DIRICHLET-TO-NEUMANN

BOUNDARY CONDITIONS FOR

MULTIPLE SCATTERING PROBLEMS
Marcus J. Grote, Christoph Kirsch

Department of Mathematics
University of Basel
Rheinsprung 21
CH - 4051 Basel
Switzerland

Preprint No. 2004-4

March 2004
www.math.unibas.ch

Dirichlet-to-Neumann Boundary Conditions

for Multiple Scattering Problems
Marcus J. Grote , Christoph Kirsch
Department of Mathematics, University of Basel,
Rheinsprung 21, CH-4051 Basel, Switzerland

Abstract
A Dirichlet-to-Neumann (DtN) condition is derived for the numerical solution of
time-harmonic multiple scattering problems, where the scatterer consists of several
disjoint components. It is obtained by combining contributions from multiple purely
outgoing wave elds. The DtN condition yields an exact nonreecting boundary condition for the situation, where the computational domain and its exterior articial
boundary consist of several disjoint components. Because each sub-scatterer can
be enclosed by a separate articial boundary, the computational eort is greatly
reduced and becomes independent of the relative distances between the dierent
sub-domains.
The DtN condition naturally ts into a variational formulation of the boundary
value problem for use with the nite element method. Moreover, it immediately
yields as a by-product an exact formula for the far-eld pattern of the scattered
eld.
Numerical examples show that the DtN condition for multiple scattering is as
accurate as the well-known DtN condition for single scattering problems [6,7], while
being more ecient due to the reduced size of the computational domain.

Introduction

For the numerical solution of scattering problems in innite domains, a wellknown approach is to enclose all obstacles, inhomogeneities and nonlinearities
with an articial boundary B. A boundary condition is then imposed on B,
To whom correspondence should be addressed.
Email addresses: [email protected] (Marcus J. Grote),
[email protected] (Christoph Kirsch).

Preprint submitted to Journal of Computational Physics

March 12, 2004

which leads to a numerically solvable boundary value problem in a nite domain . The boundary condition should be chosen such that the solution of
the problem in coincides with the restriction to of the solution in the
original unbounded region.
If the scatterer consists of several obstacles, which are well separated from
each other, the use of a single articial boundary to enclose the entire scattering region, becomes too expensive. Instead it is preferable to enclose every
sub-scatterer by a separate articial boundary Bj . Then we seek an exact


boundary condition on B = Bj , where each Bj surrounds a single computational sub-domain j . This boundary condition must not only let outgoing
waves leave j without spurious reection from Bj , but also propagate the
outgoing wave from j to all other sub-domains
, which it may reenter subsequently. To derive such an exact boundary condition, an analytic expression
for the solution everywhere in the exterior region is needed. Neither absorbing
boundary conditions [1,2], nor perfectly matched layers [3,4,5] provide us with
such a representation. Instead we shall seek a Dirichlet-to-Neumann (DtN)
boundary condition, which is based on a Fourier series representation of the
solution in the exterior region.
Exact DtN conditions have been derived for various equations and geometries,
but always in the situation of a single computational domain, where the scattered eld is purely outgoing outside [6,7,8,9,10]. In a situation of multiple
disjoint computational domains, however, waves are not purely outgoing out

side the computational domain = j , as they may bounce back and forth
between domains. We shall show how to overcome this diculty and derive an
exact DtN condition for multiple scattering. The derivation presented below
for the Helmholtz equation in two space dimensions readily extends to multiple scattering problems in other geometries and also to dierent equations.
Because this exact boundary condition allows the size of the computational
sub-domains, j , to be chosen independently of the relative distances between
them, the computational domain, , can be chosen much smaller than that
resulting from the use of a single, large computational domain.
There is an extended literature on the solution of multiple scattering problems see Martin [11] for an introduction and overview. Due to the diculties
mentioned above, numerical methods used for multiple scattering so far have
mainly been based on integral representations [12,13], while in the single scattering case many alternative methods, such as absorbing boundary conditions,
perfectly matched layers, or the DtN approach are known. To our knowledge
this work constitutes the rst attempt to generalize the well-known DtN approach to multiple scattering.
Some of the analytical techniques we shall use, have been known in the classical scattering literature for quite some time. For instance, in 1913 Zaviska
2

[14] considered multiple scattering from an array of parallel circular cylinders.

He derived an innite linear system for the unknown Fourier coecients of the
scattered eld, which involve Fourier expansions of the purely outgoing wave
elds about individual cylindrical obstacles. This method can be generalized
to cylinders with non-circular cross-sections [15]. Another class of methods is
based on single and double layer potentials, which involve integration with
the Greens function over the articial boundary. From this representation,
systems of integral equations can be derived for multiple scattering problems
see Twersky [16] and Burke and Twersky [17] for an extensive overview of
previous work until 1964, and [11] for more recent references.
In Section 2, we derive the DtN and modied DtN map for two scatterers. We
show that the solution to the boundary value problem in , with the DtN condition imposed on B, coincides with the restriction to of the solution in the
unbounded region . The formulation is generalized to an arbitrary number
of scatterers in Section 3. In Section 4, we state a variational formulation of
the articial boundary value problem for use with the nite element method.
An explicit formula for the far-eld pattern of the solution, based on the decomposition of the scattered eld into multiple purely outgoing wave elds,
is derived in Section 5. Finally, in Section 6, we consider a nite dierence
implementation of the multiple-DtN method and demonstrate its accuracy
and convergence. We also compare the multiple-DtN approach to the wellknown (single-)DtN method and show that both the numerical solutions and
the far-eld patterns, obtained by these two dierent methods, coincide.

