IMA Journal of Applied Mathematics (2009) 115

doi:10.1093/imamat/hxp002
A new study of the Burton and Miller method for the solution of a 3D
Helmholtz problem
KE CHEN
Department of Mathematical Sciences, University of Liverpool, Peach Street,
Liverpool L69 7ZL, UK
JIN CHENG
Department of Mathematics, Fudan University, Shanghai 200433, China
AND
PAUL J. HARRIS
School of Computing, Mathematical and Information Sciences, University of Brighton,
Lewes Road, Brighton, East Sussex BN2 4GJ, UK
[Received on 24 June 2005; accepted on 16 December 2008]
The exterior Helmholtz problem can be effciently solved by reformulating the differential equation as
an integral equation over the surface of the radiating and/or scattering object. One popular approach for
overcoming either non-unique or non-existent problems which occur at certain values of the wave number
is the so-called Burton and Miller method which modifes the usual integral equation into one which can
be shown to have a unique solution for all real and positive wave numbers. This formulation contains
an integral operator with a hypersingular kernel function and for many years, a commonly used method
for overcoming this hypersingularity problem has been the collocation method with piecewise-constant
polynomials. Viable high-order methods only exist for the more expensive Galerkin method. This paper
proposes a new reformulation of the BurtonMiller approach and enables the more practical collocation
method to be applied with any high-order piecewise polynomials. This work is expected to lead to much
progress in subsequent development of fast solvers. Numerical experiments on 3D domains are included
to support the proposed high-order collocation method.
Keywords: exterior Helmholtz; boundary integral equation; BurtonMiller; Green theorem; hypersingular
operators; collocation method.
1. Introduction
This paper is concerned with using the boundary integral method (BIM) to solve the exterior Helmholtz
problem. Such problems arise in the mathematical modelling of the steady-state single-frequency acous-
tic feld outside an object immersed in a fuid; see Amini et al. (1992) and Colton & Kress (1983).
It is advantageous to reformulate the underlying partial differential equation (PDE) as an integral
equation because the domain problem is reduced to one on its fnite boundary. However, it is well
known that for certain frequencies, the solution to integral equation formulations either does not exist

Email: [Link]@[Link]

Email: jcheng@[Link]

Corresponding author. Email: [Link]@[Link]

c The Author 2009. Published by Oxford University Press on behalf of the Institute of Mathematics and its Applications. All rights reserved.
IMA Journal of Applied Mathematics Advance Access published February 24, 2009
2 of 15 K. CHEN ET AL.
or is non-unique (when the original PDE has a unique solution). Over the years, a number of methods
have been proposed for overcoming these problems, as surveyed in Amini et al. (1992). This paper will
study the particular unique formulation of the problem devised by Burton & Miller (1971) which we
shall refer to as the Burton and Miller method. Although the Burton and Miller method is widely used,
due to the presence of a hypersingular operator, the usage is so far limited to the piecewise-constant
collocation framework, where it is relatively easy to overcome the hypersingularity (Amini et al., 1992;
Chen & Harris, 2001). For the theoretically appealing Galerkin method, methods exist for transforming
the hypersingular integral into one which is at worst weakly singular (see Giroire & Nedelec, 1978;
Hackbusch, 1995; Harris & Chen, 2003).
This paper proposes a new and high-order collocation method that overcomes this hypersingular
nature of operator without using fnite-part integration (refer the latter to Harris, 1992; Salvadori, 2001;
Schwab & Wendlend, 1998; Aimi & Diligenti, 2002). In what follows, we shall frst introduce the exte-
rior Helmholtz problemand the direct method for reformulating it as a simple boundary integral equation
which suffers from the non-uniqueness problems outlined above. We then discuss an alternative BIM
which possess a unique solution for all wave numbers, but which includes the hypersingular integral
operator. The main part of the paper is dedicated to describing a method for evaluating hypersingular
operator with polynomial basis functions for any order. Our new idea is to remove the hypersingularity
by using domain integrals and then to avoid domain integrals by using singularity subtraction. Finally,
we present some numerical results for some typical test problems to show that the method proposed
here can be used to obtain a very accurate solution.
We remark in passing that the methods presented in this paper are mainly for solving wave problems
with Neumann boundary conditions on the surface of the structure (which is quite common in practice).
However, for Dirichlet boundary conditions, while this work still applies, it is possible to avoid the
hypersingular operator completely by using an indirect layer potential integral equation formulation
(Amini et al., 1992) if the values of the normal derivative on the surface are not specifcally required.
We also note that the methods presented here are intended for use in solving practical problems, such
as the acoustic radiation from a sonar transducer or a loudspeaker, where it is a physical requirement
that the acoustic pressure is continuous throughout the fuid-flled domain. This allows us to compute
the required domain gradient of the pressure at all points in the fuid domain and on its boundaries.
2. The boundary integral equation formulation of the exterior Helmholtz problem
We frst introduce the basic problem formulation and discuss the associated and known problems. Con-
sider the problem of solving the Helmholtz equation

