IOP P
UBLISHING
I
NVERSE
P
ROBLEMS
Inverse Problems
24
(2008) 025007 (9pp)
doi:10.1088/02665611/24/2/025007
A simple method for solving the inverse scatteringproblem for the difference Helmholtz equation
Yuri A Godin and Boris Vainberg
Department of Mathematics and Statistics, University of North Carolina at Charlotte, Charlotte,NC 28223, USA
Received 8 November 2007, in ﬁnal form 27 December 2007Published 1 February 2008Online atstacks.iop.org/IP/24/025007
Abstract
The inverse scattering problem of determining the boundary impedance fromknowledge of the time harmonic incident wave and the farﬁeld pattern of thescattered wave is considered. We use the ﬁnitedifference approximation fortheHelmholtzequationalongwiththeexactradiationconditionforthediscreteequation. The approach involves two steps. First, we reduce the problem to awellposedsystemoflinearequationsforamodiﬁedpotential. Wenextﬁndtheboundary impedance using the modiﬁed potential through an explicit formula.Thus, the computational part of the nonlinear problem of reconstruction of theboundary impedance is reduced to the solution of a linear system. Numericalexamples are given to demonstrate efﬁciency of the new approach.(Some ﬁgures in this article are in colour only in the electronic version)
1. Introduction
Consider the Helmholtz equation in the half space
R
n
,x
n
>
0
,
−
u
=
k
2
u, x
n
>
0
,
(1)with the boundary condition
u
x
n
−
γ(
x
)u

x
n
=
0
=
0
,
x
=
(x
1
,x
2
,...,x
n
−
1
,x
n
=
0
),
(2)where
γ(
x
)
isthesurfaceimpedancewithaboundedsupportand
u
=
u(
x
)
isthesuperpositionof an incident, reﬂected and scattered wave
u(
x
)
=
e
i
ω
·
x
+ e
i
ω
∗
·
x
+
ψ(
x
).
(3)Here,
ω
=
(ω
1
,ω
2
,...,ω
n
)
is a vector such that

ω
 =
k,
ω
∗
=
(ω
1
,ω
2
,...,ω
n
−
1
,
−
ω
n
)
and the function
ψ(
x
)
satisﬁes the radiation condition
ψ(
x
)
=
e
i
k

x


x

n
−
12
f
ω
,
x

x

+
O
1

x

.
(4)
02665611/08/025007+09$30.00 © 2008 IOP Publishing Ltd Printed in the UK 1
Inverse Problems
24
(2008) 025007 Y A Godin and B Vainberg
The inverse scattering problem for(1) and (2)consists in determining the impedance
γ(
x
)
by the scattering amplitude
f
.Onecanreducetheproblemtothewholespace
R
n
byextendingfunction
u
evenlythroughthe boundary
x
n
=
0 for
x
n
<
0. Then equations (1)and (2) are replaced by the Schr¨odinger
equation
(
−
+
q)u
=
k
2
u,
x
∈
R
n
,
(5)where
q(
x
)
=
2
γ(
x
)δ(x
n
)
and
δ(x)
is the Dirac deltafunction.However, instead of pursuing the continuous case, we consider the ﬁnitedifferencecounterpart of (3)–(5). Equation(5) is then replaced by the difference equation
(
−
+
q)u
=
k
2
u, u
=
u(
ξ
),
ξ
∈
Z
nh
,
(6)with
q(
ξ
)
=
2
γ(
ξ
)δ(ξ
n
)
, and the difference Laplacian
being deﬁned on the lattice
Z
nh
withstep
h
as
u(
ξ
)
=
h
−
2
η
:

