The boundary element method : errors and gridding for

problems with hot spots

Citation for published version (APA):
Kakuba, G. (2011). The boundary element method : errors and gridding for problems with hot spots. Technische
Universiteit Eindhoven. https://doi.org/10.6100/IR696955

DOI:
10.6100/IR696955

Document status and date:

Published: 01/01/2011

Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:
[email protected]
providing details and we will investigate your claim.

Download date: 04. avr.. 2021

 The Boundary Element Method:
Errors and gridding for problems with
hot spots
Copyright c 2011 by Godwin Kakuba, Eindhoven, The Netherlands.
All rights are reserved. No part of this publication may be reproduced, stored
in a retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without prior permission of
the author.

CIP-DATA LIBRARY TECHNISCHE UNIVERSITEIT EINDHOVEN

Kakuba, Godwin

The Boundary Element Method:

Errors and gridding for problems with hot spots /
by Godwin Kakuba. -
Eindhoven, Eindhoven University of Technology, 2011.
A catalogue record is available from the Eindhoven University of Technology
Library

Proefschrift. - ISBN: 978-90-386-2443-3

NUR 919
Subject headings: boundary element methods / integral equations; numerical
methods /
boundary value problems /
boundary elements; numerical methods
2000 Mathematics Subject Classification: 65N38, 65N50, 65N55, 65N06, 74S15,
65R20, 35Q30, 80A25
 The Boundary Element Method:
Errors and gridding for problems with
hot spots

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan de

Technische Universiteit Eindhoven, op gezag van de
rector magnificus, prof.dr.ir. C.J. van Duijn, voor een
commissie aangewezen door het College voor
Promoties in het openbaar te verdedigen
op donderdag 31 maart 2011 om 16.00 uur

door

Godwin Kakuba

geboren te Rwengoma, Oeganda

Dit proefschrift is goedgekeurd door de promotor:

prof.dr. R.M.M. Mattheij

Copromotor:
dr.ir. M.J.H. Anthonissen
Acknowledgements

Having done a general and theoretical maths program in my bachelors degree

at Makerere university, I was very curious and interested in applied mathemat-
ics when I came to the Netherlands in 2003 for a masters in industrial and
applied mathematics. I was not disappointed. I liked and enjoyed the program
and therefore the choice of continuing with my PhD at the same University
was an easy one for me. I joined the CASA group as a PhD student and my
research with the group has led to this thesis. There are a lot of people whose
contribution to this cause is indispensable and I would like to register my ac-
knowledgements here to some of them.

In the first place, I would like to express my sincere gratitude to prof.dr. R.M.M.
Mattheij my promotor. He has given me tremendous support without which
this thesis would not have been a reality. This includes the support to attend
the many international conferences which were a delight. Then I would like
to thank dr.ir. M. J. H Anthonissen for the great support he has offered me
in working on this thesis. I benefited a lot from his expertise in local defect
correction techniques plus his comments as he read through my thesis.

I would like to thank the staff at the CASA group who never got tired of my
questions. In particular am very grateful to dr.ir. ter Morche, prof Jan de Graaf
and dr. M.E. Hochstenbach for the several fruitful discussions we had together.
To dr.ir. B.J. van der Linden, thank you for attending to my several IT issues.
I thank the CASA secretary mw. Enna van Dijk for the great all round support
she has extended to me since the time I got interested in studying at TU/e.

I have generally enjoyed working in the CASA group over the years, and I would
like to thank my colleagues for the pleasant working environment. I would like
to thank Hans Groot, Zoran Ilievski, Darcy Hou, Maria Ugryumova, Agnieszka
Lutowska, Mirela Darau and Roxana Ionotu for being great officemates over
the years. I also thank Zoran and dr. Christina Giannopapa for being excellent
housemates. I will be haunted if I do not mention having greatly enjoyed the
many outings, dinners and poker and sportscentrum games with my fellow PhD
students who included: Erwin Vondehoff, Peter in ’t pan Huijs, Jan Willem,
Hans Groot, Agnieszka Lutowska, Mirela Darau, Zoran Ilievski, Patricio Rosen,
vi Acknowledgements

Yves van Genip, Mark van Krij, John Businge, Yeneneh Yimer Yalew, Temesgen
Markos, Valeriu Savcenco, Ali Etaati, Maxim Pisarenco, Miguel Patricio, Fan
Yabin, Maria Rudnaya, Maria Ugryumova, Kundan Kumar, Tasnim Fatima,
Kho Changhi, Sudhir Srivastava, the list goes on and on. I also thank Willem
Dijkstra for the several helpful BEM discussions we had. Special thanks to
Mirela and Hans for helping me with the printing process.

Finally, I would like to express my great appreciation to my family and friends

at home for their support over the years. In particular I would like to regis-
ter my special thanks to my friends Doreen Karungi, Tumps Ireeta, Asbjorn
Atuhaire and John Kitayimbwa for all the support and courage they have al-
ways given me. In addition I thank Doreen for reading through and providing
lots of grammar corrections. I am very grateful to my colleagues at Makerere;
Fred Mayambala, Ismail Mirumbe and David Dumba for helping me get more
time to finish up the writing, not forgetting the head Dr. J. Kasozi and the
dean of science Prof J. Y. T. Mugisha for their understanding and support in
the process. Let me also register my thanks to Dr. B. S. Kato and his family
for the several holiday trips I enjoyed with them in London. I totally appreciate
the unyielding support of my parents and I hope I have usefully enjoyed the
freedom you gave me to do whatever I want in and with my education.
Contents

List of notations ix

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Outline of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 The boundary integral equations 7

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Integral equation for points inside Ω . . . . . . . . . . . . . . . . . 10
2.3 Integral equation for boundary points . . . . . . . . . . . . . . . . 11
2.4 Integral equation for points outside Ω . . . . . . . . . . . . . . . . 13
2.5 Operator theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Eigenvalues of Ks and Kd : A circular boundary example . . 17
2.5.2 Uniqueness of the solution to the Neumann problem . . . . 18

3 The boundary element method 21

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Constant elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Linear elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Boundary corner singularities . . . . . . . . . . . . . . . . . . . . . 25
3.5 Matrix elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.6 Numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4 Local errors in BEM for potential problems 35

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 A survey of error sources . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.1 Defining local errors . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Dirichlet problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.1 Constant elements . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.2 Linear elements . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Neumann problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4.1 Constant elements . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4.2 Linear elements . . . . . . . . . . . . . . . . . . . . . . . . . 58
viii Contents

4.5 Mixed boundary conditions . . . . . . . . . . . . . . . . . . . . . . . 60

4.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.7 Equidistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5 Global errors 73
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Spectral decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 Dirichlet problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.4 Neumann problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5 Mixed problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6 Local Defect Correction for BEM 87

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 LDC formulation with an introductory example: A Neumann prob-
lem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3 LDC formulation: A Dirichlet problem . . . . . . . . . . . . . . . . 95
6.4 Complexity of the algorithm . . . . . . . . . . . . . . . . . . . . . . 97
6.5 LDC algorithm as a fixed point iteration . . . . . . . . . . . . . . . 99
6.6 Continuous formulation of the LDC steps . . . . . . . . . . . . . . 107

7 The potential problem for the impressed current cathodic protection

system 117
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.3 BEM-LDC for the ICCP problem . . . . . . . . . . . . . . . . . . . . 122

Bibliography 136

Index 137

Summary 139

Curriculum vitae 141

List of Notations ix

Ω: a problem domain 7

∂Ω : the boundary of Ω 7

χ: an arclength coordinate along the boundary ∂Ω 9

r: coordinates in R2 9

Hr (Ω) : the Sobolev space of order r on a domain Ω 15

C(Ω) : a space of continuous functions on a two dimensional space Ω 15

N
∂Ωj : a partitioning of ∂Ω into N elements such that ∪ ∂Ωj = ∂Ω 21
j=1

Γ: a numerical boundary representing ∂Ω 21

Γj : a partitioning of Γ into N elements each representing ∂Ωj 21

L, l : a grid of size L or l 22

uj , qj : the values u(rj ) and q(rj ) respectively in a node rj 23

a BEM approximation of uj and qj respectively on

uLj , qLj :
a grid of size L 24

lj : the length/size of element Γj 26

fu (r), fq (r) : shape function for u(r) and q(r) respectively 41

r(χ) : a coordinate in R2 at a local coordinate χ 45

rj (χ) : the distance to a point χ on the j-th element 45

Ωlocal : a region of Ω where the solution has high activity 88

Γlocal : the boundary of Ωlocal 89

Γinside : a part of Γlocal that is contained inside Ω 89

a part of Γlocal that contains the high activity and

Γactive :
belongs to Γ as well 89

Γc : a part of Γ that is outside of Γactive 89

ΓL : a uniform grid discretisation of Γ into elements ΓjL of size L 90

x Notation

l
l
Γlocal : a uniform grid discretisation of Γlocal into elements Γlocal,j each of
size l 90

Γlocal : the boundary of Ωlocal 89

l l
Γactive : part of the uniform grid Γlocal that belongs to Γinside 90

rllocal : nodes of the local fine grid of size l 90

rlactive : nodes of the local fine grid that belong to Γactive 90

rlinside : nodes of the local fine grid that belong to Γinside 90

a composite grid discretisation of Γ into elements of size l in Γactive

Γ l,L :
and size L in Γc 91

rl,L : nodes of the composite grid 91

Chapter 1

Introduction

1.1 Background

The boundary element method (BEM) is a numerical method for approximating

solutions of boundary value problems (BVPs). The most important and unique
feature of BEM is that only the boundary of a domain needs to be discretised.
This is also its most important advantage over its main competitor finite element
methods (FEM) and other methods like finite difference methods(FDM). It makes
BEM suitable for exterior problems. Also problems in which singularities or
discontinuities occur in the domain can be handled efficiently by BEM solving
is at the boundary. Another advantage of the BEM is that the functions at the
boundary and their normal derivatives are solved for at the same time and with
the same degree of accuracy.

The main disadvantage of BEM is that it requires the knowledge of fundamen-

tal solutions for a problem partial differential equations in order to be used and
these are not readily available for all problems. This factor contributed to the
initially relatively slower development of BEM. The method also involves com-
puting singular integrals. However, today these fundamental solutions and the
singular integrals involved with the method are well understood and the BEM
has undergone rapid advancement in recent years [16]. The basic foundation
for BEM are integral equations. It has been known for well over a century that
BVPs whose fundamental solutions are known such as the Laplace equation
and the equations of linear elastostatics can be formulated as a boundary in-
tegral equation (BIE), see for instance [36]. The unknown boundary data then
depends on the prescribed data in the BIE. Once all the data becomes avail-
able through the solution of the BIE, then the solution to the BVP is obtained
through the application of Green’s identity.
2 Introduction

A BEM formulation can be classified as either a direct formulation or indi-

rect formulation. The basis for the indirect formulation can be credited to
Fredholm. In 1903 he already used discretised integral equations for po-
tential problems [20]. The basis of the direct formulation can be credited
to Somigliana who in 1886 presented an integral equation relating displace-
ments and stresses [69]. Since then a number of books and papers have
been published including Kellog [36], Mushkhelishvili [55], Mikhlin [50] and
Kupradze [42]. Their results are, however, limited to simple problems as the
integral equations had to be solved with analytical procedures and without the
aid of computers.

The real breakthrough came in the nineteen sixties when the integral equa-
tions for two dimensional potential problems were discretised using rectilinear
elements [27, 75] by Jawson and Symm. At each line element the functions are
approximated by constants. Their method cannot be classified as exactly a di-
rect formulation because the functions needed to be differentiated or integrated
to obtain physical quantities [16]. A direct formulation was introduced by
Rizzo [63], who also used discretised integral equations to relate displacements
and tractions in two-dimensional elasticity theory. The extension to three di-
mensions was given by Cruse [14], using triangular elements to describe the
domain boundary. In 1963 Jawson and Symm [27, 75] demonstrated that the
associated integral equation can be solved accurately and reliably using the
numerical methods. In 1978, CA Brebbia formally introduced the terminol-
ogy boundary-element method (BEM) for the first time in contrast to what had
already been established terminology, finite element method.

Presently, the BEM is well established as an effective alternative in the solu-

tion of engineering problems from a variety of application fields which include
acoustics, fracture mechanics, potential theory, elasticity theory and viscous
flows. However, estimating the BIEs numerically raises the need for error and
convergence analysis. For BEM, the mathematical foundation for such analy-
sis was laid down in Hsiao and Wendland [26], who performed such error and
convergence analysis for the Galerkin formulation. Error analyses for the collo-
cation formulation were pioneered by Arnold, Saranen and Wendland [2, 3, 64]
during the eighties. They mostly provide an indirect way of measuring error
and are usually based on the Galerkin formulation. Techniques include using
solutions obtained by using different collocation points or different formula-
tions, [22, 29, 44].

In this thesis direct error estimates for collocation BEM are derived and more-
over in the maximum norm. Starting from its basic formulation, error esti-
mates for collocation BEM are derived by tracking down the interpolation error
committed at the elements.

The issue of error analysis becomes even more crucial when considering prob-
lems with hot spots where specialised gridding is necessary. Hot spots refers
to small regions of high activity. Hot spots appear in a variety of problems. For
1.2 Outline of this thesis 3

example in combustion problems [1] or in corrosion problems, which often ex-

hibit regions of high activity where the corrosion is taking place, the microwave
heating problem [62], antenna transmissions [46], [82] and also in fluid dy-
namics [80]. If these spots are located well inside the domain, then standard
BEM has no problem since it is the boundary that is discretised. However such
spots can be in or very close to the boundary and as such require special at-
tention in a BEM discretisation. In this thesis we consider one such problem,
viz corrosion, and look at the potential problem of impressed current cathodic
protection system.

In particular, functions with hot spots are an instant of multi-scales problems,

that is, the problem is such that there is a large region where the variation is
relatively benign and small regions of rapid variation. This poses the problem
of forming a grid that accurately captures all the activity but with a high degree
of computational efficiency. The distribution of the elements therefore becomes
a very important decision. In the competing field of finite elements such a need
has generated a large amount of literature of which a good review can be found
in [71] and [72]. While adaptive meshing in finite elements is quite advanced,
in boundary element methods the research is only at the beginning. Quite a
few approaches, including element residual methods, flux projection estimates,
extrapolation error estimates and estimates measuring the sensitivity of the
numerical solution to a shift in collocation points have been considered. A good
review of the various approaches is presented in [45]. In this thesis a new form
of gridding for problems with hot spots in the name of Local Defect Correction
(LDC) is presented. This method has been tested and is well documented in the
other numerical methods such as Finite Difference Methods and Finite Volume
Methods [1, 21, 51]. It is a proven efficient approach for solving problems with
hot spots with a control on complexity. An approach for using LDC in BEM is
presented here.

1.2 Outline of this thesis

The basis for the boundary element method is a boundary integral equation.
In Chapter 2, boundary integral equations for boundary value problems using
the direct formulation are discussed. As is elaborated there, the BEM operator
has different properties depending on the kind of boundary conditions, that
is either Dirichlet, Neumann or mixed boundary conditions. In this chapter
properties of these operators that are helpful for the work covered in this thesis
are introduced and discussed. A boundary element discretisation uses either
constant elements, linear elements, quadratic elements or even higher order
elements depending on the assumptions followed on the given functions during
the formulation. In Chapter 3, constant and linear elements formulations are
introduced and discussused. These are the formulations we analyse and use
throughout this thesis. They also turn out to be the most used formulations in
4 Introduction

literature for the direct approach. Also in this chapter a set of examples that
will be used to verify our findings in this thesis are discussed.

Chapter 4 is dedicated to discussing local errors in BEM for both the constant
and linear elements formulations. The question of local errors in BEM is an in-
teresting one because by its nature, the boundary integral is a global method.
However, we show that we are able to define and use the error committed on
each element to define the local error in a point at the boundary. For the first
time we derive direct error estimates, that is, we start from the exact assump-
tions made in the formulation of BEM for a local error analysis in this chapter.
Local error behaviour for the constant and linear element formulations is pre-
dicted and verified. A sublocal error is introduced and it is shown that it is
third order in grid size. Using this sublocal error it is shown that the local error
is second order in grid size for constant elements as well as linear elements.
Then this can be exploited to obtain an error equidistributing grid.

In any numerical solution, the error that the user is interested in is the global
error. Having understood the local error in Chapter 4, good information is now
available to go on and analyse the global error. This is the subject of Chapter 5.
The problem on a circle is considered since in this case, all the analytical eigen-
functions and eigenvectors of the integral operators are available. Therefore a
spectral analysis and the information on local errors is used to derive bounds
for the global error.

Now that information on both the local and global errors is available, we present
in Chapter 6 a local defect correction (LDC) formulation for the boundary el-
ement method. It is shown here that LDC for BEM is a reliable strategy for
obtaining solutions on a composite grid but with less complexity. This is en-
tertaining news since BEM matrices are usually full matrices that would rather
require more memory and computation time as compared to when LDC is used.

The local Local defect correction formulation discussed here can be applied to
several problems with hot spots where BEM is used such as in electromagnet-
ics, fluid mechanics and structural mechanics. Chapter 7 discusses such an
application to a potential problem for an impressed current cathodic protection
(ICCP) system. ICCP is a system used to protect steel structures from corro-
sion. Such structures include submarines such as ships, underground pipes,
over ground storage tanks. Chapter 7 discusses an LDC model for ICCP for a
storage tank.
“But those who hope in the Lord will renew their strength. They will
soar on wings like eagles; they will run and not grow weary, they will
walk and not be faint.”–Isaiah 40:31
6 Introduction
Chapter 2

The boundary integral

equations

2.1 Introduction

Consider a closed domain Ω with boundary ∂Ω. On Ω, consider the potential

equation

∇2 u(r) = f(r), r ∈ Ω. (2.1.1)

The solution u = u(r) of (2.1.1) represents the potential produced at a point r in

a domain Ω due to a source f(r) distributed over Ω. Denote by n the outward
unit normal at ∂Ω.

n
Ωc n
Ωc

Ω Ω
∂Ω1
2
∂Ω
∂Ω
(b) Domain illustration for a mixed
(a) Domain illustration for a Dirichlet problem
or Neumann problem

Figure 2.1: Definition of terms for the potential problems given

in (2.1.3a) to (2.1.3c).
8 The boundary integral equations

Throughout this thesis we consider the Laplace equation for which f(r) = 0,
that is,

∇2 u(r) = 0, r ∈ Ω, (2.1.2)

for which the following boundary conditions may be defined on ∂Ω:

(1) Dirichlet boundary conditions;

u(r) = g(r), r ∈ ∂Ω, (2.1.3a)

where g is a given function.

(2) Neumann boundary conditions;
∂u
(r) = n · ∇u(r) = h(r), r ∈ ∂Ω, (2.1.3b)
∂n
where h is a given function.
(3) Mixed boundary conditions;

 u(r) = g(r)
 r ∈ ∂Ω1
(2.1.3c)
 ∂u (r) = h(r)

r ∈ ∂Ω 2
∂n
where ∂Ω1 ∪ ∂Ω2 = ∂Ω and ∂Ω1 ∩ ∂Ω2 = ∅ and g and h are given functions.

Subsequently a problem with Neumann boundary conditions will be referred

to as a Neumann problem, that with Dirichlet boundary conditions a Dirichlet
problem and that with mixed boundary conditions a Mixed problem. The funda-
mental solution of the Laplace equation is the solution of the singularly forced
Laplace equation

∇2r v(s; r) + δ(s; r) = 0, s, r ∈ Ω∞ , (2.1.4)

where r is the variable field point, s is the fixed location of the source point or
pole and Ω∞ denotes the infinite domain which is the whole plane in 2D. The
subscript r on the operator ∇ means differentiation is with respect to r.

The Dirac delta distribution δ(s; r) satisfies the following properties [60, p 12], [35,
p 21]:

Z  1, if s ∈ Ω,
δ(s; r) dΩ = (2.1.5)
Ω 
0, if s ∈ / Ω.

The Dirac delta is not truly a function but rather a generalised function. For
example, any function that is zero everywhere except in a single point has total
2.1 Introduction 9

integral zero but the Dirac delta has integral one [39, p 271]. It is used to model
phenomena with sufficiently concentrated properties such as a point charge or
point force.

For a two-dimensional case, introducing polar coordinates and using the prop-
erties of the Dirac delta distribution, the solution of (2.1.4) is found to be
1 1
v(s; r) = log (2.1.6)
2π r
p
where r = (x, y), s = (xs , ys ) and r := (x − xs )2 + (y − ys )2 is the Euclidean
distance from s to r. Let us also introduce a boundary arclength coordinate
defined along ∂Ω is shown in Figure 2.2.

∂Ω

χ0 χ
r(χ)

Figure 2.2: Definition of arclength coordinate

χ and the position r(χ) which is r at χ, χ0 is the
origin.

In this chapter we derive integral relations for the potential u(s) for s at differ-
ent locations of the domain Ω. These relations have been abundantly derived
in literature and are readily available in various books on boundary element
methods such as [57, p. 38], [35, p. 28], [59]. They are presented here for
completeness. The relations are derived starting from the following Green’s
identity:

Theorem 2.1.1 (Green’s second identity) Let φ, defined in Ω, and ψ, defined

in Ω × Ω, be two scalar functions which are continuous and admit continuous
partial derivatives, then
Z  
φ(z)∇2z ψ(s; z) − ψ(s; z)∇2 φ(z) dΩ(z)
Ω
Z 
∂ψ ∂φ

= 
 φ(r(χ)) (s; r(χ)) − ψ(s; r(χ))  dχ, s ∈ Ω, (2.1.7)
(r(χ))
∂Ω ∂n ∂n
where χ is an arc length coordinate in ∂Ω.

In (2.1.7), n is the outward unit normal at ∂Ω as shown in Figure 2.1. To

derive the integral relations for the Laplace equation, use is made of the iden-
10 The boundary integral equations

tity (2.1.7). We take φ as the unknown function u and ψ the fundamental

solution (2.1.6).

2.2 Integral equation for points inside Ω

Starting from Green’s second identity (2.1.7), an integral relation for a point
s ∈ Ω is derived, see Figure 2.3. Since the fundamental solution v is singular at
the point s, the domain of integration where Green’s second identity is applied
must be defined isolating the point s. We construct a ball Ωǫ of radius ǫ around
it. Then the new domain of integration is now Ω − Ωǫ with boundary ∂Ω + ∂Ωǫ ,
see Figure 2.3.

r(χ) n

z
n ∂Ω
s
Ωǫ ǫ
Ω − Ωǫ
∂Ωǫ

Figure 2.3: Definition of the domain of

integration for internal points.

Then apply (2.1.7) on Ω − Ωǫ and take the limit as ǫ → 0. Thus, with z ∈ Ω − Ωǫ

and s ∈ Ωǫ , we have
Z  
lim u(z)∇2 v(s; z) − v(s; z)∇2 u(z) dΩ(z)
z
ǫ→ 0
Ω−Ωǫ
Z 
∂v ∂u

= 
 u(r(χ)) (s; r(χ)) − v(s; r(χ)) (r(χ))
 dχ
∂n ∂n
∂Ω
Z 
∂v ∂u

+ lim u(r(χ)) (s; r(χ)) − v(s; r(χ)) (r(χ))
  dχ. (2.2.1)
ǫ→ 0 ∂n ∂n
∂Ωǫ

Since the point s is outside Ω − Ωǫ , use (2.1.5) to conclude that integrating the
first term on the left hand side yields 0. For the integral of the second term on
2.3 Integral equation for boundary points 11

the left hand side, note that

v(s; z) = v(z; s)

and for the Laplacian

∇2 u(z) = f(z) = 0

so that
Z Z Z
lim v(s; z)∇2 u(z)dΩ(z) = lim v(z; s)f(z)dΩ(z) = v(z; s)f(z)dΩ(z) = 0.
ǫ→ 0 ǫ→ 0
Ω−Ωǫ Ω−Ωǫ Ω
(2.2.2)

Now consider the second integral on the right hand side of (2.2.1). The bound-
ary ∂Ωǫ is a circumference of radius ǫ, the normal n is along the radius of Ωǫ
and towards the point s. So

∂v ∂v 1
dχ = ǫdθ, =− = .
∂n ∂r r=ǫ 2πǫ

Then
Z 
∂v ∂u

lim u(r(χ)) (s; r(χ)) − v(s; r(χ)) (r(χ))
  dχ
ǫ→ 0 ∂n ∂n
∂Ωǫ
2π
Z 
1 1

= lim u(r(χ))
 ǫ+  dθ = u(s). (2.2.3)
ǫ log ǫ
ǫ→ 0 2πǫ 2π
0

Therefore (2.2.1) becomes

Z  
∂u ∂v
u(s) = v(s; r(χ)) (r(χ)) − u(r(χ)) (s; r(χ)) dχ, s ∈ Ω. (2.2.4)
∂n ∂n
∂Ω

The integral relation (2.2.4) expresses the value of u at any point s in the domain
Ω in terms of its values and normal derivatives at the boundary. Thus if u and
∂u/∂n are known at the boundary, u can be computed at any point in the
domain.

2.3 Integral equation for boundary points

Consider the case s ∈ ∂Ω. In a similar fashion as above, draw a ball of radius
ǫ around the point s in ∂Ω, see Figure 2.4. Note that only part of the ball lies
within Ω; denote this intersection by Ωǫ . For generality, consider a point s
located at a corner. Let ∂Ω1ǫ be the part of ∂Ω from point A to s and s to B. Let
12 The boundary integral equations

Ω − Ωǫ

∂Ω2ǫ
B
A Ωǫ n
ǫ θ2 θ1
s x

Figure 2.4: Definition of the domain of integration

for boundary points.

∂Ω2ǫ be the arc AB. Applying Green’s identity (2.1.7) in Ω − Ωǫ , the left hand
side is zero since the Laplace equation is considered and the singular point for
the fundamental solution is outside Ω − Ωǫ . Then
Z  
∂v ∂u
0 = lim u(r(χ)) (s; r(χ)) − v(s; r(χ)) (r(χ)) dχ
ǫ→ 0 ∂Ω−∂Ω1
ǫ
∂n ∂n

Z  
∂v ∂u
+ lim u(r(χ)) (s; r(χ)) − v(s; r(χ)) (r(χ)) dχ. (2.3.1)
ǫ→ 0 ∂Ω2
ǫ
∂n ∂n

The limit of the first integral of (2.3.1) is the integral over the whole of ∂Ω.
Consider the first term of the second integral, we have

∂v 1 ∂r 1
(s; r(χ)) = − = . (2.3.2)
∂n 2πr ∂n 2πr
In this case r = ǫ and dχ = −ǫdθ. So

Z θ
Z2
∂v 1 θ1 − θ2
lim u(r(χ)) dχ = lim u(r(θ)) (−ǫdθ) = u(s) (2.3.3)
ǫ→ 0 ∂n ǫ→ 0 2πǫ 2π
∂Ωǫ θ1

because the point r(χ) approaches s as ǫ → 0.

The second term of the second integral of (2.3.1) gives,

Z Z θ2 !
∂u ǫ ln ǫ ∂u
lim v(s; r(χ)) (r(χ)) dχ = lim (r(θ)) dθ = 0. (2.3.4)
ǫ→ 0 ∂Ω
ǫ
∂n ǫ→ 0 2π θ1 ∂n
2.4 Integral equation for points outside Ω 13

Substituting (2.3.3) and (2.3.4) in (2.3.1) gives the integral relation

Z  
θ1 − θ2 ∂u ∂v
u(s) = v(s; r(χ)) (r(χ)) − u(r(χ)) (s; r(χ)) dχ. (2.3.5)
2π ∂Ω ∂n ∂n
In the limit as ǫ → 0 the angle (θ1 −θ2 ) is the internal angle at s which we denote
α(s). For a flat surface this angle is π and therefore the coefficient (θ1 − θ2 )/2π
in (2.3.5) is 1/2.

2.4 Integral equation for points outside Ω

If the point s is outside Ω, again due to the properties of v(s; r) and since the
singularity point is now outside Ω then the left hand side of (2.1.7) is zero. So
we have
Z  
∂u ∂v
0= v(s; r(χ)) (r(χ)) − u(r(χ)) (s; r(χ)) dχ, s ∈ Ωc . (2.4.1)
∂n ∂n
∂Ω

Equations (2.2.4), (2.3.5) and (2.4.1) can be summarised as

Z  
∂u ∂v
c(s)u(s) = v(s; r(χ)) (r(χ)) − u(r(χ)) (s; r(χ)) dχ, (2.4.2)
∂n ∂n
∂Ω

where the coefficient c(s) is given by


 1, s ∈ Ω,





α(s)
c(s) := , s ∈ ∂Ω, (2.4.3)

 2π




0, s ∈ Ωc .
Using the properties of the fundamental solution v introduced in Section 2.1 it
can be shown that


 −1, s ∈ Ω,
Z 

∂v 
(s; r(χ)) dχ = −1/2, s ∈ ∂Ω, (2.4.4)
∂n 

∂Ω 


0, s ∈ Ωc ,
where n, Ω and ∂Ω are as defined in Figure 2.1. We now introduce the following
definitions of the single and double layer potentials respectively:
Z
Ks q(s) := v(s; r(χ))q(r(χ))dχ, (2.4.5a)
∂Ω

Z
d ∂v
K u(s) := (s; r(χ))u(r(χ))dχ. (2.4.5b)
∂Ω ∂n
14 The boundary integral equations

These integrals are called single and double layer potentials respectively by
making an analogy with the corresponding boundary distributions of electric
charges and charge dipoles in electrostatics, see [5, p 319], [31, p 42], [38, p
67], [60, p 21], [66]. Using (2.4.5), the integral equation (2.4.2) can be written
as

(cI + Kd )u(s) = Ks q(s). (2.4.6)

The operators Ks and Kd are called the single and double layer operator re-
spectively and I is the identity operator. When s is located at the boundary
the integral equation is referred to as a boundary integral equation (BIE). The
functions

ks (s; r) := v(s; r), (2.4.7a)

∂v
kd (s; r) := (s; r), (2.4.7b)
∂n
are called the kernel of the integral operators Ks and Kd respectively.

