Advanced Boundary Element Methods: Joachim Gwinner Ernst Peter Stephan

Springer Series in Computational Mathematics 52
Joachim Gwinner · Ernst Peter Stephan
Advanced
Boundary
Element
Methods
Treatment of Boundary Value,
Transmission and Contact Problems
Springer Series in Computational Mathematics
Volume 52
Editorial Board
R.E. Bank
R.L. Graham
W. Hackbusch
J. Stoer
R.S. Varga
H. Yserentant
More information about this series at http://www.springer.com/series/797
Joachim Gwinner • Ernst Peter Stephan
Advanced Boundary Element

Methods
Treatment of Boundary Value, Transmission
and Contact Problems
123
Joachim Gwinner Ernst Peter Stephan
Fakultät für Luft- und Raumfahrttechnik Institut für Angewandte Mathematik
Universität der Bundeswehr München Leibniz Universität Hannover
Neubiberg/München Hannover, Germany
Germany
ISSN 0179-3632 ISSN 2198-3712 (electronic)

Springer Series in Computational Mathematics
ISBN 978-3-319-92000-9 ISBN 978-3-319-92001-6 (eBook)
https://doi.org/10.1007/978-3-319-92001-6
Library of Congress Control Number: 2018946112
Mathematics Subject Classification (2010): 35-XX, 45-XX, 49-XX, 65-XX
© Springer International Publishing AG, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by the registered company Springer International Publishing AG part
of Springer Nature.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To our wives Hannelore and Karin Sabine for
their love and understanding.
Preface
The boundary element method (BEM) has become an important tool to provide
approximate solutions for boundary integral equations covering a rich area of
applications in engineering and physics. Today, there exist many books and
survey articles on boundary integral equations and on boundary element methods
[98, 112, 225, 259, 260, 276, 304, 359, 362, 391]. However, we believe that modern
topics like adaptive methods; treatment of general transmission, screen, crack,
and contact problems; and the hp-version of the BEM are dealt with in special
research papers only. In this book, we collect some of the key results of these
topics, prove them in detail, and describe the most important approaches. We
elaborate on the mathematical analysis of both the boundary integral equations
and the BEM and demonstrate the power of the BEM with numerical results for
representative problems from various applications in acoustics, electromagnetics,
and solid mechanics covering Laplace, Helmholtz, Navier–Lame, and Maxwell
partial differential equations.
Our book introduces the reader into the classical setting of boundary integral
equations and standard boundary element methods in Chaps. 1–3 and Chap. 6.
The book covers advanced boundary element methods in recent research areas as
mentioned above in Chaps. 4, 5, 7–13.
In Chap. 4, we apply the modern tool of pseudodifferential operators to mixed
boundary value problems and transmission problems.
In Chap. 5, we focus on the Signorini problem and more nonsmooth BVPs,
dealing with unilateral contact without and with friction and nonmonotone contact
in delamination.
In Chap. 6, we collect basic issues of BEM, covering Galerkin and collocation
methods with modifications and extensions (augmented boundary elements, duality
estimates, and qualocation).
In Chap. 7, we turn to boundary value problems (BVPs) in nonsmooth
domains and present improved BEM with graded meshes and higher polynomial
approximation.
In Chap. 8, we investigate in detail the exponential convergence of the hp-version
BEM on geometrically graded meshes.
vii
viii Preface
In Chap. 9, we employ the Mellin transform and analyze the boundary integral
operators on polygonal domains in depth.
In Chap. 10, we study the adaptive BEM using error estimators of residual type
and of hierarchical type, and also we give results on the convergence of adaptive
boundary element schemes.
In Chap. 11, we extend the BEM to unilateral contact problems without and with
friction and nonmonotone contact problems from delamination.
In Chap. 12, we analyze the symmetric FEM–BEM coupling for various
transmission problems in applications.
The final Chap. 13 is devoted to the time-dependent BEM (TD-BEM). We
treat the scattering of waves at screens and time-dependent contact problems using
retarded potentials.
In the Appendix, we collect some fundamental concepts of linear operator
theory and also provide some supplementary material on Fourier transform and
pseudodifferential operators. Further, we present a short course on convex and
nonsmooth analysis leading to linear and nonlinear variational inequalities and their
approximation. Also, some aspects of implementations of BEM are given.
For the ease of the reader, the chapters are self-contained; hence, it is unavoidable
that the text has some repetitions.
Different from standard textbooks and monographs on BEM, we stress on
first kind integral equations, adaptive methods, the hp-version of BEM, and the
application of BEM to contact problems with recent developments for the dynamic
case. Our book is addressed to mathematicians and engineers as well as to graduate
students. Therefore, we provide the necessary foundations of BEM and demonstrate
the applicability of BEM via prototype problems. We put specific emphasis on
numerical approaches underlined by representative numerical simulations.
One of the main concerns of the book is the abstract setting of the convergence
of the boundary element method. This is dealt with by the key theorems on the
convergence of the projection method (Theorems 1.1, 1.2, 6.1, 6.11). Another
prime topic of the book is the regularity of solutions of elliptic boundary value
problems in polygonal and polyhedral domains and hence of solutions of the
corresponding boundary integral equations on polygonal curves and polyhedral
surfaces. Here, the reduced regularity of the solution near corners and edges requires
special boundary element methods like enrichment by singularity functions or the
use of graded meshes or hp-techniques. The latter are investigated in detail, and
especially exponentially fast hp-methods are described. Another way to tackle the
loss of regularity of the solutions is to use adaptive boundary element methods,
also described in detail for h, and p-versions. A further prime topic is the use of
BEM for unilateral contact problems and thus the analysis of boundary variational
inequalities. Furthermore, the symmetric FEM/BEM coupling is analyzed and
various applications are given. Also, the time-domain boundary element method
is investigated for the time-dependent acoustic scattering. Important mathematical
tools for the analysis, presented here, are Fourier and Mellin transform together with
pseudodifferential operator techniques.
Preface ix
Chapter 1 introduces to the theory of approximation methods for the solution of

operator equations and for the solution of related variational problems. Chapter 2 is
of introductory character and gives the standard approach from potential theory to
boundary integral equations. Chapter 3 introduces the concept of periodic Sobolev
spaces with the help of Fourier series and constructs the solution to the interior
and exterior Dirichlet problems for the Laplacian on the unit sphere by a Fourier
series approach. The mapping properties of weakly singular and hypersingular
boundary integral operators are analyzed by Fourier series; with this tool, a Gårding
inequality is derived for the first time. A perturbation argument allows us to go
from the unit circle to smooth curves. Chapter 4 deals first with smooth surfaces
and uses the concept of Fourier transformation and pseudodifferential operators
to treat the mixed Dirichlet–Neumann BVPs for the Laplacian and the acoustic
interface problem with the Helmholtz equation as well as crack/screen problems and
time-harmonic exterior Maxwell problems. Generalizations to Lipschitz curves and
surfaces are done for interface problems in linear elasticity in Sect. 4.4. In Chap. 5,
we present the boundary integral approach for the scalar Signorini problem with the
Laplacian and for unilateral contact problems without and with friction; in addition,
we treat nonmonotone contact problems from delamination by a combination of
boundary integral methods and regularization techniques from nondifferentiable
optimization. Chapter 6 starts with an abstract setting for the Galerkin method
for strongly elliptic operator equations. The h-version BEM (Galerkin and collo-
cation) is presented in the frame of general projection methods. Sections 6.1–6.5
provide some fundamental facts of BEM, including BEM on quasiuniform meshes,
Aubin–Nitsche duality estimate, superapproximation, and local/L∞ error estimates.
Sections 6.6–6.10 cover special topics like discrete collocation, augmentation of
the boundary element space by special singular functions, and modified collocation
and qualocation. In Sect. 6.11, a meshless method with radial basis functions is
presented for integral equations of the first kind; herewith scattered satellite data
can be accounted for. Chapter 7 is devoted to the hp-version BEM on polygonal
and polyhedral domains using first uniform meshes. Then the results for the h-
version are extended to graded meshes. Chapter 8 presents the hp-version of the
BEM on geometrically refined meshes and shows its exponentially fast convergence.
In Chap. 9, the notion of Mellin symbols for the boundary integral operators is
introduced and their mapping properties in countably normed spaces are derived.
Chapter 10 is devoted to adaptive boundary element methods. The results on residual
error estimators for integral equations on curves are given in Sects. 10.1–10.2 and
on surfaces in Sects. 10.3–10.4. Special emphasis is given to the two-level approach
with hierarchical error estimator in Sect. 10.5 for the h-version and Sect. 10.6 for the
p-version. The convergence of adaptive BEM schemes is investigated in Sect. 10.7.
Chapter 11 extends the BEM for contact problems with special emphasis on the use
of Gauss–Lobatto–Lagrange basis functions for the hp-version in Sects. 11.2,11.3
and of biorthogonal basis functions in the mixed scheme in Sect. 11.4. Sec-
tions 11.5, 11.5 combine regularization techniques of nondifferentiable optimization
with h-BEM or hp-BEM for delamination problems. Chapter 12 overviews the
symmetric FEM/BEM coupling method. Interface problems together with contact
x Preface
conditions and strongly nonlinear operators in the FEM domain are analyzed.
Also, different mixed formulations (primal/dual) are considered. Moreover, least
squares coupling methods are studied. Further, the symmetric coupling for the
time-harmonic eddy current problem from electromagnetics is addressed. Also, for
a parabolic-elliptic interface problem a FEM/BEM coupling is given. The final
Chap. 13 considers dynamic scattering and contact problems and uses the tool of
retarded potentials to obtain Galerkin approximations with the TD-BEM based on
marching-on-in-time (MOT) schemes. The Appendix supports reading of the book
and has 4 parts: In Appendix A, we give the fundamentals of linear operator theory.
In Appendix B, we present a short introduction into pseudodifferential operators.
In Appendix C, we collect some aspects on variational inequalities and convex and
nonsmooth analysis. Finally in Appendix D, we describe the implementation of the
BEM for some representative examples on curves and surfaces.
The introductory part of this monograph (Sects. 1.1–6.5 and Appendix) grew out
of lecture notes from courses given by the authors at the Universität der Bundeswehr
München and at the Leibniz Universität Hannover, whereas the other sections deal
with research topics.
First of all, we want to thank our wives Hannelore Raith and Karin Sabine
Stephan for their great understanding and support during the work-intensive time, it
took us to write our book.
The authors thank their colleagues L. Banz, C. Carstensen, A. Chernov, M.
Costabel, J. Elschner, G. Gatica, H. Gimperlein, N. Heuer,F. Leydecker, M.
Maischak, P. Mund, N. Ovcharova, D. Praetorius, T. von Petersdorff, T. Tran, and
W.L. Wendland for their cooperation which has highly influenced the contents of
the book. Especially we thank C. Özdemir for his continuous, generous, and very
pleasant support in producing the manuscript.
München, Germany Joachim Gwinner

Hannover, Germany Ernst Peter Stephan
2018
Preface xi
first kind BIE’s for general projection methods

BVP’s, transmission, contact Chapter 1,
Chapter 2,4,5 Section 6.1,6.2,6.8,10.1,12.1
standard BEM advanced BEM

h-version h,p,hp-versions
Chapter 6 Chapter 7,8
Tools:
Fourier series (Chapter 3)
Fourier transformation (Chapter 4)
Mellin transformation (Chapter 9)
linear operator theory (Appendix A)
A-BEM BEM for contact
pseudodifferential operators
Chapter 10 Chapter 11
(Appendix B)
variational inequalities, convex and
nonsmooth analysis (Appendix C)
some implementations
for BEM (Appendix D)
FE/BE coupling TD-BEM

Chapter 12 Chapter 13
Contents
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 The Basic Approximation Problems . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Convergence of Projection Methods . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
2 Some Elements of Potential Theory . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.1 Representation Formulas .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.2 Single- and Double-Layer Potential .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
2.2.1 Some Remarks on Distributions . . . . . . .. . . . . . . . . . . . . . . . . . . . 17
2.2.2 Jump Relations . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
2.3 Mapping Properties of Boundary Integral Operators .. . . . . . . . . . . . . . 25
2.4 Laplace’s Equation in R3 . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 30
2.4.1 Representation Formula . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 32
2.5 Calderon Projector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
2.6 Use of Complex Function Theory . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
2.6.1 Representation Formula Again.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
2.6.2 Applicable Representation of the Hypersingular
Integral Operator.. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
3 A Fourier Series Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
3.1 Fourier Expansion—The Sobolev Space H s [0, 2π].. . . . . . . . . . . . . . . 43
3.2 The Sobolev Space H s (Γ ) . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 48
3.3 Interior Dirichlet Problem . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
3.4 The Boundary Integral Operators in a Scale of Sobolev Spaces . . . 52
3.4.1 The Operators V and W . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 52
3.4.2 The Operators K and K . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 55
3.5 Solution of Exterior Dirichlet Problem by BIE .. . . . . . . . . . . . . . . . . . . . 57
3.6 A First Gårding Inequality .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 60
4 Mixed BVPs, Transmission Problems and Pseudodifferential
Operators .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
4.1 Mixed Boundary Value Problems . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
4.2 The Helmholtz Interface Problems . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 70
xiii
xiv Contents
4.3 Screen Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 81

4.4 Interface Problem in Linear Elasticity. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
4.5 A Strongly Elliptic System for Exterior Maxwell’s Equations.. . . . 89
4.5.1 A Simple Layer Procedure . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 89
4.5.2 Modified Boundary Integral Equations .. . . . . . . . . . . . . . . . . . . 91
5 The Signorini Problem and More Nonsmooth BVPs and Their
Boundary Integral Formulation . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
5.1 The Signorini Problem in Its Simplest Form .. . .. . . . . . . . . . . . . . . . . . . . 95
5.2 A Variational Inequality of the Second Kind Modelling
Unilateral Frictional Contact.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 102
5.3 A Nonmonotone Contact Problem from Delamination.. . . . . . . . . . . . 106
6 A Primer to Boundary Element Methods . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 115
6.1 Galerkin Scheme for Strongly Elliptic Operators.. . . . . . . . . . . . . . . . . . 116
6.2 Galerkin Methods for the Single-Layer Potential.. . . . . . . . . . . . . . . . . . 119
6.2.1 Approximation with Trigonometric Polynomials . . . . . . . . . 119
6.2.2 Approximation with Splines . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121
6.3 Collocation Method for the Single-Layer Potential . . . . . . . . . . . . . . . . 124
6.4 Collocation Methods—Revisited .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 126
6.4.1 Periodic Splines as Test and Trial Functions . . . . . . . . . . . . . . 128
6.4.2 Convergence Theorem for Projection Methods . . . . . . . . . . . 131
6.5 BEM on Quasiuniform Meshes .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 140
6.5.1 Periodic Polynomial Splines . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 140
6.5.2 The Approximation Theorem . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 141
6.5.3 Stability and Inverse Estimates . . . . . . . .. . . . . . . . . . . . . . . . . . . . 147
6.5.4 Aubin-Nitsche Duality Estimate
and Superapproximation . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
6.5.5 Numerical Quadrature .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 155
6.5.6 Local H −1/2-Error Estimates . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
6.5.7 Local L2 -Error Estimates . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 163
6.5.8 The K-Operator-Method . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 165
6.5.9 L∞ -Error Estimates for the Galerkin Approximation . . . . 168
6.6 A Discrete Collocation Method for Symm’s Integral
Equation on Curves with Corners . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 172
6.7 Improved Galerkin Method with Augmented Boundary
Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 182
6.8 Duality Estimates for Projection Methods . . . . . .. . . . . . . . . . . . . . . . . . . . 185
6.8.1 Application to Galerkin Methods . . . . . .. . . . . . . . . . . . . . . . . . . . 186
6.8.2 Application to Collocation Methods .. .. . . . . . . . . . . . . . . . . . . . 189
6.9 A Collocation Method Interpreted as (GM) .. . . .. . . . . . . . . . . . . . . . . . . . 192
6.10 Modified Collocation and Qualocation . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 198
6.11 Radial Basis Functions and Spherical Splines . .. . . . . . . . . . . . . . . . . . . . 205
Contents xv
7 Advanced BEM for BVPs in Polygonal/Polyhedral Domains:

h- and p-Versions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 223
7.1 The Dirichlet Problem . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 224
7.1.1 Regularity on a Polygon .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 225
7.1.2 BEM on a Polygon . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 226
7.1.3 Regularity on a Polyhedron . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 230
7.2 The Neumann Problem.. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 236
7.2.1 Regularity on a Polyhedron . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 240
7.3 1D-Approximation Results . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 242
7.3.1 hp-Method with Quasiuniform Mesh on Polygons . . . . . . . 242
7.3.2 Approximation of the Normal Derivative on a
One Dimensional Boundary—The h-Version on a
Graded Mesh .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 247
7.4 2D-Approximation Results . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
Two-dimensional Boundary—The h-Version
on a Graded Mesh . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 250
7.4.2 Approximation of the Trace on a Two-Dimensional
Boundary—The h-Version on a Graded Mesh . . . . . . . . . . . . 257
7.5 Augmented BEM for Screen/Crack Problems . .. . . . . . . . . . . . . . . . . . . . 264
8 Exponential Convergence of hp-BEM . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 269
8.1 The hp-Version of BEM on Polygons . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 270
8.1.1 Application to Acoustic Scattering . . . .. . . . . . . . . . . . . . . . . . . . 279
8.2 The hp-Version of BEM on Surfaces . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 282
8.3 The hp-Version of BEM on a Geometrical Mesh for Mixed
BVP on a Polygonal Domain . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 288
9 Mapping Properties of Integral Operators on Polygons . . . . . . . . . . . . . . . 295
9.1 Mellin Symbols.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 295
9.1.1 Mapping Properties in Weighted Sobolev Spaces.. . . . . . . . 299
9.2 Properties of the Mellin Transformation . . . . . . . .. . . . . . . . . . . . . . . . . . . . 306
9.2.1 Local Regularity at Vertices . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 310
9.3 A Direct Boundary Element Method for Interface Crack
Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 313
9.4 Mixed BVP of Potential Theory on Polygons . .. . . . . . . . . . . . . . . . . . . . 317
9.5 Boundary Integral Operators in Countably Normed Spaces . . . . . . . 323
10 A-BEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 333
10.1 General Frame for A Posteriori Error Estimates for
Boundary Element Methods . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 334
10.1.1 Symm’s Integral Equation .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 336
10.2 Adaptive Boundary Element Methods . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 337
10.2.1 Reliability of A Posteriori BEM Error Estimates . . . . . . . . . 340
10.2.2 Efficiency of A Posteriori BEM Error Estimates (2D) .. . . 343
xvi Contents
10.3 The Weakly Singular Integral Equation in 3D . .. . . . . . . . . . . . . . . . . . . . 349

10.3.1 Adaptive Algorithms . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
10.3.2 Numerical Example . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 355
10.4 The Hypersingular Integral Equation in 3D . . . . .. . . . . . . . . . . . . . . . . . . . 357
10.5 Two-Level Adaptive BEM for Laplace, Lamé, Helmholtz .. . . . . . . . 362
10.5.1 A Stable Two-Level Subspace Decomposition for
the Hypersingular Operator . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 370
10.6 Two-Level Subspace Decomposition for the p-Version BEM . . . . . 377
10.7 Convergence of Adaptive BEM for Estimators Without
h-Weighting Factor .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 381
11 BEM for Contact Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 389
11.1 h-BEM for the Signorini Problem .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 390
11.1.1 Discretization of the Boundary Variational Inequality .. . . 390
11.1.2 The Convergence Result . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 392
11.2 hp-BEM with Hierarchical Error Estimators for Scalar
Signorini Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 395
11.3 hp-BEM for a Variational Inequality of the Second Kind
Modelling Unilateral Contact and Friction. . . . . .. . . . . . . . . . . . . . . . . . . . 403
11.3.1 The hp-Version Galerkin Boundary Element Scheme .. . . 405
11.3.2 A Céa-Falk Lemma for Variational Inequalities
of the Second Kind . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 412
11.3.3 A Priori Error Estimate for hp-Approximation . . . . . . . . . . . 414
11.4 Mixed hp-BEM for Frictional Contact Problems .. . . . . . . . . . . . . . . . . . 420
11.4.1 Boundary Integral Formulation for Contact Problem .. . . . 420
11.4.2 hp-Boundary Element Procedure with Lagrange
Multiplier and Fast Solver.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 422
11.4.3 Error Controlled hp-Adaptive Schemes . . . . . . . . . . . . . . . . . . . 425
11.4.4 Stabilized hp-Mixed Method—A Priori Error
Estimate .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 430
11.4.5 A Priori Error Estimates for hp-Penalty-BEM for
Contact Problems in Elasticity . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 431
11.5 h-Version BEM for a Nonmonotone Contact Problem from
Delamination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 436
11.6 hp-BEM for Delamination Problems . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 443
12 FEM-BEM Coupling .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 451
12.1 Abstract Framework of Some Saddle Point Problems .. . . . . . . . . . . . . 452
12.2 Galerkin Approximation of Saddle Point Problems . . . . . . . . . . . . . . . . 455
12.2.1 Symmetric FE/BE Coupling for a Nonlinear
Interface Problem .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 459
12.3 Symmetric FE/BE Coupling—Revisited .. . . . . . .. . . . . . . . . . . . . . . . . . . . 464
12.3.1 Convergence Analysis . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 468
12.3.2 Adaptive FE/BE Coupling: Residual Based Error
Indicators.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 473
Contents xvii
12.3.3 Adaptive FE/BE Coupling with a Schur

Complement Error Indicator . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 478
12.3.4 Convergence of Adaptive FEM-BEM Couplings . . . . . . . . . 488
12.3.5 Other Coupling Methods .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 489
12.4 Least Squares FEM/BEM Coupling for Transmission
Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 491
12.4.1 The Discretized Least Squares Formulation.. . . . . . . . . . . . . . 497
12.5 FE/BE Coupling for Interface Problems with Signorini
Contact .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 499
12.5.1 Primal Method .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 499
12.5.2 Dual Mixed Method . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 503
12.6 Coupling of Primal-Mixed FEM and BEM for Plane Elasticity . . . 506
12.7 Adaptive FE/BE Coupling for Strongly Nonlinear Interface
Problems with Tresca Friction .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 515
12.8 Adaptive FE-BE Coupling for the Eddy-Current Problem in
R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 519
12.8.1 p-Hierarchical Estimator .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 530
12.9 Parabolic-Elliptic Interface Problems . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 533
13 Time-Domain BEM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 537
13.1 Integral Equations and Anisotropic Space-Time Sobolev
Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 538
13.2 A Priori and A Posteriori Error Estimates. . . . . . .. . . . . . . . . . . . . . . . . . . . 543
13.2.1 Adaptive Mesh Refinements . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 547
13.3 Time Domain BEM for Contact Problems .. . . . .. . . . . . . . . . . . . . . . . . . . 549
13.4 Algorithmic Aspects of Time Domain BEM . . . .. . . . . . . . . . . . . . . . . . . . 552
13.4.1 MOT Algorithm . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 552
13.4.2 An hp-Composite Quadrature of Matrix Elements . . . . . . . 554
13.5 Screen Problems and Graded Meshes . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 557
A Linear Operator Theory .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 563
B Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 569
C Convex and Nonsmooth Analysis, Variational Inequalities .. . . . . . . . . . . 581

C.1 Convex Optimization, Lagrange Multipliers . . . .. . . . . . . . . . . . . . . . . . . . 581
C.1.1 Convex Quadratic Optimization in Finite Dimensions . . . 582
C.1.2 Convex Quadratic Optimization in Hilbert Spaces. . . . . . . . 586
C.1.3 Lagrange Multipliers for Some Inequality
Constrained Variational Inequalities .. .. . . . . . . . . . . . . . . . . . . . 589
C.2 Nonsmooth Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 593
C.2.1 Nonsmooth Analysis of Locally Lipschitz Functions.. . . . 593
C.2.2 Regularization of Nonsmooth Functions . . . . . . . . . . . . . . . . . . 595
xviii Contents
C.3 Existence and Approximation Results for Variational

Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 601
C.3.1 Existence Results for Linear VIs. . . . . . .. . . . . . . . . . . . . . . . . . . . 601
C.3.2 Approximation of Linear VIs . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 605
C.3.3 Pseudomonotone VIs—Existence Result. . . . . . . . . . . . . . . . . . 608
C.3.4 Mosco Convergence, Approximation of
Pseudomonotone VIs . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 610
C.3.5 A Hemivariational Inequality as a
Pseudomonotone VI . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 611
D Some Implementations for BEM . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 615

D.1 Symm’s Equation on an Interval.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 615
D.2 The Dirichlet Problem in 2D. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 616
D.3 Symm’s Equation on a Surface Piece . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 618
D.3.1 Implementation of hp-BEM on Surfaces . . . . . . . . . . . . . . . . . . 622
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 631
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 651
Chapter 1
Introduction
This chapter gives an introduction to the theory of approximation methods for the
solution of operator equations and for the solution of related variational problems.
In the first section we formulate the basic approximation problems and their setting.
Then in the following section we present a collection of various examples and model
applications in a simplified way. In the following chapters we shall elaborate at
these examples at the more deeper level of boundary value problems that arise from
diverse fields of mathematical physics. Then we shall reformulate these boundary
value problems as first kind integral equations and focus to boundary element
methods for their numerical treatment.
In here a heuristic approach is given in order to show briefly the fundamental
questions in the theory of approximation methods. This chapter should motivate
the reader to go into the next chapters with at least some knowledge about what is
going on.
Furthermore one can read this chapter a second time after the development of
the boundary element methods to learn how these methods are related to general
methods, as for example with Galerkin‘s method for the solution of operator
equations.
We write f g provided there exists a constant C such that f ≤ Cg.
1.1 The Basic Approximation Problems
Problem 1.1 Let A : X → Y be a continuous linear operator between two

separable Banach spaces X and Y . The question is how to find a u ∈ X such that,
for given f ∈ Y , Au = f holds ? The idea of projection methods is to solve the
© Springer International Publishing AG, part of Springer Nature 2018 1

J. Gwinner, E. P. Stephan, Advanced Boundary Element Methods,
Springer Series in Computational Mathematics 52,
https://doi.org/10.1007/978-3-319-92001-6_1
2 1 Introduction
above equation in certain subspaces XN ⊂ X, YN ⊂ Y :
AN uN = fN ∈ YN for N ∈ N ,
where AN is an approximation to A and to hope that uN ∈ XN is a “reasonable”

approximation to u ∈ X.
Remark 1.1 Obvious requirements for a “successful” approximation method such
that uN ∈ XN converges to u for N → ∞ are
1. The perturbed right hand sides fN should converge to the given right hand side
f ∈ Y.
2. The subspaces XN should exhaust the entire space X in the sense that

XN = X .
N∈N
A more delicate question is how the operators AN are defined on the subspaces XN
and how these operators should approximate the given operator A. Let us note that
the simple choice of AN := A|XN , the restriction of A to the subspace XN seldom
works in the applications.
Linear operator equations abound in applied mathematics. They result from a
linear problem modelling or – more often – within a linearization procedure, as e.g.
Newton’s method, for genuine nonlinear problems. Here we concentrate on well-
posed operator equations that arise from the classical problems of mathematical
physics. Thus as an introductory example we have Symm’s integral equation:

1
V ψ(x) = − ln |x − y|ψ(y)dsy = f (x), (1.1)
π Γ
where Γ is the boundary of a bounded domain in R2 , such that (1.1) corresponds to

A = V , X = H −1/2 (Γ ), Y = H 1/2(Γ ) (see Chapter 2). An equally basic problem
is that of a variational problem, that is the minimization of a functional without or
under some constraints. We start with the most simple variational problem, namely
with the unconstrained one.
Problem 1.2 Assume that A : X → X is a continuous linear operator from
a Banach space X into its dual X . Assume that the quadratic form Ax, x is
nonnegative. Further let some l ∈ X be given. The question now is how to find
a x ∈ X that minimizes the functional
1
F (x) = Ax, x − l(x)
2
in X. The idea of approximation methods (in particular the so-called Ritz method)
is to solve the above variational problem in certain subspaces XN ⊂ X and likewise
to “hope” that these (approximate) solutions uN converge to u.
1.1 The Basic Approximation Problems 3
Note for Symm’s integral equation the quadratical functional becomes
1
F (ψ) = V ψ, ψ − l(ψ), ψ ∈ H −1/2(Γ ),
2
where ·, · denotes the duality between H −1/2 (Γ ) and H 1/2(Γ ) and l(ψ) =
f, ψ . Considering a polygon Γ we can take XN as space of piecewise constants
on quasi-uniform mesh where the vertices of Γ belong to the mesh points.
A variational constrained problem in general terms is the following.
Problem 1.3 Assume that F : X → R is a continuous functional on a Banach
space X. Moreover, let C be a closed subset of X. The question now is how to find
a u ∈ C that minimizes F in C. The idea of projection methods is to solve the
above variational problem in certain subsets CN ⊂ XN , where again XN ⊂ X are
subspaces (of finite dimension), and likewise to hope that the associated minimizers
uN of F in CN converge to u.
Here F may be of the form of Problem 1.2. In the most simple case, the set C is
an affine subspace of X, but may be more generally a convex cone or a convex
subset. The approximation problem becomes more delicate if CN ⊂ C does not
hold (so-called nonconforming approximation).
A typical example is:

1 1 ∂u
Minimize F (u) = u, Su − l(u) = u ds − f, u (1.2)
2 2 Γ ∂n
subject to u ≤ g on Γc . (1.3)
Here Γc is a part of Γ and S : H 1/2(Γ ) → H −1/2 (Γ ) is the Dirichlet-to-Neumann

map (Poincaré-Steklov operator), see Chap. 5.
When the functional F is convex and differentiable, the minimization of F on C
is equivalent to the variational inequality: Find x ∈ C such that
F (x), y − x ≥ 0, ∀y ∈ C.
Here F (x), z denotes the directional derivative of F at x in direction z.

In the model application above we obtain the variational inequality (VI):
Find u ∈ C such that
Su, v − u ≥ f, v − u ∀v ∈ C,
where C := {v ∈ H 1/2 (Γ )|v|Γc ≤ g}. Now we may choose XN as continuous

piecewise linear functions on a quasi-uniform mesh and CN is given as those
functions uN ∈ XN that satisfy uN ≤ g in the mesh points, see Chap. 11.
4 1 Introduction
1.2 Convergence of Projection Methods
Now we consider the situation of the general Galerkin method, i.e. for Hilbert spaces
X, Y and a linear, continuous and bijective mapping, A : X → Y we want to find
an approximation uN of the solution u ∈ X of
Au = f (1.4)
for f ∈ Y given. Thus let XN ⊂ X and TN ⊂ Y be the spaces of trial and test
functions, respectively, with dim TN = dim XN = N < ∞ . Then we want to find
uN ∈ XN such that
t, AuN = t, f ∀ t ∈ TN . (1.5)
Theorem 1.1 (Galerkin Method) Let X = Y , Y = X and TN = XN for the

above situation. If A is positive definite, i.e. ∃ α > 0 : x, Ax ≥ α||x||2X ∀ x ∈ X,
the following holds:
1. Existence of a unique solution
∀ N ∃! uN ∈ XN : t, AuN = t, f = t, Au ∀ t ∈ TN
2. Stability of the method
∃ M(independent of N) : ||uN ||X ≤ M||u||X
3. Quasioptimal error estimate
∃ C(independent of N) : ||u − uN ||X ≤ C inf ||u − v||X =: C d(u, XN )
v∈XN
4. Convergence of the method
N→∞ N→∞
d(v, XN ) −→ 0 ∀ v ∈ X ⇒ ||u − uN ||X −→ 0
Proof As already suggested above, let {b1 , . . . , bN } and {t1 , . . . , tN } be a basis of
XN and TN respectively, leading to

N
αk tj , Abk = tj , f , j = 1, . . . , N . (1.6)
k=1

1. With A being positive definite, the matrix bj , Abk j,k=1...N is positive definite,
too. Hence it is invertible, implying the existence of a unique solution uN .
2. The assumption of A being positive definite further yields
1 1 1
uN 2 ≤
uN , AuN = uN , f = uN , Au
α α α
1 1
≤ uN X AuX ≤ uN X AX→X uX
α α
1
⇒ uN X ≤ AX→X uX =: MuX
α
1.2 Convergence of Projection Methods 5
3. First of all, the triangle inequality yields for all v ∈ XN
u − uN X ≤ u − vX + v − uN X ,
where uN is defined
by t, AuN = t, Au ∀ t ∈ XN ,
X −→ XN
Let GN : be the so-called Galerkin projector.
u → uN
This linear operator is continuous by 2., further a projector onto XN , since
GN (v) = v , for all v ∈ XN (because with t, AvN = t, Av ∀ t ∈ XN and
GN v := vN for vN ∈ XN unique, it follows v = vN ).
For the Galerkin projector we have
AX→X AX→X
uN X = GN uX ≤ uX ⇒ |||GN ||| ≤ M =
α α
Hence for all v ∈ XN :
u − uN X = u − v + v − uN X = u − v + GN v − GN uX
= (1 − GN )(u − v)X ≤ (1 + |||GN |||)u − vX

AX→X
⇒ u − uN X ≤ 1 + u − vX
α
Thus 3.,

AX→X
u − uN X ≤ C inf u − vX , C := 1 + .
v∈XN α
4. By 3., lim inf u − vX = 0 ⇒ lim u − uN X = 0.

N→∞ v∈XN N→∞
For the situation of the general Petrov-Galerkin method, i.e. with X = Y and
XN = TN we shall consider the following stability criteria:
a)
∃ α > 0(independent of N) ∀ v ∈ XN ∃ t ∈ TN \{0} :

(1.7)
|t, Av | ≥ αvX · tY
b)
|t,Av |
∃α > 0 : inf sup ≥ α
v∈XN \{0} t ∈TN \{0} vX t Y (1.8)
Babuška − Brezzi − condition
6 1 Introduction
c)
∃ QN : XN → TN , ∃ M > 0 ∃ α > 0(independent of N) :

|QN v, Aw | ≤ MvX wX ∀ v ∈ XN , w ∈ X (1.9)
|QN v, Av | ≥ αv2X ∀ v ∈ XN
Remark 1.2 It is easily verified that the above three conditions are equivalent. For
the Babuška − Brezzi − condition (1.8) see also e.g. [56, 60, 65].
Theorem 1.2 (A More General Projection Method) Let one (and thus all) of the
above stability criteria (1.7)–(1.9) be satisfied, then the statements of Theorem 1.1
also hold for the projection method (1.5).
Proof
1. We will show that the kernel of the matrix defined in (1.6) only consists of 0: Let
αk tj , Abk = 0, for j = 1, . . . , N , what is
k

t, AuN = 0 ∀ t ∈ TN (with uN = αk bk ).
k
By (1.7) ∃ t ∈ TN \{0} : 0 = t, AuN ≥ αuN X tY . Hence uN = 0. Thus

the matrix is injective. Since it is also quadratic it must be bijective, assuring the
existence of a unique solution uN .
2. uN X ≤ α1 t 1 |t, AuN | (using (1.7) with v = uN )
Y
= α t Y |t, Au | ≤ α t Y tY AuY
1 1 1 1
= α AuY ≤ α |||A|||uX
1 1
3. The Galerkin projector GN is well-defined by 1. Furthermore, 2. yields |||GN ||| ≤

α |||A||| . Thus, for all v ∈ XN
1
u − uN X ≤ u − vX + GN (u − v)X

≤ (1 + |||GN |||)u − vX
|||A|||
≤ 1+ α u − vX
=: Cu − vX ∀ v ∈ XN
4. Follows directly from 3.

The following result by Hildebrandt and Wienholtz [244] generalizes Theorem 1.1
to strongly elliptic operators.
Theorem 1.3 Let X be a Hilbert space with dual X and A, D : X → X
isomorphisms such that T = A − D : X → X compact. Let {Sh }h>0 be a family of
subspaces of X such that the equations
Dwh , v = Dw, v for all v ∈ Sh (1.10)

1.2 Convergence of Projection Methods 7
define an operator GhD : w ∈ X → wh ∈ Sh with the property
GhD w − w → 0 as h → 0 for all w ∈ X. (1.11)
Then for small h > 0 the equations
Auh , v = Au, v for all v ∈ Sh (1.12)
define an operator GhA : u ∈ X → uh ∈ Sh such that
GhA ≤ C
with C independent of h.
Proof From (1.11) and the compactness of T follows A−1 T (1 − GhD ) → 0
(h → 0). Therefore for small h G̃hA := GhD [1 − A−1 T (1 − GhD )]−1 exists and
G̃hA is uniformly bounded. From equations (1.10) and (1.12) it is easily verified
that G̃hA = GhA .

Chapter 2
Some Elements of Potential Theory
In this chapter we collect well-known concepts and results of classical potential

theory that are necessary for the understanding of BEM. We focus to the elementary
level of the Laplace equation in R2 , respectively in R3 in Sect. 2.4.
First by the Gauss divergence theorem and classical limit arguments we derive
the representation formula. This leads to the definition of the single- and double-
layer potential. Then based on a distribution approach we provide the jump
relations of the associated boundary integral (BI) operators and study their mapping
properties, first in classical spaces of smooth functions, then by extension in the
relevant Sobolev spaces of fractional order. Here we are concerned with the capacity
of a Lipschitz curve what provides a sufficient condition for positive definiteness
of the single-layer potential operator. Also we discuss in detail the bijectivity of
the hypersingular operator. All 4 singular BI operators (single-layer, double-layer,
adjoint of double-layer, hypersingular) enter the Calderon projector what expresses
the jump relations in a compact form.
We conclude this chapter with another elementary approach based on complex
function theory. Here again we derive the representation formula and are concerned
with the single-layer and the hypersingular operator; for the latter we give an
applicable representation.
2.1 Representation Formulas
Let Ω be a bounded domain in R2 with smooth boundary Γ . We can later relax this
assumption on Γ .

10 2 Some Elements of Potential Theory
Let us first recall the basic Green formula for the Laplacian Δ:

∂u ∂v
(Δu v − u Δv) dx = v−u ds , (2.1)
Ω Γ ∂n ∂n
∂
where ∂n denotes the outer normal derivative, that is, the directional derivative in
the direction of the outer normal unit vector n that points towards the exterior of Ω.
In view of x log(|x − y|) = 0, for x = y, |x| → log |x| is a fundamental solution
of the Laplacian in R2 .
For later use we note
x−y ∂ x − y, n(x)
∇x log(|x − y|) = , log(|x − y|) = .
|x − y|2 ∂n(x) |x − y|2
We now want to prove the representation formula for the Laplace equation.
Theorem 2.1 (Representation Formula) Let Ω be a bounded simply connected
region with a smooth boundary Γ = ∂Ω, and let u ∈ C 2 (Ω̄). Then there holds
the following representation for u in Ω :

1
u(x) = ln |x − y|Δu(y) dy + (2.2)
2π
Ω

1
+ u(y)∂ny ln |x − y| − ∂ny u(y) ln |x − y| dsy , x∈Ω
2π
Γ
Proof Consider the following Fig. 2.1.

Let J (x) := ln |x − y| Δu(y) dy. Then using (2.1)
Ω\B (x)
⎛ ⎞

⎜ ⎟
J (x) = ⎝ln |x − y| Δu(y) − u(y) Δy ln |x − y|⎠ dy

Ω\B (x) =0
Fig. 2.1 Geometrical setting
Ω
x
B (x)
2.1 Representation Formulas for Laplace’s Equation 11

= ln |x − y| ∂n u(y) − u(y) ∂ny ln |x − y| dsy
∂(Ω\B (x))

= ln |x − y| ∂n u(y) − u(y) ∂ny ln |x − y| dsy + J ∗ ,
∂Ω

where J ∗ (x) := − ln |x − y| ∂n u(y) − u(y) ∂ny ln |x − y| dsy .
∂B (x)
With u ∈ C 2 (Ω) , we further have
2π
ln |x − y|∂n u(y) dsy = ln · d
dr u(x + reiϕ )|r= · dϕ
∂B (x) 0
2π
= · ln · ∂r u(x + reiϕ )|r= dϕ
0
→0
≤ · ln · sup ∇u(x) · 2π −→ 0 ,
x∈B
2π
u(y)∂ny ln |x − y| dsy = u(x + · eiϕ ) dr
d
ln r|r= · dϕ
∂B (x) 0
2π
→0
= u(x + · eiϕ )dϕ −→ 2πu(x) .
0
Thus, the assertion follows with ln |x − y| Δu(y) dy = lim J .

Ω →0
Remark 2.1 Setting G(x, y) = 2π 1

ln |x − y|, we leave it as an exercise to the reader
to show that for f ∈ C0 (R ) (space of continuous functions with compact support
0 2
supp (f )) the function

u(x) = G(x, y)f (y)dy
R2
satisfies the Poisson equation Δu = f in R2 .

Corollary 2.1 Let G define the integral operator f ∈ C00 (R2 ) → Gf by

1
Gf (x) := G(x, y)f (y) dy = ln |x − y|f (y) dy
2π
R2 R2
Then we have for u ∈ C02 (R2 )
u = GΔu = ΔGu. (2.3)

Proof The equation u = ΔGu follows from Remark 2.1. Take Ω such that
supp (u) ⊂⊂ Ω, then the representation formula (2.2) further yields u(x) =
GΔu(x), ∀x ∈ Ω.

Let Ω := R2 \Ω denote the ”exterior” domain. Then the basic ”Green’s represen-
tation formula” for a harmonic function (see (2.5) below) can be stated as follows.
Theorem 2.2 Let u ∈ C 2 (Ω) ∩ C 2 (Ω ). Assume there exist the limits
u|int Γ (x) = lim u(z), u|ext Γ (x) := lim u(z) ,

z→x∈Γ z→x∈Γ
z∈Ω z∈Ω
∂u

and the analogously defined limits ∂n int Γ , ∂u
∂n ext Γ . Let
[u(x)] := u(x)|int Γ − u(x)|ext Γ

∂u(x) ∂u(x) ∂u(x)
:= |int Γ − |ext Γ
∂n ∂n ∂n
be the jump of the trace and of the normal derivative of u, respectively.

Moreover let u satisfy
⎧
⎪
⎪ Δu = 0 in Ω ∪ Ω
⎪
⎪
⎪
⎪ ⎫
⎨ 1 ⎪
u(y) = O ⎪
⎬ (2.4)
⎪
⎪ |y|
⎪ for |y| → +∞
⎪
⎪ 1 ⎪ ⎪
⎪
⎩ |∇u(y)| = O ⎭
|y|2
Then there holds for y ∈ Ω ∪ Ω ,

1 ∂u(x)
u(y) = − log(|x − y|) ds(x) (2.5)
2π Γ ∂n
"
∂
− [u(x)] (log(|x − y|) ds(x) ,
Γ ∂n(x)
and for y ∈ Γ ,

u(y)|int Γ + u(y)|ext Γ 1 ∂u(x)
=− log(|x − y|)ds(x) (2.6)
2 2π Γ ∂n
"
∂
− [u(x)] log(|x − y|) ds(x) .
Γ ∂n(x)
Proof Let y ∈ Ω ∪ Ω . Choose ε > 0 such that the ε-ball B(y, ε) with boundary
Sε is contained in Ω, respectively in Ω , and moreover Ω ∪ B(y, ε) is contained in
the R–ball B(0, R) with boundary SR for large enough R > 0.
Since Δu(x) = 0, Δ log(|x − y|) = 0 for x = y, we can apply Green’s formula
both in Ω \ B(y, ε) and in Ω ∩ B(0, R) \ B(y, ε). Thus we obtain for y ∈ Ω ∪ Ω
the following equations
"
∂ ∂u(x)
u(x)|int Γ log(|x − y|) − |int Γ log(|x − y|) ds(x)
Γ ∂n(x) ∂n(x)
"
∂ ∂u(x)
+ u(x) log(|x − y|) − log(|x − y|) ds(x) = 0
Sε ∂n(x) ∂n(x)
"
∂ ∂u(x)
−u(x)|ext Γ log(|x − y|) + |ext Γ log |x − y|) ds(x)
Γ ∂n(x) ∂n(x)
"
∂ ∂u(x)
SR (∪Sε ) ∂n(x) ∂n(x)
By addition

∂ ∂u(x)
[u(x)] log(|x − y|) ds(x) − log(|x − y|) ds(x)
Γ ∂n(x) Γ ∂n(x)
"
∂ ∂u(x)
+ u(x) log(|x − y|) − log(|x − y|) ds(x)
SR ∂n(x) ∂n(x)
"
∂ ∂u(x)
Sε ∂n(x) ∂n(x)
Let ε → 0. Then

∂u(x)
| log(|x − y|) ds(x)| ≤ C log ε · 2πε → 0,
Sε ∂n

∂ −1
u(x) log(|x − y|) ds(x) = −ε u(x) ds(x) → −2πu(y),
Sε ∂n(x) Sε
where the minus sign results from the normal −

→n pointing to the interior of Sε and
where we can apply the integral mean value theorem. Further let R → ∞. Then
"
∂ ∂u(x)
u(x) log(|x − y|) − log(|x − y|) ds(x) → 0,
SR ∂n(x) ∂n

x−y
since by (2.4), u(x) |x−y|2 ≤ CR ·
1 1 u(x) x − y ds(x) ≤ 2πC → 0 and
R, |x − y|2 R

∂u(x) 1 2πC

∂n log(|x − y|) ≤ C R 2 log R, | . . . |ds(x) ≤ R log R → 0.
Thus we arrive at (2.5). To obtain (2.6) note that for y ∈ Γ , Γ smooth we have

∂ u(y)|int Γ + u(y)|extΓ
u(x) log(|x − y|) ds(x) → −2π .
Sε ∂n(x) 2
This can be seen as follows. Approximate Γ by the tangent in y that separates Sε in

two half-balls. Introduce polar coordinates. Then for = ε, 0 < ϕ < π we have
u( , ϕ) ≈ u(y)|int Γ , whereas for π < ϕ < 2π we have u( , ϕ) ≈ u(y)|ext Γ .

Remark 2.2 1. Formula (2.5) holds also for a Lipschitz domain with corners. If
y ∈ Γ is a vertex with inner angle δ (= angle of the two tangents in x along Γ )
and outer angle 2π − δ, then the left hand side of (2.6) is to be changed to
δ 2π − δ
u(y)|int Γ + u(y)|ext Γ
2π 2π
2. Consider the special case u = 0 on Ω . Then there holds for y ∈ Ω

"
1 ∂u(x) ∂
u(y) = − log(|x − y|) − u(x) log(|x − y|) ds(x).
2π Γ ∂n ∂n(x)
∂n |Γ determine u on Ω; however u|Γ and ∂n |Γ

This means that u|Γ and ∂u ∂u
cannot be prescribed independently (remind the Dirichlet–Problem and the

Neumann problem).
From Theorem 2.2 we derive the following formulas in a ”weak” sense.
Theorem 2.3 Let u satisfy the assumptions
# $ of∂uTheorem∂u 2.2. In addition, suppose
u|int Γ − u|ext Γ = 0. Then using q = ∂u
∂n = ∂n | int Γ − ∂n |ext Γ there holds for all
y∈Γ

∂u(y) 1 1 ∂
|ext Γ = − q(y) − q(x) log(|x − y|) ds(x), (2.7)
∂n 2 2π Γ ∂n(y)

∂u(y) 1 1 ∂
|int Γ = q(y) − q(x) log(|x − y|) ds(x), (2.8)
∂n 2 2π Γ ∂n(y)
Remark 2.3 Obviously the formulas (2.7) and (2.8) are equivalent to
"
1 ∂u(y) ∂u(y) 1 ∂
|ext Γ + |int Γ =− q(x) log(|x − y|) ds(x)
2 ∂n ∂n 2π Γ ∂n(y)
(2.9)
Proof (of Theorem 2.3) Let ϕ ∈ C0∞ (Rn ), i.e. infinitely differentiable with compact
support supp ϕ. Then by the Gauss divergence theorem on Ω ∩ supp ϕ (note the
orientation of −
→
n ) using (2.5) in Theorem 2.2

∂u(y)
− |ext Γ ϕ(y) ds(y) = ∇u(z), ∇ϕ(z) d(z1 , z2 )
Γ ∂n Ω

1 x − z, ∇ϕ(z)
= q(x) ds(x) dz,
2π Ω Γ |x − z|2
where the latter integral exists on Ω × Γ , since q is continuous as a difference of

continuous functions, (Ω ∩ supp ϕ) × Γ compact, and

1
ds(x) dz < +∞ ,
Ω ∩supp ϕ Γ |x − z|
as seen as follows: Introduce polar coordinatesz − x = ρ(cos θ, sin θ )T ,

1 ,z2 )
(z1 , z2 )T = (x1 + ρ cos θ, x2 + ρ sin θ )T , ∂(z
∂( ,θ) = and obtain

1
ds(x) dθ < +∞ .
Ω ∩supp ϕ Γ
Interchanging the integrations according to the theorem of Fubini leads to

∂u(y) 1 x − z, ∇ϕ(z)
− |ext Γ ϕ(y) ds(y) = q(x) dz ds(x) .
Γ ∂n 2π Γ Ω |x − z|2
With Bε := B(x; ε) and Sε := ∂Bε we have as an improper integral

x − z, ∇ϕ(z) x − z, ∇ϕ(z)
dz = lim dz .
Ω |x − z|2 ε→0 Ω \(Ω ∩Bε ) |x − z|2
Once more we apply the divergence theorem to − log(|x − z|) (z = x, x fixed on

Ω \ Bε ∩ supp ϕ). By ∇z (− log |x − z|) = |x−z|
x−z −
→
2 , by the orientation of n (y), and
by −
→
n (y) = x−y
|x−y| on Sε we have

x − z, ∇ϕ(z) x − y, −
→
n (y)
dz = − ϕ(y) ds(y)
Ω \Bε |x − z|2 Γ \Γ ∩Bε |x − y|2

1
+ ϕ(y) ds(y) .
Sε ∩Ω |x − y|
Taking ε → 0 we arrive at

x − z, ∇ϕ(z) x − y, −
→
n (y)
dz = − ϕ(y) ds(y) + πϕ(x)
Ω |x − z|2 Γ |x − y|2
and hence ( y instead of x)

∂u(y) 1 x − y, −
→n (y)
− |ext Γ ϕ(y) ds(y) = − q(x)ϕ(y) ds(x)ds(y)
Γ ∂n 2π Γ Γ |x − y|2

1
+ q(y)ϕ(y) ds(y)
2 Γ
Thus (2.7) holds true in a weak sense. The proof of (2.8) follows similar lines.

2.2 Single- and Double-Layer Potential
Definition 2.1 Let Γ ∈ Lip ,i.e.Γ is locally the graph of a Lipschitz function, and
ϕ ∈ C(Γ ) . Then we define for x ∈ Γ
i) the single-layer potential Sϕ with density ϕ by

1
(Sϕ)(x) := Sϕ(x) := − ϕ(y) ln |x − y| dsy , x ∈ Γ (2.10)
π
Γ
ii) the double-layer potential Dϕ with density ϕ by

1
(Dϕ)(x) := Dϕ(x) := − ϕ(y)∂ny ln |x − y| dsy , x ∈ Γ (2.11)
π
Γ
Corollary 2.2
i) Let Ω be bounded, Γ ∈ Lip, u ∈ C 2 (Ω) and Δu = 0 in Ω. Defining γ0 u :=
u|Γ , γ1 u := ∂n u|Γ and Ω − , Ω + (corresponding to Fig. 2.2) denoting the
2.2 Single- and Double-Layer Potential 17
Fig. 2.2 Exterior domain

Ω + and interior domain Ω −
Γ Ω− Ω+
interior domain, exterior domain, respectively, there holds the representation

formula
1 1
u = − D(γ0 u) + S(γ1 u), in Ω = Ω − . (2.12)
2 2
ii) Let Γ ∈ Lip and u ∈ C02 (R2 \Γ ) , i.e. u|Ω ∈ C 2 (Ω) and u|R2 \Ω ∈
C02 (R2 \Ω) . Setting u+/− := u|Ω +/− , we define:
a) [γ0 u] := γ0 u+ − γ0 u− = u+ |Γ − u− |Γ
b) [γ1 u] := γ1 u+ − γ1 u− = ∂n u+ |Γ − ∂n u− |Γ
This yields with f := Δu the representation
1 1
u = Gf + D[γ0 u] − S[γ1 u] in R2 \Γ . (2.13)
2 2
This is just the statement of Theorem 2.2, since in the definition of the jump [·, ·]
the role of ± are interchanged.
Remark 2.4 The notation potential in Definition 2.1 is justified, since for ϕ ∈ C(Γ )
the identity Δx ln |x − y| = 0 for x = y yields in R2 \Γ with Δ and
interchanged
ΔSϕ = 0 = ΔDϕ . (2.14)
2.2.1 Some Remarks on Distributions
We now want to introduce some basic definitions and results on distributions, in

order to derive the classical theorems on boundary integral equations with modern
methods. For this, we first need the notion of a test function, i.e. a function ϕ ∈
C0∞ (R2 ) , as for example
%
exp 1
: |x| < R,
ϕ(x) = |x|2 −R 2
0 : |x| ≥ R.
Some further example can be given as follows. Let A ⊂ O be a closed, bounded

(thus compact) subset of an open set O. Then we may define a function ϕ ∈ C ∞
by:

1 : x ∈ A,
ϕ(x) =
0 : x ∈ O.
Definition 2.2 (Distributions) We define D(Ω) := C0∞ (Ω) , endowed with the
family of seminorms
sup |D α ϕ(x)| < ∞ , ∀ α ∈ N20 ∀K ⊂⊂ Ω.

x∈K
Then we denote with D (Ω) the space of distributions on Ω , i.e. the set of linear
continuous functionals f on D(Ω).
Any function f ∈ C(R2 ) defines a regular distribution

ϕ −→ f, ϕ := f (x)ϕ(x)dx ∀ ϕ ∈ C0∞ (R2 ) .
R2
Example 2.1 Let 0 ∈ Ω ⊂ R2 . The Dirac Delta-function δ0 ∈ D (Ω) is defined by
δ0 , ϕ = ϕ(0), ϕ ∈ D(Ω) .
Definition 2.3 The support of a distribution ϕ ∈ D (Ω) is defined as the set of all
points x in Ω for which for any η > 0 the restriction of ϕ to the domain Ω ∩{y||y −
x| < η} differs from the zero distribution.
Thus, for the Dirac Delta-function we have that supp (δ0 ) = {0} .
Example 2.2 A further example is given as follows:
For ψ ∈ C(Γ ) we define γ0∗ ψ by

γ0∗ ψ, ϕ := ψ(x)ϕ(x)dsx , ∀ ϕ ∈ C0∞ (R2 )
Γ
and γ1∗ ψ by

γ1∗ ψ, ϕ := ψ(x)∂n ϕ(x)dsx , ∀ ϕ ∈ C0∞ (R2 ) .
Γ
Definition 2.4 (Derivatives of Distributions) Let t ∈ D (R2 ) . Then we define

the partial derivative ∂j t by
∂j t, ϕ := − t, ∂j ϕ , ∀ ϕ ∈ C0∞ (R2 ).

We may also define the operator G as given in Corollary 2.1 for distributions t ∈
D (R2 ) with compact support as follows:
First let f ∈ C00 (R2 ) and ϕ ∈ C0∞ (R2 ) . Then we may define for

Gf, ϕ = (Gf )(x)ϕ(x)dx = G(x, y)f (y) dy ϕ(x) dx
R2 R2 R2

= f (y) G∗ (y, x)ϕ(x) dxdy = f, G∗ ϕ = f, χG∗ ϕ ,
R2 R2
where χ ∈ C0∞ is any cut-off function with χ ≡ 1 on supp (f ). Thus, for ϕ ∈ C0∞
we have χG∗ ϕ ∈ C0∞ since ∂j G∗ ϕ = −G∗ ∂j ϕ , implying that G∗ ϕ ∈ C ∞ and
thus χG∗ ϕ ∈ C0∞ . This leads to the following
Definition 2.5 For t ∈ D (R2 ) with supp (t) ⊆ {χ ≡ 1} and ϕ ∈ C0∞ we define
Gt, ϕ := t, χG∗ ϕ
Application of Definition 2.5

1. Let χ ∈ C0∞ with χ ≡ 1 in some neighbourhood of the origin. Then we may
show that Gδ is regular:
Gδ, ϕ = δ, χG∗ ϕ = (χG∗ ϕ)(0) = χ(0) · (G∗ ϕ)(0)

= G(y, 0)ϕ(y) dy = G(·, 0), ϕ
R2
1
⇒ (Gδ)(y) = G(y, 0) = ln |y|
2π
2. We want to consider G(γ0∗ ψ) . Therefore, we now assume that χ ≡ 1 in some

neighbourhood of Γ . With ϕ ∈ C0∞ we have:

Gγ0∗ ψ, ϕ = γ0∗ ψ, χG∗ ϕ = ψ(x)(χG∗ ϕ)(x) dsx
Γ

= ψ(x)(G∗ ϕ)(x) dsx = ψ(x) G(y, x)ϕ(y)dy dsx
Γ Γ R2
Hence by Fubini with single-layer potential Sψ, see (2.10)

1
Gγ0∗ ψ, ϕ = ϕ(y) G(y, x)ψ(x) dsx dy = ϕ(y)(−Sψ)(y) dy
2
R2 Γ R2
Lemma 2.1 Let ψ ∈ D (R2 ) be a distribution with compact support , then
ΔGψ = GΔψ = ψ .
Proof From Definition 2.4 and the definition of G we have, since G∗ = G for all
ϕ ∈ C0∞ , using (2.3)
ΔGψ, ϕ = Gψ, Δϕ = ψ, G∗ Δϕ = ψ, GΔϕ = ψ, ϕ ,

GΔψ, ϕ = Δψ, G∗ ϕ = ψ, ΔGϕ = ψ, ϕ .

As an example let us now assume that u ∈ C02 (Ω − )∪C02 (Ω + ) , i.e. u ∈ C 2 (R2 \Γ )

with supp (u) compact. Then we want to find a representation for the Laplacian Δu
of u. Using the same notations as in Corollary 2.2, the second Green formula implies
for all ϕ ∈ C0∞ ,
Δu, ϕ = u, Δϕ = uΔϕ dx = u− Δϕ dx + u+ Δϕ dx

R2 Ω− Ω+
= Δu− ϕ dx + Δu+ ϕ dx + (u− ∂n ϕ − ∂n u− ϕ − u+ ∂n ϕ + ∂n u+ ϕ) ds .
Ω− Ω+ Γ
Setting f := Δu− + Δu+ , this yields
Δu, ϕ = f, ϕ + [∂n u]ϕ ds − [u]∂n ϕ ds

Γ Γ
= f, ϕ + γ0∗ [∂n u], ϕ − γ1∗ [u], ϕ ,
using γ1∗ ψ, ϕ = ψ∂n ϕ ds and γ0∗ ψ, ϕ = ψϕ ds . Thus, we finally obtain

Γ Γ
the result
Δu = f + γ0∗ [γ1 u] − γ1∗ [γ0 u] . (2.15)
Therefore, Lemma 2.1 implies that
u = GΔu = Gf + Gγ0∗ [∂n u] − Gγ1∗ [u] .
Comparing this result with the representation of u in (2.13), we find a new relation
for the double- and single-layer potential,
Sψ = −2Gγ0∗ ψ (2.16)
Dψ = −2Gγ1∗ ψ . (2.17)
2.2.2 Jump Relations
In this subsection we want to derive the jump relations for the single- and double-
layer potentials.
In the following we will make extensive use of the following boundary integral
operators
Definition 2.6 Let x ∈ Γ , then we denote by

1
V ϕ(x) := − ϕ(y) ln |x − y| dsy (2.18)
π
Γ
the single layer potential operator and the double layer potential operator is
given by

1
Kϕ(x) := − ϕ(y)∂ny ln |x − y| dsy . (2.19)
π
Γ
Moreover, we define the adjoint double layer potential operator as

1
K ϕ(x) := − ϕ(y)∂nx ln |x − y| dsy (2.20)
π
Γ
and introduce the so called hypersingular operator by
W ϕ(x) := − ∂nx Kϕ(x) (2.21)
From (2.16) and (2.17) it follows using Lemma 2.1 that
Δ(Sψ) = −2γ0∗ ψ Δ(Dψ) = −2γ1∗ ψ . (2.22)
If now we apply (2.15) for u = 12 Sψ and u = 12 Dψ , respectively, a comparison

with (2.22) will prove the following jump relations.
Lemma 2.2 (Jump Relations) Suppose Γ = ∂Ω ∈ C 3 and ψ ∈ C 2 (Γ ) , such
that we have Sψ, Dψ ∈ C 2 (Ω + ∪ Ω − ) . Then there holds
i) [γ0 (Sψ)] = 0
[γ1 (Sψ)] = −2ψ
ii) [γ0 (Dψ)] = 2ψ
[γ1 (Dψ)] = 0
Proof From (2.22), (2.15) and (2.14) we deduce
1 1 1
−γ0∗ ψ = Δ( Sψ) = 0 + γ0∗ [γ1 ( Sψ)] − γ1∗ [γ0 ( Sψ)] ,
2 2 2
hence by uniqueness of representation
−2ψ = [γ1 (Sψ)], 0 = [γ0 (Sψ)] .
Similarly from
1 1 1
−γ1∗ ψ = Δ( Dψ) = 0 + γ0∗ [γ1 ( Dψ)] − γ1∗ [γ0 ( Dψ)]
2 2 2
it follows
2ψ = [γ0 (Dψ)], 0 = [γ1 (Dψ)] .

Remark 2.5 Using Lemma 2.1 one observes that the proof of Lemma 2.2 remains
valid for Γ ∈ Lip and ψ ∈ D (R2 ), since Green’s formulas hold for Lipschitz
domains, i.e. domains Ω with Γ = ∂Ω ∈ Lip. Note that for Ω ⊂ Rd , Γ = ∂Ω ∈
Lip if every point on Γ has a neighborhood N ⊂ Rd such that, after an affine
change of coordinates (translation and rotation), Γ ∩ N is described by the equation
xd = ϕ(x1 , . . . , xd−1 ), where ϕ is uniformly Lipschitz continuous. Moreover,
Ω ∩ N is on one side of ∂Ω ∩ N, e.g. Ω ∩ N = {x ∈ N : xd < ϕ(x1 , . . . , xd−1 )}
(see [327]).
Lemma 2.3 Under the assumptions of the above lemma there holds,
i) γ1 (Sψ)+ = K ψ −ψ
γ1 (Sψ)− = K ψ +ψ
ii) γ0 (Dψ)+ = Kψ +ψ
γ0 (Dψ)− = Kψ −ψ
Proof We consider ψ ∈ C02 (R2 ) and u satisfying the equation Δu = 0 in Ω. Then
for ϕ ∈ C 1 (Ω) ∩ C(Ω),

∂u
∇u∇ϕ dx = ϕ ds
Ω Γ ∂n
and setting u = ( 12 Sψ)− we have

1
γ1 u γ0 ϕ = ∇( Sψ)∇ϕ
Γ Ω 2

1
= ∇x lim − ln |x − y|ψ(y) dsy · ∇x ϕ(x)dx
Ω →0 2π y∈Γ
|y−x|≥

1
=− ψ(y) lim ∇x ln |x − y| · ∇x ϕ(x)dx dsy
2π Γ →0 x∈Ω
|y−x|≥
where the last identity is obtained with the Fubini’s theorem. Now using Green’s
first formula, we obtain

∇x ln |x − y| · ∇x ϕ(x)dx
x∈Ω
|y−x|≥

= γ1,x ln |x − y|γ0 ϕ(x)dsx − (x ln |x − y|)ϕ(x)dx
∂(Ω\B ) Ω\B

= γ1,x ln |x − y|ϕ(x) dsx + γ1,x ln |x − y| γ0ϕ(x) dsx
x∈Ω x∈Γ
|x−y|= |x−y|≥

= γ1,x ln |x − y|(ϕ(x) − ϕ(y)) dsx + ϕ(y) γ1,x ln |x − y| dsx
x∈Ω x∈Ω
|x−y|= |x−y|=

+ γ1,x ln |x − y| γ0 ϕ(x) dsx
x∈Γ
|x−y|≥
We use polar coordinates

∂ 1
γ1,x ln |x − y| =− ln r =−
r= ∂r r=
such that

1
γ1,x ln |x − y| dsx = − dsx = −π for → 0.
x∈Ω x∈Ω
|x−y|= |x−y|=
On the other hand

γ1,x ln|x−y|(ϕ(x)−ϕ(y))dsx ≤ max |ϕ(x)−ϕ(y)| γ1,x ln|x−y| dsx
x∈Ω
x∈Ω |x−y|= x∈Ω
|x−y|= |x−y|=
Thus, when → 0, the above term tends to zero. Therefore

1
γ1 u γ0 ϕds = − ψ(y)(−πϕ(y))dsy
Γ 2π Γ

1
− ψ(y) lim γ1,x ln |x − y|γ0 ϕ(x)dsx dsy
2π Γ →0 x∈Γ
|x−y|≥

1 1
= ψ(y)ϕ(y)dsy + ψ(y) − γ1,x ln |x − y|γ0 ϕ(x) dsx dsy
2 Γ Γ 2π Γ

1 1
= ψ(y)ϕ(y)dsy + ϕ(x) − γ1,x ln |x − y|ψ(y)dsy dsx
2 Γ Γ 2π Γ
or shortly,
1 1 1
γ1 ( Sψ)− = ψ + K ψ.
2 2 2
In the same way we can also prove
γ0 (Dψ)± = Kψ ± ψ
γ1 (Sψ)+ = K ψ − ψ

Let us look again at the homogeneous Laplace problem. The representation formula
(2.2) yields for all x ∈ Ω and for u with u = 0 :

1 1
u(x) = u(y)∂ny ln |x − y| dsy − ∂n u(y) ln |x − y| dsy
2π 2π
Γ Γ
1 1 ∂u(x)
= − Du(x) + S .
2 2 ∂n
Making use of the jump relations, letting x → Γ we have for x ∈ ∂Ω = Γ :

1 u(x) 1
u(x) = u(y)∂ny ln |x − y| dsy + − ∂n u(y) ln |x − y| dsy
2π 2 2π
Γ Γ
1 1 ∂u(x)
= − Ku(x) + u(x)/2 + V
2 2 ∂n
leading with u = g for the Dirichlet problem to

1 1
f (x) := g(x) − g(y)∂ny ln |x − y| dsy = − ∂n u(y) ln |x − y| dsy .
π π
Γ Γ
Thus, for the inhomogeneous Dirichlet problem we finally obtain Symm’s integral
equation
Vq = f
for the unknown q = ∂u∂n .

The above topic was extended to a Lipschitzian boundary curve in the works by
Verchota, Mitrea and others (see [307, 421]) on one hand and by Costabel on the
other hand (see [114]).
2.3 Mapping Properties of BI Operators 25
2.3 Mapping Properties of Boundary Integral Operators
In this section we follow [116].

Lemma 2.4 Let Γ = ∂Ω , Ω ⊂ R2 , and ϕ ∈ C(Γ ) with ϕ = 0 such that

ϕ(x) dsx = 0 . (2.23)
Γ
Then we have

ϕ, V ϕ := ϕ(x)(V ϕ)(x) dsx > 0 .
Γ
Proof A simple calculation yields for large |y|:
1 (x, y) |x|2
ln |x − y| = ln |y| + ln(1 − 2 + 2)
2 |y|2 |y|
(x, y) 1 |x|2 1
= ln |y| − + + O( 2 ) .
|y| 2 2 |y| 2 |y|
Hence for x ∈
/ Γ,

1 1
−πSϕ(y) = ln |y| ϕ(x)dsx − {y1 x1 ϕ(x)dsx +y2 x2 ϕ(x)dsx }+O( )
|y|2 |y|2
Γ Γ Γ
implies by (2.23)
Sϕ(y) = O |y|−1 .
Also we have
∇Sϕ(y) = c ln |y| + O |y|−1 for |y| −→ ∞ ,
since

x−y
= 1 |y| ≤ 1 1 + |x| = O |y|−1 .
|x − y|2 |y| |x − y| |y| |y − x|
Using the jump relations given in Lemma 2.2 and Lemma 2.3 and setting u := Sϕ ,
we obtain using Green’s first formula
2ϕ, V ϕ = γ1 u− , γ0 u− − γ1 u+ , γ0 u+

= |∇u|2 + |∇u|2 = |∇Sϕ|2 .
Ω− Ω+ R2
Suppose |∇Sϕ|2 = 0 , hence Sϕ = const. and thus 2ϕ = −[γ1 Sϕ] = 0, a

R2
contradiction. Hence it follows ϕ, V ϕ > 0 for ϕ = 0 .

Definition 2.7 For m(φ) := 1
L ϕ ds with L := ds and ϕ0 := ϕ − m(ϕ) we
Γ Γ
define for ϕ ∈ C(Γ ) a norm by
ϕ2V := ϕ0 , V ϕ0 + m2 (ϕ) .
Definition 2.8 We define the space HV to be the completion or closure of L2 (Γ )

(or of C(Γ ) or of C ∞ (Γ ) ) with respect to the norm .V , i.e. HV is a Hilbert space
with inner product (ϕ, ψ)V := ϕ0 , V ψ0 + m(ϕ)m(ψ).
Remark 2.6 The spaces C(Γ ), C ∞ (Γ ) and L2 (Γ ) are dense in HV . - The
mapping V : L2 (Γ ) → L2 (Γ ) can be extended continuously to V : HV → HV ,
HV denoting the dual space which lies dense in L2 (Γ ) . Thus, L2 (Γ ) is self-dual,
HV ⊂ L2 (Γ ) ⊂ HV .
Lemma 2.5 For Γ = ∂Ω ∈ Lip the following statements are equivalent:
(i) V : HV → HV is bijective.
(ii) The equation V ψ = 1 has a solution in HV .
(iii) The equation V ψ = 0 has only the trivial solution.
Proof (i) ⇒ (ii): clear.
(ii) ⇒ (i): Let V e ≡ 1 . Then the operator V : HV → HV is
• injective, since: Let V ψ ≡ 0 , then for all ϕ ∈ HV
ϕ, V ψ = V ϕ, ψ = 0.
Thus

0 = V e, ψ = 1, ψ = ψ ds ,
Γ
hence by Lemma 2.4 ψ = 0, since V ψ, ψ = 0

• surjective, since:
For t ∈ HV the Riesz representation theorem yields: ∃ψ ∈ HV such that
∀ϕ ∈ HV
t, ϕ = (ψ, ϕ)V = ψ0 , V ϕ0 + m(ϕ)m(ψ)

= V ψ0 , ϕ + m(ψ)m(ϕ) − V ψ0 , m(ϕ)

m(ψ)
= V ψ0 + − m(V ψ0 ) ϕ(x) dsx .
L
Γ

m(ψ)
Hence, t = V ψ0 + − m(V ψ0 ) e , withV e ≡ 1 .
L
(i) ⇒ (iii): clear.
(iii) ⇒ (i): The boundary integral operator V is a Fredholm operator (see
Definition A.8 in Appendix A) with zero index, see Sections 4.2 and 4.3, see also
[259, Section 10.3].

As we shall see below, the boundary integral operator V is bijective and even
positive definite, if Ω is contained in a disk with radius < 1, what can always
be arranged by scaling.
More precisely, due to Gaier [185, Satz 11], Sloan and Spence [380, Section
2] based on analytic function theory [245, Chapter 16] the equation V ψ = 1
admits a (unique) solution ψ = e, provided cap(Γ ) = 1, where cap(Γ ) denotes
the logarithmic capacity or transfinite diameter of a Lipschitz curve Γ. Note, that
if e exists, we always have that m(e) = 0 , since: m(e) = 0 ⇒ e, V e > 0 by
Lemma 2.4, but e, V e = e, 1 = Lm(e) = 0, which is a contradiction. Further,
cap(Γ ) scales linearly, i.e. cap(Γr ) = r cap(Γ ), where for some scalar r > 0,
& x '
Γr := r · Γ = x ∈ R2 : ∈Γ .
r
In what follows, the single-layer operator V on Γ will be denoted by VΓ .
Theorem 2.4 Let Γ ∈ Lip contained in R2 . Then
(i) If cap(Γ ) = 1, then cap(Γ ) = r10 with r0 = exp( Lm(e)
π
)
(ii) VΓr is positive definite if and only if cap(Γ ) < 1.
Proof
(i) We first transform VΓr acting on Γr onto the boundary Γ as follows:
For u defined on Γ let ur : Γr −→ R be given by ur (x) := u xr .
Further, for x ∈ Γ let (VΓr u)(x) := (VΓr ur )(rx) . Then, VΓr is bijective if
and only if VΓr is bijective. One calculates:
(VΓr )u(x) = − π1 ln |rx − y|ur (y) ds(y)

Γr
y=rz
= − π1 ln |rx − rz|ur (rz)r ds(z)
Γ
= − π1 ln(r|x − z|)u(z)r ds(z)
(Γ )
= r (VΓ u)(x) − ln r
π u(z) ds(z)
Γ
= r (VΓ u(x) − ln r · c(u))
with c(u) = 1
π u(z) ds(z) . Let e solve VΓ ψ ≡ 1 . Then VΓr e = r(1 −
Γ
c(e) ln r). Hence in particular VΓr0 e = 0 for r0 = exp( c(e)
1
). Thus, we have that
for VΓ bijective, the operator VΓr0 becomes not bijective and thus
cap Γr0 = r0 cap Γ = 1 .
With r0 as given above, c(e) = 1

π e ds , and m(e) = π
L ln r0 the first part of
Γ
the theorem is proved.
m(ϕ)
(ii): To show that V is positive definite, let ϕ ∈ HV and ϕ0 := ϕ − m(e) e . By
Lemma 2.4, without loss of generality, m(ϕ) = 0. Hence
m(ϕ)
m(ϕ0 ) = m(ϕ) − m(e) = 0 ,
m(e)
2
m(ϕ) m(ϕ) m(ϕ)
ϕ, V ϕ = ϕ0 , V ϕ0 + ϕ0 , V e + e, V ϕ0 + e, V e .
m(e) m(e) m(e)
Since e, V ϕ0 = ϕ0 , V e = ϕ0 , 1 = 0 , the second summand and third

summand vanish. Further with μ := m(ϕ) m(e) = 0 and again by Lemma 2.4,
ϕ, V ϕ > Lμ2 m(e). Now, Lm(e) = lnπr0 > 0 ⇔ r0 > 1 ⇔ cap(Γ ) < 1.
Conversely let V be positive definite. Then V is injective. By Lemma 2.5,
V is bijective and there exists e with V e = 1 . Since V is positive definite,
e, V eϕ = Lm(e) > 0 . Hence by definition r0 > 1 ⇔ cap(Γ ) < 1. The
assertion of (ii) follows.

Now from
cap (Γ ) ≤ cap ∂B(0; a) = a ,
whenever Γ is contained in a disk B(0; a) with radius a, diam Γ < 1 is a sufficient

condition for V to be bijective and positive definite.
Theorem 2.5 Let Γ be a Lipschitz curve. Then for the operators V , K, K and W
1 1
(i) V :H − 2 (Γ )−→H 2 (Γ ) is continuous, positive def inite ⇐⇒cap(Γ )<1
(ii)
% 1 1
H 2 (Γ ) −→ H 2 (Γ )
1+K :
L2 (Γ ) −→ L2 (Γ )
is continuous with ker (1 + K) = R = constants

(iii)
% 1 1
H − 2 (Γ ) −→ H − 2 (Γ )
1+K : is continuous and inj ective with
L2 (Γ ) −→ L2 (Γ )
(1 + K )u = g solvable ⇔ m(g) = 1
L g=0
Γ
(iv)
% 1 1
H 2 (Γ ) −→ H 2 (Γ )
1−K : is continuous and bij ective
L2 (Γ ) −→ L2 (Γ )
(v)
% 1 1
H − 2 (Γ ) −→ H − 2 (Γ )
1−K : is continuous and bij ective
L2 (Γ ) −→ L2 (Γ )
(vi)
1 1
W : H 2 (Γ ) −→ H − 2 (Γ ) is continuous with either
ker W = {constants} or coker W = {constants}
1 1 1
⇒ W : H0 (Γ ) := {ϕ ∈ H 2 (Γ ) : m(ϕ) = 0} −→ H − 2 (Γ )/R
2
is continuous and bijective with

1
W ϕ, ϕ ≥ 0 ∀ ϕ ∈ H 2 (Γ ), W ϕ, ϕ > 0 if m(ϕ) = 0 and ϕ = 0
Proof We only prove the last assertion (vi) and leave the other items as an exercise.
Let u(x) := 12 Dϕ(x) for x ∈ Γ . Then by the jump relations given in Lemma 2.2
(ii) we have
[γ0 u] = ϕ , [γ1 u] = 0 andγ1+ u = γ1− u = −W ϕ

Thus,

ϕ, W ϕ = −[γ0 u], γ1− u −
= γ0 u , γ1 u − +
−γ0 u , γ1 u +
= |∇u| +
2
|∇u|2
Ω− Ω+
using Green’s formulas, since u = O |x| as x → ∞. It

1
follows ϕ, W ϕ ≥ 0 .
Let now u ∈ 1 (R2 ) ,
Hloc i.e. u ∈ H 1 (K) ∀ K ⊃⊃ Ω − , Kcompact. With
u=O 1
|x| as x → ∞ it follows that

|∇u|2 + |∇u|2 < ∞ .
Ω− Ω+
1
By the trace theorem we further have u+/− |Γ =: γ0+/− u ∈ H 2 (Γ ) and thus
[γ0 u], γ1 u+/− < ∞ .
Furthermore, W is continuous, since
ϕ, W ϕ
||W ϕ|| −1 := sup <∞.
H 2 (Γ ) 1 ||ϕ|| 1
ϕ∈H 2 (Γ ) H 2 (Γ )
And W is bijective, since
ϕ, W ϕ = 0 ⇐⇒ u+/− = c(constant)

⇒ u+ = 0 (since u = O 1
|x| as x → ∞ )
⇒ u− =c ⇒ γ1 u− = −W ϕ = 0
With ϕ = [γ0 u] = −c and m(ϕ) = 0 it follows ϕ ≡ 0 .

2.4 Laplace’s Equation in R3
In this chapter we want to consider the situation that Ω ⊆ R3 . We may assume that
for a neighbourhood U of any point x0 ∈ Γ = ∂Ω there exists a local parameter
representation (Fig. 2.3).
Thus, R3 ⊇ O # (u, v, w) → x(u, v, w) ∈ U with x0 ∈ U , such that
∂x
det = 0 ∀ (u, v, w) ∈ O .
∂(u, v, w)
2.4 Laplace’s Equation in R3 31
R3
Γ = ∂Ω
x0
Ω
U
O
We can arrange the local parameter representation such that
w = 0 ⇔ x(u, v, w)(= x(u, v)) ∈ Γ

w > 0 ⇔ x(u, v, w) ∈ Ω− = Ω
w < 0 ⇔ x(u, v, w) ∈ Ω + = R3 \Ω
Hence, we have that Γ ∈ C m if and only if ∀ x0 ∈ Γ there exists a local parameter

representation in C m (O).
Let the tangential and normal vectors to the curve Γ be denoted by ( ∂x ∂x T
∂u , ∂v )
and −→
n = +/− ∂u ∂v , such that − →
∂x ∂x
×
n points from Ω − to Ω + . Then, the normal
∂x
∂u × ∂v
∂x
derivative is given by ∂n u := −
→
n · grad(u) . For the surface measure we have
1
ds := (det g) 2 du dv for gij = ∂i x × ∂j x
(with ∂1 x = ∂x
∂u , ∂2 x = ∂v ), which is defined on Ũ ⊂ Γ independent of the
∂x
parameter representation. Thus we have

J
1
ϕ(x) ds(x) := ϕj (u, v)(det (g(u, v))) 2 du dv (2.24)
j =1
Γ Oj ∩IR 2
*
J
with Γ ⊆ Oj , supp (ϕj ) ⊆ Oj ∩ Γ, ϕj = ϕ .
j =1
For f ∈ C01 (R3 ) integration by parts yields

∂j f dx = nj (x)f (x) ds(x) . (2.25)
Ω Γ
This leads to the following conclusions. We obtain by partial integration for f, g ∈

C01

(∂j f · g + f · ∂j g) dx = nj (x)f (x)g(x) ds(x) , (2.26)
Ω Γ
further by the first Green’s formula for f ∈ C02 , g ∈ C01

(∇f · ∇g + Δf · g) dx = ∂n f · g ds(x) , (2.27)
Ω Γ
and by the second Green’s formula for f, g ∈ C02

(Δf · g − f · Δg) dx = (∂n f (x) · g(x) − f · ∂n g(x)) ds(x) . (2.28)
Ω Γ
For a formulation of Green’s formulas in a distributional form we shall consider

− +
(u− , u+ ) ∈ C02 (Ω ) × C02 (Ω ), where shortly Ω − = Ω − ∪ Γ , Ω + = Ω + ∪ Γ ,
and let
−
u on Ω −
u := , f := Δu|IR 3 \Γ .
u+ on Ω +
Then for any χ ∈ C0∞ (R3 ) , γ0 χ = χ|Γ and γ1 χ = ∂n χ|Γ .

With γ0∗ , γ1∗ defined by
γ0∗ ϕ, χ := ϕ · γ0 χ dx := ϕ · χ ds ,
R3 Γ
γ1∗ ϕ, χ := ϕ · γ1 χ dx := ϕ · ∂n χ ds
R3 Γ
respectively, we obtain as in Sect. 2.2.1 for the Laplacian Δu the representation

analogous to (2.15)
Δu = f + γ0∗ [γ1 u] − γ1∗ [γ0 u] (2.29)
2.4.1 Representation Formula
Analogous to Sect. 2.1 we now want to derive a representation formula for u ∈

C02 (Ω) . Let
1 1
G(x, y) = − ,
4π |x − y|
2.4 Laplace’s Equation in R3 33

R3
B (x)
x Ω
BR
which is a fundamental solution, since it is easy to show (Exercise) that

Δx G(x, y) = 0, ∀ x = y . Consider now the following Fig. 2.4:
Let u ∈ C02 (R3 ) , f = Δu and

G(x, y)f (y) dy := lim G(x, y)f (y) dy .
→0
R3 <|x−y|<R

:=J
Then by 2. Green formula (2.28) in Ω\B (x) ,

J = (∂n u · G − u · ∂ny G) dsy + 0
∂B
and

1 →0
∂n u(y) · G(x, y) dsy ≤ sup |∇u| · · 4π 2
−→ 0 .
4π
∂B (x)
We further have
u(y)∂ny G(x, y) dsy = u(x + y)[∂r (− 4πr

1
)]r= dsy
∂B (x) ∂B (0)
= 4π1 2 u(x + y) dsy
∂B (0)
→0
= 4π
1
u(x + y) dsy −→ u(x) .
∂B1 (0)
This gives

G(x, y)f (y) dy = u(x)
R3
and
ΔGu = GΔu = u, ∀ u ∈ C02 (R3 ) (2.30)
Thus, we have equation (2.30) also for distributions u ∈ D (R3 ) with compact
support, where D (R3 ) = {ϕ | ϕ : C0∞ (R3 ) −→ R, linear and continuous} .
With (2.29) and (2.30) we finally obtain the desired representation formula for u ∈
+ −
C02 (Ω ∪ Ω ) :
u = GΔu = Gf + Gγ0∗ [γ1 u] − Gγ1∗ [γ0 u] .
In correspondence with Sect. 2.2 we define the potential operators as follows.

Definition 2.9 For x ∈ Γ the operator

1 ϕ(y)
Sϕ(x) := −2Gγ0∗ ϕ(x) = ds(y) (2.31)
2π |x − y|
Γ
is called the single-layer potential with density ϕ ;

∗ 1 1
Dϕ(x) := −2Gγ1 ϕ(x) = ϕ(y)∂ny ds(y) (2.32)
2π |x − y|
Γ
is called the double-layer potential with density ϕ .

With these definitions the above representation formula reads
1 1
u = Gf + D[γ0 u] − S[γ1 u].
2 2
Remark 2.7 We may again prove the jump relations: (i) [γ0 Sψ] = 0 and [γ1 Sψ] =
−2ψ (ii) [γ0 Dψ] = 2ψ and [γ1 Dψ] = 0 .
Remark 2.8 The operators (1 ± K) in H 1/2(Γ ) and (1 ± K ) in H −1/2 (Γ ) are
contractions and they give rise to coercive bilinear forms. This allows to solve
the corresponding boundary integral equations by the use of Banach’s fixed point
theorem; and Neumann’s series always converges [118, 394].
2.5 Calderon Projector
As seen above the solution u ∈ C 2 (Ω) of Δu = 0 in Ω = Ω − has a representation

u = − 12 Dv + 12 Sφ with v = γ0 u, φ = γ1 u. Then the jump relations yield
2γ0 u = −γ0 Dv + γ0 Sφ = v − Kv + V φ
2γ1 u = −γ1 Dv + γ1 Sφ = W v + φ + K φ.
2.5 Calderon Projector 35
or more compactly written

γ0 u v 1 v v
≡ = (I d + A ) =: C (2.33)
γ1 u ϕ 2 φ φ
with

−K V
A :=
W K
Definition 2.10 The operator 12 (I d + A ) =: C (= C − ) is called Calderon

projector (for the domain Ω − and the Laplacian).
Theorem 2.6 Suppose Γ ∈ C ∞ .
1. Let v, φ ∈ C ∞ (Γ ). Then there ∞
exists u ∈ C (Ω) with Δu = 0 in Ω, v = γ0 u
v v
and φ = γ1 u if and only if =C .
φ φ
2. C 2 = C .
3. C − + C + = I d.
Proof
1. Let v, φ ∈ C ∞ (Γ ) be given and set u := − 12 Dv + 12 Sφ. Then u ∈ C ∞ (Ω),

γ0 u v
Δu = 0 and =C as shown above. If also there holds
γ u φ
1
v v v γ0 u
=C then it follows that = .
φ φ φ γ1 u
The inverse assertion is just (2.33).
∞ γ0 u
2. Let φ, v both in C (Γ ) and set u := − 2 Dv + 2 Sφ. Then there holds
1 1
=
γ1 u

v γ0 u γ0 u v
C and =C = C2
φ γ1 u γ1 u φ

v v
Hence C = C2 implying C 2 = C
φ φ
3. The jump relations yield C + = 1

2 (I d − A ).
Remark 2.9 C2 = C ⇐⇒ A2 = Id
Using the Calderon projector boundary integral equations of first and second kind
for mixed boundary and transmission problems can be derived.
Following e.g. [127, 405], consider the mixed boundary value problem
∂u
Δu = 0 in Ω, u = g1 on Γ1 , = g2 on Γ2 ,
∂−
→n
where
Γ = Γ¯1 ∪ Γ¯2 , Γ1 ∩ Γ2 = ∅
for given g1 , g2 .
Sought are v := u on Γ2 , ψ := ∂u
∂−
→
n
on Γ1 . Obtain following system of boundary
integral equations

W22 K12 v −W12 I − K22 g1
=
−K12 V11 ψ I + K11 V21 g2
where Wij etc. means integration on Γj and evaluation on Γi .
2.6 Use of Complex Function Theory
In this section we provide another elementary approach in R2 which is based on

complex function theory using in particular the Cauchy formula and the Cauchy-
Riemann differential equations. Here again we derive the representation formula
and give an applicable representation. of the hypersingular operator.
2.6.1 Representation Formula Again
Let the boundary Γ of a two dimensional domain Ω be a sufficiently smooth, simply

connected curve given by a 1-periodic vector function
%
x = x(t)
x = x(t) = for 0 ≤ t ≤ 1.
y = y(t),
Let us consider the interior Dirichlet problem for the Laplacian: Given u0 on Γ ,
find u satisfying
%
Δu = 0 in Ω
(2.34)
u = u0 on Γ.
In order to reduce this boundary value problem to an integral equation on Γ , firstly

we will deduce the representation formula for harmonic functions from the Cauchy
formula of complex function theory.
Due to Δu = 0 the function u is the real part of a holomorphic function
u = % F (z), F (z) = u + iv, z = x + iy

2.6 Use of Complex Function Theory 37
For u and v there hold the Cauchy-Riemann differential equations
ux = vy , −uy = vx .
Hence
p p
v(p) − v(p0 ) = dv = (−uy dx + ux dy).
p0 p0
Here the integral on the right hand side is path independent, since
uxx = −uyy , i.e. Δu = 0
We apply the Cauchy integral formula to F (z) and compute its real part in order to
express u
+
1 F (ζ )
u(z) = % F (z) = % dζ , ζ = ξ + iη, ζ ∈ Γ, z ∈ Ω (2.35)
2πi Γ z−ζ
Due to the Cauchy Riemann equations also the boundary values of u and v are
connected by

dv ∂u
v̇ = = v ξ̇ + v η̇ = u η̇ − u ξ̇ = n · u =
ds Γ ∂n Γ
x y x y
Here and in the following s denotes the arc length on Γ and ξ̇ the derivative of ξ
w.r.t. s; n is the normal on Γ pointing outward of Ω.
We change (2.35) to
+ +
1 u(ζ ) dζ
ds ds v(ζ ) dζ
ds ds
u(z) = % +i (2.36)
2πi Γ z−ζ Γ z−ζ
Now we have
+
1 v(ζ ) dζ
ds ds 1 d
% =% ln(z − ζ(s))ds
v(ζ ) (2.37)
2π Γ z−ζ 2π Γ ds
L
1 L dv(ζ )
=% (v(ζ ) ln(z − ζ ))0 − % ln(z − ζ(s))ds
2π s=0 ds
L
1 L 1 ∂u
= v(ζ ) ln |ζ − z|0 − (ζ )% ln(z − ζ )ds
2π 2π 0 ∂n
L
1 ∂u
= 0− (ζ ) ln |z − ζ |dsζ .
2π 0 ∂n
We rewrite the other integral in (2.36) as

+
1 u(ζ ) dζ
ds ds 1 d
% =% u(ζ ) ln(z − ζ(s))ds
2πi Γ z−ζ 2πi Γ ds

1 d
=% u(ζ ) {ln |z − ζ | + i arg (z − ζ )}ds.
2πi Γ ds
For the normal derivative

∂
ln |ζ − z| = (ln |z − ζ |)ξ η̇ − (ln |z − ζ |)η ξ̇
∂n
we use the Cauchy-Riemann equations for ln(z − ζ ):
∂ d
ln |ζ − z| = arg(z − ζ )η η̇ + arg(z − ζ )ξ ξ̇ = arg(z − ζ )
∂n ds
Thus, we obtain
+
1 u(ζ ) dζ
ds ds 1 d ∂
% )=% u(ζ ){ (ln |z − ζ |) + i (ln |z − ζ |)}ds
2πi Γ z−ζ 2πi Γ ds ∂nζ

1 ∂
= u(ζ ) (ln |ζ − z|)ds . (2.38)
2π Γ ∂nζ
Next, we combine (2.37) and (2.38) and obtain the representation formula

1 ∂ 1 ∂u
u(x, y) = u(s) (ln |z − ζ |) − ln |z − ζ |ds .
2π Γ ∂nζ 2π Γ ∂nζ
Hence the solution of (2.34) is a combination of the potential of a double layer (with
density u) and the potential of a single layer (with density ∂u
∂n ).
We apply the formulas of Plemelj-Sochozki (see e.g. [43, 203]) to this represen-
tation of u which yield for z → Γ

1 1 ∂ 1 ∂u(ζ )
u(z) = u(ζ ) ln |z − ζ (s)|ds − ln |z − ζ(s)|ds, z∈Γ
2 2π Γ ∂nζ 2π Γ ∂nζ
Since u|Γ = u0 must hold, this yields an equation with ∂u

∂n = q as unknown function
on the boundary Γ (z ∈ Γ ) :

∂
− ln |z − ζ (s)|q(s)ds = f (z) = πu0 (z) − u0 (ζ ) ln |z − ζ(s)|ds).
Γ Γ ∂nζ
2.6.2 Applicable Representation of the Hypersingular Integral

Operator
In this subsection we provide a representation of the hypersingular integral

operator W introduced in Definition 2.6 as a Cauchy principal value and further a
representation via the single layer potential on the boundary which is applicable to
numerical simulations with the boundary element method. To this end we work in
the complex plane
x = (x1 , x2 ) ↔ ζ = x1 + ix2 ; y = (y1 , y2 ) ↔ z = y1 + iy2 ,
use the complex logarithm
φ = ln(ζ − z) = ln(|x − y|) + iϑ(ζ − z)
involving the argument or angle ϑ(z) of z, and apply the Cauchy–Riemann

differential equations
(%φ)x1 = ((φ)x2 , ((φ)x1 = −(%φ)x2 ,
where %φ and (φ are respectively, the real part and imaginary part of φ.
Thus we can show the following
Proposition 2.1 Let f ∈ C 1 (Γ ). Then with t(x) denoting the tangent unit vector
in x, there holds in the weak sense

∂ ∂ df (y) x − y, t(x)
f (y) ln(|x − y|) dsy = dsy . (2.39)
∂nx ∂ny ds |x − y|2
Γ Γ
Here the first integral has to be understood in the weak (distributional) sense, the
second integral exists as a Cauchy principal value, that is with Γ parametrized by
x = Z(τ ), respectively by y = Z(σ ) with τ, σ ∈ (0, 2π),
2π
df (y) x − y, t(x) df (Z(σ )) Z(τ ) − Z(σ ), t(Z(τ ))
dsy = dσ
ds |x − y|2 dσ |Z(τ ) − Z(σ )|2
Γ 0
⎛ τ −ε ⎞
2π
= lim ⎝ ... d σ + . . . d σ⎠ .
ε→0
0 τ +ε
Moreover, the hypersingular operator W can be represented for any ϕ, ψ ∈

H 1/2(Γ ) via the single layer potential V on the boundary,

dϕ(x) dψ
W ψ, ϕ = V( )(x) dsx = ϕ̇, V ψ̇ . (2.40)
Γ dsx dsx
Proof Let g ∈ C0∞ (R2 ) arbitrarily. Then we have by partial integration on the curve
Γ without (topological) boundary, hence without boundary terms,

∂ ∂
I := g(x) f (y) ln(|x − y|) dsy dsx
∂nx ∂ny
Γ Γ

∂g(x) ∂
= − f (y) ln(|x − y|) dsy dsx .
∂n ∂ny
Γ Γ
For fixed x ∈ Γ there holds with the tangent unit vector t = (−n2 , n1 )
∂ ∂ ∂
ln |x − y| = % ln(ζ − z), % ln(ζ − z) , (n1 (y), n2 (y))
∂ny ∂y1 ∂y2
∂ ∂
= ( ln(ζ − z), − ( ln(ζ − z) , (n1 (y), n2 (y))
∂y2 ∂y1
d
= ∇y ( ln(ζ − z), t (y) = ( ln(ζ − z).
dsy
Again by partial integration, we obtain

∂g(x) df (y)
I= ( ln(ζ − z) dsy dsx .
∂nx dsy
Γ Γ
Now we rewrite for fixed y ∈ Γ , by partial integration and by Cauchy–Riemann

equations,

∂g(x) ∂g ∂g
( ln(ζ − z)dsx = ( , ), (n1 , n2 ) · ( ln(ζ − z)dsx
Γ ∂nx Γ ∂x 1 ∂x2

∂ ∂
=− g(x) ( ln(ζ − z), ( ln(ζ − z) , (n1 , n2 ) dsx
Γ ∂x1 ∂x2

∂ ∂
=− g(x) − % ln(ζ − z), % ln(ζ − z) , (n1 , n2 ) dsx
Γ ∂x2 ∂x1

= g(x)∇x % ln(ζ − z), (−n2 , n1 ) ) dsx
Γ
Hence we arrive at

df (y)
I = g(x) ∇x ln |x − y|, t(x) dsy dsx
ds
Γ Γ

df (y) x − y, t(x)
= g(x) dsy dsx ,
ds |x − y|2
Γ Γ
∂ϕ
what shows (2.39) in the weak sense. and setting ϕ = g, ψ = f with ϕ̇ = ∂s ,

1 ∂ ∂
W ψ, ϕ = (W ψ)(x)ϕ(x)dsx = ln |x−y|ψ(y)dsy ϕ(x)dsx
Γ Γ 2π Γ ∂nx ∂ny

1 dϕ(x) dψ(y)
=− ln |x − y|dsy dsx
2π Γ Γ dsx dsy

dϕ(x) dψ
= V( )(x) dsx = ϕ̇, V ψ̇ .
Γ dsx dsx
Of course, by completion, this formula holds for φ, ψ ∈ H 1/2(Γ ).

This latter representation (2.40) in 2D extends to an analogous representation of the
hypersingular operator in 3D via the single layer potential, see [322].
For further reading concerning the topics of this chapter see [206, 225, 259, 276].
Chapter 3
A Fourier Series Approach
The aim of this chapter is to guide the reader from elementary Fourier series
expansion to periodic Sobolev spaces on a simply connected smooth curve in R2 .
In this tour we detail on dual spaces and compact embedding. This leads to the
compactness of the double-layer operator and its adjoint. Moreover in the scale of
Sobolev spaces we prove the mapping property of the single-layer and hypersingular
operators. Then we treat the exterior Dirichlet problem for the Laplacian and derive
its explicit solution on the unit circle in terms of the Fourier coefficients. The Fourier
tour concludes with the first Gårding inequality for a bilinear form which is basic in
the BEM.
3.1 Fourier Expansion—The Sobolev Space H s [0, 2π ]
The Hilbert space L2 [0, 2π] is the completion of the space C[0, 2π] of the 2π-
periodic complex-valued continuous function on [0, 2π] with respect to the square
mean norm that is defined by the L2 [0, 2π] scalar product
2π
(f, g) := f (x) ḡ(x) dx .
0
Any function ϕ ∈ L2 [0, 2π] can be expanded into the Fourier series
∞

ak eikt
k=−∞

44 3 A Fourier Series Approach
with the Fourier coefficients
2π
1
ak := ϕ(t) e−ikt dt .
2π
0
This involves the trigonometric monoms
χk (t) := eikt (t ∈ R, k ∈ Z)
Note that {χk }k∈Z builds up an orthogonal system in L2 [0, 2π]:
2π
0 (k = l)
(χk , χl ) = ei(k−l)t dt =
2π (k = l)
0
In virtue of the Weierstrass approximation theorem the trigonometric polynomials

are dense in C[0, 2π] with respect to the maximum norm, hence with respect to
the square mean norm. Hence by construction of L2 [0, 2π], the orthogonal system
{χk }k∈Z is complete and the above Fourier series converges to ϕ with respect to the
square mean norm.
Since χk 22 = 2π, Parseval’s equation for any function ϕ ∈ L2 [0, 2π], ϕ =
k ak χk reads as follows
2π ∞
1 1 1
ϕ22 = |ϕ(t)|2 dt = (ak χk , al χl ) = |ak |2 . (3.1)
2π 2π 2π
0 k,l k=−∞
Moreover, classical analysis tells that the Fourier series of a continuously differen-
tiable 2π-periodic function even converges absolutely and uniformly.
Integration by parts for such a function ϕ gives for k = 0,
2π 2π
−ikt 1
2πak = ϕ(t) e dt = ϕ (t) e−ikt dt (3.2)
ik
0 0
2π
whereas for k = 0, 1
2π ϕ (t) dt = 0 by periodicity of ϕ. Hence (3.1) applied to ϕ
0
leads to
∞
2π
1
k |ak | =
2 2
|ϕ (t)|2 dt < ∞ .
2π
k=−∞ 0
3.1 Fourier Expansion—The Sobolev Space H s [0, 2π] 45
This suggests to define function spaces of 2π-periodic functions by prescribing a

certain decay of the Fourier coefficients ak for |k| → ∞.
Definition 3.1 Let 0 ≤ s < ∞. Then the Sobolev space H s [0, 2π] is the space of
all functions ϕ ∈ L2 [0, 2π] with the property
∞

(1 + k 2 )s |ak |2 < ∞
k=−∞
for the Fourier coefficients ak of ϕ.

Example 3.1 For s = 0, clearly H 0 [0, 2π] = L2 [0, 2π].
By the very definition of the function space H s [0, 2π] it makes sense to define
the norm
% ,1/2

ϕs = (1 + k ) |ak |
2 s 2
and the associated scalar product

(ϕ, ψ)s := (1 + k 2 )s ak bk ,
k∈Z
where ak , respectively bk are the Fourier coefficients of ϕ, respectively of ψ. Note

that in view of the simple inequalities k 2 ≤ 1 + k 2 ≤ 2k 2 , this norm is equivalent to
the norm of the space H s on [0, 1] that was already given in the introduction.
Proposition 3.1 H s [0, 2π] endowed with the scalar product (., .)s is a Hilbert
space. The trigonometric polynomials are dense in H s [0, 2π].
Sketch of proof: Clearly, H s [0, 2π] is a linear space. (ϕ, ψ)s is finite by the
Cauchy-Schwarz inequality for sums and thus defines a scalar product. To verify
the completeness of H s [0, 2π] use the completeness of C.
For an arbitrary ϕ ∈ H s with Fourier coefficients ak , consider the partial sum
ϕn = |k|≤n ak χk of the Fourier series. Then
∞

ϕ − ϕn 2s = (1 + k 2 )s |ak |2 → 0
|k|=n+1
for n → ∞. Thus the trigonometric polynomials are dense in H s .

Theorem 3.1 For s2 > s1 (≥ 0) there holds
H s2 [0, 2π] ⊂ H s1 [0, 2π]
with dense and compact embedding.

Proof From
(1 + k 2 )s1 ≤ (1 + k 2 )s2
(since s1 log(1+k 2 ) ≤ s2 log(1+k 2 )) it follows H s2 ⊂ H s1 with ϕs1 ≤ ϕs2 for

ϕ ∈ H s2 ; thus the identity I : H s2 → H s1 is continuous. Density of the embedding
is a consequence of the density of the trigonometric polynomials ⊂ H s2 in H s1 .
To verify the compactness of I , introduce the operator In : H s2 → H s1 by

n
In ϕ := a k χk for any ϕ ∈ H s2 with Fourier coefficients ak .
k=−n
In has finite dimensional range and thus In is compact. Now

(In − I )ϕ2s1 = (1 + k 2 )s1 |ak |2
|k|>n
1
≤ (1 + k 2 )s2 |ak |2
(1 + n2 )s2 −s1
|k|>n
1
≤ ϕ2s2 .
(1 + n2 )s2 −s1
Hence In − I → 0 in L (H s2 , H s1 ) for n → ∞ and I is compact.

In the following we provide another norm equivalent to · s . Let at first s = l ∈ N
and let C l [0, 2π] denote the space of l times continuously differentiable 2π-periodic
functions.
Theorem 3.2 For ϕ ∈ C l [0, 2π] ⊂ H l [0, 2π] with l ∈ N,
& 2π '1/2

ϕl,0 := [|ϕ(t)|2 + |ϕ (l)(t)|2 ] dt
0
defines an equivalent norm to · l .

Proof Analogously to the calculation (3.2) , we apply l-times integration by parts
2π 2π
−ikt
(l)
ϕ (t) e dt = (+ik) l
ϕ(t) e−ikt dt
0 0
3.1 Fourier Expansion—The Sobolev Space H s [0, 2π] 47
and use (3.1) for ϕ and ϕ (l) to obtain

∞

ϕ2l,0 = 2π (1 + k 2l ) |ak |2 .
k=−∞
Since for k = 0,
(1 + k 2l ) ≤ (1 + k 2 )l ≤ (2k 2 )l ≤ 2l (1 + k 2l ) ,
the claimed norm equivalence follows.

Proposition 3.2 For ϕ ∈ C 1 [0, 2π], 0 <p<1
& 2π 2π 2π

|ϕ(t) − ϕ(τ )|2 '1/2
2
ϕ0,p = |ϕ(t)| dt + t −τ 2p+1 dτ dt .
| sin 2 |
0 0 0
defines an equivalent norm to · p .

Exercise: Prove Proposition 3.2, see [Theorem 8.5][276].
Remark 3.1 The right hand integral exists, since C 1 [0, 2π] ⊂ C 0,1 [0, 2π], that is,
∀ϕ ∈ C 1 [0, 2π] ∃Cϕ : |ϕ(t) − ϕ(τ )| ≤ Cϕ |t − τ |
and because of
2x π
≤ sin x for 0 ≤ x ≤
π 2
there holds
|ϕ(t) − ϕ(τ )|2 |t − τ |2

≤ C̃ϕ = C̃ϕ |t − τ |−1+ε
| sin t −τ
2 | 2p+1 |t − τ | 2p+1
for some ε > 0.

Corollary 3.1 Let s = k + p, k ∈ N, 0 < p < 1. Then for ϕ ∈ C k+1 [0, 2π]
& '1/2
ϕk,p := ϕ20 + ϕ (k) 20,p .
defines an equivalent norm to · s .

3.2 The Sobolev Space H s (Γ )
In what follows, let Γ be the boundary of a simply connected bounded domain

Ω ⊂ R2 of class C k for some k ∈ N. Thus there exists a parameter representation
(x1 , x2 ) = Z(t), t ∈ [0, 2π],

that is k-times continuously differentiable with dZ
dt ≥ γ0 > 0 for all t ∈ [0, 2π].
Given the parameter representation Z, we define for any s ∈ [0, k] the Sobolev
space
H s (Γ ) = {ϕ ∈ L2 (Γ ) : ϕ ◦ Z ∈ H s [0, 2π]} .
where the scalar product on H s (Γ ) stems from the scalar product via
(ϕ, φ)H s (Γ ) := (ϕ ◦ Z, φ ◦ Z)H s [0,2π] .
Of course we want to admit different regular parameter representations for Γ .

Therefore we have to show the invariance of our definition with respect of a change
of the parameter representation.
Theorem 3.3 Let x = Z(t), x = Z̃(t) (t ∈ [0, 2π]) (the same parameter interval
without reduction of generality) be two different parameter representations for Γ .
Then for any s ∈ [0, k]
H̃ s (Γ ) := {ϕ ∈ L2 (Γ ) : ϕ ◦ Z̃ ∈ H s [0, 2π]}
is homeomorphic to H s (Γ ) defined above.

To show this theorem it is enough to apply the following lemma in the case f =
Z −1 ◦ Z̃.
Lemma 3.1 Let f be a diffeomorphism of the interval [0, 2π] onto itself of class
k ∈ N, that is, f is a bijection, f and f −1 belong to C k [0, 2π]. Let 0 ≤ s ≤ k.
Then for any ϕ ∈ H s [0, 2π], we have ϕ ◦ f ∈ H s [0, 2π] with
ϕ ◦ f s ≤ Cϕs ,
where the constant C only depends on f, k and s.

For a proof of this lemma using appropriate equivalent norms we refer to [276,
Lemma 8.14].
3.3 Interior Dirichlet Problem 49
3.3 Interior Dirichlet Problem
Similar to [409], where the exterior Dirichlet problem is considered, we apply

Fourier series techniques to solve the interior Dirichlet problem for the Laplacian in
the unit sphere.
−u = 0 in Ω = {z ∈ C : |z| < 1}

u = u0 on Γ = ∂Ω
We start from Symm’s boundary integral equation of the first kind
V q = (I + K)u0
for the Cauchy data q = ∂u

∂n |Γ , u0 = u|Γ , that is more detailed

1
− q(ζ ) ln |z − ζ | dγ (ζ ) = f (z) (z ∈ Γ ) ,
π
Γ
where

1 ∂
f (z) := u0 (z) − u0 (ζ ) ln |z − ζ | dγ (ζ ) .
π ∂n(ζ )
Γ
Since (K 1)(z) = −1 (z ∈ Γ ), (Exercise, use Fourier expansion and Calderon

projector) we have
- .
1 ∂
f (z) dγ (z) = u0 (z) 1 − ln |ζ − z|dγ (ζ ) dγ (z) = 0 . (3.3)
π ∂n(z)
Γ Γ Γ
Now we use complex coordinates
z = x + iy, ζ = ξ + iη,
the standard parametrization of the unit sphere S(0, 1)
x = cos τ, ξ = cos t
0 ≤ t, τ ≤ 2π ,
y = sin τ, η = sin t
and Fourier expansion of the sought 2π-periodic function q : [0, 2π] → C,
2π
q(t) = q̂k e ikt
, q̂k = q(τ ) e−ikτ dτ .
k∈Z 0
1
Thus we can rewrite the left hand side πL of the boundary integral equation as
follows:
2π
L := − q(t) ln |z − ζ (t)| dt
0
2π
= − q̂k eikt ln |z − ζ(t)| dt
0 k∈Z
2π
t − τ
= − q̂k eikt ln 2 sin dt ;
2
0 k∈Z
since
|z − ζ |2 = (x − ξ )2 + (y − η)2
= (cos τ − cos t)2 + (sin τ − sin t)2
t +τ τ −t τ +t τ −t
= 4 sin2 sin2 + 4 cos2 sin2
2 2 2 2
t −τ
= 4 sin2
2
we have indeed
t − τ

|z − ζ | = 2 sin . (3.4)
2
The substitution t = t − τ leads to
2π
t
L=− e ikτ
q̂k eikt ln 2 sin dt .
2
k∈Z 0
Since by Fourier expansion

∞
cos νx x
= − ln 2 sin für 0 < x ≤ π
ν 2
ν=1
3.3 Interior Dirichlet Problem 51
we have for k ∈ Z
2π π ∞

ikt t cos νt
− e ln(2 sin ) dt = e ikt
+e −ikt
dt
2 ν
0 0 ν=1
∞ 2π
11 k−ν k+ν
= cos t + cos t dt
2 ν 2 2
ν=1 0
%
π
|k| if |k| ∈ N,
= (3.5)
0 if k = 0
Consequently
π
L= eikτ q̂k .
|k|
k∈Z
k=0
With

f (τ ) = fˆk eikτ
k∈Z
we obtain (by (3.5) we have fˆ0 = 0)
1 1
L= eikτ q̂k = eikτ fˆk
π |k|
k∈Z k∈Z
k=0 k=0
what results in
q̂k = |k| fˆk for k = 0 .
Finally we note that by the first Green formula we have

0= Δu dx = q(z) dγ (z) = q̂0 .
Ω
Γ
Thus the unique solution of the boundary integral equation is completely

determined by its Fourier sum as above.
Exercise Let ûk be the Fourier coefficients of u. Then for s ∈ R

u2H s [0,2π] = |k|2s |ûk |2 + |û0 |2 (3.6)
k∈Z
k=0
is an equivalent norm on H s [0, 2π]. For l ∈ N0 ,
& '1/2
uH l (Γ ) ∼
= |D α u|2 dγ
|α|≤l Γ
and for s ∈ R, s = l + p, 0 < p < 1, l ∈ N there holds

& |D α u(x) − D α u(y)|2 '1/2
uH s (Γ ) ∼
= u2H l (Γ ) + dγ (x) dγ (y) ,
|x − y|1+2p
|α|≤l Γ Γ
which is known as the (Aronszajn -) Slobodeckij norm, see e.g. [171, 259].
3.4 The Boundary Integral Operators in a Scale of Sobolev

Spaces
3.4.1 The Operators V and W
Theorem 3.4 Let Γ ∈ C ∞ . Then for any σ ∈ R, the boundary integral operators
V : H σ (Γ ) → H σ +1 (Γ ) ,
W : H σ (Γ ) → H σ −1 (Γ )
are continuous.
Proof Let Γ be parametrized by x = Z(t), respectively by y = Z(τ ) with t, τ ∈
[0, 2π]. For any ϕ ∈ H σ (Γ ), that is τ → ϕ(Z(τ )) ∈ H σ [0, 2π], we have to show
V [ϕ] ◦ Z ∈ H σ +1 [0, 2π]. We start from

1
(V [ϕ] ◦ Z)(t) = − ϕ(y) ln |Z(t) − y| dγ (y)
π
Γ
2π
1
=− ϕ(Z(τ )) |Ż(τ )| ln |Z(t) − Z(τ )| dτ ,
π
0
3.4 Integral Operators in a Scale of Sobolev Spaces 53
and decompose
2π 2π Z(t) − Z(τ )

1 τ −t 1
(V [ϕ] ◦ Z)(t) = − f (τ ) ln |2 sin | dτ − f (τ ) ln τ − t dτ
π 2 π 2 sin
0 0 2
=: V0 [f ](t) + V1 [f ](t) .
Here
f (τ ) := |Ż(τ )| ϕ(Z(τ ))
is a product of a C ∞ -function and a H σ -function, hence lies in H σ [0, 2π] what can
be seen by using an appropriate equivalent norm of H σ .
In virtue of the above calculation (see in particular the above calculation of L)
we have
- .
V0 [f ](t) = V0 fˆk eik. (t)
k∈Z
1
= fˆk eikt ,
|k|
k=0
with
- .
V0 [f ] = 0 .
0
To analyse V0 we use appropriate equivalent norms (see the exercise above!) and
obtain
fˆ 2
k
V0 [f ]2σ +1 ∼
= |k|2(σ +1)
|k|
k∈Z
k=0

≤ |fˆk |2 |k|2σ + |fˆ0 |2 ∼
= f 2σ .
k=0
To treat V1 we use
t 1
Z(t) − Z(τ ) = Ż(η) dη = (t − τ ) Ż(τ + ξ(t − τ )) dξ
τ 0
and the identity (see the above calculation of L)

t −τ
|eit − eiτ |2 = (cos t − cos τ )2 + (sin t − sin τ )2 = 4 sin2 .
2
Thus we can see that the function
1
Z(t) − Z(τ ) t −τ
ζ (t, τ ) := τ −t = τ −t Ż(τ + ξ(t − τ )) dξ,
2 sin 2 sin 0
2 2
which is 2π-periodic in t and τ , belongs to C ∞ . Furthermore by the regularity of
Γ , we have
Z(t) − Z(τ )

|ζ (t, τ )| = it = 0 ,
e − e−iτ
hence also ∈ C ∞ , what results in V1 [f ] ∈ C ∞ . This proves the claim concerning
V.
By Proposition 2.1

1 dϕ(y) x − y, t (x)
W [ϕ](x) = − dsy
π ds |x − y|2
Γ
we obtain with y = Z(τ ), dsy = dγ (y) = |Ż(τ )| dτ
2π
1 d(ϕ ◦ Z)(τ ) Z(t) − Z(τ ), t (Z(t))
(W [ϕ] ◦ Z)(t) = − dτ
π dτ |Z(t) − Z(τ )|2
0
2π
1 d& 1 d(ϕ ◦ Z)(τ ) '
= − ln |Z(t) − Z(τ )| dτ
|Ż(t)| dt π dτ
0
1 d & - d(ϕ ◦ Z) . - d(ϕ ◦ Z) . '

= V0 (t) + V1 (t) .
|Ż(t)| dt dτ dτ
Now we can apply the mapping properties of V0 , and V1 above and conclude
d
ϕ ∈ H σ (Γ ) → ϕ ◦ Z ∈ H σ [0, 2π] → (ϕ ◦ Z) ∈ H σ −1 [0, 2π]
dτ
- d(ϕ ◦ Z) . d - d(ϕ ◦ Z) .
→ V0 ∈ H σ [0, 2π] → V0 ∈ H σ −1
dτ dt dτ
1 d - d(ϕ ◦ Z) .
→ V0 ∈ H σ −1 [0, 2π]
|Ż(t)| dt dτ
3.4 Integral Operators in a Scale of Sobolev Spaces 55
respectively
1 d - d(ϕ ◦ Z) .
V1 ∈ C∞ .
|Ż(t)| dt dτ
This shows the claim concerning D.

The mapping properties
V : H −1/2(Γ ) → H 1/2(Γ ),
W : H 1/2(Γ ) → H −1/2(Γ )
hold true also under weaker assumptions, e.g. Γ ∈ C 2 (see [276], Theorem 8.21),
even on Lipschitz curves.
3.4.2 The Operators K and K
We already know that K and K are adjoint operators in L2 (Γ ). By a density

argument this extends to arbitrary dual pairs (H s (Γ ), H −s (Γ ))L2 (Γ ) for s ∈ R,
in particular for s = 12
(Kϕ, ψ)L2 (Γ ) = (ϕ, K ψ)L2 (Γ ) (∀ϕ ∈ H 1/2(Γ ), ψ ∈ H −1/2(Γ )) .
Let us note that the parameter s = 12 plays a particular role with the boundary
integral approach to elliptic boundary value problems of second order.
3.4.2.1 A Geometric Interpretation of the Kernel of the Double Layer

Potential
Proposition 3.3 For any x, y ∈ Γ there holds
∂
(ln |y − x|) dsy = dϑx (y) ,
∂ny
where the polar coordinates y − x = r(cos ϑ, sin ϑ) are used.

Proof Any parametrization y = Z(τ ) leads to functions r = r(τ ), ϑ = ϑ(τ ) with
respect to the above polar coordinates. With the components
y1
y= y2 ,
dy ẏ1
dτ = ẏ2
we have
/
−
→ ẏ1 −n2 (y)
t = / ẏ1 + ẏ2 =
2 2 ,
ẏ2 n1 (y)
/
→
−
n =
ẏ2
/ ẏ12 + ẏ22
−ẏ1
and
/
ds = ẏ12 + ẏ22 dτ .
Hence
∂ 1
(ln |y − x|) dsy = 2 n1 (y)(y1 − x1 ) + n2 (y)(y2 − x2 ) dsy
∂ny r
1
= (y1 − x1 ) dy2 − (y2 − x2 ) dy1
r2
1
= 2 r cos ϑ d(r sin ϑ) − r sin ϑ d(r cos ϑ)
r
1
= 2 r cos ϑ sin ϑ dr + r 2 cos2 ϑ dϑ
r
−r sin ϑ cos ϑ dr + r 2 sin2 ϑ dϑ
= dϑx (y)

One can show for Γ ∈ C 2 that

dϑx
=: k(s, σ ) ,
ds
(where x = z(σ ), y = z(s) are related to the arc lenght parameter s) is continuous
on R2 and lim k(s, σ ) = 12 · (curvature of Γ in s ). Consequently, for Γ ∈ C ∞ ,
σ →s
the integral kernel k belongs to C ∞ and the integral operators K, K : H s (Γ ) →
H t (Γ ) are continuous for any s, t ∈ R) and because of compact embedding, K, K :
H s (Γ ) → H s (Γ ) (∀s ∈ R) moreover compact.
Here we like to mention the fundamental article by Radon [347] where he
introduces the boundaries of bounded rotation. In Rn with n ≥ 2 this idea is further
developed by J. Kral in [272] and Maz’ya in [301].
3.5 Solution of Exterior Dirichlet Problem by BIE 57
3.5 Solution of Exterior Dirichlet Problem by BIE
In this section we treat the following exterior Dirichlet problem for the Laplacian by
boundary integral methods (BIE) and provide its explicit solution on the unit circle
via Fourier series (see [409]).
c
Example 3.2 For sufficiently smooth curve Γ = ∂Ω find u ∈ C 2 (Ω c ) ∩ C 0 (Ω )
with
Δu = 0 in Ω c := R2 \ Ω, u|Γ = g (3.7)
where we demand the decaying condition at infinity: For given B ∈ R there exist
some constants c, k such that for all z ∈ R2 with |z| ≥ k:
|u(z) − B ln |z| | ≤ c < ∞. (3.8)
This means that for some a ∈ R
u(z) = a + B ln |z| + o(1) as |z| → ∞ .
Note that this decaying condition is weaker than that used in the representation
theorem. The exterior boundary value problem (3.7) arises in many applications:
potential flow, solid mechanis, conformal mappings.
Due to the considerations above, we can try to find the solution u in the form

1
u(z) = − ψ(ζ ) ln |z − ζ |dsζ − ω. (3.9)
π Γ
with an unknown constant ω and the unknown density ψ.

Therefore with

1 1 ζ
u(z) = − ln |z| ψ(ζ )ds − ψ(ζ ) ln |1 − |dsζ − ω
π Γ π Γ z
and |z| → ∞ we have due to (3.8)

1
− ψ(ζ )dsζ = B. (3.10)
π Γ
The BVP (3.7) with the decaying condition is thus reduced to the system of
boundary integral equations
g = − π1 Γ ψ(ζ ) ln |z − ζ |dsζ − ω on Γ
(3.11)
Γ ψ(ζ )dsζ = −πB
Fig. 3.1 Setting of unit circle z
ζ x = cos 2πτ ξ = cos 2πt

y = sin 2πτ η = sin 2πt
for the unknown function ψ on Γ and the unknown constant ω ∈ C with a given
function g on Γ and a given constant B. If ψ and ω in (3.11) are found then (3.9)
yields for all z ∈ Ω c the desired potential u.
For Γ , the unit circle, (3.11) can be solved explicitly via Fourier series. This will
be the basis of the later on given analysis of (3.16) for general boundaries Γ and
the finite element approximation. With the coordinates for z = x + iy, ζ = ξ + iη
on Γ (Fig. 3.1) we obtain as before by the addition theorem of the sin function,
|z − ζ | = |2 sin π(t − τ )|.
Next we expand ψ ∗ = 2πψ into a Fourier series:
∞
1
ψ ∗ (t) = e ik2πt
ψ̂k , ψ̂k = ψ ∗ (t)e−ik2πt d t. (3.12)
k=−∞ 0
Substitution into (3.11) yields with ds = 2πdt, t = t − τ ,

1 ∞

− ψ(ζ ) ln |z − ζ |dsζ − πω = − ψ̂k e2π ikt ln |2 sin π(t − τ )|dt − πω
Γ 0 k=−∞
∞
1

=− ψ̂k e2π ikt ln |2 sin πt |e2π ikτ dt − πω.
k=−∞ 0
For the integral we compute explicitly as above using periodicity

1 %
ik2πt − 2|k|
1
, k = 0
ln |2 sin πt |e dt = (3.13)
0 0 , k=0
Hence (3.11) becomes

1
ψ̂k eik2πτ − πω = πg, ψ̂0 = −πB. (3.14)
2|k|
k=0
Inserting the Fourier expansion

∞
1
g= e 2πikτ
ĝk , ĝk = g(t)e−2πikt d t
k=−∞ 0
3.5 Solution of Exterior Dirichlet Problem by BIE 59
of the given function g into (3.14) and equating coefficients in the expansion yields
− ω = ĝ0 , ψ̂k = 2|k|π ĝk for k = 0, ψ̂0 = −πB. (3.15)
Thus for the unit circle the solution ψ of (3.16) is completely determined by its
Fourier series.
On a smooth curve Γ the integral equations look like

−1
(V + C)ψ(z) = g(z) + ω, z ∈ Γ; ψds = B (3.16)
π Γ

1
V ψ(z) := − ln |z − ζ |ψ(ζ )dsζ ,
π Γ

Cψ(z) := L(z, ζ )ψ(ζ )dsζ , z ∈ Γ.
Γ
with some smoother kernel L(z, ζ ).

In order to analyze the solvability of (3.16) we introduce the Sobolev spaces
H m (Γ ) even for non-integers m. Therefore, we define a norm via the Fourier
coefficients of the corresponding functions:
⎧ ⎫1/2
⎨
∞ ⎬
f H m (Γ ) := |j |2m |fˆj |2 + |fˆ0 |2 . (3.17)
⎩ ⎭
j =−∞
For the integral equations in (3.16) we have the following well posedness result in
Sobolev spaces:
Theorem 3.5 ([409]) Let s ∈ R be fixed and let the solution of (3.16) be unique.
Then for any s ∈ R
"
(V + C)ψ − ω = g
(3.18)
− π1 Γ ψgds = B
is a bijective, continuous mapping from H s (Γ ) × R onto H s+1(Γ ) × R.

Proof The Fourier approach allows to give the following short argument for the
continuity of V for s = −1/2, L = 0 and Γ being the unit circle. With (3.17) and
(3.14) and a generic constant c > 0 we have
∞
∞

V ψ2H 1/2 (Γ ) = |j | |ĝj |2 ≤ c |j |−1 |ψ̂j |2 ≤ cψ2H −1/2 (Γ ) .
j =−∞ j =−∞
j =0 j =0
The general case can be derived from the Fredholm alternative see [257].

3.6 A First Gårding Inequality
Due to the above mapping properties the bilinear form
a(ψ, φ) := ((V + C)ψ, φ)L2 (Γ ) (3.19)
is continuous on H −1/2(Γ ) × H −1/2 (Γ ). It is also coercive in the sense of a

Gårding’s inequality as seen from the following.
Theorem 3.6 ([409]) For sufficiently smooth Γ there exists constants γ > 0, c >
0 such that for all ψ ∈ H −1/2(Γ )
a(ψ, ψ) ≥ γ ψ2H −1/2 (Γ ) − cψ2H −1 (Γ ) . (3.20)
Proof First we show (3.20) for L = 0 and Γ = {z ∈ C : |z| = 1}. Hence

1
a(ψ, ψ) = (V ψ, ψ)L2 (Γ ) = (V ψ)(τ )ψ (τ )dτ.
0
As in (3.12) – (3.15) there holds

1 ∞

1
(V ψ, ψ)L2 (Γ ) = ψ̂k eik2πτ · e−i2πτ ψ̂ dτ
0 k∈Z, k=0 2|k|π
=−∞
∞
1
1
= ψ̂k ψ̂ ei2πτ (k−) dτ .
2π|k| 0
k=0 =−∞
Thus
1
1 1
(V ψ, ψ)L2 (Γ ) = ψ2H −1/2 (Γ ) − | ψdt|2 .
2π 2π 0
But since
1 1
| ψ(t)dt|2 ≤ |ψ̂0 |2 + |ψ̂k |2 = ψ2H −1 (Γ )
0 k2
k=0
we obtain from (3.16) the inequality
1 1
(V ψ, ψ)L2 (Γ ) ≥ ψ2H −1/2 (Γ ) − ψ2H −1 (Γ ) . (3.21)
2π 2π
3.6 A First Gårding Inequality 61
In order to prove (3.20) in the general case we use (3.18) and proceed as follows
with a smooth kernel L(z, ζ ):
1
1 1
((V + C)ψ, ψ)L2 (Γ ) = − ψ(t) log |2 sin π(t − τ )|dt ψ(τ )dτ
π 0 0
1
1 1 z(t) − ζ(t)
+ − ψ(t) log dt
0 π 0 2 sin π(t − τ )
1
+ L(z, ζ )ψ(t)dt ψ(τ )dt.
0
For the first integral, we have already derived (3.21). In order to obtain an estimate
for the last two integrals we write them as
1 1
ψ(t)η(t, τ )ψ(τ )dtdτ
0 0
and estimate:
0 1 0
1 0 0

ψ, η(t, τ )ψ(t)dt ≤ cψH −1 (Γ ) 0
0 η(t, τ )ψ(t)dt 0
0 1
0 L2 (Γ ) 0 H (Γ )
0 02 1 1 "2
0 1 0
0 η(t, τ )ψ(t)dt 0 := η(t, τ )ψ(t)dt dτ
0 0
0 H 1 (Γ ) 0 0

∂η 1 1 "2
+ (t, τ )ψ(t)dt dτ
0 0 ∂τ
( % 1 2,
1 1 1 ∂η 2
≤c ψH −1 (Γ )
2
|η| dt +
2 dt dτ
∂t
0 0 0
1 % 1 1 2 2 2 )
1 ∂η 2 ∂ η
+ ψH −1 (Γ )
dt + dt }dτ
2

0 0 ∂τ 0 ∂τ ∂t
% 1 2 2 2 2 2 ,
1 1 ∂η ∂η ∂ η
= cψH −1 (Γ )
2
η + + +
2 dtdτ .
0 0 ∂t ∂τ ∂τ ∂t
Hence, since for sufficiently smooth Γ the terms in the brackets are bounded, we
have
0 1 0
0 0
0 η(t, τ )ψ(t)dt 0 ≤ cψH −1 (Γ )
0 0
0 H 1 (Γ )
and therefore
1
1
− ψ(t)η(t, τ )ψ(t)dtdτ ≥ −c2 ψ2H −1 (Γ ) .
0 0
This yields together with (3.21) the desired estimate (3.20).

For further reading see [276, 324, 343, 356, 409].
Chapter 4
Mixed BVPs, Transmission Problems
and Pseudodifferential Operators
This chapter uses Fourier transform and the modern theory of pseudodifferential
operators, see Appendix B. It brings a deeper insight in mixed boundary value
problems in the interior and exterior of a connected surface in 3D. In particular, the
Helmholtz interface problem from acoustics is studied in the presence of an obstacle
in 3D. In Sect. 4.1 we consider a direct boundary integral equation method for the
mixed boundary value problems (bvp). Then in Sect. 4.2 we look at the transmission
problem and first treat it by the indrect method based on a single layer potential
ansatz yielding a Riesz-Schauder system of second kind integral equations. Then
we treat the transmission problem by the direct method giving a strongly elliptic
system of boundary integral operators on the transmission manifold. In Sect. 4.3 we
consider screen problems. In Sect. 4.4 the smoothness assumption of an analytic
interface is relaxed to only Lipschitz continuity. in Sect. 4.5 we present a strongly
elliptic system of pseudodifferential operators for the exterior Maxwell’s equations.
4.1 Mixed Boundary Value Problems
In this section we report on the paper [397]. Let Ω1 denote a bounded simply
connected domain in R3 and Ω2 = R3 \Ω 1 , where Γ = ∂Ω1 = ∂Ω2 is assumed to
be a sufficiently smooth, connected surface, for brevity C ∞ . Γ is divided into two
disjoint pieces Γ1 and Γ2 such that Γ 1 ∩ Γ 2 = ∂Γ1 = ∂Γ2 = γ defines a simple
closed, smooth curve on Γ (see Fig. 4.1).
The interior j = 1 (exterior: j = 2) mixed boundary value problem reads as: To
given g1 on Γ1 and g2 on Γ2 find complex-valued uj in Ωj such that
∂uj
(Δ + kj2 )uj = 0 in Ωj , uj = g1 on Γ1 and = g2 on Γ2 (4.1)
∂n

64 4 Mixed BVPs, Transmission Problems

[397]
holds, where the solution u2 of the exterior problem has to satisfy at infinity:

∂u2 (x) 1
If k2 = 0 : − ik2 u2 (x) = o as |x| → ∞. (4.2)
∂|x| |x|
(SOMMERFELD’S radiation condition)

1
If k2 = 0 : u2 (x) = 0 as |x| → ∞. (4.3)
|x|
Here k1 , k2 ∈ C with (kj ≥ 0. ∂u

∂n means the normal derivative with respect to the
outward unit normal n to Ω1 . We make the general assumption
k12 , k22 are different from the eigenvalues of the interior and the (4.4)
exterior mixed boundary value problem, respectively.
Before we give the variational formulation of the mixed boundary value problem
(4.1)–(4.3) let us introduce the Sobolev spaces H s (Ωj ), H s (Γ ), H s (Γj ) for smooth
Γ, Γj as defined in the usual way [284],
& '
H s (Ωj ) = u|Ωj : u ∈ H s (R3 ) (s ∈ R)
⎧3 4
⎨ u|Γ : u ∈ H s+1/2(R3 ) (s > 0)
H s (Γ ) = L2 (Γ ) (s = 0) (4.5)
⎩ −s
H (Γ ) (dual space) (s > 0)
3 4
H s (Γj ) = u|Γj : u ∈ H s (Γ ) (s ≥ 0)
3 4
H̃ s (Γj ) = u ∈ H s (Γ ) : supp u ⊂ Γ j , H s (Γ1 ) = H s (Γ )/H̃ s (Γ2 )

H s (Γj ) = H̃ −s (Γj ) (s < 0), H̃ s (Γj ) = H −s (Γj ) (s < 0).
The spaces are endowed with their natural norms [284] which we shall recall later.
The most general case where (4.1)–(4.3) can be converted into a variational
problem is the following:
g1 ∈ H 1/2(Γ1 ), g2 ∈ H −1/2(Γ2 ) are given, and we look for u ∈ Hloc

1
(Ωj ).
4.1 Mixed Boundary Value Problems 65
In this case ∂u
∂n ∈ H̃ −1/2 (Γ2 ) ⊂ H −1/2(Γ ) is defined by GREEN’S formula:
Lemma 4.1 Let u ∈ Hloc 1 (Ω ) with Δu ∈ L2 (Ω ) and v ∈ H 1 (Ω ) with
j loc j j
−1/2 (Γ ) is defined by
∂n Γ ∈ H
bounded support. Then ∂u
5 6
∂u
v · Δu dx + ∇v · ∇u dx = (−1)j +1 , v| (j = 1, 2) (4.6)
∂n Γ
Γ
Γ
Ωj Ωj
Here ·, · Γisthe duality between H −1/2 (Γ ) = (H 1/2(Γ )) and H 1/2(Γ ), given
by f, g Γ = f (z)g(z)dsz for smooth functions f and g.
Γ
The mapping u → ∂u
∂n Γ is an extension by continuity of the corresponding trace
mapping for smooth functions.
& '
Now, let u ∈ Lj = uj ∈ H 1 (Ωj ) : (Δ + kj2 ) = uj in Ωj be the variational

solution of (4.1)–(4.3) with u|Γ1 = g1 and ∂u
∂n Γ2 = g2 . Then with arbitrary
u|
extensions lg1 ∈ H 1/2(Γ ) and lg2 ∈ H −1/2(Γ ) the Cauchy data ψv = ∂u Γ|
∂n Γ
admit the form
v = v 0 + lg1 , ψ = ψ 0 + lg2 with v 0 ∈ H̃ 1/2 (Γ2 ) and ψ 0 ∈ H̃ −1/2 (Γ1 ),
because v 0 |Γ1 = 0 and ψ 0 |Γ2 = 0.

0
Definition 4.1 ψv 0 ∈ H̃ 1/2(Γ2 ) × H̃ −1/2 (Γ1 ) are called the unknown CAUCHY
data of the variational solution of (4.1)–(4.3).
They are the unknown layers of our system of boundary integral equations in the
following.
In order to formulate the jump relations for the single and double layer potentials
we define the following boundary integral operators.
Definition 4.2 Let ψ ∈ C ∞ (Γ ), Γ a bounded closed C ∞ -surface. Then for z ∈
Ωj , j = 1, 2 we define with
⎧ −1
⎪
⎨ 4π|z−ζ | for kj = 0
ϕj (z, ζ ) =
⎪
⎩ e j
ik |z−ζ |
−4π|z−ζ | for kj = 0

VΩj ψ(z) := −2 ψ(ζ )ϕj (z, ζ )dsζ
Γ

∂
KΩj ψ(z) := −2 ψ(ζ ) ϕj (z, ζ )dsζ
∂nζ
Γ
and for z ∈ Γ the operators

Vj ψ(z) := −2 ψ(ζ )ϕj (z, ζ )dsζ (4.7)
Γ

∂
Kj ψ(z) := −2 ψ(ζ ) ϕj (z, ζ )dsζ (4.8)
∂nζ
Γ

∂
Kj ψ(z) := −2 ψ(ζ ) ϕj (z, ζ )dsζ (4.9)
∂nz
Γ
∂
Wj ψ(z) := − KΩj ψ(z). (4.10)
∂nz
For a distribution ψ we define Vj ψ and Kj ψ by approximating ψ by smooth

functions and Kj ψ by duality using the relation
7 8 9 :
Kj ψ, ω = ψ, Kj ω Γ , for all ω ∈ C ∞ (Γ )
Γ
which for smooth ψ is obviously valid.

