Integral Equations
MATHEMATICAL AND ANALYTICAL
TECHNIQUES WITH APPLICATIONS
TO ENGINEERING
Series Editor
Alan Jeffrey
The importance of mathematics in the study of problems arising from the real world,
and the increasing success with which it has been used to model situations ranging
from the purely deterministic to the stochastic, in all areas of todays Physical
Sciences and Engineering, is well established. The progress in applicable mathematics
has been brought about by the extension and development of many important
analytical approaches and techniques, in areas both old and new, frequently aided
by the use of computers without which the solution of realistic problems in modern
Physical Sciences and Engineering would otherwise have been impossible. The
purpose of the series is to make available authoritative, up to date, and selfcontained
accounts of some of the most important and useful of these analytical approaches
and techniques. Each volume in the series will provide a detailed introduction to a
specific subject area of current importance, and then will go beyond this by reviewing
recent contributions, thereby serving as a valuable reference source.
Series Titles:
THE FAST SOLUTION OF BOUNDARY INTEGRAL EQUATIONS
Sergej Rjasanow & Olaf Steinbach, ISBN 9780387340418
THEORY OF STOCHASTIC DIFFERENTIAL EQUATIONS WITH JUMPS AND
APPLICATIONS
Rong Situ, ISBN 9780387250830
METHODS FOR CONSTRUCTING EXACT SOLUTIONS OF PARTIAL
DIFFERENTIAL EQUATIONS
S.V. Meleshko, ISBN 9780387250601
INVERSE PROBLEMS
Alexander G. Ramm, ISBN 9780387231952
SINGULAR PERTURBATION THEORY
Robin S. Johnson, ISBN 9780387232003
INVERSE PROBLEMS IN ELECTRIC CIRCUITS AND ELECTROMAGNETICS
N.V. Korovkin, ISBN 9780387335247
The Fast Solution of Boundary
Integral Equations
Sergej Rjasanow
Universitt des Saarlandes
Olaf Steinbach
Technische Universitt Graz
Sergej Rjasanow
Fachrichtung 6.1 Mathematik
Universitt des Saarlandes
Postfach 151150
D66041 Saarbrcken
GERMANY
Olaf Steinbach
Institut fr Numerische Mathematik
Technische Universitt Graz
Steyrergasse 30
A8010 Graz
AUSTRIA
9 8 7 6 5 4 3 2 1
springer.com
Preface
dence of the solution on the given boundary data can be deduced. Chapter
2 is concerned with boundary element methods, especially with the Galerkin
method. The discrete version of the boundary integral equations from Chapter
1 and their variational formulations lead to systems of linear equations with
dierent matrices. The entries of these matrices are explicitly derived for all
integral operators involved. Chapter 3 describes the Adaptive Cross Approx
imation of dense matrices and provides, in addition to the theory, some rst
numerical examples. The largest part of the book, Chapter 4, contains some
results of numerical experiments. First, the Laplace equation is considered,
where we study Dirichlet, Neumann, and mixed boundary value problems as
well as an inhomogeneous interface problem. Then, two mixed boundary value
problems of linear elastostatics will be presented, and, nally, many examples
for the Helmholtz equation are described. We consider again Dirichlet and
Neumann, interior and exterior boundary value problems as well as multifre
quency analysis. Many auxiliary results are collected in three appendices.
The chapters are relatively independent of one another. Necessary nota
tions and formulas are not only crossreferred to other chapters but usually
repeated at the appropriate places.
In 2003, Prof. Allan Jerey approached us with the idea to write a book
about fast solutions of boundary integral equations. It has been delightful to
write this book and we are also very thankful for his providing the opportunity
to get this book published.
We would like to thank our colleagues from the BEM community for many
useful discussions and suggestions. We are grateful to our home institutions,
the University of Saarland in Saarbrucken and the Technical University in
Graz, for providing an excellent scientic environment and nancial funding
to our research.
We appreciate the help of Jurgen Rachor, who read the manuscript and
made valuable comments and corrections. Furthermore, the authors would
very much like to express their appreciation to Richard Grzibovski for his
help in performing numerical tests.
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
1
Boundary Integral Equations
3
2
u(x) = 2 u(x) = 0 for x R .
3
(1.1)
i=1
xi
This equation is used for the modelling of, for example, the stationary heat
transfer, of electrostatic potentials, and of ideal uids.
In (1.1), R3 is a bounded, multiply or simply connected domain with
a Lipschitz boundary = .
Multiplying the partial dierential equation (1.1) with a test function v,
integrating over , and applying integration by parts, this gives Greens rst
formula
(u(y)) v(y)dy = a(u, v) 1int u(y)0int v(y)dsy (1.2)
Here, n(y) is the outer normal vector dened for almost all y .
From Greens rst formula (1.2) and by the use of the symmetry of the
bilinear form a(, ), we deduce Greens second formula
v(y) u(y)dy + 1int v(y)0int u(y)dsy (1.3)
= u(y) v(y)dy + 1int u(y)0int v(y)dsy .
the solution of the Laplace equation (1.1) is given by the representation for
mula
int int u (x, y) int u(y)ds
u(x) = u (x, y)1 u(y)dsy 1,y 0 y (1.6)
which denes a continuous map from a given density function w on the bound
ary to a harmonic function V w in the domain . In particular,
V : H 1/2 ( ) H 1 ()
V = 0int V
with the following mapping properties, see for example [24, 71, 105].
4 1 Boundary Integral Equations
is bounded with
V w, w
cV1 w2H 1/2 ( ) for all w H 1/2 ( ).
for x , which again denes a continuous map from a given density func
tion v on the boundary to a harmonic function W v in the domain . In
particular,
W : H 1/2 ( ) H 1 ()
is continuous and W v H 1 () is a weak solution of the Laplace equation
(1.1) for any v H 1/2 ( ). Applying the interior trace operator
0int : H 1 () H 1/2 ( ) ,
is bounded with
and
1 1
(x) = lim dsy for x .
0 4 2
y:yx=
for x , or,
1 1 int 1 int 1 (x y, n(y)) int
1 u(y)dsy = 0 u(x) + 0 u(y)dsy .
4 x y 2 4 x y3
Instead of the interior trace operator 0int , we may also apply the interior
conormal derivative operator 1int to the representation formula (1.6). To do
6 1 Boundary Integral Equations
so, we rst need to investigate the interior conormal derivatives of the single
and double layer potentials V w and W v, which are both harmonic in . Then,
1int : H 1 (, ) H 1/2 ( ),
where
H 1 (, ) = 1 () .
v H 1 () : v H
is bounded with
int V
1int V wH 1/2 ( ) c21 wH 1/2 ( ) for all w H 1/2 ( ).
In particular, we have
1 1
1int V w, v
= w, v
+ K w, v
= w, v
+ w, Kv
2 2
for all v H 1/2 ( ).
In the same way as for the single layer potential V w, we now consider the
interior conormal derivate of the double layer potential W v.
is bounded with
DvH 1/2 ( ) cD
2 vH 1/2 ( ) for all v H 1/2 ( ),
Dv, v
cD
1 vH 1/2 ( )
2
for all v H 1/2 ( ).
where
curl u(x) = n(x) x u
(x) for x
is some (locally dened) extension of u into
is the surface curl operator and u
the neighbourhood of .
does not exist. Therefore the boundary integral operator D is called hyper
singular operator and it requires some appropriate regularisation procedure.
Since u(x) = 1 for x is a solution of the Laplace equation u(x) = 0,
the representation formula (1.6) reads for this special choice
1,yint u (
x, y)dsy = 1 for x .
Thus, we have
x int u (
1,y .
x, y)dsy = 0 for x
Then,
8 1 Boundary Integral Equations
(Dv)(x) = 1int (W v)(x) = lim n(x), x (W v)(
x)
, x
x x
= lim n(x), x 1,y x, y)v(y)dsy
int u (
, x
x x
= lim n(x), x 1,y x, y) v(y) v(x) dsy
int u (
, x
x x
(Dv)(x) =
1 (x y, n(y))(x y, n(x)) (n(x), n(y))
3 v(y) v(x) dsy .
4 x y5 x y3
Applying now the interior conormal derivative operator 1int to the represen
tation formula (1.6),
or,
1 (y x, n(y))(y x, n(x)) (n(x), n(y))
3
4 x y5 x y3
int int 1 int (y x, n(x)) int
0 u(y) 0 u(x) dsy = 1 u(x) 1 u(y)dsy .
2 x y3
1.1 Laplace Equation 9
Combining (1.8) and (1.10), we can write both boundary integral equations
by the use of the Calderon projector as
0int u 2I K
1
V 0int u
= 1
. (1.11)
1int u D 2I + K 1int u
From the boundary integral equations of the Calderon projector (1.11) one
can derive some important properties of boundary integral operators, i.e.
1 1
V K = KV, V D = I K I +K .
2 2
SteklovPoincare Operator
Since the single layer potential V is H 1/2 ( )elliptic and therefore invertible,
we obtain from the rst equation in (1.11) the Dirichlet to Neumann map
1
1int u(x) = V 1 I + K 0int u(x) = (S int 0int u)(x) for x (1.12)
2
which denes the SteklovPoincare operator
1int u(x) =
1 1
D+ I + K V 1 I + K 0int u(x) = (S int 0int u)(x)
2 2
with the symmetric representation of the SteklovPoincare operator
1 1
S int = D + I + K V 1 I + K . (1.13)
2 2
Note that also the representation (1.12) of the SteklovPoincare operator is
symmetric. However, due to Lemma 1.1 and Lemma 1.4, we conclude from
the symmetric representation (1.13)
1 1
S int v, v = Dv, v + V 1 I + K v, I +K v
2 2
Dv, v cD
1 vH 1/2 ( )
2
We rst consider the interior Dirichlet boundary value problem for the Laplace
equation, i.e.,
Using the representation formula (1.6), the solution of the Dirichlet boundary
value problem (1.14) is given by
int u (x, y)g(y)ds
u(x) = u (x, y)t(y)dsy 1,y y for x ,
Using the rst equation in the Calderon projector (1.11), we have to solve a
rst kind boundary integral equation to nd t H 1/2 ( ), such that
1
(V t)(x) = g(x) + (Kg)(x) for x , (1.15)
2
or,
1 t(y) 1 1 (x y, n(y))
dsy = g(x) + g(y)dsy .
4 x y 2 4 x y3
1
0 = V t g Kg 1/2 = sup .
2 H ( ) 0=wH 1/2 ( ) wH 1/2 ( )
1.1 Laplace Equation 11
Theorem 1.6. Let g H 1/2 ( ) be given. Then there exists a unique solution
w H 1/2 ( ) of the variational problem (1.18). Moreover,
1
wH 1/2 ( ) gH 1/2 ( ) .
cV1
Note that both boundary integral equations (1.15) and (1.17) are of the same
structure, while they are dierent in the denition of the right hand side. In
fact, the boundary integral equation (1.15) of the direct approach involves
the application of the double layer potential K to the given Dirichlet datum
g, while the right hand side of the boundary integral equation (1.17) of the
indirect approach is just the given Dirichlet datum g itself.
which leads, by applying the interior trace operator 0int and by the use of
Lemma 1.2, to a second kind boundary integral equation to nd v H 1/2 ( )
such that
1
v(x) (Kv)(x) = g(x) for x , (1.19)
2
or,
1 1 (x y, n(y))
v(x) v(y)dsy = g(x) for x .
2 4 x y3
The solution of the second kind boundary integral equation (1.19) is given by
the Neumann series
1
v(x) = I +K g(x) for x . (1.20)
2
=0
The convergence of the Neumann series (1.20) and therefore the unique solv
ability of the boundary integral equation (1.19) can be established when using
an appropriate norm in the Sobolev space H 1/2 ( ), see [108].
1.1 Laplace Equation 13
Theorem 1.7. Let g H 1/2 ( ) be given. Then there exists a unique solution
v H 1/2 ( ) of the boundary integral equation (1.19). Moreover,
1
vV 1 gV 1
1 cK
where cK < 1 is the contraction rate,
1
I + K z 1 cK zV 1 for all z H 1/2 ( )
2 V
with respect to the norm induced by the inverse single layer potential,
z2V 1 = V 1 z, z
for all z H 1/2 ( ).
Note that the solution of the Neumann boundary value problem (1.21) is only
unique up to an additive constant.
Using the representation formula (1.6), a solution of the Neumann bound
ary value problem (1.21) is given by
u(x) = u (x, y)g(y)dsy 1,y
int u (x, y) int u(y)ds , x .
0 y
Using the rst equation in (1.11), we have to solve a second kind boundary
integral equation to nd u = 0int u H 1/2 ( ) such that
1
u(x) + (K u)(x) = (V g)(x) for x , (1.23)
2
or,
1 1 (x y, n(y)) 1 g(y)
u(x) + u(y)dsy = dsy for x .
2 4 x y3 4 x y
As for the second kind boundary integral equation (1.19) for the Dirichlet
boundary value problem (1.14), a solution of the second kind boundary inte
gral equation (1.23) is given by the Neumann series
1
u(x) = I K (V g)(x) for x . (1.24)
2
=0
Since the given Neumann datum g H 1/2 ( ) has to satisfy the solvability
condition (1.22), and since v0 1 is the eigenfunction corresponding to the
zero eigenvalue of 1/2 I + K, all members of the Neumann series (1.24), and,
1/2
therefore, u are in the subspace H ( ) H 1/2 ( ) dened as follows
H ( ) = v H 1/2 ( ) : V 1 v, 1
= 0 .
1/2
The general solution of the second kind boundary integral equation (1.23) is
then given by u = u + c where c R is an arbitrary constant. To x the
constant, we may require the scaling condition
u , 1
= u (y)dsy = , (1.25)
is satised for all w H 1/2 ( ). Note that the bilinear form of the ex
tended variational problem (1.26) is regular due to the additional term
u , 1
w, 1
, which regularises the singular operator 1/2 I + K. Summaris
ing the above, we obtain the following result:
1.1 Laplace Equation 15
1/2
H ( ) = v H 1/2 ( ) : v, 1
= 0 .
Then the variational problem of the boundary integral equation (1.27) reads
1/2
to nd u H ( ) such that
1
Du, v = I K g, v (1.28)
2
1/2
is satised for all v H ( ). The general solution of the rst kind boundary
integral equation (1.27) is then given by u = u + c where c R is a constant
which can be determined by the scaling condition (1.25) afterwards.
1/2
Instead of solving the variational problem (1.28) in the subspace H ( )
and nding the unique solution afterwards from the scaling condition (1.25),
we can formulate an extended variational problem to nd u H 1/2 ( ) such
that
1
Du , v + u , 1 v, 1 = I K g, v + v, 1 (1.29)
2
S int u , v
+ u , 1
v, 1
= g, v
+ v, 1
(1.31)
The convergence of the series (1.33) follows as in Theorem 1.7 due to the
contraction estimate, see [108],
1
I K w cK wV for all w H 1/2 ( ) : w, 1
= 0
2 V
with cK < 1.
1.1 Laplace Equation 17
which is of the same structure and hence can be handled like the hypersingular
boundary integral equation (1.27); we skip the details.
where we have to nd the yet unknown Cauchy data 0int u on N and 1int u
on D . As we have seen in the two previous subsections on the Dirichlet and
on the Neumann problem, there exist dierent approaches leading to dierent
boundary integral equations to nd the unknown Cauchy data. However, we
consider here only two direct methods, which seem to be the most convenient
approaches to solve mixed boundary value problems by boundary element
methods. The denition of the Sobolev spaces H 1/2 (D ) can
1/2 (N ) and H
be seen in Appendix A.1.
The symmetric formulation (cf. [103]) is based on the use of the rst kind
boundary integral equation (1.15) to nd the unknown Neumann datum 1int u
on the Dirichlet part D , while the hypersingular boundary integral equation
(1.27) is used to nd the unknown Dirichlet datum 0int u on the Neumann
part N :
18 1 Boundary Integral Equations
1
(V 1int u)(x) =g(x) + (K0int u)(x) for x D ,
2
1
(D0int u)(x) = f (x) (K 1int u)(x) for x N .
2
Let g H ( ) and f H
1/2 1/2
( ) be some arbitrary, but xed extensions
of the given boundary data g H 1/2 (D ) and f H 1/2 (N ), respectively.
Then, we have to nd
1/2 (N ),
= 0int u g H
u t = 1int u f H
1/2 (D )
( ) H
t, u 1/2 (N )
1/2 (D ) H
such that
a( ; w, v) = F (w, v)
t, u (1.35)
is satised for all (w, v) H 1/2 1/2
(D ) H (N ) with
a( ; w, v) = V
t, u t, w
D K u, w
D + K
t, v
N + Du, v
N ,
1 1
F (w, v) = g + Kg V f, w + f K f D
g, v .
2 D 2 N
a(w, v; w, v) = V w, w
D + Dv, v
N ,
the unique solvability of the variational problem (1.35) follows from the map
ping properties of the single layer potential V and of the hypersingular integral
operator D.
(S int u
)(x) = f (x) (S int g)(x) for x N , (1.36)
S int u, v
+ u, v
= g, v
for all v H 1/2 ( ). (1.38)
S int u, v
+ f (u, ), v
= g, v
for all v H 1/2 ( ). (1.39)
The unique solvability of the nonlinear variational problem (1.39) follows from
appropriate assumptions on the nonlinear function f , see, e.g., [32, 95].
Theorem 1.14. Let g H 1/2 ( ) be given and let f be strongly monotone
satisfying
f (u, ) f (v, ), u v c u v2L2 ( ) for all u, v L2 ( ).
One of the main advantages in using boundary element methods for the ap
proximate solution of boundary value problems is their applicability to prob
lems in exterior unbounded domains. As a rst model problem we consider
the exterior Dirichlet boundary value problem
the exterior conormal derivative of u on . Note that the outer normal vector
n(x) is still dened with respect to the interior domain .
For a xed y0 and R > 2 diam , let BR (y0 ) be a ball of radius R
with centre in y0 and including . The solution of the boundary value problem
(1.40) is then given by the representation formula, see (1.6), for x BR (y0 )\
ext ext u (x, y)g(y)ds
u(x) = u (x, y)1 u(y)dsy + 1,y y
+ u (x, y)1int u(y)dsy int u (x, y) int u(y)ds .
1,y 0 y
BR (y0 ) BR (y0 )
As for the direct and the indirect approach for the interior Dirichlet boundary
value problem, we can conclude the unique solvability of the rst kind bound
ary integral equation (1.43) from Lemma 1.1. We then obtain the Dirichlet to
Neumann map
1
1ext u(x) = V 1 I + K 0ext u(x) + (V 1 u0 )(x) for x (1.44)
2
associated to the exterior Dirichlet boundary value problem (1.40).
Applying the exterior conormal derivative to the representation formula
(1.42), and inserting the Dirichlet to Neumann map (1.44), this gives
1
1ext u(x) = I K 1ext u(x) (D0ext u)(x)
2
1 1
1 1
= I K V I + K 0 u(x) + (V u0 )(x) (D0ext u)(x)
ext
2 2
1
= (S ext 0ext u)(x) + I K (V 1 u0 )(x) (1.45)
2
with the SteklovPoincare operator (cf. (1.13))
1 1
S ext = D + I + K V 1 I + K : H 1/2 ( ) H 1/2 ( ) (1.46)
2 2
associated to the exterior boundary value problem (1.40).
Instead of the exterior Dirichlet boundary value problem (1.40), we now con
sider the exterior Neumann boundary value problem
is satised for all v H 1/2 ( ). Using the test function v = v0 1, this gives
the trivial equality
1
Du, v0 = u, Dv0 = g, I + K v0 = 0.
2
This shows that the variational problem (1.52) has to be considered in a sub
space of H 1/2 ( ) which is orthogonal to constants. In particular, the solution
of the variational problem (1.52) is only unique up to a constant. Since the
hypersingular boundary integral operator D : H 1/2 ( ) H 1/2 ( ) is only
H 1/2 ( )semi elliptic, see Lemma 1.4, a suitable regularisation of the hyper
singular boundary integral operator has to be introduced. As in (1.29), we
obtain an extended variational problem to nd u H 1/2 ( ) such that
1
Du, v + u, 1 v, 1 = g, I +K v (1.53)
2
u, 1
= 0. Since u(x) = 1 for x e is a solution of the Laplace equation
u(x) = 0 with the radiation condition (1.41) for u0 = 1, the representation
formula (1.48) reads
u(x) = u0 + 1,y ext u (x, y)ds
y
implying
ext u (x, y)ds = 0 for x e .
1,y y
This shows that the scaling condition for the solution u of the extended vari
ational problem (1.53) can be chosen in an arbitrary way, the representation
formula (1.48) describes the correct solution for any scaling parameter.
From Greens second formula (1.3), we then obtain the representation formula
u(x) = u (x, y)t(y)dsy 1,y int u (x, y)g(y)ds +
y u (x, y)f (y)dy
for x , where t = 1int u is the yet unknown Neumann datum. As for the
interior Dirichlet boundary value problem (1.14), we have to solve a rst kind
boundary integral equation to nd t H 1/2 ( ) such that
1
(V t)(x) = g(x) + (Kg)(x) (N0 f )(x) for x , (1.55)
2
where
(N0 f )(x) = u (x, y)f (y)dy for x
is the Newton potential entering the right hand side. Hence, the unique solv
ability of the boundary integral equation (1.55) follows, as in Theorem 1.5,
for the rst kind boundary integral equation (1.15), which is associated to the
Dirichlet boundary value problem (1.14).
The drawback in considering the boundary integral equation (1.55) is the
evaluation of the Newton potential N0 f . Besides a direct computation there
exist several approaches leading to more ecient methods.
1.1 Laplace Equation 25
Integration by Parts
0int ui (x) = 0ext ue (x) , i 1int ui (x) = e 1ext ue (x) for x , (1.57)
for x and
ue (x) = u0 u (x, y)1ext ue (x)dsx + ext u (x, y) ext u (y)ds
1,y 0 e y
int/ext int/ext
for x e . To nd the unknown Cauchy data 0 u and 1 u, which
are linked via the transmission conditions (1.57), we have to solve appropriate
boundary integral equations on the interface boundary . Using the Dirichlet
to Neumann map associated to the interior Dirichlet boundary value problem
(1.54), in particular, solving the boundary integral equation (1.55),
1 int 1
(V 1int ui )(x) = ui (x) + (K0int ui )(x) (N0 f )(x) for x ,
2 0 i
we obtain
1 1
1int ui (x) = V 1 I + K 0int ui (x) V 1 (N0 f )(x) for x .
2 i
Let us assume that there is given a particular solution up satisfying
Hence, we obtain
1 int
(N0 f )(x) = up (x) + (K0int up )(x) (V 1int up )(x) for x ,
2 0
and, therefore,
1int ui (x) =
1 1 1 1
V 1 I + K 0int ui (x) + 1int up (x) V 1 I + K 0int up (x) =
2 i i 2
int int 1 int 1 int int
(S 0 ui )(x) + 1 up (x) (S 0 up )(x)
i i
is satised for all v H 1/2 ( ). The unique solvability of (1.59) nally follows
from the ellipticity estimates for the interior and exterior SteklovPoincare
operators S int and S ext .
3
ij (u, x) = 0 for x , i = 1, 2, 3 , (1.60)
j=1
xj
where R33 denotes the stress tensor. For a homogeneous isotropic ma
terial, the linear stressstrain relation is given by Hookes law
E 3
E
ij (u, x) = ij ekk (u, x) + eij (u, x)
(1 + )(1 2) 1+
k=1
for i, j = 1, 2, 3. Here, E > 0 is the Young modulus, and (0, 1/2) denotes
the Poisson ratio. The strain tensor e is dened as follows,
1
eij (u, x) = uj (x) + ui (x) for i, j = 1, 2, 3 .
2 xi xj
Inserting the strain and stress tensors, we obtain from (1.60) the Navier system
Multiplying the equilibrium equations (1.60) with some test function vi , in
tegrating over , applying integration by parts, and taking the sum over
i = 1, 2, 3, this gives the rst Betti formula
3
ij (u, y)vi (y)dy = a(u, v) 1int u(y), 0int v(y) dsy (1.61)
i,j=1
yj
3
(1int u)i (y) = ij (u, y)nj (y) for y ,
j=1
(1int u)(y) = div u(y) n(y) + 2 u(y) + n(y) curl u(y) for y .
n(y)
From (1.61) and using the symmetry of the bilinear form a(, ), we can deduce
the second Betti formula
3
ij (v, y)ui (y)dy + 1int v(y), 0int u(y) dsy (1.62)
i,j=1
yj
3
= ij (u, x)vi (x)dy + 1int u(y), 0int v(y) dsy .
i,j=1
yj
Let
1 0 0 x2 0 x3
R = span 0 , 1 , 0 , x1 , x3 , 0 (1.63)
0 0 1 0 x2 x1
be the space of the rigid body motions which are solutions of the homogeneous
Neumann boundary value problem
3
ij (v, x) = 0 for x , (1int v)i (x) = 0 for x ,
j=1
xj
where
1 wk (y)
(V wk )(x) = dsy
4 x y
0int : H 1 () H 1/2 ( )
for (V Lame w) , = 1, 2, 3, this denes a continuous boundary integral oper
ator V Lame = 0int V Lame .
is bounded with
V Lame w[H 1/2 ( )]3 cV2 w[H 1/2 ( )]3 for all w [H 1/2 ( )]3
1.2 Lame Equations 31
V Lame w, w
cV1 w2[H 1/2 ( )]3 for all w [H 1/2 ( )]3 ,
where
1 wk (y)
(V wk )(x) = dsy
4 x y
Note that the single layer potential V Lame of linear elastostatics can be
written as
where the rst part corresponds to the single layer potential V Stokes of the
Stokes problem (see Section 1.3). From V Stokes n = 0, we then obtain (cf.
[106])
1 1+ 3
(V Lame n, n) = (1 2) (V nk , nk ) ,
E 1
k=1
showing that the ellipticity constant behaves like O(1 2) for 1/2.
cV1
In particular, we have
lim cV1 () = 0.