The operators Ks and Kd together with their kernels are popular and have been
extensively studied. In Section 2.5, some of their properties that will be helpful
for the topics studied in this thesis are presented.

2.5 Operator theory

In this section, some properties of the operators Ks and Kd that will be needed
in later chapters are presented. Most important, we discuss their spectral prop-
erties and present proven theorems on their compactness and continuity. As
discussed below, properties of Ks are important in the solution for a Dirichlet
as those of Kd are for a Neumann problem. We show that the operator Kd has
a zero eigenvalue, which results in a singular system and hence a non unique
solution. Here we discuss how to circumvent this problem to obtain a unique
solution. These properties will be helpful in the investigation of local and global
errors in Chapters 4 and 5. The continuity and boundedness properties will
also be helpful in the investigation of the local defect correction algorithm in-
troduced in Chapter 6.

Consider the integral equation (2.4.6). If the function q(r) at the boundary is
known, then the right hand side is known. Define

f(s) := Ks q(s),

then (2.4.6) becomes

(cI + Kd )u(s) = f(s). (2.5.1)

2.5 Operator theory 15

Equation (2.5.1) is known as a Fredholm integral equation of the second kind,

see for instance [5, 3], [24, 42], [38, 1]. Therefore, a Neumann problem re-
sults into a Fredholm integral equation of the second kind, which sometimes is
simply known as a second kind integral equation.

On the other hand, if the function u(r) is known at the boundary, then define
g(s) := (cI + Kd )u(s) and equation (2.4.6) becomes
Ks q(s) = g(s). (2.5.2)
Equation (2.5.2) is a Fredholm integral equation of the first kind which is some-
times simply called a first kind integral equation.

A considerable amount of theory for both the first and second kind integral
equations has already been established. Also, several numerical methods for
determining the unknowns approximately have been discussed, an increasingly
popular one being BEM. In contrast to the properties possessed by the second
kind integral equation, many of the properties of the first kind integral equation
are “very unpleasant and often surprising to even mathematicians,” [81, pp
1,2]. These properties are determined by the corresponding operators and in
this section a few of them are highlighted.

Introduce Hr (Ω) which denotes the Sobolev space V r,2 (Ω) which is given by
V r,2 = {f : Ω → R ||f||k,2 < ∞}
where
 1/2
X Z
||f||k,2 =  |Dα f|2 dx
|α|≤k Ω

and Dα f denotes the mixed partial derivatives of f of order α. The integer k

is the largest order derivative that the functions in W k,2 (Ω) all admit. This
discussion and in more detail is presented in [9, pp. 19-21].

Theorem 2.5.1 For Ω a simply connected bounded domain in R2 with smooth

boundary ∂Ω, the operators Ks and Kd map functions from the Sobolev space
Hr (∂Ω) to Hr+1 (∂Ω), r ∈ R.

Proof. See [9, pp. 249, 286-287]. 

In fact, for Ω smooth, Theorem 2.5.1 can actually be strengthened to mapping

from Hr (∂Ω) to C∞ (∂Ω) for any r ∈ R, see [9, p. 249].

In this thesis, we restrict ourselves to C(Ω), a space of continuous functions on

a two dimensional space Ω with a positive measure µ such that
Z
(f, g) = f(r)g(r) dµ. (2.5.3)
Ω
16 The boundary integral equations

So over the boundary ∂Ω, the inner product (2.5.3) is defined as

Z
(f, g) := f(r(χ))g(r(χ)) dχ. (2.5.4)
∂Ω

For vectors, we use the median 2-norm which for a vector u ∈ RM is defined
as, [48, p. 98],
M 1/2
1 X 
2
||u||2,M := √  ui  (2.5.5)
 
 
M i=1 

and the infinity norm ||d||∞ = max |di | where d = (d1 , d2 , . . . , dM )T . These
i=1,...,M
norms are used in Chapters 4 and 5 to estimate and compare local and global
errors in BEM solutions.

Theorem 2.5.2 (Compactness) The single layer operator Ks is a compact and

self-adjoint operator.

Proof. See [16, p. 16], [30, p. 121]. 

From the spectral theory of compact operators, it follows that the eigenvalues
of Ks have an accumulation point at zero. In fact, on a smooth boundary,
both the double layer and the single layer are compact operators, see [5, p.
439], [38, Theorem 2.22]. Thus, the possibility of a zero eigenvalue implies the
need for regularisation in the solution of the integral equations. Section 2.5.1
uses the particular case of a circle to demonstrate the eigenvalues of the single
and double layer operators. In Section 2.5.2 it is shown that the operator
associated with the second kind integral equation of the Neumann problem
has a zero eigenvalue and thus is singular. A regularisation strategy used to
obtain a unique solution is discussed.

First the following theorem implies that an integral operators can be expanded
in terms of its eigenvalues and eigenfunctions. This will be helpful in the
regularisation strategy discussed here and in the analysis of global errors in
Chapter5.

Theorem 2.5.3 Let A 6= 0 be a compact operator in a Hilbert space H and (λn )

its ordered sequence of eigenvalues. Then there exists an orthonormal sequence
of eigenvectors aj corresponding to the λj . For every u ∈ H the expression
X
u = u0 + (u, aj )aj
j

holds where u0 is in the null space of A and

X
Au = λj (u, aj )aj .
j

Proof. See [30, pp. 121,122]. 

2.5 Operator theory 17

2.5.1 Eigenvalues of Ks and Kd: A circular boundary example

A number λ is an eigenvalue of an operator K with eigenfunction f if

λf = K f. (2.5.6)
In what follows in this section we establish the eigenvalues and eigenfunctions
for the integral operators Ks and Kd . The case of a general boundary is rather
complicated and we consider here the case of a circular boundary. Consider
a circle of radius R. Let us introduce polar coordinates so that the distance r
between two points, a fixed point s and any other point r, at polar angles θs
and θ respectively is given by
r2 = ||s − r||2 = 2R2 [1 − cos(θs − θ)]. (2.5.7)
Also the normal at r is given by [cos θ, sin θ]T . So we have
1 1
Ks (s; r) = − log r = − [2 log R + log(2 − 2 cos(θs − θ))], (2.5.8)
2π 4π
1
Kd (s; r) = − . (2.5.9)
4πR
Then using the Fourier series for the function f(θ) = log(2 − 2 cos θ) gives [16, p
25]:

Ks q(θs ) =
Z " ∞
#
R 2π X 2
− 2 log R − (cos(nθs ) cos(nθ) + sin(nθs ) sin(nθ)) q(θ) dθ.
4π 0 n
n=1
(2.5.10)
So
Z 2π
R
Ks 1 = − 2 log R dθ = −R log R, (2.5.11)
4π 0

Z 2π Z 2π
R R
Ks cos kθs = − log R cos(kθ) dθ + cos(kθs ) cos2 (kθ) dθ
2π 0 2πk 0
Z 2π
R
+ sin(kθs ) sin(kθ) cos(kθ) dθ. (2.5.12)
2πk 0

Carrying out the integrals in (2.5.12) yields zero for the first and third integrals
and the second one yields Rπ cos(kθs )/(2πk). So
R
Ks cos kθs = cos(kθs ). (2.5.13)
2k
Z 2π Z 2π
s R R
K sin kθs = − log R sin(kθ) dθ + cos(kθs ) cos(kθ) sin(kθ) dθ
2π 0 2πk 0
Z 2π
R
+ sin(kθs ) sin2 (kθ) dθ (2.5.14)
2πk 0
18 The boundary integral equations

The first and second integrals in (2.5.14) yield zero where as the third integral
yields Rπ sin(kθs )/(2πk). Then
R
Ks sin kθs = sin(kθs ). (2.5.15)
2k

For the double layer

Z 2π
1
Kd 1 = − dθ = −1/2, (2.5.16a)
4π 0
Z 2π
1
Kd cos kθs = − cos kθ dθ = 0, (2.5.16b)
4π 0
Z 2π
1
Kd sin kθs = − sin kθ dθ = 0. (2.5.16c)
4π 0

From (2.5.11), (2.5.13) and (2.5.15) we conclude that −R log R is an eigenvalue

of Ks with eigenfunction 1 and R/2k is an eigenvalue of Ks with eigenfunctions
cos kθ and sin kθ. The later eigenvalue indeed shows that the eigenvalues of
Ks converge at zero. From (2.5.16a) to (2.5.16c) we conclude that −1/2 is an
eigenvalue of Kd with eigenfunction 1 and 0 is an eigenvalue of Kd with eigen-
functions cos kθ and sin kθ. In Section 2.5.2 it is shown that −1/2 is always an
eigenvalue of the double layer operator making the second kind integral equa-
tion on smooth boundaries always singular. How to go around this to obtain
unique solutions is also discussed.

2.5.2 Uniqueness of the solution to the Neumann problem

Result (2.4.4) implies that −1/2 is an eigenvalue of the double layer operator Kd
with eigenfunction 1. From (2.4.6), the Neumann problem can be represented
as the second kind integral equation
1
( I + Kd )u(s) = Ks q(s) =: f(s), (2.5.17)
2
where f(r) is a known function at ∂Ω. Since −1/2 is an eigenvalue of Kd with
1
eigenfunction 1, it implies that 0 is an eigenvalue of the operator ( I + Kd )
2
with eigenfunction 1. Thus the space of functions spanned by 1, that is the
1
constant functions, form the eigenspace of the operator ( I+Kd ) corresponding
2
to eigenvalue 0. Let us denote this space by W0 , that is,
W0 := span{1}. (2.5.18)
Let
W1 := W0⊥ (2.5.19)
2.5 Operator theory 19

be the orthogonal complement of W0 in C2 (Ω) ∩ C1 (Ω̄).

Consider the operator

1
K := I + Kd ,
2
which is also compact and Hermitian with a zero eigenvalue.

Consider the result of Theorem 2.5.3. Let {φj } be an orthonormal basis for
the eigenspace of K corresponding to the eigenvalues λj . Then expanding u in
terms of the basis {φj } and substituting into (2.5.17) gives,
X X
λj (u, φj )φj = λj αj φj = f,
j j

where

αj := (u, φj ).

Each αj can then be computed as

(f, φj )
αj = , λj 6= 0.
λj

Now,
∞
X
λ0 α0 φ0 + λj αj φj = f. (2.5.20)
j=1

Since λ0 is chosen to be zero, we have the freedom to choose u0 = (u, φ0 )φ0 , the
component of u in the direction of φ0 . This is responsible for the nonuniqueness
of the solution. Fixing α0 would imply fixing u and thus we obtain a unique
solution. The easiest option is to choose α0 = 0. What is more general and used
in practice is to choose u(s) = us for some s a point at the boundary, see for
instance [9, p 13]. Suppose it is given that u(s) = us for some point s at the
boundary. Then
∞
X
α0 φ0 (s) + αs φj (s) = us . (2.5.21)
j=1

So we have
X ∞
1
α0 = (us − αj φj (s)). (2.5.22)
φ0
j=1

The solution is the original solution but shifted.

“The purpose of computation is insight, not numbers. ”–Richard Ham-
ming
Chapter 3

The boundary element

method

3.1 Introduction

The integral equation (2.4.2) expresses the value of the potential u at any point s
in terms of its values and normal derivative at the boundary. A discretisation of
this equation leads to the Boundary Element Method (BEM) system of algebraic
equations. To this end, the physical boundary ∂Ω is partitioned into N parts
∂Ωj , j = 1, 2, . . . , N, see Figure 3.1. Each physical partition ∂Ωj is represented

∂Ω3 ∂Ω2
∂Ω4 Γ3 Γ2
∂Ω1
Γ4 Γ1

∂Ω5 Γ5
Γ9
∂Ω9
Γ6
Γ8
∂Ω6 Γ7
∂Ω8
∂Ω7

Figure 3.1: A discretisation of an ellipse

into N = 9 elements.

N
by a numerical partition Γj . The union Γ := ∪ Γj is what is called a numerical
j=1
boundary. For Γj we use rectilinear elements in which the ends of each partition
22 The boundary element method

∂Ωj are connected by a straight line of length lj to be called a boundary element,

here simply called element and lj the element size. A grid that has the same
element size l for all the elements is called a uniform grid of size l .

The elements are numbered according to the standard BEM convention: in-
creasing in the anticlockwise sense. On each of the elements Γj an assumption
is made on the functions u and q. Basically these functions are assumed to
vary as polynomials which are called shape functions. Depending on the order
of the shape functions on each element, the type of elements used is said to
be either constant, linear or even higher order. In this thesis examples in both
constant elements and linear elements are considered. In constant elements,
constant shape functions are used, that is, the functions on each element are
assumed constant. In linear elements linear shape functions are used, that is,
the functions on each element are assumed to vary linearly. In other cases,
quadratic or even higher order shape functions may be used.

3.2 Constant elements

fu or fq
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
u2 or q2
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
r3 r2
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

r4
r1

r5
r9

r6 r8
r7

Figure 3.2: Constant elements nodes r1 , . . . , r9 corre-

sponding to the discretisation in Figure 3.1 and dis-
cretisation of functions, fu , fq are the shape functions.

For constant elements an element is represented by a single node placed at the

midpoint of the element. The element containing the j-th node is denoted Γj .
These nodes are also used as the collocation points, that is, the points where
the integral equation is applied. The BIE is then applied at the i-th node and
integration over ∂Ω is estimated by the sum of the integrations on all the N
3.2 Constant elements 23

elements. That is,

N Z
X X N Z
∂v
c(ri )u(ri ) + u(r(χ)) (ri ; r(χ)) dχ = q(r(χ))v(ri ; r(χ)) dχ. (3.2.1)
Γj ∂n
j=1 j=1 Γj

Here the nodes ri , i = 1, 2, . . . , N, are the midpoints of the elements, see Fig-
ure 3.2. Next is the discretisation of the functions u and q. These functions are
assumed to be constant on each element and equal to their nodal values where
the nodes are the midpoints of the elements. So let us introduce the definitions

uj := u(rj ), qj := q(rj ), (3.2.2)

where rj is the midpoint of Γj . Then assume that

u(r) = uj and q(r) = qj for r ∈ Γj . (3.2.3)

Then using (3.2.3) in (3.2.1) gives

N
X Z XN Z
∂v
ci ui + uj (ri ; r(χ)) dχ = qj v(ri ; r(χ)) dχ, (3.2.4)
Γj ∂n Γj
j=1 j=1

for all the collocation points i = 1, 2, . . . , N. Define

Z Z
^ ij := ∂v
H (ri ; r(χ)) dχ, and Gij := v(ri ; r(χ)) dχ. (3.2.5)
Γj ∂n Γj

These are called influence coefficients because their values express the contri-
bution of the nodal values uj and qj to the formation of ci ui , [35, p. 49]. With
this notation equation (3.2.4) becomes
N
X N
X
ci ui + ^ ij uj =
H Gij qj . (3.2.6)
j=1 j=1

Setting

^ ij ,
Hij := ci δij + H (3.2.7)

where δij is the Kronecker delta, (3.2.6) can be written as

N
X N
X
Hij uj = Gij qj . (3.2.8)
j=1 j=1

Let H and G be N × N matrices whose elements are given by (3.2.7) and (3.2.5).
Also introduce the following vectors of length N :

u := (u1 , u2 , . . . , uN )T , q := (q1 , q2 , . . . , qN )T . (3.2.9)

24 The boundary element method

Then applying equation (3.2.8) for all the collocation points ri , i = 1, 2, . . . , N,

yields

Hu = Gq. (3.2.10)

In (3.2.10) we have a system of N algebraic equations in 2N unknowns uj and

qj , j = 1, 2, . . . , N, which, as it is now, is an underdetermined system. The extra
N relations needed to solve the system uniquely must come from the boundary
conditions. For each element j, either uj or qj is known through boundary
conditions.

For generality, let us assume a part ∂Ω1 of the boundary on which u(r) is given
and a part ∂Ω2 on which q(r) is given. Let these two parts be discretised into
N1 and N2 constant elements respectively such that N1 + N2 = N. Thus, there
are still N unknowns, N − N1 values of u(r) on ∂Ω2 and N − N2 values of q(r)
on ∂Ω1 which are to be determined from the system (3.2.10). Before solving the
system, the unknown need to be separated from the known quantities. To this
end, partition the matrices H and G and write (3.2.10) as

    
ũ1  q1 

 H1 H2 
 
  =  G1
 G2 
 
 ,
 (3.2.11)
u2 q̃2

where ũ1 and q̃2 denote the known quantities on Γ 1 (representing ∂Ω1 ) and Γ 2
(representing ∂Ω2 ) respectively. Then carry out the multiplications and move
all the unknowns to the left hand side of the equation to obtain

Ax = b (3.2.12)

where
 
q1 
 
A := −G1 H2 , x :=   , b := −H1 ũ1 + G2 q̃2 . (3.2.13)
  
 
u2

The solution of the system (3.2.12) gives a BEM approximation of the unknowns
in x in the grid nodes rj . We denote by xL a BEM approximation on a grid of
size L. Thus uLj (or qLj ) is a BEM approximation of uj (or qj ) using a grid of size
L.

If the N1 points and the N2 points where the values of u and respectively q are
prescribed are not consecutive then the partitioning of the matrices in (3.2.11)
is preceded by an appropriate rearrangement of columns in H and G. Solving
(3.2.12) gives the unknown boundary quantities of u and q. Therefore we now
have all the boundary quantities. The solution u(r) can then be computet at
any point r ∈ Ω using (3.2.4) with c(r) = 1.
3.3 Linear elements 25

fu or fq
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

u3 or q3 xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

u2 or q2
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxx
xxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

3 xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxx
xxxxx
xxxxx
xxxxx

4 2

5
1

6 9
7 8

Figure 3.3: Linear elements nodes and

discretisation of the functions, fu and fq
are shape functions.

3.3 Linear elements

In the case of linear elements the potential and its derivative are assumed to
vary linearly over each element. This thesis discusses continuous linear ele-
ments in which the nodes are placed at the extremes of the element, see Fig-
ure 3.3. So fu and fq are linear interpolation functions expressed in terms of
the values of u and q respectively at the extremes of the element. That is

fu (ξ) = a + bξ, fq (ξ) = c + dξ, (3.3.1)

where ξ is a local coordinate on an element (see (3.5.4)) and a, b and c, d

are constants to be determined by substituting the values of u and q respec-
tively at the extremes of the elements. For details on obtaining the system of
equations (3.2.12) when linear elements are used see [35, p. 111], [57, p. 78].

3.4 Boundary corner singularities

Most of this thesis is devoted to discussing errors in BEM. In Chapter 4 we

discuss sources of error that are due to the approximations that we make in
a BEM formulation. One inherent source of error in BEM solutions may be
due to the boundary on which we are solving. In particular boundaries with
corners, in two dimensions, and edges, in three dimensions, deserve special
attention due to the errors that occur in these regions. These errors are due
to the singular behaviour of the functions involved in these locations, see for
26 The boundary element method

instance [13], [54], [70]. Of course the potential at any point, be it a corner
or otherwise, can only have one value. However, there are two distinct normal
gradients at every corner, and several combinations of boundary conditions at
a corner are now possible, see [78].

The spatial derivative of the potential field governed by the Laplace and Pois-
son equations can become infinite at corners and edges rendering the standard
BEM as discussed in this thesis to give inaccurate results [24, p 282], [28, p
164], [56]. In [65] the authors have suggested a formulation of mixed elements
to overcome this problem. Due to this corner problem, in the course of measur-
ing error order in the subsequent chapters of this thesis we avoid boundaries
with corners. This is because any results obtained with such boundaries will
also reflect how well this corner problem is being circumvented.

3.5 Matrix elements

Forming the system (3.2.12) requires the formation of the matrices H^ and G for
which we need to evaluate the integrals in (3.2.5). In two dimensions we have
Z Z
1
Gij = v(ri ; r(χ)) dχ = − log ||ri − r(χ)|| dχ (3.5.1a)
Γj 2π Γj
and
Z Z
^ ij = 1 ∂v 1 (ri − r(χ), n(χ))
H (ri ; r(χ)) dχ = dχ. (3.5.1b)
2π Γj ∂n 2π Γj ||ri − r(χ)||2
We note that the integrals are singular when r → ri and therefore need special
treatment when i = j. For i 6= j, the integrals are nonsingular and are evaluated
numerically using Gaussian quadrature. To this end suppose we use rectilinear
elements and rj− 1 and rj+ 1 are the extremes of element j of length lj as shown
2 2
in Figure 3.4. We denote the grid size of element Γj by
lj := ||rj+1/2 − rj−1/2 ||. (3.5.2)
We do a parameterization of the element j in terms of a local coordinate ξ on Γj ,
that is
1 (rj+ 1 − rj− 1 )
r(ξ) := (rj− 1 + rj+ 1 ) + 2 2
ξ, (3.5.3)
2 2 2 lj
where
−lj /2 ≤ ξ ≤ lj /2. (3.5.4)
The Jacobian of this transformation is
s
 2  2
dx dy
J(ξ) = + = 1,
dξ dξ
3.5 Matrix elements 27

lj /2
rj+ 1
2

lj /2
rj

rj− 1
2

ri

Figure 3.4: Illustration of integration

terms on element Γj

where, from (3.5.3),

1 (xj+ 1 − xj− 1 )
2 2
x(ξ) = (xj− 1 + xj+ 1 ) + ξ,
2 2 2 lj
1 (yj+ 1 − yj− 1 )
2 2
y(ξ) = (yj− 1 + yj+ 1 ) + ξ.
2 2 2 lj
Then using the above parameterization the integrals (3.5.1) become
Z
1 lj /2
Gij = − log ||ri − r(ξ)|| dξ (3.5.5a)
2π −lj /2

Z lj /2
^ ij = 1 (ri − r(ξ), n)
H dξ. (3.5.5b)
2π −lj /2 ||ri − r(ξ)||2
Since Gaussian quadrature points are given for the interval [−1, 1] introduce a
new coordinate η defined as η := ξ/(lj /2). Then the integrals (3.5.5) become
Z
lj 1
Gij = − log ||ri − r(η)|| dη (3.5.6a)
4π −1

Z1
^ ij = lj (ri − r(η), n)
H dη, (3.5.6b)
4π −1 ||ri − r(η)||2
28 The boundary element method

where
1 1
r(η) := (rj− 1 + rj+ 1 ) + (rj+ 1 − rj− 1 )η (3.5.7)
2 2 2 2 2 2

and η is a local coordinate on the element such that −1 ≤ η ≤ 1.

When i 6= j the integrands in (3.5.6) are nonsingular and the integrals can be
evaluated using standard Gaussian quadrature. However for i = j the inte-
grands are singular and the integrals are evaluated analytically. We have
Z lj /2 Z lj /2
log ||ri − r(ξ)|| dξ = 2 lim log ξ dξ = lj (log(lj /2) − 1).
−lj /2 ǫ→ 0 ǫ

So (3.5.5a) becomes
li
Gii = − (log(li /2) − 1). (3.5.8)
2π
The integrand of H ^ ij has the inner product (ri − r, n) in the numerator. When
i = j the vector ri − r is in the same element as the normal n and the two are
perpendicular. Consequently (ri − r, n) = 0 and hence
Z
^ 1 (ri − r(χ), n)
Hii = dχ = 0. (3.5.9)
2π Γi ||ri − r(χ)||2

3.6 Numerical integration

To complete the system of equations to be solved in BEM, the integrals in (3.5.6)

have to be evaluated. When the integrands are singular the integrals are eval-
uated analytically as shown in (3.5.8) and (3.5.9). However, the nonsingular
integrals can be approximated using standard Gauss-Legendre quadrature. Af-
ter parametrisation in (3.5.6) the integrals are of the form
Z1
f(η) dη. (3.6.1)
−1

For f(η) nonsingular, the standard Gauss-Legendre quadrature gives

Z1 m
X
f(η) dη ≈ ωi f(ηi ), (3.6.2)
−1 i=1

where ηi are the knots and ωi the weights, see [73] for tables of knots and their
corresponding weights.

Other than resorting to analytical expressions, if f has a weak or logarithmic

singularity over the interval [−1, 1], the integral can also be evaluated using
3.7 Examples 29

special numerical integration schemes for integrals with a weak or logarithmic

singularity. For an integral with a logarithmic singularity the following Gauss-
log approximation is used [15, 73],
Z1 m
X
f(η) log(η) dη ≈ ω̃i f(η̃i ), (3.6.3)
0 i=1

where η̃i are the knots and ω̃i the weights for the quadrature rule with a log-
arithmic singularity. In addition to [15, 73], the weights can also be found
in [43, p. 513]. More on numerical computation of integrals with a logarithmic
singularity can be found in [67].

3.7 Examples

In this thesis we like to use some reference cases. They are given by Exam-
ples 3.7.1 to 3.7.4. For the problem domain, either a disc or a square is used.

Example 3.7.1 (Problem (a): Dirichlet problem, smooth solution)

 2
 ∇ u(r) = 0, r ∈ Ω,
(3.7.1)

u(r) = x2 − y2 , r ∈ ∂Ω.
Taking Ω = {r ∈ R2 : ||r|| ≤ 1.2}, Example 3.7.1 has the continuous solution
shown in Figure 3.5b. Figure 3.5a shows the domain considered: a circular disc
of radius R = 1.2 centred at the origin. Since u(x, y) itself satisfies the Laplace
equation in Ω we can easily compute the analytic solution q(x, y) = n · ∇u(x, y).
The boundary function u(x, y) in polar coordinates centred at the origin is given
by u(R, θ) = R2 cos 2θ and the analytic solution is q(R, θ) = 2R cos 2θ. Figure 3.5b
shows the solution plotted against the polar angle.

Example 3.7.2 (Problem (b): Dirichlet problem, locally active solution)

 2
 ∇ u(r) = 0, r ∈ Ω,
(3.7.2)

u(r) = log(||r − r0 ||), r ∈ ∂Ω,
where r0 is a fixed point outside Ω. Again taking the disc shown in Figure 3.6a
as the domain and r0 = (0.36, 1.8), Example 3.7.2 can be solved analytically.
The continuous solution, shown in Figure 3.6b, is q(r) = (r − r0 ) · n(r)/||r − r0 ||2
where n(r) is the unit outward normal at r. The solution is plotted against the
angle θ around the circle, note the local high activity near θ = 1.4.

Example 3.7.3 (Problem (c): Neumann problem, smooth solution)

 2

 ∇ u(r) = 0, r ∈ Ω,

2(x2 − y2 ) (3.7.3)

 , r ∈ ∂Ω.
 q(r) = p 2
x + y2
30 The boundary element method

1 2

0.5 R=1.2 1

q(x(θ),y(θ))
0 0
Ω
−0.5 −1

−1 −2

−1 −0.5 0 0.5 1 0 2 4 6
0≤ θ ≤ 2π

(a) Disc domain. (b) Solution on circle

Figure 3.5: A disc domain and solution to Example 3.7.1

on the circular boundary of radius R = 1.2.

1
0
0.5 R=1.2
q(x(θ),y(θ))

−0.5
0
Ω
−0.5 −1

−1 −1.5

−1 −0.5 0 0.5 1 0 2 4 6
0≤ θ ≤ 2π

(a) Disc domain. (b) Solution on circle

Figure 3.6: A disc domain and solution to Example 3.7.2

with r0 = (0.36, 1.8) on the circular boundary of radius R =
1.2.
3.7 Examples 31

Take again as domain the disc of radius R = 1.2 shown in Figure 3.7a. The
solution to this problem can be computed analytically. For our circular domain
the solution is easily expressed in polar coordinates centred at the origin and
is given by u(R, θ) = R2 cos 2θ. Figure 3.7b is a plot of this solution against θ.

1 1

0.5 R=1.2 0.5

u(x(θ),y(θ))
0 0
Ω
−0.5
−0.5
−1
−1

−1 −0.5 0 0.5 1 0 2 4 6
0≤ θ ≤ 2π

(a) Disc domain. (b) Solution on circle

Figure 3.7: A disc domain and solution to Example 3.7.3

on the circular boundary of radius R = 1.2.

Example 3.7.4 (Problem (d): Neumann problem, locally active solution)



 ∇2 u(r) = 0, r ∈ Ω,

(3.7.4)

 (r − r0 ) · n(r)
 q(r) = , r ∈ ∂Ω,
||r − r0 ||2

where r0 is a fixed point outside Ω and n(r) is the outward normal at r. The
solution to this problem can be computed analytically. Again for the disc shown
in Figure 3.8a as domain and r0 = (0.36, 1.8), Example 3.7.4 has the continuous
solution shown in Figure 3.8 where θ is the angle around the circle. Note again
the relatively high local activity near θ = 1.4.

For these examples, the analytic expressions for the solutions are known. This
will help us compare exact analytic and numerical solutions and thus compute
global and local errors as we will see in Chapters 4 and 5. Observe in these
examples that the solution to Dirichlet problem 1 is the boundary condition in
Neumann problem 1 and vice-versa. Also the solution to Dirichlet problem 2 is
the boundary condition in Neumann problem 2 and vice-versa.
 1 1

0.5 R=1.2
u(x(θ),y(θ))

0.5
0
Ω
−0.5 0

−1
−0.5
−1 −0.5 0 0.5 1 0 2 4 6
0≤ θ ≤ 2π

(a) Disc domain. (b) Solution on circle

Figure 3.8: A disc domain and solution to Example 3.7.4

with r0 = (0.36, 1.8) on the circular boundary of radius R =
1.2.
“The study of error is not only in the highest degree prophylatic, but
it serves as a stimulating introduction to the study of truth.” – Walter
Lipmann
34 The boundary element method
Chapter 4

Local errors in BEM for

potential problems

4.1 Introduction

The BEM results from a numerical discretisation of a BIE. The subject of errors
in BEM is still a very interesting one and some aspects have not yet been as
explored as they are in other numerical methods like finite element methods
(FEM) and finite difference methods (FDM). Errors in BEM solutions may be
due to discretisation or to inaccuracies in the solver that involves the use of
BEM matrices with high condition numbers. Although a BIE is an exact repre-
sentation for the solution of a BVP, errors will occur in BEM because the BIE
is applied at only a selected set of collocation points.

