Principles of

Boundary Element Methods
Martin Costabel
Technische Hochschule Darmstadt
1 Introduction
1.1 Definition
Boundary integral equations are a classical tool for the analysis of boundary value problems
for partial differential equations. The term “ boundary element method” (BEM) denotes
any method for the approximate numerical solution of these boundary integral equations.
The approximate solution of the boundary value problem obtained by BEM has the distin-
guishing feature that it is an exact solution of the differential equation in the domain and is
parametrized by a finite set of parameters living on the boundary.
1.2 Advantages
The BEM have some advantages over other numerical methods like finite element methods
(FEM) or finite differences:
1. Only the boundary of the domain needs to be discretized. Especially in two dimensions
where the boundary is just a curve this allows very simple data input and storage
methods.
2. Exterior problems with unbounded domains but bounded boundaries are handled as
easily as interior problems.
3. In some applications, the physically relevant data are given not by the solution in
the interior of the domain but rather by the boundary values of the solution or its
derivatives. These data can be obtained directly from the solution of boundary integral
equations, whereas boundary values obtained from FEM solutions are in general not
very accurate.
4. The solution in the interior of the domain is approximated with a rather high conver-
gence rate and moreover, the same rate of convergence holds for all derivatives of any
0
Lectures given at the first graduate summer course in computational physics
“Finite Elements in Physics”
Lausanne 1–10 September 1986
1
order of the solution in the domain. There are difficulties, however, if the solution has
to be evaluated close to, but not on the boundary.
1.3 Difficult parts
Some main difficulties with BEM are the following:
1. Boundary integral equations require the explicit knowledge of a fundamental solution of
the differential equation. This is available only for linear partial differential equations
with constant or some specifically variable coefficients. Problems with inhomogeneities
or nonlinear differential equations are in general not accessible by pure BEM. Some-
times, however, a coupling of FEM and BEM proves to be useful, see section 7.
2. For a given boundary value problem there exist different boundary integral equations
and to each of them several numerical approximation methods. Thus every BEM appli-
cation requires that several choices be made. To evaluate the different possibilities, one
needs a lot of mathematical analysis. Although the analysis of BEM has been a field
of active research in the past decade, it is by no means complete. Thus there exist no
error estimates for several methods that are widely used. From a mathematical point
of view, these methods, which include very popular ones for which computer codes are
available, are in an experimental state, and there might exist problems of reliability.
3. The reason for the difficulty of the mathematical analysis is that boundary integral
equations frequently are not ordinary Fredholm integral equations of the second kind.
The classical theory of integral equations and their numerical solution concentrates
on second kind integral equations with regular kernel, however. Boundary integral
equations may be of the first kind, and the kernels are in general singular. If the
singularities are not integrable, one has to regularize the integrals which are then defined
in a distributional sense. The theoretical framework for such integral equations is the
theory of pseudodifferential operators. This theory was developed 20 years ago and is
now a classical part of Mathematical Analysis [16, 19], but it is still not very popular
within Applied Mathematics.
4. If the boundary is not smooth but has corners and edges, then the solution of the
boundary value problem has singularities at the boundary. This happens also if the
boundary conditions are discontinuous, e.g. in mixed boundary value problems. BEM
clearly have to treat these singularities more directly than FEM. Because the precise
shape of the singularities frequently contains important information, e.g. stress intensity
factors in fracture mechanics, this is a positive aspect of BEM. But besides practical
problems with the numerical treatment of these singularities, non-smooth domains also
present theoretical difficulties. These have so far been satisfactorily resolved only for
two-dimensional problems. The analysis of BEM for three-dimensional domains with
corners and edges is still in a rather incomplete stage.
1.4 Literature
The usefulness of BEM in engineering problems is documented in a literature that amounts
to several thousand pages every year. I will not attempt to give an overview of the wide
range of possible applications. One can find such an overview e.g. in one of the proceedings
2
of the annual conference on Boundary Elements [5] or in the books[4] or [8]. A review of
BEM computer codes can be found in [17]. Introductions to the theory and applications of
BEM can e.g. be found in the books [7, 6, 9, 29] or in[25, 24].
1.5 Contents
In the following sections I shall first describe the general structure of a BEM application and
illustrate it with the simple example of the Dirichlet problem from potential theory.
Section 3 contains a general method for deriving boundary integral equations for general
elliptic boundary value problems.
Section 4 describes boundary integral equations for examples from scattering theory, elas-
ticity theory, and heat conduction.
Discretization methods and their convergence are described in section 5, and section 6
contains some remarks on the treatment of singularities in BEM. The final section 7 contains
a short description of the coupling of FEM and BEM in terms of a (new) symmetric method,
for which also a convergence proof is given.
Acknowledgement. The present notes could only be prepared on the basis of several
survey articles and lectures by Prof. Dr. W. L. Wendland. Section 7 was inspired by the
stimulating atmosphere and discussions at the summer school in Lausanne, whose organizers
and participants I am indebted to.
2 The structure of a BEM application
2.1 A scheme
A typical application of BEM consists of the following parts:
• Mathematical model
• Representation formula
• Boundary integral equation
• Boundary elements
• Discrete equations
• Solution of the linear system
• Interpretation
2.2 Discussion of the scheme
For the following discussion of these parts, the Dirichlet problem for Laplace’s equation will
be used as an illustration.
3
2.2.1
The mathematical model is a boundary value problem for a partial differential equation.
An example is the Dirichlet problem
∆u = 0 in a domain Ω ⊂ R
n
(2.1)
u = g on the boundary Γ := ∂Ω. (2.2)
It is convenient to have a homogenenous differential equation and inhomogeneous boundary
data. A necessity is to know explicitly a fundamental solution of the differential equation. In
the example one has the fundamental solution
γ(x, y) = −
1

log |x −y| (x, y ∈ R
2
) for n = 2
γ(x, y) =
1
4π|x−y|
(x, y ∈ R
3
) for n = 3.
Fundamental solutions are known for equations with constant coefficients where they
can be computed by Fourier transformation, and for some elliptic equations with analytic
coefficients, e.g. the Laplace-Beltrami equation on the sphere.
2.2.2
The next step is to represent the solution of the partial differential equation in the domain
by means of boundary potentials. Such representation formulas are well known for the
classical boundary value problems of mathematical physics, e.g. Green’s third identity for
potential theory, Betti’s formula for elasticity theory, and the Stratton-Chu formula for elec-
trodynamics. In potential theory we have
u(x) =
_
Γ

n(y)
γ(x, y)[u(y)]
Γ
do(y) −
_
Γ
γ(x, y)[∂
n
u(y)]
Γ
do(y) (2.3)
for x ∈ R
n
\ Γ.
Here ∂
n
denotes the normal derivative taken with respect to the exterior normal on Γ,
do is the surface measure on Γ, and u is supposed to be a harmonic function regular both in
Ω and the exterior domain R
n
\ Ω, having i.g. different boundary values on both sides of Γ,
whose jump across Γ is denoted by [·]
Γ
:
[v(x)]
Γ
:= v
|R
n
\Ω
(x) − v
|Ω
(x) (x ∈ Γ).
Thus u is represented as the sum of a double layer and a simple layer potential. If consid-
ered in Ω alone, the representation formula (2.3) can take various different forms depending
on assumptions on u in R
n
\ Ω. Each of these forms gives rise to a different BEM, so one has
to make a choice here. Most common are the following:
1. Supposing u
|R
n
\Ω
≡ 0, one obtains
u(x) = −
_
Γ

n(y)
γ(x, y)u(y)do(y) +
_
Γ
γ(x, y)∂
n
u(y)do(y) (x ∈ Ω) (2.4)
The BEM derived from this is called “method of Green’s formula” or “direct method”.
The densities of the simple and double layer on the boundary are the normal derivative
of the solution and the solution, respectively, on the boundary. Thus these layers have
a “direct” interpretation in terms of the solution.
4
2. Supposing [u]
Γ
≡ 0, one obtains a simple layer representation
u(x) =
_
Γ
γ(x, y)ψ(y)do(y) (x ∈ Ω) (2.5)
with an unknown density ψ.
3. Supposing [∂
n
u]
Γ
≡ 0, one obtains a double layer representation
u(x) =
_
Γ
v(y)∂
n
γ(x, y)do(y) (x ∈ Ω) (2.6)
with an unknown density v.
Sometimes also a suitable linear combination of simple and double layer potentials will
be useful.
2.2.3
The representation formulas give the solution u in the interior of the domain Ω. If one takes
boundary values in the representation formula, one obtains boundary integral equations.
Here one can i.g. also choose between different possibilities. In our example, these are as
follows:
In the “indirect methods” 2. and 3. above, one will insert the representation formulas
(2.5) or (2.6) into the boundary condition (2.2). One has to take into account the well-known
“jump relations” which arise from the fact that the gradient of γ has a strong singularity and
is not integrable on Γ.
The simple layer potential
_
Γ
γ(x, y)ψ(y)do(y) is continuous across Γ, thus (2.5) yields
the integral equation
V ψ(x) :=
_
Γ
γ(x, y)ψ(y)do(y) = g(x) on Γ. (2.7)
This is a Fredholm integral equation of the first kind with a weakly singular kernel.
For the double layer potential (2.6) there holds
u(x) = −
1
2
v(x) +
_
Γ
v(y)∂
n(y)
γ(x, y)do(y) =: −
1
2
v(x) +Kv(x) on Γ.
Here the integral on Γ has to be understood in the principal value sense, but it turns out
that for smooth boundaries the kernel is continuous. Therefore, the integral equation arising
from (2.6),
(−
1
2
+K ) v = g (2.8)
is a Fredholm integral equation of the second kind. (In fact it was, together with its adjoint,
the starting point of the whole theory of integral equations as developed by C.Neumann,
I.Fredholm, and D.Hilbert in 1870–1910.) Note, however, that in the presence of edges and
corners, (2.8) has a strongly singular kernel and is no longer a Fredholm integral equation.
The representation formula (2.4) admits two ways of approaching the boundary:
If one considers just u

, one obtains
u =
1
2
u −Ku +V (∂
n
u) on Γ.
5
But one can also take ∂
n
u

and observe the jump relations:
The normal derivative of the double layer potential is continuous across Γ, if the boundary
and the density are smooth, and for x ∈ Γ

n(x)
_
Γ
γ(x, y)ψ(y)do(y) =
1
2
ψ(x) +
_
Γ

n(x)
γ(x, y)ψ(y)do(y) =:
1
2
ψ(x) +K

ψ(x).
Here, K

is the adjoint operator of K above. Thus one obtains

n
u = Du +
1
2

n
u +K


n
u on Γ,
where
Du(x) := −∂
n(x)
_
Γ

n(y)
γ(x, y)u(y)do(y) on Γ
is the normal derivative of the double layer potential. The “integral operator” D has a
hypersingular kernel.
We have now found the two relations on Γ
(
1
2
+K ) u = V (∂
n
u) (2.9)
Du = (
1
2
−K

)(∂
n
u). (2.10)
Both relations together express simply the fact that the two functions u

and ∂
n
u

, the
“Cauchy data” of u, are derived from a function u that satisfies the differential equation (2.1)
in Ω. A method to derive such relations for more general elliptic equations is described in
section 3.
Now either one of the equations (2.9), (2.10) together with the boundary condition (2.2)
can be used to determine ∂
n
u

and thus, with the representation formula (2.4), to solve the
Dirichlet problem. From (2.9) one obtains
V (∂
n
u) = (
1
2
+K ) g, (2.11)
an integral equation of the first kind with the same kernel as in (2.7). From (2.10) one obtains
(
1
2
−K

)(∂
n
u) = Dg, (2.12)
an integral equation of the second kind with the adjoint kernel of (2.8).
Any one of the four integral equations (2.7), (2.8), (2.11), or (2.12) can equally well be
used for the numerical solution of the Dirichlet problem.
It is an easy exercise to write down the corresponding integral equations for the Neumann
problem where ∂
n
u is given on Γ, and also for the Robin problem where a linear combination
of u and ∂
n
u is given on Γ, and even for mixed boundary value problems where on different
parts of Γ different boundary conditions are satisfied. In each case, only the four operators
V , D, K, and K

appearing in (2.9) are involved.
Two final remarks on the variety of boundary integral equations:
• In representation formulas like (2.4), γ can be any fundamental solution. Thus if Γ has
a simple geometry that allows to find a fundamental solution which vanishes on parts
of Γ, the integrals may simplify considerably. For instance, if γ is the Green function of
Ω (which is known e.g. for balls or half spaces), then γ(x, y) ≡ 0 for y ∈ Γ, hence the
direct method in this case coincides with the method of double layer representation.
6
• If Γ is an open surface which does not separate R
n
into two disjoint domains, one has
no choice but has to work with (2.3) directly. This rules out all integral equations of
the second kind and leaves for the Dirichlet problem the integral equation of the first
kind with the operator V and for the Neumann problem the integral equation of the
first kind with the operator D (see [22, 23]). Such situations occur in crack problems
or in screen scattering problems.
2.2.4
For the numerical solution of the boundary integral equations one needs functions on the
boundary that depend on a finite set of parameters, i.e. a finite dimensional function space
on Γ. One can choose globally defined functions like spherical harmonics or polynomials,
but for the proper BEM these trial functions are finite element functions on the boundary,
the boundary elements. One assumes that Γ can be decomposed into a finite number of
subsets each of which has a regular parameter representation by some parameter domain
in R
n−1
. Then one chooses regular partitions of the parameter domains and correspending
finite element functions, e.g. a S
k,m
h
family in the sense of Babuˇska and Aziz [2]. Using the
parameter representations, one transfers these functions to the boundary. If the parameter
representation is not explicitly known but the boundary itself has also to be approximated,
one uses the approximate coordinate representations. The order of the polynomial basis
functions in these approximations of course has to be chosen in correspondence to the order
of the Sobolev spaces in which the solution of the boundary integral equation is looked for and
also to the kind of discretization scheme that is chosen. Here it is useful to have asymptotic
error estimates in order to make the right choice. This will be discussed in section 5.
If, in the presence of corners and edges, one knows which singularities appear in the
solution, then one can build them into the space of trial functions as in FEM, either by
using a suitable nonuniform partition or by augmenting the space of trial functions by some
singular functions, i.e. using singular boundary elements.
2.2.5
The finitely many parameters determining the approximate solution are computed from
finitely many linear equations. There are three main methods to generate this system of
discrete equations: Collocation, Galerkin’s method, and the least squares method.
If we have a boundary integral equation
Au = f on Γ (2.13)
and we seek an approximate solution
u
h
(x) =
N

j=1
γ
j
µ
j
h
(x) (2.14)
with basis functions {µ
j
h
|j = 1, . . . , N}, then these methods for obtaining a linear system for
the unknown coefficients {γ
j
|j = 1, . . . , N} are the following:
1. For the collocation method one chooses a suitable set {x
j
|j = 1, . . . , N} ⊂ Γ of col-
location points and requires that the equation (2.13) is satisfied in these points. This
7
gives the system
N

j=1
(Aµ
j
h
)(x
k
) γ
j
= f(x
k
) (k = 1, . . . , N). (2.15)
2. For the Galerkin method one multiplies (2.13) for u
h
with test functions from a finite
dimensional function space, integrates over Γ and equates the integrals. In the proper
Galerkin method, test and trial functions are the same, hence this gives the system
N

j=1

k
h
, Aµ
j
h
) γ
j
= (µ
k
h
, f) (k = 1, . . . , N) (2.16)
where the brackets (·, ·) denote the L
2
(Γ) inner product, i.e.
(f, g) :=
_
Γ
f(x)g(x)do(x).
3. In the least squares method one minimizes Au
h
−f in the L
2
(Γ) norm. The resulting
equations are those for a Galerkin (-Petrov) method with {Aµ
k
h
|k = 1, . . . , N} as a basis
for the test functions:
N