1/2
32 1 Boundary Integral Equations
0int : H 1 () H 1/2 ( )
is bounded with
for x . Instead of the interior trace operator 0int , we may also apply the
interior boundary stress operator 1int to the representation formula (1.65).
To do so, we rst need to investigate the application of the boundary stress
operator to the single and double layer potentials V Lame w and W Lame v which
are both solutions of the homogeneous equilibrium equations (1.60).
is bounded with
int V
1int V Lame w[H 1/2 ( )]3 c21 w[H 1/2 ( )]3 for all w [H 1/2 ( )]3 .
In the same way as for the single layer potential V Lame w, we now consider the
application of the boundary stress operator 1int to the double layer potential
W Lame v.
is bounded with
1/2
and HR ( )elliptic,
34 1 Boundary Integral Equations
DLame v, v
cD
1/2
1 v[H 1/2 ( )]3 for all v HR ( ) ,
2
1/2
where HR ( ) is the space of all vector functions which are orthogonal to the
space R of rigid body motions. In particular, there holds
DLame u, v
=
1
3
u(y), v(x) dsy dsx +
4 x y Sk (y) Sk (x)
k=1
(M (x , n(x))v(x))
I
42 U (x, y) M (y , n(y))u(y)dsy dsx +
2 x y
3
1
Mkj (x , n(x))vi (x) Mki (y , n(y))vj (y)dsy dsx
4 x y
i,j,k=1
Applying the interior boundary stress operator 1int to the Somigliana identity
(1.65),
Since the single layer potential V Lame is [H 1/2 ( )]3 elliptic and therefore
invertible, we obtain from the rst equation in (1.70) the Dirichlet to Neumann
map
1int u(x) = (S Lame 0int u)(x) for x (1.71)
with the SteklovPoincare operator
1 1
S Lame = V Lame I + K Lame
2
1 1 1
=D Lame + I + K Lame V Lame I + K Lame .
2 2
Note that it holds
(S Lame 0int v)(x) = 0 for all v R.
in [H 1/2 ( )]3 for all test functions w [H 1/2 ( )]3 . Since the single layer
potential V Lame is [H 1/2 ( )]3 elliptic, the unique solvability of the varia
tional problem (1.72) follows due to the LaxMilgram theorem.
36 1 Boundary Integral Equations
3
ij (u, x) = 0 for x , i = 1, 2, 3,
j=1
x j (1.73)
1int u(x) = g(x) for x
Note that the solution of the Neumann boundary value problem (1.73) is only
unique up to the rigid body motions v R.
Using the Somigliana identity (1.65), a solution of the Neumann boundary
value problem (1.73) is given by the representation formula
u (x) = U (x, y), g(y) dsy int U (x, y), int u(y) ds
1,y 0 y
is satised in a weak sense, in particular, in the sense of [H 1/2 ( )]3 . Since the
hypersingular boundary integral operator DLame has the nontrivial kernel
of the rigid body motions, we have to consider the boundary integral equation
(1.75) in suitable subspaces. To this end, we dene
1/2
HR ( ) = u [H 1/2 ( )]3 : u, v
= 0 for all v R .
1/2
is satised for all v HR ( ). The general solution of the hypersingular
boundary integral equation (1.75) is then given by
1.2 Lame Equations 37
6
u (x) = u(x) + ck v k (x) ,
k=1
is satised for all v [H 1/2 ( )]3 . The extended variational problem (1.78) is
uniquely solvable for any given g [H 1/2 ( )]3 . If g satises the solvability
conditions (1.74), then u is the unique solution of the hypersingular boundary
integral equation (1.75) satisfying the scaling conditions (1.77).
for i = 1, 2, 3 is satised.
Using the Somigliana identity (1.65), the solution of the mixed boundary
value problem (1.79) is given by the representation formula
38 1 Boundary Integral Equations
3
3
u (x) =
fi (x)Ui (x, y)dsy int U ) (x, y)ds +
gi (y)(1,y i y
i=1 i=1
N,i D,i
3
3
(1int u)i (x)Ui (x, y)dsy int U ) (x, y)ds
0int ui (y)(1,y i y
i=1 i=1
D,i N,i
1/2 (N,i ),
i = 0int ui gi H
u ti = (1int u)i fi H
1/2 (D,i )
(V Lame
t)i (x) (K Lame u
)i (x) =
1
gi (x) + (K Lame g)i (x) (V Lame f)i (x)
2
for x D,i , and
(DLame u
)i (x) + K Lame t (x) =
i
1
fi (x) K Lame f (x) (DLame g)i (x)
2 i
3
3
( )
t, u 1/2 (D,i )
H 1/2 (N,i )
H
i=1 i=1
such that
a( ; w, v) = F (w, v)
t, u (1.80)
is satised for all
3
3
(w, v) 1/2 (D,i )
H 1/2 (N,i )
H
i=1 i=1
a( ; w, v) =
t, u
3 3
(V Lame t )i , wi (K Lame u
)i , wi +
D,i D,i
i=1 i=1
3
3
ti , (K Lame v )i + (DLame u
)i , vi
N,i N,i
i=1 i=1
F (w, v) =
3
1 Lame Lame
gi , wi + (K g)i , wi (V f )i , wi +
i=1
2 D,i D,i D,i
3
1
fi , vi fi , (K Lame v)i (DLame g)i , vi .
i=1
2 N,i N,i N,i
3
ij (u, x) = 0 for x , i = 1, 2, 3
j=1
x j
(1.81)
(0int u(x), n(x)) = g(x) for x ,
(0int u(x), n(x)) = g(x) , 1int u(x) (1int u(x), n(x))n(x) = 0 for x .
40 1 Boundary Integral Equations
Using
1/2
uT HT ( ) = v [H 1/2 ( )]3 : (v(x), n(x)) = 0 for x
S Lame uT , v T
= S Lame gn, v T
,
we get
u(x) + p(x) = 0 for x ,
as well as
1 2
div u(x) = p(x) = (1 + )(1 2)p(x) = 0
+ E
3
ti (u(y), p(y)) = p(y)ni (y) + 2 eij (u, y)nj (y), y , i = 1, 2, 3.
j=1
From Greens rst formula (1.84), we now derive Greens second formula which
reads for the solution (u, p) of (1.82) as
3
vi (y) + q(y) ui (y)dy p(y)div v(y)dy
yi
i=1
3
3
= ti (u(y), p(y))vi (y)dsy ti (v(y), q(y))ui (y)dsy .
i=1 i=1
for k = 1, 2, 3. As before,
denes a continuous map. Combining this with the mapping properties of the
interior trace operator
0int : H 1 () H 1/2 ( ) ,
and for k = 1, 2, 3 as a weakly singular surface integral; see also Lemma 1.15.
When considering the interior Dirichlet boundary value problem for the
Stokes system
u(x) + p(x) = 0 for x ,
div u(x) = 0 for x , (1.86)
0int u(x) = g(x) for x ,
1.3 Stokes System 43
On the other hand, it is obvious, that the pressure p satisfying the rst equa
tion in (1.86) is only unique up to an additive constant. In particular, the
homogeneous Dirichlet boundary value problem
where
V : H 1/2 ( ) H 1/2 ( )
is the single layer potential of the Laplace operator. Considering the boundary
1/2
integral equation (1.88) in H ( ), this can be rewritten as an extended
variational problem to nd t [H 1/2 ( )]3 such that
1
V Stokes t, w + t, n w, n = I + K Stokes g, w (1.89)
V V 2
is satised for all w [H 1/2 ( )]3 . Note that there exists a unique solution
t [H 1/2 ( )]3 of the extended variational problem (1.89) for any given
Dirichlet datum g [H 1/2 ( )]3 . If g satises the solvability condition (1.87),
1/2
we then obtain t H ( ).
44 1 Boundary Integral Equations
1 2
U (t, x) = U (t, x) for t > 0 , x . (1.90)
c2 t2
The equation (1.90) is valid for the wave propagation in a homogeneous,
isotrop, frictionfree medium having the constant speed of sound c. The most
important examples are the acoustic scattering and the sound radiation.
The time harmonic acoustic waves are of the form
U (t, x) = Re u(x)e t , (1.91)
From Greens formula (1.93) and by the use of the symmetry of the bilinear
form a( , ), we deduce Greens second formula,
(u(y) u(y))v(y)dy + 1int u(y)0int v(y)dsy =
2
(v(y) v(y))u(y)dy +
2
1int v(y)0int u(y)dsy .
which denes a continuous map from a given density function w on the bound
ary to a function V w, which satises the partial dierential equation (1.92)
in . In particular,
V : H 1/2 ( ) H 1 ()
is continuous and V w H 1 () is a weak solution of the Helmholtz equation
(1.92) for any w H 1/2 ( ). Using the mapping properties of the interior
trace operators
0int : H 1 () H 1/2 ( )
and
1int : H 1 (, + 2 ) H 1/2 ( ) ,
we can dene corresponding boundary integral operators, e.g. the single layer
potential operator
V = 0 V : H 1/2 ( ) H 1/2 ( ) ,
as follows:
1 e xy
(V w)(x) = u (x, y)w(y)dsy = w(y)dsy for x .
4 x y
Note that
H 1 (, + 2 ) = 1 () .
v H 1 () : v + 2 v H
V V0 : H 1/2 ( ) H 1/2 ( )
(Vk + C)w, w
= V0 w, w
cV1 0 w2H 1/2 ( ) (1.97)
for x , which again denes a continuous map from a given density function
v on the boundary to a function W v satisfying the Helmholtz equation
(1.92). In particular,
W : H 1/2 ( ) H 1 ()
is continuous and W v H 1 () is a weak solution of the Helmholtz equation
(1.92) for any v H 1/2 ( ). Using the mapping properties of the interior trace
operator
0int : H 1 () H 1/2 ( )
1.4 Helmholtz Equation 47
and
1int : H 1 (, + 2 ) H 1/2 ( ) ,
we can dene corresponding boundary integral operators, i.e. the trace
1
0int W = I + K
2
with the double layer potential operator
e xy
1
(K v)(x) = lim y , n(y) v(y)dsy for x .
0 4 x y
y :yx
D D0 : H 1/2 ( ) H 1/2 ( )
In addition to the interior boundary value problem for the Helmholtz equation
(1.92), we also consider the exterior boundary value problem
48 1 Boundary Integral Equations
For a xed y0 and R > 2 diam , let BR (y0 ) be a ball of radius R with
centre y0 and including . Let u be a solution of the exterior boundary value
problem for the Helmholtz equation (1.100) satisfying the radiation condition
(1.101). Considering Greens rst formula (1.93) with respect to the bounded
domain R = BR (y0 )\ and choosing v = u as test function, we obtain
u(y)2 dy k 2 u(y)2 = 1int u(y)0int u(y)dsy
R R R
= 1int u(y)0int u(y)dsy 1ext u(y)0ext u(y)dsy ,
BR (y0 )
when taking into account the opposite direction of the normal vector n(x) for
x . Since the left hand side of the above equation is real, we conclude
Im 1int u(y)0int u(y)dsy = Im 1ext u(y)0ext u(y)dsy .
BR (y0 )
By the use of this property, the Sommerfeld radiation condition (1.101) implies
2
int
0 = lim 1 u(y) 0int u(y) dsy
R
BR (y0 )
= lim 1int u(y)2 dsy + 2 0int u(y)2 dsy
R
BR (y0 ) BR (y0 )
2 Im 1int u(y)0int u(y)dsy
BR (y0 )
= lim 1int u(y)2 dsy + 2 0int u(y)2 dsy
R
BR (y0 ) BR (y0 )
2 Im 1ext u(y)0ext u(y)dsy
and, therefore,
1.4 Helmholtz Equation 49
2 Im 1ext u(y)0ext u(y)dsy
= lim 1int u(y)2 dsy + 2 0int u(y)2 dsy 0.
R
BR (y0 ) BR (y0 )
and, therefore,
1
u(x) = O as x . (1.102)
x
For the bounded domain R , we can apply the representation formula (1.95)
to obtain
int int u (x, y) int u(y)ds
u(x) = u (x, y)1 u(y)dsy + 1,y 0 y
+ u (x, y)1int u(y)dsy int u (x, y) int u(y)ds
1,y 0 y
BR (y0 ) BR (y0 )
We rst consider the interior Dirichlet boundary value problem for the Helm
holtz equation, i.e.
Using the representation formula (1.95), the solution of the above Dirichlet
boundary value problem is given by
u(x) = u (x, y)t(y)dsy 1,y
int u (x, y)g(y)ds
y for x ,
When applying the interior normal derivative 1int to the representation for
mula, this gives a second kind boundary integral equation to nd t H 1/2 ( )
such that
1
t(x) (K t)(x) = (D g)(x) for x . (1.107)
2
To investigate the unique solvability of the variational problem (1.106), and,
therefore, of the boundary integral equation (1.105) as well as of the boundary
integral equation (1.107), we rst consider the Dirichlet eigenvalue problem
for the Laplace operator,
Next we consider the interior Neumann boundary value problem for the Helm
holtz equation, i.e.
1.4 Helmholtz Equation 51
From the representation formula (1.95), we can obtain the solution of the
above boundary value problem as
u(x) = u (x, y)g(y)dsy 1,y
int u (x, y) int u(y)ds
0 y for x .
Applying the interior trace operator 0int to the above representation formula,
this gives a rst boundary integral equation to nd u = 0int u H 1/2 ( ) such
that
1
u(x) + (K u)(x) = (V g)(x) for x . (1.110)
2
When applying the conormal derivative operator 1int to the above represen
tation formula, this gives a second boundary integral equation,
1
(D u)(x) = g(x) (K g)(x) for x . (1.111)
2
Hence, u H 1/2 ( ) is a solution of the variational problem
1
D u, v = I K g, v for all v H 1/2 ( ). (1.112)
2
The exterior Dirichlet boundary value problem for the Helmholtz equation
reads
Since the single layer potential V of the exterior Dirichlet boundary value
problem coincides with the single layer potential of the interior Dirichlet
boundary value problem, V is not invertible when 2 = is an eigenvalue of
the Dirichlet eigenvalue problem (1.108). However, we have
1 1
I + K g, t = g, I K t = 0,
2 2
and, therefore,
1
I + K g Im V .
2
In fact, the variational problem (1.116) of the direct approach is solvable, but
the solution is not unique. As for the Neumann problem (1.21) for the Laplace
equation, we can use a stabilised variational formulation to obtain a unique
solution t H 1/2 ( ) satisfying some prescribed side condition, e.g.,
V0 t, t
= 0 ,
1.4 Helmholtz Equation 53
for all w H 1/2 ( ). Since this formulation requires the a priori knowledge of
the eigensolution t , this approach does not seem to be applicable in general.
If 2 is not an eigenvalue of the interior Dirichlet eigenvalue problem
(1.108), then the unique solvability of the boundary integral equation (1.115)
follows, since V is coercive and injective.
Instead of a direct approach, we may also consider an indirect single layer
potential approach,
u(x) = (V w)(x) = u (x, y)w(y)dsy for x e .
Then, applying the exterior trace operator, this leads to a boundary integral
equation to nd w H 1/2 ( ) such that
and, therefore,
1
I + K g Im D .
2
In fact, the variational problem (1.122) of the direct approach is solvable,
but the solution is not unique. Again, one can formulate a suitable stabilised
variational problem; we skip the details.
If 2 is not an eigenvalue of the Neumann eigenvalue problem (1.113), then
the unique solvability of the variational problem (1.122) follows, since D is
coercive and injective.
When applying the exterior trace operator 0ext to the representation for
mula, this gives a second kind boundary integral equation to be solved,
1
u(x) + (K u)(x) = (V g)(x) for x . (1.123)
2
If 2 = is an eigenvalue of the Dirichlet eigenvalue problem (1.108), the
operator
1
I K : H 1/2 ( ) H 1/2 ( )
2
is singular and, therefore, not invertible. Then, the adjoint operator
1
I K : H 1/2 ( ) H 1/2 ( )
2
is also not invertible.
As for the exterior Dirichlet boundary value problem, one may formulate
a combined boundary integral equation in L2 ( ), i.e., a linear combination of
the boundary integral equations (1.121) and (1.123) gives (cf. [19])
1 1
u(x) + (K u)(x) + (D u)(x) = (V g)(x) g(x) + (K g)(x)
2 2
for x , which is uniquely solvable due to the coercivity of the underly
ing boundary integral operators when assuming sucient smoothness of the
boundary .
56 1 Boundary Integral Equations
In the most simple case, we assume that is piecewise polyhedral and that
each boundary element mesh (2.1) consists of N plane triangular boundary
elements with mid points x . Using the reference element
( )
= R2 : 0 < 1 < 1, 0 < 2 < 1 1 ,
the boundary element = ( ) with nodes xi for i = 1, 2, 3 can be described
via the parametrisation
x() = () = x1 + 1 (x2 x1 ) + 2 (x3 x1 ) for .
For the area of the boundary element , we then obtain
*
1*
= dsx = EG F 2 d = EG F 2 ,
2
60 2 Boundary Element Methods
where
3 2
E= xi () = x2 x1 2 ,
i=1
1
3 2
G= xi () = x3 x1 2 ,
i=1
2
3
F = xi () xi () = (x2 x1 , x3 x1 ) .
i=1
1 2
The sequence of boundary element meshes (2.1) is called globally quasi uni
form if the mesh ratio
hmax
cG
hmin
is uniformly bounded by a constant cG which is independent of N N. Finally,
we introduce the element diameter
d = sup x y .
x,y
We assume that all boundary elements are uniformly shape regular, i.e.,
there exists a global constant cB independent of N such that
d c B h for all = 1, . . . , N.
With
x2 ,1 x1 ,1 x3 ,1 x1 ,1
J = x2 ,2 x1 ,2 x3 ,2 x1 ,2 R32
x2 ,3 x1 ,3 x3 ,3 x1 ,3
and using the parametrisation = ( ), a function v dened on can be
interpreted as a function v with respect to the reference element ,
v(x) = v(x1 + J ) = v () for , x = ().
Vice versa, a function v dened in the parameter domain implies a function
v on the boundary element ,
v() = v(x1 + J ) = v (x) for , x = ().
Hence, we can dene boundary element basis functions on by dening
associated shape functions on the reference element .
2.2 Basis Functions 61
0 () = 1 for
N
wh = w Sh0 ( ), w R for = 1, . . . , N.
=1
or,
N
w (x)k (x)dsx = w(x)k (x)dsx for k = 1, . . . , N.
=1
62 2 Boundary Element Methods
Due to
for k = ,
(x)k (x)dsx = ,
0 for k =
we obtain
1
w = w(x)dsx for = 1, . . . , N.
From this explicit representation of w , one can prove the error estimate, see
Appendix B.2,
N
w Qh w2L2 ( ) c wH s ( ) c h wHpw
h2s 2 2s 2
s ( ) (2.4)
=1
and
w2H 1 ( ) =  w( ())2 d for s = 1.
From the above variational formulation, we conclude the Galerkin orthogo
nality
w(x) (Qh w)(x) vh (x)dsx = 0 for all vh Sh0 ( ) ,
inf w wh H ( ) c hs wHpw
s ( ) (2.5)
wh Sh
0 ( )
With respect to the reference element , we may also dene local polynomial
shape functions of higher order. In particular, we introduce the linear shape
functions
11 () = 1 1 2 , 21 () = 1 , 31 () = 2 for . (2.6)
These shape functions imply globally discontinuous piecewise linear basis func
tions
i1 () for x = () ,
,i (x) =
0 elsewhere
for = 1, . . . , N , i = 1, 2, 3, and, therefore, the global trial space
N
Sh1,1 ( ) = span ,1 (x), ,2 (x), ,3 (x) , dim Sh1,1 ( ) = 3N.
=1
N
3
wh = w,i ,i Sh1,1 ( ) .
=1 i=1
N
3
Qh w = w,i ,i Sh1,1 ( ),
=1 i=1
N
w Qh w2L2 ( ) c h4 w2H 2 ( ) c h4 w2Hpw
2 ( )
=1
w Qh wL2 ( ) wL2 ( ) ,
w Qh wL2 ( ) c hs wHpw
s ( )
w Qh wH ( ) c hs wHpw
s ( ) (2.7)
inf w wh H ( ) c hs wHpw
s ( ) (2.8)
1,1
wh Sh ( )
for all [2, 0]. Moreover, the approximation property (2.8) remains valid
for all s 2 with < 1/2.
Note that the restrictions of j onto a boundary element k for k I(j) can
be represented by the linear shape functions j1k ,
Ph w, vh
H 1 ( ) = w, vh
for all vh Sh1 ( )
we can show the error estimate
w Ph wH ( ) c hs wHpw
s ( )
for all [2, 1]. Moreover, the approximation property (2.10) remains valid
for all s 2 with < 3/2.
Collocation Method
Inserting (2.11) into the boundary integral equation (1.15), and choosing the
boundary element mid points xk as collocation nodes, we have to solve the
collocation equations
1 1 1 1 (xk y, n(y))
t h (y)dsy = g(x ) + g(y)dsy (2.12)
4 xk y 2 k
4 xk y3
N
1 1 1 1 (xk y, n(y))
t dsy = g(xk ) + g(y)dsy
4 xk y 2 4 xk y3
=1
for k = 1, . . . , N . With
1 1
Vh [k, ] = dsy
4 xk y
for k, = 1, . . . , N , and
1 1 (xk y, n(y))
fk = g(xk ) + g(y)dsy
2 4 xk y3
Vh t = f .
follows. Combining this with the approximation property (2.5) for = 1/2,
we get the error estimate
t th H ( ) c hs tHpw
s ( ) , (2.13)
for some R follows. Combining this with the error estimate (2.13) for the
minimal possible value = 1, we obtain the pointwise error estimate
Galerkin Method
with
1 1
F (w) = V w, w I + K g, w .
2 2
N
th = t Sh0 ( )
=1
The solution th Sh0 ( ) of the above minimisation problem is dened via the
Galerkin equations
1
V th , k = I + K g, k for k = 1, . . . , N. (2.15)
2
With (2.11) and by using the denition (2.3) of the piecewise constant basis
functions , this is equivalent to
N
1 1
t dsy dsx =
4 x y
=1 k
1 1 (x y, n(y))
g(x)dsx + g(y)dsy dsx
2 4 x y3
k k
for k = 1, . . . , N . With
1 1
Vh [k, ] = dsy dsx
4 x y
k
for k, = 1, . . . , N , and
1 1 (x y, n(y))
fk = g(x)dsx + g(y)dsy dsx
2 4 x y3
k k
Vh t = f . (2.16)
The Galerkin stiness matrix Vh is symmetric and positive denite. Therefore,
one may use a conjugate gradient scheme for an iterative solution of the linear
system (2.16). Since the spectral condition number of Vh behaves like O(h1 ),
i.e.,
max (Vh ) 1
2 (Vh ) = Vh 2 Vh1 2 = c ,
min (Vh ) h
an appropriate preconditioning is sometimes needed. Moreover, since the sti
ness matrix Vh is dense, fast boundary element methods are required to con
struct more ecient algorithms, see Chapter 3.
From the H 1/2 ( )ellipticity and the boundedness of the single layer
potential
V : H 1/2 ( ) H 1/2 ( ) ,
see Lemma 1.1, we conclude the unique solvability of the Galerkin variational
problem (2.15), or, correspondingly, of the linear system (2.16), as well as the
quasi optimal error estimate, i.e. Ceas lemma,
cV2
t th H 1/2 ( ) inf t wh H 1/2 ( ) .
cV1 wh Sh
0 ( )
Combining this with the approximation property (2.5) for = 1/2, we get
1
t th H 1/2 ( ) c hs+ 2 tHpw
s ( ) ,
t th H ( ) c hs tHpw
s ( ) , (2.17)
for some R follows. Combining this with the error estimate (2.17) for the
minimal value = 2, we obtain the pointwise error estimate
Note that the error estimate (2.19) involves the position of the observation
point x again. In particular, the error estimate (2.19) does not hold in
the limiting case x .
The computation of the right hand side f in the linear system (2.16)
requires the evaluation of the integrals
1 1 (x y, n(y))
fk = g(x)dsx + g(y)dsy dsx
2 4 x y3
k k
M
gh = gj j Sh1 ( )
j=1
M
gh = g(xj ) j , (2.20)
j=1
or by the L2 projection,
M
gh = gj j ,
j=1
M
gj j , i
L2 ( ) = g, i
L2 ( ) for i = 1, . . . , M . (2.21)
j=1
This leads to
2.3 Laplace Equation 71
1
M M
1 (x y, n(y))
fk = gj j (x)dsx + gj j (y)dsy dsx
2 j=1 j=1
4 x y3
k k
M
1
= gj Mh [k, j] + Kh [k, j]
j=1
2
t
th H ( ) c1 t th H ( ) + c2 g gh H +1 ( )
follows with [2, 0], when the L2 projection (2.21) is used to dene
gh Sh1 ( ). Note that [1, 0] in the case of the interpolation (2.20).
Assuming t Hpw s
( ) and g Hpws+1
( ) for some s [0, 1], we then obtain
the error estimate
t
th H ( ) hs c1 tHpw
s ( ) + c2 g s+1
Hpw ( ) .
u(x) u
(x) c(x, t, g) h3 , (2.23)
u(x) u
(x) c(x, t, g) h2
follows.
72 2 Boundary Element Methods
is satised for all v H 1/2 ( ). Note that from the solvability condition (1.22),
we reproduce the scaling condition (1.25). Since the bilinear form of this vari
ational problem is strictly positive, the variational problem is equivalent to
the minimisation problem
F (u ) = min F (v)
vH 1/2 ( )
with
1 2 1
F (v) = Dv, v + v, 1 I K g, v v, 1 .