For a given discretisation, there are several ways to implement BEM because of
the choice in collocation and nodal points and the shape functions. These will
all influence the resulting error in the solution. Amongst the error sources also
is the fact that the choice of shape functions at the boundary elements may
not satisfy the smoothness requirements of the original BIE. In most cases
where error measurement has been performed, like in adaptive refinement,
the main focus has been a guiding measure of the error. In this chapter we
present recent results on an analysis of actual local errors in BEM solutions.
The results presented will not only be helpful in choosing an implementation
strategy but also in guiding adaptive refinement techniques. We focus our
attention to obtaining infinity norm estimates of the error.

Several techniques have been used to measure BEM errors in the area of adap-
tive refinement. In [45] and [37] a review of error estimation and adaptive
36 Local errors in BEM for potential problems

methods can be found. Most of these methods are a posteriori providing an

indirect way of estimating the error. For instance the discretisation error is
estimated by the difference between two solutions obtained using different col-
location points but the same discretisation [22]. Another technique uses the
first (singular integral equation) and the second (hypersingular integral equa-
tion) kind formulations to provide an error estimate [44]. Data from the first
kind of equation is substituted into the second kind to obtain a residual which
is then used to estimate the error. In [29] the authors have used a so called
gradient recovery procedure to develop a local error estimate. The error func-
tion is generated from differences between smoothed and non-smoothed rates
of change of boundary variables in a local tangent direction. Some authors
have used a posteriori error estimation in FEM as a guiding tool to develop
error estimates for BEM [8]. In [66] residual based estimators are used in the
case of Galerkin solutions. An error estimator is also obtained by solving an
error equation as accurately as possible, see [66]. Unfortunately usually such
techniques are restricted to Galerkin BEM and literature is scarce for colloca-
tion BEM formulations and BEM in general as compared to FEM [8, 17, 29, 45].
In our case we would like to start from the basics of the BIE discretisation to
develop error analyses for collocation BEM for potential problems. Though the
ideas could easily be adapted to 3D, 2D problems and the Laplace equation in
particular are discussed.

When a value xj is approximated by xLj using BEM, a global error ej , which is

the difference between the exact solution xj and its BEM approximation xLj , is
committed. That is,

e := x − xL . (4.1.1)

For instance for a Dirichlet problem the BEM solution is a vector of values of
the normal derivative qL and the global error is given by

e := q − qL . (4.1.2)

To assess this we first have to investigate the local discretisation error. Local
discretisation errors are usually defined as the residual that remains if the
exact solution is substituted into the discretised equation. In operator notation,
suppose we have an operator A and a discretisation of A, say AL , both defined
on the same space. Let the continuous solution u satisfy

Au = b. (4.1.3)

Let the discretised problem be

AL uL = bL , (4.1.4)

where uL is the discrete approximation of u. Then the local error dL is defined

by

AL u = bL + dL , (4.1.5)
4.2 A survey of error sources 37

where u is the projection of the full exact solution on the grid. As a conse-
quence, the global error satisfies

AL e = dL . (4.1.6)

Alternatively we can define the local error through

AuL = b + d, (4.1.7)

that is, by substituting the approximation into the exact problem. This gives

Ae = d. (4.1.8)

Both (4.1.6) and (4.1.8) can be useful to estimate global error. However, inver-
sion in order to get the global error in (4.1.8) is hampered by the fact that we
need information of the solutions everywhere. Nevertheless (4.1.7) is appealing
as it is based on the original operator.

In this chapter local errors are discussed. Global errors will be discussed in
Chapter 5. Section 4.2 presents a brief survey of error sources in BEM and
Section 4.2.1 presents a discussion on local errors due to interpolation in BEM
and a theory for the expected convergence rates of the local error for both con-
stant and linear elements. It is shown here that the local errors for both formu-
lations are of second order. Numerical results for both Dirichlet and Neumann
examples are presented to verify the theory presented on local errors.

For the theories on local errors we assume a uniform grid. However it might
not always be necessary to have uniform grids. For problems with localised
regions of high activity, a composite coarse-fine grid is useful. In Section 4.7
the process of equidistribution in BEM for problems whose solutions have small
regions of high activity is discussed.

4.2 A survey of error sources

In a BEM implementation, the first step is to discretise the boundary of the do-
main. Then in the application of the integral equation we have the freedom to
choose collocation points and the nodal points. This suggests that for a given
discretisation there is more than one way one can choose to implement the
BEM, each of which will have its own advantages and disadvantages. Through-
out this thesis we consider discretisations using rectilinear elements. This re-
sults into the numerical boundary being a polygon, see for instance Figure 4.1
in the case of a circular boundary.

Besides choosing the position of collocation and nodal points on the numerical
boundary, the way boundary conditions are treated is also important. In the
traditional constant elements for example, the nodes are the midpoints of the
38 Local errors in BEM for potential problems

∂Ωj Γj

Figure 4.1: Discretisation of a circle into a six elements polygon.

elements and are also the collocation points. For the discretisation shown in
Figure 4.1, the numerical boundary does not coincide with the physical bound-
ary and so we may talk of exact points ri on ∂Ω and their numerical represen-
tations r̄i on Γ as shown in Figure 4.2. One may thus choose collocation points

r̄i on Γi

ri on ∂Ωi

Figure 4.2: Numerical representation of boundary points; ri the exact point

and r̄i its numerical representation.

on the physical boundary or on the numerical boundary. Likewise one may

choose to compute the integrals of the known functions on the exact physical
boundary or on the numerical boundary. In Table 4.1 we give a summary of
the possibilities one may consider in the case of a Dirichlet problem when im-
plementing constant elements. Similar choices arise in the case of a Neumann
or mixed problem.

The different cases arise from the location of collocation points on either the nu-
merical boundary or the physical boundary, treatment of boundary conditions
and the choice between the numerical and physical boundary for the integrals
involved in the formation of the matrix A and vector b in (3.2.12).

Consider for instance, cases (1) and (5) in Table 4.1. In case (1) the collocation
points are on the numerical boundary Γj , the boundary function u is assumed
constant on Γj taking on the midpoint value u(rj ), the right hand side b and the
matrix A are formed by integrating on Γj . This is the traditional implementation
4.2 A survey of error sources 39

position of boundary condition form b by form A by

collocation points on the element integrating integrating
either on ∂Ωj or Γj either on ∂Ωj or Γj
Γj (1)
Γj
∂Ωj (2)
u(rj )
Γj (3)
∂Ωj
∂Ωj (4)
on Γi
Γj (5)
Γj
∂Ωj (6)
u(r)
Γj (7)
∂Ωj
∂Ωj (8)
- - (9)
- (10)
u(rj )
- (11)
∂Ωj
∂Ωj (12)
on ∂Ωi
- - (13)
- (14)
u(r)
- (15)
∂Ωj
∂Ωj (16)

Table 4.1: Summary of possibilities on how one could implement constant

elements BEM.

of constant elements. That is,

X N Z
1 ∂v
bi := − ui − uj (r̄i ; r(χ)) dχ, (4.2.1)
2 Γj ∂n
j=1
Z
Aij := − v(r̄i ; r(χ)) dχ. (4.2.2)
Γj

In case (5) the collocation points are on the numerical boundary, the boundary
function u(r) is not assumed constant on Γj but used directly as given in the
formation of b, the right hand side b and the matrix A are formed by integrating
on Γj . That is

XN Z
1 ∂v
bi := − u(r̄i ) − u(r(χ)) (r̄i ; r(χ)) dχ, (4.2.3)
2 ∂n
j=1 Γj
Z
Aij := − v(r̄i ; r(χ)) dχ. (4.2.4)
Γj

The other cases can also be explained by following their respective trends in
the table. Cases (9) to (16) are obtained by choosing collocation on the physical
boundary. The dashes in the table are for cases that are not feasible in the
sense that once we have chosen to collocate on the physical boundary then we
should implement the integral equation on the physical boundary by carrying
out the integrations in the formation of b and A on the same boundary.

In Table 4.2 results for some of the cases in Table 4.1 are presented. In par-
ticular, in Table 4.2a is a comparison of cases (1) and (5) and in Table 4.2b a
40 Local errors in BEM for potential problems

comparison of cases (12) and (16).

||d||∞ ||d||∞
N case (1) case (5) N case (12) case (16)
5 1.20 1.02E-01 5 1.76E-02 1.81E-01
15 1.66E-01 3.87E-02 15 3.53E-02 2.86E-02
45 1.90E-02 5.12E-03 45 4.99E-03 4.74E-03
135 2.13E-03 5.98E-04 135 5.93E-04 5.84E-04
405 2.38E-04 6.75E-05 405 6.73E-05 6.70E-05
1215 2.65E-05 7.57E-06 1215 7.54E-06 7.52E-06
(a) (b)

Table 4.2: Local errors in cases (1) and (5) in (a) and cases (12) and (16) in (b)
when the problem in Example 3.7.1, which is a Dirichlet problem, is solved on
a disc domain. The local error d is defined in (4.1.5).

||d||∞ ||d||∞
N case (1) case (5) N case (12) case (16)
5 3.11E-01 2.63E-02 5 5.34E-03 5.49E-02
15 4.81E-02 1.12E-02 15 1.04E-02 8.40E-03
45 5.67E-03 1.52E-03 45 1.49E-03 1.42E-03
135 6.39E-04 1.79E-04 135 1.78E-04 1.75E-04
405 7.13E-05 2.02E-05 405 2.02E-05 2.01E-05
1215 7.93E-06 2.26E-06 1215 2.26E-06 2.26E-06
(a) (b)

Table 4.3: Local errors in cases (1) and (5) in (a) and cases (12) and (16) in (b)
when we the problem in Example 3.7.1, which is a Dirichlet problem, is solved
on a disc domain. The local error d is defined in (4.1.5).

It is expected that the results of case (5) are better than those of case (1), since
case (5) should capture the variation of the boundary function better. For the
same reasons, the results of case (16) are expected to be better than those
of case (12). Since in cases (12) and (16) the numerical boundary coincides
with the physical boundary, the corresponding results should be better than
those of cases (1) and (5). The results in Table 4.2 agree with this observation.
Although the difference in results might seem not that much, we note that
the problems in real life will usually involve more complicated geometries for
the physical boundary and more wild boundary conditions. This will make the
difference between the choices in Table 4.1 more important than it looks in the
above example.

Although some work has already been done in estimating errors in BEM, there
is not yet an error analysis that traces down all the error sources above. In [28,
4.2 A survey of error sources 41

p. 141] the authors have remarked that: ‘‘any error analysis which seeks
to trace the accumulation of error as it arises from the approximation of the
physical boundary by a numerical boundary, approximation of the boundary
quantities by interpolation and approximation of the integrals by quadrature
rules is likely to be very complicated, if indeed it is possible at all.” In the next
sections we concentrate on analysing the error that is due to the approximation
of the boundary quantities by interpolation in constant and linear elements.

4.2.1 Defining local errors

Section 4.2 above introduced different ways in which BEM can be implemented.
These induce sources of error in BEM for potential problems: approximation of
the physical boundary by a numerical boundary, approximation of the bound-
ary quantities by interpolation, and approximation of the integrals by quadra-
ture rules. Thus, the local error in BEM will be an accumulation of the following
errors from each of the above sources respectively:

(i) Boundary discretisation error: This error is not there if Γj and ∂Ωj coincide.

(ii) Interpolation error: This is the most important one. The unknown func-
tions are assumed to vary as certain interpolation polynomials, the so
called shape functions fu (r) and fq (r) for u(r) and q(r) respectively. For
instance for constant elements, assuming the functions u(r) and q(r) are
assumed piecewise constant on each Γj .

(iii) Quadrature error: This error is also very important since all the integrals
in (3.2.5) are usually evaluated using numerical quadrature. However, it
can be minimised by choosing suitable high order quadrature rules.

Let us again consider a general point s at the boundary where the integral
equation is applied. Then constant elements assumes:
Z Z
∂v . ∂v
u(r(χ)) (s; r(χ)) dχ = u(r(χ)) (s; r(χ)) dχ, (4.2.5a)
∂Ωj ∂n Γj ∂n

Z Z Z
∂v . ∂v ∂v
u(r(χ)) (s; r(χ)) dχ = fu (r(χ)) (s; r(χ)) dχ = uj (s; r(χ)) dχ,
Γj ∂n Γj ∂n Γj ∂n
(4.2.5b)

Z
∂v ∂v
 
.
uj (s; r(χ)) dχ = uj Qj
 (s; r(χ))
. (4.2.5c)
Γj ∂n ∂n
42 Local errors in BEM for potential problems

 
∂v
Here Qj ∂n (s; r(χ)) denotes a quadrature approximation of the integral. Like-
wise for the single layer integrals we have
Z Z
.
q(r(χ))v(s; r(χ)) dχ = q(r(χ))v(s; r(χ)) dχ, (4.2.6a)
∂Ωj Γj

Z Z Z
.
q(r(χ))v(s; r(χ)) dχ = fq (r(χ))v(s; r(χ)) dχ = qj v(s; r(χ)) dχ, (4.2.6b)
Γj Γj Γj

Z  
.
qj v(s; r(χ)) dχ = qj Qjv(s; r(χ)) . (4.2.6c)
Γj

If Dirichlet boundary conditions are given on Γj , then uj = u(rj ) in (4.2.5b) and

likewise qj = q(rj ) in (4.2.6b), in the case of Neumann boundary conditions.

Note that the expressions in (4.2.5) and (4.2.6) will be similar in linear elements
except that now the functions fu and fq will be order one polynomials as in-
troduced in (3.3.1). The local error per element is an accumulation of the local
errors in (4.2.5) and/or (4.2.6), depending on the boundary conditions given
and how they are treated. So we define the errors on the j-th element for a
source node s as:
Z
∂v ∂v
 
u
dj (s) := u(r(χ)) (s; r(χ)) dχ − uj Qj
 (s; r(χ))
, (4.2.7a)
∂Ωj ∂n ∂n

Z  
dq
j (s) := q(r(χ))v(s; r(χ)) dχ − qj Qjv(s; r(χ)) . (4.2.7b)
∂Ωj

q
The errors duj and dj in (4.2.7) are what we will call the sublocal errors. Then
the local error at s due to contributions from all the elements is defined as
N
X N
X
d(s) := − du
j (s) + dq
j (s). (4.2.8)
j=1 j=1

When s = ri in the discretisation introduced in Section 3.2, then the error d(ri )
given by (4.2.8) is the local error in the i-th equation. That is if we were to write
the exact equation of (3.2.1) using BEM approximations uj and qj we would
obtain
N
X Z N
X Z
∂v
c(ri )u(ri ) + uj (ri ; r(χ)) dχ = qj v(ri ; r(χ)) dχ + d(ri ). (4.2.9)
∂Ωj ∂n
j=1 ∂Ωj j=1

We will mainly be interested in the error due to interpolation and will minimise
the error due to quadrature by using high order adaptive quadrature rules.
It is also important to note that in some problems the physical boundary is
4.3 Dirichlet problems 43

quite regular or indeed coincides exactly with the numerical boundary hence
eliminating one inherent source of error.

The treatment of boundary conditions also has a significant contribution to the

local error we are assessing. Consider the case of a Dirichlet problem. Although
in this case the continuous function u is known, usually u is still assumed to be
constant like in (4.2.5b). Although this leaves the method cheaper and easier to
implement, it introduces an error in the right hand side that could have rather
been avoided. The contribution du ij in (4.2.7a) to the local error can be skipped
by avoiding this assumption on u and thus have an ”exact right hand side”.

In what follows we assess further local errors for both the constant and linear
element formulations and in Chapter 5 global errors are assessed. Out of the
possibilities presented in Table 4.1, we concentrate on the traditional constant
elements. However, the constant u assumption is avoided on elements where
the continuous function is known. Thus, possibility (5) in the table is examined.
We use here a better in depth approach than that used in [33] by using the local
arc coordinate and sublocal error.

4.3 Dirichlet problems

Consider the case of Dirichlet boundary conditions. This case results in the
first kind integral equation

Ks (s; r)q(r) = f(r) (4.3.1)

where the right hand side

1
f(r) = I + Kd (s; r)u(r) (4.3.2)
2
is known since u is given. The solution of (4.3.1) in BEM therefore involves
estimating integrals of the form
Z
1
I := v(s; r(χ))q(r(χ)) dχ. (4.3.3)
2π ∂Ωj

The following two sections discuss local errors due to a BEM solution of (4.3.1)
using constant elements and linear elements.

4.3.1 Constant elements

In this section attention is focussed on constant elements and assess the in-
terpolation error. As one may expect the error will depend on the size of the
44 Local errors in BEM for potential problems

elements where the interpolation is done. We will derive a relation between the
local error and the grid size.

Suppose the boundary Γ is discretised into elements Γj each of size lj . For

an element with Dirichlet boundary conditions, if we perform exact integration
for u, then the only contribution to the sublocal error (4.2.7) is that due to
interpolation of q. Let us introduce d0 j (s), the error due to interpolation on the
j-th element when the source node is s, which is defined as:
Z Z
d0 j (s) := q(r(χ))v(s; r(χ)) dχ − qj v(s; r(χ)) dχ. (4.3.4)
Γj Γj

The subscript 0 is used to denote constant elements where a zeroth order poly-
nomial is used for interpolation. The error d0 j (s) is the sublocal error due to
interpolation and the local error d0 (s) due to interpolation is the sum of all the
sublocal errors. That is,
X
d0 (s) = d0j (s). (4.3.5)
j

The following lemma is based on a discretisation of the boundary into a polygon,

that is, the element Γj is a straight line of length lj say.

Lemma 4.3.1 The truncation error in (4.3.4) is third order in grid size lj and is
given by


 1 ∂q 1 ∂rj 1 ∂2 q

 − (χj ) (χj )l3j − (rj )l3j log (||s − rj ||2 ) + O(l4j ), s ∈
/ ∂Ωj ,
 24π ∂χ rj (χj ) ∂χ 48π ∂χ2
d0j (s) =



 1 ∂2 q
 − (ri )l3i (1/3 − log(li /2)) + O(l4i ), s ∈ ∂Ωj , j = i.
48π ∂χ2
(4.3.6)

where χ is an arc length coordinate in ∂Ω and ∂q/∂χ is the tangential derivative

of q.

Proof. Let χj be the midpoint of ∂Ωj , see Figure 4.3.

Consider the “exact” integral in (4.3.4), that is,

Z
I := v(s; r(χ))q(r(χ)) dχ. (4.3.7)
Γj

This integral can be rewritten in terms of χ. Suppose we have a Taylor series

expansion of q(χ) about χj within the element Γj , that is,

∂q 1 ∂2 q 1 ∂3 q
q(χ) = q(χj )+ (χj )(χ−χj )+ 2
(χj )(χ−χj )2 + (χj )(χ−χj )3 +· · · , (4.3.8)
∂χ 2 ∂χ 6 ∂χ3
4.3 Dirichlet problems 45

rj (χ) χj
∂Ωj
s
χ

Figure 4.3: Distance rj (χ) from s to an el-

ement Γj of length lj on a boundary curve.

∂q ′′ ∂2 q
For simplicity we may write q ′ (χj ) instead of , q (χj ) instead of , etcetera,
∂χ ∂χ2
′ ′′
so that q (χj ), q (χj ), and so on, on ∂Ωj are actually the tangential derivatives
of q at χj . In general we shall use the notation
r(χ) := ||s − r(χ)|| (4.3.9)
for a fixed point s and any point χ ∈ ∂Ω. We also define the distance rj (χ) from
a point s to ∂Ωj as
rj (χ) := ||s − r(χ)||, χ ∈ ∂Ωj . (4.3.10)
If r(χ) is a point on ∂Ωj , we write
rj (χj ) = ||s − r(χj )|| = ||s − rj || since r(χj ) = rj , (4.3.11)
see Figure 4.3. Using (4.3.8) in (4.3.7) and substituting the expression for v(s; r)
gives
Z lj /2 Z lj /2
∂q
− 2πI = log[rj (χ)]q(χj ) dχ + log[rj (χ)] (χj )(χ − χj ) dχ+
−lj /2 −lj /2 ∂χ
Z lj /2
1 ∂2 q
log[rj (χ)] (χj )(χ − χj )2 dχ + · · · . (4.3.12)
−lj /2 2 ∂χ2

In constant elements BEM only the first term of (4.3.12) is used, that is,
Z
. 1 lj /2
I=− log[rj (χ)]q(χj ) dχ. (4.3.13)
2π −lj /2
This results in a truncation error given by
Z lj /2
1 ∂q
d0j (s) = − log[rj (χ)] (χj )(χ − χj ) dχ
2π −lj /2 ∂χ
Z lj /2
1 1 ∂2 q
− log[rj (χ)] (χj )(χ − χj )2 dχ + · · · . (4.3.14)
2π −lj /2 2 ∂χ2
46 Local errors in BEM for potential problems

Consider the first term of (4.3.14). That is,

Z
1 lj /2 ∂q
T1 := − (χj ) log[rj (χ)](χ − χj ) dχ. (4.3.15)
2π −lj /2 ∂χ

For s ∈ ∂Ωj , j = i, then s = ri , r(χ) = |χ| and we see that the cauchy principal
value integral in (4.3.15) evaluates to zero. Otherwise, for χ − χj small, or
equivalently lj small, the distance rj (χ) can be expanded about the point χj as,

rj (χ) = rj (χj ) + rj′ (χj )(χ − χj ) + O((χ − χj )2 ). (4.3.16)

Since rj (χj ) 6= 0 we can write (4.3.16) as

 
.  + 1 ∂rj (χj )(χ − χj )
rj (χ) = rj (χj ) 1

 + O((χ − χj )2 ).
 (4.3.17)
rj (χj ) ∂χ

Substituting (4.3.17) in (4.3.15) gives

Z lj /2
. 1 ∂q
T1 = − (χj ) log(rj (χj ))(χ − χj ) dχ
2π −lj /2 ∂χ
Z
1 lj /2 ∂q
 
1 ∂rj
− (χj ) log 1 + (χj )(χ − χj )
 (χ − χj ) dχ (4.3.18)

 
2π −lj /2 ∂χ rj (χj ) ∂χ

Consider ξ a local coordinate in ∂Ωj such that, see Figure 4.4a,

χ = χj + ξ, −lj /2 ≤ ξ ≤ lj /2. (4.3.19)

Then the first integral on the right of (4.3.18) becomes

ξ r(ξ)
rj (ξ)
−ξ
χj
∂Ωj s
s ξ
(ξs , ηs ) η
χ
(a) Local coordinate ξ in ∂Ωj .
(b) Local coordinates ξ and η.

Figure 4.4: Local coordinates in (ξ, η) at the boundary.

Z lj /2
∂q
(χj ) log(rj (χj ))ξ dξ = 0. (4.3.20)
−lj /2 ∂χ

Define ξ as a local coordinate in ∂Ωj and η the inward normal coordinate, see
the illustration in Figure 4.4. The distance rj (ξ) introduced in (4.3.10) is then
4.3 Dirichlet problems 47

given by

ξ − ξs
rj (ξ) = [(ξ − ξs )2 + η2j ]1/2 so rj′ (ξ) = .
[(ξ − ξs )2 + η2s ]1/2

Then we have, since ξj = 0,

rj′ (ξj ) −ξs ξ

(ξ − ξj ) = 2 .
rj (ξj ) ξs + η2s

Note that ξ ∈ ∂Ωj so that |ξ| ≤ lj /2. Hence, since η2s ≥ 0 we find

rj′ (ξj ) ξs ξ
(ξ − ξj ) = 2 < 1. (4.3.21)
rj (ξj ) ξs + η2s

A natural logarithm series expansion can therefore be used for the logarithm
term in the second integral on the right hand side of (4.3.18). Therefore, the
second integral, with the help of natural logarithm series for log(1 + x), becomes

ljZ/2  
1 ∂q 1 ∂rj
− (χj ) log 
1 + (χj )(χ − χj )
 (χ − χj ) dχ =
 
2π ∂χ rj (χj ) ∂χ
−lj /2
ljZ/2
1 ∂q 1 ∂rj 1 ∂q 1 ∂rj
− (χj ) (χj )ξ2 dξ = − (χj ) (χj )l3j .
2π ∂χ rj (χj ) ∂χ 24π ∂χ rj (χj ) ∂χ
−lj /2

Hence, we have

. 1 ∂q 1 ∂rj
T1 = − (χj ) (χj )l3j + O(l4j ). (4.3.22)
24π ∂χ rj (χj ) ∂χ

Now consider the second order term of (4.3.14),

Z lj /2
1 1 ∂2 q
T2 := − (χj ) log[rj (χ)](χ − χj )2 dχ. (4.3.23)
2π −lj /2 2 ∂χ2

Recall that since rj (χj ) > 0, the distance rj (χ) introduced in (4.3.10) can be
written as
 
. 1 ∂rj
rj (χ) = rj (χj ) 
1 + (χj )(χ − χj )
. (4.3.24)
 
rj (χj ) ∂χ

Using the local coordinate notation introduced in Figure 4.4 and the result (4.3.21)
, we can again use natural logarithm series for the term in parentheses in (4.3.24)
to obtain
rj′ (ξj )
log(rj (ξ)) = log(rj (ξj )) + (ξ − ξj ) + O((ξ − ξj )2 ). (4.3.25)
rj (ξj )
48 Local errors in BEM for potential problems

Using (4.3.25) and (4.3.21) in (4.3.23) we have

Z lj /2
1 ∂2 q
 
2 −ξs ξ 
T2 = − (ξ j ) (ξ − ξ j ) (ξ
log(rj j
 )) +  dξ.

4π ∂χ2 −lj /2 ξ2s + η2s
Finally, if the original ∂Ωj is not a straight line but a curve then it is easy to
see that the midpoint of the approximate Γj (a straight line) is only O(l2j ) away
from rj and so are the derivatives of q. Hence we may conclude that q ′′ (ξj ) is
∂2 q
equal to (rj ) up to O(l2j ) and conclude that, for a general point s and using
∂χ2
the notation in (4.3.11),
. 1 ∂2 q
T2 = − (rj )l3j log(||s − rj ||) + O(l4j ), (4.3.26)
48π ∂χ2
which is third order in grid size lj .

When s ∈ ∂Ωj , j = i, r = |ξ| and this integral becomes

Z
. 1 li /2 ∂2 q
T2 = − (ri ) log[ri (ξ)](ξ − ξi )2 dξ, (4.3.27)
4π −li /2 ∂χ2
Z li /2 Z li /2
1 ∂2 q 2 1 ∂2 q
=− (s) log[|ξ|]ξ dξ = − (s) log(ξ2 )ξ2 dξ,
4π ∂χ2 −li /2 8π ∂χ2 −li /2
Z lj /2
1 ∂2 q
=− (s) lim log(ξ2 )ξ2 dξ,
4π ∂χ2 ǫ→ 0 ǫ

1 ∂2 q
=− (s)l3i (log(li /2) − 1/3). (4.3.28)
48π ∂χ2
Putting (4.3.22), (4.3.26), and (4.3.28) in (4.3.14) gives the result (4.3.6). 

Apparently, the sublocal error is third order in grid size. Moreover for a mild
∂2 q
and a given grid size, the most important factor in the sublocal error is
∂χ2
log[rj (χj )], the log of the distance ||s − rj ||. The sublocal error is large close to the
source node s and decays logarithmically away from this point. This way, the
most important contributions to the local error d0 (s) will be those from elements
close to s. Since the local error is a sum of the sublocal errors, Lemma 4.3.1
puts us in a position to state the following theorems for the local error:

Theorem 4.3.1 (Constant elements local error) The local error (4.3.5) in con-
stant elements BEM applied to a boundary curve ∂Ω with midpoint collocation on
a grid of size lj , j = 1, 2, . . . , N, where N is the number of elements used, is given
by

1 X ∂q 1 ∂rj 1 X 3 ∂2 q
d0 (s) = − (χj ) (χj )l3j − lj 2 (rj ) log(||s − rj ||)
24π ∂χ rj (χj ) ∂χ 48π ∂χ
j ∈∂Ωj
s/
2
1 ∂ q
− (s)l3i (log(li /2) − 1/3) + O(l4j ). (4.3.29)
48π ∂χ2
4.3 Dirichlet problems 49

Proof. The local error over ∂Ω is the sum of all the sublocal errors (4.3.6). So
we have

1 X ∂q 1 ∂rj 1 X 3 ∂2 q
d0 (s) = − (χj ) (χj )l3j − lj 2 (rj ) log(||s − rj ||)
24π ∂χ rj (χj ) ∂χ 48π ∂χ
∈∂Ωj
s/ ∈∂Ωj
s/
2
1 ∂ q
− (s)l3i (log(li /2) − 1/3), (4.3.30)
48π ∂χ2

which is (4.3.29). 

Corollary 4.3.1 The local error (4.3.29) is second order in grid size and is given
by
Z
. 1 1
d0 (s) = D dχ. (4.3.31)
48π log(||s − r(χ)||)
∂Ω

where D is second order in grid size.

Proof. Consider the contribution from the first order derivative term in (4.3.29).
Let
1 X ∂q 1 ∂rj
S1 := − (χj ) (χj )l3j . (4.3.32)
24π ∂χ rj (χj ) ∂χ
j

∂q ∂q
Suppose l2j (χj ) is constant, that is, l2j (χj ) = C over ∂Ω for some constant
∂χ ∂χ
C. Further, if Ω is convex, then there is an extremal point ^s ∈ ∂Ω such that
^s − s is in the direction of the normal, see Figure 4.5. Then the sum S1 can be
computed as follows:

1 X ∂q 1 ∂rj
S1 = − (χj ) (χj )l3j
24π ∂χ rj (χj ) ∂χ
j
 
 χ(^
Z s) Zs)
χ(^ 
X 1 ∂rj  1 ∂r 1 ∂r 
. .
=D (χj )lj = D (χ) dχ − (χ) dχ = 0,
rj (χj ) ∂χ 
 r(χ) ∂χ r(χ) ∂χ 

j
li /2 li /2
(4.3.33)

∂q
where D = −C/(24π). So if l2j (χj ) = C, then the local error (4.3.30) becomes
∂χ

. 1 X 3 ∂2 q
d0 (s) = − lj 2 (rj ) log(||s − rj ||)
48π ∂χ
∈∂Ωj
s/

1 ∂2 q
− (s)l3i (log(li /2) − 1/3). (4.3.34)
48π ∂χ2
50 Local errors in BEM for potential problems

^s
n

^s − s
s

dχ

Figure 4.5: Extremal point s^ corresponding to

point s on a convex boundary.