Two scatterers

We consider acoustic wave scattering from two bounded disjoint scatterers in

unbounded two-dimensional space. Each scatterer may contain one or several
obstacles, inhomogeneities, and nonlinearity. We let denote the piecewise
smooth boundary of all obstacles and impose on a Dirichlet-type boundary
condition, for simplicity. In , the region outside , the scattered eld u =
u(r, ) then solves the exterior boundary value problem
u + k 2 u = f in R2 ,
u = g on ,



lim r
ik u = 0.
r
r

(1)
(2)
(3)

The wave number k and the source term f may vary in space, while f may be
nonlinear. The Sommerfeld radiation condition (3) ensures that the scattered
eld corresponds to a purely outgoing wave at innity.
3

Next, we assume that both scatterers are well separated, that is we assume
that we can surround them by two non-intersecting circles B1 , B2 centered at
c1 , c2 with radii R1 , R2 , respectively. In the unbounded region D, outside the
two circles, we assume that the wave number k > 0 is constant, and that f
vanishes. In D the scattered eld u thus satises
u + k 2 u = 0 in D, k > 0 constant,



lim r
ik u = 0.
r
r

(4)
(5)

We wish to compute the scattered eld, u, in the computational domain =

\ D, which consists of the two disjoint components 1 and 2 . A typical
conguration with two obstacles is shown in Fig. 1. Here the computational
domain is internally bounded by = 1 2 , and externally by B = D,
which consists of the two circles B1 and B2 .
To solve the scattering problem (1)(3) inside , a boundary condition is
needed at the exterior articial boundary B. This boundary condition must
ensure that the solution in , with that boundary condition imposed on B,
coincides with the restriction of the solution in the original unbounded region
.

2.1 Derivation of the DtN map

On B we shall now derive a DtN map, which establishes an exact relation

between the values of u and its normal derivative. In contrast to the case of a
single circular articial boundary, as considered for example by Givoli [7] and
Grote and Keller [8], we cannot simply expand u outside B in a Fourier series.
First, there is no separable coordinate system outside D for the Helmholtz
equation [18] and second, u is not purely outgoing in D. Indeed, part of the
scattered eld leaving 1 will reenter 2 , and vice-versa. Hence the boundary
condition we seek on B must not only let outgoing waves leave 1 without
spurious reection from B1 , but also propagate the outgoing wave eld from
1 to 2 , and vice-versa, without any spurious reection.
We begin the derivation of an exact nonreecting boundary condition on B =
B1 B2 by introducing a local polar coordinate system (rj , j ) outside each
circle Bj , centered at cj (see Fig. 1). Next, we denote by D1 the unbounded
domain outside B1 with r1 R1 , and by D2 the unbounded domain outside
B2 with r2 R2 . We now decompose the scattered eld u in D into two purely
outgoing wave elds u1 and u2, which solve the following problems:
4

R1

r1
c1

1
x1

B1

r
2d

R2
r2

c2
B2

x2

Figure 1. A typical conguration with two obstacles bounded by 1 and 2 is shown.

The computational domain = 1 2 is externally bounded by the articial
boundary B = B1 B2 . In each domain component j , we use a local polar coordinate system (rj , j ), while (r, ) denotes the global polar coordinate system centered
at the origin.

u1 + k 2 u1 = 0 in D1 ,


ik u1 = 0,
lim r
r
r

(6)

u2 + k 2 u2 = 0 in D2 ,



ik u2 = 0.
lim r
r
r

(8)

(7)

and

(9)

Each wave eld is inuenced only by a single scatterer and completely oblivious
to the other. Therefore, u1 and u2 are entirely determined by their values on
B1 or B2 , respectively; they are given in local polar coordinates (r1 , 1 ), (r2 , 2 )
by


Hn(1) (krj )
1
uj (rj , j ) =
n=0 Hn(1) (kRj )

2

uj (Rj ,  ) cos n(j  ) d ,

rj Rj ,

(10)

for j = 1, 2. Here the prime after the sum indicates that the term for n = 0
is multiplied by 1/2, while Hn(1) denotes the n-th order Hankel function of the
rst kind. We now couple u1 and u2 with u by matching u1 + u2 with u on
B = B1 B2 :
(11)
u1 + u2 = u on B.
Both u and u1 + u2 solve the homogeneous Helmholtz equation (4) in D =
D1 D2 , together with the Sommerfeld radiation condition (5) at innity.
Since u and u1 + u2 coincide on B, they coincide everywhere in the exterior
region D. We summarize this result in the following proposition. Moreover,
before proceeding with the derivation of the DtN map, we shall also prove
that such a decomposition always exists and is unique.
Proposition 1 Let u be the unique solution to the free space problem (1)(3)
and assume that u satises (4), (5) in the exterior region, D. Then
u u1 + u2 ,

in B D,

(12)

where u1 and u2 are solutions to the problems (6), (7) and (8), (9), respectively,
together with the matching condition (11). The decomposition of u into the two
purely outgoing wave elds u1 and u2 is unique.

PROOF. By the argument above, we have already shown that if u = u1 + u2

on B, and u1 and u2 solve (6)(9), then u u1 + u2 everywhere in D. We
shall now show that u1 and u2 exist and, in fact, are unique.
Existence: In the exterior domain D we use the Kirchho-Helmholtz formula
[19] to write
 

u(x) =
B

(x, y) u
u(y)

(y)(x, y) ds(y),
n(y)
n

x D.

(13)

Here is the fundamental solution of the Helmholtz equation,

i (1)
(x, y) = H0 (k|x y|),
4
6

x = y,

(14)

while n denotes the outward normal on the articial boundary B. Let

 

uj (x) :=
Bj

(x, y) u
u(y)

(y)(x, y) ds(y),
n(y)
n

x Dj ,

(15)

for j = 1, 2. Then, a straightforward calculation shows that u1 satises (6), (7)

whereas u2 satises (8), (9). Clearly, u(x) = u1 (x) + u2(x), x D = D1 D2 .
The expressions (13) and (15) can be continuously extended up to the articial
boundaries B and B1 , B2 , respectively ([20], Theorem 2.13). Thus, u1 and u2
also satisfy the matching condition (11).
Uniqueness: Following a suggestion of Tordeux [21], we let u v1 + v2 be
another decomposition in B D, where v1 and v2 solve (6), (7) and (8), (9),
respectively. We shall now show that v1 u1 and that v2 u2 throughout D.
To do so, we let w1 := u1 v1 and w2 := u2 v2 . Hence, w1 and w2 satisfy
(6), (7) and (8), (9), respectively. Because w2 is regular throughout D2 , it is
also regular, and therefore bounded, everywhere inside B1 , including the local
origin, c1 . Thus, in the vicinity of B1 , w1 and w2 can be written in the local
polar coordinates, (r1 , 1 ), as
w1 (r1 , 1 ) =
w2 (r1 , 1 ) =


nZ

nZ

an Hn(1) (kr1 )ein1 ,

(16)

bn Jn (kr1 )ein1 ,

(17)

for r1 I := [R1 , R1 + ], with = |c2 c1 | (R1 + R2 ) > 0, because the

scatterers are assumed to be well separated. From the uniqueness of u we
obtain w1 + w2 = u1 + u2 (v1 + v2 ) 0 in B D. Therefore
an Hn(1) (kr1 ) + bn Jn (kr1 ) = 0,

n Z, r1 I.