η
−
ξ
=
h
[
u(
η
)
−
u(
ξ
)
]
.
(7)The special feature of this paper is related to the radiation condition. We do not discretizethe continuous radiation condition (4). Instead, we use the exact radiation condition for the
discrete Schr¨odinger equation(6) that provides a unique solvability of the discrete problem
on the lattice [1,2]. Thus, the problem from the very beginning is discrete and, as we will
show, is well posed for a ﬁxed step
h
. This makes the method differ from other approachessuch as the linear sampling method [3–6], the leastsquares method [7,8], or the factorization
method[9] (see[10,11] for review of other methods). Of course, the standard procedure of
approximation of the continuous model by the discrete counterpart has been used extensively.Thenoveltyofourapproachisrelatedtotheradiationconditionforthediscreteproblemwhichwas found quite recently[1,2] and is not widely known.
We do not consider the limiting transition
h
→
0 because no one can expect that thesolution of the difference problem has the limit as
h
→
0 without restrictions on the set of possible data. However, one can expect that some intermediate range of values of
h
can givea good approximation of the solution of the srcinal continuous problem.The Fourier transform
−
u
of the negative difference Laplacian is the operator of multiplication by
φ(
σ
)
=
2
h
−
2
n
−
n
i
=
1
cos
σ
i
.
(8)That means
−
u
=
φ(
σ
)
u(
σ
),
u(
σ
)
=
1
(
2
π)
n/
2
ξ
∈
Z
nh
u(
ξ
)
e
i
ξ
·
σ
,
σ
∈
T
,
(9)where
T
is [
−
πh
−
1
,πh
−
1
]
n
.Itfollowsfrom(9)thatthespectrumSp
(
−
)
of
−
isabsolutelycontinuousandcoincideswith the interval 0
k
2
4
nh
−
2
.Similar to (1)–(5), the scattering solution of the difference equation on the lattice
Z
nh
,ξ
n
>
0 is deﬁned using the reﬂection principle: it is a solution
u(
ξ
)
of equation (6)that has the form
u(
ξ
)
=
e
i
·
ξ
+ e
i
∗
·
ξ
+
ψ(
ξ
),
=
(
,
n
),
∗
=
(
,
−
n
),
=
(
1
,
2
,...,
n
−
1
,
n
=
0
),
(10)
2
Inverse Problems
24
(2008) 025007 Y A Godin and B Vainberg
where function e
i
·
ξ
obeys the homogeneous equation
(
−
−
k
2
)
e
i
·
ξ
=
0 (11)and
ψ(
ξ
)
satisﬁes the same equation and the radiation condition.Equation(11) implies the following relation between
and
k
:2
n
−
2
n
i
=
1
cos
i
=
(kh)
2
.
(12)The radiation condition for the difference equation (6)can be found in [1,2]. The
form of the radiation condition depends essentially on the value of
k
2
in the continuousspectrum Sp
(
−
)
=
[0
,
4
nh
−
2
]. We assume that
k
2
belongs to one of the intervals
(
0
,
4
h
−
2
)
or
((
4
n
−
4
)h
−
2
,
4
nh
−
2
)
. In the complement part of the spectrum, the dispersion surface
S
= {
σ
:
φ(
σ
)
=
k
2
}
is not strictly convex. This results in the existence of several scatteringwaves with different frequencies propagating in the same direction [2]. Thus, if
n
=
2 thenSp
(
−
)
=
[0
,
8
h
−
2
] and
k
2
∈
(
0
,
8
h
−
2
)
\{
4
h
−
2
}
. If
n
=
3 then Sp
(
−
)
=
[0
,
12
h
−
2
] and
k
2
∈
(
0
,
4
h
−
2
)
∪
(
8
h
−
2
,
12
h
−
2
)
.Under those restrictions on
k
2
, the radiation condition is similar to that for the continuousLaplacian: there is only one scattering wave. However, its frequency depends on the direction
θ
=
ξ
/

ξ

ψ(
ξ
)
=
e
i
µ(
θ
,k)

