ANISOTROPIC INHOMOGENEOUS BACKGROUND

PU-ZHAO KOW AND JENN-NAN WANG

Abstract. In this paper, we are interested in the problem of determining the shape of sound-soft or impedance obstacles in the acoustic wave scattering with an anisotropic inhomogeneous medium. The main theme is to extend the factorization method to this setting. Precisely, we can reconstruct the obstacle by the eigenvalues and eigenfunctions of the far-eld operator. We also provide some numerical simulations in the paper.

1. Introduction

Let an obstacle be embedded inside an anisotropic inhomogeneous background. This system causes the acoustic wave to scatter. In this work, we are interested in reconstructing the shape of the obstacle by the far-eld information assuming that the inhomogeneity is known. There are several existing methods for other types of inverse scattering problems. Here we will focus on the factorization method, which was rst introduced by Kirsch in [Ki98, Ki99] to treat the inverse scattering problems for the scalar Helmholtz equation. For the detailed development of the factorization method for the Helmholtz and time-harmonic Maxwell equations, we refer the reader to the monograph [KG08].

The factorization method shares the same spirit as the linear sampling method, where both methods try to reconstruct the shape of the scatterer (an obstacle or the support of the inhomogeneity) by determining whether a point is inside or outside the scatterer.

We now briey describe the problem considered in the paper. Let A(x) = (aij(x))be
a real-symmetric matrix with C^{∞} entries, which satises the following uniform elliptic
condition: there exists a constant 0 < c < 1 such that

c|ξ|^{2} ≤X

ij

a_{ij}(x)ξ_{i}ξ_{j} ≤ c^{−1}|ξ|^{2}

for all x ∈ R^{d}, ξ ∈ R^{d}, d = 2 or 3. Moreover, we assume that supp(A − I) is compact
in R^{d}. Suppose that the acoustic refraction index n ∈ L^{∞}(R^{d}) with n ≥ c, such
that supp(n − 1) is compact in R^{d}. Let BR be the ball centered at origin with radius
R > 0 which contains supp(A − I) and supp(n − 1). Let D be an open bounded
domain where ∂D is Lipschitz, D ⊂ BR, and R^{d}\ D is connected. Let u^{to}_{D}(x, ˆz) =
u^{inc}_{A,n}(x, ˆz) + u^{sc}_{D}(x, ˆz)with an appropriate incident eld u^{inc}_{A,n}(x, ˆz)and the corresponding
scattered eld u^{sc}D(x, ˆz), ˆz = z/|z|, satisfy the following acoustic equation

∇ · (A(x)∇u^{to}_{D}) + k^{2}n(x)u^{to}_{D} = 0 in R^{d}\ D,

Bu^{to}_{D} = 0 on ∂D,

u^{sc}_{D} satises the Sommerfeld radiation condition at |x| → ∞,

(1.1)

1

where Bu^{to}D is either Dirichlet or impedance boundary conditions. Let u^{∞}D(ˆx, ˆz) be
the far-eld pattern of u^{sc}D. The inverse obstacle problem is to determine D from the
knowledge of u^{∞}D(ˆx, ˆz)for all ˆx, ˆz ∈ S^{d−1}.

This type of inverse obstacle problem including theoretical and numerical investigations has been studied extensively. We will not try to exhaust all the related works here.

Instead, we refer to recent monographs [CCH16, KG08] and references therein for the detailed development of the problems and reconstruction methods. To put our problem in perspective, we mention several closely related results.

• Penetrable obstacle in an inhomogeneous background. In [KP98], the authors considered the acoustic scattering from a penetrable obstacle inside an inhomogeneous structure. In their work, they showed that, if the reference medium is known, then the penetrable obstacle can be uniquely determined by the far-eld data at a xed energy. The approach used in [KP98] was only for the uniqueness proof. Later, in [GMMR12], the authors designed a reconstruction algorithm in the spirit of the factorization method. For the case of anisotropic penetrable obstacle embedded inside of a homogeneous medium, a factorization method was developed in [KL13]. On the other hand, when there was an unknown cavity hidden inside of the penetrable obstacle, a reconstruction result based on the factorization method was investigated in [YZZ13].

• Impenetrable obstacle in a homogeneous background. A factorization method for reconstructing an impenetrable obstacle in a homogeneous medium (Helmholtz equation) using the spectral data of the far-eld operator was developed in [Ki98]

(see also [KG08, Chapter 1]).

• Impenetrable obstacle in an inhomogeneous background. When the inhomogeneous background was described the Schrödinger equation with an inhomogeneous potential function, the theoretical study of the factorization method was carried out in [NPT07]. Moreover, the obstacle can be reconstructed without knowing the boundary condition on the boundary of the obstacle (either Dirichlet or Robin boundary conditions). Our work is closely related to this paper. In our paper, we consider an impenetrable obstacle in an anisotropic inhomogeneous background.

We would like to remark that, in [KG08, KP98], the far-eld operator is dened in terms of the far-eld pattern induced by the incident plane waves. In our case, we take a dierent incident eld. Roughly speaking, the incident eld used in our approach is the far-eld pattern of the outgoing fundamental solution for the background equation.

This idea is inspired by the work [GMMR12]. The main dierence between our work and [GMMR12] is that they consider penetrable obstacles (involving volume potentials);

while in our case, we consider impenetrable obstacles (involving surface potentials).

Indeed, in the case where the reference medium is homogeneous, the far-eld pattern of the standard outgoing fundamental solution is simply the plane waves. Since we assume that the background medium is known, the far-eld pattern of the outgoing fundamental solution for the background equation can be determined. However, from the numerics perspective, the computation of the far-eld of the outgoing fundamental solution directly is not ecient. Luckily, thanks to the mixed reciprocity relation (see Lemma 2.2), we can compute the far-eld of the outgoing fundamental solution by solving the total eld of the background equation with the incident plane waves.

Another focal point of this work is the numerical demonstration of increasing stability phenomenon with respect to the wave number k in the reconstruction of the obstacle. It is known that most of inverse problems are ill-posed including the problem we discuss here. However, in the case of identifying the potential in the Schrödinger equation by the Dirichlet-to-Neumann map, the stability increases as we increase the wave number (the logarithmic part decreases as the wave number increases) [INUW14]. A similar property was also established for the determination of the refraction index in the acoustic equation [NUW13]. We need to point out a major dierence between [INUW14]

and [NUW13]. In [INUW14], the constant appears in the Hölder part of the stability estimate depends polynomially in k. Hence, the stability of determining the potential in the Schrödinger equation tends a Hölder type as k increases. While, in [NUW13], the constant in the Hölder part of the stability estimate depends exponentially in k.

Consequently, the stability estimate in the acoustic case remains a logarithmic type when k is too large.