(18)

Since Hn(1) and Jn are two linearly independent solutions of Bessels dierential
equation, we conclude that an = bn = 0 for all n Z. Thus, v1 u1 and
v2 u2 in B D. 2

As a consequence of the proposition, we can now explicitly determine a DtN

map for u by dierentiating u with respect to the outward normal n on B1
and B2 as follows:
n u = M[u1 ] + T [u2 ] on B1 ,
n u = M[u2 ] + T [u1 ] on B2 ,
u1 + P [u2 ] = u on B1 ,
P [u1 ] + u2 = u on B2 .
7

(19)
(20)
(21)
(22)

Here the operator M corresponds to the standard single-DtN operator



kHn(1) (kRj )
1
M[uj ](j ) :=
n=0 Hn(1) (kRj )

2

uj (Rj ,  ) cos n(j  ) d ,

(23)

j = 1, 2. The transfer operator T and propagation operator P are given by

u1
u2
(R2 , 2 ), T [u2 ](1 ) :=
(R1 , 1 ),
r2
r1
P [u1 ](2 ) := u1 (R2 , 2 ), P [u2 ](1 ) := u2 (R1 , 1 ).
T [u1 ](2 ) :=

(24)
(25)

The expressions on the right-hand sides of (19), (20) and on the left-hand
sides of (21), (22) can be evaluated explicitly by using the denitions (23)
(25) and the (exact) Fourier representation (10), valid in each local coordinate
system. These calculations involve some technical but straightforward coordinate transformations. For instance, in the particular situation shown in Fig. 1,
T [u2 ] and P [u2] are explicitly given (in local polar (r1 , 1 )-coordinates) on B1
for 1 [0, 2) by
T [u2 ](1 ) =


1
kHn(1) (kr2 )
1
(R1 + 2d sin 1 )
r2
n=0 Hn(1) (kR2 )



1
nHn(1) (kr2 )
1
+ 2d cos 1
r2
n=0 Hn(1) (kR2 )

2

2

u2 (R2 ,  ) cos n(2  ) d +

u2 (R2 , ) sin n(2 ) d ,

(26)

Hn(1) (kr2 )
1
P [u2 ](1 ) =
n=0 Hn(1) (kR2 )

2

u2 (R2 ,  ) cos n(2  ) d,

(27)

where

r2 = R12 + 4dR1 sin 1 + 4d2 ,

1
sin 2 = (R1 sin 1 + 2d) ,
r2
1
cos 2 = R1 cos 1 .
r2

(28)
(29)
(30)

The expressions for T [u1] and P [u1] on B2 are similar to (26)(30), with r2
replaced by r1 , 2 by 1 , etc.
The matching condition (21), (22) cannot be inverted explicitly, and u1 and
u2 thereby eliminated from the DtN condition (19)(22). Instead, we shall
8

compute the values of u1 on B1 and u2 on B2 , in addition to the values of

u. These auxiliary values are also useful during post-processing, as they yield
explicit expressions both for u everywhere outside and for its far-eld pattern
see Section 5.
With the DtN condition given by (19)(22), we now state the boundary value
problem for u inside the computational domain = 1 2 :
u + k 2 u = f in
u = g on
n u = M[u1 ] + T [u2 ],
n u = M[u2 ] + T [u1 ],
u1 + P [u2 ] = u on B1
P [u1 ] + u2 = u on B2

(31)
(32)
(33)
(34)
(35)
(36)

on B1
on B2

We now show that this boundary value problem has a unique solution, which
coincides with the solution to the original problem (1)(3).
Theorem 2 Let u be the unique solution to the free space problem (1)(3)
and assume that u satises (4), (5) in the exterior region, D. Then the double
scattering boundary value problem (31)(36) has a unique solution in , which
coincides with the restriction of u to .

PROOF. Existence: We shall show that u| is a solution to (31)(36). Since

u satises (1), (2) it trivially satises (31), (32). To show that u| satises the
DtN condition (33)(36) on B, we consider in B D the unique decomposition
u u1 +u2 , provided by Proposition 1. Since u1 +u2 satises the DtN boundary
condition (33)(36) on B, by construction, so does the restriction of u to .
Therefore, u| is a solution to the boundary value problem (31)(36).
Uniqueness: We extend the argument of Harari and Hughes [22] for a single
scatterer to the case of two scatterers. Let v, together with v1 |B1 and v2 |B2 ,
denote another solution of (31)(36). We shall show that v u| . First, we
denote by


Hn(1) (krj )
1
vj (rj , j ) :=
n=0 Hn(1) (kRj )

2

vj (Rj ,  ) cos n(j  ) d,

(37)

the two purely outgoing wave elds, dened for rj Rj , j = 1, 2. Next, we

construct an extension
v :=

v,

in

1 + v2 , in B D

(38)

of v into the exterior region D. We shall now show that w := u v vanishes

in . To begin, we remark that w and its normal derivative are continuous
everywhere in , while w satises w + k 2 w = 0 in and w = 0 on . By
using integration by parts we now nd that


|w|2 k 2 |w|2 dx =

w
B

w
ds,
n

(39)

from which we infer that



w
B

w
w
w
ds = 0.
n
n

(40)

Let Br denote the sphere of radius r > 0 centered at the origin. Again we use
integration by parts, (40) and the fact that w is a solution of (4) to obtain


0=
B

w
w
w
ds =
w
n
n

= lim

r
Br

ww ww dx lim

r
Br

w
w
w
ds
r
r

w
w
w
ds.
r
r

(41)

From the radiation condition (5) and (41) we now infer that
  


0 = lim  r

r
r

ik

Br

2
 

 w 
= r
lim r  
r
Br

2
 

 w 
= lim r  
 r 
r

2


w


ds


w
w
+ k |w| ik w
w
r
r
2

+ k 2 |w|2 ds.

ds
(42)

Br

Since k 2 > 0 we conclude that



lim
r

|w|2 ds = 0.