The extension to distributions makes sense since all arising operators are
pseudodifferential operators (see Sect. 4.2).
Since Γ is assumed to be C ∞ it follows that the function X is C ∞ and that an
asymptotic expansion holds
∞

|X(U ) − X(u)| ∼ Mν (U, u − U ) for |u − U | → 0
ν=1
where Mν is positive homogeneous of degree ν in u − U . We have

∞
−1 δl l
ϕj (r) = r , δl (kj ) ∈ C, r = |X(U ) − X(u)| (4.11)
4πr l!
l=0
and therefore
−1 ∞

ϕj X(U ) − X(u) ∼ + Llj (U, u−U ) for |u−U | → 0 (4.12)
4π|u − U |
l=0
with Llj positive homogeneous of degree l in u − U . Substituting (4.12) into (4.7)

and applying Fourier transform shows that for smooth Γ , Vj is a pseudodifferential
operator of order −1 with principal symbol
σ (Vj )(ξ ) = |ξ |−1 . (4.13)
Using expansion (4.11) together with surface polar coordinates one can show that
the operator of the double layer potential Kj is a pseudodifferential operator of order
−1, too (see [259, 287]).
Now we define the matrix of operators
( )
−Kj Vj
Aj := . (4.14)
Wj Kj
Due to the mapping properties of the operators we have
H 1/2(Γ ) H 1/2(Γ )
Aj : × → × is continuous,
H −1/2 (Γ ) H −1/2 (Γ )
and by the Calderon projector there holds the following result [129]:
Theorem 4.1
v
a) The statements (i) and (ii) on ψ ∈ H 1/2(Γ ) × H −1/2 (Γ ) are equivalent:

(i) ψv are CAUCHY data of some u ∈ Lj
(4.15)
(ii) I + (−1)j Aj ψv = 0

b) The operators 12 I − (−1)j Aj are projection operators, the so-called
;
“CALDERON-projectors”. They project H 1/2(Γ ) H −1/2 (Γ ) onto the
CAUCHY data of the weak solutions in Lj . This means in particular A2j = I ,
yielding the relations
Kj2 + Vj Wj = I = Wj Vj + Kj2
(4.16)
−Kj Vj + Vj Kj = 0 = −Wj Kj + Kj Wj .
Whereas the operators Vj , Kj and Kj are weakly singular integral operators on
Γ , the operator Wj of the normal derivative of the double layer potential is a
hypersingular integral operator, its kernel is 0(|z − ζ |−3 ) as z → ζ . Wj is a
pseudodifferential operator of order +1 [129, 395]. The relation (4.16) shows that Vj
is a regularizer to Wj since Kj2 , Kj2 are lower order pseudodifferential operators and
therefore compact perturbations. Thus (4.16) together with (4.13) gives the principal
symbol of Wj as
σ (Wj )(ξ ) = |ξ | (4.17)

/
with |ξ | = ξ12 + ξ22 , ξ ∈ R2 \{0}. Obviously σ (Wj )(ξ ) can also be computed by
using local coordinate systems on the smooth manifold Γ and transformation to the
case Γ = R2 .
Now we give a solution procedure for the interior mixed boundary value problem
via the direct method by inserting the boundary data into the system
"
1 ∂u
u= (I − K) u + V ,
2 ∂n
∂u 1 ∂ 1 ∂u
=− Ku + I + K .
∂n 2 ∂n 2 ∂n
This gives

W22 K12 v −W12 I − K22 g1
= , (4.18)
−K21 V11 ψ I + K11 −V21 g2
where the subscripts at Wj k etc. mean: integration over Γj and evaluation on Γk .

In order to describe the solvability of the above system we first give results on the
mapping properties of the involved operators Wj k , Vj k , Kj k , Kj k , (j, k = 1, 2). We
want to use the fact that W, V , K, K are pseudodifferential operators (and hence
bounded mappings in Sobolev spaces). Now the operators Wj k etc. act only on
pieces of the manifold Γ , therefore their layers v, ψ have to be extended by zero on
the remaining part of Γ .
Lemma 4.2 For s ∈ R and i, k = 1, 2 the mappings are continuous:
Vik : H̃ (Γi ) → H s+1(Γk ) Kik : H̃ s (Γi ) → H s+1 (Γk )

: H̃ s (Γ ) → H s+1 (Γ ). (4.19)
Wik : H̃ s+1 (Γi ) → H s (Γk ) Kik i k
Proof By definition of H̃ s (Γ1 ) in (4.5) the extension of ψ ∈ H̃ s (Γ1 ) by zero

ψ on Γ1
ψ ∗ := belongs to H s (Γ ).
0 on Γ2
Therefore the continuity of the mappings (4.19) is seen by estimating the symbols
of the pseudodifferential operators V , K, K and W from above: Neglecting the
local charts we have that the simple layer potential V is a continuous mapping from
H s (Γ ) into H s+1 (Γ ) (see definition of Sobolev spaces via Fourier transform in
Appendix B) since |σ (V )(ξ )| ≤ c(1 + |ξ |)−1 :

< ∗
V ψ ∗ 2H s+1 (Γ ) = (1 + |ξ |2 )s+1 |V ψ (ξ )|2 dξ

≤c (1 + |ξ |2 )s+1 (1 + |ξ |)−2 |ψ ∗ (ξ )|2 dξ

≤ c̃ (1 + |ξ |2 )s |ψ̃ ∗ (ξ )|2 dξ = c̃ψ2H̃ s (Γ ) .
1
Matching up the local results and restriction to Γk yields (4.19) since
ψH̃ s (Γ1 ) = ψ ∗ H s (Γ ) . (4.20)
The other assertions in (4.19) are shown analogously. This is standard in the theory
of pseudodifferential operators [376] and [415].

In order to use Lemma 4.2 to obtain information on the solvability of the system
(4.18) we rewrite it in a form appropriate to apply the result (4.19).
Substituting v = v 0 + lg1 , ψ = ψ 0 + lg2 into the system we obtain

W22 K12 v0 −WΓ 2 (I − KΓ )2 lg1
A1 U 0 := =
−K21 V11 ψ0 (I + KΓ )1 −VΓ 1 lg2
:= B1 lG (4.21)
Here WΓ 2 etc. denotes integration on Γ and evaluation on Γ2 .

Theorem 4.2 The mappings
A1 : H̃ s (Γ2 ) × H̃ s−1(Γ1 ) → H s−1 (Γ2 ) × H s (Γ1 );

B1 : H s (Γ ) × H s−1(Γ ) → H s−1(Γ2 ) × H s (Γ1 ) (4.22)
are continuous for any real s.

Proof The mapping property of A1 is a direct consequence of Lemma 4.2 whereas
that of B1 follows directly from the continuity of the extension lgi in H s (Γ ) for
gi ∈ H s (Γi ) together with the mapping properties of the simple and the double
layer potential and their respective normal derivatives.

The system satisfies a Gårding inequality because it is a strongly elliptic system
of pseudodifferential equations in appropriate Sobolev spaces.
0
v
Theorem 4.3 There exists a constant γ1 > 0 such that for all U =
ψ0
& '
(A1 + C1 )U, U 0 ≥ γ1 v 0 2H̃ 1/2 (Γ ) + ψ 0 2H̃ (−1/2) (Γ ) , (4.23)
2 1
here
C1 : H̃ 1/2(Γ2 ) × H̃ −1/2 (Γ1 ) → H −1/2 (Γ2 ) × H 1/2(Γ1 )
is compact and
7 8 7 8

A1 U, U 0 := W22 v 0 + K12 ψ 0 , v0 + −K21 v 0 + V11 ψ 0 , ψ 0
L2 (Γ2 ) L2 (Γ1 )
Proof We use a partition of unity to reduce the global inequalities to local ones,
i.e. to the inequality (4.23) for the individual terms χk v 0 , χk ψ 0 (k = 1, . . . , N)
(with χk ∈ C0∞ (Sk ) and patches Sk covering Γ ) instead of v 0 , ψ 0 (see [397] for
: H̃ −1/2 (Γ ) → H 1/2 (Γ ) and K : H̃ 1/2 (Γ ) → H 3/2 (Γ ) are
details). Since K12 1 2 21 2 1
continuous mappings, they are compact mappings H̃ −1/2(Γ1 ) → H −1/2 (Γ2 ) and
H̃ 1/2(Γ2 ) → H 1/2(Γ1 ), respectively, by Rellich’s embedding theorem. Therefore
K12 and K are compact perturbations and the principal symbol of A has the form
21 1
( )
|ξ | 0
σ (A1 )(ξ ) = , ξ ∈ R2 \{(0, 0)}.
0 |ξ1|
Now standard arguments yields the assertion (see [397, Theorem 3.3] ).

4.2 The Helmholtz Interface Problems
If a sound wave meets an obstacle, it is partially reflected from it and partially

transmitted through it. Let us consider a steady-state sound wave that is set up in a
homogeneous medium Ω characterized by a density ρ, a damping coefficient α and
sound velocity c in which there is a homogeneous body Ω of density ρi , damping
coefficient β and sound velocity ci . We shall characterize the sound wave by the
pressure v and the angular frequency ω of the acoustic vibrations. Let the medium
occupy all space R3 with the exception of the bounded domain Ω occupied by
the obstacle. We denote by vo , vi , ve the complex-valued pressure of the incident,
refracted and scattered wave, respectively, satisfying the homogeneous Helmholtz
equations
ω(ω + iβ)
Δvi + ki2 vi = 0, ki2 = in Ω ,
ci2
(4.24)
ω(ω + iα)
2
Δve + k ve = 0, k = 2
in Ω = R 3
\ Ω .
c2
Both the total acoustic field v = ve + vo and the incident field vo satisfy the
homogeneous Helmholtz equation in the exterior domain Ω. At infinity the scattered
4.2 The Helmholtz Interface Problems 71
wave ve fulfills the Sommerfeld radiation condition

∂ve
lim r − ikve = 0, lim ve = 0. (4.25)
r→∞ ∂r r→∞
Finally, on the boundary S of the obstacle, the pressure and the velocity of
vibrations in the body and the medium must coincide, yielding the transmission
conditions

1 ∂vi 1 ∂ve ∂vo
vi = ve + vo , = + , on S (4.26)
ρi ∂n ρ ∂n ∂n
∂
where ∂n denotes differentiation with respect to the outer normal n to S. Thus
the scattering of sound is described by the interface problem (4.24)–(4.26).
For higher damping the constant β is usually large leading to the total reflection
of a plane wave at an absolutely rigid immovable obstacle. Formally this means
solving only the Helmholtz equation (4.24)2 in Ω for the scattered field and
requiring that the normal derivative of the total acoustic field vanishes on S, that is
Δve + k 2 ve = 0 in Ω = R3 \ Ω
∂ve ∂vo (4.27)
=− on S
∂n ∂n
where ve satisfies (4.25) at infinity.
In the following we assume for simplicity that S is a closed analytic surface
which divides R3 into simply connected domains, an interior Ω (bounded) and an
exterior Ω (unbounded).
In order to avoid additional difficulties we assume:
k 2 = 0 is not an eigenvalue of the interior Dirichlet problem. (4.28)
The uniqueness of the solution of the interface problem (4.24)–(4.26) and of the
exterior Neumann problem (4.27) is wellknown. For brevity we give here only the
uniqueness result for the interface problem (see [129, 395]).
Theorem 4.4 Let k, ki ∈ C \ {0} with 0 ≤ arg k, arg ki ≤ π and let μ = 1
ρ,
μi = 1
ρi ∈ C \ {0} be such that
2 2
μi k i ρki
κ= 2
= 2
∈R
μk ρi k
where κ ≥ 0 (< 0) if %k · %ki ≥ 0 (< 0). Then the only solution of the
homogeneous transmission problem (4.24)–(4.26) is ve = vi = 0.
In the following we first give a boundary integral equation method based on

simple layers for solving both the interface problem (4.24)–(4.26) and the exterior
Neumann problem (4.27). Then we give the corresponding double layer procedure.
To this end we introduce the simple layer Vγ with the continuous density ψ on the
surface S by

Vγ (ψ)(x) = ψ(y)φγ (|x − y|)dSy , x ∈ R3 (4.29)
S
Here
eiγ |x−y|
φγ (|x − y|) = (4.30)
4π|x − y|
is the fundamental solution of the Helmholtz equation Δw = −γ 2 w satisfying

the Sommerfeld radiation condition for %γ = 0. There hold the following well-
known properties of the simple layer potential [287].
Lemma 4.3 For any complex γ , 0 ≤ arg γ ≤ π
2 and any continuous ψ on S:
(i) Vγ (ψ) is continuous in R3
(ii) ΔVγ (ψ) = −γ 2Vγ (ψ) in Ω ∪ Ω
(iii) Vγ (ψ)(x) = O |x|−1 eiγ |x| as |x| → ∞
±
(iv) ∂∂ñ Vγ (ψ) (x) = ∓ 12 ψ(x) + S Kγ (x, y)ψ(y)dSy on S

where the kernel Kγ is O |x − y|−1 as y → x and ± denotes the limit to S from
Ω and Ω , respectively.
In order to describe the mapping properties of Vγ and Kγ as pseudodifferential
operators acting in Sobolev spaces we first discuss some geometric ideas (see
[287]). We introduce coordinate systems for S. These consist of a finite number of
coordinate patches S1 , . . . , SN covering S. For each patch there is a region Γk ⊂ R2
and a map Xk such that x = Xk (u), u = (u1 , u2 ) ∈ R2 , covers Sk . The mappings are
compatible on overlapping regions. To say that S is a regular analytic surface means
that the individual maps from Γk to Γe on overlaps are analytic and that Xk,u1 and
Xk,u2 are linearly independent.
We use the Xk to generate local coordinate systems in R3 and set
ẽ1 (u) = Xu1 , ẽ2 (u) = Xu2 , ẽ3 (u) = ẽ1 (u) × ẽ2 (u) (4.31)
Then the equations
x = X(u) + u3 ẽ3 (u), u ∈ Γ, |u3 | < δ
will define a coordinate system for a region Uk ⊂ R3 with u3 = 0 corresponding

to Sk . We will assume that u3 > 0 corresponds to Ω.
For simplification we further assume that the coordinate systems are orthonor-
mal, that is, ẽi (u) · ẽj (u) = δij .
Following the ideas of [376] we introduce a partition of unity k ξk ≡ 1
subordinate to the Sk and define Vγ (ψ) by

Vγ (ψ)(x) = ψ (Xk (u)) ξk (u)φγ (|x − Xk (u)|) du (4.32)
k Γk
Here the orthonormality of the coordinate system implies that the surface element
is unity. For x ∈ S, (4.32) gives

Vγ (ψ)(x) = ξj ψ (Xk (u)) ξk (Xk (u))φγ (|x − Xk (u)|) du (4.33)
j k Γk
Formula (4.33) is the basis for the idea of pseudodifferential operators on S. If

ψ ∈ C0∞ (Sk ) for some patch Sk then Vγ (ψ) will be in C ∞ (Sk ). The idea is to extend
that definition to ψ’s which need not to be C ∞ but lie in some Sobolev space on
S. It is clear from (4.33) that one needs concentrate only on the quantities χVγ (ψ)
where χ and ψ have support in the same patch Sk .
Let χ, ψ ∈ C0∞ (Sk ). Then we have

χVγ (ψ) = χ(X(U )) ψ(X(u))φγ (|X(U ) − X(u)|)du
Γk
(4.34)
= =(u)Kγ (U, u − U )du
ψ
R2
with the kernel
Kγ (U, u − U ) = χ(X(U ))φγ (|X(U ) − X(u)|).
> of ψ
Introducing the Fourier transform ψ = by

>(ξ ) =
ψ =(u)e−iξ ·u du
ψ (4.35)
R2
we can write

χVγ (ψ) = (2π) −2 >(ξ )aγ (U, ξ )dξ
eiξ ·x ψ (4.36)
R2
with

aγ (U, ξ ) = χ(X(U )) e−iξ ·η Kγ (U, η)dη
R2
Now, aγ (U, ξ ) is called the symbol of Vγ .

Suppose that Kγ (U, η) has an asymptotic expansion of the form
∞

Kγ (U, η) ∼ Kγn (U, η) (4.37)
n=r
where Kγn is homogenous of degree n in η. Then aγ , the (distributional) Fourier

transform of Kγ , has the form
∞

aγ (U, ξ ) ∼ aγn (U, ξ )
n=r
where aγn is homogeneous of degree −n−2 in ξ . If (4.37) holds then Vγ obtained

by (4.36) is called a pseudodifferential operator of order r and aγr (U, ξ ) is its
principal symbol. Vγ is called elliptic if aγr (U, ξ ) = 0 for ξ = 0.
Before we cite some results from [376] on pseudodifferential operators on S we
recall the definition of Sobolev spaces on compact manifolds S. Via diffeomorphism
χ mapping any domain U ⊂ S onto open sets Uχ in R2 the Sobolev space H r (S)
is the completion of C ∞ (S), the space of infinitely differentiable functions on S, in
the norm

χψ2H r = ?
(1 + |ξ |2 )r |χ ψ(ξ )|2 dξ, ψ ∈ C0∞ (S). (4.38)
R2
defined by a partition of unity subordinate to a covering of S by domains of charts

[253].
Lemma 4.4 ([376]) Let A be a pseudodifferential operator of order r on S. Then
(i) A is a continuous map from H t (S) into H t −r (S) for any t
(ii) If A is elliptic the map A : H t (S) → H t −r (S) is Fredholm
(iii) If A is elliptic then ψ ∈ H t (S) and Aψ ∈ H s (S) implies ψ ∈ H s+r (S) and
there is a constant Ct,s such that
ψs+r ≤ Ct,s (Aψs + ψt )
Now we apply the above ideas to Vγ and show first that the expansion (4.37)
holds. Since S is assumed to be analytic it follows that the functions X are analytic
and that
∞

|X(U ) − X(u)| = Mν (U, u − U )
ν=1
where Mν is homogeneous of degree ν in u − U . Moreover the orthonormality

of the coordinate system yields
M1 (U, u − U ) = |u − U |
Then (4.30) gives

∞ j
δ
φγ (r) = r −1 r j , δ ∈ C, r = |x − y| (4.39)
j!
j =0
Thus we obtain
∞

φγ (|X(U ) − X(u)|) = |u − U |−1 + kγν (U, u − U ) (4.40)
ν=0
with kγν homogeneous of degree ν in u − U . Substituting (4.40) into (4.34) yields

(4.37) with r = −1 and
Kγ−1 (U, η) = χ(X(U ))|η|−1 (4.41)
Hence application of Fourier transform (4.35) (η → ξ ) gives the principal

symbol
1
aγ−1 (U, ξ ) = χ(X(U )) |ξ |−1 (4.42)
2
of the pseudo-differential operator Vγ . From (4.39) follows
∞ k
δ̃
φγ (r) = φi (r) + (iγ + 1) + Φγ (r), Φγ (r) = rk (4.43)
k!
k=1
yielding the following result (cf. [287, 395]), where Vi has kernel ϕi (r).
Lemma 4.5 There holds Vγ = Vi + W =γ where W
=γ is a continuous map from H t (S)
t +3
into H (S). Vi maps bijectively H (S) onto H r+1(S) for any r ∈ R.
r
Proof Due to (4.43) the first assertion follows from the decomposition
Vγ (ψ) = Vi (ψ) + Γγ (ψ) + Wγ (ψ) (4.44)
with

1 1
Γγ = (iγ + 1) ψdSψ , Wγ (ψ)(x) = ψ(y)Φγ (|x − y|)dSγ ,
4π S 4π S
since Wγ is a pseudodifferential operator of order -3 and Γγ takes H r (S) into H t (S)

for any t.
From (4.41) and (4.42) we see that Vi is an elliptic pseudodifferential oper-
ator of order -1. Thus by Lemma 4.4 Vi is a Fredholm operator from H r (S)
into H r+1(S) for any r. Moreover Vi is self-adjoint from H − 2 (S) to H 2 (S)
1 1
= dual space of H − 2 (S) since for any ψ, χ ∈ C0∞ (S) there holds
1

ψ(x)Vi (χ)(x)dSx = χ(x)Vi (ψ)(x)dSx
S S
1
because φi depends only on |x − y|. Therefore Vi is bijective from H − 2 (S) onto
1
H 2 (S) if Vi (ψ) = 0 implies ψ = 0. Then by Lemma 4.4 (iii) the assertion holds for
any r. The injectivity of Vi follows by standard arguments: Suppose Vi (ψ) = 0 for
ψ ∈ H − 2 (S). Then by Lemma 4.4 (iii) we have ψ ∈ H r (S) for any r and hence ψ
1
is continuous. Thus due to Lemma 4.3 the potential v(x) = S ψ(y)φi (|x − y|)dSy

is continuous in R3 satisfying Δv − v = 0 in Ω ∪ Ω , moreover v = O |x|−1 e−|x|
as |x| → ∞ and v ≡ 0 on S. Application of Green’s theorem over ΩR = Ω ∪
{x, |x| < R} gives

∂v
0= (Δv − v) vdx = − |grad v|2 + |v|2 dx + v
ΩR ΩR ΓR ∂n
Thus

∂v 2
|grad v|2 + |v|2 dx = v R dω
ΩR ΓR ∂n
and the integral on the right side vanishes as R → ∞, because v and ∂v

∂n are both
e−R
O R as R → ∞. Hence ||v||H 1 (R3 ) ≡ 0 implies v ≡ 0 in R3 and ∂n = 0 on S.
∂v
− ∂v +
Now the jump relations (Lemma 4.3 (iv)) give ψ = ∂v ∂n − ∂n = 0.

Via Lemma 4.3 (iv) there is defined an operator Kγ by

1 ∂ eiγ |x−y|
Kγ (ψ)(x) = − ψ(y)dSy (4.45)
4π S ∂nx |x − y|
which is the adjoint to the operator of the double layer potential

1 ∂ eiγ |x−y|
Nγ (ψ)(x) = − ψ(y)dSy (4.46)
4π S ∂ny |x − y|
Lemma 4.6 I + 2Kγ is bijective from H r (S) onto H r (S). Moreover,
(I + 2Kγ )−1 = I + Rγ (4.47)
where Rγ is continuous from H r (S) into H r+1 (S).

Proof Lemma 4.3 (iv) shows that Kγ is a pseudodifferential operator of order -1
hence takes H r (S) into H r+1 (S). Thus I + 2Kγ is a Riesz-Schauder operator. To
show that it is bijective it suffices to show that (I + 2Kγ )ψ = 0 implies ψ = 0. If
it is bijective the formula (4.47) follows from the theory in [376].
Suppose, then, that (I + 2Kγ )ψ = 0. As before, we can use Lemma 4.4
to conclude that ψ is smooth. Now define v by v = Vγ (ψ). We will have

Δ + γ 2 v = 0 in Ω and (I + 2Kγ )ψ = 0 on S implies ∂n ∂v
= 0 on S. Now,
uniqueness of this exterior Neumann problem gives v ≡ 0 in Ω. But we can also
set v = Vγ (ψ) in Ω . Assuming that γ 2 = 0 is not an eigenvalue of the interior
Dirichlet problem we deduce v ≡ 0 in Ω . Then by the jump relations in Lemma 4.3
(iv) we have
− +
∂ ∂
ψ(x) = Vγ ψ (x) − Vγ ψ (x) = 0, x ∈ S
∂n ∂n

Now we are in the position to solve (4.27) by a simple layer method. Namely,
setting ve = Vγ (ψ) the exterior Neumann problem (4.27) is transformed into a
Fredholm integral equation of the second kind on S for the unknown layer ψ,

∂vo
ψ(x) + 2 Kγ (x, y)ψ(y)dSy = 2 (x), x ∈ S, (4.48)
S ∂n
which we abbreviate with the notation (4.45) by
∂vo
(I + 2Kγ )ψ = 2 . (4.49)
∂n
As a consequence of Lemma 4.3 and Lemma 4.6 there holds the following result.
Theorem 4.5 If ψ ∈ C 0 (S) is a solution of (4.48) then ve = Vγ (ψ) yields a

(classical) solution of (4.27). For any real r there exists exactly one solution of
(4.48) for given data ∂v
∂n ∈ H (S).
o r
With formula ve = Vγ (ψ) for the exterior pressure we set for the total accoustic
field
v = Vγ (ψ) + v0 in Ω, v = Vγi (χ) in Ω . (4.50)
We obtain from the boundary conditions (4.26) a coupled system of pseudodiffer-

ential equations for the unknown layers (ψ, χ) on S:
Vγi (χ) = Vγ (ψ) + v0 ,

∂vo ρi (4.51)
(I − 2Kγi )χ + ν(I + 2Kγ )ψ = 2ν , ν= ∈ R.
∂n ρ
But by evaluating the kernel function r −1 eiγ r for small r one verifies as above that
both Vγi and Vγ are pseudodifferential operators of order -1. Hence there holds (cf.
Lemma 4.5)
Vγ (ψ) = Vγi (ψ) + W (ψ)
with a pseudodifferential operator W of order −3. Therefore multiplication of

(4.51)1 with the bijective operator Vγ−1
i
yields
χ − ψ = Vγ−1
i
W (ψ) + Vγ−1
i
(v0 ). (4.52)
Since furthermore
Kγ (ψ) = Kγi (ψ) + L(ψ)
with a pseudodifferential operator L of order −2 the second equation in (4.51) gives
∂vo
χ + νψ = 2Kγi (χ − νψ) − 2L(ψ) + 2 ν. (4.53)
∂n
The equation (4.52) and (4.53) form a Riesz-Schauder system on H r (S)×H r (S),
r ∈ R. Each of the operators occuring on the right sides is of order at most −1
and the forcing terms Vγ−1i
(v0 ) and ∂v
∂n belong to H (S) for given v0 ∈ H
o r r+1 (S).
A reversal of the steps shows that if (ψ, χ) satisfy (4.52), (4.53), then they also
satisfy (4.51). But the uniqueness result for (4.24)–(4.26) (Theorem 4.4) shows that
the only solution of the homogeneous equations (4.52), (4.53) vanishes identically.
Hence we have the following existence result for the interface problem (4.24)–
(4.26) governing the scattering of sound (for a corresponding approach to Maxwell’s
interface problem see [287]):
Theorem 4.6 Let v0 ∈ H r+1 (S) for arbitrary r ∈ R. Then the equations (4.52)
and (4.53) have a unique solution with χ, ψ ∈ H r (S).
Now we relax the regularity assumption on the interface Γ , whereas above for the
treatment of the interface problem with pseudodifferential operators we assumed
Γ to be analytic. This allowed to apply Riesz-Schauder theory for the existence
proof of the solution of second kind integral equations (Theorem 4.6); we now only
require Γ to be Lipschitz.
Next following [129], we convert the interface problem (4.24)–(4.26) via the
direct method to an equivalent strongly elliptic system of pseudodifferential equa-
tions on the interface Γ . For simplicity of notation we write (4.24)–(4.26) as
(Δ + kj2 )uj = 0 in Ωj (j = 1, 2)
∂u1 ∂u2
u1 = u2 + v0 , μ = + ψ0 on Γ
∂n ∂n

1 ∂u2 1
u2 (x) = O , − ik2 u2 (x) = O , |x| → ∞,
|x| ∂|x| |x|
where Ω1 a bounded simply connected domain (= Ω ) in R3 , Ω2 = R3 \ Ω1

and given v0 = u0 |Γ , ψ0 = ∂u
∂n |Γ with (Δ + k2 )u0 = 0 in Ω1 . By Theorem 4.1 this
0 2
transmission
problem is equivalent to the following relations for the Cauchy data
vj
of uj :
ψj

v1
(1 − A1 ) = 0, (4.54)
ψ1

v2
(1 + A2 ) = 0, (4.55)
ψ2

v2 v1 v0 10
=M − , with M = (4.56)
ψ2 ψ1 ψ0 0μ

v0
and (1 − A2 ) = 0, (4.57)
ψ0
with Aj as in (4.14). Now from the above system of six equations for four unknows

v v1
we derive a system of two equations for two unknows: Writing :=
ψ ψ1
and inserting (4.56) into (4.55) gives

v v0
(1 + A2 )M = (1 + A2 ) .
ψ ψ0
Then multiplying by M −1 from the left and subtracting (4.54) gives the boundary
integral equation

v 1 −1 v 1 −1 v0
H := (A1 + M A2 M) = M (1 + A2 ) . (4.58)
ψ 2 ψ 2 ψ0

v0
If satisfy (4.57), this simplifies to
ψ0

v v0
H = M −1 .
ψ ψ0

v
Now any solution of (4.58) generates a solution of the original transmis-
ψ
sion problem (see [129] for details).
For the system (4.58) there holds the following Gårding inequality: There exists
; ; 1
a compact operator C : H 2 (Γ ) H − 2 (Γ ) → H − 2 (Γ ) H 2 (Γ ) and a constant
1 1 1
γ > 0 such that for μ = −1 and smooth Γ there holds

5 6

(H + C) v ,
v ≥ γ v21 + φ2 1 , (4.59)
φ φ 2 −2
Γ
for all v ∈ H 2 (Γ ), φ ∈ H − 2 (Γ ). Furthermore for %(1 +

1 1 1
μ) > 0 and
%(1 + μ) > 0 there holds also for a polygon Γ in R2
5 6
v v
% (H + C) , ≥ γ v21 + φ2− 1 .
φ φ Γ 2 2
Here the operator

( )
− 12 (K1 + K2 ) 1
2 (V 1 + μV2 )
H =

1
2 W1 + μ1 W2 1
2 K 1 + K2

0 −1
is elliptic in the Agmon-Douglas-Nirenberg sense with order and
1 0
principal symbol
( )
0 1
2 (1 + μ) |ξ1|
σ (H )(ξ ) = .
1
2 1+ 1
μ |ξ | 0
4.3 Screen Problems 81
For μ = −1 H is strongly elliptic yielding the Gårding inequality (4.59). Now

for smooth Γ system (4.58) of boundary integral equations is an elliptic system of
pseudodifferential equations.
The standard
regularity theory for pseudodifferential operators shows that
v0 ; s−1 v
for given ∈ H s (Γ ) H (Γ ), any solution of (4.58) is in
ψ0 ψ
;
H s (Γ ) H s−1(Γ ) for any s ∈ R. Under the assumption of Theorem 4.4 the
solution of the transmission problem is unique. This implies that the operator H
is injective. Now, by Gårding’s inequality the operator H is Fredholm of index
zero, and hence bijective yielding the existence of the solution of (4.58) (see [129]).
Therefore insertion of that solution in the representation formula gives the solution
of the original interface problem for the Helmholtz equation.
4.3 Screen Problems
For open boundary curves or surfaces S the correct setting of integral equations
needs a refined analysis where the solutions of the integral equations must (in a
weak sense) be extendable by zero from the open surface S to a closed surface S̃
(including S), i.e. for real s [253]
H̃ s (S) = {ϕ : ϕ ∗ = ϕ on S, ϕ ∗ = 0 on S̃\S, ϕ ∗ ∈ H s (S̃)}.
1/2
Note H̃ 1/2(S) = H00 (S) in [284].
For given g(h) we consider the Dirichlet (Neumann) screen problem (k ∈ C \
{0}):
( + k 2 )u = 0 in ΩS := R3 \S
u=g on S
∂u
( = h on S)
∂n
∂u 1
− iku = o( ) as r = |x| → ∞
∂r r
where S is a bounded, simply connected, orientable,open surface in R3 with a

smooth boundary curve γ which does not intersect itself. Extend S to an arbitrary
smooth, simply connected, closed, orientable manifold ∂G1 enclosing a bounded
domain G1 (see Fig. 4.2).
∂
Let ∂n denote the exterior normal derivative to ∂G1 . Let [v] denote the jump
v− − v+ where the subscript +(−) means the limit from R3 \G1 (from G1 ) to ∂G1 .
Furthermore, let B denote a sufficiently large ball with radius R including G1 and
let G2 := B ∩ (R3 \G1 ) and ∂B denote the boundary of B.

[398]
S
∂G1
∂G2
1 eik|x−y|
Let ϕ(x, y) = 4π |x−y| and

VGj u(x) := −2 ϕ(x, y)u(y)dsy (x ∈ Gj ),
Γ

∂
KGj u(x) := −2 ϕ(x, y)u(y)dsy (x ∈ Gj , Γ = ∂Gj )
Γ ∂ny
Application of the representation formula
∂u
u(x) = (−1)j 1/2(KGj u(x) − VGj (x))
∂n
gives for x ∈ G1

1 ∂u
u(x) = − KG1 u(x) − VG1 (x) (4.60)
2 ∂n

1 ∂u
0=− KG2 u(x) − VG2 (x)
2 ∂n
∂n ]|∂Gj \S = 0, addition yields with the outer boundary ∂B = {y ∈

Since [ ∂u
R3 , |y| = R}

∂ ∂u ∂u
u(x) = u(y) ϕ(x, y)dsy − (y)ϕ(x, y)dsy − [ ](y)ϕ(x, y)dsy
|y|=R ∂ny |y|=R ∂n S ∂n
For x → S the trace theorem yields with u|S = g

∂ ∂u ∂u
g(x) = {u(y) ϕ(x, y) − (y)ϕ(x, y)}dsy − [ ](y)ϕ(x, y)dsy
|y|=R ∂n y ∂n S ∂n
4.3 Screen Problems 83
Since the radiation condition holds for u and ϕ, the integral over |y| = R vanishes
as R → ∞ and therefore the foregoing expression becomes

∂u ∂u
2g(x) = −2 [ ]ϕ(x, y)dsy =: VS [ ](x) , x ∈ S (4.61)
S ∂n ∂n
Taking in (4.60) the normal derivative gives for x ∈ G1 (note [u]|∂Gj \S = 0 )

∂u ∂ ∂ ∂u ∂
(x) = {u(y) ϕ(x, y) − (y) ϕ(x, y)}dsy
∂n |y|=R ∂nx ∂ny ∂n ∂nx

∂2
− [u](y) ϕ(x, y)dsy
S ∂nx ∂ny
Since ϕ and its derivatives satisfy the decay condition, letting x → S , R → ∞

∂n |S = h gives
and taking ∂u

∂2
− 2h(x) = 2 [u](y) ϕ(x, y)dsy =: WS [u](x) , x ∈ S (4.62)
S ∂nx ∂ny
In [398] it is shown that (4.61) is equivalent to the Dirichlet screen problem and
(4.62) to the Neumann screen problem and that for (k ≥ 0 these integral equations
are uniquely solvable with [ ∂u =−1/2(S) for given g ∈ H 1/2(S) and [u]|S ∈
∂n ]|S ∈ H
=
H (S) for given h ∈ H
1/2 −1/2 (S), respectively (for Lipschitz screens see [124]).
Now we come to the singularity of the densities of the integral equations (4.61)
and (4.62) near the edge γ of the screen S. The analysis in [398] follows the
procedure in [167] by (i) mapping locally S onto R2+ ,(ii) applying the Wiener-Hopf
technique in the halfspace R2+ and (iii) patching together the local results.
Theorem 4.7 (Theorem 2.9 in [398])
(i) Let g ∈ H 3/2+σ (S) be given. Then the solution of the integral equation (4.61)
has the form
∂u
[ ] = β(s)ρ −1/2 χ(ρ) + ψr on S (4.63)
∂n

with β ∈ H 1/2+σ (γ ) , ψr ∈ H̃ 1/2+σ (S), 0 < σ < σ < 1/2
(ii) Let h ∈ H 1/2+σ (S) be given. Then the solution of the integral equation (4.62)
has the form
[u] = α(s)ρ 1/2 χ(ρ) + vr on S (4.64)

with α ∈ H 1/2+σ (γ ), vr ∈ L2 (I ; H 1/2+σ (γ )) ∩ H̃ 3/2+σ (I ; L2 (γ )), 0 < σ <
σ < 1/2, where S is identified with I × γ , I = [0, 1].
(Here s denotes the parameter of arclength of γ , ρ corresponds to the Euclidean

distance to γ , χ is a C ∞ cut-off function with χ ≡ 1 for |ρ| < 1/2 and χ ≡ 0 for
|ρ| > 1).
This result on the singularity of the screen problem provides the basis for the
augmented BEM described in section 7.5. Crack problems can be dealt with like
the screen problems above yielding an efficient solution procedure with boundary
integral equations and boundary elements (see [130, 432]). Recently there has been
intensive research on multiple screens (see the work of Claeys and Hiptmair [107,
108]).
4.4 Interface Problem in Linear Elasticity
Next, we want to relax the smoothness assumptions on the interface Γ and only
require Γ ∈ Lip. We will show that the above approach still works and derive a
Gårding inequality for the boundary integral operators related to linear elasticity
problems. The reported results are taken from the paper [137] by Costabel and
Stephan. Here we like to mention the celebrated fundamental book on three-
dimensional potential theory of linearized elasticity [278].
The transmission problem in 3D in steady state elastodynamics reads (TMP):
For given vector fields u0 and t0 on the boundary Γ find vector fields uj in Ωj ,
j = 1, 2, satisfying the equations of linear elasticity
Pj uj − ρj ω2 uj = 0 in Ωj , j = 1, 2
and the transmission conditions
u1 = u2 + u0 , t1 = t2 + t0 on Γ
Here the differential operators Pj are given by
Pj u = −(μj u + (λj + μj )grad div u)
ρj > 0 is the density of the medium Ωj , and ω > 0 is the frequency of the incident
wave. We are interested in solutions uj ∈ Hloc
1 (Ω ) and define
j
L1 = {u1 ∈ H 1 (Ω1 ) : P1 u1 = ρ1 ω2 u1 in Ω1 }
1 (Ω ) : P u = ρ ω2 u in Ω , u satisfies a decay condition [137]}

L2 = {u2 ∈ Hloc 2 2 2 2 2 2 2
Here and in the following all function spaces, including all Sobolev spaces are
considered vectorial containing 3D vector fields.
4.4 Interface Problem in Linear Elasticity 85
Lemma 4.7 Let u ∈ Hloc 1 (Ω ) with compact support satisfy P u ∈ L2 (Ω ) and

j j loc j
let v ∈ H 1 (Ωj ) with bounded support. Then Tj uΓ ∈ H −1/2(Γ ) is defined with
f, g := Γ f · gds by

Pj u · vdx = (−1)j Tj u, v + Φj (u, v) (4.65)
Ωj
with

3
j j
Φj (u, v) = aihkl kl (u) ih (v)dx , aihkl = λj δih δkl + μj (δik δhl + δil δhk ) ,
Ωj i,h,k,l=1
where λj and μj denote the Lame constants in Ωj (j = 1, 2).

From (4.65) one obtains, with the symmetry of Φj , the second Green formula

(Pj u · v − u · Pj v)dx = (−1) j
(v · Tj (u) − u · Tj (v))ds (4.66)
Ωj Γ
This gives in Ωj with the fundamental solution Gj (x, y, ω) of (Pj − ρω2 )uj = 0,
the Somigliana representation formula for x ∈ Ωj :

3 4
uj (x) = (−1)j Tj (x, y, ω)vj (y) − Gj (x, y, ω)φj (y) ds(y) (4.67)
Γ
where vj = uj , φj = Tj (uj ) = tj on Γ . Here Gj is the 3 × 3 matrix function

"
−1 1 ikjT r T L
(Gj )ik = e δik + (kjT )−2 ∂i ∂k [(eikj r − eikj r )r −1 ]
4πμj r
with r = |x − y| and Tj (x, y, ω) = Tj,y (Gj (x, y, ω))T , kjL longitudinal (dilational)
wave number , kjT transverse (shear) wave number.
Lemma 4.8 Let uj ∈ Lj . Then (4.67) holds for uj in Ωj . For any vj ∈ H 1/2(Γ )
and any φ ∈ H −1/2 (Γ ) the formula (4.67) defines a vector field uj ∈ Lj .
Taking Cauchy data in (4.67) yields on Γ

vj vj
= Cj
φj φj
where the Calderón projector

( )
1/2 + (−1)j Λj −(−1)j Vj
Cj =
−(−1)j Wj 1/2 − (−1)j Λj
is defined via the boundary integral operators

Vj v(x) = Gj (x, y, ω)v(y)ds(y) Λj v(x) = Tj (x, y, ω)v(y)ds(y)
Γ Γ

Wj v(x) = −Tj,x Tj (x, y, ω)v(y)ds(y) Λj v(x) = Tj (y, x, ω)T v(y)ds(y)
Γ Γ
Lemma 4.9
(a) The statements (i) and (ii) on (v, ψ) ∈ H := H 1/2(Γ ) × H −1/2(Γ ) are
equivalent:
(i) (v, ψ) are Cauchy
data of some uj ∈ Lj
(ii) (I − Cj ) ψv = 0
(b) The operators Cj are projection operators mapping H on its subspace of
Cauchy data of weak solutions in Lj
Thus we can write the transmission problem (TMP) in the equivalent form

v1
(I − C1 ) =0 (4.68)
φ1

v2
(I − C2 ) =0 (4.69)
φ2

v2 v1 v0
= − (4.70)
φ2 φ1 φ0
This is a system of 6 vector equations

for 4 vector unknowns. From it we can extract
a square subsystem by inserting φv22 from (4.70) in (4.69) and subtracting (4.68)
from the resulting equation. We obtain

v1 v0
A = (I − C2 ) with A := C1 − C2 (4.71)
φ1 φ0
We have the following theorem.

Theorem 4.8 Let φv00 ∈ H = H 1/2(Γ )×H −1/2 (Γ ) be given. Then there holds
u1 |Γ
(i) If uj ∈ Lj solve the (TMP), then φv := φv11 = T (u ∈ H solves (4.71)
v 1 )|Γ
(ii) If φ ∈ H solves (4.71), then, with

v1 v v2 v v0
:= C1 and := C2 −
φ1 φ φ2 φ φ0
and uj defined by (4.67), uj ∈ Lj (j = 1, 2) solves the (TMP).

4.4 Interface Problem in Linear Elasticity 87
Proof
(i) follows from the derivation
of (4.71).
(ii) From the definition of φvj and the projection property of Cj follows
j

vj
(I − Cj ) = 0,
φj

hence φvj are Cauchy data of certain uj ∈ Lj which are then given by (4.67).
j
It remains to show that the transmission condition is satisfied:

v2 v1 v v0 v v0
− = (C2 − C1 ) − C2 = −A − C2
φ2 φ1 φ φ0 φ φ0

v0 v0 v0
= (I − C2 ) − C2 =−
φ0 φ0 φ0

Theorem 4.9 The operator A satisfies a Gårding inequality: There exist γ > 0 and
a compact operator T : H → H with
5 6
v v v
% (A + T ) , ≥ γ (v2H 1/2 (Γ ) + φ2H −1/2 (Γ ) ) ∀ ∈H
φ φ φ
(4.72)
Here the brackets denote the natural (anti)-duality of H with itself:

5 6
v w v w
, := (vψ + wφ)ds for , ∈H.
φ ψ Γ φ ψ
Proof We write
( )
−Λj Vj
A = A1 + A2 with Aj = (−1) 1/2 I − Cj =
j
Wj Λj
Since the sum of two strongly elliptic operators is strongly elliptic, it suffices to
show the strong ellipticity of the operators A1 and A2 . Proof for A1 : Due to density
arguments, one needs to show the Gåding inequality (4.72) only for smooth (v, φ).
Let then uj , j = 1, 2, be defined by

uj (x) = χ(x) {T1 (x, y)v(y) − G1 (x, y)φ(y)}ds(y), x ∈ Ωj .
Γ
Here we choose χ ∈ C0∞ (R3 ) satisfying χ ≡ 1 is a neighborhhood of Ω 1 . Then, by

definition of the Calderón projectors (i.e. the classical jump relations for the elastic
potentials), the Cauchy data vj := uj |Γ and φj := T (uj ) satisfy

vj v
= (−1)j 1/2 I − (−1)j A1 .
φj φ
By adding and subtracting these two equations, we find

v v1 v2
A1 =− −
φ φ1 φ2

v v1 v2
=− +
φ φ1 φ2
Thus the bilinear form defined by A1 is given by

5 6 5 6
v v v1 v2 v1 v2
2 A1 , = + , −
φ φ φ1 φ2 φ1 φ2
5 6 5 6
v1 v1 v2 v2
= , − ,
φ1 φ1 φ2 φ2
5 6 5 6
v2 v1 v1 v2
+ , − ,
φ2 φ1 φ1 φ2

= 2% (v 1 φ1 − v 2 φ2 )ds + 2i( (v 2 φ1 − v 1 φ2 )ds
Γ Γ
Hence
5 6
v v
% A1 , = % (v 1 φ1 − v 2 φ2 )ds (4.73)
φ φ Γ
Now we need the first Green formulas for P1 in Ω1 and Ω2 . This leads to

Φ̃j (uj , uj ) − uj · (P1 − ρ1 ω2 )uj dx = −(−1)j v j φj ds (4.74)
Ωj Γ
where

1 2 2
Φ̃j (uj , uj ) := aihkl kl (uj ) ih (uj ) − ρ1 ω|uj | dx.
Ωj
Now P1 u1 − ρ12 ωu1 = 0 and P1 u2 − ρ12 ωu2 = f2 , where f2 ∈ C0∞ (Ω2 ). (f2 ≡ 0
whenever χ ≡ 1 or χ ≡ 0 holds.) From (4.73) and (4.74) together we find
5 6
v v
% A1 , = %{Φ̃1 (u1 , u1 ) + Φ̃2 (u2 , u2 ) − u2 · f2 dx} (4.75)
φ φ Ω2
4.5 Exterior Maxwell’s Equations 89
As the support of f2 is disjoint from Γ , there is a compact operator T1 on H 1/2(Γ )×

H −1/2(Γ ) such that
5 6
v v

u2 · f2 dx ≤ T1 ,
φ φ
Ω2
From Korn’s inequality and the trace lemma we find that there exist compact
quadratic forms kj on H 1 (Ωj ) and hence a compact operator T2 on H =
H 1/2(Γ ) × H −1/2 (Γ ) such that
Φ̃1 (u1 , u1 ) + Φ̃2 (u2 , u2 ) ≥ γ1 u1 2H 1 (Ω ) + u2 2H 1 (Ω − k1 (u1 ) − k2 (u2 )
1 1)
5 6
v v
≥ γ2 v2H 1/2 (Γ ) + φ2H −1/2 (Γ ) − T2 ,
φ φ
Finally we get
5 6
v v
% (A1 + T1 + T2 ) , ≥ γ2 v2H 1/2 (Γ ) + φ2H −1/2 (Γ )
φ φ

4.5 A Strongly Elliptic System for Exterior Maxwell’s

Equations
This section reports of an approach by [286] and [287]. In [285] a simple layer
potential method for the three-dimensional eddy current problem is introduced. In
[286] solution procedures for the perfect conductor problem are given. E.g. different
sets of Maxwell equations are solved in the obstacle and outside while the tangential
components of both electric and magnetic fields are continuous across the obstacle
surface. In [287] it is shown, that the integral equation system resulting from the
three-dimensional Maxwell’s equations in air in the exterior of a perfect conductor
is coercive and thus asymptotic convergence of Galerkin’s method is established.
The purpose of this section is to show the coercivity of the system of equations
belonging to the three-dimensional conductivity problem in an exterior unbounded
domain using pseudodifferential operators.
4.5.1 A Simple Layer Procedure
We consider the eddy current problem: Let Ω be a bounded interior and Ω =

R3 \ Ω . Ω is to represent a perfect conductor characterized by constants ε, μ and
σ = ∞ denoting permitivity, permeabilty and conductivity. Ω is to represent air

characterized by ε0 , μ0 and σ0 = 0. S = ∂Ω = ∂Ω is a closed analytic surface
dividing R3 into the disjoint domains Ω and Ω . The total electromagnetic field
(E, H) consists of the sum of the incident (E0 , H0 ) and the scattered (ES , HS ) field.
Thereby, (E0 , H0 ) is assumed to originate in Ω.
The time harmonic Maxwell’s equations are given by
curl E = iωμ0 H , curl H = −iωε0 E in Ω (4.76)

curl E = iωμ H , curl H = (−iωε + σ )E in Ω . (4.77)
By appropriate rescaling of (4.76) and (4.77) one obtains
curl E =H , curl H = α 2 E in Ω (4.78)

curl E =H , curl H = iβ E in Ω , (4.79)
with α 2 = ω2 ε0 μ0 and β = (ωμσ − iω2 με) and β = ωμσ > 0, if ε = 0. Across

S the tangential components of the fields
(n × E)+ = (n × E)− , (n × H)+ = (n × H)− (4.80)
must be continuous. At higher conductivity β is large, which leads to the perfect

conductor approximations. This means solving only (4.78) and requiring that the
tangential component of the total electric field n × E = 0 vanishes on S, leading to
curl ES = HS , curl HS = α 2 ES in Ω
(4.81)
(n × ES ) = −(n × E0 ) on S.
In [285] it is shown that for (4.78)–(4.80) at most one solution exists for any α > 0
and 0 < β ≤ ∞.
By introducing the simple layer potential

1 eiα|x−y|
Vα (M)(x) = M(y) dSy (4.82)
4π S |x − y|
one can display the electric and magnetic fields in the Stratton-Chu representation
formulas [413]
,
E = Vα (n × H) − curl Vα (n × E) + grad Vα (n · E)
in Ω. (4.83)
H = curl Vα (n × H) − curl curl Vα (n × E)
Now setting n × E = 0 in (4.83) and replacing n × H and n · E by unknowns J and

M, yields
E = Vα (J) + gradT Vα (M) , H = curl Vα (J). (4.84)
Now div H is automatically zero, whereas we must guarantee that div E = 0 in Ω.

iα|x−y|
It suffices to require div E = 0 on S. It follows from (4.84) and the kernel e|x−y|
that E = −α 2 E. Hence div E = −α 2 div E in Ω. Moreover div E satisfies the
radiation condition. Hence, by uniqueness for the scalar exterior Dirichlet problem
div E = 0 on S implies div E = 0 in Ω. Hence we require div E = 0 on
S. Therefore applying the boundary condition of (4.81) and div E = 0 on S in
(4.84) one obtains a coupled system of pseudodifferential equations on the boundary
surface S with the unknowns J and M:
Vα (J)T + grad Vα (M) = −(n × E0 ) = −E0T

(4.85)
Vα (divT J) − α 2 Vα (M) = 0,
where Vα (J)T denotes the tangential component of the vector function Vα (J) and
div Vα (J) = Vα (divT J).
4.5.2 Modified Boundary Integral Equations
Furthermore in [285] it is shown that there exists a continous map Jα (J)T from
Hr (S) into H r+1 (S), r ∈ R, such that
divT Vα (J) = Vα (divT J) + Jα (J)T . (4.86)
Therefore by applying divT onto (4.85)1 and subtracting the result from (4.85)2 one
gets a new equivalent system:
Vα (J)T + gradT Vα (M) = −E0T

(4.87)
−Jα (J)T − (T + α 2 )Vα (M) = divT E0T .
In [287] it is mentioned, that (4.85) is not satisfying the Gårding’s inequality but
(4.87) does, so convergence for Galerkin’s procedure is guaranteed.
In order to show the claimed Gårding inequality for the system (4.87), we
consider the half-space case as in [287]. The equation system (4.87) becomes in
the half-space case Ω = {x ∈ R3 |x3 > 0}:
∂ ∂
φα ∗ J + φα ∗ Me1 + φα ∗ Me2 = −4π(e3 × E0 )
∂x1 ∂x2
( ) (4.88)
∂2 ∂2
− + + α 2 φα ∗ M = 4π div E0T .
∂x12 ∂x22
Here
eiα|x−y|
φα (|x − y|) = (4.89)
|x − y|
is the fundamental solution of the Helmholtz equation. In [285] it is shown that the
series expansion
∞
δj
1
φα (r) = + iα + rj , δ ∈ C (4.90)
r j!
j =1
with r = |x − y| leads to the existence of a smoothing pseudodifferential operator

Wα of the order −3 such that
Vα = V0 (M) + Wα (M). (4.91)
The system (4.88) can be written as a 3 × 3-matrix of operators

⎛ ⎞ ⎛ 1⎞ ⎛ 0 ⎞
Vα |1 0 grad1 Vα J −ET |1
Aα U := ⎝ 0 Vα |2 grad2 Vα ⎠ ⎝J 2 ⎠ = ⎝−E0T |2 ⎠ := F (4.92)
0 0 −( + α 2 )Vα M div E0T
with J = J 1 e1 + J 2 e2 . One can show that the difference between Aα and

⎛ ⎞
V0 |1 0 grad1 V0
A0 = ⎝ 0 V0 |2 grad2 V0 ⎠ (4.93)
0 0 − V0
is compact. The principle symbol σ (Aα )(ξ ) is obtained by the two-dimensional

Fourier transformation F̃ : (x1 , x2 ) → (ξ1 , ξ2 ) of A0 . In [287] it is shown that
1
φ̂α (ξ ) = (F̃ φα )(ξ ) = (|ξ |2 − α 2 )− 2 . (4.94)
Therefore the principle symbol can be displayed as

⎛ 1
⎞
0 iξ1 |ξ1|
|ξ |
⎜ ⎟
σ (Aα )(ξ ) = ⎝ 0 |ξ1| iξ2 |ξ1| ⎠ (4.95)
0 0 |ξ |
with |ξ |2 = ξ12 + ξ22 . Finally it can be proven that there exist constants γ > 0 and
κ > 4 such that
⎛ ⎞ ⎛ ⎞
10 0 ζ̄1
%(ζ1 , ζ2 , ζ3 ) ⎝0 1 0⎠ σ (Aα )(ξ ) ⎝ζ̄2 ⎠ ≥ γ (ζ1 ζ̄1 + ζ2 ζ̄2 + ζ3 ζ̄3 ) (4.96)
00 κ ζ̄3
for all ζ ∈ C3 and all ξ ∈ R3 with |ξ | = 1. From that it follows that Aα is strongly
elliptic (see Definition B.7 and [259]) and hence satisfies a Gårding inequality.
The Galerkin procedure for the modified system (4.87) is analyzed in [286]. For
a different approach see [41, 42].
In [135] a boundary integral equation method for transmission problems for
strongly elliptic differential operators is analysed, which yields a strongly elliptic
system of pseudodifferential operators and which therefore can be used for numer-
ical computations with Galerkin’s procedure. The method is shown to work for
the vector Helmholtz equation with electromagnetic transmission conditions. The
system of boundary values is slightly modified so that the corresponding bilinear
form becomes coercive over H 1 . The concept of the principal symbol of a system
of pseudodifferential operators is used to derive existence and regularity results for
the solution.
Chapter 5
The Signorini Problem and More
Nonsmooth BVPs and Their Boundary
Integral Formulation
In this chapter we deal with unilateral and nonsmooth boundary value problems,
in particular Signorini problems without and with Tresca friction and nonmontone
contact problems from adhesion/delamination in the range of linear elasticity. We
show how the boundary integral techniques developed in the previous chapters can
be used to transform those problems to boundary variational inequalities. This opens
the way to the numerical treatment of these nonlinear problems by the BEM as
detailed in Chap. 11.
5.1 The Signorini Problem in Its Simplest Form
In this section we follow [214] and introduce the Signorini boundary value problem
in its simplest form taking the Laplace equation as elliptic equation. The Signorini
problem is a unilateral boundary value problem, where the unilateral constraint
lives on the boundary. Since the domain is governed by a linear pde with constant
coefficients, a fundamental solution is available and integral equation methods
apply. Here modifying the approach of H. Han [227] we derive an equivalent
boundary variational inequality in the Cauchy data as unknows, where the associated
bilinear form is shown to satisfy a Gårding inequality in appropriate Sobolev spaces
on the boundary. Finally we turn to the convex cone of feasible solutions and provide
a density result that is useful for the convergence analysis of the boundary element
method to follow in Sect. 11.1.
Let Ω ⊂ R2 be a bounded plane domain with the Lipschitz boundary Γ [327].
Then n, the outward normal to Γ , exists almost everywhere and n ∈ [L∞ (Γ )]2
(see [327, Lemma 2.4.2]). Here we consider the simple elliptic equation
− Δu = 0 in Ω . (5.1)

96 5 Signorini Problem, More Nonsmooth BVPs
Thus we have the Cauchy data u and ϕ := ∂u ∂n on Γ .

To formulate the boundary conditions, let Γ = Γ D ∪ Γ N ∪ Γ S , where the
open parts ΓD , ΓN , and ΓS are mutually disjoint. We prescribe
u = 0 on ΓD , (5.2)
ϕ=g on ΓN . (5.3)
On the remaining part ΓS , Signorini boundary conditions are imposed, i.e.
u≤0, ϕ ≤h, u(ϕ − h) = 0 , (5.4)
where g ∈ H −1/2(ΓN ) and h ∈ H −1/2 (ΓS ) are given. We point out that a priori
it is not known where u = 0 changes to ϕ = h and the boundary part ΓS is
only taken large enough to contain this free boundary. Thus to make this free
bounday problem meaningful we assume meas (ΓS ) > 0 , but we do not require
meas(ΓD ) > 0 . Note there is no loss of generality to assume homogeneous
conditions above. Indeed, more general conditions can be reduced to the form given
above by a superposition argument that uses the solution of the linear boundary
value problem
−Δu = f in Ω
u = u0D on ΓD , ϕ=0 on ΓN , u = u0S on ΓS
and an appropriately redefined right hand side h in (5.4). To give the variational
formulation of the boundary value problem (5.1)–(5.4) we introduce the bilinear
form
2
∂v ∂w
β(v, w) := grad v · grad w dx = dx
Ω ∂xk ∂xk
k=1 Ω
and the linear form

(v) := g v ds + h v ds
ΓN ΓS
on the function space
HΓ1D ,0 (Ω) := {v ∈ H 1 (Ω) : v = 0 on ΓD } (5.5)

5.1 The Signorini Problem in Its Simplest Form 97
and the convex cone
K := {v ∈ HΓ1D ,0 (Ω) : v ≤ 0 on ΓS } . (5.6)
Then the variational formulation of (5.1)–(5.4) in the domain Ω is easily obtained

by Green’s formula (see e.g. [266, 267] for more details) as the following variational
inequality:
(P) Find u ∈ K such that
β(u, v − u) ≥ (v − u) ∀v ∈ K .
To derive a boundary integral formulation let a fundamental solution of (5.1) by

given by
1
F (x, y) := ln |x − y| .
2π
Now let u ∈ K be a solution of (P), hence −Δu = 0. According to the

representation formula (Sect. 2.1, Theorem 2.1) we have

∂F (x, y)
u(x) = u(y) dsy − F (x, y) ϕ(y) dsy ∀x ∈ Ω . (5.7)
∂ny
Γ Γ
By the jump relations, respectively continuity properties of the simple layer poten-
tial, respectively of the double layer potential (see Sects. 2.2.2, 2.4.1, 4.1), (5.7)
implies

1 ∂F (x, y)
u(x) = u(y) dsy − F (x, y) ϕ(y) dsy ∀x ∈ Γ , (5.8)
2 ∂ny
Γ Γ

1 ∂ 2 F (x, y) ∂F (x, y)
ϕ(x) = u(y) dsy − ϕ(y) dsy ∀x ∈ Γ . (5.9)
2 ∂nx ∂ny ∂nx
Γ Γ
Here, the first integral in (5.9) is a hypersingular integral (partie finie following
Hadamard); by partial integration twice using the Cauchy–Riemann equations (see
Sect. 2.6 , Proposition 2.1) one obtains

∂ 2 F (x, y) d du(y)
u(y) dsy = F (x, y) dsy .
∂nx ∂ny dsx dsy
Γ Γ
Testing (5.8) by ψ ∈ H −1/2(Γ ) leads to

1 ∂F (x, y)
− u(x)ψ(x) dsx + u(y) ψ(x) dsy dsx
2 ∂ny
Γ ΓΓ

− F (x, y) ϕ(y) ψ(x) dsy dsx = 0 ,
ΓΓ
or shortly,
− b(ψ, u) + a(ψ, ϕ) = 0 , ∀ψ ∈ H (−1/2)(Γ ) (5.10)
where

a(ψ, ϕ) := − F (x, y) ϕ(y) ψ(x) dsy dsx
ΓΓ

1 ∂F (x, y)
b(ψ, u) := u(x) ψ(x) dsx − u(y) ψ(x) dsy dsx .
2 ∂ny
Γ ΓΓ
On the other hand, for any v ∈ HΓ1D ,0 (Ω), by Green’s formula,

β(u, v) = ∇u · ∇u dv = ϕv ds . (5.11)
Γ Γ
Testing (5.9) by v and plugging in (5.11) yields

du(y) dv(x)
β(u, v) = − F (x, y) dsy dsx
Γ Γ dsy dsx

∂F (x, y) 1
− ϕ(y) v(x) dsy dsx + ϕ · v ds
Γ Γ ∂nx 2 Γ
du dv
=: a , + b(ϕ, v) . (5.12)
ds ds
By (5.10) and (5.12), introducing the convex cone
K := {v ∈ H 1/2(Γ ) : v = 0 on ΓD , v ≤ 0 on ΓS }
we arrive at the following variational problem: Find [u, ϕ] ∈ K × H −1/2(Γ ) such

that
⎧
⎨ a du , dv − du + b(ϕ, v − u) ≥ l(v − u) ∀v ∈ K ,
(π) ds ds ds
⎩
−b(ψ, u) + a(ψ, ϕ) = 0 ∀ψ ∈ H −1/2(Γ ) .
This problem (π) is equivalent to the former variational problem (P), since
conversely, for any solution [u, ϕ] to (π), we can define u in Ω by means of (5.7),
and for any v ∈ HΓ1D ,0 (Ω) we can consider its trace v|Γ to obtain β(u, v − u) ≥
(v − u). Note that (π) is equivalent to the single boundary variational inequality:
Find [u, ϕ] ∈ K × H −1/2 (Γ ) such that for all [v, ψ] ∈ K × H −1/2(Γ ),
A([u, ϕ] , [v, ψ] − [u, ϕ]) ≥ (v − u) , (5.13)
where the bilinear form A is given by
du dv
A([u, ϕ], [v, ψ]) := a , + a(ψ, ϕ) + b(ϕ, v) − b(ψ, u) .
ds ds
Indeed, since the variational equality in (π) is equivalent to the variational inequality
a(ψ − ϕ, ϕ) − b(ψ − ϕ, u) ≥ 0
on the space H −1/2(Γ ) , the implication (π) ⇒ (5.13) is immediate. On the other
hand, (π) follows from (5.13) by the choices ψ = 0, v = u .
Remark A is not symmetric (although a, β are symmetric), hence the problem (π)
is not equivalent to a minimization problem on K. A is positive semidefinite; indeed

du du
A([u, ϕ], [u, ϕ]) = a , + a(ϕ, ϕ) ≥ 0 .
ds ds
Now our aim is to establish a Gårding inequality for the bilinear form A(·, ·)
in the space H 1/2 (Γ ) × H −1/2(Γ ), i.e. positive definiteness up to a compact
perturbation term. The boundary integral operators that give rise to the bilinear
form A(·, ·) can be understood as pseudodifferential operators. Since coordinate
transformations do not affect their principal symbol, thus contribute only to compact
perturbation terms (see e.g. [256] for more detailed arguments of this kind) we need
only consider the case of a smooth domain in the subsequent reasoning.
Lemma 5.1 There exist a constant c0 > 0 and a compact operator C0 : H 1/2(Γ ) →
0 dv 02
0 0
0 0 ≥ c0 v21/2,Γ − C0 v, v H −1/2 ×H 1/2 , ∀v ∈ H 1/2 (Γ ) . (5.14)
ds −1/2,Γ
Proof Let θ = 2πs/L , where L is the boundary length, and we can assume without
loss of generality that Γ is the unit circle. Then we can argue similar to [227] with
the only difference that due to the nontrivial kernel of β an extra term enters. More
detailed using the Fourier expansion for a smooth function v – what by density
suffices to consider –
∞
a0
v= + (an cos nθ + bn sin nθ ) ,
2
n=1
∞

dv
= (nbn cos nθ − nan sin nθ ) ,
dθ
n=1
one finds
∞
a02
v21/2,Γ = + (1 + n2 )1/2 (an2 + bn2 ) ,
2
n=1
0 dv 02 ∞

0 0
0 0 = (1 + n2 )−1/2 n2 (an2 + bn2 )
dθ −1/2,Γ
n=1
∞
1
≥ (1 + n2 )1/2 (an2 + bn2 ) ,
2
n=1
- 1 2π .2
a02 = v(θ )dθ ≤ c v20,Γ (c > 0) .
2π 0
Hence
0 dv 02
0 0 1 2 c 2
0 0 ≥ v1/2,Γ − v0,Γ . (5.15)
dθ −1/2,Γ 2 4
Since
H 1/2(Γ ) ⊂ H 0 (Γ ) ≡ L2 (Γ ) ⊂ H −1/2(Γ )
forms a Gelfand triple with compact and dense embeddings, the last term in (5.15)
can be replaced by C0 v, v with C0 : H 1/2(Γ ) → H −1/2 (Γ ) compact concluding
the proof.

Lemma 5.2 The bilinear form A(·, ·) is bounded in [H 1/2(Γ ) × H −1/2(Γ )]2 ;
moreover satisfies a Gårding inequality, i.e. there exist a positive constant c and
a compact operator C : H 1/2 (Γ ) × H −1/2 (Γ ) → H −1/2(Γ ) × H 1/2(Γ ) such that
A([v, ψ], [v, ψ]) + C[v, ψ], [v, ψ] [H −1/2 (Γ )×H 1/2 (Γ )]×[H 1/2 (Γ )×H −1/2 (Γ )]
≥ c[v, ψ]H2 1/2 (Γ )×H −1/2 (Γ ) := c{v2H 1/2 (Γ ) + ψ2H −1/2 (Γ ) }
∀[v, ψ] ∈ H 1/2 (Γ ) × H −1/2 (Γ ) . (5.16)
Proof We have
dv dv
A([v, ψ], [v, ψ]) = a , + a(ψ, ψ) .
ds ds
By Theorem 2.5, [114, Theorem 1]
|a(ψ, ψ)| ≤ const ψ2H −1/2 (Γ ) .
Since for any v ∈ H 1/2(Γ ) , dv

ds = ∂v
i ∂xi ẋi ∈ H −1/2 (Γ ) , it follows
0 dv dv 0
0 0
0a , 0 ≤ const v2H 1/2 (Γ ) .
ds ds
Therefore it remains to prove (5.16). By Theorem 3.6, [114, Theorem 2] the bilinear
form a(·, ·) satisfies a Gårding inequality on [H −1/2(Γ )]2 in the general case of a
Lipschitz domain, i.e.
a(ψ, ψ) ≥ ca ψ2H −1/2 (Γ ) − CA ψ, ψ H 1/2 (Γ )×H −1/2 (Γ ) ∀ψ ∈ H −1/2 (Γ ) ,

(5.17)
where ca > 0 , CA : H −1/2(Γ ) → H 1/2(Γ ) is compact. Hence

0 dv 02
dv dv 0 0 dv dv
a , ≥ ca 0 0 −1/2 − CA , ∀v ∈ H 1/2(Γ ) . (5.18)
ds ds ds H (Γ ) ds ds
Combining (5.17) and (5.18) with Lemma 5.1 yields (5.16).

Finally we are concerned with the density relation
K ∩ C ∞ (Γ ) = K , (5.19)
which is essential for our convergence analysis to come. Since the embedding
H 1/2(Γ ) ⊂ L1 (Γ ) is continuous and L1 -convergence implies pointwise conver-
gence almost everywhere for a subsequence, K is closed. Therefore it remains to

show
K ⊂ K ∩ C ∞ (Γ ) .
To this end one uses the continuity and surjectivity of the trace operator γ :
H 1 (Ω) → H 1/2 (Γ ) and applies the analogous inclusion
K ⊂ K ∩ C ∞ (Γ ) ,
which in [215, section 4] is proved using Friedrich’s regularization and the fact that
with Ω a Lipschitz domain, H = H 1 (Ω) is a Dirichlet space and hence in particular
the map w ∈ H → w+ = max(0, w) is a continuous map into H .
To conclude this section we refer the interested reader to [222] to see how the
boundary integral approach described above extends to unilateral contact of a linear
elastic body against a rigid foundation in the range of linear elasticity.
5.2 A Variational Inequality of the Second Kind Modelling

Unilateral Frictional Contact
Let Ω ⊂ Rd (d ≥ 2) be a bounded Lipschitz domain with its boundary ∂Ω and

mutually disjoint parts ΓD , ΓN , and ΓC such that ∂Ω = Γ D ∪ Γ N ∪ Γ C and
meas (ΓC ) > 0. Let the data f ∈ H −1/2(ΓN ∪ ΓC ), γ ∈ H 1/2(ΓD ) ∩ C 0 [ΓD ], χ ∈
H 1/2(ΓC ) ∩ C 0 [Γ C ], g ∈ L∞ (ΓC ) be given, where g ≥ 0 and γ |Γ D ∩ Γ C ≤
χ|Γ D ∩ Γ C . Introduce

a(û, v̂) := ∇ û · ∇ v̂ dx ,
Ω
the bilinear form associated to the Laplacian, the convex closed set
& '
K̂ := v̂ ∈ H 1 (Ω) : v̂|ΓD = γ a.e. and v̂|ΓC ≤ χ a.e. ,
the linear form

l(v̂) := f v̂ ds ,
ΓN ∪ΓC
5.2 Modelling Unilateral Frictional Contact 103
and the continuous, positively homogeneous and sublinear, hence convex functional

j (v̂) := g|v̂| ds
ΓC
that describes Tresca friction. Then consider the variational inequality problem (π)
of the second kind: Find û ∈ K̂ such that for all v̂ ∈ K̂,
a(û, v̂ − û) + j (v̂) − j (û) ≥ l(v̂ − û). (5.20)
There exists a unique solution û (see e.g [145, 146, 249]), if ΓD has positive
measure and hence the bilinear form is coercive by the Poincaré inequality. In the
semicoercive case, when ΓD = ∅, a necessary condition for existence of a solution
is the recession condition
j (ρ) ≥ l(ρ), ∀ρ = const. ≤ 0 ,
what is equivalent to

g ds + f ds ≥ 0 .
ΓC ΓN ∪ΓC
If this condition is strengthened to

g ds + f ds > 0 ,
ΓC ΓN ∪ΓC
then existence of a solution is guaranteed (see Appendix C.3.2, [146, 201]).

With g ≡ 0 the variational problem (5.20) specializes to the domain variational
inequality of the first kind that is studied in [297]. As is well-known, this latter
variational inequality is the variational formulation of the mixed unilateral Dirichlet-
Neumann-Signorini boundary value problem for the Laplacian. On the other hand,
with χ ≡ +∞ formally, the unilateral constraint disappears and we arrive at the
variational inequality of the second kind in [221] and in a similar form (with the
Laplacian replaced by the Helmholtz operator) in [205]. To exhibit the relation
of (5.20) to unilateral contact with friction in linear elasticity we insert the following
remark.
Remark 5.1 In linear elasticity, instead of the unknown scalar field û, there is the
displacement field u which decomposes in its normal component un = u · n and
∂u
its tangential component ut = u − un n. Similarly as dual variable, the flux is
∂n
to be replaced by the stress field T with its normal component Tn and its tangential
component Tt . Then unilateral contact with a rigid foundation together with friction
according to Coulomb’s friction law requires the following conditions (see [249,
266]) on the contact surface ΓC :
un ≤ χ, Tn ≤ 0, (un − χ)Tn = 0
and
|Tt | ≤ F |Tn |, F |Tn | − |Tt | ut = 0, ut · Tt ≤ 0,
where F ≥ 0 is the friction coefficient. The latter condition expresses the obvious
law that the modulus of the tangential component is limited by a multiple of the
modulus of the normal component; if it is attained, then the body can slip off in the
direction opposite to Tt ; otherwise, the body sticks.
The fixed point approach to unilateral frictional contact as employed in the
existence proofs [261, 319] leads to a approximating sequence of unilateral
problems with given friction. In these approximations the unknown normal
component is replaced by a given slip stress gn ≥ 0, such that the latter condition
above reduces to
|Tt | ≤ F gn , F gn − |Tt | ut = 0, ut · Tt ≤ 0 .
The weak formulation of the unilateral contact problem with given friction (also
known as unilateral Tresca friction problem) is the following variational inequality
(see [249, section 7] for the proof of the formal equivalence of the classical and
weak formulation): Find u ∈ K such that for all v ∈ K

a(u, v − u) + F gn |vt | − |ut | ds ≥ f · (v − u) ds ,
ΓC ΓN
where f is the surface force, a(·, ·) is the bilinear form of strain energy in linear
elasticity, and K is the appropriately defined convex set. In this sense, (π), (5.20)
gives a simplified (scalar) model of the unilateral contact problem with given
friction.
Here we use potential theory and reduce our variational problem (5.20) on the
domain to the boundary Γ = ∂Ω. We shall obtain two different, but equivalent
boundary variationalinequalities of the second kind: a mixed variational inequality
∂u
in the Cauchy data u|Γ , as unknowns and a primal variational inequality
∂n
in the unknown u|Γ involving the Poincaré–Steklov operator (the Dirichlet-to-
Neumann map).
5.2 Modelling Unilateral Frictional Contact 105
To this end we list the relevant boundary integral operators and recall their
mapping properties. With the fundamental solution for the Laplacian,
1
G(x, y) = − ln |x − y| if d = 2,
2π
1 1
G(x, y) = if d = 3,
4π |x − y|
the operators of the single layer potential V , the double layer potential K, its formal
adjoint K , and the hypersingular integral operator W can be defined for z ∈ Γ, φ ∈
C ∞ (Γ ) as follows:

∂
V φ(z) := 2 G(z, x)φ(x) dsx , Kφ(z) := 2 G(z, x)φ(x) dsx ,
∂nx
Γ Γ

∂ ∂
K φ(z) := 2 G(z, x)φ(x) dsx , W φ(z) := − Kφ(z).
∂nz ∂nz
Γ
From Sects. 2.3, 2.4 we know that the linear operators
V : H −1/2+σ (Γ ) → H 1/2+σ (Γ ), K : H 1/2+σ (Γ ) → H 1/2+σ (Γ )

K : H −1/2+σ (Γ ) → H −1/2+σ (Γ ), W : H 1/2+σ (Γ ) → H −1/2+σ (Γ )
1
are well-defined and continuous for |σ | < .
2
Similarly as with Han [227] for the Signorini problem with the Helmholtz
operator in 3D and with Gwinner and Stephan [222] for the unilateral contact
problem in 2D elasticity, we obtain as an equivalent reformulation of (π,(5.20))
the following boundary variational equality: Find (u, ϕ) ∈ K Γ × H −1/2(Γ ) such
that for all (v, ψ) ∈ K Γ × H −1/2(Γ )
1
B(u, ϕ; v − u, ψ) + j (v) − j (u) ≥ l(v − u), (5.21)
2
where
& '
K Γ := v ∈ H 1/2(Γ ) : v|ΓD = γ |ΓD , v|ΓC ≤ χ
and the bilinear form B is given by
B(u, ϕ; v, ψ) := W u, v − (I + K) u, ψ + V ϕ, ψ + (I + K) ϕ, v .

Note that B is positive semidefinite on H 1/2(Γ )/R×H −1/2(Γ ), but non-symmetric.

Indeed, this mixed variational inequality characterizes a saddle point in K Γ ×
H −1/2(Γ ) with a Lagrangian function
appropriately
defined.
Note that using the duality L1 (Γ ), L∞ (Γ ) there holds (see [151, chapter 4.3])
⎧ ⎫
⎪
⎨ ⎪
⎬
j (v) = g|v| ds = max gvw ds | w ∈ L∞ (Γ ), |w| ≤ 1 a.e. .
⎪
⎩ ⎪
⎭
ΓC ΓC
This leads to another saddle point characterization with a Lagrange multiplier w ∈

L∞ (Γ ), |w| ≤ 1 and a suitable Lagrangian. This latter duality relation will be a
key argument in the convergence analysis to come.
Here we eliminate ϕ in (5.21) (as in [85] for unilateral problems, see [136, 257]
for earlier application of the Schur complement and its boundary integral operator
representation) and obtain as another equivalent reformulation of (π,(5.20)) and
(π,(5.21)) the primal boundary variational inequality: Find u ∈ K Γ such that for all
v ∈ KΓ
Su, v − u + j (v) − j (u) ≥ l(v − u) (5.22)
with the symmetric Poincaré–Steklov operator S for the interior problem,
1- .
S := W + K + I V −1 (K + I ) : H 1/2 (Γ ) → H −1/2 (Γ )
2
which is positive definite on H 1/2(Γ )/R.

For further related boundary variational inequalities that arise from unilateral
contact without friction and with Tresca friction for hemitropic solids in micropolar
elasticity see [182–184].
5.3 A Nonmonotone Contact Problem from Delamination
In this section we describe a nonmonotone contact problem that models the delami-
nation behaviour in bonded lightweight structures. We treat such nonlinear boundary
value problems by a combination of boundary integral methods and regularization
techniques from nondifferentiable optimization based on the investigations in
[332, 335].
Let Ω ⊂ Rd (d = 2, 3) be a bounded domain with Lipschitz boundary ∂Ω. We
assume that the boundary is decomposed into three disjoint parts ΓD , ΓN , and ΓC
such that ∂Ω = Γ D ∪ Γ N ∪ Γ C and, moreover, the measures of ΓC and ΓD are
positive. Zero displacements are prescribed on ΓD , surface tractions t ∈ (L2 (ΓN ))d
5.3 A Nonmonotone Contact Problem from Delamination 107
act on ΓN , and on the part ΓC a nonmonotone, generally multivalued boundary

condition holds. The elastic body Ω is subject to a volume force f ∈ [L2 (Ω)]d
and g ∈ H 1/2(ΓC ), g ≥ 0, is a gap function associating every point x ∈ ΓC with
its distance to the rigid obstacle measured in the direction of the unit outer normal
vector n(x).
Further, ε(u) = 12 (∇u + ∇uT ) denotes the linearized strain tensor and σ (u) =
C : ε(u) stands for the stress tensor, where C is the Hooke tensor, assumed
to be uniformly positive definite with L∞ coefficients. The stress vector on the
surface can be decomposed further into the normal, respectively, the tangential
stress:
σn = σ (u)n · n, σt = σ (u)n − σn n.
Our benchmark problem is a two- or three-dimensional symmetric laminated

structure with an interlayer adhesive under loading (see Fig. 5.1 below for the
2D benchmark problem). Because of the symmetry of the structure and the same
forces applied to the upper and lower part, it suffices to consider only the upper half
of the specimen, represented by Ω ⊂ Rd , d = 2, 3.
Problem (P) Find a displacement u ∈ H1 (Ω) := [H 1(Ω)]d such that
−div σ (u) = f in Ω (5.23)

u = 0 on ΓD ; σ (u)n = t on ΓN
un ≤ g, σt (u) = 0 on Γc ,
−σn (u) ∈ ∂f (un ) on Γc (5.24)
The contact law (5.24), written as a differential inclusion by means of the

Clarke subdifferential ∂f (see Appendix C.2) of a locally Lipschitz function
f , describes the nonmonotone, multivalued behaviour of the adhesive. More
precisely, ∂f is the physical law between the normal component σn of the stress
boundary vector and the normal component un = u · n of the displacement u
on ΓC . A typical zig-zagged nonmonotone adhesion law is shown in Fig. 5.2
below.
Fig. 5.1 Reference configuration for the 2D benchmark under loading [333]
Fig. 5.2 A nonmonotone

adhesion law [333]
To give a variational formulation of the above boundary value problem we define
H1ΓD ,0 = {v ∈ H1 (Ω) : v|ΓD = 0},
K = {v ∈ H1ΓD ,0 : vn ≤ g a. e. on ΓC }
and introduce the H1ΓD ,0 (Ω)-coercive and continuous bilinear form of linear
elasticity

a(u, v) = σ (u) : ε(v) dx.
Ω
Multiplying the equilibrium equation (5.23) in Problem (P) by v − u, integrating

over Ω and applying the divergence theorem yields

σ (u) : ε(v − u) dx = f · (v − u) dx + σ (u)n · (v − u) ds.
Ω Ω Γ
From the definition of the Clarke subdifferential (see Appendix C.2), the nonmono-
tone boundary condition (5.24) is equivalent to
−σn (un )(vn − un ) ≤ f 0 (un ; vn − un ).
Here, the notation f 0 (x; z) stands for the generalized directional derivative of f at
x in direction z defined by
f (y + tz) − f (y)
f 0 (x; z) = lim sup .
y→x,t →0+ t
Substituting σ (u)n by t on ΓN , using on ΓC the decomposition
σ (u)n · (v − u) = σt (u) · (vt − ut ) + σn (u)(vn − un )

and taking into account that on ΓC no tangential stresses are assumed, we obtain the
hemivariational inequality: Find u ∈ K such that

a(u, v − u) + f 0 (un (s); vn (s) − un (s)) ds ≥ f · (v − u) dx
ΓC Ω

+ t · (v − u) ds ∀v ∈ K . (5.25)
ΓN
Using the Poincaré-Steklov operator S we can give a boundary integral formulation

and rewrite (5.25) as a hemivariational inequality defined only on the boundary. To
this end, we introduce the free boundary part Γ0 = Γ \Γ D = ΓN ∪ ΓC and recall
the Sobolev spaces [259]:
H 1/2 (Γ ) = {v ∈ L2 (Γ ) : ∃ v ∈ H 1 (Ω), tr v = v},

H 1/2 (Γ0 ) = {v = v |Γ0 : ∃ v ∈ H 1/2(Γ )},
H̃ 1/2 (Γ0 ) = {v = v |Γ0 : ∃ v ∈ H 1/2(Γ ), supp v ⊂ Γ0 }
with the standard norms
uH 1/2 (Γ0 ) = inf vH 1/2 (Γ ) and uH̃ 1/2 (Γ0 ) = u0 H 1/2 (Γ ) ,
v∈H 1/2 (Γ ),v|Γ0 =u
where u0 is the extension of u onto Γ by zero. The Sobolev space of negative order
on Γ0 are defined by duality as
H −1/2 (Γ0 ) = (H̃ 1/2(Γ0 ))∗ and H̃ −1/2 (Γ0 ) = (H 1/2(Γ0 ))∗ .
Moreover, from [259, Lemma 4.3.1] we have the inclusions
H̃ 1/2(Γ0 ) ⊂ H 1/2(Γ0 ) ⊂ L2 (Γ0 ) ⊂ H̃ −1/2 (Γ0 ) ⊂ H −1/2(Γ0 ).
For the solution u(x) of (5.23) with x ∈ Ω\Γ we have the Somigliana representation
formula, see e.g. [229]:

u(x) = G(x, y) Ty u(y) dsy − Ty G(x, y)u(y) dsy + G(x, y)f(y) dy,
Γ Γ Ω
(5.26)
where G(x, y) is the fundamental solution of the Navier-Lamé equation defined by

⎧
λ+μ (x−y)(x−y).
⎨ 4πμ(λ+2μ)
λ+3μ
log |x − y|I + , if d=2
λ+3μ |x−y|2
G(x, y) =
⎩ λ+3μ
|x − y|−1 I + λ+μ (x−y)(x−y).
, if d=3
8πμ(λ+2μ) λ+3μ |x−y|3
with the Lamé constants λ, μ > 0 depending on the material parameters, i.e. the
modulus of elasticity E and the Poisson’s ratio ν:
Eν E
λ= , μ= .
1 − ν2 1+ν
Here, Ty stands for the traction operator with respect to y defined by Ty (u) :=
σ (u(y)) · ny , where ny is the unit outer normal vector at y ∈ Γ .
1 1
Thus we have the symmetric Poincaré–Steklov operator S : H 2 (Γ ) → H− 2 (Γ )
represented by
1
S= {W + (K + I )V −1 (K + I )}
2
Here, for x ∈ Γ

.
V v(x) = 2 G(x, y)v(y)dsy , Kw(x) = 2 Ty G(x, y) w(y)dsy
Γ Γ

.
K . v(x) = 2Tx G(x, y)v(y)dsy , W w(x) = −2Tx Ty G(x, y) w(y)dsy
Γ Γ
denote the single layer potential, the double layer potential, its adjoint operator, and
the hypersingular operator, respectively.
The Newton potential N is given by

1
Nf = K + I V −1 N0 f − N1 f,
2
where N0 , N1 are given for x ∈ Γ by

N0 f = G(x, y)f(y) dsy , N1 f = Tx G(x, y)f(y) dsy .
Γ Γ
Using the boundary function space, respectively the subset
=1/2(Γ0 )
V =H and K Γ
= {v ∈ V : vn ≤ g a.e. on ΓC },
multiplying Su by v−u, integrate on Γ0 , and using thereby again the decomposition

of σ n on ΓC into a tangential and normal part,we obtain as in the domain based case
the boundary hemivariational inequality, Problem (P): Find u ∈ K Γ such that

Su, v − u Γ0 + f 0 (un (s); vn (s) − un (s))ds ≥ t, v − u ΓN
ΓC
+Nf, v − u Γ0 ∀v ∈ K Γ. (5.27)
To shorten the right hand side we introduce the linear functional

F, v Γ0 = t · v ds + Nf, v Γ0 .
ΓN
To settle the existence of solutions to such hemivariational inequalities we impose

the following growth condition on ∂f : there exist positive constants c1 and c2 such
that for all ξ ∈ R and η ∈ ∂j (ξ ) the following inequalities hold
(a) |η| ≤ c1 (1 + |ξ |);
(b) ηT ξ ≥ −c2 |ξ |
Under these growth condition the functional ϕ : H1 (Ω) × H1 (Ω) → R defined by

ϕ(u, v) = f 0 (un (s); vn (s) − un (s)) ds, ∀u, v ∈ H1 (Ω). (5.28)
ΓC
is well-defined, and as proved in the Appendix C.3.3, ϕ(·, ·) is pseudomonotone

and weakly upper semicontinuous. We recall that the functional ϕ : X × X → R,
where X is a real reflexive Banach space, is pseudomonotone, if un ! u (weakly)
in X and lim inf ϕ(un , u) ≥ 0 imply that, for all v ∈ X, we have lim sup ϕ(un , v) ≤
n→∞ n→∞
ϕ(u, v). Thus the existence of solutions follows from the theory of pseudomonotone
variational inequalities (see C.8, also[163, Theorem 3.1], [212, Theorem 3])
To treat the nonsmoothness in the variational problem we sketch the regulariza-
tion techniques of nondifferentiable optimization that are described in more detail
in Appendic C.2 In order to smooth the functional ϕ we first approximate a locally
Lipschitz function f : R → R via convolution by the function f˜ : R++ × R given
by

f˜(ε, x) = f (x − εy)ρ(y) dy,
R
where ε > 0 is a small regularization parameter and ρ : R → R+ is a probability

density function such that

κ= |t| ρ(t) dt < ∞
R
and
R+ = {ε ∈ R : ε ≥ 0}, R++ = {ε ∈ R : ε > 0}.
In general, the smoothing function f˜ is not easily applicable in practice, since

multivariate numerical quadrature is in generally involved, but for a special class
of functions like a maximum, a minimum or a nested max-min function - what is
sufficient for our applications -, it can be explicitly computed.
For example, if f (x) = max{g1 (x), g2 (x)}, then f (x) = g1 (x) + p[g2 (x) −
g1 (x)], where p : R → R+ is the plus function defined by p = x + = max{x, 0}.
Using, for example, the Zang probability density function
%
1 if − 1
2 ≤t ≤ 1
2
ρ(t) =
0 otherwise
for the smoothing approximation p̃(ε, t) of p(t) defined via convolution, we obtain
⎧
⎪
⎨0 if t < − 2ε
p̃(ε, t) = 1
(t + 2ε )2 if − ε
≤t≤ ε
⎪
⎩
2ε 2 2
t if t > 2ε .
Hence, the smoothing function S̃ : R++ × R → R of f defined by
S̃(ε, x) = g1 (x) + P (ε, g2 (x) − g1 (x))
takes the explicit form

⎧
⎪
⎨ g1 (x) if (i) holds
S̃(ε, x) := 2ε1
[g2 (x) − g1 (x)]2 + 12 (g2 (x) + g1 (x)) + ε
if (ii) holds
⎪
⎩
8
g2 (x) if (iii) holds,
where the cases (i), (ii), (iii) are defined below, respectively, by
(i) g2 (x) − g1 (x) ≤ − 2ε
(ii) − 2ε ≤ g2 (x) − g1 (x) ≤ ε
2
(iii) g2 (x) − g1 (x) ≥ 2ε .
Also in the more general case of a maximum function f (x) = max{g1 (x), . . .
, gm (x)} of smooth functions gj , the smoothing approximation S̃ can be explicitly
constructed (see Appendix C.2).
Thus we introduce Jε : H1 (Ω) → R by

Jε (u) = S̃(un (s), ε) ds.
ΓC
and arrive at the regularized problem (Pε ) of (5.27): Find uε ∈ K Γ such that
Suε , v − uε Γ0 + DJε (uε ), v − uε ΓC ≥ F, v − uε Γ0 ∀v ∈ K Γ, (5.29)

where DJε : H1/2(Γ ) → H1/2 (Γ ) is the Gâteaux derivative of the functional Jε

and is given by

∂
DJε (u), v ΓC = S̃(un (s), ε)vn (s) ds.
ΓC ∂x
We conclude this section with the following uniqueness result. Let cS be the
coerciveness constant of S. Assume now that there exists an α0 ∈ [0, cS ) such that
for any u, v ∈ V it holds
ϕ(u, v) + ϕ(v, u) ≤ α0 u − v2V . (5.30)
Theorem 5.1 Under the assumption (5.30), there exists a unique solution of
problem (P), which depends Lipschitz continuously on the linear form given by
the right hand side.
Proof Assume that u, ũ are two solutions of (P). Then the inequalities below hold:
Su − F, v − u Γ0 + ϕ(u, v) ≥ 0 ∀v ∈ K Γ
S ũ − F, v − ũ Γ0 + ϕ(ũ, v) ≥ 0 ∀v ∈ K Γ.
Setting v = ũ in the first inequality and v = u in the second one, and summing up
the resulting inequalities, we get
Su − S ũ, ũ − u Γ0 + ϕ(u, ũ) + ϕ(ũ, u) ≥ 0. (5.31)
From the coercivity of the operator S and the assumption (5.30) we obtain
cS u − ũ2V ≤ ϕ(u, ũ) + ϕ(ũ, u) ≤ α0 u − ũ2V .
Hence, since α0 ∈ [0, cS ), if u = ũ we receive a contradiction.

Now let Fi ∈ V ∗ and denote ui = uFi , i = 1, 2. Analogously to (5.31), we find
that
Su1 − F1 − Su2 + F2 , u2 − u1 Γ0 + ϕ(u1 , u2 ) + ϕ(u2 , u1 ) ≥ 0.
Hence,
cS u1 − u2 2V ≤ ϕ(u1 , u2 ) + ϕ(u2 , u1 ) + F2 − F1 , u2 − u1 Γ0
and by (5.30),
(cS − α0 )u1 − u2 2V ≤ F2 − F1 , u2 − u1 Γ0 ≤ F1 − F2 V ∗ u1 − u2 V .

Also, since α0 < cS we deduce that
1
u1 − u2 V ≤ F1 − F2 V ∗ ,
cS − α0
which concludes the proof of the theorem.

Next we asssume that the assumption that there exists a constant α0 ≥ 0 (in general
depending on ε > 0) such that

∂ ∂
S̃(x1 , ε) − S̃(x2 , ε) (x1 −x2 ) ≥ −α0 |x1 −x2 |2 ∀ x1 , x2 ∈ R. (5.32)
∂x ∂x
Hence, for any u, v ∈ V , we have
DJε (u) − DJε (v), v − u ΓC

∂ ∂
= S̃(un (s), ε) − S̃(vn (s), ε) (vn (s) − un (s)) ds
ΓC ∂x ∂x
≤ α0 un − vn 2L2 (Γ ≤ α0 u − v2V .
C)
Due to Theorem 5.1, we have the following uniqueness result for the regularized
problem.
Theorem 5.2 Under the assumption (5.32) with α0 < cS , there exists a unique
solution to the regularized problem (Pε ), which depends Lipschitz continuously on
the right hand side F ∈ V ∗ .
The solution of unilateral nonsmooth boundary value problems with
monotone/nonmonotone boundary conditions via multivalued boundary integral
equations, boundary variational inequalites, respectively boundary hemivariational
inequalities can be traced back to the work of Haslinger and Panagiotopoulos.
While Haslinger et al. [232] study the unilateral Poisson problem of steady-state
flow through a semipermeable membrane of infinite thickness, the vectorial linear
elastic contact problem is treated by a reciprocal (dual) approach in [336]. In all
their work, without using potential theory, the Poincaré–Steklov operator (or rather
its inverse) has to be constructed by the solution of appropriate linear boundary
value problems in the domain.
Chapter 6
A Primer to Boundary Element Methods
This chapter introduces the BEM in its h−version. First we make Fourier expansion
of Chap. 3 more precise by asymptotic error estimates. Then we prove direct
and inverse approximation estimates for periodic spline approximation on curves.
Hence we develop the analysis of Galerkin methods and collocation methods for
Symm’s integral equation towards optimal a priori error estimates. Moreover, we
subsume Galerkin and collocation methods as general projection methods. To this
end we extend the above treatment of positive definite bilinear forms to the analysis
of a sequence of linear operators that satisfy a uniform Gårding inequality and
establish stability and optimal a priori error estimates in this more general setting.
Interpreting several variants of collocation methods that combine collocation and
quadrature as extended Galerkin methods we include their numerical analysis as
well. Then augmenting the boundary element ansatz spaces by known singularity
functions the Galerkin method is shown to converge with higher convergence
rates. Finally to obtain higher convergence rates in weaker norms than the energy
norm the Aubin–Nitsche duality estimates of FEM are extended to BEM so
that it allows the incorporation of the singular solution expansion for nonsmooth
domains. Sections 6.1–6.4 are based on the classroom notes by M. Costabel [116]
whereas Sects. 6.5.1–6.5.6 are based on the classroom notes by W.L. Wendland
[430]. Improved estimates of local type, pointwise estimates and postprocessing
with the K-operator are considered in Sects. 6.5.7–6.5.9. Discrete collocation
with trigonometric polynomials, where the concept of finite section operators is
used, is a subject of Sect. 6.6. In Sect. 6.7 the standard BEM is enriched by
special singularity functions modelling the behaviour of the solution near corners,
thus yielding improved convergence. In Sect. 6.8 Galerkin-Petrov methods are
considered. Section 6.9 presents the Arnold-Wendland approach to reformulate a
collocation method as a Galerkin method whereas qualocation is investigated in
Sect. 6.10. In Sect. 6.11 the use of radial basis functions (a meshless method) and of
spherical splines in the Galerkin scheme is demonstrated for problems on the unit
sphere. Integral equations of the first kind with the single layer and double layer

116 6 A Primer to Boundary Element Methods
potentials are our main subject. Integral equations of the second kind are studied
only briefly, e.g. at the end of Sect. 6.4.
There has been a tremendous amount of research on spline collocation and
Galerkin methods. We want to mention the works by J. Schmidt [365, 366], S.
Prössdorf and B. Silbermann [345] and J. Saranen and A. Vainikko [356]. The hp-
version of the BEM is one of the main subjects of this book and therefore considered
separately in Chaps. 7 and 8.
For further reading we refer to the seminal papers by Hsiao and Wendland [257]
and Nedelec and Planchard [324] and to the survey articles by W.L. Wendland
(Part III of [362, 429] and [428]) and the lecture notes of J.C. Nedelec [320].
6.1 Galerkin Scheme for Strongly Elliptic Operators
As before let X, Y be Hilbert spaces and A : X −→ Y a continuous, linear and

bijective operator. In Chap. 2 we saw that cap(Γ ) < 1 is a sufficient condition
for the operator V to be positive definite. In order to cover also examples like
A = I + C , C compact, or A = V with cap(Γ ) > 1 we need a more general
sufficient criterion and therefore the notion of a compact operator, see Definition A.6
in Appendix A. We shall now consider again the situation of the general Galerkin
method.
Theorem 6.1 Under the assumptions:
1. ∃ PN : Y −→ TN , bounded and linear, converging on Y strongly to the identity
operator, i.e.
N→∞
PN η − ηY −→ 0 ∀η ∈ Y ,
2. ∃ QN : XN −→ TN , ∃ M (independent of N ) :
|QN v, Aw | ≤ M vX wX , ∀ v ∈ XN , w ∈ X, N ∈ N,
3. ∃ C : X −→ X compact , ∃ α > 0 (independent of N ):
|QN v, Av + Cv, v | ≥ αv2X
(i) Existence of a unique solution
∃ N0 ∈ N ∀ N ≥ N0 ∃! uN : t, AuN = t, Au = t, f ∀ t ∈ TN ,
(ii) Stability of the method
∃ c > 0, N0 ∈ N ∀ N ≥ N0 : uN ≤ cu ∀ u,
(iii) Quasioptimal error estimate
∃ c > 0 , N0 ∈ N ∀ N ≥ N0 : u − uN ≤ c · inf u − χ,
χ∈XN
6.1 Galerkin Scheme for Strongly Elliptic Operators 117
(iv) Convergence of the method

N→∞ N→∞
d(u, XN ) −→ 0 ⇒ ||u − uN ||X −→ 0 .
Proof In Sect. 1.2 we have already proved the theorem for the case C ≡ 0 . We
now want to reduce the general case to this, i.e. we want to have
|Q̃N v, Av | ≥ α̃ v2 , ∀ v ∈ XN . (6.1)
This would clearly be satisfied due to assumption 3., if we had
Q̃N v, Av = QN v, Av + Cv, v . (6.2)
However, this would imply
A Q̃N = A QN + C
⇔ Q̃N = QN + (A )−1 C
but raises the problem that we do not have a mapping from XN into TN . Here we
define the operator
Q̃N := QN + PN (A )−1 C.
Then,
Q̃N = QN + (A )−1 C + (PN − 1)(A )−1 C ,
where
(PN − 1)(A )−1 C ⇒ 0 (in operator norm)
by Lemma A.2 in Appendix A, since (A )−1 C is compact and PN − 1 → 0 by

assumption. Thus we have
(a)
|Q̃N v, Aw | ≤ |QN v, Aw | + |PN (A )−1 Cv, Aw |

≤ M̃v w, ∀ N .
(b)
|Q̃N v, Av | = |QN v, Av + Cv, v + (PN − 1)(A )−1 Cv, Av |

≥ |QN v, Av + Cv, v | − |(PN − 1)(A )−1 Cv, Av |
≥ α v2X − (PN − 1)(A )−1 C · A · v2X
≥ α2 v2X , N ≥ N0 .

Corollary 6.1 Let A, B : X −→ Y be two bijective, continuous operators with

the operator A − B being compact. If the general Galerkin method of Sect. 1.2
converges for the operator A, it converges for B, too.
Definition 6.1 An operator A : X −→ X is said to be strongly elliptic, if there
exists a decomposition A = D + C with D positive definite and C compact.
Remark 6.1 The operator A is strongly elliptic if and only if there exists a constant
α > 0 and a compact operator C : X −→ X such that
Av, v ≥ αv2X + Cv, v ∀ v ∈ X. (6.3)
In the case of complex functions we have:
%Av, v ≥ αv2X + %Cv, v ∀ v ∈ X. (6.4)
The above inequalities are due to Gårding and thus usually referred to as Gårding
inequalities.
As a consequence of Theorem 6.1 we have
Theorem 6.2 Let A : X −→ X be a strongly elliptic and bijective operator. Then
every Galerkin scheme for A is convergent.
Example 6.1 Consider the single layer potential V for Γ ∈ Lip. Then we may
choose
1 1
X = H − 2 (Γ ), and X = H 2 (Γ ) .
Theorem 6.3 Let Γ be a Lipschitz curve with cap(Γ ) = 1 . Then every Galerkin
scheme for A = V is convergent.
Proof Consider first the case that cap(Γ ) < 1 . Then the operator V : HV :=
H − 2 ((Γ ) −→ HV is positive definite.
1
For cap(Γ ) > 1 , use the integral mean m(ϕ), see Definition 2.7. Then setting
ϕ0 = ϕ − m(ϕ) yields
V ϕ, ϕ = ϕ0 , V ϕ0 + Cϕ, ϕ
with Cϕ = m(ϕ)
L . Hence, the assertion of the theorem follows by application of
Theorem 6.2.

6.2 Galerkin Methods for the Single-Layer Potential 119
6.2 Galerkin Methods for the Single-Layer Potential
6.2.1 Approximation with Trigonometric Polynomials
We now want consider an approximation on Γ = ∂B1 (0) with trigonometric

∞
polynomials. Note, that for u(t) = uk e2πikt with coefficients uk :=
k=−∞
1
u(t)e−2πikt dt we have
0
∞

u2s := (1 + k 2 )s |uk |2 .
k=−∞
We then define
• TN :=⎧span{e2πikx : |k| ≤ N} with dim(TN ) = 2N + 1
⎨ H s −→ TN
• ΠN :
⎩ u → (ΠN u)(x) := uk e2πikx ∈ TN .
|k|≤N
Theorem 6.4 With the above definitions we have

N→∞
i) u − ΠN us −→ 0 for u ∈ H s (Γ ) .
ii) Approximation Property
∀ r ≤ s ∃ cr,s (independent of N ) :
u − ΠN ur ≤ cr,s · N r−s us , ∀ u ∈ H s ∀ N ∈ N .
iii) Inverse Estimate
∀ r ≤ s ∃ c̃r,s (independent of N ) :
vs ≤ c̃r,s · N s−r vr ∀ v ∈ TN , ∀ N ∈ N .
Proof
i) clear.
ii) u − ΠN u2r = (1 + k 2 )r |uk |2
|k|>N
≤ (1 + N 2 )r−s (1 + k 2 )s |uk |2
|k|>N
≤ 2r−s N 2(r−s) u2s .
iii) v2s = (1 + k 2 )s |vk |2

|k|≤N
≤ (1 + N 2 )s−r v2r .
As we have seen in Chap. 3 there holds
Theorem 6.5 For Γ = ∂B1 (0) the mapping V : H s (Γ ) −→ H s+1 (Γ ) is
continuous.
For the Galerkin scheme with trigonometric polynomials for V u = f on

∂BR (0), R = 1 , we now have by application of Theorem 6.1 (or Theorem 6.2)
the following result.
Theorem 6.6 Let uN be a solution to V u = f using the Galerkin method with
trigonometric polynomials TN as trial and test functions. For Γ ∈ C ∞ , cap(Γ ) =
1 and u ∈ H s (Γ ) , s ≥ − 12 we then have
u − uN − 1 ≤ N −(s+ 2 ) us .
1
Furthermore let r, s ∈ R with r ≤ s be arbitrary.Then there exists a constant c

independent of N such that u − uN r ≤ c · N r−s us
Proof Using the approximation property in Theorem 6.4, we have
− s+ 12
u − uN − 1 ≤ c · inf u − w− 1 ≤ c · N us .
2 w∈TN 2

We may also think about estimates in other norms, for instance
a) Norms above the energy-norm (r ≥ − 12 , s ≥ r) :
u − uN r ≤ u − ΠN ur + ΠN u − uN r
1
≤ c · N r−s us + c · N r+ 2 · ΠN u − uN − 1

2
− s+ 1
c·us N 2
≤ c · N r−s us .
b) Norms below the energy-norm (r ≤ − 12 ) :
u − uN r = sup u−u N ,v
v|r|
v∈H |r|
|u − uN , v | = |V (u − uN ), V −1 v | V −1 v|r|−1 ∼
=v|r|
= |V (u − uN ), V −1 v − t | , ∀ t ∈ ΠN
≤ V (u − uN )r +1 · V −1 v − t−r −1
(since: −r ≤|r| ⇔ r ≥r )
−|r|−r −1
≤ c · u − uN r · c N V v|r|−1

N r−r v−r
(by the approximation property of the above theorem)

⇒ u − uN r ≤ c · N r−r u − uN r , ∀ r ≥ r .
For r = − 12 it follows ∀ s ≥ − 12 :
u − uN r ≤ c · N r−s us .
Corollary 6.2 For any real number s there exists a constant c such that for u ∈
H s (Γ ),
|m(u) − m(uN )| ≤ c · N −s and |cap(Γ ) − capN (Γ )| ≤ c · N −s .

2π
2π − Lm(e
Note, for V e = 1 we have: cap(Γ ) = e− Lm(e) , capN (Γ ) = e N) .
Proof With the above theorem we have
|m(u) − m(uN )| = |1, u − uN |

≤ 10 u − uN 0
≤ cN −s us .

6.2.2 Approximation with Splines
Let a mesh ΔN on Γ be defined analogously to section 6.3, i.e.
j 1
xj = e2πi N = x(j h) =: x(sj ) , h=
N
d ≡ S d , as defined there with d ≥ −1 . We may assume that N is odd,
and let SΔ N h
e.g. N = 2M + 1 . Each φ ∈ Shd is a polynomial of degree d on each interval, so
d + 1 coefficients have to be determined. As φ is (d − 1)-times differentiable in the
nodes of ΔN , we get that dim Shd = N.
Theorem 6.7 With the above definitions there holds:
i) v ∈ Shd ⇔ vk · k d+1 = vk+N · (k + N)d+1 , ∀ k ∈ Z .
ii) v ∈ SΔd
N
is uniquely determined by {vk : |k| ≤ M} , i.e.
ΠM : SΔ d → T is a bijection with inverse Q := Π −1 : T → S d .
N N M M N ΔN
iii) For s < d + 2 there exists constants c1 , c2 such that
1
c1 ΠM vs ≤ vs ≤ c2 ΠM vs , ∀ v ∈ SΔ

d
N
.
−1
Proof We first note that w ∈ SΔ ⇔ wk+N = wk ∀ k ∈ N .
N
d d+1 −1
To prove i), we have : v ∈ SΔN implies ds
d v ∈ SΔ N
⇔ (2πik) vk is N-periodic ⇔ (2πi(k + N)) vk+N = (2πik)d+1vk .
d+1 d+1
To prove ii) write k ∈ Z as k = r + l · N , |r| ≤ M . Hence
r d+1 r d+1
vk = vr · = vr .
(r + lN) d+1 k
This gives the desired mapping

QN : vr e2πikx −→ vk e2πikx .
|r|≤M k∈Z
To prove iii) first note ΠM vs ≤ vs . Moreover,
v2s = (1 + k 2 )s |vk |2
k∈Z
M s
= 1 + (r + lN)2 |vr+lN |2
r=−M l∈Z
M s r 2(d+1)
= 1 + (r + lN)2 · |v |2
(r+lN)2(d+1) r
r=−M l∈Z
M s r 2(d+1)
= |vr |2 1+ (r + lN)2 · (r+lN)2(d+1)
r=−M l∈Z
Now for r = 0,
(1 + (r + lN)2 )s
r 2(d+1) ≤ 2 (r + lN)2(s−d−1)r 2d+2
(r + lN)2(d+1)
l l
is finite , if and only if, s − d − 1 < −1/2

We define
PN := QN ΠM , (6.5)
i.e., v = PN u ∈ SΔ
d
N
is uniquely determined by vr = ur , ∀ |r| ≤ M .
Theorem 6.8 With the above definition (6.5) the following holds:
i) Approximation Property
∀ r ≤ s , r < d + 12 , s ≤ d + 1 ∃ cr,s (independent of N ) :
u − PN ur ≤ cr,s N r−s us ∀u ∈ Hs .
ii) Inverse Property

∀ r ≤ s < d + 12 ∃ cr,s (independent of N ) :
vs ≤ cr,s N s−r vr ∀ v ∈ SΔ

d
N
.
Proof ii) follows directly from assertion (i) of Theorem 6.7 combined with the
inverse property for trigonometric polynomials.
To prove i) let k = r + lN. Then

d+1 2
u k − u r r
τ
u − PN u2τ = 1+k 2
k
r=0,|r|≤M l=0,l∈Z
and u − PN uτ ≤ u − ΠM uτ + ΠM u − PN u2τ

≤cM τ −s us
τ r 2(d+1)
⇒ ΠM u − PN u2τ = |ur |2 1 + k 2 .
k
r=0,|r|≤M l=0,l∈Z
τ r 2(d+1)
With 1 + k2 ≤ c · k 2τ −2d−2 · r 2d+2
k 2τ −2d−2
= c · r 2τ · kr
2τ −2d−2 2τ −2d−2
= c · r 2τ Nr · l + Nr

N 2s−2d−2 2τ −2d−2
= c · r 2s N 2(τ −s) · l + Nr
r

≤c for s≤d+1
it follows
r 2τ −2d−2
ΠM u − PN u2τ ≤ c · |ur |2 r 2s N 2(τ −s) · l+ .
N
r=0 l=0

≤ΠM u2
s ≤c for τ <d+ 1
2
Thus, we finally obtain
ΠM u − PN u2τ ≤ cu2s · N 2(τ −s) ⇒ (i) .

For the Galerkin method using spline functions to approximate the solution of
the integral equation of first kind V u = f on Γ = ∂BR (0), R = 1 we obtain
Theorem 6.9
i) Let f ∈ H s+1(Γ ) , s ≥ − 12 . Then, for N sufficienly large there exists a
unique uN ∈ Shd , d ≥ 0 such that:
v, V uN = v, f ∀ v ∈ Shd .
ii) For r ≤ s , −d − 2 ≤ r < d + 1

2 and −d − 3
2 < s ≤ d + 1 there holds
for u ∈ H s
u − uN r ≤ c · N r−s us .
iii) For f ∈ H d+2 we have
u − uN − 1 ≤ c · N − 2 −d ud+1
1
2
and u − uN −d−2 ≤ c · N −2d+3 · ud+1 .
iv) The above also holds for Γ ∈ C ∞ with cap(Γ ) = 1 .

Proof We only want to prove (iv) and leave the other items as an exercise to the
reader.
We have

x(s) − x(t)
log |x(s) − x(t)| = log |e 2πis −e 2πit
| + log 2πis
e − e2πit

=0, ∈C ∞
implies VΓ = V∂B0 + T , with T : Hs −→ C ∞ ∀s
Thus, T : H s (Γ ) −→ H t (Γ ) ∀ s, t ∈ R is compact.

6.3 Collocation Method for the Single-Layer Potential
For further reading see [305, 356] and also for general strongly elliptic equations
[345]. As an example of the general Galerkin scheme in Sect. 6.1 we want to
consider in this section a collocation method for the single layer potential V .
Let Γ = ∂Br (0) be a circle of positive radius r < 1 with the centre in the origin.
1
Then we know that V is positive definite in H − 2 (Γ ). We define a mesh ΔN on Γ
as follows:
j
xj = r · e2πi N = x(j h) , h = N1
x(s) = r · e2πis , |ẋ| = r · 2π = 1
Example 6.2 For N = 8 we would have Fig. 6.1.
Fig. 6.1 Break points

x2
x3 x1
x4 x0
x5 x7
x6
6.3 Collocation Method for the Single-Layer Potential 125
We define the spaces
d ≡ S d := {ϕ ∈ C d−1 (Γ ) : ϕ| s , s
SΔ N h [ i i+1 ] is a polynomial of degree d }
−1
SΔ N
:= span{δ xj : j = 0, · · · , N − 1}
and consider the nodal collocation for linear continuous functions for the equation
Vu = f
i.e. we are looking for a solution uN ∈ SΔ

1
N
of
(V uN )(xj ) = f (xj ) , j = 0, · · · , N − 1 . (6.6)
3
Theorem 6.10 Let f ∈ H 2 (Γ ) be given. Then the collocation method (6.6)
1
converges in H 2 (Γ ) , i.e.
∃ N0 ∈ N ∀ N ≥ N0 ∃! uN ∈ SΔ
1 :
N
uN 1 ≤ cu 1
H 2 (Γ ) H 2 (Γ )
and u − uN 1 ≤ c̃ inf u − v 1
H 2 (Γ ) 1
v∈SΔ H 2 (Γ )
N
for some constants c and c̃ .

1 3
Proof Consider V as a mapping X := H 2 (Γ ) −→ H 2 (Γ ) =: Y . Since the
mapping V: H s (Γ ) −→ H s+1(Γ ) is continuous and bijective for smooth Γ , we
know that V is an isomorphism. We therefore first have to show that
1
(i) XN := SΔ
1 ⊆ H 2 (Γ )
N
−1
⊆ H − 2 (Γ ) .
3
(ii) TN := SΔ N
1
1 ⊆ H 1 (Γ ) ⊆ H 2 (Γ ) . For (ii) by Sobolev’s embedding theorem,
For (i) note SΔ N
H s (Γ ) ⊆ C(Γ ) for s > 12 . Thus,
1
u2s = c · |uk |2 (1 + k 2 )s , with uk = u(s)e−2πiks ds .
k∈ZZ 0
In a second step we now have to construct the operators QN : XN → TN . In the

distributional sense, let QN = −( ds
d 2
) . Then Q := QN gives QN SΔ1 ⊆ S −1 .
N ΔN
To verify the assumptions of Theorem 6.1 we have to show that
1. |QN v, Aw | ≤ MvX wX , for v ∈ XN , w ∈ X
2. |QN v, Av + Cv, v | ≥ αv2X , for v ∈ X
with A = V . The assertion 1. follows, since
1
|Qv, V w | = | − v (s)(V w)(s) ds|
0
1
= | v (s)(V w) (s) ds|
0
1
= | v (s)(V w )(s) ds| ≤ Mv −1 · w −1
0 H 2 H 2
Setting w := v above further yields 2., since
v , V v ≥ γ v 2 1 ≥ γ̃ v2 1 ,
H − 2 (Γ ) H 2 (Γ )
which completes the proof of the theorem.

6.4 Collocation Methods—Revisited
After having briefly discussed the collocation method for the single layer potential
in the previous section, we now want to investigate this method in some more detail
and a more general setting. In the previous chapters we have already seen that the
single layer potential V , given by

1
V u(x) = − ln |x − y|u(y) dsy ,
π
Γ
yields a solution of the Laplace equation. To solve the corresponding Dirichlet prob-
lem, we have to find a function u such that the boundary condition V u(x) = f (x)
is satisfied for all x ∈ Γ . Since an analytic solution can rarely be given, one is
looking for a approximate solution in a finite dimensional trial space of dimension
N. Of course, we will not achieve that V uN = f holds for all boundary points.
If we choose N boundary points x1 , x2 , . . . , xN and find a function uN such that at
these so-called collocation points the equation
V uN (xj ) = f (xj ) , j = 1, . . . , N
holds, we obtain an approximate solution for the problem, which may approach the
exact solution u with N growing. This method is called collocation.
Although this method is not very sophisticated and good results have been
attained by its practical application, the proof of convergence is delicate and for
some simple cases it still remains open. For convergence estimates for regions with
corners see Sect. 6.10.
6.4 Collocation Methods—Revisited 127
In this chapter, we now want to prove the convergence of the collocation method
for the single layer potential with smooth boundary, i.e. Γ ∈ C ∞ .
Let {μ1 , . . . , μN } be a basis of the trial space VN with x1 , . . . , xN being the
collocation points. Then, the approximate solution uN has a representation of the
form

N
uN = αk μk .
i=1
The problem to be solved is given as follows:
Find uN ∈ VN such that

V uN (xj ) = f (xj ) , for j = 1, . . . , N
which can be rewritten by
− π1 ln |xj − y|uN (y) dsy = f (xj ) j = 1, . . . , N

Γ
N
⇐⇒ − π1 ln |xj − y| αk μk (y) dsy = f (xj ) j = 1, . . . , N
1Γ k=1 2
N
⇐⇒ αk − π1 ln |xj − y|μk (y) dsy = f (xj ) j = 1, . . . , N
k=1 Γ
This yields a linear system of equations for the unknowns α1 , . . . , αN . Thus, for the
calculation of the approximate solution, the collocation method leads to well-known
numerical tasks.
For the further studies we want to recall again the δ-distribution, as already
introduced in Chap. 2. The δ-distribution was defined by
δxj (f ) = f (xj ) ∀ f ∈ C0∞ (Ω) .
However, one may conceive the δ-distribution δxj as the derivative of a piecewise
constant function with jump at xj . This enables to link with the Galerkin-method as
follows:
Galerkin: Find uN such that t, V uN = t, f ∀ t ∈ VN

Collocation: Find uN such that V uN (xj ) = f (xj ) j = 1, . . . , N
Thus, both methods are projection methods of the form

, (6.7)
t, V uN = t, f ∀ t ∈ TN
where an approximate solution uN in a trial space VN is to be found, i.e. the exact

solution u is projected onto the trial space VN and tested against a testfunction
t ∈ TN . While for the Galerkin-method the test and trial space are identical, for
the collocation-method the test space TN is spanned be the δ−distributions of the
collocation points.
Remark 6.2 Choosing in (6.7) for the testfunction t a δ-distribution δxj , we obtain
t, V uN = δxj (V uN ) = V uN (xj )
for the left hand side; and for the right hand side
t, f = δxj (f ) = f (xj ) .
In the subsections to follow we now want to discuss some important properties

of the test and trial space of the collocation-method, before proving a convergence
theorem for projection methods. Finally, we will show that the collocation method
satisfies the assumptions of Theorem 6.11 and is therefore convergent.
6.4.1 Periodic Splines as Test and Trial Functions
We want to consider a region Ω with smooth boundary Γ = ∂Ω ∈ C ∞ . It

is well known that such a region can be transformed to the unit circle, apart from
a compact perturbation of the solution. Hence, we will consider the problem only
on the unit circle. To simplify the calculations further on, the boundary of the unit
circle is mapped onto the interval [0, 1] . Let the boundary of the circle and the unit
interval be partitioned uniformly according to
3 4
ΔN = x 0 , . . . , x N−1 x j = e2πij/N
ΔN = Δ = {x0 , . . . , xN−1 } xj = j/N .
Let the number of grid points be odd, i.e. N = 2M + 1 and let the trial functions
be splines of degree d . We may therefore define
"
ϕ ∈ C d−1 (globally) is a spline function of degree d
d
SΔ := ϕ :
with respect to ΔN = Δ , continued periodically on R
Since inside every interval the spline function is a polynomial of degree d, i.e. d + 1
coefficients have to be determined, and it is further d − 1-times differentiable in the
grid points, the dimension of the trial spaces will be
d
dim SΔ = N · (d + 1) − N · d = N .
−1
Let SΔ := span{δ(xj )| j = 1, . . . , N} be the space being spanned by the δ-
distributions corresponding to the grid points xj . Here, δ(xj ) is to be understood
as the derivative of a piecewise constant function with a jump at xj . Since the
δ-distributions are tempered, we have the following properties of the Fourier
coefficients of spline functions as defined above:
(i)
1
(δ@
x j )k = e−2πikx · δxj dx = e−2πikxj = e−2πikj/N
0
(δ@
xj )k+N = e
−2πi(k+N)j/N = e −2πikj/N · e −2πij = e −2πikj/N

=1
Thus, for j = 1, . . . , N there holds
(δ@
xj )k = (δ@
xj )k+N ,
and therefore also

−1
w ∈ SΔ ⇒ >k = w
w >k+N ,
i.e. we have periodic Fourier coefficients.

(ii)
d+1
d −1
v ∈ SΔ
d
⇒ v ∈ SΔ (6.8)
ds
d d
⇒ [( )d+1 v] = [( )d+1 v] .
ds k ds k+N
For any arbitrary function f there holds

d m
f = (2πik)m f>k
ds k
and thus
k d+1
(2πik)d+1>
vk = (2πi(k + N))d+1>
vk+N , hence >
vk+N = >
vk .
(k + N)d+1
(iii) Any integer k ∈ Z

Z has a representation of the form
k = r + lN with |r| ≤ M , l ∈ ZZ .
Thus, there holds ∀v ∈ SΔ

d
:
(r + (l − 1)N)d+1 r d+1
>
vk = >
vr+lN = >
vr+(l−1)N = >
vr . (6.9)
(r + lN) d+1 (r + lN)d+1
Here, one can easily see the reason for N being chosen to be odd and the fact
that the spline functions are already uniquely determined by their first r Fourier
vr , |r| ≤ M .
coefficients >
(iv) For the Sobolev norm we recall

v2H s = |k|2s |>
vk |2 + |>
v0 |2 .
k=0
Since the spline functions are determined by only a few coefficients, it may be
sufficient to consider only these in the definition of the norm. This abbreviated
Sobolev-norm can be estimated trivially by the original one. Both norms would
be equivalent, if the original norm could also be estimated by the abbreviated
one. For the sake of simplicity we will only write:

v2H s = |k|2s |>
vk |2 .
k
Here, we note that one obtains the same results by treating the coefficient >
v0
seperately. Now (6.9) gives for any v ∈ SΔ
d,
M
v2H s = |k|2s |>
vk |2 = |r + lN|2s |>
vr+lN |2
k r=−M l∈Z
M 2(d+1)
= |r + lN|2s (r+lN)
r
2(d+1) |>
vr |2
r=−M l∈Z
M
(r+lN)2s r 2d+2
= |>
vr |2 r 2s r 2s
· (r+lN) 2d+2
r=−M l∈Z
M r+lN 2(s−d−1)
= |>
vr |2 r 2s r .
r=−M l∈Z
r+lN 2(s−d−1)
The second factor of the last term, i.e. r , is only bounded in
l∈Z
the case
2(s − d − 1) < −1 ,
since r+lN
r = 1 + l Nr ≥ 1 + 2l is not bounded. In this case, both norms are
equivalent for v ∈ SΔ d
and s < d + 12 . In the following we will always use
the abbreviated norm without changing the notation.
As already mentioned above, the collocation points are distributed uniformly

over the circle and the interval, respectively, i.e. let
=N,ε = {x̃ 0 , . . . , x̃ N−1 } ,

Δ ε ∈ [0, 1)
=N,ε = Δ= = {x̃0 , . . . , x̃N−1 } , j +ε
Δ x̃j = xj + ε/N = N
−1
SΔ= := span{δ(x − x̃j ) | j = 1, . . . , N}
We therefore have
1

δ(x − x̃j )k = e−2πikx δ(x − x̃j ) dx = e−2πik x̃j
0

j+ε
and δ(x − x̃j )k+N = e−2πi(k+N)x̃j = e−2πik x̃j · e−2πiN N

= e−2πik x̃j · e−2πij · e−2πiε = δ(x − x̃j )k · e−2πiε
and thus altogether

−1
> vr · e−2πilε
vr+lN = > ∀ v ∈ SΔ
= .
6.4.2 Convergence Theorem for Projection Methods
Here we need some results on the convergence of projection methods, including

compact perturbations and spaces with two norms. Here the concept of collectively
compact operators (see Appendix A, Definition A.7) introduced by Anselone [2]
is an important tool. Such results are well-known [202, 273, 345] but we present a
formulation from [131] that is particularly adapted to the present case. This version
includes Theorems 1.1, 1.2, 1.3, and 6.1 as special cases.
Let X and Y be Banach-spaces and A : X → Y a continuous and bijective
operator. Let (TN )N and (VN )N be sequences of test and trial spaces with VN ⊂

X, TN ⊂ Y and dim VN = dim TN < ∞ ∀ N ∈ N . We consider the problem:

(6.10)
t, AuN = t, f , ∀ t ∈ TN
Theorem 6.11 (Lemma 1.1 in [131]) Let the following assumptions be satis-
fied:

1. There exist bounded linear operators PN : Y → TN , converging on Y strongly
to the identity operator, i.e.
N→∞
PN v − vY −→ 0 ∀v ∈ Y .
2. Let X0 be a Banach space, continuously embedded in X with the norm · X0 ,

i.e.
xX ≤ CxX0 ∀ x ∈ X0 .
3. Let VN ⊂ X0 ∀N ∈ N .
4. For all N ∈ N we are given a mapping QN : VN → TN and a constant M
independent of N such that
| < QN v, Aw > | ≤ MvX wX0 ∀ v ∈ VN , ∀ w ∈ X0 .

5. There exist a collectively compact sequence of operators CN : X → Y and a
constant γ such that
| < QN v, Av > + < CN v, v > | ≥ γ v2X ∀v ∈ VN , ∀N ∈ N.
Then we have:
i) Existence of a unique solution
There exists an N0 ∈ N such that ∀ N ≥ N0 the system (6.10) has a unique
solution uN ∈ VN for any f ∈ Y .
ii) Stability of the method
∃ C (independent of N ) : uN X ≤ CuX0 ∀ N ≥ N0 , ∀ u ∈ X0 .
iii) Quasi-optimal error estimate
∃ C (independent of N ) : u − uN X ≤ C · inf u − χX0 ∀ N ≥ N0 .
χ∈VN
Proof We first want to prove the unique solvability of (6.10) and the stability
property for the case CN = 0 for all N ∈ N , then for arbitrary CN . Eventually, the
quasi-optimality is to be derived from the first two statements.
Uniqueness: Let t, Av = 0 for all t ∈ TN . Then there holds with QN v ∈ TN :
5.
γ v2X ≤ |QN v, Av | = 0 ⇒ v ≡ 0 ,
thus, the homogeneous problem has only the trivial solution. Hence the solution
is unique.
Existence: Testing t, AuN = t, f only for a basis of TN and representing
uN in terms of a basis of VN , we obtain a N × N-system of linear equations,
which has to be regular by the uniqueness of the solution. Hence, it is solvable.
For uN ∈ VN there holds assumption 5.:
|QN uN , AuN | ≥ γ uN 2X .

We therefore have:
uN 2X ≤ γ |QN uN , AuN |

1
= γ |QN uN , f |
1
4.
= γ | < QN uN , Au >
1
| ≤ Mγ uN X · uX0 ,
Hence
M
uN X ≤ uX0 . (6.11)
γ
Now we consider the general case with nonvanishing perturbations CN ≡ 0 .

By assumption there exists operators CN and QN satisfying 4. and 5. With these,
we will now define operators Q=N and C=N that also satisfy the two estimates such
=N ≡ 0 , i.e. we reduce the general case to the special case considered
that now C
above.
Let A−1 be the inverse of the adjoint A : Y → X to A which is also
continuous and bijective. We define
=N := QN + PN A −1 CN
Q

= QN + A −1 CN − (1 − PN )A −1 CN
=N satisfies assumption 4.:

Then Q

=N v, Aw | ≤ |QN v, Aw | + |PN A −1 CN v, Aw |
|Q (by 4. for QN )

≤ MvX · wX0 + PN A −1 CN · A · vX · wX .

Furthermore, PN A −1 CN · A is bounded, since PN is bounded, A and A−1
are both bounded and CN is compact, hence bounded, too. With assumption 2. we
therefore obtain:

=N v, Aw | ≤ MvX wX0 + PN A −1 CN · A ·vX · C · wX0
|Q

=:M1
≤ (M + M1 C)vX · wX0
=N satisfies assumption 5.:

We now have to show that Q

=N v, Av | = |QN v, Av + A −1 CN v, Av − (1 − PN )A −1 CN v, Av |
|Q

= | QN v, Av + A A−1 CN v, v − A (1 − PN )A −1 CN v, v |

≥γ v2X by 5. for QN
As 1 − PN → 0 strongly on Y and the operators CN and thus A−1

CN are
−1
collectively compact, by Lemma A.2 the sequence A (1 − PN )A CN N tends

to zero. Defining δN := A (1 − PN )A −1 CN , we obtain:
=N v, Av | ≥ (γ − δN ) · v2X .
|Q
If we now choose N0 such that γ −δN > 0 for all N ≥ N0 , e.g. γ −δN ≥ γ̃ > 0 ,
there holds:
=N v, Av | ≥ γ̃ v2X .
|Q
Let the solution operator u −→ uN be denoted by GN . By (6.11), GN : X0 →

(VN , · X ) is a projection operator with bounded norm:
GN uX uN X CuX0

GN = sup = sup ≤ sup = C.
u uX0 u u X0 u uX0
Furthermore, there holds for all v ∈ VN that GN (v) = v .

Now, in order to prove quasi-optimality, let ũ ∈ VN be arbitrary. We then have:
u − uN X = u − ũ − (uN − ũ)X ≤ u − ũX + uN − ũX .
There further holds:
uN − ũX = GN (u − ũ)X ≤ GN · u − ũX0 = Cu − ũX0
and thus by assumption 2. :
u − uN X ≤ u − ũX + uN − ũX ≤ u − ũX + Cu − ũX0

≤ C̃u − ũX0 + Cu − ũX0 = Ĉu − ũX0 ,
u − uN X ≤ Ĉ inf u − ũX0 .
ũ∈VN

For the rest of this section, we want to show that the assumptions of the theorem
are satisfied for the collocation method as described above.
For a smooth boundary (Γ ∈ C ∞ ), the operator V maps H s continuously and
bijectively onto H s+1 . Furthermore, we have already seen the relationship for the
Fourier coefficients:
>
um
(V u)m = . (6.12)
|m|
d as trial space V and S −1 as the testspace. For the equivalence of the

We choose SΔ N Δ=
Sobolev-norm and the abbreviated Sobolev-norm there are two constraints to the

spaces X = H s and Y = H s+1, resp. Y = H −s−1 , namely:
d
SΔ ⊂ H s for s < d + 1/2
−1
SΔ
= ⊂ H −s−1 for −s − 1 < − 1 + 1/2 ⇔ s > − 1/2 .
The dimension of the spaces SΔd and S −1 is in both cases N. We will now show that
Δ=
the five assumptions of the Theorem 6.11 hold:
−1
ad 1.: The operator PN : H −s−1 → SΔ = is the projection onto the space of
periodical splines, by density satisfying:
PN v − v → 0 ∀v ∈ H −s−1 .
ad 2.: We have X0 := X . However, it should be noted that for problems with

corners X and X0 will be different (see Sect. 6.10).
ad 3.: There trivially holds VN ⊆ X0 ⊂ X , since s < d + 12 .
We now have to find operators QN such that assumptions 4. and 5. will hold.
ad 4.: We have
|QN v, V w | ≤ QN vH −s−1 V wH s+1 ∀ v ∈ VN , ∀ w ∈ H s .
Making use of V wH s+1 ≤ C · wH s , we only have to show:
QN vH −s−1 ≤ CvH s ∀v ∈ XN .
Now, this especially holds if we have for the Fourier coefficients:

|m|−2s−2 |(QN v)m | ≤ C · |m| |v̂m |
2 2 2s 2
resp.
|(QN v)m | ≤ C|m|
2s+1 |v̂ | .
m
Hence, if we set (Q N v)m := |m|2s+1v̂m ∀ |m| ≤ M , and in particular

(Q N v)0 := 0, the operator QN is uniquely defined and satisfies the required
property 4.
ad 5.: To prove the last assumption we will show:
%(QN v, V v ) ≥ γ v2H s − compact perturbation

⇒ |QN v, V v | ≥ γ v2H s − compact perturbation :

QN v, V v =
(QN v)k (V v)k (by change of counting and (QN v)0 = 0)
k

M
k=m+lN
= −1
(QN v)m+lN (V v)m+lN (since QN v ∈ SΔ
= cf. 6.8)
m=−M l∈Z
m=0

M @
2πilε (v)m+lN
= (QN v)m e (by definition of QN and v ∈ SΔ
d
)
|m + lN|
m=−M l∈Z
m=0

M
1 m
= |m|2s+1v̂m e2πilε ( )d+1>
vm
|m + lN| m + lN
m=−M l∈Z
m=0

M |m| m
= |m|2s |v̂m |2 e2πilε ( )d+1
|m + lN| m + lN
m=−M l∈Z
m=0
⇒ |QN v, V v | ≥ v2H s − |v̂0 |2

m
|m| d+1

× min 1 + e2πilε m N N

−M≤m≤M | N + l| N
m
+l
l∈Z
l=0

Zεd ( N
m
)
To prove 5. for the compact perturbation |v̂0 |2 which is not depending on N, i.e.
with collectively compact sequence (CN )N , it only remains to show:
m
m
1 + Zεd ≥ γ > 0 for arbitrary (6.13)
N N
With N = 2M + 1 , the term m

N will take on values in the interval [−1/2, 1/2] .
For x ∈ [−1/2, 1/2] we have:
|x|
Zεd (x) = e2πilε |l+x| x d+1
( l+x )
l∈Z
l=0
∞ −∞ 1
= |x| · x d+1 1
e2πilε |l +1 x| (l+x) d+1 + e2πilε 1
|l + x| (l+x)d+1
l=1 l=−1
= (l+x)
( )= −(l+x)
∞ −∞
= x·x d+1
|x| x
1
e2πilε (l+x) d+2 −
1
e2πilε (l+x) d+2
l=1
∞
l=−1
∞
= x d+2
|x| x
1
e2πilε (l+x) d+2 ± e−2πilε 1
(l−x)d+2
l=1 l=1
∞
= −2πilε
d+2 ± e
x d+2 1 1
|x| x e2πilε (l+x) (l−x)d+2
,
l=1
with the ‘+’–sign for d being odd and the ‘−’–sign for d being even. Now, for both,
odd and even d, there holds
Zεd (−x) = Zεd (x) .

Therefore, with Zεd (0) = 0 , we only have to examine Zεd (x) for 0 < x ≤ 1/2 :
∞

Zεd (x) = x d+2 e2πilε (l + x)−d−2 ± e−2πilε (l − x)−d−2 .
l=1
Here, we only want to consider the cases ε = 0 and ε = 1/2 , i.e. the collocation
points are chosen to be either the grid points or the midpoints of the intervals.
i) d even, ε = 0:
∞

Z0d (x) = x d+2 (l + x)−d−2 − (l − x)−d−2
l=1
( ∞ ∞
)

=x d+2
(l + x)−d−2 − (l − x)−d−2 − x d+2 (1 − x)−d−2
l=1 l=2
( ∞ ∞
)

−d−2 −d−2
=x d+2
(l + x) − (l + 1 − x) − x d+2 (1 − x)−d−2
l=1 l=1
d+2 d+2 −d−2

Z0d (1/2) = 1
2 ·0− 1
2
1
2 = −1 .
Thus, for d even and ε = 0 (6.13) does not hold.

ii) d even, ε = 1/2:
( ∞ ∞
)

−d−2 −πil −d−2
Z1/2
d
(x) =x d+2
e πil
(l + x) − e (l − x)
l=1 l=1
(∞ ∞
)

−d−2 −πil −d−2
=x d+2
e πil
(l + x) − e (l − x) −x d+2e−πi (1−x)−d−2
l=1 l=2
(∞ ∞
)

−d−2 −πi(l+1) −d−2
=x d+2
e πil
(l+x) − e (l+1−x) +x d+2(1−x)−d−2
l=1 l=1
∞
d+2
x
= x d+2 (−1)l (l + x)−d−2 + (−1)l (l + 1 − x)−d−2 +
1−x
l=1
∞
d+2
−d−2 −d−2 x
=x d+2 l
(−1) (l+x) + (l + 1 − x) + ,
1−x
l=1 ≥0
making use of eπi = − 1 .