Let us assume Ω has a radius less than one half, that is, ρ(Ω) < 1/2. This
is a nonrestrictive assumption since if not, we can introduce a scaling α such
that r → αr for all r ∈ Ω̄ and α chosen small enough. Then we will have
∂2 q
log(||s − r||) < 0 for all s, r ∈ ∂Ω. Suppose in addition that l2j 2 (χj ) = D over ∂Ω
∂χ
for some constant D, then we have

. 1 X
d0 (s) = − D lj log(||s − r(χj )||)
48π
∈∂Ωj
s/

1 ∂2 q
− (s)l3i (log(li /2) − 1/3). (4.3.35)
48π ∂χ2
Using Riemann integral interpretation for the sum and since log(||s − r||) < 0, we
have
Z
. 1
d0 (s) = − D lim log(||s − r(χ)||) dχ
48π li → 0
∂Ω−li

1 ∂2 q
− (s) lim l3i (log(li /2) − 1/3), (4.3.36)
48π ∂χ2 li → 0

which yields
Z
. 1 1
d0 (s) = D dχ. (4.3.37)
48π log(||s − r(χ)||)
∂Ω

∂2 q 2
2∂ q
If l2j (χ j ) is not constant over ∂Ω, then we take D as the average of lj (χj )
∂χ2 ∂χ2
∂2 q ∂2 q
over ∂Ω, that is, some value between max(l2j 2 (χj )) and min(l2j 2 (χj )).
∂χ ∂χ
4.3 Dirichlet problems 51

∂q ∂q
If l2j (χj ) is not constant over ∂Ω, then let D ′ be the average of l2j (χj ) over
∂χ ∂χ
∂Ω. So the sum S1 in (4.3.32) becomes, using the same argument as in (4.3.33),

1 X ∂q 1 ∂rj
S1 = − (χj ) (χj )l3j
24π ∂χ rj (χj ) ∂χ
j
X Z
. ′ 1 ∂rj . 1 ∂r
=D (χj )lj = D ′ (χ) dχ = 0. (4.3.38)
rj (χj ) ∂χ r(χ) ∂χ
j ∂Ω

The contour integral in (4.3.37) is rather complicated and of course depends on

the type of geometry under consideration. The result of Theorem (4.3.1) can be
used to derive error expressions for some regular geometries. In particular we
consider here the cases of a circle and a square. Note that we do not consider a
circle of radius unity. This is because we have existence and uniqueness of the
Dirichlet problem if and only if the logarithmic capacity of ∂Ω, which is equal to
its radius in the case of a circle, is not equal to unity. Then the BEM equation
leads to a singular system, see [11, 12, 16, 68, 76].

Corollary 4.3.2 (Constant elements local error: Circle) For ∂Ω a circle of ra-
dius R, the local error (4.3.5) is second order in grid size l and is given by

1
d0 (s) = D(2πR) log(1/R) + O(l4 ) (4.3.39)
48π

where D is second order in grid size and is as defined in Corollary 4.3.1.

Proof. Let us express the distance ||s − rj || in terms of the angle α and radius
R as shown in Figure 4.6. The contour integral over ∂Ω is now

2R

r
α

Figure 4.6: Parameterisation of a

circle for evaluating the integral
in (4.3.37).
52 Local errors in BEM for potential problems

Z   Z Zπ
1
log dχ = − log(r(α)) dα = − log(2R sin α)(2Rdα),
∂Ω ||s − r(χ)|| ∂Ω 0
Zπ
= −2πR log(2R) − 2R log(sin α) dα
0
= (2πR) log(1/R). (4.3.40)

So we obtain
1
d0 (s) = D(2πR) log(1/R) + O(l4 ), (4.3.41)
48π
where 2πR is the circumference of the circle. 

Corollary 4.3.3 (Constant elements local error: Square) For ∂Ω a square of

length L, the local error (4.3.5) is second order in grid size l and is given by

1
d0 (s) = D(4L)(1 − log(21/4 L) − π/8) + O(l4 ), (4.3.42)
48π
where D is second order in grid size and is as defined in Corollary 4.3.1.

Proof. Likewise let us start from the result of Theorem 4.3.1. We need to
compute the contour integral in (4.3.37). Now,

Γ3 L

Γ4 Γ2

ri
χ0 Γ1
L

Figure 4.7: A square boundary Γ

with source point s0 = ri at the ori-
gin.

Z   Z
1
log dχ = − log(||s − r(χ)||) dχ
∂Ω ||s − r(χ)|| ∂Ω
4 Z
X
=− log(||s − r(χ)||) dχ,
i=1 Γi

= −4L(log(21/4 L) − 1 + π/8). (4.3.43)

4.3 Dirichlet problems 53

Thus we obtain

1 ∂2 q
d0 (s) = (η)l2 (4L)(1 − log(21/4 L) − π/8) + O(l4 ), (4.3.44)
48π ∂χ2

where 4L is the perimeter of the square. 

The results above show that indeed the local error in constant elements BEM is
second order in grid size and the numerical computations in Section 4.6 agree
with these findings.

4.3.2 Linear elements

In this section we develop similar estimates for the local error due to interpola-
tion for linear elements.

Again consider an element with Dirichlet boundary conditions and assume that
all integrals involving u are computed exactly since u(r) is known. Then the
only contribution to the interpolation error is that from interpolation of q(r).
Let us introduce d1j (s), the error due to interpolation on the j-th element when
the source node is s, which is defined as:
Z Z
d1j (s) := q(r(χ))v(s; r(χ)) dχ − fq (r(χ))v(s; r(χ)) dχ (4.3.45)
Γj Γj

where fq (r) is an order one polynomial. The subscript 1 is used to denote linear
elements where an order one polynomial is used for interpolation. The error
d1 j (s) is the sublocal error due to interpolation in linear elements and the local
error d1 (s) due to interpolation for a node s is the sum of all these sublocal
errors, that is,
X
d1 (s) = d1j (s). (4.3.46)
j

Lemma 4.3.2 (Linear elements sublocal error) Let lj be the grid size of ele-
ment Γj . The sublocal error (4.3.46) is third order in lj and is given by


 1 ∂2 q
 (ζ )l3 log (||s − rj ||) + O(l4j ), ζj ∈ Γj , s ∈
/ Γj ,
 24π ∂χ2 j j

d1j (s) = (4.3.47)



 1 ∂2 q
 (ζj )l3j (log(lj ) − 5/6) + O(l4j ), ζj ∈ Γj , s ∈ Γj .
24π ∂χ2

∂2 q
where (ζj ) is the tangential double derivative of q at ζj .
∂χ2
54 Local errors in BEM for potential problems

Proof. Consider the error due to interpolation by a linear polynomial fq (r).

The interpolation error is given by, see [25, p. 324],

1 ∂2 q
q(ξ) − fq (ξ) = (ζ)(ξ − ξ0 )(ξ − ξ1 ), ζ ∈ (ξ0 , ξ1 ), (4.3.48)
2 ∂χ2
2
where ξ0 and ξ1 are the points of interpolation and ∂∂χq2 (ζ) is the tangential
double derivative of q at ζ. Since we are considering continuous linear elements
in which we interpolate at the end points of ∂Ωj , we have ξ0 =: ξleft = −lj /2 and
ξ1 =: ξright = lj /2 where ξleft and ξright are the left and right boundary points of
the element respectively. So the local truncation error is given by
Z
1
d1 j (s) = − (q(r) − fq (r(χ))) log(||s − r(ξ)||) dξ
2π ∂Ωj
Z lj /2
. 1 ∂2 q
=− (ζ) log(||s − r(ξ)||)(ξ + lj /2)(ξ − lj /2) dξ. (4.3.49)
4π ∂χ2 −lj /2

Of course, in (4.3.48), the point ζ is unknown. So at best (4.3.49) will give us a

bound on the error if the bound for ∂2 q/∂χ2 on Γj is known. Let
Z lj /2
I := log(||s − r(ξ)||)(ξ + lj /2)(ξ − lj /2) dξ.
−lj /2

For source points outside ∂Ωj the expansion (4.3.16) is used for r(ξ) and (4.3.25)
for log(||s − r(ξ)||) to obtain
Z lj /2  
I= log(||s − rj ||)(ξ + lj /2)(ξ − lj /2) + O(ξ − ξj )3  dξ
−lj /2

l3j
=− log(||s − rj ||) + O(l4j ). (4.3.50)
6
As remarked in Lemma 4.3.1, if the original ∂Ωj is not a straight line but a
curve then it is easy to see that the midpoint of the approximate Γj (a straight
line) is only O(l2j ) away from rj and so are the derivatives of q. Hence we may
∂2 q
conclude that ∂2 q(ξj )/∂χ2 is equal to (rj ) up to O(l2j ). So from (4.3.49) we
∂χ2
get

1 ∂2 q
d1j (s) = (ζj )l3j log(||s − rj ||) + O(l4j ) (4.3.51)
24π ∂χ2

which is third order in grid size. For s ∈ Γj ; r = ξ, ξleft = 0, ξright = lj , so we

have
Z lj
l3j
I= ξ(ξ − lj ) log(ξ) dξ = − (log(lj ) − 5/6). (4.3.52)
0 6
4.3 Dirichlet problems 55

Thus


 1 ∂2 q
 (ζ )l3 log(||s − rj ||) + O(l4j ), s ∈
/ ∂Ωj ,
 24π ∂χ2 j j

d1 j (s) = (4.3.53)



 1 ∂2 q
 (ζj )l3j (log(lj ) − 5/6) + O(l4j ), s ∈ ∂Ωj .
24π ∂χ2

The same remarks that were made about the sublocal error in Lemma 4.3.1
hold for the result in (4.3.53). For a mild ∂2 q/∂χ2 and a given discretisation,
the most important factor for the sublocal error is log ||s−rj ||. The sublocal error
is large close to the source node s and decays off logarithmically away from this
point. So the most important contributions to the local error d1 (s) will be those
from elements close to a source point.

Since the local error is a sum of sublocal errors, Lemma 4.3.2 puts us in a
position to state the following theorems about the local error in linear elements.

Theorem 4.3.2 (Linear elements local error) The local error (4.3.46) in linear
elements applied to a boundary curve ∂Ω with a discretisation of size lj , j =
1, 2, . . . , N, where N is the number of elements is given by

. 1 X ∂2 q 3 1 ∂2 q
d1 (s) = (χ j )lj log(||s − r j ||) + (χj )l3j (log(lj ) − 5/6). (4.3.54)
24π ∂χ2 24π ∂χ2
/ j
s∈Γ

Proof. The instruments for the proof of this theorem are similar to those for
Theorem 4.3.1. The local error is the sum of the sublocal errors so

.
X X 1 ∂2 q 1 ∂2 q
2
d1 (s) = d1 j (s) = (χ j )lj log(||s − rj ||) + (χj )l3j (log(lj ) − 5/6),
24π ∂χ2 24π ∂χ2
j ∈Γj
s/

(4.3.55)

which is the result. 

Corollary 4.3.4 The local error (4.3.54) is second order in grid size and is given
by
Z
.1
d1 (s) = D log(||s − r(χ)||) dχ, (4.3.56)
24π ∂Ω

where D is a constant second order in grid size.

56 Local errors in BEM for potential problems

∂2 q
Proof. Suppose l2j (χj ) = D over ∂Ω for some constant D. Then
∂χ2

.
X 1 ∂2 q 3 1 ∂2 q
d1 (s) = (ζ j )l log(||s − r j ||) + (ζj )l3 (log(l) − 5/6),
24π ∂χ2 24π ∂χ2
/
s∈∂Ωj

. 1 X 1
= D lj log(||s − rj ||) + Dli (log(li ) − 5/6),
24π 24π
∈∂Ωj
s/
Z
. 1 1
= D lim log(||s − r(χ)||) dχ + D lim li (log(li ) − 5/6),
24π li → 0 24π li → 0
∂Ω−li
Z
1
= D log(||s − r(χ)||) dχ, (4.3.57)
24π ∂Ω

∂2 q
where χ is the arc length coordinate in ∂Ω. If l2j (χj ) is not constant over ∂Ω,
∂χ2
∂2 q
then we take D as the average of l2j (χj ) over ∂Ω, that is, some value between
∂χ2
∂2 q 2
2∂ q
max(l2j (χ j )) and min(lj (χj )). 
∂χ2 ∂χ2

Corollary 4.3.5 (Linear elements local error: Circle) For ∂Ω a circle of radius
R, the local error in linear elements BEM is second order in grid size and is given
by
. 1
d1 (s) = D(2πR) log R, (4.3.58)
24π
where D is second order in grid size and is as defined in Corollary 4.3.4.

Proof. Using the result (4.3.40) for the contour integral in (4.3.57) on a circle
of radius R yields the result

. 1
d1 (s) = D(2πR) log R. (4.3.59)
24π


Corollary 4.3.6 (Linear elements local error: Square) For ∂Ω a square of length
L, the local error using linear elements BEM is second order in grid size and is
given by
. 1
d1 (s) = D(4L)(log(21/4 L) − 1 + π/8). (4.3.60)
24π
where D is second order in grid size and is as defined in Corollary 4.3.4.

Proof. Likewise, using the result (4.3.43) for the contour integral in (4.3.57)
on a square gives the result. 
4.4 Neumann problems 57

4.4 Neumann problems

Consider the case of Neumann boundary conditions. This case results in the
second kind integral equation
1
 
 + Kd (s; r)
  u(r) = g(r) (4.4.1)
2
where the right hand side

g(r) = Ks (s; r)q(r) (4.4.2)

is known since q is given. The solution of (4.4.1) in BEM therefore involves

estimating integrals of the form
Z
1 ∂v
I := (s; r)u(r) dχ. (4.4.3)
2π ∂Ωj ∂n

In the following two sections, local errors in the BEM solution of (4.4.1) using
constant elements and linear elements are discussed.

4.4.1 Constant elements

In the case the unknown function u in (4.4.3) is assumed to be constant over

∂Ωj and takes on the value of u at a point rj ∈ ∂Ωj . That is,
Z
. 1 ∂v
I= (s; r)u(rj ) dχ, (4.4.4)
2π ∂Ωj ∂n

and thus the truncation error committed is given by

Z Z
1 ∂v 1 ∂v
d0 j (s) := (s; r)u(r) dχ − (s; r)u(rj ) dχ. (4.4.5)
2π ∂Ωj ∂n 2π ∂Ωj ∂n

Suppose we have a Taylor series expansion of u(χ) about χj within the element
∂Ωj , that is,

∂u 1 ∂2 u 2 1 ∂3 u
u(χ) = u(χj )+ (χj )(χ−χj )+ (χ j )(χ−χ j ) + (χj )(χ−χj )3 +· · · . (4.4.6)
∂χ 2 ∂χ2 6 ∂χ3
Then using (4.4.6) in (4.4.5) yields the truncation error
Z
. 1 ∂v ∂u
d0 j (s) = (χ) (χj )(χ − χj ) dχ+
2π ∂Ωj ∂n ∂χ
Z
1 ∂v 1 ∂2 u
(χ) (χj )(χ − χj )2 dχ + · · · . (4.4.7)
2π ∂Ωj ∂n 2 ∂χ2
58 Local errors in BEM for potential problems

Recall that
∂v (s − r(χ)) · n(χ)
(χ) = (4.4.8)
∂n ||s − r(χ)||2
where n(χ) is the outward normal at r(χ). To obtain the explicit dependency
on grid size we will need to obtain the dependency of the kernel (4.4.8) on the
grid size as in Sections 4.3. This is a more complicated process but our results
show that indeed the local error is second order in grid size.

4.4.2 Linear elements

In this case the function u in the integral

Z
1 ∂v
I= (s; r)u(r) dγ(r). (4.4.9)
2π ∂Ωj ∂n

is assumed to vary as an order one polynomial over ∂Ωj and in so doing, the so
called sublocal error d1j (s) over ∂Ωj is committed. That is,
Z Z
1 ∂v 1 ∂v
d1 j (s) := (s; r)u(r) dγ(r) − (s; r)fu (r) dγ(r) (4.4.10)
2π ∂Ωj ∂n 2π ∂Ωj ∂n

where fu (r) is the order one polynomial approximating u.

Lemma 4.4.1 (Neumann linear elements sublocal error) Let lj be the grid size
of element ∂Ωj . The sublocal error (4.4.10) is third order in lj and is given by

. 1 ∂2 u ∂v
d1j (s) = (ζj ) (rj )l3j . (4.4.11)
24π ∂χ2 ∂n
∂2 u
where (ζj ) is the tangential double derivative of u at ζj a point in ∂Ω.
∂χ2

Proof. Let ξ be the local coordinate in ∂Ωj , see Figure 4.8. Then ex-
press (4.4.9) in terms of ξ so that,
Z
1 lj ∂v
I := (r(ξ))u(ξ) dξ. (4.4.12)
2π 0 ∂n
Next, replace u by an order one polynomial, that is,
. ξ lj − ξ
u(ξ) = u(ξj ) + u(ξ0 ), (4.4.13)
lj lj
to obtain
Z lj Z
1 lj ∂v ∂2 u
 
. 1 ∂v ξ lj − ξ 1
I= (r(ξ)) u(ξj ) + u(ξ0 ) dξ+ (r(ξ)) (lj −ξ)ξ 2 (ζ(ξ)) dξ.
2π 0 ∂n lj lj 2π 0 ∂n 2 ∂χ
4.4 Neumann problems 59

ξj
∂Ωj
lj
ξ0
ξ
s l0 = 0
χ

Figure 4.8: Local coordinates on

an element in linear boundary el-
ements.

(4.4.14)
The second integral on the right hand side of (4.4.14) is an error term due to
interpolation. Therefore the error when only the first term of (4.4.14) is used to
estimate the integrals in linear elements is
Z
1 lj ∂v 1 ∂2 u
d1j (s) := (ξ) (lj − ξ)ξ 2 (ζ(ξ)) dξ. (4.4.15)
2π 0 ∂n 2 ∂χ
Using the mean value theorem we obtain
. 1 ∂2 u ∂v
d1j (s) = 2
(ζj ) (rj )l3j . (4.4.16)
24π ∂χ ∂n
Note that when s ∈ ∂Ωj , ∂v/∂n = 0 and so there is no need to worry about
singular elements. 

∂2 u
Corollary 4.4.1 If (ζj )l2j = C over ∂Ω for some constant C, then the local
∂χ2
error is second order in grid size and is given by
Z
1 ∂v
d1 (s) = C (χ) dχ, (4.4.17)
24π ∂Ω ∂n
where C is second order in grid size.

Proof. The local error is the sum of all the sublocal errors. So,
X .
X 1 ∂2 u ∂v . 1 X ∂v
d1 (s) = d1j (s) = 2
(ζj ) (χj )l3j = C (χj )lj . (4.4.18)
24π ∂χ ∂n 24π ∂n
j j j

Then using Riemann integral interpretation for the sum in (4.4.18) yields
Z
. 1 ∂v
d1 (s) = C (χ) dχ. (4.4.19)
24π ∂Ω ∂n

60 Local errors in BEM for potential problems

Corollary 4.4.2 Using the properties (2.4.4) of ∂v/∂n, the local error in (4.4.19)
is
.
d1 (s) = −C/48π. (4.4.20)

Proof. From (2.4.4), the integral in (4.4.19) is −1/2 hence the result. 

4.5 Mixed boundary conditions

For a mixed problem, there is a part ∂Ω1 of the boundary ∂Ω where u is given
and a part ∂Ω2 where q is given, see Figure 4.9.

∂Ω1 , ∂Ω2 ,
u given q given

Figure 4.9:

Thus we have to estimate integrals of the form (4.3.3) on the part where u is
given and (4.4.3) on the part where q is known. Therefore, when using constant
elements, the truncation error committed per element is given by (4.3.6) in the
part where u is known and (4.4.7) in the Neumann part. When linear elements
are used, the sublocal error is given by (4.3.47) in the Dirichlet part and (4.4.11)
in the Neumann part.

Corollary 4.5.1 (Mixed problem linear elements local error) Let the Dirichlet
part ∂Ω1 be discretised using M elements and the Neumann part be discretised
using N elements. Then the local error for a mixed problem when using linear
elements BEM is given by

M
X
. 1 ∂2 q 1 ∂2 q
d1 (s) = 2
(χj )l3j log(||s − rj ||) + (χj )l3j (log(lj ) − 5/6)
24π ∂χ 24π ∂χ2
j=1,s∈Γ
/ j
N
1 X ∂2 u ∂v
+ (ζj ) (rj )l3j , (4.5.1)
24π ∂χ2 ∂n
j=M+1
4.5 Mixed boundary conditions 61

when the point s ∈ ∂Ω1 and

M
. 1 X ∂2 q
d1 (s) = (χj )l3j log(||s − rj ||)
24π ∂χ2
j=1,s∈Γ
/ j
N
X
1 ∂2 u ∂v
+ 2
(ζj ) (rj )l3j , (4.5.2)
24π ∂χ ∂n
j=M+1

when the point s ∈ ∂Ω2 .

Proof. The local error is the sum of all the local errors per element. Then
M
X N
X
d(s) = d1 j + d1 j . (4.5.3)
j=1, ∂Ωj ∈∂Ω1 j=M+1, ∂Ωj ∈∂Ω2

Then using the results (4.3.47) and (4.4.11) in (4.5.3) gives the result. 

Corollary 4.5.2 The local error (4.5.1) is second order in grid size and is given
by
Z Z
. 1 1 ∂v
d1 (s) = D log(||s − r(χ)||) dχ + D (χ) dχ, (4.5.4)
24π ∂Ω1 24π ∂Ω2 ∂n

where D is a constant second order in grid size.

∂2 q ∂2 u
Proof. Suppose l2j
2
(χj ) = D over ∂Ω1 and l2j 2 (χj ) = D over ∂Ω2 for some
∂χ ∂χ
constant D. Then we have, on ∂Ω1 ,
N
.
X 1 ∂2 q 3 1 ∂2 q
d1 (s) = (χ j )lj log(||s − rj ||) + (χi )l3i (log(li ) − 5/6),
24π ∂χ2 24π ∂χ2
j=1, s/
∈∂Ωj
N
X
. 1 1
= D lj log(||s − rj ||) + Dli (log(li ) − 5/6),
24π 24π
j=1, s/
∈∂Ωj
Z
. 1 1
= D lim log(||s − r(χ)||) dχ + D lim li (log(li ) − 5/6),
24π li → 0 24π li → 0
∂Ω1 −li
Z
1
= D log(||s − r(χ)||) dχ. (4.5.5)
24π ∂Ω1

On ∂Ω2 ,
N
X N
X X ∂v
. 1 ∂2 u ∂v 3 . 1
d1 (s) = d1 j (s) = (χ j ) (χ j )lj = D (χj )lj . (4.5.6)
24π ∂χ2 ∂n 24π ∂n
j=1 j=M+1 j
62 Local errors in BEM for potential problems

Using Riemann integral interpretation for the sum in (4.5.6) leads to

Z
. 1 ∂v
d1 (s) = D (χ) dχ. (4.5.7)
24π ∂Ω ∂n

Putting (4.5.5) and (4.5.7) together gives the result in (4.5.4).

∂2 q ∂2 u
If l2j 2
(χj )/l2j 2 (χj ) is not constant over ∂Ω1 /∂Ω2 , then take D as the average
∂χ ∂χ
2 2
2∂ q 2∂ u
of lj 2 (χj )/lj 2 (χj ) over ∂Ω1 /∂Ω2 . 
∂χ ∂χ

4.6 Examples

Below we present results to illustrate the estimates derived above using Prob-
lems (a) and (c) of Examples 3.7.1 and 3.7.3. Since in each example analytic
expression for the unknown is available, for each node ri we can compute the
exact value qi or ui of the unknown. Then the local error defined in (4.1.5) can
be computed. That is

d := Aq − b or d := Au − b.

We refine by a factor three and, to see what happens to the error after each
refinement, the ratios of the errors at consecutive grids are computed. That is,

error ratio = ||d||∞ (N)/||d||∞ (3N), (4.6.1)

where ||d||∞ (N) denotes the infinity norm of d of length N. In the formulation of
the BEM systems for each example, the right hand side is computed ”exactly”
as explained in Section 4.2.1.

Example 4.6.1 (Dirichlet Problem (a) using constant elements) The problem
in Example 3.7.1 is solved using constant elements. The results are shown in Fig-
ure 4.10.
4.6 Examples 63

Problem (a) constant elements

N ||d||∞ error ratios −2

5 2.63E-02 2.34 −4
15 1.13E-02 7.37 −6

log(||d||∞)
45 1.52E-03 8.53 −8
135 1.79E-04 8.85 −10
405 2.03E-05 8.95
−12
1215 2.26E-06 -
−14
0 2 4 6 8
log(N)

Figure 4.10: Local errors Problem (a) using constant elements.

Example 4.6.2 (Dirichlet Problem (a) using linear elements) The problem in
Example 3.7.1 is solved using linear elements. The results are shown in Fig-
ure 4.11.
Problem (a) linear elements
N ||d||∞ error ratios −4

5 1.57E-02 1.17 −6
15 1.34E-02 7.07
log(||d||∞)

−8
45 1.90E-03 8.48
135 2.24E-04 8.84 −10

405 2.53E-05 8.94 −12

1215 2.83E-06
−14
0 2 4 6 8
log(N)

Figure 4.11: Local errors Problem (a) using linear elements.

Example 4.6.3 (Neumann Problem (c) using constant elements) The problem
in Example 3.7.3 is solved using constant elements. The results are shown in Fig-
ure 4.12.
Problem (c) constant elements
N ||d||∞ error ratios −2

5 5.17E-02 3.72 −4
15 1.39E-02 8.40 −6
log(||d||∞)

45 1.65E-03 8.98 −8
135 1.84E-04 9.01 −10
405 2.04E-05 9.01
−12
1215 2.27E-06 -
−14
0 2 4 6 8
log(N)

Figure 4.12: Local errors Problem (c) using constant elements.

64 Local errors in BEM for potential problems

Example 4.6.4 (Neumann Problem (c) using linear elements) The problem in
Example 3.7.3 is solved using linear elements. The results are shown in Fig-
ure 4.13.

Problem (c) linear elements

N ||d||∞ error ratios −2

5 2.92E-02 2.54 −4
15 1.16E-02 6.39 −6

log(||d||∞)
45 1.82E-03 8.24 −8
135 2.21E-04 8.76 −10
405 2.52E-05 8.92
−12
1215 2.82E-06 -
−14
0 2 4 6 8
log(N)

Figure 4.13: Local errors Problem (c) using linear elements.

All the results, for both the Dirichlet and the Neumann problem, show a similar
error trend with respect to grid size. As we can see by taking ratios of consec-
utive errors, each time we refine by about a factor three, the local error goes
down by about a factor of nine. The behaviour is the same for both constant
and linear elements. This shows that the local error is indeed of second order
convergence with respect to grid size for both constant and linear elements as
our findings have revealed.

4.7 Equidistribution

For problems with localised regions of rapid variation, choosing a good mesh
is essential if sufficiently accurate solutions are to be obtained as cheaply as
possible. This is more important in BEM which has full matrices and uniform
fine grids will drastically increase the cost of computation. Instead we would
like to have a mesh that reflects the activity of the solution. That is the choice of
mesh considered is based on controlling the discretisation error. The procedure
discussed here is similar to that developed by [79, p. 65]. In contrast to this
approach, the quantity we intend to equidistribute is not just some quantity
related to the local error of the BEM solution but rather its asymptotically
correct estimate as developed in Section 4.2.1.

Generally the following mesh selection problem is addressed [4, p. 359]: Given
a boundary value problem and an error tolerance TOL, find a mesh

M : 0 = θ1 < θ2 < · · · < θn = 2π (4.7.1)

4.7 Equidistribution 65

with
l = max li , li = θi+1 − θi (4.7.2)
1≤i≤N

such that N is small and the error in yM (θ) as an approximate solution to y(θ)
is less than TOL. On the other hand, for a give number of elements N, find a
mesh
M : 0 = θ1 < θ2 < · · · < θn = 2π (4.7.3)
such that the local error over M is constant. Generally the mesh must be
fine in regions where the desired solution changes rapidly but can be relatively
coarse elsewhere. A mesh selection strategy is more successful if it utilises the
special properties of the particular numerical method being used, namely its
error form [4, p. 361], [10]. However, this error form is normally based upon
an asymptotic analysis that the approximation has not yet reached.

If the error we make on each element is known then the information can be
used to find a mesh that distributes this error uniformly through out the
mesh. This is what is called error equidistribution. In what is pursued here,
a good mesh is obtained by (re)distributing mesh elements through local error
equidistribution after an initial approximation has been obtained. The benefits
of equidistributing the local error are twofold. In Chapter 5 it is shown that the
global error will be equidistributed when the local error is. Besides, in [6] it is
shown that local error equidistribution will minimise the average global error.

Assume a problem whose continuous solution has a small region of high activ-
ity. Then a positive weight function φ as an indicator function for the smooth-
ness of the solution is needed, see [1, p. 77], [10], [4, p. 363]. The weight
function φ represents some error measurement of the solution and is called
a monitor function. Then error equidistribution is based on equidistribution
of this function. Therefore the success of any mesh equidistribution strategy
largely hinges on the choice of the monitor function, see [7] and the references
therein.