(43)

Br

Equation (43) then implies that w 0 in D, by Rellichs theorem ([19], Lemma

2.11). By continuity, we also have w = 0 on B. Finally we apply Proposition 1
to w, which yields the unique decomposition w w1 + w2 with w1 0 and
w2 0 in B D. Because of the DtN condition (33)(36) we conclude that
n w = 0 on B. Since the problem
10

w + k 2 w = 0
w=0
w=0
n w = 0

in
on
on B
on B

(44)
(45)
(46)
(47)

has only the trivial solution (which is veried directly by expanding the solution of (44) in a Fourier series and by using the linear independence of the
Hankel functions), w 0 in or v u| . 2

2.2 The modied DtN map

In practice the innite sums which occur in the operators M, T , and P in

the DtN condition (33)(36) have to be truncated at some nite N 0. The
corresponding truncated operators are denoted by M N , T N , and P N . Even
in the situation of a single scatterer, truncation can destroy the uniqueness
of the solution in with the truncated DtN condition imposed at B. For
single scattering, Harari and Hughes showed that uniqueness is preserved if
N is chosen large enough [22]. Alternatively, the modied DtN (MDtN) map
introduced in [8] can be used to overcome this diculty. Its generalization to
the case of two scatterers is straightforward:
n u = iku + (M ik)N [u1 ] + (T ikP )N [u2 ] on B1
n u = iku + (M ik)N [u2 ] + (T ikP )N [u1 ] on B2
u1 + P N [u2 ] = u on B1
P N [u1 ] + u2 = u on B2

(48)
(49)
(50)
(51)

Numerical results with the MDtN map applied to multiple scattering are
shown in Section 6.1. They corroborate the expected improvement in accuracy
and stability, well-known in the single scatterer case.

Multiple scattering problems

The derivation of the DtN map presented above for two scatterers is easily
generalized to the case of several scatterers. We consider a situation with J
scatterers, and surround each scatterer by a circle Bj of radius Rj . Again we


denote by B = Jj=1 Bj the entire articial boundary and by Dj the unbounded


region outside the j-th circle. Hence the computational domain = Jj=1 j ,
where j denotes the nite computational region inside Bj , whereas D =
J
j=1 Dj denotes the unbounded exterior region.
11

In D, we now split the scattered eld into J purely outgoing wave elds
u1 , . . . , uJ , which solve the problems
uj + k 2 uj = 0 in Dj ,



ik uj = 0,
lim r
r
r

(52)
(53)

for j = 1, . . . , J. Thus uj is entirely determined by its values on Bj ; it is given

in local polar coordinates (rj , j ) by (10). The matching condition is now given
by
J

j=1

uj = u on B.

(54)

In analogy to Proposition 1 we can show that

u

J

j=1

uj

in B D

(55)

and that this decomposition is unique. Therefore, we immediately nd the

DtN map for a multiple scattering problem with J scatterers:

n u = M[uj ] +

J


T [u
] on Bj ,

(56)

=1


=j

uj +

J


P [u
] = u on Bj ,

j = 1, . . . , J.

(57)

=1


=j

Here M, T and P operate on the purely outgoing wave elds uj as follows:



M : uj |Bj

uj 
u


 , T : u
|B

 , P : u
|B
 u
|Bj .
rj Bj
rj Bj

(58)

We note that no additional analytical derivations due to coordinate transformations, etc. are needed once the situation of two scatterers has been resolved.
Hence, the standard DtN operator M is given by (23), while the operators T
and P are again given by (26)(30), with 1 replaced by j and 2 by , or
vice-versa.
In practice the innite series in the operators M, T and P need to be truncated
at some nite value Nj , which can be dierent for each sub-domain j . We
denote the corresponding truncated operators by M Nj , T Nj , and P Nj , j =
1, . . . , J. For simplicity of notation we shall assume that all boundary operators
are truncated at the same value Nj = N, j = 1, . . . , J.
12

We now extend the modied DtN map (48)(51) to the situation of J scatterers:

n u = iku + (M ik)N [uj ] +

J


(T ikP )N [u
] on Bj ,

(59)

=1


=j

uj +

J


P N [u
] = u on Bj ,

(60)

=1


=j

where N 0 is the truncation index.

For J = 1, the expressions in (56), (57) and (59), (60) reduce to the wellknown DtN and modied DtN conditions for single scattering problems [6,8].
For J = 2, they correspond to the conditions derived previously in Section 2.
To further simplify the notation, we dene the (symbolic) vectors
n u|B = (r1 u|B1 , r2 u|B2 , . . . , rJ u|BJ )
u|B = (u|B1 , u|B2 , . . . , u|BJ )

uout |B = (u1 |B1 , u2|B2 , . . . , uJ |BJ )

(61)
(62)

(63)

and the operator matrices



T = T
j
P

J

j,
=1
 J
j
= P

,
j,
=1

T
j : u
|B
 rj u
|Bj

(64)

P
j : u
|B
 u
|Bj

(65)

With these notations we rewrite the DtN map (56), (57) in matrix-vector
notation as
n u = T uout on B,
P uout = u on B,

(66)
(67)

and the modied DtN (MDtN) map (59), (60) as

n u = iku + (T ikP )N uout
P N uout = u on B.

on B,

(68)
(69)

Remark: The derivation of the DtN (or MDtN) condition for multiple acoustic
scattering can easily be generalized to dierent equations (Maxwells equations
[23], linear elasticity [10], etc), to other geometries (ellipsoidal [8], wave-guide
13

[24]), or to three space dimensions. In fact, our approach can be extended to

all multiple scattering problems, for which a DtN map is already known for
single scattering.

Variational formulation

We shall now show how to combine the multiple scattering DtN condition
(66), (67) with the nite element method in . The computational domain
is bounded in part by B, the union of J disjoint circles, and in part by some
interior piecewise smooth boundary, . For simplicity we consider a Dirichlettype condition on , and assume that the acoustic medium inside is also
homogeneous and isotropic. Hence the boundary value problem in is:
u k 2 u = f in
u = g on
n u = T uout on B
P uout = u on B

(70)
(71)
(72)
(73)

Next, we introduce the function spaces

V = {v H 1 () | v| g},
V0 = {v H 1 () | v| 0}.