Thus, we have an alternating, monotonously decreasing series which is there-

fore convergent. Since the first term of the sum is of negative sign, there holds:
d+2
x
Z1/2
d
(x) ≤ ,
1−x
and further, taking into account this first term:

d+2
−d−2 −d−2 x
Z1/2
d
(x) ≥ −x d+2
(1 + x) −x d+2
(2 − x) +
1−x
= −x d+2 (1 + x)−d−2 + x d+2 (1 − x)−d−2 −x d+2(2 − x)−d−2

≥0 for 1/2≥x>0
≥ −x d+2 (2 − x)−d−2 .
1 2
d+2 d+2
x x
⇒ Zεd (x) ∈ − , ∀ x ∈ [−1/2, 1/2],
2−x 1−x
i.e. in particular we have

d+2 d+2
1/2 1
Zεd (x) ≤ 1 and Zεd (x) ≥ − = − > −1 .
2 − 1/2 3
Thus, to sum up, for even d the mid-point-collocation ( ε = 1/2 ) is convergent

whereas the break- (or grid-) point-collocation ( ε = 0 ) is not.
iii) d odd, ε = 1/2:
∞
∞ πil −d−2 +
Z1/2
d
(x) = x d+2 l=1 e (l + x) e−πil (l − x)−d−2
∞ ∞
l=1
= x d+2 eπil (l + x)−d−2 + e−πi(l+1) (l + 1 − x)−d−2

l=1 l=1
+x d+2e−πi (1 − x)−d−2 (∞ )
∞
d+2
⇒ Z1/2
d (1/2)= 1
2 (−1)l (l+1/2)−d−2− (−1)l (l+1/2)−d−2
l=1 l=1

=0
d+2 −d−2
+ 1
2 (−1) 1
2 = −1.
Thus, for ε = 1/2 we do not attain convergence.

iv) d odd, ε = 0:
( ∞ ∞
)

Z0d (x) =x d+2
(l + x)−d−2 + (l − x)−d−2 ≥ γ > 0 ∀x > 0 .
l=1 l=1
Thus, for odd d the mid-point-collocation ( ε = 1/2 ) is not convergent whereas

the break-point-collocation ( ε = 0 ) is convergent.
Now, all four cases can be summarized as follows:
The ε-collocation with spline functions of degree d for V v = f converges in
the space H s (Γ ) with s ∈ (−1/2, d + 1/2) , if for
1. d even the condition ε = 1/2
2. d odd the condition ε = 0
is satisfied. This result can be generalized to the following (proof omitted):
The ε-collocation with spline functions of degree d for V v = f converges in
the space H s (Γ ) with s ∈ [−1, d + 1/2] if and only if there holds for
1. d even the condition ε = 0 ,
2. d odd the condition ε = 1/2 .
Making use of the convergence estimates deduced from the quasi-optimality
property
u − vN ≤ C · inf {u − v} ,

v∈XN
i.e. for d being the polynomial degree of the splines and h the grid-size we have
u−uN 0 ≤ C·hd+1 ud+1 resp. u−uN r ≤ C·hd−r+1 ud+1 , −1 ≤ r ≤ d +1
with ud+1 ≤ C · f d+2 Here we used the continuity of the inverse
us = V −1 f s ≤ C · f s+1
and u − uN r ≤ C · hd−r+1 f d+2 .

The highest order of convergence in the H −1 norm is given by:
u − uN −1 ≤ C · hd+2 f d+2 .
At the end of this section we briefly look at second kind equations. Banach
algebra techniques play a dominant role in the convergence analysis of numerical
methods for second kind integral equations on curves with corners and singular
integral equations with discontinuous coefficients see S. Prössdorf and A. Rathsfeld
[344]. Chapter 7 of the book [345] by S. Prössdorf and B. Silbermann gives a good
introduction to and demonstration of the power of Banach algebra techniques in
numerical analysis. The paper [122] is concerned with approximation methods for
Neumann’s integral equation
(I − K)u = f on Γ
with the double layer potential

1 ∂
Ku(x) = − u(y) ln |x − y|dsy
π Γ ∂ny
on curves Γ with corners. In [122] necessary and sufficient conditions for the
stability of the piecewise constant − collocation and for the quadrature method,
using the rectangular rule, are given using Banach algebra techniques together with
Mellin-techniques as introduced in Chap. 9.
6.5 BEM on Quasiuniform Meshes
In Sects. 6.5.1–6.5.5 we follow [430].
6.5.1 Periodic Polynomial Splines
Let Δ = {tk }k∈Z be a partition of R with grid points tk , with t0 = 0 and tk+N = tk +1
for a fixed N ∈ N and for all k ∈ Z, i.e. {tk }N k=0 is a partition on [0, 1], which is
extended 1-periodically.
h := max{tk+1 − tk }
is called the mesh size.

For simplicity we consider 1-periodic smoothest splines of degree d :
S d (Δ) (d ∈ N0 ), i.e. Φ ∈ S d (Δ) ⇐⇒ ϕ with all derivatives up to order
≤ d − 1 is 1-periodic and continuous on R and ϕ|(tk , tk+1 ) is a polynomial at most
of degree d (∀k ∈ Z).
For any d ∈ N0 S d (Δ) is a N-dimensional space and has as basis {Bj,d }N−1 j =0 ,
the B-Splines due to de Boor, which are defined recursively as follows:
Let Qj be the characteristic function of [tj , tj +1 ). Definier Bj,0 (j ∈ Z) by
Bj,0 (t) = Qj (t) für t ∈ [tj , tj +N )
and extend Bj,0 1-periodically on R .

Then for d ≥ 1, j ∈ Z
t − tj tj +d+1 − t
Bj,d (t) = Bj,d−1 (t) + Bj +1,d−1 (t)
tj +1 − tj tj +d+1 − tj +1
for t ∈ [tj , tj +N ) (hence for support supp (Bj,d ) ⊂ [tj , tj +d+1 ]) and extend Bj,d
1-periodically on R.
There holds S d (Δ) ⊂ H s (R) ⇐⇒ s < d + 12 .
6.5 BEM on Quasiuniform Meshes 141
6.5.2 The Approximation Theorem
Theorem 6.12 Let
− ∞ < s ≤ r ≤ d + 1, s < d + 12 . (6.14)
Then there esists a constant C = C(r, s, d) and to any u ∈ H r and to any partition
Δ there exists a ϕ ∈ S d (Δ) such that
u − ϕs ≤ Chr−s ur (6.15)
For the proof we need the interpolation theorem

A: Let the linear operator L : H σ → H j be continuous with the operator norm
Lσ,j , as well as Lσ,j −1 for L : H σ → H j −1 (also continuous due to
embedding).
Then there holds for ∈ [j − 1, j ] and L : H σ → H
Lσ, ≤ (Lσ,j −1 )j − (Lσ,j ) −j +1

(6.16)
(log-convexity, j − ∈ [0, 1] und = (j − )(j − 1) + ( − j + 1)j )

B: L : H j −1 → H be bounded (and linear) with operator norm Lj −1, , as
well as Lj, for L : H j → H . Then there holds for j − 1 ≤ σ ≤ j and
L : Hσ → H
Lσ, ≤ (Lj −1, )j −σ (Lj, )σ −j +1 (6.17)
Proof (Sketch, More Detailed in [276]) A. follows from the definition of the
Sobolev spaces via Fourier series and from the Hölder inequality. B. follows from
A. by use of adjoint operators:
L∗ : H − → H 1−j resp. L∗ : H − → H −j
with
L∗ − ,1−j = Lj −1, , L∗ − ,−j = Lj, .

Proof (Approximation Theorem)
1. d = 0 = s, r = 1
We apply the equivalent norms
u2σ = u20 + u(σ ) 20 if σ ∈ N

1 1
|u(m) (t) − u(m) (τ )|
u2σ = u20 + dt dτ
|t − τ |1+2μ
0 0
for σ = m + μ, m ∈ N0 , 0 < μ < 1.

Let u ∈ H 1 . Set for k = 0, 1, . . . , N − 1
tk+1
1
ϕ(t) := u(τ ) dτ für t ∈ [tk , tk+1 )
tk+1 − tk
tk
Now with Cauchy Schwarz inequality for t ∈ (0, 1)

2
tk+1

N−1
1
|u(t) − ϕ(t)|2 = Qk (t) 1[u(t) − u(τ )] dτ
k=0 tk+1 − tk
tk

N−1 tk+1
≤ Qk (t) h−1
k |u(t) − u(τ )|2 dτ (6.18)
k=0 tk
Now we estimate further for t, τ ∈ (tk , tk+1 )

t 2

|u(t) − u(τ )| = u (σ ) dσ
2
τ
tk+1
≤ |t − τ | |u |2 dσ
tk
tk+1
≤ hk |u (σ )|2 dσ (6.19)
tk
Now integration yields

tk+1 tk+1
|u(t) − u(τ )|2 dτ ≤ h2k |u |2 dτ
tk tk
and inserting in (6.18) and integration gives

1
1 N−1 tk+1
u − ϕ20 = |u(t) − ϕ(t)|2 dt ≤ Qk (t)hk |u |2 dσ dt
0 0 k=0 tk

N−1 tk+1 1
≤ h2k |u |2 dτ ≤ h2 |u |2 dτ (6.20)
k=0 tk 0
2. d = 0 < s < 12 , r = 1
In this case (6.15) follows from the estimate that we show next
1 1
J := |u(t) − ϕ(t) − (u(τ ) − ϕ(τ ))|2 |t − τ |−1−2s dt dτ
0 0
≤ c h2−2s u 20
Here and in the following c denotes different positive constants constants which
are independent of u and Δ.
It is

J = Jk,l
0≤k,l<N
where Jk,l is the corresponding integral on (tk , tk+1 ) × (tl , tl+1 ).

With (6.19) there follows (due to d = 0 ⇒ ϕ(t) = ϕ(τ ) for t, τ ∈ (tk , tk+1 ))

Jk,k = |u(t) − u(τ )|2 |t − τ |−1−2s dt dτ
Ikk
tk+1
≤ |u |2 dτ |t − τ |−2s dt dτ
tk Ikk
tk+1
= (tk+1 − tk ) 2−2s
|t˜ − τ̃ |−2s d t˜ d τ̃ |u |2 dτ
(0,1)2 tk
tk+1
≤ ch 2−2s
|u |2 dτ (6.21)
tk
With(6.18) and (6.19) there follows for k = l due to (|A| + |B|)2 ≤ 2(|A|2 +
|B|2 )
& k+1 l+1 |u(t) − ϕ(t)|2 tk+1 tl+1

t t
|u(τ ) − ϕ(τ )|2 '
Jkl ≤ 2 dt dτ + dt dτ
|t − τ |1+2s |t − τ |1+2s
tk tl tk tl
& k+1 tl+1 ' k+1 l+1

t t t
2 2 dt dτ
≤ 2h |u | dτ + |u | dτ
|t − τ |1+2s
tk tl tk tl
Now summation gives (due to symmetry)

N−1
N−1 tk+1
Jkl ≤ 8h ckl |u |2 dτ (6.22)
k=l k=0 l=k+1 tk
where
tk+1 tl+1
dt dτ
ckl =
(τ − t)1+2s
tk tl
tl+1
1 tk+1
−2s

=− (τ − t) dt
2s tk τ =tl
1 &
= (tl+1 − tk+1 )1−2s − (tl+1 − tk )1−2s
(1 − 2s)2s
'
−(tl − tk+1 )1−2s + (tl − tk )1−2s
Hence with cs = 1
(1−2s)2s

N−1 & '
ckl = cs (tN − tk+1 )1−2s − (tN − tk )1−2s + (tk+1 − tk )1−2s
l=k+1
≤ cs (tk+1 − tk )1−2s ≤ cs h1−2s
which leads with (6.21) and (6.22) to the asserted estimate.

3. r = d + 1, s ∈ [d, d + 12 )
Introduce the following spaces
◦
1
H = {u ∈ H :
s s
u dt = 0} (s ≥ 0),
0
◦d ◦
S (Δ) := S d (Δ) ∩ H s
Then
◦ ◦
H s = H s ⊕ C, S d (Δ) = S d ⊕ C,
⎧ s+d
◦
⎨H ◦
m → Hs
D d := d
: ◦
dt ⎩ ◦d
S (Δ) → S 0 (Δ)
◦ d+1 ◦
Now it suffices to show that ∀u ∈ H ∃ψ ∈ S 0 (Δ) such that
D d u − ψs−d ≤ c hd+1−s D d u1
With s − d ∈ [0, 1/2) this follows from 1. and 2. .

4. The case of an integer: s ≤ r ≤ d + 1, s ≤ d, (s, r ∈ Z)
Since the case s = r is trivial, it remains to show the assertion for s < r what
we do inductively for s = d − k (k ∈ N0 ).
k = 0, i.e. s = d, r = d + 1 is contained in 3.
induction step from k → k + 1:
We have to show the assertion for s = d − k − 1 = j − 1 with j := d − k. For
this we use the orthogonal projection
P ,Δ : H → S d (Δ)
with
(∀u ∈ H , ϕ ∈ S d (Δ)) (u − P ,Δ u, ϕ) =0
⇐⇒ u − P ,Δ us = min{u − ϕs : ϕ ∈ S d (Δ)}
and obtain
u − Pj −1,Δ uj −1 ≤ u − Pj,Δ uj −1

& '
= sup |(ψ, u − Pj,Δ u)j | : ψj +1 ≤ 1
& '
= sup |(ψ − Pj,Δ ψ, u − Pj,Δ u)j | : ψj +1 ≤ 1
& '
= sup ψ − Pj,Δ ψj u − Pj,Δ uj : ψj +1 ≤ 1
≤ c h ψj +1 · c hr−j ur

≤ c hr−(j −1) ur (6.23)
hence the assertion of the induction.

5. Completion of the proof by interpolation
i) With 4. and (6.23) we obtain for j = r (trivial), j = r − 1, r − 2, . . . ,
r = d + 1, d, . . . , that
I − Pj,Δ : H r → H j −1
is continuous with norm ≤ c hr+1−j and
I − Pj,Δ : H r → H j
is continuous with norm ≤ c hr−j

Now Interpolation Theorem Part A gives: For s ∈ (j − 1, j ), r ≤ d and
r ∈ Z with s ≤ r there holds
I − Pj, r,s ≤ (c hr+1−j )j −s (c hr−j )s−j +1

≤ c hj −s+r−j = c hr
where
I − Pj, : H r → H s ,
also
I : H r → H s,
hence also Pj, , i.e. for r ∈ Z and s ∈ R with s ≤ r and s ≤ d there holds
∀u ∈ H r ∃ϕ := Pj, u ∈ S d (Δ) such that u − ϕs ≤ c hr−s ur

(6.24)
ii) Case r ∈
/ Z, due to assumption r ∈ [s, d + 1]
Consider
Ps,Δ : H s → S d (Δ) ⊂ H s
due to s ≤ d < d + 12 . Now
I − Ps,Δ : H s → H s
is continuous with norm ≤ 1 and
I − Ps,Δ : H d+1 → H s
is continuous with norm ≤ c hd+1−s by (6.24)

Interpolation due to B. gives for r = r−d−1
s−d−1 s +
s−r
s−d−1 (d + 1)
I − Ps,Δ : H r → H s
s−r
is continuous with norm ≤ 1(r−d−1/s−d−1) (chd+1−s ) s−d−1 = c hr−s
Finally
iii) Case d < s < d + 12 , r ∈ [s, d + 1]
Consider again
I − Ps,Δ : H s → H s
and
I − Ps,Δ : H d+1 → H s
is continuous with norm ≤ c hd+1−s by 3.

Finally interpolation by B. analogously to (ii) yields the assertion.

6.5.3 Stability and Inverse Estimates
In this section we consider in more detail 1-periodic B-splines as basis for S d (Δ)
for special families of grids Δ n.̈ We find a norm which is equivalent to the L2 -norm
on S d (Δ) where the stability constants are independent of the mesh width. Further
we show for the trial space S d (Δ) the so-called inverse property, which is important
for the convergence
& ' analysis.
Let Δ = Nj be the family of equidistant meshes with mesh width h =
j ∈Z
1
N (N ∈ N).
For fixed degree d ∈ N0 of trial functions we introduce
⎧
⎪
⎪d = 0 d = 1 d=2 for
⎪
⎪
⎪
⎨ 1 τ 1 2
0 ≤ τ < 1,
2τ
μ(τ ) = 0 2 − τ −τ + 3τ − 32
2 1 ≤ τ < 2, (6.25)
⎪
⎪
⎪
⎪ 0 0 12 τ 2 − 3τ + 92 2 ≤ τ < 3,
⎪
⎩
0 0 0 else
[compare for example: d = 2 : μ = B2,0 with h = 1, τi = ih(i = 0, . . . , 3)] where

there holds

μ(τ )dτ = 1. (6.26)
R
Further we set for l ∈ Z, N ≥ 1 + d

t
μl (t) := μ −l+1 for h(l − 1) ≤ t ≤ h(l + d), (6.27)
h
μl (t + k) = μl (t) for k ∈ Z; (6.28)

i.e. the functions μl are built from the function μ by squeezing by the factor h1 (
dilatation ), together with shifting of l − 1 ( translation ) on the t-axis and 1-periodic
extension.
Theorem 6.13 The splines μl (l = 1, . . . , N) defined in (6.25)–(6.28) build a basis
of S d (Δ). Furthermore for all

N
ϕ= γj μj (γj ∈ R)
j =1
there holds

N
N
N
c1 h γj2 ≤ γj μj 2L2 (0,1) ≤ c2 h γj2 (6.29)
j =1 j =1 j =1
with constants 0 < c1 < c2 independent of h = 1

N and γj .
Proof For the first assertion and the proof of (6.29) we consider the symmetric Gram
matrix
Mj k = Mkj := h1 (μj , μk )L2 (0,1)

1
= h1 μ ht − j + 1 μ ht − k + 1 dt
0
N−j +1
= μ(σ )μ(σ − k + 1) dσ.
−j +1
We find:
d = 0 : Mj k = δj k ;
d = 1 : Mjj = 23 , Mj,j +1 = 16 , zero else;
d = 2 : Mjj = 17
20 , Mj,j +1 = 60 , Mj,j +2 =
13 1
30 , zero else;
For these values there holds

Mkj Mj k
= = q < 1, (j = 1, . . . , N) (6.30)
Mjj Mjj
k=j k=j
i.e. the Gram matrix M is strongly

0 diagonally
0 dominant. Consequently M is regular,
0 0
since for the row-sum-norm 0 M111 M − I 0 = q < 1 .
∞
With the help of (6.30) we estmate

0 02
0 0
0N 0 N
0 γ μ 0 =h γj γk Mjk
0 j j0
0j =1 0 2 j,k=1
L (0,1)
⎧ ⎫
⎪
⎪ ⎪
⎪
⎨N
N
N ⎬
=h γj2 Mjj + γj γk Mj k
⎪
⎪ ⎪
⎪
⎩j =1 j =1 k=1 ⎭
k=1
⎧ ⎫
⎨N Mj k ⎬
= hM11 γj2 + γj γk
⎩ Mjj ⎭
j =1 j =k
from above by
⎧ ⎫
⎪
⎪ ⎪
⎪
⎨N
1
N
N
Mj k 1
N
N
Mj k⎬
≤ hM11 γj2 + γj2 + γk2
⎪
⎪ 2 Mjj 2 Mkk ⎪⎪
⎩j =1 j =1 k=1 k=1 j=1 ⎭
k=j j =k

N
= hM11 (1 + q) γj2
j =1
and from below by

⎧ ⎫
⎨N 1 M ⎬
j k
≥ hM11 γj2 − (γj2 + γk2 )
⎩ 2 Mjj ⎭
j =1 j =k

N
= hM11 (1 − q) γj2 ,
j =1
which shows (6.29).

Theorem 6.14 (Inverse Property) For arbitrary r, s ∈ R with
1
r≤s<d+
2
there exists a constant c = c(r, s), such that for all h = 1

N and for all ϕ ∈ S d (Δ)
there holds
ϕs ≤ chr−s ϕr . (6.31)

Remark Again S d (Δ) ⊂ H s implies the assumption s < d + 12 .

Proof
1. Step: We show (6.31) for d < r ≤ s < d + 21 . Since D d ϕ ∈ S 0 (Δ) for ϕ ∈ S d (Δ)
it suffices to consider the case d = 0. Using the equivalent Sobolev-Slobodecki-
Norm gives then with r = d + ρ, s = d + σ, 0 < ρ ≤ σ < 12 ; h ≤ 1
ϕ2s = D d ϕ2σ + D d ϕ20

≤ ch2(ρ−σ ) D d ϕ2ρ + D d ϕ20
≤ ch2(r−s)ϕ2r .
For 0 < ρ ≤ σ < 1/2 we have to show for arbitrary h, ϕ ∈ S 0 (Δ)
Jσ (ϕ) ≤ ch2(ρ−σ )Jρ (ϕ),
where
1 1
Jσ (ϕ) := |ϕ(x) − ϕ(y)|2 |x − y|−1−2σ dx dy.
0 0

N
With the representation ϕ = γl μ(0)
l we find
l=1
lh kh
|γl − γk |2
Jσ (ϕ) = Jlk;σ , Jlk;σ := dx dy.
|x − y|1+2σ
1≤l,k≤N (l−1)h (k−.1)h
Especially there holds Jll;σ = 0, while k = l + 1
|γl+1 − γl |2 & '

Jl,l+1;σ = 2 − 21−2σ h1−2σ ,
(1 − 2σ )2σ
consequently
Jl,l+1;σ
= c(σ, ρ)h−2(σ −ρ) .
Jl,l+1;ρ
For the remaining |l − k| ≥ 2 we estimate as
lk kh
|γl −γk |2 dx dy
Jlk;σ = |x−y|1+2ρ |x−y|2(σ −ρ)
(l−1)h (k−1)h
≤ h2(ρ−σ ) Jlk;ρ .
Hence (6.31) is shown in this case.

2. Step: We extend the validility of (6.31) to arbitrary r ≤ s with the help of

interpolation.
First let d < r < s < d + 12 . With r = 12 s + 12 (2r − s), 2r − s < s follows (
by direct application of the Hölder inequality or with the interpolation theorem
applied to the embedding H s ⊂ H 2r−s )
1/2 1/2
ϕr ≤ ϕs ϕ2r−s ,
consequently with the inequality (6.31) shown in the 1. step
ϕ2r ≤ chr−s ϕr ϕ2r−s .
The inequality ϕs ≤ chr−s ϕr once more applied yields ϕs ≤
ch2(r−s)ϕ2r−s . By induction we obtain
ϕs ≤ ch(k+1)(r−s)ϕr+k(r−s) (k ∈ N).
Hence there holds (6.31) for s ∈ d, d + 1

2 and arbitrary r ≤ s. Is s ≤ d, then
we choose a t ∈ d, d + 1
2 , interpolate
t −s s−r
s= r+ t
t −r t −r
and obtain
(t −s)/(t −r) (s−r)/(t −r)
ϕs ≤ ϕr ϕt
(t −s)/(t −r) (r−t )(s−r)/(t −r) (s−r)/(t −r)
≤ cϕs h ϕr
= ch ϕr .
r−s

Remark 6.3 Both previous theorems can be extended to quasiuniform meshes,
where there exists a constant γ ≥ 1 such that γ −1 h ≤ tk+1 − tk ≤ h for all grid
points th (see [345] ).
6.5.4 Aubin-Nitsche Duality Estimate and Superapproximation
Now we improve the error estimates with the help of the orthogonality of the
Galerkin scheme
∀ χ ∈ S m () : (Auh , χ)L2 = (f, χ)L2 (6.32)
Let A∗ : H α → H −α be the adjoint operator to A.

Lemma 6.1 ([258]) Let u solve Au = f and uh be the corresponding Galerkin

solution and
1
ρ < m + , −m − 1 + 2α ≤ ρ ≤ α (6.33)
2
and A∗ : H α → H −α continuous und A∗ : H 2α−ρ → H −ρ continuous and

bijective. Then there holds the error estimate
u − uh ρ ≤ c · hα−ρ u − uh α . (6.34)
Proof
u − un ρ = sup |(u − uh , ϕ)L2 |

ϕ−ρ ≤1
= sup |(u − uh , A∗ w)L2 |
ϕ−ρ ≤1
where
−1
w = A∗ ϕ ∈ & H
2α−ρ , w
2α−ρ ≤ ϕ−ρ '

u − uh ρ = sup u − uh , A (w − χ) 2 + (u − uh , A∗ χ)L2
∗
ϕ−ρ ≤1 L

= sup inf u − uh , A∗ (w − χ) 2 + A(u − uh ), χ 2

ϕ−ρ ≤1 χ∈Hh L L
≤ sup inf cu − uh α A∗ (w − χ)−α
ϕ−ρ ≤1 χ∈Hh
≤ sup inf cu − uh α w − χα
ϕ−ρ ≤1 χ∈Hh
≤ c hα−ρ sup w2α−ρ u − uh α
ϕ−ρ ≤1
≤ c hα−ρ sup ϕ−ρ u − uh α
ϕ−ρ ≤1

Theorem 6.15 Let A : H α → H −α be continuous, bijective and H α -coercive
without compact perturbation. A∗ : H 2α−ρ → H −ρ be continuous and bijective
with −m − 1 + 2α ≤ ρ ≤ α ≤ σ ≤ m + 1, α < m + 12 . Then there exists a h0 > 0,
such that for all h mit 0 < h ≤ h0 the Galerkin equations (6.32) are uniquely
solvable and there holds
u − uh ρ ≤ chσ −ρ uσ (6.35)
with a constant c, which does not depend on h, u, uh .

Proof The assumption of Lemma 6.1 are all satisfied, since with A also A∗ : H α →
H −α is continuous. Then (6.34) gives the assertion.

Example 6.3 Application of the Galerkin scheme to Symm’s integral equation:

1
Au := − u(y) log |x − y| dsy = f (x) on Γ
π
Γ
under the assumption diam (Γ ) < 1 gives due to α = − 12 the best order of
convergence for ρ = −m − 1 − 1, σ = m + 1 :
u − uh −m−2 ≤ ch2m+3 um+1
The Galerkin scheme for the Fredholm integral equation of the second kind

1 ∂
Au := u(x) − u(y) (log(x − y)) dsy = f (x) auf Γ
π ∂γx
Γ {x}
gives due to α = 0 the highest convergence order
u − uh −m−1 ≤ ch2m+2 um+1 .
If S m (Δ) has the inverse property, then (6.35) can be extended to the indices α <
ρ ≤ σ.
Theorem 6.16 Let S m (Δ) have the inverse property (6.31). Then for the Galerkin
solutions there holds additionally under the assumptions of Theorem 6.15
u − uh ρ ≤ chσ −ρ uσ (6.36)
for α < ρ ≤ σ ≤ m + 1 and ρ < m + 12 .

Proof Let ũΔ the best approximation of u due to Theorem 6.12 with ρ = −m − 2 +
2α. Then there holds (6.34) and (6.35)
u − uh ρ ≤ u − ũΔ ρ + ũΔ − uh ρ
≤ chσ −ρ uσ + Mhα−ρ ũΔ − uh α
≤ chσ −ρ uσ + Mhα−ρ ũΔ − uα + Mhα−ρ u − uh α
≤ chσ −ρ + cMhα−ρ+σ −α uσ + c Mhα−ρ+σ −α uσ

In the proof of the following lemma we use the Aubin-Nitsche duality estimate.
Lemma 6.2 (Superapproximation Property) Let ω ∈ C0∞ (I0 ) for any interval
I0 and t0 < d + 12 with −q ≤ t ≤ t0 , s ≤ d ,. Then there exists for any spline
I0
a b
I0
I1
ψ ∈ S d a spline ξ ∈ S d such that
ωψ − ξ H t (Γ ) ≤ c · hs+1−t ψH s (I ) .

0
Proof We only want to consider the case d = 0 . For d = 0 the proof is left to the
reader as an exercise.
First, the intervalls I0 , I0 , I1 and I shall be given according to Fig. 6.2.
Then, for ω ∈ C0∞ (I0 ) there exists a η ∈ S 0 such that
v := ωψ + η ∈ H 1 (I )
with η|I \I ∈ P0 being a polynomial of degree 0 . Defining the integral operator

0
x
D −1 f (x) := f (τ ) dτ for a < x < b , I = (a, b)
a
we have
w := D −q v ∈ H 1+q with w|I \I0 ∈ Pq .
Fixing now I2 such that I1 ⊂⊂ I2 ⊂⊂ I we have
∃ ξ ∈ S q (I ) : D k (w − ξ )L2 (I ) ≤ c · hq+1−k D q+1 wL2 (I ) ,
with ξ = w in I2 \I1 . If we define now

−(η(x) − ϕ(x)) , x ∈ I2
ϕ := D q ξ and ρ ∈ S 0 , given by ρ(x) = ,
0 , x ∈ I \I2
we have in the interval I2 :
w −
ωψ − ρ = D q
D q ξ = ωψ + η − ϕ
=v =ϕ
Thus,
D q (w−ξ ),f L2 (I2 )
ωψ − ρ−q(I2 ) := sup f q(I2 )
q
0=f ∈H0 (I2 )
(w−ξ ),D q f L2 (I2 )
= sup f q(I2 )
q
0=f ∈H0 (I2 )
≤ w − ξ L2 (I2 ) ≤ c · hq+1 D q+1 wL2 (I2 )
≤ c · hq+1 ψL2 (I ) ,
0
since D q+1 w = v = (ωψ + η) = ω ψ .

6.5.5 Numerical Quadrature
Projection and variational methods like Galerkin and collokation methods do not
only lead to the error by the method as analyzed above but to further errors given by
the numerical quadratures for the boundary and domain integrals involved, as well
as due to the approximation of a curved boundary by a polygon. The effect of such
variational crimes on the error asymptotic can be analysed similarly as in the finite
element method.
In this section we restrict ourselves to error estimates of the numerical quadrature
for boundary element Galerkin schemes on families Δ of quasiuniform meshes.
Starting point for the following error analysis is the following theorem on the
condition of the Galerkin method. Here let A be a pseudodifferential operator of
order 2α.
Theorem 6.17 Let Δ be quasiuniform and let all assumptions of Theorem 6.15 be
satisfied and additionally assume A : H 2α (Γ ) → L2 (Γ ) is continuous. Then for
the Galerkin solutions uh ∈ S d (Δ) with right hand side f ∈ S d (Δ) there holds

uh 0 ≤ ch2α f 0 and f 0 ≤ ch2α −2α uh 0 , (6.37)
where α = min{0, α}. Consequently the condition of the Galerkin equations is of

order O(h−2|α| ).
Proof i) Case α ≤ 0. Then (6.35) gives with d ≥ 0
u − uh α ≤ cuα ,
uh α ≤ (1 + c)uα ≤ c̃f −α .
Due to Thorem 6.15 and Remark 6.3 follows
uh 0 ≤ chα uh α ≤ c hα f −α ≤ c h2α f 0 .

On the other hand for f = Ph Auh there holds with L2 -orthogonal projection Ph
f 0 ≤ Auh 0 ≤ cuh 2α ≤ cuh 0 .
ii) Case α ≥ 0. Now we estimate
uh 0 ≤ uh α ≤ (1 + c)uα ≤ c̃ f −α ≤ c̃ f 0 ,

f 0 ≤ Auh 0 ≤ c uh 2α ≤ ch−2α uh 0 .
Hence we obtain in both cases (6.37).

Next we start from the numerically computed values ãjk for the entries (Aμj , μk )
of the Galerkin matrix and from the approximate values f˜k for the right hand sides
(f, μk ) of the Galerkin equations. We assume that
|(Aμj , μk )L2 − ãj k | ≤ chR , (6.38)

R
|(f, μk )L2 − f˜k | ≤ ch .
Lemma 6.3 Under the assumptions (6.38) there hold on S d (Δ) the estimates
A − ÃL2 ,L2 ≤ chR−2 , (6.39)

f − f˜L2 ≤ chR −1 ,
where Ã : S d (Δ) → S d (Δ) and f˜ ∈ S d (Δ) are defined by

N
N
(Ãw, v)L2 = λj ãj k ρk , (f˜, w)L2 = λj f˜j
j,k=1 j =1

N
N
for all w = λj μj und v = ρk μk .
j =1 k=1
Proof i) Estimates for the L2 − L2 operator norm:

3 4
|(A − Ã)w, v)L2 | = | λj ρk (Aμj , μk )L2 − ãj k |
j,k
⎧ ⎫1/2
⎨ ⎬
≤ |λj |2 · |ρk |2 · |(Aμj , μk )L2 − ãj k |2
⎩ ⎭
j k j,k
with (6.29) and (6.38)

"
1
≤c · wL2 vL2 · hR · {N 2 }1/2 = chR−2 wL2 vL2 .
h
ii) The estimate for the L2 vector norm follows analoguously:

N
|(f − f˜, w)L2 | = | λj {(f, μj )L2 − f˜j }|
j =1
⎧ ⎫1/2
⎨N
N ⎬
≤ |λj |2 · |(f, μj )L2 − f˜j |2
⎩ ⎭
j =1 j =1
R −1
≤ 1
c h1/2 wL2 hR · N 1/2 = ch wL2 .

Lemma 6.4 (Second Strang Lemma) Let (6.38) be satisfied and R −2+2α > 0.
Then there exists a h0 > 0, such that the equations

N
γ̃j ãj k = f˜j
j =1

N
are uniquely solvable for any h ∈ (0, h0 ), and for ũk = γ̃j μj there hold the
j =1
asymptotic error estimates

ũh − uh L2 ≤ (c1 hR −1 + c2 hR−2 uh L2 )h2α . (6.40)
Proof 1. Step: There holds
Ã = Ph APh + (Ã − Ph APh )

= Ph APh {I + (Ph APh )−1 (Ã − Ph APh )}.
From (6.37) and (6.39) follows

(Ph APh )−1 (Ã − Ph APh )L2 ,L2 ≤ ch2α · hR−2 → 0
for h → 0. Consequently Ã−1 can be represented as a convergent Neumann

series for sufficiently small h > 0. This gives

Ã−1 L2 ,L2 ≤ c (Ph APh )−1 L2 ,L2 ≤ c h2α (6.41)
2. Step: From
Ph APh uh = Ph f , Ãũh = f˜
we conclude
Ã(ũh − uh ) = f˜ − Ph f + (Ph APh − Ã)uh ,
and with (6.41) and with (6.40)

ũh − uh L2 ≤ c h2α {c1 hR −1 + c2 hR−2 uh L2 }.

Now we can collect the above error estimates in the following result.
Theorem 6.18 Let the assumptions of Theorem 6.15 hold. Furthermore let S d (Δ)
have the inverse property and let R − 2 + 2α > 0. Then there holds for
1 1
−d − 1 + 2α ≤ ρ ≤ σ ≤ d + 1, ρ < d + , α < d +
2 2
the asymptotic error estimate including quadrature error

u− ũh ρ ≤ c0 hσ −ρ uσ +c1 hR −1+2α +(−ρ) +c2 hR−2+2α −ρ uα + , (6.42)
where α + = max{α, 0}, where R as in Lemma 6.3

Proof First we obtain with Theorem 6.15 and 6.16 and the inverse property,
according to the sign of ρ
u − ũh ρ ≤ u − uh ρ + uh − ũh ρ

≤ c0 hσ −ρ uσ + c̃h(−ρ) uh − ũh L2 ,
where analogously to above (−ρ) = min{0, −ρ}. This gives with Lemma 6.4

u − ũh ρ ≤ c0 hσ −ρ uσ + c1 hR −1+2α +(−ρ)

+c2 hR−2+2α +(−ρ) uh L2 .
It only remains to estimate uh 0 . In case α ≥ 0 we apply Theorem 6.15 with

σ =ρ=α
uh 0 ≤ uh α ≤ u − uh α + uα ≤ (c + 1)uα .
On the other hand in case α < 0 application of Theorem 6.16 gives due to 0 < d + 12
with σ = ρ = 0
uh 0 ≤ u − uh 0 + u0 ≤ (c + 1)u0 .

6.5.6 Local H −1/2 -Error Estimates
We consider the problem
L (u − uh ) , χ L2 (Γ ) = 0 for χ ∈ Sh ,
where L = V denotes the single-layer potential on a smooth boundary Γ and

uh ∈ Sh the Galerkin approximation of u. We first want to show that for
1
e := u − uh , I0 ⊂⊂ I0 ⊆ Γ and s=k+ , k∈Z
2
there holds the estimate in the energy-norm:
"
e− 1 ,I0 ≤ c · min u − χ− 1 ,I + eH −s (Γ ) (6.43)
2 χ∈Sh 2 0
The derivation of (6.43) is motivated by the approach for the FEM Galerkin solution
of the Laplace in [363]
First we note
Lv 1 ≤ c · v− 1 (6.44)

2 2
Lv, v ∼
= v2− 1 , (with cap(Γ ) < 1) (6.45)
2
and
3
L∗ ψ = ϕ : ψ− 1 +l ≤ c · ϕ 1 +l , l = 1, 2, . . . , s − (6.46)
2 2 2
We consider a localisisation I0 ⊂⊂ I1 ⊂⊂ I2 ⊂⊂ I3 ⊂⊂ I4 ⊂⊂ I0 ⊂ Γ with
ω≡1 auf I0 , supp(ω) ⊆ I1

ω̃ ≡ 1 auf I3 , supp(ω̃) ⊆ I4 ,
for ω, ω̃ ∈ C ∞ , such that
(ωL − Lω)(ω̃v) 1 −l ≤ c · ω̃v− 3 −l (6.47)

2 2
L((1 − ω̃)v) 1 ,I2 ≤ c · v−s , (6.48)

2
for supp (1 − ω̃) ∩ I2 = ∅ . The approximation is given by:

1
∀ v ∈ H 2 (Γ ) ∃ χh ∈ Sh with supp (χh ) ⊆ I2 , such that
ωv − χh − 1 ≤ c · hv 1 ,I1 . (6.49)

2 2
For vh ∈ Sh with supp (vh ) ⊂ I2 , we have (superapproximation):
ωvh − χh − 1 ≤ c · hvh − 1 ,I2 . (6.50)

2 2
It should be noted that we may take for χh the local L2 -Projection. Eventually, if
we assume that I0 is a part of the mesh, there holds the inverse property
χ0,I0 ≤ c · h−k χ−s,I0 , ∀ χ ∈ Sh . (6.51)
Remark 6.4 All the above approximation assumptions are local. Since in (6.51)
k = s is possible, we even need not to have a quasiuniform grid; we only need
β
hmax,loc ≤ hmin,loc for a fixed β > 0 .
We now want to prove the assertion (6.43):
Let uh be the Galerkin solution, i.e.
u → uh ∈ Sh with L (u − uh ) , χ = 0 ∀ χ ∈ Sh
Then we have
e− 1 ,I0 = u − uh − 1 ,I0 ≤ ωe− 1

2 2 2
≤ ωu − (ωu)h − 1 + (ωu)h − ωuh − 1 .

2

2
=:J1 =:J2
We have
J1 ≤ c · ωu− 1 ≤ c · u− 1 ,I1 . (6.52)

2 2
For J2 we may further expand
ωuh − (ωu)h = (ωuh − (ωuh )h ) + ((ωuh )h − (ωu)h ) = Θ1 + Θ2 .
Thus, with (6.44), (6.45) and the superapproximation (6.50) it follows
Θ1 − 1 ≤ c · min ωuh − χ− 1 ≤ c · huh − 1 ,I1

2 χ∈Sh 2 2
(6.53)
≤ c · u− 1 ,I1 + c · he− 1 ,I1 .
2 2
Now, for χ ∈ Sh we have
LΘ2 , χ = −L(ωe)h , χ = − L(ωω̃e), χ

= (ωL − Lω)ω̃e, χ − ωL(ω̃e), χ (6.54)

=:T1 =:T2
Now, assuption (6.47) with l = 0 yields
|T1 | ≤ c · ω̃e− 3 · χ− 1 ≤ c · e− 3 ,I4 · χ− 1 ,

2 2 2 2
while we have for T2 :
T2 = ωL ((1 − ω̃)e) , χ − Le, ωχ .

=:T2 =:T2
For the first term we have by (6.48):
|T2 | ≤ c · e−s · χ− 1 .

2
For the second term we use e = u − uh and thus
|T2 | = |Le, ωχ − ψ |
≤ |L(ω̃e), ωχ − ψ | + |L ((1 − ω̃)e) , ωχ − ψ |
≤ c · e− 1 ,I4 · hχ− 1 + c · e−s · h · χ− 1 ,
2 2 2
making use of (6.48) and (6.50) and considering that ωχ − ψ has support in I2 .
If we now take χ = Θ2 in (6.54) and use (6.45), we obtain
Θ2 − 1 ≤ c · h · e− 1 ,I4 + e− 3 ,I4 + e−s . (6.55)

2 2 2
Thus, by (6.52) and (6.53)
e− 1 ,I0 ≤ c · u− 1 ,I1 + h · e− 1 ,I4 + e− 3 ,I4 + e−s (6.56)
2 2 2 2
For the rest of the proof we need the following

Lemma 6.5 For l = 1, 2, . . . , s − 3
2 there holds
e−l− 1 ,I0 ≤ c · h · e− 1 ,I + e−l− 3 ,I + e−s .

2 2 0 2 0
Proof First of all we have
e−l− 3 ,I0 ≤ ωe 1 = sup ωe, ϕ

2 H −l− 2 (Γ )
l+ 1
ϕ∈H 2 (Γ )
ϕl+ 1 =1
2
1
By (6.46), for all those ϕ with L∗ ψ = ϕ there holds: ψ ∈ H l− 2 (Γ ) . Thus,
ωe, ϕ = L(ωe), ψ = (L(ωω̃e), ψ

= (Lω − ωL)(ω̃e), ψ + ωL(ω̃e), ψ
= (Lω − ωL)(ω̃e), ψ + ωL((ω̃ − 1)e), ψ + Le, ωψ
=: S1 + S2 + S3 .
Here, we have
|S1 | ≤ (Lω − ωL)ω̃e−l+ 1 · ψl− 1 ≤ c · e−l− 3 ,I

2 2 2 0
and
|S2 | ≤ c · e−s
Finally, by direct application of the approximation property (6.49), for l ≥ 1 we

now have
|S3 | ≤ |Le, ωψ − χ |
≤ |L(ω̃e), ωψ − χ | + |L ((1 − ω̃)e) , ωψ − χ |
≤ c · e− 1 ,I0 + e−s · ωψ − χ− 1
2 2
≤ c · e− 1 ,I0 + e−s · h · ψ 1
2 2
For the rest of the proof of (6.43) we may now apply the above lemma inductively
in (6.56):
e− 1 ,I0 ≤ c u− 1 ,I + he− 1 ,I + e−s .

2 2 0 2 0
Iterating the estimation k−times then yields
e− 1 ,I0 ≤ c u− 1 ,I + hk e− 1 ,I + e−s .

2 2 0 2 0
Now, using (6.51) we obtain
hk e− 1 ,I ≤ cu− 1 ,I + c · hk uh 0,I0

2 0 2 0
≤ cu− 1 ,I + cuh −s,I
2 0 0
≤ c u− 1 ,I + e−s .
2 0
With u − χ = u − uh + (uh − χ) the assertion (6.43) finally follows. For further

reading see [355, 417]
6.5.7 Local L2 -Error Estimates
Again we assume Γ to be smooth and L = V Analoguous to the proof in the last

section we may show for a global quasiuniform mesh that
e0,I0 ≤ c min u − χ0,I0 + eH −s (Γ ) (6.57)
Proof The proof is similar to the above proof, but now for L2 (Γ ) instead of
H −1/2(Γ ) . First, we use the stability in L2 (Γ ) and H −1 (Γ ) :
v − vh i ≤ c · min v − χi , i = 0, −1 (6.58)

χ∈Sh
We consider the problem
Lv = f on Γ with L := V
Thus, by the orthogonality property of the Galerkin method we have
L(v − vh ), χ = 0 ∀ χ ∈ Sh .
Then we have
ωu − (ωu)h 0 ≤ c · ωu0 ≤ c · u0,I1 .
For Θ1 := ωuh − (ωu)h we obtain
Θ1 0 ≤ inf ωuh − χ0

χ∈Sh
(6.50)
≤ c · huh 0,I1 ≤ cuh −1,I1 (6.59)
≤ c · u0,I1 + c · e− 1 ,I1 .
2
Here, we use that the inverse property holds in this case.

For Θ2 := (ωe)h = (ωu)h − (ωuh )h let L∗ ψ = (ωe)h . Then, there holds:
((ωe)h 20 = (ωe)h , L∗ ψ = L(ωe)h , ψ = L(ωe)h , ψh

= L(ωe), ψh = L (ωω̃e) , ψh
= (Lω − ωL)(ω̃e), ψh + L(ω̃e, ωψh

=:T1 =:T2
Here, with (6.58) for i = −1 we have for T1 :
T1 ≤ (Lω − ωL)(ω̃e)1 · ψh −1

≤ c · ω̃e−1 · ψ−1
≤ c · e−1,I0 · (ωe)h 0 .
Correspondingly, we have for T2 :
T2 = L(ω̃e), ωψh
= ωL(ω̃ − 1)e, ψh + Le, ωψh .

=:T2 =:T2
Since ω(ω̃ − 1) = 0 , we have that ωL(ω̃ − 1) is of order k for arbitrary k. Thus,
|T2 | ≤ ωL(ω̃ − 1)e1 · ψh −1

≤ c · e−s · (ωe)h 0
and for T2 we have
T2 ≤ |L(ω̃e), ωψh − χ| + |L(1 − ω̃)e, ωψh − χ | .

=:τ1 =:τ2
For τ1 we have
τ1 ≤ (L(ω̃e)1 · ωψh − χ)−1

(6.50)
≤ cω̃e0 · hψh −1
Stability
≤ c · hω̃e0 · (ωe)h 0
≤ c · he0,I0 · (ωe)h 0
and for τ2
τ2 ≤ ωL((1 − ω̃)e)1,I0 · hψh −1

≤ c · e−s · (ωeh )0 · h
Combining the above one obtains
(ωeh )0 ≤ c e−1,I + e−s + h · e0,I

0 0
Finally, together with the results for the H − 2 −case, we have

1
e0,I0 ≤ c u0,I1 + e− 1 ,I + e−s + h · e0,I0

2 0
≤ c u0,I1 + e−s + he0,I0 .
This yields the desired estimate.

6.5.8 The K-Operator-Method
The method to be presented in this section is due to a work of J.H Bramble and A.H.
Schatz [61]. They considered the finite element method for the Dirichlet problem for
the Laplacian in a plane domain Ω. Here we consider again this problem with given
g on a smooth boundary ∂Ω = Γ with the single-layer potential ansatz

1
U (t) = − ln |t − s|z(s) ds , t ∈ Ω (6.60)
π
Γ
for some unknown density ϕ = z(s) . Thus,

1
g(t ) = − ln |t − s| · z(s) ds (6.61)
π
Γ
A parametrisation of the curve Γ shall be given by

⎧
⎨ [0, 1] → Γ
γ : x γ (x) = t with |γ | > 0
⎩ →
y γ (y) = s
We then have to solve the integral equation

1
!
Lu(x) = −2 ln |γ (x) − γ (y)| · u(y) dy = f (x) (6.62)
0
with u(x) := 2π1

z[γ (x)] · |γ (x)| and f (x) := g[γ (x)] for 0 ≤ x < 1 , which can
both be extended periodically on R2 .
Now, for the partition Δ : 0 = x0 < x1 < . . . < xN−1 < xN = 1 of the interval
[0, 1] and the space of test– and trialfunctions Sh := {ϕ : ϕ is 1-periodical and
piecewise constant on Δ} the Galerkin–method gives:
Find uh ∈ Sh such that

, (6.63)
Luh , ϕ = f, ϕ , ∀ ϕ ∈ Sh
1
where the scalar product and corresponding norm are given by v, w = v(x) ·
0
w(x) dx and w20 = w, w , respectively.
Here, we have
1
u(t) = −2 ln |t − γ (y)|u(y) dsy , t ∈Ω ∪Γ
0
1
and uh (t) = −2 ln |t − γ (y)|uh (y) dsy , t ∈Ω ∪Γ
0
and for t ∈ Ω :
1
|u(t) − uh (t)| = 2| ln |t − γ (y)|(u(y) − uh (y)) dsy |
0

G(t −γ (y)
≤ c · G(t − γ (·))2 · u − uh −2

=O(h3 )
≤c · h3 G(t − γ (·))2
For t ∈ Γ we have
|u(t) − uh (t)| ≤ c · hG(t − γ (·))0
Now, the K-operator-method is a post-processing such that
|u(t) − ũh (t)| = O h3 for t ∈ Ω ∪ Γ .
We then approximate u by u˜h = Kh ∗uh with Kh being a combination of B-Splines

such that polynomials up to a certain degree are produced by convolution:
1 x
Kh (x) := K with
h h
K(x) := − 12
1
[ψ(x + 1) + ψ(x − 1)] + 7
6 ψ(x)
⎧
⎨ x + 1 , −1 ≤ x ≤ 0
ψ(x) = 1 − x , 0 ≤ x ≤ 1
⎩
0 , else
We now want to prove some important properties of the functions Kh :

Lemma 6.6 ([61]) For the functions Kh as defined above there holds:
i) Kh ∗ u − u0 ≤ c · hs us , 0≤s≤4
ii) D α (Kh ∗ u) = Vhα ∗ ∂hα u, α = 1, 2 ,
where Vhα denotes a combination of B-splines of lower degree and ∂hα is the
central difference operator.
Using this lemma, we shall now study the application of the Kh and will therefore
distinguish two cases:
Case 1: Uniform grid
Let the grid be given by:
i 1
Δ: 0 = x0 < x1 < . . . < xN−1 < xN = 1, with xi = , h=
N N
If then uh ∈ Sh satisfies Luh , ϕ = f, ϕ ∀ ϕ ∈ Sh , and if furthermore

u ∈ H 3 := {u : D α u ∈ L2 , |α| ≤ 3} , then there holds
Kh ∗ uh − u0 ≤ c · h3 u3
Here, we have Kh ∗ uh − u ≤ Kh ∗ uh − Kh ∗ u0 + Kh ∗ u − u ,

1
where u∗h (t) := −2 ln |t − γ (y)|Kh ∗ uh (y) dy,
0
1
and thus |uh (t) − u∗h (t)| = 2| ln |t − γ (y)| (Kh ∗ uh (y) − u(y)) dy|
0
≤ c · h3 G(t − γ (·))0 , t ∈ Ω ∪ Γ .
Note that if we only have u ∈ H 3 (I1 ) ∩ H 1 (Γ ) , we shall apply local estimates.
Case 2: Quasiuniform grid
Let a quasiuniform grid be given according to Fig. 6.3.
Then, there holds:
∃h0 > 0 ∀ h ∈ (0, h0 ) ∀ ϕ ∈ Sh with supp(ϕ) ⊂⊂ I0 :
ϕ(· − 2h) ∈ Sh with supp(ϕ(· − 2h)) ⊂⊂ I1
Now, if u ∈ H 3 (I1 ) ∩ H 1 we have

Kh ∗ uh − uL2 (I0 ) ≤ c · h3 uH 3 (I1 ) + uH 1 ,
where in this case

∗
ũh := −2 ln |t − γ (y)|Kh ∗ uh (y) dy − 2 ln |t − γ (y)|uh (y) dy
I0 I −I0
and thus

|ũ∗h − u(t)| ≤ c · h3 G(t − γ (·))L2 (I0 ) + G(t − γ (·))H 2 (I \I0 ) .
For further reading we refer to [407, 416].
I0
I1

6.5.9 L∞ -Error Estimates for the Galerkin Approximation
Here we present from Rannacher and Wendland [349] the estimates for the single
layer potential. In this section, let Γ be a closed smooth curve or surface in R2 or
R3 , respectively. We want to consider L∞ -estimates for the Galerkin-error of the
single-layer potential operator, i.e. V u = f on Γ . For φh ∈ Shk,m (Γ ) ⇐⇒ φh ∈
H m (Γ ) and φh |Δ ∈ Pk−1 , the Galerkin method yields
V uh , φh L2 (Γ ) = V u, φh L2 (Γ ) ∀ φh ∈ Shk,m (Γ )
Now, for the Galerkin-error eh := u − uh we want to prove the estimate

Theorem 6.19 For a function u ∈ L∞ (Γ ) and a space of testfunctions Shk,m (Γ )
with −k ≤ − 12 there holds:

2 −1
n
1
u − uh L∞ (Γ ) ≤ c log · inf u − φh L∞ (Γ ) .
h φh ∈Sh
Let z ∈ Γ be fixed and define a weight function

1
2
σ (x) := |x − z|2 + κ 2 h2 , κ ≥1
and weighted Sobolev norms by

v2r;β := σ β (x)|D j v(x)|2 dx ,
|j |≤r K∈Πh K
with Πh being a regular triangulation. Define the Galerkin-error by e := u − uh ,

then for a smooth approximation δ of the Dirac-distribution on Γ , we will consider
the inner product e, δ , i.e. we will solve the equation V ∗ g = δ on Γ :
e, δ = e, V ∗ g = e, V ∗ (g − gh + gh ) (gh being the Galerkin-solution)

∗ ∗
= e, V η + e, V gh for η := g − gh
= u − φh , V ∗ η + e, V ∗ gh with e = u − uh + φh − φh

=0
⇒ |e, δ | ≤ |u − φh , V ∗ η, | ≤ u − φh L∞ (Γ ) · V ∗ ηL1 (Γ ) .
For the last term we have the estimate

2 −1
n
1
V ∗ ηL1 (Γ ) ≤ c log h 2 − 2 V η0;2 ,
n 3
h
⎧
⎪
⎨ − 2π
1
ln |x − y|η(y) dsy , n = 2
where V η(x) := Γ .
⎪
⎩ − 4π
1 1
|x−y| η(y) dsy ,n=3
Γ
Now, for the remainder of the proof we need the following auxiliary lemmata:
Lemma 6.7 Defining ξi := xi − zi for 1 ≤ i ≤ n , there holds for V ξi − ξi V :
1
(V ξi − ξi V )φr+1 ≤ c · φr−1 , 1 ≤ i ≤ n, − ≤r≤k.
2
Lemma 6.8 For η := g − gh there holds:
ξi ηk;0 + ξi2 ηk;−2 ≤ c · gk;2 , 1 ≤ i ≤ n.
Lemma 6.9 For α ≤ β ≤ m there exists a constant c > 0 such that
ξi ηβ ≤ c · hα−β ξi ηα + hk−β gk;2 , 1 ≤ i ≤ n.
Note, in our case we always have α = − 12 .

Lemma 6.10 With V η(x) as defined in the proof of the theorem above there holds
1
n
V η0,2 ≤ c · h 2 ξi η− 1 + c · hk+1 gk;2 .
2
i=1
Lemma 6.11 There exist constants > 0 and c > 0 (independent of ) such that

1 1 1
ξi η− 1 ≤ · h− 2 V η0;2 + c 1 + hk+ 2 gk;2 .
2
Note that for sufficiently small the last two lemmata yield V η0;2 ≤ c ·
hk+1 gk;2 , giving

2 −1
n
1 3
h 2 − 2 hk+1 gk;2 inf u − φh L∞ (Γ )
n
|e, δ | ≤ c log
h φh ∈Shk,m
+ c{hk gk + g}e−2 . (6.64)
Lemma 6.12 For δ as defined above there holds

i) gr ≤ cδ
r+1 , 0 ≤ r ≤ k "
n
ii) gk;2 ≤ c δk + hδk+1 + ξi δk+1
i=1
Lemma 6.13 ([349]) With z ∈ K there exists a function δ ∈ C0∞ (K) such that
∃ c > 0 (independent of z) : φh (z) = φh , δ ∀ φh ∈ Shk,m satisfying
1. δL1 ≤ c
2. hr δr+1 ≤ c · h− 2 − 2 , 0 ≤ r ≤ k
n 1
1 n
3. hk ξi δk+1 ≤ c · h 2 − 2 , 1 ≤ i ≤ n .
Before proving the above lemmata we first want to complete the proof of Theo-
rem 6.19:
For z ∈ K we have
|e(z)| ≤ |(u − φh )(z)| + |(φh − uh )(z)|

≤ |(u − φh )(z)| + |φh − uh , δ |
≤ |(u − φh )(z)| + |e, δ | + u − φh L∞ (K) · δL1 (K)
Thus, by Lemma 6.13 we have
|e(z)| ≤ c · inf u − φh L∞ + |e, δ | (6.65)

φh ∈Shk,m
With Lemma 6.12,(6.64) it follows

n
1 2 −1 1
inf u − φh L∞ + ch− 2 − 2 e−2
n
|e, δ | ≤ c log
h φh ∈Shk,m
completing the proof of the theorem.

In the remainder of this section we now prove the above lemmata:
Proof of Lemma 6.7: Standard property of pseudodifferential operators.
Proof of Lemma 6.8: First, we have Di (ξi ; η) = δi,j η + ξj Di η and thus
|Di (ξj η)|2 ≤ c · ξj2 |Di η|2 + |η|2

≤ c · σ 2 |Di η|2 + |η|2
⇒ |Dk Di ξj η| ≤ c · σ 4 |Dk Dj η|2 + c · |η|2 + c · σ 2 |Di η| + c · σ 2 |Dk η| .
2 2
Furthermore, there holds
ξi η2k;0 ≤ c{η2k;2 + η2k−1;0 }

ξi2 η2k;−2 ≤ cη2k;2 + c η2k−2;−2 + c η2k−1;0
If now Ph : H r −→ Sh is the interpolation operator, there holds
v − Ph vj ;β ≤ c · hr−j vr;β . (6.66)

With φh being a spline we further have
φh k,β ≤ cφh k−1,β

(6.67)
φh k,β ≤ cφh k,α
and thus:
ηk;2 ≤ η − Ph ηk;2 + Ph ηk;2

≤ cgk;2 + Ph ηk;2
≤ cgk;2 + cηk−1;0 using (6.66) and (6.67) .
ηk−j ;0 ≤ g − Ph gk−j ;0 + Ph (g − gh )k−j ;0 ≤ cgk;0 · hj
⇒ ηk−2;−2 ≤ ηk−2;0 c · h−1 ≤ c · h−1 h2 gk
The assertion of the lemma finally follows, since

g2k = |Dj g|2 σ 2 σ −2 dx ≤ c · h−2 g2k;2 with σ −2 ≤ c · h−2 .
|j |≤r K∈Πh Π
h
Proof of Lemma 6.9: The assertion follows by the three arguments (for β = 0):
1. ξj η0 ≤ ξj η − Ph (ξj η)0 + Ph (ξj η)0
2. ξj η − Ph (ξj η)0 ≤ c · ξj ηk hk ≤ c · hk gk,2 . (using Lemma 6.8)
− 12
3. Ph (ξj η)0 ≤ c · h Ph (ξj η)− 1 by the inverse property
2
− 12
≤ c·h ξj η− 1
2
Proof of Lemma 6.10: For the proof of this lemma we further need the following
estimates:
a) g2k ≤ c · h−1 g2k;2
b) f 21;−2 ≤ c ξi f 21 + cf 20 + c · h2 f 21
i
c) f k;2 ≤ c · ξi f k + f k−1 + f k
i
d) σ 2 f 1;−2 ≤ cf + f 1,2
with |Di σ 2 f |2 ≤ c · σ 4 |Di f |2 + c · σ 2 |f |2 .
Now, we have for lemma 6.10:
V ∗ η20;2 = σ 2 V ∗ η, V ∗ η
= σ 2 V ∗ η − Ph σ 2 V ∗ η , V ∗ η
≤ σ 2 V ∗ η − Ph σ 2 V ∗ η 0;−2 V ∗ η0;2
⇒ V ∗ η0;2 ≤ σ 2 V ∗ η − Ph σ 2 V ∗ η 0;−2
≤ c · hσ 2 V ∗ η1;−2
≤ c · hV ∗ η0 + c · hξj V ∗ η1 + c · h2 V ∗ η1
1
≤ 2c · hη−1 + c · h 2 ξj η− 1 + c · hk+1 gk;2
2
1
≤ c·h·h k+1
gk + c · h ξj η− 1 + c · hk+1 gk;2
2
2
1
≤ c · hk+1 gk,2 + c · h ξj η− 1 + c · hk+1 gk;2
2
2
Proof of Lemma 6.11:The assertion of the lemma can be shown by
ξi η2 ≤ ∗
γ ξi η, V ξi η
1
− 12
∗

≤ − ξi V ∗ η + γ1 ξi2 η − Ph ξi2 η , V ∗ η
γ ξi η, V ξi η
1
≤ cξi η− 1 · η− 3 + c · hk gk,2 V ∗ η0;2

2 2
≤ cξi η− 1 h 2 +k gk,2 + c · hk gk,2 V ∗ η0;2
1
Proof of Lemma 6.12: The proof is straight forward. At the end of this section we
now want to give a further estimate for the potential.
Theorem 6.20 ([349]) Under the same assumptions as made above there holds
n
2
V u − V uh L∞ (Γ ) ≤ c · hk+1 · log h1 uW k,∞ (Γ )
For pointwise estimates of pseudodifferential equations of positive order see

[350].
6.6 A Discrete Collocation Method for Symm’s Integral

Equation on Curves with Corners
Corner singularities of the solution (here of Symm’s integral equation) yield only
slow convergence for a numerical scheme like the Collocation method. This can
be overcome by an appropriate mesh grading transformation. This procedure is
described below and goes back to the initiating work [95] by Chandler and Graham.
In this section we present from [159] a collocation method with trigonometric
polynomials and its discrete counterpart for Symm’s integral equation

1
− ln | x − ξ | u (ξ ) dΓ (ξ ) = f (x) , x ∈ Γ, (6.68)
π Γ
6.6 Discrete Collocation for Symm’s Equation 173
on the boundary Γ of a simply connected bounded domain Ω in R2 . Γ is assumed

to be (infinitely) smooth, with the extension of a corner near at point x0 . Near x0
Γ should consist of two straight lines intersecting with an interior angle (1 − χ) π,
0 < |χ| < 1. We assume cap (Γ ) = 1.
For smooth Γ collocation and quadrature methods based on splines or trigono-
metric polynomials are analyzed in [379]. Here we show for a curve with a corner
that collocation and discrete collocation with trigonometric polynomials converge
with a rate as high as justified by the order of the mesh grading and the regularity of
the data .
We rewrite (6.68) using an appropriate nonlinear parametrization γ : [0, 1] → Γ
which varies more slowly than arc-length parametrization in the vincinity of x0 .
Consider a parametrization γ0 : [0, 1] → Γ such that γ0 (0) = γ1 (0) = x0 and
| γ0 (s) |> 0 for all 0 ≤ s ≤ 1. Choosing a grading exponent q ∈ N and selecting a
function υ such that
υ ∈ C ∞ [0, 1] , υ (0) = 0, υ (1) = 1, υ (s) > 0, 0 ≤ s ≤ 1, (6.69)
we define the mesh grading transformation near the corner (see also [95])
υ q (s)
γ (s) = γ0 (ω (s)) , where ω (s) = . (6.70)
υ q (s) + υ q (1 − s)
The parametrization γ is graded with exponent q near the corner. With x = γ (s) ,
ξ = γ (σ ) , equation (6.68) becomes
1
Kw (s) := −2 ln | γ (s) − γ (σ ) | w (σ ) dσ = g (s) , s ∈ [0, 1] , (6.71)
0
where
1
w (σ ) = | γ (σ ) | u (γ (σ )) , g (s) = f (γ (s)) . (6.72)
2π
The solution w of the transformed equation (6.71) may be made as smooth as
desired on [0, 1] provided f is smooth and the grading exponent is sufficiently
large. Therefore w can be optimally approximated using trigonometric polynomials
as basis functions.
We rewrite (6.71) as
Aw + Bw = g (6.73)
with
1 1
Aw (s) = −2 ln | 2e− 2 sin (π (s − σ )) | w (σ ) dσ, (6.74)
0
1
Bw (s) = b (s, σ ) dσ, (6.75)
0

γ (s) − γ (σ )

b (s, σ ) := −2 ln , 0 < s, σ < 1, s = σ. (6.76)
2e− 12 sin (π (s − σ ))
The kernel (6.76) is 1-periodic in both variables and C ∞ for 0 < s, σ < 1, but Γ
has fixed singularities at the four corners of the square [0, 1] × [0, 1] .
Next we consider trigonometric collocation: Let H t , t ∈ R, be the usual
Sobolev spaces of 1-periodic functions on the real line, with norm given by

v 2t =| v̂ (0) |2 + | m |2t | v̂ (m) |2 ,
m=0
where the Fourier coefficients of v are defined by

1
v̂ (m) = v, ei2πms = v (s) e−i2πms ds. (6.77)
0
Introduce the collocation points
h 1
sj = j h + , j ∈ Z, h := , (6.78)
2 n
and let Th denote the n-dimensional space of trigonometric polynomials with the
standard basis
& n n'
ϕk (s) = ei2πks , k ∈ Λn := j ∈ Z : − < j ≤ , (6.79)
2 2
i.e.
& '
Th = span ei2πks , k ∈ Λn , s ∈ [0, 1] .
Then, for any continuous 1-periodic function v, the interpolatory projection Qh v

onto Th is well defined by

(Qh v) sj = v sj , j = 0, . . ., n − 1. (6.80)
The following lemma shows that Qh has optimal convergence properties:

Lemma 6.14 ([5]) For r ≥ t ≥ 0 and r > 1
2 there exists c > 0 such that
v − Qh v t ≤ c hr−t v r if v ∈ H r . (6.81)
With (6.79) we have

Qh v (s) = αk ϕk (s) , αk := h v sj ϕk sj .
k∈Λn j ∈Λn
Then the collocation method for (6.73) seeks wh ∈ Th such that
Qh (A + B) wh = Qh g.
Since Qh commutes with A on Th , there holds
(A + Qh B) wh = Qh g, wh ∈ Th . (6.82)
Following [4, 155], we rewrite (6.73) as the second kind equation
(I + M) w = e, with M = A−1 B, e = A−1 g. (6.83)
Due to [4] the operator A in (6.74) can be written as
v̂ (m)
Av (s) = ϕm (s) , v ∈ Ht.
max (1, | m |)
m∈Z
and

A−1 v (s) = max (1, | m |) v̂ (m) ϕm (s) , v ∈ H t +1 .
m∈Z
Therefore the integral operator A is an isomorphism of H t onto H t +1 for any real

t. In addition
A−1 = −H D + J = −DH + J (6.84)
where J v (s) = v̂ (0) , Dv (s) = v (s) and H the (suitably normalized)Hilbert

transform
1
1
H v (s) = − p.v. cot (π (s − σ )) v (σ ) dσ,
2π 0
which is bounded in L2 . Therefore the operator M of (6.83) becomes
M = −H DB + J B.
We now recall some analytical results on equations (6.71) and (6.83) which are
needed in the convergence analysis of the trigonometric collocation method. The
first theorem was proved in [155], using a decomposition of M into a Mellin

convolution operator local to each corner and a compact operator on H 0 .
Theorem 6.21 ([155]) The operators I + M : H 0 → H 0 and K : H 0 → H 1 are
continuously invertible, and there holds the strong ellipticity estimate
Re ((I + M + T ) v, v) ≥ c v 20 , v ∈ H 0,
with some compact operator T on H 0 .

The next theorem shows that the unique solution of (6.71) is smooth provided the
given data f in (6.68) is smooth and the grading exponent q is sufficiently large. Let
1
H l (Γ ) , l > 0, denote the restriction of the usual Sobolev space H l+ 2 R2 to Γ.
Theorem 6.22 ([155]) Let l ∈ N, q > l + 1

2 (1+ | χ |) , and suppose that f ∈
l+ 25
H (Γ ) .Then the unique solution of (6.71) satisfies w ∈ H l . Moreover, there
exists δ < 12 such that
D m w (s) = O | s |l−m−δ , as s → 0, m = 0, . . ., l. (6.85)
The following theorem describes the properties of the kernel function b (s, σ )
defined in (6.76).
Theorem 6.23 ([156]) On each compact subset of R × R \ (Z × Z) , the derivates
Dsi Dσm-b (s, σ ). of order i + m ≤ q are bounded and 1-periodic. Moreover, for
s, σ ∈ − 12 , 12 \ {0} , we have the estimates
| b (s, σ ) | ≤ c | ln (| s | + | σ |) |,
| Dsi Dσm b (s, σ ) | ≤ c (| s | + | σ |)−i−m , 1 ≤ i + m ≤ q.
Next we rewrite the collocation method (6.82) as a projection method for (6.83). For
v ∈ H 0 , let Rh v ∈ Th be the solution of the collocation equations ARh v = Qh Av.
Then Rh = A−1 Qh A is a well-defined projection operator of H 0 onto Th .
Note that (6.82) is equivalent to
(I + Rh M) wh = Rh e.
It is well-known that the use of Mellin convolution operators implies that stability
can only be shown for a slightly modified Collocation method (see [95, 157]). We
introduce, for τ > 0 sufficiently small, the truncation operator
%
v (s) , if s ∈ (τ, 1 − τ )
Tτ v =
0, if s ∈ (0, τ ) ∪ (1 − τ, 1)
and consider the modified collocation method
(A + Qh BTi % h ) wh = Qh g, wh ∈ Th , (6.86)
where i % is a fixed natural number independent of h. If i % = 0 then (6.86) coincides

with (6.82). Otherwise, (6.86) can be obtained from (6.82) by a slight change to the
coefficient matrix of the corresponding linear system. Now (6.86) is equivalent to
(I + Rh MTi % h ) wh = Rh e, wh ∈ Th . (6.87)
The following theorem provides the convergence of the (modified) collocation

method with optimal order in the L2 norm.
Theorem 6.24 ([159]) Let q ≥ 2, and suppose that i % is sufficiently large.
(i) The method (6.87) is stable, that is the estimate
(I + Rh MTi % h ) v 0 ≥ c v 0 , v ∈ Th (6.88)
holds for all h sufficiently small, where c is independent of h and v.

(ii) If, in addition, the hypothesis of Theorem 6.22. holds, then (6.86) has a unique
solution for all h sufficiently small and
w − wh 0 ≤ c hl , (6.89)
where c is a constant which depends on w and i % but is independent of h.

Proof As in [155] we first verify the stability estimate (6.88). Since, by Theo-
rem 6.21, I + M is strongly elliptic and invertible on H 0 , we have stability of
the finite section operators Tτ (I + M) Tτ as τ → 0, i.e.
(I + MTτ ) v 0 ≥ c v 0 , v ∈ H 0 , τ ≤ τ0 . (6.90)
To prove (6.88), we need (6.90) and the following perturbation result:

For fixed q ≥ 2 and each ε > 0, there exists i % ≥ 1 such that for all h sufficiently
small
(I − Rh ) MTi % h v 0 ≤ ε v 0 , v ∈ H 0. (6.91)
Observe that the operator M takes the form M = −H DB + J B, where J B is a

compact operator on H 0 . Since Ti % h 0 = 1 and Rh → I pointwise on H 0 , it is
sufficient to prove (6.91) with M replaced by −H DB. From (6.93) and the fact that
I − Rh annihilates the constants, we obtain the estimate
(I − Rh ) MTi % h v 0 ≤ ch MTi % h v 1 = c h DMTi % h v 0

≤ c h D 2 BTi % h v 0 .
To prove (6.91), it is sufficient to verify that

c
D 2 BTi % h v 0 ≤ v 0 , v ∈ H 0, (6.92)
i %h
where c is indepent of i % , h and v. Using Theorem 6.23. , we now obtain

| D 2 BTi % h v (s) | ≤ | Ds2 b (s, σ ) | | v (σ ) | dσ
Ji % h

≤c (| s | + | σ |)−2 | v (σ ) | dσ
Ji % h

c |σ | 1 1
≤ | v (σ ) | dσ, s∈ − , ,
i%h Ji % h (| s | + | σ |)2 2 2
where Ji % h = − 12 , −i % h ∪ i % h, 12 . Then (6.92) follows by taking L2 norms and

using the fact that the integral operator with Mellin convolution kernel σ (s + σ )−2
is bounded on L2 (0, ∞) (see [155] Appendix).
The proof of (6.88) is complete since with (6.90) and (6.91) there holds
(I + Rh MTi % h ) v 0 ≥ v + MTi % h v 0 − MTi % h v − Rh MTi % h v 0 ≥ c v 0 .
To prove the error estimate (6.89), we note that
w − wh 0 ≤ (I − Rh ) w 0 + wh − Rh w 0 ,
where the first term is of order hl by Theorem 6.22. and
v − Rh v 0 ≤ chl v l , if v ∈ H l . (6.93)
Using (6.88), (6.87) with (6.83), and the uniform boundedness of Rh on H 0 gives
wh − Rh w 0 ≤ c (I + Rh MTi % h ) (wh − Rh w) 0
= c Rh {(I + M) w − (I + MTi % h ) Rh w} 0
≤ c (I − Rh ) w 0 +c (I − Ti % h ) w 0 .
The proof is complete since by (6.85) (with m = 0) and (I − Ti % h ) w 0 ≤ c hl

the last term is of order hl again.

The following corollary shows that the collocation solutions to the transformed
equation yield superconvergent approximations to interior potentials.
Corollary 6.3 ([159]) Under the hypothesis of Theorem 6.24.(ii), we have
w − wh −1 ≤ c hl+β ,
where β = 1 if i % = 0 and β = 1
2 if i % ≥ 1.
In the following we consider discrete collocation. To define a fully discrete version
of the collocation method (6.82), introduce the nodes
1
σr = r h, r ∈ Z, where h := . (6.94)
n
To evaluate the integral
1
I (v) = v (σ ) dσ
0
for a 1-periodic continuous function v, approximate it by the trapezoidal rule

n−1
Ih (v) = h v (σr ) . (6.95)
r=0
The integral operator B of (6.75) is now approximated by

n−1
Bh v (s) := Ih (b (s, ·) v (·)) = h b (s, σr ) v (σr ) , (6.96)
r=0
and replacing B with Bh in (6.82), the discrete collocation method can be written in
the form
(A + Qh Bh ) wh = Qh g, wh ∈ Th . (6.97)
To obtain a linear system for finding wh , let

wh (s) = αk ϕk (s)
k∈Λn
and calculate the coefficients αk from (6.97) and the definitions of A

Aϕk = max(1,|k|)
1
ϕk , Qh and Bh :
1 2
ϕk sj
+ (Bh ϕk ) sj αk = g sj , j = 0, . . ., n − 1. (6.98)
max (1, | k |)
k∈Λn
Using nodal values of wh as unknows, the following system is obtained which is

computationally less expensive:

n−1
# $
βj k + h b sj , σk wh (σk ) = g sj , j = 0, . . ., n − 1, (6.99)
k=0
where

h ϕr sj ϕk (σr )
βj k = .
max (1, | r |)
r∈Λn
For the computation of the coefficients βj k one can use the fast Fourier transform.
Our convergence analysis follows the same lines as above. That is, instead
of (6.97) we consider the modified method
(A + Qh Bh Ti % h ) wh = Qh g, wh ∈ Th . (6.100)
Setting Mh = A−1 Bh and using (6.83) and the projection Rh , (6.100) can be written
as
(I + Rh Mh Ti % h ) wh = Rh e, wh ∈ Th . (6.101)
For our analysis, the following standard estimate for the trapezoidal rule (6.95) is
needed.
Lemma 6.15 Let l ∈ N, and suppose that v has 1-periodic continuous derivates of
order < l on R and that D l v is integrabel on (0, 1) . Then
1
| I (v) − Ih (v) | ≤ c hl | D l v (σ ) | dσ,
0
where c does not depend on v and h.

The following lemma is the key to the stability of (6.101); it is used in the proof of
Theorem 6.25
Lemma 6.16 ([159]) For fixed q ≥ 2 and for each ε > 0, there exists i % ≥ 1
independent of h such that, for all v ∈ Th and all sufficiently small h,
(M − Mh ) Ti % h Mh Ti % h v 0 ≤ ε v 0 , (6.102)
(M − Mh ) Ti % h (I − Rh ) Mh Ti % h v 0 ≤ ε v 0 . (6.103)
Theorem 6.25 Assume q ≥ 2, and suppose that i % is sufficiently large. Then the
estimate
(I + Rh Mh Ti % h ) v 0 ≥ c v 0 , v ∈ Th (6.104)
holds for all h sufficiently small, where c is independent of v and h.

Proof Due to Theorem 6.24. (i) , the operators
(I + Rh MTi % h )−1 : Th → Th , h ≤ h0
exists and are uniformly bounded with respect to the H 0 operator norm if i % is large
enough. Setting
Ch := I − (I + Rh MTi % h )−1 Rh Mh Ti % h
and
Dh := (I + Rh MTi % h )−1 Rh (Mh − M) Ti % h Rh Mh Ti % h
gives
Ch (I + Rh Mh Ti % h ) = I − Dh . (6.105)
Now, using
(M − Mh ) Ti % h v 0 ≤ c v 0 , v ∈ Th
and the uniform boundedness of Rh on H 0 , we see that Rh Mh Ti % h and hence Ch

are also uniformly bounded. Furthermore, Lemma 6.16 yields for some ε ∈ (0, 1)
provided that i % is sufficiently large
Dh v 0 ≤ c Rh (Mh − M) Ti % h Rh Mh Ti % h v 0
≤ c { (M − Mh ) Ti % h Mh Ti % h v 0 + (M − Mh ) Ti % h (I − Rh ) Mh Ti % h v 0 }
≤ ε v 0 , v ∈ Th , h ≤ h0
Hence (I − Dh )−1 exists and is uniformly bounded, and (6.105) gives
(I + Rh Mh Ti % h )−1 0 = (I − Dh )−1 Ch 0 ≤ c , h ≤ h0 ,
which yields (6.104).

Finally, we show in [159] that (6.100) converges with the same optimal order as the
collocation method.
Theorem 6.26 ([159]) Let l ∈ N, q > l + 1
2 (1+ | χ |) , and suppose that f ∈
l+ 25
H (Γ ) . Suppose further that i % is sufficiently large. Then (6.100) has a unique
solution for all h sufficiently small and
w − wh 0 ≤ c hl , (6.106)
where c is independent of h.
Table 6.1 L2 error of the density [159]

q =2 q=3 q=4 q=5
n wh − w ∗ 0 EOC wh − w ∗ 0 EOC wh − w ∗ 0 EOC wh − w ∗ 0 EOC
16 8.25-2 0.69 5.08-2 1.35 4.80-2 2.25 6.37-2 3.22
32 5.11-2 0.68 1.98-2 1.22 1.00-2 1.81 6.82-3 2.39
64 3.18-2 0.72 8.47-3 1.23 2.85-3 1.80 1.29-3 2.39
128 1.91-2 0.85 3.60-3 1.28 8.16-4 1.82 2.45-4 2.40
256 1.06-2 1.24 1.47-3 1.58 2.30-4 2.05 4.67-5 2.54
512 4.46-3 4.91-4 5.53-5 8.00-6
As a numerical example we choose
γ0 (s) = sin πs (cos (1 − χ) πs, sin (1 − χ) πs) , s ∈ [0, 1] , 0 <| χ |< 1.
Here Γ is the boundary of a “teardrop-shaped” region with a single corner at s = 0

(or s = 1) and smooth elsewhere. We solve (6.68) with the quadrature-collocation
shema (6.82). The right hand side f in (6.68) is
& 1 '
f (x) = exp (x1 ) cos (x2 ) + Re (x1 + ix2 ) (1−χ) ,
with χ = −0.76, which corresponds to a re-entrant corner. Since the exact solution
of (6.68) is unknown, we compute an approximation w∗ with (6.82) for n = 1024
and use it as the exact solution (Table 6.1).
Remark 6.5 For discrete collocation of the hypersingular integral equation gov-
erning the Neumann problem see [231]. For collocation with Chebyshev/Jacobi
polynomials for first kind integral equations on an interval see [166, 230, 381].
Collocation and discrete collocation with trigonometric polynomials for the mixed
Dirichlet-Neumann problem for the Laplacian are investigated in [158, 262] and in
[406].
6.7 Improved Galerkin Method with Augmented Boundary

Elements
If the boundary curve has corners, the boundary charges and solutions develop
singularities. In this section the augmentation of boundary elements by special
singular functiones near the corners is used to derive higher convergence rate for
the Galerkin method for integral equations on polygons. The use of singularity
functions as additional test/trial functions was originally introduced in the finite
element method by G. Fix [180]. For including the singular behaviour, parts of the
pioneering work by Costabel and Stephan in [128] are presented.
6.7 Improved Galerkin with Augmented Elements 183
&
Let by A denote the set of all exceptional exponents, i.e., A= αj k |j ={1, . . . , J },
'
k ∈ N ∩ (0, 2). The space Z s is defined for all s with s − 1/2 ∈ (0, 2) \ A :

J
u ∈ Z s ⇔ u = u0 + c j k uj k χ j , (6.107)
j =1 αjk <s−1/2
where
(i) u0 ∈ H s−1 (Γ ) for s ∈ [1/2, 3/2)
u0 |Γ j ∈ H s−1(Γ j ) for s ∈ [3/2, 5/2]
(ii) cj k = 0 if uj k is not defined.
(iii) uj k = x αjk
and

J
u2Z s = u0 2H̃ s−1 (Γ ) + |cj k |2 if s ∈ [1/2, 3/2) \ A.
j =1 αjk <s−1/2
Note Z 1/2 = H̃ −1/2(Γ ).

Next, we construct augmented finite element spaces such that
p,t,k
Sh ⊂ Zk ⊂ Zs (s k).
Definition 6.2
p,t,k

J
Ũ ∈ Sh ⇔ Ũ = ũ0 + c j k uj k χ j (6.108)
j =1 αjk <p−1/2
where cj k ∈ R is arbitrary, and uj k , χj are as in the definition of Z p above and

ũ0 ∈ Sht,k .
p,t,k
Lemma 6.17 The finite element spaces Sh have the following approximation
property (6.109) and inverse property (6.110):
p,t,k
For any U ∈ Z r there exists a Ũ ∈ Sh with t r and p r and a constant
c > 0, independent of h and U , such that for q min{k, r}
U − Ũ Z q c hr−q U Z r (6.109)
For q r, ε > 0, k r there exists a constant M > 0, independent of h, such that

p,t,k
for all Ũ ∈ Sh
Ũ Z r M hk−r−ε Ũ Z q (6.110)
with ε = 0 if A ∩ [q − 1/2, r − 1/2] = ∅ where r = max{p, r}.

Proof For showing (6.109) we choose c̃j k = cj k for αj k < r − 1/2 and c̃j k = 0
for αj k ∈ (r − 1/2, p − 1/2), where c̃ are the coefficients in (6.108) and cj k those
in (6.107). Thus it remains to estimate the smooth parts, i.e. u0 − ũ0 Z q which
reduces to an ordinary Sobolev norm where we can apply the convergence property
of a Sht,k -system. Hence
U − Ũ Z q = u0 − ũZ q c hr−q u0 Z r c hr−q U Z r .

The (6.110) for q ≤ r ≤ k and A ∪ [q − 1/2, r − 1/2] = ∅
inverse property
r = max(r, p) follows immediately from the definition of norms and the inverse
property of Sht,k -systems [14]

J
J
Ũ 2Z r = ũ0 2Z r + |c̃j k | ≤ M h
2 2 2(q−r)
ũ0
Zq + |c̃j k |2
j =1 αjk <r−1/2 j =1 αjk <q−1/2
M 2 h2(q−r) Ũ 2Z q .
For the proof in the remaining cases see [128].

Now, we can perform the augmented Galerkin method for the following integral
equation with the hypersingular operator W :
For given f ∈ H −1/2(Γ ) find u ∈ H 1/2(Γ ) such that
W u = (I + K )f. (6.111)
For given f ∈ H −1/2 (Γ ) find Uh∗ ∈ Sh

p,t,k
such that
W Uh∗ , V = (1 + K )f, V

p,t,k
∀V ∈ Sh .
As shown in [128] we have

Theorem 6.27 There holds the improved error estimate
U − Uh∗ Z r c hs−r−ε U Z s c hs−r−ε f H s−1 (Γ )
where U ∈ Z s solves integral equation (6.111).

Remark 6.6 For sufficiently large s the above theorem gives explicit estimates for
the error of the stress intensity factors cj k − c̃j k (see [128]).
Remark 6.7 For corresponding results on augmented BEM applied to 2D mixed
bvp’s see [128, 433], (for numerical results see [279] ) and to 2D crack problems
see [256, 432].
6.8 Duality Estimates for Projection Methods 185
6.8 Duality Estimates for Projection Methods
For higher rates of convergence in negative norms (or norms involving less derivates
than the energy norm), Aubin-Nitsche-type duality estimates are a standard tool in
the analysis of finite element Galerkin methods [7, 106, 328, 412]. Such duality
arguments are also available for Galerkin methods for certain integral equations
and pseudodifferential equations on smooth manifolds [258] and also for some
collocation methods that can be reduced to such Galerkin methods [3, 357].
Standard formulations [258] require unique solvability of the adjoint equation
in the negative norm, and this is in general not satisfied in the presence of corners
and edges of discontinuous boundary conditions, due to singular solutions of the
homogeneous equation.
It is useful to write projection methods in the form of Galerkin-Petrov methods
[202], which is always possible. Also collocation methods can easily be written in
this form [131].
Let X and Y be Banach spaces and A : X → Y an isomorphism of norm A .
We assume that we have the subspaces
VN ⊂ X, TN ⊂ Y ,
and we replace the equation for u ∈ X :
Au = f (6.112)
by the equation for uN ∈ VN :
< t, AuN > = < t, f > ∀t ∈ TN , (6.113)
the brackets denoting the duality between the space Y and its dual Y . Here f ∈ Y
is given and uN in (6.113) is considered as an approximate solution to (6.112). One
usually has in addition
dim VN = dim TN < ∞
and uses a whole sequence of VN , TN , uN to approximate u, but for the moment we

do not need this.
We assume that we somehow know an estimate for the error u − uN in the norm
of X and want to estimate it in a smaller norm.
Thus we have a second norm · X1 on X that satisfies with some M > 0
x X1 M x X ∀x ∈ X. (6.114)
By completion of X one thus has a Banach space X1 ⊃ X with X densely

embedded, and hence the dual space X1 is contained in X in a natural way:
3 4
X1 = ξ ∈ X | ξ is continuous with respect to · X1 .
Note that the adjoint A : Y → X is bijective and continuous, so that A−1 is

well-defined on X1 .
The following lemma from [134] contains the abstract version of the Aubin-
Nitsche duality estimate.
Lemma 6.18 Assume that there is a constant cN > 0 such that
& '
inf A−1 ξ − τ Y | τ ∈ TN εN ξ X ∀ξ ∈ X1 . (6.115)
1
Then for all f ∈ Y, and u and uN satisfying (6.112) and (6.113), respectively, there
holds
u − uN X1 εN A u − uN X . (6.116)
Proof We have
& '
u − uN X1 = sup |< ξ, u − uN >| | ξ ∈ X1 , ξ X1 = 1 .
Choose ξ ∈ X1 with ξ X1 = 1 and τ ∈ TN . Then
|< ξ, u − uN >|= |< A−1 ξ, A (u − uN ) >|= |< A−1 ξ − τ, A (u − uN ) >|

A−1 ξ − τ Y A (u − uN ) Y A−1 ξ − τ Y A u − uN X .
Minimizing over τ ∈ TN we find with (6.115)
|< ξ, u − uN >| εN A u − uN X .

Remark 6.8 This is only a small modification of the statement and the proof of the
Aubin-Nitsche lemma for Galerkin methods as stated by Ciarlet [106]. Even for
Galerkin methods and for a Hilbert space X1 , however, it turns out to be useful to
distinguish between X1 and X1 , as we shall see now.
6.8.1 Application to Galerkin Methods
The well-known quasioptimality result (Céa lemma)
u − uN X C0 inf { u − v X | v ∈ VN } (6.117)
is usually applied in the following way.

One uses an approximation result for the spaces VN in X that holds for u in
some subspace of X. An example is the order of convergence of best spline (or
trigonometric) approximation in the norm of a Sobolev space (i.e. one of higher
regularity).
Additionally, one uses regularity results about the operator A that for f in
Eq. (6.112) given in a subspace Y1 of Y,the solution u is contained in a subspace
of X where the above approximation property holds, i.e
Y1 ⊂ Y with · Y M · Y1 . (6.118)
and an estimate
inf { u − v X | v ∈ VN } δN f Y1 , for all u ∈ X

with f = Au ∈ Y1 .
(6.119)
Of course, (6.117) and (6.119) together give a convergence rate O (δN ) for the error
in the norm of X.
Now suppose we have the situation of a Galerkin method:
X = Y, Y = X, V N = TN . (6.120)
We assume further that A is selfadjoint:
A = A .
Then we see immediately that the two conditions (6.115) and (6.119) are identical,
if we have εN = δN and X1 = Y1 . We can take the latter as a definition for X1 and
for X1 :
|< y, x >|
x X1 := sup for all x ∈ X. (6.121)
y∈Y1 \{0} y Y1
We see from (6.118) that (6.114) holds, and therefore we can apply Lemma 6.18.
Theorem 6.28 Let (6.117)–(6.121) be satisfied. Then there holds
u − uN X1 δN A u − uN X δN
2
C0 A f Y1 (6.122)
for all f ∈ Y1 and u and uN satisfying (6.112) and (6.113).

Remark 6.9 If A is not selfadjoint, one has to require the same regularity and
approximation result for A as for A :
inf { u − v X | v ∈ VN } δN A u Y1 for all u ∈ X with f = A u ∈ Y1 .
We now present an example of a Galerkin scheme for a boundary integral equations

where singularities of the solution are incorporated into the space of trial functions.
We consider the integral equation of the first kind with the single layer potential on
a polygon Γ

1
V u (z) = − ln | z − ζ | u (ζ ) dsζ = f (z) , z ∈ Γ.
π
Γ
If Γ is sealed such that its capacity is different from 1, then V : H − 2 +σ (Γ ) →

1
1
H 2 +σ (Γ ) is bijective for | σ |≤ 12 , where H s (Γ ) denotes the Sobolev space of
order s on Γ.
There holds the following regularity result: There exist real numbers αj k , natural
numbers rj k and explicitly known singular functions uj k which behave like | z −
zj |αjk ln | z − zj |rjk near the corner zj and are C ∞ elsewhere. If f ∈ H s+1 (Γ )
(s ≥ −1, s ∈ / A (Γ )) , where A (Γ ) ⊂ R is a certain discrete set. Then there exist
numbers Kj s depending on s and cj k depending on f, such that
Kjs

J
u= c j k uj k + u0
j =1 k=1
with u0 ∈ H s (Γ ) .There is an a-priori estimate

Kjs

J
| cj k | + u0 H s (Γ ) ≤ cf H s+1 (Γ ) .
j =1 k=1
The trial spaces Shd,s contain a regular finite element space Shd on a grid with
meshwidth h, namely the smoothest splines of degree d, plus the singular functions
uj k , their number depending on s. Then one has the following approximation
property:
& ' & '
inf u − v H t (Γ ) | v ∈ Shd,s = inf u0 − v 0 H t (Γ ) | v 0 ∈ Shd
Chs−t u0 H s (Γ )
Chs−t f H s+1 (Γ ) (6.123)
for all t s with t < − 12 + σ0 , s d + 1, and s ∈

/ A (Γ ) , and C not depending
on f and h. The Galerkin scheme reads as:
Find uh ∈ Shd,s such that
< τ, V uh > = < τ, f > for all τ ∈ Shd,s . (6.124)
Then (6.123) implies the error estimate

1
u − uh −1 Chs+ 2 f H s+1 (Γ ) . (6.125)
H 2 (Γ )
Now we make the following identifications for − 12 s d + 1, s ∈

/ A (Γ ) :
1 1
X = H − 2 (Γ ) , Y = H 2 (Γ ) , Y1 = H s+1 (Γ ) .
1
Then (6.123) implies (6.119) with δN = Chs+ 2 .
The norm in X1 is given by · H −s−1 (Γ ) , and Theorem 6.28 can be applied.
Theorem 6.29 For the Galerkin scheme (6.124) there holds the error estimate
1
u − uh H −s−1 (Γ ) Chs+ 2 u − uh −1 Ch2s+1 f H s+1 (Γ ) .
H 2 (Γ )
For fixeddegree d of the piecewise polynomials, the highest possible order is

O h2d+3 which is obtained for s = d + 1. The number Jj=1 Kj s of singular
functions has to be chosen correspondingly.
6.8.2 Application to Collocation Methods
In collocation methods, the space TN of test functions is generated by Dirac delta

functions supported by the collocation points. Let us consider a one-dimensional
problem, e.g. a boundary integral equation belonging to a two-dimensional bound-
ary value problem. Thus let Γ ⊂ R2 be a Lipschitz curve.
Let N collocation points
ΔN = {x0 , x1 , . . . , xN } ⊂ Γ with xN = x0
be chosen and
3 4
h := sup | xj − xj +1 | | j = 0, . . . , N − 1 .
Let
S −1 (ΔN ) := span {δ (x − xN ) | n = 1, . . . , N} .
Then S −1 (ΔN ) ⊂ H s (Γ ) for all s < − 12 . Here the Sobolev spaces H s (Γ ) are
defined by transfer from the parameter interval through a fixed periodic parameter
representation. It is well known that the definition of H s (Γ ) is independent of the
specific parameter representation for | s | 1 in the case of a Lipschitz curve Γ, for
| s |< 32 if Γ is piecewise smooth, and for all s if Γ is smooth.
In order to satisfy condition (6.115) for TN = S −1 (ΔN ) we need the following
approximation result [134]
Lemma 6.19 For all p ≤ q ≤ 0 with p < − 12 there exists a constant M,

independent of N , such that
& '
inf ϕ − τ H p (Γ ) | τ ∈ S −1 (N ) ≤ Mhq−p ϕH q (Γ ) (6.126)
for all h > 0 and all ϕ ∈ H q (Γ ) .

Suppose now that
A : H s (Γ ) → H t (Γ ) (6.127)
is a continuous bijective linear operator of norm A , for some s, t ∈ R. Let

f ∈ H t (Γ ) be given and let u ∈ H s (Γ ) solve
Au = f, (6.128)
whereas uh ∈ VN solves the collocation equations
Auh (xn ) = f (xn ) (n = 1, . . . , N) . (6.129)
The trial function space VN is supposed to be a N-dimensional subspace of H s (Γ ) .

The Eq. (6.129) can equivalently be written in the form
< t, Auh > = < t, f > for all t ∈ S −1 (ΔN ) , (6.130)
if t is such that S −1 (ΔN ) ⊂ H −t (Γ ) , i.e. t > 12 .

We want to apply Lemma 6.18 with
X = H s (Γ ) , Y = H t (Γ ) , TN = S −1 (ΔN ) .
In order to make (6.126) and (6.115) equivalent, we thus have to put p = −t. If we
assume −t < q 0, then H t (Γ ) ⊂ H −q (Γ ) , and hence the definition
v X1 := Av H −q (Γ ) for v ∈ H s (Γ ) (6.131)
makes sense and defines a norm, because A is injective. The dual norm is given by
ξ X = A−1 ξ H q (Γ ) , (6.132)

1
as can be seen from

" "
|< ξ, v >| |< ξ, v >|
ξ X = sup | v ∈ X = sup | v ∈ H s (Γ )
1 v X1 Av H −q (Γ )
" "
|< ξ, A−1 w >| |< A−1 ξ, w >|
= sup | w ∈ H (Γ ) = sup
t
| w ∈ H (Γ )
t
w H −q (Γ ) w H −q (Γ )
= A−1 ξ H q (Γ ).
Here we used the fact that H t (Γ ) is dense in H −q (Γ ) .

Now all hypotheses of Lemma 6.18 are satisfied, and we obtain the following
theorem.
Theorem 6.30 Let A be as in (6.127) and let u and uh satisfy (6.128) and (6.129),
respectively. Let t and q satisfy
1
t> and − t q 0,
2
and · X1 be defined by (6.131). Then
u − uh X1 Cht +q u − uh H s (Γ ) , (6.133)
where C = M A with the constant M from Lemma 6.19 for p = −t.

Remark 6.10 If A satisfies an a-priori estimate of the form
v H r (Γ ) γ Av H −q (Γ ) (6.134)
for some r and γ , then the error estimate (6.133) can also be written in the form
u − uh H r (Γ ) Cht +q u − uh H s (Γ ) . (6.135)
The a-priori estimate (6.135)holds in particular if A is a bijective elliptic pseudod-

ifferential operator of oder s − t. Then r = s − t − q, and one obtains
u − uh H r (Γ ) Chs−r u − uh H s (Γ ) (6.136)
for s − t r s.

The highest order O ht in Theorem 6.30 is obtained for q = 0. This corresponds
to r = s − t in (6.136).
This version of the duality argument based on the a-priori estimate (6.134) is
equivalent to the arguments in [258]. Note, however, that it could not be applied
in the previous section for the augmented Galerkin procedures. There, due to the
presence of singular solutions, the estimate corresponding to (6.134) does not hold.
As an example we present from [131] error estimates for the collocation method
applied to the integral equation of the second kind with the double layer pontential
on a piecewise smooth curve:
Au := (I + K) u = f,
with

1 ∂
Ku (z) = − u (ζ ) ln | z − ζ | dsζ (z ∈ Γ ) .
π ∂nζ
Γ
It is known that I +K : H s (Γ ) → H s (Γ ) is continuous and bijective for 0 ≤ s ≤ 1

if Γ is Lipschitz, and for | s − 12 |< σ0 with
"
π 1
σ0 := min | j = 1, . . ., J ∈ ,1
π+ | π − ωj | 2
if Γ is piecewise smooth .
In [131] for the nodal collocation method with piecewise linear trial functions it
is shown that
u − uh H 1 (Γ ) ≤ chσ f H 1+σ (Γ ) ,
if f ∈ H 1+σ (Γ ) and σ < σ0 − 12 . Now Theorem 6.30 yields u − uh L2 (Γ ) =

1
O h 2 +σ0 −ε ∀ε > 0.
6.9 A Collocation Method Interpreted as (GM)
Following Arnold and Wendland [3] we consider a planar Jordan curve Γ with a
regular parameter transformation,

Γ : z = z1 (t), z2 (t) ∼= z1 (t) + i z2 (t)

where z is 1-periodic on R and dz dt = 0. Thus there is a 1–1 correspondence
between functions defined on Γ and 1-periodic functions on R. Therefore the
analysis to follow is based on the periodic Sobolev spaces H s (s ∈ R) that are
defined as the closure of all smooth real-valued 1-periodic functions with respect to

f s := f H s := {|fˆ0 |2 + |fˆk |2 |2πk|2s }1/2
0=k∈Z
where fˆk = 0 e−2πikt f (t)dt (k ∈ Z) (as in Chap. 3).

1
The associated scalar product is

< f, g >s := fˆ0 · ĝ 0 + fˆk ĝ K |2πk|2s
0=k∈Z
6.9 A Collocation Method Interpreted as (GM) 193
what can be extended to a duality pairing on H s+α × H s−α for arbitrary α ∈ R. By

this duality, v ∈ H s+α
< v, w >s
vs+α = sup
w∈H s−α ws−α
holds for any v ∈ H 5+α .

Remark 6.11 Note that for j ∈ N, f (j ) = ( dt
d j
) we have by partial integration
1
ˆ) =
f (j e−2πikt f (j ) (t)dt
k
0
1
= (2πik)j e−2πikt f (t)dt = (2πik)j fˆk ,
0
ˆ ) = 0. Hence by the Parseval identity

by periodicity, in particular f (j 0

f 2j = |fˆ0 |2 + |fˆk |2 |2πk|2j
0=k∈Z

= |fˆ0 |2 + |f ˆ(j ) k |2 = f 2L2 + f (j ) 2L2 .
k∈Z
Again, we want to approximate the equation Au = f . Here we take A : H j +α →

H j −α ;
the number 2α is called the order of A.
(e.g. in the case A = V , we have α = − 12 , and the oder is −1),
whereas j ∈ R is specified later.
To begin with the approximation, we fix a mesh = {t1 = 0 < t2 . . . < tN <
1} (N fixed ∈ N), which is periodically extended by tl+N = tl + 1 (∀l ∈ Z).
Then Sd () denotes the space of 1-periodic, (d − 1)times continuous differen-
tiable spline functions of degree d subordinated to the mesh .
In the following we assume that the degree d is odd (> 0), j := (d + 1)/2 ∈ N.
Then the collocation method (CM) reads: Find u inSd () such that
(Au )(tl ) = f (tl ) l = 1, . . . , N
which is a linear equation system in the unknown ξl in the ansatz u = N l=1 ξl μl .

Now let us discuss the setting of (CM) in the scale of Sobolev spaces. We have
Au ∈ ASd () ⊂ AH s ⊂ H s−2α
provided s < d + 12 .
(CM) makes only sense, if Au is continuous in the mesh points tl , that is, for
some s < d + 12 the space H s−2α should be embedded in the space of continuous
functions. Sobolev’s embedding theorem forces s − 2α > 0 + 21 (n = 1), that is,
the following assumption
1
(A1) d = 2j − 1 > 2α ⇔ j − α >
2
(that is in the case A = V the simplest trial space is S1 () consisting of piecewise
linear spline functions)
Further we impose the following assumptions:
(A2) A : H j +α → H j −α is bijective
(A3) < Au, u >j ≥ γ u2j +α − < Ku, u >j , ∀u ∈ H j +α ,
where γ > 0, K: H j +α → H j −α is compact (linear), what is called a Gårding

inequality.
The key of the convergence analysis of Arnold and Wendland [3] is to refor-
mulate (CM) as a Galerkin method (GM) using partial integration. To this end,
introduce
1
N
1
J u := u(t) dt, Ju := (tl+1 − tl−1 ) u(tl ).
2
0 l=1
Note that the latter functional is the numerical approximation of the first integral by
the trapezoid rule, since

N
1 1 N
1
N+1
u(tl+1 )+u(tl ) tl+1 −tl = u(tl )(tl+1 −tl )+ u(tk )(tk −tk−1 ) = Ju
2 2 2
l=1 l=1 k=2
Theorem 6.31 [3] Let w ∈ H j −α . Then
(i) w(tl ) = 0 for l = 1, . . . , N
if and only if
(ii) < w − J w + Jw , χ >j = 0 for all χ ∈ Sd () .
Let us point out that in virtue of (A1), the values w in (i) and (ii) are defined; by
Sd () ⊂ H j +α , the scalar product in (ii) is defined.
Proof For any real-valued (no restriction of generality) f, g ∈ H j , (see the first
remark above)
1 1 1 1 j
2 1
j
2
d d
< f, g >j = f dt g dt + f (t) g(t) dt .
dt dt
0 0 0
Hence by partial integration in (ii),
1
j −1
< w − J w + Jw , χ >j = Jw J χ + (−1) w χ (2j −1) (t) dt.
0
On the other hand,

2j −1
d
: {v ∈ Sd () | J v = 0} → {ṽ ∈ S0 () | J ṽ = 0}
dt
is an isomorphism onto the space of piecewise constants subordinated to the mesh

with integral mean zero. In the range space we specify
⎧ −1
⎨ −hi
⎪ for t ∈ [ti−1 , ti )
χ̃(t) := h−1 t ∈ [ti , ti+1 )
⎪
⎩
i+1
0 else on [0, 1], periodically extended to R
(2j −1)
where hi = ti − ti−1 . Let χi ∈ Sd () such that χ̃ = χi and J χi = 0.
and plug these special functions in the formula above to obtain
# −1 $
< w−J w+Jw , χi >j = (−1)j −1 h−1
i+1 w(ti+1 )−w(ti ) −hi w(ti )−w(ti−1 )
Hence and from (ii) it follows for some constant κ

h−1
i w(ti ) − w(ti−1 ) = κ ∀i ∈ Z
and i = N + 1, N, . . . , 1 gives
w(tN+1 ) = κhN+1 + w(tN )

= ...
= κ(hN+1 + hN + . . . + h1 ) + w(t0 ) = κ + w(t0 )

tN+1 −t0 =1
Since w is 1-periodic, in particular w(t0 ) = w(0) = w(1) = w(tN+1 ), κ

vanishes.
Therefore (ii) implies
w(ti ) = w(t0 ) ∀i ∈ Z . (6.137)
Now (ii) does not hold only for these special χi , but also for χ = 1 ∈ Sd (), hence
< w − J w + Jw , 1 >j = Jw · 1 = 0 . (6.138)
Evidently (i) follows from (6.137) and (6.137).

Conversely (i) implies (6.137) and (6.137), hence (ii), since {χi , 1} is a basis of
Sd ().

We can understand J and j in (ii) as operators, since J u, Ju can be considered
as constant functions with the values J u, Ju respectively.
Thus formula (ii) gives rise to the operator A := (1 − J + J )A : H j +α →
H −α and by the theorem above, we have u ∈ Sd () solves the (CM) equations
j

⇔ A(u − u) (tl ) = 0 (∀l)
⇔ < A (u − u), v >j = 0 ∀v ∈ Sd ()
⇔ < A u , v > j = < A u, v >j ∀v ∈ Sd ()
what is a Galerkin method!

Lemma 6.20 The operator A = (1 − J + J )A is invertible with the inverse
A−1 = A−1 (1 + J − J ) .
Furthermore, there exists a positive constant C such that
A j +α,j −α + A−1 j −α,j +α ≤ C
for every mesh .

1
Proof For any s ∈ R, J 2s = ( Ju dt)2 = (Ju )2 ≤ u2∞ , where the latter
0
estimate is a consequence of the above formula of J (trapezoid rule).
Moreover by j − α > 21 , H j −α ⊂ C 0 and J : H j −α → H j −α , are uniformly
bounded. By continuity of A, the operators A are uniformly bounded, too.
Since
J J = JJ = J and J J = J J = J
we verify
A A−1 = (1 − J + J ) (1 + J − J )
= 1 − J + J + J − J + J − J + J − J = 1
Thus again the uniform boundedness of the operators A−1 follows from the
continuity of A−1 .

Theorem 6.32 There exist positive constants C and h0 such that, for any mesh
with h := max(tl − ll−1 ) ≤ h0 there holds the stability estimate
inf sup < Av , z >j ≥ C .

v ∈ Sd () z ∈ Sd ()
vj +α = 1 zj +α = 1
Proof Since according to (A3), A satisfies a Gårding inequality, the fundamental

Theorem 6.1 on the geneneral Galerkin method applies and yields the estimate
w j +α ≤ C wj +α (6.139)
for the solution w inSd () of the Galerkin equations
< Aw , v >j =< Aw, v >j ∀v ∈ Sd () ,
where the mesh is arbitrary, but h ∈ (0, h0 ) with appropriate positive constans
C and h0 . Thus we obtain the continuity of the Galerkin operator. Now we consider
the solution u inSd () of the equations
< A u , v >j = < Au , v >j ∀v ∈ Sd ()
and rewrite these equations as follows:
< Au , v >j = < Au − {(J − J )Au }, v >j

= < Au − {(J − J )Au − (J − J )Au}, v >j
= < A(u − A−1 {(J − J )Au − (J − J )Au}), v >j .
Thus by (6.138) we arrive at
u |j +α ≤ c (uj +α + A−1 (J − J )Au j +α + A−1 (J − J )Auj +α ))

≤ c (uj +α + c1 (J − J )Au j −α + (Jα − J )Auj −α )
μ μ
≤ c uj +α + c3 h u j −α + c3 h uj +α ,
where the approximation by the trapezoid rule provides the estimate

μ
(J − J )vj −α ≤ c2 h vj −α
for v = Au , Au respectively with μ > min(2, j − α) > 0. Therefore for any
μ
h ∈ (0, h0 ] , where c3 h ≤ min(c, 12 )„ there holds uniformly with respect to h
u j +α ≤ 3c uj +α .
Finally the fundamental Theorem 6.1 entails the claimed stability estimate.

Thus in virtue of Theorem 6.1, Part iii) we obtain the following convergence
result.
Theorem 6.33 There exist positive constants C and h0 such that, for any mesh
with h ≤ h0 , there exists a unigue solution u ∈ Sd () of the (CM) equations and
there holds the quasioptimal error estimate
u − u j +α ≤ C inf{u − vj +α : v ∈ Sd ()} .
6.10 Modified Collocation and Qualocation
Following Costabel and Stephan [131] we again consider odd degree spline
functions in the collocation method, but we dispense with the smoothness of the
boundary Γ . Instead more generally, Γ is assumed to be a connected closed planar
curve patched together from smooth arcs Γ j (j = 1, . . . , J ), that intersect each
other in the corners zj at the inner angles ωj ∈ (0, 2π).
In what follows, we use the subsequent definition of the Sobolov spaces H s (Γ ):
1
for any s > 0, the set of the restrictions of functions in H s+ 2 (R2 ) to Γ
(this makes sense by the embedding theorem that ensures u ∈ H s (Γ ) ⇔
1
∃ extension ũ ∈ H s+ 2 (R2 ))
for any s < 0, by duality H s (Γ ) := H −s (Γ ) and H 0 (Γ ) := L2 (Γ ).
Here as a simple instance of the convergence analysis we study “Symm’s” integral
equation V u = f where V ist the simple layer potential

1
V u(z) := − u(ζ ) ln |z − ζ | ds(ζ )
π
Γ
and V : H s (Γ ) → H s+1 (Γ ) is known to be continuous and bijective.

To describe the collocation method, let N = {x1 , . . . , xN } ⊂ Γ. be a mesh, which
contains the corner points and where the points xj are nodal points of the ansatz
functions and collocation points as well.
6.10 Modified Collocation and Qualocation 199
Let S 1 (N ) denote the N dimensional space of spline functions of order 1

subordinate to the mesh N ; that is, u ∈ S 1 (N ), if and only if u is continuous on Γ
and is a linear function of the arc length on each segment xn xn+1 , n = 0, . . . , N −1,
where x0 := xN . Let the mesh parameter h :=max{|xn+1 − xn | : n = 0, . . . , N −
1} → 0 (N → ∞).
Then the collocation method (CM) reads:
Find uN ∈ S 1 (N ) such that
V uN (xn ) = f (xn ) n = 1, . . . , N
that is
< V uN , tN > = < f, tN > ∀tN ∈ S −1 (N ) ,

where we use S −1 (N ) := span{δ(x − xn ) n = 1, . . . , N}, the linear space
spanned up by the Dirac functionals in the mesh points.
To obtain convergence results, we intend to apply the principal Theorem 6.11.
If we define QN = Q := D 2 , that is, the second derivative (in the sense of
distributions) with respect to the arc length, then we have
QS 1 (N ) ⊂ S −1 (N ).
Thus in this setting of Q we are led to define
X := H 1/2 (Γ ), LN := S 1 (N ), TN := S −1 (N ), A = V
however, generally < Qv, Av >= ∞, since Q does not map into Y = (AX) ⊂
H −3/2(Γ ) (compare, in contrast, assumption 2)!
Therefore we have to modify the setting and introduce the space
0 1/2
H (Γ ) := {u ∈ H 1/2 (Γ ) ∀ j = 1, . . . , J ∃ũj ∈ H 1/2(Γ ) :

ũj Γ j = u Γ j ; ũj Γ \Γ j = 0} ;
with the norm

J
u20 1/2 := ũj 2H 1/2 (Γ ) .
H (Γ ) j =1
0 1/2
Then H (Γ ) is the completion of C0∞ (Γ \{z1 , . . . , zJ }) with respect to this norm
0 1/2 dense
and H (Γ ) ⊂ H 1/2(Γ ). The associated ansatz space is now
0
S (N ) := {v ∈ S 1 (N ) v(zj ) = 0 (j = 1, . . . , J )}
0
where we need that {z1 , . . . , zJ } ⊂ N . Thus dim S (N ) = N − J . Let us fix J
functions η1 , . . . , ηJ ∈ H 3/2(Γ ) such that
ηj (zk ) = δj k (j, k = 1, . . . , J )
0 1/2
and introduce the projection R : H 3/2 → H 3/2 (Γ ) ∩ H (Γ ) by

J
Rg(z) := g(z) − g(zj )ηj (z) .
j =1
Hence the adjoint operator R acts on S −1 (N ) as follows
< R δ(z − xk ), g > = < δ(z − xk ), Rg >= (Rg)(xk )

J
= g(xk ) − g(zj )ηj (xk )
j =1
J
= < δ(z − xk ) − ηj (xk )δ(z − zj ), g > .
j =1
0
With LN := S (N ), TN := R S −1 (N ), this fits in our abstract setting of Sect. 6.4,
0
and the modified collocation method reads: Find uN ∈ S (N ) such that
J J
V uN (xn ) − V uN (zj )η(xn ) = f (xn ) − f (zj )ηj (xn ) (n = 1, . . . , N) .
j =1 j =1
The reader is cautioned that these modified collocation equations follow from the
above collocation equations, but not vice versa!
0 1/2
With X := H (Γ ), A := V , Y := AX, QN ≡ Q = R D 2 the principal
theorem 6.11 (Proof of the Gårding inequality by localization to a reference angle
Γω and by application of the Mellin transformation) yields:
0
Theorem 6.34 ∀ N ≥ N0 ∃1 uN ∈ S (N ) that solves the modified collocation
equations and satisfies uN 0 1/2 ≤ Cu 0 1/2 ;
H (Γ ) H (Γ )
0
u − uN 0 1/2 ≤ C inf{u − v 0 1/2 : v ∈ S (N )}.
H (Γ ) H (Γ )
A drawback of this convergence analysis is the required smoothness assumption
0 1/2
u ∈ H (Γ ) ⊂ H 1/2(Γ ) for the solution u of the considered integral equation. In
view of the corners, this is not always a realistic assumption, even with a smooth
right hand side f . Instead introducing
π π 1
αj := min{ , } ∈ ( , 1) for j = 1, . . . , J
ωj 2π − ωj 2
1
α0 := min{αj |j = 1, . . . , J } ∈ ( , 1)
2
the solution u behaves like O(|z − zj |aj −1 ) near the corners.

Therefore in general, u ∈ / H s (Γ ) for s ≥ α0 − 12 ∈ (0, 12 ).
To cope with this local loss of smoothness, one chooses a weight function ρ ∈
C ∞ (R2 \{z1 , . . . , zJ }) with ρ(z) = |z − zj | in a neighborhood of zj (j = 1, . . . , J )
and introduces the weighted Sobolev space
1 1
Hρ1/2(Γ ) := H 2 (Γ ) = {u ρu ∈ H 1/2(Γ )}
ρ
and accordingly the ansatz space
1 0
Sρ (N ) := S (N )
ρ
and the test space
0 −1
S (N ) = {ϕ ∈ S −1 (N )| supp ϕ ∩ {z1 , . . . , zJ } = ∅} .
By doing so, one obtains convergence and the asymptotic convergence estimate with
respect to the norm .H 1/2 .
ρ
In 1988, I. H. Sloan [378] presented a quadrature-modified collocation method
and coined the term “qualocation method” as a short name for this new method.
If for collocation methods the number of collocation points surmounts the degrees
of freedom of the trial functions then appropriate projection composed with the
overdetermined system of linear equations leads to qualocation equations [431]
where high rates of convergence can be elaborated [383–385].
In the following, we briefly sketch this method for the solution of integral
equations Au = f . In addition to the ansatz space Lh , and the test space Th a
quadrature formula Qh , comes into play, and the qualocation method reads:
Find uh ∈ Lh such that
< Auh , t >h = < f, t >h ∀t ∈ Th ,
where < v, w >h := Qh (vw) (and as usual, w denotes the conjugate complex
function to w).
More explicitly with Lh = span {v1, . . . , vN }, Th = span{t1, . . . , tN }, one has

to solve the subsequent linear system:

N
< Avj , tk >h ξj = < f, tk >h (k = 1, . . . , N) .
j =1
This extends the collocation method, for using the quadrature formula

M
Qh g = wl g(xl ) wl > 0, wl = 1, xl ∈ Γ
l=1
the qualocation equations can be rewritten in the case M = N as

N
wl [AuN (xl ) − f (xl )] t k (xl ) = 0 (k = 1, . . . , N) ,
l=1
what is equivalent to the equations
AuN (xl ) = f (xl ) (l = 1, . . . , N) ,
only provided the matrix {t k (xl )}k,l=1,...,N is nonsingular.

In [385] strongly elliptic boundary integral equations
Lu := (b+ L+ + b− L− + K)u = f
are considered on a smooth curve Γ in a space of 1−periodic functions, where

b± ∈ C and L± are operators given in Fourier series form by

L+ u(x) = û(0) + |n|β û(n)e2πinx
n=0

L+ u(x) = û(0) + (sign)|n|β û(n)e2πinx
n=0
with β ∈ R (the order of the pseudodifferential operator) and a smoothing operator

K. Note, for β = −1 L+ is the logarithmic-kernel operator. The qualocation method
in [385] is defined with a uniform grid {xk = kh, k ∈ Z} and the trial space SN of
smoothest periodic splines of order r ≥ 1 (i.e. degree ≤ r − 1 and r − 2 continuous
of order r ≥ 1. Together with a quadratic rule
derivatives) and the test space SN

N−1 J
QN g = h ωj g(xk + hξj )
k=0 j =1
where 0 ≤ ξ1 < ξ2 < . . . < ξJ < 1, Jj=1 ωj = 1, ωj > 0 there is associated a

discrete inner product u, v = QN (uv) which approximates the exact inner product
1
(u, v) = u(x)v(x)dx.
0
With this notation the qualocation method reads: Find uN ∈ SN such that
Luh , v = f, v for all v ∈ SN

(6.140)
Note if J = 1 this is equivalent to −collocation, analyzed by Schmidt [365].

Note, the Galerkin method (obtained by setting SN = S and replacing u, v
N
by (u, v)) is stable and convergent for a subset of the above equations namely for
strongly elliptic operator equations. The following definition, adapted from [408] is
taken appropriate in [385].
Definition 6.3 L0 := b+ L+ + b− L− is strongly elliptic if there exists θ ∈ C such
that
%[θ (b+ + b− )] > 0 and %[θ (b+ − b− )] > 0
As a basis of SN take {ψμ : μ ∈ ΛN }, where

N−1
ψμ = aμ e2πiμxk bk , μ ∈ ΛN
k=0
and
N N
ΛN = {μ ∈ Z : − < μ ≤ },
2 2
similarly {ψN : μ ∈ ΛN } for SN
. Then the qualocation method becomes: Find
uh = μ∈ΛN ûh (μ)ψN such that
Luh , ψμ = f, ψμ for all μ ∈ ΛN .
To derive an analysis of (6.140) first the simplified equation L0 u = f is considered.

As worked out in [96] there holds
%
1 if μ = 0
L± ψμ , ψμ =
[μ]± μ
β D± ( N ) if μ ∈ ΛN \{0}
where
% %
1, n=0 1, n=0
[n]+
β = , [n]−
β =
|n|β , n = 0 (sign)|n|β , n = 0
and

J
D± (y) = ωj [1 + Ω± (ξj , y)][1 + (ξj , y)], y ∈ [−1/2, 1/2]
j =1
with
% 0
y r Fr+0 (ξ, y), if r even
(ξ, y) =
y r0 Fr−0 (ξ, y) if r odd
e2πilx sign l
Fα+ (x, y) = Fα− (x, y) = e2πilx
|l + y|α |l + y|α
l=0 l=0
and
%
+
|y|r−β Fr−β (ξ, y) if r even +, r odd (−)
Ω + (ξ, y) = −
(−) (sign y)|y|r−β Fr−β (ξ, y) if r odd +, r even (−)
With
D(y) := b+ D+ (y) + b− (sign y)D− (y), y ∈ [−1/2, 1/2]
the qualocation method (6.140) is called stable if
inf{|D(y)| : y ∈ [−1/2, 1/2]} > 0
The qualocation method is of order r − β + b if
E(y) := b+ E+ (y) + b− (sign y)E− (y) = O(|y|r−β+b ) for y ∈ [−1/2, 1/2]
where E± (y) = Ω± (ξj , y)[1 + (ξj , y)].

As a generalization of [96] Sloan and Wendland derive in [385] the following
convergence result by first analysing the Fourier coefficients of L0 and then applying
a standard perturbation argument to include K.
Theorem 6.35 Let the qualocation method (6.140) be stable and of order r − β +
b, b ≥ 0, then for all N sufficiently large uh is uniquely defined. Moreover, for all
s, t satisfying s < r − 1/2, β + 1/2 < t, β − b ≤ s ≤ t ≤ r there holds
uh − us ≤ cht −s |u|t +max{β−s,0}
The following theorem characterises qualocation methods that are stable for
strongly elliptic operators.
6.11 Radial Basis Functions and Spherical Splines 205
Theorem 6.36 ([385]) For the qualocation method (6.140) with a symmetric
quadrature rule with positive weights there holds:
Assume r, r are of the same parity and if J = 1 that ξ1 = 1/2 if r, r even and
ξ1 = 0 if r, r odd. The method (6.140) is stable for all strongly elliptic operators
L0 if and only if
D+ (y) ≥ |D− (y)| for all y ∈ [0, 1/2].
In [385][Section 5] a list of qualocation methods is given that are stable for all
strongly elliptic operators: For example for the logarithmic-kernel operator (single
layer potential) one can take for r = r = b = 1 the 2 point rule G2,1,2 of order 3
with ξ1 = 0.2113248654051872, ξ2 = 0.7886751345948128 and ω1 = ω2 = 1/2
(This rule integrates exactly all polynomials of degree ≤ 2). For qualocation under
reduced regularity see [382, 419].
Finally, note that as long as the order α of the operator is not zero, the condition
numbers of the discrete equations are unbounded independent of the sign of α.
Hence in order to use iteration schemes for scaling the discrete conditions, suitable
preconditioners must be applied (see e.g [255, 420]). For related work of Langer
and Steinbach on boundary element tearing and interconnecting methods we refer
to [280].
6.11 Radial Basis Functions and Spherical Splines
Radial basis functions are used in [418] to define approximate solutions to boundary
integral equations on the unit sphere. These equations arise from the integral
reformulation of the Laplace equation in the exterior of the sphere, with given
Dirichlet or Neumann data, and a vanishing condition at infinity. Radial basis
functions yield a meshless method which is especially suitable to handle sattelite
data.
In the following we consider boundary integral equations on the unit sphere. Let
S denote the unit sphere in R3 , i.e., S := {x ∈ R3 : x = 1}, and Be the exterior of
the sphere, i.e., Be := {x ∈ R3 : x > 1}, where x denotes the Euclidean norm
in R3 . We now follow [418] and consider the Laplace equation
ΔU = 0 in Be , (6.141)
with either a Dirichlet boundary condition
U = UD on S, (6.142)
or else a Neumann boundary condition
∂ν U = ZN on S, (6.143)
where ∂ν = ∂/∂ν denotes differentiation in the direction of the outward unit

normal ν, and the vanishing condition at infinity for both the Dirichlet and Neumann
cases is
U (x) = O(1/x) as x → ∞. (6.144)
The solutions of these problems can be represented in terms of spherical

harmonics Yl,m , m = −l, . . . , l and l = 0, 1, . . .. They form an orthonormal basis
for L2 (S). For any function v ∈ L2 (S), its associated Fourier series,
∞
l
v= >
vl,m Yl,m (θ, ϕ), where >
vl,m = v(θ, ϕ)Yl,m (θ, ϕ)dσ,
l=0 m=−l S
(6.145)
converges in L2 (S). Here dσ is the element of surface area.
It is well-known that if the Dirichlet data UD has an expansion as a sum of
spherical harmonics
∞
l
UD (θ, ϕ) =
(U D )l,m Yl,m (θ, ϕ),
l=0 m=−l
then (see [323, Theorem 2.5.1]) the Dirichlet problem (6.141), (6.142) and (6.144)
has the unique solution
∞
l
1
U (r, θ, ϕ) = (UD )l,m Yl,m (θ, ϕ). (6.146)
r l+1
l=0 m=−l
Similarly, if
∞
l
ZN (θ, ϕ) =
(Z N )l,m Yl,m (θ, ϕ),
l=0 m=−l
then (see [323, Theorem 2.5.2]) the Neumann problem (6.141), (6.143) and (6.144)
has the unique solution
∞
l
1
U (r, θ, ϕ) = − (Z N )l,m Yl,m (θ, ϕ). (6.147)
(l + 1)r l+1
l=0 m=−l
Note that the spherical harmonic basis functions in (6.146), (6.147) are global.
In contrast, in [418] we use spherical basis functions obtained from compactly
supported radial basis functions, which are better able to capture local properties
of the solutions. We shall propose a solution process in which the boundary
value problems are reformulated in terms of boundary integral equations on S, the

solutions of which are then approximated by spherical basis functions.
Next we reformulate the boundary value problems (6.141)–(6.144) as boundary
integral equations. For s ∈ R, the Sobolev space H s (S) is defined as usual (see e.g.
[323]) with norm and Hermitian product given by
(∞ l )1/2

vs := (l + 1) |>
vl,m |
2s 2
(6.148)
l=0 m=−l
and
∞
l
v, w s := (l + 1)2s> >l,m .
vl,m w
l=0 m=−l
Note that
|v, w s | ≤ vs ws ∀v, w ∈ H s (S), ∀s ∈ R, (6.149)
and
v, w s1 +s2
vs1 = sup 2
ws2 ∀v ∈ H s1 (S), ∀s1 , s2 ∈ R. (6.150)
w∈H s2 (S)

w=0
The single-layer potential S and the double-layer potential D are defined by

1 1 1 ∂ 1
Sv(x) = v(y) dσy , Dv(x) = v(y) dσy ,
2π S x − y 2π S ∂νy x − y
for x ∈ Be . Associated with these potentials, we define the following boundary

integral operators

1 1
V v(x) = v(y) dσy
2π S x − y

1 ∂ 1
Kv(x) = v(y) dσy
2π S ∂νy x − y

1 ∂ 1
K ∗ v(x) = v(y) dσy
2π ∂νx S x − y

1 ∂ ∂ 1
W v(x) = − v(y) dσy ,
2π ∂νx S ∂νy x − y
for x ∈ S. The traces and normal derivatives on S of S and D are given by (see
Lemma 2.2 noting that the limits are taken from the exterior of S)
(Sv)|S = V v and ∂ν (Sv) = −v + K ∗ v, if v ∈ H −1/2(S),
and
(Dv)|S = v + Kv and ∂ν (Dv) = −W v if v ∈ H 1/2(S).
If U ∈ Hloc1 (B ) satisfies (6.141) and (6.144), then using the single-layer and
e
double-layer potentials, and Green’s theorem we can represent U as (see Chap. 2)
1 1
U= D(U |S ) − S(∂ν U ) in Be , (6.151)
2 2
allowing us to compute U from a knowledge of both U |S and ∂ν U . In fact, by taking
the trace on both sides of (6.151) we obtain, after rearranging the equation,
V (∂ν U ) = −U |S + K(U |S ) on S.
Similarly, by taking the normal derivative of both sides of (6.151) we find
W (U |S ) = −∂ν U − K ∗ (∂ν U ) on S.
Therefore, the Dirichlet problem (6.141), (6.142) and (6.144) is equivalent to
Vz = f on S, where f = −UD + KUD , (6.152)
and the Neumann problem (6.141), (6.143) and (6.144) is equivalent to
Wu = g on S, where g = −ZN − K ∗ ZN . (6.153)
Due to (6.151), the solution U of the Dirichlet problem can be computed from the
solution z of (6.152) by
1 1
U= DUD − Sz,
2 2
and the solution of the Neumann problem can be computed from the solution u
of (6.153) by
1 1
U= Du − SZN .
2 2
Equation (6.152) is a weakly singular integral equation and equation (6.153) is
a hypersingular integral equation. In the following we present efficient algorithms
to solve these equations . Note that V and W are pseudo-differential operators of

order −1 and 1, respectively. They have the following representations in terms of
spherical harmonics (see [323, page 122]):
∞
l
1
Vv = 2 >
vl,m Yl,m , (6.154)
2l + 1
l=0 m=−l
∞
l
l(l + 1)
W v = −2 >
vl,m Yl,m . (6.155)
2l + 1
l=0 m=−l
Now we define weak solutions to (6.152) and (6.153). It is well-known [304, 323]
that V : H −1/2 (S) → H 1/2(S) and W : H 1/2(S)/R → H −1/2(S) are bijective,
implying that (6.152) has a unique solution for all f ∈ H 1/2(S), and (6.153) has a
unique solution up to a constant for all g ∈ H −1/2 (S). Defining the bilinear forms
aV (v, w) := V v, w 0 ∀v, w ∈ H −1/2 (S)
and
aW (v, w) := −W v, w 0 ∀v, w ∈ H 1/2(S),
we seek weak solutions to equations (6.152) and (6.153) respectively as follows:
z ∈ H −1/2 (S) : aV (z, v) = f, v 0 ∀v ∈ H −1/2 (S), (6.156)
and

u ∈ H 1/2(S) : u(x) dσx = 0 and aW (u, v) = −g, v 0 ∀v ∈ H 1/2(S).
S
(6.157)
We note from (6.154) and (6.155) that
aV (v, v) 0 v−1/2 2 ∀v ∈ H −1/2 (S) , aW (v, v) 0 v1/2 2 ∀v ∈ H 1/2 (S)/R.

(6.158)
Next we shall approximate the solutions of the above equations with spherical
basis functions. These functions are defined via positive definite kernels.
A continuous function Φ : S × S → C is called a positive definite kernel on S if
it satisfies
(i) Φ(x, y) = Φ(y, x) for all x, y ∈ S;
(ii) for every set of distinct points {x1 , . . . , xM } on S, the M × M matrix A with
entries Ai,j = Φ(xi , xj ) is positive semi-definite.
If the matrix A is positive definite then Φ is called a strictly positive definite kernel.
We shall define the kernel Φ in terms of a univariate function φ : [−1, 1] → R,
Φ(x, y) = φ(x · y) ∀x, y ∈ S.
If φ has a series expansion in terms of Legendre polynomials Pl ,

∞
1
φ(t) = (2l + 1)>
φ(l)Pl (t), (6.159)
4π
l=0
where
1
> = 2π
φ(l) φ(t)Pl (t)dt, (6.160)
−1
then due to the addition formula [323]

l
2l + 1
Yl,m (x)Yl,m (y) = Pl (x · y) ∀x, y ∈ S, (6.161)
4π
m=−l
the kernel Φ can be represented as
∞

l
Φ(x, y) = >
φ (l) Yl,m (x)Yl,m (y). (6.162)
l=0 m=−l
This kernel is called a zonal kernel. The kernel Φ is strictly positive definite if and
only if >
φ (l) ≥ 0 for all l ≥ 0, and >φ (l) > 0 for infinitely many even values of l
and infinitely many odd values of l; see [97]. In the following we shall assume that
>
φ (l) > 0 for all l ≥ 0.
The native space associated with φ is defined by
∞ l
|>
vl,m |2
Nφ := {v ∈ D (S) : vφ 2 = < ∞},
>
φ (l)
l=0 m=−l
where D (S) is the space of distributions defined on S. This space is equipped with
an inner product and a norm defined by
∞ ∞
( )1/2
l
> >l,m
vl,m w l
|>
vl,m |2
v, w = and vφ = .
φ
>
φ(l) >
φ(l)
l=0 m=−l l=0 m=−l
>(l) for l = 0, 1, . . . satisfy

If the coefficients φ
c1 (l + 1)−2τ ≤ >
φ(l) ≤ c2 (l + 1)−2τ (6.163)
for some positive constants c1 and c2 , and some τ ∈ R, then the native space Nφ
can be identified with the Sobolev space H τ (S), and the corresponding norms are
equivalent.
Let X = {x1 , . . . , xM } be a set of data points on the sphere. Two important
parameters characterising the set X are the mesh norm hX and separation radius
qX , defined by
1
hX := sup min θ (xi , y) and qX := min θ (xi , xj ),
y∈S 1≤i≤M 2 i=j
where θ (x, y) := cos−1 (x · y). The spherical basis functions Φi , i = 1, . . . , M,

associated with X and the kernel Φ are defined by
∞
l
Φi (x) := Φ(x, xi ) = > l,m (xi )Yl,m (x).
φ(l)Y (6.164)
l=0 m=−l
Note that if (6.163) holds then Φi ∈ H s (S) for all s satisfying s < 2τ − 1.
Let
φ
VX := span{Φ1 , . . . , ΦM }. (6.165)
φ
We assume that (6.163) holds for some τ > 1 so that VX ⊂ Nφ = H τ (S) ⊂
φ
C(S) and study the approximation property of VX as a subspace of Sobolev spaces.
The following lemma, proven in [418] shows the boundedness of the interpolation
operator in the native space.
φ
Lemma 6.21 ([418]) The interpolation operator IX : C(S) → VX defined by
IX v(xj ) = v(xj ), j = 1, . . . , M, v ∈ C(S), (6.166)
is well-defined, and is a bounded operator in Nφ . In fact, this operator is the Nφ -

φ
orthogonal projection from Nφ onto VX .
Proposition 6.1 ([418]) Assume that (6.163) holds for some τ > 1. For any s, t ∈
R satisfying 0 ≤ t ≤ τ ≤ s ≤ 2τ , if v ∈ H s (S) then the following estimate holds
min{s−t, 2(τ −t )}
IX v − vt ≤ ChX vs .
The convergence analysis for the approximate solutions to (6.156) and (6.157)
φ
requires the following approximation property of VX .
Theorem 6.37 ([418]) Assume that (6.163) holds for some τ > 1. For any s, t ∈ R
φ
satisfying t ≤ τ and t ≤ s ≤ 2τ , if v ∈ H s (S) then there exists η ∈ VX such that
μ
v − ηt ≤ ChX vs , (6.167)
where μ = min{s − t, 2(τ − t), 2τ + |s|}, and where the constant C is independent
of v and hX .
Proof We prove the result by considering different cases of values of s and t.
Case 1: 0 ≤ t ≤ τ ≤ s ≤ 2τ Note that in this case μ = min{s − t, 2(τ − t)}.
We can choose η = IX v yielding (6.167) with s = 2τ (and with s < 2τ by
interpolation).
In the following cases, it is easy to see that s − t ≤ 2(τ − t) and thus μ =
min{s − t, 2τ + |s|}.
Case 2: 0 ≤ t ≤ s < τ
Let L = 1 h1X 2. We define for each v ∈ H s (S) a polynomial of degree L by

L
l
PL v = >
vl,m Yl,m .
l=0 m=−l
With η = IX PL v we have
v − η2t ≤ 2v − PL v2t + 2PL v − IX PL v2t

∞

l
2(τ −t )
≤2 (l + 1)2t |>
vl,m |2 + chX PL v2τ
l=L+1 m=−l
∞

l
=2 (l + 1)2(t −s)(l + 1)2s |>
vl,m |2
l=L+1 m=−l
2(τ −t )

L
l
+ chX (l + 1)2(τ −s)(l + 1)2s |>
vl,m |2
l=1 m=−l
2(τ −t )
≤ c(L + 1)2(t −s)v2s + cL2(τ −s)hX v2s ,
where in the second step we have used the result given in Case 1. Here c is a
generic constant which may take different values at different occurrences. Since
L ≤ h−1
X and (L + 1)
−1 ≤ h , we deduce (6.167) with μ = s − t.
X
Case 3: t < 0 ≤ s ≤ 2τ and Case 4: t ≤ s < 0 see [418].