The idea is now the following: Suppose di is some measure of the error on
element Γi . The error di depends on the size of the element Γi . Let li be the
length of Γi (again Γi is a straight line). Generally the error increases as li
increases. More precisely let
di = Ci lp
i (4.7.4)
be the error on Γi where the constant Ci depends on the smoothness of the
solution. It is convenient to consider a measure with a linear variation in li ,
that is,
Di := li φi
with φi independent of li . So from (4.7.4) we take
1/p 1/p
Di = di = li Ci
66 Local errors in BEM for potential problems

1/p
so that φi := Ci . The error is equidistributed if for some constant λ
1/p
li Ci = λ, i = 1, 2, . . . , N. (4.7.5)

The problem we have in BEM is that the error per element is the sublocal
error given by (4.3.6). However, what we can actually measure is the local
error, which is a sum of all the sublocal errors and therefore it depends on
all the grid element sizes. So we need to have a measure of the local error on
an element that is dependent on the grid size of that particular element. The
sublocal errors are given by

. ∂2 q
dj (s) = C (rj )l3j log(||s − rj ||), (4.7.6)
∂τ2
where C = −1/48π in constant elements and 1/24π in linear elements. That is
the local error on an element will be
N
.
X ∂2 q
d(s) = C (rj )l3j log(||s − rj ||). (4.7.7)
∂τ2
j=1

Since the logarithm decays to zero away from s, the most important contribu-
tions will be those for which ||s − rj || → 0, in particular the contribution from the
element for which j = i. Thus the following approximation can be made:


 1 ∂2 q 3
 − 48π ∂τ2 (ri )li [log(li /2) − 1/3], in constant elements,

d(ri ) ≈ di (ri ) =

 2
 1 ∂ q (r̃i )l3 [ln[li ] − 5/6],

in linear elements.
i
24π ∂τ2
(4.7.8)
For the values of l that are considered here, [log(li /2) − 1/3] and [ln[li ] − 5/6] are
simply order one constants. Then
∂2 q
d(ri ) ≈ Ci (ri )l3i , (4.7.9)
∂τ2
where

 C(log(li /2) − 1/3), in constant elements,
Ci = (4.7.10)

C(ln[li ] − 5/6), in linear elements.
The error in (4.7.9) also depends directly on the activity of the solution and so
will be large where the solution is more active and small otherwise. Therefore
we equidistribute this error as the local error per element. An interesting ob-
servation from (4.7.8) is that the ratio of the error terms for constant and linear
elements is a higher order term. That is
const (ln li − ln 2 − 1/2)
≈− (4.7.11)
linear 2(ln li − 5/6)
4.7 Equidistribution 67

which will give a higher order truncation error.

Comparing with (4.3.6) and (4.3.47), the result in (4.7.9) seems to suggest that
it is enough to equidistribute the sublocal errors. That is, when the sublocal
error is equidistributed, then so is the local error but the reverse is not neces-
sarily true.

Now, equidistributing the sublocal error means

dj (ri ) = λ, for all j, (4.7.12)

where λ is a constant. Then the local error d(ri ) is given by

X X X
d(ri ) = dj (ri ) = dj = λ = Nλ =: Λ, for all i, (4.7.13)
j j j

where Λ is a constant. Therefore the local error is also equidistributed.

On the other hand if d(ri ) is equidistributed, then

d(ri ) = β for all i, (4.7.14)

where β is some constant. That is, in the case of rapidly varying q(ξ),
X
dj (ri ) = d(ri ) = β for all i. (4.7.15)
j

This does not necessarily imply dj (ri ) is equidistributed as well.

Let us now describe strategies for error equidistribution and give some exam-
ples.

Consider the problem of Example 3.7.2. The domain and exact solution on the
circle of radius R = 1.2 are shown in Figure 4.14. Now, using (4.7.8), define

∂2 q
d(ri ) = Ci (ri )l3i (4.7.16)
∂τ2
where Ci is defined in (4.7.10). Then we have
1/3
∂2 q

i
D = Ci 2 (ri ) li . (4.7.17)
∂τ

Therefore we take the monitor function φi to be

1/3
∂2 q

φi = Ci 2 (ri ) (4.7.18)
∂τ

where Ci is defined in (4.7.10). We see that φi is a measure of the variation

of q on Γi and depends on the second order tangential derivative ∂2 q/∂τ2 . For
68 Local errors in BEM for potential problems

1.5

1 0

0.5 −0.2
R=1.2
r −0.4

q(θ)
s
0
Ω −0.6
−0.5
−0.8
−1 −1

−1.5 0 2 4 6
−1 0 1 θ

(b) Solution on circle with high activ-

(a) Disc domain.
ity around θ = π

Figure 4.14: A disc domain and solution to Example 3.7.2

with rs = (−1.4, 0) on the circular boundary of radius R =
1.2.

a BEM solution of (4.14a), we describe the equidistribution of the error in the

solution obtained by using constant elements. The strategy described is such
that given a number of elements N, we want a grid that ensures that

φi li = λ, i = 1, 2, . . . , N, (4.7.19)

for some constant λ. However, in order to use (4.7.18), we need to estimate

the second order tangential derivatives ∂2 q/∂τ2 . We can use finite difference
formulas and the initial BEM approximations to do this. Suppose for each
element Γi we have the constant element solution q0i and the linear element
solutions q1i and q2i as shown in Figure 4.15. The second order derivatives are

li /2 li /2
q1i q2i
q0i

Figure 4.15: Constant ele-

ment solution q0i and linear
element solutions q1i and q2i
on an element Γi .

estimated as
∂2 q 1
(ri ) ≈ (q1 − 2q0i + q2i ). (4.7.20)
∂τ2 (0.5li )2 i

The monitor function φi is then given by

 1/3
Ci 1 0 2
φi = (q − 2qi + qi ) . (4.7.21)
0.25l2i i
4.7 Equidistribution 69

Then from (4.7.19) we have

N
X N
1 X
φi li = Nλ ⇒ λ = φi li . (4.7.22)
N
i=1 i=1

So a new grid is obtained that ensures that

N
1 X
lnew
j φj = λ = φi li . (4.7.23)
N
i=1

Therefore we have a new grid lnew

j given by

N
!
1 X
lnew
j = φi li /φj . (4.7.24)
N
i=1

Figure 4.16b shows the solution on a uniform grid of 15 elements. We see that
the deviation from the exact solution is large in the active region than in the
rest of the boundary. Figure 4.17a shows the grid obtained by equidistributing
the error using 15 elements. As expected the grid is small in the active region
and coarse elsewhere. The solution on this grid is shown in Figure 4.17b.
Figure 4.18 shows the results of equidistribution using 45 elements. Likewise
we see that the grid is fine in the active region and coarses out as the activity
of q reduces.

1.5

1 0

0.5 −0.2

−0.4
q(θ)

0
−0.6
−0.5
−0.8
−1 −1

−1.5 0 2 4 6
−1 0 1 θ

(b) Solution on a 15 elements uniform

(a) A 15 elements uniform grid.
grid

Figure 4.16: A starting uniform grid of N = 15 elements

and the corresponding solution using constant elements.
The continuous line in 4.16b is the exact continuous solu-
tion and the circles are the BEM solution.
70 Local errors in BEM for potential problems

1.5

1 0

0.5 −0.2

−0.4

q(θ)
0
−0.6
−0.5
−0.8
−1 −1

−1.5 0 2 4 6
−1 0 1 θ

(b) Solution on a 15 elements equidis-

(a) A 15 elements equidistributing grid.
tributing grid

Figure 4.17: An equidistributing grid of N = 15 elements

and the corresponding solution using constant elements.
The continuous line in 4.17b is the exact continuous solu-
tion and the circles are the BEM solution.

1.5

1 0

0.5 −0.2

−0.4
q(θ)

0
−0.6
−0.5
−0.8
−1 −1

−1.5 0 2 4 6
−1 0 1 θ

(b) Solution on a 45 elements equidis-

(a) A 45 elements equidistributing grid.
tributing grid

Figure 4.18: An equidistributing grid of N = 45 elements

and the corresponding solution using constant elements.
The continuous line in 4.18b is the exact continuous solu-
tion and the circles are the BEM solution.
“It sounds paradoxical to say the attainment of scientific truth has
been effected, to a great extent, by the help of scientific errors.” –
Thomas Henry Huxley.
72 Local errors in BEM for potential problems
Chapter 5

Global errors

5.1 Introduction

In Chapter 4 an analysis of local errors was given. Also the behaviour of local
errors was discussed and the concept of sublocal errors introduced. However,
what is really needed in practice is the global error. In this chapter we asses
this in particular by investigating the relationship between the global error and
the local error for the Neumann and Dirichlet problems. We would like to show
that the global error behaves the same way as the local error. This is important
because it means that if we have information about one of the errors then
we can know what to expect of the other. For instance if the local error is
equidistributed, then we may hope the global error is as well.

In Section 2.5 it was noted that both the single layer operator Ks (r; r ′ ) and
the double layer operator Kd (r; r ′ ) for the model problem on a circle have the
same eigenfunctions, being sines and cosines. This provides us with a good
opportunity to study and compare solutions to integral equations of the single
and double layer operators. We can employ Fourier series for the unknown
functions and a spectral expansion of the operators to write general solutions
to the equations. This technique is used to show that, indeed, the global error
is of the same order as the local error, moreover it will be equidistributed when
the local error is equidistributed.

The type of operator in a BEM formulation will depend on whether the problem
has Dirichlet, Neumann or mixed boundary conditions. It is therefore rather
too general to say that we are looking at errors in BEM. To be more specific, the
different problems, that is, the Dirichlet, Neumann and mixed problems, are
looked at separately. In Section 5.2 we consider a Dirichlet problem and exam-
ine the global error for different cases of local error distribution. In Section 5.4
74 Global errors

we look at Neumann problems and then mixed problems in Section 5.5. We

illustrate the theory presented with results from numerical experiments. The
results show that the global errors are second order indeed.

5.2 Spectral decomposition

It was seen in Section 2.5 that a Dirichlet problem leads to a Fredholm integral
equation of the first kind,

Ks (r; r ′ )q(r ′ ) = h(r), (5.2.1)

where h(r) is a given function and the operator Ks (r; r ′ ) is defined in (2.4.5). Let
us discretise (5.2.1) using a grid formally indicated by size L. Then we have the
approximation, say

Ks L (r; r ′ )qL (r ′ ) = hL (r), (5.2.2)

where superscript L is a grid parameter denoting a numerical approximation on

the grid. If we were to use the exact function q(r ′ ) in (5.2.2) in place of qL (r ′ )
we would obtain

Ks L (r; r ′ )q(r ′ ) = hL (r) + dL (r), (5.2.3)

where the residual dL (r) is the local discretisation error. Subtracting (5.2.2)
from (5.2.3) we get

Ks L (r; r ′ )eL (r ′ ) = dL (r), (5.2.4)

where

eL (r ′ ) := q(r ′ ) − qL (r ′ ), (5.2.5)

is the global error. To obtain (5.2.3) we have in fact substituted the exact solu-
tion into the approximate integral equation. We can also consider the converse,
that is, substitute an approximate solution into the exact continuous integral
equation. Thus, suppose qL (r) is an approximation of q(r) on a grid of size L,
then we would have

Ks (r; r ′ )qL (r ′ ) = h(r) + d(r), (5.2.6)

where again d(r) is a local error . Likewise subtracting (5.2.1) from (5.2.6) gives

Ks (r; r ′ )e(r ′ ) = d(r), (5.2.7)

where

e(r) := qL (r) − q(r), (5.2.8)

5.2 Spectral decomposition 75

is a global error. Therefore given a local error d(r), the global error is the
solution of the integral equation (5.2.7) with the local error as the right hand
side.

From the spectral theory of Ks , let λk be the eigenvalues of Ks and φk the

corresponding eigenfunctions. Then we have
X
Ks e(r) = λk (e, φk )φk (r). (5.2.9)
k

where the inner product (·, ·) is defined in (2.5.4). We can then formally conve-
niently solve (5.2.7) by using an expansion of d in eigenfunctions. Note that the
eigenvalues of Ks have an accumulation point at zero. Using (5.2.9) in (5.2.7)
we have
X
λk (e, φk )φk (r) = d(r). (5.2.10)
k

Systematically taking the inner product with φk (5.2.10) gives

λk (e, φk ) = (d, φk ), (5.2.11)
where (e, φk ) are the expansion coefficients of e and (d, φk ) are the expansion
coefficients of d. Solving the global error is now equivalent to solving (5.2.11)
for its coefficients (e, φk ). This formally yields
1
(e, φk ) = (d, φk ), (5.2.12)
λk
where we assume that λk 6= 0. So
X
e(r) = (e, φk )φk (r). (5.2.13)
k

From (5.2.12) we see that since the eigenvalues λk go to zero for increasing k,
the global error coefficients might grow unboundedly large depending on the
local error coefficients. We now have the following
 α+1 !
(d, φk ) 1 P 1
Theorem 5.2.1 If d is such that ≤O , α > 0, then |(d, φk )|
λk k k λk
is a finite bound for ||e||∞ .

Proof. Substituting (5.2.12) into (5.2.13) gives

X 1
e(r) = (d, φk )φk (r). (5.2.14)
λk
k

Since the basis functions are orthonormal, we then find

X 1
||e||∞ ≤ |(d, φk )| . (5.2.15)
λk
k


76 Global errors

Corollary 5.2.1 Under the assumptions of Theorem 5.2.1 both the constant and
the linear elements schemes are convergent.

It is therefore important to have information about the λk to assess (5.2.12).

The case for a general boundary is very complicated and therefore we restrict
ourselves to a circular domain. We conjecture that the general case then fol-
lows using conformal mapping. In [16, p. 26], it has been established that,
in the case of a Dirichlet problem on a circle, the eigenvalues λk are O(1/k).
From (5.2.12) we therefore conclude that to have convergence in the coeffi-
cients of the global error the coefficients of the local error must be O(1/k2+α )
where α > 0. It is this phenomenon that is the subject of investigation in the
following subsections for different forms of d.

In order to do this we need to have the eigensolutions of the problem. We still

restrict ourselves to a Dirichlet problem on a circle, in which case the eigen-
functions are sines and cosines and expansions of functions are just Fourier
series. We begin with a problem where the local error is constant and then will
show how this can be used for almost constant problems, that is, where the
local error is properly equidistributed.

5.3 Dirichlet problems

A more detailed investigation of the global error requires assessing the Fourier
coefficients (d, φk ). As we fix our thought on a circular domain, we can use
trigonometric functions for a basis. Recall from Section 2.5.1 that the single
layer operator Ks for the Dirichlet problem has cosines and sines as eigen func-
tions. On a circle of radius R, R/2k is an eigenvalue for Ks with eigenfunctions
cos kθ and sin kθ where θ is the polar angle. Also, for a symmetric kernel K with
orthonormal eigenfunctions yk , the Fourier coefficients of an arbitrary function
h are given by hk = (h, yk ), see [41, p. 324]. Basically, for a function f of period
2π that is integrable over the period, the Fourier series of f is the trigonometric
series, see [39, p. 532],
∞
X
a0 + (an cos nχ + bn sin nχ), (5.3.1)
n=1

where an , bn , n = 1, 2, . . . are the Fourier coefficients of f given by

Z
1 π
a0 := f(χ) dχ, (5.3.2a)
2π −π
Zπ
1
an := f(χ) cos nχ dχ, n = 1, 2, . . . , (5.3.2b)
π −π
Zπ
1
bn := f(χ) sin nχ dχ, n = 1, 2, . . . . (5.3.2c)
π −π
5.3 Dirichlet problems 77

We note here that since we focus ourselves on a circular boundary, the natural
arc length coordinate χ is given by χ = Rθ. In what follows we use Fourier series
analysis to relate global error to local error.

Theorem 5.3.1 If the local error d is constant over ∂Ω, say d ≡ C over ∂Ω, then
e is also constant over ∂Ω, that is,

e(r) = C0 , (5.3.3)

for some constant C0 ∈ R.

Proof. This result follows directly from Theorem 5.2.1 noting that all Fourier
coefficients of d are zero except the one for the constant function. 

∂2 q
Corollary 5.3.1 If we use constant elements and l2j )(χj ) is equidistributed,
∂χ2
∂2 q
that is, l2j )(χj ) = C0 a constant, then the global error has the same order as
∂χ2
the local error and is given by

. 1
||e0 ||∞ = C0 (2πR) log(1/R), (5.3.4)
48π
where R is the radius of the circle.

∂2 q
Proof. Recall from Lemma 4.3.2, l2j )(χj ) = C0 results in
∂χ2

1
d0 (s) = C0 (2πR) log(1/R), (5.3.5)
48π
where C0 is second order in grid size. This result is independent of s since the
contour integral in (4.3.40) gives the same result for all positions s. So using
Theorem 5.2.1, only the Fourier coefficient for the constant mode survives to
give the result. 

∂2 q
Corollary 5.3.2 If we use linear elements and l2j )(χj ) is equidistributed, that
∂χ2
∂2 q
is, l2j )(χj ) = C1 a constant, then the global error has the same order as the
∂χ2
local error and is given by

. 1
||e1 ||∞ = C1 (2πR) log R, (5.3.6)
24π
where R is the radius of the circle.
78 Global errors

∂2 q
Proof. Again, from Lemma 4.3.5, l2j )(χj ) = C1 results in
∂χ2
1
d1 (s) = C1 (2πR) log(1/R), (5.3.7)
48π
where C1 is second order in grid size. The result is independent of s since the
contour integral in (4.3.40) gives the same result for all positions s. So using
the results of Theorem 5.2.1, only the Fourier coefficient for the constant mode
survives to give the result. 

For a more general result we can proceed along two lines. First one may wonder
what is the effect of neglecting higher order terms in local error estimates. For
this we have

Theorem 5.3.2 Let q ∈ C2 (∂Ω) and define l := max li . Then ||e||∞ = O(l2 ).
i=1,...,N

Proof. From Theorem 4.3.1 we conclude that

Z Z
1 ∂r 1 1
d0 (s) = − C1 dχ − C2 log(||s − r(χ)||) dχ + O(l3 ), (5.3.8)
24π ∂χ r 48π
where
∂q ∂q
min (χj )lj ≤ C1 ≤ max (χj )lj ,
∂χ ∂χ

∂2 q ∂2 q
min 2
(χj )l2j ≤ C2 ≤ max (χj )l2j .
∂χ ∂χ2
Let
Z Z
1 ∂r 1 1
M(χ) := − C1 dχ − C2 log(||s − r(χ)||) dχ, (5.3.9)
24π ∂χ r 48π
and

δ(χ) := O(l3 ), (5.3.10)

that is δ(χ) is the order three terms. Thus we have

d0 (s) = M(χ) + δ(χ). (5.3.11)

We can expand the δ formally in a Fourier expansion. Before doing that we note
that they can be approximated by a periodic higher order approximation poly-
nomial of degree less than or equal to N − 1, which means that the coefficients
for frequencies larger that N are zero. Hence
N−1
X
δ(χ) = (αj cos jχ + βj sin jχ) , (5.3.12)
j=0
5.4 Neumann problems 79

1
where |αj |, |βj | = O( ). We now find that
j
N−1
X 
j j
e(χ) = M + αj cos jχ + βj sin jχ (5.3.13)
2π 2π
j=0

1
Since αj , βj are O(l3 ) and we sum over N terms N = O( ) and, finally, since
l
M = O(l2 ) we thus have proven the result. 

In Table 5.1 we have the infinity norm of the global errors in the solution of
Example 3.7.1. We see that each time we increase the number of elements N by
three the error goes down by a factor of nine. The behaviour is the same for both
constant and linear elements as shown in Tables 5.1(a) and (b) respectively.
Thus the error is second order in grid size as is the local error.

N ||e||∞ error ratios N ||e||∞ error ratios

5 1.02E-01 2.63 5 7.14E-02 1.50
15 3.87E-02 7.58 15 4.74E-02 7.46
45 5.12E-03 8.56 45 6.37E-03 8.53
135 5.98E-04 8.85 135 7.46E-04 8.85
405 6.75E-05 8.92 405 8.44E-05 8.92
1215 7.57E-06 - 1215 9.45E-06

(a)Constant elements (b) Linear elements

Table 5.1: Global errors in the solution of Dirichlet Example 3.7.1 using con-
stant elements in (a), and linear elements in (b).

5.4 Neumann problems

A Neumann problem leads to a Fredholm integral equation of the second kind

given by

1
 
 I + Kd 
  u = f. (5.4.1)
2
Since we restrict ourselves to a circular domain, we have a relatively simple
spectral decomposition. Indeed, from (2.5.16) we obtain for the operator K =
1
I + Kd that
2
1
λ0 = 0, λj = , j ≥ 1, (5.4.2)
2
80 Global errors

with corresponding eigenfunctions 1, and cos jχ and sin jχ for j ≥ 1 respectively.

Following the same approach as in obtaining (5.2.7), we have

1
 
 I + Kd 
  e = d. (5.4.3)
2

Theorem 5.4.1 The global error e(r) of the solution to a Neumann problem is
second order in grid size, that is, the same order as the local error d(r). That is

e(r) = Cd(r), (5.4.4)

for some C ∈ R.

Proof. For a given d we look for solutions of (5.4.3)

 in thespace W1 introduced
1
in Section 2.5. In this space the operator K =   I + Kd 
 has all eigenvalues
2
equal to 1/2 with sines and cosines as eigenfunctions. So we have
X
λk (e, φk )φk (r) = d(r), λk = 1/2, (5.4.5)
k

for all k. Taking the inner product with a φk we obtain

(e, φk ) = 2(d, φk ), (5.4.6)

for all k. Thus we expect the global error to be of the same order as the local
error since its Fourier coefficients are only a scalar multiple of the local error
coefficients. 

The behaviour is the same for Neumann problems as shown in Table 5.2 and
as expected the error order is the same as that for the local error.

N ||e||∞ error ratios N ||e||∞ error ratios

5 5.93E-01 9.56 5 1.33E-01 4.00
15 6.20E-02 9.38 15 3.32E-02 5.38
45 6.61E-03 9.70 45 6.16E-03 8.08
135 6.81E-04 9.36 135 7.63E-04 8.74
405 7.28E-05 9.16 405 8.73E-05 8.92
1215 7.95E-06 - 1215 9.78E-06 -

(a) Constant elements (b) Linear elements

Table 5.2: Global errors in the solution of Neumann Example 3.7.3 using con-
stant elements in (a) and linear elements in (b).
5.5 Mixed problem 81

5.5 Mixed problem

The mixed problem requires some additional analysis. Suppose u(r) = g(r) on
a part ∂Ω1 of the boundary and q(r) = h(r) on a part ∂Ω2 , see Figure 5.1.

∂Ω1 ,
θ=α u = g(r) given

θ=0

∂Ω2 ,
q = h(r) given

Figure 5.1: Boundary with

mixed boundary conditions.

The integral equation for such a mixed problem can be written as

1 d s
u + K∂Ω2 u − K∂Ω1 q = f (5.5.1)
2
where
d s
f := −K∂Ω1 g + K∂Ω2 h. (5.5.2)

So we have

d s 1
K∂Ω2 u − K∂Ω1 q = f − g, r ∈ ∂Ω1 , (5.5.3a)
2
1 d s
u + K∂Ω2 u − K∂Ω1 q = f, r ∈ ∂Ω2 . (5.5.3b)
2
Consider

Kd u − Ks q = f. (5.5.4)

Now let
X
u(θ) = (αk sin kθ + βk cos kθ) on ∂Ω2 , (5.5.5)

X
q(θ) = (γk sin kθ + δk cos kθ) on ∂Ω1 , (5.5.6)
82 Global errors

where as
X
f(θ) = (ξk sin kθ + ηk cos kθ) on ∂Ω1 , (5.5.7)

X
f(θ) = (πk sin kθ + ρk cos kθ) on ∂Ω2 . (5.5.8)

Let
d s
L(u, q) = K∂Ω2 u − K∂Ω1 q.

Then we have
X

L(1, 1) = ãj cos jθ + b̃j sin jθ . (5.5.9)

Similarly

L(cos k·, cos k·) = ak + bk sin kθ + ck cos kθ (5.5.10)

and

L(sin k·, sin k·) = a ^ k sin kθ + c^k cos kθ.

^k + b (5.5.11)

Then on ∂Ω1
X
L(u, q) = ^ k sin kθ + αk c^k cos kθ
^ k + αk b
αk a
k
+ βk ak + βk bk sin kθ + βk ck cos kθ + β0 ãk cos kθ + β0 b̃k sin kθ . (5.5.12)

So
X
η0 = (αk a
^k + βk ak ) , (5.5.13a)
X
ηk = (αk c^k + βk ck ) + β0 ãk , (5.5.13b)
X
^ k + βk bk + β0 b̃k .


ξk = αk b (5.5.13c)

If β0 were known, we would find αk and βk from (5.5.13b) and (5.5.13c) since
η0 , ηk and ξk are known from the right hand side of (5.5.4) which is given. In
fact we have to solve a sparse system. If we truncate the series in (5.5.13) we
have a sparse system from which we can solve for β1 , β2 , . . . , βN and α1 , α2 , . . . , αN .

Now, for the mixed problem, the integrals in (2.5.11), (2.5.12) and (2.5.14) are
valid only on part of the circle. Thus, suppose ∂Ω1 is the arc θ ∈ [0, α] and ∂Ω2
is the arc θ ∈ [α, 2π], see Figure 5.1. Then we have, for the cos kθ eigenfunctions,
Zα Z
R R α
(Ks cos kθ)∂Ω1 = − log R cos(kθ ′ ) dθ ′ + cos(kθ) cos2 (kθ ′ ) dθ ′
2π 0 2πk 0
Z
R α
+ sin(kθ) sin(kθ ′ ) cos(kθ ′ ) dθ ′ . (5.5.14)
2πk 0
5.5 Mixed problem 83

Evaluating the integrals we get

 
s R sin kα R α sin(2kα)
(K cos kθ)∂Ω1 =− log R + cos(kθ) +
2π k 2πk 2 4k
R sin kθ
+ sin2 kα. (5.5.15)
2πk 2k
For the sin kθ eigenfunctions we have
Zα Z
R R α
(Ks sin kθ)∂Ω1 = − log R sin(kθ ′ ) dθ ′ + cos(kθ) cos(kθ ′ ) sin(kθ ′ ) dθ ′
2π 0 2πk 0
Z
R α
+ sin(kθ) sin2 (kθ ′ ) dθ ′ , (5.5.16)
2πk 0

which yields

R R cos kθ
(Ks sin kθ)∂Ω1 = − log R(1 − cos kα) + sin2 kα
2πk 2πk 2k 
R α sin 2kα
+ sin(kθ) − . (5.5.17)
2πk 2 4k

For well conditioning of (5.5.13b) and (5.5.13c), we have to look at the system

R sin2 kα
   
R α sin 2kα
  +
c^k ck  2kπ 2k 2kπ 2 4k 
^k =
 R  α sin 2kα  2
 =: A, (5.5.18)
b bk R sin kα 
−
2kπ 2 4k 2kπ 2k

which is obtained by comparing (5.5.15) and (5.5.17) with (5.5.10) and (5.5.11).
So,

R2 sin4 kα α2 sin2 2kα

 
det A = 2 2 − + , (5.5.19)
4k π 4k2 4 16k2

which is not zero unless α = 0. We are therefore able to obtain the coefficients
αk and βk . A similar process as above will show that we are able to compute
the coefficients γk and δk . So we deduce that u(θ) can be found as an infinite
Fourier series (5.5.5), and likewise q(θ) in (5.5.6). In computing the global error,
the unknowns u(θ) on ∂Ω2 and q(θ) on ∂Ω1 are instead the local errors in these
functions. So we have the following theorem

∂2 q
Theorem 5.5.1 If we use linear elements and l2j (χj ) is constant over ∂Ω1
∂χ2
∂2 u
and l2j (χj ) is constant over ∂Ω2 , then the expansion coefficients αk , βk of the
∂χ2
global error are second order in grid size, that is αk , βk = O(l2 ).
84 Global errors

Proof. The coefficients αk , βk are obtained from solving the system (5.5.13).
In (5.5.19) we have shown that the system is well conditioned. Since the right
hand side is from the local error, which in Corollary 4.5.2 we have shown that
it is second order, then the coefficients will also be second order. 

Corollary 5.5.1 For a mixed problem, if we use linear elements and have equidis-
tribution, then the global error is O(l).

Proof. Since the global error is a sum over the expansion coefficients αk , βk
which are second order, we expect this error to be of first order. 
“A good scientist is a person in whom the childhood quality of peren-
nial curiosity lingers on. Once he gets an answer, he has other ques-
tions. – Frederick Seitz.
86 Global errors
Chapter 6

Local Defect Correction for

BEM

6.1 Introduction

Often boundary value problems have small localised regions of high activity
where the solution varies very rapidly compared to the rest of the domain. This
behaviour is due to boundary conditions or due to an irregular boundary. One
therefore has to use relatively fine meshes to capture the high activity. Since
the activity is localised, one may also choose to solve on a uniform structured
grid. That is, instead of a uniform global grid, the solution is approximated
using several uniform grids with different grid sizes that cover different parts
of the domain. The size of each grid is chosen in agreement with the activity of
the solution in that part of the domain. This refinement strategy is called local
uniform grid refinement [19]. The solution is approximated on a composite grid
which is the union of the various uniform local grids. One way of approximating
this composite grid solution that is simple and less complex is by Local Defect
Correction (LDC).

In LDC, at least one grid, the global coarse grid, covers the entire domain.
Then a uniform local fine grid is used in a small part of the domain containing
the high activity. In [19, 23] LDC has been shown to be a useful way of ap-
proximating the composite grid solution in which a global coarse grid solution
is improved by a local fine grid solution through a process whereby the right
hand side of the global coarse grid problem is corrected by the defect of a lo-
cal fine grid approximation. This method has been well explored for numerical
methods such as finite differences and finite volumes, see [1, 19, 23, 51].
88 Local Defect Correction for BEM

In this chapter we explore potential analogues and develop an LDC strategy

for BEM. The first attempt on LDC for BEM was in [32] and [34] where an
algebraic approach was suggested and studied. Since in BEM we discretise the
boundary, we will be concerned with problems in which the high activity occurs
at the boundary. In Section 6.2 we develop an LDC strategy for BEM alongside
an example. In Section 6.5 we will show that indeed LDC is a fixed point
iterative method and discuss some convergence properties of the algorithm.