2 2
The solution u,h Sh1 ( ) of the above minimisation problem is then dened
via the Galerkin equations
Du,h , i + u,h , 1 i , 1 =
1
I K g, i + i , 1 (2.25)
2
M
u,j Dj , i + j , 1 i , 1 =
j=1
1
I K g, i + i , 1
2
for i = 1, . . . , M . With
1 curl j (y) , curl i (x)
Dh [i, j] = Dj , i
= dsx dsy ,
4 x y
ai = i , 1
= i (x)dsx ,
1
fi = I K g, i
2
1 int u (x, y)g(y)ds ds
= g(x)i (x)dsx i (x) 1,x y x
2
1 1 (y x, n(x))
= g(x)i (x)dsx i (x) g(y)dsy dsx
2 4 x y3
is the surface curl operator, and i is some locally dened extension of i into
the neighbourhood of . Since i is linear on every boundary element k , and
dening the extension i to be constant along n(x), we obtain curl i to be
a piecewise constant vector function. Hence, we get
Dh [i, j] = (2.27)
1 1
curl ik , curl j dsx dsy .
4 x y
k supp i supp j k
Dv, v
+ v, 1
2 cD
1 vH 1/2 ( )
2
for all v H 1/2 ( ),
holds. Combining this with the approximation property (2.10) for = 1/2,
we get
u u,h H 1/2 ( ) c hs1/2 u Hpw
s ( ) , (2.28)
when assuming u Hpw s
( ) for some s [1/2, 2]. Applying the Aubin
Nitsche trick, we also obtain the error estimate
u u,h H ( ) c hs u Hpw
s ( ) ,
for some R follows. Combining this with the error estimate (2.28) for the
minimal value = 1, we obtain the pointwise error estimate
int
u(x) u(x) c hs+1 1,y u (x, ) 1 u Hpw
s ( ) ,
H ( )
Again, the error estimate (2.29) involves the position of the observation point
x , and, therefore, it is not valid in the limiting case x .
As in the boundary element method for the Dirichlet boundary value prob
lem, we may also approximate the given Neumann datum g H 1/2 ( ) rst.
If gh Sh0 ( ) is dened by the L2 projection, i.e. if it is the unique solution
of the variational problem
gh (x)k (x) dsx = g(x)k (x) dsx for k = 1, . . . , N ,
g gh H ( ) c hs gHpw
s ( )
g gh H 1 ( ) c h2 gHpw
1 ( ) . (2.30)
Then,
76 2 Boundary Element Methods
1
fi = I K gh , i
2
N
N
1 int u (x, y)ds ds
= g i (x)dsx g i (x) 1,x y x
2
=1 =1
1
N N
1 (y x, n(x))
= g i (x)dsx g i (x) dsy dsx
2 4 x y3
=1 =1
N
1
= g Mh [, i] Kh [, i] .
2
=1
Instead of the linear system (2.26), we now have to solve a linear system with
a perturbed right hand side f yielding a perturbed solution vector u , i.e.,
we have to solve the linear system
1
Dh + a a u = Mh Kh g + a . (2.31)
2
,h Sh1 ( ), the error estimate
For the associated boundary element solution u
u u
,h H 1/2 ( ) u u,h H 1/2 ( ) + c g gh H 1/2 ( )
c h3/2 u Hpw
2 ( ) + gH 1 ( )
pw
,
u u
h H ( ) c1 u uh H ( ) + c2 g gh H 1 ( )
c h2 uHpw 2 ( ) + gH 1 ( )
pw
,
u(x) u
(x) c(x, g, u ) h2 , (2.34)
u(x) = 0 for x ,
0int u(x) = g(x) for x D ,
int u(x) = f (x)
1 for x N
1/2 (N )
= 0int u g H
u
and
t = 1int u f H
1/2 ( )
a( ; w, v) = F (w, v)
t, u
1/2 (N ) and w H
for all v H 1/2 (D ) with the bilinear form
1 1
; w, v) =
a(t, u w(x) t(y)dsy dsx
4 x y
D D
1 (x y, n(y))
w(x) (y)dsy dsx
u
4 x y3
D N
1 (y x, n(x))
+ v(x) t(y)dsy dsx
4 x y3
N D
1 (y)
curl v(x) , curl u
+ dsy dsx
4 x y
1 1
f (y)dsy dsx ,
4 x y
k
1 1 (y x, n(x))
fi = i (x)f (x)dsx i (x) f (y)dsy dsx
2 4 x y
N N
1 (curl g(y), curl i (x))
dsy dsx
4 x y
N
Inserting this into the second equation in (2.36), this gives the Schur comple
ment system
= f Kh Vh1 g
Sh u (2.37)
with the symmetric and positive denite Schur complement matrix
Sh = Dh + Kh Vh1 Kh .
is symmetric and positive denite. Hence, instead of (2.36), we now solve the
transformed linear system
Vh CV1 Vh Vh (I Vh CV1 )Kh
t
1 1
= (2.38)
Kh (I CV Vh ) Dh + Kh CV Kh
u
Vh CV1 I 0 g
1
Kh CV I f
tth 2H 1/2 ( ) + uuh 2H 1/2 ( )
c inf
t wh 2H 1/2 ( ) + inf
u vh 2H 1/2 ( )
wh Sh (D )
0 vh Sh
1 ( )
N
from Ceas lemma. Using the approximation property (2.5) for = 1/2 as
well as the approximation property (2.10) for = 1/2, this gives
t
th 2H 1/2 ( ) + h 2H 1/2 ( ) c1 h2s1 +1 
uu t2Hpw
s1
( )
+ c2 h2s2 1 
u2Hpw
s2
( )
,
t
th 2H 1/2 ( ) +
uu
h 2H 1/2 ( ) c h2(s1) u2H s () .
As for the Dirichlet and for the Neumann boundary value problem, applying
the AubinNitsche trick (for [2, 1/2)) and an inverse inequality argu
ment (for (1/2, 0]), we obtain the error estimate
t
th 2H ( ) +
uu
h 2H +1 ( ) c h2(s)3 u2H s () , (2.39)
u(x) u(x) =
u (x, y) t(y) int u (x, y) u
th (y) dsy 1,y (y) u
h (y) dsy .
N D
u(x) u
(x)
u (x, )H 1 ( )
t
th H 1 ( ) + 1int u (x, )
uu
h H 2 ( )
H 2 ( )
82 2 Boundary Element Methods
for some 1 , 2 R follows. Combining this with the error estimate (2.39)
for the minimal values 1 = 2 and 2 = 1, we obtain the pointwise error
estimate
(x) c h2s+1 u (x, )H 2 ( ) + 1int u (x, )H 1 ( ) uH s () ,
u(x) u
Note that the error estimate (2.40) is based on the exact use of the given
boundary data g H 1/2 (D ) and f H 1/2 (N ), and their extensions g
H 1/2 ( ) and f H 1/2 ( ).
Starting from an approximation uh Sh1 ( ) of the complete Dirichlet
datum 0int u,
M
MN
M
uh = u j j = u j j + h + gh ,
u j j = u
j=1 j=1 j=MN +1
M
uj j (x)i (x)dx = g(x)i (x)dsx for i = MN + 1, . . . , M.
j=MN +1 D D
N
t (x)k (x)dx = f (x)k (x)dsx for k = ND + 1, . . . , N.
=ND +1 N N
th Sh0 (D ) h Sh1 (N )
and u
a( h ; k , i ) = F(k , i )
th , u
Note that the right hand side of this system diers from the one in (2.36).
The blocks on the right have the following dimensions:
for
= ND + 1, . . . , N , k = 1, . . . , ND , j = MN + 1, . . . , M , i = 1, . . . , MN .
Sh u = f ,
2.3 Laplace Equation 85
do not allow a direct evaluation of both, the stiness matrix and the right hand
side, additional approximations are required. The application of the Steklov
Poincare operator S int related to the interior Dirichlet boundary value prob
lem can be written as
1 1
(S int u)(x) = D + I + K V 1 I + K u(x)
2 2
1
= (Du)(x) + I + K ti (x) ,
2
where 1
ti = V 1 I + K u H 1/2 ( )
2
is the unique solution of the variational problem
1
V ti , w
= I + K u, w for all w H 1/2 ( ).
2
Let ti,h Sh0 ( ) be the unique solution of the Galerkin variational problem
1
V ti,h , wh = I + K u, wh for all wh Sh0 ( ).
2
Then,
1
(Sint u)(x) = (Du)(x) + I + K ti,h (x)
2
denes an approximate SteklovPoincare operator associated to the interior
Dirichlet boundary value problem. In the same way, we dene an approximate
SteklovPoincare operator
86 2 Boundary Element Methods
1
(Sext u)(x) = (Du)(x) + I + K te,h (x) ,
2
which is associated to the exterior Dirichlet boundary value problem, and
where te,h Sh0 ( ) is the unique solution of the Galerkin equations
1
V te,h , wh
= I + K u, wh for all wh Sh0 ( ).
2
for k, = 1, . . . , N and i, j = 1, . . . , M .
Instead of the linear system (2.45) we may also solve the equivalent coupled
system
i Vh 0 i ( 12 Mh + Kh ) ti
0 e Vh e ( 2 Mh + Kh ) te
1
1 1
i ( 2 Mh + Kh ) e ( 2 Mh + Kh ) (i + e )Dh
u
( 12 Mh + Kh )up
= 0 , (2.46)
Dh up Mh tp
which is of the same structure as the linear system (2.36), i.e. block skew
symmetric but positive denite. Note that (2.45) is the Schur complement
system of (2.46).
As for the Neumann boundary value problem, we conclude the error esti
mate
h H 1/2 ( ) c1
u u inf u vh H 1/2 ( )
vh Sh
1 ( )
u u
h L2 ( ) c(u) h2 .
When the Dirichlet datum uh is known, one can compute the remaining Neu
mann datum by solving both, the interior and exterior Dirichlet boundary
value problems. Since those boundary value problems are Dirichlet boundary
value problems with approximated boundary data, the corresponding error
estimates are still valid.
3
ij (u, x) = 0 for x ,
j=1
xj
for i = 1, 2, 3. To nd the yet unknown Cauchy data (1int u)i on D,i and
0int ui on N,i , we consider the variational problem (1.80), which is related to
the symmetric formulation of boundary integral equations. Hence, we have to
nd
ti = (1int u)i fi H
1/2 (D,i )
and
1/2 (N,i )
i = 0int ui gi H
u
such that
a( ; w, v) = F (w, v)
t, u
1/2 (D,i ) and vi H
is satised for all wi H 1/2 (N,i ) for i = 1, 2, 3. Note
that the bilinear form is given by
3
3
a( ; w, v) =
t, u (V Lame
t )i , wi (K Lame u
)i , wi
D,i D,i
i=1 i=1
3
3
+ ti , (K Lame v )i + (DLame u
)i , vi ,
N,i N,i
i=1 i=1
F (w, v) =
3
1 Lame Lame
gi , wi + (K g )i , wi (V f )i , wi +
i=1
2 D,i D,i D,i
3
1
fi , vi fi , (K Lame v )i (DLame g )i , vi .
i=1
2 N,i N,i N,i
a( h ; wh , v h ) = F (wh , v h )
th , u
is satised for all wi Sh0 (D,i ) and vi Sh1 (M,i ) for i = 1, 2, 3. This
formulation is equivalent to a linear system of equations
VhLame KhLame
t = g , (2.47)
KhLame DhLame
u f
where
3
3
ND = ND,i , MN = MN,i .
i=1 i=1
While the blocks in the linear system (2.47) recover only the unknown coe
cients i,j , an implementation based on the complete stiness matrices
ti, and u
may be advantageous. Let
N
M
Sh0 ( ) = span , Sh1 ( ) = span j
=1 j=1
ND,i
N
whi (x) = wi i (x) Sh0 (D,i ), wh (x) = w (x) Sh0 ( )
=1 =1
MN,i
N
vhi (x) = vji ij (x) Sh1 (N,i ), vh (x) = vj j (x) Sh1 ( ).
j=1 j=1
where the stiness matrices VhLame , KhLame , and DhLame correspond to the
Galerkin discretisation of the associated boundary integral operators V Lame ,
90 2 Boundary Element Methods
K Lame and DLame with respect to the boundary element spaces [Sh0 ( )]3
and [Sh1 ( )]3 . In particular, for the discrete single layer potential Vh we have
the representation
VhLame = (2.48)
Vh 0 0 V11,h V21,h V13,h
1 1 1+
(3 4) 0 Vh 0 + V21,h V22,h V23,h
2E 1
0 0 Vh V31,h V32,h V33,h
KhLame = (2.51)
Kh 0 0 Vh 0 0
0 Kh 0 0 Vh 0 T + E VhLame T ,
1+
0 0 Kh 0 0 Vh
where Vh and Kh are the Galerkin matrices related to the single and double
layer potential of the Laplace operator. Furthermore, T is a transformation
matrix related to the matrix surface curl operator M (, n).
Using the representation of the bilinear form of the hypersingular bound
ary integral operator DLame as given in Lemma 1.18, one can derive a similar
representation for the Galerkin matrix DhLame , which is based on the trans
formation matrix T and on the Galerkin matrices related to the single layer
potential of both, the Laplace operator and the system of linear elastostatics.
2.5 Helmholtz Equation 91
The solution of the interior Dirichlet boundary value problem (cf. (1.104)),
N
th = t Sh0 ( )
=1
for k, = 1, . . . , N , and
92 2 Boundary Element Methods
1 1 (x y, n(y))
fk = g(x)dsx + 1 x y e xy g(y)dsy dsx
2 4 x y3
k k
for k = 1, . . . , N .
Since the single layer potential V : H 1/2 ( ) H 1/2 ( ) is coercive, i.e.
V satises (1.97), and since V is injective when 2 is not an eigenvalue of
the Dirichlet eigenvalue problem (1.108), we conclude the unique solvability
of the Galerkin variational problem (2.52), as well as the quasi optimal error
estimate, i.e. Ceas lemma,
Combining this with the approximation property (2.5) for = 1/2, we get
1
t th H 1/2 ( ) c hs+ 2 tHpw
s ( ) ,
t th H ( ) c hs tHpw
s ( ) , (2.54)
for some R follows. Combining this with the error estimate (2.54) for the
minimal value = 2, we obtain the pointwise error estimate
Hence, if t Hpw
1
( ) is suciently smooth, we obtain the optimal order of
convergence for s = 1,
Again, the error estimate (2.56) involves the position of the observation point
x , and, therefore, it is not valid in the limiting case x .
As for the Dirichlet problem for the Laplace equation, the computation of
fk requires the evaluation of the integrals
1 1 (x y, n(y))
fk = g(x)dsx + 1 x y e xy g(y)dsy dsx .
2 4 x y3
k k
and
1 (x y, n(y))
K,h [k, j] = 1 x y e xy j (y)dsy dsx (2.58)
4 x y3
k
u(x) u
(x) c(x, t, g) h3 , (2.59)
for i, j = 1, . . . , M , and
1
fi = g(x)i (x)dsx
2
1 (x y, n(y))
i (x) 1 x y e xy g(y)dsy dsx
4 x y3
2.5 Helmholtz Equation 95
for i = 1, . . . , M . Note that for the computation of the matrix entries D,h [i, j],
we can reuse the discrete single layer potential V,h for picewise constant basis
functions, but we also need to have the Galerkin discretisation with piecewise
linear continuous basis functions of the operator
xy
e
(C u)(x) = (n(x), n(y)) u(y)dsy , (2.63)
x y
D : H 1/2 ( ) H 1/2 ( )
Combining this with the approximation property (2.10) for = 1/2, we get
1
u uh H 1/2 ( ) c hs 2 uHpw
s ( ) ,
u uh H ( ) c hs uHpw
s ( ) , (2.64)
for some R follows. Combining this with the error estimate (2.64) for the
minimal value = 1, we obtain the pointwise error estimate
Hence, if u Hpw
2
( ) is suciently smooth, we get the optimal order of
convergence for s = 2,
Again, the error estimate (2.66) involves the position of the observation point
x , and, therefore, is not valid in the limiting case x .
When using a piecewise constant approximation gh Sh0 ( ) of the given
Neumann datum g H 1/2 ( ), we can compute a perturbed piecewise linear
approximation uh Sh1 ( ) from the Galerkin equations
1
D uh , i = I K gh , i for i = 1, . . . , M
2
As for the perturbed linear system (2.31) for the Neumann boundary value
problem of the Laplace equation, we obtain the error estimate
u(x) u
(x) c(x, t, g) h2 , (2.67)
The solution of the exterior Dirichlet boundary value problem (cf. (1.114))
Moreover, as for the interior Dirichlet problem, there holds the optimal error
estimate
u(x) u
(x) c(x, t, g) h3 , (2.69)
when using a L2 projection to approximate the boundary conditions, and
when assuming t Hpw
1
( ) and g Hpw
2
( ).
98 2 Boundary Element Methods
The solution of the exterior Neumann boundary value problem (cf. (1.120))
u(x) u
(x) c(x, t, g) h2 , (2.70)
When using boundary element methods for the numerical solution of boundary
value problems for threedimensional second order partial dierential equa
tions, one has to deal with two main diculties. First of all, almost all matri
ces involved are dense, i.e. all their entries do not vanish in general, leading
to an asymptotically quadratic memory requirement for the whole procedure.
Thus, classical boundary element realisations are applicable only for a rather
moderate number N of boundary elements. Fortunately, all boundary element
matrices can be decomposed into a hierarchical system of blocks which can
be approximated by the use of low rank matrices. This approximation will be
the main content of this chapter.
The second diculty is the complicated form of the matrix entries to be
generated. The Galerkin method, for example, requires the evaluation of dou
ble surface integrals for each matrix entry. This can not be done analytically
in general. Thus, combined semianalytical computations will be used to gen
erate the single entries of the matrices. A more detailed description of the
corresponding procedures is presented in Appendix C.
3.1.1 Motivation
Let K : [0, 1] [0, 1] R be a given function of two scalar variables and let
A RN M be a given matrix having the entries
with (xk , y ) [0, 1] [0, 1]. It is obvious, that the asymptotic memory re
quirement for the dense matrix A is Mem(A) = O(N M ), and the asymptotic
number of arithmetical operations required for a matrixvector multiplication
is Op(A s) = O(N M ) as N, M . This quadratic amount is too high for
modern computers, already for moderate values of N and M . However, if we
agree to store just an approximation A of the matrix A and to deal with the
product A s instead of the exact value A s, the situation may change. But
then it is necessary to control the error, i.e. to guarantee the error bound
for some prescribed accuracy , where AF denotes the Frobenius norm of
the matrix A,
1/2
N
M
AF = a2k . (3.3)
k=1 =1
where
i R+ , ui RN , vi RM , i = 1, . . . , r
are the biggest singular values and the corresponding singular vectors of the
matrix A. The rank r = r() is chosen corresponding to the condition
min(N,M )
min(N,M )
A A2F i2 2 i2 = 2 A2F . (3.5)
i=r+1 i=1
0 0
5 5
10 10
15
15
The number of signicant singular values slowly increases with the di
mension, as it can be seen in Fig. 3.2. Here the rst 32 singular values (lo
garithmic plot) for N = M = 32, 128, 256 (left plot, from below) and for
N = M = 256, 512, 1024 (right plot, from below) are shown. The accuracy of
5 5
0 0
5 5
10 10
15 15
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Fig. 3.2. First 32 singular values for N = 32, 64, 128, 256, 512, and N = 1024
the low rank approximation A(r) of the matrix A is illustrated in Fig. 3.3,
where the logarithmic plot of the function
A A(r)F
(r) =
AF
104 3 Approximation of Boundary Element Matrices
0 0
5 5
10 10
15 15
20 20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
for r = 1, . . . , 32 is depicted. The left plot in Fig. 3.3 corresponds again to the
dimensions N = 32, 64, and N = 128 (from below), while the right plot shows
the results for N = 256, 512, and N = 1024 (from below). Thus, the behaviour
of the singular values determines the quality of the low rank approximation
(3.4). The results shown do not really depend on the parameter . If becomes
smaller, the results are even better.
The situation changes if the singularity of the function K is more serious.
As a further example, let us consider the following function on [0, 1] [0, 1],
1
K(x, y) = , (3.8)
+ (x y)2
where > 0 is again some real parameter. For small values of the parameter
the function K gets an articial singularity along the diagonal {(x, x)}
of the square [0, 1] [0, 1].
In Fig. 3.4 (left plot), the rank r() for = 106 and N = M = 256 is
shown as a function of the parameter . The horizontal axis corresponds to
the values log2 (), while changes from 20 till 28 . Thus, the rank of the
matrix strongly depends on the parameter . However, if we separate the
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 2 4 6 8 0 2 4 6 8
variables x and y, i.e. consider only the quarter [0, 0.5] [0.5, 1] of the square
[0, 1][0, 1], then the situation is much better. The right plot in Fig. 3.4 shows
the same curve for separated x and y, which is more or less constant now.
3.1 Hierarchical Matrices 105
The logarithmic plots of the singular values of the matrix A for = 101
(lower curve) and for = 108 (upper curve) are shown in Fig. 3.5. The left
plot in this gure corresponds to the whole square [0, 1][0, 1], while the right
plot shows the results for the separated variables x and y, i.e. if we consider
only the quarter [0, 0.5] [0.5, 1] of the square [0, 1] [0, 1]. Now the main
0 0
5 5
10 10
15 15
[0, 0.5] [0, 0.5], [0, 0.5] [0.5, 1], [0.5, 1] [0, 0.5], [0.5, 1] [0.5, 1] ,
we will be able to approximate two of these four blocks eciently. The two re
maining, main diagonal blocks have the same structure as the original matrix,
but only half of the size and their rank will be smaller. In Fig. 3.6, the left
diagram corresponds to the whole matrix and its rank r() = 73 is obtained
for = 29 , = 106 and N = M = 256. The 2 2 block matrix together
with the ranks of the blocks is shown in the second diagram of Fig. 3.6. The
12 7
20 8 7 12 8
38 9 8 20
9 12 7 9
8
73 20 8
7 12
12 7
7 12 8
9 38 9 8 20
9 12 7
8 7 12
Degenerated approximation
In the above example, the approximation of the blocks for separated variables
x and y is based on the smoothness of the function K for x = y. However, if
the function K is degenerated, i.e. it is a nite sum of products of functions
depending only on x and y,
r
K(x, y) = pi (x)qi (y) , (3.9)
i=1
with
with
1 i K(x , y )
(ui )k = (xk x )i , (vi ) =
i! xi
for k = 1, . . . , N and = 1, . . . , M . Again, (3.10) is not the partial singu
lar value decomposition (3.4) of the matrix A. If the remainder term Rr is
uniformly bounded by the original function K satisfying
Rr (x, y) K(x, y)
for all x and y with some r = r(), then we can guarantee the accuracy of the
low rank matrix approximation
A AF AF (3.11)
for all dimensions N and M . The rank r+1 of the matrix A is also independent
of its dimension. Thus, for N M , the matrix A requires only Mem(A) =
O(N ) words of computer memory. However, an ecient construction of the
Taylor series for a given function in the threedimensional case is practically
impossible. Thus, it is rather an illustration for the fact that there exist low
rank approximations of dense matrices, which are not based on the singular
value decomposition.
A further example of a low rank approximation of a given function is the
decomposition of the fundamental solution of the Laplace equation
1 1
u (x, y) = for x, y R3
4 x y
(cf. (1.7)) into spherical harmonics in some point x with x x  < y x 
n
1 1 x x 
u (x, y) =
Pn (ex , ey ) ,
4 y x  n=0 y x 
x x y x
ex = , ey = ,
x x  y x 
108 3 Approximation of Boundary Element Matrices
where Ynm are the spherical harmonics. See [39, 93] for more details.
Large dense matrices arising from integral equations have no explicit structure
in general. As a rule, because of the singularity of the kernel function on the
diagonal, i.e. for x = y, these matrices are also not of low rank. However, it
is possible to nd a permutation, so that the matrix with permuted rows and
columns contains rather large blocks close to some lowrank matrices with
respect to the Frobenius norm (cf. (3.2)).
Cluster tree
(xk , gk ) , k = 1, . . . , N R3 R+ (3.13)
and
(y , q ) , = 1, . . . , M R3 R+ (3.14)
Cl = (xk , gk ) , k = 1, . . . , n
with n > 1 can be separated in two sons using the following algorithm.
3.1 Hierarchical Matrices 109
Algorithm 3.1
1. Mass of the cluster
n
G= gk R+ ,
k=1
1
n
X= gk xk R3 ,
G
k=1
C vi = i vi , i = 1, 2, 3 , 1 2 3 0 ,
5. Separation
5.1 initialisation
Cl1 := , Cl2 := ,
5.2 for k = 1, . . . , n
Cluster pairs
Next, cluster pairs which are geometrically well separated are identied. They
will be regarded as admissible cluster pairs, as e.g. the clusters in Fig. 3.8. An
110 3 Approximation of Boundary Element Matrices
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
where 0 < < 1 is a given parameter. Although the criterion (3.15) is quite
simple, a rather large computational eort (quadratic with respect to the
number of elements in the clusters Clx and Cly ) is required for calculating
the exact values
In practice, one can use rougher, more restrictive, but easily computable
bounds
diam(Clx ) 2 max X xk  ,
k
diam(Cly ) 2 max Y y  ,
112 3 Approximation of Boundary Element Matrices
1
dist(Clx , Cly ) X Y  diam(Clx ) + diam(Cly ) ,
2
where X and Y are the already computed centres (cf. Algorithm 3.1) of the
clusters Clx and Cly , for the admissibility condition. If a cluster pair is not
admissible, but nx > nmin and my > nmin are satised, then there exist sons
of both clusters
For simplicity, let us assume that the cluster Clx is bigger than Cly , i.e.
diam(Clx ) diam(Cly ).
Algorithm 3.2
1. Initialisation
R0 (x, y) = K(x, y) , S0 = 0 .
2. For i = 0, 1, 2, . . . compute
2.1. pivot element
The stopping criterion for the above algorithm can be realised in Step 2.1.
corresponding to the condition
Rm (xi , y) = 0 , y Y , Rm (x, yi ) = 0 , x X
Due to the stopping criterion (3.16), and due to the approximation property
(3.17), the following estimate obviously holds
A AF AF ,
Remark 3.3. Step 2.1. of Algorithm 3.2 should be discussed in more details.