Considering the discussion above, even though the inverse obstacle problems for the Schrödinger equation and for the acoustic equation at a xed wave number are mathematically equivalent, it is necessary in practice to study the acoustic equation separately at least from the viewpoint of stability. We believe that it is an important and challenging problem to solve the inverse obstacle problem in the acoustic equation stably at high wave numbers.

The paper is organized as follows. In Section 2, we rst review some properties of the outgoing fundamental solution for the background equation and its far-eld pattern.

In Section 3, we prove our main reconstruction theorem for the sound-soft (Dirichlet) obstacle. The same ideas can be generalized to the obstacle with Neumann or impedance boundary conditions. Thus, in Section4, we state the reconstruction theorems without proofs corresponding to the cases of Neumann and impedance boundary conditions.

In Section 5, some numerical simulations are presented to show the eciency of our method. Finally, in Section 6, we compare the reconstruction results for the acoustic and the Schrödinger equations at dierent wave numbers to demonstrate the increasing stability phenomena.

2. Fundamental solution and Herglotz wave functions

In this section, we discuss some properties of the outgoing fundamental solution to the background equation (that is, the reference medium).

Denition 2.1. We say that a function u ∈ Hloc^{1} (R^{d})satises the Sommerfeld radiation
condition, if

∂_{r}u(x) − iku(x) = o(|x|^{−}^{d−1}^{2} ) as |x| → ∞,
where ∂r= ˆx · ∇.

For each x ∈ R^{d}, let ΦA,n(z, x) (for x 6= z) be the outgoing fundamental solution of
the following anisotropic inhomogeneous acoustic equation

(∇_{z}· (A(z)∇_{z}Φ_{A,n}(z, x)) + k^{2}n(z)Φ_{A,n}(z, x) = δ(z − x) ∀ z ∈ R^{d},
Φ_{A,n}(z, x) satises the Sommerfeld radiation condition at |z| → ∞,

see also [DHM18] where the existence of the fundamental solution, even for elliptic
systems, was proved. Let u^{inc}ref(x, ˆz) = e^{ikx·ˆ}^{z} be an incident plane wave. The presence

of inhomogeneity (without the obstacle) in BR gives rise to a unique scattered eld
u^{sc}_{ref} ∈ H_{loc}^{1} (R^{d})which satises the Sommerfeld radiation condition. Then the total eld
u^{to}_{ref} = u^{sc}_{ref} + u^{inc}_{ref} satises

∇ · (A(x)∇u^{to}_{ref}) + k^{2}n(x)u^{to}_{ref} = 0, ∀ x ∈ R^{d}. (2.1)
We rst prove the following mixed reciprocity principle.

Lemma 2.2. Let Φ^{∞}A,n(ˆz, x) be the far-eld pattern of ΦA,n(z, x). Then we have
Φ^{∞}_{A,n}(ˆz, x) = u^{to}_{ref}(x, −ˆz) ∀ x ∈ R^{d}, ˆz ∈ S^{d−1},

where u^{to}ref(·, −ˆz) is given in (2.1).

Proof. Here we adopt the proof from [KG08]. Let Φ(z, x) = ΦI,1(z, x) be the outgoing fundamental solution of the standard Helmholtz equation. Dene

Ψ(z, x) := Φ_{A,n}(z, x) − Φ(z, x).

Then Ψ(z, x) is a smooth at all z ∈ R^{d} and

∆_{z}Ψ(z, x) + k^{2}Ψ(z, x)

= −(∇_{z}· ((A(z) − I)∇_{z}Φ_{A,n}(z, x)) + k^{2}(n(z) − 1)Φ_{A,n}(z, x)).

We can write Ψ(z, x)

= − Z

R^{d}

Φ(z, y)(∇_{y} · ((A(y) − I)∇_{y}Φ_{A,n}(y, x)) + k^{2}(n(y) − 1)Φ_{A,n}(y, x)) dy.

Since both A − I and n − 1 have compact supports, the integral equation above is well-
dened. Note that the far-eld pattern of Φ(z, y) is Φ^{∞}(ˆz, y) = e^{−ikˆ}^{z·y} = u^{inc}_{ref}(y, −ˆz).
We then see that the far-eld pattern of Ψ(z, x) is

Ψ^{∞}(ˆz, x)

= − Z

R^{d}

e^{−ikˆ}^{z·y}(∇_{y} · ((A(y) − I)∇_{y}Φ_{A,n}(y, x)) + k^{2}(n(y) − 1)Φ_{A,n}(y, x)) dy. (2.2)
Since u^{to}_{ref}(x, −ˆz) satises

∇_{x}· (A(x)∇_{x}u^{to}_{ref}(x, −ˆz)) + k^{2}n(x)u^{to}_{ref}(x, −ˆz) = 0 for all x ∈ R^{d},
we have

∇_{x}· (A(x)∇_{x}u^{sc}_{ref}(x, −ˆz)) + k^{2}n(x)u^{sc}_{ref}(x, −ˆz)

= − (∇_{x}· (A(x)∇_{x}u^{inc}_{ref}(x, −ˆz)) + k^{2}n(x)u^{inc}_{ref}(x, −ˆz))

= − (∇_{x}· ((A(x) − I)∇_{x}u^{inc}_{ref}(x, −ˆz)) + k^{2}(n(x) − 1)u^{inc}_{ref}(x, −ˆz)).

In view of the symmetry property ΦA,n(x, y) = Φ_{A,n}(y, x), using integration by parts,
we can derive

u^{sc}_{ref}(x, −ˆz)

= − Z

R^{d}

Φ_{A,n}(y, x)(∇_{y}· ((A(y) − I)∇_{y}u^{inc}_{ref}(y, −ˆz)) + k^{2}(n(y) − 1)u^{inc}_{ref}(y, −ˆz)) dy

= − Z

R^{d}

e^{−ikˆ}^{z·y}(∇_{y} · ((A(y) − I)∇_{y}Φ_{A,n}(y, x)) + k^{2}(n(y) − 1)Φ_{A,n}(y, x)) dy. (2.3)

Combining (2.2) and (2.3) implies that

Ψ^{∞}(ˆz, x) = u^{sc}_{ref}(x, −ˆz),
and hence

Φ^{∞}_{A,n}(ˆz, x) = Ψ^{∞}(ˆz, x) + Φ^{∞}(ˆz, x) = u^{sc}_{ref}(x, −ˆz) + u^{inc}_{ref}(x, −ˆz) = u^{to}_{ref}(x, −ˆz),

which is our desired lemma.