6.2 LDC formulation with an introductory exam-

ple: A Neumann problem

Consider the Neumann problem of Example 3.7.4. Let the domain be a unit
square in two dimension, that is, Ω = [0, 1] × [0, 1] and the fixed point rs =
(0.5, −0.02). So we have the problem
 2
 ∇ u(r) = 0, r ∈ Ω := [0, 1] × [0, 1],
(6.2.1)

q(r) = h(r), r ∈ Γ,
where
(r − rs ) · n(r)
h(r) = , rs = (0.5, −0.02). (6.2.2)
||r − rs ||2
The solution in Ω, shown in Figure 6.1, has a small area close to the boundary
where it changes rapidly. As a result, the solution u(r) in the boundary has a
region of high activity in a small part of the boundary, see Figure 6.1.

−0.2

−0.4
u

−0.6

−0.8

−1
0 0.2 0.4 0.6 0.8 1
(x,0)

(a) Solution in Ω with small region of (b) The solution u(r) in part of ∂Ω that
high activity. borders the high activity.

Figure 6.1: Solution in the domain and part of the solution

u(r) at the boundary for Example (3.7.4) on a unit square
with rs in h(r) equal to (0.5, −0.02).

Therefore we can identify a small region inside Ω which contains the high activ-
ity. This region we call the local domain and denote it by Ωlocal , see Figure 6.2.
6.2 LDC formulation with an introductory example: A Neumann problem 89

Its boundary Γlocal , the local boundary, consists of two parts: a part Γactive that is
also part of the global boundary and a part Γinside that is contained in the global
domain Ω, Figure 6.2b. We will call the part Γactive the local active boundary. For
instance in the problem corresponding to the solutions shown in Figure 6.1, the
boundary Γactive , may be identified as Γactive = {(x, y) : y = 0, x ∈ [0.2, 0.8]}. The part
of the global boundary Γ that is outside the active region Γactive will be denoted
Γc , that is, Γc := Γ \Γactive .

Γc

Γinside

Ωlocal

Γactive
(a) Identify local activity area Ωlocal . (b) A local domain Ωlocal

Figure 6.2: An example of a multiscaled solution with

localised high activity in 6.2a and in 6.2b, an illustra-
tion of a local problem domain. The boundary of Ωlocal
is Γlocal := Γactive ∪ Γinside .

The interest in BEM is to compute a numerical approximation of u(r) as ac-

curately as possible. For such kind of multiscaled variations one is faced with
the option of using a global uniform grid with a mesh of relatively small size l
in order to capture the high activity. This would result in very large systems
which are computationally expensive since BEM matrices are full matrices. Be-
sides, outside the local active boundary Γactive , the variation of the solution is
smooth and a relatively coarse grid would suffice. The other option is to use a
uniform structured grid designed to capture the different activities. This would
be a composite grid with a relatively fine mesh of size l in the local active region
and a coarse grid of size L elsewhere.

With LDC we approximate the solution on a composite grid in an iterative way

that involves solving a so called local problem which is a boundary value prob-
lem defined on the local domain. The local problem is solved on a fine mesh
whose size is chosen in agreement with the local activity. The solution on the
local fine grid is combined with the solution on the global coarse grid through
defect correction to obtain a composite grid solution on Γ .

The advantage of this approach is that instead of solving a large composite grid
system, two smaller systems; a global coarse grid system and a local fine grid
system, are solved independently. For problems with various local activities
the local problems can be solved separately in parallel giving a tremendously
90 Local Defect Correction for BEM

cheaper way of obtaining a composite grid solution other than solving directly
on the composite grid.

As introduced in Section 3.1, let Γ be the numerical representation of ∂Ω in

BEM. The global coarse grid Γ L is a uniform mesh of N elements each of size L
covering the whole of Γ , that is,

Γ L := {Γ1L , Γ2L , . . . , ΓN
L
} (6.2.3)

where |ΓjL | = L for all j. The local fine grid Γlocal

l
is a uniform mesh of Nl elements
each of size l covering Γlocal , that is,
l l l l
Γlocal := {Γlocal,1 , Γlocal,2 , . . . , Γlocal,Nlocal
} (6.2.4)

l
where |Γlocal,i | = l for all i. The size of the local fine grid l is chosen in agreement

9 8 7

10 6

11 5

12 4

1 2 3

Figure 6.3: Global coarse and local fine grids.

The small dots are the nodes rllocal of the local
l
fine grid Γlocal and the big circles are the nodes
r of the global coarse grid Γ L . Node 2 belongs
L

to rL ∩ rlactive .

with the activity of the solution in Γactive . Since the solution varies much more
rapidly in Γactive than elsewhere, we expect l to be much smaller than L. Part of
l
the grid Γlocal belongs to Γactive and part belongs to Γinside . The part that belongs
l
to Γactive is denoted Γactive and that that belongs to Γ inside is denoted Γinside
l
. That
is
l l l l
Γactive := {Γactive,1 , Γactive,2 , . . . , Γactive,Nactive
}, (6.2.5a)

l l l l
Γinside := {Γinside,1 , Γinside,2 , . . . , Γinside,Ninside
}, (6.2.5b)

l l l
where Γactive ∪ Γinside = Γlocal and Nactive + Ninside = Nlocal . In constant elements
that we discuss here, the collocation nodes are the midpoints of the elements,
 6.2 LDC formulation with an introductory example: A Neumann problem 91

where the solution is computed. Let us denote the nodes of the coarse grid as
rL ,
rL := {rL1 , rL2 , . . . , rLN }. (6.2.6)
Similarly we denote the nodes of the local fine grid as rllocal ,
rllocal := {rllocal,1 , rllocal,2 , . . . , rlocal,Nlocal }, (6.2.7)
and consist of rlactive and rlinside that are analogously defined.

We assume that all the grid nodes of rL ∩ rlactive belong to rlactive , see Figure 6.3.
The composite grid nodes rl,L are the union rL ∪ rlactive of the global coarse grid
nodes rL and the active local fine grid nodes rlactive . The composite grid Γ l,L
consists of the finest elements that correspond to rl,L .

First we discretise the BIE on Γ L to yield

Step (i)
N Z N Z
1 L X L ∂v X
u + uj (ri ; r(χ)) dχ = q(r(χ))v(ri ; r(χ)) dχ, (6.2.8)
2 i ΓjL ∂n L
j=1 j=1 Γj

which gives the global coarse grid system of equations

AL uL0 = bL . (6.2.9)
Once we have solved (6.2.9), the next step is to use the solution uL0 to formulate
a local problem on Ωlocal . This local problem on Ωlocal satisfies the same oper-
ator as in the global problem. The boundary conditions on Γactive are the same
as those in the global problem that is q(r) = h(r), since Γactive ⊂ Γ . On Γinside we
prescribe an artificial boundary condition g̃(r) defined below. So we have
 2

 ∇ u(r) = 0, r ∈ Ωlocal ,



q(r) = h(r), r ∈ Γactive , (6.2.10)





u(r) = g̃(r), r ∈ Γinside ,
where g̃(r) is a piecewise constant function given by
l
g̃(r) := uinside (ri ), r ∈ Γinside,i ⊂ Γinside , (6.2.11)
and

Step (ii)
N Z
X N
X Z
∂v
uinside (ri ) := q(r(χ))v(ri ; r(χ)) dχ− uLj (ri ; r(χ)) dχ, ri ∈ Γinside .
ΓjL ΓjL ∂n
j=1 j=1
(6.2.12)
 92 Local Defect Correction for BEM

Step (iii) Then a BIE for (6.2.10) on Γlocal is, for r, r(χ) ∈ Γlocal ,
Z Z
1 ∂v ∂v
u(r) + u(r(χ)) (r; r(χ)) dχ + g̃(r(χ)) (r; r(χ)) dχ =
2 Γactive ∂n Γ ∂n
Z Z inside
q(r(χ))v(r; r(χ)) dχ + q(r(χ))v(r; r(χ)) dχ. (6.2.13)
Γactive Γinside

Discretising (6.2.13) on a local fine grid defined in (6.2.4) and (6.2.5) we have
X Z X Z
1 l ∂v ∂v
ulocal,i + ulactive,j (ri ; r(χ)) dχ+ ulinside,j (ri ; r(χ)) dχ =
2 l
Γactive,j ∂n l
Γinside,j ∂n
j j

X Z X Z
qlactive,j v(ri ; r(χ)) dχ + qlinside,j v(r; r(χ)) dχ. (6.2.14)
l
Γactive,j l
Γinside,j
j j

l
In (6.2.14) we have two vectors on Γlocal : ullocal and qllocal , where
 l   l 
uactive qactive
ullocal = , ql
= . (6.2.15)
ulinside local qlinside

The vector ulinside is known through (6.2.12) and the vector qlactive is known
because q(r(χ)) is given on Γactive . So if we repeat (6.2.14) for all the local nodes
we obtain an algebraic system of Nlocal equations. We rearrange the system in
matrix by putting the known quantities on one side to obtain the local problem
system
Allocal xl0local = bl0local (6.2.16)
where
ul0active
 
xl0local = .
ql0inside
The formation of the system (6.2.16) will further be detailed in Section 6.5.

The solution ul0active is expected to be more accurate than the coarse grid so-
lution uL0 in Γactive . The next step of LDC is to use the local fine grid solution
to update the global coarse grid problem. In updating, the right hand side of
the global coarse grid problem is corrected by the defect of the local fine grid
approximation, we will call this step the defect correction step. The two approx-
imations are then used to define a composite grid approximation of u(r).

The question now is: how do we compute the defect? Consider the coarse grid
discretisation (6.2.8). If we knew the exact continuous function u(r) and hence
the exact solution uj := u(rj ) in the nodes we would use it in (6.2.8) to obtain
X N Z X N Z
1 ∂v
ui + uj (ri ; r(χ)) dχ = q(r(χ))v(ri ; r(χ)) dχ + dLi . (6.2.17)
2 ΓjL ∂n L
j=1 j=1 Γj
 6.2 LDC formulation with an introductory example: A Neumann problem 93

where di is the local defect for the i-th equation introduced in (4.2.9). We also
have the exact BIE as
N Z
X N Z
X
1 ∂v
ui + u(r(χ)) (ri ; r(χ)) dχ = q(r(χ))v(ri ; r(χ)) dχ. (6.2.18)
2 ∂Ωj ∂n ∂Ωj
j=1 j=1

Subtracting (6.2.18) from (6.2.17) gives

N
X Z X N Z
∂v ∂v
uj (ri ; r(χ)) dχ − u(r(χ)) (ri ; r(χ)) dχ = dLi . (6.2.19)
ΓjL ∂n ∂n
j=1 j=1 ∂Ωj

From (6.2.19) we define the local defect per element j as

Z Z
∂v ∂v
dLij := uj (ri ; r(χ)) dχ − u(r(χ)) (ri ; r(χ)) dχ, (6.2.20)
L
Γj ∂n ∂Ωj ∂n

so that the total defect at ri is given by

X
dLi := dLij , i = 1, 2, . . . , N. (6.2.21)
j

Therefore if we would know the exact continuous function u(r) we could com-
pute the local defect dLi , add it to the right hand side of (6.2.8) and solve for the
exact solution uj on each element. However u(r) is not known and therefore
we cannot compute the defect using (6.2.20). All we can do is estimate dLij as
accurately as possible using the best solution available, which is

Step (iv)
 L
 uj , ΓjL ⊂ Γc ,
uLbest,j = (6.2.22)

ulactive,j , ΓjL ⊂ Γactive .

So for elements in the high activity region we have the fine grid solution which
we can use to estimate the local defect as follows.

Let us consider the case of a square where ΓjL ≡ ∂Ωj . Suppose that in the local
l
fine grid Γactive a global coarse grid element ΓjL is divided into k fine elements
Γactive,jk such that ΓjL = ∪ Γactive,j
l l
k
, see an illustration in Figure 6.4 for k = 3.
k

Then the best approximations of the integrals in (6.2.20) are

Z Z
∂v ∂v
uj (ri ; r(χ)) dχ ≈ ulactive,j (ri ; r(χ)) dχ, (6.2.23a)
ΓjL ∂n ΓjL ∂n

Z X Z
∂v ∂v
u(r(χ)) (ri ; r(χ)) dχ ≈ ulactive,jk (ri ; r(χ)) dχ. (6.2.23b)
ΓjL ∂n ∂n
k Γactive,jk
 94 Local Defect Correction for BEM

l l
ΓjL Γactive,j1
l
Γactive,j Γactive,j3
2

rLj rlactive,j1 rlactive,j2 rlactive,j3

a coarse element ΓjL a refined coarse element ΓjL

Figure 6.4: A coarse element that is refined into three elements in the local fine
3
grid ΓjL = ∪ Γactive,j
l
k
.
k=1

Step (v) Therefore we have the following best approximation of the defect per ele-
ment
Z X Z
L l ∂v l ∂v
dij ≈ uactive,j (ri ; r(χ)) dχ− uactive,jk (ri ; r(χ)) dχ, (6.2.24)
Γj ∂n ∂n
k Γactive,jk

for ΓjL ⊂ Γactive and

dLij ≈ 0
for ΓjL ⊂ Γc . We can then compute the defect
X
dLi ≈ dLij , ΓjL ⊂ Γactive , for all i = 1, 2, . . . , N. (6.2.25)
j

By default integration in the BIE is global. Each node of the global coarse
grid communicates with the active region through integration. So although the
activity is local, its effect is global. The defect dLi is therefore computed for all
nodes of the global coarse grid.

Step (vi) The next step now is the updating step. The global coarse grid discreti-
sation is updated with the defect of the local fine grid solution. So we
have
AL uL1 = bL + dL . (6.2.26)
Solving (6.2.26) gives the updated coarse grid solution uL1 .

l
At this stage we use the fine grid solution on Γactive and the global coarse grid
l,L
solution to form a composite grid solution u as
 l
 u0active (r), r ∈ Γactive ,
ul,L
0,1 (r) = (6.2.27)
 L
u1 (r), r ∈ Γc .
The composite grid solution (6.2.27) can now be used to compute better bound-
ary conditions on Γinside and then form and solve the updated fine grid problem
Allocal xl1local = bl1local . (6.2.28)
 6.3 LDC formulation: A Dirichlet problem 95

Step (vii) Thus we obtain the updated composite grid solution given by
 l
 u1active (r), r ∈ Γactive ,
ul,L
1,1 (r) = (6.2.29)
 L
u1 (r), r ∈ Γc .

In Figures 6.5 and 6.6 we have the results at each of the above stages of lo-
cal defect correction for the problem (6.2.1) with boundary conditions (6.2.2).
Figure 6.7 shows how fast the global error converges. Basically the algorithm
has converged already in the first iteration since the error reduction between
successive iterations after the first one is small compared to that in the first
iteration.

initial coarse grid solution initial fine grid solution

0 0

−0.2 −0.2

−0.4 −0.4
u

−0.6 −0.6

−0.8 −0.8

−1 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(x,0) (x,0)
(a) Initial coarse grid solution (b) Initial fine grid solution

Figure 6.5: Results of a typical LDC process for a Neumann problem

in one iteration.

The process of defect correction can be repeated generating a composite grid

solution at the end of Step (vii) of each cycle. In the next section we show that
this process is a fixed point iterative process for computing the solution on a
composite grid. The solution is obtained by solving separate global coarse and
local fine grid problems.

6.3 LDC formulation: A Dirichlet problem

The process of LDC formulation for a Dirichlet problem is completely the same
as that for a Neumann problem. The only differences will be in the operators
involved. The global operator will be a single layer operator. In the case of a
Neumann problem, the local problem is a mixed problem with q given on Γactive
and u on Γinside prescribed through the use of the global solution. For a Dirichlet
96 Local Defect Correction for BEM

updated coarse grid solution updated fine grid solution

0 0

−0.2 −0.2

−0.4 −0.4
u

u
−0.6 −0.6

−0.8 −0.8

−1 −1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(x,0) (x,0)
(a) Updated coarse grid solution (b) Updated fine grid solution

Figure 6.6: Results of a typical LDC process for a Neumann problem

in one iteration.

Coarse grid solution errors

l=0.2/9
||u∗ − uL ||∞

−1
10

0 5 10
iterations

Figure 6.7: Convergence of the global coarse grid error.

6.4 Complexity of the algorithm 97

problem, the local problem will be a completely Dirichlet with u given on Γactive
and u on Γinside prescribed through the use of the global solution.

6.4 Complexity of the algorithm

In brief, the LDC iterative process can be summarised as follows:

Algorithm 6.1
LDC iteration summary

(I) Solve the global coarse grid problem on Γ .

(II) Compute u on Γinside .
(III) Solve a fine grid problem on Γlocal .
(IV) Update the global coarse grid problem.

Suppose we have p locally active small regions and thus p local problems. Let,
for each local problem, Ml be the number of elements Γlocal and Min the number
of elements on Γinside . Then the size of the local problem is M = Ml + Min . Let
Min be so small that M ≈ Ml . Let N be the size of the global problem and
NLlocal the number of global elements in Γlocal . We assume Γlocal is such a small
part of the global boundary that N − NLlocal ≈ N. Then the equivalent size of the
composite grid would be pM + N. The operational count for LU-decomposition
is N3 /3 for a size N matrix. So the complexity of the equivalent composite grid
problem would be

1 (p + 1)3 3
(pM + N)3 ≈ N if M ≈ N. (6.4.1)
3 3
The BEM-LDC algorithm converges in one step which involves solving two
coarse grid problems and p local problems and so has total complexity

1 p N3
2 · N3 + M 3 ≈ (2 + p). (6.4.2)
3 3 3
So when we compare (6.4.1) with (6.4.2) we see that the composite problem is
(p + 1)3 /(2 + p) times more expensive than BEM-LDC. Suppose instead we were
to refine globally to a grid of size equal to that of the local problems. Then if
the refinement ratio is say α, that is, L/l = α, the resulting problem would be
1
of complexity α3 N3 . So the resulting problem would be α3 /(p + 2) times more
3
complex than using LDC. For instance in the modes case of α = 2, this factor
is more than one for up to p = 5 local problems. Thus BEM-LDC is cheaper
98 Local Defect Correction for BEM

than both of its obvious alternatives of either composite gridding or refining

uniformly.

Another advantage of LDC over solving on direct composite or fine uniform

grids is the memory required. LDC requires less memory than the equivalent
composite or uniform grid problems. This is because instead of handling large
matrices and vectors of sizes say (M + N) or αN, it handles smaller vectors of
sizes N and M at a time.
6.5 LDC algorithm as a fixed point iteration 99

6.5 LDC algorithm as a fixed point iteration

In BEM, Dirichlet and Neumann problems have different properties because

they result in different integral operators. However the LDC process is the same
for both the Neumann and Dirichlet problems, with the only differences being
in the operators involved. In this chapter we consider the Neumann problem
and show that the LDC algorithm is a fixed point iterative process. The case of
a Dirichlet problem is developed along the same lines.

In general the LDC process can be summarised in the following algorithm:

Algorithm 6.2
LDC algorithm for BEM
Initialisation

(1) Solve a global coarse grid problem

AL xL0 = bL on Γ L .

(2) Compute the potential u(r) on Γinside using xL0 and the boundary
integral relation for a point inside Ω.
(3) Solve a local fine grid problem

Allocal xl0 local = bl0 local on Γlocal

l
.

Iterations i=1,2,...

(4) Compute the defect dLi−1 .

(5) Solve the global coarse grid problem.

AL xLi = bL + dLi−1 on Γ L

(6) Assemble the partly updated composite grid solution xl,L

i−1,i .

(7) Compute the potential u(r) on Γinside using xl,L

i−1,i and the bound-
ary integral relation for points inside Ω.
(8) Solve the local fine grid problem

Allocal xlilocal = blilocal on Γlocal

l
.

(9) Assemble the updated composite grid solution xl,L

i,i .

In the sequel we will formulate the above algorithm as a fixed point iteration
for a Neumann problem.
100 Local Defect Correction for BEM

Consider the Neumann problem

 2
 ∇ u(r) = 0, r ∈ Ω,
(6.5.1)

q(r) = h(r), r ∈ ∂Ω.

To formulate the LDC algorithm as a fixed point iteration, we need a vector

formulation for the steps in Algorithm 6.2. The first step of the Algorithm 6.2
is to solve a global coarse grid problem

AL xL0 = bL , (6.5.2)
for an initial solution xL0 and this involves “inversion” of the coefficient matrix
AL . The coefficient matrix for a Neumann problem is singular. In order to obtain
a unique solution we prescribe u(r) at a point at the boundary as explained in
Section 2.5.2. In particular, we prescribe Dirichlet boundary conditions in the
last node of the grid. So we use xL0 to denote the solution vector in (6.5.2) which
is a vector of u’s except the last entry which is a q. Using uL0 , the boundary
conditions and the boundary integral relation (2.4.1), we compute the potential
u(r) on Γinside . That is,
Z X Z
∂v
uinside (ri ) = v(ri ; r(χ))q(r(χ))dχ − uL0 j (ri ; r(χ))dχ, ri ∈ rlinside . (6.5.3)
∂n
Γ j Γj

L
Let us introduce a vector g and a matrix H̃ such that
Z
gi := q(r(χ))v(ri ; r(χ)) dχ, ri ∈ Γinside , (6.5.4)
Γ

Z
∂v
H̃Lij := (ri ; r(χ)) dχ, ri ∈ Γinside . (6.5.5)
ΓjL ∂n

Then we can write (6.5.3) as

L
u0 inside = g − H̃ uL0 . (6.5.6)
Using (6.5.6) we obtain Dirichlet boundary conditions on Γinside . The boundary
conditions on Γactive are the given Neumann boundary conditions since Γactive ⊂
l
Γ . Using (3.2.10) we can then write the equations on Γlocal in vector form as
 l   
u0active qactive
Hllocal = Gllocal , (6.5.7)
ul0 inside ql0inside

where Hllocal and Gllocal are matrices on Γlocal l

, ulactive and qlactive are vectors on
l l l l
Γactive and u0inside and q0inside are vectors on Γinside . The vector ul0inside is known
through (6.5.6) and the vector qactive is known through the boundary conditions.
So we rearrange (6.5.7) as
 l   
uactive qactive
[Hlactive − Glinside ] = [G l
− Hl
] (6.5.8)
ql0inside active inside ul0inside
6.5 LDC algorithm as a fixed point iteration 101

The matrix Hlactive is a block of Hl for which the column index corresponds to
l
nodes in Γactive . Similarly Hlinside is a block of Hl for which the column index
l
corresponds to nodes in Γinside . The blocks Glactive and Glinside are defined analo-
gously. The quantities on the right hand side of (6.5.8) are all known. Let
Bllocal := [Glactive − Hlinside ], (6.5.9)

Allocal := [Hlactive − Glinside ], (6.5.10)

 
qactive
bl0local := Bllocal , (6.5.11)
ul0 inside

ul0active
 
xl0local := . (6.5.12)
ql0inside
Then we have
Allocal xl0local = bl0local . (6.5.13)

The solution of (6.5.13) gives us another solution ul0active in Γactive which should
be a better approximation of u(r) than uL0 in Γactive because of the fine grid used.
Next we use this solution to compute the defect and update the global coarse
grid solution. Using (6.2.24), the defect on an element ΓjL when the collocation
node is i is given by
X Z Z
∂v ∂v
d0 ij ≈ ul0active,jk (ri ; r(χ)) dχ − ul0active,j (ri ; r(χ)) dχ, (6.5.14)
∂n ΓjL ∂n
k l
Γactive,j
k

l
where ∪ Γactive,jk
= ΓjL , see Figure 6.4. The integration in (6.5.14) is computed
k
at all the global elements that lie in Γactive so that the total defect for the i-th
collocation node is
X
d0 i = d0 ij . (6.5.15)
j,ΓjL ⊂Γactive

Since each collocation node communicates with the local active region through
integration, the defect d0i is computed for all the collocation nodes.

Let us introduce a matrix H̄ as

Z
∂v
H̄ik = (ri ; r(χ)) dχ, ri ∈ rL , Γk ∈ Γactive
l
.
∂n
Γk

Let PL,l be a restriction from the fine grid Γactive

l L
to the coarse grid Γactive in Γactive .
Then we can write the defect d0 as
L L
^ PL,l ul0active = (H̄ − H
d0 = H̄ul0active − H ^ PL,l )ul0active , (6.5.16)
102 Local Defect Correction for BEM

where the matrix H ^ is as defined in (3.2.5) with the superscript L indicating on

a coarse grid of size L. Now we have the defect for all the coarse grid nodes. We
update the coarse grid system (6.5.2) to obtain the updated coarse grid solution
xL1 . That is

^ L PL,l )ul0active .
AL xL1 = bL − (H̄ − H (6.5.17)

At this stage we can assemble a composite grid solution on Γ l,L that consists of
the initial fine grid solution and the updated coarse grid solution. So
 l 
u0 active
ul,L := , (6.5.18)
0,1 uL1 c

where uL1 c is the updated coarse grid solution on Γc outside the active region
Γactive . To complete the updated composite grid solution, we need to solve a new
local problem. To this end we use the solution in (6.5.18) to compute another
approximation of u(r) on Γinside . Thus we have
l,L l,L
u1inside = g − H̃ u0,1 , (6.5.19)
l,L
where the matrix H̃ is as defined in (6.5.6) but on the composite grid, that is,

 R ∂v

 L (ri ; r(χ)) dχ, r(χ) ∈ Γc ,
 Γc,j ∂n
l,L
H̃ij = (6.5.20)

 R ∂v

Γactive,j ∂n (ri ; r(χ)) dχ, r(χ) ∈ Γactive ,
 l

and ri ∈ Γinside . Then we formulate an updated local problem system

Allocal xl1local = bl1local , (6.5.21a)

where
qlactive
 
bl1 = Bllocal . (6.5.21b)
u1 inside

Solving (6.5.21) gives us an updated solution ul1active of u(r) on Γactive . At this

stage we have a completely updated composite grid solution given by
 l 
l,L u1 active
u1,1 = . (6.5.22)
uL1 c

This completes the first iteration that gives us the first updated composite grid
solution. The process can be repeated till there is no more change in the solu-
tion. In what follows we formulate the above process as a fixed point iterative
process.

Let Ilactive be an identity of size Nactive . Then the part of the local solution in
Γactive is given by, for the i-th iteration,

uliactive = [Iactive O]xlilocal . (6.5.23)

6.5 LDC algorithm as a fixed point iteration 103

Consider the updated composite grid solution (6.5.22) for iteration i + 1. Us-
ing (6.5.21) and (6.5.23), we have
  l  
 l  l−1 l qactive
ui+1 active [Iactive O]A B
local local uli+1 inside
ul,L
 
i+1,i+1 =
 =  . (6.5.24)
L
 
ui+1 c L
[O Ic ] ui+1

From the second block row of (6.5.24) we have

uLi+1 c = [O Ic ] uLi+1 .

This is the global coarse grid solution outside the active region. For a Neumann
problem we prescribe Dirichlet boundary conditions in the last node in order
to obtain a unique solution. Thus the last value of the solution vector will be a
q value. So in general let us write

xLi+1 c = [O Ic ] xLi+1 . (6.5.25)

Using (6.5.17) we have

−1
 
xLi+1 c = [O Ic ]AL ^ L PL,l )uliactive  ,
bL − (H̄ − H

−1
 
= xL0 c − [O Ic ]AL ^ L PL,l  uliactive ,
H̄ − H (6.5.26)

where
−1 L
xL0 = AL b .

Let us introduce a matrix M as

M := [Iactive O]Allocal Bllocal .