It can be very dicult, if not impossible, to solve the maximum problem
formulated there. There are two possibilities to proceed. First, we can look
for the maximum only in a nite set of given points (xk , y ). In this case
only the original entries of the matrix A will be used for its approximation.
The algorithm will coincide with the algebraic fully pivoted ACA algorithm as
described in Subsection 3.2.2. The other possibility is to choose some articial
points and to look for the maximum there. These points can be the zeros
of the threedimensional Chebyshev polynomials in corresponding bounding
boxes for the sets X, Y . In this case the ACA approximation will be similar
to the best possible polynomial interpolation.
3.2 Block Approximation Methods 115
Remark 3.4. The stopping criterion (3.16) can be applied only if the function
K(x, y) is smooth on (X, Y ). If this is not the case, but if the function K is
asymptotically smooth (cf. (3.19)), then we have to decompose the domains
X and Y into two systems of clusters and to approximate the function on
each admissible cluster pair (Clx , Cly ) separately using Algorithm 3.2. This
decomposition implies the corresponding decomposition of the matrix in a
hierarchical system of blocks.
for all multiindices N30 with a constant g < 0. Moreover, the matrix
A RN M with entries (3.18) is decomposed into blocks corresponding to the
admissibility condition
A AF AF ,
and
: R, = 1, . . . , M
xk , k = 1, . . . , N
be a set of collocation points. If, for example, is the union of plane triangles
(cf. (2.1)),
116 3 Approximation of Boundary Element Matrices
'
N
= ,
=1
then the most simple collocation method with piecewise constant basis func
tions
.
1 , y ,
(y) =
0,y /
can be used. In this case, the collocation points xk are the midpoints of the
triangles k . The corresponding collocation matrix A RN M with
ak = K(xk , y) (y) dsy , k = 1, . . . , N , = 1, . . . , M (3.20)
for some kernel function K can be approximated using the following fully
pivoted ACA algorithm.
Algorithm 3.6
1. Initialisation
R0 (x, y) = K(x, y) , S0 = 0 .
2. For i = 0, 1, 2, . . . compute
2.1. pivot element
(ki+1 , i+1 ) = ArgMax Ri (xk , y) (y) dsy ,
Note that the approximation property (3.17) remains valid for Algorithm
3.6, while the interpolation property (3.18) remains valid only with respect to
the variable x,
Rm (xki , y) = 0, y , i = 1, 2, . . . , m, m = 1, 2, . . . , r . (3.21)
The interpolation property with respect to y changes to the orthogonality
Rm (x, y) i (y)dsy = 0, x , i = 1, . . . , m, m = 1, 2, . . . r. (3.22)
For the analysis of Algorithm 3.6, it is useful to introduce the following func
tions
U (x) = K(x, y) (y)dsy , = 1, . . . , M ,
having the property ak = U (xk ) , (cf. (3.20)). Using the properties (3.21)
and (3.22), we can conclude that the functions
U (x) = Sr (x, y) (y)dsy (3.23)
: R, = 1, . . . , M ,
k : R, k = 1, . . . , N .
The Galerkin matrix A can be approximated using the following ACA algo
rithm.
118 3 Approximation of Boundary Element Matrices
Algorithm 3.7
1. Initialisation
R0 (x, y) = K(x, y) , S0 = 0 .
2. For i = 0, 1, 2, . . . compute
2.1. pivot element
(ki+1 , i+1 ) = ArgMax Ri (x, y) (y) k (x) dsy dsx ,
The approximation property (3.17) remains valid for Algorithm 3.7, which,
instead of the interpolation property (3.18), possesses the following orthogo
nalities for i = 1, . . . , m , m = 1, 2, . . . r:
Rm (x, y) i (y) dsy = 0 , x ,
Rm (x, y) ki (x) dsx = 0 , y .
situation, e.g. for plane triangles k and by using piecewise constant basis func
tions, some numerical integration is involved (cf. Chapter 4). If both integrals
in (3.24) are computed numerically, i.e.
ak ak = k,kx ,ky K(xk,kx , y,ky ) (y,ky )k (xk,kx ), (3.25)
kx ky
where k,kx , ,ky are the weights of the quadrature rule (including Jacobians)
and xk,kx , y,ky are the corresponding integration points, then not the exact
Galerkin matrix A, but its quadrature approximation A will be further ap
proximated by ACA. The matrix A is, corresponding to (3.25), a nite sum
of matrices as dened in (3.18), multiplied by degenerated diagonal matrices.
Therefore, Theorem 3.5 remains valid for Galerkin matrices, if the relative
accuracy of the numerical integration (3.25) is higher than the approximation
of the matrix A by ACA.
On the matrix level, all three algorithms formulated in Section 3.2.1 can be
written in the fully pivoted ACA form.
R0 = A , S0 = 0 .
2. For i = 0, 1, 2, . . . compute
2.1. pivot element
In Algorithm 3.8, ej denotes the jth column of the identity matrix I. The
whole residual matrix Ri is inspected in Step 2.1 of Algorithm 3.8 for its
maximal entry. Thus, its Frobenius norm can easily be computed in this step,
and the appropriate stopping criterion for a given > 0 at step r would be
Rr F AF .
Note that the crosses built from the columnrow pairs with the indices ki , i
for i = 1, . . . , r will be computed exactly
Sm = aki ,l , l = 1, ..., M ,
k ,l
i
Sm = ak,li , k = 1, ..., N
k,li
Table 3.1. Fully Pivoted ACA algorithm for the function (3.6)
0 8 16 24 32
0 0
8 8
100 16 16
75
30
50
25
0 20 24 24
10
10
20
32 32
30 0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
0.2
0.1 30
0
20 24 24
10
10
20
32 32
30 0 8 16 24 32
8 8
0.0004 16 16
0.0002
30
0
0.0002 20
24 24
10
10
20
30 32 32
0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
24 24
32 32
0 8 16 24 32
(cf. (3.8)), then, as we have already seen, the situation changes. The results of
the computations are presented in Table 3.2 and in the four Figs. 3.143.17.
The convergence of the ACA algorithm is now slow, the crosses chosen can
not approximate the main diagonal, because they are too small there. This
illustrates once again the necessity of the hierarchical clustering.
The next function we consider,
is degenerated corresponding to the denition (3.9), having the exact low rank
r = 7. This function is obviously innitely smooth. But it is oscillating and
the convergence of Algorithm 3.8 is slow again. The convergence is, of course,
better than for the singular function (3.26), but not really sucient. However,
after exactly 7 iterations, the error is equal to computer zero. It means that
Algorithm 3.8 has correctly detected the low rank of the function (3.27). The
numerical results can be seen in Table 3.3 and in the four Figs. 3.183.21,
3.2 Block Approximation Methods 123
Table 3.2. Fully Pivoted ACA algorithm for the function (3.8)
0 8 16 24 32
0 0
8 8
100 16 16
75
30
50
25
0 20 24 24
10
10
20
32 32
30 0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
75
50 30
25
0 20 24 24
10
10
20
32 32
30 0 8 16 24 32
8 8
16 16
60
40 30
20
0
20 24 24
10
10
20
32 32
30 0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
20
10 30
0
20 24 24
10
10
20
32 32
30 0 8 16 24 32
where the initial residual and three sequential residuals (for k = 2, 4, and
k = 6) are shown.
Table 3.3. Fully Pivoted ACA algorithm for the function (3.27)
8 8
1 16 16
0.75
30
0.5
0.25
0 20 24 24
10
10
20
32 32
30 0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
0.5 30
0
20 24 24
10
10
20
32 32
30 0 8 16 24 32
0 8 16 24 32
0 0
8 8
16 16
0.2
30
0
0.2 20 24 24
10
10
20
32 32
30 0 8 16 24 32
8 8
16 16
0.1
0.05
30
0
0.05
0.1 20 24 24
10
10
20
32 32
30
0 8 16 24 32
If the matrix A has not yet been generated, but if there is a possibility of
generating its entries ak individually, then the following partially pivoted
ACA algorithm can be used for the approximation:
Algorithm 3.9
1. Initialisation
S0 := 0 , I := , J := , c := 0 RN , r := 0 RM .
#I = N or #J = M
ki+1 := min {k : k
/ I} , CrossT ype := Row ,
3. Generate cross
3.1 Type of the cross
If
then
3.1.1 Generate row, Update control vector
3.1.2 Test
If a = 0 then GOTO 2. (zero row)
3.2 Block Approximation Methods 127
i
rv := a (um )ki+1 vm ,
m=1
i+1 := ArgMax (rv )  ,
3.1.4 Test
If rv  = 0 then GOTO 4. (linear depending row)
3.1.5 Normalising constant
i+1 := (rv )1
i+1 ,
i
ru := b (vm )i+1 um ,
m=1
ki+2 := ArgMax (ru )k  ,
else
3.2.1 Generate column, Update control vector
3.2.2 Test
If b = 0 then GOTO 2. (zero column)
3.2.3 Column of the residual and the pivot row
i
ru := b (vm )i+1 um ,
m=1
ki+1 := ArgMax (ru )k  ,
3.2.4 Test
If ru  = 0 then GOTO 4. (linear depending column)
3.2.5 Normalising constant
i+1 := (ru )1
ki+1 ,
128 3 Approximation of Boundary Element Matrices
i
rv := a (um )ki+1 vm ,
m=1
i+2 := ArgMax (rv )  ,
i
Si+1 2F = Si 2F +2 u
i+1 um vm vi+1 + ui+1 F vi+1 F .
2 2
m=1
3.5 Test
If ui+1 F vi+1 F < Si+1 F
then GOTO 4
else i := i + 1, GOTO 3
4. Check control vectors
If i
/ I and ci = 0 then
or
If j
/ J and rj = 0 then
else STOP
Algorithm 3.9 starts to compute an approximation for the matrix A by gen
erating its rst row. Then, the rst column will be chosen automatically. If
a cross is successfully computed, the next row index is prescribed, and the
procedure repeats. If a zero row is generated, then it is not possible to nd
the column, and the algorithm restarts in Step 2. Since the matrix A will not
be generated completely, we can use the norm of its approximant Si to dene
a stopping criterion. This norm can be computed recursively as it is described
in Step 3.4. However, since the whole matrix A will not be generated while
3.3 Bibliographic Remarks 129
using the partially pivoted ACA algorithm, it is necessary to check the control
vectors c and r before stopping the algorithm. Note that these vectors contain
the sums of absolute values of all elements generated. If, for example, there
is some index i / I with ci = 0 then the row i has not yet contributed to
the matrix. It can happen that this row contains relevant information, and,
therefore, we have to restart the algorithm in Step 3. The same argumen
tation is valid for the columns. The only dierence is, that the crosses after
this restart will be generated on the dierent way: rst prescribed column and
then automatically chosen row. Thus, Algorithm 3.9 can be used not only for
dense matrices but also for reducible, and even for sparse matrices containing
only few nonzero entries.
Algorithm 3.9 requires only O(r2 (N + M )) arithmetical operations and
its memory requirement is O(r(N + M )). Thus, this algorithm is perfect for
large matrices. All approximations of boundary element matrices of the next
chapter will be generated with the help of Algorithm 3.9.
In this section we describe some surfaces which will be used for numerical
examples in the following sections. We show the geometry of these surfaces
and give the number of elements and nodes. Furthermore, the corresponding
cluster structures will be shown.
The most simple smooth surface = for R3 is the surface of the unit
sphere,
= x R3 : x = 1 . (4.1)
Now we consider the TEAM problem 10 (cf. [75]). TEAM is an acronym for
Testing Electromagnetic Analysis Methods, which is a community that creates
benchmark problems to test nite element analysis software. An exciting coil
is set between two steel channels and a thin steel plate is inserted between the
channels. Thus, the domian consists of four disconnected parts. The coarsest
132 4 Implementation and Numerical Examples
1 1
0.5 0.5
0 0
0.5 0.5
1 1
1
1 1
0.5
0.5
0
0
0.5
0.5
1
1 1
0.5
0.5 0
0 0.5
0.5 1
1 1
0.5 0.5
0 0
0.5 0.5
1
1 1
1
1 1
0.5 0.5
0 0
0.5 0.5
1 1
1 1
0.5 0.5
0 0
0.5 0.5
1 1
Fig. 4.2. Discretisation of the unit sphere for N = 1280 and N = 5120
1
0.5
0
0.5
1
1
0.5
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0.5
1
1
0.5
0
0.5
1
Fig. 4.4. An admissible cluster pair for N = 1280
The test rig consists of a rotor and a stator (cf. [92]). The stator poles are tted
with coils as shown in Fig. 4.7. The rotor is locked at 22 with respect to the
stator, providing only a small overlap between the poles. The coarsest mesh of
this model is shown in Fig. 4.7 and contains N = 4640 elements. We perform
two uniform mesh renements in order to get meshes with N = 18560 and
N = 74240 elements, respectively. This model consists of four independent
parts. In Fig. 4.8 the clusters of the levels 1 and 2 for N = 4928 are shown.
4.1.4 Relay
The simply connected domain shown in Fig. 4.9 will be considered as model
for a relay. The speciality of this domain is the small air gap between the
kernel and the armature. Its surface contains N = 4944 elements. We perform
two uniform mesh renements in order to get meshes with N = 19776 and
N = 79104 elements, respectively. In Fig. 4.10 the corresponding clusters can
be seen.
4.2 Laplace Equation 135
50
100
0
50
0
100
100
50
0
50 100
100
u(x) = 0 (4.2)
50
100
0
50
0
100
100
50
0
50 100
100
50
100
0
50
0
100
100
50
0
50 100
100
Fig. 4.6. Clusters of the level 1 and 2, TEAM problem 10 for N = 4928
4.2 Laplace Equation 137
40
20
100
0
20 50
40
100
100 0
50
0 50
50
100
100
k1 ,k2 ,k3 (x) = exp k1 x1 + k2 x2 + k3 x3 , k12 + k22 + k32 = 0 ,
0,k2 ,k3 (x) = (a + b x1 ) exp k2 x2 + k3 x3 , k22 + k32 = 0 , (4.3)
0,0,0 (x) = (a1 + b1 x1 )(a2 + b2 x2 )(a3 + b3 x3 ) .
We solve the interior Dirichlet, Neumann and mixed boundary value problems,
as well as an interface problem using a Galerkin boundary element method
138 4 Implementation and Numerical Examples
40
20
100
0
20 50
40
100
100 0
50
0 50
50
100
100
40
20
100
0
20 50
40
100
100 0
50
0 50
50
100
100
Fig. 4.8. Clusters of the level 1 and 2, TEAM problem 24 for N = 4640
(cf. Section 2). Piecewise linear basis functions will be used for the approx
imation of the Dirichlet datum 0int u and piecewise constant basis functions
k for the approximation of the Neumann datum 1int u. We will use the L2
projection for the approximation of the given part of the Cauchy data. The
boundary element matrices Vh , Kh and Dh are generated in approximative
form using the partially pivoted ACA algorithm with a variable relative ac
curacy 1 . The resulting systems of linear equations are solved using some
variants of the Conjugate Gradient method (CGM) with or without precon
ditioning up to a relative accuracy 2 = 108 .
4.2 Laplace Equation 139
5
10
7.5
2.5
0
0
5
10
15
The most important matrices to be generated while using the Galerkin bound
ary element method are the single layer potential matrix Vh and the double
layer potential matrix Kh , having the entries
1 1
Vh [k, ] = dsy dsx for k, = 1, . . . , N , (4.5)
4 x y
k
and
1 (x y, n(y))
Kh [k, j] = j (y) dsy dsx (4.6)
4 x y3
k
and
140 4 Implementation and Numerical Examples
5
10
7.5
2.5
0
0
5
10
15
5
5
10
7.5
2.5
0
0
5
10
15
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
1 (x y, n(y))
Di (, x) = ,i (y) dsy for i = 1, 2, 3 . (4.8)
4 x y3
,i (xj ) = ij , j = 1, 2, 3 .
The explicit form of these functions can be seen in Appendix C.2. By the use
of the functions (4.7)(4.8), the matrix entries (4.5)(4.6) can be rewritten as
follows:
1
Vh [k, ] = S( , x) dsx + S(k , x) dsx (4.9)
2
k
for k, = 1, . . . , N and
Kh [k, j] = Di : xi ( )=xj (, x) dsx (4.10)
I(j) k
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
0.05
0.05
0.1 0.1
0
0
0.1
0.2
0 2
Fig. 4.12. Clusters of the level 1 and 2, Exhaust manifold for N = 2264
4.2 Laplace Equation 143
for k, = 1, . . . , N and
Ng
Kh [k, j] m Di : xi ( )=xj (, xk ,m ) (4.12)
I(j) m=1
Here we solve the Laplace equation (4.2) together with the boundary condition
0int u(x) = g(x) for x , where is a given surface. The variational problem
(1.16)
1
V t, w = I + K g, w for all w H 1/2 ( )
2
Unit sphere
as a test solution of the Laplace equation (4.2). The results of the computa
tions are shown in Tables 4.1 and 4.2. The number of boundary elements is
listed in the rst column of these tables. The second column contains the num
ber of nodes, while in the third column of Table 4.1, the prescribed accuracy
for the ACA algorithm for the approximation of both matrices Kh RN M
and Vh RN N is given. The fourth column of this table shows the memory
requirements in MByte for the approximate double layer potential matrix Kh .
The quality of this approximation in percentage of the original matrix is listed
144 4 Implementation and Numerical Examples
N M 1 MByte(Kh ) % MByte(Vh ) %
2
80 42 1.0 10 0.03 97.8 0.02 48.7
320 162 1.0 103 0.26 65.6 0.21 27.2
1280 642 1.0 104 2.45 39.1 1.94 15.5
5120 2562 1.0 105 20.05 20.0 15.72 7.9
20480 10242 1.0 106 149.19 9.3 115.83 3.6
81920 40962 1.0 107 1085.0 4.2 837.50 1.6
Fig. 4.13. Partitioning of the BEM matrices for N = 5120 and M = 2562
in the next column. The corresponding values for the single layer potential
matrix Vh can be seen in the columns six and seven. The partitioning of the
matrix for N = 5120 as well as the quality of the approximation of single
blocks is shown in Fig. 4.13. The left diagram in Fig. 4.13 shows the sym
metric single layer potential matrix Vh , while the rectangular double layer
potential matrix Kh is depicted in the right diagram. The legend indicates
the percentage of memory needed for the ACA approximation of the blocks
compared to the full memory. Further numerical results are shown in Table
4.2. The third column in Table 4.2 shows the number of Conjugate Gradient
iterations needed to reach the prescribed accuracy 2 . The relative L2 error
for the Neumann datum
1int u th L2 ( )
Error1 = (4.14)
1int uL2 ( )
is given in the fourth column. The next column represents the rate of con
vergence for the Neumann datum, i.e. the quotient between the errors in two
4.2 Laplace Equation 145
consecutive lines of column four. Finally, the last two columns show the ab
solute error (cf. (2.19)) in a prescribed inner point x ,
inside of the domain is shown in Fig. 4.16 for N = 80 (left plot) and
for N = 320 (right plot). The values of the numerical solution u and of the
analytical solution u have been computed in 512 points uniformly placed on
the line (4.16). The thick dashed line represents in these gures the course
of the analytical solution (4.13), while the thin solid line shows the course of
the numerical solution u. The values of the variable x1 along the line (4.16)
are used for the axis of abscissas. The next Fig. 4.17 shows these curves for
N = 1280 (left plot) and on the zoomed interval [0.2, 0.3] (right plot) with
respect to the variable x1 in order to see the dierence between them. It is
almost impossible to see any optical dierence between the numerical and
analytical curves for higher values of N . Note that the point x in (4.15) is
146 4 Implementation and Numerical Examples
5.8149E02
0.5
0 1.4856E02
1 0.5
0.5
1
0 1
0.5
0.5 0
0.5
1 1
2.8438E02
Fig. 4.14. Given Dirichlet datum for the unit sphere, N = 5120
1 3.6811E03
0.5
0
0.5
1
1
0.5
7.7897E02
0
0.5
1
1
0.5
0
0.5
1 2.1232E03
Fig. 4.15. Computed Neumann datum for the unit sphere, N = 5120
4.2 Laplace Equation 147
0
0
5 2.5
10 5
15 7.5
10
20
12.5
25
15
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.16. Numerical and analytical curves for N = 80 and N = 320, Dirichlet
problem
0 10
2.5
11
5
12
7.5
13
10
12.5 14
15 15
0.3 0.2 0.1 0 0.1 0.2 0.3 0.2 0.22 0.24 0.26 0.28 0.3
Fig. 4.17. Numerical and analytical curves for N = 1280, Dirichlet problem
chosen close to the minimum of the function u along the line, where the error
seems to reach its maximum.
TEAM Problem 10
The analytical solution is now taken in the form (4.4) with y = (0, 60, 50) .
The results of the computations are shown in Tables 4.3 and 4.4. The third
N M 1 MByte(Kh ) % MByte(Vh ) %
2
4928 2470 1.0 10 14.03 15.11 7.46 4.0
19712 9862 1.0 103 131.85 8.9 65.22 2.2
78848 39430 1.0 104 1190.00 5.0 604.56 1.3
D = diag   , = 1, . . . , N . (4.17)
148 4 Implementation and Numerical Examples
The last two columns of Table 4.4 show the absolute error (cf. (2.19)) in a
prescribed inner point x ,
for the value u(x ) obtained using the approximate representation formula
(2.18). All other entries in these tables have the same meaning as those dis
played in Tables 4.14.2. In Figs. 4.184.19 the given Dirichlet datum and the
computed Neumann datum for N = 4928 boundary elements and M = 2470
nodes are presented. The numerical curve obtained when using the approxi
8.4120E03
100
50
0
50
100
4.3880E03
50
0
0 100
10
50
0
100
3.6407E04
mate representation formula in comparison with the curve of the exact values
(4.4) along the line
4.2 Laplace Equation 149
4.6686E04
100
50
0
50
100
3.7515E05
50
0
0 100
10
50
0
100
3.9183E04
0.0 0.0
x(t) = 90.0 + t 0.0 , 0 t 1 (4.19)
49.99 99.98
inside of the domain is shown in Fig. 4.20 for N = 4928. The values of the
numerical solution u and of the analytical solution u have been computed in
512 points uniformly placed on the line (4.19). The thick dashed line represents
the course of the analytical solution (4.4), while the thin solid line shows the
course of the numerical solution u. The values of the variable x3 along the
line (4.19) are used for the axis of abscissas. The right plot in this gure
shows a zoomed picture on the interval [40, 49.99] with respect to the variable
x3 . Note that the end of the line (4.19) is very close to the boundary of the
domain, which lies at x3 = 50. Thus, the loss of accuracy of the numerical
representation formula close to the boundary can be clearly seen. The courses
of the numerical solutions obtained for N = 19712 (left plot) and N = 78848
(right plot) are shown in Fig. 4.21. They do not distinguish optically from the
course of the analytical solution on the whole interval. Thus, we show only
the zoomed pictures.
We consider the interior Neumann boundary value problem with the boundary
condition 1int u(x) = g(x) for x . The variational problem (1.29)
150 4 Implementation and Numerical Examples
0.0025 0.002675
0.00265
0.002 0.002625
0.0026
0.0015 0.002575
0.00255
0.001
0.002525
40 20 0 20 40 40 42 44 46 48 50
Fig. 4.20. Numerical and analytical curves for N = 4928, Dirichlet problem
0.00266
0.00264
0.00264
0.00262
0.00262
0.0026 0.0026
0.00258 0.00258
0.00256 0.00256
0.00254 0.00254
0.00252 0.00252
40 42 44 46 48 50 40 42 44 46 48 50
Fig. 4.21. Numerical and analytical curves for N = 19712 and N = 78848
1
Du, v + u, 1 v, 1 = I K g, v + v, 1
2
1/2
for all v H ( ), is discretised and leads to a system of linear equations
(cf. (2.31))
1
Dh + a a u = Mh Kh g + a ,
2
where the vector a RM contains the integrals of the piecewise linear basis
functions over the surface ,
1
a = supp  , = 1, . . . , M .
3
The symmetric and positive denite system is then solved using a Conjugate
Gradient method up to the relative accuracy 2 = 108 .
Unit Sphere
We consider again the harmonic function (4.13) as the exact solution. The
results for the ACA approximation of the matrix Dh RM M are presented
in Table 4.5. The corresponding results for the matrix Kh are identical to
those already presented in Table 4.1. Note that in this example, the Galerkin
matrix with piecewise linear basis functions for the hypersingular operator is
generated according to (1.9). The partitioning of the matrices Kh and Dh for
4.2 Laplace Equation 151
N M 1 MByte(Dh ) %
2
80 42 1.0 10 0.01 51.2
320 162 1.0 103 0.10 48.1
1280 642 1.0 104 1.02 32.3
5120 2562 1.0 105 8.67 17.3
20480 10242 1.0 106 64.75 8.09
81920 40962 1.0 107 446.13 3.49
Fig. 4.22. Partitioning of the BEM matrices for N = 5120 and M = 2562.
procedure is presented in Table 4.6. The numbers in this table have the same
meaning as in Table 4.2. The third column shows the number of iterations
required by the Conjugate Gradient method without preconditioning. Note
152 4 Implementation and Numerical Examples
that the convergence of the Galerkin method for the unknown Dirichlet datum
in the L2 norm
0int u u
h L2 ( )
Error1 = (4.20)
0int uL2 ( )
is now quadratic corresponding to the error estimate (2.32). Also in the inner
point x (cf. (4.15)) we now observe the quadratic convergence (7th column) as
it was predicted in (2.34) instead of the cubic order obtained for the Dirichlet
problem (cf. Table 4.2). This fact is clearly illustrated in Figs. 4.234.24, where
the convergence of the boundary element method can be seen optically. The
results obtained for N = 80 are plotted in Fig. 4.23 (left plot). The numerical
0 0
5
5
10
10
15
15
20
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.23. Numerical and analytical curves for N = 80 and N = 320, Neumann
problem
curve in Fig. 4.23 (right plot) is notedly better than the previous one. However,
its quality is not as high as the one of the corresponding curve obtained solving
the Dirichlet problem (cf. Fig. 4.16). The next Fig. 4.24 shows the same curves
for N = 1280 and for N = 5120. Here, we do not need to zoom the pictures in
order to see the dierence between the numerical and the analytical curves.