j=1
(Aµ
k
h
, Aµ
j
h
) γ
j
= (Aµ
k
h
, f) (k = 1, . . . , N). (2.17)
The coefficient matrix in these methods has to be computed using numerical integration.
Here the collocation method is the simplest one because it involves only one integral whereas
Galerkin’s method needs two and the least squares method three integrations. This may
be the reason why collocation is most frequently used in the applications. It has also some
drawbacks, however:
The coefficient matrix in the least squares method is always symmetric and, if (2.13) is
uniquely solvable, even positive definite. A careful choice of the boundary integral operator
can yield a positive definite selfadjoint operator A or at least a positive definite principal
part. For instance, the first kind operators V and D from potential theory are selfadjoint
and, in R
3
, positive definite. In the Galerkin method, then the coefficient matrix will also be
positive selfadjoint. The coefficient matrix for the collocation method is never symmetric.
Convergence proofs and asymptotic error estimates are available
• for least squares methods under the general condition of ellipticity for the operator A,
• for Galerkin methods under the somewhat stronger condition of strong ellipticity for
the operator A,
• for collocation methods only in 2 dimensions for strongly elliptic operators A and smooth
domains (apart from some very special recent results [14, 10]) and in higher dimensions
only for Fredholm integral equations of the second kind (for which, on the other hand,
there exist also a lot of other numerical methods, e.g. the well-known Nystr¨om method,
see [3]).
8
The integrations in the computation of the coefficient matrix have to be done numerically.
This requires another choice which, however, depends very much on the specific problem at
hand. A common problem is the evaluation of singular integrals, and the usual recipe is to
split off the main singularities and try to evaluate the corresponding integrals analytically,
while the less singular remainder is computed with the use of a suitable quadrature formula.
The order of this quadrature formula should be chosen in accordance with the asymptotic
order of convergence for the chosen discretization method.
2.2.6
In comparison with the computation of the matrix elements, the solution of the linear
system of equations is usually relatively uncomplicated. Note that the matrices are not
sparse, so the linear systems solvers developed for FEM are not useful in BEM. In some
specific examples, multigrid methods have been analyzed and successfully applied [20]. In
general, standard linear systems solvers should suffice, if one exploits symmetry and positive
definiteness where available.
2.2.7
The postprocessing, i.e. the interpretation of the computed data, is the last step in BEM.
The interpolation formula (2.14) gives the solution of the boundary integral equation. In the
direct method, this may be already the desired information about the boundary values of the
solution of the boundary value problem. If these Cauchy data have to be determined from
an indirect method, then another integral operator with singular kernel has to be applied.
If the solution of the boundary value problem in the interior of Ω is desired, then the
representation formula which was used to derive the boundary integral equation gives the
answer. Here only a regular integral has to be evaluated, because the kernel defined by the
fundamental solution is, for elliptic and parabolic problems, singular only if the points of
integration and of evaluation coincide. Thus this integration has a smoothing effect and sup-
presses high-frequency errors. In addition, the representation formula may be differentiated
and used to compute the derivatives of the solution in the domain.
3 Derivation of boundary integral equations
Let P be a linear elliptic differential operator of order 2m with smooth coefficients on R
n
.
Assume that a fundamental solution G of P is known, that is
• G is a two-sided inverse of P on functions (or distributions) that vanish outside a
compact set,
GPu = u = PGu for all compactly supported u.
• G is an integral operator
Gf(x) =
_
R
n
G(x, y)f(y)dy
with a kernel G(x, y) that is weakly singular for x = y.
9
In this section we describe the method of the Calder´ on projector to obtain boundary integral
equations for boundary value problems fo the following form
Pu = f in Ω ⊂ R
n
(3.1)
Rγu = g on Γ = ∂Ω. (3.2)
Here
γu := (γ
0
u, . . . , γ
2m−1
u)

: = (u, ∂
n
u, ∂
2
n
u, . . . , ∂
2m−1
n
u)


are the “Cauchy data” of u, and R is a (m×2m) matrix of differential operators on Γ. Thus
the boundary data g are a m-component vector function.
3.1 Second Green formula
Let Ω
1
:= Ω and Ω
2
:= R
n
\ Ω. For a function u on R
n
with u
|Ω
j
= u
j
we write [γ
k
u]
Γ
=
γ
k
u
2
−γ
k
u
1
if the Cauchy data γ
k
u
j
= ∂
k
n
(u
j
)
Γ
exist. Here one assumes that in a neighborhood
of Γ there exists a unit vector field n that is the normal field exterior to Ω
1
on Γ, and the
higher normal derivatives are defined using this vector field.
Introducing suitable coordinates, one can see that P has a decomposition
P =
2m

j=0
P
j

j
n
(3.3)
Here P
j
are differential operators of order 2m− j whose restrictions to Γ contain only tan-
gential derivatives.
For example, the Laplace operator P = −∆ is decomposed as
−∆ = −∆
0
−2H∂
n
−∂
2
n
,
where H =
1
2
div n is the mean curvature of Γ and ∆
0
is the Laplace-Beltrami operator on
Γ.
Now we choose a piecewise smooth function u as above such that u
j
∈ C

(Ω
j
). We apply
P to u in the distributional sense. One has
Pu = f in R
n
\ Γ with f ∈ C

(R
n
\ Γ).
Therefore the distribution Pu − f vanishes on R
n
\ Γ. It consists of multiple layers of the
form v ⊗∂
k
n
δ
Γ
defined by
< v ⊗∂
k
n
δ
Γ
, ϕ >:=
_
Γ
v ∂
′k
n
ϕdo for ϕ ∈ C

(Γ).
Here ∂
′k
n
is the formal transpose of the differential operator ∂
k
n
, hence

′k
n
= (−1)
k

k
n
+ lower order terms.
The result can be found in detail in [16]. It is the distributional version of the second
Green formula:
Pu = f +
2m−1

k=0
2m−1−k

l=0
P
k+l+1

l
u]
Γ
⊗∂
k
n
δ
Γ
. (3.4)
10
If we specialize to u
2
≡ 0, u
1
= v, and apply (3.4) to a test function ϕ, we obtain the
more familiar equivalent formula
_

(v

P

ϕ −(Pv)

ϕ)dx =
2m−1

k=0
2m−1−k

l=0
_
Γ
(P
k+l+1
γ
l
v)

(∂
′k
n
ϕ)do. (3.5)
Here the superscript

denotes transposition, because we admit throughout matrix-valued
operators P and vector-valued functions u, v etc., and P

is the formal transpose of P:
< Pu, ϕ >=< u, P

ϕ >=
_

v

P

ϕdx for ϕ ∈ C

0
(R
n
).
The formulas (3.4) and (3.5) are equivalent, but it is the distributional form (3.4) that
is the starting point for the representation formula and boundary integral equations. In the
example of the Laplace operator, (3.4) is
−∆u = f −[γ
1
u]
Γ
⊗δ
Γ
+ [γ
0
u]
Γ
⊗∂

n
δ
Γ
.
In the case of a second order operator (we abbreviate ∂
j
:= ∂/∂x
j
)
P = −
n

j,k=1

j
a
jk

k
+
n

j=1
b
j

j
+ c (3.6)
one has a first Green formula
_

(Pu)

v dx = Φ

(u, v) −
_
Γ
(∂
ν
u)

v do (3.7)
with
Φ

(u, v) :=
_

_
_
n

j,k=1
(a
jk

k
u)


j
v +
n

j=1
(b
j

j
u)

v + (cu)

v
_
_
dx
and the conormal derivative

ν
u :=
n

j,k=1
n
j
a
jk

k
u.
Here again the coefficients a
jk
, b
j
, and c may be matrix-valued. Then, for example, the
equations of linear elasticity can be treated in this way, where then ∂
ν
is the traction operator.
From (3.7) follows the second Green formula
_

(u

P

v −(Pu)

v)dx =
_
Γ
((∂
˜ ν
u)

v −u


ν
v)do (3.8)
with

˜ ν
u := ∂
ν
u −
n

j=1
n
j
b
j
u ;


ν
v :=
n

j,k=1
n
j
a

jk

k
.
In the case of elasticity, (3.8) is Somigliana’s identity.
11
3.2 Representation formula
From the second Green formula in the distributional form, one obtains the representation
formula simply by applying the inverse G of the differential operator P. Thus if u is piecewise
smooth as above and vanishes outside a compact set, then there holds the representation
formula
u = Gf +
2m−1

k=0
2m−1−k

l=0
K
k
(P
k+l+1

l
u]
Γ
) (3.9)
with the multiple layer potential operators K
k
defined by
K
k
ϕ := G(ϕ ⊗∂
k
n
δ
Γ
) for ϕ ∈ C

(Γ).
For x ∈ Γ this means
K
k
ϕ(x) =
_
Γ

′k
n(y)
G(x, y)ϕ(y) do(y).
In the example of the Laplace operator, this is just (2.3) if f = 0.
For the second order operator (3.6) one obtains from (3.8)
u(x) = Gf(x) +
_
Γ
_
(
¯

ν(y)
G(x, y)

)

[u(y)]
Γ
−G(x, y)∂
˜ ν
u(y)
_
do(y). (3.10)
For this case of a second order operator, one can allow corners and edges on Γ, even an
arbitrary Lipschitz domain [12], whereas for the general case (3.9) the boundary has to be
smooth. For the bi-Laplace operator P = ∆
2
in R
2
also representation formulas are known
in domains with corners, and they contain additional terms coming from the corners.
3.3 Calder´on projector
Taking traces on Γ in (3.9), we find an identity for the Cauchy data γu
1
, if we set u
2
≡ 0:
γu
1
= γGf
1
+ C(γu
1
) (3.11)
with
C(γu
1
) := −
2m−1

k=0
2m−1−k

l=0
γ(K
k
(P
k+l+1
γ
l
u
1
))
|Ω
.
The operator C is a (2m× 2m) matrix of pseudodifferential operators on the boundary.
In the case of the Laplace operator, it is in the notation of section 2.2.3
C =
_
1
2
−K V
D
1
2
+K

_
.
One can show that C is a projection operator: CCv = Cv for all v. The complementary
projector 1 −C is the Calder´on projector for the complementary domain.
As a pseudodifferential operator, C has well-known continuity properties in Sobolev spaces
on Γ. I will not give the definition of a Sobolev space H
s
(Γ) for arbitrary s ∈ R here, but
recall that for integer s, H
s
(Γ) is the space of all functions on Γ with finite norm
v
H
s
(Γ)
:=
_
_

k≤s
_
Γ
|D
k
v|
2
do
_
_
1
2
,
12
where D
k
v is the vector of all k-th order tangential derivatives of v. The dual space of the
Hilbert space H
s
(Γ) in the natural L
2
(Γ) duality is H
−s
(Γ).
It is known that C = (C
jk
)
2m−1
j,k=0
where C
jk
is a pseudodifferential operator of order
j − k. This means that C
jk
maps H
s
(Γ) into H
s−j+k
(Γ) continuously for any s. Thus the
boundary “integral” operators C
jk
have orders ranging from −2m+1 to 2m−1. Only those
with nonpositive order are decent integral operators, where order zero operators can contain
strongly singular kernels and thus integrals existing only in the principal value sense. The
operators with positive order have hypersingular kernels like the operator D which has order
+1.
But all of these highly nonclassical “integral” operators do appear in applications and
they can be used in BEM.
3.4 Boundary integral equations
The relation (3.11) is a faithful translation of the differential equation (3.1) to the boundary:
A 2m-vector function v satisfies v = γGf +Cv if and only if v equals γu with some solution u
of (3.1). Thus (3.11) together with the boundary condition (3.2) is an equivalent formulation
of the boundary value problem by means of functions on the boundary. It is a system of 3m
equations for the 2m unknowns γ
0
u, . . . , γ
2m−1
u.
There are several ways of making a quadratic system out of this rectangular one. One
way would be to treat the 3m× 2m system directly by least squares methods. This would
lead to a 2m× 2m system. The ellipticity of this system, i.e. the invertibility of its symbol
as a pseudodifferential operator, is equivalent to the Shapiro-Lopatinski ellipticity condition
for the boundary value problem (3.1)–(3.2).
The usual procedure, however, is to use the boundary condition to eliminate m unknowns
and find a m × m system for the remaining m unknown functions. This procedure uses a
second set S of m boundary operators with the property that the (2m×2m) matrix
M :=
_
R
S
_
of tangential differential operators on Γ has an inverse
M
−1
= (
ˇ
M
jk
)
2m−1
j,k=0
which also consist of differential operators. Then the m-component vector function
v = Sγu
contains the unknown Cauchy data, and the Cauchy data γu are known if v is known and g
is given by (3.2):
γu = M
−1
_
g
v
_
. (3.12)
In this case, the solution of the boundary value problem in Ω is given by the representation
formula (3.9):
u = Gf −KPM
−1
_
g
v
_
, (3.13)
13
where we introduced the vector of integral operators
K := (K
0
, . . . , K
2m−1
)
and the matrix of differential operators
P := (P
k+l+1
)
2m−1
k,l=0
with P
j
:= 0 for j > 2m.
Taking the “modified Cauchy data” Mγu in (3.13), we have the 2m×2m system
_
g
v
_
= MγGf − MγKPM
−1
_
g
v
_
,
which can of course be also written in the form
_
g
v
_
= MγGf +
˜
C
_
g
v
_
with the “modified Calder´on projector”
˜
C := MCM
−1
.
We write this system finally in the form
g = RγGf +RCM
−1
_
g
v
_
v = SγGf +SCM
−1
_
g
v
_
or
Av = g −RCM
−1
_
g
0
_
−RγGf (3.14)
(1 −B)v = SCM
−1
_
g
0
_
+SγGf (3.15)
where
Av = RCM
−1
_
0
v
_
= −RγKPM
−1
_
0
v
_
and
Bv = SCM
−1
_
0
v
_
= −SγKPM
−1
_
0
v
_
are the upper resp. lower right hand “(m×m) corner” of the (2m×2m) matrix
˜
C.
Now either of the two sets of m equations (3.14) or (3.15) is equivalent to the original
boundary value problem under suitable conditions which can be found in [15]:
Theorem 3.1 Let f ∈ H
s−m
(Ω) ∩ L
2
(Ω) and g ∈

m−1
j=0
H
m−µ
j

1
2
+s
(Γ) be given where s is
an arbitrary real number and µ
j
are the orders of the rows of the boundary differential operator
R. Then every solution u ∈ H
m+s
(Ω) of the boundary value problem (3.1), (3.2) defines by
(3.12) a solution of (3.14) and (3.15). Conversely, every solution v ∈

m−1
j=0
H
−m+µ
j
+
1
2
+s
(Γ)
of one of the boundary integral equations, (3.14) or (3.15), defines by the representation
formula (3.13) a solution u ∈ H
m+s
(Ω) of the boundary value problem.
14
Under some conditions on the strong ellipticity of the boundary value problem (3.1), (3.2),
one can show [15] that the operator A in (3.14) is also strongly elliptic. These conditions
are satisfied for the usual boundary value problems for the Laplace and Helmholtz equations
and for the equations of linear elasticity theory, whether formulated as the second order
system of the Navier equations or, through the Airy stress function, as the fourth order
scalar biharmonic equation. They are not satisfied for some boundary value problems from
electrodynamics.
Strong ellipticity of the operator A implies
1. Fredholm’s alternative holds for the boundary integral equation (3.14), i.e. the homo-
geneous equation has a finite number l of linear independent solutions and there exist l
linear independent solvability conditions. Thus, by adding a finite number of rows and
columns, the system can be converted into a uniquely solvable one of the form
Av +
l