Denition 2.3. For each g ∈ L^{2}(S^{d−1}), we dene the modied Herglotz wave function
as

u_{A,n,g}(x) :=

Z

S^{d−1}

Φ^{∞}_{A,n}(ˆz, x)g(ˆz) ds(ˆz) for all x ∈ R^{d}.
It is well-known that

g 7→

Z

S^{d−1}

e^{ikx·ˆ}^{z}g(ˆz) ds(ˆz) =
Z

S^{d−1}

u^{inc}_{ref}(x, ˆz)g(ˆz) ds(ˆz)

is injective. Since each u^{inc}ref(x, ˆz) leads a unique scattered eld u^{sc}ref(x, ˆz), and thus
induces a unique total eld u^{to}ref(x, ˆz). Combining these discussions and by the superposition
property, we can prove the following lemma.

Lemma 2.4. For coecients A and refraction index n prescribed above, the mapping g 7→ uA,n,g is injective.

3. Reconstruction of sound-soft obstacle

By Lemma2.2, we can choose the incident eld u^{inc}A,n(x, ˆz) = Φ^{∞}_{A,n}(ˆz, x)instead of the
usual plane wave u^{inc}ref(x, ˆz) = e^{ikx·ˆ}^{z}, also see [GMMR12] in which the similar incident

eld was used. In practice, this can be achieved by using a two-stage detection as
in Algorithm 1. By choosing this incident eld, no scattering occurs if there is no
obstacle. Now let u^{sc}Dir(x, ˆz) be the corresponding scattered eld, then the total eld
u^{to}_{Dir}(x, ˆz) = u^{inc}_{A,n}(x, ˆz) + u^{sc}_{Dir}(x, ˆz)satises

∇ · (A(x)∇u^{to}_{Dir}) + k^{2}n(x)u^{to}_{Dir} = 0 in R^{d}\ D,

u^{to}_{Dir} = 0 on ∂D,

u^{sc}_{Dir} satises Sommerfeld radiation condition at |x| → ∞,

(3.1)

that is,

∇ · (A(x)∇u^{sc}_{Dir}) + k^{2}n(x)u^{sc}_{Dir} = 0 in R^{d}\ D,

u^{sc}_{Dir} = −u^{inc}_{A,n} on ∂D,

u^{sc}_{Dir} satises Sommerfeld radiation condition at |x| → ∞.

(3.2)

Denote u^{∞}Dir be the far-eld pattern of u^{sc}Dir, which satises
u^{sc}_{Dir}(x) = γ_{d} e^{ik|x|}

|x|^{d−1}^{2} u^{∞}_{Dir}(ˆx) + O(|x|^{−}^{d+1}^{2} ) as |x| → ∞
uniformly for all directions ˆx = x/|x|, where

γ_{d} :=

(e^{iπ/4}/√

8πk if d = 2,

1/4π if d = 3, (3.3)

see e.g. [GMMR12].

The well-posedness of u^{sc}D is guaranteed by the following well-known fact:

Lemma 3.1. Given any f ∈ H^{1/2}(∂D), there exists a unique v ∈ H_{loc}^{1} (R^{d}\ D) such

that

∇ · (A(x)∇v) + k^{2}n(x)v = 0 in R^{d}\ D,

v = f on ∂D,

v satises Sommerfeld radiation condition at |x| → ∞.

(3.4)

3.1. The factorization of the far-eld operator. For each g ∈ L^{2}(S^{d−1}), recall the
modied Herglotz wave function is dened by

u_{A,n,g}(x) :=

Z

S^{d−1}

Φ^{∞}_{A,n}(ˆz, x)g(ˆz) ds(ˆz)

= Z

S^{d−1}

u^{inc}_{A,n}(x, ˆz)g(ˆz) ds(ˆz), (3.5)
which is simply the superposition of the incident elds. So, the far-eld of the scattered

eld is the superposition of u^{∞}Dir. Thus, we can dene the far-eld operator as follows:

Denition 3.2. The far-eld operator FDir: L^{2}(S^{d−1}) → L^{2}(S^{d−1}) is given by
F_{Dir}(g)(ˆx) =

Z

S^{d−1}

u^{∞}_{Dir}(ˆx, ˆz)g(ˆz) ds(ˆz) for all ˆx ∈ S^{d−1}.

With Lemma 3.1, we also dene an operator closely related to the far-eld operator:

Denition 3.3. The data-to-pattern operator GDir : H^{1/2}(∂D) → L^{2}(S^{d−1}) is dened
by GDirf = v^{∞}, where v^{∞} is the far-eld of v ∈ Hloc^{1} (R^{d}\ D) which satises (3.4).

To discuss the relation between the far-eld operator and the data-to-pattern operator, we introduce the single-layer potential operator S dened by

(Sφ)(x) :=

Z

∂D

Φ_{A,n}(x, z)φ(z) ds(z) for all x ∈ ∂D.

We rst establish the following mapping property of S.

Lemma 3.4. The potential S : H^{−1/2}(∂D) → H^{1/2}(∂D) is continuous.

Proof. Let ΦA,1(x, z) be the outgoing fundamental solution of

(∇x· (A(x)∇xΦA,1(x, z)) + k^{2}ΦA,1(x, z) = δ(x − z) ∀ x ∈ R^{d},
Φ_{A,1}(x, z) satises the Sommerfeld radiation condition at |x| → ∞.

Since the coecient A is smooth, ΦA,1(x, z) is C^{∞} for x 6= z. By the techniques of
pseudodierential operators (see [McL00, Theorem 6.11]), let

( ˜Sφ)(x) :=

Z

∂D

Φ_{A,1}(x, z)φ(z) ds(z) for all x ∈ ∂D,
then ˜S : H^{−1/2}(∂D) → H^{1/2}(∂D) is continuous. Note that

∇_{x}· (A(x)∇_{x}(Φ_{A,n}− Φ_{A,1})) + k^{2}n(x)(Φ_{A,n}− Φ_{A,1}) = k^{2}(1 − n)Φ_{A,1}.

Observe that k^{2}(1−n)Φ_{A,1} ∈ L^{2}(R^{d})(uniformly in z). By the elliptic estimate theorem,
we have that

(Φ_{A,n}− Φ_{A,1})(·, z) ∈ H^{2,2}(D).

From the trace theorem and the symmetry of the fundamental solutions, it follows that

|∇_{x}(Φ_{A,n}− Φ_{A,1})(x, z)| = |∇_{z}(Φ_{A,n}− Φ_{A,1})(x, z)| ∈ H^{1/2}(∂D)
uniformly for x ∈ ∂D. Therefore, we can show that the operator

S^{0} : φ 7→

Z

∂D

(Φ_{A,n}− Φ_{A,1})(x, z)φ(z)ds(z) (3.6)
maps H^{−1/2}(∂D)to H^{1}(∂D), in particular, S^{0} : H^{−1/2}(∂D) → H^{1/2}(∂D)is continuous.

Since S = ˜S + S^{0}, Lemma 3.4 follows.

Following the ideas in [KG08, Theorem 1.15], we can obtain the following proposition, which contains the main idea of this work.