0 0
2.5 2.5
5 5
7.5 7.5
10 10
12.5 12.5
15 15
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.24. Numerical and analytical curves for N = 1280 and N = 5120, Neumann
problem
4.2 Laplace Equation 153
Exhaust Manifold
The analytical solution is taken in the form (4.4) with y = (0, 0, 0.06) . The
results of the computations are reported in Tables 4.7 and 4.8. The third
N M 1 MByte(Kh ) % MByte(Vh ) %
2264 1134 1.0 103 6.99 35.7 3.94 10.1
9056 4530 1.0 104 62.27 19.9 34.99 5.6
36224 18114 1.0 105 500.66 10.0 282.44 2.8
The last two columns of Table 4.8 show the absolute error (cf. (2.34)) in a
prescribed inner point x ,
Error2 = u(x ) u(x ) , x = (0.0112524, 0.1, 0.05) , (4.22)
for the value u(x ) obtained using an approximate representation formula
(2.33). All other entries in these tables have the usual meaning. The quadratic
convergence of the Dirichlet datum in the L2 norm as well as the quadratic
convergence (or even slightly better) in the inner point x can be observed
again.
In Figs. 4.254.26 the given Dirichlet datum and computed Neumann da
tum for N = 2264 boundary elements and M = 1134 nodes are presented.
The numerical curve obtained when using an approximate representation
formula in comparison with the curve of the exact values (4.4) along the line
0.05 0.2
x(t) = 0.1 + t 0.0 , 0 t 1 (4.23)
0.05 0.0
154 4 Implementation and Numerical Examples
4.4393E00
0.1
0.05
2.3679E00
0
0.05
0.1
0.1
0.2
0 2
2.9645E01
2.4101E02
0.05
1.0975E02
0
0.05
0.1
0.1
0
0.1 0
0.2
0 2
2.1509E01
inside of the domain is shown in Fig. 4.27 for N = 2254 and N = 9056,
while Fig. 4.28 shows the results obtained for N = 36224 (left plot). The right
plot in this gure presents the same curve on a zoomed interval [0.05, 0.05].
The values of the numerical solution u and of the analytical solution u have
been computed in 512 points uniformly placed on the line (4.23). The thick
dashed line represents the course of the analytical solution (4.4), while the
thin solid line shows the course of the numerical solution u. The values of the
variable x1 along the the line (4.23) are used for the axis of abscissas. Again,
a quite high accuracy of the Galerkin BEM can be observed.
0.525 0.525
0.5 0.5
0.475 0.475
0.45 0.45
0.425 0.425
0.4 0.4
0.375 0.375
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.27. Numerical and analytical curves for N = 4944 and N = 19776
0.535
0.525
0.53
0.5
0.475 0.525
0.45 0.52
0.425 0.515
0.4
0.51
0.375
0.05 0 0.05 0.1 0.15 0.04 0.02 0 0.02 0.04
The matrix of the single layer potential Vh and the matrix of the double
layer potential Kh are generated in an approximative form using the partially
pivoted ACA algorithm 3.9 with increasing accuracy. The system of linear
equations (2.41) is then solved using the Conjugate Gradient method for the
Schur complement system as in (2.37) up to the relative accuracy 2 = 108 .
Note that this realisation requires an additional solution of a linear system
with the single layer potential matrix in each iteration step. This system is
solved again using a Conjugate Gradient method up to the relative accuracy
2 = 108 . The matrix of the hypersingular operator is not generated explic
itly. Its multiplication with a vector is realised using the matrix of the single
layer potential as it is described in (2.27).
Unit Sphere
In the rst example, we prescribe the Dirichlet datum on the upper part of
the unit sphere D = {x : x3 0} and the Neumann datum on the lower
part N = {x : x3 < 0}. The analytical solution is taken in the form
(4.13) and the numerical results are presented in Tables 4.94.10. Of course,
Table 4.9. Accuracy of the Galerkin method on the boundary, mixed problem
the results for the ACA approximation of the matrices Vh and Kh are the
same as for the Dirichlet problem (cf. Table 4.1). The accuracy obtained for
the mixed boundary value problem is presented in Table 4.9. The numbers in
this table have the following meaning: The third column in Table 4.9 shows
the number of Conjugate Gradient iterations without preconditioning needed
to reach the prescribed accuracy 2 for the Schur complement, while the fourth
column indicates the numbers of Conjugate Gradient iterations without pre
conditioning needed in each iteration step for the system with the single layer
potential matrix. Thus, the total number of iterations when solving a mixed
4.2 Laplace Equation 157
boundary value problem is much higher than for solving a pure Dirichlet or
Neumann boundary value problem. The error for the Dirichlet datum and the
convergence factor are shown in columns 5 and 6, while the corresponding
error for the Neumann datum can be seen in columns 7 and 8. Note that
the convergence of the Galerkin method for the unknown Dirichlet datum in
the L2 norm (4.20) is quadratic (6th column), while the convergence of the
Neumann datum (4.14) is linear (8th column). Corresponding to Table 4.1,
the matrices Kh and Vh together with some additional memory will require
more than 2 Gbyte of memory for N = 81902. Thus, we are not able to store
both these matrices on a regular workstation. The error in the inner point x
Table 4.10. Accuracy of the Galerkin method in the inner point x , mixed problem
N M Error2 CF2
0
80 42 5.82 10
320 162 4.94 101 11.77
1280 642 1.43 101 3.46
5120 2562 3.68 102 3.88
20480 10242 9.31 103 3.95
(cf. (4.15)) can be observed in Table 4.10. Here, quadratic convergence can be
observed, at least asymptotically.
TEAM Problem 10
Here, we consider the mixed boundary value problem for the Laplace equation
in the domain presented in Fig. 4.5. The Dirichlet part of the boundary is
dened by D = {x : x3 = 50}. Thus, Dirichlet boundary conditions
are given only on the top of the coil. The analytical solution is taken in
the form (4.4) with y = (0.0, 60.0, 50.0)
/ and the numerical results are
presented in Table 4.11. The meaning of the values presented in this table is
Table 4.11. Accuracy of the Galerkin method on the boundary, TEAM problem
10
N M Iter1 Iter2 D Error1 CFD N Error1 CFN
2 1
4928 2470 1061 9095 3.04 10 6.04 10
19712 9862 1732 118123 7.60 103 4.00 2.03 101 2.98
the same as in Table 4.9. The third column in Table 4.11 shows the number of
Conjugate Gradient iterations with diagonal preconditioning (4.21) needed to
reach the prescribed accuracy 2 for the Schur complement, while the fourth
158 4 Implementation and Numerical Examples
TEAM Problem 24
Here, we consider the mixed boundary value problem for the Laplace equation
in the domain presented in Fig. 4.7. The Dirichlet part of the boundary is
dened by D = {x : x3 0}. Thus, the Dirichlet boundary condition is
given on the upper part of the symmetric surface . The analytical solution
is taken in the form (4.4) with y = (0.0, 80.0, 20.0)
/ and the numerical
results are presented in Table 4.12 The meaning of the values presented in this
Table 4.12. Accuracy of the Galerkin method on the boundary, TEAM problem
24
N M Iter1 Iter2 D Error1 CFD N Error1 CFN
2 1
4640 2320 48 7278 2.56 10 3.08 10
18560 9280 71 93102 4.67 103 5.48 1.48 101 2.08
table is the same as in Table 4.11. We have used the same preconditioning as
in the previous example. The number of iterations reported in the third and
in the fourth columns of the table is now much less. In Figs. 4.294.30 the
Dirichlet datum and the Neumann datum for N = 4944 boundary elements
and M = 2474 nodes are presented. The numerical curve obtained when
using an approximate representation formula in comparison with the curve of
the exact values (4.4) along the line
0.0 0.0
x(t) = 100.0 + t 40.0 , 0 t 1 (4.25)
0 0
inside of the domain is shown in Fig. 4.31 for N = 4640. The values of the
numerical solution u and of the analytical solution u have been computed in
512 points uniformly placed on the line (4.25). The thick dashed line represents
the course of the analytical solution (4.4), while the thin solid line shows the
course of the numerical solution u. The values of the variable x2 along the
the line (4.25) are used for the axis of abscissas. The course of the numerical
approximation does not optically distinguish from the exact solution on the
left plot of Fig. 4.31. Thus, we show the zoom of these curves on the interval
[84, 76] (right plot).
4.2 Laplace Equation 159
1.1279E02
40 5.8440E03
20
100
0
20 50
40
100
100 0
50
0 50
50
100
100
4.0923E04
1.3945E03
40 5.5593E04
20
100
0
20 50
40
100
100 0
50
0 50
50
100
100
2.8372E04
0.004 0.00398
0.0038
0.00396
0.0036
0.0034 0.00394
0.0032
0.00392
0.003
Fig. 4.31. Numerical and analytical curves for N = 4640, Mixed problem
The purpose of this subsection is twofold. The rst goal is to illustrate the
numerical solution of the Poisson equation by the use of a particular solution
(cf. Subsection 1.1.7). The second goal is to illustrate the numerical solution
for an interface problem by Boundary Element Methods (cf. Subsection 1.1.8).
We consider the following interface problem
0int ui (x) = 0ext ue (x) , i 1int ui (x) = e 1ext ue (x) for x , (4.27)
then the above interface problem can be reformulated as follows (cf. Subsection
1.1.8). Introduce a new unknown function ui by
ui = ui + upi
Then, by the use of the interior and exterior SteklovPoincare operators S int
and S ext (cf. (1.13), (1.46))
1int ui = S int 0int ui , 1ext ue = S ext 0ext ue ,
we rewrite the interface problem as (cf. (1.59))
i S int + e S ext 0int ui = i 1int upi e S ext 0int upi . (4.28)
Once the Dirichlet datum 0int ui is found, we solve the interior Dirichlet
boundary value problem for the Neumann datum 1int ui . The Cauchy data
for the unknown functions ui and ue are then obtained via
e ext
0int ui = 0ext ue = 0int ui + 0int upi , 1int ui = ue = 1int ui + 1int upi .
i 1
Unit Sphere
Let be the surface of the unit sphere (4.1). The constants i , e and the
right hand side fi in (4.26) are
i = e = 1 , fi (x) = 1 , for x .
The exact solution of this simple model problem is
3 x2 1
ui (x) = , x, ue (x) = , x e . (4.29)
6 3 x
Consider the function
1
upi (x) = x21 , for x
2
as a particular solution of the Poisson equation.
The Galerkin method with piecewise linear basis functions for the
Dirichlet data 0int ui = 0ext ue and 0int ui and with piecewise constant ba
sis functions k for the Neumann data 1int ui = 1ext ue and 1int ui will be
used. The matrix of the single layer potential Vh and the matrix of the double
layer potential Kh are generated in an approximative form using the par
tially pivoted ACA algorithm 3.9 with increasing accuracy 1 . The resulting
system of linear equations (cf. (2.45)) is then solved using the Conjugate Gra
dient method without preconditioning up to the relative accuracy 2 = 108 .
The accuracy obtained for the analytical solution (4.29) is presented in Table
4.13. The numbers in this table have the following meaning. The third column
in Table 4.13 shows the number of Conjugate Gradient iterations needed to
reach the prescribed accuracy 2 for the linear system (2.45). The error for
the Dirichlet datum and the convergence factor are shown in columns 4 and 5,
while the corresponding error for the Neumann datum can be seen in columns
6 and 7. Note that the convergence of the Galerkin method for the unknown
Dirichlet datum in the L2 norm (4.20) is quadratic (5th column), while the
convergence of the Neumann datum (4.14) is linear (7th column).
162 4 Implementation and Numerical Examples
Table 4.13. Accuracy of the Galerkin method on the boundary, interface problem
In this section we consider two numerical examples for the mixed boundary
value problem of linear elastostatics (cf. (1.79))
3
ij (u, x) = 0 for x , i = 1, 2, 3,
j=1
x j
(4.30)
0int u(x) = g(x) for x D ,
The most important matrices to be generated while using the Galerkin bound
ary element method for the mixed boundary value problem (4.30) are the
single layer potential matrix VhLame and the double layer potential matrix
KhLame (cf. 2.4), having the representation (2.48) and (2.51), respectively.
Thus, in addition to the single and double layer potential matrices (Vh and
Kh ) for the Laplace operator, six additional dense matrices Vij,h RN N for
1 i j 3 have to be generated corresponding to (cf. (2.50))
1 (xi yi )(xj yj )
Vij,h [k, ] = dsy dsx .
4 x y3
k
1
Vij,h [k, ] = Sij ( , x) dsx + Sij (k , x) dsx .
2
k
The explicit form of the functions Sij can be seen in Appendix C.2.3. The
remaining integrals in the above symmetric form of the matrix entries Vij,h
can be computed numerically using a 7point quadrature rule, see Appendix
C.1.
4.3.2 Relay
The geometry of the domain is shown in Fig. 4.32. The bottom of the relay
is chosen to be the Dirichlet part D of the boundary and the boundary
condition is homogeneous, i.e.
( )
0int u(x) = 0 , for x D = x : x3 = 0 .
We choose the Young modulus E = 114 000 and the Poisson ratio = 0.24
that correspond to the values of steel. The original domain is shown in Fig.
4.32 for N = 4944 The matrix of the single layer potential Vh for the Laplace
operator (cf. (2.49)), six matrices of the single layer potential Vh for the Lame
operator (cf. (2.50)), and the matrix of the double layer potential Kh (cf.
(2.51)) for the Laplace operator are generated in an approximative form using
the partially pivoted ACA algorithm 3.9 with increasing accuracy. The system
of linear equations is then solved using the Conjugate Gradient method for
the Schur complement of the system (cf. (2.47)) up to the relative accuracy
2 = 108 . Note that this realisation requests an additional solution of a
linear system with the single layer potential matrix in each iteration step. This
system is solved again using Conjugate Gradient method up to the relative
accuracy 2 = 108 . The matrix of the hypersingular operator is not generated
explicitly. Its multiplication with a vector is realised using the matrix of the
single layer potential as it is described in Section 2.4.
The results of the approximation are presented in Tables 4.144.16. The
number of boundary elements is listed in the rst column of these tables. The
second column contains the number of nodes, while the prescribed accuracy
for the ACA algorithm for the approximation of all matrices Kh RN M
and Vh , Vk,h RN N , k, = 1, 2, 3 is given in the third column. The pairs of
further columns of these tables show the memory requirements in MByte and
the percentage of memory compared to the original matrix. The deformed
164 4 Implementation and Numerical Examples
5 0
5
0 10
15
5
15
10
N M 1 Vh % Kh %
4944 2474 1.0 104 37.24 20.0 52.96 56.8
19776 9890 1.0 105 258.65 8.7 326.45 10.9
Table 4.15. ACA approximation of the Galerkin matrices V11,h , V12,h and V13,h
Table 4.16. ACA approximation of the Galerkin matrices V22,h , V23,h and V33,h
5 0 5.3390E01
5
0 10
15
5
15
10
2.5840E01
1.7101E02
N M Iter1 Iter2
4944 2474 286 2628
19776 9890 368 2529
166 4 Implementation and Numerical Examples
domain can be seen from the same point of view in Fig. 4.33. In this gure,
the real deformation is amplied by a factor 10. In Table 4.17, the number
of iterations required by the Conjugate Gradient method is shown. The third
column of this table shows the number of iterations for the Schur comple
ment equation (2.47), while the fourth column shows the number of iterations
required for the iterative solution of the linear system for the single layer po
tential in each iteration step. The required accuracy was 2 = 108 for both
systems.
4.3.3 Foam
The geometry of the domain, which is a model for a metal foam, is shown
in Fig. 4.34. The speciality of this domain is its multiple connectivity and
rather small volume compared to its surface. There is only one discretisation
of the domain with N = 28952 surface elements. The bottom and the top of
the foam are chosen to be the Dirichlet part D of the boundary , and the
boundary condition is homogeneous on the bottom, i.e.
We choose the Young modulus E = 114 000 and the Poisson ratio = 0.24
that correspond to the values of steel. The original domain is shown in Fig.
4.34 for N = 28952. The matrix of the single layer potential Vh for the Laplace
operator (cf. (2.49)), six matrices of the single layer potential Vh for the Lame
operator (cf. (2.50)), and the matrix of the double layer potential Kh (cf.
(2.51)) for the Laplace operator are generated in an approximative form us
ing the partially pivoted ACA algorithm 3.9. The system of linear equations
is then solved using a Conjugate Gradient method for the Schur complement
system (cf. (2.47)) up to the relative accuracy 2 = 108 . Note that this real
isation requires an additional solution of a linear system with the single layer
potential matrix in each iteration step. This system is solved again using a
Conjugate Gradient method up to the relative accuracy 2 = 108 . The ma
trix of the hypersingular operator is not generated explicitly. Its multiplication
with a vector is realised using the matrix of the single layer potential as it is
described in Section 2.4.
The results of the approximation are presented in Tables 4.184.20. The
number of boundary elements is listed in the rst column of these tables.
4.3 Linear Elastostatics 167
30 0
10
20 20
30
10
40
30
20
10
The second column contains the number of nodes while in the third column
the prescribed accuracy for the ACA algorithm for the approximation of all
matrices Kh RN M and Vh , Vk,h RN N , k, = 1, 2, 3 is given. The
pairs of further columns of these tables show the memory requirements in
MByte and the percentage of memory compared to the original matrix. The
N M 1 Vh % Kh %
4
28952 14152 1.0 10 260.66 4.1 496.58 15.9
deformed domain can be seen from the same point of view in Fig. 4.35. In
this gure, the real deformation is amplied by a factor 100. The number of
iterations required by the Conjugate Gradient method is shown in Table 4.21.
In this table, the third column shows the number of iterations for the Schur
complement equation (2.47), while the number of iterations required for the
168 4 Implementation and Numerical Examples
Table 4.19. ACA approximation of the Galerkin matrices V11,h , V12,h and V13,h
Table 4.20. ACA approximation of the Galerkin matrices V22,h , V23,h and V33,h
30 0 1.0000E01
10
20 20
30
10
0
40
30
4.4986E02
20
10
1.0028E02
iterative solution of the linear system for the single layer potential in each
iteration step can be seen in the fourth column. The required accuracy was
2 = 108 for both systems.
N M Iter1 Iter2
28952 14152 253 1921
1 e xy
u (x, y) = (4.33)
4 x y
can be considered as a particular solution of the Helmholtz equation (4.31)
for both, interior (x , y e = R3 \ ), and exterior (x e , y )
boundary value problems.
We solve the interior and exterior Dirichlet and Neumann boundary value
problems using a Galerkin boundary element method (cf. Section 2). Piecewise
linear basis functions will be used for the approximation of the Dirichlet
datum 0int u and piecewise constant basis functions k for the approximation
of the Neumann datum 1int u. We will use the L2 projection for the approxi
mation of the given part of the Cauchy data. The boundary element matrices
Vh , Kh , and Ch are generated in an approximative form using the complex
valued version of the partially pivoted ACA algorithm 3.9 with a variable rel
ative accuracy 1 . The resulting systems of linear equations are solved using
the GMRES method with or without preconditioning up to a relative accuracy
2 = 108 .
170 4 Implementation and Numerical Examples
The most important matrices to be generated while using the Galerkin boun
dary element method for boundary value problems for the Helmholtz equation
(4.31) are the single layer potential matrix V,h (cf. (2.53)),
1 exy
V,h [k, ] = dsy dsx ,
4 x y
k
Furthermore, when solving the Neumann boundary value problem for the
Helmholtz equation, the matrix of the hypersingular operator (cf. (2.62))
D,h [i, j] =
xy
1 e
(curl j (y), curl i (x))dsy dsx
4 x y
2
exy
j (y)i (x)(n(x), n(y))dsy dsx
4 x y
To generate the entries of the single layer potential matrix V,h numerically,
we rewrite (2.53) as follows:
1 exy 1
V,h [k, ] = V0,h [k, ] + dsy dsx .
4 x y
k
In the above, the entries V0,h [k, ] are the entries of the single layer potential
matrix of the Laplace operator, and, therefore, can be computed as it was
discussed in Section 4.2.3. The remaining double integral has no singularity
for x y, and can be computed numerically, using the 7point quadrature
rule (cf. Appendix C.1) for each triangle.
4.4 Helmholtz Equation 171
For the double layer potential matrix, the same idea leads to the following
decomposition:
K,h [k, j] =
(x y, n(y))
1
K0,h [k, j] + (1 x y)e xy 1 j (y)dsy dsx .
4 x y3
k
Again, the rst part of this decomposition belongs to the double layer potential
matrix of the Laplace operator, while the second part has no singularity for
x y. Thus, the 7point quadrature rule can be applied again. However, the
numerical integration with respect to the variable y has to be done over all
triangles in the support of the basis function j for each integration point
with respect to the variable x. Therefore, the generation of the matrix entries
for the double layer potential matrix for the Helmholtz equation is by far more
complicated than for the Laplace equation.
However, the most complicated numerical procedure is required when gene
rating the entries of the matrix C,h corresponding to (4.34). Using the same
decomposition as for the previous matrices, we get
xy
e 1
C,h [i, j] = C0,h [i, j] + j (y)i (x)(n(x), n(y))dsy dsx .
x y
In the above, the second summand has no singularity for x y, and the 7
point quadrature rule can be applied again. Note that in this case, the quadra
ture rule has to be applied to every triangle in the support of the function
i , and, for each of its integration points, to every triangle in the support of
the function j . Furthermore, a symmetrisation is necessary in order to keep
the symmetry of the matrix C,h . Fortunately, the rst summand C0,h [i, j]
does not require some additional analytical work. Since the normal vectors
n(x) and n(y) are constant within the single triangles in the supports of the
functions i and j , these integrals can be computed using the symmetrised
combination of the analytical integration and of the 7point quadrature rule.
Here we solve the Helmholtz equation (4.31) together with the Dirichlet
boundary condition 0int u(x) = g(x) for x , where is a given surface.
The variational problem (1.106)
1
V t, w
= I + K g, w for all w H 1/2 ( )
2
The matrix V,h of this system is symmetric. This property can be used in
order to save computer memory while generating the matrix. However, the
matrix is not selfadjoint, and, therefore, the Conjugate Gradient method can
not be used. Thus, for an iterative solution of the system (4.35), the complex
GMRES method will be used instead.
Unit Sphere
which satises the Helmholtz equation (4.31) for = 2. The results of the
computations for this rather moderate wave number are shown in Tables 4.22
and 4.23. The number of boundary elements is listed in the rst column of
Table 4.22. ACA approximation of the Galerkin matrices K,h and V,h
N M 1 MByte(Kh ) % MByte(Vh ) %
2
80 42 1.0 10 0.05 99.9 0.05 50.4
320 162 1.0 103 0.66 84.0 0.55 35.0
1280 642 1.0 104 6.08 48.5 5.02 20.1
5120 2562 1.0 105 49.05 24.5 39.46 9.86
20480 10242 1.0 106 357.90 11.2 280.60 4.47
these tables. The second column contains the number of nodes, while in the
third column of Table 4.22 the prescribed accuracy for the ACA algorithm for
the approximation of both matrices K,h CN M and V,h CN N is given.
In this table, the fourth column shows the memory requirements in MByte
for the approximate double layer potential matrix K,h . The quality of this
approximation in percentage of the original matrix is listed in the next column,
whereas the corresponding values for the single layer potential matrix V,h can
be seen in the columns six and seven. The third column in Table 4.23 shows
the number of GMRES iterations needed to reach the prescribed accuracy 2 ,
while the relative L2 error for the Neumann datum,
1int u th L2 ( )
Error1 = ,
1int uL2 ( )
is given in the fourth column. The next column represents the rate of conver
gence for the Neumann datum, i.e. the quotient of the errors in two consecutive
lines of column four. Finally, the last two columns show the absolute error (cf.
(2.56)) in a prescribed inner point x ,
4.4 Helmholtz Equation 173
1 4.0924E01
0.5
0
0.5
1
1
0.5
0.0000E00
0
0.5
1
1
0.5
0
0.5
1 4.0924E01
Fig. 4.36. Given Dirichlet datum (real part) for the unit sphere, N = 1280
1 3.0617E01
0.5
0
0.5
1
1
0.5
4.4670E01
0
0.5
1
1
0.5
0
0.5
1 3.1510E01
Fig. 4.37. Given Dirichlet datum (imaginary part) for the unit sphere, N = 1280
4.4 Helmholtz Equation 175
1 1.6062E02
0.5
0
0.5
1
1
0.5
0.0000E00
0
0.5
1
1
0.5
0
0.5
1 1.6062E02
Fig. 4.38. Computed Neumann datum (real part) for the unit sphere, N = 1280
1 1.3744E02
0.5
0
0.5
1
1
0.5
2.1111E00
0
0.5
1
1
0.5
0
0.5
1 1.3322E02
Fig. 4.39. Computed Neumann datum (imaginary part) for the unit sphere, N =
1280
176 4 Implementation and Numerical Examples
0 6
1 5
2 4
3 3
4
2
5
1
6
0
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.40. Numerical and analytical curves for N = 80, Dirichlet problem
0 3
1 2.5
2 2
3 1.5
4
1
5
0.5
6
0
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.41. Numerical and analytical curves for N = 320, Dirichlet problem
1
3
2
2.8
3
4 2.6
5
2.4
6
0.2 0.22 0.24 0.26 0.28 0.3 0.2 0.22 0.24 0.26 0.28 0.3
Fig. 4.42. Numerical and analytical curves for N = 1280, Dirichlet problem
ing linear systems, and, of course, the accuracy of the whole procedure are of
special interest. We will now solve the Helmholtz equation for a xed discreti
sation of the surface but for a sequence of wave numbers [min , max ].