j=1
ω
j
v
0
j
= ψ
_
Γ
λ

k
v do = β
k
(k = 1, . . . , l),
(3.16)
where ψ is the right hand side in (3.14), v
0
j
and λ
k
are certain given functions related
to the cokernel and kernel of A, β
k
are arbitrarily given numbers, e.g. zero, and ω
j
are numbers that have to be determined from the boundary integral equation. The
enlarged system is also equivalent to the boundary value problem. Such an addition of
rows and columns is necessary in elasticity theory to eliminate the rigid body motions,
and also sometimes in potential and scattering theory.
2. Every Galerkin procedure for (3.16) is stable and converges [21]. The asymptotic or-
der of convergence is quasioptimal, i.e. it is the same order as for the best possible
approximation of the solution by means of the boundary element trial functions.
3. In the case of smooth 2-dimensional domains, also collocation methods are stable and
converge quasioptimally.
The same result of strong ellipticity for the first kind operator A in (3.14) holds for second
order systems even on arbitrary Lipschitz domains Ω, i.e. in the presence of corners and edges
[11].
For the second kind operator 1 − B in (3.15) strong ellipticity is known in special cases,
mainly on smooth domains, but in some cases on non-smooth domains it fails.
4 Examples of boundary integral equations
4.1 Acoustic scattering
We consider a screen Γ ⊂ R
3
, i.e. an open surface. The total field u = u
in
+ u
s
as well as
the incident (plane or spherical) wave u
in
and the scattered field u
s
satisfy the Helmholtz
equation. If u satisfies homogeneous Neumann conditions on Γ, we have for the scattered
field u
s
the Neumann problem
(∆ +k
2
) u
s
= 0 in R
n
\ Γ (4.1)

n
u
s
= g on Γ (4.2)
15
with g = −∂
n
u
in
.
The direct method as well as the method of double layer potential representation (they
coincide due to [∂
n
u
s
]
Γ
= 0 ) both lead to the first kind integral equation
Dv = −g (4.3)
with the hypersingular integral operator D introduced in section 2. Here, of course, the
fundamental solution γ is the outgoing one for the Helmholtz equation:
γ(x, y) =
e
ik|x−y|
4π|x −y|
.
The solution v has a direct interpretation as
v = [u
s
]
Γ
.
The operator D is (for real wave numbers k) a positive selfadjoint operator on L
2
(Γ)
with domain of definition
˜
H
1/2
(Γ) which is the dual space of H
−1/2
(Γ), and D maps
˜
H
1/2
(Γ)
bijectively onto H
−1/2
(Γ). Thus every Galerkin method for the equation (4.3) is convergent.
One has to choose the boundary elements in such a way that the solution v ∈
˜
H
1/2
(Γ) is
approximated as good as possible. In general the solution will have unbounded derivatives
near the screen edge. This allows an asymptotic order of convergence of at most O(h) for
regular finite elements. There exist methods using singular boundary elements (see section
6) that allow an order of O(h
3
) [22].
4.2 Elasticity theory
For the system of partial differential equations
µ∆u + (λ +µ) grad div u = 0 in Ω (4.4)
one has the representation formula
u(x) =
_
Γ
(F(x, y)

t(y) − T(y, x) ϕ(y))do(y) for x ∈ Ω, (4.5)
where
ϕ = u

;

t = T(u)
with the traction operator
T(u) := λn(x) div u + 2µ∂
n
u + µn ∧ curl u.
F is the fundamental solution of (4.4)
F(x, y) =
λ + 3µ
2µ(λ + 2µ)
_
γ(x, y) +
λ +µ
λ + 3µ
(x −y)(x −y)

|x −y|
n
_
(n = 2, 3)
with γ as in section 2.2.1, and
T(y, x) = (T
(y)
(F(y, x)))

.
16
Using the jump relations for elastic potentials, one obtains on Γ [28]
u(x) =
1
2
ϕ(x) −
_
Γ
(T(y, x) ϕ(y) +F(x, y)

t(y))do(y) (4.6)
T(u)(x) = −T
(x)
(
_
Γ
T(y, x) ϕ(y)do(y)) +
1
2

t(x) +
_
Γ
T(x, y)

t(y)do(y). (4.7)
Here the kernels on the right hand side are strongly singular, and the integrals have to
be understood in the Cauchy principal value sense.
The equations (4.6), (4.7) immediately give boundary integral equations for the two
boundary value problems
T(u) =

t given on Γ (4.8)
or
u = ϕ given on Γ. (4.9)
Let us consider the first one (4.8). Here the unknown is u

= ϕ.
The equation (4.6) gives for n = 3 [28]
1
2
ϕ(x) +
_
Γ
T(y, x) ϕ(y)do(y) +M(x)ω =
_
Γ
F(x, y)

t(y)do(y)
_
Γ
M(y)

ϕ(y)do(y) = 0 (4.10)
with the matrix
M(x) =
_
_
_
1 0 0 −x
2
0 x
3
0 1 0 x
1
−x
3
0
0 0 1 0 x
2
−x
1
_
_
_.
Here the extra 6 unknowns ω and the extra 6 side conditions are added in order to remove
the rigid body motions, thus making the system (4.10) uniquely solvable.
The integral operator in (4.10) is strongly singular. Thus (4.10) is not a Fredholm integral
equation of the second kind. For smooth domains, however, it has been shown that the
integral operator in (4.10) is a strongly elliptic pseudodifferential operator of order zero.
Therefore any Galerkin method for (4.10) is convergent.
If we take the equation (4.7), we obtain the first kind equation
T
(x)
(
_
Γ
T(y, x) ϕ(y)do(y)) +M(x)ω =
1
2

t(x) −
_
Γ
T(x, y)

t(y)do(y)
_
Γ
M(y)

ϕ(y)do(y) = 0. (4.11)
Here the kernel is hypersingular. But one can show that even for any Lipschitz domain the
operator defines a positive definite hermitian bilinear form, and hence any Galerkin method
converges for (4.11) even if the domain has corners and edges.
BEM for three-dimensional elasticity are of course very important in engineering applica-
tions. Frequently the equations are discretized using a collocation method (BETSY, BEASY
and other programs). It is interesting to note, however, that all attempts to prove convergence
for these methods have been, to the best of the author’s knowledge, until now unsuccessful.
This gives a theoretical argument in favor of Galerkin methods.
17
4.3 Heat conduction
Stationary heat conduction processes are governed by Laplace’s equation which was treated
above. Therefore here we concentrate on transient processes. The heat equation is
Lu := ∂
t
u −∆u = 0 in Q = Ω ×(0, T) ⊂ R
n+1
. (4.12)
There exist several BEM for initial-boundary value problems for parabolic equations:
• One can choose a time-discretization scheme that requires at each time-step the solution
of an elliptic boundary value problem in the space domain Ω.
• Also the application of Laplace transformation leads to a (parameter-dependent) elliptic
boundary value problem. The elliptic boundary value problems are then solved by one
of the BEM described above.
• With the help of a time-dependent fundamental solution one can perform BEM directly
on the boundary of the space-time cylinder Q.
The latter method is described here along the lines of the general scheme described in
section 3.
A fundamental solution of (4.12) is given by
G(x, t, y, s) =
_
_
_
(4π(s −t))

n
2
e

|x−y|
2
4(s−t)
for s > t
0 for s < t.
We apply L in the distributional sense to a function u which is smooth in Q and in
R
n+1
\ Q, satisfies Lu = f in R
n+1
\ ∂Q, and has jumps across
∂Q = Ω
0
∪ Ω
T
∪ (Γ ×[0, T]) (Ω
t
:= Ω ×{t}).
The result is the distributional version of the second Green formula:
Lu(x, t) = f(x, t)−u(x, T)δ(t−T)+u(x, 0)δ(t)+[u(x, t)]
Γ
⊗∂

n
δ
Γ
(x, t)−[∂
n
u(x, t)]
Γ
⊗δ
Γ
(x, t).
(4.13)
Convolution with G gives the representation formula for (x, t) ∈ ∂Q:
u(x, t) =
_
Q
G(x, t, y, s)f(y, s)dyds +
_

(G(x, t, y, 0)u(x, 0) −G(x, t, y, T)u(x, T))dy
+
_
T
0
_
Γ
(∂
n(y)
G(x, t, y, s)[u(y, s)]
Γ
−G(x, t, y, s)[∂
n
u(y, s)]
Γ
)do(y)ds.
Taking into account that G = 0 for s < t, this is equivalent to
u(x, t) =
_
t
0
_

G(x, t, y, s)f(y, s)dyds +
_

G(x, t, y, 0)u(x, 0)dx (4.14)
+
_
t
0
_
Γ
(∂
n(y)
G(x, t, y, s)[u(y, s)]
Γ
−G(x, t, y, s)[∂
n
u(y, s)]
Γ
)do(y)ds.
If we consider the initial value problem
u(x, 0) = u
0
(x) (4.15)
18
for the equation (4.12) then on the right hand side of (4.14) the first term is absent and the
second is known. (For general initial values u
0
, it has to be computed by a domain integral,
which, due to the smoothing nature of the kernel G(x, t, y, 0), may be discretized relatively
crudely.) The remaining double and simple layer potentials resemble those for the Laplace
equation. Thus one may take traces of u and ∂
n
u on Γ and obtain, as in section 2.2.3, two
boundary integral equations for both the Dirichlet and the Neumann boundary conditions.
We leave this as an exercise.
In the case of smooth Ω, the respective second kind integral equations have weakly sin-
gular kernels and can therefore be treated by various numerical methods. The operators are
no ordinary pseudodifferential operators, however, L being parabolic, and therefore little is
known about the first kind equations.
5 Convergence
The state of the art of asymptotic error estimates for BEM is described in several detailed
articles by W. L. Wendland, contained and quoted in the books mentioned in section 1.4. In
particular, [25] and [28] give a complete survey of the known results, ranging from abstract
approximation schemes for general pseudodifferential operators to fully discretized bound-
ary integral equations including numerical integration methods. [28] contains tables showing
relations which have to be satisfied by the approximation orders of the coordinate represen-
tation of Γ and the approximation orders in the numerical integrations in order to get highest
convergence rates.
Here are some remarks on the basic rules of error estimates for BEM. They are gener-
alizations of those for FEM, and are particularly easy to state for Galerkin methods, so we
consider only these here.
Let A : H
α
(Γ) → H
−α
(Γ) be a bijective and continuous linear operator for some real α.
Assume that A satisfies a G˚arding inequality:
There exists a compact operator T : H
α
(Γ) → H
−α
(Γ) and a γ > 0 such that
Re(v, (A +T)v) ≥ γv
2
H
α
(Γ)
∀v ∈ H
α
(Γ). (5.1)
If A is a pseudodifferential operator of order 2α, then (5.1) is equivalent to the strong
ellipticity of A. The first kind boundary integral operators for the Dirichlet respectively
Neumann problem for second order strongly elliptic systems ( the operators V resp. D in
potential theory) satisfy (5.1) with α = −
1
2
resp. α =
1
2
, even on domains with corners and
edges.
For the Galerkin procedure, one has a sequence of finite-dimensional subspaces
H
h
⊂ H
α
(Γ) (0 < h < h
0
), (5.2)
and one looks for u
h
∈ H
h
satisfying
(v, Au
h
) = (v, f) ∀v ∈ H
h
, (5.3)
where f = Au ∈ H
−α
(Γ).
If for all u ∈ H
α
(Γ),
inf{u − ˜ u
H
α
(Γ)
| ˜ u ∈ H
h
} −→ 0 as h → 0,
19
then (5.1) implies stability and quasioptimal convergence for the Galerkin procedure:
There exists h
1
> 0 such that for all h < h
1
, (5.3) defines a unique u
h
∈ H
h
, and there is
a constant c independent of u and h with
u
h

H
α
(Γ)
≤ cu
H
α
(Γ)
and
u −u
h

H
α
(Γ)
≤ c inf{u − ˜ u
H
α
(Γ)
| ˜ u ∈ H
h
}. (5.4)
Now, as is well known e.g. from FEM for shells, the boundary element trial spaces H
h
as
described in 2.2.4 above, have the following approximation property:
There are numbers k and m such that for any s < t with s < m and t ≤ k and any
u ∈ H
t
(Γ) there exists an approximating ˜ u ∈ H
h
with
u − ˜ u
H
s
(Γ)
≤ ch
t−s
u
H
t
(Γ)
(5.5)
where c does not depend on u or h. m measures the interelement differentiability and k the
degree of the polynomials.
This together with (5.4) gives a convergence order
u −u
h

H
α
(Γ)
= O(h
t−α
) if u ∈ H
t
(Γ). (5.6)
Using an inverse assumption and duality arguments, one can estimate the error also in
norms different from the “energy norm” H
α
(Γ):
u −u
h

H
s
(Γ)
= O(h
t−s
) if u ∈ H
t
(Γ), (5.7)
where s is restricted by
2α −k ≤ s < m.
Note that for (5.6) to hold, one needs α < m. This is just the conformity condition (5.2).
In particular, if α ≤ 0, one can use discontinuous trial functions.
The highest order of convergence is obtained for s = 2α −k and t = k:
u −u
h

H
2α−k
(Γ)
= O(h
2(k−α)
) if u ∈ H
k
(Γ). (5.8)
Here it is an advantage to have negative α. Then, however, the L
2
-condition number of
the discrete equations blows up like O(h
−2|α|
). This reflects the fact that equations with
such operators are traditionally called ill-posed. It depends on the size of the linear system
and the machine precision, which of the two effects, the theoretical Galerkin error or the
round-off error magnified by a big condition number, becomes dominant. If the size of the
system is chosen so that both effects balance, then the approximation is of the same order
as for approximation methods specifically created for ill-posed problems, e.g. Tychonov
regularization.
When the approximate solution of the boundary integral equations is inserted into the
representation formula, then the highest convergence rate (5.8) on the boundary is inherited
by the solution of the boundary value problem away from the boundary:
One has for the multiple layer potentials K
k
from 3.2:
K
k
ϕ
H
s
(Ω
0
)
≤ cϕ
H
t
(Γ)
for any s and t
20
for bounded subdomains Ω
0
with Ω
0
⊂ R
n
\ Γ.
A similar analysis can be given for the least squares method, because this is identical to
the Galerkin method for the operator A

A. Thus α has to be replaced by 2α, and then (5.1)
is always satisfied.
For collocation methods, the corresponding analysis is available for two-dimensional prob-
lems.
6 Singularities
In BEM one has to deal with singularities in two places: In the numerical computation
of integrals with singular kernels, and in the case of non-smooth boundaries or boundary
conditions, with singularities of the solution that cause large approximation errors.
6.1 Singular integrals
The numerical approximation of singular integrals has been studied since a long time, but the
results so far do not cover BEM completely. Thus in the preparation of every BEM application
program, this point seems to be a challenge to skill and phantasy. There is a general recipe,
however, namely to subtract the main singularity and compute the corresponding integrals
analytically if possible. For the remaining integrals one uses quadrature formulas that are, if
necessary, adapted to the still present weak singularities. The main singularities frequently
are contained in a kernel of convolutional structure. Especially in two-dimensional problems
the resulting translation invariance allows very useful simplifications.
If hypersingular integrals appear, they can sometimes be reduced to weakly singular
ones by partial integration. For example in R
2
, the bilinear form that defines the matrix
elements in the Galerkin method (2.16) for the first kind integral equation (2.10) with the
hypersingular integral operator D, is equal to the same bilinear form for the weakly singular
integral operator V , evaluated with the first derivatives of the respective trial functions. A
similar idea works also for the equation (4.3) in R
3
[18].
A review of the state of the analysis of numerical integration in BEM is given in [28].
6.2 Non-smooth boundaries
The convergence rates in (5.8) are limited by to numbers: The degree of the polynomials
chosen as a basis for the boundary element functions, and the Sobolev index describing the
regularity of the solution u. The first one is a question of algorithmic complexity.
The second one, the highest value of t such that u ∈ H
t
(Γ), is no problem for smooth
domains: From the regularity theory for elliptic pseudodifferential operators one knows u ∈
H
t
(Γ), as soon as f = Au ∈ H
t−2α
(Γ). Frequently the right hand sides are even in C