(74)
(75)

To derive a variational formulation of (70)(73) we multiply (70) by a test

function v V0 and integrate over . Then we use integration by parts,
together with (71)(73), which yields the following variational formulation for
(70)(73):
Find u V such that
(u, v) (k 2 u, v) (T uout , v)B = (f, v) ,
(P uout , v)B = (u, v)B ,

(76)
(77)

for all v V0 .
Here, (, ) and (, )B denote the standard L2 -inner products on and B,
respectively.
For the nite element discretization of (76), (77) we choose a triangulation Th
of , with mesh size h > 0 and nodes N (Th ) = N N NB . Then we choose
a subspace VN V of nite dimension N = |N (Th )| = N + N + NB , and
14

nodal basis functions

{i }N
i=1 VN ,

i (xj ) = ij ,

xj N (Th ).

(78)

We denote by uh the values of the nite element solution on N , by uhB its

values on NB and by uhout the values of uout see (63) on NB , which yields
from (76), (77) the following linear system of equations:

K
0 I

uhB

0 uh
T
P

uhout

= .
0

(79)

Here I denotes the NB NB identity matrix, while the other entries are given
by
Kij = (j , i ) (k 2 j , i ) , i, j : xi , xj N NB ,
Tij = (T j , i )B , i, j : xi , xj NB ,
Pij = (P j , i )B , i, j : xi , xj NB ,
fi = (f, i )

j:xj N

i : xi N NB .

g(xj )Kij ,

(80)
(81)
(82)
(83)

Because the nodal basis functions {j }N

j=1 are local, K is a sparse real ((N +
NB ) (N + NB ))-matrix. The (NB NB )-matrices T and P , however, have
complex valued entries and are full, because the DtN condition couples all unknowns on B. A typical sparsity pattern for a nite dierence discretization is
shown in Fig. 2. Additional information on nite element analysis for acoustic
scattering can be found in [25].

Far-field evaluation

Once the scattered eld u has been computed inside , it is usually of interest
to evaluate u also outside during a post-processing step, either at selected
locations (receivers) or in a broader region. If integral representations that
involve integration over B with the Greens function, such as (13), are used, the
evaluation of u outside becomes rather cumbersome and expensive. However,
if the multiple-DtN approach is used, the evaluation of u at some location x in
D, the region outside , is inexpensive and straightforward. Indeed, since the
purely outgoing wave elds u1 and u2 are known on B1 and B2 , respectively,
they are known everywhere outside via the Fourier representation (10). In
fact, we can rewrite (10) as
15

1
1
uj (rj , j ) =
Hn(1) (krj ) cos(nj ) (1)
uj (Rj ,  ) cos(n ) d +
n=0
Hn (kRj )





1
1
Hn(1) (krj ) sin(nj ) (1)
n=0
Hn (kRj )

2

uj (Rj ,  ) sin(n ) d,

(84)

where the two integrals correspond to the cosine and sine Fourier coecients
of uj |Bj , j = 1, 2. Thus, to compute u(x) = u1 (x) + u2 (x) at some x D,
it suces to compute the Fourier coecients of uj on Bj , j = 1, 2, yet only
once. Then u1 and u2 , and thereby u = u1 + u2 , can be evaluated anywhere
by summing a few terms in the Fourier representation (84) of u1 and u2 .
Yet another quantity which is often of interest is the far-eld pattern of the
scattered eld u. The asymptotic behavior of any solution u to the exterior
Dirichlet problem (1)(3) is
eikr
u(r, ) f (),
kr

r .

(85)

The function f is called the far-eld pattern of the solution. The value f ()
is the far-eld response from the scatterer in a direction for a given incident
wave. We shall now show how to directly compute f from the values of u1 |B1
and u2 |B2 .
Let cj = (cxj , cyj ) denote the center of Bj . The local coordinates (rj , j ), relative
to cj , of a point (r, ) D given in (global) polar coordinates are

rj = (r cos cxj )2 + (r sin cyj )2 ,

1
1
cos j = (r cos cxj ), sin j = (r sin cyj ).
rj
rj

(86)
(87)

By combining the contributions from the various purely outgoing wave elds
uj |Bj , j = 1, . . . , J, we can then derive an explicit formula for the far-eld
pattern of u, given by (88) below. We summarize this result as a theorem.
Theorem 3 The far-eld pattern f dened in (85) of the solution u to the
free space problem (1)(3) is entirely determined by the values of the purely
outgoing wave elds uj , j = 1, . . . , J, on the components Bj of the articial
boundary B, which appear in the DtN condition (56), (57). It is given by



y
1i 
x
(i)n
f () =
eik(cj cos +cj sin )
(1)
j=1
n=0 Hn (kRj )

16

2
0

uj (Rj ,  ) cos n(  ) d .
(88)

PROOF. We examine the asymptotic behavior of the Fourier representation

(10) of each purely outgoing wave eld uj , j = 1, . . . , J, for r . By Taylor
expansion of (86), (87) we observe that
rj = r (cxj cos + cyj sin ) + O(r 1 ),
1

cos j = cos + O(r ),

sin j = sin + O(r 1),

(89)

r
r

(90)
(91)

Because the angle j [0, 2) is uniquely determined by the pair

(cos j , sin j ) (cos , sin ),

r ,

(92)

we conclude that j , as r , and therefore that

cos n(j  ) cos n(  ),

r ,  [0, 2).

(93)

The asymptotic behavior of the Hankel functions [26] is given by



Hn(1) (krj )

2 i(krj 1 n 1 )
eikrj 1 i
2
4
(i)n ,
e
=

krj
krj

From (89) we conclude

y
x
krj kr and eikrj eikr eik(cj cos +cj sin ) ,

r .

r .

(94)

(95)

Each purely outgoing wave eld uj , given by (10), therefore has the asymptotic
behavior
eikr 1 i ik(cxj cos +cyj sin )
e
uj (rj , j )

kr
(i)n

n=0

Hn (kRj )

Since u =

2




(1)

J

j=1 uj ,

uj (Rj ,  ) cos n(  ) d ,

r .

(96)

the result follows by summing over j. 2

Numerical examples

We shall now combine the multiple-DtN (66), (67) and -MDtN (68), (69)
condition with a nite dierence scheme. We shall also compare the scattered
elds obtained either with the double-DtN approach or with the single-DtN
approach in a very large computational domain and demonstrate their high
accuracy and convergence properties via numerical examples.
17

We consider the following two scatterer model problem with two obstacles,
where the computational domain = 1 2 , the obstacle boundary =
1 2 , and the articial boundary B = B1 B2 :
u + k 2 u = f in ,
u = g on ,
n u = T uout on B,
P uout = u on B.