2
(p) +k
2
(p) = 0, p
+
S, (2.1)
in some unbounded 3D region
+
exterior to a closed surface S, where k > 0 is the wave number,
subject to a Neumann boundary condition on S and the Sommerfeld radiation condition
lim
r
r
_

r
ik
_
= 0. (2.2)
Since the above exterior problem cannot be easily solved by discretizing the infnite domain using fnite
differences or fnite elements, we choose to reformulate the problem as an integral equation over the
surface S; see Amini et al. (1992), Colton & Kress (1983) and Chen & Harris (2001). However, domain
methods based on infnite elements or coupled fnite and infnite elements have been devised; see Gerdes
& Demkowicz (1996) and Gerdes (1998), e.g.
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 3 of 15
There are essentially two methods for reformulating the exterior Helmholtz problem as an boundary
integral equation, the direct method which is used here and the indirect method. The direct method
of reformulation as an integral equation is obtained by applying Greens theorem to get a relationship
between the acoustic pressure and its normal derivative in the form (Burton, 1973; Hackbusch, 1995)
_
S
(q)
G
k
(p, q)
n
q
G
k
(p, q)
(q)
n
q
dS
q
=
_
1
2
(p), p S,
(p), p
+
,
(2.3)
where
G
k
(p, q) =
e
ik|pq|
4|p q|
(2.4)
is the free-space Greens function, or the fundamental solution, for Helmholtz equation and n
q
is the
unit outward normal to S at q. If the normal derivative of the acoustic feld is given on the surface S,
then (2.3) for p S gives a Fredholm integral equation of the second kind which can be solved for
the surface pressure . The acoustic pressure can then be computed at any point in
+
using (2.3).
However, the solution of (2.3) does not posses a unique solution for certain discrete values of the wave
number (called characteristic or irregular or resonance wave numbers), although it can be shown that the
underlying differential equation has a unique solution for all real and positive wave numbers (Burton,
1973). A number of different methods have been proposed for overcoming this non-uniqueness problem
and a survey of these is given in Amini et al. (1992). One of the most robust is the so-called Burton and
Miller method and for this reason, it is the method we shall use here. It is worth noting at this point that
had we used an indirect integral equation formulation, then similar problems with the existence of the
solution would have occurred at the characteristic wave numbers.
The Burton & Miller (1971) method proposed using the alternative integral equation formulation

1
2
(p) +
_
S
(q)
_
G
k
(p, q)
n
q
+

2
G
k
(p, q)
n
p
n
q
_
dS
q
=

2
(p)
n
p
+
_
S
(q)
n
q
_
G
k
(p, q) +
G
k
(p, q)
n
p
_
dS
q
, (2.5)
where is a non-zero constant. It can be shown that provided that the imaginary part of is non-
zero, then (2.5) has a unique solution for all real and positive k (Burton & Miller, 1971). However, this
formulation has introduced an integral operator with the kernel function

2
G
k
(p,q)
n
p
n
q
which has a 1/|pq|
3
singularity. For brevity, we note that (2.5) can be written as

1
2
(p) +
_
S
(q)
_
G
k
(p, q)
n
p
+

2
G
k
(p, q)
n
p
n
q
_
dS
q
= g(p), p S, (2.6)
where g(p) denotes the right-hand side associated with the Neumanns boundary condition.
Two ways which can be used to overcome the problems associated with the hypersingularity are to
use a simple piecewise-constant collocation method (which is widely used in practise) or a higher-order
Galerkin method. The collocation method has the advantages of being relatively simple to implement
and cheap to use in computational terms (see Chen & Harris, 2001, e.g.). However, for any given bound-
ary element mesh, it is not very accurate if only a low-order method, such as piecewise constant, is
4 of 15 K. CHEN ET AL.
used. The Galerkin method has the advantage that it is much more accurate than a low-order collocation
method since high-order methods are available, but it is a lot more expensive in computational terms
(and less used in engineering practice). An effcient high-order Galerkin method is discussed in Harris
& Chen (2003). In Section 3, we present a method which can be used to reformulate the integral in-
volving the second derivative of the Greens function as integrals which are at worst weakly singular for
any order of basis functions. This reformulation will allow the use of the simpler collocation method
with high-order basis functions, giving a level of accuracy which is comparable to the more expensive
Galerkin method.
3. The weakly singular formulation
We now introduce our reformulation of the above (2.5) with a view to remove the hypersingular term.
Without loss of generality, we shall assume that the unknown function lies in the space of H
1/2
(S),
which is a weak assumption as we aim for high-order methods in Section 4. In order to obtain a weakly
singular formulation, we need to apply a transformation to the hypersingular term
_
S
(p)