Optimal estimates are afterwards obtained in[338].
For the approximation of the hypersingular equation (6.157) we use radial basis
functions suggested by [427, page 128]. First we define a smoothing operator I on
the space CK [0, ∞) of continuous functions in [0, ∞) with compact supports by
∞
I : CK [0, ∞) → CK [0, ∞), I v(r) = sv(s)ds, r ≥ 0.
r
For any non-negative integer m, let

%
(1 − r)m+2 , 0 < r ≤ 1,
ρ̃m (r) =
0, r > 1,
and
ρm (r) = I m ρ̃m (r), r ≥ 0.
We define
√
φ (W ) (t) = ρm ( 2 − 2t), t ∈ [−1, 1], (6.168)
and denote by Φi(W ) , i = 1, . . . , M, the corresponding spherical basis functions;

see (6.164). Here the superscript N indicates that the functions are specifically
chosen for equation (6.157) arising from the Neumann problem. We suppress the
(W )
dependence on m in the notation of φ (W ) and Φi because m will be chosen once
(W )
and for all during the whole solution process. The functions Φi , i = 1, . . . , M,
are locally supported radial basis functions. It is proved in [315, Proposition 4.6]
that >
φ (W ) (l) satisfies (6.163) with
τ (W ) = m + 3/2. (6.169)
In Fig. 6.4 we plotted l 2m+3 >φ (W ) (l) to observe the asymptotic behaviour of >
φ (W ) (l)
for m = 0, 1, 2, 3, with >φ (W ) (l) computed by the MATLAB function quadl which
uses an adaptive Lobatto quadrature.
φ (W )
For given X = {x1 , . . . , xM } ⊂ S, let VN := VX . We will solve (6.157)
approximately by solving the Galerkin scheme

uX ∈ V N : uX (x) dσx = 0 and aW (uX , vX ) = −g, vX 0 ∀vX ∈ VN .
S
(6.170)
Using (6.155) and (6.161), one obtains the following formula to compute the
entries of the stiffness matrix from (6.170):
∞

l
l(l + 1)
|φ?
(W ) (W )
aW (Φi , Φj )=2 (W ) (l)|2 Yl,m (xi )Yl,m (xj )
2l + 1
l=0 m=−l
∞
1
= l(l + 1)|φ?
(W ) (l)|2 P (x · x ).
l i j (6.171)
2π
l=0
m=0 m=1
14 1200
1100
12
l 3 φ(N) (l)
l 5 φ(N) (l)
1000
10
900
8 800
0 100 200 300 400 500 0 100 200 300 400 500
l l
5 m=2 7 m=3
x 10 x 10
7 8
6 6
l 7 φ(N) (l)
(l)
9 (N)
5 4
l φ
4 2
3 0
0 100 200 300 400 500 0 50 100 150 200 250
l l
Fig. 6.4 Asymptotic behaviour of >

φ (W ) (l) [418]
The right-hand side in (6.170) is computed by using (6.153), noting K ∗ = −S/2

(see [323, page 122]),
∞
l
l
(ZN )l,m >
(W )
g, Φi 0 = φ (W ) (l)Yl,m (xi ). (6.172)
2l + 1
l=0 m=−l
Theorem 6.37 yields the following a priori error estimate.

Theorem 6.38 ([418]) Let φ (W ) be defined by (6.168) for some non-negative
integer m, and let τ (W ) = m + 3/2. If u is the solution to (6.157) satisfying
u ∈ H s (S), 1/2 ≤ s ≤ 2τ (W ) , and uX the solution to (6.170) then
min{s−1/2, 2τ (W ) −1}
u − uX 1/2 ≤ ChX us .
Proof The condition

u(x) dσx = uX (x) dσx = 0
S S
yields > ?
u0,0 = (u X )0,0 = 0, implying (together with (6.158))
u − uX 21/2 0 aW (u − uX , u − uX ) = aW (u − uX , u − wX )
≤ cu − uX 1/2 u − wX 1/2
for all wX ∈ VN . Now the required estimate follows from Theorem 6.37.

The corresponding optimal results can be found in Theorem 5.4 in [338].
In [418] the weakly singular integral equation is treated as follows. The following
univariate function is used:
∞
1
φ (V )
(t) = (2l + 1)(l + 1)>
φ (W ) (l)Pl (t), (6.173)
4π
l=0
(V )
Let Φi , i = 1, . . . , M, denote the corresponding spherical basis functions. It is
clear that >
φ (V ) (l) satisfies (6.163) with
τ (V ) = τ (W ) − 1/2 = m + 1;
see (6.169).
Letting
φ (V )
VD := VX = span{Φ1(V ) , . . . , ΦM
(V )
},
we approximate the solution z of (6.156) by
zX ∈ VD : aV (zX , vX ) = f, vX ∀vX ∈ VD . (6.174)
The resulting stiffness matrix has entries given as (cf. (6.171))
∞

l
(l + 1)2
aV (Φi(V ) , Φj(V ) ) = 2 |φ?
(W ) (l)|2 Yl,m (xi )Yl,m (xj )
2l + 1
l=0 m=−l
∞
1
= (l + 1)2 |φ?
(W ) (l)|2 P (x · x ).
l i j
2π
l=0
The right-hand side of (6.174) is computed by using (6.152) (see [323, page 122]),
∞
l
(l + 1)2
f, Φi(V ) =− (UD )l,m >
φ (W ) (l)Yl,m (xi ). (6.175)
2l + 1
l=0 m=−l
A priori error estimates similar to those in Theorem 6.38 can be proved.

Theorem 6.39 ([418]) Let φ (W ) be defined by (6.168) for some positive integer
m, φ (V ) be defined by (6.173), and τ (V ) = m + 1. If z is the solution to (6.156)
satisfying z ∈ H s (S), −1/2 ≤ s ≤ 2τ (V ) , and zX the solution to (6.174), then
s+1/2
z − zX −1/2 ≤ ChX zs .
Proof We note that approximation property requires m > 0, and that
1 1 1
min{s + , 2(τ (V ) + ), 2τ (V ) + |s|} = s + .
2 2 2
Since
z − zX 2−1/2 0 aV (z − zX , z − zX ),
the remainder of the proof is similar to that of Theorem 6.38, and is therefore
omitted.

Next we present numerical results obtained from experiments with the set of
scattered points X generated by a simple algorithm [354] which partitions the sphere
into equal areas; see Fig. 6.5; for detais see . The sets of points we used have number
of points M = 20, 30, 40, 50, 100, 500, and 1000.
The spherical basis functions Φi(W ) , i = 1, . . . , M, are defined by (6.164) using
the univariate function φ (W ) given by (6.168) with m = 0, 1, 2. The coefficients
>
φN (l) with l = 1, . . . , 500 are computed by the MATLAB function quadl
(V )
which uses an adaptive Lobatto quadrature. The spherical basis functions Φi ,
i = 1, . . . , M, are defined by (6.164) with φ (V ) given by (6.173).
In [418] the exterior Neumann problem (6.141), (6.143) and (6.144) is considered
with a boundary data given by
0.5x3 − 1
ZN (x) = ,
(1.25 − x3 )3/2
so that the exact solution is

1
U (x) = with p = (0, 0, 0.5).
x − p
Here x = (x1 , x2 , x3 ). Due to (6.151) and (6.153), the exact solution to (6.157) is
given by u = U |S . Let n = (0, 0, 1). By using the identity (see [323, page 20])
∞

2 −1/2
(1 − 2t cos θ + t ) = t l Pl (cos θ ), t < 1,
l=0
and the addition formula (6.161), one obtains,

Fig. 6.5 Scattered points in equally partitioned areas
for x = (sin θ cos ϕ, sin θ sin ϕ, cos θ ) ∈ S,
1 ∞
1 1
u(x) = =√ = Pl (x · n)
x − p 1 − cos θ + 1/4 2l
l=0
∞
l
4π
= Yl,m (n)Yl,m (x),
+ 1)
2l (2l
l=0 m=−l
so that
4π
>
ul,m = Yl,m (n). (6.176)
+ 1)
2l (2l
Now (6.170) is solved and the approximate solution uX is compared with the
exact solution u. Note that

M
?
(u X )l,m = φ
?(W ) (l) ci Yl,m (xi ).
i=1
Table 6.2 Errors in the M hX uX − u1/2 EOC

H 1/2 -norm with m = 0
20 0.6514 0.60872542
40 0.4418 0.18859512 3.0179
50 0.3750 0.13264247 2.1469
100 0.2672 0.05752634 2.4649
500 0.1237 0.00738320 2.6658
1000 0.0849 0.00303414 2.3627
Table 6.3 Errors in the M hX uX − u1/2 EOC

H 1/2 -norm with m = 1
20 0.6514 0.93582688
40 0.4418 0.17405797 4.3322
50 0.3750 0.07695943 4.9784
100 0.2672 0.02026597 3.9369
500 0.1237 0.00044098 4.9701
1000 0.0849 0.00008591 4.3459
The error uX − u is computed by

( 500 l )1/2

uX − u1/2 ≈ ?
(l + 1)|(u X )l,m − >
ul,m |2 .
l=1 m=−l
It is expected from the theoretical result (Theorem 6.38) that the order of
convergence for the H 1/2-norm of the error is 2(m + 1). The estimated orders of
convergence (EOC) shown in Tables 6.2, 6.3 appear to agree with the theoretical
results.
In [418] also the exterior Dirichlet problem (6.141), (6.142) and (6.144) is solved
with boundary data
1
UD (x) = .
(1.25 − x3 )1/2
The exact solution is given by
1
U (x) = with p = (0, 0, 0.5),
x − p
and hence, due to (6.151) and (6.152), the exact solution to (6.156) is
−1 + x · p (0.5x3 − 1)
z(x) = ∂ν U (x) = = .
x − p 3 (1.25 − x3 )3/2
It follows from (6.146), (6.147), and (6.176) that
4π(l + 1)
>
zl,m = − Yl,m (n).
2l (2l + 1)
Table 6.4 Errors in the M hX uX − u−1/2 EOC

H −1/2 -norm with m = 0
20 0.6514 0.63633932
40 0.4418 0.18846925 3.1339
50 0.3750 0.13291921 2.1301
100 0.2672 0.05752374 2.4712
500 0.1237 0.00738510 2.6654
1000 0.0849 0.00303428 2.3632
Table 6.5 Errors in the M hX uX − u−1/2 EOC

H −1/2 -norm with m = 1
20 0.6514 0.93557363
40 0.4418 0.18160356 4.2222
50 0.3750 0.07737696 5.2042
100 0.2672 0.02040513 3.9327
500 0.1237 0.00044099 4.9790
1000 0.0849 0.00008591 4.3460
Now (6.174) is solved and the approximate solution zX is compared with the
exact solution z. Note that

M
?
(zX )l,m = (l + 1)φ
?(W ) (l) ci Yl,m (xi ).
i=1
The error zX − z is approximated by

( 500 l )1/2
|(z?
X )l,m −>
zl,m |2
zX − z−1/2 ≈ .
l+1
l=0 m=−l
The theoretical result (Theorem 6.39) requires m > 0 and an order of

convergence of 2m + 5/2 is shown in the H −1/2-norm. The results for m = 0, 1 are
listed in Tables 6.4, 6.5
In [243] the authors analyze the approximation by radial basis functions of a
hyper singular integral equation on an open surface. In order to accommodate the
homogenous essential boundary condition along the surface boundary, scaled radial
basis functions on an extended surface and Lagrangian multipliers on the extension
are used.
Spherical splines are used in [339] to define approximate solutions to strongly
elliptic pseudodifferential equations on the unit sphere. These equations arise from
geodesy. The approximate solutions are found by using the Galerkin method. The
authors prove optimal convergence (in Sobolev norms) of the approximate solution
by spherical splines to the exact solution. A pseudodifferential operator L is a linear
operator that assigns to any v ∈ D (S) a distribution
∞

Lv := > v,m Y,m .
L()> (6.177)
=0 m=−
The sequence {>

L()}≥0 is referred to as the spherical symbol of L. Let K (L) :=
> = 0}. Then
{ : L()
kernel of L = span{Y,m : ∈ K (L), m = −, . . . , }.
For ease of presentation we consider here only the case kernel of L = ∅ (see
[339] for the general case). We look for the problem:
Find u ∈ H α such that
Lu, v = g, v ∀v ∈ H α (6.178)
where L is a strongly elliptic pseudodifferential operator of order 2α whose symbol

is given by
%
>
L() if ∈
/ K (L)
>
L() =
(1 + )2α if ∈ K (L).
Let {v1 , v2 , v3 } be linearly independent vectors in R3 . The triheron T generated by

{v1 , v2 , v3 } is defined by
T := {v ∈ R3 : v = b1 v1 + b2 v2 + b3 v3 with bi ≥ 0, i = 1, 2, 3}.
The intersection τ := T ∩ S is called a spherical triangle. For each v ∈ τ , where

τ is a spherical triangle having vertices v1 , v2 , v3 , there exist unique b1 (v), b2 (v),
b3 (v) satisfying
v = b1 (v)v1 + b2 (v)v2 + b3 (v)v3 , (6.179)
which are called the spherical barycentric coordinates of v with respect to τ .

Let Δ = {τi : i = 1, . . . , T } be a set of spherical triangles. If Δ satisfies
*T
• i=1 τi = S,
• each pair of distinct triangles in Δ are either disjoint or share a common vertex
or an edge,
then Δ is called a spherical triangulation of the sphere S.
Given X = {x1 , . . . , xN } a set of points on S, we can form a spherical
triangulation Δ which contains triangles whose vertices are elements of X (see
[339]).
Given nonnegative integers r and d, the set of spherical splines of degree d and
smoothness r associated with Δ is defined by
Sdr (Δ) := {s ∈ C r (S) : s|τ ∈ Pd , τ ∈ Δ}.

Here, Pd is the space of restrictions to S of homogeneous polynomials of degree d

in R3 . If τ has vertices v1 , v2 , v3 , then s|τ can be written as

s|τ (v) = cijτ k Bijd,τk (v), v ∈ τ,
i+j +k=d
where the coefficients cijτ k are real numbers and the functions
d! i j
Bijd,τk (v) := b (v)b2 (v)b3k (v), i + j + k = d,
i!j !k! 1
are called the spherical Bernstein–Bézier basis polynomials of degree d relative
to τ . Here, bi (v), i = 1, 2, 3, are given by (6.179). For more details, see [339]
For any spherical triangle τ , we denote by |τ | the diameter of the smallest
spherical cap containing τ , and by ρτ the diameter of the largest spherical cap inside
τ . Here the diameter of a cap is, as usual, twice its radius.We define
|Δ|
|Δ| := max{|τ |, τ ∈ Δ}, ρΔ := min{ρτ , τ ∈ Δ} and hΔ := tan .
2
(6.180)
Definition 6.4 Let β be a positive real number. A triangulation Δ is said to be β-
quasiuniform provided that
|Δ|
≤ β.
ρΔ
In the following we briefly introduce the construction of a quasi-interpolation

operator Q : L2 (S) → Sdr (Δ) which is introduced in [318]. First we introduce the
set of domain points of Δ to be
& iv1 + j v2 + kv3 '
D := ξijτ k = .
d i+j +k=d
τ =v1 ,v2 ,v3 ∈Δ
Here, τ = v1 , v2 , v3 denotes the spherical triangle whose vertices are v1 , v2 , v3 .

We denote the domain points by ξ1 , . . . , ξD , where D = dim Sd0 (Δ). Let {Bl :
l = 1, . . . , D} be a basis for Sd0 (Δ) such that the restriction of Bl on the triangle
containing ξl is the Bernstein-Bézier polynomial of degree d associated with this
point, and that Bl vanishes on other triangles.
A set M := {ζl }M l=1 ⊂ D is called a minimal determining set for Sd (Δ) if, for
r
every s ∈ Sdr (Δ), all the coefficients νl (s) in the expression s = D l=1 νl (s)Bl are
uniquely determined by the coefficients corresponding to the basis functions which
are associated with points in M . Given a minimal determining set, the authors of
[339] construct a basis {Bl∗ }M r
l=1 for Sd (Δ) by requiring
νl (Bl∗ ) = δl,l , 1 ≤ l, l ≤ M.
Using the Hahn–Banach Theorem the linear functionals νl , l = 1, . . . , M, are

extended to all of L2 (S). Now, the quasi-interpolation operator Q : L2 (S) → Sdr (Δ)
is defined by

M
Qv := νl (v)Bl∗ , v ∈ L2 (S).
l=1
The following theorem is shown in [339] with an analysis similar to [418] but
with Q instead of the interpolation operator IX .
Theorem 6.40 Assume that Δ is a β-quasiuniform spherical triangulation with
|Δ| ≤ 1, and that there holds
%
d ≥ 3r + 2, if r > 1
(6.181)
d ≥ 1, if r = 0.
Then for any v ∈ H s , there exists η ∈ Sdr (Δ) satisfying
v − ηt ≤ Chs−t
Δ vs ,
where t ≤ r + 1 and t ≤ s ≤ d + 1. Here C is a positive constant depending only

on d and the smallest angle in Δ.
We consider the Galerkin equation: Find =
u ∈ Sdr (Δ) such that
L=
u, v = f, v ∀v ∈ Sdr (Δ). (6.182)
Theorem 6.41 Assume that Δ is a β-quasiuniform spherical triangulation with

|Δ| ≤ 1 and that (6.181) hold. If the order 2α of the pseudodifferential operator
L satisfies α ≤ r + 1, and if u and =
u satisfy, respectively, (6.178) and (6.182), then
ut ≤ Chs−t
u − = Δ us ,
where s ≤ d + 1 and 2α − d − 1 ≤ t ≤ min{s, α}. Here C is a positive constant

depending only on d and the smallest angle in Δ.
Chapter 7
Advanced BEM for BVPs in
Polygonal/Polyhedral Domains: h- and
p-Versions
This chapter presents, h−, p−BEM on graded meshes and hp−BEM on quasiuni-
form meshes for the numerical treatment of boundary value problems in polygonal
and polyhedral domains. For ease of presentation we also introduce here the
hp−version on geometrically graded meshes (for details and proofs see Chap. 8).
For the solutions of Dirichlet and Neumann problems we present decompositions
into a sum of special singularity terms (describing their edge and corner behaviors)
and in regular parts (see Theorem 7.3, Theorem 7.12 for two-dimensions and
Theorem 7.7, Theorem 7.16 for three dimensions). These regularity results by von
Petersdorff, Stephan [425] are based on the seminal works of Dauge [141] and
Kondratiev [270]. Chapter 7 is organized as follows: The results for the single layer
integral equation covering the Dirichlet problem are presented in Sect. 7.1 ; those
for the hypersingular integral equation covering the Neumann problem in Sect. 7.2.
Then in Sect. 7.3 the proofs for the results for the integral equations on curves
are given, whereas in Sect. 7.4 the results for the integral equations on surfaces .
We present approximation results for solutions of the integral equations on graded
meshes in 2D and 3D from the PhD thesis by von Petersdorff [423], see also [426].
Also in detail we investigate the hp−version of BEM on quasi uniform meshes on
polygons based on the paper by Suri and Stephan [405]. For the p-version BEM
with quasi uniform meshes on polyhedra we refer to [51, 52, 374].
There has been much work on the regularity of elliptic problems. The interested
reader might also look into the key papers by Maz’ya, Nazarov and Plamenevsky
[302, 303], and into their text books [271, 317]. Recently the concept of detached
asymptotics has proved to be very fruitful, see [316].

224 7 BEM in Polygonal/Polyhedral Domains
7.1 The Dirichlet Problem
In this section we consider integral equation methods for solving boundary value
problems in non-smooth domains Ω. First we deal with the standard Dirichlet
problem
Δu = 0 in Ω,
(7.1)
u=g on Γ = ∂Ω.
Here Ω is either a plane, curvilinear polygon or the exterior of an open arc or a

polyhedron with a piecewise C ∞ −surface or an open surface piece. We use the
convention that the normal vector always points away from Ω, with corresponding
modifications when Ω is the exterior of an arc or an open surface.
As shown in Chaps. 2 and 4, (7.1) is converted into the integral equation of the
first kind for φ = ∂u
∂n |Γ :
V φ = (I + K)g on Γ (7.2)
The connection between the boundary value problem (7.1) and the integral equa-
tion (7.2) is as follows.
Theorem 7.1 Let g ∈ H 1/2(Γ ).
(i) Then there exists exactly one solution φ ∈ H −1/2(Γ ) of the integral equa-
tion (7.2).
(ii) The problem (7.1) and the equation
(7.2) are equivalent, i.e. let u ∈ Hloc
1 (Ω)
solve (7.1) then φ = ∂u −1/2 (Γ )

∂n Γ solves (7.2), conversely, let φ ∈ H
solves (7.2) then u defined by (2.2) with u|Γ = g solves (7.1)
Remark 7.1 The above method remains valid for general elliptic boundary value
problems and even transmission problems with differential operators with constant
coefficients: For problems in elasticity, acoustics, electromagnetics etc. see Chap. 4
(These boundary integral equations have similar properties as the original differen-
tial equations; they allow a variational formulation and are strongly elliptic.)
We observe that the integral equation (7.2) can be solved approximately by the
Galerkin’s method using conforming subspaces {XN } of H −1/2(Γ ):
For given g ∈ H 1/2(Γ ) find φN ∈ XN such that
V φN , χ = (1 + K)g, χ ∀χ ∈ XN . (7.3)
Here the bracket w, v denotes the duality between H 1/2(Γ ) and H −1/2 (Γ ), which
can be identified with the L2 −inner product. V is strongly elliptic by Theorems 4.3
and 6.3.
7.1 The Dirichlet Problem 225
Thus Theorem 6.1 gives the following result for the Galerkin procedure (7.3):
Theorem 7.2 Let N N0 . Then the equation (7.3) has a unique solution φN ∈
XN . Furthermore there holds
φ − φN H −1/2 (Γ ) C inf φ − χH −1/2 (Γ ) (7.4)

χ∈XN
where the constant C is independent of N,φN and the solution φ of (7.2).

Remark 7.2 The proof of Theorem 7.2 is based on the strong ellipticity of the
integral operator V . This property means that V is coercive in the sense of a
Gårding inequality, i.e. there exists a constant η > 0 and a compact operator T
from H −1/2(Γ ) into H 1/2(Γ ) such that
(V + T )ψ, ψ η ψ2H −1/2 (Γ ) ∀ψ ∈ H −1/2(Γ ). (7.5)
This inequality corresponds directly by integration by parts to the Dirichlet bilinear

form for the variational solution u ∈ Hloc
1 (Ω) of (7.1).
Theorem 7.2 guarantees the convergence of any Galerkin scheme for solving
integral equation (7.2) by use of conforming subspaces XN ⊂ H −1/2(Γ ). Due to
the quasioptimality estimate (7.4) the rate of convergence of the used scheme is
determined by the choice of approximating subspaces XN and the regularity of the
exact solution φ of the integral equation.
For an appropriate choice of XN it is crucial to know the behavior of the solution
φ near crack trips, corners and edges. There φ becomes singular which corresponds
to the behavior of the solution u of the original problem (7.1).
7.1.1 Regularity on a Polygon
We consider next the case of a plane polygon Γ with straight line segments Γ i . By
tj (j = 0, . . . , J ) we denote the corner points where Γ j and Γ j +1 meet (tJ = t0 ).
The interior angle at tj is denoted by ωj .
The following explicit regularity result for φ is obtained in [128] using localiza-
tion and Mellin transformation.

kπ
Theorem 7.3 Let g ∈ H (Γ ), 1/2 s < 3/2, s ∈
s / A = αj k = ,1 j
" ωj
J, k ∈ N . The solution φ of (7.2) has the form
⎛ ⎞

J α −1
φ= ⎝ cj k ρj jk ⎠ χj + φ0 , φ0 ∈ H s−1(Γ ), cj k ∈ R.
j =1 αjk <s−1/2
Fig. 7.1 Ω = L−shaped t3

domain with reentrant corner • • t2
at vertex t1 with ω1 = 32 π
•
t1
• t6
•
t4
•
t5
Fig. 7.2 Ω = R2 \Γ , • •
Γ = [−1, 1] t1 t2
Here ρj denotes the Euclidean distance between x ∈ Γ and the vertex tj whereas
χj is a C ∞ -cut-off function concentrated near tj . For s > 3/2 and s − 1/2 ∈
/ N, the
corresponding theorem still holds.
Remark 7.3 The analogous result for a general curvilinear polygon Ω is given in
[126]. There, additional singularity terms like ρ αjk −1 log ρ due to the curvature of
the axis Γ j may appear.
We illustrate the regularity result for some canonical domains.
Example 7.1 For Γ being the boundary of the L-shaped domain in Fig. 7.1 we
obtain with g ∈ H 3/2−ε (Γ ), ε > 0 arbitrary,
−1/3
φ = c11 ρ1 χ 1 + φ0 , φ0 ∈ H 1/2−ε (Γ ).
Here the singularity terms located at the vertices t2 , . . . , t6 are included in φ0 .

Example 7.2 For the slit Γ in Fig. 7.2 and for g ∈ H 3/2−ε (Γ ), ε > 0 arbitrary, we
have φ ∈ H̃ −1/2(Γ ) with
−1/2 −1/2
φ = c11 χ1 ρ1 + c21 χ2 ρ2 + φ0 , φ0 ∈ H̃ 1/2−ε (Γ ).
Next, we comment on various choices for the approximating subspaces XN .
7.1.2 BEM on a Polygon
7.1.2.1 h-p Method with Quasiuniform Mesh on a Polygon
j
Construction: Let for each Γ j ⊂ Γ there be given a family of grids {Th } which
*h j
N
Γ¯h,i such that Γh,i is an open interval.
j j
partition each Γ j into Nh pieces, Γ j =
i=1
j j j
We assume that {Th } is quasi uniform, in the sense that with hi = meas Γh,i and
j
h = max hi , there exists a constant τ independent of h such that
i,j
j j
h/ hi τ for all intervals Γh,i (7.6)
For p 0 , Sp,h (Γ j ) will denote the set of all functions v defined on Γ j such that
j
the restriction v|Γ j belongs to Pp Γh,i , the space of polynomials of degree p
h,i
j
on Γh,i . For p 0 , h > 0, we define Sp,h (Γ ) to be the set of functions on Γ whose
restrictions to Γ j ⊂ Γ belong to Sp,h (Γ j ).
In the standard h-method one decreases the mesh size h and keeps p = p0 fixed,
i.e. one takes as approximating subspaces XN = Sp0 ,h (Γ ) where N is proportional
to 1/ h.
On the other hand, in the p-version, one uses a fixed mesh h = h0 and increases
the degrees p of the approximating polynomials, i.e. one takes XN = Sp,h0 (Γ )
where N is proportional to p.
In the h-p method one combines the two approaches, thus one takes as approxi-
mating subspaces XN = Sp,h (Γ ) with two functions p(N) and h(N). For example,
one may take p(N) ∼ N a1 and h(N) ∼ N −a2 where a1 , a2 > 0. Note, that we do
not impose continuity for the elements of Sp,h (Γ ) at the mesh points, since we need
only Sp,h (Γ ) ⊂ H −1/2(Γ ) which is guaranteed already if Sp,h (Γ ) ⊂ L2 (Γ ).
There holds the following convergence result for the Galerkin solution φp,h ∈
Sp,h (Γ ) of (7.3). Its proof follows by combining Proposition 7.2, Theorem 7.20
and Theorem 7.2
Theorem 7.4 ([405]) Let (7.6) hold and let p be sufficiently large and h be small
enough. Then the Galerkin equations (7.3) are uniquely solvable in Sp,h (Γ ). Let
φ ∈ H −1/2(Γ ) be the solution of the integral equation (7.2) with right hand side
f = (1 + K)g ∈ H s (Γ ) and φp,h ∈ Sp,h (Γ ) be the Galerkin solution, then we
have for s 1/2 , s ∈
/ A, with α = min αj k
0 0
0φ − φp,h 0 −1/2 C max{e1 , e2 }
H (Γ )
%
e1 = max{hα p−2α , hmin{α,p−α+1/2} p−2α } log1/2 p
with
e2 = hmin{s−1/2, p+1/2}p−(s−1/2) log1/2 p
where the constant C depends on φ but is independent of h and p.

Again we illustrate the convergence rates when Ω is as in Example 7.1 and 7.2. For
the h-version we obtain
%
0 0 O(h2/3 ) if Ω is L-shaped
0φ − φp,h 0 −1/2 = .
H (Γ )
O(h1/2 ) if Ω = R2 \ [−1, 1]
For the p−version we have

%
0 0 O p−4/3+ log1/2 p if Ω is L-shaped
0φ − φp,h 0 = −1+ .
H −1/2 (Γ )
O p 1/2
log p if Ω = R2 \ [−1, 1]
In both cases (Example 7.1 and 7.2) for the h-p version the quantity e1 dominates
e2 yielding
E = O hα p−2α log1/2 p where α = 2/3 in Example 7.1
α = 1/2 in Example 7.2.
Note that increasing the dimension of the subspaces used by changing p is twice as
efficient (in terms of the asymptotic rate of convergence) as changing h. In order
to compare the h, p, and h-p-method we introduce the degree of freedom Nf =
dim XN . For the h-method Nf ∼ h−1 , hence the rate of convergence is Nf−α−ε .
For the p-method we have Nf ∼ p, thus the rate of convergence is Nf−2α−ε . In
the case of the h-p-method the rate convergence depends on the choice of h(N)
and p(N). For the above example p(N) ∼ N a2 , h(N) ∼ N −a1 we have Nf ∼
p(N)h−1 (N) = N a1 N a2 . Hence the convergence rate is
a1 +2a2
−α a1 +a2 −ε
h(N)α p(N)−2α ∼ N −α(a1 +2a2 ) ∼ Nf
This convergence rate lies between the rate Nf−α−ε of the h−version and the rate
− 3α
2 −ε
Nf−2α−ε of the p−version. For example, for a1 = a2 = 1 we have Nf .
7.1.2.2 h-Method with Graded Mesh on a Polygon
The mesh and the boundary element space are constructed as follows: Bisect all
sides of the polygon. For each corner tj (j = 1, . . . , J ) of the polygon identify each
of the 2 adjacent parts with the interval I = [0, 1] such that tj corresponds to 0 and
the endpoints correspond to 1. Then choose the mesh points corresponding to
βj
k 1
xk = , k = 0, . . . , N − 1 , xk = (kh)βj , h = (7.7)
N N
where βj ≥ 1 is called the grading exponent (βj = 1 means a uniform mesh)

+
(Fig. 7.3). Thus we obtain the graded meshes {Zj,k ; k = 0, . . . , N − 1} and
−
{Zj,k ; k = 0, . . . , N − 1} on the adjacent sides of each corner tj , j = 1, . . . , J .
β
For p 0, β = (β1 , . . . , βJ ), Sp,h (Γ ) denotes the set of piecewise polynomials
of degree p on the graded mesh described above which might be refined differently
+
Zj,k
tj
−
Zj,k
0 xk 1
Fig. 7.3 Example of graded mesh
at various vertices. Thus the simplest choice for an approximating subspace in the
β
Galerkin scheme (7.3) is S0,h (Γ ), i.e. the set of piecewise constant functions defined
on the graded mesh (for simplicity we consider only the case β = βj ). There holds
as application of Proposition 7.3 to Theorem 7.2 the following convergence result.
Theorem 7.5 ([426]) Let (7.7) hold and h be sufficiently small. Let φ ∈ H −1/2(Γ )
solve (7.2) with g ∈ H s (Γ ) , s 1/2 , s ∈ / A , s + 1/2 ∈ / N. Then the
β
Galerkin equations (7.3) are uniquely solvable in Sp,h (Γ ). Moreover we have for
α = min αj k
%
2p+3
hαβ−ε if β ≤
φ − φh H −1/2 (Γ ) C 2p+3
2α
2p+3
h 2 if β > 2α
where the constant C = C(β) is independent of h, φ and φh .

We illustrate the convergence rates when Ω is as in Example 7.1 and 7.2. With
E = φ − φh H −1/2 (Γ ) we obtain with p = 0
%
O h2β/3−ε if β < 9/4
E= , if Ω is L-shaped
O h3/2 if β > 9/4
and
%
O hβ/2−ε if β < 3
E= , if Ω is R2 \ [−1, 1]
O h3/2 if β > 3
7.1.2.3 h-p Method with Geometric Mesh on a Polygon
Near each corner introduce a geometric mesh as follows: We setΓ k = (0, 1) for
simplicity and consider the mesh near tk = Γ k−1 ∩ Γ k defined with a parameter
0 < σ < 1 by the grid points
xj = σ j (j = 0, 1, . . . , N − 1), xN = 0.
Now the space XN,σ consists of discontinuous piecewise polynomials with degree
N − 1 − j on the interval [xj +1 , xj ]. Thus we have in the first interval at the corner
polynomials of degree zero, and in the next interval polynomials of degree one and
so on, i.e. with the interval Ijk = [xj +1 , xj ] on Γ k
& '
XN,σ = v ∈ L2 (Γ ) : v|I k ∈ Ppj , pj = N − 1 − j
j
Here the total degree of freedom Nf = dim XN is proportional to N 2 .

In many practical cases the given data g in (7.1) is analytic on each piece Γ k of
the polygon Γ Two different proofs of the following result are presented in Sects. 8.1
and 8.2.
Theorem 7.6 ([20, 240]) Let N be sufficiently large. Then the Galerkin equa-
tions (7.3) are uniquely solvable in XN,σ . Let φN,σ denote the Galerkin solution
and φ ∈ H −1/2(Γ ) be the exact solution of the integral equation (7.2) with analytic
g, then we achieve the exponential convergence
0 0 √
0φ − φN,σ 0 −1/2 C e −β N
H (Γ )
where the constants C and β are independent of N.

Note that the choice of the geometric mesh is independent of the order of singularity
of the solution of (7.1). Therefore we use in our canonical examples (L-shaped
domain, slit domain) the same mesh near the vertices and the same subspaces XN,σ .
7.1.3 Regularity on a Polyhedron
*
J
Next we consider the case of a polyhedron Ω with the surface Γ = ∂Ω = Γj
j =1
in R3 with plane faces Γ j . We describe Dirichlet data g ∈ H s (Γ ) where H s (Γ ) is
defined as follows:
& '
H s (Γ ) := u|Γ |u ∈ H s+1/2(R3 ) .
Then the Neumann data φ of the solution has regularity H s−1 away from the edges
and corners. Near an edge with opening ω there are edge singularities of the form
c(y)ρ mν+2p−1. Here ν = πω , m > 0 and p 0 integers and ρ denotes the distance
to the edge, while the stress intensity factor c(y) is a function defined on the edge.
Near the corners we get additional corner singularities of the form
r λk wk (ξ ), ξ ∈ Γ0
where r denotes the distance to the vertex and wk is a function on the spherical
polygon Γ0 = Γ ∩ S2 . S2 is a sphere centered in the vertex, θ and φ are polar
coordinates on S2 . The exponent λk and the function wk are obtained as follows:
Consider the eigenvalue problem for the Laplace- Beltrami operator Δθ,φ on S2 ,
and let μk be the k−th eigenvalue with corrsponding eigenfunction vk
Δθ,φ vk (θ, φ) = μk vk (θ, φ)

vk (θ, φ)|Γ0 = 0
then
A
1 ∂
λk := −1/2 + μ2k + , wk := vk (θ, φ) .
4 ∂n Γ0
We state the decomposition theorem for the neighbourhood of a vertex t0 :

Theorem 7.7 ([425]) Let ωj denote the openings of the edges γj meeting in t0 ,
νj = ωπj (j = 1 . . . J ). Choose one of the edge exponents s1 = mνj + 2p where
m, j, p are some integers, let s2 be the next larger value of the form (mνj + 2p) or
λk + 1/2
3 4
s2 = min (mνj + 2p), λk + 1/2 | mνj + 2p > s1 , λk + 1/2 > s1 , j, p, k integers
This corresponds to the first singularity not occuring in the decomposition. Let s >
s1 with s − 1/2 ∈
/ N, s = mνj , s = λk + 1/2. Then
Δu = 0 in Ω
u|Γ = g ∈ H s+1/2(Γ )

∂u
implies for φ := ∂n Γ
mνj +2p−1
ak r λk −1 wk +
j
φ = φ 0 + χ(r) χj (θ ) Cm,p θj
0<λk <s2 −1/2 j =1,...,J mνj +2ps1

am,k r λk −1 + C̃m,0 , r · C̃m,0 ∈ H−mνj j (R+ )
j j j j s−mν
r · φ 0 ∈ H s̃ (Γ ), Cm,0 =
s1 <λk +1/2<s
s−mνj −2p
(R+ )
j
r · Cm,p ∈ H−mν for p > 0
s̃ = min {s − s1 , −1/2 + s2 − ε} , ε > 0
Here θj denotes the angle to the edge γj , H s̃ (Γ ) are the functions φ on Γ such that
their restrictions φ|Γ k to each face Γ k of the polygon have the regularity φ|Γ k ∈
B
J
H s̃ (Γ k ), i.e. H s̃ (Γ ) = H s (Γ k ). H s̃ (Γ0 ) is defined analogously.
k=1
Next we illustrate Theorem 7.7 with an example:
Example 7.3 Let Ω be the exterior of a square in R3 . Then ωj = 2π and the edge
exponents are νj = 1/2 and the vertex exponents λk are (see [235])
λ1 = 0.297 , λ2 = 1.426 , λ3 = 2.06
Now assume for the given Dirichlet data g ∈ H 3−ε (R3 ). Then we have near a corner
of the square
π −1/2
φ = φ 0 + a1 r .297−1 w1 (θ ) + C11 (r)θ −1/2 + C12 (r) −θ
2
j j j j j 3−ε +
φ 0 ∈ H 1−ε (Γ ) C1 = a1 r 1.426−1 + a2 r 2.06−1 + C̃1,0 r C̃1,0 ∈ H−1/2 R
Note that the function w1 (θ ) also becomes singular near the edges. We can state
an alternative form of the decomposition with a smoother function w̃1 (θ ) instead
of w1 (θ ) and where the edge singularities are expressed with the distance ρj to the
edge rather than θ (Fig. 7.4).
−1/2 −1/2
φ = φ0 + a1 r .297−1w̃1 (θ) + e11 (y1 )ρ1 + e12 (y2 )ρ2

φ0 ∈ H 1−ε (Γ ) , e1 (yj ) = a1 yj−.203 + a2 yj.926 + a3 r 1.56 + e1,0 , e1,0 ∈ H03−ε R+
j j j j j j
y1
P1
P2
θ
Fig. 7.4 Screen

j
Here e1 (yj ) are the physically relevant stress intensity factors. They blow up
towards the corners of the square.
Remark 7.4 Since ρ νj ∈ H 1/2+νj −ε (Γ ) and r λk ∈ H 1+λk −ε (Γ ) the regularity of φ
for sufficiently smooth g is φ ∈ H −1/2+α (Γ ) with
3 4
α = min νj , λk + 1/2 . (7.8)
Next, we consider various boundary element methods for the polyhedron.
7.1.3.1 h-p-Method with Quasi Uniform Mesh on a Polyhedron
Construct a quasi uniform mesh with width h analogously as above and define the
space Sp,h (Γ ) of piecewise polynomials of degree p on this mesh. Then the rate
of convergence is determined by the regularity of the solution, i.e. the parameter α
in (7.8):
Theorem 7.8 ([51]) Let p be sufficiently large and h be small enough. Then the
Galerkin equations (7.3) are uniquely solvable in Sp,h (Γ ). Let φ ∈ H −1/2 (Γ ) be
the solution of the integral equation (7.2) with sufficiently smooth right hand side g,
and let φp,h the Galerkin solution, then we have with α in (7.8)
0 0
0φ − φp,h 0 −1/2 C · hα−ε pε−2α . (7.9)
H (Γ )
Remark 7.5
(i) If Nf = dim XN denotes the degrees of freedom , then Theorem 7.8 gives for
the h-method the rate O Nf−α and for the p-method the rate O Nf−2α .
(ii) Method of the proof: By Theorem 7.2, we have to show, how fast the solution
φ can approximated in the spaces XN . Theorem 7.7 gives the decomposition
of φ in singularities and a regular part. Then the approximation result for the
h-method is contained in the proof of Theorem 7.9 for β = 1. The result for the
p-method of FEM in 3D was shown in [16]. The estimate (7.9) can be proved
by splitting the solution into regular/singular terms as in Theorem 7.7. Then one
treats the various terms (edge, vertex and edge-vertex singularities) separately
(modifiying the 2D analysis in section 7.4) by extending the 1D results for the
quasi uniform hp-version from section 7.3 (see also [51, 405] and [50] for the
p-version for a three-dimensional crack problem).
{1/2,
In Example 7.3 we have α = min .297 + 1/2} = 1/2, hence
Theorem
7.8
yields the convergence rate O h1/2 for the h-method and O p−1 for the p-
method.
7.1.3.2 h-Method with Graded Mesh on a Polyhedron
For simplicity we only consider the case where all the faces meeting at the vertex
are convex. Then each of these sectors can be mapped linearly on the quadrant
R+ × R+ ⊂ R2 . On the quadrant we introduce a graded mesh given by the lines
x1 = (i h)β , x2 = (j h)β , h = 1/N , i, j = 0 . . . N
β
Let S0,h be the space of piecewise constant functions on this mesh. As in
Theorem 7.5 we can compensate the effect of the singularities by a appropriately
graded mesh and get the convergence rate h3/2 . Theorem 7.9 follows by combining
Theorem 7.2 with the 2D approximation results in Sect. 7.4.1.
Theorem 7.9 ([423, 426]) Let ψ ∈ H −1/2 (Γ j ) have on every corner of the edge
Γ j a decomposition of the form as in Theorem 7.7 with νj m > 0, λk > 0. Let
1
α0 := min{λk + , νim }
2
β
Then we can approximate ψ for β ≥ 1 by the spaces Sh on Γ j in the following
way:
β
Let φh ∈ Sh be the piecewise constant function which coincides on every
subdomain with the mean value of ψ there. Then it holds for all > 0
ψ − φh H −1/2 (Γ j ) ≤ Cha−
with
%
min{α0 β, 3/2} for α0 ≥ 1/2
a :=
min{α0 β 23 (1 + α0 ), 1 + α0 } for α0 < 1/2
Here C depends on β and but not on h.

For the construction of graded meshes on individual faces on Γ , we can assume that
all faces of Γ are triangles. On general polygonal faces the construction is similar,
or one can first subdivide the polygon into triangles. On a triangular face F ⊂ Γ ,
we first draw three lines through the centroid and parallel to the sides of F . This
makes F divided into three parallelogams and three triangles (see Fig. 7.5 ). Each
of the three parallelograms can be mapped onto the unit square Q̂ = (0, 1)2 by a
linear transformation such that the vertex (0, 0) of Q̂ is the image of a vertex of F .
Analogously, each of the three sub-triangles can be mapped onto the unit triangle
T̂ = {(x1 , x2 ); 0 < x1 < 1, 0 < x2 < x1 } ⊂ Q̂ such that the vertex (1, 1) of T̂ is
the image of the centroid of F . Then, the graded mesh on Q̂ (and therefore on T̂ ) is
Fig. 7.5 Graded mesh on the triangular face F ⊂ Γ . The triangular (resp. parallelogram) block
of elements TF (resp. QF ) is the image of the graded mesh on the unit triangle T̂ (resp. the unit
square Q̂)
generated by the lines

β β
i j
x1 = x2 = i, j = 0, 1, . . . , N.
N N
Here β ≥ 1 is the grading parameter, and N ≥ 1 corresponds to the level of

refinement. Mapping each cell of these meshes back onto the face F , we obtain
a graded mesh of triangles and parallelograms on F (see Fig. 7.5). The diameter of
the largest element of this mesh is proportional to βN −1 . Thus, h = 1/N defines
the mesh paramter.
We illustrate this result with two examples analogous to the examples after
Theorem 7.5.
3π
(i) Let Ω be the exterior of a cube. Then the angles of the edges are , i.e.
2
2
νj = . The first vertex singularity is r λ1 with λ1 + 1/2 ≥ 2/3, hence α =
3 3 4
min νj , λ1 + 1/2 = 23 and Theorem 7.9 gives the convergence rate
O h2β/3−ε if β < 9/4
O h3/2 if β > 9/4

(ii) Let Ω be the exterior of a square as in Example 7.3. Then νj = 1/2 , λ1 =

0.297, thus α = 1/2. By Theorem 7.9 we have the convergence rate
O hβ/2−ε if β < 3
O h3/2 if β > 3
Remark 7.6 In most cases the edge singularities are “stronger” than the vertex
singularities, i.e. νj < λ1 + 1/2 for some j . An example for the case λ1 + 1/2 < νj
is the exterior of a pyramid with sufficiently small opening at the tip.
7.2 The Neumann Problem
In this section we consider integral equation methods for the Neumann problem in
non-smooth domains Ω above.
Δu = 0 in Ω,
(7.10)
∂u
=f on Γ = ∂Ω. f =0
∂n Γ

With the Cauchy data v = u|Γ and φ = ∂u
∂n Γ using the jump relations (see
Sect. 2.2.2) one obtains for x ∈ Γ the equation
W v = (I − K )φ on Γ (7.11)
with the integral operators

∂ ∂ ∂
K w(x) := 2 G(x, y)w(y)ds(y), Ww(x) := −2 G(x, y)w(y)ds(y).
∂nx ∂nx ∂ny
Γ Γ
Insertion of the given data ∂u

∂n = f in (7.10) into (7.11) gives the integral equation
of the first kind for φ:
W v = (I − K )f on Γ (7.12)
The connection between the boundary value problem (7.10) and the integral
equation (7.12) is as follows.
7.2 The Neumann Problem 237
Theorem 7.10 Let f ∈ H −1/2(Γ ).

(i) Then there exists exactly one solution v ∈ H 1/2 (Γ ) of the integral equa-
tion (7.12).
(ii) The problem (7.10) and the equation (7.12) are equivalent,
We observe that the integral equation (7.12) can be solved approximately by
Galerkin’s method using conforming subspaces {YN } of H 1/2(Γ ):
For given f ∈ H −1/2 (Γ ) find vN ∈ YN such that
W vN , ζ = (I − K )f, ζ ∀ζ ∈ YN . (7.13)
Theorem 6.1 gives the following result for the Galerkin procedure (7.13):
Theorem 7.11 Let N N0 . Then the equation (7.13) has a unique solution vN ∈
YN . Furthermore there holds
v − vN H 1/2 (Γ ) C inf v − ζ H 1/2 (Γ ) (7.14)

ζ ∈YN
where the constant C is independent of N,vN and the solution v of (7.12).

Remark 7.7 The proof of Theorem 7.11 is based on the strong ellipticity of the
integral operator W . This property means that W is coercive in the sense of a
Gårding inequality, i.e. there exists a constant η > 0 and a compact operator T
from H 1/2(Γ ) into H −1/2(Γ ) such that (see Theorem 4.3)
(W + T )v, v η v2H 1/2 (Γ ) ∀v ∈ H 1/2(Γ ). (7.15)
This inequality corresponds directly by integration by parts to the Dirichlet bilinear

form for the variational solution u ∈ Hloc 1 (Ω) of (7.10).
Theorem 7.11 guarantees the convergence of any Galerkin scheme for solving
integral equation (7.12) by use of conforming subspaces YN ⊂ H 1/2 (Γ ). Due to
the quasioptimality estimate (7.14) the rate of convergence of the used scheme is
determined by the choice of approximating subspaces YN and the regularity of the
exact solution v of the integral equation.
For an appropriate choice of YN it is crucial to know the behavior of the solution
v near crack trips, corners and edges. There v becomes singular which corresponds
to the behavior of the solution u of the original problem (7.10).
Now we consider the case of a plane polygon Γ with straight line segments
Γ i . By tj (j = 0, . . . , J ) we denote the corner points where Γ j and Γ j +1 meet
(tj = t0 ). The interior angle at tj is denoted by ωj .
The following explicit regularity result for v is obtained in [128] using localiza-
tion and Mellin transformation (see Lemma 9.5).

kπ
Theorem 7.12 Let f ∈ H s−1(Γ ), 1/2 s < 3/2, s ∈
/ A = αj k = ,1
" ωj
j J, k ∈ N . The solution v of (7.12) has the form
⎛ ⎞

J
v= ⎝ cj k ρj jk ⎠ χj
α
+ v0 , v0 ∈ H s (Γ ), cj k ∈ R. (7.16)
j =1 αjk <s−1/2
Next, we comment on various choices for the approximating subspaces YN .
h-p Method with Quasiuniform Mesh on a Polygon
On the mesh constructed as above we introduce the boundary element space as

follows. For p 1 , Sp,h (Γ j ) will denote the set of all functions v defined on Γ j
j
such that the restriction v|Γ j belongs to Pp Γh,i , the space of polynomials of
h,i
j
degree p on Γh,i . For p 1 , h > 0, we define Sp,h (Γ ) to be the set of functions
on Γ whose restrictions to Γ j ⊂ Γ belong to Sp,h (Γ j ).
In the standard h-method one decreases the mesh size h and keeps p = p0 fixed,
i.e. one takes as approximating subspaces YN = Sp0 ,h (Γ ) where N is proportional
to 1/ h.
On the other hand, in the p-version, one uses a fixed mesh h = h0 and increases
the degrees p of the approximating polynomials, i.e. one takes YN = Sp,h0 (Γ )
where N is proportional to p.
In the h-p method one combines the two approaches, thus one takes as approx-
imating subspaces YN = Sp,h (Γ ) with two functions p(N) and h(N). Note, that
we do impose continuity for the elements of Sp,h (Γ ) at the mesh points, since
Sp,h (Γ ) ⊂ H 1/2 (Γ ). There holds the following convergence result for the Galerkin
solution vp,h ∈ Sp,h (Γ ) of (7.13). Its proof follows by combining Proposition 7.1
and Theorem 7.11.
Theorem 7.13 ([405]) Let (7.6) hold and let p be sufficiently large and h be small
enough. Then the Galerkin equations (7.13) are uniquely solvable in Sp,h (Γ ). Let
v ∈ H 1/2(Γ ) be the solution of the integral equation (7.12) with right hand side
g = (1 + K )f ∈ H s−1 (Γ ) and vp,h ∈ Sp,h (Γ ) be the Galerkin solution, then we
have for s 1/2 , s ∈
/ A, with α = min αj k
0 0
0v − vp,h 0 C max{e1 , e2 }
H 1/2 (Γ )
%
e1 = max{hα p−2α , hmin{α,p−α+1/2} p−2α } log1/2 p
with
e2 = hmin{s−1/2, p+1/2}p−(s−1/2) log1/2 p
where the constant C depends on v but is independent of h and p.

The proof follows by combining Proposition 7.1 and Theorem 7.19 in Sect. 7.3.1.
Optimal converges rates for the p-version without log-terms are derived in [208] and
for the hp-version in [209].
h-Method with Graded Mesh on a Polygon
Mesh and boundary element space are constructed as above with p ≥ 1. There holds
as application of Theorem 7.11 the following convergence result.
Theorem 7.14 ([426]) Let (7.7) hold and h be sufficiently small. Let v ∈ H 1/2(Γ )
solve (7.12) with f ∈ H s−1(Γ ) , s 1/2 , s ∈
/ A , s +1/2 ∈/ N. Then the Galerkin
β
equations (7.13) are uniquely solvable in Sp,h (Γ ). Then we have for α = min αj k
%
2p+3
hαβ−ε if β <
v − vh H 1/2 (Γ ) C 2p+3
2α
2p+3
h 2 if β > 2α
Proof Left as exercise for the reader, c.f. Proposition 7.3.
h-p Method with Geometric Mesh on a Polygon
Near each corner introduce a geometric mesh as in Sects. 8.1 or 8.3: We set Γ k =
(0, 1) for simplicity and consider the mesh near tk = Γ k−1 ∩ Γ k defined with
0 < σ < 1 by the grid points
xj = σ j (j = 0, 1, . . . , N − 1), xN = 0.
Now the space YN,σ consists of continuous piecewise polynomials with degree
N − j on the interval [xj +1 , xj ]. Thus we have in the first interval at the corner
polynomials of degree one, and in the next interval polynomials of degree 2 and so
on, i.e. with the interval Ijk = [xj +1 , xj ] on Γ k
& '
YN,σ = v ∈ C 0 (Γ ) : v|I k ∈ Ppj , pj = N − j
j
Here the total degree of freedom Nf = dim YN is proportional to N 2 .

In many practical cases the given data f in (7.10) is analytic on each side Γ k of
the polygon Γ .
Theorem 7.15 ([21, 240]) Let N be sufficiently large. Then the Galerkin equa-
tions (7.13) are uniquely solvable in YN,σ . Let vN,σ denote the Galerkin solution
and v ∈ H 1/2(Γ ) be the exact solution of the integral equation (7.12) with piecewise
analytic f , then we achieve the exponential convergence

0 0 √
0v − vN,σ 0 1/2 C e−β N
H (Γ )
where the constants C and β are independent of N.
7.2.1 Regularity on a Polyhedron
Next we consider the Neumann problem on a polyhedron Ω. The following

regularity result follows from [426] where the mixed Dirichlet-Neumann BVP is
considered.
Theorem 7.16 Choose s1 = mνj + 2p where m, j, p are some integers, let s2
be the next larger value of the form (mνj + 2p) or λk + 1/2. Let s > s1 with
s − 1/2 ∈
/ N, s = mνj , s = λk + 1/2. Then

∂u
Δu = 0 in Ω, = f ∈ H s−1/2(Γ )
∂n Γ
implies (if s ≥ s1 + s2 − 1/2 − )

j mνj +2p
v = v 0 + χ(r) ak r λk wk + χj (θ ) hm,p θj ,
0<λk <s2 −1/2 j =1,...,J mνj +2ps1
s−mνj −2p
am,k r λk −mνj −2p + km,p , (R+ )
j j j j
wk ∈ H s̃ (Γ ), hm,p (r) = km,p ∈ H0
λk +1/2<s
s1 +1/2−
v ∈H
0
(Γ ), s̃ = min {s − s1 , −1/2 + s2 − ε} , ε > 0.
Next we illustrate Theorem 7.16 with an example:

Example 7.4 Let Ω be the exterior of a square in R3 . Let f ∈ H 1 (Γ ) given
in (7.10). Then the unique solution of (7.12) has at the corner the decomposition
π 1/2
v = v 0 + a1 r .297 w1 (θ ) + h1 (r)r 1/2θ 1/2 + h2 (r)r 1/2 −θ
2
j j j
v 0 ∈ H̃ 2−ε (Γ ) hi (r) = b1 r 0.297−1/2 + b2 r 1.426−1/2 + ki (r) , bi ∈ R
ki ∈ H 3/2−ε and r 3/2−ε ki ∈ L2 (R+ )
Next, we consider various boundary element methods for the polyhedron.

7.2.1.1 h-p- Method with Quasiuniform Mesh on a Polyhedron
Construct a quasi uniform mesh with width h analogously as above and define the
space Sp,h (Γ ) of piecewise polynomials of degree p on this mesh. Then the rate
of convergence is determined by the regularity of the solution, i.e. the parameter α
in (7.8):
Theorem 7.17 Let p be sufficiently large and h be small enough. Then the Galerkin
equations (7.3) are uniquely solvable in Sp,h (Γ ). Let v ∈ H 1/2(Γ ) be the solution
of the integral equation (7.2) with sufficiently smooth right hand side f , and let vp,h
the Galerkin solution, then we have with α in (7.8)
0 0
0v − vp,h 0 1/2 C · hα−ε pε−2α . (7.17)
H (Γ )
Remark 7.8 The proof for the h−version (p fixed) is given by the proof of
Theorem 7.18 in Sect. 7.4.2 by setting β = 1. By a refined analysis based on [24]
Schwab and Suri in [374] showed the p−version result (7.17) (h fixed) with = 0.
In [52] Heuer and Bespalov prove (7.17) with = 0 and a possible log p/ h-term.
In Example 7.4 we have α = min {1/2,
.297 + 1/2} = 1/2, hence
Theorem
7.17
yields the convergence rate O h1/2 for the h-method and O p−1 for the p-
method. Here there holds (7.17) with = 0.
7.2.1.2 h-Method with Graded Mesh on a Polyhedron

β
For the construction of the graded mesh see above. Let Sh be the space of
piecewise linear, continuous functions on this mesh. As in Theorem 7.14 we
can compensate the effect of the singularities by an appropriately graded mesh
and get the convergence rate h3/2 . The following theorem follows by combining
Theorem 7.11 with 2D approximation results for the trace v in Sect. 7.4. There we
present its proof from the Thesis [423] of T. von Petersdorff.
Theorem 7.18 ([426]) Let h be sufficiently small and α as in (7.8). Then we have
%
3
hαβ−ε if β <
v − vh H 1/2 (Γ ) C 2α
3
h3/2 if β > 2α
where the constant C = C(β) is independent of h.

7.3 1D-Approximation Results
7.3.1 hp-Method with Quasiuniform Mesh on Polygons
0 (Γ j ) denote those functions in S

Here we report from [405]. Let Sh,p j
h,p (Γ ) which
j
are continuous over Γ .
j
Proposition 7.1 For v ∈ H r (Γ j ), r > 1/2 there exists vh,p ∈ Sh,p
0 (Γ j ), p 1,
such that
j
vh,p (tl ) = v(tl ) for l = j − 1, j
v − vh,p H̃ 1/2 (Γ j ) C hμ−1/2 p−(r−1/2) log1/2 p vH r (Γ j )

j
(7.18)
where μ = min{r, p + 1}.

j
Proof Consider Γh,1 , the first interval of Γ j , assumed to be Ih = [0, h]. Take
1/2
v̂(x) ≡ v( xh ) ∈ H r (I ), I = [0, 1]. By [25] there exists a projection Pp :
H r (I ) → Pp (I ) s.t. ∀ŵ ∈ H r (I )
1/2
Pp ŵ = ŵ at x = 0, x = 1
1/2
Pp ŵ = ŵ for ŵ ∈ Pp (I )
ŵ − Pp ŵH̃ 1/2 (I ) C p−(r−1/2) log1/2 p ŵH r (I )

1/2
Hence for any Ŝ ∈ Pp (I )
1/2 1/2
v̂ − Pp v̂H̃ 1/2 (I ) = (v̂ − Ŝ) − Pp (v̂ − Ŝ)H̃ 1/2 (I )
C p −(r−1/2) log1/2 p inf v̂ − ŜH r (I ) C p −(r−1/2) hμ−1/2 log1/2 p vH r (Ih )
Ŝ∈Pp (I )
(7.19)
Repeating this over each subinterval gives the assertion. In (7.19) we have used that
for k ≥ 0
inf v̂ − p̂H k (I ) ≤ Chμ−1/2 vH k (Ih )

p̂∈Pp (I )
where μ = min(p + 1, k) and C depends on k but is independent of p, h and μ (see

[23] ).

α
Next, we consider the singular functions vj k = cj k ρj jk χj . We look at the
approximating polynomials, which vanish at the endpoints of Γ j , of a function v
defined on Γ j = (−1, 1) by v(x) = (1 + x)α χ(x), where α > 0 and χ is a C ∞
7.3 1D-Approximation Results 243
cut-off function satisfying χ = 1 for x −1/2, χ = 0 for x 0. Let us consider

the weighted spaces W s (μ, ν) with μ, ν ∈ R, integer s > 0 with norm
1 1 2 2
2 −μ ∂su 2 −ν 2
u2W s (μ,ν) = (1 − x ) + (1 − x ) u dx
∂x s
−1
The use of W s (μ, ν) is essential to show that the p-version has twice the conver-
gence rate of the h-version for singular functions.
Lemma 7.1 Let ŵ(x) = (1 + x)α χ(x) for x ∈ I = (−1, 1) with α > 0. Then there
exists ŵp ∈ Pp (I ) with
ŵp (±1) = ŵ(±1) = 0,

ŵ − ŵp H̃ 1/2 (I ) C p−2α+ε
with p 1, min( 32 , 2α) > ε > 0and C independent of p.

◦
Proof First, one sees that ŵ ∈ W s (μ − s, μ), μ = 1
2 + ε3 , for any s < 2α + 1
2 − ε
3
◦
provided α > − 14 + 6ε . Here the weighted Sobolev space W s (μ, ν) is the completion
◦
of the set {u ∈ C0∞ |uW s (μ,ν) < ∞}. Due to [22] ∀ u ∈ W s (μ − s, μ), s >
μ, ∃up ∈ Pp s.t. up = u = 0 at ±1 and
1
u − up ◦ 1/2+ε̃ C p−(s−1/2)+ε̃ u ◦ , 0 < ε̃
H (I ) W s (μ−s,μ) 2
Choosing s = 2α + 1
2 − 2ε̃, ε̃ = ε
3 yields with up = ŵp
ŵ − ŵp ◦ 1/2+ε̃ C p−2α+ε

H (I )
which gives the assertion.

We note that in [22] u is expanded in a series of Jacobi polynomials and the weighted
◦
spaces W s (μ − s, μ) are the right setting.
Next let us consider
v(x) = x α χ̃(x) on I = (0, 1) (7.20)
with χ̃ ∈ C0∞ s.t. χ̃ = 1 for x 1/4, χ̃ = 0 for x 1/2.

Theorem 7.19 For v in (7.20) on Γ j = (0, 1) and α > 0 ∃vh,p ∈ Sh,p

0 (Γ j ) s.t.
v − vh,p H̃ 1/2 (Γ j ) C max{hα p−2α+ε , min{hα , hp+1/2 p−2α log1/2 p}}
with C > 0 independent of h and p.

Proof Split v into w1 , w2 with w1 (x) = v(x)χ( xh ), w2 (x) = v(x)(1 −χ( xh )). Then
w1 (x) = x α χ( xh ) = hα ŵ( xh ) with ŵ as in lemma above. Then ∃ wp (x) = ŵp ( xh ) ∈
Pp (Ih ), Ih = (0, h) with wp = 0 at x = 0, x = h and
w − wp H̃ 1/2 (Ih ) C p−2α+ε
1 (x) = hα w (x) ∈ P (I ) gives

Now taking wp,h p p n
1
w1 − wp,h H̃ 1/2 (Ih ) C hα p−2α+ε (7.21)
1 by 0, we get a function in S 0 (Γ j ) s.t. (7.21) holds in H̃ 1/2 (Γ j ).

Extending wp,h h,p
Then approximating w2 in [h/4, 1] yields the assertion (see [405]).

Next we observe that the antiderivative v of ψs can be approximated in the H̃ 1/2(Γ i )
norm by a polynomial vp,h . Therefore the ψp,h , defined to be the derivative of vp,h
(with respect to arc length) will approximate ψs in the H̃ −1/2 (Γ i )norm with the
same accuracy.
j
Proposition 7.2 ([405]) For ψ ∈ H r (Γ j ), r > −1/2 there exists ψh,p ∈
Sh,p (Γ j ), such that
ψ − ψh,p H̃ −1/2 (Γ j ) C hμ+1/2 p−(r+1/2) log1/2 p ψH r (Γ j )

j
(7.22)
where μ = min{r, p + 1}.

j h
Proof Take Γh,1 = Ih and let ψ ∈ H r (Ih ) with r > 0 and let ψ̄ = h1 0 ψ(t) dt
x
and define v(x) = 0 (ψ − ψ̄)(t) dt.Then v ∈ H r+1 (Ih )∩ H̃ 1/2 (Ih ). By (7.19) there
1/2
exists a polynomial Pp v ∈ Pp+1 (Ih ) such that
v − Pp vH̃ 1/2 (Ih ) C p−(r+1/2) hμ+1/2 log1/2 p vH r+1 (Ih )

1/2
where μ = min{r, p + 1}. Now taking ψp = (Pp v) + ψ̄ we have

1/2
ψ −ψp H̃ −1/2 (Ih ) ≤ Cv −Pp vH̃ 1/2 (Ih ) C p−(r+1/2) hμ+1/2 log1/2 p ψH r (Ih )
1/2
Repeating this over each subinterval completes the proof.

For the convenience of the reader we want to give a further detail where u is
expanded in a series of Chebyshev polynomials. Let
u(x) = (x + 1)1/2χ(x) , x ∈ I = [−1, 1] (7.23)
with χ ∈ C ∞ satisfying χ(x) = 1 for −1 ≤ x ≤ −1/2, χ(x) = 0 for 1/2 <

x < 1. (Other singularity functions (x + 1)α can be treated similarly [22].) We
consider the approximation of u in the H̃ 1/2(I )-norm by functions in Pp (I ). Let
u be transformed to the periodic function û on Iˆ = [−π, π] by the mapping x =
cos(ξ ), i.e. û(ξ ) = u(x). Then
√
û(ξ ) = (1 + cos(ξ ))1/2 χ(cos(ξ )) = 2 χ(cos(ξ )) cos(ξ/2)
Theorem 7.20 ([404]) Let u be defined by (7.23). Then for p = 1, 2, . . . there

exists a polynomial u0p ∈ Pp (I ) s.t.
u0p (±1) = u(±1) (7.24)
u − u0p H̃ 1/2 (I ) ≤ Cp−1 log1/2 p (7.25)
∞
Proof Write û(ξ ) = ak cos(kξ ) and set u0p := up +u where u is a linear function
k=0
s.t u0p satisfies (7.24) and up ∈ Pp (I ) is defined in terms of Chebyshev polynomials
p
Tk (x) = cos(k cos−1 (x)) of degrees ≤ p by ûp = ak Tk (cos(ξ )). Now
k=0
π
ak = c û cos(kξ )dξ
0
satisfies
C
|ak | ≤ .
k2
Therefore
∞
∞
(1 + k 2 )1/2
u − up 2H 1/2 (I ) = û − ûp 2 =C ak2 (1 + k 2 )1/2 ≤ C
H 1/2 (Iˆ) k4
k=p+1 p+1
which behaves like

∞ C C
dx = 2 .
p+1 x3 p
Hence
C
u − up H 1/2 (I ) ≤ .
p
Now for any x we have

∞
∞
C C
|(u − up )(x)| ≤ |ak | ≤ 2
≤ (7.26)
k p
k=p+1 k=p+1
Furthermore
u − u0p H̃ 1/2 (I ) ≤ u − u0p H 1/2 (Iˆ) + (1 − x 2 )−1/2 (u − u0p )H 0 (I )
Hence we must bound the second term. We have
1 1/p π−1/p
π
2 −1
(1 − x ) (u − u0p )2 dx =( + + )(û − û0p )2 sin(ξ ))−1 dξ
−1 0 1/p π−1/p
Now 1/ sin(ξ ) is bounded on [1/p, π − 1/p]. Hence using (7.26)

π−1/p
π−1/p
C C
(û − û0p )2 sin(ξ ))−1 dξ ≤ 2 sin(ξ ))−1 dξ ≤ log p.
p p2
1/p 1/p
Furthermore (see [404], p. 38 for details) since
1/p
C
(û − û0p )2 sin(ξ ))−1 dξ ≤ 2
p
0
altogether we have
C log1/2 p
(1 − x 2 )−1/2 (u − u0p )H 0 (I ) ≤
p

In the framework of Jacobi-weighted Besov and Sobolev spaces in 1D Guo and
Heuer analyze in [208] lower and upper bounds for approximation errors in the p-
version BEM for hypersingular and weakly singular integral operators on polygons.
They prove optimal convergence rates.
7.3.2 Approximation of the Normal Derivative on a One

Dimensional Boundary—The h-Version on a Graded
Mesh
Here we consider the approximation of singular functions y ν−1 on I = [0, 1].

Similar results hold for y ν−1 log(y).
h
Lemma 7.2 ([423]) Let ν > 0, 1
2 − ν < σ 1, σ 0, p := h−1 y ν−1 dy.
0
Then ∃Cν ∈ R, independent of h, such that
y ν−1 − pH̃ −σ ([0,h]) Cν hν+σ −1/2 (7.27)
h
Proof Let q := h−1 g dy. Then
0
y ν−1 − p, g y ν−1 , g − q

y ν−1 − pH̃ −σ ([0,h]) = sup = sup
g∈H σ ([0,h]) gH σ ([0,h]) g∈H σ ([0,h]) gH σ ([0,h])
For ν > 1/2 there hold y ν−1 ∈ L2 ([0, h]) and hence
y ν−1 − pH̃ −σ ([0,h]) Cν hν−1/2 Chσ (7.28)
For 0 < ν 1/2 choose s with 1

2 − ν < s < σ . Using ỹ := h−1 y gives
g(h ỹ) − qH s ([0,1])

y ν−1 − pH̃ −σ ([0,h]) Cν hν sup
g∈H σ ([0,h]) gH σ ([0,h])
Furthermore interpolating between L2 ([0, 1]) and H 1 ([0, 1]) one obtains
g(h ỹ) − qH s ([0,1]) c hσ −1/2 gH σ ([0,h]) .

Proposition 7.3 ([423]) Let ψ ∈ H̃ −1/2 ([0, 1]) have the form

K
ψ(x) = ak x νk −1 + ψ 0 (x) with νk > 0 , ψ 0 (x) ∈ H 1 ([0, 1])
k=1
β
with ν0 := min{νk }.Let Sh be space of piecewise constant functions on mesh xk =
k β
N , k = 0, . . . , N with β 1, h = 1/N. Then ∀ε > 0 and −1 s ν0 − 1/2
there holds
%
h(ν0 −s−1/2)β−ε , 1β 1−s
ν0 −s−1/2
inf ψ − ph H̃ s ([0,1]) c 1−s
(7.29)
β
ph ∈Sh h1−s , β> ν0 −s−1/2
Proof Note for 1 − 1/β γ̃ 1 we have
γ̃
hk = xk − xk−1 βhβ k β−1 h(1−γ̃ )β βxk
We approximate seperately the regular and singular parts of ψ. Let p0 denote on

each Ik = [xk−1 , xk ] the mean value of ψ 0 . Then

N
N
dψ 0
ψ 0 − p0 2H̃ s (I ) ψ 0 − p0 2H̃ s (I ) C h−2s+2 2
k k
dx L (Ik )
k=1 k=1
−2s+2
Ch ψ 0 2H 1 (I )
Next we consider f (x) = x νk −1 . On Ik , k > 2, we proceed as above and get:

N
N
f − p2H̃ s (I ) C f 2L2 (I
γ̃ (−2s+2)
h(1−γ̃ )β(−2s+2)xk
k k)
k=2 K=2
Since for k 2, and x ∈ Ik , we have

β
k
xk = (k h)β = xk−1 < 2β xk−1 < 2β x
k−1
this yields

N 1
f − p2H̃ s (I ) ch (1−γ̃ )β(−2s+2)
|f (x)|2 x γ̃ (−2s+2)dx
k
k=2 0
if the intergral exists, i.e. if
νk − s − 1/2
1 − γ̃ < . (7.30)
1−s
Finally on the interval I1 we take (7.27) on [0, h1 ] with h1 = hβ and obtain with
σ = −s
x νk −1 − pH̃ s ([0,h1 ]) C hνk −s−1/2)β .

Now, choose γ̃ by 1 − γ̃ = min{ ν0 −s−1/2

1−s − ε̃, β1 } with ε̃ > 0. Hence γ̃ ≥ 1 − β1
and (7.30) holds. Altogether we obtain (7.29).

7.4 2D-Approximation Results
Lemma 7.3 ([423]) Let Q, Qj (j = 1, . . . , N) be Lipschitz domains with Q =

*N
Qj , ũ ∈ H̃ s (Q), u ∈ H s (Q), s ∈ [−1, 1]. Then there holds
j =1

N
u2H s (Qj ) u2H s (Q) (7.31)
j =1

N
ũ2H̃ s (Q) ũ2H̃ s (Q (7.32)
j)
j =1
B
N
Proof 0 s 1 (larger s analogously). Consider the map T : H̃ s (Qj ) →
j =1
H˜ s (Q) which extends uj on Qj to u on Q. T is continuous for s = 0, 1 with norm
BN
1 when H̃ s (Qj ) (s = 0, 1) has norm
j =1

N
(uj )j =1,...,N 2 = uj 2H̃ s .
(Qj )
j =1
Now interpolation yields (7.32), since there holds

N
(uj )j =1,...,N 2[s] uj 2[s]
j =1
(Here · [s] denotes the respective interpolation norm).

Inequality (7.31) is obvious for s = 0, 1. By duality this yields (7.32) for s = 1:
u, v H̃ −s (Q)×H s (Q)

uH̃ −s (Q) = sup
v∈H s (Q) vH s (Q)
N
u|Qj , v|Qj H̃ −s (Qj )×H s (Qj )
j =1
= sup
v∈H s Q vH s (Q)
N
uH̃ −s (Qj ) vH s (Qj )
j =1
sup
v∈H s (Q) vH s (Q)
⎛ ⎞1/2
N
⎜N v2
H (Qj ) ⎟
s
⎜ j =1 ⎟
⎜ u2H̃ −s (Q ) sup ⎟
⎝ j
v∈H s (Q) v2H s (Q) ⎠
j =1
⎛ ⎞1/2

N
⎝ u2H̃ −s (Q ) ⎠
j
j =1
With interpolation one obtains now (7.32) for −1 −s 0. Finally duality

yields (7.31) for 0 s 1 (see also Remark 2.2 in [86]).

The next lemma allows to estimate tensor product functions.
Lemma 7.4 ([423]) Let Ij = [0, hj ], 0 sj 1, fj ∈ H̃ −sj (Ij ) for j = 1, 2.
Then there holds
f1 (x)f2 (y)H̃ −s1 −s2 (I1 ×I2 ) f1 H̃ −s1 (I1 ) f2 H̃ −s2 (I2 )
Next we approximate H̃ s -functions on rectangles by constants.

Lemma 7.5 ([423]) Let −1 s 0, R = [0, h1 ] × [0, h2 ], u ∈ H 1 (R) and
p = h11h2 u(x, y)dy dx. Then holds
R

u − pH̃ s (R) C max{h1 , h2 }−s h1 ux L2 (R) + h2 uy L2 (R) (7.33)
If u(x, y) = u1 (x)u2 (y), uj ∈ H 1 ([0, hj ]) (j = 1, 2) then
u − pH̃ (R) c h1−s 1−s

1 ux L2 (R) + h2 uy L2 (R)

Two-dimensional Boundary—The h-Version
on a Graded Mesh
Here we prove Theorem 7.9 (see the thesis by T. von Petersdorff [423]). The results
are derived for the h-version on graded meshes and contain automatically the case
of a quasiuniform mesh by setting the grading parameter β = 1.
Fig. 7.6 Mesh on a square y
x3
R13
x2
R12 R22
x1
R11 R21 R31
0 x1 x2 x3 1 x
Proof Approximation of regular part ψ 0 : We approximate ψ 0 on each rectangle

Rkl (see Fig. 7.6) by the L2 -projection p|Rkl = hk1hl ψ 0.
Rkl
Then we use (7.32) for s = −1/2 and estimate the approximation error on each
rectangle with (7.33) and hk ≤ βh:

N
ψ 0 − p2H̃ −1/2 (Q) c ψ 0 − p2H̃ −1/2 (R
kl )
k,l=1

N
c max{hk , hl } h2k ψx0 2L2 (R + h2l ψy0 2L2 (R c h3 ψ 0 2H 1 (Q)
kl ) kl )
k,l=1
Approximation of corner singularity f = r λk −1 wk (θ ), wk ∈ H 1 ([0, π/2]):

With p|Rkl := hk1kl f (x, y)dy dx and (7.32) and (7.33) one obtains
Rkl

N
f − p2H̃ −1/2 (Q) c max{hk , hl } h2k fx 2L2 (R + h2l fy 2L2 (R
kl ) kl )
k,l=1
k+l=2
+ cf − pH̃ −1/2 (R11 ) (7.34)
Estimate for k 2, l 2: Note for k 2, x ∈ [xk−1 , xk ] there holds |hk |

β2βγ h x γ with γ = 1 − β1 > 0. Therefore
max{hk , hl }h2k fx 2L2 (R cfx max{x γ , y γ }1/2 x γ 2L2 (R · h3

kl ) kl )
and

N 1 1
f − p2H̃ −1/2 (R ) ch 3
|fx (x, y)|2 max{x γ , y γ }x 2γ dy dx
kl
h,l=2 0 0
if the integral exists. Note
|fx (x, y)| r λk −2 w̃(θ ), (7.35)
with w̃ ∈ L2 ([0, π/2]); further max{x γ , y γ } r γ . Hence the above integral

exists if
3
β> . (7.36)
2(λk + 1/2)
Proceeding analogously for fy we get under this condition that

N
f − p2H̃ −1/2 (R c h3
kl )
k,l=2
Estimate for k = 1, l > 1(analogously k > 1, l = 1):(7.34) gives

N
N
f − p2H̃ −1/2 (R ) c max{h1 , hl } h21 fx 2L2 (R + h2l fy 2L2 (R
1l 1l ) 1l )
l=2 l=2
The term with fx is bounded by

N
N
N
c h3
3γ
c h3l fx 2L2 (R xl−1 fx 2L2 (R c h3 fx (x, y)y 3γ /22L2 (R
1l ) 1l ) 1l )
l=2 l=2 l=2
h1 1
ch 3
|fx (x, y)|2 y 3γ dy dx
x=0 y=0
if the integral exists. With (7.35) this is bounded by

√
2
π/2
ch 3
r 2λk −4 r 3γ rdr dφ.
r=0 φ=0
3
This integral exists for β > 2(λk +1/2) . The fy -term is handled analogously yielding

N
f − p2H̃ −1/2 (R c h3 .
1l )
l=2
Estimate for k = 1, l = 1: f ∈ L2 (R11 ) because λk > 0. Now
f − pL2 (R11 ) f L2 (R11 ) = c hλ1 k
For any constant q there holds
f − p, g f − p, g − q
f − pH̃ −1 (R11 ) = sup = sup
g∈H 1 (R11 ) gH 1 (R11 ) g∈H 1 (R11 ) gH 1 (R11 )
Choose q as L2 projection of g. Then
g − qL2 (R11 )
f − pH̃ −1 (R11 ) f − pL2 (R11 ) sup C hλ1 k h1
g∈H 1 (R11 ) gH 1 (R11 )
Hence interpolation gives with (7.36)
λ +1/2
f − pH̃ −1/2 (R11 ) c h1 k = c hβ(λk +1/2) ≤ Ch3/2
Approximation of edge singularities: There are two types of edge singularities:

ν −1
(1) f (x, y) = χi (θi )bim ρi im with regular edge intensity factor bim ∈ H01 (R+ )
λ −ν ν −1
(2) f (x, y) = χi (θi )yi k im ρi im with corner singularities in the edge intensity
factor.
We consider Q = [0, 1]2 with singularity at the x-axis, i.e. yi = x, ρi = y.
With (7.32) we have

N
f − p2H̃ −1/2 (Q) f − p2H̃ −1/2 (R (7.37)
kl )
k,l=1
First we consider case (2):

Estimate for l k, l 2: Define χ̃ ∈ C ∞ ([0, π/2]) with χ̃ (θ ) = 0 for θ ∈ [0, φ20 ]
and = 1 for θ ∈ [φ0 , π/2] where φ0 is sufficiently small, namely tan φ0 2−β .
Then g := χ̃f = f on Rkl , l k, l 2. Now, g = r λk −1 w(θ ), w ∈
H 1 ([0, π/2]), and the proof for the corner singularity yields an approximation q
Fig. 7.7 The domains Rj∗ y
x3
x2
x1
R2∗ R3∗ R4∗ R5∗
0 x1 x2 x3 1 x
on Rkl with

f − q2H̃ −1/2 (R g − q2H̃ −1/2 (R c h3 .
kl ) kl )
lk k,l=1
l2
Estimate for l < k. Here θ < π/4 and χi (θ ) = 1. Hence f (x, y) = f1 (x)f2 (y).
* −1
j*
Divide R ∗ := Rkl into subdomains Rj∗ = Rj l , j = 2, . . . , N (Fig. 7.7).
1lk l=1
Setting f1 (x) = x λ−ν , f2 (y) = y ν−1 , ,Ik = [xk−1 , xk ], Ik∗ = [0, xk ] we have

N
f − p2H̃ −1/2 (R ∗ ) f − p2H̃ −1/2 (R ∗ )
j
j =2
For ν > 1/2 there holds (for ν < 1/2 see [423] )
f −pH̃ −1/2 (R ∗ ) f1 −p1 H̃ −1/2 (Ij ) f2 L2 (I ∗ +f2 −p2 H̃ −1/2 (I ∗ p1 L2 (Ij )
j j−1 ) j−1 )
with p1 , p2 mean value of f1 , f2 on Ij , Ij∗−1 respectively.

Now

N
N j −1

f2 − p2 2H̃ −1/2 (I ∗ p1 2L2 (I ) f1 2L2 (I f2 − p2 2H̃ −1/2 (I
j−1 ) j j) k)
j =2 j =2 k=1
1 x
ch 3
x 2λ−2ν y 2ν−4+3γ dy dx
x=0 y=0
1
3
=c h x (2λ−3+3γ )dx ≤ c h3
x=0
These integrals exist for 2λk − 3 + 3γ > −1 and 2νim − 4 + 3γ > −1, i.e.
1 3
β= > (7.38)
1−γ 2 min{λk + 12 , νim }
On the other hand for

2 min{(λk + 1/2, νim )}
γ >1− (7.39)
3
we have

N
N
f1 − p1 2H̃ −1/2 (I ) f2 2L2 (I ∗ C h3j f1 2L2 (i ) |xj2ν−1
−1 |
j j−1 ) j
j =2 j =2

N
c h3 x (λ−ν−1)+3γ /2+(ν−1/2)2L2 (I
j)
j =2
1
ch 3
x 2λ+3γ −3dx ≤ c h3
x1
Estimate for k = l = 1 Consider λ − ν > −1/2 then
f −pH̃ −1/2 (R11 ) x λ−ν y ν−1 −qH̃ −1/2 (R11 ) +(1−χ (θ )x λ−ν y ν−1 −q̃H̃ −1/2 (R11 )
(7.40)
Here q, q̃ denote the respective mean values on R11 . Then for 0 < ε < λ
x λ−ν y ν−1 − qH̃ −1/2 (R11 ) x λ−ν L2 (I1 ) y ν−1 − q2 H̃ −1/2 (R11 )
+ y ν−1 ν− 1 −ε x λ−ν − q1 H̃ ε−ν (R11 ) C hν1

H̃ 2 (I1 )
with the mean values q1 , q2 of x λ−ν and y ν−1 on I1 . Now (1 − χ(θ ))x λ−ν y ν−1 =
r λ−1 w̃(θ ), w̃ ∈ L2 ([0, π/2]) hence, the second term in (7.40) is bounded by
λ+1/2
C h1 . For λ − ν 1/2 there holds ν > 1/2 and f = χ (θ )x λ−ν y ν ∈ L2 (R11 ).
Hence
1/2 λ+1/2
χ (θ )x λ−ν y ν−1 − pH̃ −1/2 (R11 ) h1 χ (θ )x λ−ν y ν−1 L2 (R11 ) C h1
In both cases we thus have with h1 = hβ and (7.38)
f − pH̃ −1/2 (R11 ) C h3/2 .
Next we consider case (1)(Edge singularity with regular edge function):
f (x, y) = b(x)y ν−1 + (χ(θ ) − 1)b(x)y ν−1 =: f1 + f2
Note f2 ∈ H 1 (Q), and thus is approximated with order h3/2 like the regular part
ψ 0 . f1 has tensor product form. Let q denote the mean value of f1 on each Rj k ,
q1 the mean value of b(x) on [xk−1 , xk ] and q2 for y νim −1 . Then for ν ≤ 1/2, any
ε̃ > 0 and (7.38)
b(x)y ν−1 − q1 (x)q2(y)H̃ −1/2 (Q) bL2 (I ) y ν−1 − q2 H̃ −1/2 (I )
+ y ν−1 H̃ ν−1/2−ε (I ) b − q1 H̃ ε−ν (I ) C h3/2−ε̃ + c h1+ν−ε
For ν > 1/2 there holds
b(x)y ν−1 − q1 (x)q2(y)H̃ −1/2 (Q) bL2 (I )y ν−1 − q2 H̃ −1/2 (I )
+ y ν−1 L2 (I ) b − q1 H̃ −1/2 (I ) c h3/2−ε̃ + C h3/2 .
Collecting the estimates for the various parts of ψ on the subdomains gives
with (7.37) the assertion of Theorem 7.9.

Finally, we consider the case β < 3/2α0 . Then
γ̃ 1
hk β hβ(1−γ̃ ) xk , 1 − γ̃ 1.
β
Now, we can take γ̃ instead of γ = 1 − β1 and perform the proof analogously by

γ
choosing γ̃ = 1 − 2α0 /3 + ε, ε > 0. Before we obtained with hk βxk the order
h3/2 , now we get only
hβ(1−γ̃ )3/2 = hβ(2α0 /3−ε)3/2 = hα0 β−ε̃ .

7.4.2 Approximation of the Trace on a Two-Dimensional

Boundary—The h-Version on a Graded Mesh
In this section we prove Theorem 7.18. Here it helps to use anisotropic Sobolev
spaces.
Definition 7.1 Let H (1,0)(Ω) denote the closure of C ∞ (Ω) in the norm
u2H (1,0) (Ω) := u2L2 (Ω) + ux 2L2 (Ω) .
We define the anisotropic space H (s,0)(Ω) by interpolation:
H (s,0)(Ω) := L2 (Ω), H (1,0)(Ω) , H (0,s)(Ω) := L2 (Ω), H (0,1)(Ω)

[s] [s]
Note for the homogeneous Sobolev spaces we have for s 0
H s (Ω) = H (s,0)(Ω) ∩ H (0,s)(Ω)
and
uH s (Ω) CuH (s,0) (Ω) + CuH (0,s) (Ω) .
Further we need Sobolev spaces, where the functions satisfy only on a part of the
boundary a condition like in the H̃ -spaces.
Definition 7.2 Let Ω be Lipschitz, Γ0 ∩ Ω ⊂ ∂Ω part of the boundary,
◦1 & '
H Γ0 (Ω) := u ∈ H 1 (Ω)| u|Γ0 = 0

◦1
H̃Γs 0 (Ω) := L2 (Ω), H Γ0 (Ω)
[s]
Herewith for s = 1/2 we can estimate the norm of a function by the norms in the
subdomains, but with constants depending on the domains.
*
Lemma 7.6 ([423])* Let Ω, Ω1 , Ω2*be Lipschitz with Ω = Ω11 Ω2 , Γ0 = ∂Ω1 ∩
∂Ω2 , ∂Ω1 = Γ0 Γ1 , ∂Ω2 = Γ0 Γ2 , 0 s 1, s = 2 . Then there exists a
constant c > 0 such that ∀u ∈ H s (Ω), ũ ∈ H̃ s (Ω) there holds
uH s (Ω) cuH s (Ω1 ) + cuH s (Ω2 )

ũH̃ s (Ω) cuH̃ s + cuH̃ s
Γ1 (Ω1 ) Γ2 (Ω2 )
Proof Consider the mapping T which maps u ∈ H s (Ω) into (u|Ω1,u|Ω2) ∈

H s (Ω1 ) × H s (Ω2 ). For s = 12 the range of T is closed in H s (Ω1 ) × H s (Ω2 )
◦
which is equivalent to H s (Ω) = H̃ s (Ω) for s = 12 (cf. [415][(5.22)]). Hence the
assertion of the lemma follows from the graph theorem.

First, we consider approximation with bilinear functions on rectangles.
Lemma 7.7 ([423]) Let Q = [0, h1 ] × [0, h2 ], u ∈ H 3 (Q), p the bilinear
interpolant of u at the vertices of Q. Then there holds
u − pL2 (Q) C h21 uxx L2 (Q) + h22 uyy L2 (Q) + h21 h2 uxxy L2 (Q)
(7.41)
(u − p)x L2 (Q) C h1 uxx L2 (Q) + h22 uxyy L2 (Q) (7.42)
Proof First we note that for u ∈ H 2 (I ), I = [0, 1], with linear interpolant Πu at 0
and 1 there holds:
u − ΠuL2 (I ) u L2 (I ) (7.43)

ΠuL2 (I ) uL2 (I ) + Cu L2 (I ) (7.44)
On Q = [0, 1]2 , Πx denotes the partial interpolation operator in x,

i.e. (Πx u) (x, y) := (Πu(·, y)) (x) for all y ∈ I . Now p := Πy Πx u and
u − Πy Πx uL2 (Q) u − Πy uL2 (Q) + Πy (u − Πx u) L2 (Q)
(7.43) yields
1
u − Πy xL2 (Q) uyy (x, ·)2L2 (I ) ∂x uyy 2L2 (Q) (7.45)
0
Further for fixed x ∈ I we have with (7.44) and (7.45)
∂
Πy (u − Πx u) (x, ·)L2 (I ) ≤ (u − Πx u) (x, ·)L2 (I ) +C (u − Πx u) (x, ·)L2 (I )
∂y
cuxx 2L2 (Q) + cuxxy 2L2 (Q) .
Hence, (7.41) holds for Q = [0, 1]2. The substitution x̃ = h1 x, ỹ = h2 y

gives (7.41) for Q = [0, h1 ] × [0, h2 ]. (7.42) follows analogously.

Now we prove Theorem 7.18.
β
Approximation of regular part v 0 : Let p ∈ Sh denote the interpolant of v 0 , where
β
Sh are the linear functions on the graded mesh (see Fig. 7.6) xk = (kh)β , yl =
(lh)β . Hence

N
N
v 0 − p2L2 (Q) = v 0 − p2L2 (R h4j vxx
0 2
L2 (R + h4k vyy
0 2
L2 (R
jk ) jk ) jk )
j,k=1 j,k=1
+ h4j h2k vxxy

0
2L2 (R h4 v 0 2H 3 (Q)
jk )

N
N
v 0 − p2H 1 (R = v 0 − p2H 1 (R c h2j vxx
0 2
L2 (R + h2k vyy
0 2
L2 (R
jk ) jk ) jk ) jk )
j,k=1 j,k=1
+ h4j vxxy
0
2L2 (R + h4k vxyy
0
2L2 (R h2 v 0 2H 3 (Q)
jk ) jk )
Interpolation yields v 0 − pH 1/2 (Q) h3/2 v 0 H 3 (Q).

Approximation of corner singularity u = r λ w(θ ) with w ∈ H 3 ([0, π/2]). We
devide the square Q = [0, 1]2 in N − 1 overlapping domains Aj , j = 1, . . . , N − 1
(see Fig 7.8).

Aj = {Rkl |j k j + 1, l k or j l j + 1, k l}
Further we take a partition of unity {χj } on Q w.r.t. the sets Aj , e.g. let χkl the
piecewise bilinear function
3 on the mesh, which 4 is 1 at (xk , xl ) and 0 at all other
nodes. We take χj = χkl |suppχkl ⊂ Aj . Then |(χj )x | h−1 −1
j , |(χj )y | hj .
Fig. 7.8 The domains Aj 1
xj
Aj
x2
A1
x1
x 1 x2 xj 1
We estimate the approximation error in each Aj in L2 and H 1 ; then interpolation

gives an upper bound for χj (u − p) in H̃ 1/2. This is then used to derive the estimate
on Q.
For j > 1 with hk ≤ hj +1 when Rkl ⊂ Aj we have

u − p2L2 (A ) c h4k uxx 2L2 (R + h4l uyy 2L2 (R + h4k h2l uxxy 2L2 (R
j kl ) kl ) kl )
Rkl
u − p2H 1 (A ) c h2j +1 uxx L2 (Aj ) + h2j +1 uyy L2 (Aj )

j
+ h4j +1 uxxy 2L2 (A ) + h4j +1 uxyy 2L2 (A

j j)
This implies with u = r λ w(θ )

u − p2L2 (A ) C hj +1 xj −1 h4j +1 xj2λ−4 2λ−6
−1 + hj +1 xj −1
6
j

u − p2H 1 (A ) C hj +1 xj +1 h2j +1 xj2λ−4 4 2λ−6
−1 + hj +1 xj −1
j
Note that χj vanishes on ζj := ∂Aj \ ∂Q, hence interpolation yields

χj (u − p)2 1/2 C hj +1 xj +1 h3j +1 xj2λ−4
−1 .
H̃ζ (Aj )
j
Since r λ ∈ H 1 (Q) we have for j = 1
u − p2L2 (A ) Cx22λ+2 , u − p2H 1 (A ) C x22λ,

1 1
hence
χ1 (u − p)2 1/2 C x22λ+1

H̃ζ (A1 )
1
Since the terms χj (u − p) with even and odd j have different supports and vanish
on ζj , there holds

u − p2H 1/2 (Q) 2 χj (u − p)2 1/2 +2 χj (u − p)2 1/2
H̃ζ (Aj ) H̃ζ (Aj )
j odd j j even j
γ
Due to hk βxk h with γ := 1 − 1/β the terms for j = 2, . . . , N − 1 are bounded
by

N−1 1
3
C h4j +1 xj2λ−3
+1 Ch 3
x 2λ−3+3γ dx = O(h3 ) for β > .
2(λ + 1/2)
j =2 x2
Also the term with j = 1 is of same order. Hence in total for the corner singularity
we have convergence rate h3/2 in H 1/2(Q).
Approximation of singular edge functions:
f (x, y) = x λ−ν y ν χ(x) (7.46)
with corner exponent λ, edge exponent ν and cut-off function χ with χ (θ ) = 1 near
θ = 0. Divide Q = (0, 1)2 into 2 triangles
A: = {(x, y) ∈ Q|y x} , B := {(x, y) ∈ Q|y > x} ,
and estimate the interpolation error on each triangle in H 1/2+ε , 0 ≤ ≤ 1/2.

Lemma 7.6 yields error bound in H 1/2+ε (Q).
On triangle B there holds f (x, y) = r λ w(θ ) with w(θ ) smooth on [ π4 , π2 ].
Extend w(θ ) to a smooth function w̃(θ ) on [0, π2 ] and define f˜(x, y) := r λ w̃(θ ).
Now f˜ can be approximated on Q like a corner singularity and for its interpolant p̃
there holds
f˜ − p̃H 1/2+ε (Q) c h3/2−ε
yielding, by restriction to B,
f − pH 1/2+ε (B) c h3/2−ε .
On triangle A we estimate the approximation error separately in the anisotropic

spaces H (s,0) and H (0,s) with s > 1/2. Consider the larger domain Ã with (see
Fig. 7.9)

A ⊂ Ã := {Rkl |l k + 1}
Fig. 7.9 The rectangles Aj 1
Ã
Aj
x2
x1
x 1 x2 xj 1
Let χj (x) be piecewise linear on 0 < x1 < · · · < xN−1 < 1 with
χj (xj ) = 1 and vanishing in the other nodes, j = 2, . . . , N − 2. χ1 pw. linear
with χ1 (0) = χ1 (x1 ) = 1 and χ1 (xj ) = 0 for j > 1. χN−1 pw. linear with
χN−1 (xN−1 ) = χN−1 (1) = 1 and χN−1 (xj ) = 0 for j < N − 1. Divide Ã into
overlapping rectangles Aj = [xj −1 , xj +1 ] × [0, xj +1 ], j = 1, . . . , N − 1. With
%
(f (x, y) − p(x, y)) χj (x) on Aj
gj := (7.47)
0 else
N−1
there holds f − p = gj . With χ (θ ) ≡ 1 on A2 , . . . , AN−1 in (7.46) we have
j =1
f (x, y) = f1 (x)f2 (y) with f1 (x) := x λ−ν , f2 (y) := y ν . Let ζj denote the “left
and right” boundary of Aj for j = 2, . . . , N − 2; ζ1 the “right boundary” of A1 ,
ζN−1 the “left boundary” of AN−1 . Then

N−1
gj 2H (s,0) (Ã) 2 gj H̃ (s,0) (A ) + 2 gj H̃ (s,0) (A
ζj j ζj j)
j =1 j odd j even
Now for j 2 we have with the interpolant pj of fj with s > 1/2
gj 2 χ1 f1 2H̃ s ([x f2 − p2 2L2 ([0,x

j−1 ,xj+1 ]) j+1 ])
(s,0)
H̃ζ (Aj )
j
+ χj (f1 − p1 )2H̃ s ([x p2 2L2 ([0,x =: B1 + B2

j−1 ,xj+1 ]) j+1 ])
Let us first consider the term B1 : Interpolation gives
1/2−s λ−ν
χj f1 H̃ s ([xj−1 ,xj+1 ]) c hj xj
On the other hand we use the one dimensional approximation result

2
xj +1 −2γ
f2 − p2 L2 ([0,xj+1 ]) c xj +1 f2 (x)x 2γ L2 ([0,xj+1 ])
j +1
With xj +1 /(j + 1) hj +1 this implies for f2 (x) = x ν
−4γ 2ν−3+4γ
f2 − p2 2L2 ([0,x c h4j xj +1 xj +1
j+1 ])
if 2ν − 4 + 3γ > −1, i.e. γ > 1 − 2ν/3, hence β > 3/(2ν). Altogether

N−1 1
4−2s
χj f1 2H̃ s ([x ,x ]) f2 −p2 2L2 ([0,x ]) ch x 2λ−3+(4−2s)γ dx < ∞
j−1 j+1 j+1
j −2 0
(7.48)
for s < 12 + ε with ε sufficiently small and β > 3

2(λ+1/2) .
Next, we consider the term B2 : First we have
2(λ−ν−2)
χj (f1 − p1 )2H̃ s ([x c hj4−2s hj xj
j−1 ,xj+1 ])
Hence with f2 (x)) = x ν

N−1
N−1
χj (f1 − p1 )2H̃ s ([x ,x ]) p2 2L2 ([0,x ]) C hj5−2s xj2λ−2ν−4xj2ν+1 ,
j−1 j+1 j+1
j =2 j =2
which can be bounded like (7.48) above.

Finally, we must bound g1 in (7.47) on A1 . This is done by a similarity argument:
A1 = [0, x2 ]2 is mapped with x̃ = x2−1 x, ỹ = x2−1 y onto [0, 1]2 and |g1 (x, y)| =
x2λ |g̃(x, y)| with g̃(x, y) = χ̃(x, y) r λ w(θ ) − p̃ with functions χ̃, p̃ independent
of h. Interpolation gives for 0 s 1
g1 (x, y)H (s,0) [0,x2 ]2 C x21−s+λ C hβ(1−s+λ).
This yields for s = 1

2 + ε, β > 3/(2λ + 1) and restriction of Ã to A for ε̃ > 0 (suff.
small)
f − pH (1/2+ε,0) (A) C h3/2−ε̃
Next, we estimate f − p in H (0,s)(Ã). We subdivide Ã into disjoint rectangles

Ãj = [xj −1 , xj ] × [0, xj0 ], j0 := min{j + 1, N}, j = 1, . . . , N; see Fig. 7.10.
In x-direction only L2 -regularity is used (Fig. 7.10). Hence

N
f − p2H (0,s) (Ã) f − p2
H (0,s) Ãj
j =1
Fig. 7.10 The rectangles Ãj 1
x2
Ã1 Ã2 ÃN
x1
x 1 x2 xj 1
On the rectangles Ãj (j 2) we have f (x, y) = f1 (x)f2 (y), p(x, y) =

p1 (x)p2 (y). Hence with
Ij = [xj −1 , xj ],Ij∗ = [0, xj0 ], j0 = min{j + 1, N}, 0 s 1 :

f − p p1 L2 (Ij ) f2 − p2 H s (Ij∗ ) + f1 − p1 L2 (Ij ) f2 H s (Ij∗ )
H (0,s) Ãj
Hence

N
N
2λ−3+(4−2s)γ
f − p2 c h4−2s hj xj
H (0,s) Ãj
j =2 j =2
which we already estimated in (7.48). For details see [423], p. 83.

The remaining term on Ã1 can be estimated by the above similarity arguments,
leading to
f − pH (0,s) (Ã1 ) C h3/2−ε̃
Altogether we have f − pH (0,1/2+ε) (A) c h3/2−ε̃ and collecting the estimates we
*
have with Lemma 7.6 on Q = A B: f − pH 1/2+ε (Q) c h3/2−ε̃ .
Approximation of regular edge functions: f˜(x, y) = χ (θ )b(x)y ν , b(x) ∈ H03 (I ))
with χ ≡ 1 near θ = 0. Due to b(x) ∈ H03 (I ) there holds (1 − χ(θ ))b(x)y ν ∈
H 3 (Q) and hence it can be approximated like the regular part. We set f (x, y) =
b(x)y ν =: f1 (x)f2 (y) and p(x, y) = p1 (x)p2 (y) with pw. linear interpolants pj
of fj . Hence for 0 s < 12 + ν
f − pH s (Q) p1 L2 (I ) f2 − p2 H s (I ) + p1 H s (I ) f2 − p2 L2 (I )

+ f1 − p1 L2 (I ) f2 H s (I ) + f1 − p1 H s (I ) f2 L2 (I )
yields (by application of 1D approximation results cf. Theorem 7.14, Proposi-

tion 7.3) for β > 3/(2ν)
f − pH 1/2 (Q) c h3/2 .
Collecting all the above results completes the proof of Theorem 7.18.

7.5 Augmented BEM for Screen/Crack Problems
In this section we report from [398] and treat the screen problems of Sect. 4.4 by
the augmented BEM. We solve the boundary integral equations (4.61) and (4.62) in
finite dimensional subspaces Sh1 , Sh2 of the Sobolev spaces H̃ −1/2 (S) and H̃ 1/2(S),
7.5 Augmented BEM for Screen/Crack Problems 265
respectively. For conformity we assume the families of finite element subspaces Sh1 ,
Sh2 satisfy for integers t, k
Sh1 = Sht −1,k−1 (S) ⊂ H k−1 (S) ⊂ H̃ −1/2 (S)

(7.49)
Sh2 = Sht,k (S) ⊂ H k (S) ∩ H̃ 1 (S) ⊂ H̃ 1/2(S), t >k≥1
The surface S is given by local representations such that regular partitions in

the parameter domains are mapped onto a corresponding partition of S. On the
partitions in the parameter domains we use a regular (t, k)-system , called Sht,k of
finite elements. The parameters mean: h is the mesh size of the partition, e.g. longest
side of a triangle of a triangulation; t − 1 is the degree of piecewise polynomials; k
describes the conformity Sht,k ⊂ H k (S). Now the Galerkin procedures for the screen
problems read:
For the Dirichlet problem find ψh ∈ Sht −1,k−1 (S) such that
VS ψh , φ L2 (S) = 2g, φ L2 (S) (7.50)
for all φ ∈ Sht −1,k−1 (S) with t, k as in (7.49). For the Neumann problem find vh ∈
Sht,k (S) such that
WS vh , wh L2 (S) = −2h, wh L2 (S) (7.51)
for all wh ∈ Sht,k (S) with t, k as in (7.49). The solvability of the above Galerkin
schemes and their convergence are based on the Gårding inequalities for VS and WS
and the uniqueness of the integral equations. Application of the general results on
the Galerkin procedure for strongly elliptic pseudodifferential operators yields the
following Theorem.
Theorem 7.21 There exists a h0 > 0 such that (7.50) and (7.51) are uniquely
solvable for any h, 0 < h ≤ h0 and
ψ − ψh −1/2,S ≤ c inf ψ − χ (7.52)

χ∈Sht−1,k−1 (S)
v − vh 1/2,S ≤ c inf v − w (7.53)

w∈Sht,k (S)
with c independent of h. Here ·r,S denotes the norm in H̃ r (S)

Due to Theorem 4.7 in Chap. 4 the exact solutions of VS ψ = 2g and of WS v = −2h
behave like ρ −1/2 and ρ 1/2 ,respectively, where ρ is the distance to the boundary
γ of the screen S. Since ρ 1/2 ∈ H 1− for some > 0 the above estimates give at
most convergence of order 1/2 − . As in the two-dimensional case [410, 411]

we can improve the asymptotic rate of convergence by using so-called singular
elements in the Galerkin procedures. This gives the augmented finite element spaces
1/2 3/2
Zh (S), Zh (S) on S:
◦ t −1,k−1
Zh (S) := {ψ̃ = ψ̃r + β̃ρ −1/2 χ : β̃ ∈ Sht ,l (γ ), ψ̃r ∈ S h
1/2
(S)} (7.54)
3/2 ◦ t,k
Zh (S) := {ṽ = ṽr + α̃ρ 1/2 χ : α̃ ∈ Sht ,l (γ ), ṽr ∈ S h (S)} (7.55)
◦ t −1,k−1
where α̃, β̃ ∈ Sht ,l (γ ) ⊂ H 1 (γ ); ψ̃r ∈ S h (S) ⊂ H k−1 (S) ∩ H̃ 1 (S);
◦ t,k
ṽr ∈ S h (S) ⊂ H k (S) ∩ H̃ 2 (S)
with t > l ≥ 1; t , l ∈ N and t, k as in (7.49).

The improved Galerkin schemes now read as:
For the Dirichlet screen problem find ψh = βh ρ −1/2 χ + ψhr ∈ Zh (S) such that
1/2
VS ψh , φ̃ L2 (S) = 2g, φ̃ L2 (S) (7.56)
for all φ̃ = β̃ρ −1/2 χ + φ̃ r ∈ Zh (S).

1/2
3/2
For the Neumann screen problem find vh = αh ρ 1/2 χ + vhr ∈ Zh (S) such that
WS vh , w̃ L2 (S) = −2h, w̃ L2 (S) (7.57)
3/2
for all w̃ = α̃ρ 1/2 χ + ṽ r ∈ Zh (S).
3/2 1/2
The above Galerkin equations with test functions w̃ ∈ Zh (S), φ̃ ∈ Zh (S)
define quadratic system of linear equations for the unknown coefficients of αh , βh ∈
◦ t,k ◦ t −1,k−1
Sht ,l (γ ), vhr ∈ S h (S) and ψhr ∈ S h (S). In [398] the following result is proven:
Theorem 7.22 The Galerkin equations (7.56) and (7.57) are uniquely solvable for
sufficiently small h and there holds
ψ − ψh −1/2,S ≤ c inf ψ − η−1/2,S ≤ ch1+σ ψZ 1/2+σ (S) (7.58)

1/2
η∈Zh (S)
v − vh 1/2,S ≤ c inf v − w1/2,S ≤ ch1+σ vZ 3/2+σ (S) . (7.59)

3/2
w∈Zh (S)
7.5 Augmented BEM for Screen/Crack Problems 267
with c independent of h . The arising norms are defined as follows ( ·q,γ denotes
the Sobolev norm in H q (γ )):
% √
α ρχ(ρ) + vr q,S , 1/2 ≤ q < 1 − , > 0 arbitrary
vZ q (S) :=
αq,γ + vr q,S , 1 ≤ q ≤ 3/2 + σ.
%
− α2 ρ −1/2 + ψr p,S , −1/2 ≤ p < − ,
ψZ p (S) :=
αp,γ + ψr p,S , 0 ≤ p ≤ 1/2 + σ.
Similar results can be shown for crack problems in linear elasticity, see [396].
Chapter 8
Exponential Convergence of hp-BEM
The first section of this chapter collects results from [240] which gives a further
contribution to the analysis of the hp-version of the boundary element method
(BEM) by presenting a more general result for Dirichlet and Neumann problems
than [21] allowing the use of a general geometric mesh refinement on the polygonal
boundary Γ . Here as in [240] we prove the exponential convergence of the hp-
version of the boundary element method by exploiting only features of the solutions
of the boundary integral equations. The key result in this approach is an asymptotic
expansion of the solution of the integral equations in singularity functions reflecting
the singular behaviour of the solutions near corners of Γ . With such expansions
we show that the solutions of the integral equations belong to countably normed
spaces. Therefore these solutions can be approximated exponentially fast in the
energy norm via the hp- Galerkin solutions of those integral equations. This result
is not restricted to integral equations which stem from boundary value problems for
the Laplacian but applies to Helmholtz problems as well. Further applications are
2D crack problems in linear elasticity. For numerical experiments with hp-version
(BEM) see [165, 340].
In Sect. 8.2 we consider the hp-version of the boundary element method (BEM)
for Dirichlet and Neumann screen problems of the Laplacian in R3 \Γ , where Γ is a
planar surface piece with polygonal boundary (for details see also the survey paper
[400]).
For the pure Dirichlet and the pure Neumann problems of the Laplacian the
exponential convergence of the corresponding hp-version of the Galerkin scheme
was already shown in Babuška, Guo, Stephan [21]. Here in Sect. 8.3 we extend their
analysis to the mixed bvp of the Laplace equation. A short version of this section
is the conference paper [211]. By a further approach J. Elschner [154] has shown
exponential convergence for the Galerkin solution for Mellin convolution equations
(of second kind) on the interval (0, 1).

270 8 Exponential Convergence of hp-BEM
8.1 The hp-Version of BEM on Polygons
We consider boundary integral equation methods for solving Dirichlet and Neumann
boundary value problems for the Laplacian in a polygonal domain with boundary
Γ . Let us assume that Γ has conformal radius less than one; this can always be
achieved by an appropriate scaling. Then the problems under consideration are the
following ones:
Dirichlet Problem For given f ∈ H 1/2(Γ ) find u ∈ H 1 (Ω) such that
Δu = 0 in Ω, u = f on Γ (8.1)
Neumann Problem For given g ∈ H −1/2 (Γ ) with Γ gds = 0 find u ∈ H 1 (Ω)

such that
∂u
Δu = 0 in Ω, = g on Γ (8.2)
∂n
Here ∂u∂n denotes the normal derivative of u with respect to the outer normal n. It
is well-known [114] that problems (8.1) and (8.2) can be converted into boundary
integral equations of the first kind on Γ . With v = uΓ , ψ = ∂u
∂n Γ we have for
(8.1) and (8.2), respectively,
V ψ = (1 + K)f on Γ (8.3)
W v = (1 − K )g on Γ (8.4)
with the integral operators (for w ∈ H −1/2(Γ ), z ∈ H 1/2(Γ ))

1 1 ∂
V w(x) := − ln |x−y|w(y)dsy , Kw(x) := − (ln |x−y|)w(y)dsy
π Γ π Γ ∂ny

1 ∂ 1 ∂ ∂
K w(x) := − (ln |x − y|)w(y)dsy , Wz(x) := (ln |x − y|)z(y)dsy .
π Γ ∂nx π ∂nx Γ ∂ny
It is also well-known that there exist unique solutions ψ ∈ H −1/2 (Γ ) of (8.3) and
1/2 3 4
v ∈ H0 (Γ ) = w ∈ H 1/2(Γ ) : Γ wds = 0 of (8.4). The boundary integral
operators V and W are strongly elliptic pseudodifferential operators satisfying a
Gårding inequality on H −1/2 (Γ ) and H0 (Γ ), respectively. Therefore due to [408]
1/2
any conforming Galerkin scheme for (8.3) and (8.4) converges quasioptimally in the
energy norm. Let XN , YN denote subspaces of dimension N of X := H −1/2(Γ )
1/2
and Y := H0 (Γ ) then the Galerkin schemes read:
8.1 The hp-Version of BEM on Polygons 271
Find ψN ∈ XN satisfying
V ψN , φ = (1 + K)f, φ ∀ φ ∈ XN , (8.5)
find vN ∈ YN satisfying
W vN , w = (I − K )g, w ∀ w ∈ YN . (8.6)
Then for the Galerkin solutions ψN , vN and the true solutions ψ and v there holds
[408], by Theorem 6.1
ψ − ψN H −1/2 (Γ ) ≤ c1 ψ − φH −1/2 (Γ ) ∀ φ ∈ XN (8.7)
and
v − vN H 1/2 (Γ ) ≤ c2 v − wH 1/2 (Γ ) ∀w ∈ YN (8.8)
where the constants c1 , c2 > 0 are independent of N.

It is shown in [405], Chap. 7 that the h-version and the p-version of (8.5) and
(8.6) on a quasiuniform mesh have only algebraic rate of convergence whereas it is
shown in [21] that the h-p version on a geometric mesh converges exponentially fast.
However in [21] the boundary element mesh is the trace on Γ of a geometric mesh
in Ω and the boundary elements on Γ must be the traces or normal derivatives on Γ
of finite element functions in Ω. This means a restriction on the choice of boundary
elements and on the construction of the geometric mesh refinement on the boundary
Γ . Here we give a new proof of the exponential convergence of the h- p version of
the boundary element method which does not require these restrictions. The analysis
given here can be extended, e.g. to curved polygons Γ and to the Helmholtz operator
in (8.1), (8.2) (instead of the Laplacian) as shown in [240].
To describe the hp- version we introduce the geometric mesh Γσn on Γ =
*J j j j
j =1 Γ , Γ being open arcs, with endpoints tj −1 , tj . First, we bisect each side Γ
j j
with length dj into two pieces Γ1 (containing the vertex tj −1 ) and Γ2 (containing
j,k
the vertex tj ). Then for the distance of the subarc Γ1 to the vertex tj −1 there holds
j,k dj
dist tj −1 , Γ1 = 2 σ n−k+1 for k ≤ n + 1 where σ ∈ (0, 1) and n is an integer.
On this geometric mesh Γσn the boundary element space S P , (Γσn )( = 1 or 2) is
given by
3
S P , (Γσn ) := ψ ∈ H −1 (Γ j ), 1 ≤ j ≤ J, ψ ∈ C 0 (Γ )
j,m 4
if = 2; |ψ|Γ j,m ∈ Pp (Γk ), k = 1, 2, m = 1, . . . , n + 1 (8.9)
k
j,m
where Pp (Γk ) denotes the space of polynomials of degree ≤ p on the subarc
j,m
Γk .
With the choice XN := S P ,1 (Γσn ) in the Galerkin scheme (8.5) we have the
following results on exponential convergence.
Theorem 8.1 Provided the given data f in (8.1) is piecewise analytic, then there
holds the estimate
√
ψ − ψN H −1/2 (Γ ) ≤ Ce−b N
(8.10)
for the error between the Galerkin ψN ∈ S P ,1 (Γσn ) of (8.5) and the solution ψ
of (8.3) if the degrees P are suitably chosen. Here the positive constants C and b
depend on the mesh parameter σ but not on the dimension N of S P ,1 (Γσn ).
With the choice YN := S P ,2 (Γσn ) in the Galerkin scheme (8.6) we have
Theorem 8.2 Provided the given data g in (8.2) is piecewise analytic, then there
holds the estimate
√
v − vN H 1/2 (Γ ) ≤ Ce−b N
(8.11)
for the error between the Galerkin solution vN ∈ S P ,2 (Γσn ) of (8.6) and the solution
v of (8.4) if the degrees P are suitably chosen here; the positive constants C, b
depend on σ but are independent of N = dimS P ,2 (Γσn ).
Remark 8.1 The functions in XN need not to be continuous on Γ since XN ⊂
H −1/2(Γ ) whereas the constraint YN ⊂ H 1/2(Γ ) requires continuity for the
functions in YN .
The proofs of Theorem 8.1 and 8.2 are based on regularity results for the solutions
of the integral equations and on approximation results for splines on the geometric
mesh.
From [128] we know that for f ∈ H t (Γ j ), 1 ≤ j ≤ J the solution ψ of (8.3)
has the form (see Chap. 9)

J
n
π ωj
ck x αkj −1 χj (x) + ψ0 (x),
j j
ψ(x) = ck ∈ R, αkj = k , n≤ (t − 3/2)
ωj π
j =1 k=1
(8.12)

where ψ0 Γ j ∈ H t −1(Γ j ) whereas the solution v of (8.4) for g Γ j ∈
H τ −1 (Γ j ), 1 ≤ j ≤ J , has the form

J
n
j αkj
v(x) = dk x χj (x) + v0 (x) (8.13)
j =1 k=1

with v0 Γ j ∈ H τ (Γ j ), dk ∈ IR. Here χj is a C ∞ – cut off-function concentrated at
j
the j th corner, with opening angle ωj . When ωπj is an integer then the singularity
functions in (8.12), (8.13) have the forms x αkj −1 ln |x| and x αkj ln |x|, respectively.
Note if the boundary Γ is curvilinear there appear also terms of the form
k ωπ +m
x j , m integer, in the above expansions (see [126]).
For the local singularity terms we have the following result using the countably
normed spaces Bβ .
Lemma 8.1 Let R > 0 and ϕμ (x) := x μ , ϕμ,k (x) := x μ logk x for x ∈ (0, R), k
pos. integer. Then
(i) ϕμ ∈ Bβ (0, R) for μ > − 1/2 − s,
(ii) ϕμ,k ∈ Bβ (0, R) for μ > − 1/2 − β, where u ∈ Bβ (0, R) if and only
if u ∈ Hβm, (0, R) ∀m ≥ and ∃C > 0, d > 1 such that
x β+k− u(k) L2 (0,R) < Cd k− (k − )!, k = , + 1, . . .
and u ∈ Hβm, (0, R) if and only if
u ∈ H −1 (0, R) and|u|m,

Hβ(0,R) := x u L2 (0,R) < ∞, ≤ k ≤ m.
β+k− (k)
The proof of the Lemma 8.1 follows immediately by inspection.

Lemma 8.2 Let |cn | < C < ∞ then
∞
π
ϕ(x) := cn x n ω ∈ Bβ (0, R) for β > − π/ω − 1/2
n=1
Proof We have to show ϕ (k) x β+k− L2 (0,R) ≤ Cd k− (k − )! (k ≥ ).
With ϕN := N n=1 cn x , αn = n ω and (α)k := α(α − 1) · · · (α − k + 1) we
αn π
have
R
(k) (k)
x β+k− ϕN 2L2 (0,R) = |ϕN (x)|2 x 2(β+k−)dx
0

N
R 2(αn −+β)+1
≤C cn2 (αn )2k
2(αn + β − ) + 1
n=1
⎡
⎢ R 2αn
≤ CR 2(β−)+1 ⎢
⎣ Γ (αn + 1)2 Γ (k − αn )2
2(αn + β − ) + 1
n≥1
αn <k
⎤
R 2αn ⎥
+ αn2k ⎥
2(αn + β − ) + 1 ⎦
n≤N
αn ≥k
⎡ ⎤
⎢ 2 R 2αn C ⎥
≤ CR 2(β−)+1 ⎢
⎣ k! + ⎥ ≤ Ck!2
⎦
2(αn + β − ) + 1 1 − α∞ R
2 2π/ω
n≥1
αn <k
Here we have used

(i) |(αn )k | ≤ (αn )k for αn ≥ k, |(αn )k | ≤ Γ (αn + 1)Γ (k − αn ) for αn < k,
(ii) Γ (αn + 1)Γ (k − αn ) ≤ ck! for αn < k,
αn
R 2αn
(iii) n∈N αn 2(α +β−)+1 ≤
2k
n
C
1−α 2 R 2π/ω
with α∞ := limn→∞ ααn+1 n
= eπ/ω .
αn ≥k ∞
2αn
For the proof of (iii) we observe that with an := αn2k 2(αn +β−)+1
R
there holds
an+1
limn→∞ sup an ≤ α∞ 2 R 2π/ω < 1 for suitably chosen R.
Hence with some C > 0
∞
R 2αn C
αn2k ≤C ) =
2 2π/ω n
(α∞ R .

2(αn + β − ) + 1 1 − α∞ R 2π/ω
2
u∈N n=0
αn ≥k
Remark 8.2 For a more general version of Lemma 8.2 see [240]. Inspection of the
proof of Lemma 8.2 shows ∞ n=1 cn x
nπ/ω ln |x| ∈ B (0, R) for β > −π/ω−1/2.
β
j j
From [128] we know that the coefficients ck and dk in (8.12), (8.13) are continuous
functionals on the given data f and g. Therefore these coefficients are bounded
satisfying the assumption of Lemma 8.2. Hence if f ∈ H t (Γ j ), 1 ≤ j ≤ J , for any
t then the solution ψ of (8.3) has the form
∞
J

ck x αkj −1 ln |x|χj (x) + ψ0 , ψ0 ∈ HΓt −1
j
ψ= j
j =1 k=1 Γj
with the notation in (8.12) and therefore with Lemma 8.2 we have ψ ∈ Bβ1 (Γ ) for
β > 12 − πω . Analogously, if g ∈ H τ −1 (Γ j ), 1 ≤ j ≤ J , for any τ then the solution
v of (8.4) has the form
∞

J

dk x αkj χj (x) + v0 , v0
j
v= ∈ H τ (Γ j )
j =1 k=1 Γj
with the notation in (8.13) and therefore with Lemma 8.2 we have v ∈ Bβ2 (Γ ) with
β > 32 − πω . Next we need some properties of Legendre polynomials.
Lemma 8.3
(i) Let I = (−1, 1), u(x) = ∞ j =0 cj j (x), j Legendre polynomial of degree j .
Then
2 (j + k)!
|u(k)(x)|2 (1 − x 2 )k dx = cj2
I 2j + 1 (j − k)!
j ≥k
(ii) Let I = (−1, 1) and u ∈ H k+1 (I )(k ∈ N0 ). Then there exists a ϕ ∈ Pk (I )

and a constant c > 0 such that
0 02 (k − s)! 0 0
0 0 0 (s+1)02
0(u − ϕ)(m) 0 2 ≤ C 0u 0 2
L (I ) (k + s + 2 − 2m)! L (I )
(m = 0, 1, 0 ≤ s ≤ k, s ∈ N0 , k > 0 or m = s = k = 0) where ϕ(−1) =

u(−1), ϕ(1) = u(1) for k > 0.
(iii) Let J = (a, b), h = b − a and u ∈ H k+1 (J )(k ∈ N0 ). Then there exist a
ϕ ∈ Pk (J ) and a constant C > 0 and that
0 02 2(s+1) 0 0
0 (m) 0 −2m h (k − s)! 0 (s+1)02
0(u − ϕ) 0 2 ≤ Ch 0u 0 2
L (J ) 2 (k + s + 2 − 2m)! L (J )
(m = 0, 1, 0 ≤ s ≤ k, k > 0 or m = s = k = 0) and ϕ(a) = u(a), ϕ(b) =

u(b) (k > 0).
(iv) Let I = (0, R) for R > 0, J = (a, b), J ⊂ I and λ > 0 be a fixed number with
h = b − a ≤ λa. Then for u ∈ Hβk+1,(I ) there exists a polynomial ϕ ∈ Pk (J )
and a constant c > 0 such that for n = 0(k = 0) and n = 0, 1(k > 0),
respectively, there holds
0 02 2s
0 (n) 0 Γ (k − s + 1) λ
0(u − ϕ) 0 2 ≤ Ca 2(−n−β)
|u|2 s+1,
L (J ) Γ (k + s + 3 − 2n) 2 Hβ (I )
(8.14)
(n < s + 1, 1 ≤ ≤ s + 1 ≤ k + 1, s ∈ IR) with ϕ(a) = u(a), ϕ(b) =

u(b)(k > 0).
Proof Assertion (i) is wellknown (see e.g. [17]). (ii) follows from (i) by expanding u
and u in Legendre series (see [210]). (iii) follows from (ii) via affine transformation
(see [210]). Assertion (iv) can be seen as follows.
0 02
By definition |u|2 s+1, ≥ a 2(β+s+1−) 0u(s+1) 0L2 (J ).
Hβ (I )
By (iii) there exists ϕ ∈ Pk (J ) with
0 02 2(s+1)
0 0 (k − s)! h
0(u − ϕ)(n) 0 2 ≤ Ch−2n a −2(β+s+1−) |u|2
L (J ) (k + s + 2 − 2n)! 2 Hβs+1, (I )
yielding (8.14).

Next we consider a geometric mesh on I = (0, 1) with n subintervals Ij =
Iσn
[xj −1 , xj ],
x0 = 0, xj = σ n−j , hj = xj − xj −1 , 1 ≤ j ≤ n.
For a degree vector p = (p1 , . . . , pn ) of nonnegative integers we set
& '
S p, (Iσn ) = q ∈ H (I ) : q|Ij ∈ Ppj (Ij ) (8.15)
Lemma 8.4 Let I = (0, 1), u ∈ Bβ (I ), = 1, 2, then there exists a ϕ ∈

S p,−1 (Iσn ) with 0 < σ < 1 p1 = − 1, pi = max{, [μi]}(i = 2, . . . , n) such that
√
u − ϕH −1 (I ) ≤ Ce−b N
(8.16)
where the positive constants C and b depend on σ but are independent of N =

dimS p,−1 (Iσn ).
Proof First we use Lemma 8.3 (iv) on each subinterval Ii (i > 1): Thus we have a
ϕi ∈ Ppi (Ii ) with
0 02 2si
0 0 2(−n−β) Γ (pi − si + 1) λ
0(u − ϕi )(n) 0 2 ≤ Cxi−1 |u|2 s +1,
L (Ii ) Γ (pi + si + 3 − 2n) 2 Hβi (I )
(n < si + 1, 1 ≤ ≤ si + 1 ≤ pi + 1, si ∈ IR) ∈ HΓt −1

j since u ∈ Bβ (I ) implies

u ∈ Hβsi +1 (I ) (si + 1 ≥ ).

On the first interval I1 (i = 0) we have (see [210])
2(1−β)
u − ϕ1 2H −1 (I ) ≤ Ch1 |u|2 .
1 Hβ, (I1 )
Thus there exists ϕ ∈ S p,−1 (Iσn ) with

1 2

n
2(1−β) Γ (pi − si + 1) λ 2si
u − ϕ2H −1 (I ) ≤C σ 2(1−β)n
+ xi−1 |u|2 si +1,
Γ (pi + si + 5 − 2) 2 Hβ (I )
i=2
With the estimate
|u|s+1,
Hβ(I ) ≤ C()d Γ (s + 1) (s ∈ IR+ )
s
and xi − xi−1 ≤ λxi−1 = 1−σ n−i+1

σ σ (2 ≤ i ≤ n) we obtain
-
u − ϕ2H −1 (I ) ≤ C σ 2(1−β)n

n 2si .
Γ (pi − si + 1) 2 ρd
+ σ 2(n−i+1)(1−β)
Γ (si + 1)
Γ (pi + si + 5 − 2) 2
i=2
1 2

n
≤ C σ 2(1−β)n + σ 2(n−i+1)(1−β)pi (F (ρd, αi ))pi
i=2
where
2α "
αd (1 − α)1−α 1 2
F (d, α) := and α i := max , α min , αmin := √ .
2 (1 + α) 1+α pi 4 + λ2 d 2
There holds
infα∈(0,1) F (d, α) = Fmin = F d, √ 2 < 1 with Fmin := F (αmin ).
4+d 2
Taking pi = max{, [μi]} ([x] means the smallest integer greater or equal to x)
(i = 2, . . . , n) with
"
2(1 − β) log σ
μ > max 1, (8.17)
log Fmin
- .
and defining i0 by pi0 = 1
αmin + 1, then pi0 = [μi0 ] ≤ 1
αmin + 2 and thus i0 is
bounded.
Hence
-
i0
u − ϕ2H −1 (I ) ≤ C σ 2(1−β)n + σ 2(n−i+1)(1−β)pi F (ρd, αi )pi
i=2

n .
+ σ 2(n−i+1)(1−β)pi (Fmin )pi
i=i0 +1
⎡ ( )pi

i0 F (ρd, p1i )
2(1−β)n ⎣ 2(1−i)(1−β)
≤ Cσ 1+ σ pi
(Fmin ) pi max
1≤i≤i0 Fmin
i=2
⎤

n
+ σ 2(1−i)(1−β)pi (Fmin )pi ⎦ .
i=i0 +1
μ
With pi = [μi] and q := σ 2(1−β)
F
< 1 due to (8.17) we have i>i0 iq < ∞ since
i
(iq i )1/ i → q < 1 as i → ∞.

Hence the term in the bracket is bounded yielding with a positive constant c
u − ϕ2H −1 (I ) ≤ c e2(1−β)n. (8.18)
Next we observe for = 1 : N = dim S P ,0 (Kσn ) = 1 + ni=2 (pi + 1) = 1 +

n n
i=2 ([μi] + 1) ≤ cμn and for = 2 : N = dim S
P ,1 (I n ) = 2 +
i=2 (pi +
2
σ
1) − n + 1 ≤ cμn . 2
Hence we obtain from (8.18) ( = 1, 2)

√
u − ϕH −1 (I ) ≤ Ce−b N
with
1−β 1
b = √ log . (8.19)
μ σ

Corollary 8.1 Let I = (0, 1), u ∈ Bβ2 (I )
for some β < 1, then there exists a
ϕ ∈ S (Iσ ) with 0 < σ < 1, p1 = 1, pi = [μi], 2 ≤ i ≤ n, such that
P ,1 n
√
u − ϕH 1/2 (I ) ≤ ce−b N
with constants c, b > 0 independent of N = dim S p,1 (Iσn ).

Proof The assertion follows by interpolation directly from Lemma 8.4.

The corollary can be generalised from the interval I to the polygon Γ in a
straightforward manner.
Now the proofs of Theorem 8.1 and 8.2 are completed as follows:
Proof (of Theorem 8.2) First we observe with Lemma 8.2 that the analyticity of g
on Γ j implies v ∈ Bβ2 (Γ ) for 1 > β > 3/2 − π/ω where v satisfies (8.4). Hence
j j j
by Lemma 8.4 there exists for each boundary piece Γk a ϕk ∈ S Pj,k ,1 (Γk ) with
j,m
degree Pj,k,m on Γk such that ( = 1 or 2)
0 0 √
0 j0
0v − ϕk 0 j ≤ Ce−bj,k Nj,k
,
H −1 (Γk )
Nj,k = dimS Pj,k ,1 (Γjk ), k = 1, 2, j = 1, . . . , J
j j
where ϕk coincides with v at the endpoints of Γk .
Let
% %
j j j
j ϕk on Γk j v on Γk
=
ϕk = and vk = .
0 elsewhere 0 elsewhere
Then for = 1 and 2

J 2 0
J 0
0 2
=j 0 0 j j0
0v − ϕk 0H −1 (Γ ) ≤ 0vk − =
ϕk 0
H −1 (Γ )
j =1 k=1 j =1 k=1
2 0
J 0
0 j0
√
= 0v Γ j − ϕk 0 j
≤ Ce−b N
(8.20)
k H l−1 (Γk )
j =1 k=1
3 4 3 4
with b = min 1≤j≤J bj,k , N = min 1≤j≤J Nj,k . Note the estimate (8.20) holds
1≤k≤2 1≤k≤2
j j j j j
since vk − =
ϕk ∈ C 0 (Γ ) and vk − = ϕk ≡ 0 on Γ \ Γk . Hence the assertion of
Theorem 8.2 follows from (8.20) by interpolation.

Proof (of Theorem 8.1) First we observe with Lemma 8.2 that the analyticity of f
on Γ j implies ψ ∈ Bβ1 (Γ ) for 1 > β > 3/2 − π/ω where ψ satisfies (8.3). Hence
j j j
by Lemma 8.4 there exists for each boundary piece Γk a ϕk ∈ S Pj,k ,0 (Γk ) with
j,m
degree Pj,k,m − 1 on Γk such that
0 0 √
0 j0
0ψ − ϕk 0 j ≤ Ce−bj,k Nj,k
, Nj,k = dimS Pj,k ,0 (Γjk )
L2 (Γk )
Hence the assertion of Theorem 8.1 follows.

8.1.1 Application to Acoustic Scattering
We consider for μ, k1 , k2 ∈ C\{0} and μ = −1 the transmission problem
(Δ + k12 )u1 = 0 in Ω1 , (Δ + k12 )u2 = 0 in Ω2 := IR2 \Ω 1
∂u1 ∂u2
u1 = u2 + v0 , μ = + ψ0 on Γ (8.21)
∂u ∂u
subject to the Sommerfeld radiation condition
∂u2
− ik2 u2 = o(R −1/2 ), u2 = O(R −1/2 ) as |x| = R → ∞.
∂R
In the case of scattering problems, u1 (u2 ) denote the refracted (scattered) field and
v0 and ψ0 are the boundary trace and the normal derivative of the incident field u0 .
In [129] the above transmission
problem is reduced on Γ = ∂Ω1 for the Cauchy
1
data v1 = u1|Γ , ψ1 = ∂u ∂u Γ :
( ) ( )
v1 1 −(K1 + K2 ) V1 + μV1 v1 v0
H := = (8.22)
ψ1 2 W1 + μ1 W2 K1 + K2 ψ1 1
μ ψ0
where (j = 1 or 2)

∂
Vj ϕ(z) = −2 γj (z1 ζ )ϕ(ζ )dsζ , Kj ϕ(z) = −2 ϕ(ζ ) γj (z, ζ )dsζ , z ∈ Ωj
Γ Γ ∂nζ
∂
Wj uj = − Kj uj Γ and Kj is the adjoint operator of Kj
∂n
and
i (1) 1
γj (z, ζ ) = − H0 (kj |z − ζ |) = ln |z − ζ | + O(|z − ζ |−1 ) (8.23)
4 2π
is the fundamental solution of the Helmholtz equation Δw = −kj2 w in Ωj where

H0(1) is the Hankel function of first order and degree zero.
It is shown in [129] that the operator H from H 1/2(Γ ) × H −1/2(Γ ) →
H (Γ ) × H −1/2(Γ ) is bijective if and only if the homogeneous transmission
1/2
problem (8.21) as well as the adjoint problem – obtained by interchanging Ω1 and

Ω2 – have only the trivial solution. This is assumed in the following. From the
regularity results in [129] follows that for piecewise analytic data v0 , ψ0 the solution
(v1 , ψ1 ) of (8.22) has expansions of the form (8.12), (8.13) with αkj = kαj and αj
being a zero of the transcendental equation

sin(π − ωj ) μ+1
=± . (8.24)
sin πωj μ−1
The boundary element Galerkin scheme for (8.22) reads. Find (vN , ψN ) ∈ YM ×
XN such that
5 6 5 6
vN w v0 w
H , = , ∀(w, φ) ∈ YM × XN (8.25)
ψN φ Γ μψ0 φ Γ
where YM , XN are finite dimensional subspaces of H 1/2 (Γ ) and H −1/2(Γ ) with

dim XN = N and dim YM = M.
Since the operator H satisfies a Gårding’s inequality in H 1/2(Γ )×H −1/2 (Γ ) this
boundary element Galerkin scheme converges quasioptimally in the energy norm,
i.e.
vN − v1 H 1/2 (Γ ) + ψM − ψ1 H −1/2 (Γ )

"
≤ C inf v1 − wH 1/2 (Γ ) + inf ψ1 − φH −1/2 (Γ ) (8.26)
w∈XN φ∈YM
Next we choose XN = S p−1,1 (Γσn ) and YM = S P ,2 (Γσn ) as in Section 2 and obtain

the exponential convergence of the hp- version of the Galerkin scheme (8.25) for
the transmission problem (8.21) (see [210, 241]).
Proposition 8.1 Let v0 and ψ0 in (8.21) be piecewise analytic, then for the error
between the Galerkin solution vN ∈ S P ,2 (Γσn ).ψN ∈ S P −1,1 (Γσn ) and the exact
solution of (8.22) there holds
√
v1 − vN H 1/2 (Γ ) + ψ1 − ψM H −1/2 (Γ ) ≤ Ce−b N
if the degrees P are suitably chosen. Here N is the number of degrees of freedom of
S P ,2 (Γσn ), C and b are constants depending on σ but not on N.
Proof Firstly, we observe that for piecewise analytic data v0 , ψ0 the solution
(v1 , ψ1 ) of (8.22) belong to Bβ2 (Γ )×Bβ1 (Γ ) with 1 > β > 1/2 −αmin where αmin
is the smallest zero of (8.24). Therefore application of the above analysis yields the
assertion of the proposition.