If Im = 0, then the inner Dirichlet boundary value problem is uniquely
solvable. The situation is dierent for Im = 0. In this case the uniqueness
holds only if 2 is not an eigenvalue of the Laplace operator subjected to
homogeneous Dirichlet boundary conditions,
(cf. Section 1.4). For general , the eigenvalues of the Laplace operator are
not known and it can happen that one or even several of them belong to the
interval [min , max ]. In this case some diculties will occur when solving
the discrete problem. Now we are going to illustrate the situation. The exact
eigenfunctions and eigenvalues on the unit ball are analytically known and
can be represented in spherical coordinates
cos sin
x = sin sin , 0 < 1 , 0 < 2 , 0
cos
as follows
Jn+1/2 (n,m )
uk,n,m (, , ) = Pn,k (cos ) e k , (4.39)
with
m N, n N0 , k n .
In (4.39), n,m are the zeros of the Bessel functions Jn+1/2 . Pn,k are the
associated Legendre polynomials
k/2 dk
Pn,k (u) = (1)k 1 u2 Pn (u)
dxk
dened for
u 1 , k = 0, . . . , n , n N0 .
The Legendre polynomials Pn are given in (3.12). The corresponding eigen
values are
n,m = 2n,m .
For n = k = 0 and m N, the eigenvalues and the eigenfunctions are of an
especially simple form. In this case we use
/
2 sin z
J1/2 (z) =
z
and obtain
178 4 Implementation and Numerical Examples
sin(0,m )
u0,0,m (, , ) = cm , m N. (4.40)
Thus, the corresponding critical values of are
= 0,m = m , m N.
In particular, = is a critical value.
We solve the boundary value problem for the Dirichlet boundary value
problem for the Helmholtz equation (4.31) having the analytical solution
(4.33) with y = (1.1, 0.0, 0.0)
/ . We will use 17 uniformly distributed val
ues of on the interval [3.1, 3.2]. The discretisation of the boundary will be a
polyhedron having N = 320 boundary elements (cf. Fig. 4.1). The following
gures illustrate the results: In Fig. 4.43, the L2 error of the Neumann datum
is shown as a function of in the left plot. The right plot shows the number
of GMRES iterations needed to reach the relative accuracy 2 = 108 of the
numerical solution of the linear system (2.57). Thus, a signicant jump of the
42
0.74
41
0.735
40
0.73
39
0.725
38
3.1 3.12 3.14 3.16 3.18 3.2
3.1 3.12 3.14 3.16 3.18 3.2
0.02 0.01
0 0.02
0.02 0.03
0.04
0.04
0.06
0.05
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.44. Numerical and analytical curves for = 3.15 and N = 320, Dirichlet
problem
0.01
0.02
0.02
0
0.02 0.03
0.04 0.04
0.06
0.05
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.45. Numerical and analytical curves for = 3.175 and N = 320, Dirichlet
problem
0.02
0.02
0
0.03
0.02
0.04 0.04
0.06
0.05
0.08
0.3 0.2 0.1 0 0.1 0.2 0.3 0.3 0.2 0.1 0 0.1 0.2 0.3
Fig. 4.46. Numerical and analytical curves for = 3.2 and N = 320, Dirichlet
problem
In the next example, we solve the Dirichlet boundary value problem on the
polyhedron h with N = 1280 elements for 65 values of the wave number
uniformly distributed on the interval [0, 16]. Thus, the rst value corresponds
to the Laplace operator.
In Figs. 4.474.48 we show how the ACA approximation quality of the
matrices K,h and V,h depends on the wave number. The left plots in these
gures present the memory requirements in MByte, while the right plots show
the same result in percentage compared to the full memory for 1 = 104 .
The linear dependence of the memory requirement of the wave number is
180 4 Implementation and Numerical Examples
clearly indicated by these numerical tests. Also this example shows the loss
100
11
80
10
60
9
8 40
7 20
6
0 2.5 5 7.5 10 12.5 15 0 2.5 5 7.5 10 12.5 15
Fig. 4.47. Approximation of the double layer potential matrix K,h for N = 1280
100
9
80
8
60
7
40
6
20
5
Fig. 4.48. Approximation of the single layer potential matrix V,h for N = 1280
of the accuracy close to the critical values of the parameter . In Fig. 4.49, we
present again the L2 norm of the error for the Neumann datum (left plot) and
the number of GMRES iterations (right plot) as functions of the wave number
. The left plot in Fig. 4.49 clearly shows a total loss of accuracy close to the
1.6
1.4
200
1.2
150
1
0.8
100
0.6
50
0.4
0 2.5 5 7.5 10 12.5 15 0 2.5 5 7.5 10 12.5 15
wave number = 13.0. If we plot the analytical and the numerical values
4.4 Helmholtz Equation 181
of the solution of the boundary value problem for three subsequent points
12.75, 13.0 and 13.25 for the parameter we can see this loss of accuracy
optically. The results are presented in Figs. 4.504.52. It is remarkable that
0.06 0.08
0.04 0.06
0.04
0.02
0.02
0
0
0.02
0.02
0.04
0.04
0.06
0.06
0.4 0.2 0 0.2 0.4 0.4 0.2 0 0.2 0.4
Fig. 4.50. Numerical and analytical curves for = 12.75 and N = 1280, Dirichlet
problem
0.04 0.1
0.02 0.05
0
0
0.02
0.05
0.04
0.06 0.1
0.4 0.2 0 0.2 0.4 0.4 0.2 0 0.2 0.4
Fig. 4.51. Numerical and analytical curves for = 13.0 and N = 1280, Dirichlet
problem
0.04 0.06
0.04
0.02
0.02
0
0
0.02
0.02
0.04
0.04
0.06
0.06
0.4 0.2 0 0.2 0.4 0.4 0.2 0 0.2 0.4
Fig. 4.52. Numerical and analytical curves for = 13.25 and N = 1280, Dirichlet
problem
only one critical value (close to 4) of the wave number was detected on
182 4 Implementation and Numerical Examples
the interval [0, 16]. This fact is due to the rather big step of 0.25 with respect
to , which was used in the above example.
Exhaust Manifold
The analytical solution is taken in the form (4.33) with y = (0, 0, 0.06) and
= 80, which is moderate compared to the rather small dimension of the
domain (cf. Fig. 4.11). The results of the computations are shown in Tables
4.24 and 4.25. The third column shows the number of iterations required
Table 4.24. ACA approximation of the Galerkin matrices K,h and V,h
N M 1 MByte(K,h ) % MByte(V,h ) %
3
2264 1134 1.0 10 16.62 42.4 11.82 15.1
9056 4530 1.0 104 137.86 22.0 97.08 7.8
36224 18114 1.0 105 1046.80 10.5 696.65 3.5
1.2070E00
0.05
5.9464E01
0
0.05
0.1 0.1
0
0
0.1
0.2
0 2
2.3963E00
Fig. 4.53. Given Dirichlet datum (real part) for the exhaust manifold
4.7306E00
0.05
1.5706E00
0
0.05
0.1 0.1
0
0
0.1
0.2
0 2
1.5894E00
Fig. 4.54. Given Dirichlet datum (imaginary part) for the exhaust manifold
184 4 Implementation and Numerical Examples
3.6341E02
0.05
1.3407E02
0
0.05
0.1 0.1
0
0
0.1
0.2
0 2
9.5273E01
Fig. 4.55. Computed Neumann datum (real part) for the exhaust manifold
1.9587E02
0.05
6.7482E01
0
0.05
0.1 0.1
0
0
0.1
0.2
0 2
6.0909E01
Fig. 4.56. Computed Neumann datum (imaginary part) for the exhaust manifold
4.4 Helmholtz Equation 185
0.05 0.2
x(t) = 0.1 + t 0.0 , 0 t 1 (4.42)
0.05 0.0
inside of the domain is shown in Figs. 4.574.58 for N = 2264 and corre
spondingly for N = 9056. The values of the numerical solution u and of the
analytical solution u have been computed in 512 points uniformly placed on
the line (4.42). The thick dashed line represents the course of the analytical
solution (4.33) while the thin solid line shows the course of the numerical so
lution u. The values of the variable x1 along the line (4.42) are used for the
axis of abscissas. Note that the numerical solution for N = 9056 perfectly
0.4 0.4
0.2 0.2
0 0
0.2
0.2
0.4
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.57. Numerical and analytical curves for the exhaust manifold for N = 2264
0.4
0.4
0.2 0.2
0 0
0.2
0.2
0.4
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.58. Numerical and analytical curves for the exhaust manifold for N = 9056
Unit Sphere
We consider again the harmonic function (4.36) as the exact solution. The
results for the ACA approximation of the matrix C,h CM M are presented
in Table 4.26. The corresponding results for the matrices K,h for the compu
tation of the right hand side of the above system and V,h , which will be used
for the multiplication with the matrix D,h are identical to those already pre
sented in Table 4.22. Note that in this example the complex valued Galerkin
Table 4.26. ACA approximation of the Galerkin matrix C,h , Neumann problem
N M 1 MByte(C,h ) %
80 42 1.0 102 0.01 51.2
320 162 1.0 103 0.20 49.1
1280 642 1.0 104 2.15 34.2
5120 2562 1.0 105 17.53 17.5
20480 10242 1.0 106 126.29 7.89
matrix C,h with piecewise linear basis functions is generated. This is a rather
time consuming procedure. Thus, a quite good approximation of this matrix is
especially important when using the ACA algorithm. The accuracy obtained
for the whole numerical procedure is presented in Table 4.27. The numbers in
this table have the usual meaning. The third column shows the number of it
erations required by the GMRES method without preconditioning. Note that
4.4 Helmholtz Equation 187
the convergence of the Galerkin method for the unknown Dirichlet datum in
L2 norm,
int u u
h L2 ( )
Error1 = 0 int ,
0 uL2 ( )
is now quadratic corresponding to the estimate (2.67). In the inner point x ,
we now observe quadratic convergence (7th column), as it was predicted in
(2.67), instead of the cubic order obtained for the Dirichlet problem (cf. Table
4.23). This fact is clearly illustrated in Figs. 4.594.61, where the convergence
of the boundary element method can be seen optically. The results obtained
for N = 80 are plotted in Fig. 4.59, where the left plot shows the course of the
real part of the solution, while the imaginary part is presented on the right.
The numerical curves in Fig. 4.60 are notedly better than the previous ones.
However, their quality is not as high as of the corresponding curves obtained
while solving the Dirichlet problem (cf. Fig. 4.41).
0
4
1
2 3
3
2
4
5 1
6
0
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.59. Numerical and analytical curves for N = 80, Neumann problem
0
3
1
2.5
2
2
3
1.5
4
1
5
0.5
6
0
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.60. Numerical and analytical curves for N = 320, Neumann problem
Exhaust Manifold
The analytical solution is taken in the form (4.33) with y = (0, 0, 0.06) and
= 10, which is rather small compared with the dimension of the domain (cf.
188 4 Implementation and Numerical Examples
0 3
1 2.5
2 2
3 1.5
4 1
5
0.5
6
0
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.61. Numerical and analytical curves for N = 1280, Neumann problem
Fig. 4.11). We will consider bigger values of the wave number, when studying
the multifrequency behaviour of the problem. This small value of is well
situated to demonstrate convergence properties of the Galerkin BEM for a
regular value of the wave number. The results of the computations are shown
in Tables 4.28 and 4.29. The quality of the ACA approximation of the ma
Table 4.28. ACA approximation of the Galerkin matrices K,h and V,h
trix C,h can be seen in columns eight and nine of the Table 4.28. The third
column shows the number of iterations required by the GMRES method with
diagonal preconditioning (4.21). The fourth column displays the L2 error of
the computed Dirichlet datum, and the next column shows its quadratic con
vergence. Column six of Table 4.29 displays the absolute error (cf. (2.67)) in
a prescribed inner point x ,
0.05
0.5
0
0.05 0.45
0.1 0.4
0.15
0.35
0.2
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.62. Numerical and analytical curves for the Neumann Problem, N = 2264
0.05
0.5
0
0.05 0.45
0.1 0.4
0.15
0.35
0.2
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.63. Numerical and analytical curves for the Neumann Problem, N = 9056
the numerical solution for N = 9056 perfectly coinsides with the analytical
curves.
190 4 Implementation and Numerical Examples
92.5
0.3
90
0.25
87.5
0.2
85
0.15
82.5
0.1
80
0.05 77.5
16 18 20 22 24 16 18 20 22 24
wave numbers = 17.5 and = 20.5. If we plot the analytical and the numer
ical values of the solution of the boundary value problem for three subsequent
points 17.0, 17.5, and 18.0 for the parameter , we can see this loss of accuracy
optically. The results are presented in Figs. 4.654.67. Thus, the numerical
0.34 0.3
0.36
0.2
0.38
0.4 0.1
0.42
0
0.44
0.46 0.1
0.48
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.65. Numerical and analytical curves for = 17.0 and N = 2264, Neumann
problem
solution is quite wrong for = 17.5, while it is acceptable for = 17.0 and
= 18.0.
The picture is similar, if we consider the numerical and the analytical
solutions for = 20.0, 20.5, and = 21.0.
Note that the quality of the approximation of the boundary element ma
trices K,h , V,h , and C,h is more or less constant for all values of the wave
number on the whole interval [16, 24].
4.4 Helmholtz Equation 191
0.3
0.25
0.2
0.3
0.1
0.35
0
0.4
0.1
0.45
0.2
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.66. Numerical and analytical curves for = 17.5 and N = 2264, Neumann
problem
0.2
0.3
0.1
0.35
0
0.4
0.1
0.45
0.2
0.05 0 0.05 0.1 0.15 0.05 0 0.05 0.1 0.15
Fig. 4.67. Numerical and analytical curves for = 18.0 and N = 2264, Neumann
problem
Unit Sphere
Table 4.30. ACA approximation of the Galerkin matrices K,h and V,h
N M 1 MByte(K,h ) % MByte(V,h ) %
2
80 42 1.0 10 0.05 100.0 0.05 50.6
320 162 1.0 103 0.75 94.5 0.64 40.8
1280 642 1.0 104 6.88 54.9 5.66 22.7
5120 2562 1.0 105 53.85 26.9 42.91 10.7
20480 10242 1.0 106 379.09 11.8 302.35 4.72
for the wave number = 4, which is still moderate, are shown in Tables 4.30
and 4.31. The number of boundary elements is listed in the rst column of
these tables. The second column contains the number of nodes, while in the
third column of Table 4.30, the prescribed accuracy for the ACA algorithm
for the approximation of both matrices K,h CN M and V,h CN N is
given. The dierence in the ACA approximation of these matrices for = 4
(Table 4.30) and for = 2 (Table 4.22) can be clearly seen. In Table 4.31, the
1int u th L2 ( )
Error1 = ,
1int uL2 ( )
is given in the fourth column. The next column represents the rate of conver
gence for the Neumann datum, i.e. the quotient of the errors in two consecutive
lines of column four. We can see that linear convergence can be observed as
ymptotically. Finally, the last two columns show the absolute error (cf. (2.69))
in the point x e ,
for the value u(x ) obtained using the approximate representation formula
(2.68). Again, a rather high accuracy of the ACA approximation is necessary in
4.4 Helmholtz Equation 193
7.6908E01
0.5
0 3.2795E01
0.5
5 1

0.5
0
0
1 0.5
1
0.5
0 1
0.5
1
1.1317E01
Fig. 4.68. Given Dirichlet datum (real part) for the unit sphere, N = 320
is shown in Figs. 4.70 (real part) and 4.71 (imaginary part). The numerical
curves obtained when using an approximate representation formula in com
parison with the curve of the exact values (4.33) along the line
1.1 0.0
x(t) = 0.0 + t 0.0 , 0 t 1 (4.45)
4.0 8.0
inside of the domain e is shown in Fig. 4.72 for N = 80 and in Fig. 4.73 for
N = 320. The values of the numerical solution u and of the analytical solution
u have been computed in 512 points uniformly placed on the line (4.45). In
these gures, the thick dashed line represents the course of the analytical
solution (4.36), while the thin solid line shows the course of the numerical
solution u. The values of the variable x3 along the line (4.38) are used for
the axis of abscissas. The left plots in these gures correspond to the real
parts of the solutions, while the imaginary parts are shown on the right. The
next Fig. 4.74 shows these curves for N = 1280, but on the zoomed interval
194 4 Implementation and Numerical Examples
3.3070E01
0.5
0 1.2859E01
0.5
5 1

0.5
0
0
1 0.5
1
0.5
0 1
0.5
1
7.3514E02
Fig. 4.69. Given Dirichlet datum (imaginary part) for the unit sphere, N = 320
2.6568E01
0.5
0 1.2298E00
0.5
5 1

0.5
0
0
1 0.5
1
0.5
0 1
0.5
1
2.7254E00
Fig. 4.70. Computed Neumann datum (real part) for the unit sphere, N = 320
4.4 Helmholtz Equation 195
1.8124E01
0.5
0 3.6691E02
0.5
5 1

0.5
0
0
1 0.5
1
0.5
0 1
0.5
1
2.5462E01
Fig. 4.71. Computed Neumann datum (imaginary part) for the unit sphere, N =
320
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
0.05
0.1
4 2 0 2 4 4 2 0 2 4
Fig. 4.72. Numerical and analytical curves for N = 80, Dirichlet problem
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
0.05
0.1
4 2 0 2 4 4 2 0 2 4
Fig. 4.73. Numerical and analytical curves for N = 320, Dirichlet problem
196 4 Implementation and Numerical Examples
[1, 1] with respect to the variable x3 , in order to see the very small dierence
between them. It is almost impossible to see any optical dierence between
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
0.1 0.05
1 0.5 0 0.5 1 1 0.5 0 0.5 1
Fig. 4.74. Numerical and analytical curves for N = 1280, Dirichlet problem
the numerical and the analytical curves for higher values of N . Note that the
point x in (4.44) is chosen close to the maximum of the function Re u along
the line where the error seems to reach its maximum.
Thus, for the moderate value of the wave number = 4, the quality of
the numerical results on the unit sphere is perfect. The ACA approximation
is good, the number of GMRES iterations is low without any preconditioning
and it grows corresponding to the theory, and, nally, the theoretical linear
convergence order of the Neumann datum on the surface as well as the cubic
convergence order in the inner points of the domain e is perfectly illustrated.
We consider the exterior Neumann boundary value problem for the Helmholtz
equation in e = R3 \ with the boundary condition 1ext u(x) = g(x) for
x . The variational problem (1.122)
1
D u, v = I + K g, v for all v H 1/2 ( )
2
Unit Sphere
We consider again the analytical solution in the form (4.33) for an interior
point y = (0.9, 0, 0) and = 4 as the exact solution. The results for the
4.4 Helmholtz Equation 197
Table 4.32. ACA approximation of the Galerkin matrix C,h , Neumann problem
N M 1 MByte(C,h ) %
2
80 42 1.0 10 0.01 51.2
320 162 1.0 103 0.20 50.3
1280 642 1.0 104 2.41 38.3
5120 2562 1.0 105 19.17 19.1
20480 10242 1.0 106 135.47 8.46
ACA approximation of the matrices K,h and V,h are the same as for the
exterior Dirichlet boundary value problems and can be seen in Table 4.30. The
approximation results for the matrix C,h are shown in Table 4.32. In Table
4.33, the accuracy obtained for the whole numerical procedure is presented
and the numbers in this table have the usual meaning. The third column shows
the number of iterations required by the GMRES method without precondi
tioning. Note that the convergence of the Galerkin method for the unknown
Dirichlet datum in the L2 norm,
0int u u
h L2 ( )
Error1 = ,
0int uL2 ( )
0.4
0.2
0.3
0.1
0.2
0 0.1
0
0.1
0.1
4 2 0 2 4 4 2 0 2 4
Fig. 4.75. Numerical and analytical curves for N = 80, exterior Neumann problem
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
0.05
0.1
4 2 0 2 4 4 2 0 2 4
Fig. 4.76. Numerical and analytical curves for N = 320, exterior Neumann problem
0.25
0.2
0.2
0.15
0.1
0.1
0.05
0
0
0.05
0.1
4 2 0 2 4 4 2 0 2 4
Fig. 4.77. Numerical and analytical curves for N = 1280, exterior Neumann problem
A
Mathematical Foundations

D u(x) = u(x1 , . . . , xd ).
x
1
1
. . . x
d
d
= = (R3 \).
we then obtain
 u, v
 1 1
vLq () = sup for 1 p < , + = 1.
0=uLp () uLp () p q
A.1 Function Spaces 201
v= u Lloc
1 () ,
xi
if it satises the equality
v(x)(x)dx = u(x) (x)dx for all C0 (). (A.3)
xi
k () = max D vL () .
vW
k
Now we are able to dene the Sobolev spaces Wpk () as the closure of C ()
with respect to the Sobolev norms as introduced above,
W k ()
Wpk () = C () p .
Up to now, the above Sobolev spaces are dened only for k N0 . However,
the denition of the Sobolev norms Wpk () , and, therefore, of the Sobolev
spaces, can be generalised for arbitrary s R. For s > 0 with s = k + ,
k N0 , (0, 1), we dene the SobolevSlobodeckii norm
202 A Mathematical Foundations
1/p
vWps () = vpW k () + vpW s ()
p p
For s < 0 and 1 < p < , the Sobolev space Wps () is the dual space of
Accordingly, W ps () is the dual space of Wqs ().
Next we will collect some properties of the Sobolev spaces needed later.
Theorem A.1 (Sobolevs imbedding theorem). Let R3 be a boun
ded domain with Lipschitz boundary = , and let s 3 for p = 1, and
s > 3/p for p > 1. Then every function v Wps () is continuous, v C(),
satisfying
vL () c vWps () for all v Wps ().
In particular, we are interested in the Sobolev spaces W2s (), i.e. for p = 2.
For s = 1, the norm in W21 () is given by
1/2
vW21 () = v2L2 () + v2L2 () .
Now, we will derive equivalent norms in W21 (). A norm W21 (),f is called
equivalent to the norm W21 () , if there are some positive constants c1 and
c2 such that
A.1 Function Spaces 203
and 2
v2W 1 (), = v(x)dsx + v2L2 () .
2
Let
P0 () = {q(x) = q0 : x Rd }
be the space of constant polynomials and let
P1 () = {q(x) = q0 + q1 x1 + . . . + qd xd : x Rd }
Recall that the denition of the Sobolev spaces W2s () is based on the gen
eralised derivatives (cf. (A.3)) in the sense of Lloc
1 (). In what follows, we
will give a second denition of Sobolev spaces, which is based on derivatives
of distributions. A distribution T D () is a complex continuous linear
functional with respect to D() = C0 (). For v Lloc
1 (), the equality
Tv () = v(x)(x)dx for D()
The equality
Rd
Now we are in a position to dene the Sobolev space H s (Rd ) as a space of all
distributions v S (Rd ) with J s v L2 (Rd ), and with the inner product
u, v
H s (Rd ) = J s u, J s v
L2 (Rd ) ,
It turns out that H s (Rd ) = W2s (Rd ) for all s R. For a bounded domain
Rd , the Sobolev space H s () is dened by restriction
A.1 Function Spaces 205
( )
H s () = v = v : v H s (Rd )
s () = C ()H s (Rd ) ,
H H0s () = C0 ()
H s ()
0
s () = H0s () 1 3 5
H for s = , , ,... .
2 2 2
Moreover,
s
s () = H s () ,
H ()
H s () = H for all s R.
'
J
( )
= i , i = x R3 : x = i (), Ti R2 .
i=1
J
i (x) = 1 for x , i (x) = 0 for x \i .
i=1
J
J
v(x) = i (x)v(x) = vi (x) for x ,
i=1 i=1
with
206 A Mathematical Foundations
For the above dened functions vi (), for Ti R2 , and s 0, we now con
sider the Sobolev space H s (Ti ) associated with the norms vi H s (Ti ) . Hence,
we can dene the Sobolev spaces H s ( ) equipped with the norm
J 1/2
vHs ( ) =
vi 2H s (Ti ) .
i=1
Moreover,
2 1/2
v(x) v(y) 2
vH 1/2 ( ), = v(x)dsx + dsx dsy
x y3
w, v
wH s ( ) = sup ,
0=vH s ( ) vH s ( )
and
vH s (0 ) = inf
v H s ( ) .
H s ( ):
v v0 =v
Correspondingly,
( )
s (0 ) =
H v = v0 : v H s ( ), supp v 0 .
'
J
= i, i j = for i = j ,
i=1
For s < 0, the corresponding Sobolev spaces are given as a product space
J
s
Hpw ( ) = s (i ) ,
H
i=1
wH s ( ) wHpw
s ( ) .
At the end of this subsection, we state some relations between the Sobolev
spaces H s () in the domain and H s ( ) on the boundary = .
208 A Mathematical Foundations
0int : H s () H s1/2 ( )
satisfying 0int Ev = v.
The fundamental solution (1.7) of the Laplace equation can be found from
the relation (1.5), which can be written as a partial dierential equation in
the distributional sense,
y u (x, y) = 0 (y x) for x, y R3 ,
where 0 denotes the Dirac distribution. Since the Laplace operator is in
variant with respect to translations and rotations, we may use the transfor
mation z = y x to nd v satisfying
Dening
+
U (x, y) = w (x, y) grad div w (x, y) for x, y R3 ,
+ 2
the solution of the inhomogeneous system of linear elastostatics is equivalent
to a system of scalar BiLaplace equations
or
(z) = 0 (z), v(z) = (z) .
From this, we rst get
1 1 1
(z) = .