−1

Then from the first block row of (6.5.24) we have

 l 
l qactive
ui+1 active = M (6.5.27)
uli+1 inside

Note that the matrix M is rectangular in size. Let us break it into two blocks: a
l
square block Mactive that operates on Γactive and a block Minside that operates on
l
Γinside . Then we can write (6.5.27) as

qlactive
 
uli+1 active = [Mactive Minside ] = Mactive qlactive +Minside uli+1 inside . (6.5.28)
uli+1 inside

From (6.5.19) we see that

l,L
ui+1 inside = g − H̃ ul,L
i,i+1 . (6.5.29)
104 Local Defect Correction for BEM

l,L
Let us also break down the operator H̃ in (6.5.29) into a part that operates
l
on Γactive and another that operates on ΓcL , so that we can write (6.5.29) as
l,L l,L
ui+1 inside = g − H̃active uliactive − H̃c uLi+1 c . (6.5.30)

Using (6.5.30) in (6.5.28) we have

 
l,L l,L
uli+1 active = Mactive qlactive + Minside g − H̃active uliactive − H̃c uLi+1 c  ,

l,L l,L
= Mactive qlactive + Minside g − Minside H̃active uliactive − Minside H̃c uLi+1 c .
(6.5.31)

We introduce the following operators

l,L
R := −Minside H̃active ,

l,L
T := −Minside H̃c ,

so that we can write (6.5.31) as

uli+1 active = Mactive qlactive + Minside g + Ruliactive + TuLi+1 c . (6.5.32)

l
In (6.5.32) the updated solution on the active grid Γactive is expressed in terms
of the previous solution there and the updated solution outside the active grid.
To have an expression for the iteration that takes place on the active region
alone, we use (6.5.26) to replace uLi+1,c . So,

uLi+1 c = D1 xLi+1 c + D2 bc , (6.5.33)

where bc is the vector of boundary conditions outside the active region. Because
the last entry of xLi+1 c is a q-value and that of bc is a u-value, the matrices D1
and D2 are the projections
   
I 0 O 0
D1 := , D 2 := . (6.5.34)
0T 0 0T 1

So we have
−1
 
uLi+1 c = D1 (xL0 c − [O Ic ]AL ^ L PL,l  uliactive ) + D2 bc .
H̄ − H (6.5.35)

Let us introduce the following notation

−1 L
W := [O Ic ]AL ^ PL,l )
(H̄ − H (6.5.36)

so that we can write(6.5.35) as

uLi+1 c = D1 xL0 c − D1 Wuliactive + D2 bc . (6.5.37)

6.5 LDC algorithm as a fixed point iteration 105

Now we use (6.5.37) in (6.5.32) to obtain

uli+1 active = Mactive qlactive + Minside g + Ruli active + T(D1 xL0 c − D1 Wuli active + D2 bc )
which can be rearranged as
uli+1 active = (R−TD1 W)uliactive +TD1 xL0 c +Mactive qlactive +Minside g+TD2 bc . (6.5.38)
We introduce a vector v as
v := TD1 xL0 c + Mactive qlactive + Minside g + TD2 bc .
We can now write equation (6.5.38) as
uli+1 active = (R − TD1 W)uliactive + v. (6.5.39)
The vector v remains fixed throughout the iteration since qlactive , g, bc , remain
−1
fixed and xL0 c = [O Ic ]AL bL remains the same throughout the iteration. Equa-
l
tion (6.5.39) expresses the iteration that takes place on the fine grid Γactive as a
fixed point iteration with iteration matrix Q defined as
Q := R − TD1 W. (6.5.40)
Thus we have
uli+1 active = Quliactive + v. (6.5.41)
This iteration will converge if the spectral radius of the iteration matrix Q is less
than unity. In Tables 6.1 and 6.2 we have the spectral radii of Q for different
combinations of L and l.
l
L 0.2 0.2/3 0.2/9 0.2/27 0.2/81
0.2 2.137E-01 2.308E-01 2.402E-01 2.457E-01 2.492E-01
0.2/3 2.253E-01 2.342E-01 2.399E-01 2.436E-01
0.2/9 2.316E-01 2.367E-01 2.401E-01
0.2/27 2.357E-01 2.388E-01
0.2/81 2.385E-01

Table 6.1: Spectral radius of the iteration matrix Q for a Neumann problem for
different combinations of fine and coarse grid sizes l and L respectively. The
local problem domain is the rectangle [0.2, 0.8] × [0, 0.4] and Ω = [0, 1] × [0, 1].

Recall again the problem of Example 3.7.4 on page 31 with Ω = [0, 1] × [0, 1]
and rs = (0.5, −0.02) as summarised in (6.2.1). We identify Ωlocal as Ωlocal :=
[0.2, 0.8] × [0, 0.4], see Figure 6.8. LDC is then used with various sizes of coarse
grid size L and fine grid size l. We also expect the ratios ||uli+1 − uli ||/||(uli − uli−1 )||
to be a measure of the convergence speed. In Table 6.3 we have computed
these ratios for five iterations and different combinations of grid sizes. The
results in Tables 6.1, 6.2 and 6.3 all indicate that the spectral radius of the
LDC algorithm iteration matrix is smaller than one and therefore we expect the
algorithm to converge.
106 Local Defect Correction for BEM

l
L 0.2 0.2/3 0.2/9 0.2/27 0.2/81
0.2 1.756E-01 2.165E-01 2.338E-01 2.426E-01 2.476E-01
0.2/3 2.086E-01 2.255E-01 2.349E-01 2.405E-01
0.2/9 2.219E-01 2.306E-01 2.361E-01
0.2/81 2.341E-01

Table 6.2: Spectral radius of the iteration matrix Q for a Neumann problem for
different combinations of grid sizes L and l. The local problem domain is the
rectangle [0.4, 0.6] × [0, 0.2] and Ω = [0, 1] × [0, 1].

1
0

−0.2

−0.4
u

−0.6 0.4

−0.8
Ωlocal
−1
0 0.2 0.4 0.6 0.8 1
(x,0) 0 0.2 0.8 1
(a) The solution u(r) in part of ∂Ω (b) Domain demarcations for LDC

Figure 6.8: Part of the solution at the boundary and the

LDC domain demarcations.

L l i ρi L l i ρi
0.2 0.2/3 1 0.1574 0.2 0.2/9 1 0.2024
2 0.1848 2 0.2228
3 0.1883 3 0.2295
4 0.1886 4 0.2312
5 0.1886 5 0.2316

L l i ρi L l i ρi
0.2 0.2/27 1 0.2177 0.2 0.2/81 1 0.2226
2 0.2334 2 0.2372
3 0.2393 3 0.2428
4 0.2416 4 0.2452
5 0.2425 5 0.2463

Table 6.3: The ratios ρi = ||uli+1 − uli ||2 /||(uli − uli−1 )||2 when we use LDC to solve
the problem in (6.2.1).
 6.6 Continuous formulation of the LDC steps 107

6.6 Continuous formulation of the LDC steps

The integral operators Ks and Kd have somewhat been extensively studied in

continuous form. We can make use of this available information to draw an
analysis for our LDC algorithm if we can have its equivalent formulation in
continuous form. That is the focus of this section.

Basically we would like to study the error convergence during the iteration pro-
cess. So we develop a continuous formulation of the LDC algorithm in terms
of the error. For this purpose we need the equations which the initial and con-
verged solutions satisfy. The Algorithm 6.2 can be summarised in the following
six steps:

Step (i) Initialisation: solve the global coarse grid problem

X N Z
1 L ∂v
u0 i + uL0 j (ri ; r(χ)) dχ = bLi . (6.6.1)
2 ΓjL ∂n
j=1

l l l
Step (ii) After discretisation of Γlocal into Γlocal = Γinside ∪ Γactive , see (6.2.4), (6.2.5) and
l
Figure (6.3), compute u(r) on Γinside using the integral relation

Z N
X Z
∂v
ul0inside,i = v(ri ; r(χ))q(r(χ)) dχ − uL0 j (ri ; r(χ)) dχ, ri ∈ rlinside .
Γ ΓjL ∂n
j=1
(6.6.2)

This gives a vector ulinside of u’s on Γinside as introduced in (6.2.15).

Step (iii) Now we have Dirichlet boundary conditions on Γinside . On Γactive q(r) is
known. Next is to solve the local problem:

X Z
1 l l ∂v
u + u0active,j (ri ; r(χ)) dχ+
2 0local,i l
Γactive,j ∂n
j
X Z
l ∂v
u0inside,j (ri ; r(χ)) dχ =
l
Γinside,j ∂n
j

Z X Z
v(ri ; r(χ))qactive (r(χ)) dχ + ql0inside,j v(ri ; r(χ)) dχ.
Γactive l
Γinside,j
j
(6.6.3)

Here qactive (r) is known from the boundary conditions of the global prob-
lem since Γactive ⊂ Γ . Also ulinside is known through (6.6.2). Solving (6.6.3)
gives ql0inside and ul0active .
 108 Local Defect Correction for BEM

Step (iv) We need to compute the defect. First we need the best solution available,
which is:
 L
 u0 c,j , ΓjL ⊂ Γc ,
uL0 best,j = (6.6.4)
 l
u0active,j , ΓjL ⊂ Γactive .

Step (v) Using the fine grid solution, we estimate the defect per element dL0 ij in
(6.2.24) as
 R ∂v

 ul0active,j Γ L (ri ; r(χ)) dχ−

 j ∂n

 R
 P l ∂v
. u0 active,jk (ri ; r(χ)) dχ, ΓjL ⊂ Γactive ,
dL0 ij = k l ∂n
 Γactive,j

 k




0 Γj ⊂ Γc .
(6.6.5a)

Then the defect dLi for each equation is estimated as

X
dL0 i = dL0 ij
j
 
X
 Z X Z 
∂v ∂v

.
 
l
ul0active,jk
 
= 
 u0active,j (ri ; r(χ)) dχ − (ri ; r(χ)) dχ

L ∂n ∂n
 
Γ
 
j  k
 j

l

Γactive,j
k

(6.6.5b)

l
for all ri . Here ∪ Γactive,jk
= ΓjL for ΓjL ⊂ Γactive , see Figure 6.4.
k

Step (vi) Defect correction: add the defect to the right hand side of the coarse grid
equation and solve for the updated global coarse grid solution uL1 . That is

X N Z
1 L ∂v
u1,i + uL1,j (ri ; r(χ)) dχ = bLi + dL0 i . (6.6.6)
2 ΓjL ∂n
j=1

Step (vii) Assemble the composite grid solution and go back to step (ii), replace uL0
with uL1 and continue.

We assume that the algorithm (i) to (vi) converges. Let the unknowns in steps (ii)
to (vi) converge to their respective fixed points denoted by an asterisk, namely
u∗ L , u∗ linside , u∗ lactive and q∗ linside . The fixed point should satisfy the equations
in (6.6.2) to (6.6.6). So we have:
 6.6 Continuous formulation of the LDC steps 109

Step (ii) Computing u(r) on Γinside :

Z N Z
X ∂v
u∗inside,i = v(ri ; r(χ))q(r(χ)) dχ − u∗j L (ri ; r(χ)) dχ, ri ∈ rlinside .
Γ ΓjL ∂n
j=1
(6.6.7)

Step (iii) The local problem equations for ulactive and qlinside :
X l Z
1 ∗l ∂v
u local,i + u∗ active,j (ri ; r(χ)) dχ
2 Γactive,j ∂n
j
X l Z
∂v
+ u∗ inside,j (ri ; r(χ)) dχ =
Γinside,j ∂n
j
Z X l Z
v(ri ; r(χ))qactive (r(χ)) dχ + q∗ inside,j v(ri ; r(χ)) dχ.
Γactive j Γinside,j

(6.6.8)

Step (iv) Computing the defect; Best solution available:

 ∗L
 u c,j , ΓjL ⊂ Γc ,
∗L
u best,j = (6.6.9)
 ∗l
u active,j , ΓjL ⊂ Γactive .

Step (v) Fixed point for the defect:

X
 Z
∗ L ∗l ∂v
di ≈ u active,j (ri ; r(χ)) dχ −


 L
Γj ∂n
j

X Z 
∂v


u∗ lactive,jk

(ri ; r(χ)) dχ
 (6.6.10a)
∂n 

k

l

Γactive,j
k

for all i, ΓjL ⊂ Γactive .

Step (vi) The fixed point for the updated global coarse grid equations:
N Z
1 ∗L X ∗L ∂v
u + uj (ri ; r(χ)) dχ = bLi + d∗ Li . (6.6.11)
2 i L
Γj ∂n
j=1

Let us define the errors:

δuLj := u∗j L − uLj ,
δulactive,j := u∗ lactive,j − ulactive,j ,
δulinside,j := u∗ linside,j − ulinside,j ,
δqlinside,j := q∗ linside,j − qlinside,j ,
δdLi := d∗i L − dLi .
 110 Local Defect Correction for BEM

Then subtracting equations (6.6.2) to (6.6.6) from (6.6.7) to (6.6.11) yields the
following equations for the errors:

Step (ii) In computing u on Γinside :

N
X Z
∂v
δulinside,i = − δuLj (ri ; r(χ)) dχ, ri ∈ rlinside . (6.6.13)
ΓjL ∂n
j=1

Step (iii) In solving the local problem:

Z
1 l X l ∂v
δu + δuactive,j (ri ; r(χ)) dχ
2 i l
Γactive,j ∂n
j
X Z
∂v
+ δulinside,j (ri ; r(χ)) dχ =
l
Γinside,j ∂n
j
X Z
l
δqinside,j v(ri ; r(χ)) dχ. (6.6.14)
l
Γinside,j
j

Step (iv) In the best solution available:


 δuLc,j , ΓjL ⊂ Γc ,
δuLbest,j = (6.6.15)

δulactive,j , ΓjL ⊂ Γactive .

Step (v) In computing the defect:


X Z
L l ∂v
δdi ≈ δuactive,j (ri ; r(χ)) dχ −


∂n

j ΓjL

X Z 
∂v


δulactive,jk

(ri ; r(χ)) dχ
 . (6.6.16)
∂n 

k

l

Γactive,j
k

Step (vi) In updating the global solution:

N Z
1 L X L ∂v
δu + δuj (ri ; r(χ)) dχ = δdLi . (6.6.17)
2 i ΓjL ∂n
j=1
6.6 Continuous formulation of the LDC steps 111

Like we did in investigating global errors in Chapter 5, we can also model the
above error equations in continuous form. Then we can use the available prop-
erties of the operators to derive some convergence properties for the LDC algo-
rithm.

Consider the exact BIE on Γ

Z Z
1 ∂v
u(r) + u(r(χ)) (r; r(χ)) dχ = q(r(χ))v(r; r(χ)) dχ (6.6.18)
2 Γ ∂n Γ

Let u0 (r) be a BEM solution of (6.6.18) on Γ with error δu0 (r). That is
Z Z
1 ∂v
u0 (r) + u0 (r(χ)) (r; r(χ)) dΓ = q(r(χ))v(r; r(χ)) dΓ + f0 (r).
2 Γ ∂n Γ

where f0 (r) is a source term incorporating discretisation errors. Since

u(r) = u0 (r) + δu0 (r), (6.6.19)

we can write
Z Z
1 ∂v
(u0 (r) + δu0 (r)) + (u0 (r(χ)) + δu0 (r(χ))) (r; r(χ)) dχ = q(r(χ))v(r; r(χ)) dχ.
2 Γ ∂n Γ
(6.6.20)

So the error term f0 (r) is given by

Z
1 ∂v
f0 (r) := δu0 (r) + δu0 (r(χ)) (r; r(χ)) dχ, r, r(χ) ∈ Γ. (6.6.21)
2 Γ ∂n

Theorem 6.6.1 For points on Γ in , if we eliminate the error in integration, the

error in u is bounded by the error in u on Γ .

Proof. Let u0 (r) be the global solution on Γ . Let uinside (r) be u on Γ in computed
using the solution u0 (r). Then
Z Z
∂v
uinside (r) = q(r(χ))v(r; r(χ)) dχ − u0 (r(χ)) (r; r(χ)) dχ, r ∈ Γ in , r(χ) ∈ Γ.
Γ Γ ∂n
(6.6.22)

If δuinside (r) is the error in uinside (r) and δu0 (r) is the error in u0 (r), then
Z Z
∂v
uinside (r)+δuinside (r) = q(r)v(r; r(χ)) dχ− (u0 (r(χ))+δu0 (r(χ))) (r; r(χ)) dχ.
Γ Γ ∂n
(6.6.23)

So the error δuinside (r) is given by

Z
∂v
δuinside (r) = − δu0 (r(χ)) (r; r(χ)) dχ, r ∈ Γ in , r(χ) ∈ Γ. (6.6.24)
Γ ∂n
112 Local Defect Correction for BEM

In (6.6.24) we have a continuous equivalent of (6.6.13). Taking the maximum

norm (6.6.24) gives
Z
∂v
||δuinside (r)||∞ = max δu0 (r(χ)) (r; r(χ)) dχ
Γ ∂n

Z
∂v
≤ ||δu0 (r(χ))||∞ || (r; r(χ))||∞ dχ,
Γ ∂n

≤ ||δu0 (r(χ))||∞ , (6.6.25)

since ∂v/∂n does not change sign over Γ and the integral of ∂v/∂n over Γ is 1 for
source points inside the domain. 

Now, let ulocal and qlocal be functions on Γlocal . Let u0 local and qlocal be the initial
BEM solutions on Γlocal . The functions u0local and qlocal can also be divided into
u0 active and u0 inside and q0 active and q0 inside respectively. Then we have
Z Z
1 ∂v ∂v
u0 local (r)+ u0 active (r(χ)) (r; r(χ)) dχ+ u0 inside (r(χ)) (r; r(χ)) dχ =
2 ∂n ∂n
Γactive Γinside
Z Z
q active (r(χ))v(r; r(χ)) dχ + q0 inside (r(χ))v(r; r(χ)) dχ; r, r(χ) ∈ Γlocal .
Γactive Γinside
(6.6.26)
Again using δf to denote a perturbation of a function f, we have
1  Z 
u0 active (r(χ)) + δu0 active (r(χ)) ∂v (r; r(χ)) dχ+

u0 local (r) + δu0 local (r)+
2 ∂n
Γactive
Z  Z
u0 inside (r(χ)) + δu0 inside (r(χ)) ∂v (r; r(χ)) dχ =

q active (r(χ))v(r; r(χ)) dχ+
∂n
Γinside Γactive
Z  
q0 inside (r(χ)) + δq0 inside (r(χ)) v(r; r(χ)) dχ; r, r(χ) ∈ Γlocal . (6.6.27)
Γinside

Comparing (6.6.27) with (6.6.26) gives a BIE for the error:

Z Z
1 ∂v ∂v
δu0 local (r)+ δu0 active (r(χ)) (r; r(χ)) dχ+ δu0 inside (r(χ)) (r; r(χ)) dχ =
2 ∂n ∂n
Γactive Γinside
Z
δq0 inside (r(χ))v(r; r(χ)) dχ; r, r(χ) ∈ Γlocal . (6.6.28)
Γinside

Theorem 6.6.2 The fine grid solution error δu0 active (r) on Γactive is bounded by
the error committed on Γinside , that is,
||δu0 active (r(χ))||∞ (Γactive ) ≤ ||δu0 inside (r(χ))||∞ (Γinside ) . (6.6.29)
6.6 Continuous formulation of the LDC steps 113

Proof. Equation (6.6.28) is a BIE for an error function δu0 local (r) on Γlocal , the
boundary of Ωlocal , with homogeneous Neumann boundary conditions on Γactive .
According to the weak maximum principle, the maximum of δu0 local (r) occurs
at the boundary. That is, there is an r0 ∈ Γlocal such that

δu0 local (r) ≤ δu0 local (r0 ) = max δu0 local (r). (6.6.30)
Ωlocal

We suppose that the maximum δu0local (r0 ) is on Γinside . Suppose r0 ∈ Γactive .

According to Hopf’s boundary point lemma, see [61, p43], if

- δu0 local (r0 ) is continuous at r0 ,

- δu0local (r) < δu0local (r0 ) ∀ r ∈ Ωlocal ,
- there is a ball BR of radius R, BR ⊂ Ωlocal , with r0 ∈ ∂BR ,

∂
then δu0 local (r0 ) > 0. This is a contradiction since we have homogeneous
∂n
Neumann boundary conditions on Γactive . If δu0local (r) is a negative function
then starting with the weak minimum principle would give a similar result. So
we have

||δu0 active (r)||∞ ≤ max |δu0 local (r)| = |δu0 local (r0 )| = ||δu0 inside (r)||∞ . (6.6.31)
Ωlocal

However we note that Hopf’s lemma does not rule out the corner points. 

The next step in the LDC algorithm is to compute the defect. The defect d(r) is
the integral of the double layer with a perturbation in the exact u, that is
Z
∂v
d0 (r) = δu0 (r(χ)) (r; r(χ)) dχ, r, r(χ) ∈ Γ. (6.6.32)
Γ ∂n
In our simulations δu0 (r(χ)) is the difference between the local problem and
global problem solutions. Once we have the defect then we apply defect cor-
rection to the initial BIE and solve for the updated solution. Let u1 denote the
updated solution on Γ , then
Z
1 ∂v
u1 (r) + u1 (r(χ)) (r; r(χ)) dχ + d0 (r) =
2 Γ ∂n
Z
q(r(χ))v(r; r(χ)) dΓ + f1 (r); r, r(χ) ∈ r(χ). (6.6.33)
Γ

Using notations already introduced above, we have

Z
1 ∂v
(u1 (r) + δu1 (r)) + (u1 (r(χ)) + δu1 (r(χ))) (r; r(χ)) dχ + d0 (r) =
2 Γ ∂n
Z
q(r(χ))v(r; r(χ)) dχ. (6.6.34)
Γ
114 Local Defect Correction for BEM

Comparing (6.6.34) with (6.6.33) we have

Z
1 ∂v
δu1 (r) + δu1 (r(χ)) (r; r(χ)) dχ + d0 (r) = 0. (6.6.35)
2 Γ ∂n

That is,
1 ∂v
 
 I+
  δu1 (r) = −d0 (r), r ∈ Γ.
(r; r(χ)) (6.6.36)
2 ∂n

Theorem 6.6.3 (The combined process) The combined LDC process will con-
verge since the error of the updated solution is bounded by the error of old solu-
tion provided the shifted double layer operator inverse is bounded by a constant
not more than 2 in absolute size.

Proof. Let the error in the initial global coarse grid solution be δu0 (r), that is,

u(r) = u0 (r) + δu0 (r). (6.6.37)

Then (6.6.24) implies that the error in u on Γ in is given by

Z
∂v
δuinside (r) = δu0 (r(χ)) (r; r(χ)) dχ, r ∈ Γ in ⊂ Ω, r(χ) ∈ Γ. (6.6.38)
Γ ∂n

In the space W1 defined in (2.5.19), equation (6.6.36) has a unique solution.

Thus the operator (0.5I + Kd ) is invertible in that space. Using (6.6.36) we have
−1
1

 I + Kd 
δu1 (r) = −   d0 (r), r ∈ Γ. (6.6.39)
2
Also recall that the defect is given by
Z Z
∂v ∂v
d0 (r) = δu0 (r(χ)) (r; r(χ)) dχ = δu0active (r(χ)) (r; r(χ)) dχ, r ∈ Γ.
Γ ∂n Γactive ∂n
(6.6.40)

So we have,
−1 Z
1 ∂v

 I + Kd 
δu1 (r) = −   u0 active (r(χ)) (r; r(χ)) dχ. (6.6.41)
2 Γactive ∂n

Let a scalar C be such that

 −1
0.5I + Kd  ≤ C. (6.6.42)
∞

Then

||δu1 (r)||∞ ≤ C||d0 (r)||∞ . (6.6.43)

6.6 Continuous formulation of the LDC steps 115

For the defect d0 (r) we have

Z
∂v
||d0 (r)||∞ = max δu0active (r(χ)) (r; r(χ)) dχ
Γactive ∂n
Z
∂v
≤ |δu0active (r(χ))| (r; r(χ)) dχ
Γactive ∂n
Z
∂v
≤ ||δu0active (r(χ))||∞ (r; r(χ)) dχ,
Γactive ∂n
1
≤ ||δu0 active (r(χ))||∞ , (6.6.44)
2
since ∂v/∂n is of the same sign around Γ and the integral of ∂v/∂n around Γ is
1/2 for boundary points. Using (6.6.44) in (6.6.43) yields

C
||δu1 (r)||∞ ≤ ||δu0active (r(χ))||∞ . (6.6.45)
2
Then we use (6.6.31) and (6.6.25) to obtain
C
||δu1 (r)||∞ ≤ ||δu0 (r)||∞ . (6.6.46)
2


“Education is what remains after one has forgotten everything one
learned in school.” – Albert Einstein
Chapter 7

The potential problem for the

impressed current cathodic
protection system

7.1 Introduction

Usually steel structures are protected from corrosion by painting. But not all
parts can be painted, like for instance propellers of a ship. Also, due to damage,
some parts lose their paint hence creating more exposed parts of the steel. Ca-
thodic protection effectively protects underground or submerged metallic struc-
tures through the use of a negative potential applied by an external source to
the structure. The method is typically applied to iron or steel structures such
as underground pipelines, storage tanks, submarine structures, ocean pilings,
and electrical transmission towers. Cathodic protection is a proven technology
for controlling corrosion on the bottoms of above ground storage tanks [40].

There are two types of cathodic protection systems used against corrosion: a
galvanic or sacrificial anode cathodic protection system and an impressed current
cathodic protection system (ICCP) . The galvanic anode system is based upon the
natural potential difference which exists between the structure being protected
and the auxiliary electrode (anode) which is installed in the electrolyte. The
current that prevents corrosion is then due to the potential difference between
the structure and the anode. Materials commonly used for galvanic anode
systems are magnesium, zinc and aluminum.

The ICCP system uses anodes in conjunction with an external direct current
(DC) power source. The structure to be protected is connected to the nega-
118 The potential problem for the impressed current cathodic protection system

tive terminal of a direct current power source and electrical current is forced
to flow from the positive terminal to the anodes through the electrolyte to the
structure. This type of cathodic protection system uses long life anode ma-
terials such as high silicon chromium, cast iron, graphite, and mixed metal
oxide coated titanium. In water storage tanks, cathodic protection systems are
usually designed to protect the interior wetted surfaces of the tank [47, p. 40].
However, in some cases, the exterior of a tank bottom or shell is in contact with
corrosive soils and in that case cathodic protection can also be used [40].

A properly designed, installed and operated cathodic protection system will

eliminate the corroding areas by passing direct current to the metal surface.
This therefore necessitates the solution of the ICCP potential problem. In this
chapter we discuss the solution of the potential problem associated with im-
pressed current cathodic protection (ICCP) of the external surfaces of tank bot-
toms which are in contact with corrosive soils. The potential problem has
localised regions of high activity. These are the anodes or the cathodes which
are the corroded parts of the surface to be protected. Because of the small re-
gions of high activity associated with the problem, we would like to use of LDC
for BEM to solve the potential problem model and that is the subject of this
chapter. In the sequel we begin with a brief description of the problem. Then
in Section 7.2 we introduce the model and an example of a full BEM solution
and in Section 7.3 we give the BEMLDC solution.

Corrosion is an electrochemical process that involves passage of electrons from

one substance, usually metallic, called the anode to another substance called
the cathode. It is at the anode that the oxidation reaction which is responsible
for the corrosion of the metal involved takes place according to the following
ionic equation, in the case of steel

Fe(s) → Fe2+ −
(aq) + 2e , (7.1.1)

where Fe represents a ferrous atom, Fe2+ a ferrous II ion, e an electron and the
subscripts s and aq imply the solid state and the aqueous state respectively.
The electrons so produced pass on to the cathode where they are used up in a
reduction reaction

2H+ −
(aq) + 2e → H2(g) (7.1.2)

in acidic solutions, or:

O2(g) + 2H2 O(l) + 4e− → 4OH−

(aq) (7.1.3)

in neutral solutions, see Figure 7.1. Metal II is a metal that is more electropos-
itive than steel.
In equations (7.1.2) and (7.1.3), H+ represents a hydrogen ion, O2 oxygen gas,
H2 O a water molecule and OH− a hydroxyl ion. The subscripts l and g imply
the liquid state and the gaseous state respectively. Therefore corrosion basi-
cally occurs at the anode. Cathodic protection involves attaching an anode to
the surface to be protected and supplying a direct current through it such that
7.1 Introduction 119

electrolyte
current flow
2Fe → 2Fe2+ + 4e− O2 + 2H2 O + 4e− → 4OH−
(corrosion)

electron flow
Anode Cathode
(Steel) (Metal II)

Figure 7.1: Bimetallic corrosion cell.

all the other parts of the surface become cathodic and therefore do not corrode,
see Figure 7.2.

Anodes

Rectifier

TANK

(b) ICCP system for a storage tank,

source: [40].
(a) ICCP system for a submarine ship,
source: [31].

Figure 7.2: Examples of ICCP systems for ship’s hull and exterior
of a water storage tank.
120 The potential problem for the impressed current cathodic protection system

7.2 Modelling

We will now describe a two dimensional model for ICCP. The ICCP problem
for the cathodic protection of a ship in water is an exterior problem, see [31].
For continuity of the analysis in this thesis we will solve a model for the ICCP
protection of water storage tanks, which is an interior problem. One typical
situation of a tank under cathodic protection is shown in Figure 7.2b. This
problem can be modelled in two dimensions as shown in Figure 7.3. The do-
main consists of soil of electric conductivity σ. Further, the potential satisfies
boundary conditions at the anode surface, cathode surface and insulating sur-
face. The insulating surface is the perfectly (not damaged) painted tank surface
and the anodes interconnectivity. The boundary conditions are derived by tak-
ing into account the electric field E and the electric current density J in the
domain, which are related by

E = −∇u, J = σE, (7.2.1)

where σ represents the soil conductivity and u is the potential. For the insulat-
ing surface we have

n · J = σn · E = −σn · ∇u. (7.2.2)

Since no current flows through the insulating surface, this results in the con-
dition
∂u
(r) = n · ∇u(r) = 0, (7.2.3)
∂n
at the insulating surface. At the cathode surface, the relation between current
density and potential difference is given by a polarisation curve
∂u
J · n = −σ = fcathode (u), (7.2.4)
∂n
with the general shape of fcathode (u) determined by the oxidation and reduction
currents of Iron and Oxygen, [31, p. 13]. At the anode surface, perfect contact
with the domain medium is assumed. Therefore either one of the following
relations may be assumed at the anode surface:

u(r) = uanode , (7.2.5a)

I
σn · ∇u = − , (7.2.5b)
A
∂u
J · n = −σ = fanode (u). (7.2.5c)
∂n
The first boundary condition expresses that each point of the anode surface
has the same potential, see [77]. The second boundary condition expresses a
homogeneously distributed current density flowing out of the anode surface.
Here I is the total current flowing through the anode surface and A is the total
7.2 Modelling 121

surface area , see [74]. The third condition is a more general polarisation curve
for anodic surfaces used by [18, 49, 52, 58]. Boundary conditions involving
polarisation curves are non-linear. That is, the relation between the potential
and the current density is non-linear, but it can be linearised [49, 53]. For
our model, we shall use (7.2.5a). Similarly, we also assume at the cathode
fcathode (u) = ucathode . The current J flows from the anode to the cathode as
depicted in the illustration Figure 7.3. The domain is the soil area between

Γ 2 , Cathode

J = σE

Γ 3 , Insulating surface

Soil, σ

∇2 u = 0

Γ 1 , Anode

Figure 7.3: Schematic two dimension model for

the ICCP of a water storage tank shown in Fig-
ure 7.2b.

the tank surface and the anodes interconnectivity. Apart from the anodes and
the damaged parts, the rest of the painted surface is assumed to be a perfect
insulator. Also, for simplicity we assume one anode and cathode. However, the
solution process is still the same for more active regions. At the electrodes we
consider the potential to be a prescribed value.