(97)
(98)
(99)
(100)

To precisely describe the typical structure of the resulting discrete linear system, we consider a polar equidistant grid along B1 and B2 . Inside 1 and
2 , we discretize the solution with step size hr in the r-direction and h in
the -direction. Then we use second order centered nite dierences in r- and
(1)
(2)
-direction to discretize (97). The vectors uN and uN denote the values of
the numerical solution on the articial boundary. The discretization of (97)
(1)
(2)
involves the values uN +1 and uN +1 at ghost points, which lie outside the
computational domain . These unknown values are eliminated by using a
second order nite dierence discretization of (99), (100). Next, we let the
vectors u1 and u2 denote the values of the purely outgoing wave elds on their
respective boundary components. Then the discretization of the multiple-DtN
condition (99), (100) is given by

2
I
h2r

(1)

(1)
P (2)

I
(2)

I P (1)

(2)

0 M (1) T (2)
uN 1

(1)
(2)

0 Q(2) T (1) M (2)

u

0 Q(1)
2
I
h2r

(1)

u
N 1

(2)
uN

u1

0


0
,

0

(101)

u2
with identity matrices I, all-zero matrices 0 and tridiagonal matrices Q. The
matrices M , T and P are full matrices obtained by discretizing the integral
operators with the second order trapezoidal quadrature rule.
A typical sparsity pattern of the entire nite dierence matrix, including the
discretization of (97) in the interior is shown in Fig. 2, for the special case of
two circular obstacles with an equidistant polar mesh throughout 1 and 2 .
Here the ordering of the interior and boundary nodes is chosen by starting from
the innermost layers in both domains and moving outward with increasing
index. The 6 small full blocks in the lower right corner correspond to the
full block-matrices in (101). The numerical solution of the system of linear
equations corresponding to a single-scattering DtN problem has been treated
in [27].
18

Figure 2. The sparsity pattern of the nite dierence matrix for a two scatterer
problem. There are 21 layers of 240 grid points in each domain 1 and 2 . Hence
the total number of unknowns is 2 (21 240) for u plus 2 240 for u1 |B1 and
u2 |B2 .

6.1 Accuracy and convergence study

To demonstrate the accuracy and convergence of our method, we consider the

following test problem: We let an incident plane wave impinge on a circular
disk shaped obstacle centered at (0, d), with radius 0.5 and distance d = 1.5
from the origin see Fig. 1 for an illustration. The obstacle is located inside 1
and is bounded by 1 . In 2 , no physical obstacle is present. The sound-soft
boundary condition requires that the total eld be zero on 1 , while the JacobiAnger expansion (see for example [19], p. 67) yields the exact solution for the
scattered eld everywhere outside 1 . Then we prescribe its values on the
boundary of a second virtual obstacle, centered at (0, d) with radius 0.75, and
compute the numerical solution in the two (disjoint) computational domains
19

Table 1
The maximal relative errors for plane wave scattering from a single obstacle, with the
values of the exact solution prescribed on the boundary of the second obstacle.
Incidence angle , wave number k = 2, DtN expansion truncated at N = 50,
comparison with exact solution. Grids with Nr N cells in r- and -direction,
respectively.
relative error in the solution

5 60

10 120

20 240

40 480

7.53 102

1.77 102

4.38 103

1.09 103

/4

7.85 102

1.84 102

4.54 103

1.13 103

/2

9.77 102

2.24 102

5.49 103

1.37 103

relative error in the far-eld pattern

5 60

10 120

20 240

40 480

4.68 102

1.11 102

2.76 103

6.87 104

/4

6.05 102

1.45 102

3.60 103

8.97 104

/2

7.69 102

1.85 102

4.57 103

1.14 103

1 , 2 , bounded by circles B1 and B2 with radii R1 = 1 and R2 = 1.25,

respectively. We then compare the numerical result with the exact solution
for single scattering. We choose k = 2 for the wave number and truncate
the DtN expansion at N = 50. We also compute the exact far-eld pattern
and compare it with that given by our numerical result. The maximal relative
errors for dierent grids and incidence angles are shown in Table 1. We observe
second order convergence of our method in every case, as expected, as the mesh
size h 0.

To study the eect of the truncation parameter N on the error we choose

= /4 for the incidence angle and compute the solution with varying N,
either with the DtN and MDtN condition imposed at B. The relative error is
shown in Fig. 3. We observe that the modied DtN condition leads to better
accuracy, even for small truncation indices N. When N max{kR1 , kR2 },
the two solutions computed with DtN and MDtN essentially coincide for this
model problem. This behavior of the DtN and MDtN conditions illustrated
in Fig. 3 is typical, and has been reported previously for single scattering
problems [22,8].
20

10

DtN
MDtN

max. rel. error

10

10

10

10

5
N

10

Figure 3. The maximal relative error in the solution vs. the truncation index N ,
for k = 2 and incidence angle = /4, on the 20 240 grid. Comparison of DtN
(squares) and MDtN (circles).

6.2 Comparison with the single-DtN FE approach

Here we consider the scattering of a plane wave with incidence angle = 3/8
on two obstacles with sound-soft elliptic boundaries. The numerical solution
obtained by using our nite dierence scheme with the multiple-DtN condition
on the articial boundaries is compared with a numerical solution obtained
by using a nite element scheme in a larger domain, which contains both
obstacles, with the single-DtN condition imposed at the articial boundary
r = 3. The wave number is k = 2 and the resolutions are comparable, with
about 45 grid points per wavelength. Here the modied DtN map is used and
the truncation index is set to N = 50. The contour lines of the real part of
the total eld are shown for both solutions in Fig. 4. Note that the size of
the computational sub-domains in the multiple-DtN case is independent of
the relative distance between them, leading to a much smaller computational
domain, in comparison with the single-DtN case.

21

finite differences, multiDtN

finite elements, singleDtN

3
3

3
3

Figure 4. Scattering from two ellipses, k = 2, = 3/8. Contour lines of the

real parts of the total wave elds for two solutions are shown. Left: the numerical
solution obtained by a second-order nite dierence method combined with the
multiple-DtN condition; Right: the numerical solution obtained by a (piecewise
linear) nite element method combined with the single-DtN condition.