ξ


ξ

n
−
12
f(
θ
,k)
+
O
1

ξ

.
(13)Here,
µ(
θ
,k)
=
σ
(
θ
,k)
·
ξ

ξ

is the projection of the vector
σ
=
σ
(
θ
,k)
onto the directionof vector
ξ
, and
σ
(
θ
,k)
is a point on the dispersion surface
S
= {
σ
:
φ(
σ
)
=
k
2
}
where theexternal normal vector is parallel to
ξ
.The goal of this paper is to recover the impedance
γ
(or the potential
q
) by measuring thescattering amplitude
f(
θ
,k)
for a ﬁxed frequency
k
. Observe that the impedance
γ
is deﬁnedin a ﬁnite number of points on the lattice
Z
n
−
1
h
, while the amplitude
f(
θ
,k)
is a function of continuous argument
θ
. We will obtain an exact description of a ﬁnitedimensional space of admissible functions
f
. We assume that
f
is known (measured) in a number of points
θ
m
which exceeds substantially the number of points on the support of
γ
. Typically, the numberof measurements should be by an order of magnitude larger than the number of points on thesupport
γ
. It also depends on the noise level and its optimal choice is discussed in the sectionon numerical results. Then the leastsquares method is used to recover the amplitude in theadmissible ﬁnitedimensional space and then recover
γ
.There is a crucial observation in [12] related to the inverse difference problem for theSchr¨odinger equation in
R
n
with an arbitrary compactly supported potential
q
(which is notnecessarily supported on
x
n
=
0). It was shown that the nonlinear operator
F
:
f
→
q
can berepresented in the form
F
=
F
1
F
2
, where
F
2
is a linear operator and the nonlinear operator
F
1
admits the explicit inverse
F
−
11
. This result did not have further development because theoperator
F
2
is highly degenerate: the matrix of this operator has order
N
n
, while the rank is
O(N
n
−
1
)
, where
N
the side of the cube supported the potential. We use this approach, and inour case the operator
F
2
is not degenerate. Hence, we need to solve only the linear problem(ﬁnd
F
−
12
). Of course, the problem becomes bad if
h
→
0 and when the size of the matrix
F
2
grows.The same approach works in the problem of ﬁnding the boundary of unknown convexobstacle in
R
n
, since the number of unknowns in this case is not
O(N
n
)
, but rather
O(N
n
−
1
)
,as in the case with unknown impedance. Here we consider only the impedance problem inthe half space. The problem of recovery of the boundary of an obstacle will be discussedelsewhere.
3
Inverse Problems
24
(2008) 025007 Y A Godin and B Vainberg
2. Modiﬁed potential
c
(
ξ
)
Substituting(10) into(6), we obtain that the scattering solution
ψ(
ξ
)
satisﬁes the equation
(
−
+
q(
ξ
)
−
k
2
)ψ
= −
q(
ξ
)(
e
i
·
ξ
+ e
i
∗
·
ξ
)
= −
2
q(
ξ
)
e
i
·
ξ
.
(14)Equation (14) is uniquely solvable if
ψ
satisﬁes the radiation condition (13)and
k
2
belongs to
((
4
n
−
4
)h
−
2
,
4
nh
−
2
)
(see [2]). From (14)it follows
(
−
−
k
2
)ψ
= −
q(
ξ
)(
2e
i
·
ξ
+
ψ).
(15)Let us denote by
c(
ξ
)
the righthand side of (15)
c(
ξ
)
= −
q(
ξ
)(
2e
i
·
ξ
+
ψ).
(16)In this notation, equation (15)has the form
(
−
−
k
2
)ψ
=
c(
ξ
).
(17)Observe that the coefﬁcient
c(
ξ
)
vanishes outside the support of
q(
ξ
)
, and hence the solutionof the difference equation (17)can be written as a ﬁnite sum
ψ
=
j
G(
ξ
−
ξ
j
)c(
ξ
j
),
ξ
j
∈
supp
q,
(18)where
G(
ξ
−
ξ
j
)
is the Green’s function of (17) which is determined as a limit
z
→
k
+ i0 of the square integrable Green’s function for the operator
−
−
z
2
. Substituting (18) into(16),
we obtain equation for determining
q(
ξ
)c(
ξ
)
= −
q(
ξ
)
⎛⎝
2e
i
·
ξ
+
j
G(
ξ
−
ξ
j
)c(
ξ
j
)
⎞⎠
,
(19)from which one can ﬁnd
q(
ξ
s
)
= −
c(
ξ
s
)
2e
i
·
ξ
s
+
j
G(
ξ
s
−
ξ
j
)c(
ξ
j
),
ξ
j
,
ξ
s
∈
supp
q.
(20)Thus, (20)allows to ﬁnd explicitly
q(
ξ
s
)
if
c(
ξ
s
)
is known. In the following section, we willdescribe a method of determining
c(
ξ
s
)
from the scattering amplitude
f(
θ
,k)
with ﬁxed
k
.Note that the mapping
c
→
f
is linear. Hence, the initial nonlinear inverse problem is splitinto two steps: solution of a linear problem (restoring
c
from
f
) and application of the explicitformula (20).
3. Calculation of the modiﬁed potential
c
(
ξ
)
TheGreen’sfunction
G(
ξ
)
ofthedifferenceequation(6)hasthefollowingasymptoticbehavior
at inﬁnity [2]:
G(
ξ
)
=
e
i
µ(
θ
,k)