Proposition 3.5. Let G^{∗}Dir : L^{2}(S^{d−1}) → H^{−1/2}(∂D) be the adjoint of GDir and S^{∗} :
H^{−1/2}(∂D) → H^{1/2}(∂D) be the adjoint of S. Then

F_{Dir} = −G_{Dir}S^{∗}G^{∗}_{Dir}. (3.7)
Proof. We consider an auxiliary operator HDir : L^{2}(S^{d−1}) → H^{1/2}(∂D) dened as the
restriction of the modied Herglotz wave function on ∂D, that is,

(H_{Dir}g)(x) :=

Z

S^{d−1}

Φ^{∞}_{A,n}(ˆz, x)g(ˆz) ds(ˆz) for all x ∈ ∂D.

From (3.2), we know that

FDir = −GDirHDir. (3.8)

Also, it is not dicult to check that the adjoint HDir^{∗} : H^{−1/2}(∂D) → L^{2}(S^{d−1}) of HDir

is given by

(H_{Dir}^{∗} φ)(ˆz) =
Z

∂D

Φ^{∞}_{A,n}(ˆz, x)φ(x) dx

for all φ ∈ H^{−1/2}(∂D). Observe that (HDir^{∗} φ)(ˆz) is nothing but the far-eld pattern of
v(z) =

Z

∂D

ΦA,n(z, x)φ(x) dx,
which shows that HDir^{∗} = G_{Dir}S, that is,

H_{Dir} = S^{∗}G^{∗}_{Dir}. (3.9)

Putting together (3.8) and (3.9) implies (3.7).

Similar to [KG08, Theorem 1.12], we can show that

Proposition 3.6. Φ^{∞}_{A,n}(·, z) ∈ range(G_{Dir}) if and only if z ∈ D.

Proof. First of all, we assume that Φ^{∞}_{A,n}(·, z) ∈ range(G_{Dir}), i.e., there exists f ∈
H^{1/2}(∂D) such that GDirf = Φ^{∞}_{A,n}(·, z). Let w ∈ Hloc^{1} (R^{d}\ D)be the unique solution to
(3.4). Since both

wand ΦA,n(·, z) have the same far-eld pattern Φ^{∞}A,n(·, z),
by Rellich's lemma, there exists r0 > 0 such that

w(x) = Φ_{A,n}(x, z) for |x| > r0.

Since R^{d}\ D is connected, and both w and ΦA,n satisfy the anisotropic inhomogeneous
acoustic equation (2.1) in R^{d}\ D, by the unique continuation property of this equation,
we have that

w(x) = Φ_{A,n}(x, z) for all x ∈ R^{d}\ D.

Note that w ∈ H_{loc}^{1} (R^{d}\ D). In other words, w has no singularity in R^{d}\ D and thus
z ∈ D.

Conversely, let z ∈ D. Dene w = ΦA,n(·, z) in R^{d}\ D and f = w|∂D, we can easily
see that GDirf = Φ^{∞}_{A,n}(·, z), that is, Φ^{∞}A,n(·, z) ∈ range(GDir).
With Proposition 3.5 and 3.6 at hand, it is helpful to study the operators S, GDir,
and FDir.

3.2. The properties of the single-layer potential S. In this section, we shall study the properties of the far-eld operator. With the help of [McL00, Theorem 7.5], we can further improve Lemma3.4.

Lemma 3.7. If k^{2} is not an eigenvalue of

(∇ · (A(x)∇u) + k^{2}n(x)u = 0 in D,

u = 0 on ∂D, (3.10)

then S : H^{−1/2}(∂D) → H^{1/2}(∂D) is an isomorphism.

Next, by mimicking the proofs of [KG08, Lemma 1.14(b),(c),(d)], we can obtain the following lemmas. Since we are considering a dierent equation, we include the proofs here for the sake of completeness.

Lemma 3.8. If k^{2} is not an eigenvalue of (3.10), then =hφ, Sφi ≤ 0 for all φ ∈
H^{1/2}(∂D), where h·, ·i is the H^{−1/2}(∂D) × H^{1/2}(∂D) duality pair and = denotes the
imaginary part of a complex number. Moreover,

=hφ, Sφi = 0 if and only if φ ≡ 0.

Proof. For each φ ∈ H^{−1/2}(∂D), dene
v(x) = (SL φ)(x) =

Z

∂D

Φ_{A,n}(x, y)φ(y) ds(y) for x ∈ R^{d}\ ∂D.

Then, by [McL00, Theorem 6.11], we have v ∈ H^{1}(D) ∩ H_{loc}^{1} (R^{d}\ D) and the traces on

∂D with the outer normal normal ν(x),
v_{±}(x) = lim

h→0+

v(x ± hν(x)),
ν(x) · A(x)∇v_{±}(x) = lim

h→0+

ν(x) · A(x ± hν(x))∇v(x ± hν(x)), exist and satisfy the jump condition

v = v± = Sφ for all φ ∈ H^{−1/2}(∂D).

Choose φ = ν · A∇v−− ν · A∇v_{+}, then
hφ, Sφi = hν · A∇v−− ν · A∇v_{+}, vi

= Z

∂D

(ν · A∇v−)v ds − Z

∂D

(ν · A∇v_{+})v ds

= Z

∂D

(ν · A∇v−)v ds + Z

∂(BR\D)

(ν · A∇v_{+})v ds −
Z

|x|=R

(ν · ∇v)v ds

= Z

D∪(BR\D)

(∇ · (A∇v))v dx + Z

D∪(BR\D)

∇¯v · A∇v dx − Z

|x|=R

(∂rv)v ds.

Since ∇ · (A∇v) + k^{2}n(x)v = 0 in R^{d}\ ∂D, together with the Sommerfeld radiation
condition, then

hφ, Sφi = Z

D∪(BR\D)

(∇¯v · A∇v − k^{2}n(x)|v|^{2}) dx −
Z

|x|=R

(∂_{r}v)v ds

= Z

D∪(BR\D)

(∇¯v · A∇v − k^{2}n(x)|v|^{2}) dx − ik
Z

|x|=R

|v|^{2}ds + o(R^{−}^{d−1}^{2} )

(3.11)

as R → ∞. Then

=hφ, Sφi = −k Z

|x|=R

|v|^{2}ds + o(R^{−}^{d−1}^{2} ),
and thus

=hφ, Sφi = −k lim

R→∞

Z

|x|=R

|v|^{2}ds = −k lim

R→∞

Z

S^{d−1}

|R^{(d−1)/2}v|^{2}ds(ˆx)

= −|γ_{d}|^{2}k
Z

S^{d−1}

|v^{∞}|^{2}ds(ˆx) ≤ 0,

(3.12)

where γd is given in (3.3).