(Γ),
e.g. in scattering problems, so any t is allowed.
If corners and edges are present, then the solution has singularities there. Consider for
instance the screen scattering problem of section 4.1 or a crack problem in elasticity. There
it is known that the solution u behaves like ρ
1/2
and ∂
n
u like ρ
−1/2
near the edge, where ρ
denotes the distance to the edge. Therefore u is contained at most in H
1−ǫ
(Γ) for any ǫ > 0,
but not in H
1
(Γ). Then the Galerkin method for equation (4.3) has from (5.8) with α = 1 a
highest convergence rate of O(h
1−2ǫ
).
21
If one knows the precise shape of the singularities, one can design special boundary element
functions in order to obtain a better approximation than given by (5.5). Suppose it is known
that
u = βu
s
+u
0
, (6.1)
where u
s
is explicitly known, only the constant β and u
0
depend on the given data, and u
0
is regular: u
0
∈ H
t
(Γ). Then one can take trial functions of the form
˜ u =
˜
βu
s
+ ˜ u
0
, (6.2)
with the same u
s
and regular boundary elements ˜ u
0
.
Then for the best approximation, β =
˜
β, and hence
u − ˜ u
H
s
(Γ)
≤ ch
t−s
u
0

H
t
(Γ)
.
Thus the convergence rate is solely determined by the regular part u
0
.
Decompositions as in (6.1) are known for two-dimensional domains with corners [13],
where the singularities have the form
u
s
=
J

j=1
k
j

k=0
c
jk
ρ
α
j
(log ρ)
k
.
Here ρ is the distance to the corner points and α
j
are some complex numbers depending on
the corner angles and the “Mellin symbol” of the integral operator. They are given as zeros
of some analytic function, the multiplicities being k
j
+ 1. As an example, in the Dirichlet
and Neumann problem in potential theory, α
j
has the form
α
j
=

ω
i
+m,
where l and m are integers and ω
i
is the angle at the i-th corner point.
Singularities in three-dimensional problems are more complicated and are the subject of
a currently very active field of research. As an example, consider the screen scattering or
crack problem as above. There one has (6.1) with
u
s
= ρ
1/2
with ρ the distance to the edge, but now β = β(s) is a function of the parameter s on the
edge curve. Thus β itself has to be approximated by one-dimensional finite element functions
on this curve. One can show [23] that for smooth data one has u
0
∈ H
2−ǫ
(Γ) for any ǫ > 0.
Therefore from (5.8) one can obtain a highest convergence rate of O(h
3−2ǫ
) for the augmented
Galerkin procedure with trial functions (6.2).
7 Coupling of FEM and BEM
If the coefficients of the differential operator are not constant, then in general a fundamental
solution is not known and therefore pure BEM cannot be applied. Frequently, however,
the inhomogeneities are contained in a subdomain Ω
1
, whereas the system in the other part

2
of the domain Ω is described by a differential operator with constant coefficients whose
22
fundamental solution is known. For example, in problems on unbounded domains Ω the
inhomogeneities are frequently appearing only in a bounded domain Ω
1
.
In this situation, a coupling of FEM on Ω
1
and BEM on Ω
2
can be used. Such couplings
are popular in applications, see [9]. For the case of potential theory, also a mathematical
analysis including error estimates is available. Recently, more general equations and systems,
in particular those of linear elasticity, have been investigated, see [27, 26] for a survey.
Most coupling methods yield nonsymmetric matrices which is unpleasant for both the
application and the error analysis. In the case of elasticity theory in [26] for example, the
assumption has to be made that the ratio of the meshwidths in the domain Ω
1
for FEM
and on the coupling boundary for BEM tends either to zero or to infinity in order to prove
convergence.
Here I describe a symmetrical method which, in case of selfadjoint boundary value prob-
lems, yields symmetric matrices and allows a very simple error analysis. For simplicity, only
the Dirichlet problem for second order strongly elliptic systems will be treated, but the gen-
eral case of higher order operators and other boundary conditions as in section 3 presents no
principal difficulty. More precisely, the problem is as follows:
The domain Ω is decomposed into
Ω = Ω
1
∪ Γ
c
∪ Ω
2
,
where Γ
c
:= ∂Ω
1
∩∂Ω
2
is the coupling boundary. We write further Γ
j
:= ∂Ω
j

c
(j = 1, 2),
and allow Γ
1
or Γ
2
to be empty, as e.g. if Ω
1
and Ω
2
are the interior and exterior of a closed
boundary Γ
c
. On the other hand, also nonclosed Γ
1
, Γ
2
, and Γ
c
are allowed, as e.g. if Ω
1
and

2
are two adjacent rectangles.
We consider the differential equations
P
j
u
j
= f
j
in Ω
j
(j = 1, 2) (7.1)
with boundary conditions
u
j
= g
j
on Γ
j
(j = 1, 2) (7.2)
and the interface conditions
u
1
= u
2

ν
1
u
1
= ∂
ν
2
u
2
_
on Γ
c
. (7.3)
Here P
1
and P
2
are second order (matrix-valued) operators as in (3.6), and ∂
ν
1
and ∂
ν
2
are
the respective conormal derivatives, which may be different. The normal vector is exterior
to Ω
1
and interior to Ω
2
. The given data are f
j
on Ω
j
and g
j
on Γ
j
. For simplicity, we
assume that f
2
≡ 0, g
1
≡ 0 and that P
2
contains no first order differential operators. More
essentially, we also require that a fundamental solution G
2
for P
2
is known, so that in Ω
2
we
have the representation formula (3.10)
u
2
(x) =
_
∂Ω
2
_
(
¯

ν(y)
G(x, y)

)

v(y) −G(x, y)ϕ(y)
_
do(y), (7.4)
where v := u
2|∂Ω
2
, ϕ := ∂
ν
2
u
2|∂Ω
2
. Taking Cauchy data in (7.4), one finds the relations on
∂Ω
2
v = −Aϕ +Bv (7.5)
ϕ = Cϕ −Dv (7.6)
23
which define the integral operators A, B, C, D, constituting the Calder´on projector on Ω
2
.
In the case of potential theory, we have with the notation of section 2
_
−A B
C −D
_
=
_
−V
1
2
+K
1
2
−K

−D
_
.
On Ω
1
we have the first Green formula (3.7)
Φ

1
(u
1
, w) −
_
∂Ω
1
(∂
ν
1
u
1
)

wdo =
_

1
f

1
wdx (7.7)
for all w ∈ H
1
(Ω
1
). If we assume that u
1
satisfies the boundary condition (7.2) on Γ
1
and
the test function w satisfies the corresponding homogeneous boundary condition
w ∈ H
1
0
(Ω
1
) := {w ∈ H
1
(Ω
1
) | w

1
= 0}, (7.8)
then (7.7) is just the weak formulation of the boundary value problem on Ω
1
, if ∂
ν
1
u
1
is given
on Γ
c
.
Now the common coupling method [9] can be described as follows: Solve (numerically)
one of the equations (7.5), (7.6) for ϕ in terms of v and insert the resulting expression for

ν
2
u
2|Γc
= ∂
ν
1
u
1|Γc
in terms of u
1|Γc
into (7.7). Then (7.7) has a familiar variational form
to which FEM can be applied. This works particularly nice if a Green function G
2
for Ω
2
is
known. Then (7.6) is simply
ϕ = −Dv,
and inserting this into (7.7) gives
Φ

1
(u
1
, w) +
_
Γc
w

D(u
1|
Γc
)do =
_

1
f

1
wdx −
_
Γc
w

Dg
2
do. (7.9)
Here the bilinear form on the left hand side satisfies a G˚arding inequality in H
1
(Ω
1
) if P
1
and
P
2
are strongly elliptic, and it is symmetric and positive definite if P
1
and P
2
are in addition
selfadjoint. Thus one obtains immediately that every conforming Galerkin method for (7.9)
converges with optimal order. This method is successfully applied e.g. in plasma physics, see
[1].
If a Green function for Ω
2
is not known then one introduces
ϕ = ∂
ν
2
u
2|∂Ω
2
as an additional unknown. The common method then takes a weak formulation of the first
equation(7.5) on the boundary together with (7.7) and the coupling conditions (7.3) to form
a system which is discretized by approximating u
1
with finite elements in the domain Ω
1
and
ϕ with boundary elements on the boundary ∂Ω
2
. The resulting matrix is not symmetric, and
one does not have a G˚arding inequality except for scalar equations.
One obtains a symmetrical method if one adds to (7.7) weak forms of both equations (7.5)
and (7.6) on the boundary. Thus we define the following bilinear form
a(u, ϕ; w, ψ) := Φ

1
(u, w)
+
_
Γc
_
w

D(u
|Γc
) −w

Cϕ −ψ

u +ψ

B(u
|Γc
)
_
do (7.10)

_
∂Ω
2
ψ

Aϕdo
24
for u, w ∈ H
1
0
(Ω
1
); ϕ, ψ ∈ H

1
2
(∂Ω
2
). The given data f
1
∈ H
−1
(Ω
1
), g
2

˜
H
1
2

2
) (note
f
2
= 0, g
1
= 0) define the linear form
ℓ(w, ψ) :=
_

1
f

1
wdx +
_
Γ
2
ψ

g
2
do −
_
∂Ω
2
ψ

Bg
2
do −
_
Γc
w

Dg
2
do. (7.11)
We consider the variational equation for (u, ϕ) ∈ H
1
0
(Ω
1
) ×H

1
2
(∂Ω
2
)
a(u, ϕ; w, ψ) = ℓ(w, ψ) ∀(w, ψ) ∈ H
1
0
(Ω
1
) ×H

1
2
(∂Ω
2
). (7.12)
The corresponding Galerkin scheme is:
Find (u
h
, ϕ
h
) ∈ H
1
h
×H

1
2
h
such that
a(u
h
, ϕ
h
; w, ψ) = ℓ(w, ψ) ∀(w, ψ) ∈ H
1
h
×H

1
2
h
. (7.13)
We can then show the following theorem, and we give a proof here, because the method
is apparently new.
Theorem 7.1 Let P
1
and P
2
be strongly elliptic, let P
2
have a selfadjoint principal part,
and let the problem (7.1)–(7.3) be uniquely solvable. Then the weak solution of (7.12) exists
and defines a solution of (7.1)–(7.3), where u
1
= u and u
2
is given by (7.4). Furthermore
every Galerkin scheme (7.13) with approximating finite dimensional spaces H
1
h
⊂ H
1
0
(Ω
1
),
H

1
2
h
⊂ H

1
2
(∂Ω
2
) converges with optimal order. If P
1
and P
2
are selfadjoint, then so are
the bilinear form a in (7.12) and the stiffness matrix of the Galerkin scheme (7.13).
Proof. We assume everything to be real-valued, and we make a simplifying assumption
which in general is satisfied modulo compact perturbations:
P
2
is selfadjoint, and the Dirichlet problem in Ω
2
is solvable.
This implies that the operators A and D are selfadjoint and that the following lemma
holds (which is evident in potential and elasticity theory):
Lemma 7.2 Under the present assumptions, the adjoint B

of the operator B in (7.5) is
1 −C with C from (7.6).
We omit the proof which uses the first Green formula in Ω
2
.
Now assume that (7.12) holds. For w = 0 this is
_
∂Ω
2
ψ

{(B −1)v −Aϕ}do = 0 ∀ψ ∈ H

1
2
(∂Ω
2
),
where
v =
_
u on Γ
c
g
2
on Γ
2
.
Thus v and ϕ satisfy the relation (7.5) and hence are Cauchy data of the function u
2
defined
by (7.4). Hence also the relation (7.6) is satisfied.
From this we find
a(u, ϕ; w, ψ) −ℓ(w, ψ) = Φ

1
(u, w) +
_
Γc
_
w

D(u
Γc
) −w

(ϕ +Dv)
_
do

_

1
f

1
wdx +
_
Γc
w

Dg
2
do
= Φ

1
(u, w) −
_
Γc
w

ϕdo −
_

1
f

1
wdx.
25
Thus (7.7) holds for all w ∈ H
1
0
(Ω
1
), with ∂
ν
1
= ϕ on Γ
c
. Hence with u
1
= u we have found
a solution of (7.1)–(7.3).
The selfadjointness of the bilinear form a follows from that of Φ

1
and of the operators
A and D and from the lemma.
For the convergence of the Galerkin method we show Babuˇska’s inf-sup-condition. If
(u, φ) is in the trial space then (u, −φ) is in the test space, and we have with the lemma
a(u, ϕ; u, −ϕ) = Φ

1
(u, u) +
_
Γc
u

D(u
|Γc
)do +
_
∂Ω
2
ϕ

Aϕdo
≥ c
_
u
2
H
1
(Ω
1
)

