means ∥Rα A − I∥ → 0 as α → 0.
Then there exists one α > 0 such that ∥Rα A − I∥ < 1/2 and then

∥A−1 f ∥ = ∥A−1 f − Rα AA−1 f + Rα f ∥

≤ ∥A−1 f − Rα AA−1 f ∥ + ∥Rα f ∥
≤ ∥I − Rα A∥∥A−1 f ∥ + ∥Rα f ∥
≤ 1/2∥A−1 f ∥ + ∥Rα f ∥.

Thus ∥A−1 f ∥ ≤ 2∥Rα ∥∥f ∥, which implies A−1 is bounded. This leads again to a contra-
diction.
The aim of the regularization scheme is to approximate the solution φ of Aφ = f by

φδα := Rα f δ .

We want to estimate the error thus

φδα − φ = Rα f δ − Rα f + Rα Aφ − φ,

then
∥φδα − φ∥ ≤ δ∥Rα ∥ + ∥Rα Aφ − φ∥.
It is difficult, since when α → 0, ∥Rα ∥ is very large but ∥Rα Aφ − φ∥ is small; while
when α > 1, the first term is small but the second is large. So the choice of α is the key
question. A reasonable strategy is to find a proper δ-dependent parameter α = α(δ) such
that φδα → φ when δ → 0.

Definition 5.2.3. A strategy for a regularization scheme Rα (α > 0), i.e., a method for
choosing the regularization paremeter α = α(δ), is called regular if for every f ∈ A(X)
and all f δ ∈ Y such that ∥f δ − f ∥ ≤ δ we have

Rα(δ) f δ → A−1 f, as δ → 0.

5.3 Singular value decomposition

Let A : X → Y be a compact operator between two infinite dimensional Hilbert spaces.
The operator A∗ A is self-adjoint and compact, thus by the Hilbert-Schimidt theorem,
5.3. SINGULAR VALUE DECOMPOSITION 45

there exist a countable set of eigenvalues {λn : n = 1, 2, . . . } and eigenfunctions {φn : n =

1, 2, . . . } such that A∗ Aφn = λn φn . Thus
(A∗ Aφn , φn ) = λn ∥φn ∥2 ≥ 0
then λn ≥ 0. Then the square roots of λn are called the singular values of A.
Theorem 5.3.1. Let {µn : n = 1, 2, . . . } be the sequence of nonzero singular values of the
compact operator A such that
µ1 ≥ µ2 ≥ µ3 ≥ · · · .
Then there exist orthonormal sequences {φn : n = 1, 2, . . . } ⊂ X and {gn : n =
1, 2, . . . } ⊂ Y such that
Aφn = µn gn , A∗ gn = µn φn .
For every φ ∈ X we have the singular value decompostion
∞
X
φ= (φ, φn )φn + P φ
n=1

where P : X → N (A) is the orthogonal projection operator from X to N (A) and

∞
X
Aφ = µn (φ, φn )gn .
n=1

The system (µn , φn , gn ) is called a singular system of A.

Proof. Let {φn : n = 1, 2, . . . } be the orthonormal eigenelements of A∗ A w.r.t. {µn : n =
1, 2, . . . }, i.e.,
A∗ Aφn = µ2n φn .
Define
1
gn := Aφn ,
µn
then
1 ∗
Aφn = µn gn , A∗ gn = A Aφn = µn φn .
µn
From the projection,
∞
X
φ= (φ, φn )φn + P φ
n=1
where P : X → N (A A) is the orthogonal projection of X onto N (A∗ A). Now we show
∗

that N (A∗ A) = N (A).

First if ψ ∈ N (A) then Aψ = 0. Thus A∗ Aψ = 0 which implies N (A) ⊂ N (A∗ A).
Second if ψ ∈ N (A∗ A) then A∗ Aψ = 0. Then (A∗ Aψ, ψ) = ∥Aψ ∥2 = 0. Thus ψ ∈ N (A)
then N (A∗ A) ⊂ N (A). Finally N (A) = N (A∗ A).
Apply A to the expansion,
∞
X ∞
X
Aφ = (φ, φn )Aφn = µn (φ, φn )gn .
n=1 n=1
46 CHAPTER 5. INVERSE PROBLEMS

Note that from above decompostions, we also have:

∞
X
∥φ∥2 = ∥(φ, φn )∥2 + ∥P φ∥2 ; (5.3.1)
n=1
∞
X
2
∥Aφ∥ = µ2n ∥(φ, φn )∥2 . (5.3.2)
n=1

Based on the singular value decomposition, the solution to the problem Aφ = f is

concluded in the following theorem. Before that, a lemma is necessary.
Lemma 5.3.2. For a bounded linear operator A, we have

A(X)⊥ = N (A∗ ) and N (A∗ )⊥ = A(X).

Proof. • First we show A(X)⊥ = N (A∗ ). Suppose g ∈ A(X)⊥ , it means that (Aφ, g) =
0 for all φ ∈ X. Thus (φ, A∗ g) = 0, which implies that A∗ g = 0 thus g ∈ N (A∗ ).
This implies A(X)⊥ ⊂ N (A∗ ).
Suppose g ∈ N (A∗ ), then A∗ g = 0 thus (φ, A∗ g) = 0 for all φ ∈ X. Thus (Aφ, g) = 0
which implies g ∈ A(X)⊥ . Then N (A∗ ) ⊂ A(X)⊥ .

• Then we show N (A∗ )⊥ = A(X). It comes directly from N (A∗ )⊥ = A(X)⊥ ⊥ = A(X).

Theorem 5.3.3 (Picard). Let A : X → Y be a compact linear operator with singular

system (νn , φn , gn ). Then the equation is solvable if and only if f ∈ N (A∗ )⊥ and satisfies
∞
X 1
2
|(f, gn )|2 < ∞. (5.3.3)
µ
n=1 n

In this case a solution is given by

∞
X 1
φ= (f, gn )φn .
n=1
µ n

Proof. If the problem is solvable, then f ∈ A(X) ⊂ N (A∗ )⊥ . Moreover,

(f, gn ) = (Aφ, gn ) = (φ, A∗ gn ) = µn (φ, φn ).

From (5.3.1),
∞ ∞
X 1 2
X
2
∥(f, gn )∥ = ∥(φ, φn )∥2 < ∞.
µ
n=1 n n=1

The necessity is then proved.

Now assume that f ∈ N (A∗ )⊥ and (5.3.3) is fullfilled. Let
∞
X 1
ψN := (f, gn )φn .
n=1
µ n
5.3. SINGULAR VALUE DECOMPOSITION 47

Then ∥ψN ∥ is uniformly bounded thus the sequence converges. Apply A to ψN we get:
N N
X 1 X
AψN = (f, gn )Aφn = (f, gn )gn .
n=1
µ n n=1
P∞ 1
Then ∥AψN ∥ is also uniformly bounded thus converges. Then let n → ∞, φ = n=1 µn (f, gn )φn
solve the problem Aφ = f .
The Picard’s theorem explains the nature of the ill-posed problems. Suppose f ∈
N (A∗ ). Let f δ := f + δgn for some large n and a non-zero value δ. Then it still lies in
N (A∗ ) with the convergent result (5.3.3). From direct computation, the solution φδ has
the form of ∞
δ
X 1 δ
φ = (f, gn )φn + φn .
µ
n=1 n
µn
Thus
∥φ − φδ ∥ 1
=
∥f − f δ ∥ µn
can be very large since µn → 0 as n → ∞. Moreover, if the convergence is slow, then the
problem is slightly ill-posed; if the convergence is fast, then it is severely ill-posed.

Example 5.3.4. Consider the inverse scattering problem. Let the scatterer be a disk D
centered at 0 and with radius R > 0. The problem is modelled by:

∆u + k 2 u = 0 in D; u = φ on ∂D,

and it satisfies the Sommerfeld radiation condition. The problem is to reconstruct the
boundary data φ ∈ L2 (∂D) from the far-field pattern u∞ (θ).
First we rewrite φ by the Fourier series:
X
φ(θ) = an einθ , θ ∈ [0, 2π).
n∈Z

Then the radiating solution u has the form of

(1)
X Hn (k|x|)
u(x) = an (1)
einθ , θ ∈ [0, 2π).
n∈Z Hn (kR)

From the asymptotic behaviour of the Hankel functions,

r
2 −iπ/4 X 1
u∞ (θ) = e (−1)n an (1) einθ .
π n∈Z Hn (kR)

Thus A which maps φ to u∞ has the singular values

r
(kR)n
 
2 1
µn = =O
π |Hn(1) (kR)| 2n (n − 1)!
48 CHAPTER 5. INVERSE PROBLEMS

From Stirling’s approximation,

√  n n
n! ∼ 2πn ,
e
then  n
ekR
µn = o
2n
which converges to 0 very rapidly. Thus it is a severely ill-posed problem.

Now we introduce an abstract description of the regularization scheme based on the

singular value decomposition.

Theorem 5.3.5. Let A : X → Y be an injective compact linear operator with singular

system (µn , φn , gn ), n ∈ N, and let q : (0, ∞) × (0, ∥A∥] → R be a bounded function such
that for each α > 0 there exists a positive constant c(α) with

|q(α, µ)| ≤ c(α)µ, 0 < µ ≤ ∥A∥,

and
lim q(α, µ) = 1, 0 < µ ≤ ∥A∥.
α→0

Then the bounded linear operators Rα : Y → X, α > 0 defined by

∞
X 1
Rα f := q(α, µn )(f, gn )φn , f ∈Y
n=1
µ n

describes a regularization scheme with ∥Rα ∥ ≤ c(α).

Proof. From direct computation,

∞ ∞
2
X 1 2 2 2
X
∥Rα f ∥ = 2
|q(α, µn )| |(f, gn )| ≤ c (α) |(f, gn )|2 = c2 (α)∥f ∥2 ,
µ
n=1 n n=1

thus ∥Rα ∥ ≤ c(α). Next we need to check that

lim Rα Aφ = φ.
α→0

From direct computation,

∞ ∞
X 1 X
Rα Aφ − φ = q(α, µn )(Aφ, gn )φn − (φ, φn )φn
µ
n=1 n n=1
∞  
X 1 ∗
= q(α, µn )(φ, A gn ) − (φ, φn ) φn
n=1
µ n
∞
X
= [q(α, µn ) − 1](φ, φn )φn .
n=1
5.3. SINGULAR VALUE DECOMPOSITION 49

Then ∞
X
2
∥Rα Aφ − φ∥ = |q(α, µn ) − 1|2 |(φ, φn )|2 .
n=1
Next we will prove that the above series converge to 0 as α → 0. For any fixed φ ∈ X,
given ε > 0, there is a large positive integer N such that
∞
X ε
|(φ, φn )|2 < ,
n=N +1
2(M + 1)2

where |q(α, µ)| ≤ M holds. Then

∞
X ε
|q(α, µn ) − 1|2 |(φ, φn )|2 ≤ .
n=N +1
2

For the first N terms, there is an α0 := α0 (ε) > 0 such that for any 0 < α < α0 ,
ε
|q(α, µn ) − 1|2 ≤
2∥φ∥2
holds for n = 1, 2, . . . , N . Thus
∞ N
X
2 2 ε X ε
|q(α, µn ) − 1| |(φ, φn )| ≤ 2
|(φ, φn )|2 + ≤ ε.
n=1
2∥φ∥ n=1 2

From the above theorem, a regularization scheme can be developed by choosing a

proper filter/damping function q.
Theorem 5.3.6. Let A : X → Y be an injective compact linear operator with singular
system (µn , φn , gn ), n ∈ N. Then the spectral cut-off
X 1
Rm f := (f, gn )φn
µ ≥µ
µ n
n m

1
describeds a regularization scheme with regularization parameter m → ∞ and Rm = µm
.
Proof. From the definition, the function q satisfies q(m, µ) = 1 when µ ≥ µm and q(m, µ) =
0 for µ < µm . Then
X 1 1 X 1
∥Rm f ∥2 = 2
|(f, g n )|2
≤ 2
|(f, gn )|2 ≤ 2 ∥f ∥2 .
µ ≥µ
µn µm µ ≥µ µm
n m n m

1 1 1
Thus ∥Rm ∥ ≤ µm
. Let f = gm , then Rm f = φ .
µm m
Thus ∥Rm ∥ = µm
.
The regularization parameter here is m instead of α, which determines the terms in
the sum. To make the approximation accurate, m should be large; on the other hand, it
becomes unstable when m is large. So we need a criteria to choose a proper m, which is
explained in the following discrepancy principle.
50 CHAPTER 5. INVERSE PROBLEMS

Theorem 5.3.7. Let A : X → Y be an injective compact linear operator with dense range
in Y . Let f ∈ Y and δ > 0. Then there is a smallest integer m such that

∥ARm f − f ∥ ≤ δ.

Proof. The range of A is dense thus N (A∗ )⊥ = A(X) = Y . Thus A∗ is injective. Decom-
pose f as
∞
X
f= (f, gn )gn
n=1

then ∞
X 1 X X
ARm f − f = (f, gn )Aφn − (f, gn )gn = (f, gn )gn .
µn ≥µm
µn n=1 µ <µ
n m

When m → ∞, it comes directly that ∥ARm f − f ∥ → 0. Thus for any δ > 0, there is a
smallest integer m = m(δ) such that ∥ARm f − f ∥ ≤ δ.

5.4 Tikhonov regularization

The SVD based regularization is not easily carried out - the SVD is difficult and time-
consuming itself. To this end, we need further techniques which are efficient and conve-
nient. In this section, we will learn the Tikhonov regularization, which is much easier to
implement.
Still consider the problem we want to solve:

Aφ = f

where A : X → Y is a compact linear operator between two Hilbert spaces X and Y .

Then A∗ is also compact and we want to solve the following problem instead:

A∗ Aφ = A∗ f.

The advantage to the second equation is that, the operator A∗ A is self-adjoit and can be
diagonalized. It is always a benefit to solve this problem. However, when A is compact,
A∗ A is also compact and ill-posedness gets worse. The idea for the Tikhonov regularization
is to solve the slightly perturbed problem

(αI + A∗ A)φα = A∗ f, (5.4.1)

then
Rα := (αI + A∗ A)−1 A∗ f.
In this case, the eigenvalues for A∗ A are not smaller than α > 0, so the ill-posedness is
not so severe.
Theorem 5.4.1. Let A : X → Y be a compact linear operator. Then for each α > 0, the
operator αI + A∗ A is bijective and has a bounded inverse. Further more, if A is injective
then Rα describes a regularization scheme with ∥Rα ∥ ≤ 2√1 α .
5.4. TIKHONOV REGULARIZATION 51

Proof. First we prove the injectivity. Suppose (αI +A∗ A)φ = 0. From direct computation,

0 = ((αI + A∗ A)φ, φ) = α∥φ∥2 + ∥Aφ∥2 ≥ α∥φ∥2 .

This implies φ = 0 thus the operator is an injection.

Then prove the surjectivity, i.e., for any ψ ∈ X, we look for a solution (αI +A∗ A)ξ = ψ.
Recall the singular system (µn , φn , gn ), the element ξ and ψ are decomposed as
∞
X ∞
X
ξ= (ξ, φn )φn + P ξ, ψ= (ψ, φn )φn + P ψ
n=1 n=1

where P : X → N (A) is the orthogonal projection. Then

∞
X X∞
∗ ∗ ∗
A Aξ = A ( (ξ, φn )Aφn + AP ξ) = A ( µn (ξ, φn )gn )
n=1 n=1
∞
X ∞
X
= µn (ξ, φn )A∗ gn = µ2n (ξ, φn )φn .
n=1 n=1

Thus ∞
X
∗
(αI + A A)ξ = (α + µ2n )(ξ, φn )φn + αP ξ.
n=1

From (αI + A∗ A)ξ = ψ, we have:

∞
X 1 1
ξ= 2
(ψ, φn )φn + P ψ,
n=1
α + µn α

then it solves the problem. The surjectivity is also proved.

When A ins injective, then N (A) = {0} and X = span{φn : n = 1, 2, . . . }. Any
element ψ ∈ X is written as
X∞
ψ= (ψ, φn )φn .
n=1

Consider
(αI + A∗ A)φα = A∗ f.
Then (A∗ f, φn ) = (f, Aφn ) = µn (f, gn ) thus
∞
X
∗
Af= µn (f, gn )φn
n=1

Then the solution ∞

X µn
φ= (f, gn )φn .
n=1
α + µ2n
Thus the function is chosen as:
µ2
q(α, µ) = .
α + µ2
52 CHAPTER 5. INVERSE PROBLEMS

It is clear that 0 < q(α, µ) < 1 for (α, µ) ∈ (0, ∞) × (0, ∞) and limα→0 q(α, µ) = 1.
Moreover, since √
α + µ2 ≥ 2 αµ,
we have:
µ2 µ2 µ
q(α, µ) = 2
≤ √ = √ .
α+µ 2 αµ 2 α
1
Let c(α) = √
2 α
, then we get the norm of Rα .

The Tikhonov regularization has also another interpretation. This is given in the
following theorem.

Theorem 5.4.2. Let A : X → Y be a compact linear operator and α > 0. Then for each
f ∈ Y there exists a unique φα ∈ X such that

∥Aφα − f ∥2 + α∥φα ∥2 = inf ∥Aφ − f ∥2 + α∥φ∥2 .


φ∈X

The minimizer is given by the Tikhonov regularization and depends continuously on f .

Proof. We compare the functions on both side:

∥Aφ − f ∥2 + α∥φ∥2 = ∥Aφα − f ∥2 + α∥φα ∥2

+ 2Re (φ − φα , αφα + A∗ (Aφα − f ))
+ ∥A(φ − φα )∥2 + α∥φ − φα ∥2

which is valid for all φ ∈ X.

If the condition (5.4.1) is satisfied, then the third term on the right hand side vanishes.
Thus obviously φα is a minimizer of the problem. So the condition is sufficient.
If φα is a minimizer, then it is required that

2Re (φ − φα , αφα + A∗ (Aφα − f )) + ∥A(φ − φα )∥2 + α∥φ − φα ∥2 ≥ 0

for all φ. Let ψ := αφα + A∗ (Aφα − f ) and define

φ := φα + tψ

with a parameter t ∈ R. Then

2t∥ψ∥2 + t2 ∥Aψ∥2 + αt2 ∥ψ∥2 ≥ 0 for all t ∈ R.

Let t = − ∥A∥12 +α , then

1
0 ≤ 2t∥ψ∥2 + t2 ∥Aψ∥2 + αt2 ∥ψ∥2 = − ∥ψ∥2
∥A∥2 +α

thus ψ = 0 which implies the condition (5.4.1).

Thus the minimization problem is equivalent to the Tikhonov regularization.
5.4. TIKHONOV REGULARIZATION 53

From the interpretation, the problem is a modification to the problem, i.e., to find the
minimizer of the residual
min ∥Aφ − f ∥2
φ∈X

(least square). The penalty term α∥φα ∥2 is to strenthen the stability.