2
G
k
(p, q)
n
p
n
q
dS
q
. (3.1)
In order to do, this we make use of the following lemma.
LEMMA 1 Let vector a R
3
and S be a piecewise C
2
-closed surface enclosing some interior region
in R
3
. Assume that p is a point such that n
p
is unique and well defned. Then,
_
S
a (q p)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
a n
q
G
k
(p, q)
n
p
dS
q
k
2
a
_

(q p)
G
k
(p, q)
n
p
dV
q

a n
p
2
. (3.2)
(The proof is shown in Appendix A.)
We now consider how to utilize Lemma 1 to express (3.1) in terms of weakly singular integrals (and
involving domain integrals). First, we rewrite the hypersingular integral as
_
S
(q)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
[(q) (p) (p)(q p)]

2
G
k
(p, q)
n
p
n
q
dS
q
+(p)
_
S

2
G
k
(p, q)
n
p
n
q
dS
q
+
_
S
(p) (q p)

2
G
k
(p, q)
n
p
n
q
dS
q
, (3.3)
where (p) denotes the domain gradient of at the point p. The frst term on the right-hand side of
(3.3) contains what is essentially the difference between (q) and the terms up to and including the frst
derivative terms of its Taylors series. Hence, the remainder term in (3.3) is of order |q p|
2
and so the
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 5 of 15
whole frst integral behaves as |qp|
1
which means it is weakly singular. Clearly, for this to work, we
require that is at least two times differentiable at the point p, but this is not a major restriction as in
most physical problems, the acoustic pressure will satisfy this condition. In addition, it is a requirement
of Lemma 1 that the point p be such that S has a well-defned normal, and a consequence of this is
that will be well defned at such points. Hence, the frst term on the right-hand side of (3.3) can be
evaluated using an appropriate quadrature rule.
The second term can be evaluated using (Meyer et al., 1978)
_
S

2
G
k
(p, q)
n
p
n
q
dS
q
= k
2
_
S
n
p
n
q
G
k
(p, q)dS
q
(3.4)
to yield that
_
S
(q)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
[(q) (p) (p) (q p)]

2
G
k
(p, q)
n
p
n
q
dS
q
+ k
2
(p)
_
S
n
p
n
q
G
k
(p, q)dS
q
+
_
S
(p) (q p)

2
G
k
(p, q)
n
p
n
q
dS
q
, (3.5)
where the frst two integrals on the right-hand side are now weakly singular. It now remains to address
the singularity in the fnal integral on the right-hand side.
In order to overcome this singularity, set a = (p) (remembering that p S must be a point at
which the normal is well defned) and use Lemma 1 to obtain that
_
S
(p) (q p)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
(p) n
q
G
k
(p, q)
n
p
dS
q
k
2
_

(p) (q p)
G
k
(p, q)
n
p
dV
q

1
2
(p) n
p
. (3.6)
Substituting (3.6) into (3.5) yields that
_
S
(q)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
[(q) (p) (p) (q p)]

2
G
k
(p, q)
n
p
n
q
dS
q
+k
2
(p)
_
S
n
p
n
q
G
k
(p, q)dS
q
+
_
S
(p) n
q
G
k
(p, q)
n
p
dS
q
6 of 15 K. CHEN ET AL.
k
2
_

(p) (q p)
G
k
(p, q)
n
p
dV
q

1
2
(p) n
p
, (3.7)
where every integral on the right-hand side is now weakly singular.
Theoretically, we have thus completed the task of reformulating (3.1). However, this new formula-
tion has introduced a volume integral over the domain interior to S, although is only required on
S and there is no hypersingularity. This integral can be expensive and diffcult to evaluate, so we apply
further reformulation.
Clearly, this problem can be avoided by setting k = 0 as this will cause the volume integral to
disappear along with the second integral on the right-hand side of (3.7). In order to make use of this, we
shall use the singularity subtraction technique by rewriting the integral involving the second derivative
of the Greens function as
_
S
(q)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
(q)
_