Thus, let Ω be the region between the tank surface and the anodes intercon-
nectivity, Γ 1 the anode surface, Γ 2 the cathode surface and Γ 3 the insulating
surface. Let u(r) be the potential at a point r ∈ Ω ⊂ R2 , then we have the
following problem:
 2
 ∇ u(r) = 0,


r ∈ Ω,




 u(r) = uanode ,

 r ∈ Γ 1,

2
(7.2.6)


 u(r) = ucathode , r ∈ Γ ,





 ∂u (r) = 0,

r ∈ Γ 3.
∂n
122 The potential problem for the impressed current cathodic protection system

where n(r) is the unit normal vector at r pointing into Ωc , see Figure 7.4.

Γ2

dχ
Γ3 Ω r
n at r ∈ ∂Ω

Γ1
Ωc

Figure 7.4: Diagram for a two di-

mensional mathematical model for
an ICCP.

So we have the following BIE representation for the problem:

Z Z
∂v ∂u
c(s)u(r) = u(r) (s; r)dχ − (r)v(s; r)dχ (7.2.7)
Γ ∂n Γ ∂n

where Γ = Γ 1 ∪ Γ 2 ∪ Γ 3 . As we will see in Section 7.3, the solution of the inte-

gral equation (7.2.7) together with the boundary conditions in (7.2.6) has rapid
variation of the potential around the electrodes and the solution is smooth else-
where. This is why the use of LDC is recommended for such a problem in order
to use a grid that captures the activity well yet keeping the computational costs
low. In the next section we present the application of LDC to solve this problem.

7.3 BEM-LDC for the ICCP problem

In this section we give an LDC formulation and solution for the ICCP potential
problem model presented in Section 7.2. In practice the number of active re-
gions p can be one, two, or even more, that is, several anodes and cathodes.
Here we consider two active regions, that is p = 2, one anode and one cathode.
The procedure for even more anodes and cathodes is the same and the gain in
complexity for a given p is discussed on page 97. After solving a global coarse
grid problem on the global boundary Γ , we formulate a local problem at each of
1 2
the electrodes Γactive and Γactive , see Figure 7.5.
7.3 BEM-LDC for the ICCP problem 123

2
Γactive

n
Ω2local 2
Γinside

Ω1local Γinside
1

1
Γactive

Figure 7.5: Global and local domains for

a LDC formulation.

On each Ωplocal , p = 1, 2, see Figure 7.5, we have the interior problem

 2

 ∇ u(r) = 0, r ∈ Ωp local ,



p
u(r) = uelectrode (r), r ∈ Γactive , (7.3.1)




 p
u(r) = g̃(r), r ∈ Γinside ,

where g̃(r) is a piecewise constant function given by

p
g̃(r) = uinside (ri ), r ∈ rlinside ⊂ Γinside , r = ri , (7.3.2)

and
 1
 uanode , r ∈ Γactive ,
uelectrode (r) = (7.3.3)
 2
ucathode r ∈ Γactive .

The LDC process is what has been described in Section 6.2. Here we only add
in the twist of having more than one local problem. So we have the following
process:

(i) Initialisation: we solve the global coarse grid problem on Γ using a grid of
124 The potential problem for the impressed current cathodic protection system

size L to obtain the initial solution qL0 , uL0 ;

X N Z X N Z
1 L ∂v
u0,i = uL0,j (ri ; r(χ))dχ − qL0,j v(ri ; r(χ))dχ. (7.3.4)
2 ΓjL ∂n ΓjL
j=1 j=1

p p,l p,l
(ii) For each local problem, after discretisation of Γlocal into Γlocal = Γinside ∪
p,l p,l
Γactive , compute u(r) on Γinside using the global coarse grid solution and the
integral relation

N
X Z X N Z
∂v
up,l
inside,i = uLj (ri ; r(χ))dχ − qLj p
v(ri ; r(χ))dχ, ri ∈ Γinside .
ΓjL ∂n ΓjL
j=1 j=1
(7.3.5)

p p
(iii) Now we have Dirichlet boundary conditions on Γinside . On Γactive the bound-
p
ary conditions are as given on the global Γ since Γactive ⊂ Γ, p = 1, 2. So
p
next is to solve the local problems on Γlocal :

X p,l Z
1 p,l ∂v
ulocal,i + uactive,j (ri ; r(χ)) dχ
2 l
Γactive,j ∂n
j

X Z
∂v
+ up,l
inside,j (ri ; r(χ)) dχ =
l
Γinside,j ∂n
j

X Z X Z
qp,l
active,j v(ri ; r(χ)) dχ + qp,l
inside,j v(ri ; r(χ)) dχ.
l
Γactive,j l
Γinside,j
j j
(7.3.6)

Note that the p problems are independent of one another and can be solved
in parallel.
p
(iv) To compute the defect: we now have a better solution on Γactive , p = 1, 2
and
 L
 uc,j , ΓjL ⊂ Γc ,
uLbest,j = (7.3.7)
 p,l p
uactive,j , ΓjL ⊂ Γactive ,
 L
 qc,j , ΓjL ⊂ Γc ,
qLbest,j = (7.3.8)

qp,l L p
active,j , Γj ⊂ Γactive .

Using the fine grid solution, we approximate the defect di for each node of
7.3 BEM-LDC for the ICCP problem 125

the global coarse grid as follows. We compute

 2 
 P R
 qp,l

 active,j ΓjL v(ri ; r(χ)) dχ−

 p=1



 P R 
dLij ≈ qp,l p

v(r ; r(χ)) dχ , ΓjL ⊂ Γactive ,

i

 active,jk



 k Γ p,l 

 active,jk




0 Γj ⊂ Γc .
(7.3.9)
p,l
Here ∪ Γactive,jk
= ΓjL , see Figure 6.4. Then
k
X
dLi = dLij (7.3.10)
j

for ΓjL ⊂ Γactive . The defect di has contributions from all the p active regions
hence the summation for p = 1, 2, in (7.3.9).
(v) We now do the defect correction: the defect (7.3.10) is added to the right
hand side of the coarse grid equation and then we solve for the updated
global coarse grid solution uL1 , qL1 on the global coarse grid,

X N Z N
X Z
1 L ∂v
u1,i = uL1,j (ri ; r(χ))dχ − qL1,j v(ri ; r(χ))dχ + dLi . (7.3.11)
2 ΓjL ∂n ΓjL
j=1 j=1

(vi) Go back to step (ii) with uL replaced by uL1 and qL replaced by qL1 .

To test this LDC formulation for more than one local problem we solve the fol-
lowing problem, whose continuous solution is known. We consider the function

f(r) = log(r1 ) + log(r2 ) (7.3.12)

where

r1 = ||r − s1 ||2 , r2 = ||r − s2 ||2 (7.3.13)

and s1 , s2 are source points just outside the square. Then we solve the problem
 2
 ∇ u(r) = 0,
 r ∈ Ω = [0 1] × [0 1],
(7.3.14)
 q(r) = ∂f (r), r ∈ Γ.

∂n

The solution to (7.3.14) is shown in Figure 7.6.

Clearly the solution has two small regions, one on the lower side and another
on the upper side of the square where u varies rapidly. In Figures 7.7 and 7.8
126 The potential problem for the impressed current cathodic protection system

−0.8

−1

−1.2

−1.4

−1.6

−1.8

−2

−2.2

−2.4

−2.6

Figure 7.6: Profile of the potential

u inside a unit square due to prob-
lem (7.3.14). The source points s1 and s2
are (0.5, −0.02) and (0.5, 1.02) respectively.

0 exact continuous
first solution 0
−0.2
updated solution
−0.2
−0.4
−0.4
−0.6
u

−0.6
−0.8
−0.8 exact continuous
−1 first solution
−1 updated solution
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(x,0) (1,y)

(a) Solutions on y = 0. (b) Solutions on x = 1.

Figure 7.7: LDC solutions to problem 7.3.14 on a unit

square after the first iteration.

0 exact continuous
first solution 0
−0.2 updated solution
−0.2
−0.4
−0.4
−0.6
u

−0.6
−0.8
−0.8 exact continuous
−1 first solution
−1 updated solution
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(x,1) (0,y)

(a) Solutions on y = 1. (b) Solutions on x = 0.

Figure 7.8: LDC solutions to problem 7.3.14 on a unit

square after the first iteration.
7.3 BEM-LDC for the ICCP problem 127

we have the BEM solutions to the problem using local defect corrections with a
local problem at each active region.

The updated solution agrees with the continuous solution better than the initial
problem as expected. We can therefore proceed to solve the ICCP problem
whose continuous solution is not known. Once we have the potential u in Ω we
use (7.2.1) to compute the electric field in Ω. Figure 7.2.6 shows the potential
and the electric field lines for the ICCP potential problem (7.2.6) with uanode = 1
and ucathode = −uanode . Now for the ICCP problem we let

1 1

0.5

0
y

−0.5

−1 0
x 0 1

(a) Potential inside Ω. (b) Electric field lines inside Ω.

Figure 7.9: The potential and field lines of the ICCP model 7.2.6 in a
region inside Ω where uanode = 1 and ucathode = −uanode .

1
Γactive = {(x, y) : x ∈ [0.4, 0.6], y = 0},
2
Γactive = {(x, y) : x ∈ [0.4, 0.6], y = 1}.

Then we set, see Figure 7.10,

Ω1local = {(x, y) ∈ [0.4, 0.6] × [0, 0.4]},

Ω2local = {(x, y) ∈ [0.4, 0.6] × [0.6, 1]}.

LDC solutions for the problem are shown in Figures 7.11 and 7.12 .

Since we do not have the exact continuous solution to (7.3.1) we cannot com-
pute the actual error. However, assuming that the solution converges to a fixed
point xfixed , we can then compute the errors ||xi − xfixed ||∞ where xi is the solu-
tion after the i-th iteration to demonstrate how fast we reach the fixed point.
The results are shown in Figure 7.13. The results indeed show that one or two
iterations is just enough.
128 The potential problem for the impressed current cathodic protection system

Ω2local

0.6

0.4

Ω1local

0 0.4 0.6 1

Figure 7.10: Domain demarcations for

the LDC solution of the model problem.

0.5

0
y

−0.5

−1
x

(a) Solution of local problem on Ω1local . (b) Field lines due to potential of local prob-
lem solution on Ω1local .

Figure 7.11: LDC solutions to first local problem.

7.3 BEM-LDC for the ICCP problem 129

0.5

0
y

−0.5

−1
x

(a) Solution of local problem on Ω2local . (b) Field lines due to potential of local prob-
lem solution on Ω2local .

Figure 7.12: LDC solutions to the second local problem.

0.8
fixed ∞
||

0.6
||x − x

0.4
i

0.2

0
0 1 2 3 4 5 6 7
iteration, i

Figure 7.13: Variation of the

global coarse grid solution er-
ror from the fixed point solution
during iteration. The fine grid
is of size l = L/9 and the global
coarse grid of size L = 0.2.
Bibliography

[1] M. J. H. Anthonissen, Local Defect Correction Techniques: Analysis and Ap-

plication to Combustion, Ph.D. thesis, Eindhoven University of Technology,
Eindhoven, The Netherlands, 2001.

[2] D. N. Arnold and W. L. Wendland, On the asymptotic convergence of collo-

cation methods, Math. Comp. 41 (1983), 349 – 381.

[3] D. N. Arnold and W. L. Wendland, The convergence of spline collocation for

the strongly elliptic equations on curves, Numer. Math. 47 (1985), 317 –
341.

[4] U. M. Ascher, R. M. M. Mattheij, and Robert D. Russel, Numerical Solu-

tion of Boundary Value Problems for Ordinary Differential Equations, SIAM,
Philadelphia, 1995.

[5] K. E. Atkinson, The numerical solution of integral equations of the second

kind, Cambridge University Press, 1997.

[6] T. J. Baker, Mesh adaptation strategies for problems in fluid dynamics,

Finite Elements Analysis and Design 25 (1997), 243 – 273.

[7] W. Cao, W. Huang, and R. D. Russel, A study of monitor functions for two
dimensional adaptive mesh generation., SIAM J. Sci. Computing. 20 No. 6
(1999), 1978–1994.

[8] C. Carsten and E. P. Stephan, A posteriori error estimates for boundary

element methods., Mathematics of Computation. 64 (1995), 483–500.

[9] G. Chen and J. Zhou, Boundary Element Methods, Academic press, Lon-
don, San Diego, 1992.

[10] K. Chen, Error equidistribution and mesh adaptation, SIAM Journal on

Scientific Computing 15, No. 4 (1994), 798–818.

[11] C. Constanda, On the solution of the Dirichlet problem for the two-
dimensional laplace equation., Proceedings of the American Mathematical
Society. , Vol. 119, No. 3 (Nov., 1993), 877–884.
132 BIBLIOGRAPHY

[12] M. Costabel and M. Dauge, Invertibility of the biharmonic single layer po-
tential operator., Inter Equat Oper Th 24 (1996), 46–67.

[13] M. Costabel, M. Dauge, and S. Nicaise, Boundary value problems and in-
tegral equations in nonsmooth domains, Marcel Decker, INC, New York,
1995.

[14] T. A. Cruse, Numerical solutions in three dimensional elastostatics., Int. J.

Solids Structures. 5 (1969), 1259 – 1274.

[15] P. J. Davis and P. Rabinowitz, Numerical integration, Blaisdell Publishing

Company, Waltham, 1967.

[16] W. Dijkstra, Condition Numbers in the Boundary Element Method: Shape

and Solvability, Technische Universiteit Eindhoven, Eindhoven, 2008.

[17] W. Dijkstra, G. Kakuba, and R. M. M. Mattheij, Condition numbers and

local errors in the boundary element method, In BOUNDARY ELEMENT
METHODS IN ENGINEERING AND SCIENCES (Imperial College Press Lon-
don, 2010), M H Aliabadi and P H Wen, COMPUTATIONAL AND EXPERI-
MENTAL METHODS IN STRUCTURES 4 (2010), 412.

[18] P. Doig and P.E.J. Flewitt, A finite difference numerical analysis of galvanic
corrosion for semi-infinite linear coplanar electrodes, Journal of The Elec-
trochemical Society 126(12) (1979), 2057–2063.

[19] P. J. J. Ferket and A. A. Reusken, Further Analysis of the Local Defect

Correction Method, Computing 56 (1996), 117 – 139.

[20] I. Fredholm, Sur une classe d’equations fonctionnelles, Acta Numer. 27

(1903), 365 – 390.

[21] M. Graziadei, Using local defect correction for laminar flame simulation,
Ph.D. thesis, Eindhoven University of Technology, 2004.

[22] M. Guiggiani and F. Lombardi, Self-adaptive boundary elements with h-

Hierachical shape functions., Advances in Engineering Software, Special
issue on error estimates and adaptive meshes for FEM/BEM. 15 (1992),
269–277.

[23] W. Hackbusch, Local defect correction method and domain decomposition

technique, Computing [Suppl.] 5 (1984), 89 – 113.

[24] W. Hackbusch, Integral Equations: Theory and Numerical Treatment,

Birkhauser Verlag, Basel, 1995.

[25] M. T. Heath, Scientific computing, an introductory survey, McGraw Hill,

New York, 2002.

[26] G.C. Hsiao and W. L. Wendland, A finite element method for some integral
equations of the first kind., J. Math. Anal. Appl. 58 (1977), 449 – 481.
BIBLIOGRAPHY 133

[27] M.A. Jaswon, Integral equation methods in potential theory i, Series a, vol.
275, 1963.

[28] M.A. Jaswon and G. Symm, Integral Equation Methods in Potential Theory
and Elastostatics, Academic Press, London, 1977.

[29] A. B. Jorge, G. O. Ribero, and T. S. Fischer, New approaches for error

estimation and adaptivity for 2d potential boundary element methods, Int.
J. Numer. Meth. Engng 56 (2003), 177 – 144.

[30] K. Jorgens, Linear Integral Operators, Pitman Advanced Publishing Pro-

gram, Boston, London, Melbourne, 1982.

[31] G. Kakuba, The Impressed Current Cathodic Protection System, Master’s

thesis, Eindhoven University of Technology, Eindhoven, The Netherlands,
2005.

[32] G. Kakuba and R. M. M. Mattheij, Convergence analysis of Local Defect

Correction for the BEM, Advances in Boundary Integral Methods. Proceed-
ings of the sixth UK conference on Boundary Integral Methods 15 (2007),
251 – 261.

[33] G. Kakuba and R. M. M. Mattheij, Local errors in the constant and linear
boundary element method for potential problems, Advances in boundary
element techniques X. Proceedings of the 10th International conference
(2009), 367 – 374.

[34] G. Kakuba, R. M. M. Mattheij, and M. J. Anthonissen, Local Defect Correc-

tion for the Boundary Element Method, Computer Modeling in Engineering
Sciences 15 (2006), 127 – 135.

[35] J. T. Katsikadelis, Boundary elements theory and applications, Elsevier,

2002.

[36] O. D. Kellogg, Foundations of potential theory, Springer, Berlin, 1929.

[37] E. Kita and N. Kamiya, Adaptive mesh refinement in boundary element

method, an overview, Engineering Analysis with Boundary Elements 25
(2001), 479–495.

[38] R. Kress, Linear integral equations, Springer-Verlag, 1989.

[39] E. Kreyszig, Advanced Engineering Mathematics, John Wiley and Sons,

New York, 1999.

[40] D. H. Kroon and M. Urbas, Cathodic protection of above ground storage

tank bottoms, Corrpro Companies, Inc and Harco technologies corporation
respectively. http://www.corrpro.co.uk/pdf/TechnicalPapers/40CP
Cathodic protection of above ground storage tank bottoms.pdf (2010
October 7).

[41] A. Kufner and J. Kadlec, Fourier series, London: Iliffe Books, 1971.
134 BIBLIOGRAPHY

[42] V. D. Kupradze, Potential methods in theory of elasticity, Daniel Davy, New

York, 1965.

[43] P. K. Kythe and M. R. Schferkotter, Handbook of computational methods

for integration, Chapman & Hall/CRC, Florida, 2005.

[44] M. T. Liang, J. T. Chen, and S. S. Yang, Error estimation for boundary

element method., Engineering analysis with boundary elements. 23 (1999),
257–265.

[45] S. Liapis, A review of error estimation and adaptivity in the boundary el-
ement method., Engineering analysis with boundary elements. 14 (1994),
315–323.

[46] Z. Liu, M. D. Thoreson, A. V. Kildishev, and V. M. Shalaev, Translation of

nanoantenna hot spots by a metal-dielectric composite superlens, Applied
physics letters 95 (2009), 033114.

[47] American Water Works Association MANUAL M42, Steel water storage
tanks, American Water Works Association, 1999.

[48] R. M. M. Mattheij, S. W. Rienstra, and J. H. M. ten Thije Boonkkamp, Par-

tial differential equations: Modeling, analysis, computation, SIAM, Philadel-
phia, 2005.

[49] E. McCafferty, Distribution of potential and current in circular corrosion cells

having unequal polarisation parameters, Journal of The Electrochemical
Society 124(12) (1977), 1869–1878.

[50] S. D. Mikhlin, Integral equations, Pregamon Press, Oxford, 1957.

[51] R. Minero, M. J. H. Anthonissen, and R. M. M. Mattheij, A Local Defect

Correction Technique for Time-Dependent Problems, Numerical Methods for
Partial Differential Equations 22 (2006), 128 – 144.

[52] R. Morris and W. Smyrl, Galvanic interactions on periodically regular het-

erogeneous surfaces, AIChE Journal 34(5) (1988), 723–732.

[53] R. Morris and W. Smyrl, Current and potential distribution in thin elec-
trolyte layer galvanic cells, Journal of The Electrochemical Society 136(11)
(1989), 3229–3236.

[54] P. Mund, E. P. Stephan, and J. W. Hannover, Two-level mthods for the

single layer potential in r3 ., Computing 60 (1998), 243 – 266.

[55] N. I. Muskhelishvili, Some basic problems of mthematical theory of elastic-

ity, Noordholl, Holland, 1953.

[56] E. T. Ong and K. M. Lim, Three-dimensional singular boundary elements for

corner and edge singularities in potential problems., Engineering analysis
with Boundary Elements 29 (2005), 175–189.
BIBLIOGRAPHY 135

[57] F. Paris and J. Canas, Boundary Element Method: Fundamentals and Ap-
plications, Oxford University Press, Oxford, 1997.

[58] R. Oltra Ph. Bucaille and T. Warner, Theoretical and experimental studies of
galvanic corrosion between aluminium and Al-Cu alloys, Materials Science
Forum 242 (1997), 207–212.

[59] C. Pozrikidis, Boundary integral and singularity methods for linearised vis-
cous flow, Cambridge University Press, 1992.

[60] C. Pozrikidis, A Practical Guide to Boundary Element Methods with the soft-
ware BEMLIB, Chapman & Hall/CRC, London, 2002.

[61] P. Pucci and J. Serrin, The maximum principle, Birkhauser Verlag AG,
2007.

[62] C. G. Reimbert and A. A. Minzoni, Effect of radiation losses on hotspot

formation and propagation in microwave heating, IMA Journal of Applied
Mathematics 57 (1996), 165 – 179.

[63] F. J. Rizzo, An integral equation approach to boundary-value problems of

classical elastostatics., Q. Applied Math. 25 (1967), 83 – 95.

[64] J. Saranen and W. L. Wendland, On the asymptotic convergence of colloca-

tion methods with spline functions of even dgree, Math. Comp. 45 (1985),
91 – 108.

[65] K. Schimmanz and A. Kost, Formulation of mixed elements for the 2d-bem,
The International Journal for Computation and Mathematics in Electrical
and Electronic Engineering 23 No. 4 (2004), 866–875.

[66] H. Schulz and O. Steinbach, A new a posteriori error estimator in adap-

tive direct boundary element methods: the Dirichlet problem, CALCOLO
Springer-Verlag 37 (2000), 79–96.

[67] V. Sladek, J. Sladek, and M. Tanaka, Numerical integration of logarith-

mic and nearly logarithmic singularity in bems, Applied mathematical mod-
elling 25 (2001), 901 – 992.

[68] I. H. Sloan, Error analysis of boundary integral methods, Acta Numerica

(1991), 287–339.

[69] C. Somigliana, Sopra l’equilibrio di un corpo elastico isotrope, Il Nuovo

Ciemento (1886), 181–185.

[70] H. Strese, Remarks concerning the boundary element method in potential

theory, Appl. Math. Modelling 8 (1984), 40 – 44.

[71] T. Strouboulis and K. A. Hague, Recent experiences with error estimation

and adaptivity. part i: Review of error estimators for scalar elliptic problems,
Comp. Meth. Appl. Mech. Eng. 97 (1992), 399 – 436.
136 BIBLIOGRAPHY

[72] T. Strouboulis and K. A. Hague, Recent experiences with error estimation

and adaptivity. part ii: Error estimation for h-adaptive approximation on
grids of triangles and quadrilaterals., Comp. Meth. Appl. Mech. Eng. 100
(1992), 359 – 430.
[73] A. H. Stroud and D. Secrest, Gaussian quadrature formulas, Prentice-Hall,
Englewood Cliffs, NJ, 1966.
[74] W. Sun and K. M. Liu, Numerical solution of cathodic protection systems
with nonlinear polarisation curves, Journal of The Electrochemical Society
147(10) (2000), 3687–3690.
[75] G. T. Symm, Integral equation methods in potential theory - ii, Series a.,
Proc. Roy. Soc. Lond. 275 (1963), 33 – 46.
[76] R. Vodicka and V. Mantic, On invertibility of elastic single-layer potential
operator, Journal of Elasticity 74 (2004), 147 – 173.
[77] J. T. Waber, Mathematical studies on galvanic corrosion, Journal of The
Electrochemical Society 101(6) (1954), 271–276.
[78] S. P. Walker and R. T. Fenner, Treatment of corners in BIE analysis of poten-
tial problems., International journal for numerical methods in engineering.
28 (1989), 2569–2581.
[79] K. Wang, Bem simulation for glass parisons, Ph.D. thesis, Eindhoven Uni-
versity of Technology, 2002.
[80] J. A. Whitehead, Fluid models of geological hotspots, Annual review of Fluid
Mechanics 20 (1988), 61 – 68.
[81] G. M. Wing, A primer on integral equations of the first kind: the problem of
deconvolution and unfolding., SIAM, Baltimore, 1991.
[82] D. J. Wu, Y. Cheng, and X. J. Liu, Hot spots induced near-field enhance-
ments in au nanoshell and au nanoshell dimer, Applied Physics B: Lasers
and Optics 97, Number 2 (2009), 497 – 503.
Index

anode, 117 finite element methods, 1

aqueous, 118 Fourier series, 76
arclength coordinate, 9 Fredholm integral equation of the sec-
ond kind, 15
boundary discretisation error, 41 Fredholm integral equation of the first
boundary element, 22 kind, 15
boundary element method, 21
boundary integral equation, 14 galvanic cathodic protection, 117
Gauss-Legendre quadrature, 28
cathode, 118 global coarse grid, 87, 90
cathodic protection, 117 global error, 36, 37, 73–75
collocation point, 22 grid size, 22, 26
constant elements, 22
continuous linear elements, 25, 54 impressed current cathodic protection,
corrosion, 117, 118 4, 117
integration knots, 28, 29
defect correction, 92 integration weights, 28, 29
Dirac delta distribution, 8 interpolation error, 2, 41, 54
direct current, 117
Dirichlet problem, 8 kernel, 14
double layer operator, 14
double layer potential, 13 linear elements, 22
local active boundary, 89
eigenfunctions, 18 local boundary, 89
eigenvalues, 17, 18 local defect correction, 87
electric conductivity, 120 local discretisation error, 36, 74
electrochemical process, 118 local domain, 88
electrode, 117 local error, 36, 44, 53
electrolyte, 117 local fine grid, 87
electron, 118 local problem, 89
element size, 22
elements, 21 median 2-norm, 16
equidistribution, 37, 64, 65 mesh selection, 64
exact right hand side, 43 Mixed problem, 8
expansion coefficients, 75 monitor function, 65

Ferrous, 118 Neumann problem, 8

finite difference methods, 1 numerical boundary, 21
138 Index

oxidation reaction, 118

quadrature error, 41

shape functions, 22, 41

single layer operator, 14
single layer potential, 13
Sobolev space, 15
sublocal error, 42, 44, 53, 58

tangential derivative, 45

uniform grid, 22

weight function, 65
Summary

Adaptive gridding methods are of fundamental importance both for industry

and academia. As one of the computing methods, the Boundary Element Method
(BEM) is used to simulate problems whose fundamental solutions are avail-
able. The method is usually characterised as constant elements BEM or linear
elements BEM depending on the type of interpolation used at the elements. Its
popularity is steadily growing because of its advantages over other numerical
methods most important of which being that it only involves obtaining data at
the boundary and computation of the solution in the domain is merely a case
of post processing through the use of an identity. This results in the reduction
of the problem dimension by unity.

Although there is a reduction in dimension when we use BEM, the method

usually results in full matrices which can be expensive to solve. This makes
the method costly for problems that require very fine grids like those with hot
spots as in the example of impressed current cathodic protection systems (ICCP).
The BEM is a global method in nature in that the solution in one node depends
on the solutions in all the other nodes of the grid. Hence an error in one node
can pollute the solution in all the other nodes.

In this thesis, we first focus our attention on defining and studying both the
local and global errors for the BEM. This is not a completely new study as
the literature suggests, however, our approach is different. We use the basic
foundations of the the method to define the errors. Since the method is a global
method, first we use the interpolation error on each element to define what we
have called a sublocal error. Then using the sublocal error we have defined the
local error. Understanding the local errors enabled us study the global error.
Theoretical and numerical results show that these errors are second order in
grid size for both the constant and linear element cases.

Then, having explored errors, we study a method for adaptive grid refinement
for the BEM. Rather than using a truly nonuniform grid, we present a method
called local defect correction (LDC) that is based on local uniform grid refine-
ment. This method is already developed and documented for other numerical
methods such as finite difference and finite volume methods but not for BEM.
140 Summary

In the LDC method, the discretization on a composite grid is based on a combi-

nation of standard discretizations on several uniform grids with different grid
sizes that cover different parts of the domain. At least one grid, the uniform
global coarse grid, should cover the entire boundary. The size of the global
coarse grid is chosen in agreement with the relatively smooth behaviour of the
solution outside the hot spots. Then several uniform local fine grids each of
which covers only a (small) part of the boundary are used in the hot spots. The
grid sizes of the local grids are chosen in agreement with the behaviour of the
continuous solution in that part of the boundary.

The LDC method is an iterative process whereby a basic global discretisation

is improved by local discretisations defined in subdomains. The update of the
coarse grid solution is achieved by adding a defect correction term to the right
hand side of the coarse grid problem. At each iteration step, the process yields a
discrete approximation of the continuous solution on the corresponding com-
posite grid. We have shown how this discretisation can be achieved for the
BEM. We apply the discretisation to an academic example to demonstrate its
implementation and later show how to use it for an application as the ICCP
system. The results show that it is a cheaper method than solving on a truly
composite grid and converges in a single step.
Curriculum vitae

The author of this thesis was born in Rwengoma, Uganda. He finished his
high school at St. Leo’s College, Kyegobe. He later joined Makerere University
where he graduated in June 2001 with a Bachelor of Science degree in the
subjects Maths, Physics and Education. In 2002 he joined and worked at
the Department of Mathematics at Makerere University as a teaching assistant
until September 2003 when he joined Eindhoven University of Technology in
the Netherlands to pursue a masters course. In August 2005 he graduated cum
laude as a master of science in computational science and engineering. His
master’s thesis, written under the supervision of prof.dr.ir. R. M. M. Mattheij
and dr. ir. M.J. H. Anthonissen was titled The impressed current cathodic
protection system.

Late in 2005, Godwin returned to the Netherlands, and started working as

a PhD student in the Centre for Analysis and Scientific Computing group at
Eindhoven University of Technology under the supervision of prof.dr. R. M. M.
Mattheij, the group’s chair for scientific computing. His research on boundary
element methods has lead to this thesis.

Godwin is at the moment working as an assistant lecturer in the department of

mathematics at Makerere University, Kampala.