The regularity of the Tikhonov regularization is concluded in the following theorem.

Theorem 5.4.3. Let A : X → Y be an injective compact linear operator with dense range
in Y . Let f ∈ A(X) and f δ ∈ Y satisfy

∥f δ − f ∥ ≤ δ < ∥f δ ∥

where δ > 0 is the error level. Then there exists a unique parameter α = α(δ) such that

ARα(δ) f δ − f δ = δ

and
Rα(δ) f δ → A−1 f, δ → 0.

Proof. For the first argument, we will show the function F : (0, ∞) → R defined by
2
F (α) := ARα(δ) f δ − f δ − δ2

has a unique zero. Since A(X) = Y = N (A∗ )b ot, A∗ is an injection, f δ is decomposed into
∞
X
δ
f = (f δ , gn )gn .
n=1

Then we have:
∞ X µ2∞
δ
X µn n
ARα f = (f, g n )Aφn = (f, gn )gn .
n=1
α + µ2n n=1
α + µ2n

Then
∞
X α2
F (α) = − δ2.
n=1
(α + µ2n )2

Since when α → 0, F (α) → 0 and when α → ∞, F (α) → ∥f δ ∥ − δ 2 > 0, the function has
a unique zero which determines α(δ).
Then we prove the convergence. From direct computation,
2
φδ − A−1 f = ∥φδ ∥2 − 2Re φδ , A−1 f + ∥A−1 f ∥2 .

To prove the above formula tends to 0, we need to prove the following statements: ∥φδ ∥ ≤
∥A−1 f ∥ and φδ weakly convergent to A−1 f .
54 CHAPTER 5. INVERSE PROBLEMS

To prove the first inequality, recall that φδ := Rα(δ) f δ minimizes the Tikhonov func-
tional. Thus

δ 2 + α∥φδ ∥2 = ∥Aφδ − f δ ∥2 α∥φδ ∥2

≤ ∥AA−1 f − f δ ∥2 + α∥A−1 f ∥2
≤ δ 2 + α∥A−1 f ∥2

thus ∥φδ ∥ ≤ ∥A−1 f ∥.

To prove the weak convergence, we estimate

(Aφδ − f, g) ≤ ∥Aφδ − f δ ∥ + ∥f δ − f ∥ ∥g∥ ≤ 2δ∥g∥ → 0,

δ → 0.

Thus Aφδ ⇀ f . Since A is injective, it is invertible in A(X) thus φδ ⇀ A−1 f .

With these statements, we are prepared to prove the final convergence result.

∥φδ − A−1 f ∥2 ≤ 2 ∥A−1 f ∥2 − Re (φδ − A−1 f ) = 2Re A−1 f − φδ , A−1 f

which converges to 0 due to the weak convergence. The proof is finished.

Chapter 6

Inverse scattering problems

In this chapter, we focus on the first kind of inverse scattering problems, i.e., the inverse
obstacle problems.
Recall the problems we have discussed. Given an incident field, now required to be
plane waves:
ui (x) = eix·d ,
where d = (cos θ, sin θ) is the direction of propagation of the incident wave. Then ui
satisfies
∆ui + k 2 ui = 0 in R2 .
Given an impenetrable bounded obstacle D ⊂ R2 (also in R3 ), the incident is scattered
by D and produces the scattered field us in the exteriour of D which also satisfies the
Helmholtz equation
∆us + k 2 u2 = 0 in R2 \ D,
with the Sommerfeld radiation condition:
√ ∂us
 
s
lim r − iku = 0.
r→∞ ∂r
The boundary condition is given in the form of Bu = 0 on ∂D. The following types are
named after the physical properties:
• Sound-soft, i.e., u = 0 and us = −ui ;
∂us i
• Sound-hard, i.e., ∂u
∂ν
= 0 and ∂ν
= − ∂u
∂ν
;
∂us i
• Impedance, i.e., ∂u
∂ν
+ iλu = 0 and ∂ν
+ iλus = − ∂u
∂ν
− iλui .
The direct problems are, when we know the incident field ui and the obstacle as well
as the boundary condition, then we want to solve the equation to find the scattered field
us . The inverse problems are the opposite, when we no longer know the obstacle, or the
boundary condition, or neither. But we can do the experiment and measure the scattered
field us (near-field data) or its far field pattern u∞ , and then to predict the unknown
obstacle/boundary condition from the measurement. In this chapter, the measure ments
are fixed to be the far field patterns. Explicity, there are the following three kinds of
inverse problems:

55
56 CHAPTER 6. INVERSE SCATTERING PROBLEMS

• Given u∞ and B, determine D.

• Given u∞ and D, determine B.

• Given u∞ , determine D and B.

We will first introduce the uniqueness results for the inverse obstacle problems, and
then introduce different methods to solve these problems.

6.1 Uniqueness of the inverse obstacle problems

In this section, we consider the uniqueness of the inverse problems, i.e., in which condition
the obstacles/boundary conditions are uniquely determined by the measurements.
We start with the known sound-soft boundary conditions. Suppose D1 and D2 are
two sound soft obstacles, when are they identical to each other? Before that, we need the
follow lemma.

Lemma 6.1.1. Suppose u∞,1 (x̂) and u∞,2 (x̂) are the far-field patterns of two scattered
waves us1 and u2s , which satisfy the Helmholtz equation in the exterior of D1 and D2 . If
u∞,1 (x̂) = u∞,2 (x̂), then
us1 = us2 in R2 \ (D1 ∪ D2 ).

Proof. Let w := us1 − us2 in R2 \ (D1 ∪ D2 ), then it is real-analytic. Let B(0, R) be a

sufficiently large ball such that D1 ∪D2 ⊂ B(0, R). Then when r ≥ R, w has the expansion
in terms of the Hankel functions of the first kind:
′
X
w(r, θ) = an Hn(1) (kr)ein theta .
n∈Z

Then the far-field pattern is given by the asymptotic behaviour of the Hankel functions:
r
2 −iπ/4 X
0 = u∞,1 (x̂) − u∞,2 (x̂) = w∞ (θ) = e (−1)n an einθ .
π n∈Z

This implies that an = 0 for all n ∈ Z, thus w = 0 when r ≥ R. From the real analyticity
of w in R2 \ (D1 ∪ D2 ), w = 0 which implies us1 = us2 in R2 \ (D1 ∪ D2 ).

We still need an additional formula, i.e., the Jacobi-Anger expansion:

∞
X
eiz cos θ = in Jn (z)einθ .
n=−∞

On the other hand,

Z 2π Z 2π ∞
X
iz cos θ
e dθ = in Jn (z)einθ dθ = 2πJ0 (z).
0 0 n=−∞
6.1. UNIQUENESS OF THE INVERSE OBSTACLE PROBLEMS 57

Theorem 6.1.2. Assume D1 and D2 are two sound soft scatteres and the wave number
k > 0 is fixed. If the far field pattern coincide for incident plane waves with all directions
d, then D1 = D2 .

Proof. We will prove by contradiction. Suppose for a given ui = eix·d , the related scattered
fields be us1 (x; d) and us2 (x; d) w.r.t D1 and D2 . And the far field patterns be u∞,1 (x̂; d)
and u∞,2 (x̂; d). Then from u∞,1 (x̂; d) = u∞,2 (x̂; d) we have us1 (x; d) = us2 (x; d) for x ∈
R2 \ (D1 ∪ D2 ) and infinite number of d.
Without loss of generalicity, suppose D∗ := D1 \ D2 ̸= ∅. The boundary of D∗ is
composed of two parts: one part ∂D1 \ D2 and the other part ∂D2 ∩ D1 . Consider the
scattered field us2 and the related total field u2 . Since D∗ ⊂ R2 \ D2 , u2 satisfies

∆u2 + k 2 u2 = 0 in D∗ .

On ∂D2 ∩ D1 , u2 = 0 from the boundary condition; on ∂D1 \ D2 , since us1 = us2 and u1 = 0,
we have u2 = us2 + ui = us1 + ui = u1 = 0. Thus u2 satisfies

u2 = 0 on ∂D∗ .

It means that u2 is a Dirichlet eigenfunction in D∗ and k 2 is a Dirichlet eigenvalue of −∆

in D∗ .
From the condition, there are infinite number of d’s which means there are infinite
number of eigenfunctions. On the other hand, for a bounded domain D∗ , there are only
finitely many linearly independent Dirichlet eigenfunctions in H01 (D∗ ). We will establish
the contradiction by proving that the solutions u2 (x, d) are linearly independent with
different d’s. To this end, we need to prove the following two facts:

• The solutions u2 (x, d) are linearly independent with different d’s.

• There are only finitely many linearly independent Dirichlet eigenfunctions in H01 (D∗ )

Assume that we have the following argument:

N
X
cn u2 (·, dn ) = 0
n=1

for N distinct incident directions dn where n = 1, 2, . . . , N . The above relationship is

expanded to R2 \ D from the analyticity of the functions. Since us2 (x, dn ) = O(|x|−1 ), we
have:
N  
X
ikx·dn 1
cn e =O √ , R → ∞.
n=1 R
We want to apply the Jacobi-Anger formula, thus multiply the left hand side by e−ikx·dm
(for any m = 1, 2, . . . , N ) and integral on the circle |x| = R,
N Z  
1X ikx(dn −dm ) 1
cn e ds(x) = O √ .
R n=1 |x|=R R
58 CHAPTER 6. INVERSE SCATTERING PROBLEMS

Let x = R(cos θ, sin θ) and dn − dm = |dn − dm |(cos φn , sin φn ), then when n ̸= m,

Z Z 2π
ikxdn
e ds(x) = exp(ikR|dn − dm |(cos θ cos φn + sin θ sin φn ))Rdθ
|x|=R 0
Z 2π
=R exp(ikR|dn − dm | cos(θ − φn ))dθ = 2πRJ0 (kR|dn − dm |).
0

When n = m, the above integral equals to 2πR. Thus Then

 
X 1
2πcm + 2πJ0 (kR) cn = O √ .
n̸=m
R
 
This implies that cm = O √1R . Let R → ∞, cm = 0 for all m = 1, 2, . . . , N . Thus the
functions u2 (·, dn ) where n = 1, 2, . . . , N are linearly independent.
Then we assume that there are infinitely many linear independent eigenfunctions {un }
in H01 (D∗ ). We can require that
Z
un um dx = δnm .
D∗

Then it is easily deduced that

Z Z
2 2
|∇un | dx = k |un |2 dx = k 2
D∗ D∗
√
thus ∥un ∥H01 (D∗ ) = 1 + k 2 for all n ∈ N. It means that {un : n ∈ N} is a uniformly
bounded set. Note that the compact embedding H01 (D∗ ) ⊂ L2 (D∗ ) implies a convergent
subsequence unk , still denoted by un , in the norm of L2 (D∗ ). Thus it implies that for any
ε > 0, there is a sufficiently large integer N such that for any m, n > N ,

∥um − un ∥L2 (D∗ ) ≤ ε.

On the other hand,

∥um − un ∥2L2 (D∗ ) = (um , um ) − (um , un ) − (un , um ) + (un , un ) = 2

which contradicts with the convergece result. Thus there are only finitely many linear
independent eigenfunctions.
With above two results, it is impossible to have a non-empty set D∗ defined in this
way. Thus D∗ = D1 ⊂ D2 = ∅. Similarly we can also prove D2 ⊂ D1 = ∅. This implies
that D1 = D2 . The proof is finished.
The following results will be listed with only brief ideas of proofs, or even without
proofs.
Theorem 6.1.3. Suppose D1 , D2 ⊂ B(0, R), then the far field patterns from N different
incident plane waves uniquely determines the sound soft obstacle, when N is sufficiently
large.
6.1. UNIQUENESS OF THE INVERSE OBSTACLE PROBLEMS 59

Proof. We need to estimate the number of eigenfunctions related to the eigenvalue k 2 .

Since D∗ ⊂ B(0, R), it is known that the eigenvalues of −∆ is always smaller in the larger
domain. This is, if the eigenvalues in D∗ are

λ1 ≤ λ2 ≤ · · · ≤ λm = k 2

and in B(0, R), these are

µ1 ≤ µ2 ≤ · · · ≤ µm < λm = k 2 .

So the dimension of the eigenspace related to the eigenvalue k 2 in D∗ is bounded by the

number of eigenvalues of B(0, R) which are smaller or equal to k 2 .
The eigenfunctions in B(0, R) have the form of
√
Jn ( µm |x|)einθ , n ∈ Z,
√
thus we need to find out all the zeros of Jn ( µm R). This number determines N .

When D1 , D2 are sufficiently small, even one incident field determines the domains.

Corollary 6.1.4. Suppose D1 , D2 ⊂ B(0, R) when kR < π, the the domain D1 = D2 if

the far field patterns coincide for one incident field.

Theorem 6.1.5. A sound-soft ball is uniquely determined by the far field pattern from
one incident field.

Theorem 6.1.6. A sound-soft convex polyhedron is uniquely determined by the far field
pattern from one incident field.

Now we move on to the general boundary condition Bu = 0. The next question is, how
many incident plane waves can uniquely determine both the scatterer D and the boundary
condition B? Before that, we need the following reciprocity and mixed reciprocity relations.
Let ui (x; d) = eikx·d is the plane wave with the direction d ∈ S. The scattered field is
denoted by us (x; d), and the total field u(x; d) and u∞ (x̂; d).

Lemma 6.1.7. The far field patterns for the same boundary condition satisfy:

u∞ (x̂; d) = u∞ (−d, −x̂), x̂, d ∈ S.

Proof. Recall that the radiating solutions have the boundary integral representations:

∂us (y; d)
Z  
s s ∂Φ(x, y)
u (x; d) = u (y; d) − Φ(x, y) ds(y)
∂D ∂ν(y) ∂ν(y)

and
∂us (y; −x̂)
Z  
s s ∂Φ(x, y)
u (x; −x̂) = u (y; −x̂) − Φ(x, y) ds(y)
∂D ∂ν(y) ∂ν(y)
60 CHAPTER 6. INVERSE SCATTERING PROBLEMS

From the far field pattern of Φ(x, y),

∂e−ikx̂·y s
Z  
1 s −ikx̂·y ∂u (y; d)
u∞ (x̂; d) = u (y; d) −e ds(y)
c0 ∂D ∂ν(y) ∂ν(y)
and
∂eikd·y s
 
iky·d ∂u (y; −x̂)
Z
1 s
u∞ (−d; −x̂) = u (y; −x̂) −e ds(y)
c0 ∂D ∂ν(y) ∂ν(y)
√
8πk
where c0 := eiπ/4
.
From the Green’s theorem, we have the following equations:

∂ui (y; −x̂) ∂ui (y; d)

Z  
i i
u (y; d) − u (y; −x̂) ds(y) = 0
∂D ∂ν(y) ∂ν(y)
and
∂us (y; −x̂) ∂us (y; d)
Z  
s s
u (y; d) − u (y; −x̂) ds(y) = 0
∂D ∂ν(y) ∂ν(y)
Then

c0 u∞ (x̂; d) − c0 u∞ (−d; −x̂)

∂ui (y; −x̂) ∂us (y; d)
Z  
s i
= u (y; d) − u (y; −x̂) ds(y)
∂D ∂ν(y) ∂ν(y)
∂ui (y; d) ∂us (y; −x̂)
Z  
s i
− u (y; −x̂) − u (y; d) ds(y)
∂D ∂ν(y) ∂ν(y)
∂ui (y; −x̂)
Z  
i ∂u(y; d)
= u(y; d) − u (y; −x̂) ds(y)
∂D ∂ν(y) ∂ν(y)
∂us (y; −x̂)
Z  
s ∂u(y; d)
− u (y; −x̂) − u(y; d) ds(y)
∂D ∂ν(y) ∂ν(y)
Z  
∂u(y; −x̂) ∂u(y; d)
= u(y; d) − u(y; −x̂) ds(y).
∂D ∂ν(y) ∂ν(y)
It is easily checked that the above equation equals to 0 with the same boundary condition.
The proof is finished.
Let wi (x, z) := Φ(x, z) be a point source located at z ∈ R2 \ D, then the scattered field
is denoted by ws (x, z) and the total field w(x, z). Let the far field pattern is w∞
s
(b
x, z).
Lemma 6.1.8. For obstacle scattering of point sources and plane waves we have the fol-
lowing mixed reciprocity relation:
s
c0 w ∞ (−d; z) = us (z; d), z ∈ R2 \ D, d ∈ S.