Finally, we want to show that =hφ, Sφi = 0 leads to φ ≡ 0. If =hφ, Sφi = 0, then
v^{∞} ≡ 0 from (3.12). Since R^{d} \ D is connected, from Rellich's lemma and unique
continuation property, we know that v = 0 in R^{d}\ D. Therefore, by the trace theorem,
we have Sφ = 0. By Lemma3.7, we know that S is an isomorphism and thus φ ≡ 0.
Lemma 3.9. Let Si be the single-layer operator corresponding to the wave number k =
i = √

−1. Then Si : H^{−1/2}(∂D) → H^{1/2}(∂D) is self-adjoint and coercive. Moreover,
S − S_{i} : H^{−1/2}(∂D) → H^{1/2}(∂D) is compact.

Proof. Substituting k = i into (3.11) gives
hφ, S_{i}φi =

Z

D∪(BR\D)

(∇¯v · A∇v + n(x)|v|^{2}) dx +
Z

|x|=R

|v|^{2}ds + o(R^{−}^{d−1}^{2} )
as R → ∞. By taking R → ∞, we reach

hφ, S_{i}φi =
Z

R^{d}\∂D

(∇¯v · A∇v + n(x)|v|^{2}) dx,
which shows that Si is self-adjoint.

Since A is uniform elliptic and n(x) ≥ c, then

hφ, S_{i}φi ≥ ckvk^{2}_{H}1(R^{d}\∂D).

By trace theorem, together with Lemma3.7, there exist constants c1, c_{2} with 0 < c2 <

c_{1} < c such that

hφ, S_{i}φi ≥ c_{1}kvk^{2}_{H}1/2(∂D) = c_{1}kS_{i}φk^{2}_{H}1/2(∂D) ≥ c_{2}kφk^{2}_{H}−1/2(∂D),

that is, Si is coercive. Note that the operator S − Si has the same mapping property
as S^{0} in (3.6). We thus immediately conclude that

S − S_{i} : H^{−1/2}(∂D) → H^{1/2}(∂D)

is compact since the embedding H^{1}(∂D) into H^{1/2}(∂D)is compact.
3.3. The properties of the data-to-pattern operator GDir. We can prove following
two lemmas about GDir.

Lemma 3.10. GDir : H^{1/2}(∂D) → L^{2}(S^{d−1}) is injective.

Proof. This is a direct consequence of Rellich's lemma and unique continuation property.

Lemma 3.11. GDir : H^{1/2}(∂D) → L^{2}(S^{d−1}) is compact.

Proof. This lemma can be proved following the line of [KG08, Lemma 1.13]. Let R > 1
be a large number. For each f ∈ H^{1/2}(∂D), let v ∈ H_{loc}^{1} (R^{d}\ D) satisfy (3.4). By the
representation formula

v^{∞}(ˆx) =
Z

∂D

(ν · A(z)∇Φ^{∞}_{A,n}(ˆx, z))v(z) − Φ^{∞}_{A,n}(ˆx, z)(ν · A(z)∇v(z))

ds(z) (3.13)
for ˆx ∈ S^{d−1}, we can decompose GDir into GDir = G2◦ G1, where

G_{1} : H^{1/2}(∂D) → C(∂B_{R}) × C(∂B_{R})
G_{2} : C(∂B_{R}) × C(∂B_{R}) → L^{2}(S^{d−1})
are dened by

G_{1}f =

v|_{∂B}_{R}, ν · A∇v|_{∂B}_{R}

,
G_{2}(g, h) =

Z

∂BR

(ν · A∇Φ^{∞}_{A,n}(ˆx, z))g(z) − Φ^{∞}_{A,n}(ˆx, z)h(z)

ds(z).

By interior regularity results for elliptic equation, we know that G1 : H^{1/2}(∂D) →
C(∂B_{R}) × C(∂B_{R}) is bounded. Moreover, from the analyticity of Φ^{∞}A,n(ˆx, z) in ˆx, it
immediately follows that G2 : C(∂BR) × C(∂BR) → L^{2}(S^{d−1}) is compact and the proof

is completed.

Modify the ideas in [AC06, Theorem 3.1], we can establish the denseness of range(GDir).
Lemma 3.12. ran(GDir) is dense in L^{2}(S^{d−1}).

Proof. Let the orthogonal complement of range(GDir)

ran(G_{Dir})^{⊥} := g ∈ L^{2}(S^{d−1}) (g, ψ)_{L}^{2}_{(S}^{d−1}_{)} = 0 for all ψ ∈ ran(GDir) .
Our aim is to show that ran(GDir)^{⊥} = 0. Recall that

ran(G_{Dir})^{⊥} = ker(G^{∗}_{Dir}).

Given any f ∈ H^{1/2}(∂D), let v ∈ H_{loc}^{1} (R^{d}\ D) be the unique solution to (3.4) and
GDirf = v^{∞} be its far-eld pattern. Again, using the representation formula (3.13), we
have

(g, G_{Dir}f )_{L}2(S^{d−1})= (g, v^{∞})_{L}2(S^{d−1})

= Z

S^{d−1}

g(ˆx)

Z

∂D

(ν · A(z)∇Φ^{∞}_{A,n}(ˆx, z))v(z) ds(z)

ds(ˆx)

− Z

S^{d−1}

g(ˆx)

Z

∂D

Φ^{∞}_{A,n}(ˆx, z)(ν · A(z)∇v(z)) ds(z)

ds(ˆx)

= Z

∂D

ν · A(z)∇

Z

S^{d−1}

Φ^{∞}_{A,n}(ˆx, z)g(ˆx) ds(ˆx)

v(z) ds(z)

− Z

∂D

Z

S^{d−1}

Φ^{∞}_{A,n}(ˆx, z)g(ˆx) ds(ˆx)

ν · A∇v(z) ds(z)

= Z

∂D

(ν · A(z)∇u_{A,n,g})f − u_{A,n,g}(ν · A(z)∇v)

ds. (3.14) Let u be the unique solution to (3.4) with boundary data u = uA,n,g on ∂D, that is,

∇ · (A(x)∇u) + k^{2}n(x)u = 0 in R^{d}\ D,

u = u_{A,n,g} on ∂D,

u satises Sommerfeld radiation condition at |x| → ∞.

(3.15)

Since both v and u satises ∇·(A(x)∇u)+k^{2}n(x)u = 0in R^{d}\Dand satisfy Sommerfeld
radiation condition, then

0 = Z

∂D

(ν · (A∇u))f − u_{A,n,g}(ν · (A∇v))

ds (3.16)

Subtracting (3.14) by (3.16) yields
(g, GDirf )_{L}^{2}_{(S}^{d−1}_{)}=

Z

∂D

(ν · (A∇(uA,n,g − u)))f ds, that is,

G^{∗}_{Dir}g = ν · A∇(u_{A,n,g}− u).