2
G
k
(p, q)
n
p
n
q

2
G
0
(p, q)
n
p
n
q
_
dS
q
+
_
S
(q)

2
G
0
(p, q)
n
p
n
q
dS
q
. (3.8)
By using (3.7) with k = 0 to evaluate the fnal integral on the right-hand side of (3.8), we obtain that
_
S
(q)

2
G
k
(p, q)
n
p
n
q
dS
q
=
_
S
(q)
_

2
G
k
(p, q)
n
p
n
q

2
G
0
(p, q)
n
p
n
q
_
dS
q
+
_
S
_
{(q) (p) (p) (q p)}

2
G
0
(p, q)
n
p
n
q
_
dS
q
+
_
S
(p) n
q
G
0
(p, q)
n
p
dS
q

1
2
(p) n
p
, (3.9)
where every integral on the right-hand side is at worst weakly singular. Hence, we have reformulated
(3.1) into a form which only involves boundary or surface integrals which are at worst weakly singular.
Substituting (3.9) into (2.6) yields our reformulated BurtonMiller method for (2.1):

1
2
(p) +
_
S
(q)
G
k
(p, q)
n
p
dS
q
+
_
_
S
(q)
_

2
G
k
(p, q)
n
p
n
q

2
G
0
(p, q)
n
p
n
q
_
dS
q
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 7 of 15
+
_
S
_
{(q) (p) (p) (q p)}

2
G
0
(p, q)
n
p
n
q
_
dS
q
+
_
S
(p) n
q
G
0
(p, q)
n
p
dS
q

1
2
(p) n
p
_
= g(p), p S, (3.10)
which has not only an unique solution (like the well-known BurtonMiller formulation) but also no
strong singularises (unlike the well-known BurtonMiller formulation). Hence, subsequent solution of
(3.10) by any standard method can be considered (note that, in contrast, most standard methods cannot
be applied to solve the well-known BurtonMiller formulation).
Finally, we consider the technical point of computing the gradient of the surface function without
involving interior values. Let the surface be parameterized locally in terms of the two variables u and v
so that we can make use of the chain rule to give
_

x
x
u
+

y
y
u
+

z
z
u
=

u
,

x
x
v
+

y
y
v
+

z
z
v
=

v
,

x
n
x
+

y
n
y
+

z
n
z
=

n
.
(3.11)
Thus, at any surface point (x, y, z), corresponding to local coordinates (u, v) in the reference space, a
translation of reference space gradients to physical space gradients can be made.
4. Application to the collocation method
In general, (3.10) cannot be solved analytically. One effective numerical method is the collocation
method. It solves (3.10) by approximating the solution by a function of the form
(q) =
m

j =1

j
(q) (4.1)
in the subspace of piecewise polynomials, where {
1
,
2
, . . . ,
m
} are a set of linearly independent
basis functions which are assumed to be polynomials of any order, and then picking m collocation
points p
1
, p
2
, . . . , p
m
at which the residual due to the use of the approximated solution is forced to
be zero. Further, as in many similar applications, we assume that the surface S is approximated by
M boundary elements which are interpolated using a parametric mapping to a reference element in the
(u, v)-plane. The parametric mapping is used to calculate surface-related quantities such as the Jacobian
and unit normal vector at points which are interior to the elements.
However, the use of the collocation method means that we are also required to work out the normal
at the collocation points. This presents a problem if the usual continuous approximation based on the
8 of 15 K. CHEN ET AL.
element vertices and mid-side points is used as the unit normal is usually not well defned at such points
due to the jump in the derivatives of the interpolating functions at the element boundaries. Further,
our new method for evaluating the hypersingular operator requires that the collocation point lies on a
smooth part of S which has a well-defned normal vector. For smooth surfaces S, it is possible to use the
unit normal to the underlying exact surface (assuming that it is simple to calculate) and ignore the fact
that the node lies on a vertex of the approximate surface. However, for surfaces which are not smooth
(i.e. they contain edges and/or vertices), it is not possible to compute a unit normal to the surface at
any node which lies on an edge or vertex. To overcome this problem, we propose using discontinuous
interpolation where the collocation points are chosen to lie inside any given element, corresponding
to prescribed points (u, v) in the reference element. As these points will lie on a smooth part of the
approximate surface, there are no problems with computing the unit normal at these points. The exact
choice of the interior points will be discussed in the Section 5 on numerical results. Here, we will assume
that there are L basis functions and collocation points associated with each element. The exact choice
of basis functions and collocation points used in this work will be discussed later.
A further problem encountered when attempting to discretize (3.10) using (3.11) to fnd the gradient
of the solution is the complexity of keeping track of the various contributions to the system matrix and
right-hand side vector. For example, if there are L basis functions associated with the element, then
each integral involving (p) appearing in (3.10) will give a contribution to L
2
entries in the system
coeffcient matrix and make L contributions to the right-hand side vector.
A simpler alternative is to substitute (4.1) into (2.6) to give
m