Proof. The integral representation of w∞ (x̂; z) is given by:

∂eikd·y s
Z  
s ikd·y ∂w (y; z)
c0 w∞ (−d; z) = w (y; z) −e ds(y).
∂D ∂ν(y) ∂ν(y)
6.1. UNIQUENESS OF THE INVERSE OBSTACLE PROBLEMS 61

Similarly, we have:
∂us (y; d)
Z  
s s ∂Φ(x, y)
u (z; d) = u (y; d) − Φ(x, y) ds(y)
∂D ∂ν(y) ∂ν(y)
From the Green’s theorem:
∂ui (y; d) ∂wi (y; z)
Z  
i i
w (y; z) − u (y; d) ds(y) = 0
∂D ∂ν(y) ∂ν(y)
and
∂us (y; d) ∂ws (y; z)
Z  
s s
w (y; z) − u (y; d) ds(y) = 0
∂D ∂ν(y) ∂ν(y)
Then
s
c0 w ∞ (−d; z) − us (z; d)
∂ui (y; d) ∂ws (y; z)
Z  
s i
= w (y; z) − u (y; d) ds(y)
∂D ∂ν(y) ∂ν(y)
∂wi (y; x) ∂us (y; d)
Z  
s i
− u (y; d) − w (y; x) ds(y)
∂D ∂ν(y) ∂ν(y)
∂ui (y; d)
Z  
i ∂w(y; z)
= w(y; z) − u (y; d) ds(y)
∂D ∂ν(y) ∂ν(y)
∂us (y; d)
Z  
s ∂w(y; x)
− u (y; d) − w(y; x) ds(y)
∂D ∂ν(y) ∂ν(y)
Z  
∂u(y; d) ∂w(y; z)
= w(y; z) − u(y; d) ds(y) = 0.
∂D ∂ν(y) ∂ν(y)
The proof is finished.
Theorem 6.1.9. Assume D1 and D2 are two scatterers with boundary conditions B1 and
B2 . When the far field patterns coincide for all d ∈ S and a fixed wavenumber, then
D1 = D2 and B1 = B2 .
Proof. Since u∞,1 (x̂; d) = u∞,2 (x̂; d), with Lemma 6.1.1, us1 (z; d) = us (z; d) for all x ∈
R2 \ (D1 ∪ D2 ). From Lemma 6.1.8, w∞,1 (−d; z) = w∞,2 (−d, z) for all d ∈ S and z ∈
R2 \ (D1 ∪ D2 ). Thus ws (x; z) = w( x; z) for all x, z ∈ R2 \ (D1 ∪ D2 ).
Assume that D1 ̸= D2 . Without loss of generality, there is an x∗ ∈ ∂D1 \ D2 . Let µ(x∗ )
be the unit normal vector at the point x∗ directed to the exterior of D1 ∪ D2 , then
1
zn := x∗ + µ(x∗ ) ∈ R2 \ (D1 ∪ D2 ), n = 1, 2, . . . ,
n
for sufficiently large n. Then let wni := (x; zn ) and it is related to w1s (x; zn ) and w2s (x; zn ).
From above arguments, w1s (x; zn ) = w2s (x; zn ) for all x ∈ R2 \ (D1 ∪ D2 ).
First focus on w2s (x; zn ). Since dist(zn , D2 ) ≥ C > 1, wni (·, zn ) is uniformly bounded on
∂D2 . Thus
lim B1 w2s (x∗ , zn ) = B1 w2s (x∗ , x∗ )
n→∞
62 CHAPTER 6. INVERSE SCATTERING PROBLEMS

is bounded.
On the other hand, the scattered field w1s (x; zn ) satisfies the boundary condition
B1 w1s (x∗ ; zn ) = −B1 wi (x∗ ; zn ).
Take n → ∞, we have
lim B1 w1s (x∗ ; zn ) = ∞,
n→∞
which contradicts with the fact that w1s (x∗ ; zn ) = w2s (x∗ ; zn ). Thus D1 = D2 .
Now we turn to B1 = B2 with the unique domain D. Assume that the boundary
conditions are:
∂u
B1 u = + ikλ1 u = 0
∂ν
and
∂u
B2 u = + ikλ2 u = 0.
∂ν
Assume λ1 ̸= λ2 . Suppose u1 and u2 are two total fields with one incident field ui and
two boundary conditions B1 and B2 . Since u1∞ = u2∞ , us1 = us2 when x ∈ R2 \ D. Thus
u1 = u2 := u in R2 \ D. Thus
(λ1 − λ2 )u = 0 on ∂D.
It implies that u = 0 on ∂D thus ∂u
∂ν
= 0. It results in that u = 0 thus ui = −us satisfies
the Sommerfeld radiation condition. However, when ui (x) = eikx·d , it does not satisfy the
Sommerfeld radiation codition. So λ1 = λ2 .

Example 6.1.10. Suppose D = B(0, R) where R is the unknown radius, which we will
reconstruct. The incident field is given by the superposition of the incident plane waves:
Z
v (x) = eikx·d ds(d) = 2π|x|J0 (k|x|).
i
S
s
The scattered field v is given by the superposition of the scattered fields
Z
v (x) = us (x; d)ds(d)
s
S
and the far field pattern Z
v∞ (x) = u∞ (x; d)ds(d).
S
When the boundary is sound soft, then
v i (x) + v s (x) = 0
which gives
(1)
2πRJ0 (kR) + cH0 (kR) = 0
thus
2πRJ0 (kR) (1)
v s (x) = − (1)
H0 (kx)
H0 (kR)
and
eiπ/4 2πRJ0 (kR)
v∞ (x̂) = − √
8kπ H0(1) (kR)
6.2. THE LINEAR SAMPLING METHOD 63

6.2 The linear sampling method

In this section, we introduce a simple algorithm to determine the obstacle D from the
knowledge of the far field pattern. The algorithm, simply speaking, is to solve the following
far field equation: Z ∞
u∞ (θ, φ)g(φ)dφ = γe−ikrz cos(θ−θz ) (6.2.1)
0
eiπ/4
for any sampling point z = rz (cos θz , sin θz ), where γ = √
8πk
. Or in simpler notation,

(F g)(x̂) = Φ∞ (x̂, z) = γe−ikx̂·z .

When z ∈ D, ∥gz ∥ is small; when z ∈ R2 \ D, ∥gz ∥ is large. This phenomenan determines

the obstacle D.
To introduce the method, first we have to define the Heglotz wave function, the far
field operator and its properties. For a function g ∈ L2 [0, 2π], the Herglotz function is
defined by Z Z 2π
vg (x) := eikx·d g(d)ds(d) = eikr cos(θ−φ) g(φ)dφ.
S 0
It is a linear superposition of plane waves with different propagation directions. For each
plane wave, there is a far field pattern u∞ (θ, φ), then define the far field operator

F : L2 [0, 2π] → L2 [0, 2π]

Z 2π
g 7→ u∞ (θ, φ)g(φ)dφ.
0

Theorem 6.2.1. Assume the Herglotz wave function v with kernel g vanishes in R2 , then
g = 0.
Proof. Recall the expansion of the plane wave:
∞
X
eikr cos θ = in Jn (kr)einθ .
n=−∞

Thus
Z 2π
0 = vg (x) = eikr cos(θ−φ) g(φ)dφ
0
Z 2π ∞
X
= in Jn (kr)ein(θ−φ) g(φ)dφ
0 n=−∞
X Z 2π
= n
i Jn (kr)e inθ
g(φ)e−inφ dφ.
n∈Z 0

This implies that ĝ(n) = 0 for all n ∈ Z thus g = 0.

Theorem 6.2.2. Suppose k 2 is not an eigenvalue for the −∆ operator with the boundary
condition Bφ = 0 in D. Then the far field operator is an injection with dense range.
64 CHAPTER 6. INVERSE SCATTERING PROBLEMS

Proof. First we show that

(F ∗ g)(θ) = (F h)(θ + π)
then we only need to prove that F is injective. If F is injective, then F ∗ is injective. Thus

N (F ∗ )⊥ = R(F ) = L2 [0, 2π].

So we prove the above equation first, and then prove F is injective.

From the definition,
Z 2π Z 2π
⟨F g, h⟩ = u∞ (θ, φ)g(φ)dφh(θ)dθ
0 0
Z 2π Z 2π Z 2π
= g(φ) u∞ (θ, φ)h(θ)dφdθ
0 0 0
= ⟨g, F ∗ h⟩ .

Thus Z 2π
∗
(F h)(φ) = u∞ (θ, φ)h(θ)dθ.
0
From the reciprocity relation,
Z 2π
∗
(F h)(θ) = u∞ (−φ, −θ)h(θ)dθ
0
Z 2π
= u∞ (φ + π, θ + π)h(θ)dθ
0
Z 2π
= u∞ (φ + π, θ)h(θ − π)dθ
0
= (F h(θ − π))(φ + π),

it implies that
(F ∗ h)(θ) = (F h(θ − π))(φ + π).
In the second step, we prove that F is an injection. Let F g = 0 for some g ∈ L2 [0, 2π],
then vg is the Herglotz wave function with kernel g. Then

v∞ (x̂) = (F g)(x̂) = 0.

Then we know that v s (x) = 0 for all x ∈ R2 \ D thus Bvg = 0. Since vg satisfies the
Helmholtz equation in D with the vanishing Dirichlet boundary condition, vg = 0 in D.
Since vg is a real-analytic function, vg = 0 in R2 . This implies that g = 0. Thus F is
injective, so is F ∗ . The proof is finished.

General idea for the linear sampling method. When z ∈ D, then Φ(·, z) is a radiating
solution to the Helmholtz equation with the boundary condition given by BΦ(·, z) = 0 on
∂D. In this case, there is a g ∈ L2 [0, 2π] such that vgs = Φ(·, z) which is equivalent to
6.2. THE LINEAR SAMPLING METHOD 65

v∞ = Φ∞ (·, z). Thus ∥g∥ is bounded. When z ∈ R2 \ D, Φ(·, z) is not a radiating solution
due to the singularity. Thus we can not find a kernel g such that v∞ = Φ∞ (·, z). Thus by
solving
F g = Φ∞ (·, z)
we can determine if the point z ∈ D. For a rigorous proof, we still need the factorization
of the operator F .
For simplicity, we consider the forward problem
∂v
∆v + k 2 v = 0 in R2 \ D with + iλu = 0 on ∂D
∂ν
and the scattered field v s := v − v i where v i is the Herglotz wave function vg , satisfies the
Sommerfeld radiation condition. Here λ is an unknow positive and continuous function.
Note that in this case, there is no eigenvalue for the interior problem.
We begin with the following general scattering problem:

∆w + k 2 w = 0 in R2 \ D (6.2.2)
∂w
+ iλw = f on ∂D (6.2.3)
 ∂ν
√ ∂w

lim r − ikw = 0 (6.2.4)
r→∞ ∂R

where f ∈ H −1/2 (∂D). In particular, the scattered field v s satisfies the boundary condition
when
∂vg
f =− − iλvg .
∂ν
Then we define the following boundary operator

B : H −1/2 (∂D) → L2 [0, 2π]

f 7→ w∞

Thus  
∂vg
v∞ = −B + iλvg .
∂ν
Moreover, define the operator H as follows:

H : L2 [0, 2π] → H −1/2 (∂D)

 
∂vg
g 7→ + iλvg
∂ν ∂D

With these definitions, we have the following factorization:

F = −BH.

To be prepared for the study of the method, we need to show the properties of the operators
B and H, respectively.
66 CHAPTER 6. INVERSE SCATTERING PROBLEMS

Theorem 6.2.3. The boundary operator B : H −1/2 (∂D) → L2 [0, 2π] is compact, injective
and has dense range in L2 [0, 2π].

Proof. Recall that the modified single-layer potential,

Z
w(x) = Γ(x, y)φ(y)ds(y), x ∈ R2 \ D,
∂D

where Γ(x, y) is given by:

iX
Γ(x, y) = Φ(x, y) + an Hn(1) (kr)Hn(1) (kry )ein(θ−θy ) .
4 n∈Z

Then Z
w∞ (x̂) = Γ∞ (x̂, y)φ(y)ds(y)
∂D

where Γ∞ (x̂, y) is an analytic function for both variables thus Γ∞ (x̂, y) is smooth for both
variables (x̂, y) ∈ S × ∂D. Thus for any fixed x̂ ∈ S, Γ(x̂, ·) ∈ H 1 (∂D). Now we prove that
B is a compact from H −1/2 (∂D) to L2 [0, 2π]. From direct computation:
Z Z
|w∞ (x̂1 ) − w∞ (x̂2 )| = Γ∞ (x̂1 , y)φ(y)ds(y) − Γ∞ (x̂2 , y)φ(y)ds(y)
∂D ∂D
≤ ∥Γ∞ (x̂1 , ·) − Γ∞ (x̂2 , ·)∥H 1 (∂D) ∥φ∥H −1 (∂D)

which implies that w∞ (x̂) ∈ C[0, 2π]. Thus B is a bounded operator from H −1 (∂D) to
C[0, 2π]. From the compact embedding from H −1/2 (∂D) to H −1 (∂D) and from C[0, 2π]
to L2 [0, 2π], it is compact from H −1/2 (∂D) to L2 [0, 2π].
Now prove that B is an injection. Let Bf = 0, then w to the boundary value problem
(6.2.2)-(6.2.4) is 0. Thus f = 0, which proves the injectivity.
Then we prove that B has dense range. Since the set {einθ : θ ∈ [0, 2π]} is complete
in L2 [0, 2π], we only need to check that for any finite series
n
X
wn,∞ (θ) = aj eijθ ,
j=−n

there is a function fn such that Bfn = wn,∞ . From the far field pattern of the Hankel
(1)
functions Hj (kr)einθ , the radiating solution
n
X (1)
wn (x) = aj γj−1 Hj (kr)eijθ
j=−n

where r
2 −i(jπ/2+pi/4)
γj = e .
π
Let fn := ∂wn
∂ν
+ iλwn then it is a function in H −1/2 (∂D). Thus B has dense range.
6.2. THE LINEAR SAMPLING METHOD 67

Theorem 6.2.4. The boundary operator H is bounded, injective and has dense range in
H −1/2 (∂D).
Proof. It is clear that H is bounded, from the definition of the Herglotz wave function.
Suppose Hg = 0, thus vg is a global solution to the Helmholtz equation with ∂v
∂ν
g
+iλvg =
0. Then vg = 0 in D which implies that g = 0. Thus H is an injection.
Recall that a function is a Heglotz wave function if and only if it has the form of
X X
an Jn (kr)einθ := an un (x), r ≥ 0, θ ∈ [0, 2π).
n∈Z n∈Z

Thus to prove that H has a dense range, it is sufficient to show that

  
∂un
+ iλun : n∈Z
∂ν ∂D

is complete in H −1/2 (∂D), i.e., its orthogonal complement is trivial. Thus we need to show
that if for some g ∈ H 1/2 (∂D),
Z  
∂un
g(y) + iλun ds(y) = 0, ∀ n ∈ Z,
∂D ∂ν
then g = 0.
Suppose we have such a g ∈ H 1/2 (∂D), define
Z  
∂Φ(x, y)
u(x) := g(y) + iλΦ(x, y) ds(y).
∂D ∂ν(y)
From the following formula, when |x| >> |y|:
(1)
X X
H0 (k|x − y|) = Hn(1) (k|x|)Jn (k|y|)ein(θy −θ) := un (y)Hn(1) (k|x|)einθ
n∈Z n∈Z

, we have:
X Z 
∂un
 
u(x) = g(y) + iλun ds(y) Hn(1) (k|x|)einθ = 0,
n∈Z ∂D ∂ν

wheren |x| ≥ R >> |y| for all y ∈ ∂D. Thus u = 0 in R2 \ D. Thus

∂u +
u|+
∂D = =0
∂ν ∂D

Extend the solution u into D, then it satisfies the Helmholtz equation with the jump
conditions:
+ − ∂u + ∂u −
u|∂D − u|∂D = g, − = −iλg.
∂ν ∂D ∂ν ∂D
Thus we have the following boundary conditions:
∂u −
u|−
∂D = −g, = iλg.
∂ν ∂D
68 CHAPTER 6. INVERSE SCATTERING PROBLEMS

With the Green’s formula,

Z Z Z
2 2 2 ∂u
0 = (|∇u| − k |u| )dx = uds = −i λ|g|2 ds,
D ∂D ∂ν ∂D

which implies that g = 0 on ∂D. The proof is finished.

Theorem 6.2.5. Let Φ∞ (x̂, z) be the far field pattern of Φ(x, z), then Φ∞ (x̂, z) ∈ R(B) if
and only if z ∈ D.

Proof. If z ∈ D, then Φ(·, z) is the solution of (6.2.2)-(6.2.4) with f = ∂Φ

∂ν
+ iλΦ ∂D ,
then Φ∞ = Bf .
If z ∈ R2 \ D but Φ∞ = Bf for some f ∈ H −1/2 (∂D). Thus Φ(·, z) = w in R2 \ D,
which is impossible due to the singularity. Thus Φ∞ ̸∈ R(B).

Theorem 6.2.6. Let u∞ be the far field pattern for the scattered field u.

• If z ∈ D. For every ε > 0, there exists gzε := gz ∈ L2 [0, 2π] satisfying

∥F gz − Φ∞ (·, z)∥L2 [0,2π] < ε

such that
lim ∥gz ∥L2 [0,2π] = ∞
z→∂D

and
lim ∥vgz ∥H 1 (D) = ∞.
z→∂D

• If z ̸∈ D. The for every ε > 0 and δ > 0, there exists gzε,δ := gz ∈ L2 [0, 2π] satisfying
the inequality
∥F gz − Φ∞ (·, z)∥L2 [0,2π] < ε + δ
such that
lim ∥gz ∥L2 [0,2π] = ∞
δ→0

and
lim ∥vgz ∥H 1 (D) = ∞.
δ→0

∂
∈ H −1/2 (∂D) such


Proof. When z ∈ D. Thus there is an fz = − ∂ν
+ iλ Φ(·, z)
∂D
that Bfz = −Φ∞ (·, z). Since H has a dense range in L2 [0, 2π], for any ε > 0, there is a
gzε ∈ L2 [0, 2π] such that
∥Hgz − fz ∥H −1/2 (∂D) < ε.
Since B is continuous and F = −BH,

∥F gz − Φ∞ (·, z)∥H −1/2 (∂D) = ∥BHgz − Bfz ∥H −1/2 (∂D) < Cε.

From the definition, when z → ∂D, the norm ∥fz ∥H −1/2 (∂D) → ∞. Thus ∥Hgz ∥H −1/2 (∂D) →
∞ thus ∥vg ∥H 1 (D) → ∞. Thus ∥gz ∥L2 [0,2π] → ∞.
6.3. THE ITERATIVE METHOD 69

/ D. In this case, −Φ∞ (·, z) ̸∈ R(B) but R(B) = L2 [0, 2π]. Thus from the
When z ∈
Tikhonov regularization technique, for any δ > 0, there is a regularized solution which
solves
(α(δ)I + B ∗ B)fzα(δ) = −B ∗ Φ∞ (·, z),
or more explicity,
∞
X µn
fzα(δ) := − (Φ∞ (·, z), gn )φn
n=1
α + µ2n
where (µn , φn , gn ) is a singular system for B such that

∥Bfzα(δ) + Φ∞ (·, z)∥L2 [0,2π] < δ.

From Picard’s theorem, we notice that since Φ∞ (·, z) ̸∈ R(B),

∥fzα(δ) ∥H −1/2 (∂D) → ∞ as α → 0.

α(δ)
For each fixed α, we can again find a gz such that

∥Bfzα − BHgzα ∥L2 [0,2π] < ε.

Thus we have

∥F gzα − Φ∞ (·, z)∥L2 [0,2π] = ∥BHgzα + Φ∞ (·, z)∥L2 [0,2π]

≤ ∥BHgzα − Bfzα ∥L2 [0,2π] + ∥Bfzα + Φ∞ (·, z)∥L2 [0,2π] ≤ ε + δ.

It is clear that when α → 0 (δ → 0), ∥gzα ∥L2 [0,2π] → ∞ and ∥vgzα ∥H 1 (D) → ∞.
Note that although F is an injection in L2 [0, 2π] and Φ∞ (·, z) ∈ R(F ) when z ∈ D,
if the far field data u∞ (·, d) is perturbed the far field equation is still possible to have no
solution. Thus the regularization technique is still required to solve the equation. The
linear sampling method will be solved by the following algorithm.

• Suppose D is known to be contained in a bounded set S, then select the grid points
zn ∈ S.

• For each zn , use a regularization technique (e.g., Tikhonov) to solve the problem
F gn = Φ∞ (·, zn ).

• Determine a cutoff value C such that zn ∈ D if and only if ∥gn ∥ ≤ C.