If g ∈ ker(G^{∗}Dir), then ν · A∇uA,n,g = ν · A∇u. Combining this with (3.15), we can
extend u in D by dening u = uA,n,g, and then u satises

∇ · (A∇u) + k^{2}n(x)u = 0 in R^{d}.

Since u satises Sommerfeld radiation condition at |x| → ∞, by uniqueness result, we
see that u ≡ 0, which gives uA,n,g = 0 in D. By the unique continuation property, we
hence have uA,n,g ≡ 0. Finally, Lemma2.4 implies g ≡ 0, that is, ran(GDir)^{⊥} = 0.
3.4. The properties of the far-eld operator FDir. In view of the argument in
[KG08, Theorem 1.8], we can prove an important property of FDir.

Lemma 3.13. The far-eld operator FDir : L^{2}(S^{d−1}) → L^{2}(S^{d−1}) satises

F_{Dir}− F_{Dir}^{∗} = i2kγ^{2}F_{Dir}^{∗} F_{Dir}, (3.17)

where FDir^{∗} : L^{2}(S^{d−1}) → L^{2}(S^{d−1}) is the adjoint operator of FDir, and the constant γ is
given in (3.3).

Proof. Given any g, h ∈ L^{2}(S^{d−1}), choose v^{inc} = u_{A,n,g} and w^{inc} = u_{A,n,h}, which are
modied Herglotz wave functions given in (3.5). Let v and w be the solutions of
the scattering problem (3.1), with corresponding scattered elds v^{sc} = v − v^{in} and
w^{sc}= w − w^{in}, which are solution to (3.2), and the far-eld patterns of v^{sc} and w^{sc} are
v^{∞} and w^{∞}, respectively. By denition of FDir, we have

v^{∞}= FDirg and w^{∞}= FDirh.

Note that 0 =

Z

BR\D

(v∇ · (A∇w) − w∇ · (A∇v)) dx = Z

|x|=R

v(∂_{r}w) − w(∂_{r}v)

ds. (3.18) Clearly, for each R > 0,

Z

|x|=R

v^{inc}(∂rw^{inc}) − w^{inc}(∂rv^{inc})

ds = 0. (3.19)

By the Sommerfeld radiation condition and the asymptotic expansion of the scattered

eld, we can compute

v^{sc}(x)(∂_{r}w^{sc}(x)) − w^{sc}(x)(∂_{r}v^{sc}(x))

= −ikv^{sc}(x)w^{sc}(x) − ikw^{sc}(x)v^{sc}(x) + o(|x|^{−}^{d−1}^{2} )

= −ik

γ e^{ik|x|}

|x|^{d−1}^{2} v^{∞}(ˆx) + O(|x|^{−}^{d+1}^{2} )

γe^{−ik|x|}

|x|^{d−1}^{2} w^{∞}(ˆx) + O(|x|^{−}^{d+1}^{2} )

− ik

γe^{−ik|x|}

|x|^{d−1}^{2} w^{∞}(ˆx) + O(|x|^{−}^{d+1}^{2} )

γ e^{ik|x|}

|x|^{d−1}^{2} v^{∞}(ˆx) + O(|x|^{−}^{d+1}^{2} )

+ o(|x|^{−}^{d−1}^{2} )

= −i2kγ^{2}

|x|^{d−1}v^{∞}(ˆx)w^{∞}(ˆx) + O(|x|^{−d})
uniformly for all ˆx = x/|x|. Hence we have

R→∞lim Z

|x|=R

v^{sc}(∂_{r}w^{sc}) − w^{sc}(∂_{r}v^{sc})

ds

= −i lim

R→∞

Z

|x|=R

2kγ^{2}

|x|^{d−1}v^{∞}(ˆx)w^{∞}(ˆx) ds

= −i lim

R→∞

2kγ^{2}
R^{d−1}

Z

|x|=R

v^{∞}(ˆx)w^{∞}(ˆx) ds

= −i lim

R→∞

2kγ^{2}
R^{d−1}

Z

S^{d−1}

v^{∞}(ˆx)w^{∞}(ˆx)R^{d−1}ds

= −i2kγ^{2}
Z

S^{d−1}

v^{∞}(ˆx)w^{∞}(ˆx) ds

= −i2kγ^{2}(F_{Dir}g, F_{Dir}h)_{L}2(S^{d−1}). (3.20)

Next, for each R > 1, by Fubini's theorem and the representation formula of the far-eld pattern, we can obtain that

Z

|x|=R

v^{inc}(∂rw^{sc}) − w^{sc}(∂rv^{inc})

ds(x)

= Z

S^{d−1}

g(ˆz)

Z

|x|=R

Φ^{∞}_{A,n}(ˆz, x)(∂rw^{sc}) − w^{sc}(∂rΦ^{∞}_{A,n}(ˆz, x)) ds(x)

ds(ˆz)

= − Z

S^{d−1}

g(ˆz)w^{∞}(ˆz) ds(ˆz)

= −(g, F_{Dir}h)_{L}^{2}_{(S}^{d−1}_{)}, (3.21)

and similarly,

Z

|x|=R

v^{sc}(∂_{r}w^{inc}) − w^{inc}(∂_{r}v^{sc})

ds(x)

= (F_{Dir}g, h)_{L}2(S^{d−1}). (3.22)

Combining (3.18), (3.19), (3.20), (3.21), and (3.22) gives

−i2kγ^{2}(F_{Dir}g, F_{Dir}h)_{L}2(S^{d−1})− (g, F_{Dir}h)_{L}2(S^{d−1})+ (F_{Dir}g, h)_{L}2(S^{d−1}) = 0,

which is our desired result.

With the help of (3.17), we can investigate the range property of the far-eld operator
F_{Dir}.

Lemma 3.14. Let SDir : L^{2}(S^{d−1}) → L^{2}(S^{d−1}) be the scattering operator dened by
S_{Dir} = I + i2kγ^{2}F_{Dir},

then SDir is unitary, i.e., SDir^{∗} S_{Dir} = S_{Dir}S_{Dir}^{∗} = I. Moreover, FDir is normal.

Proof. It is straightforward to see that
S_{Dir}^{∗} S_{Dir} =

I − i2kγ^{2}F_{Dir}^{∗}

I + i2kγ^{2}F_{Dir}

= I + i2kγ^{2}(F_{Dir}− F_{Dir}^{∗} ) + 4k^{2}γ^{4}F_{Dir}^{∗} F_{Dir}.

(3.23)

Combining this and (3.17), we obtain SDir^{∗} S_{Dir} = I. Since FDir : L^{2}(S^{d−1}) → L^{2}(S^{d−1})
is a compact operator, then SDir is a compact perturbation of the identity, and so SDir

is bijective. Hence, S_{Dir}^{−1} = S_{Dir}^{∗} , which shows that SDir is unitary. Note that
S_{Dir}S_{Dir}^{∗} =

I + i2kγ^{2}F_{Dir}

I − i2kγ^{2}F_{Dir}^{∗}

= I + i2kγ^{2}(F_{Dir}− F_{Dir}^{∗} ) + 4k^{2}γ^{4}F_{Dir}F_{Dir}^{∗} .