Remark 8.3 For the transmission problem (8.21) with k1 = k2 = 0 the exponential
convergence of the hp- version of the bem is shown in [210].
Two-dimensional crack problems in linear elasticity can be converted into first kind
integral equations (see [256, 432]). for example, let us consider the Neumann crack
problem for the domain ΩΓ exterior to an arc Γ : find u ∈ Hoc 1 (Ω ) such that
p
Δ∗ u ≡ μΔu + (λ + μ)grad div u = 0 in ΩΓ = IR2 \ Γ

T (u) = ψ1 , T (u) = ψ2
Γ1 Γ2
−1
for given ψi ∈ H 2 (Γ ), i = 1, 2, where T denotes the traction operator on the
sides Γ1 and Γ2 of Γ and λ,μ are the given Lamé constants. Under appropriate
conditions, e.g. assuming a decaying condition for u at infinity, this problem can be
converted into the integral equation

W φ(x) = −Tx (Ty (E(x, y)))T φ(y)dsy = f (x) (8.27)
Γ
for the jump φ ≡ [u] = u|Γ1 − u|Γ2 with the fundamental solution of the Navier
operator Δ∗
"
λ + 3μ 1 = λ + μ (x − y)(x − y)T
E(x, y) = ln I+
4πμ(λ + zμ) |x − y| λ + 3μ |x − y|2
Here T denotes the transposed tensor and = I is the identity matrix and f is given
via ψ1 and ψ2 . It is shown in [432] that the solution φ of the hypersingular integral
1
equation (8.27) behaves like x 2 (d1 + d2 x + d3 x 2 + · · · ), dj ∈ IR, near the crack tip,
i.e. like v in (8.13) with αk = 12 + k, k, k integer > 0. Therefore φ ∈ Bβ2 (Γ ) for β >
1 since in the case of a crack ω = 2π. The operator W in (8.27) satisfies a Gårding’s
inequality in H = 12 (Γ ) (see [432]) and therefore the corresponding Galerkin scheme
converges quasioptimally in H = 12 (Γ ). Therefore the above analysis applies also to
the integral equation (8.27) yielding exponentially fast convergence for the Galerkin
solution of the hp- version for (8.27).
8.2 The hp-Version of BEM on Surfaces
In this section we report from [235] on the hp-version of the Galerkin boundary
element method for Dirichlet and Neumann screen problems in R3 when the screen
Γ is a smooth open surface piece with piecewise smooth boundary.
That is, given f or g on Γ find u ∈ R3 \Γ satisfying
Δu = 0 in R3 \Γ¯
∂u
u = f ∈ H 1/2(Γ ) (Dirichlet) or = g ∈ H −1/2(Γ ) (Neumann)
∂n
and
u = O(|x|−1 ) as |x| → ∞.
These exterior boundary value problems are called screen problems and can be
formulated equivalently as first kind integral equations with weakly singular and
hypersingular kernels, namely (see Sect. 4.3)

1 1
V ψ(x) := ψ(y) dsy = 2f (x), x ∈ Γ (Dirichlet) (8.28)
2π Γ |x − y|

1 ∂ ∂ 1
W v(x) := − v(y) dsy = 2g(x), x ∈ Γ (Neumann).
2π ∂nx Γ ∂ny |x − y|
(8.29)
As we have shown in [398] (see also Section 5.3) these integral equations have
unique solutions ψ ∈ H̃ −1/2 (Γ ), v ∈ H̃ 1/2(Γ ) = H00 (Γ ).
1/2
The Galerkin boundary element schemes for (8.28) and (8.29) read with the L2 -
duality on Γ ·, · :
Find ψN ∈ Sh,p0
V ψN , φN = 2f, φN 0
∀φN ∈ Sh,p ⊂ H̃ −1/2 (Γ ) (8.30)
and find vN ∈ Sh,p

1
W vN , wN = 2g, wN ∀wN ∈ Sh,p

1
⊂ H̃ 1/2(Γ ) (8.31)
8.2 The hp-Version of BEM on Surfaces 283
Since the operators V and W define coercive, continuous bilinear forms we

immediately have quasi-optimality of the Galerkin errors:
0

ψ − ψN H̃ −1/2 (Γ ) dist ψ, Sh,p (Γ )
and

v − vN H̃ 1/2 (Γ ) dist v, Sh,p
1
(Γ ) .
In [235] we prove the Theorem 8.3 using the setting of countably normed spaces
together with a detailed investigation of the special singular behaviour of the
solutions of the screen problems for the Laplacian at corners and edges, see
Examples 7.3 and 7.4. When these problems are converted via the direct method
into boundary integral equations then the solutions of the latter possess these corner
and corner-edge singularities. For the screen problems above these estimates yield
only very low convergence rate of order O(h1/2−ε p−1+2ε ) with arbitrary ε > 0 (see
[51, 374, 426] and Chap. 7).
0 1 (Γ ) refer
The indices h and p in the notation for the trial spaces Sh,p (Γ ) and Sh,p
to h- and p-versions, respectively; where in the h-version a more accurate Galerkin
solution is obtained by mesh refinement (and the polynomial degree p is kept fixed)
whereas in the p-version a higher accuracy is obtained by increasing the polynomial
degree on the same mesh. The implementation of the h-version is standard. In the
p-version BEM for the weakly singular integral equation we use tensor products
of Legendre polynomials on rectangular meshes and for the hypersingular integral
equation we take instead antiderivatives of Legendre polynomials.
If one uses a geometric mesh refinement together with a properly chosen
polynomial degree distribution one obtains even exponentially fast convergence
rates for the Galerkin errors of the above integral equations. Numerical experiments
are presented at the end of this section which show exponential convergence. For
application of our error analysis to Helmholtz screen problems see [250]. It is
only for ease of presentation that we consider screen problems. In case of a closed
surface Γ = ∂Ω a similar analysis can be performed leading also to exponential
convergence; the interested reader might look at [251, 290, 295].
For the finite element method the exponential convergence of the hp-version
was proposed for three dimensional problems in [207] making use of the setting of
countably normed spaces. Whereas the analysis [207] requires the use of special
meshes our analysis allows to use much simpler meshes due to the tensor product
structure of our approximated subspaces of the boundary element hp-version. Our
approach uses regularity results of the solutions of the underliying integral equations
which follow from [425, 426] and are based on [141]. Those regularity results
show that the solutions have decompositions (see Sects. 7.1.3, 7.2.1) into special
edge and corner-edge singularities which on the other hand belong to countably
normed spaces see [235]. For smooth given data the solutions of the screen problems
admit improved decompositions into additional edge and corner-edge singularities
plus arbitrarily smooth remainders. This is why we can show that the error of the
4 Q40.5
I0.5
0.5
4
I0.5
0.25
p1 p2 p3 p4
0
0 0.25 0.5 1 0 0.25 0.5 1
Fig. 8.1 Geometric mesh on the square plate (σ = 0.5, n = 4)
Galerkin solution in the hp-version of the boundary element method consists of

one term which decays exponentially fast with a number of unknowns N and of
a term O(N −α ) with arbitrarily large positive α. The O(N −α )−term results from
the approximation of the smooth remainder of the solution of the integral equation.
Our numerical experiments show no contribution of this O(N −α )−term and clearly
demonstrate the exponentially fast convergence of the hp−version of the BEM.
For simplicity we take Γ = [0, 2]2 × {0} and introduce the geometric mesh
Γσ (cf. Fig. 8.1) with the reference mesh Qnσ where Iσn consists of pieces
n
[xk−1 , xk ], x0 = 0, xk = σ n−k, k = 1, . . . , n. With Qnσ we associate a degree vector

0
and define Sh,p−1 1 (Γ n )) as the vector space of all piecewise (continuous)
(Γσn ) (Sh,p σ
polynomials on Γσ having degree pk in x and pl in y on [xk−1, xk ] × [xl−1 , xl ], 1 ≤
n
k, l ≤ n. We have hk = xk − xk−1 ≤ ( σ1 − 1)x ∀x ∈ [xk−1 , xk ]. The detailed proof

of the following theorem is given in [235].
Theorem 8.3 For given piecewise analytic functions f , g in (8.28) and (8.29) and
corresponding Galerkin solutions ψN ∈ Sh,p−1
0 (Γσn ), vN ∈ Sh,p
1 (Γ n ) of (8.30) and
σ
(8.31) on the geometric mesh Γσn there holds
,
ψ − ψN H̃ −1/2 (Γ ) √
≤ C exp(−b N) + O(N −α )
4
v − vN H̃ 1/2 (Γ )
with constants C, b > 0 independent of the dimension N of the trial space and
arbitrary α > 0.
In order to give a flavour of the proof let us assume Bβ2 (Q) regularity; this is
satisfied for the higher order terms in the expansions (see Examples 7.3 and 7.4),
whereas the lower order terms must be treated separately (see [235] for details). As
shown in [251] the solution of the Neumann problem (up to an additional term) has
also this regularity.
The local mesh at a right angle corner of Γ is given in Fig. 8.1. The proof of the
theorem is based on analysing the error in countably normed spaces and is based on
the following lemma showing exponential convergence.
Lemma 8.5 For u ∈ Bβ2 (Q), 0 < β < 1, there exists a spline uN ∈ Sh,p
1 (Qn ) and
σ
constants C, b > 0 independent of N, but dependent on σ, μ, β such that
√
4
u − uN H 1 (Q) ≤ C e−b N
(8.32)
with p1 = 1, pk = max(2, [μ(k − 1)] + 1) (k > 1) for μ > 0.

In the above lemma we need the countably normed function space Bβ2 (Q) which we
introduce now for the square Q = [0, 1]2 with the help of weighted Sobolev spaces
Hβk,2 (Q) as
&
Bβ2 (Q) = u : u ∈ Hβk,2 (Q), ∀k ≥ 2, Φβ,α,2 D α uL2 (Q) ≤ C d k−2 (k − 2)!
'
for |α| = k = 2, 3, . . . , with C ≥ 1, d ≥ 1 indpt. of k .
⎧ β+α1 −2
⎪ x , α1 ≥ 2, α2 =0
⎪
⎪
⎪
⎪ xβ + yβ , α1 = 1, α2 =1
⎪
⎨ β+α1 −2
x y + x β+α1 −1 + y β , α1 ≥ 2, α2 =1
Φβ,(α1 ,α2 ),2 (x, y) = −2
⎪x
⎪
β+α 1 y2α + (x β + y β )x α1 −1 y α2 −1 + x1α y β+α2 −2 , α1 ≥ 2, α2 ≥2
⎪
⎪
⎪
⎪
β + xy β+α 2 −2 + y β+α2 −1 , = 1, α2 ≥2
⎩ x β+α α1
y 2 −2 , α1 = 0, α2 ≥2
whereas the weighted Sobolev spaces Hβk,2 (Q) are given by
k

|u|2 = |∂xα1 ∂yα2 u(x, y)|2Φβ,α,2
2
(x, y) dy dx.
Hβk,2 (Q)
|α|=2 Q
u2 = u2H 1 (Q) + |u|2 .

Hβk,2 (Q) Hβk,2 (Q)
Proof
1.) In element R11 at the origin: Due to u ∈ Hβ2,2 (Q) there exists a bilinear
interpolant φ11 ∈ P11 (R11 ) with u(0, 0) = φ11 (0, 0), u(0, h1 ) = φ11 (0, h1 ),
u(h1 , 0) = φ11 (h1 , 0), u(h1 , h1 ) = φ11 (h1 , h1 ) (h1 = x1 = σ n−1 )
2(1−β)
u − φ11 2H 1 (R ≤ C h1 u2 .
11 ) Hβ2,2 (Q)
2.) On strips near edges {(x, y) | h1 ≤ x ≤ 1, 0 ≤ y ≤ h1 } ∪ {(x, y) | 0 ≤

x ≤ h1 , h1 ≤ y ≤ 1} there exist polynomials φk1 ∈ Ppk 1 (Rk1 ) and
φ1l ∈ P1pl (R1l ), coinciding with u at vertices (0 < β < 1):
2(1−β)
u − φk1 2H 1 (R ≤ C h1 |u|2
k1 ) Hβ2,2 (Q)
2(sk +1)
2(1−β) Γ (pk − sk + 1) λ
+C xk−1 |u|2 s +2,2 (k ≥ 2)
Γ (pk + sk + 1) 2 Hβk (Q)
2(1−β)
u − φ1l 2H 1 (R ≤ C h1 |u|2
1l ) Hβ2,2 (Q)
2(sl +1)
2(1−β) Γ (pl − sl + 1) λ
+C xl−1 |u|2 s +2,2 (l ≥ 2). (8.33)
Γ (pl + sl + 1) 2 Hβl (Q)
Therefore (corresponding estimates hold away from the edges) on Rkl (2 ≤

k, l ≤ n) with 1 ≤ sk ≤ pk for 0 ≤ α1 , α2 ≤ 1 there holds:
D α (u − φkl )2L2 (R ) ≤
kl
2sk
2(2−α1 −β) Γ (pk − sk + 1) λ
≤ C xk−1 |u|2 sk +3,2
Γ (pk + sk + 3 − 2|α|) 2 Hβ (Q)
2sl "
2(2−α −β) Γ (pl − sl + 1) λ
+ xl−1 2 |u|2 sl +3,2
Γ (pl + sl + 3 − 2|α|) 2 Hβ (Q)
3.) Combining 1.) and 2.) we obtain (1 ≤ sk ≤ pk )

n
u − φkl 2H 1 (R
kl )
k,l=1
2(1−β) 2(1−β)
≤ C h1 u2 + (2n − 2) C h1 |u|2
Hβ2,2 (Q) Hβ2,2 (Q)

n 2(sk +1)
2(1−β) Γ (pk − sk + 1) λ
+2nC xk−1 |u|2 s +3,2
Γ (pk + sk + 1) 2 Hβk (Q)
k=2
Now with h1 = σ n−1 and
|u| s +3,2 ≤ Cd sk +1 Γ (sk + 2) (8.34)

Hβk (Q)
we obtain (8.32). Note: u ∈ Bβ2 (Q) implies (8.34).

Figures 8.2 and 8.3 show numerical experiments (cf. [300]) obtained with the
integral equations for linear elasticity treating crack problems with the open surface
piece Γ as crack surface. The operators are here given with the Green’s function for
the Navier–Lamé equation
"
λ + 3μ 1 λ + μ (x − y)(x − y)t
G(x, y) = I+
4πμ(λ + 2μ) |x − y| λ + 3μ |x − y|3
The legends for Figs. 8.2 and 8.3 have the following meanings: conf-uni-h-4
and conf-uni-p-4 mean conforming h-version of BEM and conforming uniform
p-version of BEM on uniform rectangular meshes, respectively. conf-grad-h-4-
beta=4.0 stands for conforming h-version of the BEM on graded meshes graded
algebraically towards the edges of Γ = [−1, 1]2 with grading parameter β = 4.
geo-sigma=0.5-mu=0.5 and geo-sigma=0.17-mu=0.5 stand for two hp-versions of
the BEM with geometric mesh parameter geo-sigma and parameter mu for the
polynomial degree distribution. Figures 8.2 and 8.3 show clearly the exponentially
fast convergence of the hp-version on the geometric mesh with optimal mesh
grading parameter σ = 0.17. The parameter μ = 0.5 describes the increase of
the polynomial degree, namely (q, p), (q, p), (q, p + 1), (q, p + 1), (q, p + 2),
(q, p + 2),. . . in the x2 -direction and correspondingly in the x1 -direction, for a
geometric mesh consisting of rectangles only and refined towards the edges. Very
100
conf-uni-h-4
conf-uni-p-4
conf-grad-h-4-beta=4.0
geo-sigma=0.5-mu=0.5
10
error in energy norm
0.1
10 100 1000 10000
degrees of freedom
Fig. 8.2 Weakly singular integral equation (Lamé) [300].

0.1
conf-uni-h-4
conf-uni-p-4
conf-grad-h-4-beta=4.0
0.01
0.001
1e-04
1 10 100 1000 10000
degrees of freedom
Fig. 8.3 Hypersingular integral equation (Lamé) [300]
good results are also obtained for the h-version on an algebraically graded mesh;
this is in agreement with the theoretical results in [426]. Also Figs. 8.2 and 8.3
show that the uniform p-version converges twice as fast as the uniform h-version
[51, 374].
Recently for the hp-version of the FEM exponentially fast convergence was
shown in [368] for axis-parallel domains – based on anisotropic analytic estimates
for boundary value problems for the Laplacian in polyhedra derived in [119].
8.3 The hp-Version of BEM on a Geometrical Mesh for

Mixed BVP on a Polygonal Domain
Let Ω ⊂ R2 be a bounded polygonal domain with boundary Γ = Γ 1 ∪ Γ 2 , vertices

Aj , j = 1 . . . M, AM+1 := A1 and straight sides Γ j with endpoints Aj , Aj +1 . Let
ωj be the interior angle at the vertex Aj . We consider the mixed boundary value
problem of the Laplacian
Δu = 0 in Ω,
u = g1 on Γ1 , (8.35)
∂u
= g2 on Γ2 .
∂n
8.3 hp-Version of BEM on a Geometrical Mesh 289
If g1 ∈ H 1/2(Γ ), g2 ∈ H 1/2(Γ2 ) then (8.35) has a unique solution u ∈ H 1 (Ω).

Here the spaces H s (Γ ), H s (Γj ), H̃ s (Γj ) for j = 1, 2 are defined as follows:
⎧
⎪
⎨ {u|Γ : u ∈ H
⎪ s+1/2(R2 )} for s > 0
H s (Γ ) := L2 (Γ ) for s = 0
⎪
⎩H −s (Γ )
⎪
for s < 0
%
u|Γj : u ∈ H s (Γ ) for s 0
H s (Γj ) :=
(H̃ −s (Γ j )) for s < 0
%
u|Γ : u ∈ H s (Γ ), u|Γ /Γ j = 0 for s 0
H̃ s (Γj ) := j
H −s (Γj ) for s < 0
and similarly for Γ j instead of Γj . In order to formulate a regularity result for

piecewise analytic data g1 , g2 we introduce weighted Sobolev spaces on Ω and Γ .
Let Hβm,l (Ω), m ≥ l ≥ 0 integers, be the completion of the set of all infinitely
differentiable functions under the norm

k=m
u2 = u2H l−1 (Ω) + Φβ+k−l |D α u|2L2 (Ω) , for l ≥ 1 (8.36)
Hβm,l (Ω)
|α|=k,k=l

k=m
u2 = Φβ+k |D α u|2L2 (Ω) , (8.37)
Hβm,0 (Ω)
|α|=k,k=0
where B
Φβ+k (x) = M βi +k , x ∈ Ω and r (x) = dist (x, A ) = |x − A |,
i=1 |ri (x)| i i i
x ∈ Ω, denotes the Euclidean distance between the point x and the vertex Ai .
Let
Bβl (Ω) = {u ∈ Hβl,l (Ω), Φβ+k−l |D α u|L2 (Ω) ≤ Cd k−1 (k − l)!, (8.38)
k = l, l + 1, . . . , C ≥ 1, d ≥ 1, independent of k}
For investigation of the singularities at corners we introduce weighted Sobolev

spaces and countable normed spaces on the boundary Γ .
Let I = (a, b) and for x ∈ (a, b), r̂1 = |x − a|, r̂2 = |x − b|, Φ̂β̂+k (x) =
B2 β̂i +k
i=1 r̂i (x), β̂ = (β̂1 , β̂2 ), 0 < β̂1 , β̂2 < 1, k integer. Now we define for k ≥
l ≥ 0 and integer l ≥ 0 the spaces H k,l (I ), B l (I ) to (8.36),(8.38) with I instead of
β̂ β̂
Ω (for details see [21]).
By B l,l+1 (Γj ) we denote the space of all functions with restrictions on Γ i ⊂ Γj ,

β̂
1 ≤ i ≤ M, j = 1, 2 belonging to B l (Γ i ), 0 < β̂i < 1
2 or B l+1 (Γ i ), 1
2 < β̂i < 1.
β̂i β̂i
Exploring the analysis by Babuška, Guo in [19],[18] one obtains the following
regularity result for the mixed bvp (8.35), cf also [21].
Theorem 8.4 Let g1 ∈ C(Γ1 ) ∩ B 1,2 (Γ1 ), g2 ∈ B 0,1 (Γ2 ) with β̂ = (β̂1 , . . . , β̂M ),
β̂ β̂
β̂i = (β̂i,1 , β̂i,2 ), 0 < β̂i,j ≤ 1, 1 ≤ i ≤ M, 1 ≤ j ≤ 2. Then u ∈ Bβ2 (Ω) with
β = (β1 , . . . , βM ) satisfying (8.39), where
⎧
⎪ π
⎨= βi∗ if βi∗ > 1 −
βi 2ωi for Ai ∈ Γ 1 ∩ Γ 2 (8.39)
⎪ π π
⎩> 1 − if βi∗ 1 −
2ωi 2ωi
⎧
⎪ π
⎨= βi∗ if βi∗ > 1 −
βi ωi for Ai ∈
/ Γ1 ∩ Γ2
⎪ π π
⎩> 1 − if βi∗ 1 −
ωi ωi
with βi∗ = max(β i−1,2 , β i,1 ), β i,j := β̂ − 12 sign(β̂i,j − 1/2)

Next we give an equivalent boundary integral equation formulation of Problem
(8.35). Let G(z, ζ ) be the fundamental solution of the Laplacian
1
G(z, ζ ) = − ln |z − ζ |
2π
and define the following boundary integral operators: Let fj ∈ C0∞ (Γj ), j, k = 1, 2.
Then for z ∈ Γk

∂
Vj i fj (z) := −2 fj (ζ )G(z, ζ ) dsζ , Kj i fj (z) := −2 fj (ζ ) G(z, ζ ) dsζ
Γj Γj ∂n ζ

∂ ∂ ∂
Kj i fj (z) := −2 fj (ζ ) G(z, ζ )dsζ , Wij fj (z) := 2 fj (ζ ) G(z, ζ )dsζ .
Γj ∂n z ∂n z Γj ∂n ζ
For the distribution fj on Γj we define Vj k fj and Kj k fj by approximating fj with

smooth functions and Kj k fj by duality using the relation

Kj k fj , f˜k L2 (Γk ) = fj , Kkj f˜k L2 (Γj ) , ∀f˜k ∈ C ∞ (Γk ).
Here the subscript of Dj k etc means integration over Γj and evaluation on Γk .

Define the extension operator l : H 1/2(Γ1 ) → H 1/2(Γ )in the following way:
Assume 3that Γ is parametrized
4 by a3 piecewise linear4 function φ(x) : [−1, 1]
with Γ1 = φ(x)|x ∈ [−1, 0] , Γ2 = φ(x)|x ∈ [0, 1] and let v ∈ H 1/2(Γ1 ) be
expressed with respect to this parametrization. Then define
%
v(x) if x ∈ [−1, 0]
l v(x) :=
v(−x) if x ∈ [0, 1].
Next we introduce the extension operator l˜ : H −1/2(Γ2 ) → H − 2(Γ ) for ψ ∈

1
1
H − 2(Γ2 ) by
%
˜lψ(x) := −ψ(−x) if x ∈ [−1, 0]
ψ(x) if x ∈ [0, 1]
Then we have from [127, 128] the following result:Here we need for s ∈ R the
space
H̃ s (Γk ) = {f ∈ H s (Γ ) : suppf ⊂ Γk }
Theorem 8.5 ([127, 128].) The boundary integral equation

∗ ( )
v g1 l g1|Γ2
A =B −A ˜ (8.40)
ψ∗ g2 l g2
Γ1
with the boundary integral operators

W22 K12 −W12 1 − K22
A := , B :=
−K21 V11 1 + K11 −V21

v∗
has a unique solution in H̃ 1/2(Γ2 ) × H̃ −1/2(Γ1 ). This solution yields the
ψ∗
∂u
unknown boundary data u|Γ2 , of the unique solution u ∈ H 1 (Ω) of problem
∂n Γ1
(8.35) by

∂u
∗
u|Γ2 = v + l g1 |Γ2 , = ψ ∗ + l˜ g2 . (8.41)
∂n Γ1 Γ1
Now we define the boundary element spaces for the h-p-method. Let dj be the
j
length of the side Γ j . First bisect each side Γ j into two parts Γ− (containing
j
Aj ) and Γ+ (containing Aj +1 ). Choose a mesh parameter 0 < σ < 1, and an
dj
integer n 0. Introduce the points Aj,k
_ on Γ_ with dist(A_ , Aj ) = σ
j j,k n−k
2
for k = 1 . . . n and Aj,0
_ := Aj . This defines n subintervals Γ_
j,k with endpoints
j,k j
Aj,k−1
_ _ for k = 1 . . . n. Analogously define the points A+ on Γ+ with
and Aj,k
j,k dj j,k j,k−1
dist(A+ , Aj +1 ) = σ n−k yielding n subintervals Γ+ with endpoints A+
2
j,k
and A+ for k = 1 . . . n.
Define the spaces Sn,2 and Sn,1 on Γ2 and Γ1 as follows
& '
j,k j,k
Sn,2 := v ∈ C 0 (Γ2 ) v|Γ j,k ∈ Pk (Γ± ) for Γ± ⊂ Γ j ⊂ Γ2 , k = 1 . . . n
±
(8.42)
& '
j,k j,k
Sn,1 := ψ | ψ|Γ j,k ∈ Pk (Γ± ) for Γ± ⊂ Γ j ⊂ Γ1 , k = 1 . . . n (8.43)
±
j,k j,k
where Pk (Γ± ) denotes the space of polynomials of degree k on Γ± . Then let
Sn := Sn,2 × Sn,1 . (8.44)
There holds dimSn M(n + 2)2 . Then we obtain exponential convergence for the
Galerkin method (8.45).
Next we describe the Galerkin method for the approximation of the solution of
the integral equation (8.40). Choose a sequence of finite dimensional subspaces
Sn ⊂ H̃ 1/2(Γ2 ) × H̃ −1/2 (Γ1 ) with
∞

Sn = H̃ 1/2(Γ2 ) × H̃ −1/2 (Γ1 ).
n=1
Then find a solution (vn∗ , ψn∗ ) ∈ Sn satisfying

5 ∗ 6 I ( )
l g1|Γ2 wn
J
wn vn wn g1
∀ ∈ Sn A , = B −A ˜ , .
φn ψn∗ φn g2 l g2 φn
Γ1
(8.45)
Then the strong ellipticity of the operator A [127] gives the quasioptimality of the
Galerkin solution:
Theorem8.6 For sufficiently large n the Galerkin equations (8.45) have a unique
vn∗
solution ∈ Sn . Furthermore there holds
ψn∗
0( ) ( )0 0( ) ( )0
0 ∗ 0 0 0
0 vn v ∗ 0 0 w v ∗ 0
0 ∗ − ∗ 0 C inf 0 n − ∗ 0
0 ψn ψ 0 0 φn ψ 0
H̃ 1/2 (Γ2 )×H̃ −1/2 (Γ1 ) H̃ 1/2 (Γ2 )×H̃ −1/2 (Γ1 )
(8.46)

wn
with C independent of n, where the infimum is taken for ∈ H̃ 1/2(Γ2 ) ×
φn
H̃ −1/2(Γ1 ).
Theorem 8.7 Assume the boundary data in (8.35) satisfy g1 ∈ B 1,2(Γ1 ), g2 ∈
∗ β̂
0,1 vn
B (Γ2 ). Let ∈ Sn be the solution of the Galerkin equation (8.45) for
β̂ ψn∗
sufficiently large n, with Sn given by (8.42)–(8.44) and let u be the exact solution
of
˜
the boundary value problem (8.35). Then vn := l g1 |Γ2 + vn , ψn := l g2 + ψn∗
∗
Γ1
satisfy
0 0
0 0 0 ∂u 0 √
0 vn − u|Γ 0 0 0
+ 0 ψn − 0 C e−b N
(8.47)
2 H̃ 1/2 (Γ2 ) 0 ∂n Γ1 0
H̃ −1/2 (Γ1 )
where N = dim Sn = dim Sn,z + dim Sn,1 and C, b > 0 are constants independent
of n
Proof By Theorem 8.4, u ∈ Bβ2 (Ω). By [18, 19] there exists a function ũn in an h-p
finite element space S̃n on a geometric mesh in Ω satisfying
u − ũn H 1 (Ω) C e−b̃(dimS̃n )

1/3
with b̃ > 0, and C independent of n. The geometric mesh in Ω is here given such
that its nodes on the boundary Γ create the above introduced geometric mesh. Now
we will take the traces on Γ2 and the normal derivatives on Γ1 . The mapping T :
H 1 (Ω) → H̃ 1/2 (Γ2 ) × H̃ −1/2(Γ1 ) given by
( ( ) )
∂f ∂f
f → f |Γ2 − (l f |Γ1 )Γ , − l˜

2 ∂n Γ1 ∂n Γ2
Γ1
is continuous. This gives with f := u − ũn

0 0
0 0
0 u|Γ2 − ( l u|Γ1 )Γ2 − ũn |Γ2 − ( l ũn |Γ1 )Γ2 0 C e−b̃(dimS̃n )
1/3
H̃ 1/2 (Γ2 )
(8.48)
0( ( ) ) ( ( ) )0
0 ∂u ∂ ũn 0
0 − l˜ ∂u ˜ ∂ ũn 0
0 ≤ Ce−b̃(dimS̃n ) .
1/3
0 − − l
0 ∂n Γ1 ∂n Γ2 ∂n Γ1 ∂n Γ1 0
Γ1 Γ2 H̃ −1/2 (Γ1 )
(8.49)
By Theorem 8.5 we have for the exact solution (v ∗ , ψ ∗ ) of (8.40)

( )
∂u ∂u
∗
v = u|Γ2 − l u|Γ1 Γ , ∗
ψ = − l˜ (8.50)
2 ∂n Γ1 ∂n Γ2
Γ1

∂u
using the boundary conditions u|Γ1 = g1 , = g2 in (8.35). By the
∂n Γ2
( )
∂ ũn
construction of the spaces S̃n and Sn there holds ũn |Γ2 , ∈ Sn and we
∂n Γ1
obtain with the definition of the operators l, l˜
( )
∂ ũn ∂ ũ
vn∗ − l ũn |Γ1 Γ ∈ Sn,2 ψn∗ − l˜
n
:= ũn |Γ2 , := ∈ Sn,1
2 ∂n Γ1 ∂n Γ2
Γ1
(8.51)
Thus (8.48),(8.49),(8.50),(8.51) yield the existence of (vn∗ , ψn∗ ) ∈ Sn with

0 ∗ 0 0 ∗ 0
0v − v ∗ 0 1/2 0 ∗0 −b̃(dimS̃n )1/3
n H̃ (Γ ) + ψ − ψn H̃ −1/2 (Γ ) C e
2 1
(8.52)
Using dimS̃n C1 n3 and dimSn C2 n2 with suitable constants C1 , C2 we obtain
C e−b̃(dimS̃n ) C e−b(dimSn )
1/3 1/2
(8.53)
with b > 0 independent of n. Now the quasioptimality (8.46) gives the result (8.47)
using the definitions of un , ψn and (8.41).

Chapter 9
Mapping Properties of Integral
Operators on Polygons
In this chapter we introduce the analysis of boundary integral operators on a

polygon with the tool of the Mellin transformation from the original paper [128].
The interested reader may also look into [241] where the Mellin calculus is used
to analyse the mapping properties of the integral operators in countably normed
spaces. These results are crucial for deriving exponentially fast convergence of
the hp−version of the boundary element method (see Chap. 8). The results of the
subsection describing the regularity of the solution near the vertices were originally
published in [138]. The Mellin calculus is used in Sect. 9.3 to analyze the regularity
of the solution at the tip of an interface crack, in Sect. 9.4 to analyze the mixed
boundary value problem for the Laplacian with the hypersingular operator and
the singular behaviour of its solution at the point where Dirichlet and Neumann
conditions meet and in Sect. 9.5 to analyze the mapping propeties of boundary
integral operators with countably normed spaces. In the framework of these spaces
the analysis of the exponential convergence of the hp Galerkin approximation is
presented in Sect. 8.1.
9.1 Mellin Symbols
In this section (following [128]) we now have to cope with the problem that the
integral operators are now defined on curves with corners so that we cannot directly
apply the Fourier transformation. We will consider a polygon Γ as follows: We split
it into sectors Γ ω and pieces of straight lines (see Fig. 9.1) and apply now Mellin
techniques on Γ ω and pseudodifferential operators on the straight lines.

296 9 Mapping Properties on Polygons
For φ ∈ C0∞ (0, ∞) we define the Mellin transformed ϕ̂ of φ by:
∞ ∞
φ̂(λ) := x iλ−1
φ(x) dx = φ(e−t )e−iλt dt ,
0 −∞

R → R+
making use of the Euler transformation
t → e−t := x .
We have the inverse formula:

1
φh (x) = φh (e−t ) = eiλt φ̂(λ) dλ .
2π
((λ)=h
Defining now the operator (single-layer potential)

V−− V+−
V ∼
=
V−+ V++
with u|Γ ω ∼
= (u− , u+ ) on R+ , we have for φ ∈ C0∞ [0, ∞)
∞
1
V++ φ(x) := − ln |x − y|φ(y) dy
π
0
∞ ∞
1 1 x
=− ln(y)φ(y) dy −
ln 1 − φ(y) dy , x ∈ R+ = Γ+
π π y
0 0

=:l(φ) V0 φ(x)
9.1 Mellin Symbols 297
∞
1
V−+ φ(x) := − ln |x − eiω y|φ(y) dy , x ∈ R+ = Γ+
π
0
∞ ∞
1 1 x −iω
V+− φ(x) := − ln |xe iω
− y|φ(y) dy = l(φ) − ln |1 − e |φ(y) dy
π π y
0 0
=: l(φ) − Vω φ(x) , x ∈ Γ−
It is easily verified that there holds V++ = V−− , V+− = V−+ and on Γ ω :

ll V0 Vω
V = + .
ll Vω V0
Consider now the double-layer potential

∼ K−− K+−
K= .
K−+ K++
Due to the geometric interpretation of K as the variation of the angle, i.e. Kg(z) :=
− π1 Γ g(ζ ) ∂n∂ ζ ln |z − ζ |dsζ = − π1 Γ g(ζ )dθζ (z), where θζ (z) is the angle
between ζ − z and some fixed direction, we have K−− = K++ = 0 . Furthermore,
there holds
∞
1 ∂
K+− φ(x) := − ln |xeiω − y|φ(y) dy
π ∂ny
0
∞
1
= ((xeiω − y)−1 φ(y) dy =: Kω φ(x) .
π
0
With K−+ = K+− = Kω we therefore have on Γω :

0 Kω
K= .
Kω 0
@
In the following lemma we provide the Mellin symbols V>0 (λ), V @ω (λ).
ω (λ), K
Lemma 9.1 For the operators V and K as given above there holds:
i) Let φ ∈ C0∞ [0, ∞) . Then, for ((λ) ∈ (−1, 0) we have:
cosh(πλ) >
V? >0 (λ) · >
0 φ(λ) = V φ(λ − i) := φ(λ − i),
λ sinh(πλ)
cosh(π − ω)λ
V? @
ω φ(λ) = V ω (λ) · >
φ (λ − i) := >
φ (λ − i) .
λ sinh(πλ)
ii) Let φ ∈ C0∞ (0, ∞) . Then, for ((λ) ∈ (−1, 1) we have:
sinh(π − ω)λ >

K? @ω (λ) · >
ω φ(λ) = K φ(λ) := − φ(λ) .
sinh(πλ)
Proof Before proving the statements of the lemma, we leave it as an exercise to the
reader to show that for ω ∈ (0, 2π) and ((λ) ∈ (−1, 0) the following identity
holds (see [167]):
∞
e−iλτ e±λ(ω−π)
dτ = iπ . (9.1)
e−τ ±iω −1 sinh(πλ)
−∞
We now want to start with the equation for the double-layer potential, i.e.
ii):
∞
−t 1 1
Kω φ(e ) = ( φ(e−τ )e−τ dτ
π e−t +iω− e−τ
−∞
∞
1 1 1
= − φ(e−τ ) dτ
2πi e−(t −τ )+iω −1 e−(t −τ )−iω −1
−∞
∞
=: f (t − τ )φ(e−τ ) dτ . (9.2)
−∞
∞
Defining Ff (λ) := e−iλt f (t) dt , (9.1) yields:
−∞
eλ(ω−π ) −λ(ω−π )
Ff (λ) = 1
2πi iπ sinh(πλ) − iπ esinh(πλ)
= − sinh(π−ω)
sinh(πλ)
λ
@ω (λ) .
=: K
Thus, by the convolution theorem for the Fourier transformation and the above defi-
nition of the Mellin transformation, the assertion of ii) directly follows from (9.2).
i):
∞
Vω φ(e−t ) = − π1 ln |1 − eτ −t −iω |φ(e−τ )e−τ dτ
−∞
∞
=: f (t − τ )g(τ ) dτ .
−∞
Here we have
∞
F g(λ) = e−iλτ e−τ φ(e−τ )dτ = >
φ (λ − i) .
−∞
Now, integration by parts yields

∞
Ff (λ) = − π1 e−iλt % ln(1 − e−t −iω ) dt
−∞
=f (t )
∞
e−iλt e−t−iω
=− iπλ % 1−e−t−iω
dt
−∞
= cosh(π−ω)λ @
=: V ω (λ) ,
λ sinh(πλ)
for Imλ ∈ (−1, 0). For V? 0 φ one shall argue correspondingly and the assertion
follows analogously to i).

9.1.1 Mapping Properties in Weighted Sobolev Spaces
In this subsection we will consider in some more detail the weighted Sobolev spaces.
◦
Definition 9.1 We define the weighted Sobolev space W0s (R+ ) to be the comple-
tion of the space C0∞ (0, ∞) with respect to the norm:

1 s
φ2◦ := 1 + |λ|2 |>
φ(λ)|2 dλ for s ∈ R ,
W0s (R+ ) 2π
((λ)=s− 21
∞
where >
φ(λ) := x iλ−1 φ(x) dx .
0
For φ ∈ S we have the Fourier transformation:
∞
=
φ (ξ ) := e−iξ x φ(x) dx for ξ ∈ R ⇒=
φ∈S
−∞
Here, we recall the definition of the standard Sobolev space on R (see Appendix B)
Definition 9.2 The space H s (R) is defined as the completion of C0∞ (R) with
respect to the norm

s
φ2H s (R) := 1 + |ξ |2 |=
φ (ξ )|2 dξ for s ∈ R .
R
In order to investigate the mapping properties of the integral operators (e.g. the
single-layer potential V ) on a smooth boundary Γ , it therefore suffices to consider
its action on a function φ ∈ C0∞ (R) . For the theorem to follow we will need the
notion of the principal symbol of an operator which is defined as the leading term
of the Fourier transformation of the kernel of the operator and, furthermore, the
convolution theorem, stating that
γ γ (ξ ) · =
∗ φ(ξ ) = = φ (ξ )
holds ∀ φ ∈ C0∞ (R) (see Appendix B, Definition B.6).

Theorem 9.1 For a smooth boundary Γ there holds:
cont inuously
V : H s (Γ ) −→ H s+1(Γ ) ∀s ∈ R (9.3)
(or equivalently:
cont inuously
VR : H s (R) −→ H s+1(R) ∀ s ∈ R ). (9.4)
Proof We will show (9.4):

s+1 2
VR φ2H s+1 (R) = 1 + |ξ |2 (V<
R φ)(ξ ) dξ
R
= 1
|=
φ (ξ )|2
|ξ|2
(by the convolution theorem)
Now, for |ξ | large enough there holds: |ξ1|2 − 1+|ξ

1
|2
∼ |ξ1|4 . Let χ be C ∞ - cut-off
function with χ(ξ ) = 0 for |ξ | < and χ(ξ ) = 1 for |ξ | > 2 . Then, for the
principal symbol of χVR we have:

⎧
1 ⎪
|ξ | , |ξ | > 2
⎨
1
σ (χVR )(ξ ) = χ = 0 , |ξ | < .
|ξ | ⎪
⎩ smooth , else
We note that VR may be replaced by χVR , since the corresponding principal

symbols are the same apart from |ξ | < 2 . We may therefore replace the symbol
1
of VR by 1 .
(1+|ξ |2 ) 2
Thus, we have
VR φ2H s+1 (R) ≤ c2 φ2H s (R) ,

Corollary 9.1 The following operators are continuous:
◦ s ◦ s+1
(i) V0 : W0 (R+ ) −→ W0 (R+ ) for |s| < 1
2
◦ s ◦ s+1
(ii) Vω : W0 (R+ ) −→ W0 (R+ ) for |s| < 1
2
◦ s ◦ s
(iii) Kω : W0 (R+ ) −→ W0 (R+ ) for − 1
2 <s< 3
2
Proof We leave it as an exercise to the reader to show:
(i) V?φ(λ) = cosh(π−ω)λ >
λ sinh(πλ) φ (λ − i)
ω
⇒ @Vω (λ) ∼ 1+|λ| on every line ((λ) = h ∈ (−1, 0) .
1
>ω (λ)| < C on ((λ) = h ∈ (−1, 1) .

(ii) ∃ C > 0 (independent of λ) : |K
◦ s
We then have for φ ∈W0 (R+ ):
s+1
Vω φ2◦ s+1 ∼ 1 + |λ|2 >ω (λ)|2 · |>
|V φ (λ − i)|2 dλ
W0 (R+ ) ((λ)=s+ 21
s
1 + |λ|2 |>
φ (λ − i)|2 dλ
((λ)=s+ 21
s
1 + |λ|2 |>
φ(λ)|2 dλ 0 φ2◦ s ,
((λ)=s− 21 W0 (R+ )
For the other operators one shows the assertions correspondingly.

◦ s
We want to show next the relation between the weighted Sobolev spaces (R+ ) W0
=s s
and the Sobolev spaces H (R+ ) of H (R+ ) functions that have a zero continuation
on R− in H s (R) :
Lemma 9.2 ([284]) Let χ ∈ C0∞ [0, ∞) . Then, the mapping

% ◦ s
Ks (R+ )
W0 (R+ ) −→ H
u → χu
◦ s
and its inverse Ks (R+ ) −→ W0 (R+ )
H are both continuous for s ≥ 0 , i.e.
◦ s
Ks (R+ ) and
the norms of H W0 (R+ ) for s ≥ 0 are equivalent on compact
intervals.
Remark 9.1 It was shown in [204] that the mappings u → χu from H s (R+ ) into
◦ s
W0 (R+ ) and vice versa are continuous for s ≤ 0 , too.
Lemma 9.3 For 0 ≤ s < 32 there holds:
3 4
=s (R+ )
i) H s (Γ ω ) = u = (u1 , u2 ) ∈ H s (R+ )2 | u− − u+ ∈ H
3 4
ii) H −s (Γ ω ) = u = (u1 , u2 ) ∈ H −s (R+ )2 | u− + u+ ∈ H=−s (R+ )
Furthermore, the mappings

% , % ,
=s (R+ )
R : H s (Γ ω ) → H s (R+ ) × H R : H −s (Γ ω ) → H =−s (R+ )×H −s (R+ )
and
(u− , u+ ) → (u− + u+ , u− − u+ ) (u− , u+ ) → (u− + u+ , u− − u+ )
are isomorphisms.
Proof For the proof of the lemma we refer to [128, 204].

We now want to show that the mapping
χV χ : H s−1(Γ ω ) −→ H s (Γ ω ) (9.5)
is continuous for s ∈ − 12 , 32 , where χ shall denote the C ∞ -cut-off-function,

concentrated at {0}. By Lemma 9.3 we have that (9.5) is equivalent to the continuity
of the mapping:

=s (R+ ) for s ∈ − 1 , 3 .
=s−1(R+ )×H s−1(R+ ) −→ H s (R+ )× H
RχV χR −1 : H
2 2
=s for |s| <

Note that H s = H 1
and
2

1 1 2l + V0 + Vω 0
=
R> and thus RχV χR −1 =
>χ χ.
1 −1 0 V0 − Vω
Hence, (9.5) will be shown by the following lemma:

Lemma 9.4 ([128]) Let χ ∈ C0∞ [0, ∞) with supp(1 − χ) ⊂⊂ (0, ∞) .
Then the following mappings are continuous:

s +
= (R ) −→ H s+1 (R+ )
H
i) for s ∈ − 32 , 12
u → χ(l + Vω )χu
s +
H (R ) −→ H =s+1 (R+ )
ii) for s ∈ − 32 , 12
u → χ(V0 − Vω )χu
⎧ s + s +
⎨ H = (R ) H= (R )
+ −→
iii) s
H (R ) H s (R+ ) for s ∈ − 12 , 32
⎩
u → χKω χu
In analogy to the above lemma and with the help of the following exercise one
can show that the mapping
χ(I + K)χ : H s (Γ ω ) −→ H s (Γ ω ) (9.6)
is continuous for s ∈ − 12 , 32 .
Exercise 9.1 For the operators as defined above there holds

1 Kω
χ(I + K)χ =χ
ˆ χ
Kω 1
and thus

−1 1 + Kω 0
Rχ(I + K)χR =χ
ˆ χ.
0 1 − Kω
Thus, (9.6) follows from the fact that the mapping
=s (R+ ) −→ H s (R+ ) × H
Rχ(I + K)χR −1 : H s (R+ ) × H =s (R+ ) (9.7)
is continuous for s ∈ − 12 , 32 , which itself is a consequence of Lemma 9.4.
Theorem 9.2 For the single and double layer potential operators V and K there
holds:
i) ∃ γ = γ (I, ω) > 0 ∀ v ∈ L2 (Γ ω ) with supp(v) ⊂⊂ I ⊂⊂ Γ ω :
% (v, (I + K)v ) ≥ γ v2L2 (Γ ω ) .
1
ii) ∃ γ = γ (I, ω) > 0 ∀ ψ ∈ H − 2 (Γ ω ) with supp(ψ) ⊂⊂ I ⊂⊂ Γ ω :
% (ψ, V ψ ) ≥ γ ψ2 −1
.
H 2 (Γ ω )
Proof We first note that on Γ ω the operator K maps even (resp. odd) functions onto
even (resp. odd) functions.
ad i):
We have:
5 6
−1 I + Kω 0
v, (I + K)v L2 (Γ ω ) = v, R Rv
0 I − Kω
5 6 L (Γ )
2 ω
I + Kω 0
= 12 Rv, Rv
0 I − Kω
7 8 2 (Γ ω )
L
= 2 v+ + v− , (I + Kω )(v+ + v− ) 2 + 2 +
1
7 8L (R )×L (R )
+ 12 v− − v+ , (I − Kω )(v− − v+ ) 2 + 2 + .
L (R )×L (R )
Now, making use of Parseval’s equation for the Mellin transformation we obtain for
arbitrary v ∈ C0∞ (0, ∞)

% v, (I ± Kω )v L2 (Γ ω ) = 1 ? (I ±
v(λ) Kω )v(λ) dλ
2π
((λ)=− 21
sinh(π−ω)λ
= 1
2π 1∓ sinh(πλ) |>
v (λ)|2 dλ .
((λ)=− 21
There further holds

sinh(π − ω)λ
≤ sin π − ω =: q ω=0,2π
< 1 ∀ λ with ((λ) = −
1
sinh(πλ) 2 2
and thus

1−q
% (v, (I ± K)v ) ≥ |>
v (λ)|2 dλ = c(1 − q) v2L2 (Γ ω ) .
2π
((λ)=− 21 =:γ
Here, we note that on polygons the operator K is a contraction map, i.e.
KvL2 (Γ ω ) ≤ ηvL2 (Γ ω ) with η < 1 .
ad ii)
Analogously to i) we now have:
7 8
ψ, V ψ −1/2 ω 1 = 1/2 (ψ−+ψ+ ), (V0 +Vω )(ψ−+ψ+ ) 1 1
H (Γ )×H 2 (Γ ω ) =− 2 (R+ )×H 2 (R+ )
H
7 8
+ (ψ− − ψ+ ), (V0 − Vω )(ψ− − ψ+ ) − 1 + 1
+ H 2 (R = 2 (R )
)×H
By Parseval’s equation there holds for arbitrary ψ ∈ C0∞ [0, ∞) :
ψ, (V0 ± Vω )ψ L2 (Γ ω ) = 1
ψ >0 (λ) ± V
>(λ) V >ω (λ) ψ
>(λ − i) dλ
2π
((λ)=− 21
= 1 > − i)|2 dλ
V̂0 (λ) ± V̂ω (λ) |ψ(λ
2π
((λ)=0
and thus

1
ψ, (V0 ± Vω )ψ L2 (Γ ω ) = >(λ − i)|2 dλ
m± (λ)|ψ
2π
((λ)=0
cosh(πλ) ± cosh(π − ω)λ

for m± (λ) = .
λ sinh(πλ)
Note that for λ ∈ R we have

1 + |λ| 1
m+ (λ) ∼ and m− (λ) ∼ .
|λ|2 1 + |λ|
Hence,
(

% (ψ, V ψ ) ≥ γ 1+|λ| ψ>− (λ − i) + ψ >+ (λ − i)2 dλ
|λ|2
((λ)=0 )
1 >
2
+ ψ− (λ) − ψ
>+ (λ) dλ
1+|λ|
( ((λ)=−1 )
≥γ ψ− + ψ+ 2 1 + ψ− − ψ+ 2 1
=− 2 (R+ )
H ◦ −2
W 0 (R+ )
≥ γ · Rψ2 1 1
=− 2 (R+ )×H − 2 (R+ )
H
≥γ · ψ2 − 12
,
H (Γ ω )
since there holds (see [128]):

1 + |λ|
∃C > 0 ∀u ∈ C0∞ (0, ∞) : u2 − 1 ≤ C · |û(λ − i)|2 dλ ,
= 2 (R+ )
H |λ|2
((λ=0)
if the integral exists.

9.2 Properties of the Mellin Transformation
Let u ∈ C0∞ (0, ∞) be given. Then the Mellin transformed of u, defined by
∞
>
u(λ) = x iλ−1 u(x) dx
0
is an entire, analytic function. Defining

1
uh (x) := x −iλ û(λ) dλ ,
2π
((λ)=h
the residue theorem yields for uh1 and uh2 with h2 > h1 as in Fig. 9.2

uh2 = uh1 − i Res û(λ)x −iλ ,
((λ)∈(h1 ,h2 )
for uh2 being the meromorphic continuation of uh1 .

Here, we assume that > u(λ) only has poles in the range |λ| < M < ∞ and that
>
u(λ) is rapidly decaying for |λ| > M=>M.
Lemma 9.5 ([128]) Let u ∈ H =comp

s (R+ ) be given, i.e. for s < t , >
u(λ) is
meromorphic for ((λ) < t − 12 =: k with poles of order mk + 1 at
the points λl = iαl for s − 12 < αl < k , 1 ≤ l ≤ L and
s
1 + |λ|2 |>u(λ)|2 dλ < ∞ .
((λ)=s−1/2
Then there holds:
=t (R+ ) and χ ∈ C ∞ [0, ∞) we have
i) For uk (x) ∈ H 0

L
ml
u(x) = clm x αl logm (x)χ(x) + uk (x) (9.8)
l=1 m=0
i m+1
3 4
with χ ≡ 1 at x = 0 and clm = m! Resλ=λl (λ − λl )m û(λ) .
Fig. 9.2 Domain for Mellin (λ)

transform
h2
h1
(λ)
9.2 Properties of the Mellin Transformation 307
L ml
ii) |clm |2 ≤ c u2H=s + uk 2H=t .
l=1 m=0
=comp
iii) On the contrary, if for all u ∈ H s (R+ ) with a decomposition as in (9.8)
=
there holds that uk ∈ H (R+ ) , then the Mellin-transformed function has the
t
above properties.
Proof Note that for u(x) = x α lnl xχ(x) we have
>
φ (x)
>
u(x) =
(λ − iα)l+1
where >
φ is an entire function of exponential type which is rapidly decreasing for
%λ → ±∞. Therefore the inverse Mellin transform

1
uh (x) := eiλt û(λ)dλ (x = e−t ∈ R+ )
2π
(λ=h
exists for h ∈
/ {α1 , . . . , αn } and the path of integration may be shifted if we take
into account the residues of eiλt û(λ). Thus we get uh = u for h < α1 and uh2 (x) −
uh1 (x) = −i (λ∈(h1 ,h2 ) Res{û(λ)eiλt }(h1 < h2 ).
Now

lk
−i d lk iλt
−iResλ=iαk {eiλt fˆ(λ)} = [e û(λ)(λ−iαk )lk +1 ] = − ckl x αk logl x
(lk )! dλ x=iαk
l=0
gives i). For further details see [128].

With this lemma we have:
=comp 1
u∈H s
(R+ ) ⇒ >
u(λ) is holomorphic for ((λ) < s − ,
2
where u = uh for h < s − 12 , i.e. if the solutions do lie in the energy-space, the
parts below the energy-norm will cause smooth perturbations.
Lemma 9.6 Let k, φ ∈ C0∞ (0, ∞) and α, β ∈ C . Then there holds:
∞
x φ(y)
u(x) := α β
x y k dy ⇒ u(λ) = >
> k(λ − iα)>
φ λ − i(α + β) .
y y
0
Exercise 9.2 Prove Lemma 9.6 by using the result for the Fourier transform
of the convolution of two functions in the Appendix together with the Euler
transformation.
For the rest of this section we want to consider again the Dirichlet problem
−Δu = 0 in Ω,
u=g on Γ := ∂Ω,
given as an integral equation of the form
∂u
V = (I + K)g on Γ (9.9)
∂n
with
A Ψ = BG on Γ ω
for

V0 Vω ll I Kω
A=
> + , B>
= ,
Vω V0 ll Kω I
Ψ := (ψ− , ψ+ )t , G := (g− , g+ )t and Γ ω corresponding to Fig. 9.3

We assume that g± ∈ C0∞ [0, ∞), g+ (0) = g− (0) and will then show that a
solution Ψ of (9.9) has a representation of the form:
⎛ ⎞

n
lk
ψ± (x) = ⎝ ± αk −1
ckl x logl x ⎠ χ(x) + ψ±0
(x) .

k=1 l=0
=s (R+ ), s<3/2
∈H
Let A Ψ = BG =: H . We then have
?
A Ψ (λ) = A>(λ)Ψ >(λ)
> (λ − i) = H
with

>ω (λ)
V>0 (λ) V 1 cosh(πλ) cosh(π − ω)λ
A>(λ) = >ω (λ) V>0 (λ) =
V λ sinh(πλ) cosh(π − ω)λ cosh(πλ)
−Δu = 0
u= g Γ−
ω Γ+
Γω
χ
Fig. 9.3 Localization

>(λ − i) is the meromorphic extension of χ

for ((λ) ∈ (−1, 0) . Here, Ψ ? ψ(λ − i)
for the region from ((λ) ∈ (−1, 0) to ((λ) ∈ (−1, s − 12 ) . For all h ∈
>(λ − i) := A>(λ)−1 · H
−1, s − 1 , h = ((λp ) with λp being a pole of Ψ >(λ) ,
2
the function

1
Ψh (x) = >(λ − i)x −iλ−1 dλ
Ψ
2π
((λ)=h
exists. For h ∈ (−1, 0) there further holds that Ψh (x) = χψ(x) . For all other h
Cauchy’s integral theorem yields:

Ψh (x) = χψ(x) − i >(λ − i)x −iλ−1 .
Res Ψ
((λ)∈(−1,h)
>(λ) and A>−1 (λ) will be given by:

The residuals at the poles of H
(i) λ = 0 >(λ)
pole of H
(ii) λ = ik , k ∈ N >(λ)
pole of H
(iii) zeroes of det A>(λ) poles of A>−1 (λ) .
For the last item we have

sinh(2π − ω)λ · sinh(λω)
det A>(λ) =
λ2 sinh2 (πλ)
and

λ sinh(πλ) cosh(πλ) − cosh(π − ω)λ
A>−1 (λ) = .
sinh(2π − ω)λ · sinh(λω) − cosh(π − ω)λ cosh(πλ)
!
⇒ % sinh(2π − ω)λ · sinh(λω) = 0
λ=iα, α∈(0,2) (i) sinh(2π − ω)iα = 0 ⇔ α = 2π−ω lπ
, l = 1, 2, 3
⇒
(ii) sinh(iαω) = 0 ⇔ α = ω kπ
, k = 1, 2, 3
For more details the interested reader is referred to [128], where also the Neumann
problem and the mixed Dirichlet-Neumann problem are considered.
9.2.1 Local Regularity at Vertices
In order to obtain a local representation of the solution of integral equations with

singularity functions at the vertices of a polygon, we first need the regularity of the
solution on the smooth parts of the boundary. This regularity is characterised by
some standard a priori estimates making use of pseudodifferential operators. In this
subsection we report from [138]
Lemma 9.7 Let χ ∈ C0∞ be the cut-off-function with support inside of a segment
1
Γ j of Γ . Let f ∈ H s (Γ ) , s ≥ 12 and ψ ∈ H − 2 (Γ ) be a solution of V ψ = f
on Γ . Then we have the a-priori estimate:
"
χψH s−1 (Γ ) ≤ C · f H s (Γ ) + ψ −1 (9.10)
H 2 (Γ )
Proof We have χV ψ = χf and will consider the following situation:

Then there holds (Fig. 9.4):
χV χ1 ψ = − χV (1 − χ1 )ψ + χf =: χh . (9.11)
Here we note that with χ(1 − χ1 ) ≡ 0 the integral-kernel of χV (1 − χ1 ) is a

C ∞ -function. Thus,
"
χhH s (Γ ) ≤ C · f H s (Γ ) + ψ − 1 (9.12)
H 2 (Γ )
Therefore (9.11) may be conceived as an equation on a simply connected C ∞ -curve

Γ=, containing Γ j with cap(Γ=) = 1 . On Γ= we will now consider V −1 which is a
pseudo-differential operator of order 1. We then have
χ=χχ1
V −1 χV χ1 ψ = χχ1 ψ + V −1 χ − χV −1 V χ1 ψ = V −1 χh

pseudo-diff. op. of order 0
This yields the estimate

3 4
χψH s−1 (Γ ) ≤ C · χ1 ψH s−2 (Γ ) + χhH s (Γ ) .
for χ = χχ1 , χ1 ∈ C0∞ (Γ j )

χ χ1 χ2
Fig. 9.4 Cut-off functions

Repeating the above arguments gives

3 4
χ1 ψH s−2 (Γ ) ≤ C · χ2 ψH s−3 (Γ ) + χ1 hH s−1 (Γ ) .
with χ2 ∈ C0∞ (Γ j ) such that χ1 = χ1 χ2 , χ1 (1 − χ2 ) ≡ 0 . After a finite

number of applications of these arguments we obtain:
"
χψH s−1 (Γ ) ≤ C · χn ψ − 12 + χn h
H s (Γ ) ,
H (Γ )
which, in combination with (9.12), proves the lemma. For further details see [138].

For the regularity at the vertices we will use the Mellin-transformed equations and
Cauchy’s integral theorem. One may observe that the singularity-functions are given
by:
ψ ∼
= (ψ− , ψ+ ) = (c− , c+ )x −iλ−1 χ(x) , (9.13)
for λ being the zero of the transcendental equation
sinh(2π − ω)λ · sinh(ωλ) = 0 (9.14)
and (c− , c+ ) ∈ C2 the corresponding eigenvector of the Mellin-symbol of V . In

the case that λ is a double zero of (9.14) there holds:

1
ψ ∼
= (cl− , cl+ )x −iλ−1 logl (x)χ(x) . (9.15)
l=0
The local regularity results for the solution of the integral equation may be summed
up as follows:
Theorem 9.3 Let f ∈ H s (Γ ) , s > 1/2 , s = 2 + ((λ)
1
for all zeroes λ of (9.14).
− 21
Let further ψ ∈ H (Γ ) be a solution of V ψ = f on Γ . Then ψ has the local
representation

χψ = χψ (s) + ck vk , ck ∈ R,
0<((λk )<s− 21
for vk being functions of the form of either (9.13) or (9.15) with λ being a zero
of (9.14) or λk = i · m , m = 1, 2, . . . and χψ (s) ∈ H s−1 (Γ ω ) . Further there
holds
"
χψ (s) H s−1 (Γ ω ) + |ck | ≤ C · f H s (Γ ) + ψ − 1 .
H 2 (Γ )
0<((λk )<s− 21
Proof First we will transform the integral equation V ψ = f on Γ such that

Mellin transformation can be applied. To this end we take the cut-off-functions
χ1 , χ2 , χ3 ∈ C0∞ (C) , which only depend on |z| with

1 for |z| ≤ αj
0 ≤ χj ≤ 1 and χj ≡
0 for |z| ≥ βj ,
with j = 1, . . . , 3 ; αj < βj and β1 < α2 , β2 < α3 . Furthermore the support

of χj contains only the corner at the origin. In the neighbourhood of this vertex we
have
V ψ = f ⇐⇒ χ2 V χ1 ψ = χ2 f − V (1 − χ1 )ψ =: χ2 F on Γ .
Obviously, the same will hold on Γ ω . The right hand side of the above equation
satisfies the relation
"
χ2 F H s (Γ ω ) ≤ C · f H s (Γ ) + ψ − 1 ,
H 2 (Γ )
since
χ2 V (1 − χ1 )ψ = χ2 V (1 − χ1 )χ3 ψ + χ2 V (1 − χ1 )(1 − χ3 )ψ ,
where on one hand

"
χ2 V (1−χ1 )χ3 ψH s (Γ ω ) ≤ C·(1−χ1 )χ3 ψH s−1 (Γ ) ≤ C· f H s (Γ ) +ψ 1
H − 2 (Γ )
(by Lemma 9.7 and χ = (1 − χ1 )χ3 )

and on the other hand we have
χ2 V (1 − χ1 )(1 − χ3 )ψH s (Γ ω ) ≤ C · ψ −1 ,

H 2 (Γ )
since the operator has a C ∞ -kernel. Thus for the rest of the proof we only need
to give a lower estimate for χ2 V χ1 ψH s (Γ ω ) . For that purpose we approximate
χ1 ψ in H − 2 (Γ ω ) by φ ∈ C0∞ (Γ ω ) . Then >φ(λ − i) converges to χ?
1
1 ψ(λ − i) for
((λ) < 0 and thus V χ1 ψ(λ) = V >(λ)χ? 1 ψ(λ − i) for ((λ) ∈ (−1, 0) .
⇒ V χ1 ψ = (1 − χ2 )V χ1 ψ + χ2 F =: H on Γ ω .
9.3 Interface Crack 313
Application of the Mellin-transformation now yields:
V >(λ)
χ1 ψ(λ) = H ∀ ((λ) ∈ (−1, 0) ,
which can be continued meromorphically for all λ with ((λ) ∈ (−1, s − 12 ) .

The rest of the proof easily follows with the estimates of the last section.

9.3 A Direct Boundary Element Method for Interface Crack

Problems
We present from [424] a Galerkin boundary element method to solve the transmis-
sion problem with a crack in the interface for the Helmholtz equation. We investigate
the regularity for polygons and improve the rate of convergence by various methods:
(i) refinement of the mesh toward the crack tips, (ii) augmenting the finite element
space by singular functions, (iii) increasing the order of polynomials for fixed
mesh (p-method). For other boundary value problems or elasticity problems one
can proceed analogously, applying the convergence analysis in [128] to the general
method in [422].
We consider a bounded Lipschitz domain Ω1 ⊂ R2 and its exterior Ω2 = R2 \
Ω 1 . The boundary ∂Ω1 consists of a crack and a transmission part. The two sides
of the crack facing Ω1 and Ω2 are denoted by Γ1 and Γ4 ,resp., and the two sides
of the interface by Γ2 and Γ3 . The normal vector on Γ1 and Γ2 points to Ω2 , the
normal vector on Γ3 and Γ4 points to Ω1 . The crack tips are denoted by z1 and
z2 . We look for solutions of the following crack transmission problem: Find uj ∈
1 (Ω ), j = 1, 2; satisfying
Hloc j
Δ + kj2 uj = 0 in Ωj , j = 1, 2 (9.16)

1 ∂u1 1 ∂u2
u1 |Γ2 = u2 |Γ3 + v2 , =− + ψ2 (9.17)
ρ1 ∂n Γ2 ρ2 ∂n Γ3

1 ∂u1 1 ∂u2
= ψ1 = ψ4 (9.18)
ρ1 ∂n Γ1 ρ2 ∂n Γ4
Here we assume for the wave numbers Im kj2 0, kj = 0 and for the densities
ρj > 0. The it is shown in [422] that for given boundary data ψ1 , v2 , ψ2 , ψ4 in the
space
3
Y = (ψ1 , v2 , ψ2 , ψ4 ) ∈ H −1/2(Γ1 )×H 1/2(Γ2 ) × H −1/2(Γ2 ) × H −1/2(Γ4 ) |
4
(ψ1 + ψ4 , ψ2 ) ∈ H −1/2 (∂Ω1 ) (9.19)
there exists a unique solution u1 , u2 of (9.16)–(9.18). The unknown boundary data

1
u1 |Γ1 , ρ1−1 ∂u
∂n , u1 |Γ2 , u2 |Γ4 are an element of the affine space
Γ2
3
X = (w1 , φ2 , w2 , w4 ) ∈ H 1/2 (Γ1 ) × H −1/2 (Γ2 ) × H 1/2 (Γ2 ) × H 1/2 (Γ4 ) |
4
(w1 , w2 ) ∈ H 1/2 (∂Ω1 ), (w4 , w2 − v2 ) ∈ H 1/2 (∂Ω2 ), (ψ1 , φ2 ) ∈ H −1/2 (∂Ω1 ) .
(9.20)
Following the method in [422] we derive the following boundary integral equation
for the unknown boundary data ζ = (w1 , φ2 , w2 , w4 ) ∈ X
Hζ = g (9.21)
where g is determined by the given boundary data (ψ1 , v2 , ψ2 , ψ4 ) ∈ Y and

⎡ −1 ⎤
ρ1 W11 K12 ρ1−1 W12 0
⎢ −K21 ρ1 V22 + ρ2 V33 −K22 + K33 K34 ⎥
H =⎢
⎣ρ W21 K − K
−1 −1 −1 −1
⎥
⎦ (9.22)
1 22 33 ρ1 W22 + ρ2 W33 ρ2 W34
−1 −1
0 −K43 ρ2 W43 ρ2 W44
The boundary integral operators in H are the traces and normal derivatives of the
single and double layer potential, which are defined with the fundamental solution
(1)
Gj (x, y) = 4i H0 (kj |x − y|) of (9.16). Because of the singularities these integrals
are understood as finite part integrals: (r = 1 for i, j 2, r = 2 for i, j 3, x ∈
Γi )

∂
Vi,j ψj (x) = 2 Gr (x, y)ψj (y)dsy , Kij ψj (x) = 2 Gr (x, y)ψj (y)dsy ,
∂nx
Γj Γj

∂ ∂ ∂
Kij vj (x) = 2 Gr (x, y)vj (y)dsy , Wij vj = −2 Gr (x, y)vj (y)dsy .
∂ny ∂nx ∂ny
Γj Gj
There hold the following results for (9.21) (see [424]):

9.3 Interface Crack 315
Theorem 9.4 The interface crack problem (9.16)–(9.18) and the boundary integral
equation (9.21) are equivalent: Let uj ∈ Hloc
1 (Ω ) be a weak solution of (9.16)–
j
(9.18), then
( )
∂u1
ζ = u1 |Γ1 , ρ1−1 , u2 |Γ2 , u2 |Γ4
∂n Γ2
belongs to X and satisfies (9.21). If on the other hand ζ = (w1 , φ2 , w2 , w4 ) satisfies

(9.21), then w1 , w2 , ψ1 , φ2 define by Green’s representation formula a function
u1 ∈ Hloc1 (Ω ) and w := w − v , w , φ := −φ − ψ , ψ define a function
1 3 2 2 4 3 2 2 4
u2 ∈ Hloc (Ω2 ) such that u1 and u2 satisfy (9.16)–(9.18).
1
Theorem 9.5
(I) Let (ψ1 , v2 , ψ2 , ψ4 ) ∈ Y be given. Then there exists exactly one solution ζ ∈ X
of the integral equation (9.21).
(II) There exists a compact operator T0 : X̃ → Y and γ > 0 such that for all
ζ ∈ X̃
9 :
% (H + T0 ) ζ, ζ Y ×X̃ γ ζ 2X̃ (9.23)
where X̃ is the homogeneous space corresponding to X. This is the dual space

of Y and ·, · Y ×X̃ denotes the natural duality.
A general Galerkin procedure involves a family
* 0 of finite dimensional affine
subspaces Sh0 ⊂ X for h ∈ (0, h0 ) such that Sh is dense in X, and solving
h>0
the problem: Find ζh ∈ Sh0 such that for all ξh ∈ S̃h0
H ζh − g, ξh = 0 (9.24)
where S̃h0 denotes the homogeneous space corresponding to Sh0 .

Now Gårding’s inequality (9.23) and the invertibility of H together imply the
following quasioptimal error estimate for the Galerkin procedure (9.24) by standard
arguments [408] (see Theorem 6.1):
Proposition 9.1 There exists h0 > 0 such that for any h ∈ (0, h0 ) the Galerkin
equations (9.24) have a unique solution ζh , and there exists C > 0 such that for the
exact solution ζ ∈ X of (9.21) holds
ζ − ζh X̃ C inf ζ − ξh X̃ . (9.25)

ξh ∈Sh0
We choose S t,k -systems of finite elements for Sh0 :
Sh := Sht,k (Γ1 ) × Sht −1,k−1 (Γ2 ) × Sht,k (Γ2 ) × Sht,k (Γ4 ), t ∈ N, k ∈ N, t > k
(9.26)
and define Sh0 as the affine subspace of Sh which satisfies the inhomogeneous
compatibility conditions in the crack tips:
w1h (zi ) = w2h (zi ), w2h (zi ) = w4h (zi ) + v2 (zi ), i = 1, 2.
In order to derive convergence rates in (9.25) we need to know the regularity of the
B
nj
exact solution ζ of (9.21). Let H s (Γj ) := H s Γjl where Γjl , l = 1, . . . , nj ,
l=1
are the sides of the polygon contained in Γj . Then define
& '
Y s = (ψ1 , v2 , ψ2 , ψ4 ) ∈ Y |v2 ∈ H 1/2 (Γ2 ), ψj ∈ H −1/2+s (Γj ), j = 1, 2, 4
& '
Xs = (w1 , φ2 , w2 , w4 ) ∈ X|wj ∈ H 1/2+s (Γj ), j = 1, 2, 4, φ2 ∈ H −1/2+s (Γj )
Proceeding as in[128],[129] we use the Mellin transformation and obtain, see

Sect. 9.1
Theorem 9.6 Let (ψ1 , v2 , ψ2 , ψ4 ) ∈ Y s be given. Then the solution ζ of (9.21) has
the form
L∗

J j
ζ = Cj l ζj l + ζs (9.27)
j =1 l=1
with ζs ∈ Xs , s ∈
/ A where A is the set of singular exponents described below. Here
jl jl jl jl
the singular functions ζj l = w1 , φ2 , w2 , w4 are of the form
α α−1
wj l (x) = x − zj logr x − zj χj (x), φ j l (x) = x − zj logr x − zj χj (x)
where zj , j = 1 . . . J , are the crack tips and the corners of the polygon, χj (x) are
cut-off functions. The singular exponents α are either in N or a zero of det Ĥ(j ) (iα).
H(j ) is the boundary integral operator describing the problem in the neighbourhood
of zj , and Ĥ(j ) (λ) is its Mellin symbol.
For corners zj in the crack with angle ωj one gets the well-known singularities
for Neumann problems, namely α = k ωπj , k ∈ N (see [128]). For corners in the
interface the results of [129] apply and yield the transcendental equation (ρ1 +
ρ2 )2 sin2 πα = (ρ1 − ρ2 )2 sin2 (π − ωj )α. For the singularity in the crack tips we
have to calculate the determinant of the Mellin symbol of H in (9.22). This gives
the transcendental equation
(ρ1 + ρ2 )2 sin2 π2α = (ρ1 − ρ2 )2 sin2 (π − ω)2α (9.28)
Hence the singular exponents are half the exponents of a corner in the interface. For
ω = π or ρ1 = ρ2 we get α = 12 , the known singularity of a crack. For ω = π
9.4 Mixed BVP of Potential Theory on Polygons 317
let ω = min{ω, 2π − ω}, αo = π

2(2π−ω ) > 14 , α o = min{ 2ω
π π
, (2−ω ) } 4 . Then
3
(9.28) has in (0, α o ) exactly two real solutions α1 , α2 with αo < α1 < 12 < α2 < α o
and α1 → αo , α2 → α o as ρρ12 tends to zero or to infinity.
Applications
of Theorem 4 to the estimate (9.25) yields the convergence rate
O hα0 −ε where α0 = min A, 14 < α0 12 . T his low convergence rate can be
improved by various methods:
(i) Refinement of the mesh towards the crack tips by a grading paramter β 1:
For β > 2α3 0 we obtain for t = 2, k = 1 in (9.26) a convergence rate of

O h3/2−ε (see [422]).
(ii) By augmenting the boundary elements by all singular functions withexponents
smaller than some α1 ∈ A, we get a convergence rate of O hα1 −ε if t > α1
in (9.26) (see [128], [129].
(iii) Performing the p−method where the degree p of the polynomials
on a fixed
mesh is increased gives a convergence rate of O p−2α0 +2 , using results in
[405], [404]. This is twice the rate of the h−version with uniform mesh.
9.4 Mixed BVP of Potential Theory on Polygons
The following boundary integral operators (and some closely related to them) appear
in these applications:

Vf (z) = f (ζ )G(z, ζ )dsζ ;
Γ

∂
Kf (z) = f (ζ ) G(z, ζ )dsζ ;
∂nz
Γ

∂
K f (z) = f (ζ ) G(z, ζ )dsζ ;
∂nz
Γ

∂ ∂
Wf (z) = − f (ζ ) G(z, ζ )dsζ ;
∂nz ∂nz
Γ

1
Sf (z) = (ζ − z)−1 f (ζ )dζ,
iπ
Γ
where 12 G(z, ζ ) is the fundamental solution −(1/2π) ln |z − ζ | of the Laplace or

(i/4π)H0(l)(k|z − ζ |) of the Helmholtz equation.
As an example, we consider the mixed Dirichlet-Neumann problem with given
Cauchy data g = u on Γ1 and h = ∂u ∂n on Γ2 , and unknown Cauchy data v := u on
Γ2 and ψ := ∂u
∂n on Γ1 . The representation formula is

∂ ∂u
2u(z) = u(ζ ) G(z, ζ )dsζ + (ζ )G(z, ζ )dsζ for z ∈ Ω1 ,
∂nζ ∂nζ
Γ Γ
where ∂u/∂n means the derivative with respect to the normal pointing from Ω1 to
Ω2 .
Taking the limit of u(z) for z ∈ Γ2 and the normal derivative ∂u(z)/∂n for z ∈ Γ1
in this formula and using the jump relations, one finds the system

W22 K12 v −W12 1 − K22 g
= (9.29)
−K21 V11 ψ 1 + K11 −V21 h
where the subscripts at Wj k etc. mean: integration over Γj and evaluation on Γk .

This is a system of first kind integral equations for v and ψ. The same problem
has been studied in [128] using a second kind integral equation for v instead of the
first line in (9.29). This was obtained by taking the limit of u(z) for z ∈ Γ1 instead
∂u/∂n. The system then had to be modified in order to satisfy a Gårding inequality,
whereas the system (9.29) does this without modification. Another advantage is that
the natural bilinear form associated with (9.29) is equivalent to the “energy norm”,
i.e., the natural norm for the Cauchy data of the weak (or variational) solution of the
problem. On the other hand, (9.29) contains the operator W of the normal derivative
of the double layer potential having a hypersingular kernel that therefore was (for
obvious reasons) used to be avoided in applications.
Local properties of the boundary integral operators are studied using an infinite
sector {z ∈ C|0 < arg z < ω} as a model domain. The boundary is then
Γ ω = Γ˙− ⊂ Γ + with Γ− = eiω · R+

and Γ+ = R+ ,
and the equations on Γ are reduced to Γ ω by means of standard localizing

techniques, [128], Sect. 9.1 . In the localized forms of the operators appear integral
operators on R+ with kernels on the following form
1 x
Vω (x, y) = − ln |1 − eiω |;
π y

1 1 1 eiω
Kω (x, y) = ( ; Kω (x, y) = (
π xeiω − y π xe − y
iω
1 ∂
Wω (x, y) = − Kω (x, y).
x ∂ω
These kernels are functions positively homogeneous in (x, y) and thus for the
corresponding operators the Mellin convolution theorem (Lemma 9.6) can be
applied.
If the differential operator contains lower order terms or the boundary Γ is curved
near the corners, then the kernels contain in addition an expansion

x k y (xeiω − y)−n (9.30)
l,,n
into terms of ascending degree of homogeneity. Finally, the localization procedure

produces additional smoothing operators thus introducing smooth or compact
perturbations of the results obtained by the symbolic calculus.
The convolution theorem (Lemma 9.6) gives for f ∈ C0∞ (0, ω) the Mellin
symbols:
cosh(π − ω)λ ˆ
V?
ω f (λ) = f (λ − i) =: V̂ω (λ)fˆ(λ − i)
λ sinh πλ
((λ ∈ (0, 1));
sinh(π − ω)λ ˆ

Kω f (λ) = − f (λ) =: K̂ω (λ)fˆ(λ)
sinh πλ
((λ ∈ (−1, 1));

K ˆ
ω f (λ) = K̂ω (λ + i)f (λ) ((λ ∈ (−2, 0));

Wω f (λ) = −(λ + i) V̂ω (λ + i)fˆ(λ + i)
2
((λ ∈ (−2, 0)).
Introducing this into the localized form of the boundary integral equations
AU = F , and including smoothing parts into the right hand side Γ , one finds the
transformed equation

m
Â(λ)Û (λ) = F̂ (λ) − Âk (λ)Û (λ − ki). (9.31)
k=l
The additional terms on the right side correspond to the lower order terms (9.30) in
the kernels [128].
In our example we have locally, at a corner where the boundary condition
changes, Γ− corresponding to Γ2 and Γ+ to Γ1 :

−W0 Kω
A= , (9.32)
−Kω V0
hence

λ2 V̂0 (λ) K̂ω (λ)
Â(λ) = ;
−K̂ω (λ) V̂0 (λ)

V̂ (λ)
Û (λ) = ; (9.33)
ψ̂(λ − i)

fˆ1 (λ − i)
F̂ (λ) = .
fˆ2 (λ)
The right hand side in (9.31) is converted by local Mellin transforms into

λ2 V̂ω (λ) 1 g(λ)
F̂ (λ) = .
1 −V̂ω (λ) ĥ(λ − i)
Having calculated the Mellin symbols, one uses the Parseval relation and the
connection between singular expansion at the origin and poles of the Mellin
transform (Lemma 9.5) to deduce from equation (9.31) results of the following
kind:
(i) Continuity of the operators in Sobolev spaces H s (Γ ), H s (Γj ) or H̃ s (Γj ),
(j = 1, 2), where, as usual, for s > 0 H s (Γ ) is the space of traces
of H s+1/2(R2 ) on Γ, H 0 (Γ ) = &L2 (Γ ), and H −s (Γ is the
' dual space of
H (Γ ). Furthermore, H̃ (Γ1 = u ∈ H (Γ )|u|Γ2 = 0 and H s (Γ1 ) =
s s s
H s (Γ )/H̃ s (Γ2 ) have to be used.

(ii) A Gårding inequality on the boundary.
(iii) Regularity of the solution. This includes an expansion into explicitly given
singular functions at the corners and a more regular part, and an a-priori
estimate for the coefficients of the singular functions and for the regular part in
terms of Sobolev norms.
These results are then used to reach the following conclusions: Firstly, Fredholm’s
alternative holds, yielding existence of a solution of the boundary value problem
if uniqueness is known. (Uniqueness, in general, has to be deduced from different
principles).
Secondly, standard arguments for approximation schemes yield convergence and
quasi-optional asymptotic error estimates for the general Galerkin procedure applied
to the integral equations. The regularity results (iii) show that higher convergence
rates are obtained by including the singular functions in the approximate solution.
In our example, the above results are obtained in the following way (we always
localize at a “mixed” corner, the cases of “Dirichlet” or “Neumann” corners have
been treated earlier (see Sect. 9.2):
(i) Here, the symbol Â(λ) has to be estimated from above: For s ∈ − 12 , 32 , one
finds for (λ = s − 1
2 the qualitative behaviour:
−1/2

V̂0 (λ) ≈ 1 + |λ|2 ,

K̂ω (λ) ≈ e− min{ω,2π−δ}·|λ| ,
hence,
2 2

Â(λ)Û (λ) ≤ C 1 + |λ|2 v̂(λ)
−1 2
+ 1 + |λ|2 ψ̂(λ − i) .
By Parseval’s relation, this means continuity from H s (R+ ) ⊕ H s−1 (R+ ) to

H s−1(R+) ⊕H̃ s (R+ ). Duality and interpolation extends the range to s ∈ − 12 , 32 ,
and patching up the local results, one obtains:

W22 K12
Theorem 9.7 Let s ∈ − 12 , 32 . The operator B = maps
−K21 V11
H̃ s (Γ2 ) H s−1(Γ2 )
⊕ continuously into ⊕ .
s−1
H̃ (Γ1 ) s
H (Γ1 ),
(ii) For Gårding’s inequality, the symbol has to be estimated from below on the
appropriate line (λ = s − 12 . In our example s = 12 . For real λ, we obtain
%Â(λ)Û (λ) · Û (λ)

&
= % (λ2 V̂0 (λ)v̂(λ) + K̂ω (λ)ψ̂(λ − i))v̂(λ)
'
+ −K̂ω (λ)v̂(λ) + V̂0 (λ)ψ̂ (λ − i) ψ̂(λ − i)
= λ2 V̂0 (λ)|v̂(λ)|2 + V̂0 (λ)|ψ̂(λ − i)|2

1 + |λ| 2

≥ γ (1 + |λ|2 )1/2 |v̂(λ)|2 + ψ̂(λ − i) .
λ2
Therefore,
∞
% 2π AU (x) · U (x)dx = % ÂÛ (λ) · Û (λ)dλ
0 (λ=0
(
≥γ (1 + |λ|2 )1/2 |v̂(λ)|2 dλ
(λ=0
)
1+|λ|
+ |λ+i|2
|ψ̂(λ)|2 dλ
(λ=−1
≥ γ v2 + ψ2 .
H̃ 1/2 (R+ ) H̃ −1/2 (R+ )
Note that after applying Parseval’s relation for s = 0, we shifted the path of
integration from (λ = − 12 to (λ = 0, e.g., :

1
2π (−W0 v(x))v(x)dx = −W0 v(λ)v̂(λ)dλ
2
(λ=0 (λ=− 12

= (λ + i)2 V̂0 (λ + i)v̂(λ + i)v̂(λ + i)dλ
(λ=− 12

= λ2 V̂0 (λ)|v̂(λ)|2 dλ.
(λ=0
This shifting is allowed if we assume v ∈ C0∞ (0, ∞) so that v̂ is analytic.

Moreover, we used that H̃ −1/2 = (H 1/2) , and therefore from

|λ|2
|û(λ)|2 dλ = C|u|21/2 ≤ Cu2H 1/2 (R+ )
1 + |λ|
(λ=0
follows

1 + |λ|
u2H̃ −1/2 (R ≤C |û(λ)|2 dλ.
+) |λ + i|2
(λ=−1
Together with the local Gårding inequalities for Dirichlet and Neumann
corners shown earlier, we thus obtain a global Gårding inequality:
Theorem 9.8 There exists a constant γ > 0 and a compact operator C :
H̃ 1/2(Γ2 ) ⊕H̃ −1/2 (Γ2 ) → H −1/2 (Γ2 ) ⊕ H 1/2(Γ1 ) such that for all U =
(v, ψ) = H̃ 1/2(Γ2 ) ⊕ H̃ −1/2(Γ1 ) there holds
%(B + C)U, Ū ≥ γ (v2H̃ 1/2 (Γ ) + ψ2H̃ −1/2 (Γ ) .

2 1
Here ·,· means the natural duality between H̃ 1/2(Γ2) ⊕ H̃ −1/2(Γ1) and
H −1/2(Γ2 ) ⊕H 1/2 (Γ1 ), which is, for smooth functions U = (v, ψ) and
9.5 BI Operators in Countably Normed spaces 323
W = (f1 , f2 ), given by

W, U := f1 vds + f2 ψds.
Γ2 Γ1
(iii) For a singular expansion and an a priori estimate, one has, according to
Lemma 9.5 , to find the poles and the residues of the meromorphic function
Û (λ). In equation (9.31), one may assume F to be smooth which means that
F̂ is meromorphic whith poles only at λ ∈ iZ. Then from (9.31) there arise
two kinds of poles for Û , namely at the poles of Â(λ)−1 and additionally at
poles of Û (λ − ki), k = 1, . . . , m, due to the lower order terms Âk (λ). The
latter give in this way, for every pole of Â(λ)−1 at λ = λ0 , rise to infinitely
many poles at λ = λ0 + i, λ0 + 2i, . . .. In our example, no lower order terms
are present, so we only have to find the poles of Â(λ)−1 , i.e., the zeros of
det Â(λ) = λ2 V̂0 (λ)2 + K̂ω (λ)2

cosh ωλ cosh(2π − ω)λ
= .
sinh2 πλ
For cosh ωλ = 0 we find the poles at λ = iα, α = 2k−1 2 ω , k ∈ N, which

π
give the well known functions |z − zj |α log |z − zj |r for v(z) and |z − zj |α−1
log |z−zj |r for ψ(z)(r = 0, 1). Here zj is the corner point where the boundary
conditions change. Taking into account the poles of Û (λ) for (λ ∈ 0, s − 12
at each corner point zj (j = 1, . . . , J ) and the corresponding singular functions
vj , ψj ( = 1, . . . , Lj ), one finds for the Cauchy data of the weak solution
of the mixed boundary value problem the expansion
J Lj
v vj vs
= cj +
ψ j =l =l ψj ψs
with (vs , ψs ) ∈ H̃ s (Γ2 ) ⊕ H s−1 (Γ1 )
and a corresponding a priori estimate.
9.5 Boundary Integral Operators in Countably

Normed Spaces
Our regularity investigations below will be based on the weighted Sobolev spaces
and the countably normed spaces as introduced in the following.
Let I = (0, 1). By Hβm,l (I ) (m ≥ l ≥ 1, integers and 0 < β < 1) we denote the
completion of the set of all infinitely differentiable functions under the norm

m
φ2 m,l = φ2H l−1 (I ) + |φ|2 j,l (9.34)
Hβ (I ) Hβ (I )
j =l
where
|φ|H j,l (I ) := x β+j −l φ (j ) L2 (I ) . (9.35)

β
The countably normed spaces Bβl (I ) on I are defined as
Bβl (I ) = {φ ∈ Hβm,l (I ), m = l, l + 1, . . . ; ∃C ≥ 0, d ≥ 1 ∀j = l, l + 1, . . .
(9.36)
|φ|H j,l (I ) ≤ Cd (j −l) (j − l)!} (l ≥ 1, integer).
β
On Γ these spaces are defined as the product spaces
j
Hβm,l (Γ ) = ΠjJ=1 Πk=1
2
Hβm,l (Γk ),
j
Bβl (Γ ) = ΠjJ=1 Πk=1
2
Bβl (Γk ) ∩ H l−1 (Γ ) (9.37)
j
where each boundary piece Γk has to be mapped onto I such that the vertex tj +k−2
falls onto 0 in order to apply the definition (9.36). If we want to emphasize the
l
dependence on the constants C and d we will write Bβ,C,d instead of Bβl .
For technical reasons we need the following representation of the countable
normed spaces:
∞
L l,L
l
Bβ,C,d (I ) = Bβ,C,d (I ) (9.38)
L=l
where
l,L
Bβ,C,d (I ) := {φ ∈ HβL,l (I ) ; |φ|H j,l (I ) ≤ Cd (L−l) (j − l)!, j = l, l + 1, . . . , L}.
β
(9.39)
The spaces Bβl,L (Γ ) are defined accordingly to (9.37). For localization techniques
one needs to introduce cut-off functions. These turn out not to be comprised by the
general countable normed spaces Bβl . But, evidently, for the spaces Bβl,L there exist
partitions of unity. Furthermore, they can be chosen such that the constants C and d
l,L
of Bβ,C,d (Γ ) do not depend on the parameter L.
Lemma 9.8 ([241, Lemma 2.1]) Let U ⊂ Γ be an open set and Uδ := {x ∈ Γ :

l,L
dist (x, U ) ≤ δ} for δ > 0. Let φ ∈ Bβ,C,d (Uδ ) for all L ≥ l. Then there exists for
∞
each L ≥ l a cut-off function χL ∈ C (Γ ) such that
χL | U ≡ 1 and χL |Γ \Uδ ≡ 0
and
χL φ ∈ B l,L (Γ )
β,C̃,d̃
with constants C̃ and d̃ independent of L.

For φ ∈ C0∞ (0, ∞) the Mellin transformation is defined by
∞
M (φ)(λ) := φ̂(λ) := x iλ−1 φ(x)dx. (9.40)
0
The seminorm |φ|H j,l (I ) can be characterized by using this transformation.

β
Lemma 9.9 ([241, Lemma 2.4]) Let φ ∈ C0∞ (I ) and 0 < β < 1. Then

|φ|2 j,l 0 |fj (λ)|2 |φ̂(λ)|2 dλ (j ≥ l),
Hβ (I ) ((λ)=l−1/2−β
where fj (λ) := iλ · (iλ + 1) · · · (iλ + j − 1). The constants in the mutual estimates
do not depend on j .
Now we present our main results concerning the Poincaré–Steklov operator in
countable normed spaces from [242] . Before doing so we need to recall the
respective results for standard Sobolev spaces.
For Lipschitz domains continuity and regularity of the integral operators (2.18)–
(2.21) as mappings between usual Sobolev spaces have been investigated by
Costabel, see [114]. Using these estimates and noting that the Poincaré-Steklov
operator maps the Dirichlet datum onto the Neumann datum, we obtain the
following proposition.
Proposition 9.2 ([114]) For all σ ∈ [0, 1/2] S : H 1/2+σ (Γ ) → H −1/2+σ (Γ )
is continuous. For σ ∈ [0, 1/2] let v ∈ H 1/2(Γ ) satisfy Sv ∈ H −1/2+σ (Γ ). Then
v ∈ H 1/2+σ (Γ ), and there holds the a priori estimate
vH 1/2+σ (Γ ) ≤ C (SvH −1/2+σ (Γ ) + vH 1/2 (Γ ) ).
Following Costabel and Stephan in [129] we use the method of Mellin trans-
formation to investigate the Poincaré-Steklov operator acting on countable normed
spaces, see also [241]. First we look at the local properties on an infinite angle Γ ω .
In a second step we apply these results to the boundary Γ of a polygonal domain. Let
Γ ω = Γ − ∪ {0} ∪ Γ + with Γ − = eiω R+ and Γ + = R+ (ω ∈ (0, 2π)). A function

φ on Γ ω can be identified with the pair (φ− , φ+ ) of functions on R+ defined by
φ− (x) = φ(xeiω ), φ+ (x) = φ(x) (x > 0). We will choose the representation of φ
by its even and odd parts (in a formal sense) which are defined by
1 1
φe = (φ− + φ+ ), φo = (φ− − φ+ ).
2 2
This induces for any operator A acting on functions on Γ ω a representation by a
2 × 2-matrix of operators acting on functions on R+ :

Aee Aoe (Aφ)e = Aee φ e + Aoe φ o ,
A=
ˆ A := where
Aeo Aoo (Aφ)o = Aeo φ e + Aoo φ o .
We need the following operators acting on functions on R+ :

1 ∞ x
Vω φ(x) := − ln |1 − e−iω |φ(y) dy, V0 = Vω for ω = 0,
π 0 y
∞
1 1
Kω φ(x) := ( φ(y) dy,
π 0 xeiω − y

1 ∞ eiω
Kω φ(x) := ( φ(y) dy, (9.41)
π 0 xe − y
iω
1 ∂
Wω φ(x) := − Kω φ(x), W0 = lim Wω .
x ∂ω ω→0
Then, with the exception of finite dimensional operators which are negligible in
our theory, the integral operators (9.41) can be represented by the following matrices
(see [129]):

V0 + Vω 0
V=
ˆV = ,
0 V0 − Vω

Wω − W0 0
W=
ˆW = ,
0 −(W0 + Wω )

Kω 0 Kω 0
K=
ˆ K = , K =
ˆK = .
0 −Kω 0 −Kω
Using these representations we also obtain the representation of the Poincaré-

Steklov operator acting on even and odd functions on the infinite angle:

See Soe
S=
ˆS =
Seo Soo
with
See = Wω − W0 + (I + Kω )(V0 + Vω )−1 (I + Kω ),

(9.42)
Soo = W0 − Wω + (I − Kω )(V0 − Vω )−1 (I − Kω ),
and Seo = Soe = 0.

The Mellin symbols of all the components are explicitly known (see [129]
Sect. 9.1):
cosh[(π − ω)λ]
M (Vω φ)(λ) = V̂ω (λ)φ̂(λ − i) := φ̂(λ − i), ((λ) ∈ (0, 1),
λ sinh πλ
(9.43)
M (Wω φ)(λ) = Ŵω (λ + i)φ̂(λ + i) (9.44)

cosh[(π − ω)(λ + i)]
:= −(λ + i) φ̂(λ + i), ((λ) ∈ (−2, 0),
sinh[π(λ + i)]
sinh[(π − ω)λ]
M (Kω φ)(λ) = K̂ω (λ)φ̂(λ) := − φ̂(λ), ((λ) ∈ (−1, 1),
sinh πλ
(9.45)
M (Kω φ)(λ) = K̂ω (λ + i)φ̂(λ), ((λ) ∈ (−2, 0). (9.46)
On an infinite angle the continuity with respect to seminorms is as follows.

Lemma 9.10 Let ρ < β < 1 for ρ := 3/2 − min{ 2π−ω
π
, ω }.
π
Then there holds
|Sφ|H j,1 (Γ ω ) ≤ C|φ|H j+1,2 (Γ ω ) (j ≥ 1)

β β
The constant C > 0 does not depend on j .

Proof We use the representation (9.42) of the Poincaré-Steklov operator and
Lemma 9.9 to handle the norms of the weighted Sobolev spaces. To calculate the
seminorms | · |H j,l (R ) it suffices to concentrate on test functions with compact
β +
support in (0, ∞): Let φ ∈ C0∞ [0, ∞) with supp (1 − φ) ⊂ (0, ∞). Then we
automatically have
supp (φ (j ) ) ⊂ C0∞ (0, ∞), j ≥ l ≥ 1,
and the values φ(x) for x near 0 are not taken into account for calculating the
seminorms | · |H j,l (R ) .
β +
ˆ e , φ o ) ∈ C0∞ (R+ )2 . Then we have

Therefore we take φ =(φ
S φ = (See φ e + Soe φ o , Seo φ e + Soo φ o )T = (See φ e , Soo φ o )T
and it remains to estimate the two components of the right hand side by means of
the seminorm |φ|H j,l (R ) .
β +
Using the representations (9.43)–(9.46) we obtain

M (See φ e )(λ − i)

= M (Wω − W0 )(λ)φ e (λ) + M (I + Kω )(λ)M (V0 + Vω )−1 (λ)M (I + Kω )(λ)φ̂ e (λ)
sinh πλ − sinh(π − ω)λ e
= 2λ φ̂ (λ) =: M (See )(λ)φ̂ e (λ). (9.47)
cosh πλ + cosh(π − ω)λ
Therefore, by Lemma 9.9,

|See φ e |2 j,1 0 |fj (λ)|2 |M (See φ e )(λ)|2 dλ
Hβ (R+ ) ((λ)=1/2−β

sinh πλ − sinh(π − ω)λ 2 e
=4 |fj +1 (λ)|2 |φ (λ)|2 dλ.
((λ)=3/2−β cosh πλ + cosh(π − ω)λ
Here we used the relation |fj (λ)(λ + i)| = |fj +1 (λ + i)|.

Now, cosh πλ + cosh[(π − ω)λ] does not vanish if ((λ) = 3/2 − β <
π
min{ 2π−ω , πω }, i.e. if β > 3/2 − min{ 2π−ω
π
, πω } = ρ.
Thus, since
sinh πλ − sinh(π − ω)λ

is bounded for |%(λ)| → ∞ when avoiding the roots of the denominator, there holds

sinh πλ − sinh(π − ω)λ 2 e
2
|See φ e |2 j,1 0 |fj +1 (λ)| |φ (λ)|2 dλ
Hβ (R+ ) ((λ)=3/2−β cosh πλ + cosh(π − ω)λ

≤C |fj +1 (λ)|2 |φ e (λ)|2 dλ 0 |φ e |2 j+1,2
((λ)=3/2−β Hβ (R+ )
for β > ρ and j ≥ 1.

Analogously, we obtain

sinh πλ + sinh(π − ω)λ 2 o
|Soo φ o |2 0 |fj +1 (λ)|2 |φ (λ)|2 dλ
cosh πλ − cosh(π − ω)λ
j,1
Hβ (R+ ) ((λ)=3/2−β

≤C |fj +1 (λ)|2 |φ o (λ)|2 dλ 0 |φ o |2 j+1,2
((λ)=3/2−β Hβ (R+ )
for β > ρ and j ≥ 1.

Therefore
|S φ|2 j,1 ≤ C(|φ e |2 j+1,2 + |φ o |2 j+1,2 )

Hβ (Γ ω ) Hβ (R+ ) Hβ (R+ )
≤ c(|φ− |2 j+1,2 + |φ+ |2 j+1,2 ) = c|φ|2 j+1,2

Hβ (R+ ) Hβ (R+ ) Hβ (Γ ω )
and the assertion of the lemma is proved.

Using this lemma we obtain the continuity of S within countable normed spaces on
the whole polygon Γ .
Theorem 9.9 For ρ < β < 1 with ρ := 3/2 −min{ 2π−ω
π
, ω } let φ
π
∈ Bβ2 (Γ ). Then
there holds Sφ ∈ Bβ (Γ ).
1
Proof Let φ ∈ Bβ,C,d

2 (Γ ). Due to the definition of the countable normed spaces
in (9.36), (9.37) and the respective norms (9.34) we have to consider the global
norm SφL2 (Γ ) and the seminorms |Sφ|H j,1 (Γ i ) (j ≥ 1) for i = 1, . . . , J and
β k
k = 1, 2. The continuity with respect to the L2 -norm is proved by Proposition 9.2.
To show the continuity with respect to the seminorms (9.35) we need a partition of
2,L
unity which exists due to Lemma 9.8 for each Bβ,C,d (Γ ) (L ≥ 2), cf. (9.38). Let
χi ∈ C ∞ (Γ ), i = 1, . . . , J , such that supp χi ⊂ Γi ∪ {ti } ∪ Γi+1 and J
i=1 χi =1
and
χi φ ∈ B 2,L (Γi ∪ {ti } ∪ Γi+1 )

β,C̃,d̃
(χi φ is supposed to be extended by 0 outside supp χi ). Then for β > ρ, Lemma 9.10
yields
|Sχi φ|H j,1 (Γ ∪{t }∪Γ ≤ c|χi φ|H j+1,2 (Γ ∪{t }∪Γ (9.48)
β i i i+1 ) β i i i+1 )
≤ cC̃ d̃ L−2 (j − 1)!, j = 1, . . . , L − 1
where the second inequality is caused by the regularity of χi φ and definition (9.39).
Now, the already known boundedness of Sχi φL2 (Γ ) and the estimate (9.48) yield
due to (9.39) the local regularity
Sχi φ ∈ B 1,L−1 (Γi ∪ {ti } ∪ Γi+1 ). (9.49)

β,cC̃,d̃
Again, there exists a partition of unity {ζi ; i = 1, . . . , J } for Sχi φ and the index
L − 1. Due to (9.49) there holds
1,L−1
ζi Sχi φ ∈ Bβ,C ,d (Γi ∪ {ti } ∪ Γi+1 ) (9.50)
and due to the analyticity of the kernel of the Poincaré-Steklov operator aside the
diagonal x = y we also have
1,L−1
ζj Sχi φ ∈ Bβ,C ,d (Γ \ Γi ∪ {ti } ∪ Γi+1 ) (j = i). (9.51)
Putting (9.50) and (9.51) together for each of the corners {ti } we obtain

m
m
1,L−1
Sφ = ζj Sχi φ ∈ Bβ,C % ,d % (Γ ) (9.52)
j =1 i=1
where C % and d % are the largest numbers of the different C s and d s, respectively.
Since the constants C and d do not depend on the parameter L and the partitions
of unity {ζi } and {χi } corresponding to L, eq. (9.52) finally yields together with the
representation (9.38)
1
Sφ ∈ Bβ,C % ,d % (Γ ).

We now investigate the inverse of the Poincaré-Steklov operator. First let us

formulate the local regularity result which corresponds to the continuity given by
Lemma 9.10.
Lemma 9.11 Let ρ = 3/2 − min{ 2π−ω
π
, πω }. For φ ∈ H 1 (Γ ω ) such that Sφ ∈
j −1,1 j,2
Hβ (Γ ω ) for j ≥ 2 there holds φ ∈ Hβ (Γ ω ) for ρ < β < 1 and
|φ|H j,2 (Γ ω ) ≤ C|Sφ|H j−1,1 (Γ ω ) .

β β
The constant C does not depend on j .

Proof Again we use the representation of functions on Γ ω by their even and odd
parts on Γ − and Γ + and the induced matrix representation of the Poincaré-Steklov
operator as in the proof of Lemma 9.10, cf. (9.42). With regard to (9.47) there holds

φ̂ e (λ) = M (See φ e )(λ − i), ((λ) ∈ (−1, 1).
2λ sinh πλ − sinh(π − ω)λ
Noting that |fj (λ)/λ| = |fj −1 (λ − i)| we obtain by Lemma 9.9 for β ∈ (1/2, 1)

|φ e |2 j,2 0 |fj (λ)|2 |φ̂ e (λ)|2 dλ (9.53)
Hβ (R+ ) ((λ)=3/2−β
cosh π(λ + i) + cosh(π − ω)(λ + i) 2
1
= |fj −1 (λ)|2 |M (See φ e )(λ)|2 dλ.
4 sinh π(λ + i) − sinh(π − ω)(λ + i)
((λ)=1/2−β
Since
cosh π(λ + i) + cosh(π − ω)(λ + i)

sinh π(λ + i) − sinh(π − ω)(λ + i)
is bounded for %(λ) → ±∞ provided the denominator does not vanish at the
horizontal strip under consideration, we obtain by (9.53) and Lemma 9.9
|φ e |2 j,2 0
Hβ (R+ )
cosh π(λ + i) + cosh(π − ω)(λ + i) 2

|fj −1 (λ)|2 |M (See φ e )(λ)|2 dλ
sinh π(λ + i) − sinh(π − ω)(λ + i)
((λ)=1/2−β

≤C |fj −1 (λ)|2 |M (See φ e )(λ)|2 dλ 0 |See φ e |2 j−1,1
((λ)=1/2−β Hβ (R+ )
for β > 3/2 − π/(2π − ω). Analogously, we obtain for the odd part
cosh πλ − cosh(π − ω)λ

φ̂ o (λ) = M (Soo φ o )(λ − i), ((λ) ∈ (−1, 1)
2λ sinh πλ + sinh(π − ω)λ
and therefore,
|φ o |2 j,2 0
Hβ (R+ )
cosh π(λ + i) − cosh(π − ω)(λ + i) 2

|fj −1 (λ)|2 |M (Soo φ o )(λ)|2 dλ
sinh π(λ + i) + sinh(π − ω)(λ + i)
((λ)=1/2−β

≤C |fj −1 (λ)|2 |M (Soo φ o )(λ)|2 dλ 0 |Soo φ o |2 j−1,1
((λ)=1/2−β Hβ (R+ )
for β > 3/2 − π/ω.

j,2
Altogether, since φ ∈ H 1 (Γ ω ) by assumption, we proved that φ e ∈ Hβ (R+ )
j,2 j,2
and φ o ∈ Hβ (R+ ) for j ≥ 2 and therefore we have φ ∈ Hβ (Γ ω ) and the proof
of the lemma is finished.

Now we use again the partitions of unity to prove the regularity of the Poincaré-
Steklov operator on the whole polygon.
Theorem 9.10 Let ρ = 3/2 − min{ 2π−ω
π
, π ; j = 1, . . . , J } and ρ < β < 1.
j ωj
Then there holds φ ∈ Bβ2 (Γ ) if Sφ ∈ Bβ1 (Γ ) with Γ Sφ ds = 0 where φ is unique
up to a constant.
Proof The proof is analogous to the proof of Theorem 9.9. The regularity with
respect to the global Sobolev norms is given by Proposition 9.2. The boundedness
with respect to the seminorms | · |H j,l (Γ j ) follows from Lemma 9.11. Here again, we
β k
have to use a partition of unity as in the proof of Theorem 9.9.

For Mellin convolution equations on an interval (second kind integral equations)
stability and exponential convergence in the Lq -norm 1 ≤ q ≤ ∞ for Galerkin and
Collocation methods are proved for piecewise polynomials on geometrically refined
meshes in [154].
For the use of Mellin techniques for integral operators over polyhedral domains
see [152, 153, 351, 367].
In [125] Mellin techniques are applied to first kind integral equations for linear
elasticity in polygonal domains.
Chapter 10
A-BEM
First in this chapter we give a general framework of adaptive Petrov–Galerkin

methods for the solution of operator equations in Banach spaces. This approach
is made precise in the application to Symm’s integral equation. Then we present
more general adaptive BEM. Here we use the residual error estimator and prove
reliability and efficiency in 2D. Finally we analyze the hierarchical error estimator
and demonstrate its applicability in two-level adaptive BEM for scalar and vector
boundary value problems. Special emphasis is given to the 3D case for the weakly
singular integral equation (Sect. 10.3) and for the hypersingular integral equation
(Sect. 10.4). In Sect. 10.5 we present a two-level adaptive BEM for the weakly
singular operator and the h-version on surface pieces. In Sect. 10.6 based on a
two-level subspace decomposition for the p-version BEM we give hierarchical
error estimators for the hypersingular integral operator on curves. Finally recent
developments on the convergence of the adaptive BEM for the h-version are given
in Sect. 10.7.
We do not want to miss to mention further work on adaptive boundary elements.
Faermann in [170] provides local a posteriori error indicators for the Galerkin
discretization of boundary integral equations of the Babuşka-Rheinboldt type. These
error indicators are for a wide class of integral operators reliable on special meshes
(like K−meshes) and efficient for arbitrary meshes. The proof of efficiency is
problematic for BEM: Carstensen derives efficiency only for uniform meshes (see
Sect. 10.2.2) and Mund et al. show efficiency and reliability only for uniform meshes
and under the assumption of a saturation condition (see Sect. 10.5). Faermann also
used local double-integral seminorms as estimators in [171, 172] and their relation
to weighted residual error indicators and multilevel (hierarchical) error indicators
shown in the survey paper [78]. The earliest suggestions for error indicators
and adaptive BEM [348, 434] used the concept of an influence index. They are
equivalent to Faermann’s Babuşka-Rheinboldt error indicators. Super convergence
properties and gradient recovering techniques for the purpose of adaptive mesh
refinements are used in [375] and [88]. For the adaptive method there are also some

334 10 A-BEM
different concepts as in [369–371, 389, 390]. In the following we concentrate on

estimators of residual type and hierarchical type and present some of the respective
analysis.
10.1 General Frame for A Posteriori Error Estimates for

Boundary Element Methods
In this section we consider the analogue of adaptive feedback algorithms developed

within finite element methods and present their analogue within boundary element
methods. This can be generally described as follows (see [92]):
Let X, Y be Banach spaces and let the operator A : X → Y be linear, bounded
and bijective. We want to approximate the continuous solution u := A−1 f ∈ X of
the problem
Au = f for f ∈ Y given
by uh ∈ Xh ⊂ X such that:
t, Auh = t, f ⇐⇒ 0 = t, f − Auh ∀ t ∈ Th ⊂ Y ∗ . (10.1)
If we define the residual R by R := f − Auh ∈ Y, and assume that it can be

computed, at least numerically, we have by the Banach’s inverse mapping theorem
that A−1 is bounded and there holds
u − uh X = A−1 RX ≤ |||A−1 |||L(Y ;X) · RY (10.2)
Let uh be obtained by some Galerkin procedure such that there exists a finite
dimensional subspace Th of Y ∗ , the dual of Y and let , Y ∗ ×Y denote the dual
pairing between Y ∗ and Y . Then
0 = th , R Y ∗ ×Y ∀ th ∈ Th ⊂ Y ∗ (10.3)
By a consequence of the Hahn-Banach theorem there exists ρ̃ ∈ Y ∗ with norm 1

and RY = ρ̃, R Y ∗ ×Y . Setting ρ = RY ρ̃ shows the following assertion.
Lemma 10.1 There exists some ρ ∈ Y ∗ with
ρ2Y ∗ = R2Y = ρ, R Y ∗ ×Y .
Choosing some ρ as in the lemma, we have with (10.3)
R2Y = ρ, R Y ∗ ×Y = inf ρ − th , R Y ∗ ×Y ≤ RY · inf ρ − th Y ∗ ,

th ∈Th th ∈Th
10.1 General Frame for A Posteriori Error Estimates 335
yielding
RY ≤ inf ρ − th Y ∗ . (10.4)

th ∈Th
Comparing this with (10.2) yields the abstract a posteriori error estimate
u − uh X ≤ |||A−1 |||L(Y ;X) · inf ρ − th Y ∗ .

th ∈Th
Combining interpolation with this estimate gives:

Theorem 10.1 Let X, Y0 , Yθ , Y1 be Banach spaces with Y1 ⊆ Yθ ⊆ Y0 . Let cθ
be some positve constant such that for some θ ∈ (0, 1)
yYθ ≤ cθ · y1−θ
Y0 · yY1
θ
∀y ∈ Y1 . (10.5)
Let A : X → Yθ be linear, bounded and bijective. Let f ∈ Yθ , u := A−1 f , and

let uh ∈ X with R := f − Auh ∈ Y1 satisfying
0 = th , R Y0∗ ×Y0 ∀ th ∈ Th , (10.6)
where , Y0 ×Y0 denoting the dual pairing between Y0∗ and Y0 , Th ⊆ Y0∗ . Let further
ρ ∈ Y0∗ be defined as in Lemma 10.1 (with Y0 replacing Y ), i.e. with
ρ2Y ∗ = R2Y0 = ρ, R Y0∗ ×Y0 . (10.7)

0
Then there holds:
u − uh X ≤ cθ · |||A−1 |||L(Yθ ;X) · RθY1 · inf ρ − th 1−θ

Y∗
.
th ∈Th 0
Proof Using (10.2) with Yθ replacing Y and estimating RYθ via (10.5) we obtain
u − uh X ≤ cθ · |||A−1 |||L(Yθ ;X) · RθY1 · R1−θ

Y0 .
By Lemma 10.1 we have some ρ satisfying (10.7). Using (10.4) with Y0 replacing
Y concludes the proof.

We now want to apply the above Theorem to Symm’s integral equation and
explicitly determine ρ for this example.
336 10 A-BEM
10.1.1 Symm’s Integral Equation
Given a bounded Lipschitz domain Ω ⊆ Rn with boundary Γ the Dirichlet

problem for the Laplacian is related with the Symm’s integral equation for the
unknown density φ on Γ :
V φ(x) = g(x), x∈Γ (10.8)
with the weakly singular operator V .

Let Th = Sh◦ ⊆ L2 (Γ ) denote the vector space of piecewise polynomials with
respect to a ‘triangulation’ of the boundary Γ = ∪N j =1 Γj , where the ‘elements’
Γ1 , . . . , ΓN are either equal or have at most one common point or side, respectively.
For x ∈ Γj we define the piecewise constant function

Γ −→ [0, ∞)
h:
x → |Γj | =: hj ,
|Γj | > 0 denoting the element size. Let φh ∈ Th denote the Galerkin solution
of (10.8), i.e.
th , R = 0 ∀ th ∈ Th (10.9)
where R = g − V φh . Assume that g ∈ H 1 (Γ ) so that, according to Th ⊆ L2 (Γ ),

R = V (φ − φh ) ∈ H 1 (Γ ) due to the mapping properties of the single-layer
potential. The following result gives an a posteriori error estimate using the residual
R on the right hand side.
Theorem 10.2 ([92])
Under the above assumptions, there holds for 0 ≤ s ≤ 1 :
φ − φh H −s (Γ ) ≤ |||V −1 |||L(H 1−s (Γ );H −s (Γ )) · R1−s

H 1 (Γ )
· h · ∇Γ RsL2 (Γ ) .
(10.10)
Proof Note that H 1 (Γ ) ⊆ H 1−s (Γ ) ⊆ L2 (Γ ) = H 0 (Γ ) and (10.5) holds with
cθ = 1, therefore we can apply Theorem 10.1 with A = V , X = H −s (Γ ), Y0 =
L2 (Γ ), Y1 = H 1 (Γ ), θ = 1 − s and Yθ = H 1−s (Γ ) . Note that ρ = R ,
identifying Y0∗ withY0 by the Riesz representation theorem, satisfies (10.7) and there
holds
inf R − th L2 (Γ ) ≤ h · ∇Γ RL2 (Γ )

th ∈TN
by standard interpolation and approximation arguments. Hence Theorem 10.1 yields

the assertion.

10.2 Adaptive Boundary Element Methods 337
We can obtain an adaptive feedback algorithm as follows (see [92]): For a given
‘triangulation’ of the boundary as above we can compute an approximation of the
contribution aj of one element Γj by numerical integration. Defining for Symm’s
integral equation aj := R L2 (Γj ) , we have the error estimate in the energy-norm:
⎛ ⎞1 ⎛ ⎞1
4 4

N
N
φ − φh 1 ≤ c·⎝ aj2 ⎠ · ⎝ h2j aj2 ⎠ .
H − 2 (Γ )
j =1 j =1
The mesh may be steered be the following, adaptive feedback algorithm where 0 ≤
θ ≤ 1 is a global parameter:
Algorithm Given some coarse, e.g. uniform mesh, refine it successively by halving
some of the elements due to the following rule. For any triangulation define
a1 , . . . , aN as above and refine the element Γj if and only if
hj · aj ≥ θ · max hk · ak .
1≤k≤N
Note for θ = 0 we have a uniform triangulation whereas the number of refined

elements decreases with increasing θ .
Remark 10.1 The above abstract setting allows also to handle hp-adaptive bound-
ary element methods. In [82] the following estimates are derived for the hypersin-
gular operator W and the weakly singular operator V , respectively; see [265] for
a different approach with a γ -shape regular mesh on Γ and a γp -shape regular
polynomial degree distribution,
u − uN H 1/2 (Γ ) ≤ (h/p)1/2 RN L2 (Γ )
for RN = g − W uN , where W u = g on Γ ,
φ − φN H −1/2 (Γ ) ≤ (h/p + 1)1/2 ∇Γ RN L2 (Γ )
for RN = f − V φN , where V φ = f on Γ .
10.2 Adaptive Boundary Element Methods
Let A : H → H ∗ with H ⊆ H m (Γ ), m ∈ R, H ∗ the dual of H , be some

pseudodifferential operator which is bounded, linear and positive definite on the
closed subspace H of H m (Γ ), i.e., we have a constant α > 0 with
< Au, u > ≥ αu2H m (Γ ) for all u ∈ H ⊆ H m (Γ ).

338 10 A-BEM
Here, <, > extends the L2 (Γ ) scalar product to the duality between the Sobolev
spaces H and H ∗ . Due to the Lax–Milgram lemma we then have a unique solution
u ∈ H ⊆ H m (Γ ) of
Au = f (10.11)
for any given right hand side f ∈ H −m (Γ ) ⊆ H ∗ .

For the numerical approximation of u let Γ = ∪N j =1 Γj be partitioned into
N pairwise disjoint elements Γ1 , . . . , ΓN and let Shk denote a finite dimensional
subspace of H ⊆ H m (Γ ) of piecewise polynomials, i.e., vh |Γj is a polynomial of
degree at most k for any vh ∈ Shk and j = 1, . . . , N. Then, the Galerkin equations
< Auh , vh > = < f, vh > for all vh ∈ Shk (10.12)
have a unique solution uh ∈ Shk . Due to the Céa lemma we have
u − uh H m (Γ ) ≤ C · inf u − vh H m (Γ ) (10.13)
vh ∈Shk
where C denotes a generic positive constant which is independent of the data f and
Shk . In [93] we prove a posteriori error estimates of the form

N
u − uh H m (Γ ) ≤ C · aj · h1−r
j (10.14)
j =1
via “local interpolation” for the class of K-meshes where the quotient of the mesh
sizes of two neighbouring elements is uniformly bounded by some constant K > 1.
Let Ω ⊂ R2 be a bounded domain with Lipschitz boundary Γ= and Γ ⊂ Γ= be
a connected piece of Γ=. The Dirichlet problem for the Laplacian is equivalently
related to Symm’s integral equation
V φ(x) = g(x) (x ∈ Γ ) (10.15)
where

1
V φ(x) := − φ(y) ln |x − y|dsy (10.16)
π Γ
Note if cap(Γ ) < 1 (see Section 2.3), then A := V is bijective between H :=

H̃ −1/2(Γ ) and H ∗ = H 1/2(Γ ). Letting m = −1/2, k = 0 , the equations (10.11)
and (10.12), i.e., (10.15) and
< V φh , ψh > = < g, ψh > for all ψh ∈ Sh0 , (10.17)

Sh0 being the piecewise constant functions on Γ = ∪N j =1 Γj , have unique solutions

φ and φh , respectively, and and there holds a quasi-optimal convergence estimate
(10.13).
Remark 10.2 Note that, for given g in H 1 (Γ ) the residual R := g − V φh lies also
in H 1 (Γ ) for φh ∈ L2 (Γ˜ ). Moreover,
< R, ψh >= 0 for all ψh ∈ Sh0 (10.18)
by the Galerkin procedure.

The Neumann problem for the Laplacian is equivalently related to the hypersingular
integral equation
W v(x) = f (x) (x ∈ Γ ), (10.19)
where

1 ∂ ∂
W v(x) := v(y) ln |x − y|dsy (10.20)
π ∂nx Γ ∂ny
is hypersingular and f can be computed from the given normal derivative of the
displacement field u which gives v = u|Γ˜ on Γ .
Thus, A := W is bijective between H = H 1/2(Γ )/R and its dual. Letting
m = 1/2, k = 1, the problems (10.11) and (10.12), i.e., (10.19) and
< W vh , wh > = < f, wh > for all wh ∈ Sh1 ,
have unique solutions v and vh , respectively. Here, Sh1 are continuous piecewise
linear functions on Γ = ∪Nj =1 Γj with support in Γ , e.g., the trial functions vanish at
the endpoints of Γ if Γ is an open arc. Again we have a quasi-optimal convergence
estimate (10.13).
Remark 10.3 Note that, for given f in L2 (Γ ) ∩ H ∗ the residual f − W vh lies also
in L2 (Γ ) for vh ∈ H̃ 1 (Γ ). Moreover,
< f − W vh , wh > = 0 for all wh ∈ Sh1 (10.21)
by the Galerkin procedure.

340 10 A-BEM
10.2.1 Reliability of A Posteriori BEM Error Estimates
With the above notations we have the following a posteriori error estimates.
Their proofs are based on “local interpolation” whereas (10.10) is based on
“global interpolation” (see [92]).
Let K ≥ 1 denote the maximum of all quotients of sizes h1 , . . . , hN of
neighbouring elements Γ1 , . . . , ΓN which describe the discretization within the
Galerkin method.
Theorem 10.3 ([93]) Let g ∈ H 1 (Γ ) and cap(Γ ) < 1. For Symm’s integral
equation we have for 0 ≤ s ≤ 1 with R = ∂s
∂
R
√
N
φ−φh H̃ −s (Γ ) ≤ 2K 1−s V −1 L(H 1−s (Γ );H̃ −s (Γ )) hsj (h2j +1)(1−s)/2R L2 (Γj )
j =1
(10.22)
Theorem 10.4 ([93]) Let f ∈ L2 (Γ ) and cap(Γ ) < 1. For the hypersingular
integral equation we have for 0 ≤ s ≤ 1
√ s
N
v − vh H s ≤ 2K W −1 L(H s−1 ;H s ) · h1−s
j (hj + 1)
2 s/2
· f − W vh L2 (Γj ) .
j =1
(10.23)
The proofs of the above theorems use the following proposition where Sh0 denotes
the piecewise constant functions with respect to a partition Γ1 , . . . , ΓN of Γ =
∪Nj =1 Γj .
Proposition 10.1 ([93]) If f ∈ H 1 (Γ ) satisfies
< f, ψh > = 0 for all ψh ∈ Sh0 , (10.24)
then for 0 ≤ σ ≤ 1 there holds
√
N
f H σ (Γ ) ≤ 2 · Kσ f L2 (Γj ) · h1−σ
j (1 + h2j )σ/2 .
j =1
Proof Since Γjf ds = 0 we have at least one zero yj of the continuous function
f in the interior of Γj , j = 1, . . . , N. Define f˜j ∈ H 1 (Γ ) to be equal to f on the
part of Γ between yj and yj +1 and equal to 0 on the remaining part of Γ (note that
f˜j is continuous, e.g., at yj and hence absolute continuous on Γ and piecewise in
H 1 ; thus f˜j ∈ H 1 (Γ )). Here we set y0 = yN and M := N if Γ is closed; y0 and
yN+1 as the starting point and endpoint of Γ , respectively, and M := N + 1 if Γ is
an open arc. Then, f˜j ∈ H σ (Γ ) and we conclude from the triangle inequality that

M−1
f H σ (Γ ) ≤ f˜j H σ (Γ ) . (10.25)
j =0
Since Γ˜j := supp f˜j ⊆ Γj ∪ Γj +1 , we have by definition of f˜j , and interpolation

of H 1 (Γ ) and L2 (Γ ) [44]
f˜j H σ (Γ ) ≤ f˜j σH 1 (Γ ) · f˜j 1−σ

L2 (Γ )
= f σH 1 (Γ˜ ) · f 1−σ
2 ˜
. (10.26)
j L (Γj )
Note that in any nonempty Γj ∩ Γ˜k =: Γj k (hence k = j or k = j − 1 , because

everywhere else Γj ∩ Γ˜k =: ∅) there is at least one zero of f . Using the fundamental
theorem of calculus gives:
f L∞ (Γjk ) ≤ f L1 (Γjk )
yielding
f L2 (Γjk ) ≤ hj k · f L2 (Γjk )
where hj k := |Γj ∩ Γ˜k | is the length of Γj ∩ Γ˜k . Hence, (10.26) gives
f˜j H σ (Γ ) ≤ f σH 1 (ΓK) f 1−σ Kj )

L2 (Γ
j
-
f˜j H σ (Γ ) ≤ f σH 1 (ΓK) f 1−σ
K) ≤ (f L2 (Γjj ) + f L2 (Γj+1j ) )
L2 (Γ
2
(10.27)
j j
.σ/2
+ (hjj f L2 (Γjj ) + hj +1j f L2 (Γj+1j ) )2
1−σ
· hjj f L2 (Γjj ) + hj +1j f L2 (Γj+1j ) .
Use the following

√ M estimate for positive a, b, α, β: (a + b)2 + (αa + βb)2 ≤
(a 1 + α + b 1 + β 2 )2 , (10.27) shows with hj k ≤ hj
2
/ / σ
f˜j H σ (Γ ) ≤ f L2 (Γjj ) 1 + h2j + f L2 (Γj+1j ) 1 + h2j +1
1−σ
· hj f L2 (Γjj ) + hj +1 f L2 (Γj+1j ) . (10.28)
Now
(a + b)1−s · (αa + βb)s ≤ K 1−s · (aα s + bβ s ) (10.29)
for any a, b, α, β ≥ 0 with 1/K ≤ α/β ≤ K and 0 ≤ s ≤ 1.

342 10 A-BEM
/ /
Now we set α = hj / 1 + h2j and β = hj +1 / 1 + h2j +1 . Since 1/K ≤
hj / hj +1 ≤ K (by the definition of K as the maximum of all quotients of
neighbouring elements)
N
O
α O h2j / h2j +1 − 1
= P1 +
β h2j + 1
belongs to the interval [1/K, K] as well.

Hence, we may use (10.29) and obtain
f˜j H σ (Γ ) ≤ K σ f L2 (Γjj ) h1−σ

j (1 + h2j )σ/2
+ f L2 (Γj+1j ) h1−σ

j +1 (1 + hj +1 )
2 σ/2
. (10.30)
Estimating each summand in (10.25) with (10.30) leads to

N
f L2 (Γjj ) + f L2 (Γjj−1 ) h1−σ
σ
f H σ (Γ ) ≤ K σ j (1 + h2j ) 2
j =1
√ N
≤ Kσ 2 f L2 (Γj ) · h1−σ
j (1 + h2j )σ/2
j =1
which proves the proposition.

Proof (of Theorem 10.3) Due to Remark 10.2 we have that R := g −V φh ∈ H 1 (Γ )
satisfies (10.24). Now Proposition 10.1 leads to
φ − φh H̃ −s (Γ ) = V −1 RH̃ −s (Γ ) ≤ V −1 L(H 1−s (Γ );H̃ −s (Γ )) · RH 1−s (Γ )

√
≤ 2 · K 1−s· V −1 L(H 1−s (Γ );H̃ −s (Γ ))· N
j =1 hj · (hj + 1)
s 2 (1−s)/2 · R
L2 (Γj ) .
This concludes the proof.

Proof (of Theorem 10.4) Due to Remark 10.3 we have that f − W vh ∈ L2 (Γ )
Hence, an application of Proposition 10.1 leads to (see [93])
√
N
RH σ (Γ ) ≤ 2 · Kσ f − W vh L2 (Γj ) · h1−σ
j (1 + h2j )σ/2 .
j =1
∂
Finally, we consider ∂s : H σ → H σ −1 (i.e., differentiation with respect to the
arc length) which is bounded by 1 in the operator norms for σ = 1 and as well for
σ = 0 which follows by duality when ∂s ∂

is defined in the distributional sense. By
interpolation this shows f − W vh H σ −1 ≤ RH σ = RH σ (Γ ) for 0 ≤ σ ≤ 1.
Since W : H s → H s−1 is linear, bounded, and bijective, one concludes this
proof as in the proof of Theorem 10.3.

10.2.2 Efficiency of A Posteriori BEM Error Estimates (2D)
In this section we report on the paper of Carstensen [74]. Let π = {Γ1 , . . . , Γn } be

a partition of the polygon Γ in intervals Γ1 , . . . , Γn . Then

Sπ0 := {vh ∈ L∞ (Γ ) : vh Γ ∈ R ∀j = 1, . . . , N}
j
denotes the linear space of piecewise

constant functions and h(π) ∈ Sπ0 is defined
as the local mesh size, i.e., h(π)Γ := |Γj | := length of Γj .
j
In the following we consider for Symm’s integral equation the Galerkin method
with piecewise constants on quasi uniform meshes on Γ , i.e. meshes for which there
exists a global constant cu such that for all meshes π under consideration,
max h(π)/ min h(π) = max |Γj |/|Γk | ≤ cu (10.31)

j =k
with max h(π) := h(π)L∞ (Γ ) and min h(π) := min{h(π)(x) : x ∈ Γ }.

In [74] Carstensen shows the following efficiency result.
Theorem 10.5 ([74]) If f is continuous and smooth on each side of Γ , there exist
constants c0 , h0 > 0 (depending only on Γ, f and cu ) such that for all partitions π
of Γ with max h(π) < h0 and (10.31), and for s ∈ [0, 1], one has
max h(π)s · Rh L2 (Γ ) ≤ c0 · φ − φh H −s (Γ ) (10.32)
where φ solves Symm’s integral equation V φ = f on Γ and φh denotes its Galerkin

solution; Rh = f − V φh .
Let us first introduce an abstract setting and give some results which are used to
prove the above theorem. Let X1 ⊂ X0 and Y1 ⊂ Y0 be real Banach spaces, and
let Xθ := [X0 , X1 ]θ , Yθ := [Y0 , Y1 ]θ , 0 ≤ θ ≤ 1 be defined by interpolation. We
assume, that there are positive constants cθ,X and cθ,Y such that for all x ∈ X1 and
y ∈ Y1 (10.5) is satisfied:
xXθ ≤ cθ,X ·x1−θ

X0 ·xX1 ,
θ
yYθ ≤ cθ,Y ·y1−θ
Y0 ·yY1 .
θ
(10.33)
Let L(X; Y ) denote the Banach space of linear bounded mappings between the
Banach spaces X and Y , and let · L(X;Y ) be the corresponding operator norm.
344 10 A-BEM

Then, for each Aj ∈ L(Xj ; Yj ), j = 0, 1 with A0 X = A1 , the restriction Aθ =
1
A0 X belongs to L(Xθ ; Yθ ) and
θ
Aθ L(Xθ ;Yθ ) ≤ Aθ 1−θ

L(X0 ;Y0 ) · Aθ L(X1 ;Y1 )
θ
(10.34)
Let A = Aθ be such a mapping and assume, in addition, that Aθ : Xθ → Yθ is

bijective. Then, fix a right-hand side f ∈ Y1 and the solution u ∈ Xθ of
Au = f. (10.35)
We apply the Galerkin method to approximate u. Let Sh ⊂ X1 and Th ⊂ Y0∗ be

finite-dimensional subspaces such that there exists some uh ∈ Sh satisfying (10.1).
We define the residual Rh := f − Auh and the error eh := u − uh . Note that
now there holds the assertion of Theorem 10.1 with Xθ instead of X. We consider a
family of Galerkin methods described by a family of discrete subspaces (Sh : h ∈
I ) of X1 and (Th : h ∈ I ) of Y0∗ , where I is an index set. We assume the three
properties: (i) Approximation property ,(ii) Inverse assumption, (iii) Stability, made
precise in the following.
(i) Approximation property. Assume that the solution u ∈ Xθ of (10.35), also
belongs to X1 . Then, for each h ∈ I , let
E(u, Sh ) := inf{u − vh X1 : vh ∈ Sh } = u − Πh uX1
be the best approximation error in the norm of X1 and let Πh : X1 → Sh

denote a projection such that Πh u is the best approximation in Sh . Furthermore
we define:
u − Πh uX0
F (u, Sh ) :=
u − Πh uX1
(ii) Inverse assumption. For each h ∈ I let
& v '
h X1
G(Sh ) := sup : vh ∈ Sh \{0}
vh X0
(iii) Stability. For each h ∈ I let Ph : X0 → Sh be a projection such that Ph ∈
L(X0 , X0 ) and Ph X ∈ L(X1 , X1 ) with norms
1
& Ph vX '

j
Ph j := sup : v ∈ Xj \{0} , j = 0, 1
vXj
Then for h, H ∈ I define
E(u, SH )
δ(u, Sh , SH ) := · 1 + cθ,X · PH θ1 [F (u, SH )G(SH )PH 0 ]1−θ
E(u, Sh )
(10.36)
Theorem 10.6 ([74]) Let A ∈ L(X1 , Y1 ), h, H ∈ I with Sh ⊆ SH ; consider

u, uh ∈ X1 , eh := u − uh , Rh := Aeh ∈ Y1 , and assume (i),(ii),(iii). If
δ(u, Sh , SH ) < 1, then
PH 1−θ
0 · PH θ1
Rh Y1 ≤ · AL(X1 ,Y1 ) · G(SH )1−θ · eh Xθ (10.37)
1 − δ(u, Sh , SH )
The proof in [74] is based on two further corollaries, below.

In case if Sh are spline function spaces E(u, Sh ), F (u, Sh ) and G(Sh ) can be
bounded. Now: Let (Sh : h ∈ I ) be a family of subspaces of X, where the index h
is a positive parameter, say, I ⊂ (0, 1).Suppose Sh ⊂ SH ∀h, H ∈ I with H < h
and that ∪h∈I Sh is dense in X1 . Suppose that there exist positive constants cα , cβ , cp
and real constants α, β such that ∀h ∈ I
F (u, Sh ) ≤ cα · hα (10.38)
G(Sh ) ≤ cβ · hβ (10.39)
Ph 1−θ
0 · Ph θ1 ≤ cp (10.40)
Corollary 10.1 Assume (10.38)(10.39)(10.40) and α + β ≥ 0. Define
c1 := 2 + 2cθ,X · cp · cα1−θ · cβ1−θ , c2 := 2cp · AL(X1 ;Y1 ) · cβ1−θ
Then, for each h ∈ I , we can find H ∈ I with
1
E(u, SH ) ≤ E(u, Sh ) H <h (10.41)
c1
and we have Rh Y1 ≤ c2 · H β(1−θ) · eh Xθ

As chosen in Corollary 10.1, H depends highly on h and we need more information
on E(u, Sh ) to control this in (10.41).
Corollary 10.2 In addition to the assumption of Corollary 10.1 let there exist
constants η, q, h0 , 0 < η, q < 1,such that for all h ∈ I with h < h0 we have
E(u, Sη·h ) ≤ q · E(u, Sh ) η·h ∈I (10.42)
Then, there exists c0 > 0 such that for all h ∈ I with h < h0
Rh Y1 ≤ c0 · hβ(1−θ) · eh Xθ (10.43)

346 10 A-BEM
Lemma 10.2 [74] For any vh ∈ Sh , h ∈ I , we have
vh X1 ≤ Ph 1−θ

0 Ph 1 · G(Sh )
θ 1−θ
· vh Xθ
and
u − Πh uXθ ≤ cθ,X · F (u, Sh )1−θ · u − Πh uX1
Symm’s integral equation for a polygon Γ , V φ = f has a unique solution φ ∈

L2 (Γ ) for a given right hand side f ∈ H 1 (Γ ) if cap(Γ ) < 1, which is assumed
now. We are interested in its L2 (Γ )-best approximation error:
& '
E(φ, Sh ) = min φ − ψh L2 (Γ ) : ψh ∈ Sπ0 (Γ ) = φ − Πh φL2 (Γ ) (10.44)
where, Πh is the orthogonal projection onto Sπ0 (Γ ) in L2 (Γ ) .

Proposition 10.2 ([74]) Provided φ is not constant, there exists positive constants
γ , h0 , cγ and cγ (depending only on Γ, f and cu in (10.31) ) such that 0 < γ ≤ 1
and either
cγ ≤ max h(π)−γ · E(φ, Sh ) ≤ cγ (10.45)
or γ = 1/2 and
cγ ≤ − max h(π)−1/2 · log−1/2 (max h(π)) · E(φ, Sh ) ≤ cγ (10.46)
holds for all meshes π with max h(π) < h0 and satisfying (10.31).
Proof The solution φ of V φ = f has the form (see Section 7.1)

m
φ(x) = φ0 (x) + cj · φj (x), x∈Γ (10.47)
j =1
where φ0 ∈ H 2 (Γ ) and the real constants cj depend on f . The singular functions

φj depend on the corners of the polygon with interior angle ωj as follows:
φj (x) = r βj · χj (x), βj + 1 = kj π/ωj , kj integer ≥ 0
or
φj (x) = r βj · log(r) · χj (x), βj + 1 = kj π/ωj integer, βj ≥ 1
The proof in [74] is split into several steps. φ is approximated by Sh0 (0, 1) on a quasi-
uniform mesh on (0, 1) described by a partition 0 = x0 < x1 < x2 < · · · < xn <

xn+1 = 1. Define mj = h−1
xj+1
j · xj φ(x)dx = Πh φ
(xj ,xj+1 )
and hj := xj +1 − xj
for j = 0, . . . , n.
Let φ(x) = x β for x ∈ (0, 1), and −1/2 < β < 1/2 , β = 0. Let 0 ≤ a <
a + h ≤ 1, and consider the error φ − mL2 (a,a+h) , where φ is approximated by
the constant m = h−1
a+h
a φ(x)dx. One obtains:
φ − m2L2 (a,a+h) = a 2β+1 · η(h/a) (a > 0) (10.48)
β2
φ − m2L2 (0,h) = (a = 0), (10.49)
(1 + β 2 )(2β + 1)
where
(1 + δ)2β+1 − 1 [(1 + δ 2 )β − 1]2

η(δ) := − (δ > 0)
2β + 1 δ(1 + β)2
yielding
η(δ) = c1 · δ 3 + hot (δ) (10.50)
with a positive constant c1 (depending only on β > −1/2 ) and hot (δ) denoting
higher order terms in δ. Moreover, one can conclude from (10.50) and (10.48) that
c2 ≤ η(δ) · δ −3 ≤ c3 ∀δ ∈ (0, cu ] (10.51)
with constants cj depending on β and cu . Since Γ = ∪Γj for j = 0, . . . , n and

using (10.48) and (10.49) with a = xj gives
2β+1 β2 2β+1n
φ − Πh φ2L2 (Γ ) = h0 · + xj · η(hj /xj ) (10.52)
(1 + β 2 )(2β + 1)
j =1
⎛ ⎞
n
2β+1
2β+1 ⎝ β2 xj
= h0 + · η(hj /xj )⎠
(1 + β )(2β + 1)
2 h0
j =1
Estimating with the Riemann’s Zeta function ζ gives

n ∞

j 2β−2
≤ ζ (2 − 2β) = j 2β−2
j =1 j =1
and yields (10.45) with γ = β + 1/2. Similar arguments hold for β = 1/2 and for
φ(x) = x β log x , β ≥ 1. See [74] for details.