6.3 The iterative method

In contrast to the linear sampling method which is fast, convenient but not very accurate,
the iterative method suffers from a large computational complexity and produces very
accurate solutions. For simplicity, we assume that D is starlike, i.e.,

∂D := {r(x̂) : x̂ = (cos θ, sin θ) where x ∈ [0, 2π]}

70 CHAPTER 6. INVERSE SCATTERING PROBLEMS

where r ∈ C 1 [0, 2π] is a positive periodic function.

For a fixed incident field ui , we define the operator
F : C 1 (∂D) → L2 [0, 2π]
∂ 7→ u∞
which maps the sound soft boundary ∂D to the far field pattern u∞ . We also write
F(r) := F(∂D) to indicate the dependence of the function r.
The operator is nonlinear and ill-posed in general thus a direct solver to the problem
is not possible. Thus the operator is first linearized:
F(r0 ) + Fr′0 q = u∞ .
By solving the above linear problem with the help of the regularization technique we get
q, then r ≈ r0 + q. The linearization is based on the differentiation in the normed spaces,
which is an extension of the classic differentiation of functions. First, we need to define
the Fréchet derivative.
Definition 6.3.1 (Fréchet derivative). Let X and Y be two normed spaces and the operator
B : U → Y in an open domain U ⊂ X. The operator B is Fréchet differentiable at x ∈ U ,
if there is a bounded linear operator A : X → Y such that
∥B(x + h) − B(x) − Ah∥Y
lim = 0,
∥h∥X →0 ∥h∥X
or
B(x + h) = B(x) + Ah + o(h).
The operator A is the Fréchet derivative of B at the point x, we write
DBx = Bx′ = A.
For this problem, the derivative Fr′ is also called the domain derivative. The idea is
natural: compare the difference between F(r) and F(r + εh) for some unit element h and
0 < ε << 1, and then analyze the case when ε → 0. The difficulty is, with r and r + h, the
related obstacles are different thus the problems are defined in two areas. Thus we need
to study both of the problems in one reference domain.
At the very beginning, we have to mention the weak solution in a bounded domain.
Assume that 0 ∈ D and B(0, R) is a disc with center 0 and radius R. The radius R >> 1
such that D ⊂ B(0, R/2). Thus we need the Dirichlet-to-Neumann operator related to the
sound soft scattering problem. The operator is defined as follows:
T : H 1/2 (ΓR ) → H −1/2 (ΓR )
∂w
w 7→
∂ν
where ΓR := {x ∈ R2 : ∥x∥ = R}. From the previous knowledge, a radiating solution has
the following expansion:
X
w(r, θ) = an Hn(1) (kr)einθ , D ⊂ B(0, R).
n∈Z
6.3. THE ITERATIVE METHOD 71

Thus T is defined in an explicit form:

X kHn(1)′ (kr) X
Tw = (1)
ŵn einθ , where w = ŵn einθ .
n∈Z Hn (kr) n∈Z

Example 6.3.2. The domain derivative needs a lot of computation. To have a easier ac-
cess to the variational method, let’s begin with an example which solves the inverse medium
prblem with n ∈ L∞ (D):

∆u + k 2 nu = 0 in R2 ;
us : = u − ui satisfies S.R.C.

Then we can easily get the equations for the scattered field us :

∆us + k 2 nus = k 2 (1 − n)ui in R2 ,

and assume that n − 1 is always bounded and compactly supported in a domain D. First,
let B(0, R) be a sufficiently large disk such that D ⊂ B(0, R/2). Then on the boundary
s
ΓR , ∂u
∂ν
= T us , thus
∂us
 i 
∂u i
= Tu + − T u := T u + f
∂ν ∂ν
where f ∈ H −1/2 (ΓR ). Now we discuss the problem in the disk B(0, R). Then multiply the
equation on both sides and apply the DtN map, we get
Z Z Z
2
[∇u · ∇φ − k nuφ]dx − T uφds = f φds. (6.3.1)
B(0,R) ΓR ΓR

Without proof, the problem (6.3.1) admits a unique solution in H 1 (B(0, R)). And the far
field operator is defined as
F(n) := u∞ .
We want to find its Fréchet derivative at the point n. Now let’s take a small perturbation
εg and let the new total field be uε . Then it satisfies
Z Z Z
ε 2 ε ε
[∇u · ∇φ − k (n + εg)u φ]dx − T u φds = f φds.
B(0,R) ΓR ΓR

Let the LHS in (6.3.1) be denoted by S(u, φ), then uε satisfies

Z Z
ε 2 ε
S(u , φ) − εk gu φdx = f φds.
B(0,R) ΓR

Thus Z
ε 2
S(u , φ) − εk guε φdx = S(u, φ)
B(0,R)

thus   Z
1 ε 2
S (u − u), φ = k uε φdx.
ε B(0,R)
72 CHAPTER 6. INVERSE SCATTERING PROBLEMS

When the solution depends continuously on ε, uε → u when ε → 0. Thus let û :=

limε→0 1ε (uε − u), it satisfies
  Z
1 ε 2
S (u − u), φ = k guφdx,
ε B(0,R)

which means it is a radiating solution of

∆û + k 2 nû = −k 2 gu in R2 .

By solving this problem, we can solve the nonlinear equation Fn = u∞ .

Chapter 7

Scattering by an Orthotropic
Medium

In this chapter, we will discuss the scattering problems with more complex media, i.e.,
the orthotropic media. It is more complex compared to homogeneous/isotropy media, and
the boundary integral equation method no longer works in general. We start from the
Maxwell’s equations and then apply the variational method to study these problems.

7.1 Orthotropic Medium

Recall the Maxwell’s equations:
∂H ∂E
∇ × E + µ(x) = 0, ∇ × H − ε(x) = σ(x)E,
∂t ∂t
where ε is the electric permittivity, µ is the magnetic permeability and σ is the electric
conductivity. The time-harmonic fields have the following form:
−iωt −iωt
E(x, t) = E(x)e
e , H(x, t) = H(x)e
e

where ω > 0 is the frequency. Then

∇×E
e − iωµ(x)He =0
∇×H
e − (iωε(x) − σ(x))E
e=0

Assume that the obstacle is a cylinder in R3 and it is along the x3 direction. Assume its
cross section is D then the obstacle is D × R. Assume that ε(x) = ε0 , µ(x) = µ0 and
σ(x) = 0 outside the cylinder and do not depend on x3 . Define

e int,ext = √1 E int,ext , e int,ext = √1 H int,ext , √

E H k = ω ε0 µ 0 ,
ε0 µ0

and  
1 σ(x) 1
A(x) = ε(x) + i , N (x) = µ(x),
ε0 ω µ0

73
74 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

e int is the field in the cylinder and E

where E e ext outside it. Then from direct computation,

∇ × E int − ikN H int = 0, ∇ × H int + ikAE int = 0

and
∇ × E ext − ikH ext = 0, ∇ × H ext + ikE ext = 0.
The equation for the magnetic field outside is

∇ × ∇ × H ext − k 2 H ext = 0,

and the interior

∇ × A−1 (x)∇ × H int − k 2 N H int = 0.
Suppose E i and H i are the incident fields, and E s and H s are the scattered fields, then

E ext = E i + E s , H ext = H s + H i .

Suppose the incident field propagates perpendicular to the axis of the cylinder and is
polarized:
H i (x) = (0, 0, ui ), H s (x) = (0, 0, us ), H int (x) = (0, 0, v).
Then
∇ · A(x)∇v + k 2 n(x)v = 0 D (7.1.1)
where A is a 2 × 2 matrix,
∆us + k 2 us = 0 in R2 \ D. (7.1.2)
From the transmission conditions ν ×H ext = ν ×H int and ν ×∇×H ext = ν ×A−1 ∇×H int ,
we have the following boundary conditions:

v − us = ui , ν · A∇v − ν · us = ν · ∇ui on ∂D. (7.1.3)

Finally, us satisfies the Sommerfeld radiation condition.

7.2 Mathematical framework for the variational meth-

ods
We will study the problem (7.1.1)-(7.1.3) with the help of the variational methods. To this
end, we introduce in this section the mathematical framework. We begin with the review
of some definitions. Let X be a Hilbert space with the norm ∥ · ∥ and inner product (·, ·).

Definition 7.2.1. A mapping a(·, ·) : X × X → C is a sesquilinear form if

a(λ1 u1 + λ2 u2 , v) = λ1 a(u1 , v) + λ2 a(u2 , v);

a(u, µ1 v1 + µ2 v2 ) = µ1 a(u, v1 ) + µ2 a(u, v2 )

holds for all λ1 , λ2 , µ1 , µ2 ∈ C and u, u1 , u2 , v, v1 , v2 ∈ X.

7.2. MATHEMATICAL FRAMEWORK FOR THE VARIATIONAL METHODS 75

Definition 7.2.2. A mapping F : X → C is called a conjugate linear functional if

F (µ1 v1 + µ2 v2 ) = µ1 F (v1 ) + µ2 F (v2 )

holds for all µ1 , µ2 ∈ C and v1 , v2 ∈ X.

In the variational problem, we will solve the following abstract problem: Given a
conjugate linear functional F : X → C and a sesquilinear form a(·, ·) on X × X, find
u ∈ X such that
a(u, v) = F (v) for all v ∈ X. (7.2.1)
The first answer to this question is the well-known Lax-Milgram Lemma.
Theorem 7.2.3. Assume that a(·, ·) : X × X → C is a bounded and coercive sesquilinear
form (not necessarily symmetric). This means, there exist constants α, β > 0 such that

|a(u, v)| ≤ α∥u∥∥v∥ for all u, v ∈ X

and
a(u, u) ≥ β∥u∥2 for all u ∈ X.
Then for any bounded conjugate linear functional F : X → C, there exists a unique
element u ∈ X such that

a(u, v) = F (v) for all v ∈ X.

Moreover, ∥u∥ ≤ C∥F ∥ where C > 0 is a constant independent of F .

Proof. For each fixed u ∈ X, a(u, v) is a conjugate linear functional w.r.t. v. From Riese
Representation theorem, there is an element w ∈ X such that

a(u, v) = (w, v) for all v ∈ X.

Apply the Riese representation theorem again to F , there is also an element f ∈ X such
that
F (v) = (f, v) for all v ∈ X.
Moreover, ∥f ∥ = ∥F ∥.
Define the following operator:

A: X → X
u 7→ w

then we aim to solve the problem:

(Au, v) = (f, v) for all v ∈ X.

This is equivalent to solve the equation T u = f .

In the following, we will prove that T is a bounded linear bijection thus it is invertible.
Then the solution u is given by T −1 f .
76 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

Theorem 7.2.4. Assume that a(·, ·) : X × X → C is a bounded and coercive sesquilinear

form and b(·, ·) : X × X → C is a bounded sesquilinear form which defines a compact
operator B : X → X such that
b(u, v) = (Bu, v),
then the operator A + B is Fredholm. Thus either the problem has a unique solution, or
f = 0 results in a nontrivial solution in X.

7.3 Variational method for scattering problems

We start with the scattering problem with an impenetrable obstacle:
∆u + k 2 u = 0 in R2 \ D; u = 0 on ∂D
where us := u − ui satisfies the Sommerfeld radiation considtion. Let B(0, R) be a suf-
ficiently large disk with radius R > 0, and its boundary is called ΓR . Define ΩR :=
B(0, R) \ D. Let
u(x) := us (x) + ui (x)X (|x|)
where X ∈ C ∞ (R) satisfies X = 1 when |x| < r (r is chosen such that D ⊂ B(0, r)) and
X = 0 when |x| > 2r where 2r < R. Then
∆u + k 2 u = f in ΩR ;
u = 0 on ΓR ;
∂u
= T u on ΓR .
∂ν
Recall that the DtN map T is defined as
X X
Tφ = an γn einθ with φ = an einθ
n∈Z n∈Z

(1)′
where γn := kH (1)
n
. From the asymptotic behaviour of the Hankel functions and their
Hn (kR)
derivatives, it is known that there are two positive constants c1 and c2 such that
c1 |n| ≤ |γn | ≤ c1 |n|, n = ±1, ±2, . . . .
More precisely,   
n 1
γn = − 1+O , |n| → ∞.
R n
Now we apply the variational method to the obstacle scattering problem. The space
X := {φ ∈ H 1 (ΩR ) : v|∂D = 0}. Multiply with the test function v ∈ X and apply the
Green’s theorem, we obtain:
Z Z Z
2
[∇u · ∇v − k uv]dx − T uvds = − f vdx. (7.3.1)
ΩR ΓR ΩR

We begin with the analysis of the operator T . The first lemma shows that T is a
bounded linear invertible operator from H 1/2 (ΓR ) to H −1/2 (ΓR ).
7.3. VARIATIONAL METHOD FOR SCATTERING PROBLEMS 77

Lemma 7.3.1. The operator T is bounded from H 1/2 (ΓR ) to H −1/2 (ΓR ). It is a bijection
and its inverse is also bounded.

Proof. Recall the definition of the H s (ΓR ) space:

X
∥φ∥2H s (∂D) = (1 + n2 )s |an |2
n∈Z

an einθ . For any φ ∈ C ∞ (ΓR ) which has a finite series

P
where φ = n∈Z

N
X
φ= an einθ ,
n=−N

it is a function in H 1/2 (ΓR ) thus

N
X
∥φ∥2H 1/2 (∂D) = (1 + n2 )1/2 |an |2 < ∞.
n=−N

From direct computation,

N
X
Tφ = an γn einθ ,
n=−N

thus it is a smooth function. Then since |γn | ≤ c2 |n| < c2 (1 + n2 )1/2 ,

X
∥T φ∥2H −1/2 (∂D) = (1 + n2 )−1/2 |an |2 |γn |2
n∈Z
X
≤ (1 + n2 )−1/2 |an |2 c22 (1 + n2 ) = c22 ∥φ∥2H 1/2 (∂D) .
n∈Z

For any random φ ∈ H 1/2 (∂D), the above inequaltiy also holds using a limiting process.
Thus T is a bounded operator from H 1/2 (∂D) P to H −1/2 (∂D).
The invertibility is given directly. Let ψ = Nn=−N bn e
inθ
, then

T −1 ψ = an γn−1 einθ .

Apply the previous process and use the property γn ≥ c1 |n|, we show the boundedness of
T −1 .

To formulate the variational problem in the framework, we need to define

Xn
T := T0 + T1 , where T0 φ = − an einθ ,
n∈Z
R

thus X X  n  inθ
T1 := T − T0 = an ξn einθ = an γn + e ,
n∈Z n∈Z
R
78 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

where ξn = O(1). Then similar to the above lemma, T0 is also a bounded linear invertible
operator. Furhter more,
Z X 2πn
− T uuds = |ûn |2 ≥ 0.
ΓR n∈Z
R

Define: Z Z
a(u, v) = [∇u · ∇v + uv]dx − T0 uvds,
ΩR ΓR

and Z Z
2
b(u, v) = −(k + 1) uvdx − T1 uvds.
ΩR ΓR

Thus a is a bounded, coercive sesquilinear form thus admits a bounded invertible linear
operator A. For the sesquilinear form b, it is already clear that the first term admits
R a com-
pact operator. Now the rest of the work concerns with the second term, i.e., ΓR T1 uvds.

Lemma 7.3.2. The operator T1 is compact from H 1/2 (ΓR ) to H −1/2 (ΓR ).

Proof. We still begin with φ = N inθ

P
n=−N an e , then

N
X
T1 φ = an ξn einθ
n=−N

n


where ξn = γn + R
, then ξn = O(1) when n → ∞. From direct computation,

N
X N
X
∥T1 φ∥2H 1/2 (∂D) = 2 2
|an | |ξn | = |an |2 |ξn |2 (1 + n2 )1/2
n=−N n=−N
N
X
≤ sup{|ξn |2 } |an |2 (1 + n2 )1/2
n∈Z
n=−N
2 2
≤C ∥φ∥H 1/2 (∂D) .

This implies that T1 is a bounded linear operator from H 1/2 (∂D) to H 1/2 (∂D). From
the compact embedding from H 1/2 (∂D) to H −1/2 (∂D), T1 is compact from H 1/2 (∂D) to
H −1/2 (∂D).

In this case, the sesquilinear form b admits a compact operator B. Thus the original
problem is formulated as
(A + B)u = f.
Since the operator is Fredholm, it is sufficient to prove that the problem has at most
one equation. Thus let f = 0, then u is the radiating solution in the exterior of D with
vanishing boundary value. Thus u = 0. So the uniqueness holds thus implies the solvability
of the problem.
7.4. SCATTERING BY AN ORTHOTROPIC MEDIUM 79

7.4 Scattering by an Orthotropic medium

Now we are prepared to solve the problem (7.1.1)-(7.1.3). Instead of the original problems,
we consider the following truncated equations:

∇ · A∇v + k 2 nv = 0 in D (7.4.1)
∆u + k 2 u = 0 in ΩR \ D (7.4.2)
v−u = f on ∂D (7.4.3)
∂v ∂u
− = h on ∂D (7.4.4)
∂νA ∂ν
∂u
= T u on ΓR . (7.4.5)
∂ν

Here ∂ν∂v
A
= ν · A∇v, f ∈ H 1/2 (∂D) and h ∈ H −1/2 (∂D). On D, A is a matrix valued
function A : D → C2×2 whose elements are continuously differentiable. Moreover, the
real- and imaginary parts, denoted by Re (A) and Im (A), are symmetric and satisfy

ξ · Im (A)ξ ≤ 0, ξ · Re (A)ξ ≥ γ|ξ|2

for some positive constant γ > 0. Moreover, n ∈ C(D) and Im (n) ≥ 0.

Then we seek for the solution (v, u) ∈ H 1 (D) × H 1 (ΩR \ D). The difficulty of this
problem is that v and u jump across the boundary of D. To this end, we have to choose
a proper R such that k 2 is not a Dirichlet eigenfunction for −∆ in ΩR \ D. In this case,
let uf ∈ H 1 (ΩR \ D) be the solution to the following boundary value problem:

∆uf + k 2 uf = 0 in ΩR \ D, uf = 0 on ΓR , uf = f on ∂D.

Then there is a constant C > 0 which only depends on the domain ΩR such that

∥uf ∥H 1 (ΩR \D) ≤ C∥f ∥H 1/2 (∂D) .