4 z
210 A Mathematical Foundations
+ 2
U1,1 (z) = v(z) v(z) ,
+ 2 z12
+ 2
U2,1 (z) = v(z) ,
+ 2 z1 z2
+
U3,1 (z) = v(z) ,
+ 2 z1 z3
and, hence,
1 + 3 1 1 + z12
U1,1 (z) = + ,
8 ( + 2) z 8 ( + 2) z3
1 + z1 z2
U2,1 (z) = ,
8 ( + 2) z3
1 + z1 z3
U3,1 (z) = .
8 ( + 2) z3
Doing the same computations for = 2, 3, and inserting the Lame constants,
this gives the Kelvin tensor as the fundamental solution of linear elastostatics,
1 1 1+ k (yk xk )(y x )
Uk (x, y) = (3 4) +
8 E 1 x y x y3
for k, = 1, 2, 3.
To nd the fundamental solution for the Stokes system, we have to solve the
following system of partial dierential equations
A.2 Fundamental Solutions 211
for = 1, 2, 3, and
2 w (x, y) = 0 (y x)e .
For = 1, we nd
1 1 1
w1,1 (x, y) =
4 x y
and
1 1
w1,1 (x, y) = x y, w1,2 (x, y) = w1,3 (x, y) = 0 .
8
Using
1 1 1 1 y1 x1
div w1 (x, y) = x y = ,
8 y1 8 x y
we obtain
U11 (x, y) = w1,1 (x, y) div w1 (x, y) =
y1
1 1 1 1 1 y1 x1
= =
4 x y 8 y1 x y
1 1 1 (x1 y1 )2
= + ,
8 x y x y3
U12 (x, y) = div w1 (x, y) =
y2
1 1 y1 x1 1 1 (x1 y1 )(x2 y2 )
= = ,
8 y2 x y 8 x y3
U13 (x, y) = div w1 (x, y) =
y3
1 1 y1 x1 1 1 (x1 y1 )(x3 y3 )
= = ,
8 y3 x y 8 x y3
and
212 A Mathematical Foundations
1 1 1 x1 y1
q1 (x, y) = divw1 (x, y) = = .
4 y1 x y 4 x y3
Doing the same computations for = 2, 3, we obtain the fundamental solution
of the Stokes system as
1 1 k (xk yk )(x y )
Uk (x, y) = + ,
8 x y x y3
1 x y
q (x, y) =
4 x y3
for k, = 1, 2, 3.
To nd the fundamental solution U 4 (x, y) and q4 (x, y), we have to solve
the system
we get
(x, y) = 0 (y x) for x, y R3
and therefore
1 1
(x, y) = .
4 x y
Hence,
1 xk yk
U4k (x, y) = (x, y) = , k = 1, 2, 3 .
yk 4 x y3
implying
q4 (x, y) = 0 (y x) .
Using spherical coordinates as well as v(z) = v(r) with r = z, this is equiv
alent to 1 2
1
2 r2 v(r) 2 v(r) = 0 for r > 0 .
r r r
With the substitutions
w(r) w(r) 1 1
v(r) = , v(r) = = w(r) 2 w(r) ,
r r r r r r r
we get
2
w(r) + 2 w(r) = 0 for r > 0 ,
r2
with the general solution
Choosing A = 1/4 and B = 0, this gives the fundamental solution for the
Helmholtz equation,
1 e xy
u (x, y) = for x, y R3 .
4 x y
1 (), we have N
In particular, for a given f H f H 1 () satisfying
f H 1 () c f 1
uH 1 () = N 1 ().
for all f H
H ()
214 A Mathematical Foundations
0int : H 1 () H 1/2 ( ) ,
this gives
: H
N = 0int N 1 () H 1/2 ( ) .
With this we can derive corresponding mapping properties of the single layer
potential
(V w)(x) = u (x, y)w(y)dsy for x R3 \ .
V : H 1/2 ( ) H 1 ()
is a bounded operator, and, combining this with the mapping property of the
interior trace operator
0int : H 1 () H 1/2 ( ) ,
this gives
V = 0int V : H 1/2 ( ) H 1/2 ( ).
Moreover, for x we have
(V w)(x) = u (x, y)w(y)dsy = x u (x, y) w(y)dsy = 0,
A.3 Mapping Properties 215
u (x, y) = u (y, x) .
Therefore, these considerations are valid only for selfadjoint partial dier
ential operators. For more general partial dierential operators, one has to
incorporate also all the volume and surface potentials which are dened by
the fundamental solution of the formally adjoint partial dierential operator,
see, for example, [71].
It remains to describe an explicite representation of the single layer po
tential operator
(V w)(x) = 0int (V w)(x)
for x . For w L ( ), we obtain
(V w)(x) = 0int (V w)(x) = u (x, y)w(y)dsy
as a weakly singular boundary integral (cf. Lemma 1.1). For > 0, we consider
and x with 
x x x < . Then,
lim u ( x, y)w(y)dsy u (x, y)w(y)dsy
x
x
y :yx>
lim u ( x, y) u (x, y) w(y)dsy
x
x
y :yx>
+ lim u (x, y)w(y)dsy
x
x
y :yx<
= lim u (
x, y)w(y)dsy
x
x
y :yx<
sup w(y) lim u (
x, y)dsy
y :yx< x
x
y :yx<
1 1
wL ( ) lim dsy
4 x
x 
x y
y :yx<
1 1
= wL ( ) dsy c wL ( ) ,
4 x y
y :xy<
216 A Mathematical Foundations
where we used polar coordinates to obtain the last inequality. Taking the
limit 0, this gives the denition of the single layer potential as a weakly
singular integral. In the same way we get for the exterior trace
0ext (V w)(x) = (V w)(x) for x .
Since V w H 1 () is a generalised solution of the Laplace equation for any
w H 1/2 ( ), we can compute the associated conormal derivative of V w as
1int V w H 1/2 ( ). For v H 1 (), and using Greens rst formula as well
as interpreting the surface integral as weakly singular integral, we obtain
1int (V w)(x)0int v(x)dsx = (V w)(x), v(x) dx
= u (x, y)w(y)dsy , v(x) dx
= lim u (x, y)w(y)dsy , v(x) dx
0
y :yx
= w(y) lim x u (x, y), v(x) dx dsy .
0
x:xy
x:xy x :xy
+ int u (x, y) int v(x)ds ,
1,x 0 x
x:xy=
and, therefore,
1int (V w)(x)0int v(x)dsx = w(y) lim 1,x u int
(x, y)0int v(x)dsx dsy
0
x :xy
int
+ w(y) lim 1,x u (x, y)0int v(x)dsx dsy
0
x:xy=
int
= lim 1,x u (x, y)w(y)dsy 0int v(x)dsx
0
y :yx
int
+ w(y) lim 1,x u (x, y)v(x)dsx dsy
0
x:xy=
= (K w)(x)0int v(x)dsx + w(y)(y)v(y)dsy .
A.3 Mapping Properties 217
with
int
(y) = lim 1,x u (x, y)dsx
0
x:xy=
1 (y x, n(x)) 1 1
= lim dsx = lim dsx ,
0 4 x y3 0 4 2
x:xy= x:xy=
0int : H 1 () H 1/2 ( )
gives
0int W : H 1/2 ( ) H 1/2 ( ).
It turns out (cf. Lemma 1.2) that
y :yx
+ int u (
1,y x, y) v(y) v(x) dsy
y :yx
+v(x) int u (
1,y x, y)dsy .
y :yx
In addition,
lim int u (
1,y x, y) v(y) v(x) dsy = 0.
0
y :yx
and, therefore, the jump relation of the double layer potential reads
which does not exist. In what follows we will sketch the computations to obtain
the alternative representation (1.9) (cf. Lemma 1.4). For the derivatives of the
double layer potential for x we rst have
(W v)(
x) = 1int u (
x, y)v(y)dsy
xi xi
1 1
= n(y), y v(y)dsy
4 xi 
x y
1 1
= n(y), y v(y)dsy
4 yi x y
1 1
= n(y), curly ei y v(y)dsy
4 
x y
1 1
= v(y) curl,y ei y dsy .
4 
x y
220 A Mathematical Foundations
we obtain
1 1
x) =
(W v)( curl v(y), ei y dsy
xi 4 
x y
1 1
= ei , curl v(y) y dsy ,
4 
x y
and, therefore,
1 1
x (W v)(
x) = curl v(y) y dsy
4 
x y
1 1
= curl v(y) x dsy .
4 
x y
Hence,
1 1
x), n(x)) =
(x (W v)( curl v(y) x , n(x) dsy
4 
x y
1 1
= curl v(y), n(x) x dsy .
4 
x y
x, we obtain
Taking the limes x
1 1
(Dv)(x) = curl v(y), n(x) x dsy
4 x y
1 1
= curl v(y), curl,x dsy ,
4 x y
and, therefore,
1 1
Dv, u
= u(x) curl v(y), curl,x dsy dsx
4 x y
1
1
= curl,x u(x)curl,y v(y) dsx dsy .
4 x y
With
A.3 Mapping Properties 221
curl,x u(x)curl,y v(y) = n(x), x u(x)curl,y v(y)
= n(x), x u(x) curl,y v(y)
= n(x) x u(x), curl,y v(y)
= curl,x u(x), curl,y v(y) ,
and
c2 0ext u2H 1/2 ( ) a e (u, u).
Hence, we obtain
V w, w
= w(x)(V w)(x)dsx
1
= w(x) + (K w)(x) (V w)(x)dsx
2
1
w(x) + (K w)(x) (V w)(x)dsx
2
= a (u, u) + a e (u, u)
min{c1 , c2 } 1int u2H 1/2 ( ) + 1ext u2H 1/2 ( ) .
Using
1 1 2
w2H 1/2 ( ) = w + K w w + K w 1/2
2 2 H ( )
V w, w
cV1 w2H 1/2 ( ) for all w H 1/2 ( ).
Note that the ellipticity estimate of the hypersingular boundary integral oper
ator D (cf. Lemma 1.4) follows almost in the same manner when considering
the double layer potential u = W v.
If = is the boundary of a Lipschitz domain R3 , we then can
extend the mapping properties of all boundary integral operators, see [24]. In
particular, the boundary integral operators
A.3 Mapping Properties 223
V : H 1/2+s ( ) H 1/2+s ( ),
K : H 1/2+s ( ) H 1/2+s ( ),
K : H 1/2+s ( ) H 1/2+s ( ),
D : H 1/2+s ( ) H 1/2+s ( )
,
: X X R .
Then,
f, v
AvX cA
2 vX for all v X, (B.1)
Au, v
= Av, u
for all u, v X.
Au = f (B.2)
is satised in X . Then,
Au f, v
0 = Au f X = sup ,
0=vX vX
Au, v
= f, v
for all v X. (B.3)
226 B Numerical Analysis
Av, v
0 for all v X ,
then the solution of the variational problem (B.3) is equivalent to the min
imisation problem
F (u) = min F (v) . (B.4)
vX
f, v
= u, v
X for all v X. (B.5)
uX = f X .
The Riesz representation theorem (see (B.5)) denes a linear and bounded
operator J : X X with
Jf, v
X = f, v
for all v X, Jf X = f X .
JAu = Jf in X. (B.6)
JAvX = AvX cA
2 vX for all v X,
JAXX cA
2 .
Av, v
cA
1 vX
2
for all v X . (B.7)
JAv, v
X = Av, v
cA
1 vX
2
for all v X
with
T = I JA : X X
for some positive R+ . For
cA
0 < < 2 1 2
cA
2
1 uX Au, u
= f, u
f X uX ,
cA 2
and, therefore,
1
uX f X .
cA
1
with
1
F (vM ) = AvM , vM
f, vM
2
1
M M M
= v vj A , j
v f,
= F(v) .
2 j=1
=1 =1
d
F (v) = 0 for k = 1, . . . , M ,
dvk
we then obtain
M
v A , k
f, k
= 0 for k = 1, . . . , M .
=1
M
u A , k
= f, k
for k = 1, . . . , M, (B.10)
=1
AuM , vM
= f, vM
for all vM XM . (B.11)
M
M
M
M
(AM u, v) = A[k, ]u vk = A , k
u vk = AuM , vM
(AM v, v) = AvM , vM
cA
1 vM X
2
1 uM X AuM , uM
= f, uM
f X uM X ,
cA 2
Au, vM
= f, vM
for all vM XM X.
B.1 Variational Methods 229
Subtracting from this the Galerkin variational problem (B.11), this gives the
Galerkin orthogonality
A(u uM ), vM
= 0 for all vM XM . (B.12)
Using the ellipticity estimate (B.7), the Galerkin orthogonality (B.12), and
the boundedness of A, we then obtain
1 u uM X A(u uM ), u uM
cA 2
= A(u uM ), u vM
+ A(u uM ), vM uM
= A(u uM ), u vM
A(u uM )X u vM X
cA
2 u uM X u vM X ,
and, therefore,
cA
u uM X 2
u vM X for all vM XM .
cA
1
cA
u uM X 2
inf u vM X . (B.13)
cA
1 vM XM
BgX cB
2 gY for all g Y.
fk = Bg, k
for all k = 1, . . . , M ,
N
N
fk = BgN , k
= gj Bj , k
= BN [k, j]gj ,
j=1 j=1
= BN g .
AM u
A
uM , vM
= BgN , vM
for all vM XM .
A(u
uM ), vM
= A(uM
uM ), vM
= B(ggN ), vM
for all vM XM .
1 uM u
cA M 2X A(uM u
M ), uM u
M
= B(g gN ), uM u
M
cB
2 g gN Y uM u
M X ,
and, therefore,
cB
uM u
M X 2
g gN Y .
cA
1
Hence, we nd from the triangle inequality the error estimate
u u
M X u uM X + uM uM X
B
c
u uM X + 2A g gN Y .
c1
Besides an approximation of the given right hand side f , we also have to con
sider an approximation of the given operator A, e.g., when applying numerical
integration schemes. Hence, instead of the Galerkin formulation (B.11), we
have to solve a perturbed variational problem to nd the solution u M XM
satisfying
uM , vM
= f, vM
for all vM XM .
A (B.14)
: X X is a bounded linear operator satisfying
We assume that A
X cA
Av 2 vX for all v X.
uM , vM
= 0 for all vM XM .
AuM A
we nd
From this, and by using the XM ellipticity of A,
M u
1 uM u
cA M 2X A(u M ), uM u
M
A)uM , uM u
= (A M
and, therefore,
1 M X .
uM u
M X
(A A)u
cA
1
Hence, applying the triangle inequality, we nally obtain the error estimate
u u
M X u uM X + uM u M X
1 M X
u uM X + (A A)u
cA
1
1 X + (A A)(u
u uM X + (A A)u uM )X
cA
1
cA + c A
1 X .
1 + 2 2 u uM X + (A A)u
cA
1 c A
1
'
J
= i , i j = for i = j,
i=1
'
N
= ,
=1
232 B Numerical Analysis
we further assume that for each boundary element there exists exactly one
boundary segment i with i . Hence, there also exists an element qi such
that = i (qi ). For the area of the boundary element , we then obtain
*
= dsx = EG F 2 d ,
qi
with
3 2 3 2 3
E= xi () , G = xi () , F = xi () xi ().
i=1
1 i=1
2 i=1
1 2
Hence, we obtain
c1 area qi c2 area qi .
Then,
*
v2L2 ( ) = v(x)2 dsx = v(i ())2 EG F 2 d
qi
c
c2 v(i ()) d 2
2
v(i ())2 d .
c1 area qi
qi qi
and with
1
area qi = d = det i  d = det i  ,
2
qi
we nally obtain
v2L2 ( ) 2c
v 2L2 ( ) , v () = v(i (i ())) .
Note that this result would follow directly when considering a parametrisation
of the boundary element with respect to the reference element . However,
the above approach is needed when considering higher order Sobolev spaces,
for example
B.2 Approximation Properties 233
v2H m ( ) = D v(i ())2 d, m N.
=m
qi
For the parametrisation qi = i ( ) and for  = m, we now obtain the norm
equivalence inequalities,
1
(area qi )1m D v ()2 d D v(i ())2 d ,
cm
qi
and
D v(i ())2 d cm (area qi )1m D v ()2 d .
qi
Hence, we have
c1 1m

v 2H m ( ) v2H m ( ) c2 1m

v 2H m ( ) for m N .
N
w (x)k (x)dsx = w(x)k (x)dsx for k = 1, . . . , N.
=1
as well as
1
w = w(y)dsy for = 1, . . . , N.
w Qh w2L2 ( ) 2c w
Q w
2L2 ( ) ,
() = Qh w(i ()).
where Q w
For an arbitrary but xed L2 ( ), we dene the linear functional
) =
f (w () Q w
w () ()d
satisfying
) w
f (w Q w
L2 ( ) L2 ( )
2 w
L2 ( ) L2 ( ) 2 w
H 1 ( ) L2 ( ) ,
234 B Numerical Analysis
) c w
f (w H 1 ( ) L2 ( ) .
and, therefore,
Q w
w L2 ( ) c w
H 1 ( ) .
Altogether, we nally obtain
w Qh w2L2 ( ) = 2 w
Q w
2L2 ( )
c w
2H 1 ( ) c h2 w2H 1 ( ) ,
as well as
N
w Qh w2L2 ( ) c h2 w2H 1 ( )
=1
and
w Qh wL2 ( ) c h wHpw
1 ( ) ,
where d is the element diameter (cf. Chapter 2). Integrating over , and
since the boundary element is shape regular, i.e. d cB h , this gives with
= h2
(w(x) w(y))2
(w(x) Qh w(x))2 dsx c2+2s h 2s
dsy dsx .
B
x y2+2s
w Qh wH ( ) = sup
0=vH ( ) vH ( )
w Qh w, v Qh v
= sup
0=vH ( ) vH ( )
w Qh wL2 ( ) v Qh vL2 ( )
= sup
0=vH ( ) vH ( )
c h w Qh wL2 ( ) c hs wHpw
s ( ) ,
N
3
w,i ,i (x)k,j (x)dsx = w(x)k,j (x)dsx
=1 i=1
w Qh w2L2 ( ) 2c w
Q w
L2 ( )
for the local error. The BrambleHilbert lemma now implies
w L2 ( ) c w
Q w H 2 ( ) .
Hence, we have
w Qh w2L2 ( ) c w
2H 2 ( )
c 2 w2H 2 ( ) =
c h4 w2H 2 ( ) ,
and, therefore,
N
w Qh w2L2 ( )
c h4 w2H 2 ( )
=1
as well as
w Qh wL2 ( ) c h2 wHpw
2 ( ) ,
Qh : L2 ( ) L2 ( )
w Qh wH ( ) c hs wHpw
s ( )
for [2, 0] and s [0, 2], which is the approximation property (2.8).
Finally, we consider the piecewise linear, globally continuous L2 projection
Qh w Sh1 ( ) as the unique solution of the variational problem
Qh w(x)j (x)dsx = w(x)j (x)dsx for j = 1, . . . , M.
M
Ph u = uj j Sh1 ( ),
j=1
Ph u = arg min u vh H 1 ( ) .
vh Sh
1 ( )
B.2 Approximation Properties 237
Ph u, vh
H 1 ( ) = u, vh
H 1 ( ) for all vh Sh1 ( ).
u Ph uH 1 ( ) c h uHpw
2 ( ) ,
Ph : H 1 ( ) H 1 ( )
u Ph uH 1 ( ) uH 1 ( ) .
u Ph u, v
H 1 ( )
u Ph uH ( ) = sup
0=vH 2 ( ) vH 2 ( )
u Ph u, v Ph v
H 1 ( )
= sup
0=vH 2 ( ) vH 2 ( )
u Ph uH 1 ( ) v Ph vH 1 ( )
sup
0=vH 2 ( ) vH 2 ( )
c h1 u Ph uH 1 ( ) c hs uH s ( ) ,
'
N
N = .
=1
We assume that all boundary elements are plane triangles. Using the ref
erence element
( )
= R2 : 0 < 1 < 1 , 0 < 2 < 1 1 ,
for . Using
= dsx ,
with respect to the reference element , which are exact for polynomials of
a certain order. From this, we may nd 3M parameters (i,1 , i,2 , i ) for i =
1, . . . , M . Let ( )
Pk = span 11 22 1 +2 k
be the space of polynomials of an order less or equal k. The dimensions of
Pk and the minimal number M of integration points are given in Table C.1.
The last line of this table shows the number P = 3M of free parameters.
These parameters (i,1 , i,2 , i ) for i = 1, . . . , M are solutions of the nonlinear
k 0 1 2 3 4 5
dimPk 1 3 6 10 15 21
M 1 3 4 7
P 3 9 12 21
equations
1
M
11 22 d = 1 2
i i,1 i,2 for 1 , 2 : 1 + 2 k . (C.1)
2 i=1
Note that with this choice, the second and third equations are satised. For
the remaining three equations, we obtain
4s2 8s + 3 = 0 ,
1 = , 2 = 3 = 4 .
However, the rst weight 1 is negative, and, therefore, the above quadrature
can not be used for practical computations.
For k = 5 and M = 7, the nonlinear equations (C.1) read
1
7
11 22 d = 1 2
i i,1 i,2 for all 1 , 2 : 1 + 2 5.
2 i=1
and
1 1 t 1 1 1 t 1 1 0 t 2
5 = + , 6 = + , 7 = + ,
2 0 6 2 2 1 6 1 2 1 6 1
as well as the associated integration weights as
1 = , 2 = 3 = 4 , 5 = 6 = 7 .
1 = 1 32 35 .
2402 (s 1)2 (3s2 10s + 13) + 155 (t 1)2 (3t2 + 2t + 19) = 176 ,
33602 (s 1)2 (8 11s + 6s2 s3 ) + 1055 (13 t + 3t2 + t3 ) = 1184 .
1 35 (t 1)2 10s 6
2 = , 5 = .
12(s 1)2 15(t 1)2 (2s + t 3)
244 C Numerical Algorithms
and, therefore,
t1
s = 3 .
5t 9
The fourth equation is nally equivalent to
7t2 18t + 3 = 0 .
Thus,
9 2 15
t1/2 = .
7
Hence, we have
9 2 15 6 15
t = , s = .
7 7
For the integration weights, we nd
155 + 15 155 15 9
5 = , 2 = , 1 = ,
1200 1200 40
while for the integration nodes, we nally obtain
1 6 15 1 9 + 2 15 1 6 15
2 = , 3 = , 4 =
21 6 15 21 6 15 21 9 + 2 15
and
1 6 + 15 1 6 + 15 1 9 2 15
5 = , 6 = , 7 = .
21 6 + 15 21 9 2 15 21 6 + 15
'
N
N = .
=1
We assume that all boundary elements are plane triangles. In order not to
overload the subsequent formulae, we consider a plane triangle R3 given
via its three corner points x1 , x2 , and x3 (cf. Fig. C.2), having in mind that
this triangle is one of the boundary elements , = 1 . . . , N .
We rst dene a suitable local coordinate system connected to the bound
ary element with the origin in x1 and having the basis vectors
C.2 Analytic Integration 245
x3
r2
n
r1
x1
x
x2
r1 , r2 , n .
t = (x1 x2 , r2 ) .
Then,
1
r1 = (x x1 ) , s = x x1  .
s
Finally, the corresponding normal vector is dened by
n = r1 r2 .
Two further characteristic quantities, we will need for the analytical integra
tion, are the tangents of the angles between the height x1 x and the sides
x1 x2 and x1 x3 , correspondingly. These quantities can be computed as
t t t
1 = , 2 = .
s s
Thus, an appropriate parametrisation of the boundary element would be
x = x1 + sx r1 + tx r2 + ux n
with
sx = (x x1 , r1 ), tx = (x x1 , r2 ), ux = (x x1 , n) .
x y2 = (s sx )2 + (t tx )2 + u2x .
with parameters
tx + sx (tx sx )2
p = , q 2 = u2x + .
1 + 2 1 + 2
Integration by parts gives
*
F (s, ) = (s sx ) log s tx + (1 + 2 )(s p)2 + q 2 s
*
(tx sx ) (1 + 2 )(s p)2 + q 2 + (1 + 2 )(s p)(p s ) q 2
* ds .
(1 + )2 (s p)2 + q 2 + (s tx ) (1 + 2 )(s p)2 + q 2
C.2 Analytic Integration 247
1 + 2 (tx sx ) cosh u + (tx sx ) sinh u 1 + 2 q
=q du.
q(1 + 2 ) cosh u + 1 + 2 q sinh u + (sx tx )
and, therefore,
1 B2 v 2 + B1 v + B0
I = 2q dv ,
1 v 2 A2 v 2 + A1 v + A0
with parameters
*
B2 = 1 + 2 (tx sx + q),
B1 = 2(tx sx ),
*
B0 = 1 + 2 (tx sx q),
A2 = (1 + 2 )q (sx tx ),
*
A1 = 2 1 + 2 q,
A0 = (1 + 2 )q + (sx tx ).
Due to
A21 4A0 A2 = 4(1 + 2 )u2x 0 ,
we decompose
1 B2 v 2 + B1 v + B0 C1 C2 C3 v + C4
= + +
1 v A2 v + A1 v + A0
2 2 1 v 1 + v A2 v + A1 v + A0
2
to obtain
248 C Numerical Algorithms
1 B0 + B1 + B2
C1 = ,
2 A0 + A1 + A2
1 B0 B1 + B2
C2 = ,
2 A0 A1 + A2
C3 = A2 (C1 C2 ),
C4 = B0 (C1 + C2 )A0 ,
and, therefore,
tx sx 1 + 2 u2x
C1 = C 2 = , C3 = 0, C4 = .
2q 1 + 2 q
Hence,
tx sx 1 1 * 1
I = + 2 1 + ux
2 2
dv .
1 + 2 1 v 1 + v A2 v 2 + A1 v + A0
Integrating the rst part and resubstituting, this gives
1 1 1 + v
dv = log v + 1 log v 1 = log
1+v v1 1 v
1 + 2 (s p)
= 2 arctanh v = u = arcsinh
q
5
1 + 2 (s p) (1 + 2 )(s p)2
= log + +1
q q2
* *
= log 1 + 2 (s p) + (1 + 2 )(s p)2 + q 2 log q .