(3.24)

By comparing (3.23) and (3.24), we conclude that FDir is normal.
Lemma 3.15. If k^{2} is not an eigenvalue of (3.10), then FDir : L^{2}(S^{d−1}) → L^{2}(S^{d−1})
is injective and ran(FDir) is dense in L^{2}(S^{d−1}).

Proof. We rst establish the injectivity. Given g ∈ ker(FDir). Recall that FDirg is the far-eld pattern corresponding to the incident eld

v^{inc}(x) = u_{A,n,g}(x).

Let v^{sc}be the associated scattered eld. As above, applying Rellich's lemma and unique
continuation property, we obtain v^{sc} = 0 in R^{d}. This shows that v^{inc} satises (3.10).

Since k^{2} is not an eigenvalue for (3.10), then v^{inc} = 0in D. By the unique continuation
property for the acoustic equation, we then have v^{inc} = uA,n,g ≡ 0in R^{d}. The injectivity
of g → uA,n,g in Lemma 2.4 implies g ≡ 0.

Next, to prove that ran(FDir) is dense in L^{2}(S^{d−1}), it suces to show that FDir^{∗} :
L^{2}(S^{d−1}) → L^{2}(S^{d−1}) is injective. For each g, h ∈ L^{2}(S^{d−1}), the reciprocity relation for
the far-eld pattern gives

(g, F_{Dir}h)_{L}^{2}_{(S}^{d−1}_{)}=
Z

S^{d−1}

g(ˆx)

Z

S^{d−1}

v_{Dir}^{∞}(ˆx, ˆz)h(ˆz) ds(ˆz)

ds(ˆx)

= Z

S^{d−1}

g(ˆx)

Z

S^{d−1}

v_{Dir}^{∞}(−ˆz, −ˆx)h(ˆz) ds(ˆz)

ds(ˆx)

= Z

S^{d−1}

Z

S^{d−1}

v_{Dir}^{∞}(−ˆz, −ˆx)g(ˆx) ds(ˆx)

h(ˆz) ds(ˆz)

=

Z

S^{d−1}

v^{∞}_{Dir}(−ˆz, −ˆx)g(ˆx) ds(ˆx), h

L^{2}(S^{d−1})

, that is,

(F_{Dir}^{∗} g)(ˆz) =
Z

S^{d−1}

v_{Dir}^{∞}(−ˆz, −ˆx)g(ˆx) ds(ˆx).

Therefore, if g ∈ ker(FDir^{∗} ), then by the injectivity of F , we conclude that g ≡ 0.
3.5. The (F^{∗}F )^{1/4}-method. Now we are going to derive our main reconstruction
method. The result is based on the following theorem in [KG08].

Theorem 3.16. [KG08, Theorem 1.23] Let H be a Hilbert space and X be a reexive Banach space. Assume that the compact operator F : H → H have a factorization of the form

F = B ˜AB^{∗}
with operators B : X → H and ˜A : X^{∗} → X such that

(1) =hϕ, ˜Aϕi 6= 0 for all 0 6= ϕ ∈ ran(B^{∗}).

(2) ˜A = ˜A_{0} + C for some compact operator C and some self-adjoint operator ˜A_{0}
which is coercive on ran(B^{∗}).

If F is injective and I + irF is unitary for some r > 0, then ran(B) = ran((F^{∗}F )^{1/4}).
Putting together Proposition3.5, Lemma3.83.11,3.14,3.15, and applying the above
theorem, we can immediately obtain the following lemma.

Lemma 3.17. If k^{2} is not an eigenvalue of (3.10), then ran(GDir) = ran((F_{Dir}^{∗} FDir)^{1/4}).
Finally, we want to characterize ran((F_{Dir}^{∗} F_{Dir})^{1/4}) as in [KG08] for the purpose of
numerical simulations. Since FDir : L^{2}(S^{d−1}) → L^{2}(S^{d−1}) is compact and normal,
there exists a set of (complex) eigenvalues {λ^{j}Dir}_{j∈N} with corresponding eigenfunctions
{φ^{j}_{Dir}}_{j∈N} in L^{2}(S^{d−1}). Furthermore, the set of eigenfunctions forms an orthonormal

basis of L^{2}(S^{d−1}), see e.g. [Zim90, Corollary 3.2.9]. For g ∈ L^{2}(S^{d−1}), we can write
(F_{Dir}^{∗} F_{Dir})^{1/4}g =X

j∈N

q

|λ^{j}_{Dir}|(g, φ^{j}_{Dir})_{L}2(S^{d−1})φ^{j}_{Dir},
Φ^{∞}_{A,n}(·, z) =X

j∈N

(Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)}φ^{j}_{Dir}.

Therefore, if Φ^{∞}A,n(·, z) ∈ ran((F_{Dir}^{∗} F_{Dir})^{1/4}), then there exists a g ∈ L^{2}(S^{d−1}) such that
X

j∈N

q

|λ^{j}_{Dir}|(g, φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)}φ^{j}_{Dir}=X

j∈N

(Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)}φ^{j}_{Dir}.
Equivalently,

q

|λ^{j}_{Dir}|(g, φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)} = (Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)} for all j ∈ N,
or

(g, φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)} = (Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)}
q

|λ^{j}_{Dir}|

for all j ∈ N.

Hence, Φ^{∞}A,n(·, z) ∈ ran((F_{Dir}^{∗} F_{Dir})^{1/4}) if and only if
X

j∈N

|(Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}2(S^{d−1})|^{2}

|λ^{j}_{Dir}| < ∞.

Combining this with Lemma3.17and Proposition3.6, we get the following key theorem.

Theorem 3.18. If k^{2} is not an eigenvalue of (3.10), then the following are equivalent:

(1) z ∈ D;

(2) Φ^{∞}_{A,n}(·, z) ∈ ran((F_{Dir}^{∗} F_{Dir})^{1/4});
(3) WDir(z) :=

X

j∈N

|(Φ^{∞}_{A,n}(·, z), φ^{j}_{Dir})_{L}^{2}_{(S}^{d−1}_{)}|^{2}

|λ^{j}_{Dir}|

−1

> 0.

In other words, the characteristic function of the sound-soft obstacle D is given by
χ_{D}(z) = sign(W_{Dir}(z)).