Moreover, uf has the weak formulation

Z Z Z
2 ∂uf ∂uf
[∇uf · ∇φ − k uf φ]dx = φds − φds.
ΩR \D ΓR ∂ν ∂D ∂ν

Let w be defined by u + uf in ΩR and v in D. Then w satisfies

∇ · A∇w + k 2 nw = 0 in D (7.4.6)
∆w + k 2 w = 0 in ΩR \ D (7.4.7)
w|− − w|+ = 0 on ∂D (7.4.8)
− +
∂w ∂w ∂uf
− = h− on ∂D (7.4.9)
∂νA ∂ν ∂ν
∂w ∂uf
= Tw + on ΓR . (7.4.10)
∂ν ∂ν
80 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

Multiply the equation on both sides by φ and apply the Green’s theorem,
Z Z −
2 ∂w
[A∇w · ∇φ − k nwφ]dx = φds (7.4.11)
D ∂D ∂νA

and Z Z Z +
2 ∂w ∂w
[∇w · ∇φ − k wφ]dx = φds − φds. (7.4.12)
ΩR \D ΓR ∂ν ∂D ∂ν
Add these two equations together and apply the boundary conditions,
Z Z Z
2 2
[A∇w · ∇φ − k nwφ]dx + [∇w · ∇φ − k wφ]dx − T wφds (7.4.13)
D ΩR \D ΓR
Z Z  
∂uf ∂uf
= φds + h− φds (7.4.14)
ΓR ∂ν ∂D ∂ν
Z Z Z
∂uf ∂uf
= hφds + ds − ds (7.4.15)
∂D ΩR \D ∂ν ∂D ∂ν
Z Z
= hφds + [∇uf · ∇φ − k 2 uf φ]dx. (7.4.16)
∂D ΩR \D

Now we summarize the above weak formulation as follows. Define

a(w, φ) := a1 (w, φ) − a2 (w, φ)

where
Z Z Z
a1 (w, φ) = [A∇w · ∇φ + wφ]dx + [∇w · ∇φ + wφ]dx − T0 wφds;
D ΩR \D ΓR
Z Z Z
2 2
a2 (w, φ) = (k n + 1)wφdx + (k + 1)wφdx − T1 wφds.
D ΩR \D ΓR

The right hand side

Z Z
F (φ) = hφds + [∇uf · ∇φ − k 2 uf φ]dx.
∂D ΩR \D

From the boundedness of A and n, the boundedness of a1 and a2 are clear. For F , it is
given by

|F (φ)| ≤ ∥h∥H −1/2 (∂D) ∥φ∥H 1/2 (∂D) + C∥uf ∥H 1 (ΩR \D) ∥φ∥H 1 (ΩR \D)
≤ C∥h∥H −1/2 (∂D) ∥φ∥H 1 (ΩR \D) + C∥f ∥H 1/2 (∂D) ∥φ∥H 1 (ΩR \D)


≤ C ∥f ∥H 1/2 (∂D) + ∥h∥H −1/2 (∂D) ∥φ∥H 1 (D)

Thus


∥F ∥ ≤ C ∥f ∥H 1/2 (∂D) + ∥h∥H −1/2 (∂D) .
7.5. INVERSE PROBLEMS 81

Since Re (A) is positive definite,

Z Z
2 2
a1 (w, w) = [A∇w · ∇w + |w| ]dx + ∥w∥H 1 (ΩR \D) − T0 wwds
D ΓR
≥ γ∥∇w∥L2 (D) + ∥w∥2L2 (D) + ∥w∥2H 1 (ΩR \D) ≥ C∥w∥2H 1 (ΩR ) .

It is also easily checked that a2 (w, φ) defines a compact operator. Thus uniqueness implies
the unique solvability. We only need to check if (f, h) = (0, 0) results in (v, u) = (0, 0).
From assumption, ξ · Im (A)ξ ≤ 0. We also further assume that Im (n) > 0 in D.
Assume that w is the solution related to f = 0 and h = 0, then
Z Z Z
2 2 2 2 2
[A∇w · ∇w − k n|w| ]dx + [|∇w| − k |w| ]dx − T wwds = 0.
D ΩR \D ΓR

Take the imaginary part,

Z Z Z
 2 2
 ∂w
∇w · Im (A)∇w − k Im (n)|w| dx = Im T wwds = Im wds.
D ΓR ΓR ∂ν
The first term is negative or 0 and the second term is also negative. Thus
Z
∂w
Im w ds ≥ 0.
ΓR ∂ν

Then w = 0 in ΩR \ D from Rellich’s lemma. Since Im (n) < 0, w = 0 in D. This proves

the uniqueness.

7.5 Inverse Problems

Suppose we have incident plane waves: ui = eikx·d , then the problem is given by

∇ · A∇v + k 2 nv = 0 in D
∆us + k 2 us = 0 in R2 \ D
v − us = eikx·d on ∂D
∂v ∂us ∂eikx·d
− = on ∂D
∂νA ∂ν ∂ν
∂us
= T us on ΓR .
∂ν
Here d = (cos φ, sin φ) for some φ ∈ [0, 2π). Recall that when |x| → ∞,

eik|x|
us (x) = p u∞ (θ, φ) + O |x|−3/2


|x|
where x = |x|(cos θ, sin θ). The properties for the far field pattern with bounded impene-
trable obstacles still hold in this case, such as the reciprocity:

u∞ (θ, φ) = u∞ (φ + π, θ + π);
82 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

and the integral representation:

eiπ/4 ∂e−ikx̂·y s
Z  
s −ikx̂·y ∂u (y)
u∞ (θ, φ) = √ u (y) −e ds(y).
8πk ∂B ∂ν ∂ν

Also the uniqueness results:

Theorem 7.5.1. Suppose the far field pattern u∞ = 0 for a fixed incident field, then
us = 0 in R2 \ D.

The related inverse problems are, given u∞ (θ, φ) for all incident angles φ ∈ [0, 2π] and
observation angles θ ∈ [0, 2π], how to determine D?
To solve the inverse problems, we recall the Herglotz wave function:
Z 2π
vg (x) = eikx·d g(φ)dφ
0

and the far field operator:

Z 2π
(F g)(θ) = u∞ (θ, φ)g(φ)dφ.
0

Recall that we always need to solve the following ill-posed equation:

F g = Φ∞

to solve the inverse problems. The solver to this problem is related to the properties of F :
injection with dense range.
If there is a g ∈ L2 [0, 2π] such that F g = 0 (not an injection), then the scattered us
field related to F g also vanishes. Thus

∇ · A∇v + k 2 nv = 0 in D
∆us + k 2 us = 0 in R2 \ D
v = vg on ∂D
∂v ∂vg
= on ∂D.
∂νA ∂ν
Note that vg satisfies the Helmholtz equation globally, the above second equation is re-
placed by
∆vg + k 2 vg = 0 in D.
Then the above problem is written as a system in D. If F g = 0 is related to a non-
vanishing g, then (v, vg ) ̸= (0, 0) is a nontrivial solution to the homogeneous problem, thus
it is an eigenfunction. So the properties of this problem is crucial for the properties of F .

Theorem 7.5.2. The far field operator is injective with dense range if and only if the
problem does not have nontrivial solutions in H 1 (D) × H 1 (D).
7.6. INTERIOR TRANSMISSION PROBLEMS 83

Proof. Recall that we have already shown that

(F ∗ g)(θ) = (F h(φ + π))(θ + π)

thus F is injective if and only if F ∗ is injective. Also since

N (F ∗ )⊥ = R(F ),

F is has a dense range if and only if F ∗ is an injection. Thus we only need to show that
F is an injection if and only if the problem does not have nontrival solutions. From the
above arguments, we already know that F g = 0 has a nontrivial solution g is equivalent
to the nontrivial solution to the problem.

From above arguments, we can see that the properties of F is closely related to the
following Interior Transmission Problem. Given (f, h) ∈ H 1/2 (∂D) × H −1/2 (∂D), find
(v, w) ∈ H 1 (D) × H 1 (D) satisfying

∇ · A∇v + k 2 nv = 0 in D (7.5.1)
∆w + k 2 w = 0 in D (7.5.2)
v−w = f on ∂D (7.5.3)
∂v ∂w
− = h on ∂D (7.5.4)
∂νA ∂ν
When (f, h) = (0, 0), then (7.5.1)-(7.5.4) is a homogeneous interior transmission prob-
lem.

Definition 7.5.3. The values k 2 for which the homogeneous interior transmission problem
has a nontrivial solution are called transmission eigenvalues.

7.6 Interior Transmission Problems

We first need to consider the unique solvability of the iterior transmission problem.

Theorem 7.6.1. If either Im (n) > 0 or ξ · Im(A)ξ < 0 for a point x0 ∈ D, then the
interor transmission problem (7.5.1)-(7.5.4) has at most one solution.

Proof. Suppose (v, w) be a solution to the homogeneous interior transmission problem.

Multipliy the equation (7.5.1) by v on both sides and apply the divergence theorem,
Z Z
2 2 ∂v
[A∇v · ∇v − k n|v| ]dx = vds.
D ∂D ∂νA

Similarly, we do the same thing for (7.5.2), then

Z Z
2 2 2 ∂w
[|∇w| − k |w| ]dx = wds.
D ∂D ∂ν
84 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

From the boundary conditions, we have:

Z Z
[A∇v · ∇v − k n|v| ]dx = [|∇w|2 − k 2 |w|2 ]dx.
2 2
D D

Take the imaginary parts, we have:

Z Z
2
[Im (A)∇v · ∇v]dx − k Im (n)|v|2 dx = 0.
D D

If for some x0 ∈ D, Im (n) < 0. Since n is continuous, there is a small ball B(x0 , ε) ⊂ D
such that Im (n) < 0 in it. Since for all x ∈ D, ξ · Aξ ≤ 0 and Im (n) ≥ 0, it means that
v = 0 in B(x0 , ε) thus ∇v = 0 in it as well. From the unique continuation principle, v = 0
in D. If for some x0 ∈ D, ξ · Aξ < 0, then there is a small ball such that the above
inequality holds. Thus ∇v = 0 in the ball, and so does v.
Since v = 0 in D, w satisfies ∆w + k 2 w = 0 in D with w = ∂w ∂ν
= 0. From integral
representation formula, w = 0 in D. Thus the solution is unique.
Now we need consider the solvability of the interior transmission problem. However, it
is convenience to consider a modified problem.

∇ · A∇v − mv = ϱ1 in D (7.6.1)
∆w − w = ϱ2 in D (7.6.2)
v−w = f on ∂D (7.6.3)
∂v ∂w
− = h on ∂D (7.6.4)
∂νA ∂ν

Here ϱ1 ∈ L2 (D), ϱ2 ∈ L2 (D), f ∈ H 1/2 (∂D) and h ∈ H −1/2 (∂D). We consider the unique
solvability of this problem first.
First, multiply with (7.6.1) on both sides by φ, then
Z Z Z
∂v
[A∇v · ∇φ + mvφ]dx = φds − ϱ1 φdx.
D ∂D ∂νA D

Let s := ∇w, then ∆w = ∇ · ∇w = ∇ · s. Recall we consider the solution w ∈ H 1 (D),

then s = ∇w ∈ (L2 (D))2 . Moreover, ∇ · s = ∆w = ϱ2 + ∈ L2 (D) and ∇ × s = 0. We are
looking for the vector in the Hilbert space

W (D) := {s ∈ (L2 (D))2 : ∇ · s ∈ L2 (D), ∇ × s = 0}

equipped with the norm

∥s∥2W (D) = ∥s∥2(L2 (D))2 + ∥∇ · s∥2L2 (D) .

Multiply (7.6.2) on both sides by ∇ · ψ for ψ ∈ W (D),

Z Z
[(∇ · s)(∇ · ψ) − w(∇ · ψ)]dx = ϱ2 (∇ · ψ)dx.
D D
7.6. INTERIOR TRANSMISSION PROBLEMS 85

Use the divergence theorem,

Z Z
∇ · (wψ)dx = wψ · µds,
D ∂D

which implies Z Z Z
∇w · ψdx + w(∇ · ψ)dx = wψ · νds.
D D ∂D

Thus Z Z Z
[(∇ · s)(∇ · ψ) + s · ψ]dx = ϱ2 (∇ · ψ)dx + wψ · νds.
D D ∂D

Then
Z Z
[A∇v · ∇φ + mvφ]dx + [(∇ · s)(∇ · ψ) + s · ψ]dx
D
Z Z ZD Z
∂v
= φds − ϱ1 φdx + ϱ2 (∇ · ψ)dx + wψ · νds
∂D ∂νA D D ∂D
Z Z Z Z Z Z
∂w
=− ϱ1 φdx + ϱ2 (∇ · ψ)dx + φds + hφds + vψ · νds − f ψ · νds.
D D ∂D ∂ν ∂D ∂D ∂D

∂w
Since ∂ν
= ν · ∇w = ν · s,
Z Z Z Z
[A∇v · ∇φ + mvφ]dx + [(∇ · s)(∇ · ψ) + s · ψ]dx − ν · sφds − vψ · νds
DZ Z D Z Z ∂D ∂D

=− ϱ1 φdx + ϱ2 (∇ · ψ)dx + hφds − f ψ · νds

D D ∂D ∂D

The above equation is simplified as

A(U, V ) = L(V ), ∀ v ∈ H 1 (D) × W (D), (7.6.5)

here U = (v, s) ∈ H 1 (D) × W (D) and the test function V = (φ, ψ).

Theorem 7.6.2. The problem (7.6.1)-(7.6.4) has a unique solution (v, w) ∈ H 1 (D) ×
H 1 (D) if and only if the problem (7.6.5) has a unique solution U = (v, s) ∈ H 1 (D)×W (D).
Moreover, if (v, w) is the unique solution to (7.6.1)-(7.6.4) then U = (v, s) is the unique
solution to (7.6.5). Conversely, if U = (v, s) is the unique solution of (7.6.5), then (v,w)
is the unique solution of (7.6.1)-(7.6.4) where s = ∇w.

Proof. It has already been seen from above that if (7.6.1)-(7.6.4) has a solution (v, w) ∈
H 1 (D) × H 1 (D), then U = (v, s) ∈ H 1 (D) × W (D) is a solution to (7.6.5). Now we show
that if U = (v, s) ∈ H 1 (D) × W (D) is a solution to (7.6.5), then (v, w) where ∇w = s is
a solution to (7.6.1)-(7.6.4).
Let the test function be (φ, 0) where φ ∈ H01 (D), then with the Green’s theorem,
Z Z Z
[A∇v · ∇φ + mvφ]dx = − ϱ1 φdx = − (∇ · A∇v − mv)φdx.
D D D
86 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

From the abitrary choice of φ,

∇ · A∇v − mv = ϱ1 in D.

Similarly let the test function be (0, ψ) where ψ|∂D = 0, let s := ∇w where w is decided
up to a constant vector. Then
Z Z Z
[∆w(∇ · ψ) + (∇w) · ψ] = [(∇ · s)(∇ · ψ) + s · ψ] = ϱ2 (∇ · ψ)dx.
D D D

With the following equation:

Z Z Z Z
(∇w) · ψdx + w(∇ · ψ)dx = ∇(wψ)dx = w(ν · ψ)ds = 0,
D D D ∂D

we have: Z Z
(∆w − w)(∇ · ψ)dx = ϱ2 (∇ · ψ)dx.
D D
Thus we obtain the equation for w:

∆w − w = ϱ2 .

Now we turn to the boundary terms. From above arguments, if we choose a random test
function (φ, ψ) ∈ H 1 (D) × W (D), we can easily show that
Z Z Z Z
∂v
φds + w(ν · ψ)ds − ν · sφds − vψ · νds
∂D ∂νA ∂D ∂D ∂D
Z Z Z
= ϱ2 (∇ · ψ)dx + hφds − f ψ · νds
D ∂D ∂D

∂w
Note that ν · s = ν · ∇w = ∂ν
, then

∂v ∂w
v − w = f, − = h.
∂νA ∂ν
Thus (v, w) is a solution of (7.6.1)-(7.6.4). The equivalence of the existence of solutions
between (7.6.1)-(7.6.4) and (7.6.5) is proved. Then we prove the equivalence of the unique-
ness.
Suppose (7.6.1)-(7.6.4) has at most one solution, i.e., ϱ1 = ϱ2 = 0 and f = h = 0
results in (v, w) = 0. Suppose (v, s) be a solution of (7.6.5) with vanishing right hand
side, then (v, w) where s = ∇w solves (7.6.1)-(7.6.4) with vanishing right hand side. Then
(v, w) = 0 thus s = 0. This proves the uniquenss of solutions of (7.6.5).
Suppose (7.6.5) has at most one solution, i.e., vanishing right hand side results in
(v, s) = 0. Suppose (v, w) is a solution to (7.6.1)-(7.6.4) with vanishing right hand side,
then (v, ∇w) is a solution to (7.6.5) with vanishing right hand side. Then v = 0 and
∇w = 0. Since ∆w − w = 0, we have w = ∇ · ∇w = 0. This proves the uniqueness.
In the following we will study the solutions to the problem (7.6.5).
7.6. INTERIOR TRANSMISSION PROBLEMS 87

Theorem 7.6.3. Suppose there is a constant γ > 1 such that for any x ∈ D,

ξ · Re (A)ξ ≥ γ|ξ|2 for all ξ ∈ C2 , m(x) ≥ γ,

then the problem (7.6.5) has a unique solution U = (v, s) ∈ H 1 (D) × W (D). Moreover,

γ+1
∥v∥H 1 (D) + ∥s∥W (D) ≤ 2C (∥ϱ1 ∥ + ∥ϱ2 ∥ + ∥f ∥ + ∥h∥) .
γ−1

Proof. From the trace theorem, it is easily seen that

∥L∥ ≤ C (∥ϱ1 ∥ + ∥ϱ2 ∥ + ∥f ∥ + ∥h∥) .

The boundedness of A is also obvious. Now we show the coercive.

Z Z Z Z
2 2 2
A(U, U ) = [A∇v · ∇v + m|v| ]dx + [|∇ · s| + |s| ]dx − (ν · s)vds − v(ν · s)ds
D D
Z ∂D ∂D
2 2
≥ γ∥v∥H 1 (D) + ∥s∥W (D) − 2Re v(ν · s)ds.
∂D

It is easily estimated that (exercise)

Z
v(ν · s)ds ≤ ∥v∥H 1 (D) ∥s∥W (D) .
∂D

Thus

A(U, U ) ≥ γ∥v∥2H 1 (D) + ∥s∥2W (D) − 2∥v∥H 1 (D) ∥s∥W (D)

γ−1 γ−1
≥ ∥v∥2H 1 (D) + ∥s∥2W (D) (exercise)
2 γ+1
γ−1 
≥ ∥v∥2H 1 (D) + ∥s∥2W (D) .
γ+1

The unique solvability of this problem as well as the boundedness of the solution comes
directly from Lax-Milgram lemma.

This theorem shows the existence and uniqueness of solutions to the variational problem
(7.6.5). We have to get back to the modified interior transmission problem (7.6.1)-(7.6.4).