j =1

j
_

1
2

j
(p
i
) +
_
S
_

j
(q)
G
k
(p
i
, q)
n
p
+

2
G
k
(p
i
, q)
n
p
n
q
_
dS
q
_
= g(p
i
). (4.2)
We now use (3.7) with replaced by
j
to evaluate the hypersingular integral to give
m

j =1

j
_

1
2

j
(p
i
) +
_
S

j
(q)
G
k
(p
i
, q)
n
p
dS
q
+
_
S

j
(q)
_

2
G
k
(p
i
, q)
n
p
n
q

2
G
0
(p
i
, q)
n
p
n
q
_
dS
q
+
_
S
{
j
(q)
j
(p
i
)
j
(p
i
)}

2
G
0
(p
i
, q)
n
p
i
n
q
dS
q
_
S

j
(p
i
) n
q
G
0
n
p
dS
q

1
2

j
(p) n
p
_
= g(p
i
). (4.3)
At this point, it is worth noting that by choosing the basis functions to be piecewise-constant functions,
at any point p
i
that is in the interior of one of the surface elements (and recall that this a requirement
on the points p
i
),
j
(p
i
) = 0 and (4.3) reduces to a variation of the standard piecewise-constant
approximation.
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 9 of 15
The gradient of each basis function can be computed from
_

j
x
x
u
+

j
y
y
u
+

j
z
z
u
=

j
u
,

j
x
x
v
+

j
y
y
v
+

j
z
z
v
=

j
v
,

j
x
n
x
+

j
y
n
y
+

j
z
n
z
= 0,
(4.4)
where the last equation in (4.4) is due to the fact that the basis functions do not vary in the direction
perpendicular to the surface element.
Using the discontinuous basis functions discussed above, assume that the solution over each element
is interpolated using L basis functions and hence we will require L collocation points within each
element. Hence, the total number of basis functions and collocation points is given by m = ML. The
approximation to the integral equation (4.3) can be written in terms of matrix notation as
(A +C) = f , (4.5)
where = [
1
,
2
, . . . ,
ML
] is the vector of the coeffcients appearing in (4.1). In general, A will be
block matrix consisting of M M blocks each of size L L and f will be a block vector consisting
of M blocks of size L. Denote the L collocation points in element i by p
[i ]
l
, l = 1, . . . , L, and let
[i ]
l
,
l = 1, . . . , L, be the L basis functions for element i . Then, the (i, j ) blocks of A and C which we shall
denote A
[i, j ]
and C
[i, j ]
, respectively, are given by
A
[i, j ]
l,m
=
1
2

[ j ]
m
(p
[i ]
l
) +
_
S
j
G
k
(p
[i ]
l
, q)
n
q

[ j ]
m
(q)dS
q
,
C
[i, j ]
l,m
=
_
S
j

2
G
k
(p
[i ]
l
, q)
n
p
n
q

[ j ]
m
(q)dS
q
(4.6)
for 1 l, m L. Clearly, if i = j , then the integrals appearing in (4.6) are non-singular and can be
evaluated using an appropriate quadrature rule. In the case i = j , the integrals for A
[i, j ]
are weakly
singular about the point p
[i ]
l
, but this is not a problem provided that special quadrature rules for dealing
with this weak singularity are used (see Amini et al., 1992, for a discussion of appropriate quadrature
rules). To evaluate the blocks on the diagonal of C, we make use of (3.9) to give
C
[i,i ]
l,m
=
_
S
i
_

2
G
k
(p
[i ]
l
, q)
n
p
n
q

2
G
0
(p
[i ]
l
, q)
n
p
n
q
_

[i ]
m
(q)dS
q
+
_
S
i
(
[i ]
m
(q)
[i ]
m
(p
[i ]
l
)
[i ]
m
(p
[i ]
l
) (q p
[i ]
l
))

2
G
0
(p
[i ]
l
, q)
n
p
n
q
dS
q
+
M

j =1, j =i
_
S
j
(
[i ]
m
(p
[i ]
l
)
[i ]
m
(p
[i ]
l
) (q p
[i ]
l
))