348 10 A-BEM
Proposition 10.3 There holds F (φ, Sh ) ≤ max h(π) with X0 = H −1 (Γ ), X1 =

L2 (Γ ).
Proof Using η − Πh ηL2 (Γ ) ≤ max h(π) · η L2 (Γ ) for all η ∈ H 1 (Γ ) one
obtains
Γ (φ − Πh φ)(η − Πh η)ds
φ − Πh φH −1 (Γ ) = sup ≤ max h(π )·φ − Πh φL2 (Γ )
η∈H 1 (Γ ) ηH 1 (Γ )
yielding the assertion.

Next we define
∂wh
Sπ1 (Γ ) := {wh ∈ H 1 (Γ ) : ∈ Sπ0 (Γ )} (10.53)
∂s
the linear space of continuous and piecewise linear functions with respect to a mesh
π.
The following result shows G(Sh ) ≤ C/ h.
Proposition 10.4 [74] There exists a constant c > 0 such that for all meshes π
with max h(π) < 1,
wh H 1 (Γ ) ≤ c · min h(π)−1 · wh L2 (Γ ) ∀ wh ∈ Sπ1 (Γ ), (10.54)
ψh L2 (Γ ) ≤ c · min h(π)−1 · ψh H −1 (Γ ) ∀ wh ∈ Sπ1 (Γ ). (10.55)
Proof We already know the inverse inequality (10.54), which also can be easily
proved by direct calculations on each element. Let I (f ) be defined ny integrating f
along Γ with respect to the arclength, then I : Sπ0 (Γ ) ∩ L20 (Γ ) → Sπ1 (Γ ) ∩ L20 (Γ )
is an isomorphism and (10.55) follows from (10.54).

For each ψ ∈ H −1 (Γ ) define Ph ψ ∈ Sπ0 (Γ ) by
∂ 1
Ph ψ := ψ 0 + Π I (ψ − ψ 0 )
∂s h
for ψ 0 := |Γ |−1 Γ ψds, where Πh1 is the L2 -Projection onto Sπ1 (Γ )

Proposition 10.5 [74] The operator Ph is a projection onto Sπ0 (Γ), which is
bounded as a mapping between H −1 (Γ ) and H −1 (Γ ) or between L2 (Γ ) and
L2 (Γ )
Now let us prove Theorem 10.5
Proof With the above notation and the Propositions 10.3, 10.4 and 10.5 we get the
properties (10.38) , (10.39) and (10.40), where the index parameter h is identified
with max h(π) for a mesh π satisfying (10.31). For a given mesh π, let η = 1/k for
10.3 The Weakly Singular Integral Equation in 3D 349
an integer k and define a new mesh by dividing each element Γj of π in k pieces

of length |Γj |/k. Then, the new mesh also satisfies (10.31). Moreover, according to
Proposition 10.2 we obtain (10.42) of Corollary 10.2 with some q which depends
on k, cγ , cγ and γ . If we choose k large enough, we can obtain 0 < q < 1. Note
that η depends only on cγ , cγ and γ . Application of Corollary 10.2 concludes the
proof.

A corresponding result holds for the hypersingular integral equation W v = g on
Γ when the Galerkin solution vh ∈ Sπ1 (Γ )∩H01 (Γ ) is computed on a quasi uniform
mesh: Assume g is smooth on each side of the polygon. Then there exist constants
c0 , h0 > 0 (depending on Γ , g and cu ) such that for all partitions with (10.31) and
for 0 < s < 1 there holds
max h(π)1−s · Rh L2 (Γ ) ≤ c0 · v − vh H s (Γ ) . (10.56)
10.3 The Weakly Singular Integral Equation in 3D
Here we report from [87]. Let us consider Symm’s integral equation which is
equivalent to interior or exterior Dirichlet problems for the Laplacian in a bounded
Lipschitz domain Ω ⊂ R3 with boundary ∂Ω or on the open surface Γ ⊂ ∂Ω:
Given f find ψ with

1 ψ(y)
V ψ(x) := dsy = f (x) (x ∈ Γ ). (10.57)
4π Γ |x − y|
A Galerkin discretisation provides ψN and a partition T = {Γ1 , . . . , ΓN } of Γ in

elements Γ1 , . . . , ΓN with mesh sizes h1 , . . . , hN with the property that the residual
RN ,
RN (x) := f (x) − V ψN (x) (x ∈ Γ ),
satisfies a Poincaré inequality on Γj , i.e., there holds
RN L2 (Γj ) ≤ C(Γj ) ∇RN L2 (Γj ) (10.58)
Piecewise constant test functions cause Γj RN ds = 0 and so (10.58) follows with

C(Γj ) ≤ diam (Γj ). Following [87] we show below that (10.58) is sufficient for an
a posteriori error estimate with localized residuals

N
1/2
ψ − ψN H̃ −α (Γ ) ≤ c(α, T ) j ∇RN L2 (Γ
h2α 2
, (10.59)
j)
j =1
350 10 A-BEM
where ∇ is the surface gradient on Γ , (0 ≤ α ≤ 1) (see also Theorem 10.3 in

Sect. 10.2.1 for the two-dimensional situation). The upper bound consists of a sum
of computable residuals ηj := hαj ∇RN L2 (Γj ) , which serve as error indicators
in an adaptive mesh refinement algorithm (Algorithm A) in Sect. 10.3.1. The
underlying meshes are shape-regular and locally uniform.
Next a localisation is performed by multiplication with functions from a partition
of unity with local supports.
Definition 10.1 Let Ω be a bounded Lipschitz domain in R3 . A finite partition
of unity of ∂Ω is a finite sequence Φ := (ϕ1 , . . . , ϕM ) of Lipschitz functions
ϕ1 , . . . , ϕM : ∂Ω → R such that on ∂Ω
1 = ϕ1 + . . . + ϕM and ϕ1 , . . . , ϕM ≥ 0. (10.60)
The overlap K(Φ) is defined by (card{S} denotes the number of elements in a set S)
K(Φ) := max card{k ∈ {1, . . . , M} : ϕk ϕj = 0 on ∂Ω}.

j =1,...,M
K(Φ) may be much smaller than M, even bounded, while M is increasing to

infinity as the mesh-size tends to zero. Hence, one distributes ϕ1 , . . . , ϕM into a
minimal number K ≤ K(Φ) of groups to apply the following lemma (Lemma 7.3
in Section 7.4) in order to derive Theorem 10.7, below.
Lemma 10.3 ([405, 423]) Let f1 , . . . , fn ∈ H α (∂Ω), 0 ≤ α ≤ 1, such that
fj fk = 0 on ∂Ω whenever 1 ≤ j < k ≤ n. Let ωj := interior(supp fj ) satisfy
ωj = supp fj . Then

n
n
fj 2H α (∂Ω) ≤ C1 fj 2H α (ωj ) .
j =1 j =1
The constant C1 depends on ∂Ω but does not depend on fj or on n.

Furthermore one needs the following result.
Lemma 10.4 ([87]) Let Φ be a finite partition of unity of ∂Ω with overlap K(Φ).
Then there exists a partition of {1, . . . , M} into K ≤ K(Φ) non-empty subsets
M1 , . . . , MK ,

K
Mj = {1, . . . , M} and Mj ∩ Mk = ∅ if j = k (j, k = 1, . . . , K),
j =1
such that, for all ∈ {1, . . . , K} and j, k ∈ M with j = k,
ϕj ϕk = 0 on ∂Ω. (10.61)
Theorem 10.7 Let Γ be a connected subpiece of ∂Ω and let Φ be a finite partition

of unity of ∂Ω with overlap K(Φ). Then, for any f ∈ H α (∂Ω), 0 ≤ α ≤ 1, we
have

M
f 2H α (Γ ) ≤ K(Φ) f ϕj 2H α (ωj ) . (10.62)
j =1
Let Φ be a finite partition of unity on ∂Ω with overlap K(Φ) and let ωj the
interior of supp ϕj and dj := width(ωj ) for each j ∈ {1, . . . , M}. Then, for 0 <
α < 1, Γ ⊂ ∂Ω, and f ∈ H 1 (∂Ω), we have

M
2(1−α)
f 2H α (Γ ) ≤ K(Φ) dj (1 + dj2 )α ∇(ϕj f )2L2 (ω ) . (10.63)
j
j =1
Next, let ψN ∈ L2 (Γ ) such that RN := f − V ψN satisfies Γk RN ds = 0 for

each element Γk with diameter hk of a partition T = {Γ1 , . . . , ΓN } of Γ . Suppose
that the supports of hat functions ϕj are matched exactly by a finite number of
elements. Then, the Galerkin condition Γk RN ds = 0 (for all k = 1, . . . , N)
implies ωj RN ds = 0 (for all j = 1, . . . , M).
Corollary 10.3 On a locally uniform mesh there exists a constant C > 0 that
depends only on 0 < α < 1, Γ , ∂Ω, and the shape (not the size) of the elements and
patches such that for any ψN ∈ L2 (Γ ) and RN := f − V ψN with ωj RN ds = 0
for all j = 1, . . . , M, we have

N
Ψ − ΨN 2H̃ −α (Γ ) ≤ C j ∇RN L2 (Γ ) .
h2α 2
(10.64)
j
j =1
Proof First with ψ − ψN = V −1 RN and the boundedness of V −1 we observe that

(10.63) implies: There exists a constant C > 0 that depends only on 0 < α < 1, Γ ,
and ∂Ω such that for any ψN ∈ L2 (Γ ) and RN := f − V ψN we have

N
Ψ − ΨN 2H̃ −α (Γ ) ≤ C dj2α ∇(ϕj RN )2L2 (ω ) . (10.65)
j
j =1
Furthermore, a Poincaré inequality gives
RN 2L2 (ω ) ≤ C(ωj )∇RN 2L2 (ω ) .

j j
352 10 A-BEM
Thus
∇(ϕj RN )2L2 (ω ≤ RN ∇ϕj 2L2 (ω ) + ϕj ∇RN 2L2 (ω

j) j j)
≤ Lip(ϕj )RN 2L2 (ω ) + ∇RN 2L2 (ω ) ≤ C∇RN 2L2 (ω ) .

j j j
since Lip(ϕj ) C(ωj ) ≤ C for a locally uniform mesh. Here Lip(ϕj ) denotes the
Lipschitz constant of the function ϕj .

The above corollary motivates the isotropic error indicator
1/2
μj := hj ∇RN L2 (ωj ) (10.66)
since C( N 2 1/2 is a computable upper error bound with respect to the energy
j =1 μj )
norm.
Open surfaces yield singularities near the edge (see Subsection 7.3.1) which limit
the regularity of the exact solution, so in general ψ ∈/ L2 (Γ ). Here an anisotropic
error indicator such as
1/2
μj,k := hj,k ∂RN /∂xk L2 (Γj ) (k = 1, 2) (10.67)
with an axes parallel rectangle Γj with edge-lengthes hj,1 and hj,2 reflects the
singular behaviour better than μj .
A two-level ansatz and a saturation assumption are used in [314](see Sec-
tion 10.5) to see for quasi-uniform meshes, that, up to multiplicative constants
( N j =1 ηj,1 + ηj,2 + ηj,3 )
2 2 2 1/2 is a lower and upper error bound, where,
RN , ξj,k
ηj,k := − (k = 1, 2, 3)
V ξj,k , ξj,k 1/2
and the ansatz functions ξj,k are defined for one rectangle by dividing it into four
congruent rectangles, see Fig. 10.4, where ξj,k is denoted by βj +1,·,k .
The error indicators μj , μj,k , νj := RN H α (ωj ) and ηj,k can used for steering
automatic mesh-refinements.
Theorem 10.8 ([87]) There exist constants c1 , c2 , c3 > 0 which depend on the
|Γ |
aspect ratio ( max |Γjl | ) of the elements Γ1 , . . . , ΓN in T and on Γ but not on
j =l
f , RN := f − V ψN ,or ψ = V −1 f and neither on the sizes nor numbers of
elements in T . We have
ηj,k ≤ c1 μj,k ≤ c1 μj (k = 1, 2; j = 1, . . . , N) (10.68)
⎛ ⎞1/2 ⎛ ⎞1/2

⎝ 2
(η1,l 2 ⎠
+ η2,l ) ≤ c2 νj ≤ c3 ⎝ μ2l ⎠ (j = 1, . . . , M) (10.69)
Γl ∈ωj Γl ∈ωj
Estimates with ηj,k require α = 1/2 while the other holds for all α ∈ [0, 1].
Proof ξj,k can be written as the derivative of a hat function φj,k with height hj,k /2.
Then
RN , ξj,k = −∂xk RN , φj,k ≤ φj,k L2 (Γj ) ∂xk RN L2 (Γj )
Now we note that for each triangulation and hat function ϕz at node z with hz :=
diam(supp(ϕz)) there holds
ϕz H̃ s (Γ ) ∼ h1−s

z for 0 ≤ s ≤ 1 (10.70)
Therefore taking a hat function ϕ with supp (ϕ) ⊂ ω and ϕL2 (Γ ) ∼ hω :=

diam (ω) we have with the characteristic function χ of ω
χ, η 2 χ, ϕ 2
V χ, χ ≈ χ2H=−1/2 (Γ ) ≥ sup ≥ ≈ hω |ω|
η∈H 1/2 (Γj ) ηH 1/2 (Γ φ2H 1/2 (Γ )
j 2
j)
η=0
Thus
RN , ξj,k φj,k L2 (Γj )

≤ c1 ∂xk RN L2 (Γj ) μj,k
V ξj,k , ξj,k 1/2 (hj,k |Γj |)1/2
This implies the first estimate in (10.68). For (10.69) see [87] .

10.3.1 Adaptive Algorithms
The error estimators derived above lead to the following three algorithms, Algorithm
A without direction control, Algorithm B with a direction control and (hierarchical)
Algorithm C.
Let the parameter 0 ≤ θ ≤ 1 and an initial partition π0 of Γ be given. Let
S0,N0 (Γ ) be the finite dimensional space of piecewise constant functions of π0 .
The adaptive algorithms read as follows:
Algorithm A For m = 0, 1, 2, . . .
(i) Compute the Galerkin solution ψN ∈ S0,Nm (Γ ) according to the partition
πm = {Γ1 , . . . , ΓNm }.
(ii) For each element Γi ∈ πm compute the local error indicator
1/2
μi := hi ∇RN L2 (Γi )
where hi = diam Γi . Check for a stopping criterion.

354 10 A-BEM
(iii) Compute μmax := max{μi ; i = 1, . . . , N} and refine the rectangle Γi into

four equal sized rectangles iff
μi ≥ Θμmax
This defines the new partition πm+1 and the refined Nm+1 -dimensional space
S0,Nm+1 (Γ ) ⊃ S0,Nm (Γ ).
Algorithm B For m = 0, 1, 2, . . .
πm = {Γ1 , . . . , ΓNm }.
(ii) For each element Γi ∈ πm compute the local error indicators
1/2
μi,k := hi,k ∂xk RN L2 (Γi )
where hi,k = diameter of Γi in xk -direction (k = 1, 2). Check for a stopping

criterion.
(iii) Compute μmax,k := max{μi,k ; i = 1, . . . , N} (k = 1, 2) and refine the
rectangle Γi along the xk -axis iff
μi,k ≥ Θμmax,k (k = 1, 2).
S0,Nm+1 (Γ ) ⊃ S0,Nm (Γ ).
Algorithm C For m = 0, 1, 2, . . .
πm = {Γ1 , . . . , ΓNm }.
(ii) For each element Γi ∈ πm compute the local error indicators
|f − V ψN , ξi,1 | |f − V ψN , ξi,2 |
H
ηi,1 := , H
ηi,2 :=
V (ξi,1 , ξi,1 )1/2 V (ξi,2 , ξi,2 )1/2
where ξi,1 = βi+1,·,1 , ξi,2 = βi+1,·,2 see Fig. 10.4 and the global error indicator
⎛ ⎞1/2
m
4N
Σn := ⎝ ηi2 ⎠ .
i=Nm +1
The algorithm stops if Σn < ε0 where ε0 > 0 is a given constant.

H
(iii) Compute ηmax := max{ηi,1H , ηH ; i = 1, . . . , N} (k = 1, 2) and refine the
i,2
rectangle Γi along the xk -axis iff
H
ηi,k ≥ Θηmax
H
(k = 1, 2).
S0,Nm+1 (Γ ) ⊃ S0,Nm (Γ ).
10.3.2 Numerical Example
Let Γ be the L-Shape and f = 1. The energy norm of the solution ψ of V ψ = 1

is known to be ψH̃ −1/2 = 2.878293. We start with a uniform mesh containing 12
elements and apply our algorithms A,B and the algorithm C using the hierarchical
error indicator. Table 10.1 gives the results for Algorithm B and Table 10.2 gives
the results for the hierarchical error indicator. Figure 10.1 gives the error curves
for the adaptive algorithms, the uniform h-version and for algebraically graded
meshes with different grading parameters. We see that Algorithm B and Algorithm
C (hierarchical error indicator) are giving nearly identical curves with optimal
convergence rates. The adaptive schemes generate the same curves for Θ ∈ [0.5, 1],
only the number of newly generated meshes raises with increasing Θ.
The error indicators of Algorithms A and B have been computed by numerical
integration of the analytically computed gradient of the residual by an 4 × 4-
Gaussian quadrature rule. The error indicators of Algorithm C have been calculated
analytically using the same algorithms as in the computation of the Galerkin matrix.
All computations have been performed on a Sun Ultrasparc-II (300MHz) at
the Institute for Applied Mathematics, University of Hannover, using the program
system maiprogs [293].
Table 10.1 Residual error N Indicator Error Effectivity Rate

indicators and efficiency for
θ = 0.5 (Alg. B) [87] 12 0.9558277 0.6646427 1.438108
33 0.7157455 0.4917142 1.455613 0.29789
64 0.5231094 0.3603837 1.451534 0.46911
105 0.3791923 0.2621146 1.446666 0.64310
148 0.2857298 0.1968658 1.451394 0.83396
184 0.2220168 0.1548582 1.433678 1.10237
245 0.1641732 0.1158337 1.417319 1.01408
306 0.1275614 0.0909866 1.401980 1.08598
371 0.0984725 0.0707287 1.392257 1.30757
450 0.0789983 0.0579766 1.362589 1.02987
540 0.0638628 0.0465304 1.372496 1.20629
656 0.0509174 0.0376706 1.351647 1.08548
805 0.0411453 0.0308533 1.333580 0.97535
968 0.0340271 0.0260617 1.305636 0.91532
1243 0.0266476 0.0207189 1.286148 0.91749
1543 0.0217106 0.0173712 1.249805 0.81513
1902 0.0179114 0.0144254 1.241652 0.88835
2418 0.0145247 0.0118803 1.222581 0.80867
3127 0.0116897 0.0097656 1.197027 0.76231
3926 0.0096766 0.0081669 1.184856 0.78566
356 10 A-BEM
Table 10.2 Hierarchical N Indicator Error Effectivity Rate

error indicators and efficiency
for θ = 0.5 (Alg. C) 12 0.3460879 0.6646427 0.520713
33 0.2516634 0.4917142 0.511808 0.29789
64 0.1843478 0.3603837 0.511532 0.46911
105 0.1343572 0.2621146 0.512590 0.64310
158 0.0961964 0.1898212 0.506774 0.78970
194 0.0755879 0.1485726 0.508760 1.19363
255 0.0576133 0.1104080 0.521822 1.08590
324 0.0451620 0.0838526 0.538589 1.14883
382 0.0376194 0.0670436 0.561118 1.35852
469 0.0320552 0.0526239 0.609138 1.18028
606 0.0251860 0.0410832 0.613048 0.96602
759 0.0206936 0.0328849 0.629274 0.98873
977 0.0168162 0.0262459 0.640718 0.89313
1240 0.0138612 0.0212237 0.653100 0.89097
1667 0.0109750 0.0167291 0.656044 0.80418
1670 0.0108687 0.0165264 0.657656
2487 0.0079173 0.0120928 0.654716 0.78429
3188 0.0063593 0.0096686 0.657722 0.90097
4053 0.0053212 0.0077425 0.687276 0.92541
1
Error in Energy norm
0.1
h-version, p=0
graded mesh(5), beta=2.0
Algorithm A (theta=0.5)
Algorithm A (theta=0.8)
Algorithm B (theta=0.5)
0.01 Algorithm C (theta=0.5)
10 100 1000
Number of Unknowns
Fig. 10.1 Dirichlet problem on the L-shape in R3 [87]

10.4 The Hypersingular Integral Equation in 3D 357
Remark 10.4 For uniform and graded meshes, the superiority of a grading with
a parameter (j/J )4 for j = 0, . . . , J gives the best result. Algorithm (A) is not
competitive, there is a need for anisotropic mesh-refinement. Algorithms (B) and (C)
yield similar results and can compete for higher degrees of freedom with optimally
graded meshes.
10.4 The Hypersingular Integral Equation in 3D
Here we report from [86]. Each element Γj of the triangulation T is supposed to

be a closed (flat) triangle or parallelogram in R3 .
The set of all nodes in the triangulation is denoted by N and the free nodes by
K := N \γ , where γ is the boundary of Γ . The set of all edges in the triangulation
is denoted by E and E is split into edges on the boundary γ , namely Eγ := {E ∈
E : E ⊆ γ }, and interior edges EΓ := {E ∈ E : E ∩ γ = ∅}.
For each free node z ∈ K there is a hat function ϕz which equals zero on each
element Γk (if z ∈ Γk ) or one of the nodal local basis functions (if z is a vertex of
Γk ) such that ϕz (z) = 1 and ϕz (x) = 0 for all x ∈ N \ {z}. The hat functions are
Lipschitz continuous and form a partition of unity. Their linear hull
S := span{ϕz : z ∈ K } ⊆ H01 (Γ ) (10.71)
satisfies proper boundary conditions. Since {ϕz : z ∈ K } is, in general, not a

partition of unity, we choose a node ζ (z) ∈ K for each z ∈ N \ K and define
ζ (z) = z for z ∈ K . We get a partition of N into card{K } classes I (z) := {z̃ ∈
N : ζ (z̃) = z}, z ∈ K . For each z ∈ K set

ψz := ϕz̃ (10.72)
z̃∈I (z)
and notice that {ψz : z ∈ K } is a partition of unity. It is required that
Ωz := {x ∈ Γ : 0 < ψz (x)} (10.73)
is connected and contains only a limited number of elements. We remark that ψz =

ϕz implies that γ ∩ ∂Ωz has a positive surface measure. Those definitions follow
and adapt [77] in order to employ an approximation operator (0 ≤ s ≤ 1)
J : H̃ s (Γ ) → S ⊆ H̃ s (Γ ). (10.74)
358 10 A-BEM
Ωz gψz ds
For each g ∈ L1 (Γ ) let gz ∈ R be gz := for z ∈ K and then set
Ωz ϕz ds

J g := gz ϕz ∈ S. (10.75)
z∈K
Lemma 10.5 ([77]) There is a constant c4 > 0 that depends on Γ and the aspect
ratio of the elements (but not on their sizes) such that for all z ∈ K and g ∈ H01 (Γ )
=H =1 (Γ ) we have
3 4
ψz g − ϕz gz L2 (Ωz ) ≤ c4 min gL2 (Ωz ) , hz ∇gL2 (Ωz ) , (10.76)
where hz = diam (Ωz ).

Suppose that the residual R := f − W uh = W (u − uh ) ∈ L2 (Γ ) with the
hypersingular operator W satisfies the Galerkin conditions
R, ϕz = 0 for all z ∈ K , (10.77)
where ·, · denotes the duality pairing on H̃ −s and H s . Then we have the following
residual-based a posteriori error estimate.
Theorem 10.9 There is a constant c5 > 0 such that for all R ∈ L2 (Γ ) with (10.77)
and 0 ≤ s ≤ 1 there holds
1/2
RH s−1 (Γ ) ≤ c5 h2−2s
z R2L2 (Ω . (10.78)
z)
z∈K
Proof Since (ψz : z ∈ K ) defines a partition of unity, we have R = z∈K ψz R.

This combined with (10.77) shows for g ∈ H01 (Γ )

R, g = R, gψz = R, gψz − gz ϕz (10.79)
z∈K z∈K
with gz ∈ R defined as above. Now
gψz − gz ϕz L2 (Ωz ) ≤ c1 h1−s

z gH 1−s (Ωz ) (10.80)
D
and (10.79) imply

1/2 1/2
R, g ≤ c2 hz2−2s R2L2 (Ω g2 1−s (10.81)
z) HD (Ωz )
z∈K z∈K
A coloring argument (see proof of Lemma 3.1 [87] ) completes the proof.

The following result guarantees reliability of the estimator

1/2
η := ηT2 with ηT := h1−s
T RL2 (T ) and hT := diam (T ) for T ∈ T .
T ∈T
Theorem 10.10 For 0 < s < 1, there is a constant c6 > 0 that depends on s, Γ ,
and the aspect ratio of the elements (but not their size) such that
u − uh H̃ s (Γ ) ≤ c6 c7 h1−s
T RL2 (Γ ) , (10.82)
where the T -piecewise constant hT ∈ L∞ (Γ ) is defined by hT (x) = hT for x ∈ T

and c7 := max{hz / hT : z ∈ K , T ∈ T with T ⊆ Ωz }.
Proof W : H̃ s (Γ ) → H s−1 (Γ ) is a continuous, linear bijection and by the open
mapping theorem the inverse map W −1 is continuous i.e., W −1 < ∞. Since
R = W (u − uh ) we obtain
u − uh H̃ s (Γ ) ≤ W −1 RH s−1 (Γ ) . (10.83)
Theorem 10.9 and the finite overlap of the {Ωz : z ∈ K } imply

1/2
RH s−1 (Γ ) ≤ c5 h1−s R2L2 (Ω ≤ Mc5 c7 h1−s
T RL2 (Γ ) (10.84)
z)
z∈K
with a bounded number M and h(x) := max{hz : x ∈ Ωz }.

Now we compare η with the multilevel error estimator μ from Section 10.5.1 for
pw. linears on triangles and from[313] for pw. bilinears on rectangles.
We need at least two meshes where one triangulation Th is a refinement of TH .
The set of free nodes Kh and KH give rise to hat functions (ϕz : z ∈ Kh ) and
(ϕz : z ∈ KH ) with respect to Th and TH , respectively. If Sh and SH denote the
respective discrete spaces, SH ⊆ Sh and
Sh = SH ⊕ span{ϕz : z ∈ Kh \KH }, (10.85)
with respective discrete solutions uh and uH . For the practical computation, only
uH is required, Sh plays the role of a fictitious larger space. However, the saturation
assumption,
u − uh W ≤ κu − uH W , (10.86)
for some fixed 0 ≤ κ < 1, plays an essential role. In contrast to the finite
element context for partial differential equations [84, 143, 144], a proof of (10.86)
is unknown for boundary element problems.
360 10 A-BEM
Definition 10.2 For each z ∈ Kh , let μz := R, ϕz /ϕz W with R := f − W uH .

Theorem 10.11 ([313]) Under the saturation assumption (10.86) we have with
mesh independent constants c8 , c9

c8 μ2z ≤ u − uH 2W ≤ c9 μ2z . (10.87)
z∈Kh \KH z∈Kh \KH
The proof of the above theorem is given in Subsection 10.5.1.

Remark 10.5 The constant c8 in (10.87) depends on 1/(1 − κ) and (possibly)
degenerates as κ → 1. The constant c7 is robust with 0 ≤ κ < 1. The two-level
error estimator
1/2
μ := μ2z (10.88)
z∈Kh \KH
performs very accurately in practice.
Theorem 10.12 There is an hT -independent constant c10 such that, for each z ∈
Kh \KH and supp ϕz ⊆ T1 ∪ T2 with T1 , T2 ∈ TH , we have
1/2
μ2z ≤ c10 hT RL2 (T1 ∪T2 ) ≤ c10 (ηT1 + ηT2 ). (10.89)
Proof Since ϕz 2W ≈ ϕz 2 ≈ hz due to (10.70) we have

H̃ (Γ )1/2
μ2z ≤ R2L2 (T ϕz 2L2 (T ≤ hz R2L2 (T

1 ∩T2 ) 1 ∩T2 ) 1 ∩T2 )
In our numerical experiments we use the estimators
1/2
ηN := hT RL2 (Γ ) for the residual R := f − W uN
1/2
μN := μ2z from (10.88)
z∈Kh \Kk
for the current coarse mesh TH = Tk and one (fictitious) refinement Th with
the new nodes Kh \Kk on all edges. We perform the refinement strategy of
Fig. 12.2. In [86] we have performed numerical experiments for the Neumann
screen problem of the Laplacian with boundary data g(x) = 1 on the L-shaped
screen (Example 1) and with g(x) = |x−(0.1,0.1)|
1
on Γ = [−1, 1]2 (Example
2). The numerical experiments are compactly displayed in Figs. 10.2 and 10.3,
where we plotted the energy norm EN versus the number of degrees of freedom N.
Both axes are scaled logarithmically. For description of algorithms (AR ) (residual)
and (AH ) (hierarchical) see [86]. In both examples the experimental convergence
rate of the uniform h-version is ≈ 1/2. The examples validate the reliability of
1
h-uniform, Q1
p-uniform, Qp
graded-Q1, β =4.0
2 h-uniform, P1
1 p-uniform, Pp
(AR), θ =0.8
(AR), θ =0.9
(AH), θ =0.8
(AH), θ =0.9
1
1
3
0.1
1 10 100 1000 10000

degrees of freedom
Fig. 10.2 Neumann problem on the L-shape in Example 1 [86]
h-uniform, Q1
h-uniform, P1
p-uniform, Qp
2
1 p-uniform, Pp
1 graded-Q1 , β =4.0
(AR), θ =0.8
(AH), θ =0.8
(AR), θ =0.9
(AH), θ =0.9
1
3
2
0.1
1 10 100 1000 10000

degrees of freedom
Fig. 10.3 Neumann problem on the unit square in Example 2 [86]

362 10 A-BEM
the estimator ηN and the efficiency of μN . The numerical experiments show the
superiority of automatic adaptive over uniform mesh-refinements. An experimental
convergence rate higher than one is not observed in our examples; the conforming
triangles require too many degrees of freedom to resolve the anisotropic layer
structure as they refine isotropically. The numerical results in Sect. 10.3 for Symm’s
integral equation on rectangles show that indeed a speed-up of a factor 3/2 in
the convergence rate is possible when anisotropic refinement is used. The optimal
grading parameter is β = 3 + and we choose β = 4 in the experiments according
to Theorem 7.18.
10.5 Two-Level Adaptive BEM for Laplace, Lamé,

Helmholtz
In this section we consider hierarchical adaptive refinement strategies for boundary

element discretizations of the Laplace, the Lamé and the Helmholtz equation with
Dirichlet boundary conditions on surface pieces in R3 modelling screens and cracks.
This leads to weakly singular integral equations of the first kind with the single layer
potential. To obtain discrete approximations of the integral equations we apply the
h-version Galerkin boundary element method with piecewise constant test and trial
functions. We define a two-level decomposition of the discrete spaces for the case
of the Laplacian and use the stability of this decomposition to derive a-posteriori
error estimators for the case of the Laplace, Lamé and Helmholtz operators which
estimate the Galerkin error in the energy norm from above and from below and
which consist of easily computable local error indicators. Here, stability of the
subspace decomposition means that the condition number of the corresponding two-
level additive Schwarz operator is bounded independently of the mesh size h. For
the theoretical results we assume a saturation condition and uniform refinements.
The latter restriction is relaxed in the adaptive algorithm where we allow local
directional refinements of the quadrilateral elements which lead to anisotropic
boundary element meshes. These meshes are essential for optimal approximations
since the solutions of our model problems have edge and corner singularities. The
theoretical results are illustrated by numerical experiments for the case of the Lamé
equation.
Our approach to hierarchical adaptivity is related to the results by Bank and
Smith [32]. The first results for 3D adaptive boundary element methods were
obtained by Mund, Stephan and Weisse [314].
The Model Problem
Let Γ ⊂ R3 denote an open surface piece. For simplicity we assume that Γ is

a right angled polygonal domain in the (x1 , x2 )-plane. In the unbounded domain
10.5 Two-Level Adaptive BEM 363
Ω = R3 \ Γ we consider the following Dirichlet problem: Given a function g ∈

(H 1/2(Γ ))d find u ∈ (Hloc
1 (Ω))d such that
Du = 0 in Ω
u=f on Γ
where the differential operator D is either the Laplace operator Δ or the Lamé
operator Δ∗ or the Helmholtz operator Δ + κ 2 (κ > 0). If D = Δ∗ we have d = 3
and otherwise d = 1. Furthermore we assume the radiation condition u = O(r −1 )
and additionally in the case of the Helmholtz equation
∂u
− iku = o(r −1 ) for r = |x| → ∞ .
∂r
According to Lions and Magenes [284] we define the Sobolev space H d =

=−1/2(Γ ))d as the dual space of (H 1/2 (Γ ))d .
(H
The problem can be reformulated as the following boundary integral equation
[396] and [398]: Find φD ∈ H d such that
V D φD = g (10.90)
where g = 2f and VD : H d → (H 1/2(Γ ))d is the single layer potential operator

(VD ψ)(x) = −2 GD (x, y) ψ(y) dσy .
Γ
Here, GD (x, y) ∈ Rd×d denotes the fundamental solution of the differential

operator D, i.e.
⎧
⎨ 1 exp(iκ|x−y|)
4π |x−y| if D = Δ + κ 2
GD (x, y) = & .
'
⎩ λ+3μ 1
+ λ+μ (x−y)(x−y)
if D = Δ∗ .
8πμ(λ+2μ) |x−y| I3×3 λ+3μ |x−y|3
with κ ≥ 0 and the Lamé constants λ and μ satisfy μ > 0 and 2λ + μ > 0.
The variational formulation of (10.90) is given by: Find φD ∈ H d such that
VD φD , ψ = g, ψ (10.91)
for all ψ ∈ H d where ·, · denotes the duality pairing of (H 1/2(Γ ))d and H d . We
define the symmetric bilinear form VD (χ, ψ) = VD χ, ψ . If D = Δ or D = Δ∗
then VD (·, ·) is an inner product in H d and in this case there is a constant η > 0
such that
VD (ψ, ψ) ≥ η ψ2H d (D ∈ {Δ, Δ∗ }) (10.92)

364 10 A-BEM
for all ψ ∈ H d . The single layer potential for the Helmholtz operator satisfies
|(VΔ+κ 2 − VΔ )χ , ψ | ≤ c χL2 (Γ ) ψH d (10.93)
for all χ, ψ ∈ H d and for the constant c > 0.

Now, let τ0 be a an initial partition of Γ into regular quadrilaterals of size h0
and let {τj }∞
j =1 be a sequence of uniform refinements of τ0 where τj is obtained
by refining all elements in τj −1 into four equally sized quadrilaterals of size hj =
h0 2−j . The number of elements in τj is denoted by nj . We seek approximations to
the exact solution φD of (10.91) in the spaces
Sj,D = {ψ : Γ → Δd | ψ piecewise constant w.r.t. τj } (10.94)
where Δ = C if D = Δ + κ 2 (κ > 0) and Δ = R otherwise.

The Galerkin method at level j reads as follows: Find φj,D ∈ Sj,D such that
VD (φj,D , ψ) = g, ψ (10.95)
for all ψ ∈ Sj,D .
A Stable Two-Level Decomposition
In this section we consider a two-level decomposition by Mund et al. [314] for the
Laplacian Δ. The stability of this decomposition will be used in the next section to
derive a-posteriori estimates for the Galerkin error in the cases of the Laplace, Lamé
and Helmholtz equations.
For J ∈ N we consider hierarchical basis functions βj,i,l in SJ,Δ where j ∈
{0, . . . , J } denotes the level, i ∈ {1, . . . , nj } denotes the number of the element and
l ∈ {1, 2, 3}. We start with the standard basis of piecewise constant brick functions
j
in S0,Δ and define the hierarchical basis recursively. Whenever an element Γi ∈ τj
j +1 j +1 j +1 j +1
(j ≥ 0) is divided into the four elements Γi1 , Γi2 , Γi3 and Γi4 we extend
the basis of Sj,Δ by the basis functions βj +1,i,1 , βj +1,i,2 and βj +1,i,3 which are
defined as
⎧ j +1 j +1
⎪
⎨ 1 if x ∈ Γi1 ∪ Γi2
βj +1,i,1 (x) = −1 if x ∈ Γij +1 ∪ Γij +1
⎪
⎩ 3 4
0 otherwise
⎧ j +1 j +1
⎪
⎨ 1 if x ∈ Γi1 ∪ Γi4
j +1 j +1
βj +1,i,2 (x) = −1 if x ∈ Γi ∪ Γi3
⎪
⎩ 2
0 otherwise
βj +1,i,3 (x) = βj +1,i,1 (x) · βj +1,i,2 (x) (see Fig. 10.4) .

j
Fig. 10.4 Refinement of Γi j +1 j +1
into four new elements and x2 Γi4 Γi3
j
the additional basis functions Γi
x1 j +1 j +1
Γi1 Γi2
−1 −1 +1 −1 −1 +1
+1 +1 +1 −1 +1 −1
βj +1,i,1 βj +1,i,2 βj +1,i,3
We define the one-dimensional spaces Yj +1,i,l = span{βj +1,i,l } and note that
j
adding Yj +1,i,1 to the discrete space Sj,Δ corresponds to a bisection of Γi (the
i-th element at level j ) along the x1 -axis and adding Yj +1,i,2 corresponds to
j
a bisection of Γi along the x2 -axis. Now, we consider the following two-level
subspace decomposition of Sj +1,Δ :
nj
Q Q
3
Sj +1,Δ = Sj,Δ ⊕ Yj +1,i,l . (10.96)
i=1 l=1
To define the corresponding two-level additive Schwarz operator we need the

Galerkin projections Pj : Sj +1,Δ → Sj,Δ and Pj +1,i,l : Sj +1,Δ → Yj +1,i,l which
are defined by
VΔ (Pj χ, ψ) = V (χ, ψ) ∀ ψ ∈ Sj,Δ
and
VΔ (Pj +1,i,l χ, ψ) = V (χ, ψ) ∀ ψ ∈ Yj +1,i,l
with χ ∈ Sj +1,Δ .
The two-level additive Schwarz operator PjΔ+1 : Sj +1,Δ → Sj +1,Δ is now
defined as
nj 3

PjΔ+1 = Pj + Pj +1,i,l .
i=1 l=1
The following result was proved by Mund et al. [314]:

Theorem 10.13 There are constants μ1 , μ2 > 0 independent of j such that
μ1 VΔ (ψ, ψ) ≤ VΔ (PjΔ+1 ψ, ψ) ≤ μ2 VΔ (ψ, ψ) ∀ψ ∈ Sj +1,Δ . (10.97)

366 10 A-BEM
The above result shows that the condition number of the operator PjΔ+1 is bounded
independently of j and, hence, the subspace decomposition (10.96) is termed as
being stable.
A-Posteriori Error Estimates
In this section we introduce a-posteriori error estimates based on the hierarchical

structure of the discrete spaces. We estimate the difference between the exact
solution φD of (10.91) and the Galerkin solution φj,D of (10.95) in the H d -norm for
D = Δ+κ 2 (κ ≥ 0) and d = 1 or D = Δ∗ and d = 3. The estimate is proved for the
case of uniform refinements and under the assumption of the following saturation
condition:
Assumption 10.1 (AD ) There exists an integer j0 and a constant < 1 such that
φD − φj +1,D H d ≤ φD − φj,D H d ∀ j ≥ j0 . (10.98)
This assumption is certainly satisfied if φD − φj,D H d ∼ n−α

j for some constant
α > 0 and where nj denotes again the number of elements in τj . From (10.98) we
conclude that
1 1
φj +1,D − φj,D H d ≤ φD − φj,D H d ≤ φj +1,D − φj,D H d
1+ρ 1−ρ
(10.99)
and, hence, it remains to estimate the difference between two successive Galerkin
solutions.
For the Laplacian, i.e. D = Δ, we obtain from Theorem 10.13 and from the
orthogonality of the Galerkin projections Pj +1,i,l that
nj 3

χ2H d ∼ VΔ (PjΔ+1 χ, χ) = VΔ (Pj χ, χ)+ VΔ (Pj +1,i,l χ, χ) (10.100)
i=1 l=1
for all χ ∈ Sj +1,Δ .

For χ = φj +1,Δ − φj,Δ the terms on the right hand side of (10.100) can be
calculated explicitly: From the Galerkin property of φj +1,Δ and φj,Δ we obtain
VΔ (Pj (φj +1,Δ − φj,Δ ), φj +1,Δ − φj,Δ ) = 0 . (10.101)
Since Pj +1,i,l projects the space Sj +1,Δ onto the one-dimensional space Yj +1,i,l it
is easy to verify that
|g − VΔ φj,Δ , βj +1,i,l |
Δ
ηj,i,l := VΔ (Pj +1,i,l (φj +1,Δ − φj,Δ ), φj +1,Δ − φj,Δ ) = .
VΔ (βj +1,i,l , βj +1,i,l )1/2
(10.102)
The following a-posteriori error estimate follows now from (10.99)–(10.102):

Theorem 10.14 Under Assumption AΔ there exist constants C1 , C2 > 0 and an
integer j0 such that for all j ≥ j0 the following holds:
nj 3 nj 3
2 2
C1 Δ
ηj,i,l ≤ φΔ − φj,Δ 2H ≤ C2 Δ
ηj,i,l .
i=1 l=1 i=1 l=1
Δ are defined in (10.102).

The local error indicators ηj,i,l
A corresponding result for D = Δ + κ 2 with κ > 0 was obtained by Maischak

et al. [294] by taking advantage of (10.93):
Theorem 10.15 Under Assumption AΔ+κ 2 there exist constants C1 , C2 > 0 and an
nj 3 nj 3
2 2
Δ+κ 2 2
C1 ηj,i,l ≤ φΔ+κ 2 − φj,Δ+κ 2 2H ≤ C2 Δ+κ
ηj,i,l .
i=1 l=1 i=1 l=1
The local error indicators are defined as
Δ+κ 2 |g − VΔ+κ 2 φj,Δ+κ 2 , βj +1,i,l |

ηj,i,l = .
VΔ (βj +1,i,l , βj +1,i,l )1/2
In the case of the Lamé operator, D = Δ∗ , we use the fact that

3
VΔ∗ (ψ, ψ) ∼ VΔ (ψk , ψk ) ∀ ψ = (ψ1 , ψ2 , ψ3 ) ∈ H 3
k=1
which follows from (10.92) and from the continuity of the single layer potential
operators VΔ and VΔ∗ . By applying the stability result of Theorem 10.13 to each
component of the vector φj +1,Δ∗ − φj,Δ∗ we obtain local error indicators similar to
those in (10.102). Since the spaces spanned by the basis functions βj +1,i,l are now
three-dimensional we have to solve a 3 × 3 linear system (see (10.103) below) to
Δ∗ .
obtain the error indicators ηj,i,l
Theorem 10.16 Under Assumption AΔ∗ there exist constants C1 , C2 > 0 and an
nj 3 nj 3
2 2
Δ∗ ∗
C1 ηj,i,l ≤ φ Δ∗ −φ j,Δ∗ 2H ≤ C2 Δ
ηj,i,l .
i=1 l=1 i=1 l=1
368 10 A-BEM
The local error indicators are defined as

∗
Δ
ηj,i,l = VΔ∗ (ej,i,l , ej,i,l )1/2
and ej,i,l ∈ (span{βj +1,i,l })3 is obtained by solving
VΔ∗ (ej,i,l , ψ) = g − VΔ∗ φj,Δ∗ , ψ (10.103)
for all ψ ∈ (span{βj +1,i,l })3 .
The Adaptive Algorithm
Based on the above a-posteriori estimates we formulate a refinement strategy for

h-adaptivity. We relax the regularity assumptions on the meshes τj and allow in this
section non-uniform and anisotropic refinements. The following adaptive algorithm
is closely related to those used by Maischak et al. [294] and Mund et al. [314].
Algorithm 10.1 Let the parameter 0 ≤ θ ≤ 1 and an initial subdivision τ0 of Γ
be given. Let S0,D be the finite dimensional space of piecewise constant functions
over τ0 (cf. (10.94)) where D denotes the Laplace, Helmholtz or Lamé operator. For
j = 0, 1, 2, . . . we perform the following:
(1) Compute the Galerkin solution φj ∈ Sj,D .
j
(2) For each element Γi ∈ τj compute the local error indicators ηj,i,1D D
and ηj,i,2
D indicates the error along the x -axis (l ∈ {1, 2}). Check
where we note that ηj,i,l l
the stopping criterion.
D , ηD } and refine Γ j along the x -axis iff
(3) Compute ηmax := maxΓ j ∈τ {ηj,i,1 j,i,2 i l
i j
D
ηj,i,l ≥ θ ηmax .
This defines the subdivision τj +1 and the refined space Sj +1,D ⊃ Sj,D . Go back
to Step 1.
We stop the algorithm in Step 2 if the criterion
( nj 3 )1/2

Σj := D
(ηj,i,l )2 < ε0 (10.104)
i=1 l=1
is satisfied, where ε0 > 0 is a given constant. This stopping criterion is motivated

by the a-posteriori error estimates in Theorems 10.14–10.16.
j
In Step 3 of the algorithm, refinement of Γi along the xl -axis means bisection of
j j
Γi (l ∈ {1, 2}) and refinement along both axes means subdivision of Γi into four
D appears in (10.104) but is
elements (cf. Fig. 10.4). Note that the third indicator ηj,i,3
j
not used to decide whether or not to refine the element Γi since the corresponding
j
basis function βj +1,i,3 only appears if Γi is refined along both axes.
D we can use the same subroutines as for
To compute the local error indicators ηj,i,l
the computation of the Galerkin matrix and of the right hand side vector. Hence, an
existing boundary element code which allows basis transformations and local mesh
refinements can be easily equipped with the above algorithm.
Numerical Results
Let Γ denote the L-shaped surface piece modelling a screen or crack with corners at
(−1, −1, 0), (1, −1, 0), (1, 0, 0), (0, 0, 0), (0, 1, 0) and (−1, 1, 0). We consider the
Dirichlet boundary value problem (10.90) with the Lamé operator D = Δ∗ and right
hand side f = −1/2 and this leads to the boundary integral equation VΔ∗ φ = 1 on
Γ (cf. (10.90)). Let φj = φj,Δ∗ ∈ Sj,Δ∗ denote the Galerkin approximation of φ
at level j (cf. (10.95)). As in the previous section Sj,Δ∗ is the space of piecewise
constant functions on the locally refined mesh τj .
We apply Algorithm 10.1 with θ = 0.7 and the initial mesh τ0 consisting of
12 equally sized squares and we need 10 adaptive refinement steps to reach the
given accuracy ε0 = 5 · 10−2 in (10.104). At each step we compute the Galerkin
error (VΔ∗ (φ − φj , φ − φj ))1/2 in energy norm which is equivalent to the norm in
−1/2
H 3 = (H00 (Γ ))3 . From the Galerkin property of φj we obtain
1/2
Ej := (VΔ∗ (φ − φj , φ − φj ))1/2 = VΔ∗ (φ, φ) − VΔ∗ (φj , φj ) . (10.105)
To compute (10.105) we replace the quantity VΔ∗ (φ, φ) by the value 3.72844, which
was obtained from the values VΔ∗ (φj , φj ) by extrapolation.
It is well known that the exact solution of our model problem has edge and corner
singularities at the boundary of Γ , with the exception of the incoming corner at
(0, 0, 0). Hence, we may expect anisotropically refined meshes at the edges of Γ .
The optimal convergence rate of our adaptive Galerkin method is given by Ej ∼
−3/4+
nj where nj denotes the number of elements in τj and > 0 is arbitrarily
−1/4
small. However, for uniform refinements it is only Ej ∼ nj .
The values of Ej as a function of the number of unknowns Nj = 3nj (0 ≤
j ≤ 10) are plotted in Fig. 10.5. For comparison we also plotted the optimal curve
−3/4
cNj for some constant c > 0. We observe that both curves have approximatly the
same slope for Nj sufficiently large and this indicates the reliablity of the adaptive
370 10 A-BEM
1
error in energy norm (adaptive h-version)
optimal convergence
0.1
100 1000
Nj
Fig. 10.5 Error Ej in the energy norm plotted versus the number of unknowns Nj
algorithm and of the a-posteriori error estimate in Theorem 10.16 (for locally refined
meshes). The sequence of refined meshes is shown in Fig. 10.6.
10.5.1 A Stable Two-Level Subspace Decomposition for the

Hypersingular Operator
Let Sh1 (Γ ) denote the space of continuous, piecewise linear functions on Γ . For
a partition of unity {θj ; j = 1, . . . , Jν } which consists of continuous, piecewise
linear functions:

θj = 1, supp θj = Γj , 0 ≤ θj ≤ 1. (10.106)
j
there holds
∂ ∂
| θj |, | θj | ≤ c/ h.
∂x ∂y
where the domain Γj is the union of the elements Γi , which are adjacent to the
node xj .
Fig. 10.6 The sequence {τj }8j =0 of refined meshes
Lemma 10.6 For any w ∈ Sh1 (Γ ) there holds for the linear interpolation operator
Π1 in the nodal points
Π1 θj wH̃ 1/2 (Γ ) ≤ Cθj wH̃ 1/2 (Γ )
uniformly in h and for all cut-off functions θj in (10.106). Here, C is an arbitrary

constant ≥ 4.07.
Proof We first prove the continuity of Π1 with respect to the L2 - and H01 -norms
and then interpolate these results.
372 10 A-BEM
Fig. 10.7 Configuration: (2h, 2h)

Hat function θ concentrated
at (h, h)
ϕ3
Γ1
ϕ1 ϕ2
(0, 0)
Let us consider the reference rectangle (0, 2h)2 and the part Γ1 = {(x, y) ∈
(0, h)2 | y < x} therein. By θ we denote the hat function concentrated in (h, h),
i.e. θ (x, y) = 1 for (x, y) = (h, h) and θ (x, y) = 0 at the adjacent 6 nodes, see
Fig. 10.7. On Γ1 the space of continuous, piecewise linear functions is spanned
by the hat functions ϕ1 , ϕ2 , ϕ3 which are 1 at the nodes (0, 0), (h, 0), (h, h), respec-
tively. Then, on Γ1 , any w ∈ Sh1 (Γ1 ) can be represented by w = χ1 ϕ1 +χ2 ϕ2 +χ3 ϕ3
with χ1 , χ2 , χ3 ∈ R and ϕ1 (x, y) = 1− xh , ϕ2 (x, y) = xh − yh , ϕ3 (x, y) = yh . Further
2
we have θ w = ϕ3 w on Γ1 , Π1 θ w = χ3 ϕ3 on Γ1 and Π1 θ w2L2 (Γ ) = χ32 h12 ,
h2
1

θ w2L2 (Γ ) = 180 10χ12 + 4χ22 + 6χ32 − 11χ1 χ2 + 15χ1χ3 − 9χ2 χ3 .
1
Next we show that there exists a constant C1 > 0 such that
C1 Π1 θ w2L2 (Γ ) ≤ θ w2L2 (Γ ) . (10.107)

1 1
For
180
fC (χ1 , χ2 , χ3 ) := (θ w2L2 (Γ ) − CΠ1 θ w2L2 (Γ ) )
h2 1 1
we find
⎛ ⎞⎛ ⎞
20 −11 15 χ1
∇fC (χ1 , χ2 , χ3 ) = ⎝ −11 8 −9 ⎠ ⎝ χ2 ⎠ .
15 −9 2(6 − 15C) χ3
Due to

20 −11
det(20) = 20 > 0, det = 39 > 0
−11 8
and
⎛ ⎞
20 −11 15
det ⎝ −11 8 −9 ⎠ = 18 − 1170C > 0, if C < 1
65
15 −9 2(6 − 15C)
it is obvious that fC (χ) tends to +∞ for |χ| → +∞ if C < 65 1

. Since ∇fC is a
multi-linear function the only extreme value of fC in R is a minimum at 0. Thus
3
1
fC (χ) ≥ inf fC (χ) = fC (0) = 0 (C < )
χ∈R3 65
1
and therefore (10.107) holds for C1 < 65 .
Now we show that there exists a constant C2 > 0 such that
C2 Π1 θ w2H 1 (Γ ) ≤ θ w2H 1 (Γ ) . (10.108)

0 1 0 1
Analogously as before we find Π1 θ w2 = 12 χ32 and

H01 (Γ1 )
θ w2H 1 (Γ )
0 1

1 2 1 2 1 2 1 1 1
= χ1 + χ2 + χ3 − χ1 χ2 + χ1 χ3 − χ2 χ3
3 3 3 2 2 2
For gC (χ1 , χ2 , χ3 ) := θ w2 − CΠ1 θ w2

H01 (Γ1 ) H01 (Γ1 )
we find
⎛ ⎞⎛ ⎞
2/3 −1/2 1/2 χ1
∇gC (χ1 , χ2 , χ3 ) = ⎝ −1/2 2/3 −1/2 ⎠ ⎝ χ2 ⎠ .
1/2 −1/2 2/3 − C χ3
It is obvious that gC (χ) tends to +∞ for |χ| → +∞ if C < 21 5

. Since ∇gC is a
multi-linear function the only extreme of gC in R is a minimum at 0. Thus
3
5
gC (χ) ≥ inf gC (χ) = gC (0) = 0 (C < )
χ∈R3 21
5
and therefore (10.108) holds for C2 < 21 .
Obviously (10.107) and (10.108) are valid on the whole reference rectangle and
we obtain via interpolation
Π1 θj wH̃ 1/2 (Γ ) ≤ Cθj wH̃ 1/2 (Γ ) , j = 1, . . . , Jν ,

374 10 A-BEM
for a constant
1 5 −1/4
C ≥ (C1 C2 )−1/4 > ( )
65 21
which is independent of h.
10.5.1.1 2-Level Method
We decompose the ansatz space as

& '
Sh1 = span SH
1
∪ Sh,1
1
∪ . . . ∪ Sh,J
1
ν
, (10.109)
1 is the space of continuous linear functions on a coarser mesh with size

where SH
H = 2h and where
1
3 4
Sh,j = span φhj
Let P = PH + Jj ν=1 Pi be the two-level additive Schwarz operator belonging to

the subspace decomposition (10.109) and the bilinear form < W., . >, i.e.,
W PH ϕ, ψ = W ϕ, ψ ∀ψ ∈ SH
1
, ϕ ∈ Sh1
W Ph ϕ, ψ = W ϕ, ψ ∀ψ ∈ Sh,i
1
, ϕ ∈ Sh1 .
Then there holds

Theorem 10.17 There exist constants c1 , c2 > 0 , independent of h, such that for
all u ∈ Sh1
c1 W u, u ≤ W P u, u ≤ c2 W u, u . (10.110)
Proof First we show that for all v ∈ Sh1 we can find a representation v = vH +
vh,1 + . . . + vh,Jν such that

Jν
0 0
vH 2H=1/2 (Γ ) + 0vh,j 02=1/2 ≤ c v2H=1/2 (Γ )
H (Γ )
j =1
This is done by use of cut-off-functions θj and the interpolation operator Π1 .

Let us define the L2 -projection: H =1/2(Γ ) → S 1 by vH = QH v ∈ S 1 (ΓH )
H 1
and vh,j = Π1 (θj wh ) with wh = v − vH . Then supp Π1 θj wh = Γj , and
1 (Γ ). Furthermore
Π1 θj wh ∈ C 0 is piecewise linear. Thus vh,j ∈ Sh,j j
⎛ ⎞

Jν
Jν
vh,j = Π1 ⎝ Θj wh ⎠ = wh
j =1 j =1
since wh and Jj ν=1 θj wh have the same nodal values. Next we define the localiza-
tion operator by
R
1
Λ : Sh,1 1
(Γ ) → SH × 1
Sh,j Γj
j =1

v → (Λv)Jj=0 = QH v, Π1 θ1 wh , . . . , Π1 θJnu wh
Then Lemma 10.6 implies

Jν
0 0
0(Λv)j 02 1/2
= (Γ ) ≤ c vH
2
H =1/2 (Γ )
j =1
yielding the lower bound in (10.110).

To estimate the maximal eigenvalue of P we show for all v ∈ Sh1 (Γ ) and for all
representations v = vH + vh,1 + . . . + vh,Jν :
⎛ ⎞

Jν
0 0
v2H=1/2 (Γ ) ≤ c ⎝vH 2H=1/2 (Γ ) + 0vh,j 02=1/2 ⎠
H (Γ )
j =1
Using the triangle inequality together with a colouring argument we obtain

⎛ 0 02 ⎞
0Jν 0
⎜ 0 0 ⎟
v2H=1/2 (Γ ) ≤ 2 ⎝vH 2H=1/2 (Γ ) + 0
0 v 0
h,j 0 ⎠
0j =1 0 =1/2
H (Γ )
⎛ ⎞

Jν
0 0
≤ 2 ⎝vH 2H=1/2 (Γ ) + c 0vh,j 0 =1/2 ⎠
2
H (Γ ) j
j =1
⎛ ⎞

Jν
0 0
≤ 2 ⎝vH 2H=1/2 (Γ ) + c 0vh,j 0 =1/2 ⎠
2
H (Γ )
j =1
This completes the proof.

Next we prove Theorem 10.11 for the above setting of piecewise linear functions
on a uniform partition γh of Γ into triangles (Fig. 10.7).
376 10 A-BEM
Proof (of Theorem 10.11) The saturation assumption (10.86) yields the equivalence
of norms
uh − uH 1/2 ∼ u − uH 1/2 .
Due to Theorem 10.17 we have
c1 uh − uH 21/2

Jν
≤ PH (uh − uH )21/2 + Pi (uh − uH )21/2 ≤ c2 uh − uH 21/2 .
i=1
First, we observe that since uH and uh satisfiy the Galerkin equation there holds for
any w ∈ SH 1
W PH uh , w = W uh , w = f, w = W uH , w = W PH uH , w .
Hence
PH (uh − uH )21/2 = 0 .
1 . The
The error indicator μj is obtained by solving a linear problem in the space Sh,j
function vh,j = Pj (uh − uH ) ∈ Sh,j
1 solves for any v ∈ S 1 :
h,j
W vh,j , v = f − W uH , v . (10.111)
Hence, firstly one solves (10.111) for 0 ≤ j ≤ Jν and then one computes the terms
μj = W vh,j , vh,j 1/2 . Since Sh,j
1 = span{φhj } is a one-dimensional space, we
have vh,j = cφhj with coefficient
f − W uH , φhj
c= .
W φhj , φhj
Hence
1/2
μj = |c|W φhj , φhj
and there holds with constants c1 , c2 independent of h,

Jν
Jν
c1 μ2j ≤ u − uH 21/2 ≤ c2 μ2j ,
j =1 j =1
which is Theorem 10.11.

10.6 Subspace Decomposition for p-Version BEM 377
10.6 Two-Level Subspace Decomposition for the p-Version

BEM
Here we consider the p-version of the boundary element method for the hypersingu-
lar integral equation in two dimensions. We present from [238] an a-posteriori error
estimate that is based on a stable two-level subspace decomposition of the enriched
ansatz space. The Galerkin error is estimated by inverting local projection operators
that are defined on small subspaces of the second level. We consider an enriched
space on Γ = ∪Jj=1 Γj
H̃N := Sh,p (Γ ) + Z1 + · · · + ZJ , (10.112)
where
Sh,p (Γ ) = {φ ∈ C 0 (Γ ); φ|∂Γ = 0, φ|Γj ∈ Ppj (Γj ), j = 1, . . . , J }.
Here, the local enrichment is given by adding bubble functions on the elements
Zj = span(ψpj +1 ◦ Tj−1 ), j = 1, . . . , J,
with the affine map Tj : I = (−1, 1) → Γj and

A
2j − 1 x
ψj (x) = Lj −1 (t) dt (2 ≤ j ≤ pj ),
2 −1
where Lj −1 is the Legendre polynomial of degree j − 1 and
1−x 1+x
ψ0 (x) = , ψ1 (x) = .
2 2
We use the notation φ̃j = φj ◦ Tj for functions φj defined on Γj .

Now for the hypersingular integral equation on a given curve Γ we consider the
problem: given f ∈ H −1/2(Γ ), find u ∈ H̃ 1/2(Γ ) such that
W u, v L2 (Γ ) = f, v L2 (Γ ) ∀ v ∈ H̃ 1/2(Γ ), (10.113)
where u2W = W u, u L2 (Γ ) .
Lemma 10.7 There exist positive constants c1 and c2 , independent of the mesh and
p, such that

J
J
c1 (1 + log pmax )−2 φj 2W ≤ φ1∗ 2W ≤ c2 φj 2W
j =1 j =1
378 10 A-BEM
for all φ1∗ = J

j =1 φj ∈ Z, where φj ∈ Zj , j = 1, . . . , J . Here, pmax =
max{p1 , . . . , pJ }.
Proof First, we observe that

J
J
φ1∗ 2W ≤ C φ1∗ 2H̃ 1/2 (Γ ) ≤C φj 2H̃ 1/2 (Γ ) ≤C φj 2W .
j
j =1 j =1
On the other hand there holds

J
J
φj 2H̃ 1/2 (Γ ) ≤ C φ̃j 2H̃ 1/2 (I )
j
j =1 j =1

J
J
≤C (1 + log(pj + 1))2 φ̃j 2H 1/2 (I ) ≤ C(1 + log pmax )2 φ̃j 2H 1/2 (I ) .
j =1 j =1
(10.114)
For the first inequality we used that the H̃ 1/2-norm scales with a constant that is
independent of the element size under affine transformations, see Lemma 3.1 in
[405] and Lemma 2 in [233]. The second inequality is Theorem 6.5 in [15]. Note that
φ̃j (−1) = φ̃j (+1) = 0. Therefore, we can apply the Poincaré–Friedrichs inequality
and obtain

J
J
J
φ̃j 2H 1/2 (I ) ≤ C |φ̃j |2H 1/2 (I ) ≤ C |φj |2H 1/2 (Γ ) ≤ C|φ1∗ |2H 1/2 (Γ ) .
j
j =1 j =1 j =1
(10.115)
The last inequality is a direct application of the definition of the Sobolev–Slobo-

-deckij seminorm | · |H 1/2 (Γ ) . Finally, by combining (10.114) and (10.115) and by
using the fact that W is positive definite on H̃ 1/2(Γ ), the proof is finished.

Next we present a two-level subspace decomposition.
Lemma 10.8 (subspace decomposition) There exist positive constants c3 and c4 ,
independent of the mesh and p, such that

J
J
c3 (1 + log pmax )−2 W φj , φj ≤ W φ, φ ≤ c4 W φj , φj
j =0 j =0
J
for all φ = j =0 φj ∈ H̃N , where φ0 ∈ HN and φj ∈ Zj , j = 1, . . . , J .
Proof With Lemma 10.7 we note that there holds

J
W φ, φ = φ0 + φ1∗ 2W ≤ 2 {φ0 2W + φ1∗ 2W } ≤ C W φj , φj ,
j =0
10.6 Subspace Decomposition for p-Version BEM 379
which is the right inequality of the assertion. To prove the left inequality we show
that there exists C > 0 such that
C (φ0 2W + φ1∗ 2W ) ≤ φ2W .
It suffices to show φ0 W ≤ C φW since this implies with φ1∗ = φ − φ0 that
φ1∗ W ≤ C φW by the triangle inequality. To show |φ0 |H 1 (Γ ) ≤ C |φ|H 1 (Γ ) we
prove |φ0 |2H 1 (Γ ) ≤ C |φ|2H 1 (Γ ) and then sum over j .
j j
On I = (−1, 1) :
p+1
φ̃(x) = φ|Γj ◦ Tj (x) = ci ψi (x)
i=0
p
φ̃0 (x) := φ0 |Γj ◦ Tj (x) = c0 ψ0 (x) + c1 ψ1 (x) + ci ψi (x)
i=2
and

φ̃1∗ (x) := φ1∗ |Γj ◦ Tj (x) = cp+1 ψp+1 (x).
A
d φ̃0 1 1 p
2i − 1
= − c0 + c1 + ci Li−1 (x)
dx 2 2 2
i=2
and
A
d φ̃1∗ 2p + 1
= cp+1 Lp (x)
dx 2
d φ̃0 d φ̃1∗
Note dx and dx are orthogonal in L2 (I ). Therefore
|φ|2H 1 (Γ ) = |φ0 |2H 1 (Γ ) + |φ1∗ |2H 1 (Γ ) ,

j j j
which shows that |φ0 |2H 1 (Γ ) ≤ C |φ|2H 1 (Γ ) and thus, |φ0 |H 1 (Γ ) ≤ C |φ|H 1 (Γ ) .
j j
We consider the normalized shape functions ψi∗ := ψi /ψi L2 (I ) , i =
0, . . . , p + 1. Then, we can represent
p
φ̃0 (x) := φ0 |Γj ◦ Tj (x) = v0 ψ0∗ (x) + v1 ψ1∗ (x) + vi ψi∗ (x)
i=2
380 10 A-BEM
and
p+1
φ̃(x) := φ|Γj ◦ Tj (x) = v0 ψ0∗ (x) + v1 ψ1∗ (x) + vi ψi∗ (x).
i=2
Note φ̃0 2L2 (I ) = v T Av, where A is a matrix of dimension (p + 1) × (p + 1) with

entries given by
1
aij = ψi∗ (x)ψj∗ (x) dx i, j = 0, . . . , p
−1
and v T = (v0 , . . . , vp ). We get

⎡ ⎤
1 c0 b0 −b1
⎢ c0 1 ⎥
⎢ b0 b1 ⎥
⎢ b ⎥
⎢ 0 b0 1 0 b2 ⎥
⎢ ⎥
⎢ −b b1 0 1 0 b3 ⎥
A=⎢ 1 ⎥
⎢ b2 0 1 ... ... ⎥
⎢ ⎥
⎢ ... . . . . . . . . . bp−2 ⎥
⎢ ⎥
⎣ ... ... 1 0 ⎦
bp−2 0 1
/ /
where c0 = 12 , b0 = − 5
8 , b1 =− 7
40 and
S
1 (2j − 3)(2j + 5)
bj = − ∀ j ≥ 2.
2 (2j − 1)(2j + 3)
It follows that

A bT
φ̃2L2 (I ) = ũT ũ,
b 1
where ũT = (v0 , . . . , vp , vp+1 ) and b = (0, 0, . . . , bp−1 , 0). To show that
φ0 L2 (Γj ) ≤ C φL2 (Γj ) , we need to bound the maximum eigenvalue of

A0 v A b v
=λ T .
0 0 w b 1 w
Following Pavarino [337], we deduce that |λ| < C < ∞ which shows φ0 2L2 (Γ ) ≤
j
C φ2L2 (Γ ) . This gives φ0 2L2 (Γ ) ≤ C φ2L2 (Γ ) . The assertion now follows by
j
interpolation.

10.7 Convergence of ABEM 381
Application of Lemma(10.8) yields an a posteriori error estimate for the Galerkin

solution with hierarchical error indicators when the following saturation assumption
holds: There exists σ ∈ [0, 1) such that
u − ũN W ≤ σ u − uN W
with the solution u of (10.113) and the corresponding Galerkin solutions uN ∈

Sh,p (Γ ), ũN ∈ H̃N of (10.112).
Theorem 10.18 Assume that the mesh Th is locally quasi-uniform. Then there
exist positive constants c1 and c2 such that for the error estimator based upon the
decomposition (10.112) there holds

J
c2
J
c1 θj2 ≤ u − uN 2W ≤ (1 + log pmax ) 2
θj2 .
1 − σ2
j =1 j =1
Here, uN ∈ Sh,p (Γ ) is the Galerkin approximation of the solution u of (10.113) and

σ is the saturation parameter. θj = Pj (ũN − uN )W , ũN ∈ H̃N is the Galerkin
solution of the enriched space H̃N and Pj : H̃N → Zj is definied by
W Pj φ, ψ = W φ, ψ ψ ∈ Zj .
The corresponding result for the single layer potential operator using Legendre
polynomials as bubble functions can also be found in [238] together with numerical
experiments for p- and hp- adaptive algorithms. In [239] a p-adaptive algorithm with
bubble functions is investigated for the BEM with the hypersingular operator on the
plane screen.
10.7 Convergence of Adaptive BEM for Estimators Without

h-Weighting Factor
In this section we present some results from [176].

A posteriori error estimation and related adaptive mesh-refining algorithms are
one important basis of modern scientific computing. Starting from an initial mesh
T0 and based on a computable a posteriori error estimator, such algorithms iterate
the loop
solve → estimate → mark → refine

(10.116)
to create a sequence of successive locally refined meshes T , corresponding discrete

solutions U , as well as a posteriori error estimators μ . We consider the frame
382 10 A-BEM
of conforming Galerkin discretizations, where T is linked to a finite-dimensional

subspace X of a Hilbert space H with corresponding Galerkin solution U ∈
X , where successive refinement guarantees nestedness X ⊆ X+1 ⊂ H for all
∈ N0 .
Convergence of this type of adaptive algorithm in the sense of
lim u − U H = 0 (10.117)
→∞
has first been addressed in [26] for 1D FEM and [143] for 2D FEM. We note that
already the pioneering work [26] observed that validity of some Céa-type quasi-
optimality and nestedness X ⊆ X+1 for all ∈ N0 imply a priori convergence
lim U∞ − U H = 0, (10.118)

→∞
*
where U∞ is the unique Galerkin solution in X∞ := ∈N0 X . From a conceptual
point of view, it thus only remained to identify the limit u = U∞ . Based on such
an a priori convergence result (10.118), a general theory of convergence of adaptive
FEM is devised in [308, 377], where the analytical focus is on estimator convergence
lim μ = 0. (10.119)
→∞
Moreover, the recent work [79] gives an analytical frame to guarantee convergence
with optimal convergence rates; see also the overview article [174] for the current
state of the art of adaptive BEM. Throughout, it is however implicitly assumed that
the local contributions μ (T ) of the error estimator μ are weighted with the local
mesh-size, i.e., |T |α for some appropriate α > 0, or that μ is locally equivalent to
a mesh-size weighted error estimator.
Our analysis in [176] covers the two-level error estimators for BEM considered
Sections 10.4, 10.5, 10.6 and in [164, 234, 238, 294, 313, 314] or the adaptive FEM-
BEM coupling considered in Subsection 12.3.3 and in [12, 198, 274, 312]. The
local contributions are projections of the computable error between two Galerkin
solutions onto one-dimensional spaces, spanned by hierarchical basis functions.
These estimators are known to be efficient. On the other hand, reliability is only
proven under an appropriate saturation assumption which is even equivalent to
reliability for the symmetric BEM operators [11, 161, 162]. However, such a
saturation assumption is formally equivalent to asymptotic convergence of the
adaptive algorithm [178] which cannot be guaranteed mathematically in general
and is expected to fail on coarse meshes.
Next we take an abstract setting as follows. Let H be a Hilbert space with dual
space H % and A : H → H % be a bi-Lipschitz continuous, not necessarily linear
operator, i.e.
−1
Ccont w − vH ≤ Aw − AvH % ≤ Ccont w − vH for all v, w ∈ H
(10.120)
Here, ·H % , denotes the operator norm on H % ,
|F, v |
F H % = sup for all F ∈ H % (10.121)
v∈H \{0} vH
Suppose that there exists some subspace X00 ⊆ H such that for any given
closed subspace X00 ⊆ X% ⊆ H and any continuous linear functional F ∈ H %
on H the Galerkin formulation
AU% , V% = F, V% for all V% ∈ X% (10.122)
admits a unique solution U% ∈ X% , where ·, · denote the duality bracket between
H and H % . This implies the existence of a unique solution u ∈ H of
Au = F (10.123)
We shall assume that X is a finite-dimensional subspace of H related to

some triangulation T and that U (F ) ∈ X is the corresponding Galerkin
solution (10.122) for X% = X . Starting from an initial mesh T0 , the triangulations
T are successively refined by means of the following realization of (10.116), where
for all E ⊆ T
1/2
μ (F ; E ) := μ (F ; T )2 <∞ and μ (F ) := μ (F ; T )
T ∈E
(10.124)
is a computable a posteriori error estimator. Its local contributions μ (F ; T ) ≥ 0

measure, at least heuristically, the error u(F ) − U (F ) locally on each element
T ∈ T .
Algorithm 10.2 INPUT: Right-hand side F ∈ H ∗ , initial mesh T0 with X0 ⊇
X00 , and parameter 0 < θ ≤ 1.
For = 0, 1, 2, . . . iterate the following:
(i) Compute Galerkin solution U (F ) ∈ X .
(ii) Compute refinement indicators μ (F ; T ) for all T ∈ T .
(iii) Determine some set M ⊆ T of marked elements which satisfies
θ μ (F )2 ≤ μ (F ; M )2 . (10.125)
(iv) Generate a new mesh T+1 and hence an enriched space X+1 by refinement
of at least all marked elements T ∈ M .
OUTPUT: Sequence of successively refined triangulations T as well as correspond-
ing Galerkin solutions U (F ) ∈ X and error estimators μ (F ), for ∈ N0 .
384 10 A-BEM
The convergence results of Propositions 2.4 and 2.5 in [176] require an auxiliary
error estimator
1/2
ρ (F ) := ρ (F ; T ) with ρ (F ; E ) := ρ (F ; T )2 <∞ for all E ⊆ T
T ∈E
(10.126)
with local contributions ρ (F ; T ) ≥ 0. For all ∈ N0 , we suppose that there

exists some set R ⊆ T with M ⊆ R which satisfies the following three
assumptions (A1)–(A3):
(A1). μ (F ) is a local lower bound of ρ (F ): There is a constant C1 > 0 such that
for all ∈ N0 holds
μ (F ; M) ≤ C1 ρ (F ; R ). (10.127)
(A2). ρ (F ) is contractive on R : There is a constant C2 > 0 such that for all

, m ∈ N0 and all δ > 0 holds
1
C2−1 ρ (F ; R )2 ≤ ρ (F )2 − ρ+m (F )2
1+δ
+ (1 + δ −1 )C2 U+m (F ) − U (F )2H . (10.128)
The constants C1 , C2 > 0 may depend on F , but are independent of the level ∈
N0 , i.e., in particular independent of the discrete spaces X and the corresponding
Galerkin solutions U (F ). If ρ (F ) is not well-defined for all F ∈ H ∗ , but only on
a dense subset D ⊆ H ∗ , we require the following additional assumption:
(A3). μ (·) is stable on M with respect to F : There is a constant C3 > 0 such that
for all ∈ N0 and F ∈ H ∗ holds
|μ (F ; M ) − μ (F ; M)| ≤ C3 F − F ∗H . (10.129)
Some remarks are in order to relate the abstract assumptions (A1)–(A3) to the
applications we have in mind.
Choice of ρ μ (F ) being the two-level error estimator for BEM considered in
Sections 10.4, 10.5, 10.6 , see also [11, 161, 162, 164, 234, 238, 294, 313, 314]
and for the FEM-BEM coupling in Subsection 12.3.3, see also [12, 198, 312].
ρ (F ) denotes some weighted-residual error estimator, see [76, 86, 87, 92, 93] and
Sects. 10.1–10.4 for BEM and [8, 91, 198] and Subsection 12.3.2 for the FEM-BEM
coupling.
Necessity of (A3) In these cases, the weighted-residual error estimator ρ imposes
additional regularity assumptions on the given right-hand side F . For instance, the
weighted-residual error estimator for the weakly singular integral equation [76, 87,
92, 93] requires F ∈ H 1 (Γ ), while the natural space for the residual is H 1/2 (Γ ).
Convergence (10.119) of Algorithm 10.2 for arbitrary F ∈ H 1/2(Γ ) then follows
by means of stability (A3).
Verification of (A1)–(A2) For two-level estimators, (A1) has first been observed
in [86] for BEM and follows essentially from scaling arguments for the hierarchical
basis functions. Finally, the novel observation (A2) follows from an appropriately
constructed mesh-size function and refinement of marked elements as well as
appropriate inverse-type estimates, where we shall build on the recent developments
of [9]; see e.g. the proof of Theorem 10.19 in [176].
Verification of (A3) Suppose that the operator A is linear and μ (·) is efficient
μ (F ) ≤ Ceff u(F ) − U (F )H for all F ∈ H ∗ . (10.130)
Provided μ (·) has a semi-norm structure, the corresponding triangle inequality

yields
μ (F ) ≤ μ (F ) + μ (F − F ) ≤ μ (F ) + Ceff u(F − F ) − U (F − F )H

≤ μ (F ) + Ceff Ccea u(F − F )H
≤ μ (F ) + Ceff Ccea A−1 F − F H ∗ ,
(10.131)
where A−1 denotes the operator norm of A−1 , and the (bounded) inverse exists
due to (10.120). This proves stability (A3) with C3 = Ceff Ccea A−1 .
As a model problem we consider the weakly singular integral equation

Au(x) = −2 G(x − y) u(y) dΓ (y) = F (x) for all x ∈ Γ (10.132)
Γ
on a relatively open, polygonal part Γ ⊆ ∂Ω of the boundary of a bounded,

polyhedral Lipschitz domain Ω ⊂ Rd , d = 2, 3. For d = 3, we assume that the
boundary of Γ (a polygonal curve) is Lipschitz itself. Here,
1 1
G(z) = ln |z| resp. G(z) = − |z|−1 (10.133)
2π 4π
denotes the fundamental solution of the Laplacian in d = 2, 3. The reader is referred
to Chapter 2, Sections 2.3 and 2.4for proofs of and details on the following facts:
The singe layer integral operator A : H → H ∗ is a continuous linear operator
between the fractional-order Sobolev space H = H =−1/2(Γ ) and its dual H ∗ =
H (Γ ) := {>
1/2 v |Γ : >
v ∈ H (Ω)}. Duality is understood with respect to the extended
1
L2 (Γ )-scalar product ·, · . In 2D, we additionally assume diam (Ω) < 1 which
can always be achieved by scaling. Then, the single layer integral operator is also
elliptic, i. e.
v, Av ≥ Cell v2H=−1/2 (Γ ) =−1/2(Γ )

for all v ∈ H = H (10.134)
386 10 A-BEM
with some constant Cell > 0 which depends only on Γ . Thus, A meets all
assumptions of the abstract setting at the beginning of this section, and v2A :=
Av, v even defines an equivalent Hilbert norm on H .
Next we introduce the discretization. Let T% be a γ -shape regular triangulation
of Γ into affine line segments for d = 2 resp. plane surface triangles for d = 3. For
d = 3, γ -shape regularity means
diam (T )2
sup ≤γ <∞ (10.135a)
T ∈T% |T |
with | · | being the two-dimensional surface measure, whereas for d = 2, we impose

uniform boundedness of the local mesh-ratio
diam (T )
≤ γ < ∞ for all T , T ∈ T% with T ∩ T = ∅. (10.135b)
diam (T )
To abbreviate notation, we shall write |T | := diam (T ) for d = 2. In addition, we

assume that T% is regular in the sense of Ciarlet for d = 3, i.e., there are no hanging
nodes.
With X% = P 0 (T% ) being the space of T% -piecewise constant functions, we
now consider the Galerkin formulation (10.122).
A weighted residual error estimator (see Sections 10.1–10.4) is our next concern.
According to the Galerkin formulation (10.122), the residual F − AU% (F ) ∈
H 1/2(Γ ) has T% -piecewise integral mean zero, i.e.,

(F − AU% (F )) dΓ = 0 for all T ∈ T% . (10.136)
T
Suppose for the moment that the right-hand side has additional regularity F ∈
H 1 (Γ ) ⊂ H 1/2(Γ ). Since A : H =−1/2(Γ ) → H 1/2(Γ ) is an isomorphism with
=−1/2+s (Γ ) → H 1/2+s (Γ ) for all −1/2 ≤ s ≤ 1/2 (We
additional stability A : H
note that A is not isomorphic for s = ±1 and Γ ∂Ω.), a Poincaré-type inequality
in H 1/2(Γ ) shows
=−1/2(Γ ) 0 F − AU% (F )H 1/2 (Γ )

u(F ) − U% (F )H
(10.137)
1/2
h% ∇Γ (F − AU% (F ))L2 (Γ ) =: η% (F ),
see [76, 87, 92, 93]. Here, ∇Γ (·) denotes the surface gradient, and h% ∈ P 0 (T% )
is the local mesh-width function defined pointwise almost everywhere by h% |T :=
diam (T ) for all T ∈ T% . Overall, this proves the reliability estimate
=rel η% (F ),
u(F ) − U% (F )H=−1/2 (Γ ) ≤ C (10.138)
and the constant C=rel > 0 depends only on Γ and the γ -shape regularity (10.135) of
=rel = C ln1/2 (1 + γ ), where C > 0 depends
T% ; see [87]. In 2D, it holds that C
only on Γ ; see [76]. In particular, the weighted-residual error estimator can be
localized via
1/2
η% (F ) = η% (F ; T )2
T ∈T% (10.139)
with η% (F ; T ) = diam (T )1/2∇Γ (F − AU% (F ))L2 (T ) .
Recently, convergence of Algorithm 10.2 has been shown even with quasi-optimal
rates, if η (F ) = μ (F ) is used for marking (10.125); see [175, 177]. We stress
that our approach with η (F ) = ρ (F ) = μ (F ) would also give convergence
η (F ) → 0 as → ∞. Since this is, however, a much weaker result than that
of [177], we omit the details.
Unlike reliability (10.138) of η% (F ) which is proved for general F ∈ H 1 (Γ ),
efficiency η% (F ) u(F ) − U% (F )H=−1/2 (Γ ) is only known for special right-hand
sides F ∈ H 1 (Γ ) which guarantee equivalence of the weakly singular integral
equation (10.132) to some 2D Laplace problem
−ΔU = 0 in Ω ⊂ R2 subject to U = g on Γ = ∂Ω
with smooth Dirichlet data g; see [74] for quasi-uniform meshes and the very
recent work [10] for the generalization to locally refined meshes which are γ -shape
regular (10.135b), see also Sect. 10.2.2.
Next we consider a two-level error estimator. In the frame of weakly singular
integral equations (10.132), the two-level error estimator was introduced in [314],
Sect. 10.5. Let T>% denote the uniform refinement of T% . For each element T ∈
T% , let T>% |T := {T ∈ T>% : T ⊂ T } denote the set of sons of T . Let
{χT , ϕT ,1 , . . . , ϕT ,D } be a basis of P 0 (T>% |T ) with fine-mesh functions ϕT ,j which
satisfy supp (ϕT ,j ) ⊆ T and T ϕT ,j dΓ = 0. We note that usually D = 1 for d = 2
and D = 3 for d = 3. Typical choices are Fig. 10.4 and Fig. 10.8. Then, the local
contributions of the two-level error estimator from [161, 164, 238, 294, 314] read

D
F − AU% (F ), ϕT ,j
μ% (F ; T )2 = μ%,j (F ; T )2 with μ%,j (F ; T ) = 1/2
.
j =1 AϕT ,j , ϕT ,j
(10.140)
Put differently, we test the residual F − AU% (F ) ∈ H 1/2(Γ ) with the additional
basis functions from P 0 (T>% )\P 0 (T% ). This quantity is appropriately scaled by the
corresponding energy norm ϕH=−1/2 (Γ ) 0 Aϕ, ϕ 1/2 = ϕA .
388 10 A-BEM
+1 −1 −1 +1 −1 −1
+1
+1 −1 +1 −1 +1 +1
χT T,1 T,2 T,3
Fig. 10.8 For d = 3, uniform bisection-based mesh-refinement usually splits a coarse mesh
element T ∈ T (left) into four sons T ∈ T> (right) so that |T |/4 = |T |. Typical hierarchical
basis functions ϕT ,j are indicated by their piecewise constant values ±1 on the son elements T
Theorem 10.19 ([176]) Suppose that the two-level error estimator (10.140) is
used for marking (10.125). Suppose that the mesh-refinement guarantees uniform
γ -shape regularity (10.135) of the meshes T generated, as well as that all marked
elements T ∈ M are refined into sons T ∈ T+1 with |T | ≤ κ |T | with some
uniform constant 0 < κ < 1. Then, Algorithm 10.2 guarantees
μ (F ) → 0 as → ∞ (10.141)
for all F ∈ H 1/2 (Γ ).

For the corresponding results on the convergence of adaptive schemes for
the hypersingular operator and the Johnson-Nedelec coupling and the symmetric
coupling of Mund and Stephan in [312] see [176].
For further reading we suggest [89] and [13].
Chapter 11
BEM for Contact Problems
In literature we find various finite element discretization schemes that tackle

variational inequalities that arise from scalar unilateral Signorini problems and
from contact problems without and with friction in solid mechanics, see e.g.
[199, 249, 266]. Each scheme has to overcome several challenges, mainly the
discretization of a cone, a primal one in variational inequalities or a dual one in
mixed methods, the non-differentiability of the friction functional in the classical
sense and the reduced regularity of the solution at the a priori unknown free
boundary/interface from contact to non-contact and from stick to slip.
In many cases the insufficient resolution of these interfaces is the dominant
source of error. As they lie on the boundary only, it seems to be favorable to reduce
these nonlinear boundary value problems to boundary variational inequalities as
shown in Chap. 5 and use the boundary element method. Thereby one only requires
a boundary mesh.
This chapter starts with the h-BEM for scalar Signorini problems with a first
convergence result. Then more advanced hp-versions of the BEM for boundary vari-
ational inequalities arising from frictional contact are treated. The chapter concludes
with the study of h-BEM and hp-BEM for nonmonotone contact problems from
delamination of adhesively bonded interfaces in material science.
Mixed methods are typically studied for low order finite element schemes where
the discrete Lagrange multiplier is a piecewise constant or linear discontinuous
function which allows a conforming discretization of the sign and box constraints
of the Lagrange multiplier. For higher polynomial degrees these conditions are no
longer easy to realize and are only satisfied in a discrete sense. Therefore it is
important that the missing conformity is captured in the a posteriori error estimates.
Now as shown in Sect. 11.4.2 the a posteriori error estimate can be localized
(approximately) and therefore used to recover/improve the convergence rate within
an adaptive scheme compared to a uniform scheme which has reduced convergence
rate due to the reduced regularity from the contact conditions.

390 11 BEM for Contact Problems
In adaptive schemes a sequence of solutions must be computed. Hence it is

important that the convergence rate of the adaptive scheme is significantly higher
than that of the uniform scheme. Here, hp-adaptive schemes show their superiority
over pure h-adaptive schemes.
The a posteriori error estimate strategy of Theorem 11.4 gives an upper and
lower bound for the discretization error under the saturation assumption, but a global
variational inequality over a p-enriched boundary element space needs to be solved
for the evaluation of the error estimate. Whereas the a posteriori error estimate
strategy of Theorem 11.10 gives a much easier to evaluate upper bound for the
discretization error, a proof of a corresponding lower bound seems possible only for
lowest order discretizations.
11.1 h-BEM for the Signorini Problem
11.1.1 Discretization of the Boundary Variational Inequality
In the following we start from the variational formulation of the scalar Signorini
problem as a boundary variational inequality given in the Sect. 5.1. We suppose
that for simplicity Ω ⊂ R2 is polygonal, but not necessarily convex. Let Γ be
represented by
xi = Xi (s) , 0 ≤ s ≤ L (i = 1, 2)
with Xi (0) = Xi (L) (i = 1, 2) . We partition Γ into finitely many segments by the

points Pj = (X1 (sj ), X2 (sj )) , j = 1, . . . , J , where the endpoints of Γ D and Γ S
are included and where s1 = 0 , sJ +1 = L . The partitioning of Γ is characterized
by the mesh size
h := max |sj +1 − sj | .
j =1,...,J
Note that the boundary variational inequality (5.13) splits into a variational equality
in H −1/2(Γ ) , which can be discretized in a standard way, and a novel variational
inequality in the convex cone K ⊂ H 1/2(Γ ) . As an important issue we want to
treat not only piecewise linear, but also piecewise quadratic and piecewise cubic
approximations of K . To this end we introduce the space P κ of polynomials of
degree less than or equal to κ (κ = 1, 2, 3) and the subsequent finite point sets:
Σ1h := {sj : j = 1, . . . , J }
Σ2h := {s ∈ (0, L) : s is a midpoint of an interval (sj , sj +1 )

for some j = 1, . . . , J }
Σ3h := {s ∈ (0, L) : s divides an interval (sj , sj +1 ) by the ratio 1:2}

11.1 h-BEM for the Signorini Problem 391
Moreover,
Π1h := {Pj : j = 1, . . . , J } ∩ Γ S ,
where with appropriate 1 ≤ j0 ≤ j1 < J ,
Π1h = {Pj : j = j0 , . . . , j1 + 1} ,
and for κ = 2, 3 we set

& '
Πκh := P = (X1 (s), X2 (s)) : s ∈ Σ1h ∪ Σκh ∩ Γ S .
Recall
HΓ1D ,0 (Ω) = {v ∈ H 1 (Ω) : v|ΓD = 0} ,
introduce the trace space
1/2
HΓD ,0 (Γ ) = {v ∈ H 1/2 (Γ ) : v|ΓD = 0}
which can be approximated by the finite dimensional subspace

&
h
Uκ,μ := v h ∈ C μ (Γ ) : v h ◦X | (sj , sj +1 ) ∈ P κ
(j = 1, . . . , J ) ; v h |ΓD = 0} .
Further approximate the convex cone K by the convex cone

h
Kκ,μ := {v h ∈ Uκ,μ
h
: v h (P ) ≤ 0 (∀P ∈ Πκh )} ,
imposing only finitely many inequality constraints, and H −1/2(Γ ) by the finite
dimensional subspace
&
h
Φκ−1,μ−1 := ψ h ∈ C μ−1 (Γ ) : ψ h ◦X | (sj , sj +1 ) ∈ P κ−1
'
(j = 1, . . . , J ) .
Here μ ∈ N0 with μ ≤ κ − 1 and C −1 (Γ ) denotes the space of discontinuous

h
functions. Note that K1,0 ⊂ K holds for all h > 0 .
Thus we are led to the following discretized variational problem:

Find [uh , ϕ h ] ∈ Kκ,μ
h × Φh
κ−1,μ−1 such that
⎧
⎪ h h h
⎨ a du , dv − du + b(ϕ h , v h − uh ) ≥ (v h − uh ) ∀v h ∈ Kκ,μ
h
,
(πκh ) ds ds ds
⎪
⎩ a(ψ , ϕ ) = b(ψ , u ) ∀ψ h ∈ Φκ−1,μ−1
h h h h h
;
or equivalently
A([uh , ϕ h ] , [v h , ψ h ] − [uh , ϕ h ]) ≥ (v h − uh )
∀[v h , ψ h ] ∈ Kκ,μ
h
× Φκ−1,μ−1
h
. (11.1)
Let us remark that the condition (C.28) in the Appendix guarantees the existence
and uniqueness of not only the solution [u, ϕ] of the problem (π) , but also of the
solution [uh , ϕ h ] of the approximate problems (πκh ) because our discretization does
not affect the linear form .
11.1.2 The Convergence Result
Theorem 11.1 Let solutions [u, ϕ] to (π) and [uhκ , ϕκh ] to (πκh ) (h > 0) exist.
Assume that the solution [u, ϕ] is unique. Then for κ = 1, 2, 3
lim [uhκ , ϕκh ] − [u, ϕ] H 1/2 (Γ )×H −1/2 (Γ ) = 0

h→0
Proof In virtue of Lemma 5.2, the bilinear form A(·, ·) satisfies the Gårding
inequality (5.16). Therefore the convergence theorem C.7 in the Sect. C.3.2 in
Appendix C applies and requires the following hypotheses:
H1 If {v h }h>0 weakly converges to v , where v h ∈ K h := Kκ,μ
h , then v ∈ K .
H2 There exist a subset M ⊂ H (Γ ) such that M = K and mappings

1/2
ρ h : M → U h = Uκ,μ h with the property that, for each w ∈ M , ρ h w strongly
converges to w (as h → 0+) and ρ h w ∈ K h for all 0 < h < h0 (w) .

We note that the analogous hypotheses for the approximation of ψ ∈ H −1/2(Γ ) by
ψ h ∈ Φ h are trivally satisfied in view of Φ h ⊂ H −1/2(Γ ) and well-known density
and approximation properties.
Verification of (H1). Since K1,μ h is contained in the weakly closed set K for all
h > 0 , we have only to consider the cases κ = 2 and κ = 3 with μ ∈ N0 such that
μ ≤ κ − 1.
11.1 h-BEM for the Signorini Problem 393
Let the polygonal boundary part Γ S be partitioned by

j1
ΓS = [Pj , Pj +1 ] ,
j =j0
where the closed line segment [Pj , Pj +1 ] has the intermediate point Pj + 1 ∈ Π2h ,
2
respectively the two intermediate points Pj + 1 , Pj + 2 ∈ Π3h . For any ψ ∈ C 0 (Γ S )
3 3
with ψ ≥ 0 we define

j1
ψh = ψ(Pj + 1 ) χ ,
2 j+ 1
2
j =j0
where χ denotes the characteristic function of the open segment ]Pj , Pj +1 [ .

j + 12
Then ψ h ≥ 0 on ΓS (κ = 2, 3) and by the uniform continuity of ψ on Γ S
lim ψ h − ψL∞ (ΓS ) = 0 . (11.2)

h→0
1/2
Now let {v h }j >0 be a family weakly convergent to v ∈ HΓD ,0 (Γ ) , where v h ∈ Kκ,μ
h
(h > 0 ; κ = 2 or κ = 3) . Since the embedding H 1/2(Γ ) ⊂ L1 (ΓS ) is weakly

continuous, the functions v h converge weakly to v in L1 (ΓS ) and are norm bounded.
Therefore by the estimate

| (v h ψ h − vψ)ds| ≤ v h L1 (ΓS ) ψ h − ψL∞ (ΓS ) + | (v h − v)ψ ds| ,
ΓS ΓS
using (11.2) and ψ ∈ L∞ (ΓS ) = (L1 (ΓS ))∗ , we obtain that

lim v h ψ h ds = vψ ds . (11.3)
h→0 ΓS ΓS
From Simpson’s rule it follows for v h ∈ K2,μ

h and all ψ ∈ C 0 (Γ ) with ψ ≥ 0 that
j1
sj+1
2
v h ψ h ds = ψ(Pj + 1 ) (v h ◦X)(s) ds
ΓS 2
j =j0 sj i=1
1
j1 - .
= ψ(Pj + 1 )(sj +1 − sj ) v h (Pj ) + 4v h (Pj + 1 ) + v h (Pj +1 )
6 2 2
j =j0
≤ 0, (11.4)
whereas from Newton’s pulcherrima quadrature rule [226, §7.1.5]) for v h ∈ K3,μ
h
-
1
j1
v ψ ds =
h h
ψ(Pj + 1 ) (sj +1 − sj ) v h (Pj ) + 3v h (Pj + 1 )
ΓS 8 2 3
j =j0
.
+3v h (Pj + 2 ) + v h (Pj +1 ) ≤ 0 . (11.5)
3
Combining (11.3) and (11.4), respectively (11.5) we obtain that for all ψ ∈ C 0 (Γ S )
with ψ ≥ 0

vψ ds ≤ 0 ,
ΓS
hence v ≤ 0 almost everywhere on ΓS or v ∈ K . This proves (H1).

Verification of (H2). In virtue of (5.19), we can take M = K ∩ C ∞ (Γ ) . Now we
define ρκh : H 1/2(Γ )∩C ∞ (Γ ) → Uκ,κ−1
h ⊆ Uκ,μ
h by L-periodic spline interpolation
subordinated to the partitioning of Γ . Thus in particular
ρκh w(P ) = w(P ) , ∀P ∈ Πκh (κ = 1, 2, 3) .
h
Hence ρκh w belongs to Kκ,κ−1 ⊆ Kκ,μ
h for any w ∈ M , since μ ≤ κ − 1 . Moreover
by spline interpolation theory, Uκ,κ−1 is a regular family of finite elements in the

sense of Babuška and Aziz [14] and therefore we have
w − ρκh wH 1/2 (Γ ) ≤ chκ−1/2 wH κ (Γ ) (κ = 1, 2, 3)
with c > 0 independent of h and w . Hence we conclude that
lim w − ρκh wH 1/2 (Γ ) = 0 , ∀w ∈ M ; κ = 1, 2, 3 .

h→0

Remark 11.1 By the proof above (see especially the estimates (11.4) and (11.5)) we
have shown that boundary element convergence holds true for arbitrary piecewise
polynomial approximations as long as the corresponding Newton-Cotes quadrature
formula has positive weights. This is a reasonable restriction for practical compu-
tations and is satisfied for the Newton-Cotes formulae up to the order κ = 8 [160,
§6.2.1].
In this section we considered the simplest elliptic equation. However the method
presented can be extended to the more general problems of unilateral contact
involving the Navier - Lame -system of linear elasticity, see [222]. A priori error
estimates with linear boundary elements are provided in [292, 386, 387]. Nitsche
type error estimates for variational inequalities are derived in [393].
11.2 hp-BEM with Hierarchical Error Estimators 395
11.2 hp-BEM with Hierarchical Error Estimators for Scalar

Signorini Problems
In this section we report from [297] a priori and a posteriori error estimates for the
hp− discretization of a boundary integral formulation of the Signorini problem of
the Laplacian. We present hp− convergence results for the BEM Galerkin solution
in the energy norm. The a priori error estimate shows O h 4 p− 4 convergence
1 1
rate and corresponds to the FEM result (originally derived by Falk in [173] for the
h version). The presented a posteriori error estimate is efficient and reliable. The
hierarchical error estimators used are computed by enriching the boundary element
spaces by bubble functions on each element. This enrichment defines two-level
subspace decompositions with corresponding additive Schwarz operators where the
latter have condition numbers which depend only logarithmically on the polynomial
degrees. For ease of reading, we present most of the proofs from [297]. Numerical
experiments in [297] show that a three-step adaptive algorithm (steered by the
hierarchical error estimators) leads to appropriate mesh refinement and reasonable
polynomial degree distribution. For extension to friction problems see [103, 222].
Let Ω ⊂ Rn , n ≥ 2 be a bounded domain with Lipschitz boundary Γ = ∂Ω
which is a disjoint union of ΓD , ΓN and ΓS = ∅. We consider the following problem.
Given h ∈ H −1/2(ΓN ∪ΓS ), g ∈ H 1/2 (ΓD ∪ΓS )∩C 0 (Γ¯D ∪ Γ¯S ), find û ∈ H 1 (Ω)
such that
Δû = 0 in Ω,
û = g on ΓD ,
(11.6)
∂n = h
∂ û
on ΓN ,
û ≤ g, ∂ û
∂n ≤ h, (û − g)( ∂∂nû − h) = 0 on ΓS
As we have seen in Sect. 5.1 this problem can be formulated equivalently as a

variational inequality over Ω with the convex subset K̂ := {v̂ ∈ H 1 (Ω) : v̂|ΓD =
g|ΓD , v̂|ΓS ≤ g|ΓS } ⊂ H 1 (Ω): Find û ∈ K such that

∇ û · ∇(v̂ − û) dx ≥ h(v̂ − û) ds ∀v̂ ∈ K̂. (11.7)
Ω ΓN ∪ΓS
The variational inequality (11.7) has a unique solution if ΓD = ∅ or ΓN ∪ΓS h ds <

0 (cf. [386] and Appendix C.3.1). We reformulate (11.6) as an equivalent variational
inequality involving only boundary integral operators on Γ .
With the integral operators of the single layer, double layer potential and their
normal derivatives V , K, K , W respectively we have the problem (L):
Find (u, ϕ) ∈ K Γ × H −1/2(Γ ) such that
u, W (v − u) + ϕ, (I + K)(v − u) ≥ 2l(v − u)

ψ, V ϕ − ψ, (I + K)u = 0 ∀(v, ψ) ∈ KΓ × H −1/2(Γ )
(11.8)
where
K Γ := {v ∈ H 1/2 (Γ ) : v|ΓD = g|ΓD , v|ΓS ≤ g|ΓS } (11.9)
and

l(v) = hv ds.
ΓN ∪ΓS
System (L) can be rewritten with the coercive and non-symmetric bilinear form
B(u, ϕ; v, ψ) := W u, v + (I + K)t ϕ, v + V ϕ, ψ − (I + K)u, ψ
as: Find (u, ϕ) ∈ K Γ × H −1/2(Γ ) such that
B(u, ϕ; v − u, ψ) ≥ L (v − u, ψ) ∀(v, ψ) ∈ K Γ × H −1/2 (Γ ) (11.10)
where

L (v, ψ) := 2 hv ds. (11.11)
ΓN ∪ΓS
On the other hand, eliminating ϕ in (L) leads to the equivalent problem (S): Find
u ∈ K Γ such that
Su, v − u ≥ l(v − u) ∀v ∈ K Γ (11.12)
with the symmetric Poincaré–Steklov operator S for the interior problem
1
S := (W + (K + I )V −1 (K + I )) : H 1/2(Γ ) → H −1/2 (Γ ) (11.13)
2
which is positive definite on H 1/2(Γ )/R. Existence and uniqueness of the solution
of problems (S) and (L), respectively, have been shown by Houde Han [227]; for the
corresponding elasticity problem see Gwinner and Stephan [222].
Let ωh , γh be two not necessarily identical regular partitions of Γ , such that all
corners of Γ and all “end points” Γ¯S ∩ Γ¯N , Γ¯N ∩ Γ¯D , Γ¯D ∩ Γ¯S are nodes of ωh , γh .
Let p = (pe )e∈ωh , or q = (qe )e∈γh be degree vectors which associate each
element of ωh or γh with a polynomial degree pe ≥ 1 or qe ≥ 0.
On the interval [−1, 1] we choose N + 1 Gauss-Lobatto quadrature points, i.e.
the points ξjN+1 , 0 ≤ j ≤ N, that are the zeros of (1 − ξ 2 )LN (ξ ), where LN
denotes the Legendre polynomial of degree N. It is known (cf. [46, Prop. 2.2, (2.3)])
that there exist positive weight factors jN+1 := 1
2 N+1 such that ∀φ ∈
N(N+1)LN (ξj )
N 1
P2N−1 ([−1, 1]) : N+1
j =0 φ(ξj ) j = −1 φ(ξ ) dξ .
By an affine transformation we define the set of Gauss-Lobatto points Ge,hp on

each element e of*the partition ωh of Γ , corresponding to the polynomial degree pe
and set Ghp := e∈ωh Ge,hp .
On the partition ωh of Γ we introduce σhp as the space of continuous polynomi-
als with σhp |e ⊆ Ppe (e), ∀e ∈ ωh . A suitable basis of σhp is given by the Lagrange
interpolation polynomials on the set of Gauss-Lobatto points of each element, see
Fig. 11.1, On the partition γh of Γ we introduce τhp as the space of piecewise
polynomials with τhp |e ⊆ Pqe (e), ∀e ∈ γh ; here a suitable basis is given by the
Legendre polynomials.
Therefore, our test and trial spaces are defined as
σhp := {vhp ∈ C 0 (Γ ; R) : vhp |e ∈ Ppe (e), ∀e ∈ ωh } (11.14)

τhp := {ψhp ∈ L2 (Γ ; R) : ψhp |e ∈ Pqe (e), ∀e ∈ γh }. (11.15)
Now, we chose
Γ := {v ∈ σ | ∀x ∈ G ∩ Γ : v(x) ≤ g(x) and ∀x ∈ G ∩ Γ : v(x) = g(x)},

Khp hp hp S hp D
(11.16)
Γ ⊆ K Γ for p ≥ 2 (as Fig. 11.1

which is a convex, closed subset of σhp . Note that Khp
illustrates) or for non-concave gap function g.
Based on the local set of Gauss-Lobatto points Ge,hp we define the local
interpolation operator ie,pe : C 0 (ē) → Ppe (e) by
(ie,pe ψ)(x) = ψ(x) ∀x ∈ Ge,hp , ∀ψ ∈ C 0 (ē)
Fig. 11.1 Lagrange Polynomials , N = 5

and the global interpolation operator iωh ,p : C 0 (Γ ) → σhp by

iωh ,p ψ := χe ie,pe ψ|ē ∀ψ ∈ C 0 (Γ ) (11.17)
e∈ωh
where χe is the characteristic function of e ∈ ωh .

With the above choices, the hp-version of (L) is given by the discrete problem
(Lhp ):
Find (uhp , ϕhp ) ∈ Khp
Γ ×τ
hp such that
B(uhp , ϕhp ; v − uhp , ψ) ≥ 2l(v − uhp ) ∀(v, ψ) ∈ Khp

Γ
× τhp . (11.18)
Via the canonical imbeddings jhp : σhp *→ H 1/2(Γ ) and khp : τhp *→ H −1/2(Γ )
∗ , and k ∗ the discrete Poincaré-Steklov operator S
and their duals jhp hp hp : σhp →
∗
σhp
1 ∗ ∗
Shp := (j Wjhp + jhp (I + K )khp (khp
∗
V khp )−1 khp
∗
(I + K)jhp ) (11.19)
2 hp
is well defined (see [73]). Now the hp-version of (S) given by the discrete problem
(Shp ) for the general hp-version (11.12) reads: Find uhp ∈ Khp
Γ such that
Shp uhp , vhp − uhp ≥ l(jhp vhp − jhp uhp ) ∀vhp ∈ Khp
Γ
. (11.20)
There holds the following convergence result for the Galerkin solution of (Lhp )
in the energy norm without any regularity assumptions.
Theorem 11.2 ([297]) Let (ωh , γh )h∈I be a family of quasi uniform meshes, such
that h := max{|e|, e ∈ ωh or e ∈ γh }, where I ⊂ (0, ∞) with 0 ∈ I . Let p =
(pe )e∈ωh , such that pe = p for all e ∈ ωh , and let q = (qe )e∈γh , such that qe = p−1
for all e ∈ γh .
Let the solutions (u, ϕ) of (11.10) and (uhp , ϕhp ) of (11.18) exist uniquely.
Suppose that for the polygonal domain Ω, there are only finite number of end points
Γ¯S ∩ Γ¯D , Γ¯D ∩ Γ¯N , Γ¯N ∩ Γ¯S . Then there holds

lim u − uhp , ϕ − ϕhp H 1/2 (Γ )×H −1/2 (Γ ) = 0, if h fixed
p→∞
and

lim u − uhp , ϕ − ϕhp H 1/2 (Γ )×H −1/2 (Γ ) = 0, if p fixed
h→0
For its proof see the proof of Theorem 11.6 in the subsequent section that treats
the discretization of a friction-type functional in addition.
If we assume higher regularity of the solution (u, ϕ) of (L) and of the contact
functon g in (11.6), i.e u, g ∈ H 3/2(Γ ), ϕ ∈ H 1/2(Γ ) we obtain the following a
priori error estimate which proposes a convergence rate of O(h1/4 p−1/4 ).
Theorem 11.3 ([297]) Let (ωh , γh )h be a family of quasi uniform meshes, such
that h := max{|e|, e ∈ ωh or ∈ γh }and let p = (pe )e∈ωh , pe = p , q = (qe )e∈γh ,
qe = p − 1. Let (u, ϕ) ∈ K Γ × H 1/2(Γ ) be a solution of problem (L) and let
(uhp , ϕhp ) ∈ Khp
Γ ×τ
hp be solution of problem (Lhp ) and assume u, g ∈ H
3/2 (Γ )
and h∗ − Su ∈ L2 (Γ ), with h∗ |ΓN ∪ΓS = h and h∗ |ΓD = 0. Then there holds

u − uhp , ϕ − ϕhp H 1/2 (Γ )×H −1/2 (Γ ) ≤ Ch1/4 p−1/4 uH 3/2 (Γ )
for p → ∞ or h → 0.
For its proof see the proof of Theorem 11.7 in the subsequent section that treats
the discretization of a friction-type functional in addition.
Next we consider a hierarchical boundary element method, where we extend the
finite dimensional spaces σhp and τhp by bubble functions on each element in ωh
and γh .
Each element e ∈ ωh is associated with a polynomial degree pe and the affine
mapping Fe with Fe (ξ ) = x(ξ ) ∈ e for ξ ∈ [−1, 1].
With the Legendre polynomial Lj of degree j we set
A
1−ξ 1−ξ 2j − 1 ξ
ψ0 (ξ ) := , ψ1 (ξ ) := , ψj (ξ ) := Lj −1 (t) dt, 2 ≤ j,
2 2 2 −1
and take
σe = span{ψe,pe +1 }, ψe,j (x) := ψj (Fe−1 (x)). (11.21)
On the other hand, each element e ∈ γh is associated with a polynomial degree qe .

Setting
A
1 2j + 1
φ0 (t) = , φj (t) := Lj (t), 1 ≤ j,
2 2
we take
τe = span{φe,qe +1 }, φe,j (x) := φj (Fe−1 (x)). (11.22)
In this way we obtain the subspace decompositions

σh,p+1 := σhp ⊕ Lp , Lp := σe , (11.23)
e∈ωh

τh,p+1 := τhp ⊕ λp , λp := τe . (11.24)
e∈γh
Hence the polynomial degree vector associated with σh,p+1 is (pe + 1)e∈ωh and the
polynomial degree vector associated with τh,p+1 is (qe + 1)e∈γh .
Let Php : σh,p+1 → σhp , Php,e : σh,p+1 → σe , php : τh,p+1 → τhp , php,e :
τh,p+1 → τe be the Galerkin projections with respect to the bilinear forms W ·, ·
and V ·, · . For all u ∈ σh,p+1 we define Php and Php,e by
W Php u, v = W u, v ∀v ∈ σhp (11.25)

W Php,e u, v = W u, v ∀v ∈ σe , e ∈ ωh (11.26)
and for all ϕ ∈ τh,p+1 we define php and php,e by
Vphp ϕ, φ = V ϕ, φ ∀φ ∈ τhp (11.27)

Vphp,e ϕ, φ = V ϕ, φ ∀φ ∈ τe , e ∈ γh . (11.28)
Finally, we define the two-level additive Schwarz operators

Pσ := Php + Php,e and pτ := php + php,e . (11.29)
e∈ωh e∈γh
For all v ∈ σh,p+1 , ψ ∈ τh,p+1 there holds with positive constants c1 , c2 , C1 , C2

independently of h, p and q

C1 (1 + log pmax )−2 v2W ≤ Php v2W + Php,e v2W ≤ C2 v2W (11.30)
e∈ωh

c1 (1 + log qmax )−2 ψ2V ≤ php ψ2V + php,e ψ2V ≤ c2 ψ2V (11.31)
e∈γh
with pmax := max{pe , e ∈ ωh }, qmax := max{qe , e ∈ γh } (see [239]).

For all (u, ϕ) ∈ H 1/2 (Γ )/R × H −1/2(Γ ) we define the norm (u, ϕ)H =
(u2W +ϕ2V )1/2, which is equivalent to (u2H 1/2 (Γ )/R+ϕ2H −1/2 (Γ ))1/2 . The norm
(u,ϕ)H is generated by the bilinear form
a(u, ϕ; v, ψ) := W u, v + V ϕ, ψ .
Let (u, ϕ) be the solution of the variational inequality (11.10) and let
(uhp , ϕhp ) ∈ σhp × τhp , (uh,p+1 , ϕh,p+1 ) ∈ σh,p+1 × τh,p+1 be the solutions
of the corresponding discrete problems.
As for finite element problems (see, e.g. [32]) we make the saturation assump-
tion: There exists a parameter 0 ≤ κ < 1 such that for all discrete spaces holds:
(u − uh,p+1 , ϕ − ϕh,p+1 )H ≤ κ(u − uhp , ϕ − ϕhp )H . (11.32)
Now the saturation assumption implies with the triangle inequality:
(1 − κ)(u − uhp , ϕ − ϕhp )H ≤ (uh,p+1 − uhp , ϕh,p+1 − ϕhp H

≤ (1 + κ)(u − uhp , ϕ − ϕhp )H .
Theorem 11.4 We assume that (11.32) holds for the solution (u, ϕ) of the vari-
ational inequality (11.10) and the solutions (uhp , ϕhp ) , (uh,p+1 , ϕh,p+1 ) of the
corresponding discrete problems. Then there are constants ζ1 , ζ2 > 0 such that
ζ1 ηhp ≤ (u − uhp , ϕ − ϕhp )H ≤ ζ2 (1 + log max{pmax , qmax })ηhp (11.33)
where
1/2
ηhp := Θhp
2
+ ηu,hp
2
+ ηϕ,hp
2
, Θhp := Php eh,p+1 W , (11.34)
1/2
ηu,hp := 2
Θhp,e , Θhp,e := Pe eh,p+1 W , (11.35)
e∈ωh
1/2 |B(uhp , ϕhp ; 0, ψe,qe +1 )|
2
ηϕ,hp := θhp,e , θhp,e := ,
e∈γh
ψe,qe +1 V
(11.36)
and eh,p+1 ∈ Kuhp is the solution of the variational inequality
W eh,p+1 , v − eh,p+1 ≥ L (v − eh,p+1 , 0) − B(uhp , ϕhp ; v − eh,p+1 , 0) ∀v ∈ Kuhp
(11.37)
with Kuhp := {v − uhp | v ∈ Kh,p+1

Γ } = Kh,p+1
Γ − uhp .
Proof For the defect (ẽh,p+1 , ε̃h,p+1 ) := (uh,p+1 − uhp , ϕh,p+1 − ϕhp ) of the
solution to
B(uh,p+1 , ϕh,p+1 ; v − uh,p+1 , ψ) ≥ L (v − uh,p+1 , ψ) Γ

∀(v, ψ) ∈ Kh,p+1 × τh,p+1
(11.38)
we can write
B(ẽh,p+1 , ε̃h,p+1 ; v − ẽh,p+1 , ψ) ≥ L (v − ẽh,p+1 , ψ) − B(uhp , ϕhp ; v − ẽh,p+1 , ψ)

(11.39)
for all (v, ψ) ∈ Kuhp × τh,p+1 .

We define (eh,p+1 , εh,p+1 ) ∈ Kuhp × τh,p+1 by
a(eh,p+1 , εh,p+1 ; v − eh,p+1 , ψ) ≥ L (v − eh,p+1 , ψ) − B(uhp , ϕhp ; v − eh,p+1 , ψ)
(11.40)
for all (v, ψ) ∈ Kuhp × τh,p+1 .

This variational inequality can be separated into an inequality and an equality
which are independent, i.e. (11.40) is equivalent to
W eh,p+1 , v − eh,p+1 ≥ L (v − eh,p+1 , 0) − B(uhp , ϕhp ; v − eh,p+1 , 0) ∀v ∈ Kuhp ,
(11.41)
V εh,p+1 , ψ = L (0, ψ) − B(uhp , ϕhp ; 0, ψ) ∀ψ ∈ τh,p+1 .
(11.42)
Now, application of (11.30) gives:
C1 (1 + log pmax )−2 eh,p+1 2W + c1 (1 + log qmax )−2 εh,p+1 2V

≤ (Php eh,p+1 , php εh,p+1 )2H + Pe eh,p+1 2W + pe εh,p+1 2V
e∈ωh e∈γh
≤ C2 eh,p+1 2W + c2 εh,p+1 2V
By definition of php and the Galerkin orthogonality we have
php εh,p+1 2V = V εh,p+1 , php εh,p+1 = 0 (11.43)
yielding
(Php eh,p+1 , php εh,p+1 )2H = Php eh,p+1 2W . (11.44)

11.3 hp-BEM for a VI of Second Kind 403
Furthermore
V εh,p+1 , φe,qe +1 V εh,p+1 , φe,qe +1
pe εh,p+1 = φe,qe +1 = φe,qe +1
V φe,qe +1 , φe,qe +1 φe,qe +1 2V
L (0, φe,qe +1 ) − B(uhp , ϕhp ; 0, φe,qe +1 )
= φe,qe +1
φe,qe +1 2V
−B(uhp , ϕhp ; 0, φe,qe +1 )
= φe,qe +1
φe,qe +1 2V
and hence
|B(uhp , ϕhp ; 0, φe,qe +1 )|
pe εh,p+1 V = =: θhp,e . (11.45)
φe,qe +1 V
Collecting together we get

c3 (eh,p+1 , εh,p+1 )2H ≤ Php eh,p+1 2W + Pe eh,p+1 2W + 2
θhp,e 0 ηhp
e∈ωh e∈γh
≤ c4 (eh,p+1 , εh,p+1 )2H
with c3 := min{C1 (1 + log pmax )−2 , c1 (1 + log qmax )−2 } , c4 = max{C1 , C2 }.

But (see [Lemma 16 ,[297])
(eh,p+1 , εh,p+1 )H 0 (ẽh,p+1 , ε̃h,p+1 )H = (uh,p+1 − uhp , ϕh,p+1 − ϕhp )H
(11.46)
and therefore
1 1
√ ηhp ≤ (uh,p+1 − uhp , ϕh,p+1 − ϕhp )H ≤ √ ηhp .
c4 c3
Application of the saturation assumption yields the assertion of the theorem with
ζ1 = (1+κ)√max{C
1
,c }
and ζ2 = (1−κ)√min{C
1
,c }
.

2 2 1 1
11.3 hp-BEM for a Variational Inequality of the Second

Kind Modelling Unilateral Contact and Friction
This section continues Sect. 5.2 and is based on [217, 218, 297]. Here we give a
further contribution to the analysis of the hp-version of the BEM for nonsmooth
boundary value problems.
We recall the primal boundary variational inequality (5.22), which is equivalent

to (5.20) and (5.21): Find u ∈ K Γ such that for all v ∈ K Γ
Su, v − u + j (v) − j (u) ≥ l(v − u)
with the convex closed set

& '
K Γ := v ∈ H 1/2(Γ ) : v|ΓD = χ|ΓD a.e. and v|ΓC ≤ χ|ΓC a.e. ,
the symmetric Steklov-Poincaré operator S for the interior problem,
1- .
S := W + K + I V −1 (K + I ) : H 1/2(Γ ) → H −1/2(Γ ) ,
2
the continuous, positively homogeneous and sublinear, hence convex friction-type
functional

j (v) := g|v| ds ,
ΓC
and the continuous linear form

l(v) := f v ds ,
ΓN ∪ΓC
where the data f ∈ H −1/2(ΓN ∪ ΓC ), χ ∈ H 1/2(ΓC ∪ ΓD ) ∩ C 0 [Γ C ∪ Γ D ], g ∈

L∞ (ΓC ) with g ≥ 0 are given. Here Ω ⊂ Rd (d ≥ 2) is a bounded Lipschitz
domain with its boundary Γ and mutually disjoint parts ΓD , ΓN , and ΓC such that
Γ = Γ D ∪ Γ N ∪ Γ C and meas (ΓC ) > 0.
In contrast to the preceding section taken from [297] and to a related paper of
Guediri [205] on a boundary variational inequality of the second kind modelling
friction, we take the quadrature error of the friction-type functional into account of
the error analysis. At first without any regularity assumptions, we prove convergence
of the hp-BEM Galerkin approximation in the energy norm. Then under mild
regularity assumptions, we establish an a priori error estimate that is based on
a Céa-Falk lemma for abstract variational inequalities of the second kind. This
lemma permits to split the total discretization error into three different parts: the
approximation error due to the approximation of the Steklov-Poincaré operator
by its discrete counterpart, the distance of the continuous solution to the convex
set of approximations in the trial space, and the consistency error caused by
the nonconforming approximation. Here as in [297] we apply the well-known
approximation theory of spectral methods [47], extend the approximation analysis
of Falk [173], and use interpolation arguments to obtain estimates in Sobolev
norms of fractional order. Moreover, we exploit the special structure of the friction
functional. Thus for our more general variational problem we arrive under mild
regularity assumptions at an a priori error estimate of the same convergence order
as in [297] which is suboptimal because of the appearance of the consistency error in
the nonconforming approximation scheme and because of the well-known regularity
threshold in unilateral problems [267].
11.3.1 The hp-Version Galerkin Boundary Element Scheme
For simplicity let Ω be a polygonal, planar domain and let g be a piecewise

constant function on ΓC . These are no restrictions of generality. In fact, the
approximation of nonpolygonal domains/curves by polygons is well-understood.
The analysis to follow can be extended to higher dimensional domains by tensor
product approximation.
Let SN (N ∈ N) be a sequence of regular partitions of Γ such that all corners of
Γ and all “end points” Γ C ∩ Γ N , Γ N ∩ Γ D , Γ D ∩ Γ C are nodes of SN and that
moreover, g is constant on each edge e of SN contained in ΓC . We introduce the
set of edges on the contact boundary,
Ec,N = {e : e ⊂ ΓC is an edge of SN } .
Further we denote by pN,e ≥ 1 a polynomial degree for each edge e ∈ SN . We

assume that neighboring elements have comparable polynomial degrees, i.e. there
exists a constant c > 0 such that for edges e, e ∈ SN with e ∩ e Ne∅ there holds
c−1 pN,e ≤ pN,e ≤ c pN,e .
As in [297] and in the preceding section we employ Gauss-Lobatto integration

in the discretization procedure. To this end we introduce for p ≥ 1 on the reference
p+1
interval [−1, 1] the p + 1 Gauss-Lobatto points, i.e., the zeros ξj (0 ≤ j ≤ p) of
(1 − ξ 2 )Lp (ξ ), where Lp denotes the Legendre polynomial of degree p ≥ 1. It is
known (see [47], chapter I, section 4) that there exist positive weights
p+1 1
ωj := p+1
p(p + 1)L2N (ξj )
such that the quadrature formula
1
p
p+1 p+1
φ(ξ ) dξ = ωj φ(ξj )
−1 j =0
is exact for all polynomials φ up to degree 2p − 1. By affine transformation Fe :

[−1, 1] → e we define the set of Gauss-Lobatto points Ge,N for each element e of
* p+1 p+1
SN and set GN := {Ge,N : e ∈ SN } . Note that ξ0 = −1 and ξN = 1 are
the end points of the reference interval. Therefore we can introduce the space ΣN
of continuous functions on Γ that are piecewise polynomial up to degree pN,e on
each e ∈ SN as our ansatz space for u,
& '
ΣN := vN ∈ C 0 (Γ ) : vN |e ∈ P pN,e , ∀ e ∈ SN .
A suitable basis of ΣN is given by the Lagrange interpolation polynomials on the

set of Gauss-Lobatto points of each element, since because of the essential boundary
condition we have to evaluate the ansatz functions in the sets Ge,N for each e in the
Dirichlet boundary. To approximate K Γ we choose the Gauss-Lobatto points as
control points of the unilateral constraint and define
KNΓ := {vN ∈ ΣN : vN = γ on GN ∩ Γ D , vN ≤ χ on GN ∩ Γ C } .
Clearly, KNΓ is a convex closed subset of ΣN . However, KNΓ is generally not

contained in K Γ for polynomial degree ≥ 2 or for a non-concave obstacle χ.
We also approximate the nonlinear nonsmooth functional j using the above
quadrature rule by

pN,e
p +1 pN,e +1
jN (v) = ge ωj N,e v ◦ Fe (ξj ) ,
e∈Ec,N j =0
where ge denotes the constant value of the function g on e. Note that for the
piecewise polygonal boundary Γ, v ◦ Fe is piecewise polynomial of the same degree
as v.
Associated to the Gauss-Lobatto points Ge,N we have the local interpolation
operator ie,N := ie,pN,e : C 0 (e) → P pN,e given by
(ie,N η)(x) = η(x), ∀ x ∈ Ge,N , η ∈ C 0 (e)
and the global interpolation operator iN : C 0 (Γ ) → ΣN by

iN η = 1e ie,N η|e, ∀ η ∈ C 0 (Γ ),
e∈SN
where 1e denotes the {0, 1}-valued characteristic function of e ∈ SN .

For later use we recall from [47, Theorem 13.49] the following result on the
polynomial interpolation error in the reference interval Λ = (−1, 1).
Theorem 11.5 For any real numbers r and s satisfying s > (1 + r)/2 and 0 ≤ r ≤
1, there exists a positive constant c depending only on s such that for any function
η ∈ H s (Λ) the following estimate holds:
η − iΛ,p ηH r (Λ) ≤ c pr−s ηH s (Λ) .
To approximate the dual variable ϕ ∈ H −1/2(Γ ), we take as ansatz space

& '
ΦN := ψN ∈ L2 (Γ ) : ψN |e ∈ P pN,e −1 ∀ e ∈ SN .
Here a suitable basis is given by the Legendre polynomials.

Thus we arrive at the following discrete variational problem (πN ) as approxima-
tion to our variational problem (π) given in (5.21): Find (uN , ϕN ) ∈ KNΓ ×ΦN such
that for all (v, ψ) ∈ KNΓ × ΦN ,
1
B(uN , ϕN ; v − uN , ψ) + jN (v) − jN (uN ) ≥ l(v − uN ) . (11.47)
2
Note that we only replaced the nonlinear functional j by its approximate
jN . In most computations, however, also B and l have to be replaced by some
approximations that take into account e.g. numerical integration, matrix compres-
sion, or approximation of a curved boundary. Since such approximations are well
documented in the literature of numerical analysis of linear elliptic boundary value
problems, we omit this aspect here.
Instead we take care of the inconsistency when approximating the Steklov-
Poincarè operator S by its discrete counterpart, here SN : ΣN → ΣN∗ taking
SN = Shp as in the previous section and in [297]:
1 ∗
SN := (ιN W ιN + ι∗N K + I κN (κN∗ V κN )−1 κN∗ (K + I ) ιN ) ,
2
where ιN : ΣN *→ H 1/2(Γ ), κN : ΦN *→ H −1/2(Γ ) denote the canonical embed-

dings and ι∗N , κN∗ their duals. Thus analogously to the primal formulation (5.22) of
our variational problem (π), the discrete problem (πN ) reads: Find uN ∈ KNΓ such
that for all vN ∈ KNΓ ,
SN uN , vN − uN + jN (vN ) − jN (uN ) ≥ (l ◦ ιN )(vN − uN ) . (11.48)
Recall that the primal variable u and the dual variable ϕ, respectively their hp-
approximates uN and ϕN are related by
V ϕ = (I + K)u ; κN∗ V κN ϕN = κN∗ (I + K)ιN uN .
For later use we provide the following a priori estimate.

Lemma 11.1 Let EN = ι∗N SιN − SN . Then there holds
ϕ − κN ϕN 2H −1/2 (Γ ) EN u, u + u − ιN uN 2H 1/2 (Γ ) .
Proof Define ϕ̂N ∈ ΦN as the solution of
κN∗ V κN ϕ̂N = κN∗ (I + K)u .
Then by Galerkin orthogonality and the boundedness of the Galerkin projection,
ϕ − κN ϕN 2H −1/2 (Γ ) 0 V (ϕ − κN ϕN ), ϕ − κN ϕN = ϕ − κN ϕN 2V
≤ 2ϕ − κN ϕ̂N 2V + 2κN ϕ̂N − κN ϕN 2V

= 4EN u, u + 2(I + K)(u − ιN uN ), κN (κN∗ V κN )−1 κN∗ (I + K)(u − ιN uN )
EN u, u + u − ιN uN 2H 1/2 (Γ ) .

Without any regularity assumption for the solution (u, ϕ) of (π) we can show the
following convergence result for the Galerkin BEM solution of (πN ) in the energy
norm.
Theorem 11.6 Let the solution (u, ϕ) of (π) given by (5.21) exist uniquely. Suppose
that for the polygonal domain Ω there are only a finite number of “end points”
Γ C ∩ Γ D , Γ D ∩ Γ N , Γ N ∩ Γ C and the gap function χ|ΓC belongs to H 1/2+ε (ΓC )
for some ε > 0. Then (uN , ϕN ) given by (11.47) are bounded in H 1/2 (Γ )×H − 2 (Γ )
1
independently of N. Moreover for N → ∞ with max hN,e → 0 and with pN,e = p

e∈SN
fixed for all e ∈ SN or with mine∈SN pN,e → ∞ and with hN,e = he fixed for
e ∈ SN there holds (uN , ϕN ) → (u, ϕ) in H 1/2(Γ ) × H −1/2(Γ ).
Proof Since
p
N,e
p +1 p +1
0 ≤ jN (wN ) ≤ gL∞ (ΓC ) ωj N,e |wN ◦ FE (ξj N,e )|
e∈Ec,N j =0
≤ 3gL∞ (ΓC ) meas (ΓC )1/2 wN L2 (ΓC )

wN H 1/2 (Γ ) ∀wN ∈ SN ,
by the L2 stability of Gauss Lobatto quadrature, see [48, Lemma 2.2], the indirect
argument for the a priori bound of approximate solutions of a variational inequality
of the first kind, which is given in the proof of Theorem C.7 in Sect. C.3.2 of the
Appendix, extends to the boundary variational inequality (11.47) of the second kind.
Thus the assumed uniqueness of the solution (u, ϕ) of (π) implies the boundedness
of (uN , ϕN ) in H 1/2 (Γ ) × H − 2 (Γ ).
1
To prove the claimed convergence assertions we adapt the discretization theory

of Glowinski [199] to more general semicoercive variational inequalities of the
second kind over a convex subset instead over the whole space, see Theorem C.7
in Sect. C.3.2 in the Appendix for variational inequalities of the first kind and also
[221] for variational inequalities of the second kind. Thus we have to show the
following hypotheses:
H1 If vN ! v in H 1/2 (Γ ) for N → ∞ with vN ∈ KNΓ , then v ∈ K and
lim inf jN (vN ) ≥ j (v).

N→∞
H2 There exists a subset M ⊂ K Γ dense in K Γ and mappings N : M → ΣN

such that, for each w ∈ M, N (w) → w for N → ∞,

lim jN N (w) = j (w),
N→∞
and N (w) ∈ KNΓ for all N ≥ N0 (w) for some N0 (w) ∈ N.

Note that the analogous hypotheses for the approximation of ψ by ψN are
trivially satisfied in view of the inclusion ΦN ⊂ H −1/2 (Γ ).
The convergence of h-BEM, respectively h-FEM for the variational problem
under study has been treated in [221], respectively in [215], where Newton-Cotes
formulas in numerical quadrature are used instead of Gauss-Lobatto quadrature.
Inspecting the proof of the respective convergence theorem shows that the norm
convergence for a fixed polynomial and quadrature order hinges on the positiveness
of the quadrature weights, what is satisfied for all quadrature orders with Gauss-
Lobatto quadrature. For the proof of norm convergence (in the case j = 0) for
h → 0 with pN,e = p fixed for all e ∈ SN we can also refer to [297]. Therefore in
the following we can focus to the case where for all e ∈ SN , hN,e = he is fixed for
all N and mine∈SN pN,e → ∞.
To verify H1 it is enough to show that for any λ ∈ C 0 (Γ ) with λ|ΓC ≥ 0,

(v − χ)λ ds ≤ 0, (11.49)
ΓC
and to show that for any μ ∈ C 0 (Γ ) with |μ| ≤ 1 on ΓC there holds

g vμ ds ≤ lim jN (vN ) , (11.50)
N→∞
ΓC
since by duality with respect to (L1 , L∞ ), see [151, chapter 4.3], and density
& '
j (v) = sup g vμ ds|μ ∈ C 0 (Γ ), |μ| ≤ 1 .
ΓC
Moreover, since the partition SN is here independent of N, we can simply

consider the above integrals on any fixed edge e ∈ SN such that e ⊂ ΓC with
ge = g|e and p := pN,e instead. Thus fix λ, μ ∈ C 0 [e] with λ ≥ 0, |μ| ≤ 1 and
as [297] approximate these functions by a combination of Bernstein polynomials
BN with the local mapping Fe : [−1, 1] → e to define λN := BN λ ◦ Fe , μN :=
BN μ ◦ Fe via
p

p 1 + t k 1 − t p−k 2k
λN (t) = (BN λ ◦ Fe )(t) := ( ) ( ) (λ ◦ Fe ) −1 .
k 2 2 p
k=0
Since the Bernstein operators are monotone, λN ≥ 0 and |μN | ≤ 1. By [142,

Chapter 1, Theorem 2.3],
lim λN − λL∞ (e) = lim μN − μL∞ (e) = 0 . (11.51)

p→∞ p→∞
For the obstacle function χ ∈ H 1/2+ε (ΓC ) we use the interpolate χN := ie,p χ as
1
approximation. By Theorem 11.5 with r = 0, s = + ε
2
lim χN − χL2 (e) = 0. (11.52)
N→∞
Since the embedding H 1/2 (Γ ) *→ L1 (e) is weakly continuous, vN ! v in L1 (e)

and vN L1 (e) is bounded. Therefore from
# $

(vN − χN )λpN,e −1 − (v − χ)λ dt
e
# $

≤ vN − χN L1 (e) λp−1 − λL∞ (e) + (vN − χN ) − (v − χ) λ dt ;
e
0 0

[vN μpN,e −1 − vμ] dt ≤ 0vN L1 (e) μp−1 − μ0L∞ (e) + [vN − v]μ dt ,
e e
∗
(11.51), (11.52) and using λ, μ ∈ L∞ (e) = L1 (e) , we conclude

lim (vN − χN )λpN,e −1 dt = (v − χ)λ dt , (11.53)
N→∞
e e

lim vN μpN,e −1 dt = vμ dt . (11.54)
N→∞
e e
On the other hand, (vN −χN )λp−1 and vN μp−1 are polynomials of degree 2p−1.
Hence the above integrals can be evaluated exactly by the Gauss-Lobatto quadrature
formula to obtain

p
p+1 # $ p+1
(vN − χN )λp−1 dt = ωj (vN − χN )λp−1 ◦ Fe (ξj ) ,
e j =0

p
p+1 p+1
vN μp−1 dt = ωj (vN μp−1 ) ◦ Fe (ξj ).
e j =0
p+1 p+1
Since the weights ωj > 0, λp−1 ≥ 0, (vN −χN )◦Fe (ξj ) ≤ 0 by vN ∈ KNΓ ,
respectively |μp−1 | ≤ 1, ge ≥ 0 we arrive at

(vN − χN )λp−1 dt ≤ 0,
e

p
p+1 p+1
ge vN μp−1 dt ≤ ge ωj |vN ◦ Fe (ξj )| =: je,N (vN ) ,
e j =0

je,N (vN ) = jN (vN ) .
e∈Ec,N
In view of (11.53) (11.54) this proves our claim (11.49),(11.50).

In the last step let us prove H2. By the finiteness assumption we have due to [222,
Lemma 3.3] the density relation
K Γ ∩ C ∞ (Γ ) = K Γ .
Therefore we can take M = K Γ ∩ C ∞ (Γ ) and define N : M → ΣN by N := iN .