2
H

1
2 (Ω
1
)
_
,
because under the present assumptions the bilinear form Φ

1
and the operators D and A are
selfadjoint and positive.
References
[1] R. Albanese, J. Blum, and O. De Barbieri. On the solution of the magnetic flux equation
in an infinite domain. Europhysics Conference Abstracts, 10 D:41–44, 1986.
[2] I. Babuˇska and K.Aziz. Survey lectures on the mathematical foundation of the finite
element method. In A. K. Aziz, editor, The Mathematical Foundation of the Finite Ele-
ment Method with Applications to Partial Differential Equations, pages 3–359, Academic
Press, New York, 1972.
[3] C. T. H. Baker. The Numerical Treatment of Integral Equations. Oxford University
Press, 1977.
[4] P. K. Banerjee et al., editors. Developments in Boundary Element Methods, Vols 1,2,3,
Applied Science Publishers, London, 1979 ff.
[5] C. A. Brebbia, editor. Boundary Element Methods, Proc. of the 8th Int. Conf. on BEM,
Como 1986, Springer-Verlag, Berlin, 1986.
[6] C. A. Brebbia, editor. Boundary Element Techniques in Computer-Aided Engineering,
Martinus Nijhoff Publishers, Dordrecht, 1984.
[7] C. A. Brebbia, editor. Progress in Boundary Element Methods, Vol. 1, Springer-Verlag,
Berlin, 1981.
[8] C. A. Brebbia, editor. Progress in Boundary Element Methods, Vols 1,2,3,4, Springer-
Verlag, Berlin, 1981 ff.
[9] C. A. Brebbia, J. C. F. Telles, and L. C. Wrobel. Boundary Element Techniques.
Springer-Verlag, Berlin, 1984.
[10] G. A. Chandler and I. G. Graham. Product integration-collocation methods for non-
compact integral operator equations. 1986. to appear.
[11] M. Costabel. Boundary Integral Operators on Lipschitz domains: Elementary Results.
Preprint 898, TH Darmstadt, 1985. to appear.
26
[12] M. Costabel. Starke Elliptizit¨ at von Randintegraloperatoren erster Art. Habilitations-
schrift, Technische Hochschule Darmstadt, 1984.
[13] M. Costabel and E. P. Stephan. Boundary integral equations for mixed boundary
value problems in polygonal domains and Galerkin approximation. In W. Fiszdon and
K. Wilma´ nski, editors, Mathematical Models and Methods in Mechanics 1981, Banach
Center Publications Vol. 15, pages 175–251, PWN – Polish Scientific Publishers, Warsaw,
1985.
[14] M. Costabel and E. P. Stephan. On the Convergence of Collocation Methods for Bound-
ary Integral Equations on Polygons. Preprint 982, TH Darmstadt, 1986. to appear in
Math. Comp.
[15] M. Costabel and W. L. Wendland. Strong Ellipticity of Boundary Integral Operators.
Preprint 889, TH Darmstadt, 1985. J. Reine Angew. Math. 372:34–63, 1986.
[16] J. Dieudonn´e. El´ements d’Analyse. Volume 8, Gauthier-Villars, Paris, 1978.
[17] J. Mackerle and T. Andersson. Boundary element software in engineering. Advances in
Eng. Software, 6:66–102, 1984.
[18] J. C. Nedelec. Integral equations with non integrable kernels. Integral Equations Oper.
Theory, 5:562–572, 1982.
[19] B. E. Petersen. Introduction to the Fourier Transform and Pseudodifferential Operators.
Pitman, Boston, 1983.
[20] H. Schippers. Multigrid methods for boundary integral equations. Numer. Math., 46:351–
363, 1985.
[21] E. Stephan and W. L. Wendland. Remarks to Galerkin and least squares methods with
finite elements for general elliptic problems. Manuscripta Geodaetica, 1:93–123, 1976.
[22] E. P. Stephan. Boundary Integral Equations for Mixed Boundary Value Problems, Screen
and Transmission Problems in R
3
. Habilitationsschrift, Technische Hochschule Darm-
stadt, 1984.
[23] E. P. Stephan. Boundary integral equations for screen problems in R
3
. Integral Equations
Oper. Theory, 9, 1986.
[24] W. L. Wendland. Bemerkungen zu Randelementmethoden und ihren mathematischen
und numerischen Aspekten. GAMM-Mitteilungen, 2:3–27, September 1986.
[25] W. L. Wendland. Boundary element methods and their asymptotic convergence. In
P. Filippi, editor, Theoretical Acoustics and Numerical Techniques, CISM Courses 277,
pages 135–216, Springer-Verlag, Wien, New York, 1983.
[26] W. L. Wendland. On Asymptotic Error Estimates for Combined BEM and FEM. Lecture
Notes, Udine, September 1986.
[27] W. L. Wendland. On asymptotic error estimates for the combined boundary and fi-
nite element method. In R. P. Shaw et al., editors, Innovative Numerical Methods in
Engineering, pages 55–70, Springer-Verlag, Berlin, 1986.
27
[28] W. L. Wendland. On some mathematical aspects of boundary element methods for ellip-
tic problems. In J. Whiteman, editor, Mathematics of Finite Elements and Applications
V, pages 193–227, Academic Press, London, 1985.
[29] J. Whiteman, editor. Mathematics of Finite Elements and Applications V, Academic
Press, London, 1985.
28

order of the solution in the domain. There are difficulties, however, if the solution has to be evaluated close to, but not on the boundary.

1.3

Difficult parts

Some main difficulties with BEM are the following: 1. Boundary integral equations require the explicit knowledge of a fundamental solution of the differential equation. This is available only for linear partial differential equations with constant or some specifically variable coefficients. Problems with inhomogeneities or nonlinear differential equations are in general not accessible by pure BEM. Sometimes, however, a coupling of FEM and BEM proves to be useful, see section 7. 2. For a given boundary value problem there exist different boundary integral equations and to each of them several numerical approximation methods. Thus every BEM application requires that several choices be made. To evaluate the different possibilities, one needs a lot of mathematical analysis. Although the analysis of BEM has been a field of active research in the past decade, it is by no means complete. Thus there exist no error estimates for several methods that are widely used. From a mathematical point of view, these methods, which include very popular ones for which computer codes are available, are in an experimental state, and there might exist problems of reliability. 3. The reason for the difficulty of the mathematical analysis is that boundary integral equations frequently are not ordinary Fredholm integral equations of the second kind. The classical theory of integral equations and their numerical solution concentrates on second kind integral equations with regular kernel, however. Boundary integral equations may be of the first kind, and the kernels are in general singular. If the singularities are not integrable, one has to regularize the integrals which are then defined in a distributional sense. The theoretical framework for such integral equations is the theory of pseudodifferential operators. This theory was developed 20 years ago and is now a classical part of Mathematical Analysis [16, 19], but it is still not very popular within Applied Mathematics. 4. If the boundary is not smooth but has corners and edges, then the solution of the boundary value problem has singularities at the boundary. This happens also if the boundary conditions are discontinuous, e.g. in mixed boundary value problems. BEM clearly have to treat these singularities more directly than FEM. Because the precise shape of the singularities frequently contains important information, e.g. stress intensity factors in fracture mechanics, this is a positive aspect of BEM. But besides practical problems with the numerical treatment of these singularities, non-smooth domains also present theoretical difficulties. These have so far been satisfactorily resolved only for two-dimensional problems. The analysis of BEM for three-dimensional domains with corners and edges is still in a rather incomplete stage.

1.4

Literature

The usefulness of BEM in engineering problems is documented in a literature that amounts to several thousand pages every year. I will not attempt to give an overview of the wide range of possible applications. One can find such an overview e.g. in one of the proceedings 2

and heat conduction. Discretization methods and their convergence are described in section 5. Section 3 contains a general method for deriving boundary integral equations for general elliptic boundary value problems. A review of BEM computer codes can be found in [17]. Dr. The present notes could only be prepared on the basis of several survey articles and lectures by Prof. 9. W. 6.of the annual conference on Boundary Elements [5] or in the books[4] or [8]. Introductions to the theory and applications of BEM can e. Section 7 was inspired by the stimulating atmosphere and discussions at the summer school in Lausanne. be found in the books [7. 2 2. Acknowledgement.g.1 The structure of a BEM application A scheme A typical application of BEM consists of the following parts: • Mathematical model • Representation formula • Boundary integral equation • Boundary elements • Discrete equations • Solution of the linear system • Interpretation 2. whose organizers and participants I am indebted to. The final section 7 contains a short description of the coupling of FEM and BEM in terms of a (new) symmetric method. 29] or in[25.2 Discussion of the scheme For the following discussion of these parts. L. and section 6 contains some remarks on the treatment of singularities in BEM. 1. 24]. elasticity theory. the Dirichlet problem for Laplace’s equation will be used as an illustration. for which also a convergence proof is given.5 Contents In the following sections I shall first describe the general structure of a BEM application and illustrate it with the simple example of the Dirichlet problem from potential theory. Wendland. 3 . Section 4 describes boundary integral equations for examples from scattering theory.

2 The next step is to represent the solution of the partial differential equation in the domain by means of boundary potentials. and u is supposed to be a harmonic function regular both in Ω and the exterior domain Rn \ Ω.1 The mathematical model is a boundary value problem for a partial differential equation.1) (2. the Laplace-Beltrami equation on the sphere. y)[∂n u(y)]Γ do(y) (2.g.3) can take various different forms depending on assumptions on u in Rn \ Ω. Betti’s formula for elasticity theory. (2.2.2) It is convenient to have a homogenenous differential equation and inhomogeneous boundary data.g. having i.g. y ∈ R2 ) for n = 2 for n = 3. different boundary values on both sides of Γ. and the Stratton-Chu formula for electrodynamics. e. y)u(y)do(y) + Γ γ(x. e. y)[u(y)]Γ do(y) − Γ γ(x. Supposing u|Rn \Ω ≡ 0.2. on the boundary. so one has to make a choice here. Here ∂n denotes the normal derivative taken with respect to the exterior normal on Γ. In the example one has the fundamental solution 1 γ(x.2. Green’s third identity for potential theory. 4 . y ∈ R3 ) Fundamental solutions are known for equations with constant coefficients where they can be computed by Fourier transformation. y) = − 2π log |x − y| (x.3) for x ∈ Rn \ Γ. do is the surface measure on Γ. Most common are the following: 1. Thus these layers have a “direct” interpretation in terms of the solution. 2. the representation formula (2. In potential theory we have u(x) = Γ ∂n(y) γ(x. The densities of the simple and double layer on the boundary are the normal derivative of the solution and the solution. y)∂n u(y)do(y) (x ∈ Ω) (2. Thus u is represented as the sum of a double layer and a simple layer potential. An example is the Dirichlet problem ∆u = 0 in a domain Ω ⊂ Rn u = g on the boundary Γ := ∂Ω. whose jump across Γ is denoted by [·]Γ : [v(x)]Γ := v|Rn \Ω (x) − v|Ω (x) (x ∈ Γ). If considered in Ω alone.4) The BEM derived from this is called “method of Green’s formula” or “direct method”. Such representation formulas are well known for the classical boundary value problems of mathematical physics. respectively. Each of these forms gives rise to a different BEM. A necessity is to know explicitly a fundamental solution of the differential equation. and for some elliptic equations with analytic coefficients. γ(x. one obtains u(x) = − Γ ∂n(y) γ(x. y) = 1 4π|x−y| (x.

(2.8) 2 is a Fredholm integral equation of the second kind.6) into the boundary condition (2. Therefore. 2. y)do(y) (x ∈ Ω) (2. The representation formula (2. In our example. Sometimes also a suitable linear combination of simple and double layer potentials will be useful. If one takes boundary values in the representation formula. y)do(y) =: − v(x) + Kv(x) 2 Γ on Γ. and 3. the integral equation arising from (2.2. also choose between different possibilities. y)ψ(y)do(y) = g(x) on Γ.Hilbert in 1870–1910. The simple layer potential Γ γ(x. I. (2. however.4) admits two ways of approaching the boundary: If one considers just u|Γ . one obtains boundary integral equations. Here the integral on Γ has to be understood in the principal value sense. one will insert the representation formulas (2. . For the double layer potential (2. Supposing [∂n u]Γ ≡ 0. thus (2. (In fact it was. 3.7) This is a Fredholm integral equation of the first kind with a weakly singular kernel. together with its adjoint. One has to take into account the well-known “jump relations” which arise from the fact that the gradient of γ has a strong singularity and is not integrable on Γ. that in the presence of edges and corners.2.5) yields the integral equation V ψ(x) := Γ γ(x. 1 (− + K ) v = g (2.3 The representation formulas give the solution u in the interior of the domain Ω. above.6).Neumann. Here one can i. and D. the starting point of the whole theory of integral equations as developed by C.6) there holds 1 u(x) = − v(x) + 2 1 v(y)∂n(y) γ(x.8) has a strongly singular kernel and is no longer a Fredholm integral equation.Fredholm.) Note.5) or (2. y)ψ(y)do(y) (x ∈ Ω) (2. y)ψ(y)do(y) is continuous across Γ.5) with an unknown density ψ.6) with an unknown density v.g. Supposing [u]Γ ≡ 0. but it turns out that for smooth boundaries the kernel is continuous. these are as follows: In the “indirect methods” 2. one obtains 1 u = u − Ku + V (∂n u) 2 5 on Γ. one obtains a double layer representation u(x) = Γ v(y)∂n γ(x. one obtains a simple layer representation u(x) = Γ γ(x.2).

For instance. y) ≡ 0 for y ∈ Γ. 2 Γ Here.9) one obtains 1 (2. In each case. y)u(y)do(y) is the normal derivative of the double layer potential. with the representation formula (2.10) one obtains 1 (2. hence the direct method in this case coincides with the method of double layer representation. (2. It is an easy exercise to write down the corresponding integral equations for the Neumann problem where ∂n u is given on Γ. γ can be any fundamental solution.10) 2 Both relations together express simply the fact that the two functions u|Γ and ∂n u|Γ .11).2) can be used to determine ∂n u|Γ and thus. (2. Any one of the four integral equations (2.9) 2 1 D u = ( − K ′ )(∂n u). From (2. on Γ ∂n(y) γ(x. or (2.12) ( − K ′ )(∂n u) = D g.But one can also take ∂n u|Γ and observe the jump relations: The normal derivative of the double layer potential is continuous across Γ. and K ′ appearing in (2.8). and even for mixed boundary value problems where on different parts of Γ different boundary conditions are satisfied.7). y)ψ(y)do(y) = ψ(x) + 2 Γ 1 ∂n(x) γ(x. the integrals may simplify considerably. the “Cauchy data” of u. K. 6 . A method to derive such relations for more general elliptic equations is described in section 3. 2 an integral equation of the first kind with the same kernel as in (2. From (2. D.12) can equally well be used for the numerical solution of the Dirichlet problem. Two final remarks on the variety of boundary integral equations: • In representation formulas like (2. to solve the Dirichlet problem.10) together with the boundary condition (2.9). and for x ∈ Γ ∂n(x) 1 γ(x. K ′ is the adjoint operator of K above. y)ψ(y)do(y) =: ψ(x) + K ′ ψ(x). if γ is the Green function of Ω (which is known e. Now either one of the equations (2.7). only the four operators V . The “integral operator” D has a hypersingular kernel. 2 an integral equation of the second kind with the adjoint kernel of (2. (2.11) V (∂n u) = ( + K ) g. then γ(x. if the boundary and the density are smooth. (2. are derived from a function u that satisfies the differential equation (2.8). Thus if Γ has a simple geometry that allows to find a fundamental solution which vanishes on parts of Γ. for balls or half spaces).9) are involved.1) in Ω.g.4).4). We have now found the two relations on Γ 1 ( + K ) u = V (∂n u) (2. Thus one obtains 1 ∂n u = Du + ∂n u + K ′ ∂n u 2 where Du(x) := −∂n(x) Γ on Γ. and also for the Robin problem where a linear combination of u and ∂n u is given on Γ.

14) with basis functions {µj |j = 1. one knows which singularities appear in the solution. 2. then these methods for obtaining a linear system for h the unknown coefficients {γj |j = 1. N }.2.13) is satisfied in these points. . e. and the least squares method. N } ⊂ Γ of collocation points and requires that the equation (2. but for the proper BEM these trial functions are finite element functions on the boundary. in the presence of corners and edges. If we have a boundary integral equation Au = f and we seek an approximate solution N on Γ (2. using singular boundary elements. a finite dimensional function space on Γ. .2.13) uh (x) = j=1 γj µj (x) h (2.m s finite element functions. Here it is useful to have asymptotic error estimates in order to make the right choice. One assumes that Γ can be decomposed into a finite number of subsets each of which has a regular parameter representation by some parameter domain in Rn−1 . . .e.4 For the numerical solution of the boundary integral equations one needs functions on the boundary that depend on a finite set of parameters. one has no choice but has to work with (2. Using the parameter representations. There are three main methods to generate this system of discrete equations: Collocation.g. a Sh family in the sense of Babuˇka and Aziz [2]. This 7 . This rules out all integral equations of the second kind and leaves for the Dirichlet problem the integral equation of the first kind with the operator V and for the Neumann problem the integral equation of the first kind with the operator D (see [22. then one can build them into the space of trial functions as in FEM.3) directly. one uses the approximate coordinate representations. N } are the following: 1. Then one chooses regular partitions of the parameter domains and correspending k. Such situations occur in crack problems or in screen scattering problems. If. . One can choose globally defined functions like spherical harmonics or polynomials. This will be discussed in section 5. i. . Galerkin’s method. one transfers these functions to the boundary. .e. For the collocation method one chooses a suitable set {xj |j = 1. i. . .• If Γ is an open surface which does not separate Rn into two disjoint domains. . The order of the polynomial basis functions in these approximations of course has to be chosen in correspondence to the order of the Sobolev spaces in which the solution of the boundary integral equation is looked for and also to the kind of discretization scheme that is chosen. either by using a suitable nonuniform partition or by augmenting the space of trial functions by some singular functions. 23]). If the parameter representation is not explicitly known but the boundary itself has also to be approximated. . . the boundary elements. 2.5 The finitely many parameters determining the approximate solution are computed from finitely many linear equations.

f ) h h h (k = 1.16) j=1 where the brackets (·. f ) h h h (k = 1. in R3 . (f.g. This may be the reason why collocation is most frequently used in the applications. e. A careful choice of the boundary integral operator can yield a positive definite selfadjoint operator A or at least a positive definite principal part. even positive definite. (2. N } as a basis h for the test functions: N (Aµk . Here the collocation method is the simplest one because it involves only one integral whereas Galerkin’s method needs two and the least squares method three integrations.15) j=1 2. . then the coefficient matrix will also be positive selfadjoint.13) is uniquely solvable. Aµj ) γj = (Aµk . . ·) denote the L2 (Γ) inner product. o see [3]). . i. the first kind operators V and D from potential theory are selfadjoint and. . . 10]) and in higher dimensions only for Fredholm integral equations of the second kind (for which. . integrates over Γ and equates the integrals. .13) for uh with test functions from a finite dimensional function space. there exist also a lot of other numerical methods.e. . 8 . the well-known Nystr¨m method. . In the least squares method one minimizes Auh −f in the L2 (Γ) norm. . . • for collocation methods only in 2 dimensions for strongly elliptic operators A and smooth domains (apart from some very special recent results [14.17) j=1 The coefficient matrix in these methods has to be computed using numerical integration. The coefficient matrix for the collocation method is never symmetric. (2. if (2. . 3. • for Galerkin methods under the somewhat stronger condition of strong ellipticity for the operator A. The resulting equations are those for a Galerkin (-Petrov) method with {Aµk |k = 1. on the other hand. . test and trial functions are the same. Convergence proofs and asymptotic error estimates are available • for least squares methods under the general condition of ellipticity for the operator A.gives the system N (Aµj )(xk ) γj = f (xk ) h (k = 1. however: The coefficient matrix in the least squares method is always symmetric and. In the proper Galerkin method. hence this gives the system N (µk . Aµj ) γj = (µk . N ). . In the Galerkin method. It has also some drawbacks. For the Galerkin method one multiplies (2. . positive definite. . g) := Γ f (x)g(x)do(x). N ) (2. N ). For instance.