In Fig. 5, the values of the two solutions on the articial boundary at r = 3,

which was used for the nite element solution, are shown. The multiple-DtN
solution is evaluated on that boundary by using the Fourier representation
(84) for the purely outgoing wave elds.
real part, values at r=3.0

imaginary part

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0.2

0.2

0.4

0.4

0.6

0.6

0.8

0.8

1
0

FE, singleDtN
FD, multiDtN
0

Figure 5. Comparison of multiple-DtN with single-DtN. Values of the scattered eld

on the articial boundary r = 3 used for the nite element solution shown in Fig. 4.

22

scattering cross section

20
finite elements, singleDtN
finite differences, multiDtN
15

10

10

Figure 6. Comparison of multiple-DtN with single-DtN. Values of the bistatic scattering cross section for both solutions.

The bistatic scattering cross section for plane wave scattering from two ellipses,
given by
() := 20 log10 |f ()| where f is the far-eld pattern (88) of the
solution, is displayed in Fig. 6 for the single-DtN and multiple-DtN solutions.
The two solutions coincide.
6.3 An example with ve obstacles
An important advantage of our multiple-DtN approach is that no further analytical derivation is needed to extend it to higher numbers of scatterers, once
the DtN condition is known for two domains. Here we consider the scattering
of a plane wave with incidence angle = /8 impinging on ve cylindrical
obstacles of dierent sizes with sound-soft boundaries. The wave number is
set to k = 8 and the grid consists of about 20 points per wavelength. We use
the modied DtN map and truncate the innite series at N = 50. The real
part of the total eld and the scattering cross section are shown in Figs. 7 &
8.

23

total field, real part

3

3
3

Figure 7. The total eld for plane wave scattering from ve cylinders, k = 8,
incidence angle = /8.

24

scattering cross section

90

35

120

60
25
15

150

30
5
5
15

180

330

210

300

240
270

Figure 8. The bistatic scattering cross section for the ve cylinders, k = 8, incidence
angle = /8.

Conclusion

We have derived a Dirichlet-to-Neumann (DtN) map for multiple scattering

problems, which is based on a decomposition of the scattered eld into several purely outgoing wave elds. We have proved that the corresponding DtN
boundary condition is exact. When the multiple-DtN boundary condition is
used to solve multiple scattering problems, the size of the computational domain is much smaller, in comparison to the use of one single large articial
boundary. In particular, the size of the computational sub-domains in the
multiple-DtN case does not depend on the relative distances between the components of the scatterer. Although the articial boundaries must be of simple
geometric shape (circles, ellipses, spheres, etc.), the DtN condition is not tied
25

to any coordinate system inside the computational domain; in particular, it

remains exact independently of the discretization used inside .
We have presented a variational formulation of a multiple scattering problem
with this boundary condition and also derived a formula for the far-eld of
the solution, which is obtained by exploiting auxiliary values used in the formulation. Accuracy and convergence have been demonstrated on a simple test
problem, and a comparison with single-DtN has been made in the situation
of two elliptical obstacles.
This approach is based on the decomposition of the scattered eld into several
purely outgoing wave elds. It can also be used to derive exact nonreecting
boundary conditions for multiple scattering problems for other geometries and
equations, both in two and in three space dimensions.
For large-scale applications in multiple scattering, it may be useful, or even
necessary, to solve the sequence of sub-problems in 1 , 2 , etc. iteratively,
while exchanging boundary values between the disjoint exterior boundary
components via the operators M, P , and T . Parallelism can be increased
even further by using standard domain decomposition techniques [28,29] separately within each sub-domain j . Although the convergence of such a Jacobi
or Gauss-Seidel like iterative procedure remains an open question, it could certainly be used as an ecient preconditioner.
In this work we have only treated the time-harmonic case. In the time-dependent
case, a similar approach can be used to derive exact nonreecting boundary
conditions for multiple scattering problems, by using a representation formula
derived in [30]. The authors are currently investigating the time-dependent
case and will report on their results elsewhere in the near future.

Acknowledgements

The authors would like to thank Joseph B. Keller for useful comments and
suggestions. They also wish to thank Patrick Meury for his nite element code
with a single-DtN boundary condition.

References

[1] B. Engquist, A. Majda: Absorbing Boundary Conditions for the Numerical

Simulation of Waves. Math. Comp. 31 (139), pp. 629651, 1977

26

[2] A. Bayliss, M. Gunzburger, E. Turkel: Boundary conditions for the numerical

solution of elliptic equations in exterior regions. SIAM J. Appl. Math. 42 (2),
pp. 430451, 1982
[3] J. P. Berenger: A Perfectly Matched Layer for the Absorption of Electromagnetic
Waves. J. Comp. Phys. 114 (2), pp. 185200, 1994
[4] E. Turkel, A. Yefet: Absorbing PML boundary layers for wave-like equations.
App. Num. Math. 27 (4), pp. 533557, 1998
[5] I. Harari, M. Slavutin, E. Turkel: Analytical and numerical studies of a finite
element PML for the Helmholtz equation. J. Computational Acoust. 8 (1), pp.
121-137, 2000
[6] J. B. Keller, D. Givoli: Exact non-reflecting boundary conditions.
J. Comp. Phys. 82 (1), pp. 172192, 1989
[7] D. Givoli: Numerical Methods for Problems in Infinite Domains (Elsevier,
Amsterdam 1992)
[8] M. J. Grote, J. B. Keller: On Nonreflecting Boundary Conditions.
J. Comp. Phys. 122 (2), pp. 231243, 1995
[9] D. Givoli: Recent Advances in the DtN FE Method.
Archives of Comput. Meth. Engin. 6 (2), pp. 71116, 1999
[10] G. K. G
achter, M. J. Grote: Dirichlet-to-Neumann Map for Three-Dimensional
Elastic Waves. Wave Motion 37 (3), pp. 293311, 2003
[11] P. A. Martin: Multiple Scattering: an Invitation. In Proc. Third Int. Conf. on
Math. and Numerical Aspects of Wave Propagation, Eds. G. Cohen et al.,
SIAM, Philadelphia, pp. 316, 1995
[12] P. A. Martin: Integral-equation methods for multiple-scattering problems. I.
Acoustics. Quart. J. Mech. Appl. Math. 38, pp. 105118, 1985
[13] P. A. Martin, F. J. Rizzo: Partitioning, boundary integral equations, and exact
Greens functions. Int. J. Numer. Meth. Engin. 38, pp. 34833495, 1995