Collecting all terms together and ignoring constant parts, we nally obtain
*
F (s, ) = (s sx ) log s tx + (s sx )2 + (s tx )2 + u2x s
sx tx * *
+ log 1 + 2 (s p) + (1 + 2 )(s p)2 + q 2
1 + 2
*
x tx
q s
1+2 (1 + 2 )(s p)2 + q 2 + s tx q q
+2ux arctan .
(s p)ux
For piecewise linear basis functions we have to compute the local contributions
of the integrals
1 (x y, n(y))
Di (, x) = ,i (y)dsy
4 x y3
s 2 s
1 s s ux
= dtds
4 s (t t )2 + (s s )2 + u2 3/2
0 1 s x x x
s 6 7 2 s
ux s s 1 t tx
= * ds
4 s (s sx )2 + u2x (t tx )2 + (s sx )2 + u2x
0 1 s
1
= F (s , 2 ) F (0, 2 ) F (s , 1 ) + F (0, 1 ) .
4s
Hence, it is sucient to compute the integral
(s s)(s tx )
F (s, ) = ux * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
(s sx ) (sx tx ) (s sx ) + (s sx )(sx tx ) + u2x
= ux * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
1
ux * ds
(1 + )(s p)2 + q 2
2
(s sx ) (sx tx ) (s sx ) + (s sx )(sx tx ) + u2x
= ux * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
* *
ux log 1 + 2 (s p) + (1 + 2 )(s p)2 + q 2 ,
1 + 2
with parameters p and q already used for the single layer potential:
tx + sx (tx sx )2
p= , q 2 = u2x + .
1 + 2 1 + 2
250 C Numerical Algorithms
and, thereforfe,
C 2 v 2 + C 1 v C2
I = 2ux dv ,
a1 v 4 a2 v 3 + a3 v 2 + a2 v + a1
with parameters
1
C2 = (s sx )(tx sx ) (1 + 2 )q 2 ,
1+ 2
C1 = 2q s + tx 2sx ,
a1 = (1 + 2 ) u2x + (p sx )2 = u2x + 2 q 2 > 0 ,
* 4q(tx sx )
a2 = 4q 1 + 2 (p ss ) = ,
1 + 2
a3 = 4q 2 2a1 ,
or
2ux C2 v 2 + C1 v C2
I = dv ,
u2x + 2 q 2 v4 av 3 + bv 2 + av + 1
with
4q(tx sx ) 4q 2
a = , b = 2.
1 + 2 u2x + 2 q 2 u2x + 2 q 2
For the decomposition
C.2 Analytic Integration 251
C 2 v 2 + C 1 v C2 D1 v + E 1 D2 v + E2
= 2 + 2 ,
v4 av 3 + bv 2 + av + 1 v + A1 v + B1 v + A2 v + B2
we nd the coecients
2 1 + 2 q tx sx
A1/2 = 2 q ,
ux + 2 q 2 1 + 2
2
1 + 2 tx sx
B1/2 = 2 q ,
ux + 2 q 2 1 + 2
as well as
1 2
D1 = ux + 2 q 2 ,
2
1 2
D2 = ux + 2 q 2 ,
2
1
E1 = s sx + q tx sx + (1 + 2 )q ,
2 1+ 2
1
E2 = s sx q tx sx (1 + 2 )q .
2 1 + 2
Hence, we have
2ux D1 v + E1 D2 v + E2
I = 2 dv + dv ,
ux + 2 q 2 v 2 + A1 v + B1 v 2 + A2 v + B2
and it remains to integrate
2ux D1/2 v + E1/2
I1/2 = 2 dv
ux + 2 q 2 v 2 + A1/2 v + B1/2
2ux 1
= 2 2 2
D1/2 log v 2 + A1/2 v + B1/2
ux + q 2
1 1
+ E1/2 A1/2 B1/2 dv
2 v 2 + A1/2 v + B1/2
1
= ux log v 2 + A1/2 v + B1/2
2
2ux 1 1
+ 2 2 2
E1/2 A1/2 B1/2 2
dv .
ux + q 2 v + A1/2 v + B1/2
Due to
2
1 (1 + 2 )u2x tx sx
G21/2 = B1/2 A21/2 = 2 q ,
4 u2x + 2 q 2 1 + 2
we nd
1 + 2 ux  tx sx
G1/2 = q > 0.
u2x + 2 q 2 1 + 2
252 C Numerical Algorithms
Hence, we obtain
1 1
E1/2 A1/2 D1/2 dv
2 v 2 + A1/2 v + B1/2
1 1
= E1/2 A1/2 D1/2 1 dv
2 (v + 2 A1/2 )2 + G21/2
1 1 2v + A1/2
= E1/2 A1/2 D1/2 arctan
G1/2 2 2G1/2
u2x + 2 q 2 2v + A1/2
= (s sx ) arctan .
2ux  2G1/2
Therefore,
1 ux 2v + A1/2
I1/2 = ux log v 2 + A1/2 v + B1/2 (s sx ) arctan .
2 ux  2G1/2
Recall that
* *
1 + sinh2 u 1 (1 + 2 )(s p)2 + q 2 q
v = = .
sinh u 1 + 2 (s p)
y = x1 + s r1 + t r2 , x = x1 + sx r1 + tx r2 + ux n ,
bij (s sx ) cij ux (t tx ) dij (s sx )2 eij (s sx )ux + fij u2x
3/2 + 3/2
(s sx )2 + (t tx )2 + u2x (s sx )2 + (t tx )2 + u2x
with
s 2 s
1 1 1
Sij (, x) = aij
4 1/2 dt ds,
0 1 s
(s sx )2 + (t tx )2 + u2x
s 2 s
2 1 bij (s sx ) cij ux (t tx )
Sij (, x) = dt ds,
4 (s s )2 + (t t )2 + u2 3/2
0 s x x x
1
s 2 s
3 1 dij (s sx )2 eij (s sx )ux + fij u2x
Sij (, x) =
4 3/2 dt ds.
0 1 s
(s sx )2 + (t tx )2 + u2x
1
Note that Sij (, x) corresponds to the single layer potential function for the
Laplace equation.
For the second entry we obtain
s 2 s
2 1 bij (s sx ) cij ux (t tx )
Sij (, x) =
4 3/2
0 s
(s sx )2 + (t tx )2 + u2x
1
s 6 72 s
1 bij (s sx ) cij ux
= 1/2 ds
4 (s sx )2 + (t tx )2 + u2x
0 1 s
1 2
= Fij (s , 2 ) Fij2 (s , 1 ) Fij2 (0, 2 ) + Fij2 (0, 1 ) ,
4
with the integral
bij (s sx ) cij ux
Fij2 (s, ) = 1/2 ds
(s sx )2 + (s tx )2 + u2x
b (s sx ) cij ux
= * ij ds
(1 + 2 )(s p)2 + q 2
254 C Numerical Algorithms
sp bij (p sx ) cij ux
= bij * ds + * ds
(1 + )(s p)2 + q 2
2 (1 + 2 )(s p)2 + q 2
bij 1/2
= (1 + 2 )(s p)2 + q 2
1 + 2
bij (p sx ) cij ux * *
+ log 1 + 2 (s p) + (1 + 2 )(s p)2 + q 2 .
1 + 2
For the third integral, we get
s 2 s
3 1 dij (s sx )2 eij (s sx )ux + fij u2x
Sij (, x) =
4 3/2 dt ds
0 1 s
(s sx )2 + (t tx )2 + u2x
s 6 7 2 s
t tx dij (s sx )2 eij (s sx )ux + fij u2x
= * ds
(s sx )2 + u2x (s sx )2 + (t tx )2 + u2x
0 1 s
1 3
= Fij (s , 2 ) Fij3 (s , 1 ) Fij3 (0, 2 ) + Fij3 (0, 1 )
4
with
s tx dij (s sx )2 eij (s sx )ux + fij u2x
Fij3 (s, ) = * ds
(s sx ) + ux
2 2
(s sx )2 + (s tx )2 + u2x
s tx dij (s sx )2 + u2x eij (s sx )ux + fij dij u2x
= * ds
(s sx )2 + u2x (s sx )2 + (s tx )2 + u2x
s tx
= dij * ds
(1 + 2 )(s p)2 + q 2
(s sx )ux s tx
eij * ds
(s sx ) + ux (1 + 2 )(s p)2 + q 2
2 2
u2x s tx
+ fij dij * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
= dij F 3,1 (s, ) eij F 3,2 (s, ) + fij dij F 3,3 (s, )
and
s t
F 3,1
(s, ) = * ds
(1 + 2 )(s p)2 + q 2
sp p t
= * ds + * ds
(1 + 2 )(s p)2 + q 2 (1 + 2 )(s p)2 + q 2
C.2 Analytic Integration 255
*
= (1 + 2 )(s p)2 + q 2
1 + 2
p t * *
+ log 1 + 2 (s p) + (1 + 2 )(s p)2 + q 2 .
1 + 2
The remaining integrals can be computed in the same way as for the double
layer potential of the Laplace operator. In particular, up to the sign and the
choice s = s , the integral
(s sx )ux s tx
3,2
F (s, ) = * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
coincides with the integral of the double layer potential function of the Laplace
operator.
Furthermore, using the transformations as for the Laplace operator, we
have
u2x s tx
F 3,3 (s, ) = * ds
(s sx )2 + u2x (1 + 2 )(s p)2 + q 2
q sinh u + 1 + 2 (p tx )
= u2x du
q 2 sinh2 u + 2 1 + 2 q(p sx ) sinh u + (1 + 2 )[u2x + (p sx )2 ]
q2v+ 1+2 (ptx )(1v 2 )
= 2u2x dv
4q 2 v 2+4 1+2 q(psx )v(1v 2 )+(1+2 ) u2x +(p sx )2 (1v 2 )2
C2 v 2 + C1 v C 2
= 2u2x dv ,
a 1 v a 2 v 3 + a 3 v 2 + a 2 v + a1
4
and
4q(tx sx )
a1 = u2x + 2 q 2 , a2 = , a3 = 4q 2 2a1 .
1 + 2
Hence, we get
2u2 C2 v 2 + C1 v C2
F 3,3
(s, ) = 2 x 2 2 dv
ux + q (v 2 + A1 v + B1 )(v 2 + A2 v + B2 )
with
2
2 1 + 2 q tx sx 1 + 2 tx sx
A1/2 = 2 q , B 1/2 = q .
ux + 2 q 2 1 + 2 u2x + 2 q 2 1 + 2
C2 v 2 + C1 v C2 D1 v + E1 D2 v + E2
= 2 + 2 ,
(v 2 + A1 v + B1 )(v 2 + A2 v + B2 ) v + A1 v + B1 v + A2 v + B2
we nd the coecients
(tx sx ) + (1 + 2 )q (tx sx ) (1 + 2 )q
D1 = D2 = 0, E1 = , E2 = .
2 1+ 2 2 1 + 2
Therefore,
3,3 2 2 E1 E2
F (s, ) = 2 2 2 dv + dv ,
+ q v 2 + A1 v + B1 v 2 + A2 v + B2
with
1 + 2   tx sx
G1/2 = q > 0.
u2x + 2 q 2 1 + 2
Finally,
2v + A1/2
I1/2 = ux  arctan .
2G1/2
Ax = f , A RN N , x, f RN . (C.2)
Note that the dimension N of the system is usually connected to the discreti
sation parameter h (cf. (2.2)), and, therefore, we have a sequence of linear
systems to be solved for N . We rst consider symmetric and positive
denite systems and later general systems with regular matrices.
(Ap , pk ) = 0 for k = .
Note that (Apk , pk ) > 0 since A is assumed to be positive denite. The vector
system P forms an Aorthogonal basis of the vector space RN . Hence, the
solution vector x RN of the linear system (C.2) can be written as
N 1
x = x0 p
=0
N 1
Ax = Ax0 Ap = f .
=0
N 1
(Ap , pk ) = (Ax0 f , pk ) , for k = 0, . . . , N 1 ,
=0
(Ax0 f , pk )
k = for k = 0, . . . , N 1 . (C.4)
(Apk , pk )
k
xk+1 = x0 p = xk k pk for k = 0, . . . , N 1 .
=0
(rk , pk )
k = for k = 0, . . . , N 1 , (C.5)
(Apk , pk )
where
rk = Axk f
is the residual vector induced by the approximate solution xk . Note that the
following recursion holds:
(rk , pk )
(rk+1 , pk ) = (rk k Apk , pk ) = (rk , pk ) (Apk , pk ) = 0
(Apk , pk )
for k = 0, . . . , N 2. From
k
(Awk+1 , p )
pk+1 = wk+1 k p , k = , for = 0, . . . , k . (C.7)
(Ap , p )
=0
l1
(rk+1 , w ) = (rk+1 , p ) + j (rk+1 , pj ) = 0 (C.8)
j=0
for = 0, . . . , k. In particular, if the residual vector rk+1 does not vanish, the
vectors
{w0 , w1 , . . . , wk , rk+1 }
are linear independent and we can choose wk+1 = rk+1 . If rk+1 = 0, then
the system (C.2) is solved. In general, this will not happen and we can choose
wk = rk for k = 0, . . . , N 1 .
C.3 Iterative Solution Methods 259
k1
(r , p ) = (r , r )
k k k k
k1, (rk , p ) = (rk , rk ) = k .
=0
by using the recursion (C.7) and the orthogonality relation (C.6). From > 0,
it follows > 0 for = 0, . . . , k, and, therefore, we can write
1
Ap = r r+1 .
For the enumerator of the coecients kj in (C.7), we then obtain
1 k+1
(Awk+1 , p ) = (rk+1 , Ap ) = (r , r r+1 ) = 0
for = 0, . . . , k 1 and
1 k+1 k+1 k+1
(Awk+1 , pk ) = (r ,r ) =
k k
for = k. Hence, we have k = 0 for = 0, . . . , k 1 and
k+1
kk = k = .
k (Apk , pk )
and, therefore,
k+1
pk+1 = rk+1 + k pk , k = .
k
Hence, we end up with the conjugate gradient method as summarised in Al
gorithm C.1.
Algorithm C.1
1. Compute for an arbitrary given initial solution x0 RN
r0 = Ax0 f , p0 = r0 , 0 = (r0 , r0 ).
2. Iterate for k = 0, . . . , N 2
k
sk = Apk , k = (sk , pk ), k = , xk+1 = xk k pk
k
260 C Numerical Algorithms
Stop, if
k+1 2 0
is satisfied with some prescribed accuracy .
Otherwise compute
k+1
k = , pk+1 = rk+1 + k pk .
k
N 1
xk+1 x = Ap ,
=k+1
1
k+1 2 N
x xA = A(xk+1 x), xk+1 x = 2 (Ap , p ) .
=k+1
k
uk+1 = x0 p , (C.9)
=0
1
k+1 2 k
2 N
u xA = (Ap , p ) + 2 (Ap , p )
=0 =k+1
where the minimum is taken over all vectors uk+1 of the form (C.9). From the
recursions
we nd representations
p = (A)r0
with a matrix polynomials (A) of degree . Hence, we conclude with e0 =
x0 x
C.3 Iterative Solution Methods 261
k
k
k
uk+1 x = x0 x p = e0 (A) r0 = e0 (A) Ae0 ,
=0 =0 =0
and, therefore,
uk+1 x = pk+1 (A)e0
with some matrix polynomial pk+1 (A) of degree k + 1 and having the prop
erty pk+1 (0) = 1. The polynomial pk+1 (A) obtained for the vector xk+1 is,
therefore, the solution of the minimisation problem
k+1
x xA = min pk+1 (A)e0 A , (C.10)
pk+1 (A)
where the minimum is taken over all polynomials pk+1 (A) having the property
pk+1 (0) = 1. The space
( ) ( )
Sk (A, r0 ) = span p0 , p1 , . . . , pk = span r0 , Ar, . . . , Ak r0
where the minimum is again taken over all polynomials pk+1 (A) having the
property pk+1 (0) = 1, and the maximum over the spectrum of the matrix A,
i.e. [min (A), max (A)] . The above min maxproblem will be solved by
the scaled Tschebysche polynomials Tk+1 (), and we nd
2q k+1
min max pk+1 () = max Tk+1 () = ,
pk+1 (A) 1 + q 2(k+1)
with * * *
max (A) + min (A) 2 (A) + 1
q = * * = * ,
max (A) min (A) 2 (A) 1
where
max (A)
2 (A) =
min (A)
is the spectral condition number of the symmetric and positive denite matrix
A. Using the Raleigh quotient
(Ax, x) (Ax, x)
min (A) = min max = max (A) ,
x=0 (x, x) x = 0 (x, x)
cA
2 (A) 2
.
cA
1
Since the spectral condition number of the boundary element stiness matrices
may depend on mesh parameters such as the mesh size h or the mesh ratio
hmax /hmin , an appropriate preconditioning is mandatory in many cases.
Hence, we assume that there is a symmetric and positive denite matrix
CA RN N which can be factorised as CA = CA CA , where CA is again
1/2 1/2 1/2
symmetric and positive denite. Instead of the linear system (C.2), we now
consider the transformed system
2. Iterate for k = 0, . . . , N 2
k
sk = Apk , k = (sk , pk ), k = , xk+1 = xk k pk
k
and compute the new residual
1 k+1
rk+1 = rk k sk , v k+1 = CA r , k+1 = (v k+1 , rk+1 ).
Stop, if
k+1 2 0
is satisfied with some prescribed accuracy .
Otherwise compute
k+1
k = , pk+1 = v k+1 + k pk .
k
For the preconditioned conjugate gradient scheme, we then obtain the error
estimate
k+1 2q k+1
e e0 ,
A 1+q 2(k+1) A
where
C.3 Iterative Solution Methods 263
5
+1
2 (A)
max (A) cA
q = , =
2 (A) 2 ,
1
2 (A)
min (A) cA
1
and cA A
1 , c2 are the positive constants from the spectral equivalence inequalities
x, x
cA
1 ( ) (A
x, x ) cA
2 ( ) for all x
x, x RN .
1/2
= CA x, this is equivalent to the spectral equivalence inequalities
Inserting x
1 (CA x, x) (Ax, x) c2 (CA x, x)
cA for x RN .
A
(C.11)
For a symmetric and positive denite matrix A, we have used the Krylov
space ( )
Sk (A, r0 ) = span r0 , Ar0 , . . . , Ak r0
to construct an Aorthogonal vector system P (cf. (C.3)). Formally, such a
vector system can be dened for any arbitrary matrix A RN N . However,
a nonsymmetric and possibly indenite matrix A does not induce an inner
product. Instead of an Aorthogonal vector system, we therefore dene an
orthonormal vector system
( )N 1
V = v k k=0
satisfying
k 1 for k = ,
(v , v ) =
0 for k =
using the method of Arnoldi as described in Algorithm C.3.
Algorithm C.3
1. Compute for an arbitrary given initial solution x0 RN
r0
r0 = Ax0 f , v0 = .
r0 2
2. Iterate for k = 0, . . . , N 1
k
v k+1 = Av k k v , k = (Av k , v ) .
=0
264 C Numerical Algorithms
k
xk+1 = x0 v ,
=0
where we have to nd the yet unknown coecients . To this end, we may re
quire to minimise the residual rk+1 = Axk+1 f with respect to the Euclidean
vector norm,
k+1
k
r = Axk+1 f = Ax0 f Av 2 min ,
2 2
=0
+1
Av = v +1 + j v j = j v j , +1 = v +1 2 .
j=0 j=0
Hence, we have
k
+1
rk+1 = r0 j v j = r0 Vk+1 Hk
=0 j=0
where the notation e0 = (1, 0, . . . , 0) Rk+2 has been used. Since Vk+1 is
orthogonal, we deduce
C.3 Iterative Solution Methods 265
k+1
r = Vk+1 r0 2 e0 Hk
2 2
0 0 0
= r 2 e Hk 2 = r 2 Qk e0 Qk Hk 2 ,
k+1 2
= r0 2 Qk e0 Rk
=0
k
2 2 2
= r0 2 Qk e0 Rk + r0 2 Qk e0 = r0 2 Qk e0 ,
k+1 k+1
=0
if the coecient vector Rk+1 is found from the upper triangular linear
system
Rk = r0 2 Qk e0 .
It remains to nd an orthogonal matrix Qk R(k+2)(k+2) transforming the
upper Hessenberg matrix
0,0 1,0 . . . k,0
..
0,1 1,1 .
.. R(k+2)(k+1)
Hk = 0 1,2 . . . .
.
0 . . k,k
k,k+1
This can be done by the use of the Givens rotations. Let us rst consider the
column vector
hj = (j,0 , . . . , j,j1 , j,j , j,j+1 , 0, . . . , 0) Rk+2 ,
j
is satised. For this, it is sucient to consider the orthogonal matrix G
R 22
such that
j,j j,j
Gj =
j,j+1 0
j R22 allows the general representation
is fullled. The orthogonal matrix G
aj b j
Gj = , a2j + b2j = 1 ,
bj aj
where the coecients aj and bj can be found from the condition
bj j,j + aj j,j+1 = 0
as
j,j j,j+1
aj = 8 , bj = 8 ,
2
j,j + 2
j,j+1 2
j,j 2
+ j,j+1
and, therefore, when assuming j,j+1 > 0,
8
j,j = aj j,j + bj j,j+1 = 2 + 2
j,j j,j+1 > 0. (C.12)
Qk = Gk Gk1 . . . G1 G0 R(k+2)(k+2) ,
which fulls
1 a0
0 b0
0 0
Qk e = Gk . . . G0 . = Gk . . . G1 .
0
.. ..
0 0
0 0
a0 a0
a1 (b0 ) a1 (b0 )
(b0 )(b1 ) a2 (b0 )(b1 )
= Gk . . . G2 .. = .. Rk+2 .
. .
0 ak (b0 ) (bk1 )
0 (b0 ) (bk )
From this, we nd
k
k+1 = e0 2 (Qk e0 )k+1 = e0 2 bj .
j=0
k
wk = Av k , v k+1 = wk k v , k = (wk , v ), kk+1 = v k+1 2 .
=0
Go to 3. if kk+1 = 0 is satisfied.
Otherwise compute
1
v k+1 = v k+1 .
kk+1
For = 0, . . . , k 1 compute
and
kk kk+1 8
ak = 8 , bk = 8 , kk = 2 + 2
kk kk+1
2 + 2
kk 2 + 2
kk
kk+1 kk+1
as well as
and
k
xk+1 = x0 v .
=0
References
77. J. C. Nedelec. Integral equations with non integrable kernels. Int. Eq. Operator
Th., 5:562572, 1982.
78. J. C. Nedelec. Acoustic and Electromagnetic Equations. Springer, New York,
2001.
79. J. C Nedelec and J. Planchard. Une methode variationelle delements nis
pour la resolution numerique dun probleme exterieur dans R3 . R.A.I.R.O.,
7:105129, 1973.
80. C. Neumann. Untersuchungen uber das Logarithmische und Newtonsche Po
tential. Teubner, Leipzig, 1877.
81. E. J. Nystrom. Uber die praktische Auosung von linearen Integralgleichungen
mit Anwendungen auf Randwertaufgaben der Potentialtheorie. Soc. Sci. Fenn.
Comment. Phys. Math., 4:152, 1928.
82. E. J. Nystrom. Uber die praktische Auosung von linearen Integralgleichungen
mit Anwendungen auf Randwertaufgaben. Acta Math., 54:185204, 1930.
83. G. Of, O. Steinbach, and W. L. Wendland. Applications of a fast multipole
Galerkin boundary element method in linear elastostatics. Comput. Vis. Sci.,
8(34):201209, 2005.
84. G. Of, O. Steinbach, and W. L. Wendland. The fast multipole method for the
symmetric boundary integral formulation. IMA J. Numer. Anal., 26(2):272
296, 2006.
85. N. Ortner. Fundamentallosungen und Existenz von schwachen Losungen lin
earer partieller Dierentialgleichungen mit konstanten Koezienten. Ann.
Acad. Sci. Fenn. Ser. A I Math., 4:330, 1979.
86. N. Ortner and P. Wagner. A survey on explicit representation formulae for
fundamental solutions of linear partial dierential operators. Acta Appl. Math.,
47:101124, 1997.
87. J. Ostrowski, Z. Andjelic, M. Bebendorf, B. CranganuCretu, and J. Smajic.
Fast BEMSolution of Laplace Problems with HMatrices and ACA. IEEE
Trans. on Magnetics, 42(4):627630, 2006.
88. J. Plemelj. Potentialtheoretische Untersuchungen. Teubner, Leipzig, 1911.
89. A. Pomp. The BoundaryDomain Integral Method for Elliptic Systems. With
an Application to Shells, volume 1683 of Lecture Notes in Mathematics.
Springer, Berlin, 1998.
90. B. Reidinger and O. Steinbach. A symmetric boundary element method for
the Stokes problem in multiple connected domains. Math. Methods Appl. Sci.,
26:7793, 2003.
91. F. J. Rizzo. An integral equation approach to boundary value problems of
classical elastostatics. Quart. Appl. Math., 25:8395, 1967.
92. D. Rodger, N. Allen, H. C. Lai, and Leonard P. J. Calculation of transient
3D eddy currents in nonlinear media = verication using a rotational test rig.
IEEE Transaction on Magnetics, 30(5(2)):29882991, 1994.
93. V. Rokhlin. Rapid solution of integral equations of classical potential theory.
J. Comput. Phys., 60(2):187207, 1985.
94. V. Rokhlin and M. A. Stalzer. Scalability of the fast multipole method for
the Helmholtz equation. In Proceedings of the Eighth SIAM Conference on
Parallel Processing for Scientic Computing (Minneapolis, MN, 1997), page 8
pp. (electronic), Philadelphia, PA, 1997. SIAM.
95. K Ruotsalainen and W. L. Wendland. On the boundary element method for
some nonlinear boundary value problems. Numer. Math., 53:299314, 1988.
274 References
Гораздо больше, чем просто документы.
Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.
Отменить можно в любой момент.