If these Cauchy data have to be determined from an indirect method. Note that the matrices are not sparse.2. for elliptic and parabolic problems. The order of this quadrature formula should be chosen in accordance with the asymptotic order of convergence for the chosen discretization method. GP u = u = P Gu for all compactly supported u. so the linear systems solvers developed for FEM are not useful in BEM. 2.6 In comparison with the computation of the matrix elements. standard linear systems solvers should suffice. Thus this integration has a smoothing effect and suppresses high-frequency errors. Assume that a fundamental solution G of P is known. then another integral operator with singular kernel has to be applied. is the last step in BEM. The interpolation formula (2. y)f (y)dy with a kernel G(x. 2.e.2. 9 . the solution of the linear system of equations is usually relatively uncomplicated. In addition. i.14) gives the solution of the boundary integral equation. In some specific examples. Here only a regular integral has to be evaluated.7 The postprocessing. while the less singular remainder is computed with the use of a suitable quadrature formula. this may be already the desired information about the boundary values of the solution of the boundary value problem. 3 Derivation of boundary integral equations Let P be a linear elliptic differential operator of order 2m with smooth coefficients on Rn . depends very much on the specific problem at hand. however. and the usual recipe is to split off the main singularities and try to evaluate the corresponding integrals analytically. if one exploits symmetry and positive definiteness where available. the representation formula may be differentiated and used to compute the derivatives of the solution in the domain. that is • G is a two-sided inverse of P on functions (or distributions) that vanish outside a compact set.The integrations in the computation of the coefficient matrix have to be done numerically. multigrid methods have been analyzed and successfully applied [20]. then the representation formula which was used to derive the boundary integral equation gives the answer. A common problem is the evaluation of singular integrals. y) that is weakly singular for x = y. If the solution of the boundary value problem in the interior of Ω is desired. singular only if the points of integration and of evaluation coincide. In the direct method. • G is an integral operator Gf (x) = Rn G(x. because the kernel defined by the fundamental solution is. In general. the interpretation of the computed data. This requires another choice which.

Now we choose a piecewise smooth function u as above such that uj ∈ C ∞ (Ωj ). . One has Pu = f in Rn \ Γ with f ∈ C ∞ (Rn \ Γ). where H = 1 div n is the mean curvature of Γ and ∆0 is the Laplace-Beltrami operator on 2 Γ.3) Here Pj are differential operators of order 2m − j whose restrictions to Γ contain only tangential derivatives. ∂n u)|Γ in Ω ⊂ Rn on Γ = ∂Ω. It consists of multiple layers of the k form v ⊗ ∂n δΓ defined by k < v ⊗ ∂n δΓ . and the higher normal derivatives are defined using this vector field.1) (3.In this section we describe the method of the Calder´n projector to obtain boundary integral o equations for boundary value problems fo the following form Pu = f Rγu = g Here 2 2m−1 ⊤ γu := (γ0 u. (3. . . The result can be found in detail in [16]. and R is a (m × 2m) matrix of differential operators on Γ. . γ2m−1 u)⊤ : = (u. the Laplace operator P = −∆ is decomposed as 2 −∆ = −∆0 − 2H∂n − ∂n .4) 10 . . 3. Therefore the distribution P u − f vanishes on Rn \ Γ. ϕ > := Γ ′k v ∂n ϕ do for ϕ ∈ C ∞ (Γ). Introducing suitable coordinates. . . Thus the boundary data g are a m-component vector function. For example.1 Second Green formula Let Ω1 := Ω and Ω2 := Rn \ Ω.2) are the “Cauchy data” of u. ′k k Here ∂n is the formal transpose of the differential operator ∂n . It is the distributional version of the second Green formula: 2m−1 2m−1−k k=0 l=0 Pu = f + k Pk+l+1 [γl u]Γ ⊗ ∂n δΓ . Here one assumes that in a neighborhood of Γ there exists a unit vector field n that is the normal field exterior to Ω1 on Γ. one can see that P has a decomposition 2m P = j=0 j Pj ∂n (3. . (3. For a function u on Rn with u|Ωj = uj we write [γk u]Γ = k γk u2 −γk u1 if the Cauchy data γk uj = ∂n (uj )Γ exist. ∂n u. ∂n u. We apply P to u in the distributional sense. hence ′k k ∂n = (−1)k ∂n + lower order terms.

and P ′ is the formal transpose of P : < P u..4) and (3. u1 = v. ∂ν v := j=1 j. ϕ > = < u.5) are equivalent.If we specialize to u2 ≡ 0. but it is the distributional form (3.5) Here the superscript ⊤ denotes transposition. In the case of a second order operator (we abbreviate ∂j := ∂/∂xj ) n n P =− j. and c may be matrix-valued.k=1 ∂j ajk ∂k + j=1 bj ∂j + c (3. (3.k=1 j=1 and the conormal derivative (bj ∂j u)⊤ v + (cu)⊤ v  dx ∂ν u := j. Then.k=1 nj ajk ∂k u. From (3. because we admit throughout matrix-valued operators P and vector-valued functions u. jk In the case of elasticity.6) one has a first Green formula (P u)⊤ v dx = ΦΩ (u.8) with ∂ν u := ∂ν u − ˜ nj bj u . v) := Ω n n (ajk ∂k u)⊤ ∂j v + n j. Here again the coefficients ajk . 11 .4) to a test function ϕ. the equations of linear elasticity can be treated in this way. we obtain the more familiar equivalent formula 2m−1 2m−1−k (v ⊤ P ′ ϕ − (P v)⊤ ϕ)dx = Ω k=0 l=0 Γ ′k (Pk+l+1 γl v)⊤ (∂n ϕ)do. (3.4) is ′ −∆u = f − [γ1 u]Γ ⊗ δΓ + [γ0 u]Γ ⊗ ∂n δΓ . The formulas (3. and apply (3.8) is Somigliana’s identity. P ′ ϕ > = Ω v ⊤ P ′ ϕdx ∞ for ϕ ∈ C0 (Rn ). In the example of the Laplace operator.4) that is the starting point for the representation formula and boundary integral equations.7) with ΦΩ (u.k=1 nj a⊤ ∂k . for example. (3.7) follows the second Green formula (u⊤ P ′ v − (P u)⊤ v)dx = Ω Γ n ((∂ν u)⊤ v − u⊤ ∂ν v)do ˜ n (3. bj . v) −   Ω Γ (∂ν u)⊤ v do  (3. where then ∂ν is the traction operator. v etc.

one can allow corners and edges on Γ. y)ϕ(y) do(y).3 C= 1 2 −K D 1 2 V + K′ .2. this is just (2.6) one obtains from (3. o As a pseudodifferential operator. One can show that C is a projection operator: CCv = Cv for all v. I will not give the definition of a Sobolev space H s (Γ) for arbitrary s ∈ R here. In the case of the Laplace operator. The operator C is a (2m × 2m) matrix of pseudodifferential operators on the boundary. For the second order operator (3. 3.9).3) if f = 0. Thus if u is piecewise smooth as above and vanishes outside a compact set. For x ∈ Γ this means Kk ϕ(x) = Γ ′k ∂n(y) G(x. y)∂ν u(y) do(y). whereas for the general case (3.3. then there holds the representation formula 2m−1 2m−1−k k=0 l=0 u = Gf + Kk (Pk+l+1 [γl u]Γ ) (3. and they contain additional terms coming from the corners. The complementary projector 1 − C is the Calder´n projector for the complementary domain.9) the boundary has to be smooth.11) γ(Kk (Pk+l+1 γl u1 ))|Ω .8) u(x) = Gf (x) + Γ (∂ν(y) G(x. we find an identity for the Cauchy data γu1 . if we set u2 ≡ 0: γu1 = γGf1 + C(γu1 ) with C(γu1 ) := − k=0 l=0 2m−1 2m−1−k (3. k 2 1 2 12 .10) For this case of a second order operator.2 Representation formula From the second Green formula in the distributional form. even an arbitrary Lipschitz domain [12]. but recall that for integer s. In the example of the Laplace operator. one obtains the representation formula simply by applying the inverse G of the differential operator P . C has well-known continuity properties in Sobolev spaces on Γ. y)⊤ )⊤ [u(y)]Γ − G(x. it is in the notation of section 2. For the bi-Laplace operator P = ∆2 in R2 also representation formulas are known in domains with corners.9) with the multiple layer potential operators Kk defined by k Kk ϕ := G(ϕ ⊗ ∂n δΓ ) for ϕ ∈ C ∞ (Γ). ˜ (3.3 Calder´n projector o Taking traces on Γ in (3. H s (Γ) is the space of all functions on Γ with finite norm v :=   H s (Γ) k≤s Γ |D v| do .

There are several ways of making a quadratic system out of this rectangular one.k=0 which also consist of differential operators. is to use the boundary condition to eliminate m unknowns and find a m × m system for the remaining m unknown functions.11) together with the boundary condition (3. It is known that C = (Cjk )2m−1 where Cjk is a pseudodifferential operator of order j.4 Boundary integral equations The relation (3. . 3.e. the solution of the boundary value problem in Ω is given by the representation formula (3.1) to the boundary: A 2m-vector function v satisfies v = γGf + Cv if and only if v equals γu with some solution u of (3.13) v 13 . γ2m−1 u. The ellipticity of this system. i. It is a system of 3m equations for the 2m unknowns γ0 u. Only those with nonpositive order are decent integral operators. The dual space of the Hilbert space H s (Γ) in the natural L2 (Γ) duality is H −s (Γ). . Thus the boundary “integral” operators Cjk have orders ranging from −2m + 1 to 2m − 1. the invertibility of its symbol as a pseudodifferential operator. Thus (3. The usual procedure. is equivalent to the Shapiro-Lopatinski ellipticity condition for the boundary value problem (3. This would lead to a 2m × 2m system.11) is a faithful translation of the differential equation (3. .1)–(3. where order zero operators can contain strongly singular kernels and thus integrals existing only in the principal value sense. (3.9): g u = Gf − KPM −1 . . This means that Cjk maps H s (Γ) into H s−j+k (Γ) continuously for any s.12) v In this case. however. One way would be to treat the 3m × 2m system directly by least squares methods.2) is an equivalent formulation of the boundary value problem by means of functions on the boundary. The operators with positive order have hypersingular kernels like the operator D which has order +1.k=0 j − k. and the Cauchy data γu are known if v is known and g is given by (3. This procedure uses a second set S of m boundary operators with the property that the (2m × 2m) matrix M := R S of tangential differential operators on Γ has an inverse ˇ M −1 = (Mjk )2m−1 j.2): g γu = M −1 .1).where Dk v is the vector of all k-th order tangential derivatives of v.2). But all of these highly nonclassical “integral” operators do appear in applications and they can be used in BEM. Then the m-component vector function v = Sγu contains the unknown Cauchy data. (3.

we have the 2m × 2m system g v = M γGf − M γKPM −1 g v . (3. every solution v ∈ m−1 H −m+µj + 2 +s (Γ) j=0 of one of the boundary integral equations. Now either of the two sets of m equations (3. 14 1 .15). which can of course be also written in the form g v ˜ = M γGf + C g v with the “modified Calder´n projector” o ˜ C := M CM −1 . lower right hand “(m × m) corner” of the (2m × 2m) matrix C.14) or (3.14) and (3. Then every solution u ∈ H m+s (Ω) of the boundary value problem (3.14) or (3. . defines by the representation formula (3. K2m−1 ) and the matrix of differential operators P := (Pk+l+1 )2m−1 k.12) a solution of (3.14) (3.13). We write this system finally in the form g = RγGf + RCM −1 v = SγGf + SCM −1 or Av = g − RCM −1 (1 − B)v = SCM −1 where Av = RCM −1 and Bv = SCM −1 0 v = −SγKPM −1 0 v 0 v = −RγKPM −1 0 v g 0 g 0 − RγGf (3.2) defines by 1 (3.1 Let f ∈ H s−m (Ω) ∩ L2 (Ω) and g ∈ m−1 H m−µj − 2 +s (Γ) be given where s is j=0 an arbitrary real number and µj are the orders of the rows of the boundary differential operator R.13) a solution u ∈ H m+s (Ω) of the boundary value problem.15) is equivalent to the original boundary value problem under suitable conditions which can be found in [15]: Theorem 3.15) g v g v + SγGf ˜ are the upper resp. Taking the “modified Cauchy data” M γu in (3.l=0 with Pj := 0 for j > 2m. . (3. . .15). Conversely.where we introduced the vector of integral operators K := (K0 .1).

For the second kind operator 1 − B in (3. e. The total field u = uin + us as well as the incident (plane or spherical) wave uin and the scattered field us satisfy the Helmholtz equation.1 Examples of boundary integral equations Acoustic scattering We consider a screen Γ ⊂ R3 . one can show [15] that the operator A in (3. The same result of strong ellipticity for the first kind operator A in (3.e. βk are arbitrarily given numbers. (3. zero. vj and λk are certain given functions related to the cokernel and kernel of A.Under some conditions on the strong ellipticity of the boundary value problem (3.16) is stable and converges [21].15) strong ellipticity is known in special cases. They are not satisfied for some boundary value problems from electrodynamics. 4 4. by adding a finite number of rows and columns. If u satisfies homogeneous Neumann conditions on Γ. Fredholm’s alternative holds for the boundary integral equation (3. i.14). i. i.e. as the fourth order scalar biharmonic equation. the system can be converted into a uniquely solvable one of the form l Av + j=1 0 ωj vj = ψ (3. we have for the scattered field us the Neumann problem (∆ + k2 ) us = 0 ∂n us = g 15 in Rn \ Γ on Γ (4.2) . whether formulated as the second order system of the Navier equations or. . Thus. . mainly on smooth domains. Such an addition of rows and columns is necessary in elasticity theory to eliminate the rigid body motions. an open surface. Strong ellipticity of the operator A implies 1. i. l). .2). but in some cases on non-smooth domains it fails. In the case of smooth 2-dimensional domains.e. Γ λ⊤ v do = βk k 0 where ψ is the right hand side in (3.14) holds for second order systems even on arbitrary Lipschitz domains Ω. The enlarged system is also equivalent to the boundary value problem.1) (4.e.14) is also strongly elliptic. and ωj are numbers that have to be determined from the boundary integral equation. 3. These conditions are satisfied for the usual boundary value problems for the Laplace and Helmholtz equations and for the equations of linear elasticity theory. Every Galerkin procedure for (3.g. in the presence of corners and edges [11].14). the homogeneous equation has a finite number l of linear independent solutions and there exist l linear independent solvability conditions. The asymptotic order of convergence is quasioptimal.1). it is the same order as for the best possible approximation of the solution by means of the boundary element trial functions. . 2. also collocation methods are stable and converge quasioptimally.16) (k = 1. through the Airy stress function. and also sometimes in potential and scattering theory.