Moreover, since w ∈ M satisfies the constraints in K Γ pointwise, N w ∈ KNΓ for all
w ∈ M. By Theorem 11.5, N w → w in H 1/2(Γ ). Finally by jN (w) = j ( N w),
|j (w) − jN ( N w)| ≤ |j (w) − j ( N w)| + |jN (w) − jN ( N w)|
≤ gL∞ (ΓC ) [w − N wL1 (ΓC ) + w − N wL∞ (ΓC ) ] → 0 (N → ∞).

11.3.2 A Céa-Falk Lemma for Variational Inequalities

of the Second Kind
In this subsection we provide from [217] an abstract Céa-Falk approximation lemma

for variational inequalities of the second kind, so that the above boundary variational
inequality (5.22) is included as a special case.
Consider real normal vector spaces (E, .E ), (G, .G ) and their duals E ∗ , G∗
such that E ⊂ G continuously. Let EN be a subspace of E (N ∈ N) with the
embedding ιN ∈ L (EN , E). Let C ⊆ E and CN ⊆ EN be convex sets. Let
f ∗ ∈ E ∗ and for simplicity, fN∗ = ι∗N f ∗ . Let B ∈ L (E, E ∗ ), BN ∈ L (EN , EN ∗)
uniformly bounded and positive definite with respect to , E ; i.e. there exist some
cB , cB > 0 (independent of EN ) such that
cB v2E ≤ Bv, v ≤ cB v2E (∀ v ∈ E),

cB vN E ≤ BN vN , vN ≤ cB vN 2E (∀ vN ∈ EN ).
2
In addition following the analysis in [73], we assume an estimate of ι∗N BιN − BN

in the form
(ι∗N BιN − BN )vN , wN ≤ c0 wN E eN (u) + u − vN E (11.55)
for any u ∈ E; vN , wN ∈ EN , where c0 > 0 is independent of N and eN is some

function that describes the inconsistency and satisfies eN (u) → 0 as N → ∞.
Lemma 11.2 Let the preceding assumptions of this subsection be satisfied; let u ∈
K and uN ∈ KN such that for all v ∈ K, respectively for all vN ∈ KN ,
Bu, v − u + j (v) − j (u) ≥ f ∗ , v − u , (11.56)

BN uN , vN − uN + jN (vN ) − jN (uN ) ≥ fN∗ , vN − uN . (11.57)
Assume Bu − f ∗ ∈ G∗ . Then there exists a constant c > 0 which depends on

cB , cB , c0 but not on N such that
# $2
cu − uN 2E ≤ eN (u)
& '
+ inf Bu − f ∗ G∗ uN − vG + |jN (uN ) − j (v)|
v∈K
&
+ inf u − vN 2E + |j (u) − jN (vN )|
vN ∈KN
'
+Bu − f ∗ G∗ u − vN G .
Proof Estimate similar to [297][p.432/433]
cB ιN (uN − vN )2E ≤ BN (uN − vN ), uN − vN
using fN∗ = ι∗N f ,(11.56) , and (11.57) for any v ∈ K and vN ∈ KN
≤ Bu, v − ιN uN + Bu, ιN vN − u
+ Bu, ιN uN − ιN vN + BN vN , vN − uN
−f ∗ , v − u − f ∗ , ιN (vN − uN )
+j (v) − j (u) + jN (vN ) − jN (uN )
≤ Bu − f ∗ , v − ιN uN + Bu − f ∗ , ιN vN − u
+B(u − ιN vN ), ιN uN − ιN vN
+(ι∗N BιN − BN )vN , uN − vN + J,
where we abbreviate

J := j (u) − jN (vN )| + |jN (uN ) − j (v) .
Hence using
0 0
0 0
(ι∗N BιN − BN )vN , uN − vN ≤ c0 0ιN (uN − vN )0 (eN (u) + u − ιN vN E )
E
with arbitrary ε > 0

0
0
cB 0ιN (uN − vN )2E

1 1
≤ cB ε ιN uN − ιN vN 2E + u − ιN vN 2E
2 ε
c0 1 - .2
+c0 ε ιN uN − ιN vN 2E + eN (u)
2 ε
c0 1
+ u − ιN vN 2E + J
2 ε
+Bu − f ∗ G∗ v − ιN uN G + ιN vN − uG .
Choose a suitable ε > 0 and again by Cauchy’s inequality conclude for some c =
c(cB , cB , c0 ) > 0
- .2
c u − ιN uN 2E ≤ u − ιN vN 2E + eN (u) + J
+Bu − f ∗ G∗ u − ιN vN G + v − ιN uN G .

11.3.3 A Priori Error Estimate for hp-Approximation
In this section we apply the Lemma 11.2 to obtain an a priori error estimate for
the hp-approximate (uN , ϕN ) of the variational problem (π) under the realistic
regularity assumptions of [297], in particular assuming H 3/2(Γ ) regularity of the
solution u.
Theorem 11.7 Let (u, ϕ) ∈ K Γ × H −1/2(Γ ) be the unique solution of the
problem (π) given by (5.21) with meas (ΓD ) > 0 and let (uN , ϕN ) ∈ KNΓ × ΦN
be the solution of (11.47). Assume u ∈ H 3/2 (Γ ), χ|ΓC ∈ H 3/2(ΓC ), χ|ΓD ∈
H 3/2(ΓD ), f ∗ − Su ∈ L2 (Γ ), where f ∗ |ΓN ∪ ΓC = f and f ∗ |ΓD = 0. Then
there exists c = c(u, χ, f ∗ , g) > 0, independent of N such that for N → ∞
0 0 1/4 −1/4
0(u − uN , ϕ − ϕN )0 1/2 ≤ c max hN,e pN,e .
H (Γ )×H −1/2 (Γ ) e∈SN
Proof Let us fix N ∈ N and write h = maxe∈SN hN,e , p = mine∈SN pN,e for
short. First we prove the claimed error estimate for u − uN in H 1/2(Γ ). Since
u ∈ K Γ , respectively uN ∈ KNΓ satisfies
Su, v − u + j (v) − j (u) ≥ l(v − u) ≡ f ∗ (v − u) ds, ∀ v ∈ K Γ ;

Γ
SN uN , vN − uN + jN (vN ) − jN (uN ) ≥ (l ◦ ιN ) (vN − uN ), ∀ vN ∈ KNΓ ,
respectively, we can apply the Lemma 11.2 to the boundary variational inequality
(11.48) in the setting
E := H 1/2(Γ ) ⊂ G := L2 (Γ ), K := K Γ , KN := KNΓ , B := S, BN := SN .
Moreover, the nondifferentiable convex functionals j and jN are given as previously

by

j (v) = g|v| dt = ge |v| dt ,
e∈Sh e
ΓC
e⊂ΓC

p
p+1
) .
p+1
jN (v) = ge ωj v ◦ Fe (ξj
e∈Sh j =0
e⊂ΓC
Thus the Lemma 11.2 splits the error under study into three different error terms:
cu − uN 2H 1/2 (Γ ) ≤ [eN (u)]2 + (11.58)

& '
inf Su − f ∗ L2 (Γ ) uN − vL2 (Γ ) + |jN (uN ) − j (v)|
v∈K Γ
&
+ inf u − vN 2H 1/2 (Γ ) + |j (u) − jN (vN )|
vN ∈KN
Γ
'
+ Su − f ∗ L2 (Γ ) u − vN L2 (Γ ) .
First we verify the inconsistency estimate (11.55) for EN := ι∗N SιN − SN ,
eN (u) := dist H −1/2 (Γ ) (ϕ, ΦN ) inf = V −1 (K + I ) u − ψH −1/2 (Γ ) ,

ψ∈ΦN
and bound the first term in (11.58). By the subsequent Lemma 11.3 we obtain
EN vN , wN ≤ EN vN H −1/2 (Γ ) wN H 1/2 (Γ )

≤ c0 eN (u) + EN u − vN H 1/2 (Γ ) wN H 1/2 (Γ ) ,
where EN is bounded independently of N. Moreover by [47, Theorem 6.1, Remark

6.1], [63, Theorem 4.4.20]
s− 12
h
eN (u) ≤ cI uH s (Γ ) , 1/2 < s ≤ 3/2 . (11.59)
p
To bound next the approximation error inf{. . . |vN ∈ KNΓ }, which is the third term
in (11.58), take vN = u∗N := iN u ∈ KNΓ , the interpolate of u ∈ H 3/2 (Γ ) ⊂ C 0 (Γ ).
By Theorem 11.5, [63, Theorem 4.4.20] there are constants c1 , c2 > 0 independent
of u and N such that

∗ h s
u − uN L2 (Γ ) ≤ c1 uH s (Γ ) , 1/2 < s ≤ 3/2 ,
p

h s−1
u − u∗N H 1 (Γ ) ≤ c2 uH s (Γ ) , 3/4 < s ≤ 3/2 ,
p
hence by real interpolation, see e.g. [373],
h
u − u∗N H 1/2 (Γ ) ≤ c3 uH 3/2 (Γ ) ,
p
Further, by construction,

p
p+1
jN (u∗N ) = (u∗N ◦ Fe ) (ξj )
p+1
ge ωj
e⊂ΓC j =0

p
p+1
) = jN (u) .
p+1
= ge ωj (u ◦ Fe ) (ξj
e⊂ΓC j =0
Hence using the interpolation operators ie,p , iN and the exactness of Gauss-Lobatto
quadrature

j (u) − jN (u∗ ) ≤ ge |u|ds − ie,p |u| ds
N
e⊂ΓC e e

≤ gL∞ (ΓC ) |u| − iN |u| ds
ΓC
0 0
0 0
≤ meas (ΓC )1/2 gL∞ (ΓC ) 0|u| − iN |u| 0 . (11.60)
L2 (ΓC )
Since |u| can only be guaranteed to lie in H 1 (Γ ) (as the max of the two absolutely
continuous functions u, −u, see also [267, Corollary A.6]) we can use the regularity
assumption to derive only
0 0
0|u|0 t ≤ uH t (Γ ) , 1/2 < t ≤ 1 . (11.61)
H (Γ )
what for t < 1 follows easily from the triangle inequality in the Sobolev-
Slobodetskii norm. Therefore we can conclude by Theorem 11.5,[63, Theorem
4.4.20]

h t 0 0
j (u) − jN (u∗ ) ≤ c̃ 0|u|0
N
p H t (Γ )
t
h
≤ c̃ uH t (Γ ) , 1/2 < t ≤ 1 .
p
Thus for the approximation error,
3 4 h
inf . . . |vN ∈ KNΓ ≤ cI I (u, f ∗ , g) uH 3/2 (Γ ) . (11.62)
p
3 4
To bound the consistency error inf . . . |v ∈ K Γ , which is the second term
in (11.58), let χN := iN χ and take (with min defined pointwise a.e.) v ∗ = min(uN −
χN , 0) + χ on ΓC , v ∗ = γ on ΓD , v ∗ = uN on ΓN . Clearly, v ∗ ∈ K Γ . To show that
v ∗ ∈ H 1 (Γ ) we use the arguments of [297]: For any e ⊂ ΓC ∪ ΓD , the polynomial

uN −χN | E, which is of degree pN,e ≤ p, has at most p zeros on e. Hence, the level
set {x ∈ ΓC ∪ ΓD : (uN − χN )(x) < 0} is the finite union of open subintervals and
min(uN − χN , 0) is continuous and piecewise a polynomial on ΓC ∪ ΓD . Therefore
v ∗ ∈ H 1 (Γ ) as claimed.
Moreover, following [297] write v ∗ − uN = min(uN − χN , 0) + χ − uN =
χ − χN + min(0, χN − uN ) − 0 on ΓC ∪ ΓD . Hence
v ∗ − uN L2 (ΓC ∪ΓD ) ≤ χ − χN L2 (ΓC ∪ΓD ) + min(0, χN − uN ) − 0L2 (ΓC ∪ΓD ) .
For the first term there holds
χ − χN L2 (ΓC ∪ΓD ) ≤ C1 h 2 p− 2 χ

1 1
1 .
H 2 (ΓC ∪ΓD )
By uN ∈ KNΓ , iN min(0, χN − uN ) = 0 and hence by Theorem 11.5 for the second

term there holds
min(0, χN − uN ) − 0L2 (ΓC ∪ΓD ) ≤ χN − uN L2 (ΓC ∪ΓD )
min(0, χN − uN ) − 0L2 (ΓC ∪ΓD ) ≤ C2 h1 p−1 min(0, χN − uN )H 1 (ΓC ∪ΓD )
≤ C2 h1 p−1 χN − uN H 1 (ΓC ∪ΓD ) ,
thus by real interpolation,

1 1
min(0, χN − uN ) − 0L2 (ΓC ∪ΓD ) ≤ Ch 2 p− 2 χN − uN 1 .
H 2 (ΓC ∪ΓD )
By Theorem 11.6 for some C > 0, independent of h and p,
uN 1 ≤ C u 1 .
H 2 (ΓC ∪ΓD ) H 2 (ΓC ∪ΓD )
By definition of v ∗ , v ∗ − uN L2 (ΓN ) = 0. Hence we arrive at the estimate

v ∗ − uN L2 (Γ ) ≤ Ch1/2 p−1/2 χH 1/2 (Γ ) + uH 1/2 (Γ ) . (11.63)
Further by construction,
p+1 p+1 p+1
v ∗ Fe (ξj ) = (uN − χN + χ) Fe (ξj ) = uN Fe (ξj ) ,
jN (v ∗ ) = jN (uN ) .
On the other hand, by Theorem 11.5,
v ∗ H 1 (ΓC ) ≤ uN − χN H 1 (ΓC ) + χH 1 (ΓC )
≤ uN H 1 (Γ ) + 2χH 1 (ΓC ) + cp−1/2 χH 3/2 (ΓC ) ,

v ∗ L2 (ΓC ) ≤ uN − χN L2 (ΓC ) + χL2 (ΓC )
≤ uN L2 (Γ ) + 2χL2 (ΓC ) + c p−3/2 χH 3/2 (ΓC ) .
Hence by real interpolation of the nonlocal H t -norm and by the boundedness of

uN in virtue of Theorem 11.6, v ∗ is bounded, too, in H 1/2(ΓC ). Thus with similar
arguments as above, see (11.60) and (11.61), we can conclude
∗

j (v ) − jN (uN ) ≤
ge |v |ds − ie,pN,e |v ∗ | ds
∗
e∈Ec,N e e
0 0
0 0
≤ meas (ΓC )1/2 gL∞ (ΓC ) 0|v ∗ | − iE,pN,E |v ∗ | 0
L2 (ΓC )
≤ c h p−1 v ∗ H 1 (ΓC ) .
To show the analogue estimate with respect to the L2 norm we estimate separately:

∗
j (v ) = ge |v ∗ |ds ≤ meas (ΓC )1/2 gL∞ (ΓC ) v ∗ L2 (ΓC )
e∈Ec,N e
and with Cauchy-Schwarz inequality

pN,e
∗ p +1 pN,e +1
jN (v ) = ge ωj N,E uN ◦ FE (ξj )
e∈Ec,N j =0
≤ 3gL∞ (ΓC ) meas (ΓC )1/2 uN (L2 (ΓC ))2 ,
where we use the L2 stability of Gauss Lobatto quadrature, see [48, Lemma 2.2].
Therefore by real interpolation we arrive at
∗
j (v ) − jN (uN ) ≤ c h1/2 p−1/2 v ∗
H 1/2 (ΓC ) .
Hence for the consistency error,

3 4
inf . . . |v ∈ K Γ ≤ cI I I (u, f ∗ , χ, g) h1/2 p−1/2 . (11.64)
Altogether, (11.58), (11.59), (11.62), and (11.64) yield the claimed estimate of
the error u − uN .
It remains to prove the claimed estimate of the error ϕ − ϕN in H −1/2 (Γ ).

This follows immediately from the proved estimate of the error u − uN in virtue
of Lemma 11.1 and Lemma 11.3 using (11.59).

Remark 11.2 The preceding proof is given in the case of dimension d = 2. It
extends to d > 2 by tensor product approximation on a quadrilateral mesh.
In the proof above we used the following lemma, due to Maischak and Stephan,
see [297, Lemma 15].
Lemma 11.3 Let EN = ι∗N SιN − SN . Then EN is bounded independently of N and
there holds
EN uH −1/2 (Γ ) eN (u) = inf V −1 (K + I ) u − ψH −1/2 (Γ ) ,

ψ∈ΦN
0 ≤ EN (u), u eN (u)2 .
Proof Let y := (I + K)u ∈ H 1/2(Γ ). Since V is positive definite, there exists a

1
solution z ∈ H 2 (Γ ) of V z = y and a solution zN ∈ ΦN of κN∗ V κN zN = κN∗ y, i.e.,
1
zN is the Galerkin approximation of z ∈ H 2 (Γ ) in ΦN and we have
zH −1/2 (Γ ) yH 1/2 (Γ ) and zN H −1/2 (Γ ) yH 1/2 (Γ ) .
Then it results
1
EN (u), v = z − κN zN , (I + K)v
2
yH 1/2 (Γ ) (I + K)vH 1/2 (Γ ) uH 1/2 (Γ ) vH 1/2 (Γ ) ,
i.e., EN is bounded independently of N. By Galerkin orthogonality we have
2EN (u), v = y, V −1 y − y, κN (κN∗ V κN )−1 κN∗ y

= z, V z − κN zN , V κN zN = z2V − κN zN 2V = z − κN zN 2V ≥ 0 ,
i.e., due to the quasi-optimal error estimate we have
1
0 ≤ EN (u), u = z − κN zN 2V eN (u)2 .
2
Using the boundedness of (I + K) and the quasi-optimal error estimate we have
1
EN uH −1/2 (Γ ) ≤ (I + K)z − κN zN H −1/2 (Γ ) eN (u) .
2
Therefore, the assertion of the Lemma follows.

The approach using the Steklov-Poincaré operator and the treatment of the resulting
consistency error in our analysis opens the way to attack contact friction problems
with nonlinear material behaviour by more efficient hp− methods. For such
problems hp−finite element methods can be employed in a small limited subdomain
such that a coupling procedure permits to use the advantages of the boundary
element method, specifically in the reduction of dimension.
Thus we anticipate that many of the details of our numerical analysis will be
applicable to various interesting unilateral contact problems and other free boundary
value problems.
11.4 Mixed hp-BEM for Frictional Contact Problems
11.4.1 Boundary Integral Formulation for Contact Problem
Now we report from [38] and consider the unilateral contact problem with Tresca
friction. For a given gap function g, friction threshold F ≥ 0, elasticity tensor C
and Neumann data f , find u such that
− div σ (u) = 0 in Ω ⊂ Rd , d = 2 or 3 (11.65a)

σ (u) = C : (u) in Ω (11.65b)
u=0 on ΓD (11.65c)
σ (u) · n = f on ΓN (11.65d)
σn ≤ 0, un ≤ g, σn (un − g) = 0 on ΓC (11.65e)
|σt | ≤ F , σt ut + F |ut | = 0 on ΓC (11.65f)
Here, n, t denote the outer unit normal, tangential, respectively, and σn =

(σ (u)n)n ∈ R, σt = (σ (u)n)t ∈ Rd−1 , un = u · n ∈ R, ut = u · t ∈ Rd−1
the normal, tangential component of σ (u)n or of u. The friction constraint (11.65f)
is equivalent to
|σt | ≤ F , |σt | < F ⇒ ut = 0, |σt | = F ⇒ ∃ ν ≥ 0 : ut = −νσt
where we use | · | to abbreviate the Euclidean norm. We denote ΓΣ := Γ¯N ∪ Γ¯C

where ∂Ω = Γ¯N ∪ Γ¯C ∪ Γ¯D and Γ¯C ∩ Γ¯D = ∅ for simplicity of the notation. The
space dimension d may be two or three. Since the main source of difficulties, i.e. the
Signorini (11.65e) and friction (11.65f) condition, lie on the boundary it maybe
favorable to choose a boundary integral formulation with the symmetric Steklov-
Poincaré operator S : H 2 (Γ ) → H − 2 (Γ ).
1 1
1
S := {W + (K + I )V −1 (K + I )}
2
11.4 Mixed hp-BEM for Frictional Contact Problems 421
Here, for x ∈ Γ

.
V μ(x) = 2 G(x, y)μ(y)dsy , Kv(x) = 2 Ty G(x, y) v(y)dsy
Γ Γ
(11.66)

.
K . μ(x) = 2Tx G(x, y)μ(y)dsy , W v(x) = −2Tx Ty G(x, y) v(y)dsy
Γ Γ
(11.67)
with Tn (u) := σ (u)|Γ ·n denote the single layer potential, the double layer potential,
its adjoint operator and the hypersingular operator with the fundamental solution of
the Lamé equation
⎧
⎨ λ+3μ λ+μ (x−y)(x−y).
4πμ(λ+2μ)
1
ln |x−y| I+ λ+3μ |x−y|2
, if d=2
G(x, y) = λ+μ (x−y)(x−y).
⎩ λ+3μ |x − y|−1 I + λ+3μ |x−y|3 , if d=3
8πμ(λ+2μ)
In [37] we consider the following variational inequality formulation on the boundary

only, which is a reformulation of (11.65): Find u ∈ K s.t.
Su, v − u ΓΣ + j (v) − j (u) ≥ f, v − u ΓN ∀v ∈K (11.68)
with friction functional j (v) := ΓC F |vt | ds and convex cone of admissible

& 1
'
functions K := v ∈ H̃ 2 (ΓΣ ) : vn ≤ g a.e. on ΓC . Equivalently, there holds a
mixed formulation on the boundary only, in which the Lagrange multiplier λ =
1
−σ n|ΓC is sought as an additional unknown: Find (u, λ) ∈ H̃ 2 (ΓΣ ) × M + (F ) s.t.
1
Su, v ΓΣ + λ, v ΓC = f, v ΓN ∀ v ∈ H̃ 2 (ΓΣ ) (11.69a)
u, μ − λ ΓC ≤ g, μn − λn ΓC ∀ μ ∈ M + (F ) (11.69b)
where
& 1 1
'
M + (F ) := μ ∈ H̃ − 2 (ΓC ) : μ, v ΓC ≤ F , |vt | ΓC ∀v ∈ H 2 (ΓC ), vn ≤ 0
is the set of admissible Lagrange multipliers. The connection between these three
formulations is as follows:
Theorem 11.8 ([37]) The problems (11.68) and (11.69) are equivalent and are also
equivalent to (11.65) in a distributional sense. Furthermore, there exists exactly one
solution to (11.68) and (11.69). This means:
(i) Any solution of (11.69) is also a solution of (11.68).
(ii) For the solution u ∈ K of (11.68) there exists a λ ∈ M + (F ) such that (u, λ)
is a solution of (11.69).
(iii) There exists a unique solution to (11.69).
11.4.2 hp-Boundary Element Procedure with Lagrange

Multiplier and Fast Solver
To set up the discrete problem we firstly introduce biorthogonal hp-BE-spaces,

namely: The approximation of the primal variable u is sought in
& 1
- .d
Vhp = vhp ∈ H̃ 2 (ΓΣ ) ∩ C 0 (ΓΣ ) : vhp |E ◦ ΨE ∈ PpE ([−1, 1]d−1 ) ∀ E ∈ Th ∩ ΓΣ ,
'
vhp = 0 in the endpoints of ΓΣ
for a given boundary mesh Th of Γ , consisting of line segments in 2D and

quadrilaterals in 3D, in which the nodes, edges coincide with the boundary of the
boundary parts ΓD , ΓN and ΓC , and with polynomial degree distribution p : Th →
N. Here, ΨE : [−1, 1]d−1 → E ∈ Th is the affine mapping onto the physical
element. The dual, Lagrange multiplier variable λhp is sought in
⎧ ⎫
⎨
NC
9 : ⎬
+
Mhp (F ) = μhp = μi ψi : μhp , vhp ΓC ≤ F , |(vhp )t |h ∀vhp ∈ Vhp , (vi )n ≤ 0
⎩ ⎭
i=1
N
Here, we take |(vhp )t |h := i |(vi )t |φi where vhp = i=1 vi φi , vi ∈ R and
d
where φi denotes a scalar Gauss-Lobatto-Lagrange basis function (for uhp ) and ψj

a biorthogonal basis function (for λhp ), see Fig. 11.2, i.e.

φi ψj ds = δij φi ds 1 ≤ i, j ≤ NC .
ΓC ΓC
In particular, we assume that the first NC basis functions are associated with
a Gauss-Lobatto point which lies on ΓC . Furthermore, (vi )n := v. i n, (μi )n :=
.
μi n and (vi )t := vi − (vi )n n, (μi )t := μi − (μi )n n are the normal, tangential
components of the expansion coefficients v, μ. Note that we use the same mesh and
polynomial degree distribution for uhp |ΓC and for λhp . Due to the biorthogonality,
the discrete inf-sup condition is still satisfied [36].
(a)
1.2
GLL 1
GLL 2
GLL 3
1 GL Points
0.8
0.6
0.4
0.2
−0.2
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(b)
1.5
Biorth 1
Biorth 2
Biorth 3
GL Points
0.5
−0.5
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Fig. 11.2 Basisfunctions for p = 2 [37]. (a) Gauss-Lobatto-Lagrange. (b) Biorthogonal
In [37] we analyse the following mixed hp-boundary element formulation: Find

+
(uhp , λhp ) ∈ Vhp × Mhp (F ) such that
9 :
Shp uhp , vhp ΓΣ + λhp , vhp Γ = f, vhp ΓN ∀vhp ∈ Vhp
C
(11.70a)
9 : 9 : +
uhp , μhp − λhp Γ ≤ g, (μhp )n − (λhp )n Γ ∀ μhp ∈ Mhp (F )
C C
(11.70b)
−1 ∗
where Shp := W + (K + 1/2). ihp Vhp ihp (K + 1/2) approximates S in the standard
D → H −1/2 (Γ ) and its
manner [91]. That is with the canonical embedding ihp : Vhp
∗
dual ihp where
- .d "
1
D
Vhp = φhp ∈ H − 2 (Γ ) : φhp |E ◦ ΨE ∈ PpE −1 ([−1, 1]d−1) ∀ E ∈ Th .
In particular, Vhp is the Galerkin realization of the single layer potential over Vhp
D.
There holds:
Theorem 11.9 ([37]) There exists exactly one solution of the discrete mixed prob-
lem (11.70).
+
Proof uhp ∈ Vhp , λhp ∈ Mhp (F ) are uniquely defined by their expansion
coefficients u, λ. Note that the discrete weak contact conditions are equivalent to

1
(ui )n ≤ gi := gψi ds (11.71a)
Di ΓC
(λi )n ≥ 0 (11.71b)
(λi )n ((ui )n − gi ) = 0 (11.71c)

1
|(λi )t | ≤ Fi := F φi ds (11.71d)
Di ΓC
|(λi )t | < Fi ⇒ (ui )t = 0 (11.71e)

|(λi )t | = Fi ⇒ ∃ α ∈ R : α 2 (λi )t = (ui )t (11.71f)
for all 1 ≤ i ≤ NC and Di := ΓC φi ds > 0. Now the system of linear equations

from (11.70a) has the block structure

SC C SC R uC DC fC
+ λ= .
S RC S RR uR 0 fR
Let P≥0 , PB(Fi ) be closest point projections onto R≥0 , B(Fi ), a ball with center
zero and radius Fi , respectively. Let
NC
T (λ) := P≥0 ((λi )n + r((ui )n − gi )) , PB(Fi ) ((λi )t + r(ui )t ) i=1 .
Then
T (λ1 ) − T (λ2 )22 ≤ λ1 − λ2 22 1 − 2αrγ 2 + r 2 γ 2 < λ1 − λ2 22
for r = α where α = λmin D −1 C S C C − S C R S −1

RR S RC > 0 is the minimal
eigenvalue, γ := δuC 2 / δλ2 . Thus T is a strict contraction and Banach’s fixed
point theorem yields the assertion for λhp . Since Shp is H̃ 1/2 (ΓΣ )-coercive, uhp
exists and is unique as well.

An important benefit of this approach with biorthogonal basis functions is the
componentwise decoupling of the weak contact conditions [37]. For the non-linear
complementarity functions (NCF)

ϕη : Rd·N × Rd·NC → RNC with ϕη (u, λ) i = η (λi )n + (gi − (ui )n )
/
− (λi )2n + (gi − (ui )n )2 + (1 − η) (λi )+
n (gi − (ui )n )
+
for the Signorini condition and for ct > 0
CT : Rd·N×Rd·NC → R(d−1)·NC with (CT (u, λ))i = max{Fi , |(λi )t +ct (ui )t |}(λi )t
− Fi · ((λi )t + ct (ui )t ) , 1 ≤ i ≤ NC
for the Tresca condition, there holds: ϕη (u, λ) = 0 if and only if (11.71a)–(11.71c)
hold, and CT (u, λ) = 0 if and only if (11.71d)–(11.71f) hold. Therewith, the
discrete mixed problem (11.70) is equivalent to
⎛ ⎞
Su + Dλ − f
0 = F (u, λ) = ⎝ ϕη (u, λ) ⎠ (11.72)
CT (u, λ)
where Su+Dλ−f = 0 is matrix representation of the variational equality (11.70a).

As shown in [37] a fast solver for (11.72) is the semi-smooth Newton (SSN)
method which converges locally super-linearly and even locally quadratically in the
frictionless case, i.e, when F ≡ 0.
11.4.3 Error Controlled hp-Adaptive Schemes
In [37] we derive the following a posteriori error estimate with residual type error
indicators.
Theorem 11.10 Let (u, λ), (uhp , λhp ) be the solution of (11.69), (11.70) respec-
tively. Then there exists a constant C > 0, independent of h and p s.t.
0 0 0 02
0λhp − λ02 + 0u − uhp 0 ≤C 2
ηres (E) + contact terms
− 21 1
H̃ (ΓC ) H̃ 2 (ΓΣ )
E∈Th
−1 ∗
with ψhp := ihp Vhp ihp (K + 12 )uhp
hE 0 0
0f − Shp uhp 02 2 hE 0 0
0λhp + Shp uhp 02 2
2
ηres (E) := L (E∩Γ )
+ L (E∩ΓC )
pE N pE
0 02
0∂ 0
0
+hE 0 (V ψhp − (K + I )uhp )0 0 2
∂s L (E)
where the contact terms (resulting from the violation of the contact condition by the
discrete solution of (11.70)) are
0 0 9 : 0 0
0(λhp )− 02 + (λhp )+ + 0 − 02
n −1 n , (g − (uhp )n ) ΓC + (g − (uhp )n ) 1
H̃ 2 (ΓC ) H 2 (ΓC )

0 02
+ 0((λhp )t − F )+ 0 − 1 − ((λhp )t − F )− (uhp )t
H̃ 2 (ΓC )
ΓC

+ (λhp )t (uhp )t − (λhp )t (uhp )t ds with v + = max {v, 0} , v − = min {v, 0} .
These contact terms can be interpreted as: violation of the consistency, comple-
mentarity and non-penetration condition with respect to the normal component of
the solution, and violation of the consistency, violation of the stick condition and
of having the same the sign in slip condition wrt. the tangential component of the
solution.
Proof The starting point is to consider the auxiliary problem (Braess’ trick [59]):
Let z ∈ H̃ 1/2 (ΓΣ ) such that
9 :
Sz, v ΓΣ = f, v ΓN − λhp , v Γ ∀v ∈ H̃ 1/2(ΓΣ ) (11.73)
C
for which uhp is the Galerkin approximation. The ellipticity and continuity of S
gives
0 02 0 0 0 0
α 0uhp − u0H̃ 1/2 (Γ ) ≤ C 0uhp − z0H̃ 1/2 (Γ ) 0uhp − u0H̃ 1/2 (Γ )
Σ Σ Σ
9 : 9 :
+ λn − λhp n , uhp n − un Γ + λt − λhp t , uhp t − ut Γ .
C C
Furthermore
9 :
λn − λhp n , uhp n − un Γ
C
7 + 8 0 0 0 − 0
+ 0 0
≤ λhp n , g − uhp n + 0 λhp n − λn 0H̃ −1/2 (Γ ) 0 g − uhp n 0 1/2
ΓC C H (ΓC )
0 0 0 0
0 −0 0 uhp − un 0 1/2
+ 0 λhp n 0 −1/2 n H (Γ )
H̃ (ΓC ) C
whereas
9 :
λt − λhp t , uhp t − ut Γ
C
0 0 00
0 0 0 +0 00 0 0 0
≤0 0 0
0 λhp t 2 − F 0
0 0 ut − uhp t 02 0
H̃ −1/2 (ΓC ) H 1/2 (ΓC )
5 6
0 0 − 0 0
− 0 λhp 0 − F , 0 uhp t 02
t 2
ΓC
:
9 70 0 0 0 8
− λhp t , uhp t Γ + 0 λhp t 02 , 0 uhp t 02 .
C ΓC
From the continuous inf-sup condition it follows
0 0 2C 2 0 0 2C 2 0 02
0λhp − λ02 −1/2 ≤ 0u − uhp 02 1/2 + 2 0uhp − z0H̃ 1/2 (Γ )
H̃ (Γ C) β 2 H̃ (ΓΣ ) β Σ
yielding
0 0 0 02
0λhp − λ02 + 0uhp − u0H̃ 1/2 (Γ )
H̃ −1/2 (ΓC ) Σ
( 0 0 0
0 02 0 − 0 0 0 0 + 02
0 0 0 02 0
≤ C uhp − z H̃ 1/2 (Γ ) + 0 λhp n 0 −1/2 0
+ 0 0 λhp t 0 − F 00
Σ H̃ (ΓC ) 2 H̃ −1/2 (ΓC )
5 0 6
0 0 0 − 00 0 0
7 8
− 0 λhp t 0 − F , 0 uhp t 0 − λhp t , uhp t
2 2 Γ Γ C
C
5 6
70
0 0 0 0
0 0 0
8 + +
+ 0 λhp t 0 , 0 uhp t 0 + λhp n , g − uhp n
2 2 ΓC ΓC
0 0 )
0 − 02
+0
0 g − uhp n
0
0 .
H 1/2 (ΓC )
0 02
It remains to estimate 0uhp − z0H̃ 1/2 (Γ ) ≤ C E∈Th
2 (E) with the local
ηres
Σ
contributions
0 02
2 hE 0 0
0t˜ − Shp uhp 02 2
0∂
0
0
0
ηres (E) := L (E∩Γ )
+ hE 0 (V ψhp − (K + I )u hp ) 0 2
pE Σ ∂s L (E)
−1 ∗
where t˜|ΓN = f and t˜|ΓC = −λhp . ψhp := ihp Vhp ihp (K + I )uhp is a natural side
product when estimating the error induced by the approximation of V −1 .

In [37] we replace F with F λn , F (λhp )n in the continuous and in the discrete
cases, respectively, as modifications for Coulomb friction. There we show that the
decoupling of contact constraints still holds with Fi := D1i ΓC F (λhp )n φi ds and
Fi = F (λi )n for constant F . Furthermore, if F L∞ (ΓC ) is sufficiently small, the
frictional part of the contact terms in Theorem 11.10 changes to
0 + 0 7 − 8
0 0
0 λhp t − F (λhp )n 0 − λhp t − F (λhp )n , uhp t
H̃ −1/2 (ΓC ) ΓC
9 : 9 :
− λhp t , uhp t Γ + λhp t , uhp t Γ
C C
Alternatively, we can take a bubble error estimate instead of a residual estimate for
the variational equality part, but then the saturation assumption
0 02 0 02 0 02 0 02
κ ∈ (0, 1) : 0uhp+1 − z0W + 0ψ − ψhp+1 0V ≤ κ 2 0uhp − z0W + 0ψ − ψhp 0V
must hold. For details see [37] where the following algorithm is performed.
Algorithm 11.1 (Solve-mark-refine algorithm for hp-adaptivity)
(i) Choose initial discretization Th and p, steering parameters θ ∈ (0, 1) and

δ ∈ (0, 1).
(ii) For k = 0, 1, 2, . . . do
a. solve discrete mixed problem (11.70).
b. compute local indicators Ξ 2 to current &solution.
&

c. mark all elements E ∈ N := argmin Nˆ ⊂ Th : E∈Nˆ Ξ 2 (E) ≥
''

θ E∈Th Ξ 2 (E) for refinement
d. estimate local analyticity [254], i.e. compute Legendre coefficients of

p
2i + 1 1
uhp |E (ΨE (x)) = ai Li (x), ai = uhp |E (ΨE (x))Li (x) dx
2 −1
j =0
Use a least square approach to compute the slop m of | ln |ai || = mi + b. If

e−m ≤ δ then p-refine, else h-refine marked element E. (This can be done
analogously in higher dimensions [36]).
e. refine marked elements based on the decision in 2(d).
& '
Example 11.1 We take a disc with diameter one, i.e. Γ = x ∈ R2 : |x| = 12 . The
3 4
boundary is split into ΓN = Γ ∩ x ∈ R2 : x2 ≥ 0 and ΓC = Γ \ ΓN . The rigid
(a) (b)
0.5
−0.5
−1
0.5
−1.5
−0.5 0
0
0.5 −0.5
Fig. 11.3 Solution of the Hertz problem with Coulomb-friction, uniform mesh 2048 elements,
p = 2, [37]. (a) Reference (gray), deformed (blue). (b) λy and ty (blue), λx (red)
uniform h, p=1
uniform h, p=2
h−adaptive, p=1
hp−adaptive
−1
10
Error Estimation
−2
10
−3
10
1 2 3 4
10 10 10 10
Degrees of Freedom
Fig. 11.4 Bubble error estimates for different families of discrete solutions (Hertz) [37]
body motions are set to zero by the iterative solver to obtain a unique solution.
The Young’s modul is E = 5 and the Poisson’s ratio is ν = 0.45. The Coulomb
friction coefficient is F ≡ 1, the Neumann force is f = (0, −0.2). and the gap is
dist(ΓC , −0.5). Figure 11.3a visualizes the reference configuration in grey and the
deformed state in blue. The corresponding Neumann (λ on ΓC ) data are depicted
in Fig. 11.3b. The reduction of the bubble error indicator for different families of
discrete solutions is displayed in Fig. 11.4. In particular, the uniform h-versions
with p = 1, p = 2 have a convergence rate of about 0.5 underlying the limited

regularity of the continuous solution. Not only is the hp-adaptive scheme superior
to the other methods but it also has with 2.4 a significantly higher convergence rate.
11.4.4 Stabilized hp-Mixed Method—A Priori Error Estimate
In [33] mixed hp-boundary element methods are analyzed for frictional contact
problems for the Lame equation. The stabilization technique circumvents the inf-
sup conditions for the mixed problem and thus allows to use the same mesh and
polynomial degree for primal and dual variables. A priori estimates are given
for Tresca friction using Gauss-Legendre-Lagrange polynomials as test and trial
functions for the Lagrange multiplier. In [33] we review about Coulomb friction and
present numerical experiments which underline the insensitivity of the method to the
scaling of the stabilization term. This approach is motivated by the seminal work of
Barbosa and Hughes [39]. Assuming the mesh and polynomial degree distribution
to be locally quasi-uniform we consider with the affine mapping ΨE form [−1, 1]
onto E ∈ Th the ansatz spaces
& 1 # $2 '
Vhp = v hp ∈ H̃ 2 (ΓΣ ) : v hp |E ◦ ΨE ∈ PpE ([−1, 1]) ∀ E ∈ Th ⊂ C 0 (ΓΣ )
& '
D = φ hp ∈ H − 12 (Γ ) : φ hp | ◦ Ψ ∈ #P $2
∀ E ∈ Th ,
Vhp E E pE −1 ([−1, 1])
& # $2
+
M̃k,q (F ) := μkq ∈ L2 (ΓC ) : μkq |E ◦ ΨE ∈ PqE ([−1, 1]) , μkq (x) ≥ 0,
'
− F (x) ≤ μkq |t (x) ≤ F (x) for x ∈ Gkq
where Gkq is the set of affinely transformed Gauss-Legendre points and μkq are
linear combinations of Gauss-Legendre-Lagrange basis functions. The stabilized
+
mixed method reads: Find (uhp , λkq ) ∈ Vhp × M̃k,q (F ) such that ∀v hp ∈ Vhp and
+
∀μkq ∈ M̃k,q (F )
Shp uhp , v hp ΓΣ + λkq , v hp ΓC − γ (λkq + Shp uhp ), Shp v hp ΓC = f, v hp ΓN

kq kq
μkq − λkq , uhp ΓC − γ (μkq − λkq ), λkq + Shp uhp ΓC ≤ g, μn − λn ΓC
(11.74)
1+β −2−η
Here γ is a piecewise constant function on ΓC such that γ |E = γ0 hE pE with
constants γ0 > 0, β, η ≥ 0 for all elements E ∈ Th |ΓC .
In [33] it is shown that the solution (u, λ) ∈ H̃ 1/2(ΓΣ ) × M + (F ) of (11.69) is
+
approximated by the solution (uhp , λkq ) ∈ Vhp × M̃k,q (F ) of (11.74) with h = k
sufficiently small, p = q, 0 ≤ α < 1/2 satisfying
u − uhp 2H̃ 1/2 (Γ + γ 1/2 (λ − λkq )2L2 (Γ ≤ chα/2 p−α/3

Σ) C)
(see in [33] Theorem 16 and Remark 17). For improved stabilization see [35].
11.4.5 A Priori Error Estimates for hp-Penalty-BEM for

Contact Problems in Elasticity
From [102] we report an a priori error analysis for hp−version of the penalty
Galerkin BEM for frictionless contact problems.
Let us consider an elastic body in two dimensions under small strain assumption.
We associate the body with a bounded two-dimensional polygonal domain Ω with
boundary Γ = ∂Ω. Assume that some part of the boundary ΓD ⊂ Γ is fixed, which
prevents the rigid body motions and therefore provides uniqueness of the solution.
Further, we denote by ΓN the boundary part with prescribed boundary tractions.
Finally, we call ΓC ⊂ Γ the boundary part which potentially can come into contact
with a rigid obstacle. The three parts of the boundary are assumed to be disjoint and
satisfy Γ = Γ D ∪ Γ N ∪ Γ C .
As we have seen in Sect. 5.1 the problem can be rewritten in a weak sense
as a variational inequality. Due to the non-penetration conditions on the contact
boundary the space of the admissible solutions is restricted by an inequality
constraint and forms a convex cone K . The main difficulty in deriving the discrete
Galerkin formulation of the original problem is the discretization of K .
One possibility will be to impose the inequality constraint only in the Gauss-
Lobatto points, and then solve the resulting constrained optimization problem with
e.g. generalized conjugate gradient method, cf. [297] (see also Sect. 11.2). This
approach is nonconforming since the set of discrete solutions Khp is not a subset
of the continuous cone: Khp K . Unfortunately, this allows only to prove the a
priori error estimates with reduced rate of convergence [297, Theorem 3] (see also
Sect. 11.3)
||u − U ||H 1/2 (Γ ) ≤ C (h/p)1/4 ||u||H 3/2 (Γ ) ,
where u ∈ K solves the variational inequality and U ∈ Khp solves its discrete
version.
The other way to solve the problem is to use the penalty method, [90, 149, 266].
In this approach, we approximate the variational inequality by introducing a penalty
parameter > 0 that connects the normal displacement un , the normal boundary
stress (traction) σn and the distance g to the rigid obstacle, requiring −σn :=
−1 (u − g)+ , where (·)+ denotes the positive part of the function (11.83). The
n
penetration of the body into the obstacle is now allowed, but if the penalty parameter
is very small, it will cause a large outer pressure, which pushes the body back
and prevents large penetrations. The space of the admissible displacements is now
unconstrained and, therefore, can be discretized in a conforming way.
The total error consists now of two parts. The reduction of the element size and
increasing of the polynomial degree with fixed penalty parameter will not lead to
the convergence of the method, the same as the decreasing of the penalty parameter
alone with fixed discretization parameters. Only combined changing of , h and p
provides convergence to the exact solution. We carry out the corresponding a priori
error analysis and show (Theorem 11.14) that the convergence rate O((h/p)1− ) is
achieved, if = C̃(h/p)1− for some fixed ∈ (0, 1) and C̃ > 0.
Now we consider (11.65) with σt = 0 on ΓC . Further, we introduce the functional
spaces and sets required for the forthcoming analysis
V := H̃ 1/2(Σ), (11.75)
W := H −1/2(Γ ), (11.76)
3 4
K := v ∈ V : (vn − g)|ΓC ≤ 0 , (11.77)
"
−1/2
Λ := λ ∈ H̃ (ΓC ) : ∀v ∈ V , vn |ΓC ≤ 0, λvn ds ≥ 0 . (11.78)
ΓC
The classical problem (11.65) can be reformulated in a weak form with the
Poincaré–Steklov operator S as a variational inequality (see [299] or Sect. 11.2):
Find u ∈ K :
9 :
Su, v − u ≥ L(v − u) ∀v ∈ K , (11.79)
or equivalently as the saddle point formulation with Lagrange multiplier (cf. [266],
Appendix C.1.3): Find u ∈ V , λ ∈ Λ :
9 : 9 :
9Su, v − λ, vn : = L(v) ∀v ∈ V , (11.80)
μ − λ, un − g ≥ 0 ∀μ ∈ Λ.
The existence and uniqueness of the solution of the variational inequality is a well
known result of convex analysis (e.g. [222]). Note that both formulations (11.79)
and (11.80) include the inequality constraints, which is very inconvenient for
construction of the discrete formulation, error estimation and implementation.
The penalty formulation allows to avoid such inequality constraints in the set of
admissible solutions and to obtain a variational equation. The penalty formulation
is given as follows, [149]: Find u ∈ V :
9 : 9 :
Su , v − p , vn = L(v) ∀v ∈ V , (11.81)
1
p := − (un − g)+ . (11.82)
Here the penalty parameter > 0 must be chosen in advance. Furthermore we

denote the positive and negative part of a scalar-valued function f by
f + := (|f | + f )/2 ≥ 0, f − := (|f | − f )/2 ≥ 0, (11.83)
which provides f = f + − f − .
Next we investigate, how good the solution u of the penalty formulation (11.81)
and the function p approximate the solution (u, λ) of the saddle point formula-
tion (11.80) depending on the penalty parameter .
Now, we derive an upper bound for the energy norm of the error, caused by the
approximation of the solution of the saddle point problem by the solution of the
penalty formulation.
Theorem 11.11 Let u ∈ V , λ ∈ Λ ∩ H 1/2(ΓC ) solve the Lagrange multiplier
formulation (11.80), let u ∈ V solve the penalty formulation (11.81), and let p be
defined with (11.82). Then there holds
CS
||u − u ||H̃ 1/2 (Σ) ≤ || λ||H 1/2 (ΓC ) , (11.84)
cS α
CS2
||λ − p ||H̃ −1/2 (ΓC ) ≤ || λ||H 1/2 (ΓC ) . (11.85)
cS α 2
Proof Since u and u solve (11.80) and (11.81) respectively, there holds
9 : 9 :
Su, v − λ, vn = L(v) ∀v ∈ V ,
9 : 9 :
Su , v − p , vn = L(v) ∀v ∈ V .
Subtracting those variational equations and choosing v := u − u ∈ V we obtain

9 : 9 : 9 :
S(u − u ), u − u = λ − p , un − g + p − λ, un − g
9 : 9 :
Now λ − p , un − g ≤ 0, and p − λ, (un − g)− ≥ 0. Thus,
9 : 9 :
S(u − u ), u − u ≤ p − λ, un − g
9 :
≤ p − λ, (un − g)+ .
9 : 9 :
Definition
9 (11.82) provides
: p − λ, (un − g)+ = p − λ, − p . Further, since
p − λ, (p − λ) ≥ 0, we have
9 : 9 :
S(u − u ), u − u ≤ p − λ, − p
9 : 9 :
≤ p − λ, − p + p − λ, (p − λ)
9 :
= λ−p , λ (11.86)
≤ ||λ − p ||H̃ −1/2 (ΓC ) || λ||H 1/2 (ΓC )
CS
≤ ||u − u ||H̃ 1/2 (Σ) || λ||H 1/2 (ΓC ) ,
α
Now, ellipticity of the Steklov-Poincaré operator S yields the assertion.

Next we introduce the discrete penalty formulation as follows: Find U ∈ Vhp :
9 : 9 :
Shp U , v − P , vn = L(v) ∀v ∈ Vhp , (11.87)
where
1
P := − (Un − g)+ , > 0. (11.88)
Furthermore, for u ∈ V and U ∈ Vhp we define the traction-like functions
ψ := V −1 (K + I )u ,
Ψ ∗ := V −1 (K + I )U , (11.89)
−1 ∗
Ψ := ihp Vhp ihp (K + I )U .
Lemma 11.4 Let u ∈ V , U ∈ Vhp and traction-like functions defined by (11.89).

Then the following identity holds
9 : 9 :
||u − U ||2W + ||ψ − Ψ ||2V = 2 Su − Shp U , u − U + V (Ψ ∗ − Ψ ), ψ − Ψ ,
where
9 :1/2
||u − U ||W := W (u − U ), u − U ,
9 :1/2
||ψ − Ψ ||V := V (ψ − Ψ ), ψ − Ψ .
Lemma 11.5 For Ψ ∗ , Ψ defined in (11.89) there holds

9 :
V (Ψ ∗ − Ψ ), Φ = 0, ∀Φ ∈ Whp .
Theorem 11.12 Let u solve the continuous penalty problem (11.81), let U solve
the discrete penalty problem (11.87). Let ψ, Ψ be defined by (11.89). Then there
exists C > 0 independent of h, p, such that for ∀w ∈ Vhp , ∀Φ ∈ Whp there holds
||u − U ||H̃ 1/2 (Σ) + ||ψ − Ψ ||H −1/2 (Γ ) + || 1/2

(p − P )||L2 (ΓC )
−1/2
≤ C(||u − w||H̃ 1/2 (Σ) + ||ψ − Φ||H −1/2 (Γ ) + || (wn − un )||L2 (ΓC ) ).
Assume that u ∈ H̃ 3/2(Σ) and ψ ∈ H 1/2(Γ ). According to [45, 405] the

following approximation properties hold
h
inf ||u − w||H̃ 1/2 (Σ) ≤ C ||u ||H̃ 3/2 (Σ) , (11.90)
w∈Vhp p
h
inf ||ψ − Φ||H −1/2 (Γ ) ≤ C ||ψ||H 1/2 (Γ ) , (11.91)
Φ∈Whp p

−1/2 h 3/2 −1/2
inf || (wn − un )||L2 (ΓC ) ≤ C || un ||H 3/2 (ΓC ) . (11.92)
w∈Vhp p
Let T be the Dirichlet-to-Neumann operator, which maps the function u ∈

H̃ 1/2(Γ ) to the function T u ∈ H −1/2 (Γ ), such that the prolongation of the Cauchy
data u, T u into the domain Ω satisfies the homogeneous Lamé equation (the first
equation in (11.65)), see e.g. [391]. Employing the well-known jump conditions it
can be shown that T can be written in the non-symmetric form as
T = V −1 (K + I ). (11.93)
From (11.89) we have ψ = T u .

The approximation properties (11.90)–(11.92) combined with Theorem 11.12
yield the following a priori error estimate for the solution of the penalty formula-
tion (11.81).
Theorem 11.13 Let u ∈ H̃ 3/2(Σ) be a solution of (11.81) and T u ∈ H 1/2 (Γ ).
Let U ∈ Vhp be a solution of (11.87). Then there exists a constant C > 0
independent of h, p, , such that
||u − U ||H̃ 1/2 (Σ) + ||T u − Ψ ||H −1/2 (Γ ) + || 1/2(p − P )||L2 (ΓC )
( )
h h h 3/2 −1/2
≤C ||u ||H̃ 3/2 (Σ) + ||T u ||H 1/2 (Γ ) + || un ||H 3/2 (ΓC ) .
p p p
In order to obtain an a priori error estimate for the total error between the solutions
of problems (11.80) and (11.87) in terms of the solution of the variational inequal-
ity (11.81) we need to combine the results of Theorem 11.11 and Theorem 11.12.
Theorem 11.14 Let u ∈ H̃ 3/2(Σ), λ ∈ Λ ∩ H 1/2(Γ ) be a solution of (11.80) and
let T u ∈ H 1/2 (Γ ), where T is defined by (11.93). Let U ∈ Vhp solve (11.87).
Assume that = C̃(h/p)1−˜ for arbitrary ˜ ∈ (0; 1) and C̃ > 0. Then there exists
a constant C > 0 independent of h, p, such that

h h
||u − U ||H̃ 1/2 (Σ) ≤ C ||u||H̃ 3/2 (Σ) + + ||T u||H 1/2 (Γ ) . (11.94)
p p
11.5 h-Version BEM for a Nonmonotone Contact Problem

from Delamination
This section continues 5.3 and based on [333], it presents the h-version BEM for
the considered nonmonotone contact problem.
Now let Ω ⊂ Rd , d = 2, 3, be a polygonal domain with the boundary Γ . We start
from a triangulation Th of edges in the 2D case and triangles in the 3D case on Γ
that is consistent with the decomposition of Γ into Γ0 and ΓD . For the discretization
of the displacement u we choose continuous piecewise linear functions on Th and
define
Vh = {vh ∈ C(Γ ) : vh |E ∈ [P1 ]d−1 ∀E ∈ Th , vh = 0 on Γ¯D } ⊂ V = HD (Γ ),

1/2
KhΓ = {vh ∈ Vh : (vh · n)(Pi ) ≤ 0 ∀Pi ∈ Σh , Pi ∈ Γ¯C \Γ¯D },
where Σh is the set of all nodes of Th .

To discretize the stresses we use the space of piecewise constant functions on
Th :
Wh = {ψ ∈ L2 (Γ ) : ψ|E ∈ [P0 ]d−1 ∀E ∈ Th } ⊂ H−1/2 (Γ ).
Let {ϕi }N NN
i=1 and {ψj }j =1 be the nodal bases in Vh and Wh , respectively.
D
Then the boundary element matrices associated to the boundary integral operators
V , K, K , W are given by
N ,N N ,N
Vh = {V ψi , ψj }i,jN=1 D Kh = {Kφi , ψj }i,jD=1 N
Kh = {K ψi , φj }i,jN=1 D Wh = {W φi , φj }i,jD=1 N
N ,N N ,N
The matrix Vh is symmetric and positive definite, so it can be inverted by Cholesky

decomposition. This gives the Schur complement matrix
1
Sh = Wh + Kh + Ih Vh−1 (Kh + Ih ) .
2
With the canonical embeddings
kh : Wh *→ H−1/2 (Γ )
ih : Vh *→ H1/2 (Γ )
and their duals kh∗ and ih∗ , we obtain the discrete Poincaré-Steklov operator Sh :
Vh → Vh∗ represented by
1 ∗
Sh = ih W ih + ih∗ K + I kh (kh∗ V kh )−1 kh∗ (K + I )ih .
2
11.5 h-BEM for Delamination Problems 437
Due to [73], this operator is well-defined and satisfies
Sh uh , uh ≥ cih uh 2H1/2 (Γ ) . (11.95)
Further, we define the operator Eh : H1/2 (Γ ) → H−1/2 (Γ ), reflecting the

consistency error in the discretization of the Poincaré-Steklov operator S (see also
the previous Sect. 11.3, in particular the proof of Theorem 11.7), by
1
Eh := S − Sh = (I + K )(V −1 − ih (ih∗ V ih )−1 ih∗ )(I + K).
2
Due to [73, 297], the operator Eh is bounded and satisfies
Eh (u)H−1/2 (Γ ) ≤ c inf V −1 (I + K)u − wH−1/2 (Γ ) ∀u ∈ H1/2 (Γ ) ,

w∈Wh
(11.96)
hence by periodic polynomial spline approximation theory, see Theorem 6.12,
Eh (u)H−1/2 (Γ ) ≤ c uH1/2 (Γ ) .
Lemma 11.6
(i) If uh ! u (weak convergence) and vh → v in H1/2(Γ ), then lim Sh uh , vh =
h→0
Su, v .
(ii) If uh → u and vh ! v in H1/2(Γ ), then lim Sh uh , vh = Su, v .
h→0
Proof The part (i) follows immediately from the estimate below. Indeed, there
exists a constant c0 such that

Sh vh − ih∗ Sv, wh Vh ≤ c0 wh H1/2 (Γ ) eh (v) + vh − vH1/2 (Γ )
for any v ∈ H1/2 (Γ ) and for any vh , wh ∈ Vh , where eh (v) satisfies eh (v) → 0 as
h → 0.
Hence, using the symmetry of S and Sh , we obtain
Sh uh , vh − Su, v = Sh vh − kh∗ Sv, uh + Sv, kh uh − u

≤ c0 uh H1/2 (Γ ) eh (v) + vh − vH1/2 (Γ ) + Sv, uh − u
and thus, (i) is satisfied.

The proof of (ii) follows in the same way.

Now, we turn to the discretization of the regularized problem (5.29). To this end,
we define Π : H1/2 (Γ ) → L2 (ΓC ) by
Πuh = uh · n on ΓC .
The mapping Π is linear continuous from H1/2 (Γ ) into L2 (ΓC ), i.e.
∃c > 0 : ΠvL2 (ΓC ) ≤ cvH1/2 (Γ ) ∀v ∈ H1/2(Γ ). (11.97)
Further, we denote by V=h the image of Vh with respect to Π, i.e.
V=h = {wh ∈ C(Γ C ) : wh |E ∈ P1 (E) ∀E ∈ Th |ΓC , wh = 0 on Γ¯D },
where Th |ΓC denotes the partition of ΓC induced by Th .

Let {Pi }m
i=0 be the set of all nodes of Th lying on ΓC . To approximate the Gâteaux
derivative DJε (·), · we use Kepler’s trapezoidal rule for numerical integration and
define
DJε,h (uh ), vh :=

m−1
# ∂ S̃ ∂ S̃ $
1
2 |Pi Pi+1 | (Πuh (Pi ), ε)Πvh (Pi ) + (Πuh (Pi+1 ), ε)Πvh (Pi+1 ) .
∂x ∂x
i=0
Herewith the discretization of the regularized problem (5.29) reads as follows:

Problem (Pε,h ) Find uε,h ∈ KhΓ such that for all vh ∈ KhΓ

vh − uε,h , Sh uε,h + DJε,h (uε,h ), vh − uε,h ≥ t · (vh − uε,h ) ds. (11.98)
ΓN
Let Dh be another partitioning of ΓC consisting of elements Ki joining the

midpoints Pi−1/2 , Pi+1/2 of the edges E ∈ Th lying on ΓC sharing Pi as a common
point. If Pi is a vertex of ∂Ω then Ki is half of the edge. Moreover, if the element Ki
is linked to the boundary node Pi of ΓD , it will be added to its neighbour element
Ki+1 , see Fig. 11.5. Further, on Dh we introduce the space Yh of piecewise constant
functions by
Yh = {μh ∈ L∞ (ΓC ) : μh |K ∈ P0 (K) ∀K ∈ Dh }
Fig. 11.5 Discretization on

ΓC ; P0 is a boundary point
for ΓD [333]
and define the piecewise constant Lagrange interpolation operator Lh : V=h → Yh

by

Lh (wh )(x) = wh (Pi ) χKi (x),
Pi ∈ΓC ∩Σh
where χKi is the characteristic function of the interior of Ki in ΓC .

It holds that

DJε,h (uh ), vh ΓC = Sx (ε, Lh (Πuh ))Lh (Πvh ) ds. (11.99)
ΓC
Moreover, the operator DJε,h : Vh → Vh∗ is strongly continuous, and there exists
a constant C > 0 independent of ε and h such that
∃C > 0 : DJε,h (uh ), uh ΓC ≥ −Cuh H1/2 (Γ ) , (11.100)
for the proofs see [332].

From [200], we know that
Lh (vh · n)L2 (Γ ) ≤ 2 vh · nL2 (Γ ) (11.101)
and therefore,
vh · n − Lh (vh · n)L2 (Γ ) ≤ 3 vh · nL2 (Γ ) . (11.102)
Let now Hs (Γ j ), s ≥ 0, be the standard Sobolev space from [204, 284] defined on
the open straight pieces Γ j by
Hs (Γ j ) = {u|Γ j : u ∈ Hs (Γ )}.
R
J
According to Grisvard [204], Hs (Γ ) ⊂ Hs (Γ j ) for s ∈ [1/2, 3/2) and
j =1

J
u2Hs (Γ j ) ≤ Cu2Hs (Γ ) . (11.103)
j =1
Again from [200],
vh · n − Lh (vh · n)2L2 (Γ j ) ≤ Ch2 vh · n2H 1 (Γ j ) ≤ Ch2 vh 2H 1 (Γ j ) . (11.104)
Summing over all j and using thereafter
vh · n − Lh (vh · n)L2 (Γ ) ≤ Chvh H 1 (Γ ) . (11.105)

By interpolation between L2 (Γ ) and H 1 (Γ ) we deduce from (11.102) and (11.105)

that
vh · n − Lh (vh · n)L2 (Γ ) ≤ Ch1/2 vh H 1/2 (Γ ) . (11.106)
By the compactness of H s1 (Γ ) ⊂ H s2 (Γ ) for 0 ≤ s2 < s1 (Ω ⊂ R2 ) (see

Theorem 3.1), this gives
vh ! v in H1/2 (Γ ) ⇒ Lh (vh · n) − v · nL2 (Γ ) → 0. (11.107)
Further, we introduce the functional ϕε,h : Vh × Vh → R by
ϕε,h (uh , vh ) := DJε,h (uh ), uh . (11.108)
Due to [332], see also Sect. C.3.5 in Appendix C, this functional is pseudomonotone
and upper semicontinuous with respect to the first argument. Moreover by argu-
ments similar as in [332] one can show the following assertions:
(i) If {vh } weakly converges to v in H=1/2(Γ0 ), vh ∈ K Γ , then v ∈ K Γ .
h
(ii) For any v ∈ K there exists {vh } such that vh ∈ KhΓ and vh → v in
Γ
=1/2 (Γ0 ).
H
(iii) For any {uh } and {vh } such that uh ∈ KhΓ , vh ∈ KhΓ , uh ! u and vh → v in
V we have
lim sup ϕε,h (uh , vh ) ≤ ϕ(u, v).
(iv) There exist constants c > 0, d, d0 ∈ R and α > 1 such that for some vh ∈ KhΓ
with vh → v there holds
−ϕε,h (uh , vh ) ≥ cuh αV + duh V + d0 ∀uh ∈ KhΓ .
Based on these assertions, the general approximation result [220, Theorem 3.1],
here Theorem C.9 in Appendix C, applies to arrive at the following convergence
result.
Theorem 11.15 The problem (Pε,h ) has at least one solution uε,h . Moreover, the
family {uε,h } of solutions is uniformly bounded in V = H =1/2(Γ0 ) and any weak
accumulation point of {uε,h } is a solution to the problem (P).
Here we show the uniform boundedness of {uε,h }. Indeed, the choice vh = 0
in (11.98), and the estimates (11.95) and (11.100) lead to
cuε,h 2V ≤ Sh uε,h , uε,h ≤ tV ∗ uε,h V + ϕε,h (uε,h , 0)
= tV ∗ uε,h V + DJε,h (uε,h ), −uε,h
≤ tV ∗ uε,h V + cuε,h V .

Further, in case of uniqueness we can improve the convergence result of

Theorem 11.15 and show that the weak convergence can be replaced by the strong
one.
Theorem 11.16 Let the solutions u to (P) and uε,h to (Pε,h ) exist uniquely. Then
lim uε,h − uH

=1/2 (Γ0 ) = 0.
ε→0,h→0
Proof Let {hn } and {εn } be arbitrary sequences such that hn → 0+ and εn → 0+
as n → ∞. In view of (ii), there exists a sequence {ūεn ,hn } such that ūεn ,hn ∈ KhΓn
=1/2(Γ0 ).
and ūεn ,hn → u in V := H
Using (11.95), we obtain
cūεn ,hn − uεn ,hn 2V

≤ Sh (ūεn ,hn − uεn ,hn ), ūεn ,hn − uεn ,hn
= Sh ūεn ,hn , ūεn ,hn − uεn ,hn − Sh uεn ,hn , ūεn ,hn − uεn ,hn . (11.109)
Since ūεn ,hn → u in V and uεn ,hn ! u in V , it follows from Lemma 11.6 (ii) that
the first term on the right-hand side of (11.109) tends to zero.
Using the definition of (Pεn ,hn ), inequality (11.98), the second term can be
estimated as follows:
|Sh uεn ,hn , uεn ,hn − ūεn ,hn | (11.110)

≤ |g, uεn ,hn − ūεn ,hn | + |DJεn ,hn (uεn ,hn ), ūεn ,hn − uεn ,hn | ,
where
|DJεn ,hn (uεn ,hn ), ūn,hn − un,hn |

∂ S̃
= (Lhn (Πuεn ,hn ), εn )Lhn (Π(ūεn ,hn − uεn ,hn )) ds
ΓC ∂x
∂ S̃
≤ (Lhn (Πuεn ,hn ), εn )L2 (ΓC ) Lhn (Π(ūεn ,hn − uεn ,hn ))L2 (ΓC )
∂x
converges to zero, as follows from the boundedness of {uεn ,hn } in H =1/2 (Γ0 ),
∂ S̃
(11.107) and the boundedness of { (Lhn (Πuεn ,hn ), εn )} in L2 (Γ ), what we show
∂x
next.
From

∂ S̃

∃c > 0 : (ε, x) ≤ c(1 + |x|) ∀x ∈ R ,
∂x
the elementary inequality (a + b)2 ≤ 2(a 2 + b 2 ) and integration over Γ we obtain
2
∂ S̃

(ε, Lh (Πuh (s)) ds ≤ 2c2 meas (ΓC ) + 2c2 Lh (Πuh )2L2 (ΓC ) .
ΓC ∂x
Hence,
0 0
0 ∂ S̃ 0
0 0 1/2
0 (Lh (Πuh ), ε)0 ≤ 2c2 meas (ΓC ) + 2c2 Lh (Πuh )2L2 (Γ
0 ∂x 0 C)
L2 (ΓC )
√
≤ 2c (meas (ΓC ))1/2 + Lh (Πuh )L2 (ΓC )
≤ c̃(1 + uh H1/2 (ΓC ) ), (11.111)
where we have used (11.102) and (11.97), and the elementary inequality
√
a 2 + b2 ≤ |a| + |b|.
Passing now to the limit superior in (11.111), we get
lim sup Sh uεn ,hn , uεn ,hn − ūεn ,hn ≤ 0.

n→∞
Hence, (11.109) entails in the limit
ūεn ,hn − uεn ,hn V → 0.
Finally, from the triangle inequality
uεn ,hn − uV ≤ uεn ,hn − ūεn ,hn V + ūεn ,hn − uV ,
we get the strong convergence of {uεn ,hn } to u in V .

As an advantage of the combination of regularization methods of nondifferen-
tiable optimization with the h-BEM we arrive at smooth optimization problems at
the discrete level which can be solved by standard optimization methods, like trust
region methods [111]. For numerical experiments using h-FEM instead of h-BEM
we can refer to [220].
Similar nonmontone contact problems from adhesion have been treated by the
h-BEM directly in [326]. Then special nonsmooth optimization solver, like bundle
methods, have to be employed at the discrete level. For the convergenc analysis of
the h-BEM and numerical results for a similar benchmark problem along this latter
approach we refer to [326].
For further reading we refer to [27, 147, 149, 150] where boundary integral
equations and boundary element methods for related contact problems are treated,
especially in [148] different adaptive methods are presented.
11.6 hp-BEM for Delamination Problems 443
11.6 hp-BEM for Delamination Problems
To avoid domain approximation, let Ω ⊂ Rd , d = 2, 3, be a polygonal domain.

Let Th be a sufficiently fine finite element mesh of the boundary Γ respecting the
decomposition of Γ into ΓD , ΓN and ΓC , p = (pT )T ∈Th a polynomial degree
distribution over Th , PpT (T̂ ) the space of polynomials of order pT on the reference
element T̂ , and ΨT : T̂ → T ∈ Th a bijective, (bi)-linear transformation. In 2D,
T̂ is the interval [−1, 1], whereas in 3D it is the reference square [−1, 1]2. Let
ΣT ,hp be the set of all (pT + 1)d−1 affinely transformed (tensor product based)
Gauss-Lobatto nodes on the element T of the partition Th of Γ , and set Σhp :=

ΣT ,hp , see [218, 275, 297]. Furthermore, we assume in this section that
T ∈Th |ΓC
g ∈ C 0 (ΓC ) to allow point evaluation.
For the discretization of the displacement u we use
Vhp = {vhp ∈ C0 (Γ ) : vhp |T ◦ ΨT ∈ [PpT (T̂ )]d ∀ T ∈ Th , vhp = 0 on Γ D },

Khp
Γ
= {vhp ∈ Vhp : (vhp · n)(Pi ) ≤ g(Pi ) ∀ Pi ∈ Σhp }.
In general Khp Γ ⊆ K Γ . For the approximation S

hp of the Poincaré-Steklov
operator, we need the space
Whp = {ψhp ∈ L2 (Γ ) : ψhp |T ◦ ΨT ∈ [PpT −1 (T̂ )]d ∀ T ∈ Th } ⊂ H−1/2 (Γ ).
Now, we turn to the discretization of the regularized problem (Pε ), see (5.29)
in Sect. 5.3. The discretized regularized problem (Sε,hp ) is: Find uεhp ∈ Khp
Γ such
that for all vhp ∈ Khp

Γ
Shp uεhp , vhp −uεhp Γ0 +DJε (uhp ), vhp −uhp ΓC

ε ε
≥ F, vhp −uεhp Γ0 . (11.112)
Lemma 11.7 Let uε ∈ K Γ be the solution of the problem (Pε ) and let uεhp ∈ Khp Γ
be the solution of the problem (Pε,hp ). Assume that α0 < cS in (5.30), where cS is
the coerciveness constant of S, further uε ∈ H3/2(Γ ), g ∈ H 3/2 (ΓC ) and Suε − F ∈
L2 (Γ ). Then there exists a constant c = c(uε , g, F) > 0, but independent of h and
p such that
cuε − uεhp 2H1/2 (Γ ) ≤ Ehp (uε )2H−1/2 (Γ )
+ inf {Suε − FL2 (Γ ) uεhp − vL2 (Γ ) + DJε (uε ), v − uεhp ΓC }

v∈K Γ
&
+ inf uε − vhp 2H1/2 (Γ ) + Suε − FL2 (Γ ) uε − vhp L2 (Γ )
vhp ∈Khp
Γ
'
+DJε (uεhp ), vhp − uε ΓC .
Proof Using the definitions of (Pε ) and (Pε,hp ), and estimates similar to [297,
Theorem 3], we obtain for all v ∈ K Γ , vhp ∈ Khp
Γ
cP uε − uεhp 2H1/2 (Γ ) ≤ Ehp (uε )2H−1/2 (Γ ) + uε − vhp 2H1/2 (Γ )
+ Suε − FL2 (Γ ) uε − vhp L2 (Γ ) + uεhp − vL2 (Γ ) + D,
where we abbreviate
D = DJε (uε ), v − uε ΓC + DJε (uεhp ), vhp − uεhp ΓC .
To bound the term D, we use (5.30) and estimate as follows:
D = DJε (uε ), v − uεhp ΓC + DJε (uεhp ), vhp − uε ΓC
+ DJε (uε ) − DJε (uεhp ), uεhp − uε ΓC
≤ DJε (uε ), v − uεhp ΓC + DJε (uεhp ), vhp − uε ΓC + α0 uε − uεhp 2V .
Therefore, since α0 < cS by assumption, we obtain the assertion.

Theorem 11.17 Let uε ∈ K Γ be the solution of the problem (Pε ) and let uεhp ∈
KhpΓ be the solution of the problem (P
ε,hp ). Assume that α0 < cS in (5.30), uε
∈ H (Γ ), g ∈ H (ΓC ) and Suε − F ∈ L2 (Γ ). Then there exists a constant
3/2 3/2
c = c(uε , g, F) > 0, but independent of h and p such that
uε − uεhp H1/2 (Γ ) ≤ ch1/4p−1/4 . (11.113)
Proof Taking into account the estimates obtained by Maischak and Stephan in their
Theorem 3 in [297] for the consistency error, the approximation error, and for
Ehp uH−1/2 (Γ ) , we only need to estimate
DJε (uε ), v∗ − uεhp ΓC (11.114)
and
DJε (uεhp ), vhp − uε ΓC . (11.115)
To estimate (11.114)–(11.115) we must consider the same test functions v∗ and

vhp as in [297, Theorem 3] used to estimate the standard error terms. Let v∗ ∈
K Γ ∩ H1 (Γ ) be defined by
⎧
⎪
⎨ uhp,t + [g + inf{uhp,n − ghp , 0}]n on ΓC
ε ε
∗
v := 0 on ΓD
⎪
⎩ γ uε
N hp on ΓN ,
where ghp := Ihp g is the interpolate of the gap function g, and γN is the trace
map onto ΓN .
As shown in [334, Lemma 2]

∗ ∂
DJε (uε ), v − uhp ΓC =
ε
S̃(uε,n , ε)(vn∗ − uεhp,n ) ds
ΓC ∂x

≤ c 1 + uε,n L2 (ΓC ) vn∗ − uεhp,n L2 (ΓC ) . (11.116)
The elaborate analysis in [297], see the proof of Theorem 3 there, gives
vn∗ − uεhp,n L2 (ΓC ) ≤ C2 h1/2 p−1/2 gH 1/2 (ΓC ) + uεhp H 1/2 (ΓC ) . (11.117)
Further, let vhp := Ihp uε ∈ Khp

Γ be the interpolate of u ∈ H3/2 (Γ ) ⊂ C 0 (Γ ).
ε
Analogously to (11.116), we have

∂
DJε (uεhp ), vhp − uε ΓC = S̃(uεhp,n , ε)(vhp,n − uε,n ) ds
ΓC ∂x
≤ c 1 + uεhp,n L2 (ΓC ) uε,n − vhp,n L2 (ΓC ) .

(11.118)
By [47, Theorems 4.2 and 4.5] and by the real interpolation between H 1 (Γ ) and
L2 (Γ ) there exists a constant C1 > 0 such that
uε,n − vhp,n H 1/2 (Γ ) ≤ C1 h1 p−1 uε H 3/2 (Γ ) . (11.119)
Finally, combining the error estimates for the interpolation (11.119) and the
consistency (11.117) with (11.118) and (11.116), respectively, and taking into
account the boundedness of uεhp,n in H 1/2(ΓC ), we prove the asserted bound
for (11.114) and (11.115).

To be able to split the approximation error into the discretization error of a
simpler variational equation and contributions arising from the constraints on ΓC
we introduce the mixed regularized formulation (11.120a)–(11.120b), which is
equivalent to the regularized problem (Pε ).
Find (uε , λε ) ∈ V × M(uε ) such that
9 ε : 9 :
Su , v Γ + λε , vn Γ = F, v Γ0 ∀v ∈ V (11.120a)
0 C
9 :
μ − λ , un − g Γ ≤ 0
ε ε
∀ μ ∈ M(uε ) (11.120b)
C
with the set of admissible Lagrange multipliers

& 9 : '
M(uε ) := μ ∈ X∗ : μ, η ΓC ≥ DJε (uε ), η Γ ∀ η ∈ X, η ≥ 0 a.e. on ΓC
C
where X = {w | ∃ v ∈ V , vn |ΓC = w} ⊆ H 1/2(ΓC ) and X∗ its dual space.

Lemma 11.8
(i) Let uε solve the regularized problem (Pε ), then there exists a λε ∈ M(uε ) such
that (uε , λε ) solves (11.120).
(ii) Let (uε , λε ) solve (11.120), then uε solves (Pε ).
Given the discrete solution uεhp ∈ Khp
Γ to (P
ε,hp ), we reconstruct λhp ∈
ε
span {ψi }i=1 such that

M
7 8 7 8
λεhp , vn = F, v Γ0 − Shp uεhp , v ∀ v ∈ Vhp (11.121)
ΓC Γ0
by solving a potentially over-constrained system of linear equations for an arbitrary

choice of basis {ψ}.
Following the Braess trick [59] as e.g. in [37], we define the auxiliary problem
7 8
z∈V : Sz, v Γ0 = F, v Γ0 − λεhp , vn ∀v ∈ V . (11.122)
ΓC
Subtracting (11.120a) and (11.122) yields

9 : 7 8
S(uε − z), v Γ = λεhp − λε , vn ∀v ∈ V (11.123)
0 ΓC
and additionally with the continuous inf-sup condition [101, Theorem 3.2.1] this
yields (see [37])
0 0 C0 0 0 0 C0 0
0 ε 0 0uε − z0 ≤ C 0 ε 0 0 ε 0
0λhp − λε 0 ≤ V 0uε
− uhp 0 + 0uhp − z 0 (11.124)
X∗ β β V β V
with inf-sup constant β > 0. See [101, Theorem 3.2.1] for a proof of the inf-sup
condition for the difficult case when Γ¯C ∩ Γ¯D = ∅, i.e. X∗ = H̃ −1/2 (ΓC ).
∂
Theorem 11.18 Under the assumption (5.32) and if ∂x S̃(·, ε) is Lipschitz con-
tinuous, then there exists a constant C independent of h and p such that for
0 < ς < cP −α
4
0
arbitrary

C 1 00 ε
02
0
(cP − α0 − 4ς) uε − uεhp 2V ≤ + 1 z − uεhp 2V + 0(λhp − DJε (uεhp ))− 0 ∗
ς 4ς X
0 0
1 1 1 0 ε 02
+C + 2 + 2
0(uhp,n − g)+ 0
ς β ςβ X
7 8
− (λεhp − DJε (uεhp ))+ , (uεhp,n − g)−
ΓC
with (uεhp , λεhp ) satisfying (11.121), z solving (11.122) and uε solving (Pε ).
The a-posteriori error estimate decomposes into the discretization error of a

variational equality z − uεhp 2V , which can be further estimated by e.g. residual
error estimates [75] or bubble error estimates, e.g. [37], and violation of the
0 02
0 0
consistency condition 0(λεhp − DJε (uεhp ))− 0 ∗ , violation of the non-penetration
0 02 X
0 0
condition 0(uεhp,n − g)+ 0 and violation of the complementarity condition
7 X 8
− (λεhp − DJε (uεhp ))+ , (uεhp,n − g)− . Localizing an approximation of the
ΓC
global a-posteriori error estimate gives rise to the following solve-mark-refine
algorithm for hp-adaptivity.
For the numerical experiments we choose Ω = (0, 1/2)2, ΓD = {0} × [0, 1/2],
ΓC = (0, 1/2] × {0}, ΓN = ∂Ω \ (ΓD ∪ ΓC ). The material parameters are E = 5,
ν = 0.45, f ≡ 0, t = 0.25 on [1/4, 1/2] × {1} and zero elsewhere, g = 0. The
delamination law is given via
f (un (x)) = min{g1 (g(x) − un (x)), g2 (g(x) − un (x)), g3 (g(x) − un (x))}

= − max{−g1 (−un (x)), −g2 (−un (x)), −g3 (−un (x))}
with
A1 2
g1 (y) = y , g2 (y) = b2 (y 2 − t12 ) + d2 , g3 (y) = d3
2t1
and parameters
A1 = 0.05, A2 = 0.03, t1 = 0.02, t2 = 0.04,

A2 t1
b2 = , d2 = A1 , d3 = b2 (t22 − t12 ) + d2 .
2t2 2
The regularized delamination law Sx with regularization parameter ε = 10−4 is

plotted in Fig. 11.6. The characteristic saw tooth shape is already present, but the
absolute value in the tips and the slope approximating the jump are still noticeable
coarse approximated.
The discrete Lagrange multiplier λεhp is obtain by solving (11.121) where ψi are
discontinuous, piecewise polynomials on ΓC on a one time coarsened mesh (H =
2h) with polynomial degree reduced by one (q = p − 1) compared to the mesh
and polynomial degree distribution of uεhp . Figure 11.7 displays the deformation of
the rectangle and the normal stresses on ΓC obtained from the lowest order uniform
h-method with 16384 elements and regularization parameter ε = 10−4 . The normal
stress on ΓC , Fig. 11.7b, reflects the delamination law from Fig. 11.6 well.
Figure 11.8 displays the reduction of the error in (u, λ) and of the error estimate.
Since the exact solution is not known, we compute the error approximately by
uf ine − uhp S and λf ine − λhp V , with norms induced by the Poincaré-Steklov
0.05
0.04
0.03
-Sx (u n )
0.02
0.01
0
0 0.02 0.04 0.06
g-u n
Fig. 11.6 Regularized delamination law Sx for ε = 10−4
0.6
0.05
0.5 0.04
normal contact stress
0.4 0.03
0.3 0.02
0.2 0.01
0
0.1
-0.01
0 0 0.1 0.2 0.3 0.4 0.5
-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6
ΓC
(a) Deformation (b) Contact stresses
Fig. 11.7 Uniform h-version, p=1, 16384 elements with 4096 elements on ΓC . (a) Deformation.
(b) Contact stresses
operator S and the single layer potential V acting on ΓC , respectively. The pair
(uf ine , λf ine ) is a very fine (last) approximation for each sequence of discretization.
10 -1
error approx., uniform h, p=1
error estimate, uniform h, p=1
error approx., h-adaptive
error estimate, h-adaptive
error approx., hp-adaptive
10 -2 error estimate, hp-adaptive
10 -3
10 -4
10 1 10 2 10 3 10 4 10 5
Degrees of Freedom
/
Fig. 11.8 uf ine − uhp 2P + λf ine − λhp 2V and error estimate for different families of dis-
crete solutions, ε = 10−4 [334]
Chapter 12
FEM-BEM Coupling
The BEM is well established for the solution of linear elliptic boundary value
problems. Its essential feature is the reduction of the partial differential equation in
the domain to an integral equation on the surface. Then, for the numerical treatment,
only the surface has to be discretized. This leads to a comparatively small number of
unknowns. It is possible to solve problems in unbounded domains. In contrast, the
FEM requires a discretization of the domain. However, when dealing with nonlinear
problems, the latter method is more versatile. Typical examples for which the
coupling of both methods is advantageous are rubber sealings and bearings that are
located between construction elements made of steel, concrete, or glass. For these
elements, linear elasticity often is a sufficient model, and the BEM is favorable.In
contrast, for sealings and bearings the nonlinear material behavior imply that the
FEM is preferable. Moreover, for rubberlike materials the incompressibility has
to be taken into account, which requires mixed finite elements. Thus to combine
the advantages of both discretization methods we are led to study FEM-BEM
coupling, “marriage a‘ la mode” [439]. Here we focus first to symmetric coupling
and consider two alternative approaches: (i) the abstract setting of saddle point
problems introduced by Costabel and Stephan in [136] is reported in Sect. 12.1
with application to an elastic interface problem in Sect. 12.2.1, (ii) the use of the
Poincare-Steklov operator in the variational formulation as given by Carstensen and
Stephan in [91] is reported in Sect. 12.3 together with adaptive coupling versions
using residual type error indicators in Sect. 12.3.2 and hierarchical type error
indicators in Sect. 12.3.3. In Sect. 12.3.5 we report on other coupling methods like
Johnson-Nedelec coupling and Bielak-MacCamy coupling. Other topics are least
squares FEM-BEM couplings in Sect. 12.4 and FEM/BEM coupling for Signorini
contact problems in Sect. 12.5 with a primal method in Sect. 12.5.1 and dual
methods in Sect. 12.5.2. In Sect. 12.6 we consider a primal mixed FEM-BEM
coupling for plane elasticity. An elliptic interface problem with a strongly nonlinear
differential operator is considered in Sect. 12.7. In Sect. 12.8 the time-harmonic

452 12 FEM-BEM Coupling
eddy-current problem is treated with Hiptmair’s symmetric coupling method and

a posteriori error estimates of residual type as well as of p-hierarchical type are
presented. Section 12.9 presents a discontinuous (in time) Galerkin method for a
parabolic-elliptic interface problem. For further reading we refer to the seminal
works of M.Costabel [115] and H.Han [228]; see also [121, 132, 133, 439].
12.1 Abstract Framework of Some Saddle Point Problems
In this section we report from [136]. As investigated there the direct boundary
integral
equation method for the interface problem leads to an operator of the form
P Q
where P and A are strongly elliptic operators and Q is the adjoint of
Q −A
Q. The same structure is shared by the equations that arise from the symmetric
coupling method. This typical form leads to critical points of saddle point type for
functionals that are strictly convex in one direction and strictly concave in the other
direction. As we will see, this situation can be reduced to the study of the minimum
of a strictly convex functional. Due to the strict monotonicity of both operators P
and A, we do not need any analogue of the Babuška-Brezzi stability condition.
Let X, Y be reflexive Banach spaces with duals X , Y . Let P : X → X be a
nonlinear operator. Let Q : Y → X be a continuous, linear operator. Let A : Y →
Y be a continuous, linear, bijective operator. We define P1 : X × Y → X × Y by

u P (u) + Qφ
P1 := .
φ Q u − Aφ
We identify P1 with the matrix

P Q
P1 = .
Q −A
In addition, we define the operator P2 : X → X by
P2 (u) := P (u) + QA−1 Q u
Then, if 1X and 1Y denote the respective identity mappings, there holds the
following relation:

1X QA−1 P2 0
· P1 = (12.1)
0 1Y Q −A
This relation is the key to the proof of the following equivalence theorem.
12.1 Abstract Framework of Saddle Point Problems 453
Theorem 12.1
(i) P1 is surjective if and only if P2 is surjective.
(ii) P1 is injective if and only if P2 is injective.
(iii) P1 has a bounded inverse if and only if P2 has a bounded inverse.
Proof The operator

1X QA−1
0 1Y
on X × Y is a bounded, linear, bijective operator with bounded inverse

1X −QA−1
.
0 1Y
Therefore (12.1) shows that the statements of the theorem are equivalent to the
following.
The operator P2 : X → X is surjective, injective, or boundedly invertible if and
only if the operator

P2 0
: X × Y → X × Y
Q −A
has the respective properties. The claimed equivalence is now clear, because A was
assumed to be bijective. It can be seen from the equality
−1
P2 0 P2−1 0
=

Q −A A Q P2−1
−1 −A−1
Next we assume that the problem is given in variational form, i.e. the operators
are given as derivatives of functionals. As an example we present in Sect. 12.2.1 the
elastoplastic interface problem for material obeying the Hencky-von Mises stress-
strain relation. We introduce
• J : X → R a functional defined everywhere.
• Q : Y → X linear, continuous, q : X × Y → R bilinear form with q(u, φ) :=
Qφ, u = Q u, φ .
• A : Y → Y linear, continuous, bijective, self-adjoint, a : Y → R,
a(φ) := 12 Aφ, φ .
• J1 : X × Y → R, J1 (u, φ) := J (u) + q(u, φ) − a(φ) .
• J2 : X → R, J2 (u) := J (u) + 12 q(u, A−1 Q u) = J (u) + 12 QA−1 Q u, u .
If J is Gateaux differentiable, we denote its Gateaux derivative by DJ (u) ∈ X and
its Gateaux derivative in direction w by DJ (u, w) ∈ R.
Theorem 12.2 Let J be differentiable on X. Then

(i) J1 and J2 are differentiable
(ii) For (u, φ) ∈ X × Y there holds

1X QA−1 DJ2 (u)
· DJ1 (u, φ) = ∈ X × Y (12.2)
0 1Y Q u − Aφ
(iii) For u ∈ X there holds DJ2 (u) = 0 if and only if there exists φ ∈ Y with
DJ1 (u, φ) = 0. If this is satisfied, then
φ = A−1 Q u
Proof Let u, w ∈ X. Then from the definition of J2 and the symmetry of A−1 there
follows
DJ2 (u, w) = DJ (u, w) + QA−1 Q u, w ,
hence DJ2 (u) = DJ (u) + QA−1 Q u.

Let ψ ∈ Y . Then the definition of J1 implies
DJ1 (u, φ; w, ψ) = DJ (u; w) + Qφ, w + Qψ, u − Aφ, ψ
hence

DJ (u) + Qφ
DJ1 (u, φ) = ;
Q u − Aφ
thus J1 and J2 are differentiable and (12.2) is already verified, see (12.1) with
P = DJ . In order to show (iii), assume that DJ2 (u) = 0 is satisfied and define
φ ∈ Y by φ := A−1 Q u. Then the right hand side in (12.2) vanishes, and
therefore DJ1 (u, φ) = 0. Conversely, (12.2) also shows that DJ1 (u, φ) = 0 implies
DJ2 (u) = 0 and φ = A−1 Q u.

Lemma 12.1 Assume J is twice continuously differentiable and there exist con-
stants λ, Λ > 0 such that ∀u, w ∈ X :
λw2X D 2 J (u; w, w) Λw2X . (12.3)
Assume further that
a(φ) 0 ∀φ ∈ Y. (12.4)
12.2 Galerkin Approximation of Saddle Point Problems 455
Then there is a Λ Λ such that for all u, w ∈ X
λw2X D 2 J2 (u; w, w) Λ w2X .
Proof The assertion follows from
D 2 J2 (u; w, w) = D 2 J (u; w, w) + QA−1 Q w, w = D 2 J (u; w, w) + 2a(A−1 Q w)

In the following we assume (12.3) and (9.19) to hold. As a corollary, we obtain an
existence and uniqueness result for critical points of J1 .
Theorem 12.3
(i) J2 has exactly one critical point u ∈ X. This is a minimum.
(ii) J1 has exactly one critical point (u, φ) ∈ X × Y . This is a saddle point:
J1 (u, φ + ψ) J1 (u, φ) ∀ψ ∈ Y
J1 (u + w, φ) J1 (u, φ) ∀w ∈ X
Proof From Lemma 12.1 it is clear that J2 has precisely one critical point u which
is a minimum, because J2 is coercive , lower semicontinuous, and strictly convex.
From Theorem 12.2 we see that J1 has a unique critical point (u, φ) ∈ X × Y and
that Q u = Aφ holds. Thus for ψ ∈ Y we obtain with (9.19)
1
J1 (u, φ + ψ) = J (u) + Q u, φ + ψ − A(φ + ψ), φ + ψ
2
= J (u) − a(φ) + Q u, φ − a(ψ) = J1 (u, φ) − a(ψ) J1 (u, φ).
φ
In order to show the second saddle point inequality we define a functional J1 by
φ φ
J1 (w) := J1 (w, φ). It satisfies D 2 J1 (w) = D 2 J (w) and it is strictly convex due
to (12.3). Hence its critical point u is a minimum.

12.2 Galerkin Approximation of Saddle Point Problems
Suppose (12.3) and (12.4) hold, and J1 has a unique critical point (u, φ) ∈ X × Y .
Let XN ⊂ X, YN ⊂ Y be closed subspaces of finite dimension. Let dX (w, XN ) :=
inf{w − vX | v ∈ XN }, dY (ψ, YN ) denote the distances to XN , YN . The
restriction of J1 to XN × YN inherits all relevant properties from J1 . Thus due to
Theorem 12.3 it has exactly one critical point (uN , φN ) ∈ XN × YN . There holds
Theorem 12.4 There exist exactly one (uN ,φN) ∈ XN × YN such that
DJ1 (uN , φN ;w,ψ)= 0 for all (w, ψ) ∈ XN × YN .
There exists C > 0 independent of XN , YN such that
u − uN X + φ − φN Y C (dX (u, XN ) + dY (φ, YN ))
To prove Theorem 12.4 we need some elementary consequences of the strong

ellipticity assumption (12.3).
Lemma 12.2 Let Φ : X → R be twice continuously differentiable and assume
there exist λ, Λ > 0 such that λw2X D 2 Φ(v; w, w) Λw2X ∀v, w ∈ X.
Then
λv − w2X DΦ(v; v − w) − DΦ(w; v − w) ∀v, w ∈ X (12.5)
and
λ Λ
v − w2X Φ(v) − Φ(w) − DΦ(w; v − w) v − w2X , ∀v, w ∈ X
2 2
(12.6)
The functional Φ has a unique minimum u ∈ X. The restriction of Φ to XN has a

unique minimum u∗N ∈ XN and there holds
λ Λ
u − u∗N 2X Φ(u∗N ) − Φ(u) dX (u, XN )2 (12.7)
2 2
Proof (12.5) follows from
1
λv−w2X D 2 Φ(w+t (v−w); v−w, v−w)dt = DΦ(v; v−w)−DΦ(w; v−w).
0
(12.8)
similarly (12.6) follows from
1 1
λ
v − w2X D 2 Φ(w + tτ (v − w); v − w, v − w)tdtdτ
2
0 0
= Φ(v) − Φ(w) − DΦ(w; v − w)

Λ
v − w2X
2
Finally for (12.7) we choose an arbitrary w ∈ XN and obtain from

DΦ(u; v) = 0∀v ∈ X, Φ(w) Φ(u∗N ) and (12.6)
λ Λ
u − u∗N 2X Φ(u∗N ) − Φ(u) Φ(w) − Φ(u) w − u2X

2 2
Now let J1 N be the restriction of J1 to XN × YN . Then we have for (w, ψ) ∈

XN × YN
J1 N (w, ψ) = J (w) + q(w, ψ) − a(ψ)

1
= J (w) + QN ψ, w − AN ψ, ψ
2
, A : Y → Y are defined by the relations:
where the operators QN : YN → XN N N N
QN ψ, w = Qψ, w ∀(w, ψ) ∈ XN × YN

AN ψ, χ = Aψ, χ ∀(ψ, χ) ∈ YN × YN .
Note that AN : YN → YN is invertible, see Theorem 1.1 i).

Lemma 12.3 Let t ∈ Y be given and ψ := A−1 t ∈ Y, ψN := A−1 N PN t ∈ YN
where PN : Y → YN is the natural projection, i.e. PN t, χ = t, χ ∀χ ∈ YN .
Then
ψ − ψN Y C · dY (ψ, YN ) (12.9)
Proof ψN solves the Galerkin equation
AψN , χ = t, χ ∀χ ∈ YN .
Hence (12.9) is the quasi-optimality of the Galerkin error for the self-adjoint
operator A, see Theorem 1.1 iii) in the more special situation of a Hilbert
space Y .

Proof (of Theorem 12.4) Define
1
J2N (w) := J (w) + QN A−1
N QN w, w ∀w ∈ XN
2
Note J2N does not coincide with the restriction of J2 to XN . There are unique critical
points uN , u∗N satisfying
DJ2N (uN ; w) = DJ (uN ; w) + QN A−1

N QN uN , w = 0 ∀w ∈ XN
DJ2 (u∗N ; w) = DJ (u∗N , w) + QA−1 Q u∗N , w = 0 ∀w ∈ XN (12.10)

Note
DJ2 (u; w) = DJ (u; w) + QA−1 Q u, w = 0 ∀w ∈ X.
Now by Lemma 12.2, (12.7)

A
Λ
u − u∗N X dX (u, XN ). (12.11)
λ
Thus, in order to obtain an error estimate for u − uN , we estimate u∗N − uN as

follows: With Φ = J2N on XN we get
λu∗N − uN 2N DJ2N (u∗N , u∗N − uN )

= DJ (u∗N , u∗N − uN ) + QN A−1 ∗ ∗
N QN uN , uN − uN (12.12)
With (12.10) we can rewrite this as
DJ2 (u∗N ; u∗N − uN ) + QN A−1 ∗ ∗ −1 ∗ ∗

N QN uN , uN − uN − QA Q uN , uN − uN
∗
= A−1 −1 ∗
N QN − A Q uN , Q uN − uN

= A−1
N PN − A
−1
Q u∗N , Q u∗N − uN
& '
≤ A−1N PN −A
−1
Q (u∗N − u)Y + A−1N PN −A
−1
Q uY Q (u∗N −uN )Y
& '
C · u∗N − uX + dY A−1 Q u, YN · u∗N − uN X
Here we used
ψ, QN w = QN ψ, w = Qψ, w = ψ, Q w = ψ, PN Q w ∀(w, ψ) ∈ XN × YN
and the stability of the Galerkin scheme for the operator A and (12.9).
Next, note DJ1 (u, φ) = 0 with φ = A−1 Q u by Theorem 12.2.
Then (12.11), (12.12) yield

u∗N − uN X C u∗N − uX + dY (φ, YN ) C (dX (u, XN ) + dY (φ, YN ))
and hence
u − uN X C (dX (u, YN ) + dY (φ, YN )). (12.13)
∗ + φ ∗ − φ with
In order to estimate φN − φY we write φN − φ = φN − φN N
φN∗ := A−1 P Q u, where φ = A−1 Q u = A−1 P Q u .
N N N N N N N N N
Therefore
∗
φN − φN Y = A−1
N PN Q (uN − u)Y CuN − uX
and again by (12.9,

∗
φN − φY = A−1
N PN − A
−1
Q uY C · dY (φ, YN ).
Together with (12.13) this gives
φN − φY C (dX (u, XN ) + dY (φ, YN )) .

12.2.1 Symmetric FE/BE Coupling for a Nonlinear Interface

Problem
Let us consider the interface problem (TMP): For given F, u0 find u1 , u2 satisfying
P1 (u1 ) = F in Ω1 , P2 (u2 ) := Δ∗2 u2 = 0 in Ω2 , u1 = u2 , t1 = t2 on Γc , u1 = u0
on Γ1 with the geometry as in Fig. 12.1, the Lamé operator Δ∗2 and P1 (u1 )l :=
3
∂xl (k − 2/3μ(Γ (u1 ))) div u1 + j =1 2 ∂xl μ(Γ (u))eij (u1 ), l = 1, 2, 3.
∂ ∂
Here the nonlinear material is described by the Hencky - von Mises stress-strain
relation. We set

3
Φ1 (u, w) := {k − 2/3μ(Γ (u)) div u div w + 2μ(Γ (u))eij (u)eij (w)}dx
Ω1 j =1
with the bulk modulus k, the Lamé function μ(Γ ) for the nonlinear material and the
strain tensor eij . Then the weak formulation of (TMP) reads: Find
"
3
u ∈ HΓ11 (Ω1 ) = u ∈ H 1 (Ω1 ) , u = 0 on Γ1 , φ ∈ H −1/2(Γc )
Fig. 12.1 Geometrical Ω2

setting
Γc
Γ1
Ω1
such that
b(u, φ; w; ψ) = l(w, ψ) ∀w ∈ HΓ11 (Ω1 ), ψ ∈ H −1/2 (Γc ). (12.14)
Here

z v
b(u, φ; w, ψ) = Φ1 (u, w) − φ · wds + , (1 − C2 ) (12.15)
−ψ φ
Γc
where

z v
z = w|Γc , , v = u|Γc , , := z, φ + v, ψ
ψ φ
and

1/2 + Λ2 −V2
C2 = , l(w, ψ) := F · wdx for F ∈ (L2 (Ω1 ))3
−W2 1/2 − Λ2
Ω1
Explicitly (12.15) becomes

1 1
b(u, φ; w, ψ) = Φ1 (u, w)+z, W2 v −z, − Λ2 φ − − Λ2 v, ψ −ψ, V2 φ
2 2
Here V2 , Λ2 , Λ2 , W2 are the boundary integral operators of single layer, double
layer, and it’s adjoint and it’s traction for the Lamé operator.
We use L1 , L2 .
"
3 3
L1 = u1 ∈ H 1 (Ω1 ) , P1 u1 = F ∈ L2 (Ω1 ) , u = 0 on Γ1
"
3 1
L2 = u2 ∈ Hloc
1
(Ω 2 ) , P2 u2 = 0 on Ω2 and u2 = O( ) as |x| → ∞
|x|
Inspection shows that (12.15) is a weak form of the Euler equation

DJ1 (u, φ; w, ϕ) = 0 with J1 (u, φ) := J (u) + q(u, φ) − a(φ) where v = u|Γc .
Here we have set J (u) = J0 (u) + 1/2v, W2 v with
Γ (u)
J0 (u) = {(1/2)k| div u|2 + μ(t)dt − F · u}dx,
Ω1 0
and q(u, φ) = φ, (Λ2 − 1/2)v , a(φ) = 1/2φ, V2 φ .

Then there holds

Theorem 12.5 Let F ∈ (L2 (Ω1 ))3 . Then, if uj ∈ Lj (j = 1, 2) solve (TMP),
then u = u1 and φ = T1 (u1 )|Γc solve (12.14). Conversely, if (u, φ) ∈ HΓ11 (Ω1 ) ×
H −1/2(Γc ) satisfy (12.14) and if u2 is defined by the representation formula

u2 (x) = {T2 (x, y)v2 (y) − G2 (x, y)φ2 (y)} dsy (12.16)
Γc
with v2 = v = u|Γc and φ2 = φ on Γc , then u1 = u ∈ L1 and u2 ∈ L2 solve

(TMP). Here
"
λ2 + 3μ2 1 λ2 + μ2 (x − y)(x − y)T
G2 (x, y) = I+ ,
8πμ2 (λ2 + 2μ2 ) |x − y| λ2 + 3μ2 |x − y|3
T2 (x, y) := T2,y (G2 (x, y))T .
Proof Let (u, φ) satisfy(12.14) and define u1 = u ∈ HΓ11 (Ω1 ) ,and u2 ∈ L2

via (12.16) with v2 = v = u|Γc and φ2 = φ. Then from the definition of C2
we have

1
u2 |Γc = + Λ2 v − V2 φ (12.17)
2
and

1
T2 (u2 )|Γc = −W2 v + − Λ2 φ (12.18)
2
Setting w = 0 in (12.14), and taking ψ ∈ H −1/2(Γc ) arbitrarily gives with (12.16)

the relation

1
− Λ2 v + V2 φ = 0
2
Hence with (12.17) u2 |Γc = v and thus u1 = u2 on Γc .

Next, take ψ = 0 in (12.14) and w ∈ HΓ11 (Ω1 ) arbitrarily. Then comparison with

P1 u · w∂x = Φ1 (u, w) − T1 u, w
Ω1
shows that

1
P1 u1 = F in Ω1 , T1 (u1 ) = − Λ2 φ − W2 v on Γc .
2
Therefore together with (12.18) we get T1 (u1 ) = T2 (u2 ) on Γc .

Lemma 12.4
(i) ∃γ1 > 0 ∀φ ∈ H −1/2 (Γc ) φ, V2 φ γ1 φ2H −1/2 (Γ
c)
(ii) ∀v ∈ H 1/2(Γc ) v, W2 v 0
Proof Let v ∈ H 1/2(Γc ), φ ∈ H −1/2(Γc ) be given. From the definition of the

Calderon projector C2 we have

v 1 v
φ, V2 φ + v, W2 v = , − C2 (12.19)
φ 2 φ
Let Ω2c := R3 \ Ω 2 be the complementary domain of Ω2 . Define uc2 in Ω2c and u2

in Ω2 by the representation formula (12.16) with v replacing v2 and φ replacing φ2 .
Let
c c
v2 u2 v2 u2
:= and :=
φ2 T2 (u2 ) Γ φ2c T2 (uc2 ) Γ
c c
denote the respective Cauchy data. Then the jump relations yield
c
v2 v v
− 2c =
φ2 φ2 φ
and
c
v2 v v
+ 2c = 2 (C2 − 1/2) .
φ2 φ2 φ
Therefore we can rewrite (12.19) as follows,

c
1 v2c − v2 v2 + v2
φ, V2 φ + v, W2 v = ,
2 φ2c − φ2 φ2c + φ2
13 c 4
= v2 − v2 , φ2c + φ2 + v2c + v2 , φ2c − φ2
2
= v2c , φ2c − v2 , φ2
Next, we need the first Green formula for P2 in both Ω2 and Ω2c . Then
Φ2 (u2 , u2 ) + φ2 , v2 = 0
and
Φ2c (uc2 , uc2 ) − φ2c , v2c = 0
where

3
Φ2c (u, v) = 2
aihkl kl (u) ih (v)dx
2
, aihkl = λ2 δih δkl + μ2 (δik δhl + δil δhk )
Ω2c i,h,k,l=1
with δik = 1 for i = k and δik = 0 for i different from k. Hence
φ, V2 φ + v, W2 v = Φ2 (u2 , u2 ) + Φ2c (uc2 , uc2 ) (12.20)
Now, the right hand side is nonnegative, which proves the assertion (ii). Further-
more, Korn’s inequality shows that the right hand side satisfies a Gårding inequality
in the H 1 norm. This implies that V2 and W2 satisfy Gårding inequalities in the
H −1/2(Γc ) and H 1/2(Γc ) norms, respectively. Thus V2 is positive up to a compact
perturbation. It remains to show that φ, V2 φ > 0 holds for φ = 0. Thus let v = 0.
Then v2c = v2 which shows that u2 ∈ Hloc 1 (R3 ) if we define u = uc in Ω c .
2 2 2
Therefore (12.20) yields

3
φ, V2 φ = 2
ai,h,k,l kl (u2 ) ih (u2 )dx >0
i,h,k,l=1
R3
unless u2 is a rigid body motion, then implying φ = φ2 − φ2c = 0. Thus for φ =

0
there holds φ, V2 φ > 0.

As an application of Theorem 12.3 we obtain the following existence and
uniqueness result for (TMP) if we assume that P1 is strongly monotone (see [136]).
Theorem 12.6 Let F ∈ (L2 (Ω1 ))3 be given. Then there exists exactly one solution
(u, φ) ∈ HΓ11 (Ω1 ) × H −1/2(Γc ) of (12.14) yielding precisely one solution u1 in
L1 , u2 ∈ L2 of (TMP).
Next, we choose finite-dimensional subspaces XN , YN of X = HΓ11 (Ω1 ), Y =
H −1/2(Γc ) with
dX (w, XN ) → 0, dY (ψ, YN ) → 0 as N → ∞ ∀w ∈ HΓ11 (Ω1 ), ψ ∈ H −1/2 (Γc ).

The Galerkin scheme reads: Find (uN , φN ) ∈ XN × YN such that
b(uN , φN ; w, ψ) = l(w, ψ) ∀w ∈ XN , ψ ∈ YN (12.21)
Theorem 12.7 For all N ∈ N there exists exactly one solution (uN , φN ) ∈ XN ×
YN solution of (12.21).
Furthermore there exists C > 0 independent of N such that
u − uN H 1 (Ω1 ) + φ − φN H −1/2 (Γc ) C{dX (u, XN ) + dY (φ, YN )}
where u ∈ HΓ11 (Ω1 ) and φ) ∈ H −1/2 (Γc ) solve (12.14).

Proof Direct consequence of Theorem 12.4.

12.3 Symmetric FE/BE Coupling—Revisited
Firstly, we introduce the symmetric coupling method of Costabel from [113, 115]
and show its equivalence with the original transmission problem (IP) (Theo-
rem 12.8). Secondly we report from [91] an h-adaptive procedure for the symmetric
coupling of FEM and BEM. The theoretical results are obtained via the Poincaré-
Steklov operator. For an alternative approach via saddle points see [136] and the
foregoing Sect. 12.1 . An a posteriori error estimate is presented which guarantees
a given bound for the energy norm.
Let Ω1 := Ω ⊂ Rd , d ≥ 2 be a bounded domain with Lipschitz boundary
Γ = ∂Ω1 , and Ω2 := Rd \Ω̄1 with normal n on Γ pointing into Ω2 . In the case
d = 2 we always assume cap(Γ ) < 1 in the following. This can always be achieved
by scaling. For given f ∈ L2 (Ω1 ), u0 ∈ H 1/2(Γ ), t0 ∈ H −1/2(Γ ) we consider the
model interface problem (IP): Find u1 ∈ H 1 (Ω1 ), u2 ∈ Hloc1 (Ω ) such that
2
− div A(∇u1 ) = f in Ω1 (12.22)

Δu2 = 0 in Ω2 (12.23)
u1 = u2 + u0 on Γ (12.24)
∂u2
A(∇u1 ) · n = + t0 on Γ (12.25)
∂n

b ln |x| + o(1), d = 2
u2 (x) = , |x| → ∞ (12.26)
O(|x|2−d ), d≥3
12.3 Symmetric FEM/BEM Coupling 465
where b ∈ R is a constant (depending on u2 ). The operator A is assumed to be

uniformly monotone and Lipschitz continuous, i.e. there exist positive constants α
and C such that for all η, τ ∈ L2 (Ω)d

(A(η) − A(τ )) · (η − τ ) dx ≥ αη − τ 20,Ω (12.27)
Ω
A(η) − A(τ )0,Ω ≤ Cη − τ 0,Ω . (12.28)
Here .0,Ω denotes the norm in L2 (Ω)d . Examples for operators of this type can
be found in [399] and for models of nonlinear elasticity in Section 62 of [437]. The
definition of the Sobolev spaces is as usual:
H s (Ω) = {φ|Ω ; φ ∈ H s (Rd )} (s ∈ R),

⎧
⎨ {φ|Γ ; φ ∈ H s+1/2(Rd )} (s > 0)
H (Γ ) = L2 (Γ )
s
(s = 0)
⎩ −s
(H (Γ )) (dual space) (s < 0)
In the following we often write .s,B for the Sobolev norm · H s (B) with B =
Ω or Γ . We now derive the symmetric coupling method as discussed in detail in
[121]. Green’s formula together with the decaying condition (12.26) leads to the
representation formula for the solution in the exterior domain u2 of (12.23).
"
∂ ∂u2
u2 (x) = G(x, y)u2 (y) − G(x, y) dsy , x ∈ Ω2 (12.29)
Γ ∂n(y) ∂n(y)
with the fundamental solution of the Laplacian given by

%
− ω12 ln |x − y|, d = 2
G(x, y) = (12.30)
ωd |x − y|
1 2−d , d ≥ 3
where we have ω2 = 2π, ω3 = 4π.

By using the boundary integral operators

V ψ(x) := 2 G(x, y)ψ(y) dsy , x∈Γ (12.31)
Γ

∂
Kψ(x) := 2 G(x, y)ψ(y) dsy , x ∈ Γ (12.32)
Γ ∂n y

∂
K ψ(x) := 2 G(x, y)ψ(y) dsy , x ∈ Γ (12.33)
∂nx Γ

∂ ∂
W ψ(x) := −2 G(x, y)ψ(y) dsy , x ∈ Γ (12.34)
∂nx Γ ∂ny
together with their jump conditions (see Chap. 2) we obtain from (12.29) the
following integral equations
∂u2 ∂u2
2 = −W u2 + (I − K ) (12.35)
∂n ∂n
∂u2
0 = (I − K)u2 + V . (12.36)
∂n
One observes that (12.36) gives one part of the weak formulation of problem (IP)
∂u2
−u1 , ψ − V , ψ + Ku1 , ψ = −u0 , ψ + Ku0 , ψ ∀ψ ∈ H −1/2 (Γ ).
∂n
(12.37)
The second part of the weak formulation has to couple the exterior problem (12.23)
and the interior problem (12.22). To this end we use

a(u1 , v) := A(∇u1 ) · ∇v dx = (A(∇u1 ) · n)v ds + f v dx ∀v ∈ H 1 (Ω1 ).
Ω1 Γ Ω1
(12.38)
Taking the integral equation (12.35) and substituting (12.24) and (12.25)
into (12.38) one obtains for all v ∈ H 1 (Ω1 )
∂u2 ∂u2
2a(u1, v) − , v + K , v + W u1 , v = 2(f, v) + 2t0 , v + W u0 , v ,
∂n ∂n
(12.39)
where (f, v) = Ω1 f v dx.

Note that in this way we obtain the following variational formulation (P=):
Given (f, u0 , t0 ) ∈ L2 (Ω) × H 1/2(Γ ) × H −1/2 (Γ ) find u := u1 ∈ H 1 (Ω1 ) and
φ := ∂u −1/2 (Γ ) such that for all v ∈ H 1 (Ω) and ψ ∈ H −1/2 (Γ )
∂n ∈ H
2
2a(u, v) + (K − I )φ, v + W u, v = 2t0 , v + W u0 , v + 2(f, v)

(K − I )u, ψ − V φ, ψ = (K − I )u0 , ψ .
(12.40)
There holds the following equivalence:

Theorem 12.8 The problems (IP) and (P=) are equivalent in the following sense. If
1 (Ω ) is a solution of (IP) then (u, φ) ∈ H 1 (Ω )×H −1/2(Γ )
(u, v) ∈ H 1 (Ω1 )×Hloc 2 1
solves (P=) with φ := ∂n
∂v
|Γ . If, conversely, (u, φ) is a solution of problem (P=) then
(u, v) solves (IP) with v ∈ Hloc

1 (Ω ) defined by
2

1
v(z) = − φ(ζ ) · G(z, ζ ) dsζ (12.41)
2π Γ

1 ∂
+ (u − u0 )(ζ ) · G(z, ζ ) dsζ (z ∈ Ω2 ).
2π Γ ∂nζ
Proof By deriving the coupling formulation we have already shown that if (u, v)
solves (I P ) then (u, φ) solves (P=).
Conversely, let (u, φ) solve (P=) and define v by (12.41). Then, according to
[129], v satisfies (12.23), (12.26) and with the jump relations we have

v|Γ 1 u|Γ − u0 −K V
= (I d − H ) with H := . (12.42)
∂n |Γ
∂v
2 φ W K
The first component of (12.42) together with (12.37) yields u|Γ = v|Γ + u0 . From
the second identity in (12.42) we then have
∂v 1
|Γ = − {W (u|Γ − u0 ) + (K − 1)V φ}.
∂n 2
Using this in (12.38) gives, by Green’s formula

∂v
(div(Agrad u) + f )η dΩ = (Agrad u) · n|Γ − |Γ − t0 , η|Γ
Ω ∂n
for all η ∈ H 1 (Ω). Choosing η ∈ H01 (Ω), the completion of C0∞ (Ω) in the H 1 -
norm, we get the weak form of (12.22). Hence using (12.22) we get (12.25).

Remark 12.1 We note that
W 1 = 0 and K1 = −1 (12.43)
with 1 being
the
constant function with the value one. The identities (12.43) follow
from H 10 = 10 (cf. [129, Lemma 3.5]).
For the Galerkin scheme we choose finite dimensional subspaces XM ⊂ H 1 (Ω)
and YN ⊂ H −1/2 (Γ ) and define the Galerkin solution (uM , φN ) ∈ XM × YN by
2a(uM , v) + (K − I )φN , v + W uM , v = 2t0 , v + W u0 , v + 2(f, v)

(K − I )uM , ψ − V φN , ψ = (K − I )u0 , ψ
(12.44)
for all v ∈ XM and ψ ∈ YN .

There holds the following convergence result as application of Theorem 12.4. In

the next section we will present a different proof via the Poincare-Steklov operator,
cf. Corollary 12.1.
Theorem 12.9 ([136]) Every Galerkin scheme (12.44) with approximating finite
dimensional spaces XM ⊂ H 1 (Ω) and YN ⊂ H −1/2(Γ ) converges with optimal
order, i.e. with the exact solution (u, φ) of (12.40) and the Galerkin solution
(uM , φN ) of (12.44) there holds the estimate
u − uM 1,Ω + φ − φN −1/2,Γ ≤ C { inf u − û1,Ω + inf φ − φ̂−1/2,Γ }

û∈XM φ̂∈YN
(12.45)
where the constant C is independent of M, N, u and φ.
12.3.1 Convergence Analysis
In this section we prove existence and uniqueness of the weak (variational) solution
of the interface problem (IP) in Sect. 12.3 and show convergence of the Galerkin
solution proving Theorem 12.9 above, now following [91] and using heavily the
strong coerciveness of the Poincaré-Steklov operator (for the exterior problem) and
of its discrete analogue.
Firstly, we note that the weak formulation (12.40) is
Problem (P): Find (u, φ) ∈ H 1 (Ω1 ) × H −1/2(Γ ) with

u v v
B( , )=L ∀(v, ψ) ∈ H 1 (Ω1 ) × H −1/2(Γ ). (12.46)
φ ψ ψ
Here the continuous mapping B : (H 1(Ω) × H −1/2(Γ ))2 → R and the linear form
L : H 1 (Ω) × H −1/2 (Γ ) → R are defined by

u v 1
B( , ) := A(∇u) · ∇v dx + W u|Γ + (K − I )φ, v|Γ
φ ψ Ω1 2
1
+ ψ, V φ + (I − K)u|Γ (12.47)
2

v 1 1
L := f · v dx + ψ, (I − K)u0 + t0 + W u0 , v|Γ (12.48)
ψ Ω1 2 2
for any (u, φ), (v, ψ) ∈ H 1 (Ω1 ) × H −1/2(Γ ).

Note that (12.36) is equivalent to
φ = −V −1 (I − K)(u1 − u0 ) (12.49)
which we use to eliminate φ = ∂u 2

∂n in (12.40). Thus we arrive at the following
equivalent formulation: Find u ∈ H (Ω1 ) with
1

A (u)(η) := 2 A(∇u) · ∇η dx + Su|Γ , η|Γ (12.50)
Ω1

= L (η) := 2 f v dx + 2t0 + Su0 , η|Γ (η ∈ H 1 (Ω))
Ω1
with the Poincaré-Steklov operator for the exterior problem
S := W + (I − K )V −1 (I − K) : H 1/2 (Γ ) −→ H −1/2(Γ ) (12.51)
Lemma 12.5 ([91]) (Suppose cap(Γ ) < 1 in 2D). The operator S := W + (1 −

K )V −1 (1 − K) : H 1/2(Γ ) → H −1/2 (Γ ) is linear, bounded, symmetric, and
positive definite.
Proof Due to the above mentioned properties of W, K, V , K , the operator S is
linear, bounded, symmetric, positive semidefinite, and a Fredholm operator of index
zero. Thus, it suffices to prove that the kernel ker S is trivial in order to conclude
that S is positive definite. Let u ∈ ker S, then 0 = Su, u . On the other hand
Su, u ≥ W u, u ≥ 0, so that W u, u = 0. By Theorem 2.5, u is constant.
Therefore 0 = V −1 (1 − K)u, (1 − K)u . By Theorem 2.5, V −1 is positive definite
so that (1 − K)u = 0. Using (12.43), this implies that the constant u is equal to
zero. Thus, ker S = {0}. The arguments can be extended to cover also the 3D case
making use of Lemma 12.4.

Lemma 12.6 There exists a constant β > 0 such that for all (u, φ), (v, ψ) ∈
H 1 (Ω) ×H −1/2(Γ ) we have

u−v u−v
β · H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2
φ−ψ η − δ H (Ω)×H (Γ )

u u−v v u−v
≤B , −B ,
φ η−δ ψ η−δ
(12.52)
with 2η := φ + V −1 (I − K)u|Γ , 2δ := ψ + V −1 (I − K)v|Γ ∈ H −1/2(Γ ).

Proof Some calculations show

u u−v v u−v
B( , ) − B( , )
φ η−δ ψ η−δ

= (A∇u) − (A∇v) · ∇(u − v) dx
Ω
1 1 1
+ W (u − v), u − v + S(u − v), u − v + V (φ − ψ), φ − ψ .
4 4 4
Since A is uniformly monotone, W is positive semi-definite, S and V

are positive definite, the right hand side in (12.52) is bounded below by
u−v 2
c̃ 1 −1/2 with a suitable constant c̃.
φ − ψ H (Ω)×H (Γ )
On the other hand, by definition of η, δ, we have with a constant c

u−v
η − δH −1/2 (Γ ) ≤ c 1 −1/2 ,
φ − ψ H (Ω)×H (Γ )
yielding (12.52).

Theorem 12.10 The interface problem (IP) and the problem (P) have unique
solutions.
Proof The operator A on the left hand side in (12.50) maps H 1 (Ω1 ) into its dual; it
is continuous, bounded, uniformly monotone and therefore bijective. This yields the
existence of u satisfying (12.50). Letting φ as in (12.49) we have that (u, φ) solves
problem (P). Uniqueness of the solution follows from Lemma 12.6 yielding also the
unique solvability of the equivalent interface problem (IP).

Next we treat the discretization of problem (P) in the 2D case.
−1/2
Let (Hh × Hh : h ∈ I ) for I ⊆ (0, 1) with 0 ∈ I¯ be a family of finite
dimensional subspaces of H 1 (Ω)×H −1/2 (Γ ). Then, the coupling of finite elements
and boundary elements consists in the following Galerkin procedure.
−1/2
Definition 12.1 (Problem (Ph )) For h ∈ I find (uh , φh ) ∈ Hh × Hh such that

uh vh vh
B( , )=L (12.53)
φh ψh ψh
−1/2
for all (vh , ψh ) ∈ Hh × Hh .
In order to prove a discrete Babuška–Brezzi condition if A is linear, we
need some notations and the positive definiteness of the discrete Poincaré-Steklov
operator.
−1/2
Assumption 12.1 For any h ∈ I let Hh × Hh ⊆ H 1 (Ω) × H −1/2(Γ ). Suppose
−1/2
1 ∈ Hh for any h ∈ I , where 1 denotes the constant function with value 1.
−1/2
Let ih : Hh *→ H 1 (Ω) and jh : Hh (Γ ) *→ H −1/2(Γ ) denote the canonical
−1/2
injections with their duals ih∗ : H 1 (Ω)∗ → Hh∗ and jh∗ : H 1/2(Γ ) → Hh )(Γ )∗
being projections. Let γ : H 1 (Ω) → H 1/2(Γ ) denote the trace operator, γ u = u|Γ
for all u ∈ H 1 (Ω), with the dual γ ∗ . Then, define
Vh := jh∗ Vjh , Kh := jh∗ Kγ ih , Wh := ih∗ γ ∗ W γ ih , Kh := ih∗ γ ∗ K jh

(12.54)
and, since Vh is positive definite as well as its continuous analogue V ,
Sh := Wh + (Ih∗ − Kh )Vh−1 (Ih − Kh ) : Hh → Hh∗ (12.55)
with Ih := jh∗ γ ih and its dual Ih∗ .

A key role is played by the following coerciveness of the discrete version of the
Poincaré-Steklov operator, due to [91].
Lemma 12.7 There exist constants c0 > 0 and h0 > 0 such that for any h ∈ I with
h < h0 we have
Sh uh , uh ≥ c0 · uh |Γ 2H 1/2 (Γ ) for all uh ∈ Hh .
Proof Assume that the assertion is false. Then one can construct a sequence of
functions (uhn )n=1,2,3,... in H 1 (Ω) with
1
uhn ∈ Hhn , uhn |Γ H 1/2 (Γ ) = 1, Shn uhn , uhn ≤ (n = 1, 2, 3, . . .)
n
and limn→∞ hn = 0. Due to the Banach–Alaoglu theorem (uhn |Γ )n=1,2,3,...
converges weakly towards some w ∈ H 1/2(Γ ) in H 1/2(Γ ) (a subsequence at least).
Then, by definition of Sh ,first we conclude that W uhn |Γ , uhn |Γ tends towards
zero so that W w, w = 0, i.e. w|Γ is constant. A decomposition of uhn |Γ =
1/2
vn + wn with vn ∈ H0 (Γ ) = {v ∈ H 1/2(Γ ), v, 1 = 0} and wn ∈ R
shows additionally that (vn )n=1,2,3,... tends towards zero strongly in H 1/2(Γ ), since
1/2
W is positive definite on H0 (Γ ). Hence we have also strong convergence of
(uhn |Γ )n=1,2,3,... towards the constant w ∈ R in H 1/2(Γ ).
−1/2
On the other hand we have 0 = limn→∞ V zn , zn with zn := Vh−1 n
(φn ) ∈ Hhn ⊆
−1/2
H −1/2(Γ ), φn := jh∗n yn ∈ (Hhn )∗ , yn := uhn − Kuhn ∈ H 1/2(Γ ).
Thus, 0 = limn→∞ zn H −1/2 (Γ ) whence 0 = limn→∞ φn (H −1/2 )∗ .
hn
Because of (uhn |Γ )n=1,2,3,... → w we get (yn )n=1,2,3,... → 2w (strongly) in
H 1/2(Γ ) (by (12.43) and w ∈ R). Hence,
2w1, 1 = lim 1, yn = lim jhn 1, yn = lim 1, φn = 0,

n→∞ n→∞ n→∞
i.e. w = 0. This contradicts wH 1/2 (Γ ) = limn→∞ uhn |Γ H 1/2 (Γ ) = 1.

In the above Lemma it is assumed that the initial boundary mesh is sufficiently fine.
This assumption has first been proved to be unnecessary in [13], where the original
argument of the above proof is refined.
Lemma 12.8 There exist constants β0 > 0 and h0 > 0 such that for any h ∈ I
−1/2
with h < h0 we have that for any (uh , φh ), (vh , ψh ) ∈ Hh × Hh

uh − vh uh − vh
β0 H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2 (12.56)
φh − ψh ηh − δh H (Ω)×H (Γ )

uh uh − vh vh uh − vh
≤ B( , ) − B( , ) (12.57)
φh ηh − δh ψh ηh − δh
−1/2
with 2ηh := φh + Vh−1 (Ih − Kh )uh , 2δh := ψh + Vh−1 (Ih − Kh )vh ∈ Hh .
Proof The proof is analogous to that of Lemma 12.6.Due to Lemma 12.7 the
constants are independent of h as well so that β0 does not depend on h < h0 ,
h0 chosen in Lemma 12.7. This concludes the proof.

Corollary 12.1 There exist constants c0 > 0 and h0 > 0 such that for any h ∈ I
with h < h0 the problem (Ph ) has a unique solution (uh , φh ) and with the solution
(u, φ) of (P), there holds

u − uh u − vh
H 1 (Ω)×H −1/2 (Γ ) ≤ c0 · inf H 1 (Ω)×H −1/2 (Γ ) .
φ − φh ( vh )∈Hh ×H −1/2 φ − ψh
ψh h
Proof The existence and uniqueness of the discrete solutions follows as in the proof
−1/2
of Theorem 12.10. Let (Uh , Φh ) ∈ H h × Hh be the orthogonal projections onto
−1/2
H × Hh
h of the solution (u, φ) of Problem (P) in H 1 (Ω) × H −1/2(Γ ). From
−1/2
Lemma 12.8 we conclude with appropriate (ηh , δh ) ∈ H h × Hh that

Uh − u h Uh − u h
β0 · H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2 (Γ )
Φh − φh ηh − δh H (Ω)×H

Uh Uh − u h uh Uh − u h
≤ B( , ) − B( , ).
Φh ηh − δh φh ηh − δh
Using the Galerkin equations and the Lipschitz continuity of B with constant L, the
right hand side is bounded by

Uh − u h Uh − u
L· H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2 (Γ ) ,
ηh − δh Φh − φ H (Ω)×H
what gives the assertion.

12.3.2 Adaptive FE/BE Coupling: Residual Based Error

Indicators
In this section we present a posteriori error estimates for the h-version of the
symmetric coupling method from [91].
For simplicity, we restrict ourselves to linear ansatz functions on triangles as
−1/2
finite elements in Hh and to piecewise constant functions in Hh .
Assumption 12.2 Let Ω be a two-dimensional domain with polygonal boundary
Γ on which we consider a family T := (Th : h ∈ I ) of decompositions Th =
{Δ1 , . . . , ΔN } of Ω in closed triangles Δ1 , . . . , ΔN such that Ω̄ = ∪N
i=1 Δi , and
two different triangles are disjoint or have a side or a vertex in common. Let Sh
denote the sides, i.e. Sh = {∂Ti ∩ ∂Tj : i = j with ∂Ti ∩ ∂Tj is a common side },
∂Tj being the boundary of Tj . Let Gh = {E : E ∈ Sh with E ⊆ Γ } be the set of
“boundary sides” and let Sh0 = Sh \ Gh be the set of “interior sides”. We assume
that all the angles of some Δ ∈ Th ∈ T are ≥ Θ for some fixed Θ > 0 which does
not depend on Δ or Th .
Then, define
Hh := {ηh ∈ C(Ω) : ηh |Δ ∈ P1 for any Δ ∈ Th } (12.58)

−1/2
Hh := {ηh ∈ L∞ (Γ ) : ηh |E ∈ P0 for any E ∈ Gh } (12.59)
where Pj denotes the polynomials with degree ≤ j .

For fixed Th let h be the piecewise constant function defined such that the
constants h|Δ and h|E equal the element sizes diam(Δ) of Δ ∈ Th and diam(E)
of E ∈ Sh . We assume that A(∇vh ) ∈ C 1 (Δ) for any Δ ∈ Th ∈ T and any trial
function vh ∈ Hh . Finally, let f ∈ L2 (Ω), u0 ∈ H 1 (Γ ), and t0 ∈ L2 (Γ ).
Let n be the exterior normal on Γ and on any element boundary ∂Δ, let n have
a fixed orientation so that [A(∇uh ) · n]|E ∈ L2 (E) denotes the jump of the discrete
tractions A(∇uh ) · n over the side E ∈ Sh0 . Define

R12 := diam(Δ)2 · |f + div A(∇uh )|2 dx (12.60)
Δ∈Th Δ

R22 := diam(E) · |[A(∇uh ) · n]|2 ds (12.61)
E
E∈Sh0
√ 1 1
R3 := h · t0 − A(∇uh ) · n + W (u0 − uh |Γ ) − (K − I )φh L2 (Γ )
2 2
(12.62)
∂
R4 := diam(E)1/2 · {(I − K)(u0 − uh |Γ ) − V φh }L2 (E) . (12.63)
∂s
E∈Gh
Under the above assumptions there holds the following a posteriori estimate where
(u, φ) and (uh , φh ) solve problem (P ) and (Ph ).
Theorem 12.11 There exists some constant c > 0 such that for any h ∈ I with
h < h0 (h0 from Lemma 12.7) we have

u − uh
1 −1/2 ≤ c · (R1 + R2 + R3 + R4 ).
φ − φh . H (Ω)×H (Γ )
The proof of Theorem 12.11 is divided into several lemmas. We set
1
e := u − uh , := φ − φh , δ := ( + V −1 (1 − K)e|Γ ).
2
Lemma 12.9 We have

e e
β· H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2 ≤ T1 + T2 + T3 + T4
δ H (Ω)×H (Γ )
−1/2
where, for any (eh , δh ) ∈ Hh × Hh ,

T1 := (f + div(A grad uh ))(e − eh ) dΩ
Δ∈Th Δ

T2 := − [(A grad uh ) · n](e − eh )|E ds
E
E∈Sh0
1
T3 := t0 − (A grad uh ) · n + W (u0 − uh |Γ )
2
1
− (K − 1)φh , (e − eh )|Γ
2
1
T4 := δ − δh , (1 − K)(u0 − uh |Γ ) − V φh .
2
Proof Due to the arguments of the proof of Lemma 12.6 we have

e e
β· H 1 (Ω)×H −1/2 (Γ ) · 1 −1/2
δ H (Ω)×H (Γ )

u e uh e
≤ B( , ) − B( , )
φ δ φh δ

e − eh uh e − eh
=L − B( , )
δ − δh φh δ − δh
using (12.53) and (12.46). By definition of B and L, the last term equals

(f (e − eh ) − A grad uh grad (e − eh )) dΩ
Ω
1 1
+ t0 + W (u0 − uh |Γ ) − (K − 1)φh , (e − eh )|Γ
2 2
1
+ δ − δh , (1 − K)(u0 − uh |Γ ) − V φh .
2
Using Green’s formula on all elements Δ ∈ Th we obtain

− A grad uh grad (e − eh ) dΩ
Ω

= div(A grad uh )(e − eh ) dΩ
Δ∈Th Δ

− [(A grad uh ) · n](e − eh )|E ds
E
E∈Sh0
− (A grad uh ) · n, (e − eh )|Γ .
yielding the assertion.

We note that under the Assumption 12.2 Clement interpolation can be applied and
gives the following lemma where c > 0 is a generic constant and depends only on
T but not on h, Δ, N, u, etc.
Lemma 12.10 There exists a family of interpolation operators (Ih : H 1 (Ω) →
Hh : h ∈ I ) and a constant c > 0 such that the following holds. For any Δ ∈ Th ∈
T and integers k, q with 0 ≤ k ≤ q ≤ 2 and with N := ∪{Δ ∈ Th : Δ ∩ Δ = ∅},
the union of all neighbor elements of Δ, and for all u ∈ H q (N),
|Ih u − u|2H k (Δ) ≤ c · diam(T )2(q−k) · |u|2H q (N) .
Furthermore, choosing eh := Ih e we have Ti ≤ c · |e|H 1 (Ω) · Ri , i = 1, 2, 3.

Additionally, there exists a constant c > 0 such that for any E, E is one side of
Δ ∈ Th ∈ T , and any u ∈ H 1 (Δ) there holds
diam(Δ)u2L2 (E) ≤ c · u2L2 (Δ) + diam(Δ)2 · |u|2H 1 (Δ) .
Lemma 12.11 For ψ := (1 − K)(u0 − uh |Γ ) − V φh we have with a constant c

√
ψH 1/2 (Γ ) ≤ c · h · ψ L2 (Γ ) .
Proof (of Theorem 12.11) The assertion follows from Lemmas 12.10, and 12.11
to estimate T1 ,T2 , T3 , and T4 (with δh = 0) in Lemma 12.9, respectively. Then,
division by eδ H 1 (Ω)×H −1/2 (Γ ) proves the theorem.

12.3.2.1 Adaptive Feedback Procedure
For a given triangulation Th = {Δ1 , . . . , ΔN } of Ω and the related partition

{Γ1 , . . . , ΓM } = Gh of the boundary Γ we can consider one element Δj ∈ Th
and compute its contributions aj , bk to the right hand side of the a posteriori error
estimate in Theorem 12.11

aj2 := diam(Δj )2 · |f + div(A grad uh )|2 dΩ
Δj

+ diam(E) · |[(A grad uh ) · n]E |2 ds
E
E∈Sh0 ,E⊆∂Δj
1
+ diam(Γ ∩ ∂Δj ) · t0 − (A grad uh ) · n + W (u0 − uh |Γ )
2
1
− (K − 1)φh 2L2 (Γ ∩∂Δ )
2 j
∂
bk := diam(Γk )1/2 · {(1 − K)(u0 − uh |Γ ) − V φh }L2 (Γk ) .
∂s
If we neglect the constant c > 0 in Theorem 12.11, the error in the energy norm is
bounded by
N
O
ON
M
O
P aj2 + bk . (12.64)
j =1 k=1
Note that the different nature of the coefficients aj and bk is, in general, caused
by two different discretizations: aj is related to a finite element, bk is related
to a boundary element. Because of a simple storage organization and a simple
computation of the stiffness matrices, it is convenient to use only one mesh, i.e. to
take the boundary element discretization induced by the finite element triangulation.
Therefore, we consider this case in the sequel. For any element Δj let

N
cj := aj + bk
k=1,Γk ⊆Δj
where the sum may be zero or consists of one or two summands.