Theorem 7.6.4. Suppose there is a constant γ > 1 such that for any x ∈ D,

ξ · Re (A)ξ ≥ γ|ξ|2 for all ξ ∈ C2 , m(x) ≥ γ,

then the problem (7.6.1)-(7.6.4) has a unique solution U = (v, w) ∈ H 1 (D) × H 1 (D).
Moreover,
γ+1
∥v∥H 1 (D) + ∥w∥H 1 (D) ≤ C (∥ϱ1 ∥ + ∥ϱ2 ∥ + ∥f ∥ + ∥h∥) .
γ−1
88 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

Proof. Since the existence and uniqueness of solutions between (7.6.1)-(7.6.4) and (7.6.5)
are equivalence, we only need to consider the estimation. As is already known,
γ+1
∥v∥H 1 (D) + ∥∇w∥W (D) ≤ 2C (∥ϱ1 ∥ + ∥ϱ2 ∥ + ∥f ∥ + ∥h∥) .
γ−1
Recall the definition of the W (D) norm:

∥∇w∥2W (D) = ∥∇w∥2(L2 (D))2 + ∥∇ · ∇w∥2L2 (D) = ∥∇w∥2(L2 (D))2 + ∥w + ϱ2 ∥2L2 (D) .

Since ∇ · ∇w = ∆w and ∆w − w = ϱ2 ,

∥∇w∥2W (D) ≥ ∥∇w∥2(L2 (D))2 + ∥w∥2L2 (D) − ∥ϱ2 ∥2L2 (D) = ∥w∥2H 1 (D) − ∥ϱ2 ∥2L2 (D) .

Plug this inequality into the above results then proves the final theorem.
Now we are prepared to show the existence of the original interior transmission problem
(7.5.1)-(7.5.4).
Theorem 7.6.5. Assume that either Im (n) > 0 or ξ · Im (A)ξ < 0 at a point x0 ∈ D and
there exists a constant γ > 1 such that ξ · A(x)ξ ≥ γ∥ξ∥2 for all ξ ∈ C2 and x ∈ D, then
(7.5.1)-(7.5.4) has a unique solution (v, w) ∈ H 1 (D) × H 1 (D) with the estimation

∥v∥H 1 (D) + ∥w∥H 1 (D) ≤ C (∥f ∥ + ∥h∥) .

Proof. Let s = ∇w, then the variational formulation for (7.5.1)-(7.5.4) is given by:

A(U, V ) + B(U, V ) = L(V ),

where U = (v, s) ∈ H 1 (D) × W (D), V = (φ, ψ) ∈ H 1 (D) × W (D), A is defined in the

wave way as in (7.6.4) and
Z Z
L(V ) = hφds − f (ν · ψ)ds.
∂D ∂D

The sesquilinear for B is given as:

Z Z
B(U, V ) = − (k n + m)vφdx − (k 2 + 1)s · ψdx.
2
D D

It results in a compact operator in H 1 (D) × W (D), thus uniqueness is equivalent to the

existence. We only need to prove the uniqueness, which has been proved. Thus the problem
is uniquely solvable and the solution is bounded by the right hand side.
Now we focus on the case that both A and n are real. Before that, we recall the analytic
Fredholm theorem.
Theorem 7.6.6. Let G(z) : X → X be a family of Fredholm operators in the Hilbert
space X, which depend analytically on z ∈ B where B is an open domain. Then either
• G(z) is not invertible for all z ∈ B, or
7.7. UNIQUENESS OF THE INVERSE PROBLEMS 89

• there is a discrete set S ⊂ B such that G(z) is invertible for all z ∈ B \ S.

Theorem 7.6.7. Assume that Im (n) = 0 and Im (A) = 0 in D. There is a constant γ > 1
such that for all x ∈ D,
ξ · A(x)ξ ≥ γ∥ξ∥2 ∀ ξ ∈ R2 .
Then the set of transmission eigenvalues is either empty or forms a discrete set.
Proof. Since A + B defines a family of Fredholm operators depend analytically on k, we
need to apply the analytic Fredholm theorem. Either it is not invertible for all k, or the
inverse operator is bounded except for a discrete set of k. Thus if we can show that the
problem is uniquely solvable for one k. Obviously when m = n and k = i, then B = 0
thus the problem is solvable. The proof is finished.
Note that the above results are extendable to cases when
ξ · Re (A)−1 ξ ≥ γ∥ξ∥2 for all ξ ∈ C2 .
Generally speaking, it is not known that if the transmission eigenvalues exist. The strict
condition is that either ∥Re (A)∥ > 1 or ∥Re (A)∥ < 1 for x ∈ D. For the case that A = I,
it is quite special and we will discuss in the next chapter. For the case that ∥Re (A)∥ > 1
for x ∈ D0 ⊂ D and ∥Re (A)∥ < 1 for x ∈ D \ D0 where D0 , D \ D0 ̸= ∅, it is still an open
problem.

7.7 Uniqueness of the inverse problems

We have to always assume that A satisfies
ξ · Re (A)ξ ≥ γ|ξ|2 , or ξ · Re (A−1 )ξ ≥ γ|ξ|2
for some γ > 1 and all ξ ∈ C2 . The uniqueness result is to show that the far field pattern
of all incident plane waves uniquely determines both the shape and the refractive index of
the inhomogeneous medium. However, we need a lemma to prepare for the proof.
Lemma 7.7.1. Let {vn , wn }n∈N ∈ H 1 (D) × H 1 (D) be a sequence of solutions to the
ITP (7.5.1)-(7.5.4) with boundary data (fn , hn ) ∈ H 1/2 (∂D) × H −1/2 (∂D). If fn and
hn converge in their spaces respectively, and vn and wn are bounded in H 1 (D), then there
are subsequences vnk and wnk which converge in H 1 (D).
Proof. We assume first that ξ · Re (A)ξ ≥ γ|ξ|2 (for the other case it is similar). Since vn
and wn are bounded in H 1 (D) and the identity from H 1 (D) to L2 (D) is compact, there
is a subsequence vnk and wnk which converge in L2 (D). Now we rewrite the problem as
follows:
∇ · A∇vnk − γvnk = −(γ + k 2 n)vnk in D
∆wnk − wnk = −(1 + k 2 )wnk in D
vnk − wnk = fnk on ∂D
∂vnl ∂wnk
− = hnk on ∂D
∂νA ∂ν
90 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

According to the well-posedness of the MITP,

γ+1
∥vnk ∥H 1 (D) + ∥wnk ∥H 1 (D) ≤C (γ + k 2 ∥n∥∞ )∥vnk ∥L2 (D) + (1 + k 2 )∥wnk ∥L2 (D)
γ−1


+∥fnk ∥H 1/2 (∂D) + ∥hnk ∥H −1/2 (∂D) .
Moreover, let k, ℓ ∈ N and sufficiently large, and v := vnk − vnl , w := wnk − wnl then it
solve the
∇ · A∇v − γv = −(γ + k 2 n)(vnk − vnl ) in D
∆w − w = −(1 + k 2 )(wnk − wnl ) in D
v−w = fnk − fnl on ∂D
∂v ∂w
− = hnk − hnl on ∂D
∂νA ∂ν
From the convergece of vnk and wnk in the L2 (D) space, as well as the convergece of fnk
and hnk , given any small ε > 0, there is a N >> 1 such that
(γ + k 2 ∥n∥∞ )∥vnk − vnl ∥L2 (D) + (1 + k 2 )∥wnk − wn,l ∥L2 (D)
+∥fnk − fnl ∥H 1/2 (∂D) + ∥hnk − hnl ∥H −1/2 (∂D) < ε.
This implies that for k, l > N ,
∥vnk − vnl ∥H 1 (D) + ∥wnk − wnl ∥H 1 (D) = ∥v∥H 1 (D) + ∥w∥H 1 (D) < Cε.
Thus vnk and wnk are two Cauchy sequences in H 1 (D) then they converge.
We still need one more lemma to get ready for the uniqueness result.
Lemma 7.7.2. If k 2 is not a Dirichlet eigenvalue for the bounded domain B with C 2
boundary and R2 \ B is connect, then
span{eikx·d |∂B : |d| = 1} = H 1/2 (∂B).
Proof. Suppose φ ∈ H −1/2 (∂B) such that
Z
φ(y)e−iky·d ds(y) = 0.
∂B

We define Z
u(x) := φ(y)Φ(x, y)ds(y), x ∈ R2 \ ∂B,
∂B

then u∞ = 0. This implies that u = 0 in R2 \ B. Thus u ∈ Hloc

1
(B) and u|∂B ∈ H 1/2 (∂B).
From jump relation,
∂u + 1 ′ 1
0= = K φ − φ.
∂ν 2 2
′ ′
Then φ = K φ. Recall that when φ ∈ H (∂B), K φ ∈ H 1/2 (∂B). Thus φ ∈ H 1/2 (∂B) ⊂
1/2

L2 (∂B). Use the result that K ′ is bounded from L2 (∂B) to H 1 (∂B) ⊂ C(∂B). Thus
u is continuous accross the boundary ∂B. Since u = 0 on ∂B and k 2 is not a Dirichlet
eigenvalue, u = 0 in B. With the jump relation, φ = 0. Thus the proof is finished.
7.7. UNIQUENESS OF THE INVERSE PROBLEMS 91

Now we are prepared to prove the uniqueness result.

Theorem 7.7.3. Let D1 and D2 be the domains of the inhomogenous media represented
by the matrix valued functions A1 and A2 , and the functions n1 and n2 . Assumptions
for the matrix and function are satisfied. If for all incident angle φ ∈ [0, 2π], the related
farfield pattern
u1∞ (θ, φ) = u2∞ (θ, φ), for all θ ∈ [0, 2π],
then D1 = D2 .

Proof. It is easily known that when u1∞ (θ, φ) = u2∞ (θ, φ), then us1 (x, φ) = us2 (x, φ) when
x ∈ R2 \ (D1 ∪ D2 ). The first task is to extend this result to incident point sources, i.e.,
when the incident field is Φ(x, z) with z ∈ R2 \ (D1 ∪ D2 ), then us1 (x, z) = us2 (x, z) for
x ∈ R2 \ (D1 ∪ D2 ). In this case, the boundary data are given by

∂
fj = Φ(·, z)|∂Dj , hj = Φ(·, z) , j = 1, 2.
∂ν ∂Dj

The idea is to approximate the incident field Φ(·, z) by plane waves. Choose a sufficiently
large disk Ω ⊃ D1 ∪ D2 such that z ∈/ Ω. Moreover, k 2 is not a Dirichlet eigenvalue of Ω.
Then
span{eikx·d |∂Ω : |d| = 1} = H 1/2 (∂Ω).
Let uin be a linear combination of plane waves such that

∥uin − Φ(·, z)∥H 1/2 (∂Ω) → 0, n → ∞.

Then
∥uin − Φ(·, z)∥H 1 (Ω) → 0, n → ∞.
From the continuous dependence of scattered fields on the incident waves,

∥usn,1 − us1 (·, z)∥H 1 (Ω\D1 ) , ∥usn,2 − us2 (·, z)∥H 1 (Ω\D2 ) → 0, n → ∞.

Since usn,1 = usn,2 when x ∈ Ω \ D1 ∪ D2 , we can also easily know that us1 (·, z) = us2 (·, z) in
Ω \ D1 ∪ D2 . Since the scattered fields are real-analytic, they are identical in R2 \ D1 ∪ D2 .
Again since the scattered fields are analytic w.r.t. z, us1 (x, z) = us2 (x, z) for x, z ∈ R2 \
D1 ∪ D2 .
We will prove the uniqueness by contradiction. Let D1e := R2 \ D1 and D2e := R2 \ D2 .
If D1 ̸⊂ D2 , then D1 ∩ D2e ̸= ∅. Thus there is a point z ∈ ∂D1 ∩ D2e and a small ε > 0 such
that

• B(z, 8ε) ∩ D2 = ∅.

• D1 ∩ B(z, 8ε) is contained in the connected component of D1 to which z belongs.

• The connect component of D1 to which z belongs has other points outside D1 ∩

B(z, 8ε).
92 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM

• Let zn := z + nε ν(z), then zn ∈ R2 \ D1 ∪ D2 for all n ∈ N, where ν(z) is the unit

normal of ∂D at z.

Then it is known that

∥Φ(·, zn )∥H 1 (D1 ) → ∞, as n → ∞.

Define
1
wn (x) := Φ(x, zn ), x ∈ D1 ∪ D2 .
∥Φ(·, zn )∥H 1 (D)
Thus ∥wn ∥H 1 (D) = 1 for all n and ∥wn ∥H 2 (D) → 0 as n → ∞. Let (v1n , un1 ) and (vn2 , u2n )
are solutions of the scattering problems with boundary data given by wn and its normal
derivative on ∂D1 and D2 , respectively. First from above arguments, since z ∈ R2 \
D1 ∪ D2 , u1n = u2n in R2 \ D1 ∪ D2 .
First, from the fact that ∥wn ∥H 2 (D) → 0 as n → ∞, then ∥un2 ∥H 1 (Ω\D1 ∪D2 ) → 0 as
n → ∞. Since un1 = un2 in R2 \ D1 ∪ D2 , ∥un1 ∥H 1 (Ω\D1 ∪D2 ) → 0 as well. Thus

∂un1
un1 |∂(D1 ∪D2 ) → 0, → 0, as n → ∞
∂ν ∂(D1 ∪D2 )

Choose a cutoff function X ∈ C0∞ (B(z, 8ε)) such that X (x) = 1 in B(z, 7ε), then

∂(X un1 )
X un1 |∂D1 → 0, → 0, as n → ∞
∂ν ∂D1

in H 1/2 and H −1/2 (exercise!) on the boundaries, respectively.

On the other hand, consider the domain Ω \ B(z, 2ε). Note that there is a constant
C > 0 such that
∥wn ∥H 2 (Ω\B(z,2ε)) ≤ C, ∀ n ∈ N.
vn, u
Consider the solution ((1 − X )v n ), (1 − X )un1 ) := (e en1 ). From direct computation, it is
a solution to the direct scattering problem with the right hand side given by

v n ∇ · A∇(1 − X ) + 2A∇(1 − X ) · ∇v n
un1 ∆(1 − X ) + 2∇(1 − X ) · ∇un1
(1 − X )wn
∂(1 − X ) ∂(1 − X ) ∂wn
v− u + (1 − X )
∂νA ∂ν ∂ν
From the definition of the cutoff function, the right hand side of this problem lies in
L2 (D1 ) × L2 (Ω \ D1 ) × H 3/2 (∂D1 ) × H 1/2 (∂D1 ). The solution u
en1 can reach the regularity
of H 2 (Ω \ D1 \ B(z, 2ε)). Since H 2 is compactly embedded in H 1 , there is a subsequence
en1 k such that it converges in H 1 . This implies that
u

∂(1 − X )un1 k )
(1 − X )un1 k |∂D1 ,
∂ν ∂D1
7.7. UNIQUENESS OF THE INVERSE PROBLEMS 93

converge in H 1/2 (∂D1 ) and H −1/2 (∂D1 ), respectively. Thus

∂un1 k )
un1 k |∂D1 ,
∂ν ∂D1

converge in H 1/2 (∂D1 ) and H −1/2 (∂D1 ).

Since v1nk and wnk satisfy the ITP with the right hand side given by

∂unk
f = un1 k , h=
∂ν
and convergent; the solutions are uniformly bounded, with the result of the lemma, there
is a subsequence of wnk , still denoted by wnk which converges to w in H 1 (D1 ). On one
hand,
∥wnk ∥H 1 (D1 ) , k = 1, 2, . . . .
Thus ∥w∥H 1 (D1 ) = 1. On the other hand, for any fixed point in D1 \ B(z, 2ε), Φ(·, z) is
uniformly bounded thus wn → 0. This implies that w = 0 in D1 \ B(z, 2ε). This argument
can be applied to any point in the interior of D1 thus w = 0 in D1 . This contradicts with
∥w∥H 1 (D1 ) = 1, thus D1 ̸⊂ D2 . The proof is finished.
94 CHAPTER 7. SCATTERING BY AN ORTHOTROPIC MEDIUM
Chapter 8

Scattering from inhomogeneous

medium

8.1 Lippmann-Schwinger equation

In this chapter, we consider the scattering from inhomogeneous medium:

∆u + k 2 nu = 0 in R2 , (8.1.1)

where u = ui + us and us satisfies the Sommerfeld radiation condition. Here n is the

refractive index which is always positive. The medium may also be abosrbing, that is to
say, Im (n) ≥ 0.
Assume that n piecewise continuous and m := 1 − n is compactly supported. Let

D := {x ∈ R2 : m(x) ̸= 0}.

Then we consider the volume potential:

Z
u(x) = Φ(x, y)φ(y)dy, x ∈ R2 ,
R2

where Φ(x, y) is the fundamental solution, φ is a continuous function with compact support
in R2 . The following theorem summarizes important properties of the volume potential.

Theorem 8.1.1. The volume potential u exists as an improper integral for all x ∈ R2 and
has the following properties. If φ ∈ C0 (R2 ) then u ∈ C 1,α (R2 ); if φ ∈ C0 (R2 ) ∩ C 0,α (R2 )
then u ∈ C 2,α (R2 ). In this case, u satisfies

∆u + k 2 u = −φ in R2 .

Moreover,
∥u∥2,α,R2 ≤ C∥φ∥α,R2
where C > 0 only depends on the support of φ. Moreover, if φ ∈ C0 (R2 ) ∩ C 1,α (R2 ) then
u ∈ C 3,α (R2 ).

95
96 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

However, from our assumption that n is only required to be piecewise continuous, the
solutions do not exist in C 2 spaces. Actually, the second order derivatives are no longer
continuous but still may be integrable. Thus we also need to study the problems in Sobolev
2
spaces Hloc (R2 ).
Theorem 8.1.2. Let D, G be two bounded domains in R2 . Then the volumen potential
Z
(V φ)(x) := Φ(x, y)φ(y)dy, x ∈ R2
D

is a bounded operator from L2 (D) to H 2 (G).

In the following, we will show that the scattering problem is equivalent to the problem
Z
i
u(x) = u (x) − Φ(x, y)m(y)u(y)dy, x ∈ R2 , (8.1.2)
R2

which is known as the Lippmann-Schwinger equation.