3) with γ as in section 2. the fundamental solution γ is the outgoing one for the Helmholtz equation: γ(x.with g = −∂n uin . bijectively onto H ˜ One has to choose the boundary elements in such a way that the solution v ∈ H 1/2 (Γ) is approximated as good as possible. y) = The solution v has a direct interpretation as v = [us ]Γ . x)ϕ(y))do(y) for x ∈ Ω. y)t(y) − T(y. 4π|x − y| 4. There exist methods using singular boundary elements (see section 6) that allow an order of O(h3 ) [22]. x) = (T(y) (F(y.3) is convergent. The operator D is (for real wave numbers k) a positive selfadjoint operator on L2 (Γ) ˜ ˜ with domain of definition H 1/2 (Γ) which is the dual space of H −1/2 (Γ). y) + λ + µ (x − y)(x − y)⊤ λ + 3µ |x − y|n (n = 2. t = T (u) with the traction operator T (u) := λ n(x) div u + 2µ∂n u + µn ∧ curl u. and T(y. eik|x−y| .2 Elasticity theory For the system of partial differential equations µ∆u + (λ + µ) grad div u = 0 one has the representation formula u(x) = Γ in Ω (4. Thus every Galerkin method for the equation (4.3) with the hypersingular integral operator D introduced in section 2.5) where ϕ = u|Γ . 16 . This allows an asymptotic order of convergence of at most O(h) for regular finite elements. Here. y) = λ + 3µ 2µ(λ + 2µ) γ(x.1. (4. and D maps H 1/2 (Γ) −1/2 (Γ).4) (F(x.4) F(x. of course. F is the fundamental solution of (4.2. The direct method as well as the method of double layer potential representation (they coincide due to [∂n us ]Γ = 0 ) both lead to the first kind integral equation D v = −g (4. x)))⊤ . In general the solution will have unbounded derivatives near the screen edge.

(4. we obtain the first kind equation T(x) ( T(y. one obtains on Γ [28] u(x) = 1 ϕ(x) − 2 Γ (T(y. M(x) =  0 1 0 0 0 1 0 x2 −x1  1 t(x) − 2 T(x. BEASY and other programs). It is interesting to note.8) or u=ϕ given on Γ. and the integrals have to be understood in the Cauchy principal value sense. x)ϕ(y)do(y)) + 1 t(x) + 2 T(x.10) uniquely solvable. x)ϕ(y)do(y)) + M(x)ω = Γ ⊤ Γ 1 0 0 −x2 0 x3   x1 −x3 0 . to the best of the author’s knowledge. Here the unknown is u|Γ = ϕ. The equation (4.10) is not a Fredholm integral equation of the second kind. y)t(y))do(y) Γ (4.10) is convergent. y)t(y)do(y) (4.6) gives for n = 3 [28] 1 ϕ(x) + 2 T(y. For smooth domains. and hence any Galerkin method converges for (4. Frequently the equations are discretized using a collocation method (BETSY. BEM for three-dimensional elasticity are of course very important in engineering applications. y)⊤ t(y)do(y). thus making the system (4.6) (4. y)⊤ t(y)do(y) Γ M(y) ϕ(y)do(y) = 0. Therefore any Galerkin method for (4. The equations (4. x)ϕ(y)do(y) + M(x)ω = Γ Γ F(x.8). If we take the equation (4.Using the jump relations for elastic potentials.7).10) is strongly singular.9) Let us consider the first one (4.10)  M(y)⊤ ϕ(y)do(y) = 0 Γ with the matrix Here the extra 6 unknowns ω and the extra 6 side conditions are added in order to remove the rigid body motions. Γ Here the kernels on the right hand side are strongly singular. This gives a theoretical argument in favor of Galerkin methods.10) is a strongly elliptic pseudodifferential operator of order zero. however. But one can show that even for any Lipschitz domain the operator defines a positive definite hermitian bilinear form. that all attempts to prove convergence for these methods have been. until now unsuccessful.7) T (u)(x) = −T(x) ( T(y. The integral operator in (4. x)ϕ(y) + F(x. (4.6). (4.7) immediately give boundary integral equations for the two boundary value problems T (u) = t given on Γ (4. 17 .11) Here the kernel is hypersingular.11) even if the domain has corners and edges. Thus (4. it has been shown that the integral operator in (4. however.

y. t). s)dyds + Ω G(x. t) = 0 Ω t 0 G(x. Taking into account that G = 0 for s < t. t. T )u(x. s) =  0   n − 4(s−t) |x−y|2 for s > t for s < t. t.12) is given by (4π(s − t))− 2 e G(x. 0)u(x. If we consider the initial value problem u(x. 0)dx (4. t)−u(x. s)]Γ − G(x. The heat equation is Lu := ∂t u − ∆u = 0 in Q = Ω × (0. t. satisfies Lu = f in Rn+1 \ ∂Q.13) Convolution with G gives the representation formula for (x. T ) ⊂ Rn+1 . t. t. y.3 Heat conduction Stationary heat conduction processes are governed by Laplace’s equation which was treated above. y. 0) − G(x. t) = Q G(x. A fundamental solution of (4. We apply L in the distributional sense to a function u which is smooth in Q and in Rn+1 \ Q. • With the help of a time-dependent fundamental solution one can perform BEM directly on the boundary of the space-time cylinder Q. (4. t)]Γ ⊗δΓ (x. s)f (y.15) . 0)δ(t)+[u(x.12) There exist several BEM for initial-boundary value problems for parabolic equations: • One can choose a time-discretization scheme that requires at each time-step the solution of an elliptic boundary value problem in the space domain Ω. t)−[∂n u(x. s)]Γ − G(x. t)]Γ ⊗∂n δΓ (x. t. this is equivalent to t u(x. y. y. s)[u(y. y. y. T )δ(t−T )+u(x. The latter method is described here along the lines of the general scheme described in section 3. and has jumps across ∂Q = Ω0 ∪ ΩT ∪ (Γ × [0. s)f (y. Therefore here we concentrate on transient processes.4. y.14) + Γ (∂n(y) G(x. The elliptic boundary value problems are then solved by one of the BEM described above. t. t. y. t. T ]) (Ωt := Ω × {t}). s)[∂n u(y. The result is the distributional version of the second Green formula: ′ Lu(x. 0)u(x. t. s)]Γ )do(y)ds. t) ∈ ∂Q: u(x. 0) = u0 (x) 18 (4. t) = f (x. s)[u(y. s)]Γ )do(y)ds. s)[∂n u(y. T ))dy + 0 Γ (∂n(y) G(x. s)dyds + Ω T (G(x. • Also the application of Laplace transformation leads to a (parameter-dependent) elliptic boundary value problem. (4. y.

(5. ∀v ∈ Hh . as in section 2. which. (5.14) the first term is absent and the second is known. 0). For the Galerkin procedure. (For general initial values u0 . [28] contains tables showing relations which have to be satisfied by the approximation orders of the coordinate representation of Γ and the approximation orders in the numerical integrations in order to get highest convergence rates. ranging from abstract approximation schemes for general pseudodifferential operators to fully discretized boundary integral equations including numerical integration methods. α = 1 . We leave this as an exercise. 5 Convergence The state of the art of asymptotic error estimates for BEM is described in several detailed articles by W. If for all u ∈ H α (Γ).1) with α = − 1 resp. Wendland. the respective second kind integral equations have weakly singular kernels and can therefore be treated by various numerical methods. due to the smoothing nature of the kernel G(x.for the equation (4. may be discretized relatively crudely.3) (0 < h < h0 ).) The remaining double and simple layer potentials resemble those for the Laplace equation. In particular. contained and quoted in the books mentioned in section 1. L. They are generalizations of those for FEM. and are particularly easy to state for Galerkin methods.4. (5. Let A : H α (Γ) → H −α (Γ) be a bijective and continuous linear operator for some real α. Assume that A satisfies a G˚ arding inequality: There exists a compact operator T : H α (Γ) → H −α (Γ) and a γ > 0 such that Re(v. inf{ u − u ˜ ˜ H α (Γ) | u ∈ Hh } −→ 0 19 as h → 0. so we consider only these here. In the case of smooth Ω. Here are some remarks on the basic rules of error estimates for BEM.1) is equivalent to the strong ellipticity of A. even on domains with corners and 2 2 edges. one has a sequence of finite-dimensional subspaces Hh ⊂ H α (Γ) and one looks for uh ∈ Hh satisfying (v.3. (A + T )v) ≥ γ v 2 H α (Γ) ∀v ∈ H α (Γ). two boundary integral equations for both the Dirichlet and the Neumann boundary conditions. The first kind boundary integral operators for the Dirichlet respectively Neumann problem for second order strongly elliptic systems ( the operators V resp. Thus one may take traces of u and ∂n u on Γ and obtain. t. and therefore little is known about the first kind equations.12) then on the right hand side of (4. The operators are no ordinary pseudodifferential operators. L being parabolic. D in potential theory) satisfy (5. [25] and [28] give a complete survey of the known results. Auh ) = (v.2. it has to be computed by a domain integral. then (5. f ) where f = Au ∈ H −α (Γ).1) If A is a pseudodifferential operator of order 2α. however. y.2) .

The highest order of convergence is obtained for s = 2α − k and t = k: u − uh H 2α−k (Γ) H s (Γ) = O(ht−s ) if u ∈ H t (Γ). one needs α < m.5) where c does not depend on u or h. In particular. m measures the interelement differentiability and k the degree of the polynomials. from FEM for shells.g. the L2 -condition number of the discrete equations blows up like O(h−2|α| ).8) Here it is an advantage to have negative α. It depends on the size of the linear system and the machine precision. which of the two effects. Note that for (5.6) to hold. This is just the conformity condition (5. This together with (5. the theoretical Galerkin error or the round-off error magnified by a big condition number. Tychonov regularization. however.4) Now. becomes dominant. If the size of the system is chosen so that both effects balance. ˜ (5. the boundary element trial spaces Hh as described in 2. have the following approximation property: There are numbers k and m such that for any s < t with s < m and t ≤ k and any u ∈ H t (Γ) there exists an approximating u ∈ Hh with ˜ u−u ˜ H s (Γ) ≤ cht−s u H t (Γ) (5.g. (5. if α ≤ 0. one can use discontinuous trial functions.8) on the boundary is inherited by the solution of the boundary value problem away from the boundary: One has for the multiple layer potentials Kk from 3. When the approximate solution of the boundary integral equations is inserted into the representation formula.1) implies stability and quasioptimal convergence for the Galerkin procedure: There exists h1 > 0 such that for all h < h1 . (5. then the approximation is of the same order as for approximation methods specifically created for ill-posed problems.4) gives a convergence order u − uh H α (Γ) = O(ht−α ) if u ∈ H t (Γ). (5. (5.3) defines a unique uh ∈ Hh . This reflects the fact that equations with such operators are traditionally called ill-posed. one can estimate the error also in norms different from the “energy norm” H α (Γ): u − uh where s is restricted by 2α − k ≤ s < m.4 above.2: Kk ϕ H s (Ω0 ) ≤ c ϕ H t (Γ) for any s and t 20 . Then. and there is a constant c independent of u and h with uh and u − uh H α (Γ) H α (Γ) ≤c u H α (Γ) ≤ c inf{ u − u ˜ H α (Γ) | u ∈ Hh }.6) Using an inverse assumption and duality arguments. then the highest convergence rate (5. e.2.7) = O(h2(k−α) ) if u ∈ H k (Γ).2).then (5. as is well known e.

For collocation methods.10) with the hypersingular integral operator D. A similar analysis can be given for the least squares method. however. 6.16) for the first kind integral equation (2. Especially in two-dimensional problems the resulting translation invariance allows very useful simplifications. For the remaining integrals one uses quadrature formulas that are. The main singularities frequently are contained in a kernel of convolutional structure. is no problem for smooth domains: From the regularity theory for elliptic pseudodifferential operators one knows u ∈ H t (Γ). e. If corners and edges are present. with singularities of the solution that cause large approximation errors. they can sometimes be reduced to weakly singular ones by partial integration. Frequently the right hand sides are even in C ∞ (Γ). is equal to the same bilinear form for the weakly singular integral operator V . the corresponding analysis is available for two-dimensional problems. Consider for instance the screen scattering problem of section 4. 6 Singularities In BEM one has to deal with singularities in two places: In the numerical computation of integrals with singular kernels. where ρ denotes the distance to the edge.1) is always satisfied. A similar idea works also for the equation (4. 21 . adapted to the still present weak singularities. Therefore u is contained at most in H 1−ǫ (Γ) for any ǫ > 0. the bilinear form that defines the matrix elements in the Galerkin method (2.8) are limited by to numbers: The degree of the polynomials chosen as a basis for the boundary element functions. If hypersingular integrals appear.g. in scattering problems. and in the case of non-smooth boundaries or boundary conditions. The first one is a question of algorithmic complexity. There is a general recipe. because this is identical to the Galerkin method for the operator A∗ A. For example in R2 . this point seems to be a challenge to skill and phantasy. but the results so far do not cover BEM completely. There it is known that the solution u behaves like ρ1/2 and ∂n u like ρ−1/2 near the edge. as soon as f = Au ∈ H t−2α (Γ).for bounded subdomains Ω0 with Ω0 ⊂ Rn \ Γ. evaluated with the first derivatives of the respective trial functions.3) in R3 [18].1 or a crack problem in elasticity.1 Singular integrals The numerical approximation of singular integrals has been studied since a long time. the highest value of t such that u ∈ H t (Γ). Thus in the preparation of every BEM application program. so any t is allowed.3) has from (5. Thus α has to be replaced by 2α. and then (5.2 Non-smooth boundaries The convergence rates in (5. 6. and the Sobolev index describing the regularity of the solution u. but not in H 1 (Γ). A review of the state of the analysis of numerical integration in BEM is given in [28]. Then the Galerkin method for equation (4. then the solution has singularities there. if necessary. The second one.8) with α = 1 a highest convergence rate of O(h1−2ǫ ). namely to subtract the main singularity and compute the corresponding integrals analytically if possible.