2
G
0
(p
[i ]
l
, q)
n
p
n
q
dS
q
10 of 15 K. CHEN ET AL.
M

j =1
_
S
j

[i ]
m
(p
[i ]
l
) n
q
G
0
(p
[i ]
l
)
n
p
dS
q
(4.7)
for 1 l, m L. All the integrals appearing in (4.7) are at worst weakly singular and so can be
evaluated using the same quadrature rules as for the diagonal block of A.
The entries in the right-hand side vector are given by
f
[i ]
l
=
M

j =1
_
S
j
_
G
k
(p
[i ]
l
, q) +
G
k
(p
[i ]
l
)
n
p
_
(q)
n
q
dS
q
+

2
(p
[i ]
l
)
n
p
, (4.8)
in which all the integrals are at worst weakly singular and provided we take the same care over the
choice of quadrature rule as for the matrices, there should not be any problems with evaluating them.
All the integrals in the current formulation are at worst weakly singular and so can be evaluated using
appropriate quadrature rules. The change of variables given in Duffy (1982), as used in Amini et al.
(1992), along with a product Gauss rule has been used here to evaluate the weakly singular integrals.
The non-singular integrals have been evaluated using a product Gauss rule mapped onto the reference
triangle. As recommended in Amini et al. (1992), we have used a higher-order product Gauss rule for
the weakly singular integrals than the one used for the non-singular integrals.
The choice of coupling parameter is crucial to the accuracy of the Burton and Miller method.
In Amini (1989) and Amini et al. (1992), it was suggested that the almost optimal choice of which
minimizes the condition number of the integral operators is =
i
k
, and this is the value that we shall
use here.
We remark that our new reformulation (3.9) of the second derivative of the Greens function opens
up many new possibilities of solving the fnal linear system by effcient iterative solvers, e.g. by the
fast multipole methods (Rokhlin, 1990) or the conjugate-gradient-type methods, using types of sparse
preconditioners as developed by Chen & Harris (2001) and Chen (2001). Also related is an application
of this work to solving the Helmholtz equation for the case of high wave numbers; see Chandler-Wilde
et al. (2004) and the references therein. We intend to investigate such methods in the near future.
5. Numerical results
We now demonstrate that the high-order collocation method for solving the exterior Helmholtz equa-
tion in three dimensions, proposed in this paper, can yield much higher degree of precision than the
piecewise-constant collocation method that is widely used in practice (see Amini et al., 1992; Chen &
Harris, 2001, and the references therein, e.g.). The individual test problems considered here are out-
lined below, where the point sources are used to calculate both the Neumann boundary data and the
corresponding exact solution:
1. A unit sphere with point sources at (0, 0, 0.5) and (0.25, 0.25, 0.25) with strengths 2+3i and 4i,
respectively.
2. A cylinder of length 0.537 and radius 0.2685 with point sources at (0, 0, 0.15) and (0.25, 0.25,
0.25) with strengths 2 +3i and 4 i, respectively.
3. A peanut-shaped surface defned by
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 11 of 15
x =
_
cos 2 +
_
1.5 sin
2
2 sin cos ,
y =
_
cos 2 +
_
1.5 sin
2
2 sin sin ,
z =
_
cos 2 +
_
1.5 sin
2
2 cos ,
0 ,
0 < 2,
with point sources at (0.2, 0 , 1) and (0, 0.2, 0.75) with strength 2 +3i and 4 i, respectively.
We note that the second test problem is for a non-smooth surface in the sense that it does not possess
a unique normal at every point and that the third test problem is for a surface which is not convex. For
each of these, a test mesh with 576 quadratically curved triangular elements and 1154 nodes was used.
These are the same test problems as used for the Galerkin method in Harris & Chen (2003).
The basis functions used here are the complete linear and quadratic polynomial basis functions,
giving three or six basis functions and collocation points per element, respectively. For the discontinuous
linear basis functions, the collocation points are located at (/3, /3), (12/3, /3) and (/3, 12/3),
where 0 < < 1 is a location parameter. If = 1, then the collocation points would all be located at
the element centroid, while if = 0, the collocation points would be located at the three vertices of the
element. Clearly, there would be problems with either of these cases as the resulting linear system of
equations would not be linearly independent due to two or more collocation points being coincident. For
the discontinuous quadratic basis functions, the collocation points are the three collocation points used
for the linear case and the three points which are midway between each of these. The equivalent results
that would be obtained using the usual piecewise-constant approximation are also given to illustrate the
improvement that can be obtained in the accuracy by using the higher-order basis functions. We remark
that while the results presented in this paper are for the usual quadratically curved triangular boundary
elements (see Amini et al., 1992, for further details), the methods used can be easily adapted for other
types of element.
The relative L
2
-error, defned by
E =