2
Theorem 8.1.3. If u ∈ Hloc (R2 ) solves (8.1.1), then it solves (8.1.2). Conversely, if
2 2
u ∈ C(R ) solves (8.1.2) then u ∈ Hloc (R2 ) solves (8.1.1).
2
Proof. Let u ∈ Hloc (R2 ) solves (8.1.1). Let x ∈ R2 be an arbitrary point and B a large
ball centered at 0 which contains both x and the support of m. Apply the Green’s formula
to u, then
Z   Z
∂u(y) ∂Φ(x, y)
u(x) = Φ(x, y) − u(y) ds(y) − (∆u(y) + k 2 u(y))Φ(x, y)dy
∂B ∂ν(y) ∂ν(y) B
Z   Z
∂u(y) ∂Φ(x, y) 2
= Φ(x, y) − u(y) ds(y) − k Φ(x, y)m(y)u(y)dy, x ∈ B.
∂B ∂ν(y) ∂ν(y) B

Recall the Green’s theorems in the intereior and exterior of B again, we have:
Z  i 
i ∂u (y) i ∂Φ(x, y)
u (x) = Φ(x, y) − u (y) ds(y), x ∈ B.
∂B ∂ν(y) ∂ν(y)
Moreover, since both us and Φ(x, y) satisfies the radiation condition,
Z  s 
∂u (y) s ∂Φ(x, y)
Φ(x, y) − u (y) ds(y) = 0.
∂B ∂ν(y) ∂ν(y)
Thus (8.1.2) is obtained immediately.
On the other hand, when u ∈ C(R2 ) satisfies (8.1.2), and let
Z
s 2
u (x) := −k Φ(x, y)m(y)u(y)dy, x ∈ R2 ,
R2

then us ∈ Hloc
2
(R2 ) and satisfies the Sommerfeld radiation condition since m is compactly
supported. Moreover,
∆us + k 2 us = k 2 mu.
Since ∆ui + k 2 ui = 0, we have
∆u + k 2 u = k 2 mu
thus u solves (8.1.2). The proof is finished.
8.1. LIPPMANN-SCHWINGER EQUATION 97

We will show that the Lippmann-Schwinger equation (8.1.2) is uniquely solvable for
k > 0. But this is easy only when k is small. We leave the larger k’s to the next section
and only discuss the small k’s.

Theorem 8.1.4. Consider the problem in R3 . Suppose m = 0 for |x| > a where a > 0
such that k 2 < M2a2 , where M = sup|x|≤a |m(x)|. Then the equation (8.1.2) has a unique
solution.

Proof. Let B be the circle centered at 0 with radius a > 0. When ui = 0, suppose us is
the scattered field then
Z
s 2
u (x) = −k Φ(x, y)m(y)us (y), x ∈ R3 .
R3

Define the operator

Z
(Tm u)(x) = Φ(x, y)m(y)u(y)dy, x ∈ R3 .
B

Then
us = −k 2 Tm us .
Since m is piecewise continuous, Tm is a bounded operator from C(B) to C(B). More-
over,
Z
∥Tm u∥∞ ≤ |Φ(x, y)| |mu|dy
B Z
≤ ∥u∥∞ ∥m∥∞ |Φ(x, y)|dy
B
M ∥u∥∞
Z
1
≤ dy.
4π B |x − y|

Define the function Z

1
h(x) := dy, x ∈ B.
B |x − y|
1
Recall that the function 4π|x−y|
is the fundamental solution of the Poisson equation, h
satisfies
∆h = −4π.
Moreover, since y is located in a disk, h does not depend on the angle. This is easily
checked. Use the spherical coordinate,

∂2
   
1 ∂ 2 ∂ 1 1 ∂ ∂
∆= 2 r + 2 2 + sin φ .
r ∂r ∂r r sin φ ∂θ2 r2 sin φ ∂φ ∂φ

Since h does not depend on θ,

 
1 ∂ 2 ∂h
∆h = 2 r = −4π.
r ∂r ∂r
98 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

The first integral results in that

∂h 4πr3
r2 =− + C1 ,
∂r 3
which implies
∂h 4πr C1
=− + 2.
∂r 3 r
Second integral results in
2πr2 C1
h(r) = − − + C2
3 r
where C1 , C2 are arbitrary constants. First observe that h is continuous at 0, C1 = 0.
Then Z
1
h(0) = C2 = dy = 2πa2 .
B |y|
Thus
2πr2
h(x) = − + 2πa2 ≤ 2πa2 .
3
Then
M M a2
∥Tm u∥∞ ≤ 2πa2 ∥u∥∞ = ∥u∥∞ .
4π 2
Recall that us = −k 2 Tm us , it implies that

M a2 k 2 s
∥us ∥∞ ≤ k 2 ∥Tm ∥∥us ∥∞ ≤ ∥u ∥∞ .
2
Since M k 2 a2 < 2, the above inequality holds only if us = 0 in B. Thus the uniqueness is
proved.

For general positive wavenumber k > 0, first observe that Tm is a compact operator
from C(B) to C(B), since H 2 (B) is compactly embedded in C(B). Thus the problem is
written as
(I + k 2 Tm )u = ui
which is of the Fredholm type. Thus the uniqueness implies the unique solvability. How-
ever, further techniques are needed which leaves to the next section.

8.2 The unique continuation principle

Now we focus on the general wavenumber k > 0 and consider the uniqueness of the solution
to (8.1.1) where u = ui + us with us satisfying the Sommerfeld radiation condition. To
study the uniqueness, we always want to show that when ui = 0, then us must be 0 as
well. Note that when ui = 0, u = us which saisfies the Sommerfeld radiation condition.
Then we will show that if u is a solution to (8.1.1) which satisfies the Sommerfeld radiation
condition, then u = 0.
8.2. THE UNIQUE CONTINUATION PRINCIPLE 99

Recall the corollary of Rellich’s lemma (Corollary 3.4.4): if us is a radiating solution

to the Helmholtz equation in R3 \ D and
∂us
Z
Im us dS ≥ 0,
∂D ∂ν
then us = 0 in R2 \ D.
We first apply this result to this problem. Multiply the equation (8.1.1) on both sides
by u and apply the Green’s theorem in B, we have:
Z Z
2 2 2 ∂u
[|∇u| − k n|u| ]dx − udS = 0.
B ∂B ∂ν

Thus Z Z
∂u
udS = k 2 Im n|u|2 dS ≥ 0.
∂B ∂ν ∂B
2
Thus u = 0 in R \ B. So the major question is: How to prove that u = 0 in B? To answer
this question, the unique continuation principle is introduced.
Lemma 8.2.1. Let G ⊂ R2 be a domain and u1 , . . . , up ∈ H 2 (G) be real valued functions
satisfying
P
X
|∆up | ≤ c {|uq | + |∇uq |} in G
q=1

for p = 1, 2, . . . , P and some constant c > 0. Assume that up vanishes in a neighbourhood

of some x0 ∈ G for p = 1, 2, . . . , P . Then up is identically zero in G for p = 1, 2, . . . , P .
Before the proof of this lemma, we first apply it to our proplem. Let u be the solution
of (8.1.1) with the Sommerfeld radiation condition. It is already known that u = 0 in
R2 \ B thus it is 0 in B \ D (we can always choose B a little larger). Since u is complex
valued, let
u1 := Re (u), u2 := Im (u),
and it is obvious that u1 = u2 = 0 in D\B. For simplicity, let n = n1 +in2 then n1 ≥ γ > 0
and n2 ≥ 0 are piecewise continuous functions. Moreover, take the real- and imaginary
parts of (8.1.1) we have:
∆u1 + k 2 n1 u1 − k 2 n2 u2 = 0,
∆u2 + k 2 n1 u2 + k 2 n2 u1 = 0.
Thus
|∆u1 |, |∆u2 | ≤ c(|u1 | + |u2 |)
holds with a constant c > 0 which only depends on k and n. Apply Lemma 8.2.1 we get
u1 = u2 = 0 in B thus u = 0 in R2 . This result is concluded in the final theorem in this
section.
2
Theorem 8.2.2. For each k > 0, there is a unique solution u ∈ Hloc (R2 ) to (8.1.1), and
∥u∥∞ ≤ C∥ui ∥∞ .
100 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

The rest of the section is focused on the proof of Lemma 8.2.1.

Proof of Lemma 8.2.1. Main idea: given any other point x1 ∈ G, it is connected to x0 via
a finite curve S. For x0 , there is a 0 < R ≤ 1 such that B(x0 , R) ⊂ G. We will show
that up (x) = 0 for all x ∈ B(x0 , R/2). Then for a point x2 ∈ S ∩ B(x0 , R/2), up (x2 ) = 0
thus we can repeat the proof. For x2 , there is a R2 ∈ (0, 1] such that B(x2 , R2 ) ⊂ G and
up (x) = 0 in B(x2 , R2 /2). Thus we can find a sequence xn ∈ S with radius Rn . Only finite
number of steps are necessary since S is compact and Rn ≥ R0 (the curve is away from
the boundary of G). Then we finally prove that up (x1 ) = 0.
Without loss of generality, let x0 = 0 thus up (0) = 0 for p = 1, 2, . . . , P . We need to
introduce some functions to help us in the proof. We temporarily igonro the subscripts
but simply deonte by u and v. Let
−n
v(x) := er u(x), x ̸= 0

and v(0) = 0. Here n ∈ N is a parameter. Thus v ∈ H 2 (G) since u ∈ H ( G) and vanishes

near x0 = 0. However, note that the above functions are defined in G but we are only
interested in their behaviours in B(0, R). To this end, we introduce a cutoff function. Let
φ ∈ C 2 (R2 ) such that φ(x) = 1 when |x| ≤ R/2 and φ(x) = 0 for |x| ≥ R. Let û = φu
and v̂ = φv, then both are compactly supported in B(0, R). Then
  
−r−n 2n ∂v̂ n n −n
∆û = ∆(e v̂(x)) = ∆v̂ + n+1 + n+2 n − n + 1 v̂ e−r .
r ∂r r r
Then
4n ∂v̂ 
−n n n  
(∆û)2 ≥ e−2r ∆v̂ + − n + 1 v̂ .
rn+1 ∂r rn+2 rn
Then Z Z
n+2 2r−n 2 ∂v̂  n n  
r e (∆û) dx ≥ 4n r ∆v̂ + n+2 n − n + 1 v̂ dx.
G G ∂r r r
Now we need to work on the terms on the right hand side. First, note that
∂v̂
r = x · ∇v̂,
∂r
then
∂v̂
r ∆v̂ = ∇ · ((x · ∇v̂)∇v̂) − ∇(x · ∇v̂) · ∇v̂.
∂r
Since v̂ = 0 on ∂G, Z Z
∂v̂
r ∆v̂dx = − ∇(x · ∇v̂) · ∇v̂dx.
G ∂r G
Note that
1
∇(x · ∇v̂) · ∇v̂ = |∇v̂|2 + x · ∇|∇v̂|2
2
1 1
= |∇v̂| + ∇ · (x|∇v̂|2 ) − (∇ · x)|∇v̂|2
2
2 2
1 2 1 2
= ∇ · (x|∇v̂| ) − |∇v̂| .
2 2
8.2. THE UNIQUE CONTINUATION PRINCIPLE 101

Thus Z Z
∂v̂ 1
r ∆v̂dx = |∇v̂|2 dx.
G ∂r 2 G
For the second and third term, let’s consider the general form
Z Z π Z 2π Z R
1 ∂v̂ 1 ∂v̂
v̂
m ∂r
dx = sin θdθdφ v̂ dr
m−1 ∂r
G r 0 0 0 r
Z π Z 2π Z R  
∂ v̂
=− sin θdθdφ v̂ dr
0 0 0 ∂r rm−2
Z π Z 2π Z R  
v̂ 1 ∂v̂
=− sin θdθdφ v̂ (2 − m) m−1 + m−2 dr
0 0 0 r r ∂r
|v̂|2
Z Z
v̂ ∂v̂
=− m
dx + (m − 2) m+1
dx.
G r ∂r G r

Thus
2n − 1 v̂ 2
Z Z
∂v̂ 1
v̂ 2n+1
dx = dx,
G ∂r r 2 G r2n+2
and
n−1 v̂ 2
Z Z
∂v̂ 1
v̂ n+1
dx = dx.
G ∂r r 2 G rn+2
Finally,

v̂ 2 v̂ 2
Z Z Z Z
n+2 2r−n 2 2 3 2 2
r e (∆û) dx ≥ 2n |∇v̂| dx + 2n (2n − 1) 2n+2
dx − 2n (n − 1) n+2
dx
G G G r G r
v̂ 2
Z Z
2 2 2
≥ 2n |∇v̂| dx + 2n (n + n − 1) 2n+2
dx
G G r

From the relationship between v̂ and v̂,

−n nx
∇û = e−r (∇v̂ + v̂),
r+2

2r−n 2n2
e |∇û| ≤ 2|∇v̂| + 2n+2 |v̂|2 ,
2 2
r
thus −n
e2r
Z Z Z
n+2 2r−n 2 2r−n 2 4
r e (∆û) dx ≥ n e |∇û| dx + n 2n+2
û2 dx.
G G G r
Return to the original condition, it is clear that
P 
|∇uq |2 |uq |2

2 2
X R
|∆up | ≤ 2P c + 3n+4 , |x| ≤ .
q=1
rn+2 r 2

Moreover,
|∆ûp |2 R
|∆ûp |2 ≤ , ≤ |x| ≤ R.
r3n+4 2
102 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

Then
−n
e2r
Z Z
2r−n 2 4
n e |∇up | dx + n 2n+2
u2p dx
G G r
Z
2r−n
≤ rn+2 e (∆û)2p dx
G
P −n
!
e2r
Z Z
−n
X
≤2P c2 e2r |∇uq |2 dx + 2n+2
u2q
q=1 |x|≤R/2 |x|≤R/2 r
−n
e2r (∆ûp )2
Z
+ dx.
R/2≤|x|≤R r2n+2

When n is sufficiently large,

Z 2r−n −n
e2r (∆ûp )2
Z
4 e 2
n u dx ≤ C
2n+2 p
dx
G r R/2≤|x|≤R r2n+2

for some constant C > 0. Since the function

−n
e2r
r→
7 , r>0
r2n+2
is strictly decreasing, Z Z
4
n u2p dx ≤C ∆û2p dx.
G R/2≤|x|≤R

Let n → ∞, we finally get u = 0 in B(0, R/2).

8.3 Uniqueness results for the inverse problems

Similar as before, we can still define the far field pattern. Given the incident field

ui (x) = eikx·d ; d ∈ S2 ,

then we get the scattered field us (x) which results in the far field pattern u∞ (x̂, d) where
x
x̂ = |x| ∈ S2 . Here
S2 := {d ∈ R3 : |d| = 1}.
We will study the uniqueness based on the complex geometric optics solutions.
Lemma 8.3.1. Let B be a ball containing the support of m := 1 − n. Then there exists a
constant C > 0 such that for each z ∈ C3 with z · z = k 2 and |Im (z)| > 2k 2 ∥n∥∞ , there
exists a solution v ∈ H 2 (B) of ∆v + k 2 nv = 0 in B in the form of

v(x) = eiz·x (1 − w(x))

where
C
∥w∥L2 (B) ≤ .
|Re (z)|
8.3. UNIQUENESS RESULTS FOR THE INVERSE PROBLEMS 103

To prove the uniqueness, we need a further lemma.

Lemma 8.3.2. Let B ⊂ B0 ⊂ R3 where B0 is a larger ball. Then the set
{u(·, d) : d ∈ S2 }
which satisfies the Helmholtz equation is complete in the closure of
{v ∈ H 2 (B0 ) : ∆v + k 2 nv = 0 in B0 }
with respect to the L2 (B) norm.
Proof. Let G(x, y) be the fundamental solution, i.e.,
∆x G(x, y) + k 2 n(x)G(x, y) = −δ(x − y).
Then
Z
2
G(x, y) = Φ(x, y) − k Φ(x, z)m(z)G(z, y)dz
B
eikr −ikx̂·y
 Z 
−ikx̂·z
= e −k 2
e m(z)G(z, y) + O(r−2 ).
4πr B

Since
∆us + k 2 nus = k 2 mui ,
then Z
s 2
u (x) = −k G(x, y)m(y)ui (y)dy
B
and for any fixed d, Z
ikx·d 2
u(x) = e −k G(x, y)m(y)eikd·y dy.
B
Thus
eikr
G(x, y) = u(−x̂, y) + O(r−2 ).
4πr
Let φ ∈ H 2 (B) be a solution to ∆φ + k 2 nφ = 0 which is orthogonal with all u(x, d) in B
and let Z
w(x) = G(x, y)φ(y)dy.
B
R
Then w is a radiating solution of ∆w + k 2 nw = −φ. Since B uφdx = 0,
eikr
Z
w(x) = u(−x̂, y)φdy + O(r−2 ) = O(r−2 )
4πr B
thus w∞ = 0. This implies that w = 0 in R3 \ B. Since ∆φ + k 2 nφ = 0 in B0 ,
Z Z
|φ| dx = − (∆w + k 2 nw)φdx
2
B Z B
= w(∆φ + k 2 nφ)dx = 0.
B

Thus φ = 0 in B. The proof is finished.

104 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

Now we are prepared to prove the uniqueness. Let n1 and n2 be two refractive indices
such that
u1∞ (·, d) = u2∞ (·, d),
and B be the ball that contains the supports of 1 − n1 and 1 − n2 . Then u1 (·, d) = u2 (·, d)
for all d ∈ S2 in R3 \ B.
For any arbitrary d1 , d2 ∈ S2 , since

∆u1 (·, d1 ) + k 2 n1 u1 (·, d1 ) = 0,

multiply on both sides by u2 (·, d2 ),

Z
 1
∆u (x, d1 ) + k 2 n1 u1 (x, d1 ) u2 (x, d2 )dx

0=−
Z B
 1
∇u (x, d1 ) · ∇u2 (x, d2 ) − k 2 n1 u1 (x, d1 )u2 (x, d2 ) dx.

=
B

Similarly we have
Z
 2
∇u (x, d2 ) · ∇u1 (x, d1 ) − k 2 n2 u2 (x, d2 )u1 (x, d1 ) dx = 0.

B

Thus we have Z
(n1 − n2 )u1 (x, d1 )u2 (x, d2 )dx = 0
B

for all d1 and d2 . From the density arguments, it is obtained immediately that
Z
(n1 − n2 )v 1 v 2 dx = 0
B

for all v1 , v2 ∈ H 2 (B0 ) such that

∆vj + k 2 nj vj = 0 in B0 ⊂ B.