Singularities in three-dimensional problems are more complicated and are the subject of a currently very active field of research. They are given as zeros of some analytic function. Decompositions as in (6.5). ˜ ˜ ˜ with the same us and regular boundary elements u0 . ˜ ˜ Then for the best approximation. in the Dirichlet and Neumann problem in potential theory. one can design special boundary element functions in order to obtain a better approximation than given by (5. As an example. but now β = β(s) is a function of the parameter s on the edge curve.2). Here ρ is the distance to the corner points and αj are some complex numbers depending on the corner angles and the “Mellin symbol” of the integral operator.1) with us = ρ1/2 with ρ the distance to the edge. Frequently. Therefore from (5. then in general a fundamental solution is not known and therefore pure BEM cannot be applied. Thus the convergence rate is solely determined by the regular part u0 . and hence u−u ˜ H s (Γ) (6.If one knows the precise shape of the singularities. As an example. consider the screen scattering or crack problem as above. however. and u0 is regular: u0 ∈ H t (Γ). There one has (6.2) ≤ cht−s u0 H t (Γ) . (6. Then one can take trial functions of the form u = βus + u0 . β = β. whereas the system in the other part Ω2 of the domain Ω is described by a differential operator with constant coefficients whose 22 . 7 Coupling of FEM and BEM If the coefficients of the differential operator are not constant. Suppose it is known that u = βus + u0 . the inhomogeneities are contained in a subdomain Ω1 . the multiplicities being kj + 1.1) are known for two-dimensional domains with corners [13]. αj has the form αj = lπ + m. Thus β itself has to be approximated by one-dimensional finite element functions on this curve. where the singularities have the form J kj u = j=1 k=0 s cjk ραj (log ρ)k . ωi where l and m are integers and ωi is the angle at the i-th corner point.8) one can obtain a highest convergence rate of O(h3−2ǫ ) for the augmented Galerkin procedure with trial functions (6. One can show [23] that for smooth data one has u0 ∈ H 2−ǫ (Γ) for any ǫ > 0.1) where us is explicitly known. only the constant β and u0 depend on the given data.

as e. g1 ≡ 0 and that P2 contains no first order differential operators.g. (7. also a mathematical analysis including error estimates is available. Γ2 .6) ϕ = . see [27. For simplicity. For example. we assume that f2 ≡ 0.10) u2 (x) = ∂Ω2 (∂ν(y) G(x.1) Here P1 and P2 are second order (matrix-valued) operators as in (3. We consider the differential equations Pj uj = fj with boundary conditions uj = gj and the interface conditions u1 = u2 ∂ν1 u1 = ∂ν2 u2 on Γc .4) where v := u2|∂Ω2 . in case of selfadjoint boundary value problems. Most coupling methods yield nonsymmetric matrices which is unpleasant for both the application and the error analysis.6).4).g. one finds the relations on ∂Ω2 v = −Aϕ + Bv Cϕ − Dv 23 (7. ϕ := ∂ν2 u2|∂Ω2 . see [9]. y)⊤ )⊤ v(y) − G(x. more general equations and systems. More precisely.fundamental solution is known. also nonclosed Γ1 . 2) (7.2) in Ωj (j = 1. On the other hand. in problems on unbounded domains Ω the inhomogeneities are frequently appearing only in a bounded domain Ω1 . 2). Here I describe a symmetrical method which. we also require that a fundamental solution G2 for P2 is known. if Ω1 and Ω2 are the interior and exterior of a closed boundary Γc . and ∂ν1 and ∂ν2 are the respective conormal derivatives. (7. More essentially. y)ϕ(y) do(y). For the case of potential theory. yields symmetric matrices and allows a very simple error analysis. Such couplings are popular in applications. In this situation. Recently. For simplicity. and allow Γ1 or Γ2 to be empty. Taking Cauchy data in (7. 2) (7. which may be different. only the Dirichlet problem for second order strongly elliptic systems will be treated.5) (7. but the general case of higher order operators and other boundary conditions as in section 3 presents no principal difficulty. in particular those of linear elasticity. the problem is as follows: The domain Ω is decomposed into Ω = Ω 1 ∪ Γc ∪ Ω 2 . 26] for a survey. We write further Γj := ∂Ωj \ Γc (j = 1. and Γc are allowed. where Γc := ∂Ω1 ∩ ∂Ω2 is the coupling boundary.3) on Γj (j = 1. a coupling of FEM on Ω1 and BEM on Ω2 can be used. In the case of elasticity theory in [26] for example. if Ω1 and Ω2 are two adjacent rectangles. the assumption has to be made that the ratio of the meshwidths in the domain Ω1 for FEM and on the coupling boundary for BEM tends either to zero or to infinity in order to prove convergence. as e. so that in Ω2 we have the representation formula (3. The normal vector is exterior to Ω1 and interior to Ω2 . The given data are fj on Ωj and gj on Γj . have been investigated.

5) and (7.10) − ∂Ω2 24 . One obtains a symmetrical method if one adds to (7.7) gives ΦΩ1 (u1 .7) and the coupling conditions (7. and one does not have a G˚ arding inequality except for scalar equations.9) converges with optimal order. constituting the Calder´n projector on Ω2 . Then (7. C. This method is successfully applied e.g. Now the common coupling method [9] can be described as follows: Solve (numerically) one of the equations (7.7) is just the weak formulation of the boundary value problem on Ω1 .6) on the boundary.which define the integral operators A. and it is symmetric and positive definite if P1 and P2 are in addition selfadjoint. w) + Γc w⊤ D(u|Γc ) − w⊤ Cϕ − ψ ⊤ u + ψ ⊤ B(u|Γc ) do ψ ⊤ Aϕ do (7. B.2) on Γ1 and the test function w satisfies the corresponding homogeneous boundary condition 1 w ∈ H0 (Ω1 ) := {w ∈ H 1 (Ω1 ) | w|Γ1 = 0}. The resulting matrix is not symmetric.7) ΦΩ1 (u1 . (7. w) + Γc w⊤ D(u1|Γc )do = Ω1 ⊤ f1 w dx − Γc w⊤ Dg2 do. w. and inserting this into (7. see [1]. (7. If a Green function for Ω2 is not known then one introduces ϕ = ∂ν2 u2|∂Ω2 as an additional unknown. This works particularly nice if a Green function G2 for Ω2 is known.6) is simply ϕ = −Dv.9) Here the bilinear form on the left hand side satisfies a G˚ arding inequality in H 1 (Ω1 ) if P1 and P2 are strongly elliptic. if ∂ν1 u1 is given on Γc .3) to form a system which is discretized by approximating u1 with finite elements in the domain Ω1 and ϕ with boundary elements on the boundary ∂Ω2 . On Ω1 we have the first Green formula (3. (7. Then (7. we have with the notation of section 2 −A B C −D = 1 2 −V − K′ 1 2 +K −D .7) for all w ∈ H 1 (Ω1 ). D.5) on the boundary together with (7. ϕ. in plasma physics.7) has a familiar variational form to which FEM can be applied. Thus one obtains immediately that every conforming Galerkin method for (7.8) then (7.6) for ϕ in terms of v and insert the resulting expression for ∂ν2 u2|Γc = ∂ν1 u1|Γc in terms of u1|Γc into (7. If we assume that u1 satisfies the boundary condition (7.7) weak forms of both equations (7. w) − ∂Ω1 (∂ν1 u1 )⊤ w do = Ω1 ⊤ f1 w dx (7.7). Thus we define the following bilinear form a(u. The common method then takes a weak formulation of the first equation(7. ψ) := ΦΩ1 (u.5). o In the case of potential theory.

w ∈ H0 (Ω1 ). From this we find a(u. the adjoint B ′ of the operator B in (7. Proof. ϕ.3).12) holds. ϕ. ψ) ∈ Hh × Hh 2 . For w = 0 this is ψ ⊤ {(B − 1)v − Aϕ}do = 0 ∂Ω2 −1 1 ∀ψ ∈ H − 2 (∂Ω2 ). This implies that the operators A and D are selfadjoint and that the following lemma holds (which is evident in potential and elasticity theory): Lemma 7. Hence also the relation (7. w. 1 (7.6) is satisfied.13) with approximating finite dimensional spaces Hh ⊂ H0 (Ω1 ). (7.13) We can then show the following theorem. w.5) and hence are Cauchy data of the function u2 defined by (7. ψ) = ΦΩ1 (u. w) − 25 w⊤ ϕ do − Γc . We assume everything to be real-valued. The given data f1 ∈ H −1 (Ω1 ). Furthermore 1 1 every Galerkin scheme (7. ψ ∈ H − 2 (∂Ω2 ). and we make a simplifying assumption which in general is satisfied modulo compact perturbations: P2 is selfadjoint. ϕ. Γc Γ2 . where u1 = u and u2 is given by (7.1)–(7. ψ) = ℓ(w. w) + − = Ω1 Γc w⊤ D(uΓc ) − w⊤ (ϕ + Dv) do Γc ⊤ f1 w dx + w⊤ Dg2 do Ω1 ⊤ f1 w dx.12) exists and defines a solution of (7. ψ) (7. Now assume that (7. because the method is apparently new. ψ) The corresponding Galerkin scheme is: 1 Find (uh .4).11) 1 We consider the variational equation for (u. Hh 2 ⊂ H − 2 (∂Ω2 ) converges with optimal order.3) be uniquely solvable. w. ΦΩ1 (u.2 Under the present assumptions. then so are the bilinear form a in (7. ψ) ∈ H0 (Ω1 ) × H − 2 (∂Ω2 ).1)–(7.5) is 1 − C with C from (7.6).4). −1 a(uh . ϕ) ∈ H0 (Ω1 ) × H − 2 (∂Ω2 ) 1 a(u. and the Dirichlet problem in Ω2 is solvable. ϕh . If P1 and P2 are selfadjoint. and let the problem (7. We omit the proof which uses the first Green formula in Ω2 . let P2 have a selfadjoint principal part. ψ) − ℓ(w. g1 = 0) define the linear form ℓ(w. ϕh ) ∈ Hh × Hh 1 −2 1 ∀(w.12) such that 1 ∀(w.12) and the stiffness matrix of the Galerkin scheme (7. ψ) := Ω1 ⊤ f1 w dx + Γ2 ψ ⊤ g2 do − ∂Ω2 ψ ⊤ Bg2 do − Γc w⊤ Dg2 do.1 1 ˜1 for u. g2 ∈ H 2 (Γ2 ) (note f2 = 0. ψ) = ℓ(w.13). Then the weak solution of (7. 1 where v = u g2 on on Thus v and ϕ satisfy the relation (7. and we give a proof here.1 Let P1 and P2 be strongly elliptic. Theorem 7.

Telles. Oxford University Press. pages 3–359. If s (u. Berlin. Aziz.2. because under the present assumptions the bilinear form ΦΩ1 and the operators D and A are selfadjoint and positive. [9] C.1 Thus (7. 1986. Blum.. De Barbieri. On the solution of the magnetic flux equation in an infinite domain. 26 .3). u. [2] I. Vols 1. Babuˇka and K. Boundary Element Methods. Springer-Verlag. Vol. G. 1972. [7] C. Baker. K. Graham. Proc. References [1] R. Dordrecht. Wrobel. [11] M. H. A. of the 8th Int. Boundary Element Techniques. Vols 1. [5] C. Albanese.4.7) holds for all w ∈ H0 (Ω1 ).3. 10 D:41–44. Survey lectures on the mathematical foundation of the finite s element method. For the convergence of the Galerkin method we show Babuˇka’s inf-sup-condition. and O. to appear. TH Darmstadt. K. A. 1986. SpringerVerlag. Costabel. 1981. Berlin.2. Hence with u1 = u we have found a solution of (7. Brebbia. φ) is in the trial space then (u. editor.1)–(7. 1979 ff. Brebbia. Conf. In A. Product integration-collocation methods for noncompact integral operator equations. [4] P. A. New York. Berlin.3. F. editor. J. [6] C. Developments in Boundary Element Methods. Brebbia. The selfadjointness of the bilinear form a follows from that of ΦΩ1 and of the operators A and D and from the lemma. J. T. Banerjee et al. A. [3] C. Preprint 898. 1984. The Numerical Treatment of Integral Equations. Brebbia. Boundary Integral Operators on Lipschitz domains: Elementary Results. and we have with the lemma a(u. 1981 ff. −ϕ) = ΦΩ1 (u. Chandler and I. editor. [10] G. Martinus Nijhoff Publishers. 1977. 1. Como 1986. Academic Press. Berlin.Aziz. u) + ≥ c u 2 H 1 (Ω1 ) Γc u⊤ D(u|Γc )do + 2 1 H − 2 (Ω1 ) ϕ⊤ Aϕ do ∂Ω2 + ϕ . editors. Springer-Verlag. −φ) is in the test space. and L. editor. Springer-Verlag. ϕ. Brebbia. editor. 1985. 1986. The Mathematical Foundation of the Finite Element Method with Applications to Partial Differential Equations. C. on BEM. Europhysics Conference Abstracts. A. C. [8] C. Applied Science Publishers. London. Progress in Boundary Element Methods. A. 1984. to appear. Progress in Boundary Element Methods. with ∂ν1 = ϕ on Γc . Boundary Element Techniques in Computer-Aided Engineering.

Manuscripta Geodaetica. 1986. L. Costabel. Springer-Verlag. [13] M. Wendland.. Boundary element methods and their asymptotic convergence. Stephan. Math. Mathematical Models and Methods in Mechanics 1981. TH Darmstadt. [23] E. [22] E. Stephan and W. editors. Advances in Eng. 1986. Boundary integral equations for screen problems in R3 . 2:3–27. L. 372:34–63. P. [25] W. Wilma´ski. Innovative Numerical Methods in Engineering. 1985. 1984. Springer-Verlag. 9. 1986. Preprint 982. Remarks to Galerkin and least squares methods with finite elements for general elliptic problems. P. pages 55–70. [18] J. Costabel and E. Volume 8. In P. 1984. 1976. P. On asymptotic error estimates for the combined boundary and finite element method. Udine. Schippers. Strong Ellipticity of Boundary Integral Operators. Lecture Notes. Boundary Integral Equations for Mixed Boundary Value Problems. 1:93–123. J. CISM Courses 277. Wendland. In R. Boston. Habilitationsa schrift. Math. E. Banach n Center Publications Vol. pages 135–216. 5:562–572. [21] E. Nedelec. Reine Angew. Integral Equations Oper. Berlin. Boundary element software in engineering. editor. [16] J. PWN – Polish Scientific Publishers. Technische Hochschule Darmstadt. L. Comp. Technische Hochschule Darmstadt. P. 46:351– 363. TH Darmstadt. September 1986. Dieudonn´. 1984. Numer. Bemerkungen zu Randelementmethoden und ihren mathematischen und numerischen Aspekten. New York. 1978. Costabel and E. e e [17] J. Wien. [14] M. Wendland. Gauthier-Villars. [24] W. Multigrid methods for boundary integral equations. Warsaw. Wendland. Paris. September 1986. Introduction to the Fourier Transform and Pseudodifferential Operators. Boundary integral equations for mixed boundary value problems in polygonal domains and Galerkin approximation. 15. Software. Wendland. 1983. to appear in Math. Fiszdon and K. On the Convergence of Collocation Methods for Boundary Integral Equations on Polygons. Shaw et al.[12] M. Pitman. P. editors. Theoretical Acoustics and Numerical Techniques. Andersson. Screen and Transmission Problems in R3 . L. GAMM-Mitteilungen. Habilitationsschrift. Filippi. El´ments d’Analyse. In W.. L. Theory. 27 . Starke Elliptizit¨t von Randintegraloperatoren erster Art. Stephan. Costabel and W. [15] M. C. L. Integral equations with non integrable kernels. 1982. On Asymptotic Error Estimates for Combined BEM and FEM. Stephan. 1985. 1983. Integral Equations Oper. [26] W. pages 175–251. [20] H. 6:66–102. Stephan. [27] W. 1985. [19] B. Mackerle and T. Petersen. 1986. Wendland. Theory. Preprint 889.

Academic Press. editor. 1985. L.[28] W. Mathematics of Finite Elements and Applications V. On some mathematical aspects of boundary element methods for elliptic problems. Whiteman. London. Whiteman. London. In J. editor. 28 . Mathematics of Finite Elements and Applications V. Wendland. 1985. [29] J. Academic Press. pages 193–227.

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.