2
, (5.1)
where
2
=
_
_
S
|(q)|
2
dS
q
and and

denote the exact (point-source) and approximate solutions,
respectively, is the measure of the error used here. All errors are given as a percentage. We note that the
integrals appearing in (5.1) have to be evaluated numerically, and here we have used a quadrature rule
of suffcient accuracy so that the errors in evaluating these integrals is minute compared to the error in
the approximate solution.
Figure 1 shows how the L
2
-error varies with using linear basis functions. The values of k used for
the sphere, cylinder and peanut are 3.142, 3.8 and 6.3, respectively, which are the approximate value of
the frst characteristic wave number for each surface. The corresponding results for the quadratic basis
functions are given in Fig. 2. It is clear that the choice of has a signifcant effect on the overall accuracy
of the scheme. The choice = 0.4 seems to give the best results for the linear basis functions, while the
choice = 0.25 gives the best results for the quadratic basis functions, although the appropriate value
to use is not so clear in this case. Similar results can be obtained for other values of k and for brevity
these are not shown here.
12 of 15 K. CHEN ET AL.
FIG. 1. The original (top) and modifed surface (bottom) used to prove Lemma 1.
FIG. 2. L
2
-relative error on each test surface for different values of using the linear basis functions.
Table 1 gives the L
2
-relative error in the computed solution for each surface with a piecewise-
constant approximation, a discontinuous piecewise-linear approximation and a discontinuous piecewise-
quadratic approximation.
These results show that the discontinuous linear and quadratic approximations proposed here are
considerably more accurate than the usual piecewise-constant approximation which has been widely
used in previous work.
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 13 of 15
FIG. 3. L
2
-relative error on each test surface for different values of using the quadratic basis functions.
TABLE 1 L
2
-relative error in the computed solution using constant, linear and quadratic
approximations for each surface
Sphere Cylinder Peanut
k Constant Linear Quadratic Constant Linear Quadratic Constant Linear Quadratic
1 2.63 0.11 0.01 7.18 0.88 0.36 3.77 0.45 0.08
2 3.98 0.2 0.03 10.38 1.08 0.45 5.69 0.67 0.12
3 5.54 0.32 0.04 14.15 1.55 0.57 9.34 1.32 0.17
4 7.21 0.47 0.06 17.84 2.42 0.68 12.29 1.89 0.22
5 8.94 0.66 0.08 22.15 3.64 0.84 14.17 2.52 0.31
6 10.77 0.98 0.21 27.67 5.21 1.08 17.37 3.27 0.54
7 12.65 1.27 0.18 31.32 7.15 1.45 21.78 4.68 0.92
8 14.6 1.6 0.19 35.66 9.1 1.78 24 6.15 1.05
9 16.65 2.09 0.25 45.41 12.13 2.45 27.3 10.18 1.49
10 18.78 3.11 0.38 50.78 15.81 3.52 34.36 12.68 1.9
6. Conclusions
The well-known approach for solving the exterior Helmholtz problem in 3D is the collocation method
using piecewise-constant elements. However, until now it has not been possible to use high-order ele-
ments with the collocation method, due to the integral operator with a hypersingular kernel function,
unless special fnite-part quadrature rules are used to evaluate this hypersingular operator. The method
proposed here reformulates the hypersingular integral into one which is at worst weakly singular and
which can be readily evaluated using well-known techniques.
The numerical results presented in this paper clearly show that there is a considerable improvement
in the accuracy of the newly reformulated Burton and Miller method if discontinuous linear or quadratic
basis functions are used in place of the more usual piecewise-constant basis functions. Further work on
fast solvers issues is in progress.
14 of 15 K. CHEN ET AL.
REFERENCES
AIMI, A. & DILIGENTI, M. (2002) Hypersingular kernel integration in 3D Galerkin boundary element method.
J. Comput. Appl. Math., 138, 5172.
AMINI, S. (1989) On the choice of coupling parameter in boundary integral formulations of the exterior acoustic
problem. Appl. Anal., 35, 7592.
AMINI, S., HARRIS, P. J. & WILTON, D. T. (1992) Coupled Boundary and Finite Element Methods for the
Solution of the Dynamic Fluid-Structure Interaction Problem (C. A. Brebbia & S. A. Orszag eds). Lecture
Note in Engineering, vol. 77. London: Springer.
BURTON, A. J. (1973) The solution of Helmholtz equation in exterior domains using integral equations. NPL
Report NAC 30. Teddington, Middlesex, UK: National Physical Laboratory.
BURTON, A. J. & MILLER, G. F. (1971) The application of integral equation methods to the numerical solution of
boundary value problems. Proc. R. Soc. Lond., A232, 201210.
CHANDLER-WILDE, S. N., LANGDON, S. & RITTER, L. (2004) A high wavenumber boundary element method
for an acoustic scattering problem. Proc. R. Soc. Lond., A362, 647671.
CHEN, K. (2001) An analysis of sparse approximate inverse preconditioners for boundary integral equations. SIAM
J. Matrix Anal. Appl., 22, 10581078.
CHEN, K. & HARRIS, P. J. (2001) Effcient preconditioners for iterative solution of the boundary element equations
for the three dimensional Helmholtz equation. J. Appl. Numer. Math., 36, 475489.
COLTON, D. & KRESS, R. (1983) Integral Equation Methods in Scattering Theory. New York: John Wiley.
DUFFY, M. (1982) Quadrature over a pyramid or cube of integrands with a singularity at the vertex. SIAM J. Numer.
Anal., 19, 12601262.
GERDES, K. (1998) Asummary of infnite element formulations for exterior Helmholtz problems. Comput. Methods
Appl. Mech. Eng., 164, 95105.
GERDES, K. & DEMKOWICZ, L. (1996) Solution of 3D-Laplace and Helmholtz equations in exterior domains
using hp-infnite elements. Comput. Methods Appl. Mech. Eng., 137, 239273.
GIROIRE, J. & NEDELEC, J. C. (1978) Numerical solution of an exterior Neumann problem using a double layer
potential. Math. Comput., 32, 973990.
HACKBUSCH, W. (1995) The Integral Equation Method. Basel: Birkhauser.
HARRIS, P. J. (1992) A boundary element method for the Helmholtz equation using fnite part integration. Int. J.
Comput. Methods Appl. Mech. Eng., 95, 331342.
HARRIS, P. J. & CHEN, K. (2003) On effcient preconditioners for iterative solution of a Galerkin boundary
element equation for the three dimensional exterior Helmholtz problem. J. Comput. Appl. Math., 156, 303
318.
MEYER, W. L., BELL, W. A., ZINN, B. T. & STALLYBRASS, M. P. (1978) Boundary integral solution of three
dimensional acoustic radiation problems. J. Sound Vib., 59, 245262.
ROKHLIN, V. (1990) Rapid solution of integral equations of scattering theory in two dimensions. J. Comput. Phys.,
86, 414439.
SALVADORI, A. (2001) Analytical integrations of hypersingular kernel in 3D BEM problems. Comput. Methods
Appl. Mech. Eng., 190, 39573975.
SCHWAB, C. & WENDLEND, W. L. (1998) Kernel properties and representations of boundary integral operators.
Math. Nachr., 156, 187218.
Appendix A. Proof of Lemma 1
Proof. Let S