Now we apply the geometric optics solutions. Given any z1 ∈ C3 such that z · z = 0
and |Re (z)| > 2k 2 ∥n∥∞ , there is a solution

v1 (x) = eiz1 ·x (1 − w1 (x))

with
C
∥w1 ∥L2 (B) ≤ .
|Im (z2 )|
Similarly,
v2 (x) = eiz2 ·x (1 − w2 (x))
with
C
∥w2 ∥L2 (B) ≤ .
|Im (z2 )|
8.3. UNIQUENESS RESULTS FOR THE INVERSE PROBLEMS 105

Now we construct the special z1 and z2 . For any ξ ∈ R3 , let

z1 = ξ + ζ, z2 = ξ − ζ,
where ζ ∈ C3 such that z1 · z1 = z2 · z2 = 0 and |Re (z1 )|, |Re (z1 )| >> 1. To this end, let
ζ = a + iϱb
where a, b ∈ R3 with |b| = 1 and ϱ > 0. Then
|Im (z1 )| = |Im (z2 )| = ϱa
which are sufficiently large when ϱ is large. Thus we only need to determine b. Since
z1 · z1 = z2 · z2 = k 2 ,
|ξ|2 + |a|2 − ϱ2 + 2ξ · a + 2iϱξ · b + 2iϱa · b = k 2 ;
|ξ|2 + |a|2 − ϱ2 − 2ξ · a − 2iϱξ · b + 2iϱa · b = k 2 .
Compare the real- and imaginary parts, we have that
ξ · a = 0;
ξ · b = 0;
a · b = 0;
|b|2 = k 2 + ϱ2 − |ξ|2 .
Thus b is also determined up to a sign. This means that we have already obtained z1 and
z2 , then Z
(n1 − n2 )e2iξ·x (1 − w1 )(1 − w2 )dx = 0.
B
By choosing larger ϱ, it is known that |Im (z1 )|, |Im (z2 )| = ϱ. Then
C
∥w1 ∥L2 (B) , ∥w2 ∥L2 (B) ≤ .
ϱ
Thus
Z Z
2iξ·x
(n1 − n2 )e dx = (n1 − n2 )e2iξ·x (w1 + w2 − w1 w2 )dx
B B
≤ ∥n1 − n2 ∥L2 (B) ∥w1 ∥L2 (B) + ∥n1 − n2 ∥L2 (B) ∥w2 ∥L2 (B)
+ ∥n1 − n2 ∥∞ ∥w1 ∥L2 (B) ∥w2 ∥L2 (B)
≤ Cϱ−1 ,
which tends to 0 when ϱ → ∞. Since the left hand side does not depend on ϱ,
Z
(n1 − n2 )e2iξ·x dx = 0
B
3
for any fixed ξ ∈ R . When ξ = 0, the choice of a and b becomes multiple. This implies
that Fourier transform of n1 − n2 is always 0 then n1 = n2 .
Note that the method no longer works in 2D. The reason is the choice of z1 and z2 .
Still, let
z1 = ξ + a + iϱb, z2 = ξ − a − iϱb.
From z1 · z1 = z2 · z2 = k 2 , we still have that (ξ, a, b) is a orthogonal basis in R2 , which is
impossible.
106 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

8.4 Stability estimates for the inverse problems

As we mentioned, the inverse problems are always ill-posed. But how bad it is? Let’s
answer the question in this section for the inverse medium scattering problems.
From the previous arguments, it is already known that the far-field pattern with all
incident and oberservation directions uniquely determines the refractive index n. However,
the stability is still not investigated, i.e., if the far-field pattern is perturbed a little, will
the reconstructed refractive index be a little perturbation of the exact one? Let’s check
that.
Suppose we have a refractive index n and it’s perturbation n e, where both 1 − n and
1−n e are compactly supported in a ball B centered at 0 with radius 1. Let ui := Φ(x, y)
be the incident field, then let the total field wn (x, y) := Φ(x, y) + wns (x, y) is the Green’s
function for the problem. Similarly we have wne (x, y) := Φ(x, y) + wnes (x, y). We only need
to consider the Green’s function since it contains all important informaltion. We submit
incident field on ∂BR and then receive the near-field data on ∂BR as well.
The following theorems gives the estimation.
Theorem 8.4.1. There is a constant C > 0 which depends only on s, k, R, Cn where
∥1 − n∥H s (B) , ∥1 − n
e∥H s (B) ≤ Cn such that
s

− s+3
e∥L2 (B1 ) ≤ C − ln− ∥wns − wnes ∥L2 (∂BR ×∂BR )

∥n − n ; (8.4.1)
2s−3
− 2s+3
e∥L∞ (B1 ) ≤ C − ln− ∥wns − wnes ∥L2 (∂BR ×∂BR )


∥n − n . (8.4.2)
Here ln− (t) = ln(t) when t ≤ 1/e and ln− (t) = −1 otherwise.

Figure 8.1: Decay rate of the function [− ln(t)]−1 .

The similar result for the far-field pattern w.r.t. incident plane waves is described as
follows.
s
Theorem 8.4.2. Let 0 < ε < s+3 be a fixed constant. There is a constant C > 0 which
depends only on s, k, R, Cn where ∥1 − n∥H s (B) , ∥1 − ne∥H s (B) ≤ Cn such that

− s+3s
+ε
e∥L2 (B1 ) ≤ C − ln− ∥u∞ ∞

∥n − n n − u n
e ∥ L2 (∂B ×∂B )
R R
; (8.4.3)
2s−3
− 2s+3 +ε
e∥L∞ (B1 ) ≤ C − ln− ∥u∞ ∞


∥n − n n − un e ∥L2 (∂BR ×∂BR ) . (8.4.4)
Here ln− (t) = ln(t) when t ≤ 1/e and ln− (t) = −1 otherwise.
8.4. STABILITY ESTIMATES FOR THE INVERSE PROBLEMS 107

Since any solution can be written by the Green’s function, we only focus on the proof
of the first theorem. The idea is to apply the Fourier transform. Let
Z
1
fˆ(j) := f (x)e−ix·j dx.
(2π)3 (−π,π)3

Then X
n−n
e= \
(n e)(j)eij·x ,
−n x ∈ B1 .
j∈Z3

We want to study the Fourier coefficients separately, i.e., when |j| is small and large. The
larger part is relatively easier.
Since both n and n e are functions in H s ,
X X
∥n∥2H s (B1 ) = (1 + |j|2 )s |n̂(j)|2 < ∞, ∥e
n∥2H s (B1 ) = ˆ (j)|2 < ∞
(1 + |j|2 )s |n
e
j∈Z3 j∈Z3

Then we can estimate the following term:

X 2 1 X 2
\
(n −n
e)(j) ≤ 2 s
(1 + |j|2 )s (n
\ −n
e)(j)
(1 + ϱ )
|j|>ϱ |j|>ϱ
c
≤ 2s
ϱ
where c > 0 depends on s and ∥1 − n∥H s , ∥1 − n
e∥H s < Cn .
Now we need to turn to the |j| ≤ ϱ casese. To study these coefficients, we need the
following lemma with the help of the geometrical optics solutions. First, define
Z
(Sn φ)(x) := wn (x, y)φ(y)ds(y), s ∈ ∂BR .
∂B

Here wn is the Green’s function for the problem. Similarly we can define Sne . This is easily
proved that
∥Sn − Sne ∥L2 (∂BR ) ≤ ∥wns − wnes ∥L2 (∂BR ×∂BR ) .
Lemma 8.4.3. Assume 1 < R < R′ . There is a positive constant C such that all solutions
u ∈ C 2 (BR′ ) ∩ L2 (BR′ ) to ∆u + k 2 nu = 0 and u
e with ne,
Z
(n − n udx ≤ C∥Sn − Sne ∥L2 (∂BR ) ∥u∥L2 (BR′ ) ∥e
e)ue u∥L2 (BR′ ) .
B1

Proof. Define the function v as an extension of u|BR such that v|BR = u|BR ; while

∆v + k 2 v = 0 in R3 \ BR ; v = u on ∂BR

is a radiating solution. Similar to the standard fundamental solution, we can have (exer-
cise)  
∂v − ∂v +
v = Sn − on ∂BR .
∂ν ∂ν
108 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

Similarly, we can define ve in the same way.

From the equations for u and u e,
Z Z  
2 ∂e
u ∂u
k (n − n
e)ue
udx = u −ue ds
B1 ∂BR ∂ν ∂ν
Z  
v −
∂e ∂v −
= v − ve ds
∂BR ∂ν ∂ν
Z   Z  
v − ∂e
∂e v + ∂v − ∂v +
= v − − ve − ds
∂BR ∂ν ∂ν ∂BR ∂ν ∂ν
Z    
∂v − ∂v + v − ∂e
∂e v +
= − (Sn − Sne ) − ds.
∂BR ∂ν ∂ν ∂ν ∂ν

Therefore,
Z − −
1 ∂v ∂v + ∂e
v ∂e
v +
(n − n udx ≤ 2 ∥Sn − Sne ∥L2 (∂BR )
e)ue − −
B1 k ∂ν ∂ν L2 (BR ) ∂ν ∂ν L2 (BR )

≤ C∥Sn − Sne ∥L2 (∂BR ) ∥u∥L2 (BR′ ) ∥e

u∥L2 (BR′ ) .

With this lemma, we are prepared to estimate the Fourier coefficients j ∈ Z3 .

Lemma 8.4.4. Assume ϱ ≥ 2. When t > t0 where t0 > k is sufficiently large constant,
there is a constant C > 0 such that
 
\ 4R(t+ϱ) 1
(n −n
e)(j) ≤ C e ∥Sn − Sne ∥L2 (∂BR ) + .
t

Proof. Apply the geometric optics solutions. Let j ∈ Z3 and (d1 , d2 , j) be a orthogonal
basis with |d1 | = |d2 | = 1. Define
r
1 |j|2
ζt := − j + i t2 − k 2 + d1 + td2 ∈ C3 ;
2 r 4
1 |j|2
ζet := − j − i t2 − k 2 + d1 − td2 ∈ C3 .
2 4
Then
ζt + ζet = −j, ζt · ζt = ζet · ζet = k 2 .
Moreover, |Im (ζt )|, |Im (ζet )| ≥ ct. Then there exist geometric optics solutions

U (x, ζt ) = eiζt ·x (1 + v(x, ζt )), x ∈ B2R ;

e (x, ζet ) = eiζet ·x (1 + ve(x, ζet )),
U x ∈ B2R .

Then
e (x, ζet ) = e−ij·x (1 + p(x, t))
U (x, ζt )U
8.4. STABILITY ESTIMATES FOR THE INVERSE PROBLEMS 109

where Z
c
|p(x, t)|dx ≤ .
B1 t
Then
Z
1
\
(n −n
e)(j) = 3
e)e−ij·x dx
(n − n
(2π) B1
Z Z
≤C (n − n e)U (x, ζt )U (x, ζt )dx + C
e e (n − n e)p(x, t)dx
B1 B1
 
1
≤ C ∥Sn − Sne ∥L2 (∂BR ) ∥U (x, ζt )∥L2 (B2R ) ∥Ue (x, ζt )∥L2 (B ) +
e
2R
t
 
1
≤ C e4R(t+ϱ) ∥Sn − Sne ∥L2 (∂BR ) +
t
 
4R(t+ϱ) 1
≤C e ∥Sn − Sne ∥L2 (∂BR ) + .
t

Now we are prepared to prove the main theorem. From direct computation,
X 2
e∥2L2 (B1 ) =
∥n − n \
(n −n
e)(j)
j∈Z3
c X
\
2
≤ 2s
+ (n −ne)(j)
ϱ
|j|≤ϱ

ϱ3
 
c 3 4R(t+ϱ) 1
≤ 2s + cϱ e ∥Sn − Sne ∥L2 (∂BR ) + + s
ϱ t ϱ
2
ϱ3

(4R+1)(t+ϱ) s s 1
≤c e ∥wn − wne ∥L2 (∂BR ×∂BR ) + + s .
t ϱ

Let t sufficiently large and ϱ := t1/(s+3) with ϱ ≥ 2, then

 
2
∥n − ne∥L2 (B) ≤ c e(8R+2)t ∥wns − wnes ∥L2 (∂BR ×∂BR ) + .
ts/(s+3)

Assume that ∥wns − wnes ∥ is sufficiently small, and let

1 3
t := − ln (∥wns − wnes ∥) .
s + 3 8R + 2
Then

e∥L2 (B1 ) ≤ c ∥wns − wnes ∥s/(s+3) + (− ln ∥wns − wnes ∥)−s/(s+3)

∥n − n
≤ c(− ln ∥wns − wnes ∥)−s/(s+3) .

The proof is finished.

110 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

8.5 Numerical solution to the inverse medium prob-

lem
Although the fast imaging method reconstructs shapes of the obstacles very efficiently, we
still want to know the refractive index inside the obstacle, which is not possible for the
FIM. Thus we need to give a detailed description of the algorithm. For a brief introduction
we can get back to Chapter 6.3.
For fixed wavenumber k > 0 and incident field ui , we define the operator:
F : L2 (B) → L2 [0, 2π]
n 7→ u∞ .
Suppose that n − 1 is always compactly supported in the ball centered at 0 with radius
R > 0. From previous section, it is already knownt hat the equation is severely ill-posed.
Thus the Tikhonov regularization technique is always applied. We need to rewrite the
problem as an optimization problem. Let the measured data be U (x̂) with the fixed
wavenumber and incident field, and the noise is limited:
∥U − u∞ ∥L2 [0,2π] < δ.
Then we aim to solve the equation:
F(n) = U.
This problem is not easily solvable, since we need some norm to determine the distance
between two vectors. The 2-norm is the most popular one since it is differentiable. Then
the problem is written as an optimization problem:
min ∥F(n) − U ∥2L2 [0,2π] .
n∈L2 (B)

Since the problem is ill-posed, we apply the Tikhonov regularization:

n o
2 2
min
2
∥F(n) − U ∥ L2 [0,2π] + α∥n∥L2 (B) ,
n∈L (B)

where α > 0 is the Tikhonove regularization parameter. Now the question is, how to solve
this problem.
Well known methods are available to solve the nonlinear optimization methods. Here
we only take an example of the Newton-CG method (the Newton-method combined with
the Conjugate-Gradient method). From now on we focus only on real valued refractive
indexes n. Moreover, it is convenient to represent n in terms of its Fourier series:
X
n(x) = Cj φj (x), x ∈ B,
j∈Z3

where φj is either sin(j · x) or cos(j · x). Since we are only able to compute finite series,
we look for the solution to the problem
( N
)
X
min
2
∥P (C1 , . . . , CN ) − U ∥2L2 [0,2π] + α Cj2 ,
n∈L (B)
j=1
8.5. NUMERICAL SOLUTION TO THE INVERSE MEDIUM PROBLEM 111

and let Cℓ to indicate the vector {C1 , . . . , CN } in the ℓ-th step.

Algorithm 1 Newton-CG Method

Input: Data U ; ε > 0; j = 0.
Initialization: C0 = (0, . . . , 0) ∈ RN .
PN
1: while ∥P (Cℓ ) − U ∥2L2 [0,2π] + α j=1 ∥Cℓ ∥2ℓ2 > ε∥U ∥2L2 [0,2π] do
CGN E iteration scheme to solve (DP )∗ (DP ) + αI (H) = (DP )∗ (U − P (Cℓ ))
 
2:
3: Cℓ+1 = Cℓ + H;
4: j = j + 1;
5: end while

The Conjugate Gradient method is applied to solve the equation, for simplicity,
Ax = b,
where
A = (DP )∗ (DP ) + αI ;
 

b = (DP )∗ (U − P (Cℓ )),

so A is real valued, symmetric and positive-definite. The method is described in the
following algorithm.

Algorithm 2 CGNE
Input: r0 = b − Ax
Initialization: p0 = r0 , ℓ = 0
1: while ∥rℓ−1 ∥ ≥ γ∥r0 ∥ do
r⊤ rℓ
2: αℓ = p⊤ℓAp
ℓ ℓ
3: xℓ+1 = xℓ + αℓ pℓ
4: rℓ+1 = rℓ − αℓ Apℓ
r⊤ rℓ+1
5: βℓ = ℓ+1 rℓ⊤ rℓ
6: pℓ+1 = rℓ+1 + βℓ pℓ
7: ℓ = ℓ + 1
8: end while

Other algorithms are similar, the major steps involves multiple computations of DP
and (DP )∗ ) (equivalently DF and its adjoint).
After the update of the coefficient Cj , we always need to solve the forward problem,
which is given by the Lippmann-Schwinger equation:
Z
i 2
u(x) = u (x) − k Φ(x, y)m(y)u(y)dy.
B
(1)
Since Φ(x, y) = 4i H0 (k|x − y|), Note that m is supported in D ⊂ B where B is centered
at 0 with radius R, we are only interested in the solution u with x ∈ B. Thus x − y lies
in B ′ centered at 0 with radius 2R. Thus we take the following steps.
112 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM

(1)
For simplicity, let B ′ ⊂ [−π, π]3 . Let Φ(x) := 4i H0 (k|x|) and multiply the function
with a smooth cutoff function X (|x|) which equals to 1 in B ′ and 0 near the boundary of
[−π, π]3 . Then extend the function as a periodic function in R3 . Then it has a Fourier
series: X
Φ(x) := aj eij·x .
j∈Z3

Then in B ′ , X
Φ(x, y) = aj eij·(x−y) .
j∈Z3

Then Z X
i 2
u(x) = u (x) − k aj eij·(x−y) m(y)u(y)dy.
B j∈Z3

Here u is no longer the solution to the whole scattering problem. It is the solution to the
new problem with a periodic Φ. If u is also extended into a Fourier series,
X
u= uj eij·x ,
j∈Z3

then Z
X X X
uj eij·x
= uij eij·x −k 2 ij·x
aj e e−ij·y m(y)u(y)dy.
j∈Z3 j∈Z3 j∈Z3 B

Note that Z X
e−ij·y m(y)u(y)dy = mj−n un .
B n∈Z3

Now we need to truncate the Fourier series. Let Z3N := {(j1 , j2 , j3 ) : −N ≤ j1 , j2 , j3 ≤

N } ⊂ Z and uN := j∈Z3 uj eij·x , then
3
P
N

X X X X
uj eij·x = uij eij·x − k 2 aj eij·x mj−n un
j∈Z3N j∈Z3N j∈Z3N n∈Z3N

Thus we need mj in Z32N . The problem can be written as

U + k 2 A · [M ∗ U ] = U i

where U , U i and M are the vectors of related Fourier coefficients. This is a linear system
thus there are several methods to solve it.
In the Conjugate-Gradient method, we need to compute DP and (DP )∗ several times.
From Example 6.3.2,
(DF)(g) = û∞
where û is a radiating solution of

∆û + k 2 nû = −k 2 gu in R2 .
8.5. NUMERICAL SOLUTION TO THE INVERSE MEDIUM PROBLEM 113

Thus Z
2
û = k Φ(x, y)g(y)u(y)dy in R2 .
B

From the far field pattern of Φ(x, y),

k 2 −iπ/4
Z
û∞ = √ e e−ikx̂·y g(y)u(y)dy.
8kπ B

Let Q := (DP )∗ be an opertor from L2 [0, 2π] to L2 (B), then

⟨Qh, g⟩L2 (B) = ⟨h, (DP )g⟩L2 [0,2π]

Z 2π
= h(x̂)û∞ (x̂)dθ
0
k 2 iπ/4 2π
Z Z
=√ e h(x̂) eikx̂·y g(y)u(y)dydθ
8kπ 0 B
2 Z Z 2π
k
=√ eiπ/4 eikx̂·y u(y)h(x̂)dθg(y)dy
8kπ B 0

thus
2π
k 2 iπ/4
Z
(Qh)(y) = √ e u(y) eikx̂·y h(x̂)dθ.
8kπ 0
114 CHAPTER 8. SCATTERING FROM INHOMOGENEOUS MEDIUM