4. The determination of the obstacle with other boundary conditions
We can extend the factorization developed above to the obstacle with other types of
boundary conditions on ∂D, i.e., Neumann or impedance boundary conditions. In the
case of Helmholtz equation (homogeneous medium) with the obstacle having these two
types of boundary conditions, the factorization method has been discussed in detail in
[KG08, Chapter 1,2]. For the acoustic equation considered here, once we use Φ^{∞}_{A,n}(ˆz, x)
as the incident eld, the factorization method for obstacles with Neumann or impedance
boundary conditions can be established following exactly the arguments in [KG08,
Chapter 1,2]. Since we have presented the detailed proof for the Dirichlet case in
Section3, we will not repeat the arguments here. Instead, we state the main theorems
of the reconstruction method without proofs.

We rst discuss the Neumann case (sound-hard). The (F^{∗}F )^{1/4}-method can be easily
modied to treat this case as in [KG08, Chapter 1]. Let FNeu be the far-eld operator
corresponding to the sound-hard obstacles, which is dened similarly as FDir.

Theorem 4.1. If k^{2} is not an eigenvalue of

(∇ · (A(x)∇u) + k^{2}n(x)u = 0 in D,

ν · (A(x)∇u) = 0 on ∂D, (4.1)

then the following are equivalent:

(1) z ∈ D;

(2) Φ^{∞}A,n(·, z) ∈ ran((F_{Neu}^{∗} F_{Neu})^{1/4});
(3) WNeu(z) :=

X

j∈N

|(Φ^{∞}_{A,n}(·, z), φ^{j}_{Neu})_{L}^{2}_{(S}^{d−1}_{)}|^{2}

|λ^{j}_{Neu}|

−1

> 0, where {λ^{j}_{Neu}, φ^{j}_{Neu}} is an
eigen-system of the operator FNeu.

In other words, the characteristic function of the sound-hard obstacle D is given by
χ_{D}(z) = sign(W_{Neu}(z)).

Now we consider the obstacle with impedance condition. Let FImpbe the corresponding far-eld operator. In this case, FImp fails to be normal in general, unless λ is real-valued (i.e. Robin boundary condition). However, the problem can be overcome by the F]- method, which can be found in [KG08, Chapter 2]. The self-adjoint operator F] is dened as follows:

F_{]}= |<F_{Imp}| + |=F_{Imp}|,
where

<FImp := 1

2(FImp+ F_{Imp}^{∗} ) and =FImp = 1

2i(FImp− F_{Imp}^{∗} ).

In this case, the result reads:

Theorem 4.2. Let λ ∈ L^{∞}(∂D) be a complex-valued function with non-negative the
imaginary part on ∂D. If k^{2} is not an eigenvalue of

(∇ · (A(x)∇u) + k^{2}n(x)u = 0 in D,

ν · (A(x)∇u) + λ(x)u = 0 on ∂D, (4.2)

then the following are equivalent:

(1) z ∈ D;

(2) Φ^{∞}A,n(·, z) ∈ ran(F_{]}^{1/2});
(3) WImp(z) :=

X

j∈N

|(Φ^{∞}_{A,n}(·, z), φ^{j}_{Imp})_{L}2(S^{d−1})|^{2}

|λ^{j}_{Imp}|

−1

> 0, where {λ^{j}_{Imp}, φ^{j}_{Imp}} is an
eigen-system of the operator F].

5. Numerical results

In this section, we present some numerical simulation results to show the eciency of the method developed in Section 3. For simplicity, we treat the acoustic equation in two dimensions. The simulations are obtained using MATLAB 2020a (with PDE Toolbox). We consider four shapes of sound-soft obstacles whose parametric equations are listed in Table1.

Obstacle shape Parametrization (anti-clockwise oriented) circle x(t) = 0.5 cos t

0 ≤ t ≤ 2π y(t) = 0.5 sin t

peanut [YZZ13] x(t) = cos t ·√

cos^{2}t + 0.25 sin^{2}t

0 ≤ t ≤ 2π y(t) = sin t ·√

cos^{2}t + 0.25 sin^{2}t
kite [YZZ13] x(t) = 0.5 cos t + 0.325 cos 2t − 0.325

0 ≤ t ≤ 2π y(t) = 0.75 sin t

heart x(t) = −0.5 sin^{3}t

0 ≤ t ≤ 2π
(Wolfram Mathworld) y(t) = ^{13}_{32}cos t −_{32}^{5} cos 2t −_{16}^{1} cos 3t − _{32}^{1} cos 4t

Table 1. Parametrization of the obstacles

To carry out the simulations, we rst need to compute the incident eld
u^{inc}_{A,n}(x, ˆz) = Φ^{∞}_{A,n}(ˆz, x) = u^{to}_{ref}(x, −ˆz).

By the mixed reciprocity relation described above, it suces to compute the total eld for the acoustic equation (without obstacle) with plane incident elds. After generating the required incident eld, we then simulate the scattered eld and compute the far-

eld pattern for the acoustic equation with a sound-soft obstacle. The simulation is explained in the following algorithm.

Algorithm 1 Simulation of the scattered eld u^{sc}Dir and the far-eld pattern u^{∞}Dir

1: Choose an incident direction ˆz ∈ S^{1};

2: Compute the scattered eld u^{sc}_{ref}(x, −ˆz) for the acoustic equation (1.1) without
obstacle by the incident eld u^{inc}ref(x, −ˆz) = e^{ikx·(−ˆ}^{z)};

3: Compute the total eld u^{to}ref(x, −ˆz) = u^{sc}_{ref}(x, −ˆz) + u^{inc}_{ref}(x, −ˆz);

4: Compute the scattered eld u^{sc}Dir(x, ˆz) for the acoustic equation (3.2) with incident

eld u^{inc}A,n(x, ˆz) = u^{to}_{ref}(x, −ˆz);

5: Compute the far-eld of u^{∞}Dir and generate the far-eld operator FDir.

We now discuss the algorithm in more detail. In our simulation, we take A ≡ I and n(x, y) =

(

1 + e^{−}^{1+x2+y2}^{1} for x^{2}+ y^{2} < 1

1 otherwise

which has jump discontinuities at x^{2}+ y^{2} = 1. The original problem was formulated
in R^{2}, together with the Sommerfeld radiation condition at innity. Here, we restrict
the computational domain in the ball x^{2} + y^{2} < 4 and approximate the Sommerfeld
radiation condition by the impedance condition:

(∆˜u^{sc}_{ref} + k^{2}n˜u^{sc}_{ref} = k^{2}(1 − n)u^{inc}_{ref} in x^{2}+ y^{2} < 4,

∂_{r}u˜^{sc}_{ref} − ik ˜u^{sc}_{ref} = 0 on x^{2}+ y^{2} = 4.

We solve this boundary valued problem FEM with mesh size ≤ 0.1. Then the approximated
total eld ˜u^{to}_{ref} = ˜u^{sc}_{ref} + u^{inc}_{ref} and the needed incident eld

˜

u^{inc}_{I,n}(x, ˆz) = ˜u^{to}_{ref}(x, −ˆz).