be the surface of a sphere centred on p,

be excluding the interior of S

and

S denote
the surface bounding

(see Fig. 3). We note that for q

, it holds that

2
q
G
k
(p, q)
n
p
+k
2
G
k
(p, q)
n
p
= 0. (A.1)
THE BURTON AND MILLER METHOD FOR THE SOLUTION OF A 3D HELMHOLTZ PROBLEM 15 of 15
Applying Greens second theorem,
_

S
_

2
n
q

1
n
q
_
dS
q
=
_

(
1

2
q

2

2

2
q

1
)dV
q
(A.2)
with
1
= a (q p) and
2
=
G
k
(p,q)
n
p
leads to (taking note of (A.1) in the frst integral on the
right-hand side of (A.2))
_

S
_
a (q p)

2
G
k
(p, q)
n
p
n
q
a n
q
G
k
(p, q)
n
p
_
dS
q
=
_

k
2
a (q p)
G
k
(p, q)
n
p
dV
q
. (A.3)
The surface integral in (A.3) can be expressed as the sum of integral over the relevant part of the surface
S

and the integral over the remaining part of

S, say

S

. As 0, then

S

S,

and
_
S

_
a (q p)

2
G
k
(p, q)
n
p
n
q
a n
q
G
k
(p, q)
n
p
_
dS
q

a n
p
2
. (A.4)
Hence, we obtain that
_
S
_
a (q p)

2
G
k
(p, q)
n
p
n
q
a n
q
G
k
(p, q)
n
p
_
dS
q
=
_

k
2
a (q p)
G
k
(p, q)
n
p
dV
q

a n
p
2
, (A.5)
which can be rearranged to give the desired result. Hence, the proof is completed.