ξ


ξ

n
−
12
b(
θ
,k)
1 +
O
1

ξ

,

ξ
 → ∞
.
(21)The Green’s function of a shifted argument behaves as
G(
ξ
−
ξ
j
)
=
e
i
µ(
θ
,k)

ξ


ξ

n
−
12
b(
θ
,k)
e
−
i
σ
(
θ
,k)
·
ξ
j
1 +
O
1

ξ

,

ξ
 → ∞
.
(22)Here,
b(
θ
,k)
is deﬁned by the curvature
(
θ
,k)
at the point
σ
=
σ
(
θ
,k)
on the surface
S
:
φ(
σ
)
=
k
2
b(
θ
,k)
=√
2
π
e
i
(n
+1
)
π
4
√ 
(
θ
,k)
∇
φ(
σ
(
θ
,k))

.
(23)
4
Inverse Problems
24
(2008) 025007 Y A Godin and B Vainberg
In twodimensional case
n
=
2, and the curvature
(
θ
,k)
in (23)of the curve
S
:2
−
cos
σ
1
−
cos
σ
2
=
12
h
2
k
2
at the point
(σ
1
,σ
2
)
is
=
cos
σ
1
sin
2
σ
2
+ cos
σ
2
sin
2
σ
1
(
sin
2
σ
1
+ sin
2
σ
2
)
3
.
(24)Since
∇
φ
=
(
sin
σ
1
,
sin
σ
2
),
(25)formula (23)becomes
b(
θ
,k)
=√
2
π
e
3
π
i4
(
sin
2
σ
1
+ sin
2
σ
2
)

cos
σ
1
sin
2
σ
2
+ cos
σ
2
sin
2
σ
1

.
(26)Then (18)and (22)imply that
f(
θ
,k)
=
b(
θ
,k)
j
c(
ξ
j
)
e
−
i
σ
(
θ
,k)
·
ξ
j
,
ξ
j
∈
supp
q,

ξ
j

< Mh
−
1
.
(27)Equation (27) shows that the exact value of the scattering amplitude belongs to a ﬁnitedimensional space span
j
{
b(
θ
,k)
e
i
σ
(
θ
,k)
·
ξ
j
}
of linearly independent functions, since theelements
{
e
i
σ
m
(
θ
,k)
ξ
j
}
,
1
m,j
N,
form a Vandermonde matrix which is not degenerate(see the appendix). Taking measurements of the scattering amplitude
f(
θ
,k)
, we can ﬁnd thecoefﬁcients
c(
ξ
j
)
using the leastsquares method.
4. Solution of the direct problem
Here we consider twodimensional case
n
=
2. Equation (14)implies that the function
ψ
onthe boundary
ξ
2
=
0 satisﬁes the equation
ψ(ξ
1
,
0
)
+
j
γ
ξ
j
1
G
ξ
1
−
ξ
j
1
,
0
ψ
ξ
j
1
,
0
= −
2
j
γ
ξ
j
1
G
ξ
1
−
ξ
j
1
,
0
e
i
1
ξ
j
1
,

ξ
j
1

,

ξ
1

< mh
−
1
,
(28)where
G(ξ
1
,
0
)
=
12
π
π
−
π
π
−
π
e
i
ξ
1
x
d
x
d
y
2cos
x
+ 2cos
y
−
k
2
.
(29)Thus, calculation of
ψ(ξ
1
,
0
),

ξ
1

< mh
−
1
, is reduced to the solution of the linear system of equations. This allows us to determine
ψ
on the whole lattice
Z
2
h
as follows:
ψ(
ξ
)
= −
j
γ
ξ
j
1
G
ξ
1
−
ξ
j
1
,ξ
2
ψ
ξ
j
1
,
0
+ 2e
i
1
ξ
j
1
, ξ
∈
Z
2
h
,
(30)and ﬁnd the scattering amplitude using (22)
f(
θ
,k)
= −
b(
θ
,k)
j
γ
ξ
j
1
e
−
i
σ
(
θ
,k)
·
ξ
j
ψ
ξ
j
1
,
0
+ 2e
i
1
ξ
j
1
.
(31)
5