[14] F. Z
aviska: Uber
die Beugung elektromagnetischer Wellen an parallelen,
unendlich langen Kreiszylindern. Annalen der Physik, Folge 4 (40), pp. 1023
1056, 1913
[15] B. Peterson, S. Str
om: Matrix formulation of acoustic scattering from an
arbitrary number of scatterers. J. Acoust. Soc. Am. 56 (3), pp. 771780, 1974
[16] V. Twersky: On multiple scattering of waves. J. Res. Nat. Bur. Stand. 64D,
pp. 715730, 1960
[17] J. E. Burke, V. Twersky: On scattering of waves by many bodies. J. Res. Nat.
Bur. Stand. 68D, pp. 500510, 1964
[18] F. M. Arscott, A. Darai: Curvilinear coordinate systems in which the Helmholtz
equation separates. IMA J. Appl. Math. 27, pp. 3370, 1981

27

[19] D. Colton, R. Kress: Inverse Acoustic and Electromagnetic Scattering Theory

(Springer, 1992)
[20] D. Colton, R. Kress: Integral Equation Methods in Scattering Theory (WileyInterscience, 1983)
[21] S. Tordeux, INRIA, private communication, July 2003
[22] I. Harari, T. J. R. Hughes: Analysis of continuous formulations underlying the
computation of time-harmonic acoustics in exterior domains.
Comp. Meth. Appl. Mech. Engin. 97 (1), pp. 103124, 1992
[23] J.-C. Nedelec: Acoustic and electromagnetic equations: integral representations
for harmonic problems (Springer, 2001)
[24] I. Harari, I. Patlashenko, D. Givoli: Dirichlet-to-Neumann maps for unbounded
wave guides. J. Comp. Phys. 143 (1), pp. 200-223, 1998
[25] F. Ihlenburg: Finite element analysis of acoustic scattering (Springer, 1998)
[26] M. Abramowitz, I. A. Stegun: Handbook of mathematical functions (Dover
Publications, 1970)
[27] A. A. Oberai, M. Malhotra, P. M. Pinsky: On the implementation of the
Dirichlet-to-Neumann radiation condition for iterative solution of the Helmholtz
equation. App. Num. Math. 27 (4), pp. 443464, 1998
[28] J.-D. Benamou, B. Despres: A Domain Decomposition Method for the Helmholtz
Equation and Related Optimal Control Problems. J. Comp. Phys. 136 (1), pp.
6882, 1997
[29] M. Balabane: A domain decomposition algorithm for the Helmholtz equation. 1
The dissipating case. J. Asympt. Analysis, in press
[30] M. J. Grote, C. Kirsch: Far-field evaluation via nonreflecting boundary
conditions. In Proc. Ninth Int. Conf. on Hyperbolic Problems (Hyp2002), Eds.
Th. Y. Hou, E. Tadmor, Springer, pp. 195204, 2003

28

LATEST PREPRINTS
No.

Author: Title

2004 4

Marcus J. Grote, Christoph Kirsch

Dirichlet-to-Neumann Boundary Conditions for Multiple Scattering
Problems

2004 3

Urs Hasler, Anna Schneebeli, Dominik Schtzau

Mixed Finite Element Approximation of Incompressible MHD Problems
Based on Weighted Regularization

2004 2

Karin Baur, Jan Draisma

Higher Secant Varieties of the Minimal Adjoint Orbit

2004 1

Stephan Mohrdieck
A Steinberg Cross-Section for Non-Connected Affine Kac-Moody Groups

2003 18 Assyr Abdulle, Sabine Attinger:

Separating DNA Molecules in Asymmetric Periodic Obstacle Arrays
2003 17 Paul Houston, Ilaria Perugia, Anna Schneebeli, Dominik Schtzau:
Interior Penalty Method for the Indefinite Time-harmonic Maxwell
Equations
2003 16 Marcus J. Grote:
Nonreflecting Boundary Conditions for Time Dependent Waves
2003 15 Simon Stingelin:
New Numerical Solutions for the Brezis-Nirenberg Problem on S
2003 14 Jir Hork:
Constrained Mountain Pass Algorithm for the Numerical Solution of
Semilinear Elliptic Problems
2003 13 Paul Houston, Dominik Schtzau, Thomas P. Wihler:
Energy Norm A Posteriori Error Estimation for Mixed Discontinuous
Galerkin Approximations of the Stokes Problem
2003 12 Dominik Schtzau, Christoph Schwab, Andrea Toselli:
Stabilized hp-DGFEM for Incompressible Flow
2003 11 B. M. Brown, Wolfgang Reichel:
n
Eigenvalues of the Radially Symmetric p-Laplacian in R
_______________________________________________________________________
Preprints are available under http://www.math.unibas.ch/preprints

2003 10 Bernardo Cockburn, Guido Kanschat, Dominik Schtzau:

The Local Discontinuous Galerkin Method for Linear Incompressible Fluid
Flow: A Review
2003 9

Jacqueline Fleckinger, Wolfgang Reichel:

Global Solution Branches for p-Laplacian Boundary Value Problems

2003 8

Wolfgang Reichel:
Sharp Parameter Ranges in the Uniform Anti-Maximum Principle for
Second-Order Ordinary Differential Operators

2003 7

Wolfgang Reichel:
Uniqueness Results for Semilinear Polyharmonic Boundary Value Problems
on Conformally Contractible Domains

2003 6

Wolfgang Bangerth, Marcus J. Grote, Christel Hohenegger:

Finite Element Method for Time-Dependent Scattering: Nonreflecting
Boundary Condition, Adaptivity, and Energy Decay

2003 5

Catherine Bandle:
Asymptotic Behaviour of Large Solutions of Quasilinear Elliptic Problems

2003 4

Stephan Mohrdieck, Robert Wendt:

Twisted Conjugacy Classes, Coadjoint Orbits of Loop Groups, and
D-Branes in the WZW Model

2003 3

Dominik Schtzau:
Mixed Finite Element Methods for Stationary Incompressible MagnetoHydrodynamics

2003 2

P. Joseph McKenna, Wolfgang Reichel:

Radial Solutions of Singular Nonlinear Biharmonic Equations and
Applications to Conformal Geometry

2003 1

Catherine Bandle, Stella Vernier-Piro:

Estimates for Solutions of Quasilinear Problems with Dead Cores

Preprints are available under http://www.math.unibas.ch/preprints