The meshes in our numerical examples are steered by the following algorithm
where 0 ≤ θ ≤ 1 is a global parameter:
Algorithm 12.1 ((A)) Given some coarse e.g. uniform mesh refine it successively
by halving some of the elements due to the following rule. For any triangulation
define a1 , . . . , aN as above and divide some element Γj by halving the largest side if
cj ≥ θ · max ck .
k=1,...,N
In a subsequent step all hanging nodes are avoided by further refinement in order
to obtain a regular mesh.
Remark 12.2
(i) Note that in Algorithm (A) θ = 0 gives a uniform refinement and with
increasing θ the number of refined elements in the present step decreases.
(ii) By observing (12.64) we have some error control which, in some sense, yields
a reliable algorithm. In particular, the relative improvement of (12.64) may be
used as a reasonable termination criterion.
(iii) If in some step of Algorithm (A), (12.64) does not become smaller then we
may add some uniform refinement steps (θ = 0). It can be proved that in this
case (12.64) decreases and tends towards zero. If we allow this modification
we get convergence of the adaptive algorithm.
In [91] we consider (IP) with p = 1, f = 0, Ω the L-shape region with vertices
(0, 0), (1, 0), (1, 1), (−1, 1), (−1, −1), (0, −1) and take
2 1 1 1
u = r 2/3 · sin( α) and v = ln((x + )2 + (y − )2 )
3 2 2 2
such that u0 , t0 are given by (12.24), (12.25).
In Table 12.1 we have the numerical results for the uniform mesh (θ = 0)
and for the meshes generated by Algorithm (A) for θ = 0.2, 0.4, and 0.6. Here,
we show only the number of degrees of freedom N for the finite element method
(chosen by the algorithm; a new row corresponds to a new refinement step in the
adaptive algorithm), and the corresponding relative error of the displacements eN in
the H 1 (Ω)-norm.
Let γN be the error in energy norm divided by (12.64). Hence, by Theorem 12.11,
γN is bounded which can be observed from Table 12.1 Moreover, γN is bounded
below which indicates efficiency of the estimate and hence of the adaptive scheme.
For further experiments see [91].
Table 12.1 Numerical results for the linear transmission problem [91]
Uniform mesh (A) for θ = 0.4 (A) for θ = 0.6
N eN γN N eN γN N eN γN
8 0.20434 .152 8 0.20434 .152 8 0.20434 .152
11 0.18587 .173 11 0.18587 .173 10 0.20467 .173
21 0.14485 .164 15 0.17074 .176 13 0.17286 .176
33 0.12564 .185 21 0.14520 .182 17 0.14848 .185
65 0.09563 .149 26 0.12197 .188 21 0.13954 .193
113 0.08027 .159 31 0.11007 .201 26 0.11594 .196
225 0.06230 .148 40 0.09420 .168 33 0.10579 .209
(A) for θ = 0.2 48 0.08544 .177 38 0.09402 .214
N eN γN 55 0.07824 .180 50 0.08328 .181
8 0.20434 .152 71 0.06837 .182 55 0.07744 .181
11 0.18587 .173 80 0.06260 .184 69 0.06742 .183
19 0.14621 .163 101 0.05633 .187 78 0.06448 .185
27 0.12844 .182 134 0.04959 .187 97 0.05639 .189
41 0.10297 .155 157 0.04510 .184 08 0.05448 .189
52 0.09020 .166 201 0.03904 .183 49 0.04533 .189
66 0.07554 .162 226 0.03656 .184 64 0.04367 .185
75 0.06900 .172 211 0.03783 .184
102 0.05947 .174 239 0.03562 .185
134 0.05128 .176
156 0.04646 .175
201 0.04004 .177
235 0.03604 .177
12.3.3 Adaptive FE/BE Coupling with a Schur Complement

Error Indicator
Recently, the use of adaptive hierarchical methods has becoming increasingly

popular. Using the discretization of the Poincaré-Steklov operator we present from
[274] for the symmetric FE/BE coupling method an a posteriori error estimate
with ’local’ error indicators; for an alternative method which uses the full coupling
formulation see [312]. By using stable hierarchical basis decompositions for finite
elements we have two-level subspace decompositions for locally refined meshes.
Assuming a saturation condition to hold an adaptive algorithm is formulated to
compute the finite element solution on a sequence of refined meshes in the interior
domain and on the interface boundary. At the end of this subsection we present
numerical experiments which show efficiency and reliability of the error indicators.
Let ρ ∈ C 1 (R+ ) satisfy the conditions
ρ0 ≤ ρ(t) ≤ ρ1 and ρ2 ≤ ρ(t) + tρ (t) ≤ ρ3 (12.65)

for some global constants ρ0 , ρ1 , ρ2 , ρ3 > 0. We consider the following nonlinear

interface problem (NP) (cf. [274]) in R2 :
Problem (NP): Given the functions f : Ω1 → R and u0 , t0 : Γ → R find
ui : Ωi → R, i = 1, 2, and b ∈ R such that
− div(ρ(|∇u1 |) ∇u1 ) = f in Ω1 (12.66a)
− Δu2 = 0 in Ω2 (12.66b)
u1 − u2 = u0 on Γ (12.66c)
∂u1 ∂u2
ρ(|∇u1 |) − = t0 on Γ (12.66d)
∂n ∂n
u2 (x) = b ln |x| + o(1) for |x| → ∞ (12.66e)

∂v
where ∂n is the normal derivative of v pointing from Ω1 into Ω2 .
By the symmetric coupling method the problem (12.66) is transformed into the
following variational problem (cf. [91]):
Given f ∈ (H 1 (Ω1 )) , u0 ∈ H 1/2(Γ ) and t0 ∈ H −1/2(Γ ) find u ∈ H 1 (Ω1 )
and φ ∈ H −1/2(Γ ) such that
a(u, v) + B(u, φ; v, ψ) = L (v, ψ) (12.67)
for all v ∈ H 1 (Ω1 ) and ψ ∈ H −1/2 (Γ ) where the form a(·; ·) is defined as

a(u, v) := 2 ρ(|∇u|) ∇u · ∇v dx ,
Ω1
the bilinear form B(·; ·) is defined as
B(u, φ; v, ψ) := W u|Γ + (K − I )φ , v|Γ − ψ , (K − I )u|Γ − V φ ,
and the linear form L (·) is defined as
L (v, ψ) := 2(f, v) + 2t0 + W u0 , v|Γ − ψ , (K − I )u0 .
Here, (·, ·) and ·, · denote the duality pairings between (H 1 (Ω1 )) and H 1 (Ω1 )
and between H −1/2(Γ ) and H 1/2(Γ ), respectively. The unknowns in (12.66) satisfy
u1 = u and ∂u ∂n = φ and u2 can be obtained via a representation formula (see the
2
foregoing section).
Lemma 12.12 The following problem is equivalent to (12.67):

Find u ∈ H 1 (Ω1 ) such that
a(u, v) + Su|Γ , v = F (v) ∀v ∈ H 1 (Ω1 )

(12.68)
where F (v) := 2 f vdx + 2t0 + Su0 , v|Γ ,
Ω1
a(·, ·) as in (12.67), and the Poincaré-Steklov operator for the exterior domain
represented by S := W + (K − I )V −1 (K − I ) is a continuous map from H 1/2(Γ )
into H −1/2(Γ ) and coercive on H 1/2(Γ ) for cap(Γ ) < 1.
Next, we describe the coupling of the finite element method (FEM) and the
boundary element method (BEM) to compute approximations to the solution (u, φ)
of (12.67). We consider regular triangulations ωH of Ω1 and partitions γH of Γ .
Our test and trial spaces are defined as
TH := {vH : Ω1 → R ; vH p.w. linear on ωH , vH ∈ C 0 (Ω1 )} , (12.69)

τH := {ψH : Γ → R ; ψH p.w. constant on γH } . (12.70)
For simplicity, we assume that the mesh for the discretization of the boundary
element part γH is induced by that of the finite element part. This yields the
following discretization of problem (12.67):
Find (uH , φH ) ∈ TH × τH such that

2 ρ(|∇uH |) ∇uH · ∇v dx + B(uH , φH ; v, ψ) = L (v, ψ) (12.71)
Ω1
for all (v, ψ) ∈ TH × τH .

Application of Newton’s method to (12.71) yields a sequence of linear systems
(0) (0)
to be solved. Given an initial guess (uH , φH ) we compute
(u(l) (l) (l−1)

H , φH ) = (uH
(l−1)
, φH (l) (l)
) + (dH , δH ) (l = 1, 2, . . . )
such that
(l) (l) (l) (l−1) (l−1) (l−1)
au(l−1) (dH , v) + B(dH , δH ; v, ψ) = L (v, ψ) − a(uH , v) − B(uH , φH ; v, ψ)
H
(12.72)
for all (v, ψ) ∈ TH × τH with a(·, ·), B(·, · ; ·, ·), and L (·, ·) as in (12.67). The
bilinear form aw (·, ·) is defined by

aw (u, v) := 2 ρ̃(∇w) ∇u · ∇v dx , (12.73)
Ω1
and ρ̃ ∈ R2 is the Jacobian of x → ρ(|x|)x, i.e.
x · xT
ρ̃ = ρ(|x|)I2×2 + ρ (|x|) (x ∈ R2 ). (12.74)
|x|
From the assumptions on ρ in (12.65) it follows that there exist constants ν, μ > 0
such that
aw (u, v) ≤ ν uH 1 (Ω1 ) vH 1 (Ω1 ) and μ u2H 1 (Ω ) ≤ aw (u, u) (12.75)

1
for all w, u, v ∈ H 1 (Ω1 ).

By the assumptions on ρ in (12.65) the energy functional of (12.67) is strictly
convex, and hence, Newton’s method converges locally.
For the implementation of (12.72) we define the piecewise linear basis functions
bi via
bi (νj ) := δi,j (1 ≤ j ≤ nin , nin + 1 ≤ j ≤ nT )
where νi ∈ Ω1 \Γ (1 ≤ i ≤ nin ) are the inner nodes of ωH and νi ∈ Γ

(nin + 1 ≤ i ≤ nH := dim TH ) are the boundary nodes of ωH counted along the
closed curve Γ .
The above result obviously holds for subspaces of Th . That means that ωh does
not necessarily has to be a uniform refinement of ωH . Thus (12.77) holds also for
locally refined meshes.On the boundary the following basis of τH is introduced: Let
μi ∈ γH be the boundary element induced by the nodes νnin +i , νnin +i+1 (1 ≤ i ≤
nτ − 1, nτ := dim τH ) and μnτ by the nodes νnT , νnin +1 . With each μi we associate
the basis function
%
1 if x ∈ μi
βi (x) := .
0 if x ∈ Γ \μi
With the basis functions bi and βi (12.72) yields a linear system which may be
solved with the hybrid modified conjugate residual (HMCR) scheme together with
efficient preconditioners [236].
In [312] an adaptive algorithm is given based on a posteriori error estimates of
the solution (uH , φH ) of (12.71). Here we apply a Schur complement method based
on a Galerkin discretization of the variational formulation (12.68) eliminating the
unknown vector φ. In this way we also obtain a discretization of the Poincaré-
Steklov operator which will be used to develop an a posteriori error indicator
which needs only a refinement of the mesh defining TH and does not need a finer
discretization as τH .
Next, we introduce hierarchical two-level decompositions for the finite element
space Th on ωh (cf. (12.69)) where we get ωh by the refinement shown in Fig. 12.2.
δ
s
Fig. 12.2 Refinement of δ ∈ ωH . The longest edge of δ is denoted by s. The new nodes are the
midpoints of the edges of δ [312]
These decompositions will be used to derive an a posteriori error estimate for

the Galerkin solution to (12.67) which is obtained by applying a Schur complement
to (12.71).
We take the hierarchical two-level subspace decomposition
Th := TH ⊕ Lh , Lh := T1 ⊕ T2 ⊕ . . . ⊕ Tn
with Ti := span{b̂i } where b̂i denote the piecewise linear basis functions in the new
n node-points νi of the fine grid, [312, 436]. Let PH : Th −→ TH , Pi : Th −→ Ti be
the Galerkin projections with respect to the bilinear form b(·, ·) which is defined as

b(u, v) := (∇u · ∇v + uv) dx . (12.76)
Ω1
For all u ∈ Th we define PH and Pi by
b(PH u, v) = b(u, v) ∀v ∈ TH
b(Pi u, v) = b(u, v) ∀v ∈ Ti .
The following result states that the two-level additive Schwarz operator
P := PH + m i=1 Pi has bounded condition number: There are constants
c1 , c2 > 0 which depend only on the smallest angle of the triangles in ωH and
on the diameter of Ω1 such that

m
c1 v2H 1 (Ω ) ≤ PH v2H 1 (Ω ) + Pi v2H 1 (Ω ) ≤ c2 v2H 1 (Ω ∀v ∈ Th .
1 1 1 1)
i=1
(12.77)
Now, we introduce the approximate Poincaré-Steklov operator on fine mesh

functions
∗
S̃h := Wh + (KH,h ∗
− IH,h )VH−1 (Kh,H − Ih,H ) (12.78)
where for u, v ∈ Th and φ, ψ ∈ τH
Wh u|Γ , v|Γ = W u|Γ , v|Γ

(Kh,H − Ih,H )u|Γ , ψ|Γ = (K − I )u|Γ , ψ|Γ
VH φ|Γ , ψ|Γ = V φ|Γ , ψ|Γ
∗ ∗
(KH,h − IH,h )φ|Γ , v|Γ = (K − I )φ|Γ , v|Γ .
Furthermore we consider the discrete Poincaré-Steklov operator

∗
SH := WH + (KH,H ∗
− IH,H )VH−1 (KH,H − IH,H ) (12.79)
on coarse mesh functions where the operators are defined as above by substituting
TH for Th .
With the discrete Poincaré-Steklov operators S̃h and SH we formulate discrete
problems to (12.68):
Find uH ∈ TH such that
a(uH , v) + SH uH |Γ , v|Γ = FH (v) ∀v ∈ TH , (12.80)
and
find ũh ∈ Th such that
a(ũh , v) + S̃h ũh|Γ , v|Γ = F̃h (v) ∀v ∈ Th (12.81)
where FH (·) and F̃h (·) are obtained by substituting SH for S in F of (12.68) and
S̃h , respectively.
For our analysis to derive an a posteriori error estimate (Theorem 12.12) we have
to make the following saturation assumption.
Assumption 12.3 Let u, uH , ũh be defined as in (12.68), (12.80) and (12.81).
There exists a constant κ ∈ (0, 1) independent of H, h such that
u − ũh H 1 (Ω1 ) ≤ κu − uH H 1 (Ω1 )
The foregoing assumption immediately implies
(1 − κ)u − uH H 1 (Ω1 ) ≤ ũh − uH H 1 (Ω1 ) ≤ (1 + κ)u − uH H 1 (Ω1 ) . (12.82)
The following a posteriori error estimate is proved in[274] (see also [312]).
Theorem 12.12 Assume Assumption 12.3 holds. Let T0 ⊂ T1 ⊂ T2 ⊂ . . . be a
sequence of hierarchical subspaces where T0 is an initial FEM space (cf. (12.69)).
The refinement of all triangles defining Tk according to Fig. 12.2 gives us Th,k .
Let k denote the number of the refinement level and uk the corresponding Galerkin
solution of (12.80) and u the exact solution of (12.68), then there are constants
ζ1 , ζ2 > 0, k0 ∈ N0 , such that for all k ≥ k0

n
2
1/2
n
2
1/2
ζ1 θi,k ≤ u − uk H 1 (Ω1 ) ≤ ζ2 θi,k (12.83)
i=1 i=1
where the local error indicators

|2ϑΩ (bi,k ) + ϑΓ (bi,k )|
θi,k := (12.84)
bi,k H 1 (Ω1 )
are obtained via basis functions bi,k ∈ Th,k \Tk by a domain part

ϑΩ (bi,k ) := f bi,k dx − ρ(|∇uk |) ∇uk · ∇bi,k dx (12.85)
Ω1 Ω1
and a boundary part
9 : 9 :
ϑΓ (bi,k ) := 2t0 + S̃h,k u0 , bi,k|Γ − S̃h,k uk|Γ , bi,k|Γ (12.86)
with S̃h,k defined as in (12.78) with respect to Th,k , Tk instead of Th , TH .

Proof We define the form

# $
Q(w, uk , v) = 2 ρ(|∇w|) ∇w − ρ(|∇uk |) ∇uk − ρ̃(∇uk ) ∇(w − uk ) · ∇v dx
Ω1
for all w, v ∈ Th,k . Note that
Q(w, uk , v) = a(w, v) − a(uk , v) − auk (w − uk , v) (12.87)
for all w, v ∈ Th,k . The bilinear form aw is defined in (12.73). Since the function
%
R2 −→ R2
G:
x −→ 2ρ(|x|)x
is differentiable and 2ρ̃ (see (12.74)) is the Jacobian of G we obtain
G(∇ ũh,k ) − G(∇uk ) − 2ρ̃(∇uk )(∇(ũh,k − uk ))[L2 (Ω1 )]2

δ(k) := −→ 0
∇(ũh,k − uk )[L2 (Ω1 )]2
for
ũh,k − uk H 1 (Ω1 ) → 0 (12.88)

and
|Q(ũh,k , uk , v)| ≤ δ(k)ũh,k − uk H 1 (Ω1 ) vH 1 (Ω1 ) (12.89)
where ũh,k denotes the Galerkin solution of (12.81) substituting Th by Th,k . (12.88)
is obvious for k → ∞ with (12.82).
We define ek ∈ Th,k by
b(ek , v) = F̃k (v) − a(uk , v) − S̃k uk , v ∀v ∈ Th,k (12.90)
where the bilinear form b(·, ·) is given in (12.76), a(·, ·) in (12.67), and F̃k , S̃k
in (12.81) taking Tk for TH . Next, we show that there are constants μ0 , ν0 > 0,
which are independent of k, such that
μ0 ũh,k − uk H 1 (Ω1 ) ≤ ek H 1 (Ω1 ) ≤ ν0 ũh,k − uk H 1 (Ω1 ) . (12.91)
By (12.75), (12.87), (12.81), and (12.90) we obtain
auk (ũh,k − uk , ũh,k − uk )

μũh,k − uk H 1 (Ω1 ) ≤
ũh,k − uk H 1 (Ω1 )

≤ sup auk (ũh,k − uk , v) + S̃k (ũh,k − uk ), v
v∈M

≤ sup a(ũh,k , v) − a(uk , v) + S̃k (ũh,k − uk ), v − Q(ũh,k , uk , v)
v∈M

≤ sup F̃k (v) − a(uk , v) − S̃k uk , v − Q(ũh,k , uk , v)
v∈M

≤ sup b(ek , v) − Q(ũh,k , uk , v)
v∈M
≤ ek H 1 (Ω1 ) + δ(k)ũh,k − uk H 1 (Ω1 )
where M := {v ∈ Th,k | vH 1 (Ω1 ) = 1}. The second inequality follows by the
positive definiteness of S̃k , the last by (12.89) and the Cauchy-Schwarz inequality.
Furthermore, we obtain by (12.90), (12.81), (12.87), and (12.75) that
ek 2H 1 (Ω ) = b(ek , ek ) = a(ũh,k , ek ) − a(uk , ek ) + S̃k (ũh,k − uk ), ek

1
= auk (ũh,k − uk , ek ) + Q(ũh,k , uk , ek ) + S̃k (ũh,k − uk ), ek

≤ (ν + δ(k))ũh,k − uk H 1 (Ω1 ) ek H 1 (Ω1 )
+ S̃k (ũh,k − uk )H −1/2 (Γ ) ek H 1/2 (Γ )

≤ (ν + δ(k) + νS )ũh,k − uk H 1 (Ω1 ) ek H 1 (Ω1 )
where the last inequalities follow, again, by (12.89) and the existence of a constant
νS since S̃k is uniformly bounded for all k > k0 , k0 ∈ N0 constant. Here, the
uniform boundedness of S̃k follows by the approximation properties of the discrete
BE spaces defining S̃k and the boundedness of S.
If k0 is sufficiently large (such that δ(k) ≤ δ0 < μ for all k ≥ k0 ) then (12.91)
follows with μ0 = μ − δ0 and ν0 = ν + δ0 + νS .
Since
ek 2H 1 (Ω ) = b(ek , ek )

1
where b(·, ·) is defined in (12.76), we can apply (12.77) to obtain

mk
c1 ek 2H 1 (Ω ) ≤ P (k) ek 2H 1 (Ω ) + Pi,k ek 2H 1 (Ω ) ≤ c2 ek 2H 1 (Ω ) .
1 1 1 1
i=1
(12.92)
Here P (k) : Th,k → Tk and Pi,k : Th,k → span{bi,k } are the Galerkin projections
with respect to the bilinear form b(·, ·). With the notations of (12.77) P (k) = PH .
By definition of P (k) and Pi,k , by (12.90), (12.81), and (12.80) it follows that
P (k) ek 2H 1 (Ω ) = b(ek , P (k) ek ) = F̃k (P (k) ek ) − F̃k (P (k) ek ) = 0 (12.93)

1
and
b(ek , bi,k ) F̃k (bi,k ) − a(uk , bi,k ) − S̃k uk , bi,k

Pi,k ek = bi,k = bi,k .
b(bi,k , bi,k ) bi,k 2 1 H (Ω1 )
Hence, we have
|F̃k (bi,k ) − a(uk , bi,k ) − S̃k uk , bi,k |

Pi,k ek H 1 (Ω1 ) = = θi,k . (12.94)
bi,k H 1 (Ω1 )
By (12.77), (12.93) and (12.94) we obtain

mk
c1 ek 2H 1 (Ω ≤ 2
θi,k ≤ c2 ek 2H 1 (Ω ) .
1) 1
i=1
This yields together with (12.91) that

mk 1/2
mk 1/2
1 1
√ 2
θi,k ≤ ũh,k − uk H 1 (Ω1 ) ≤ √ 2
θi,k
ν0 c2 μ0 c 1
i=1 i=1
Table 12.2 Results for adaptive algorithm based on Theorem 12.12 for (NP) with u1 , u2
from (12.95), ζ = 0.15
L nk dim Tk dim τk Ek ηk ηk /Ek κk αk
0 37 21 16 0.10608 0.13067 1.232 – –
1 55 37 18 0.07596 0.08283 1.090 0.716 0.842
2 78 58 20 0.05511 0.06495 1.179 0.725 0.919
3 109 85 24 0.04510 0.05596 1.241 0.818 0.599
4 163 129 34 0.03626 0.04373 1.206 0.804 0.542
5 454 396 58 0.02063 0.02419 1.172 0.569 0.550
6 677 595 82 0.01654 0.01936 1.171 0.802 0.554
and, finally, we obtain the assertion of the theorem by (12.82) with

1 1
ζ1 = √ and ζ2 = √ .

(1 + κ)ν0 c2 (1 − κ)μ0 c1
In Table 12.2 , we list the numerical experiment for (NP) with ≡ 1 (for (t) =
1
6 1 + 1+5t
5
see [91]) and choose Ω1 to be the L-shaped domain with corners at
(0, 0), (0, 14 ), (− 14 , 14 ), (− 14 , − 14 ), ( 14 , − 14 ), ( 14 , 0). The exact solution of the model
problem (NP) is given by
/
u1 (r, α) = r sin 3 (α − 2 ), u2 (x1 , x2 ) = ln (x1 + 18 )2 + (x2 + 18 )2 .
2/3 2 π
(12.95)
The functions u0 , t0 , f are chosen to yield the exact solution. The quantities in
Table 12.2 are given as follows: With k we denote the refinement level, with nk
the total number of unknowns and with Nk the total number of triangles defining
Tk . The error Ek is defined as
Ek := u − uk 1,Ω1 .
The global error indicator ηk is defined by

⎛ ⎞1/2

Nk
1/2
ηk = ⎝ 2 ⎠
ηi,k , ηi,k := θi21 ,k + θi22 ,k + θi23 ,k (i = 1, . . . , Nk ) .
i=1
Here i1 , i2 , i3 denote the three edges and the corresponding new base functions for
every element of the old mesh. The values of the quotient ηk /Ek , the efficiency
index, indicate the efficiency of the error indicator ηk and confirm Theorem 12.12.
The quantity
u − uk 1,Ω1
κk :=
u − uk+1 1,Ω1
estimates the saturation constant κ. Since κk is bounded by a constant less than 1

the saturation condition (Assumption 12.3) is satisfied for the sequence of meshes
which is generated by our adaptive algorithm. The experimental convergences rates
αk are given by
ln(Ek /Ek−1 )
αk = .
ln(nk−1 /nk )
From Table 12.2 we see that αk approaches 1/2, which is the convergence rate in
case of a smooth solution. This shows the quality of the adaptive algorithm. For
uniform meshes one obtains the non-optimal convergence rate α = 1/3 The above
hierarchical method is easily implemented since for the computation of the error
indicators one can use the same routine as for the computation of the entries of the
Galerkin matrix.
12.3.4 Convergence of Adaptive FEM-BEM Couplings
Let ζl denote an a posteriori estimator, e.g. the residual estimator R1 +R2 +R3 +R4
of Sect. 12.3.2. We assume ζl2 = ζl (Il )2 := ζl ()2 , where Il = TlΩ ∪ ElΩ ∪ TlΓ
∈Il
with TlΩ = {1 , . . . , N }, ElΩ the set of interior edges, TlΓ the set boundary edges.
ζl () denotes the local refinement indicator for ∈ Il .
Algorithm 12.2 (Adaptive) Input: Initial mesh T0 , li = 0, 0 < θ ≤ 1:
i) Compute discrete solution ul ∈ Hl = XNl × YNl
ii) Compute refinement indicators ζl () ∀ ∈ Il
iii) Determine Ml ⊆ Il such that Dörfler marking

θ ζl2 ≤ ζl (T )2
∈Ml
holds
iv) Compute new triangulation Tl+1 , where at least all marked elements ∈ Ml
are refined.
v) Increase counter l and go to i)
Output: Sequence of Galerkin solutions {ul }L
l=0 , sequence of error estimators {ζl }l=0
L
and sequence of triangulations {Tl }Ll=0 .

For the symmetric coupling method the reliability of the corresponding error
estimators follows from Theorem 12.11 in Sect. 12.3.2. Corresponding residual type
error estimators and a priori estimates hold for the Johnson-Nedelec and the Bielak-
MacCamy couplings. As shown in [[8] for Bielak-MacCamy coupling ], in [[9] for
the symmetric coupling], in [[181] for the Johnson-Nedelec coupling] the above
adaptive algorithm, (steered by the residual error estimator) converges. The proof
crucially needs the following inverse estimates of the boundary integral operators.
Lemma 12.13 ([264]) Let TlΓ be a regular triangulation of Γ . Let hl ∈ P0 (TlΓ )

with hl = ||1/(d−1), where d is the dimension of Ω. Then there exist constants

K , C V > 0 with
Cinv inv
1/2
hl ∇Γ Kvl L2 (Γ ) ≤ Cinv
K
vl H 1/2 (Γ )
1/2
hl W vl L2 (Γ ) ≤ Cinv
K
vl H 1/2 (Γ )
1/2
hl ∇Γ V ψl L2 (Γ ) ≤ Cinv
V
ψl H −1/2 (Γ )
hl K vl L2 (Γ ) ≤ Cinv

1/2 V
ψl H −1/2 (Γ )
for all vl ∈ S p (TlΓ ), ψl ∈ Pp (TlΓ ) (continous, respectively discontinous

K , C V only depend on the γ -regularity
polynomials of degree p). The constants Cinv inv
Γ
of Tl , the boundary mesh, and the polynomial degree p.
A further ingredience of the proof is the newest vertex bijection (NVB) as
refinement strategy. See [181] for further details. In [176] we prove convergence
of the adaptive algorithm with the hierarchical two-level estimator, considered in
Sect. 12.3.3. There we show that the usual adaptive algorithm (with the weighted-
residual error estimator terms in Theorem 12.11) drives the hierarchical estimator
to zero.
12.3.5 Other Coupling Methods
In this section we consider the Johnson-Nedelec coupling [263, 439] which is often
called the direct one-equation coupling, since only one equation of the Calderon
projector is used. The first stability results rely on the compactness of the double
layer operator K ([263]). This has the disadvantage that Γ needs to be smooth
(then K is compact in the Laplace case), this is for standard FEM resp. BEM not
optimal. Here the work by ([360]) turned out to be the breakthrough showing (for
the first time for the Laplace transmission problem (IP)) that the Johnson-Nedelec
coupling is well-defined on polygonal domains. The proof shows stability of the
adjoint problem and was applied to some problems in linear elastostatics in [186].
A different approach was developed in [392], where an explicit stabilisation is
introduced which leads to an equivalent problem in the continuous case. Steinbach
shows in [392] that this equivalent problem (with a linear operator A in Ω1 ) is
elliptic under the assumption cmon > 1/4. Here cmon is the smallest eigenvalue
of A. This condition was improved to cmon > cK /4 where cK ∈ [1/2, 1) is
the contraction constant of the double layer potential K [330]. Unfortunately with
this stabilization an additional boundary integral equation must be solved on each

discrete level whereas for the implicit stabilization, given in [181], the equivalence
to (IP) holds on the discrete level. Thus Johnson-Nedelec coupling is obtained by
taking (12.38) and testing (12.36) in H −1/2(Γ ), i.e.:
Find (u, φ) ∈ H 1 (Ω1 ) × H −1/2(Γ ) such that for all (v, ψ) ∈ H 1 (Ω1 ) ×
H −1/2 (Γ ) there holds
B(u, φ), (v, ψ) = F ((v, ψ)) (12.96)
where
B(u, φ), (v, ψ) := (A∇u, ∇v) − φ, v + ψ, (I − K)u + V φ

F ((v, ψ)) := (f, v) + t0 , v + ψ, (I − K)u0
The implicit stabilization reads: Assume there exists ξ ∈ YN with ξ, 1 = 0 there
holds
B̃(u, φ), (v, ψ) = F̃ ((v, ψ)) (12.97)
where
B̃(u, φ), (v, ψ) := B(u, φ), (v, ψ) + ξ, (I − K)u + V φ ξ, (I − K)v + V ψ
F̃ ((v, ψ)) := F ((v, ψ)) + ξ, (I − K)u0 ξ, (I − K)v + V ψ
Now (uh , φh ) ∈ XN × YN solves (12.96) ∀(vh , ψh ) ∈ XN × YN if and

only if it solves (12.97) for all ∀(vh , ψh ) ∈ XN × YN . One obtains under the
assumption (12.27), (12.28) quasi-optimality of the Galerkin scheme for (12.97)
(see [181] for details).
Our model problem (IP) can also be reformulated with the Bielak-MacCamy
coupling [54]. This method is also called indirect one-equation coupling, since an
indirect ansatz (which does not use the Calderon system) is applied to solve the
exterior problem. The Bielak-MacCamy coupling reads:
Find (u, φ) ∈ H 1 (Ω)×H −1/2(Γ ) such that for all (v, ψ) ∈ H 1 (Ω)×H −1/2(Γ )
there holds
B(u, φ), (v, ψ) = F ((v, ψ)) (12.98)
where
B(u, φ), (v, ψ) := (A∇u, ∇v) − (I − K )φ, v + ψ, V φ − u

F ((v, ψ) := (f, v) + t0 , v + ψ, u0
12.4 Least Squares FEM/BEM Coupling 491
Now the implicit stabilization reads with ξ as above:
B̃(u, φ), (v, ψ) = F̃ ((v, ψ)) (12.99)
where
B̃(u, φ), (v, ψ) := B(u, φ), (v, ψ) + ξ, V φ − u ξ, V ψ − v

F̃ ((v, ψ)) := F ((v, ψ)) − ξ, u0 ξ, V ψ − v
Analogously (uh , φh ) ∈ XN × YN solves 12.98 if and only if (uh , φh ) solves 12.99

on XN × YN and again the Galerkin solution (uh , φh ) converges quasioptimally in
the energy norm. These results hold true for polygonal Γ and coupling problems in
elastostatics (see again [181] for details).
12.4 Least Squares FEM/BEM Coupling for Transmission

Problems
Here we report from [296] a least squares formulation for the numerical solution of
second-order linear transmission problems, where in a bounded domain the second
order partial differential equation is rewritten as first-order system. The least squares
functional is given in terms of Sobolev norms of order −1 and of order 1/2 and uses
boundary integral operators. In [296] these norms are computed by approximating
the corresponding inner product using multilevel preconditioners (multigrid and
BPX) for the differential operator and weakly singular integral operator.
Let Ω1 := Ω ⊂ Rd , d ≥ 2 be a bounded domain with Lipschitz boundary Γ =
∂Ω1 , and Ω2 := Rd \Ω̄1 with normal n on Γ pointing into Ω2 . Let f ∈ L2 (Ω1 ),
u0 ∈ H 1/2(Γ ), t0 ∈ H −1/2(Γ ). We consider the model transmission problem of
finding u1 ∈ H 1 (Ω1 ), u2 ∈ Hloc1 (Ω ) such that
2
− div(a∇u1 ) = f in Ω1 (12.100)
Δu2 = 0 in Ω2 (12.101)
u1 = u2 + u0 on Γ (12.102)
∂u2
(a∇u1) · n = + t0 on Γ (12.103)
∂n

A ln |x| + o(1), d = 2
u2 (x) = , |x| → ∞ (12.104)
O(|x|2−d ), d≥3
Let aij ∈ L∞ (Ω1 ) such that a = (aij ) satisfies for some α > 0
αz2 ≤ zT a(x)z ∀z ∈ Rd and for almost all x ∈ Ω1 .

In the following, we will apply the boundary integral equation method in Ω2 and
reduce the original problem to a nonlocal transmission problem on the bounded
domain Ω. The fundamental solution of the Laplacian is given by
%
− 1 ln |x − y|, d = 2
G(x, y) = 1 ω2
ωd |x − y|
2−d , d ≥ 3
where we have ω2 = 2π, ω3 = 4π. For all x ∈ Ω2

"
∂ ∂u
u2 (x) = G(x, y)u(y) − G(x, y) dsy
Γ ∂n(y) ∂n(y)
satisfies the Laplace equation (12.101) and the radiation condition (12.104). By
using the boundary integral operators

V ψ(x) := 2 G(x, y)ψ(y) dsy , x ∈ Γ (12.105)
Γ

∂
Kψ(x) := 2 G(x, y)ψ(y) dsy , x ∈ Γ (12.106)
Γ ∂n y

∂
K ψ(x) := 2 G(x, y)ψ(y) dsy , x ∈ Γ (12.107)
∂nx Γ

∂ ∂
W ψ(x) := −2 G(x, y)ψ(y) dsy , x ∈ Γ (12.108)
∂nx Γ ∂ny
together with their well known-jump conditions we obtain the following integral
equations
∂u2 ∂u2
2 = −W u2 + (I − K ) (12.109)
∂n ∂n
∂u2
0 = (I − K)u2 + V . (12.110)
∂n
In this way, the original transmission problem (12.100) — (12.104) reduces to
the following non-local boundary value problem in Ω. Find (u, σ ) ∈ H 1 (Ω) ×
− div(a∇u) = f in Ω (12.111)
σ = (a∇u) · n on Γ (12.112)
2(σ − t0 ) = −W (u − u0 ) + (I − K )(σ − t0 ) on Γ (12.113)
0 = (I − K)(u − u0 ) + V (σ − t0 ) on Γ (12.114)
Note that the flux variable θ := a∇u belongs to the Hilbert space.
H (div; Ω) = {θ ∈ [L2 (Ω)]d : θ 2[L2 (Ω)]d + div θ 2L2 (Ω) < ∞}.
with the inner product
(θ, ζ )H (div;Ω) = (θ, ζ )[L2 (Ω)]d + (div θ, div ζ )L2 (Ω) .
Moreover, for all ζ ∈ H (div; Ω) there holds ζ ·n ∈ H −1/2 (Γ ) and ζ ·nH −1/2 (Γ ) ≤
ζ H (div;Ω) (see [196]).
With the interface conditions we can rewrite the transmission problem as follows
with a first order system on Ω:
Find (θ, u, σ ) ∈ H (div; Ω) × H 1 (Ω) × H −1/2 (Γ ) such that
θ = a∇u in Ω (12.115)
− div θ = f in Ω (12.116)
σ = θ · n on Γ (12.117)
2(σ − t0 ) = −W (u − u0 ) + (I − K )(σ − t0 ) on Γ (12.118)
0 = (I − K)(u − u0 ) + V (σ − t0 ) on Γ (12.119)
Let H̃ −1 (Ω) denote the dual space of H 1(Ω), equipped with the dual norm
(w,v)L2 (Ω)
wH̃ −1 (Ω) = supv∈H 1 (Ω) vH 1 (Ω) . Then the solution of (12.115)—(12.119) is a
solution of the following minimization problem:
Find (θ, u, σ ) ∈ X := [L2 (Ω)]d × H 1 (Ω) × H −1/2(Γ ) such that
J (θ, u, σ ) = min J (ζ, v, τ ) (12.120)

(ζ,v,τ )∈X
where J is the quadratic functional defined by
J (ζ, v, τ ) = a∇v − ζ 2[L2 (Ω)]d + (I − K)(v − u0 ) + V (τ − t0 )2H 1/2 (Γ )

1
+ div ζ + f − δΓ ⊗ (W (v − u0 ) + 2ζ · n − 2t0 − (I − K )(τ − t0 ))2H̃ −1 (Ω)
2
= a∇v − ζ 2[L2 (Ω)]d + (I − K)v + V τ − (I − K)u0 − V t0 2H 1/2 (Γ )
1
+ div ζ − δΓ ⊗ (W v + 2ζ · n − (I − K )τ )
2
1
+ f + δΓ ⊗ (W u0 + 2t0 − (I − K )t0 )2H̃ −1 (Ω) . (12.121)
2
Here δΓ ⊗ τ denotes the distribution in H̃ −1 (Ω) for τ ∈ H −1/2(Γ ) defined by
δΓ ⊗ τ, ϕ H̃ −1 (Ω)×H 1 (Ω) = (τ, ϕ|Γ )H −1/2 (Γ )×H 1/2 (Γ ) ∀ ϕ ∈ H 1 (Ω).

Due to coercivity and continuity of the corresponding variational problem the

authors obtain in [296] uniqueness of (12.120) and equivalence between (12.115)—
(12.119) and (12.120).
Defining g(ζ, v, τ ) := div ζ − 12 δΓ ⊗ (W v + 2ζ · n − (I − K )τ ) we introduce
with
B((θ, u, σ ), (ζ, v, τ )) = (a∇u − θ, a∇v − ζ )L2 (Ω)

+((I − K)u + V σ, (I − K)v + V τ )H 1/2 (Γ )
+(g(θ, u, σ ), g(ζ, v, τ ))H̃ −1 (Ω) (12.122)
and
G(ζ, v, τ ) = ((I − K)v + V τ, (I − K)u0 + V t0 )H 1/2 (Γ )

1
−(g(ζ, v, τ ), f + δΓ ⊗ (W u0 + 2t0 − (I − K )t0 ))H̃ −1 (Ω) (12.123)
2
the variational formulation for (12.120) as:
Find (θ, u, σ ) ∈ X = [L2 (Ω)]d × H 1 (Ω) × H −1/2(Γ ) such that
B((θ, u, σ ), (ζ, v, τ )) = G(ζ, v, τ ) ∀(ζ, v, τ ) ∈ X. (12.124)
Theorem 12.13 The bilinear form B(·, ·) is strongly coercive in X, i.e. there holds
B((ζ, v, τ ), (ζ, v, τ )) (ζ, v, τ )2X , ∀(ζ, v, τ ) ∈ X. (12.125)
Proof Let (ζ, v, τ ) ∈ X = [L2 (Ω)]d × H 1 (Ω) × H −1/2 (Γ ).

We can estimate ζ [L2 (Ω)]d by
ζ [L2 (Ω)]d ≤ ζ − a∇v[L2 (Ω)]d + a∇v[L2 (Ω)]d ζ − a∇v[L2 (Ω)]d + vH 1 (Ω) .
(12.126)
Using the boundedness of V −1 (as a mapping from H 1/2(Γ ) into H −1/2(Γ )) and
I − K we can estimate
τ H −1/2 (Γ ) V τ H 1/2 (Γ )
V τ + (I − K)vH 1/2 (Γ ) + (I − K)vH 1/2 (Γ )
V τ + (I − K)vH 1/2 (Γ ) + vH 1/2 (Γ )
V τ + (I − K)vH 1/2 (Γ ) + vH 1 (Ω) . (12.127)
Now we use the Poincaré-Steklov operator S : H 1/2(Γ ) → H −1/2 (Γ ) for the

exterior domain, given by
S := W + (I − K )V −1 (I − K).
From [91, Lemma 4] we know that with the L2 inner products (·, ·) and ·, · on Ω
and Γ , respectively, there holds
1
v2H 1 (Ω) (a∇v, ∇v) + Sv, v ∀v ∈ H 1 (Ω),
2
yielding
(a∇v, ∇w) + 12 Sv, w

vH 1 (Ω) sup .
w∈H 1 (Ω) wH 1 (Ω)
We can expand the expression by
1
(a∇v, ∇w) + Sv, w
2
1
= (a∇v − ζ, ∇w) − (div ζ, w) + ζ · n, w + Sv, w
2
1
= (a∇v − ζ, ∇w) − (div ζ − δΓ ⊗ [ζ · n + Sv], w)
2
and obtain
(a∇v − ζ, ∇w) (div ζ − δΓ ⊗ [ζ · n + 12 Sv], w)

vH 1 (Ω) sup + sup
w∈H 1 (Ω) wH 1 (Ω) w∈H 1 (Ω) wH 1 (Ω)
1
≤ a∇v − ζ [L2 (Ω)]d + div ζ − δΓ ⊗ [ζ · n + Sv]H̃ −1 (Ω) . (12.128)
2
Finally, writing
Sv = W v − (I − K )τ + (I − K )V −1 (V τ + (I − K)v)
we can estimate
1
div ζ − δΓ ⊗ [ζ · n + Sv]H̃ −1 (Ω)
2
1
div ζ − δΓ ⊗ [2ζ · n + W v − (I − K )τ ]H̃ −1 (Ω) + V τ + (I − K)vH 1/2 (Γ ) .
2
(12.129)
Collecting the bounds (12.126) for ζ [L2 (Ω)]d , (12.127) for τ H −1/2 (Γ ) , (12.128)
for vH 1 (Ω) and (12.129), we obtain (12.125).

Theorem 12.14 The bilinear form B(·, ·) is continuous in X × X and the linear
form G(·) is continuous on X.
Proof Following the definition of B(·, ·) we obtain first
B((θ, u, σ ), (ζ, v, τ )) ≤ a∇u − θ [L2 (Ω)]d · a∇v − ζ [L2 (Ω)]d

+(I − K)u + V σ H 1/2 (Γ ) · (I − K)v + V τ H 1/2 (Γ )
+g(θ, u, σ )H̃ −1 (Ω) · g(ζ, v, τ )H̃ −1 (Ω) .
Using the triangle inequality, the mapping properties and the trace theorem we have
a∇u − θ [L2(Ω)]d ≤ aL∞ (Ω) ∇u[L2 (Ω)]d + θ [L2 (Ω)]d uH 1 (Ω) + θ [L2 (Ω)]d
and
(I − K)u + V σ H 1/2 (Γ ) uH 1/2 (Γ ) + σ H −1/2 (Γ ) uH 1 (Ω) + σ H −1/2 (Γ ) .
Finally, there holds
1
g(θ, u, σ )H̃ −1 (Ω) ≤ div θ −δΓ ⊗θ ·nH̃ −1 (Ω) + δΓ ⊗(W u−(I −K )σ )H̃ −1 (Ω)
2
and we obtain
(div θ − δΓ ⊗ θ · n, v)
div θ − δΓ ⊗ θ · nH̃ −1 (Ω) = sup
v∈H 1 (Ω) vH 1 (Ω)
(div θ, v) − θ · n, v (θ, ∇v)
= sup = sup ≤ θ [L2 (Ω)]d
v∈H 1 (Ω) v H 1 (Ω) v∈H 1 (Ω) v H 1 (Ω)
and, analogously,
(δΓ ⊗ (W u − (I − K )σ ), v)
δΓ ⊗ (W u − (I − K )σ )H̃ −1 (Ω) = sup
v∈H 1 (Ω) vH 1 (Ω)
W u − (I − K )σ, v
= sup
v∈H 1 (Ω) vH 1 (Ω)
≤ W u − (I − K )σ H −1/2 (Γ ) uH 1 (Ω) + σ H −1/2 (Γ ) .
Collecting the individual terms, the continuity of B(·, ·) follows. The continuity of
G(·) can be seen analogously.

Now application of the Lax-Milgram lemma gives the following result (see
[296])
Theorem 12.15 There exists a unique solution of the variational least-squares
formulation (12.124), which is also a solution of (12.115) — (12.119).
12.4.1 The Discretized Least Squares Formulation
Following [62] we give an alternative representation for the norm in H̃ −1 (Ω) which
will be discretized later. Let T : H̃ −1 (Ω) → H 1 (Ω) be defined by Tf := w where
w ∈ H 1 (Ω) is the unique function satisfying
(∇w, ∇v) + (w, v) = (f, v) ∀v ∈ H 1 (Ω).
As observed in [62, Lemma 2.1], there holds
(v, θ )2
v2H̃ −1 (Ω) = sup = T v2H 1 (Ω) = (v, T v).
θ∈H 1 (Ω) θ H 1 (Ω)
2
Therefore, the inner product on H̃ −1 (Ω)× H̃ −1 (Ω) is given by (v, T w), for v, w ∈
H̃ −1 (Ω).
Let Vh ⊂ H 1 (Ω). Then let Th : H̃ −1 (Ω) → Vh be defined by Th f := w where
w ∈ Vh is the unique function satisfying
(∇w, ∇v) + (w, v) = (f, v) ∀v ∈ Vh .
In case of the space H 1/2(Γ ) we proceed analogously: Let R : H 1/2(Γ ) →

H −1/2(Γ ) be defined by Rf := w where w ∈ H −1/2 (Γ ) is the unique function
satisfying
V w, v = f, v ∀v ∈ H −1/2 (Γ ).
Then there holds
v, θ 2 v, θ 2
v2H 1/2 (Γ ) = sup ∼ sup = v, Rv .
θ∈H −1/2 (Γ ) θ H −1/2 (Γ ) θ∈H −1/2 (Γ ) V θ, θ
2
where ∼ denotes norm equivalence. Let Sh ⊂ H −1/2(Γ ). Then let Rh :

H 1/2(Γ ) → Sh be defined by Rh f := w where w ∈ SH is the unique function
satisfying
V w, v = f, v ∀v ∈ Sh .
For the numerical efficiency of the proposed scheme we replace in [296] Th by the
preconditioner Bh and Rh by the preconditioner Ch such that there holds (Th ·, ·) ∼
(Bh ·, ·) and Rh ·, · ∼ Ch ·, · . Bh and Ch are chosen in such a way that their
evaluation is much cheaper than the computation of Th vh or Rh τh .
For the discretization we assume that there exists projection operators which are
bounded independently of h
Ph : H 1 (Ω) → Vh ⊂ H 1 (Ω) (12.130)

Qh : H −1/2(Γ ) → Sh ⊂ H −1/2(Γ ). (12.131)
As a consequence also their adjoints are bounded
Ph∗ : H̃ −1 (Ω) → Vh∗ ⊂ H̃ −1 (Ω) (12.132)

Q∗h : H 1/2(Γ ) → Sh∗ ⊂ H 1/2(Γ ). (12.133)
Replacing T in the representation of the H̃ −1 (Ω) inner product by the precon-

ditioner Bh and R in the representation of the H 1/2 (Γ ) inner product by the
preconditioner Ch we obtain the discretized formulation:
Find (θh , uh , σh ) ∈ Xh such that
B (h) ((θh , uh , σh ), (ζh , vh , τh )) = G(h) (ζh , vh , τh ) ∀(ζh , vh , τh ) ∈ Xh ,

(12.134)
where Xh = Hh × Vh × Sh , Hh ⊂ [L2 (Ω)]d with
B (h) ((θ, u, σ ), (ζ, v, τ )) = (a∇u − θ, a∇v − ζ )L2 (Ω)

+Ch Q∗h ((I − K)u + V σ ), Q∗h ((I − K)v + V τ ) L2 (Γ )
+(Bh Ph∗ g(θ, u, σ ), Ph∗ g(ζ, v, τ ))L2 (Ω) (12.135)
G(h) (ζ, v, τ ) = Ch Q∗h ((I − K)v + V τ ), Q∗h ((I − K)u0 + V t0 ) L2 (Γ )

1
−(Bh Ph∗ g(ζ, v, τ ), Ph∗ (f + δΓ ⊗ (W u0 + 2t0 − (I − K )t0 )))L2 (Ω)
2
(12.136)
for all (θ, u, σ ), (ζ, v, τ ) ∈ X. Analogously to the proofs of Theorem 12.13 and
Theorem 12.14 the authors show in [296] the following result:
Theorem 12.16 For arbitrary functions (ζh , vh , τh ) ∈ Xh the following a-priori
estimate holds
vh 2H 1 (Ω) + ζh 2[L2 (Ω)]d + τh 2H −1/2 (Γ ) B (h) ((ζh , vh , τh ), (ζh , vh , τh ))
∼ a∇vh − ζh 2[L2 (Ω)]d

1
+Bh Ph∗ (div ζh − δΓ ⊗ [W vh + 2ζh · n − (I − K )τh ])2L2 (Ω)
1/2
2
+Ch Q∗h [(I − K)vh + V τh ]2L2 (Γ )
1/2
(12.137)
12.5 FE/BE Coupling for Signorini Contact 499
Furthermore for arbitrary functions (ζ, v, τ ) ∈ X the discretized bilinear

form B (h) (·, ·) and the discretized linear form G(h) (·) are continuous, i.e. there
holds B (h) ((θ, u, σ ), (ζ, v, τ )) (θ, u, σ )X · (ζ, v, τ )X , and G(h) ((ζ, v, τ ))
(ζ, v, τ )X for all (θ, u, σ ), (ζ, v, τ ) ∈ X with constants independent of h.
For finite dimensional subspaces Xh := Hh × Vh × Sh ⊂ X we assume the
usual approximation properties, e.g. for the space Vh of continuous, piecewise
linear/bilinear functions on a regular triangulation, for the space Hh of either
piecewise constant functions or continuous, piecewise linear/bilinear functions or
H (div; Ω)-conforming Raviart-Thomas elements of lowest order, and for the space
Sh of piecewise constant functions on the boundary (see [105, 352]):
There exists r > 1 such that for all u ∈ H r (Ω)
inf u − vh H 1 (Ω) hr−1 uH r (Ω) ,

vh ∈Vh
inf σ − τh H −1/2 (Γ ) hr−1 σ H r−3/2 (Γ ) hr−1 uH r (Ω) ,

τh ∈Sh
inf θ − ζh [L2 (Ω)]d hr−1 θ [H r−1 (Ω)]d hr−1 uH r (Ω) .
ζh ∈Hh
Now, application of Theorem 12.16, the Lax-Milgram lemma and the Second
Strang lemma gives the following result (see [296] for details ):
Theorem 12.17 The unique solution (θh , uh , σh ) ∈ Xh of the discretized formula-
tion (12.134) exists and there holds the following convergence estimate
u − uh H 1 (Ω) + θ − θh [L2 (Ω)]d + σ − σh H −1/2 (Γ ) hr−1 uH r (Ω) .
For numerical experiments see [296].
12.5 FE/BE Coupling for Interface Problems with Signorini

Contact
12.5.1 Primal Method
Here we report from [298] a FEM-BEM coupling procedure which is based on

reducing the given nonlinear interface problem with contact to a boundary / domain
variational inequality. In [298] also Coulomb friction is considered. For the ease of
the reader here we restrict our presentation to the simpler case of Signorini contact
(see also [292]).
Let Ω ⊂ Rd , d ≥ 2, be a bounded domain with Lipschitz boundary Γ . Let Γ =
Γt ∪ Γs where Γt and Γs are nonempty, disjoint and open in Γ . In the interior part
we consider a nonlinear partial differential equation modeling nonlinear material
behavior in elasticity, whereas in the exterior part we consider the Laplace equation
and impose a radiation condition:
− div( (|∇u|) · ∇u) = f in Ω (12.138)

−Δu = 0 in Ωc = R \Ω̄ d
(12.139)
,
u(x) = a + 2π log |x| + o(1)
b
if d = 2,
(|x| → ∞), (12.140)
u(x) = O(|x| 2−d
) for d ≥ 3,
where a, b are real constants (constant for any u but varying with u).
Further, : [0, ∞) → [0, ∞) is a C 1 [0, ∞) function with t · (t) being
monotonously increasing with t, (t) ≤ 0 , (t · (t)) ≤ 1 and further (t) +
t · min{0, (t)} ≥ α > 0. With u1 := u|Ω and u2 := u|Ωc , the tractions on Γ are
given by (|∇u1 |) ∂u ∂u2
∂n and − ∂n with normal n pointing into Ωc .
1
We consider transmission conditions on Γt
∂u1 ∂u2
u1 |Γt − u2 |Γt = u0 |Γt and (|∇u1 |) |Γt − |Γ = t0 |Γt , (12.141)
∂n ∂n t
and Signorini conditions on Γs
u1 |Γs − u2 |Γs ≤ u0 |Γs

∂u1 ∂u2
(|∇u1 |) |Γ = |Γ + t0 |Γs ≤ 0 (12.142)
∂n s ∂n s
∂u1
0 = (|∇u1 |) |Γ · (u2 + u0 − u1 )|Γs .
∂n s
Given data f ∈ L2 (Ω), u0 ∈ H 1/2(Γ ), and t0 ∈ H −1/2(Γ ) (with (f, 1)L2 (Ω) +
t0 , 1 = 0 if d = 2) we look for u1 ∈ H 1 (Ω) and u2 ∈ Hloc 1 (Ω )
c
satisfying (12.138)–(12.142) in a weak form.
Setting
t
g(t) = s · (s) ds
0
the assumptions on yield that

G(u) = 2 g(|∇u|) dx
Ω
is finite for any u ∈ H 1 (Ω) and its Fréchet derivative

DG(u; v) = 2 (|∇u|)(∇u)T · ∇v dx ∀u, v ∈ H 1 (Ω) (12.143)
Ω
is uniformly monotone, i.e., there exists a constant γ > 0 such that
γ |u − v|2H 1 (Ω) ≤ DG(u; u − v) − DG(v; u − v) ∀u, v ∈ H 1 (Ω), (12.144)
=1/2 (Γs ) where H

Let E := H 1 (Ω) × H =1/2(Γs ) := {w ∈ H 1/2(Γ ) : supp w ⊆
Γs } and set
D := {(u, v) ∈ E : v ≥ 0 a.e. on Γs and S1, u|Γ + v − u0 = 0 if d = 2},
where S denotes the Poincaré-Steklov operator for the exterior problem:
S = 1/2 W + (K − I )V −1 (K − I )
Then the primal formulation of (12.138)–(12.142), called problem (SP), consists in

finding (û, v̂) in D such that
Ψ (û, v̂) = inf Ψ (u, v).

(u,v)∈D
where

1
Ψ (u, v) := 2 g(|∇u|) dx + S(u|Γ + v), u|Γ + v − λ(u, v),
Ω 2
and λ ∈ E ∗ , the dual of E, is given by
λ(u, v) := L(u, u|Γ + v) + Su0 , u|Γ + v
with

L(u, v) := 2 f · u dx + 2 t0 · v ds
Ω Γ
for any (u, v) ∈ E.

Due to [85] there exists exactly one solution (û, v̂) ∈ D of problem (SP), which is
the variational solution of the transmission problem (12.138)—(12.142). Moreover,
(û, v̂) ∈ D is the unique solution of the variational inequality
A (û, v̂)(u − û, v − v̂) ≥ λ(u − û, v − v̂) (12.145)
for all (u, v) ∈ D, with
A (u, v)(r, s) := DG(u, r) + S(u|Γ + v), r|Γ + s . (12.146)

For the discretization we take nested regular quasi-uniform meshes (Th )h

consisting of triangles or quadrilaterals. Then, let Hh1 denote the related continuous
and piecewise affine-linear trial functions on the triangulation Th . The mesh on Ω
−1/2
induces a mesh on the boundary, so that we may consider Hh as the piecewise
constant trial functions. Assuming that the partition of the boundary leads also to a
=1/2 is then the subspace of continuous and piecewise affine-linear
partition of Γs , H h
functions on the partition of Γs which vanish at intersection points in Γ¯s ∩ Γ¯t . Then
we have Hh1 × H =1/2 × H −1/2 ⊂ H 1 (Ω) × H =1/2(Γs ) × H −1/2 (Γ ). Now, Dh is
h h
given by
=
Dh := {(uh , vh ) ∈ Hh1 × H
1/2
: v(xi ) ≥ 0, ∀xi node of the partition of Γs ,
h
and S1, uh |Γ + vh − u0 = 0 if d = 2}. (12.147)
Note that vh ≥ 0, once the nodal values of vh are ≥ 0. Therefore we have Dh ⊂ D.

With the approximation Sh as in (12.55) of S the primal FE-BE coupling method
(SPh ) reads: Find (ûh , v̂h ) ∈ Dh such that
Ah (ûh , v̂h )(uh − ûh , vh − v̂h ) ≥ λh (uh − ûh , vh − v̂h ) (12.148)
for all (uh , vh ) ∈ Dh , where
Ah (uh , vh )(rh , sh ) := DG(uh , rh ) + Sh (uh |Γ + vh ), rh |Γ + sh (12.149)
and
λh (uh , vh ) := L(uh , uh |Γ + vh ) + Sh u0 , uh |Γ + vh . (12.150)
with the discrete Steklov-Poincaré operator Sh (12.79).

There holds the following a priori error estimate for the solutions (û, v̂)
of (12.145) and (uˆh , vˆh ) of (12.148) with a positive constant C, independent of
h, for h < h0 , for some h0 > 0 ,
û − uˆh , v̂ − vˆh 2H 1 (Ω)×H̃ 1/2 (Γ ) ≤ C{ inf û − uh 2H 1 (Ω)
s uh ∈Hh1
+ inf v̂ − vh 2H̃ 1/2 (Γ ) + v̂ − vh L2 (Γ )

1/2 s
vh ∈H̃h
−1/2 2
+ dist V −1 (I − K)(û + v̂ − u0 ), Hh ) }
H −1/2 (Γ )
This error estimate shows that the solution (ûh , v̂h ) ∈ Dh of (SPh ) converges for
h → 0 towards the solution (û, v̂) ∈ D of (SP ).
In [189] we investigate an adaptive FE/BE procedure for scalar nonlinear
interface problems involving friction, where the nonlinear uniformly monotone
operator such as the p-Laplacian is coupled to the linear Laplace equation on

the exterior domain. The procedure is again to reduce the contact problem to a
boundary/domain variational inequality.
12.5.2 Dual Mixed Method
Now we consider again the Signorini problem (12.138)—(12.142) with ≡ 1

and present from [188] a dual mixed variational formulation in terms of a convex
minimization problem and an associated variational inequality.
In [188] a coupling method is proposed and analyzed for dual mixed finite
elements and boundary elements for (12.138)–(12.142) using the inverse Steklov-
Poincaré operator R, the Neumann-to-Dirichlet (NtD) map, given by
1
R := S −1 = − [V + (I + K)W −1 (I + K )] : H −1/2(Γ )] → H 1/2(Γ ) .
2
(12.151)
Define Ψ̃ : H (div; Ω) → R ∪ {∞} by
1 1 1
Ψ̃ (q) := q2[L2 (Ω)]d + q · n, R(q · n) − q · n, R(t0 ) + 2u0 , (12.152)
2 4 2
and the subset of admissible functions by
D̃ := {q ∈ H (div; Ω) : q · n ≤ 0 on Γs , − div q = f in Ω} .
K ) consists in finding q D ∈ D̃
Then the uniquely solvable dual formulation (SP
such that
Ψ̃ (q D ) = min Ψ̃ (q) . (12.153)

q∈D̃
As shown in [188] problem (SP K ) is equivalent to the original Signorini contact

problem (12.138) — (12.142) with ≡ 1.
Next a saddle point formulation (M) of (SPK ) is given with the help of H :
H (div; Ω) × L2 (Ω) × H̃ 1/2(Γs ) → R ∪ {∞} defined as

H (p, v, μ) := Ψ̃ (p) + v div p dx + f v dx + p · n, μ Γs (12.154)
Ω Ω
for all (p, v, μ) ∈ H (div; Ω) × L2 (Ω) × H̃ 1/2 (Γs ), and consider the subset of
admissible functions
1/2
H̃+ (Γs ) := {μ ∈ H̃ 1/2(Γs ) : μ ≥ 0} . (12.155)
The saddle point problem (M) reads:

1/2
Find (q̂, û, λ̂) ∈ H (div; Ω) × L2 (Ω) × H̃+ (Γs ) such that
1/2
H (q̂, u, λ) ≤ H (q̂, û, λ̂) ≤ H (q, û, λ̂) ∀ (q, u, λ) ∈ H (div; Ω) × L2 (Ω) × H̃+ (Γs ) ,
(12.156)
1/2
which is equivalent to finding a solution (q̂, û, λ̂) ∈ H (div; Ω)×L2(Ω)× H̃+ (Γs )
of the variational inequality:
a(q̂, q) + b(q, û) + d(q, λ̂) = q · n, r ∀ q ∈ H (div; Ω) , (12.157)

b(q̂, u) = − f u dx ∀ u ∈ L2 (Ω) , (12.158)
Ω
1/2
d(q̂, λ − λ̂) ≤ 0 ∀ λ ∈ H̃+ (Γs ) , (12.159)
where r = R(t0 ) + 2u0 and

a(p, q) = 2 p · q dx + q · n, R(p · n) ∀ p, q ∈ H (div; Ω) , (12.160)
Ω

b(q, u) = u div q dx ∀ (q, u) ∈ H (div; Ω) × L2 (Ω) , (12.161)
Ω
d(q, λ) = q · n, λ Γs ∀ (q, λ) ∈ H (div; Ω) × H̃ 1/2(Γs ) , (12.162)
K ) and the saddle point problem

The connection between the dual problem (SP
(M) is as follows.
Theorem 12.18 ([188]) The dual problem (SP K ) is equivalent to the mixed dual
variational inequality (M). More precisely:
1/2
(i) If (q̂, û, λ̂) ∈ H (div; Ω) × L2 (Ω) × H̃+ (Γs ) is a saddle point of H in
1/2
H (div; Ω) × L2 (Ω) × H̃+ (Γs ), then q̂ = ∇ û, û = 12 R(t0 − q̂ · n) + u0
on Γt , λ̂ = − 12 R(q̂ · n − t0 ) + u0 − û on Γs , and q̂ ∈ D̃ is the solution of
problem (SP K ).
K ), and define λ̂ := − 1 R(q D · n − t0 ) + u0 −
(ii) Let q D ∈ D̃ be the solution of (SP 2
û on Γ , where û ∈ H 1 (Ω) is the unique solution of the Neumann problem:
∂ û
−Δû = f in Ω, = q D ·n on Γ , such that μ, û + 12 R(q D ·n−t0 )−u0 ≥ 0
∂n
for all μ ∈ H −1/2(Γ ) with μ ≤ −q D · n on Γs . Then, (q D , û, λ̂) is a saddle
1/2
point of H in H (div; Ω) × L2 (Ω) × H̃+ (Γs ).
K ) is solved approximately by using mixed finite
In [188], the problem (SP
elements in Ω and boundary elements on Γ choosing finite-dimensional subspaces
−1/2
Lh ×Hh ×Hh ×Hh ×H of L2 (Ω)×H (div; Ω)×H −1/2(Γ )×H 1/2(Γ )/R×
1/2 1/2
s,h̃
H̃ 1/2(Γs ).
1/2
The subspaces (Lh , H ) and Hh are supposed to verify the usual discrete
s,h̃
Babuška-Brezzi condition, which means that there exists β ∗ > 0 such that
B(qh , (uh , λh̃ ))

inf sup ≥ β∗ . (12.163)
(uh ,λ )∈Lh ×H
1/2 q ∈H
h h
qh H (div;Ω) (uh , λh̃ )L2 (Ω)×H=1/2 (Γs )
h̃ s,h̃ q =0
h
(uh ,λh̃ )=0
where B(q, (u, λ)) = b(q, u) + d(q, λ).

−1/2
Now, for h, h̃ ∈ I let jh : Hh *→ H (div; Ω), kh : Hh *→ H −1/2(Γ )
1/2
and lh : Hh *→ H 1/2(Γ )/R denote the canonical imbeddings with their
corresponding duals jh∗ , kh∗ and lh∗ .
In order to approximate R define the discrete operators
Rh := jh∗ γ ∗ Rγjh , R̃h := jh∗ γ ∗ V γjh +jh∗ γ ∗ (I +K)lh (lh∗ W lh )−1 lh∗ (I +K )γjh ,
where γ : H (div; Ω) → H −1/2(Γ ) is the trace operator yielding the normal

component of functions in H (div; Ω).
Note that the computation of R̃h requires the numerical solution of a linear
system with a symmetric positive definite matrix Wh := lh∗ W lh . In general, there
holds R̃h = Rh because R̃h is a Schur complement of matrices from discretization
while Rh is a discretized Schur complement of operators.
In order to approximate the solution of problem (M), the authors consider in
[188] the nonconforming Galerkin scheme (Mh ):
1/2
Find (q̂h , ûh , λ̂h̃ ) ∈ Hh × Lh × H such that
s,+,h̃
ah (q̂h , qh ) + b(qh , ûh ) + d(qh , λ̂h̃ ) = qh · n, rh ∀ qh ∈ Hh , (12.164)

b(q̂h , uh ) = − f uh dx ∀ uh ∈ Lh , (12.165)
Ω
1/2
d(q̂h , λh̃ − λ̂h̃ ) ≤ 0 ∀ λh̃ ∈ H , (12.166)
s,+,h̃
where
1/2 1/2
H := {μ ∈ H : μ ≥ 0} , (12.167)
s,+,h̃ s,h̃

ah (p, q) = 2 p · q dx + q · n, R̃h (p · n) ∀ p, q ∈ Hh , (12.168)
Ω

b(q, u) = u div q dx ∀ (q, u) ∈ Hh × Lh , (12.169)
Ω
1/2
d(q, λ) = q · n, λ Γs ∀ (q, λ) ∈ Hh × H , (12.170)
s,+,h̃
and
rh := kh∗ ((V + (I + K)lh (lh∗ W lh )−1 lh∗ (I + K ))t0 + 2u0 ) .
Note that the nonconformity of problem (Mh ) arises from the bilinear form ah (·, ·)
approximating a(·, ·).
There holds the following a priori error estimate (see [188]) yielding convergence
for the solution of the nonconforming Galerkin scheme (Mh ) to the weak solution
of (M) and therefore to the weak solution of the original Signorini contact problem
due to the equivalence result of Theorem 12.18.
Theorem 12.19 ([188]) Let (q̂, û, λ̂) and (q̂h , ûh , λ̂h̃ ) be the solutions of problems
(M) and (Mh ), respectively. Define φ̂ := W −1 (I + K )(q̂ · n) and φ0 := W −1 (I +
K )t0 . Then there exists c > 0, independent of h and h̃, such that the following Cea
type estimate holds
q̂ − q̂h H (div;Ω) + û − ûh L2 (Ω) + λ̂ − λ̂h̃ H̃ 1/2 (Γs )

1/2
≤ c inf q̂ − qh H (div;Ω) + inf û − uh L2 (Ω) + inf λ̂ − λh̃
qh ∈Hh uh ∈Lh 1/2 H̃ 1/2 (Γs )
λh̃ ∈H
s,+,h̃
"
+ inf φ̂ − φh H 1/2 (Γ )/R + inf φ0 − φh H 1/2 (Γ )/R . (12.171)
1/2 1/2
φh ∈Hh φh ∈Hh
A suitable choice for finite element and boundary element spaces are Lh the set
of piecewise constant functions, Hh the space of H (div; Ω) conforming Raviart-
1/2
Thomas elements of order zero and H the set of continuous piecewise linear,
s,+,h̃
nonnegative functions of the partition τh̃ of Γs (see [188] for details).
12.6 Coupling of Primal-Mixed FEM and BEM for Plane

Elasticity
Here we report on the solution procedure in [66], where a Stokes-type mixed finite
element method with the pressure as the secondary unknown is employed (with
the displacement as the primary unknown). In the BEM domain linear elasticity
is considered. In the FEM domain an incompressible nonlinear elastic material
(governed by a uniformly monotone operator) is assumed. We present from [66]
the proofs of existence and uniqueness of the solution and the quasi optimal
convergence of a Galerkin method. Finally, we cite from [66] an a posteriori error
estimator of explicit residual type.
12.6 A Primal-Mixed FEM/BEM for Plane Elasticity 507
ΓN
Ω
ΓD Ω
BEM n n
FEM
Fig. 12.3 Geometrical setting [66]
Let Ω̃ ⊂ Rd , d = 2, 3, be a domain with Lipschitz continuous boundary. Ω̃

is partitioned as Ω̃ = Ω ∪ Γ ∪ Ω, Ω ∩ Ω = ∅ (Fig. 12.3). On the bounded
subdomain Ω we will use a finite element method whereas for Ω we will exploit
boundary integral equations such that only the boundary of Ω will be discretized.
For simplicity we assume that all of the boundary ∂Ω belongs to the coupling
interface Γ , which is also assumed to be Lipschitz continuous.
In Ω the unknowns are the displacement
u ∈ HD1 (Ω) := {v ∈ [H 1 (Ω)]d : v|ΓD = 0}
and the pressure
p ∈ L2 (Ω).
H 1 (Ω) is the usual Sobolev space with the norm v1,Ω = (v20,Ω +
grad v20,Ω )1/2 and ·0,Ω denotes the norm in L2 (Ω).
We seek (u, p) ∈ HD1 (Ω) × L2 (Ω) such that
Ω {A( (u)) : (v) + p div v}dx = Γ φ · vds + L(v) ∀v ∈ HD1 (Ω)

(12.172)
Ω q div udx = 0 ∀q ∈ L2 (Ω).
The linearized strain is (v) := 12 (grad v + (grad v)T ), and we use the notation σ :
= ij σij ij . The possibly nonlinear operator A : [L2 (Ω)]d×dsym → [L (Ω)]sym
2 d×d
yields a symmetric tensor field. The exterior loads are

L(v) := f · vdx + g · vds.
Ω ΓN
with a body force density f ∈ L2 (Ω) and a surface traction g ∈ L2 (ΓN ). In the
coupling method, the interface traction φ will be an unknown.
The operator A is assumed to be uniformly monotone and Lipschitz continuous,
i.e., there exist positive constants α and M such that for all , η ∈ [L2 (Ω)]d×d
sym

(A( ) − A(η)) : ( − η)dx ≥ α − η2L2
Ω
A( ) − A(η)L2 ≤ M − ηL2 .
With the stress σ (u, p) := A( (u)) + pI (I denotes the d × d unit matrix), the
corresponding strong form is
− div σ (u, p) = f in Ω
div u = 0 in Ω
σ (u, p)n = g on ΓN (12.173)
σ (u, p)n = φ on Γ
u = 0 on ΓD .
In the BEM-domain we consider linear elasticity. The strong form is
− div(2μ (u) + λ[div u]I ) = 0 in Ω

(12.174)
(2μ (u) + λ[div u]I )n = −φ on Γ.
We assume that the Lamé coefficients λ and μ are constant on Ω . For d = 3 we

allow Ω to be unbounded (Fig. 12.4) and in this case require the decay condition
∂u(x)
u(x) = O(1/|x|) and = O(1/|x|2), j = 1, . . . , d, for |x| → ∞.
∂xj
(12.175)
Γ
ΓN
Ω Ω
n n
BEM FEM
ΓD
Fig. 12.4 Notation for coupling with unbounded exterior domain Ω [66]
At any point x ∈ Ω , the displacement field can be represented by the Betti

formula

T
u(x) = − G(x, y) Ty u(y) dsy + Ty G(x, y) u(y) dsy .
Γ Γ
Here Ty u(y) = (2μ (u(y)) + λ[div u(y)]I ) n(y) is the traction corresponding to u
at a point y ∈ Γ , and the columns of Ty G(x, y) are the tractions of G(x, y) at y.
G(x, y) is the fundamental solution and equals
"
λ + 3μ 1 λ + μ (x − y)(x − y)T
ln I+ if d = 2,
4πμ(λ + 2μ) |x − y| λ + 3μ |x − y|2
"
λ + 3μ 1 λ + μ (x − y)(x − y)T
I+ if d = 3.
8πμ(λ + 2μ) |x − y| λ + 3μ |x − y|3
Letting x → Γ we obtain with the classical jump relations the boundary integral
equation
u = −V φ + Ku (12.176)
with φ(y) = Ty u(y) and the integral operators

(V φ)(x) = 2 G(x, y) φ(y) dsy , x ∈ Γ,
Γ

T
(Ku)(x) = 2 Ty G(x, y) u(y) dsy , x∈Γ.
Γ
Applying the traction operator Tx we get another boundary integral equation
φ = −K φ − W u (12.177)
where

(K φ)(x) = 2 Tx G(x, y) φ(y) dsy , x∈Γ ,
Γ

T
(W u)(x) = −2Tx Ty G(x, y) u(y) dsy , x∈Γ .
Γ
For d = 3, V is positive definite, i.e. there is a constant C > 0 such that
φ, V φ ≥ Cφ2−1/2,Γ ∀φ ∈ H −1/2(Γ )
For d = 2, V is positive definite when restricted to

−1/2
H0 (Γ ) := {φ ∈ H −1/2 (Γ ) : φds = 0}.
Γ
The operator W has the kernel ker W = ker |Γ , i.e., the kernel consists of the
(linearized) rigid body motions. W is positive definite on H 1/2(Γ )/ ker . For proofs
of these properties we refer to [136]. In the sequel we will use the notation
%
−1/2 H −1/2 (Γ ) if d = 3
H := −1/2
H0 (Γ ) if d = 2.
To obtain the coupled formulation we rewrite the integral equation (12.177) as
φ = (I − K )φ − W u
and insert this into the right-hand side of (12.172). The first integral equa-
tion (12.176) is weighted by a function ψ ∈ H −1/2 . Further we assume u to
be continuous across the interface Γ . Hence our method reads: Find (u, φ, p) ∈
HD1 (Ω) × H −1/2 × L2 (Ω) such that for all (v, ψ, q) ∈ HD1 (Ω) × H −1/2 × L2 (Ω)
2A (u, φ; v, ψ) + b(p, v) = L(v)

1
(12.178)
b(q, u) = 0,
where

A (u, φ; v, ψ) := 2 A( (u)) : (v) dx + v, W u − v, (I − K )φ
Ω
− ψ, V φ − ψ, (I − K)u

b(p, v) := p div v dx.
Ω
It is well known that the divergence operator div : HD1 (Ω) → L2 (Ω) is
surjective. The proof can be performed similarly to [196] taking into account that
∂Ω \ ΓD is of positive surface measure. The surjectivity is equivalent to the inf–sup
condition: There is a constant β > 0 such that
b(q, v)
inf sup ≥ β. (12.179)
q∈L2 (Ω) v∈H 1 (Ω) q0,Ω v1,Ω
D
Theorem 12.20 Let ΓD = ∅. Let A be uniformly monotone and Lipschitz

continuous. Then (12.178) has a unique solution.
Proof Let us consider the following problem: Find u ∈ ker B := {v ∈ HD1 (Ω) :
b(q, v) = 0 ∀q ∈ L2 (Ω)} and φ ∈ H −1/2 such that
A (u, φ; v, ψ) = L(v) ∀(v, ψ) ∈ ker B × H −1/2. (12.180)

For all (w, χ), (v, ψ) ∈ ker B × H −1/2 there holds
A (w, χ; w − v, −(χ − ψ)) − A (v, ψ; w − v, −(χ − ψ))

= 2{A( (w)) − A( (v))} : (w − v)dx
Ω
+ w − v, W (w − v) + χ − ψ, V (χ − ψ)
≥ αw − v21,Ω + Cχ − ψ2−1/2,Γ . (12.181)
Thus A corresponds to a nonlinear operator which maps ker B × H −1/2 onto its
dual and is uniformly monotone and Lipschitz continuous. Hence the main theorem
on monotone operators[437] implies that (12.180) has a unique solution (u, φ).
Next we show existence of p. It suffices to find a p ∈ L2 (Ω) such that
b(p, v) = L(v) − A (u, φ; v, 0) ∀v ∈ HD1 (Ω). (12.182)
The right-hand side is a continuous linear functional in v that due to (12.180) lies
in (ker B)0 , i.e., it vanishes on ker B. Concerning the left-hand side, note that the
operator

L2 (Ω) → [HD1 (Ω)]
B :
p → b(p, .)
is the adjoint of the divergence operator. Since (B = (ker B)0 , we conclude

that (12.182) has a solution. (12.182) implies that the first equation of (12.178)
is satisfied with (u, φ) being the solution of (12.180). To show uniqueness, let
(u, φ, p) and (ũ, φ̃, p̃) be solutions of (12.178). Then for all (v, ψ, q) ∈ HD1 (Ω) ×
H −1/2 × L2 (Ω)
A (u, φ; v, ψ) − A (ũ, φ̃; v, ψ) + b(p − p̃, v) + b(q, u − ũ) = 0.
Choosing v = u − ũ, ψ = φ − φ̃ and q = −(p − p̃) and exploiting uniform

monotonicity we conclude u = ũ and φ = φ̃. Now we have ker B = {0} and thus
p is unique.

We will use a Galerkin method with finite-dimensional subspaces Hh1 ⊂ HD1 (Ω),
−1/2 −1/2
Hh ⊂ H −1/2 and Hh0 ⊂ L2 (Ω): Find (uh , φh , ph ) ∈ Hh1 × Hh × Hh0 such
−1/2
that for all (vh , ψh , qh ) ∈ Hh1 × Hh × Hh0
A (uh , φh ; vh , ψh ) + b(ph , vh ) = L(vh )

(12.183)
b(qh , uh ) = 0.
Theorem 12.21 Let ΓD = ∅. Let A be uniformly monotone and Lipschitz

continuous. Let the discrete spaces Hh1 and Hh0 satisfy the inf–sup condition: There
is an h-independent constant β > 0 such that
b(qh , vh )
inf sup ≥ β. (12.184)
qh ∈Hh0 vh ∈H 1 qh 0,Ω vh 1,Ω
h
Then (12.183) has a unique solution which converges quasioptimally:
u − uh 1,Ω + φ − φh −1/2,Γ + p − ph 0,Ω

% ,
≤C inf u − vh 1,Ω + inf φ − ψh −1/2,Γ + inf p − qh 0,Ω .
vh ∈Hh1 −1/2 qh ∈Hh0
ψh ∈Hh
(12.185)
Proof Theorem 12.20 also holds for finite-dimensional subspaces. This establishes
unique solvability of (12.183).
We follow the general theory in [65]. Let ker Bh := {vh ∈ Hh1 : b(qh, vh ) =
0 ∀qh ∈ Hh0 }. Clearly uh ∈ ker Bh . As in (12.181), it can be shown that for all
−1/2
(wh , χh ) ∈ ker Bh × Hh
αw − v21,Ω + Cχ − ψ2−1/2,Γ

≤ A (wh , χh ; wh − uh , φh − χh ) − A (uh , φh ; wh − uh , φh − χh ). (12.186)
For the last term on the right-hand side we exploit the Galerkin orthogonality
A (uh , φh ; wh − uh , φh − χh ) = A (u, φ; wh − uh , φh − χh ) + b(p − ph , wh − uh ).
Since wh − uh ∈ ker Bh we have
b(p − ph , wh − uh ) = b(p − qh , wh − uh ) ∀qh ∈ Hh0 .
Now we apply the Cauchy–Schwarz inequality, the Lipschitz continuity and the
trace theorem to bound (12.186) further by
C(wh −uh 1,Ω +χh −φh −1/2,Γ )(wh −u1,Ω +χh −φ−1/2,Γ +p−qh 0,Ω ).
The triangle inequality yields
u − uh 1,Ω + φ − φh −1/2,Γ ≤ C{u − wh 1,Ω + φ − χh −1/2,Γ + p − qh 0,Ω }

(12.187)
−1/2
for all wh ∈ ker Bh , χh ∈ Hh and qh ∈ Hh0 .
The inf–sup condition (12.184) is equivalent to[65]
b(qh , vh )
sup ≥ β inf vh − wh 1,Ω = βvh H 1 / ker Bh
qh ∈Hh0 qh 0,Ω wh ∈ker Bh h
for all vh ∈ Hh1 . The equation b(q, u) = 0 ∀q ∈ L2 (Ω) implies that for every
vh ∈ Hh1 there is a wh ∈ ker Bh such that
b(qh , vh − u)
βvh − wh 1,Ω ≤ sup ≤ Cu − vh 1,Ω ,
qh ∈Hh0 qh 0,Ω
and thus
u − wh 1,Ω ≤ u − vh 1,Ω + vh − wh 1,Ω ≤ Cu − vh 1,Ω .
Therefore
inf u − wh 1,Ω ≤ C inf u − vh 1,Ω .

wh ∈ker Bh vh ∈Hh1
Together with (12.187) this implies the bound on u − uh 1,Ω + φ − φh −1/2,Γ

as claimed in the theorem.
For arbitrary qh ∈ Hh0 the inf–sup condition (12.184) implies
b(qh − ph , vh )
βqh − ph 0,Ω ≤ sup .
vh ∈Hh1 vh 1,Ω
Exploiting the Galerkin orthogonality

b(qh − ph , vh ) = − 2{A( (u)) − A( (uh ))} : (vh ) dx − vh , W (u − uh )
Ω
+ vh , (I − K )(φ − φh ) + b(qh − p, vh ) ∀vh ∈ Hh1
and using the Cauchy–Schwarz inequality, the Lipschitz continuity and the trace
theorem we obtain
βqh − ph 0,Ω ≤ C{u − uh 1,Ω + φ − φh −1/2,Γ + p − qh 0,Ω }.
Now the triangle inequality yields the bound on p − ph 0,Ω .

Remark 12.3 Finite element spaces that satisfy (12.184) are well known[65, 196].
A convenient choice is the quadrilateral or hexahedral Qk /Pk−1 element: uh is
continuous and a polynomial of degree k in each coordinate whereas ph is
discontinuous and a polynomial of total degree k − 1. For k ≥ 2, (12.184) is
fulfilled.
In this section we make a slightly stronger hypothesis on the material law by

assuming hyperelasticity:
∂Ψ ( (u))
A( (u))(x) := (x) ∀x ∈ Ω (12.188)
∂
with some stored energy function Ψ . We require that the functional

Π̃(v) := Ψ ( (v)) dx (12.189)
Ω
has continuous second-order Gâteaux derivatives and there exist positive constants
α and C such that
αv21,Ω ≤ D 2 Π̃ (u)(v, v) ≤ Cv21,Ω ∀u, v ∈ HD1 (Ω). (12.190)
This implies uniform monotonicity and Lipschitz continuity of the Fréchet deriva-
tive

D Π̃ (u) = A( (u)) dx.
Ω
For the discretization we assume that the spaces Hh1 and Hh0 consist of piecewise
−1/2
polynomial functions on a triangulation Ω̄ = ∪{T : T ∈ Th } and Hh consists
of piecewise polynomial functions on a partition of Γ . The elements T typically are
closed triangles, quadrilaterals or (in R3 ) tetrahedra. For T = T , T ∩ T is either
empty or a common vertex or edge or side. (In R2 sides coincide with edges.)
We assume the following approximation property to hold. Given an element T ∈
Th with diameter hT , let T̃ := ∪{T ∈ Th : T ∩ T = ∅}. Let S be a side of T with
diameter hS . Then for every v ∈ HD1 (Ω) there exists a vh ∈ Hh1 such that
v − vh 0,T ≤ ChT v1,T̃ (12.191)

1/2
v − vh 0,S ≤ ChS v1,T˜ (12.192)
with the constant C being independent of T .

For families of partitions into triangles with interior angles being uniformly
bounded from below, (12.191) and (12.192) follow from the analysis in [110].
For each element T with exterior unit normal nT we define the jump in the
computed traction across the side S ⊂ ∂T as
⎧1
⎨ 2 {σ (uh , ph )|T ∩S − σ (uh , ph )|T ∩S }nT if S = T ∩ T
[[[[σh nT ]] := σ (uh , ph )nT − g if S ⊂ ΓN .
⎩
σ (uh , ph )nT + W uh − ( 12 I − K )φh if S ⊂ Γ
12.7 Strongly Nonlinear Interface Problems 515
Note that [[σh nT ]] is well defined on Γ since uh |Γ ∈ H 1 (Γ ) and φh ∈ L2 (Γ ) and

thus W uh − ( 12 I − K )φh ∈ L2 (Γ ).
For the estimator we need the following residual quantities:
RT(1) := hT f + div σ (uh , ph )0,T
RT(2) := div uh 0,T

1/2
RS(3) := hS [[σh n]]0,S
1
R (4) := V φh + ( I − K)uh 1/2,Γ
2
In [66] the following result is shown.
Theorem 12.22 Let (12.190) be satisfied. Let (uh , φh, ph) be the solution
of (12.183). Then there holds the a posteriori estimate
u − uh 1,Ω + φ − φh −1/2,Γ + p − ph 0,Ω

⎧ ⎫
⎨ & ' ⎬
(1) (2) (3)
≤C RT + RT + RS + R (4)
⎩ ⎭
T ∈Th S⊂∂T \ΓD
where C is independent of uh if the mesh is sufficiently fine.

The estimator of Theorem 12.22 is similar to that of [83] for the standard
(2)
FE/BE coupling. Here the additional term RT provides for the incompressibility
constraint. For the coupling of dual-mixed finite elements and boundary elements
applied to elasticity see [187].
For further reading see [81] and [306] for coupling of mixed finite elements
and boundary element as well as see [80] for coupling with nonconforming finite
elements. In recent year a strong research on coupling with discontinuous (DG)
elements has been developped , see e.g. [237].
12.7 Adaptive FE/BE Coupling for Strongly Nonlinear

Interface Problems with Tresca Friction
Now we consider again equations (12.138)–(12.141) and write u1 = u|Ω and

u1 = u|Ω c . For given g ∈ L2 (Γs ) we set with the contact conditions on the contact
boundary Γs
− (|∇u1 |)∂ν u1 (u0 + u2 − u1 ) + g|(u0 + u2 − u1 )| =0, (12.193)

| (|∇u1 |)∂ν u1 | ≤g on Γs
with interface ∂Ω = Γs ∪ Γt with disjoint components Γs and Γt = ∅. Here (t)

denotes a function ∈ C(0, ∞) satisfying for p ≥ 2
∗
0 ≤ (t) ≤ [t δ (1 + t)1−δ ]p−2 ,
∗
| (t)t − (s)s| ≤ [(t + s)δ (1 + t + s)1−δ ]p−2 |t − s| (12.194)
and
(t)t − (s)s ≥ ∗ [(t + s)δ (1 + t + s)1−δ ]p−2 (t − s)
for all t ≥ s > 0 uniformly in x ∈ Ω ( with fixed δ ∈ [0, 1], ∗ , ∗ > 0).
In [189] the above nonlinear interface problemis reformulated as the variational
inequality: Find (û, v̂) ∈ Xp = W 1,p (Ω) × W = 12 ,2 (Γs ), W
= 12 ,2 (Γs ) = {u ∈
H 2 (∂Ω) : supp u ⊂ Γ¯s }, such that
1
G û, u − û + S(û|∂Ω + v̂), (u − û)|∂Ω + v − v̂ + j (v) − j (v̂) ≥ λ(u − û, v − v̂)
(12.195)
1
for all (u, v) ∈ Xp with the Steklov-Poincaré operator S : W 2 ,2 (∂Ω) →
1
W − 2 ,2 (∂Ω) from (12.51). Here λ(u, v) = t0 + Su0 , u|∂Ω + v + Ω f u,
j (v) = Γs g|v| for v ∈ L1 (Γs ), G u, v = Ω (|∇u|)∇u · ∇v for u, v ∈

W 1,p (Ω). G is strictly convex and G : W 1,p (Ω) → W 1,p (Ω) bounded and
uniformly monotone, hence coercive, with respect to the seminorm | · |1,p . The
variational inequality (12.195) is uniquely solvable and equivalent to the original
problem (12.193).
In order to avoid using S = W + (1 − K )V −1 (1 − K ) explicitly, the numerical
implementation involves a variant of the variational inequality (12.195) in terms of
1
the layer potentials: Find (û, v̂, φ̂) ∈ Xp × W − 2 ,2 (∂Ω) =: Y p , such that
G û, u − û + W (û|∂Ω + v̂) + (K − 1)φ̂, (u − û)|∂Ω + v − v̂

+ j (v) − j (v̂) ≥ t0 + W u0 , (u − û)|∂Ω + v − v̂ + f (u − û),
Ω
φ, V φ̂ + (1 − K )(û|∂Ω + v̂) = φ, (1 − K )u0
for all (u, v, φ) ∈ Y p . In short,
B(û, v̂, φ̂; u − û, v − v̂, φ − φ̂) + j (v) − j (v̂) ≥ Λ(u − û, v − v̂, φ − φ̂)
12.7 Strongly Nonlinear Interface Problems 517
with
B(u, v, φ; ū, v̄, φ̄) = G u, ū + W (u|∂Ω + v) + (K − 1)φ, ū|∂Ω + v̄

+φ̄, V φ + (1 − K )(u|∂Ω + v) ,

Λ(u, v, φ) = t0 + W u0 , u|∂Ω + v + f u + φ, (1 − K )u0 .
Ω
1,p
Let {Th }h∈I a regular triangulation of Ω. Let Wh (Ω) ⊂ W 1,p (Ω) the space
1
,2
of functions whose restrictions to any K ∈ Th are linear. Wh2 (∂Ω) denotes the
1
= 2 ,2 (Γs ) the subspace of
corresponding space of piecewise linear functions, and Wh
− 1 ,2 1
those supported on Γs . Finally, Wh 2 (∂Ω) ⊂ W − 2 ,2 (∂Ω) the space of piecewise
constant functions on the boundary mesh.
p 1,p
Then the discretized variational inequality reads with Xh = Wh (Ω) ×
= 2 ,2 (Γs ): Find (ûh , v̂h , φ̂h ) ∈ Y p = Xp × W − 2 ,2 (∂Ω) such that for all
1 1
W h h h h
p
(uh , vh , φh ) ∈ Yh :
B(ûh , v̂h , φ̂h ; uh − ûh , vh − v̂h , φh − φ̂h )+j (vh )−j (v̂h ) ≥ Λ(uh − ûh , vh − v̂h , φh − φ̂h ).
There holds a Céa type a priori error estimate for the solutions (û, v̂, φ̂) ∈ Y p ,
p
(ûh , v̂h , φ̂h ) ∈ Yh be the solutions of the continuous resp. discretized variational
problem, uniformly in h < h0 :
p
û− ûh , v̂ − v̂h , φ̂ − φ̂h Y p inf p
û−uh , v̂ −vh , φ̂ −φh 2Y p +v̂−vh L2 (Γs ) .
(uh ,vh ,φh )∈Yh
For adaptive error control in [189] a gradient recovery scheme in the interior with
a residual type error estimator on the boundary is given: There holds the following
a posteriori error estimate where (e, ẽ, ) denotes the error between the Galerkin
p
solution (ûh , v̂h , φ̂h ) ∈ Yh and the true solution (û, v̂, φ̂) ∈ Y p (f ∈ W 1,p (Ω)):
p
e, ẽ, Y p ηgr
2
+ ηf2 + ηS2 + η∂2 + ηg2 ,
where

2
ηgr = Gp,δ (∇ ûh , ∇ ûh − Gh ûh ),
K∈Th K

ηf2 = Gp ,1 (|∇ ûh |p−1 , hK (f − fK )),
K∈Th K
2
− 1 ,2
ηS2 = dist − 1 ,2 V −1 (1 − K)(û + v̂ − u0 ), Wh 2 (∂Ω)
W 2 (∂Ω)
p
η∂2 = ν · A (∇ ûh ) + S(ûh |∂Ω + v̂h − u0 ) − t0 1 ,p
−1+ p
W

p
ηg2 = (|σ (ûh )| − g)+ 1 + |(|σ (ûh )| − g)− ||v̂h | + (σ (ûh )v̂h )+ .
= − 2 ,2 (Γs )
W Γs Γs
where Gp,δ (x, y) = |y|2 ω(x, y)p−2 = |y|2 [(|x| + |y|)δ (1 + |x| + |y|)1−δ ]p−2
whenever |x| + |y| > 0 and 0 otherwise.
In [189] numerical experiments are presented for (12.138)–(12.141) with
(12.193) with the L-shape domain Ω as in the previous section. We set
(t) = (ε + t)p−2 , with p = 3 and ε = 0.00001, f = 0, u0 = r 2/3 sin 23 (ϕ − π2 ),
t0 = ∂ν u0 |∂Ω . The friction parameter is g = 0.5, leading to slip conditions on
the interface. To solve the variational inequality we apply the following Uzawa
algorithm with the damping parameter ρ = 25.
Algorithm 12.3 (Uzawa)
= − 2 ,2 (Γs ) : |σh (x)| ≤ 1 a.e. on Γs }.

1
(i) Choose σh0 ∈ Λh = {σh ∈ W h p
(ii) For n = 0, 1, 2, . . . find (uh , vhn ) ∈ Xh such that
n

G unh , uh + Sh (unh |∂Ω + vhn ), uh |∂Ω + vh + gσhn vh ds = λh (uh , vh )
Γs
p
for all (uh , vh ) ∈ Xh .
(iii) Set
σhn+1 = PΛ (σhn + ρgvhn ),
where for every nodal point of the mesh Th |Γs there holds δ → PΛ (δ) =
sup{−1, inf(1, δ)}.
(iv) Repeat with 2. until a convergence criterion is satisfied.
The nonlinear variational problem in the Uzawa algorithm is solved by Newton’s
method in every Uzawa-iteration step (Table 12.3).
Table 12.3 Convergence rates and Uzawa steps for uniform meshes [189]
DOF Jh (ûh , v̂h ) δJ αJ I tUzawa τ (s)
28 −0.511609 0.017249 – 2 0.190
80 −0.517938 0.010920 −0.435 2 0.640
256 −0.521857 0.007001 −0.382 2 2.440
896 −0.524293 0.004566 −0.341 2 11.05
3328 −0.525841 0.003017 −0.316 2 61.85
12800 −0.526865 0.001993 −0.308 2 437.5
50176 −0.527571 0.001287 −0.320 2 4218.0
12.8 Adaptive FE-BE Coupling for Eddy-Current 519
Here the terms arizing in the above table have the following meanings:
1
J (û, v̂) := G(û) + S(û|∂Ω + v̂), (û|∂Ω + v̂) − λ(û, v̂)
2
with its approximation Jh (ûh , v̂h ) and δJ = Jh (ûh , v̂h ) − J (û, v̂) where
t
G(u) = q(|∇u|), q(t) = sρ(s)ds.
Ω 0
Further αJ , I tUzawa and τ (s) denote the convergence rate , the number of Uzawa
iterations and the computation time, respectively ( see [189] for details).
12.8 Adaptive FE-BE Coupling for the Eddy-Current

Problem in R3
In this section we present from [282] a reliable and efficient residual based a
posteriori error estimator for the following time-harmonic eddy current problem
in R3 and furthermore we give a p-hierarchical error estimator from [283]. The
problem is discretized by edge elements inside the conductor and the exterior
region is taken into account by means of a suitable boundary integral coupling.
Given a conductor and a monochromatic exciting current, the task in eddy current
computations is to compute the resulting magnetic and electric fields, in the
conductor Ω as well as in the exterior domain ΩE , which represents air. Let
Ω ⊂ R3 be a bounded, simply connected open Lipschitz polyhedron with boundary
Γ = ∂Ω, and further set ΩE = R3 \ Ω̄. The conductor has conductivity
σ ∈ L∞ (R3 ), σ1 ≥ σ (x) ≥ σ0 > 0 and magnetic permeability μ ∈ L∞ (R3 ),
μ1 ≥ μ(x) ≥ μ0 > 0 with positive constants σ0 , σ1 , μ0 , μ1 . In ΩE , we set σ ≡ 0
and by scaling μ ≡ 1. The elementwise regularity of the material parameters reflects
the fact that Ω can consist of different conducting materials, i.e. the conductivity and
permeability can jump from one material to another. We assume a source current
J0 ∈ H(div, R3 ) with supp(J0 ) ⊂ Ω̄. Hence J · n = 0 on Γ (there is no flow of J
through Γ ).
A mathematical model of the resulting time-harmonic eddy current problem for
low frequencies (cf. Ammari, Buffa & Nédélec [1]) consists of Maxwell’s equations
curl E = −iωμH, curl H = σ E + J0 in R3 , (12.196)
the Coulomb gauge condition div E = 0 in ΩE together with the transmission

conditions
[E × n]Γ = 0, [H × n]Γ = 0, (12.197)

and the Silver-Müller radiation conditions

1 1
E(x) = O , H(x) = O | uniformly for |x| → ∞. (12.198)
|x| |x
The equations in (12.196) are just the time-harmonic Maxwell equations with
neglected displacement currents (formally setting ω = 0, where denotes the
electric permittivity). This approximation is justified in view of low frequencies ω.
Note that the second equation in (12.196) reduces to curl H = 0 in the exterior
domain ΩE . Therefore E cannot be uniquely determined in ΩE and requires
the Coulomb gauge condition. The transmission conditions (12.197) result from
requiring curl E and curl H to be in L2loc (R3 ).
In [246], Hiptmair derives an E-based coupling method for solving the prob-
lem (12.196)–(12.198) which is based on Costabel‘s symmetric coupling method
[113].The use of boundary elements for the exterior eddy current problem is not
new, we mention the early work of MacCamy & Stephan [285–288] and Nédélec
[322, 325] (see also Bossavit [58] for the eddy current problem). The unknowns
of Hiptmair’s coupled formulation, considered in this section, are u, the electrical
field E in Ω, and λ, the twisted tangential trace of the magnetic field on Γ . The
natural Sobolev space for u is H(curl, Ω), the space of L2 -fields in Ω with rotation
in L2 (Ω), and the space for λ turns out to be a trace space of H(curl, Ω). The
discretization of u uses the lowest order H(curl, Ω)-conforming finite element
space of Nédélec [321]. It is then obvious to use the corresponding trace space for
discretizing λ, which is just a generalization of the lowest order finite element space
of Raviart-Thomas on Γ . Let Ω be a simply connected polyhedron, starlike with
respect to a ball and denote the planar boundary faces by Γi , i = 1, . . . , NΓ such
* Γ
that ∂Ω = Γ = N i=1 Γi .
The complex duality pairings in Ω and on Γ will be denoted by (·, ·)Ω and
·, · Γ . We use the usual Sobolev spaces H s (Ω) for scalar functions and Hs (Ω) for
vector fields of order s ∈ R. Furthermore we use the spaces
H(curl, Ω) := {v ∈ L2 (Ω) : curl v ∈ L2 (Ω)},

H(div, Ω) : = {v ∈ L2 (Ω) : div v ∈ L2 (Ω)},
X(Ω) := H(curl, Ω) ∩ H(div, Ω),
H0 (div, Ω) := {v ∈ H(div, Ω) : v · n = 0 on ∂Ω},
−1/2 −1/2
H (divΓ , Γ ) := {ζ ∈ H (Γ ) : divΓ ζ ∈ H −1/2 (Γ )},
−1/2 −1/2
H (divΓ 0, Γ ) := {ζ ∈ H (divΓ , Γ ) : divΓ ζ = 0, ζ ∈ H −1/2 (Γ )},
−1/2 −1/2
H⊥ (curlΓ , Γ ) := {ζ ∈ H⊥ (Γ ) : curlΓ ζ ∈ H −1/2 (Γ )},
with the surface divergence operator divΓ u := − curlΓ (u × n) and the surface curl
operator curlΓ u := curl u · n, where γt× u := u × n. see also [69, 70, 246]. We
furthermore need the vectorial surface rotation for a scalar function φ defined by
−1/2
curlΓ φ := γt× (grad φ). The spaces of distributional tangential fields H (Γ )
−1/2
and H⊥ (Γ ) are introduced in [69] by duality.
In the coupling formulation we use integral operators to represent the exterior
problem. These operators are defined for x ∈ Γ as follows (for their properties see
e.g. [246] see also Chapter 4).

V (λ)(x) := γD V(λ)(x) = γD Φ(x, y)λ(y) ds(y),
Γ

K (λ)(x) := γD K(λ)(x) = γD curlx Φ(x, y)(n × λ)(y) ds(y),
Γ

KK(λ)(x) := γN V(λ)(x) = (γt× )K(λ × n)(x) = γN Φ(x, y)λ(y) ds(y),
Γ

W (λ)(x) := γN K(λ)(x) = (γt× )W(λ)(x) = γN curlx Φ(x, y)(n × λ)(y) ds(y)
Γ
with Laplace kernel Φ(x, y) = 4π|x−y|

1
and the limits γD and γN from ΩE onto Γ
of the traces γD u := n × (u × n) =: uΓ and γN u := γt× (curl u). Furthermore we
need γn u := u · n.
Following Buffa et al [67, 68, 71, 72] we introduce for 0 < s < 1 the trace spaces
Hs⊥ (Γ ) := γt× (Hs+1/2(Ω)), Hs (Γ ) := γD (Hs+1/2(Ω)).
The spaces H−s −s

⊥ (Γ ) and H (Γ ), 0 < s < 1, are then defined as the dual spaces of
Hs⊥ (Γ ) and Hs (Γ ), resp., with L2t (Γ ) := {u ∈ L2 (Γ ) : u · n = 0 a.e. on Γ } as
pivot space. For any s > 12 we define Hs− (Γ ) := {u ∈ L2t (Γ ) : u|Γj ∈ Hst (Γj ), j =
1, . . . , NΓ }, furthermore
⎧ −1/2
⎪
⎨ H (divΓ , Γ ), s = − 12 ,
H (divΓ , Γ ) := {λ ∈ Hs (Γ ), divΓ λ ∈ H s (Γ )}, − 12 < s < 12 ,
s
⎪
⎩ {λ ∈ Hs (Γ ), div λ ∈ H s (Γ )}, s > 1 ,
Γ − 2
⎧
⎪ −1/2
⎨ H⊥ (curlΓ , Γ ), s = − 12 ,
H⊥ (curlΓ , Γ ) := {λ ∈ H⊥ (Γ ), curlΓ λ ∈ H (Γ )}, − 12 < s < 12 ,
s s s
⎪
⎩ {λ ∈ Hs (Γ ), curl λ ∈ H s (Γ )}, s > 1 .
⊥ Γ − 2
The trace mappings γD and γt× can be extended to continuous mappings
γt× : Hs (curl, Ω) → H
s−1/2 s−1/2
γD : Hs (curl, Ω) → H⊥ (curlΓ , Γ ), (divΓ , Γ )
(12.199)
for all 0 ≤ s < 1, where Hs (curl, Ω) := {u ∈ Hs (Ω) : curl u ∈ Hs (Ω)}

see [68, 72].
After having collected the operators and spaces needed the coupled variational
problem for the eddy current problem introduced by Hiptmair reads as ([246].
−1/2
Find u ∈ H(curl, Ω), λ ∈ H (divΓ 0, Γ ) such that for all v ∈
−1/2
H(curl, Ω), ζ ∈ H (divΓ 0, Γ )
(μ−1 curl u, curl v)Ω + iω(σ u, v)Ω − W uΓ , vΓ Γ + KKλ, vΓ Γ = −iω(J0 , v)Ω ,

(I − K )uΓ , ζ Γ + V λ, ζ Γ = 0.
(12.200)
For brevity write (12.200) as
A (u, λ; v, ζ ) = L (v, ζ ).
The above formulation is obtained by using Green’s formula in Ω and a Stratton-

Chu representation formula for E in ΩE . The unknown u corresponds to E|Ω , and
the unknown λ on the boundary corresponds to γN E = −iωH|ΩE × n, which can
indeed be seen to be surface divergence free. Due to the transmission conditions
there holds λ = γN u. Note that the formulation (12.200) is block skew-symmetric.
As observed by Hiptmair [246], the sesquilinear form A is continuous and elliptic
−1/2
on (H(curl, Ω) × H (divΓ 0, Γ ))2 . Thus, the variational formulation (12.200)
admits a unique solution. Setting E|Ω := u, E|ΩE := curl V(n × γD E) − V(λ)
with the single layer potential V with Laplace kernel and H := iωμ 1
curl E gives a
solution to the original problem (12.196)–(12.198).
Next the eddy current problem is discretized by edge elements inside the
conductor and the exterior region is taken into account by means of a suitable
boundary integral coupling. Let Th be a regular triangulation (with tetrahedral or
hexahedral elements) of Ω and Kh = {T ∩ Γ : T ∈ Th } the induced triangulation
on Γ . For the Galerkin method we use the finite element spaces suggested in [246]
, namely the well known H(curl, Ω)-conforming finite element space ND1 (Th ) of
first kind Nédélec elements of first order [321] for discretization of the unknown
u ∈ H(curl, Ω) and RT10 (Kh ) := {λh ∈ RT1 (Kh ), divΓ λh = 0} for the
−1/2
boundary unknown λ ∈ H (divΓ 0, Γ ), where RT1 (Kh ) denotes the lowest
−1/2
order H (divΓ , Γ )-conforming finite element space of Raviart-Thomas, which
can be obtained as the image of ND1 (Th ) under the mapping γt× . Thus the Galerkin
method reads:
Find uh ∈ ND1 (Th ), λh ∈ RT10 (Kh ) such that ∀vh ∈ ND1 (Th ), ζ h ∈ RT10 (Kh )
(μ−1 curl uh , curl vh )Ω + iω(σ uh , vh )Ω

− W γD uh , γD vh Γ + KKλh , γD vh Γ = −iω(J0 , vh )Ω ,
(I − K )γD uh , ζ h Γ + V λh , ζ h Γ = 0.
(12.201)
Now the conformity of the discrete spaces and the strong ellipticity of A (·, ·)
imply that the Galerkin formulation (12.201) has a unique solution (uh , λh ) ∈
ND1 (Th ) × RT10 (Kh ).
For simplicity, let σ and μ be piecewise C ∞ . Besides the set of elements
of the interior mesh Th , we need the set of faces Fh , the set of exterior faces
FhΓ = {F ∈ Fh : F ⊂ Γ } (which coincides with the induced boundary
triangulation Kh ) and the set of interior faces FhΩ = Fh \ FhΓ . Further let hT
denote the maximal diameter of an element T ∈ Th and hF the maximal diameter of
a face F ∈ Fh . We assume shape regularity of the mesh, which in particular means
hT hT ∀T , T ∈ Th , T ∩ T = ∅ and hF hT ∀F ∈ Fh (T ), where Fh (T )
is the set of faces of the element T ∈ Th . For F ∈ FhΩ a common face of two
elements T1 , T2 and the normal n(x) pointing into T2 we define the jump [n · q]F :=
n · q|F ⊂T1 − n · q|F ⊂T2 . For F ∈ FhΓ we define [n · q]F := n · q|F . Analogously
we define the jumps [n × q]F . We assumed Γ to be simply connected. Therefore
we have RT10 (Kh ) = curlΓ S=1 (Kh ), where S1 (Kh ) denotes the finite element
space of scalar, continuous piecewise linear functions. Thus we now seek a function
ϕh ∈ S=1 (Kh ) := {ψ ∈ S1 (Kh ) : Γ ψdsx = 0} and then set λh := curlΓ (ϕh ).
−1/2
We will use the notations X := H(curl, Ω) × H (divΓ 0, Γ ) for the continuous
space of our variational problem (12.200) and Xh := ND1 (Th ) × curlΓ S=1 (Kh )
for the discrete space of the Galerkin formulation (12.201) and we define the energy
norm (v, ζ )2X := v2E + ζ 2e via
v2E := (μ−1 curl v, curl v)Ω + ω(σ v, v)Ω 0 v2H(curl,Ω) , (12.202)
ζ 2e := V ζ , ζ Γ 0 ζ 2H −1/2 (Γ ) (12.203)
−1/2
on H(curl, Ω) × H (divΓ 0, Γ ).
The following theorem gives a residual-based reliable a posteriori error estimator
for the FE-BE coupling method (12.201). Here σA and μA denote the average of σ
and μ on a face F , e.g. σA := 0.5(σT1 + σT2 ) with T1 ∩ T2 = F . We assume that σ
and μ grow only mildly on neighbouring elements.
Theorem 12.23 Let (u, λ) ∈ X and (uh , λh ) ∈ Xh denote the solutions of the
continuous resp. the discrete formulation (12.200) resp. (12.201) and let (e, ε) be
the Galerkin error, i.e. e := u − uh and ε := λ − λh . There holds the a posteriori
error estimate
1/2
(e, ε)X (η0T )2+(η1T )2+(η0F ,C )2+(η1F ,C )2+(η0F ,Γ )2+(η1F ,Γ )2+(η2F ,Γ )2
(12.204)
=: η
with
1/2 1/2
ηjT := (ηjT )2 (j = 0, 1), ηjF ,C := (ηjF,C )2 (j = 0, 1),
T ∈Th F ∈FhΩ
1/2
ηjF ,Γ := (ηjF,Γ )2 (j = 0, 1, 2) (12.205)
F ∈FhΓ
and
√ √ −1
η0T := hT ω σ (div J0 + div σ uh )0,T ,
√ √ √
η1T := hT i μ ωJ0 + i μ ωσ uh + μ curl(μ−1 curl uh )0,T ,
M √ √ −1
η0F,C := hF ω σA [σ uh · n]F 0,F ,
M √
η1F,C := hF μA [μ−1 curl uh × n]F 0,F ,
M √ √
η0F,Γ := hF ω σ uh · n0,F ,
M √ √ √
η1F,Γ := hF μ−1 curl uh × n − μ W γD uh + μ KKλh 0,F ,
M
η2F,Γ := hF curlΓ uh − curlΓ K γD uh + curlΓ V λh 0,F .
If σ, μ are constant on an element T (or on two elements with common face F ), the
error estimators can be simplified to
M
η0T = hT ωσ −1 div J0 + σ div uh 0,T ,
√
η1T = hT μ iωJ0 + iωσ uh + μ−1 curl curl uh 0,T ,
M /
η0F,C = hF ωσA−1 [σ uh · n]F 0,F ,
M √
η1F,C = hF μA [μ−1 curl uh × n]F 0,F ,
M √
η0F,Γ = hF ωσ σ uh · n0,F ,
M √
η1F,Γ = hF μ μ−1 curl uh × n − W γD uh + KKλh 0,F .
Using lowest order Nédélec elements, we even obtain:

M √
η0T = hT ωσ −1 div J0 0,T , η1T = hT ω μ J0 + σ uh 0,T .
Proof Setting e := u − uh , ε := λ − λh . The ellipticity of A yields
e2E + ε2e |A (e, ε; e, ε)| = |L (e, ε) − A (uh , λh ; e, ε)|.

But for arbitrary (eh , εh ) ∈ Xh we have the equation
A (uh , λh ; eh , εh ) = L (eh , εh ),
which we can insert in the above equation to obtain
e2E + ε2e |L (e − eh , ε − εh ) − A (uh , λh ; e − eh , ε − εh )|

= −iω(J0 + σ uh , e − eh )Ω − (μ−1 curl uh , curl(e − eh ))Ω

+W γD uh − KKλh , γD e − γD eh Γ + (K − I )γD uh − V λh , ε − εh Γ .
(12.206)
In order to analyze the error e, we decompose it into a weakly solenoidal and an

irrotational part. Using the regularized Poincaré map, as investigated in the work by
Costabel and McIntosh [123], one has for bounded domains, which are starlike with
respect to a ball, the decomposition
H(curl, Ω) = H1 (Ω) + grad H 1 (Ω). (12.207)
For this decomposition see also the proof of Lemma 5.8 in [57]. Therefore for any
v ∈ H(curl, Ω) there exist functions v⊥ ∈ H1 (Ω) and ψ ∈ H 1 (Ω)/C with v =
v⊥ + grad ψ such that there holds
|v⊥ |H1 (Ω) curl vL2 (Ω) , (12.208)
grad ψL2 (Ω) ≤ vH(curl,Ω) . (12.209)
We thus split the error term e into the two parts
e = e⊥ + grad ψ, e⊥ ∈ H1 (Ω), ψ ∈ H 1 (Ω)/C (12.210)
using (12.207). For the boundary error term ε we remark that ε = curlΓ φ for some
−1/2
φ ∈ H 1/2(Γ )/C, since ε ∈ H (divΓ 0, Γ ) (cf. [69, 70]). Next we define the
discrete functions eh and εh . We choose
eh = P1h e⊥ + grad Ph1 ψ ∈ ND1 (Th )
with e⊥ , ψ from (12.210) and the interpolation operators P1h : H1 (Ω) → ND1 (Th )
and Ph1 : H 1 (Ω) → S1 (Th ) where S1 denotes the space of continuous and
piecewise trilinear functions. On the boundary we choose φh = ph1 φ with ph1 :
H 1/2(Γ ) → S1 (Kh ) and then
εh = curlΓ φh ∈ RT10 (Kh ).

With (12.210) and the above definitions of eh and εh we obtain with (12.206) the
residual estimate

e2E + ε2e −iω(J0 +σ uh , e⊥ −P1h e⊥ )Ω − (μ−1 curl uh , curl(e⊥ −P1h e⊥ ))Ω

+ W γD uh − KKλh , γD e⊥ − γD P1h e⊥ Γ

+ −iω(J0 +σ uh , grad (ψ − Ph1 ψ))Ω + W γD uh − KKλh , grad Γ (ψ − Ph1 ψ) Γ

+ (K − I )γD uh − V λh , curlΓ (φ − ph1 φ) Γ .
(12.211)
Now integration by parts gives

(μ−1 curl uh , curl(e⊥ − P1h e⊥ ))Ω = (μ−1 curl uh , curl(e⊥ − P1h e⊥ ))T
T ∈Th

= (curl(μ−1 curl uh ), e⊥ − P1h e⊥ )T + μ−1 γN uh , γD e⊥ − γD P1h e⊥ ∂T
T ∈Th

= (curl(μ−1 curl uh ), e⊥−P1h e⊥ )Ω + [μ−1 curl uh ×n]F , γD e⊥−γD P1h e⊥ F.
F ∈Fh
(12.212)
We have used the fact that the terms μ−1 curl uh × n and γD e⊥ − γD P1h e⊥ are
in L2 (∂T ) (since uh|T is a polynomial and e⊥ , P1h e⊥ ∈ H1 (T )), such that we can
−1/2 −1/2
consider the H (divΓ , ∂T ) − H⊥ (curlΓ , ∂T )-duality ·, · ∂T as a L2 (∂T )-
duality. Furthermore we can write curl uh × n for γN uh due to the regularity of uh .
Since elementwise uh ∈ H(div) we obtain similarly
(J0 +σ uh , grad ψ −grad Ph1 ψ)Ω = −(div J0 +div σ uh , ψ −Ph1 ψ)Ω

+ [σ uh · n]F , ψ −Ph1 ψ F . (12.213)
F ∈Fh
Next, we regard the term W γD uh−KKλh , grad Γ ψ−grad Γ Ph1 ψ Γ from (12.211),
−1/2 −1/2
which constitutes a H (divΓ , Γ ) − H⊥ (curlΓ , Γ )-duality pairing (the left
−1/2 −1/2
hand side is in H (divΓ , Γ ), the right hand side is in H⊥ (curlΓ , Γ )). With
the integration by parts formula given in [69] we obtain
W γD uh − KKλh , grad Γ (ψ − Ph1 ψ) Γ = −divΓ W γD uh − divΓ KKλh , ψ −Ph1 ψ Γ.

(12.214)
−1/2
Next note that for u ∈ H(curl, ΩE ), λ ∈ H (divΓ 0, Γ ) there holds
divΓ KKλ = 0 in H −1/2(Γ ), divΓ W γD u = 0 in H −1/2(Γ ).
Thus (12.214) yields
W γD uh − KKλh , grad Γ (ψ − Ph1 ψ) Γ = 0.
The last term from (12.211) to consider is (K −I)γD uh−Vλh , curlΓ φ − curlΓ ph1 φ Γ ,
−1/2 −1/2
which is again a duality pairing between H⊥ (curlΓ , Γ ) and H ( divΓ , Γ ).
Using again the integration by parts formula from [69] we obtain
(K − I)γD uh−Vλh , curlΓ (φ −ph1 φ) Γ = curlΓ (K − I )γD uh − curlΓ V λh , φ − ph1 φ Γ.

(12.215)
Altogether we have

e2E + ε2e |(−iωJ0 − iωσ uh − curl(μ−1 curl uh ), e⊥ − P1h e⊥ )T |
T ∈Th

+ |[μ−1 curl uh × n]F , γD e⊥ − γD P1h e⊥ F|
F ∈FhΩ

+ |μ−1 curl uh × n − W γD uh + KKλh , γD e⊥ − γD P1h e⊥ F|
F ∈FhΓ

+ |ω div J0 + ω div σ uh , ψ − Ph1 ψ)T |
T ∈Th

+ |ω[σ uh · n]F , ψ − Ph1 ψ F| + |ωσ uh · n, ψ − Ph1 ψ F|
F ∈FhΩ F ∈FhΓ

+ |curlΓ (I − K )γD uh + curlΓ V λh , φ − ph1 φ F |.
F ∈FhΓ
Applying the Cauchy-Schwarz inequality and standard approximation properties

for P1h , Ph1 and ph1 yields the Theorem.

Now we present the efficiency of the residual error estimator for FE-BE coupling
of the eddy current problem (for details see [281]). We assume that the volume
mesh Th is shape-regular and that the induced boundary mesh FhΓ = Th |Γ is
quasi-uniform. We assume there holds
hΓ,max
1≤ ≤ Q(Th |Γ ) (12.216)
hΓ,min
for a certain quasi-uniformity constant Q(Th |Γ ), independent of the mesh, where

hΓ,max := max{hF , F ∈ FhΓ } and hΓ,min := min{hF , F ∈ FhΓ }.
Theorem 12.24 ([282]) Let (u, λ) ∈ X and (uh , λh ) ∈ Xh denote the solutions
of the continuous resp. the discrete formulation (12.200) resp. (12.201) and let (e, ε)
be the Galerkin error, i.e. e := u − uh and ε := λ − λh . Then there exists a constant
C > 0, depending on the quasi-uniformity constant of the boundary element mesh
Q(ThΓ ) of (12.216) and on the shape regularity constant of Th , such that there
holds for the error estimator η in (12.204) and δ > 0
3
η2 ≤ C (e, ε)2X + e0 2L2 (T ) + (osc1T )2 + u − uh 2E,T
T ∈Th T ∈TΓ
h1+2δ
+ hΓ,max u − uE 2H 1/2 (curl,Ω) + hΓ,max u − uE 2H1/2+δ (curl,Ω)
hmin
4
+ hΓ,max λ − λE 2H0 (div
Γ ,Γ )
(12.217)
with the interpolant uE := Π1h u ∈ ND1 (Th ), λE ∈ RT1 (Kh ) the orthogonal
−1/2
projection of λ with respect to the H (divΓ , Γ ) inner product. Here osc1T :=
√
hT μω(J0 −Π1h J0 )L2 (T ) denotes the oscillation term where Π1h is an interpo-
lation operator into ND1 (Th ). Furthermore, TΓ denotes the set of elements which
have at least one face on the boundary.
Example
12.1 The geometry
in this example is the L-block Ω := [−1, 1]3 \
[0, 1] ∪[0, 1] ×[−1, 0] . Here, we consider a singularity function as given current.
3 2
J0 := grad r 2/3 sin( 23 φ) in the L-block,
where r and φ are cylindrical coordinates. Hence, one expects an adaptive refine-
ment towards the re-entrant edge.
The energy norm of the unknown exact solution is extrapolated by the energy
norms on the sequence of uniform meshes. We perform an adaptive refinement
(10% of elements) using hanging nodes. The resulting meshes can be found in
Fig. 12.5 and the error in Fig. 12.6. Due of the 2/3-singularity in the interior
domain we expect a convergence rate of α = 23 with respect to the mesh size
h and a convergence rate of α = 29 with respect to the degrees of freedom.
This correspondents to the results in Table 12.4. For the adaptive refinement using
the residual error indicators we get a better convergence rate of about 0.4. The
effectivity indices are quite constant which underlines the reliability and efficiency
of the error estimator.
Fig. 12.5 The adaptive meshes (levels of refinement 7 and 9) for Example 12.1 using the residual
error estimator [282]
L-Block, f=grad (r^(2/3)sin(2/3 phi))

10
Error uni
Error adap
Est. uni
Est. adap
1
0.1
0.01
10 100 1000 10000
degrees of freedom
Fig. 12.6 Energy norm e of the Galerkin error and the residual error estimator η of Example 12.1
(L-block) [282]
Table 12.4 Values and convergence rates with respect to the total degrees of freedom DOF of
the Galerkin error e and of the residual error estimator η and the effectivity indices q := ηe for
Example 12.1 (the L-block) [282]
Uniform refinement
n DOF e α η α q
2 70 0.4186472 1.0506895 2.509725
4 334 0.2869302 0.241762 0.7324853 0.2308640 2.552834
6 902 0.2246235 0.246421 0.5789936 0.2366966 2.577618
8 1882 0.1881433 0.240962 0.4870686 0.2350675 2.588817
10 3382 0.1638018 0.236375 0.4248354 0.2332292 2.593594
12 5510 0.1462253 0.232553 0.3794070 0.2317002 2.594674
Adaptive refinement
70 0.4186472 1.0506895 2.50972537
152 0.3661693 0.172731 0.9203291 0.1708448 2.51339776
231 0.3528255 0.088695 0.8116177 0.3003362 2.30033742
362 0.2749754 0.554936 0.6961152 0.3417302 2.53155446
526 0.2319625 0.455246 0.5897867 0.4435999 2.54259503
778 0.1853135 0.573613 0.4921377 0.4624187 2.65570344
1306 0.1501191 0.406604 0.4074256 0.3646738 2.71401574
2229 0.1306073 0.260452 0.3577131 0.2434174 2.73884461
3648 0.1056281 0.430896 0.2965062 0.3809444 2.80707690
5615 0.0943108 0.262784 0.2627426 0.2803241 2.78592272
12.8.1 p-Hierarchical Estimator
In the following we present from [401] a p-hiarchical error estimator for tetrahedral
meshes. As well known, Nédélec elements on a tetrahedron T
NDk (T ) := (Pk−1 (T ))3 + {p ∈ (Pk (T ))3 : pT · x = 0} ⊂ (Pk (T ))3 ,
are determined by local degrees of freedom:

(i) u ·tq ∀q ∈ Pk−1 , e edge of T ,
e
(ii) (u × n) · q ∀q ∈ (Pk−2 )2 , F face of T ,
F
(iii) u·q ∀q ∈ (Pk−3 )3 .
T
For k = 2 we apply the stable decomposition of Nédélec edge elements from [40]
<⊥
ND2 (Th ) = ND1 (Th ) ⊕ grad S=2 (Th ) ⊕ ND 2 (Th ) (12.218)
<⊥
ND 2 (Th ) := {uh ∈ ND2 (Th ) : uh , t e = 0, ∀e edge of Th }
S=2 =S2 \ S1 (hierarchical surplus), Sk = {w ∈ C 0 : w|T ∈ Pk }.

Let (uh , λh ) ∈ ND1(Th) × curlΓ S1 (Kh) =: Xh , and (u2 , λ2 ) ∈ ND2 (Th) ×

curlΓ S=2 (Kh ) =: X2 (satisfying divΓ λ2 = 0) be Galerkin solutions of (12.201)
with exact solution (u, λ) ∈ X . We make the saturation assumption:
There exists a δ ∈ (0, 1) such that there holds
(u − u2 , λ − λ2 )X ≤ δ(u − uh , λ − λh )X . (12.219)
Theorem 12.25 ([283]) Assuming (12.219), then on a tetrahedral grid Th there

holds
1
η (u − uh , λ − λh )X η (12.220)
1−δ
with the local a posteriori estimator

M
2
N
2
m
2
η2 := Θ (ei ) + Θ (Fj ) + ϑ (ei ) ,
i=1 j =1 i=1
where
|A (uh , λh ; 0, curlΓ ϕ (e) )|

ϑ (e) := for ϕ (e) ∈ S=2 (Kh ),
curlΓ ϕ (e) e
|L (grad φ (e) , 0) − A (uh , λh ; grad φ (e) , 0)|
Θ (e) := for φ (e) ∈ S=2 (Th )
grad φ (e) E
and
Θ (F ) := κ1 b(F ) (F )
for b(F ) <⊥
1 + κ2 b2 E i ∈ ND 2 (Th ) (i = 1, 2),
where (κ1 , κ2 )T is the solution of the linear system
a(κ1 b(F ) (F ) (F ) (F ) (F )
1 + κ2 b2 , bi ) = L (bi , 0) − A (bh , λh ; bi , 0) (i = 1, 2).
and
a(u, v) := (μ−1 curl u, curl v)Ω + iω(σ u, v)Ω (12.221)
Here, M (N) denotes the number of edges (faces) in Th , m < M the number of
edges in Kh (those on Γ ).
Example 12.2 We compute the Galerkin solution (12.201) with Ω = [−1, 1]3 on
a series of uniform meshes. Setting μ = σ = ω = 1 we choose the exact solution
of (12.196)–(12.198)

1
u(x) := curl ρ(y) dy, λ = curl u × n
Ω x − y
with
2 1
ρ(x) = (1 − x12 )(1 − x22 )(1 − x32 ) x1 x2 x3 1 in Ω
1
and right hand side
J0 = −u + 4πi curl ρ in Ω and J0 = 0 in ΩE .
Figure 12.7 shows convergence of the FE/BE Galerkin coupling method (12.201) for
hexahedral and tetrahedral elements. Furthermore Fig. 12.7 shows that both residual
and hierarchical error estimators are reliable and efficient.
For further reading we recommend [247] and for the electric field integral
equation [53, 248].
e
ηhier
e
ηres
ηres
Fig. 12.7 Energy norm e of the Galerkin error (u − uh , λ − λh ), residual error estimator ηres
and reduced hierarchical error estimator ηhier for the Example. The superscript corresponds to
hexahedral elements, to tetrahedral elements
12.9 Parabolic-Elliptic Interface Problems 533
12.9 Parabolic-Elliptic Interface Problems
In [311] the authors consider the following linear parabolic - elliptic interface
problem which describes two-dimensional eddy currents in electro dynamics [289]:
Let Ω1 ⊂ R2 be a bounded simply connected polygonal domain with boundary Γ
and complement Ω2 := R2 \ Ω̄1 .
Given T > 0, λ > 0 and the functions f : Ω1 × [0, T ] → R, v0 , ψ0 : Γ ×
[0, T ] → R and q : Ω1 → R find A : [0, T ] → R and ui : Ωi × [0, T ] → R,
i = 1, 2, such that
u̇1 − Δu1 + λu1 = f in Ω1 × (0, T )
− Δu2 = 0 in Ω2 × (0, T )
u1 − u2 = v0 on Γ × (0, T ) (12.222)
∂u1 ∂u2
− = ψ0 on Γ × (0, T )
∂n ∂n
u1 (x, 0) = q(x) ∀ x ∈ Ω1
u2 (x, t) = A(t)log|x| + O( |x|

1
) for |x| → ∞ .
∂v
where ∂n is the normal derivative of v pointing from Ω1 into Ω2 .
The symmetric coupling method yields in [311] the following formulation: For
u, w ∈ H 1 (Ω1 ) and φ, ψ ∈ H −1/2 (Γ ) define the bilinear form

B(u, φ; w, ψ) = 2 (∇u·∇w+λuw) dx+W u+(K −1)φ, w +ψ, (K−1)u−V φ
Ω1
and the linear form
L(w, ψ) = 2ψ0 + W v0 , w + ψ, (K − 1)v0 + (2w, f )
where, here, ., . denotes the duality pairing between H −1/2 (Γ ) and H 1/2(Γ ). The
bilinear form B(.; .) satisfies the following Babuska-Brezzi condition:
There exists a constant β > 0 such that for all (v, ψ) ∈ H := H 1 (Ω1 ) ×
H −1/2 (Γ ):
B(v, ψ; v, −ψ) ≥ β (v, ψ)2H (12.223)
where (v, ψ)2H = v2H 1 (Ω ) + ψ2H −1/2 (Γ ) .

1
The problem (12.222) is now transformed into the following variational problem
(cf. [120]):
Given f ∈ L2 (0, T ; H 1(Ω1 ) ), v0 ∈ L2 (0, T ; H 1/2(Γ )), ψ0 ∈ L2 (0, T ; H −1/2
(Γ )) and q ∈ H 1 (Ω1 ) find u ∈ QT and φ ∈ BT such that
(u̇, w) + B(u, φ; w, ψ) = L(w, ψ) (12.224)
for all w ∈ L2 (0, T ; H 1 (Ω1 )) and ψ ∈ L2 (0, T ; H −1/2(Γ )) where
QT = {u ∈ L2 (0, T ; H 1(Ω1 )); u̇ ∈ L2 (0, T ; H 1(Ω1 ) ), u|t =0 = q} and

BT = L2 (0, T ; H −1/2(Γ )) .
For the definition of the Sobolev-spaces L2 (0, T ; X) we refer to [284]. The

existence of a unique solution (u, φ) of (12.224) is proved in [120].
Now, let (u, φ) ∈ QT × BT be the unique solution of (12.224). To obtain
a Galerkin aprroximation of (u, φ) we have to define suitable finite dimensional
subspaces of L2 (0, T ; H 1 (Ω1 )) and L2 (0, T ; H −1/2(Γ )) as follows: Let I ⊂ (0, 1)
be an indexed subset of (0, 1) with 0 ∈ I¯ and let {ΔhΩ1 ; h ∈ I } be a family of
regular triangulations of Ω1 with corresponding partitions {ΔhΓ , h ∈ I } of Γ . A
triangulation ΔΩ1 of Ω1 is called regular if
• the intersection of any two triangles of ΔΩ1 is either a common side or a common
node or empty,
• there exists a constant ρ > 0 such that ρh2Δ ≤ Δ dx for all triangles Δ ∈ ΔΩ1
with diameter hΔ ,
• there holds hΔ < 1 for all triangles Δ ∈ ΔΩ1 .
For the discretization of the space variable we need the following finite dimensional
subspaces of H 1 (Ω1 ) and H −1/2 (Γ ):
& '
Sh = vh : Ω1 → R ; vh piecewise linear on ΔhΩ1 , vh ∈ C 0 (Ω1 ) ,
3 4
S̃h = ψh : Γ → R ; ψh piecewise constant on ΔhΓ .
Let qh be the orthogonal L2 -projection of q into the space Sh , i.e. (qh , wh ) =

(q, wh ) for all wh ∈ Sh . Now, the semi-discrete Galerkin scheme reads as follows:
Find uh ∈ C 1 ([0, T ]; Sh ) and φh ∈ C 0 ([0, T ]; S̃h ) such that uh (0) = qh and
(u̇h , w) + B(uh , φh ; w, ψ) = L(w, ψ) (12.225)
for all (w, ψ) ∈ Sh × S̃h .

Choosing a finite element basis {wi }M i=1 of Sh and a boundary element
basis {ψi }mi=1 of S̃h we define U h = U h (t) = [ui (t)]i=1,...,M and Φh =
M
Φh (t) = [φi (t)]i=1,...,m where uh (x, t) = i=1 ui (t)wi (x) and φh (x, t) =
12.9 Parabolic-Elliptic Interface Problems 535
m
Hence, the Galerkin equations (12.225) are equivalent to the
i=1 φi (t)ψi (x).
following system of ordinary differential equations:
M U̇h + 2(K + λM )Uh + DUh + BΦh = F

S Φh − B . Uh = G (12.226)
where
M = [(wi , wj )]i=1,...,M
j =1,...,M K = [(∇wi , ∇wj )]i=1,...,M
j =1,...,M
D = [wi , W wj ]i=1,...,M
j =1,...,M S = [ψi , V ψj ]i=1,...,m
j =1,...,m
B = [(K − 1)psij , wi ]i=1,...,M

j =1,...,m B . = [ψi , (K − 1)wj ]i=1,...,m
j =1,...,M
F = [2ψ0 + W v0 , wi + (2wi , f )]i=1,...,M and G = [ψi , (1 − K)v0 ]i=1,...,m
For a full discretization of (12.224) we consider partitions 0 = t0 < t1 < . . . <

tn = T of the time interval [0, T ] in subintervals In = (tn−1 , tn ] of length kn = tn −
tn−1 . With each time interval In we associate a regular triangulation ΔnΩ1 = ΔhΩn1 of
Ω1 and a corresponding partition ΔnΓ of Γ . We define Shn := Shn and S̃hn := S̃hn . To
obtain a fully discrete scheme we use the discontinuous (in time) Galerkin method
with piecewise linear test and trial functions . Therefore we define the following
finite dimensional spaces:
3 4
Vhn = v : In → Shn ; v(t) = χ0 + tχ1 , χ0 , χ1 ∈ Shn , t ∈ In ,
& '
=hn = ψ : In → S̃hn ; ψ(t) = ξ0 + tξ1 , ξ0 , ξ1 ∈ S̃hn , t ∈ In .
V
and
Whk = {v ; v|In ∈ Vhn for n = 1, . . . , N} ,
=hk = {ψ ; ψ|In ∈ V
W =hn for n = 1, . . . , N} .
With the following notations
vn+ := lim v(tn + t) , vn− := lim v(tn + t) and [v]n := vn+ − vn−
t →0+ t →0−
the discontinuous Galerkin method for symmetric coupling of FEM/BEM reads as

follows: Find (U, Φ) ∈ Whk × W=hk such that for all (w, ψ) ∈ Whk × W
=hk :
T
N T
3 4 + + + +
(U̇ ,w)+B(U,Φ;w,ψ) dt+ ([U ]n−1 ,wn−1 )+(U0 ,w0 ) = (q,w0 )+ L(w,ψ)dt.
0 n=2 0
(12.227)
This is equivalent to:

=n such that for all (w, ψ) ∈ V n × V
For n = 1, . . . , N find (Un , Φn ) ∈ Vhn × V =n :
h h h

3 4 + + − +
(U̇n , w) + B(Un , Φn ; w, ψ) dt + (Un−1 , wn−1 ) = (Un−1 , wn−1 )+ L(w, ψ)dt.
In In
(12.228)
Let (U, Φ) be the corresponding (unique) solution of (12.227) and let (u, φ) be
the exact solution of (12.224). If ΔhΩ1 is a uniform triangulation of mesh size h, if
the time steps are of length k and if u ∈ C 1 ([0, T ]; H r (Ω1 )) ∩ C 3 ([0, T ]; L2 (Ω1 ))
and φ ∈ C 0 ([0, T ]; H r−3/2(Γ )), r ∈ [ 32 , 2], the following convergence result [311]
holds:
(u − U, φ − Φ)L2 (0,T ;H ) = O(hr−1 + k 2 ) .
A similar result has been proved in [120] for time discretization by the Crank-
Nicolson method. An error controlled adaptive scheme is given in [311].
For the time-dependent eddy current problem a FE/BE coupling with the
discontinuous Galerkin method in time is established in [341] with a priori error
estimates ; a posteriori error control and an adaptive algorithm are given in [342].
Chapter 13
Time-Domain BEM
Time-domain Galerkin boundary elements provide an efficient tool for numerical

solution of boundary value problems for the homogeneous wave equation. In
Sect. 13.1 we present from [193] a time-domain Galerkin BEM for the wave
equation outside a Lipschitz obstacle in an absorbing half-space.A priori error
estimates from [193] and a posteriori error estimates from [194] are given in
Sect. 13.2.
Efficient and accurate computational methods to simulate sound emission in
space and time are of interest from modeling environmental noise to the acoustics of
concert halls [34, 94, 197]. This chapter reviews a time-domain Galerkin boundary
element method for acoustic wave problems [193, 194]. Our approach proves to be
stable and accurate in long-time computations and is competitive with frequency
domain methods for realistic problems from the sound emission of tyres.
Computations in time-domain are of particular interest for problems beyond the
reach of frequency-domain methods, such as the simulation of transient dynamics,
moving sound sources or nonlinear and dynamical contact problems. They can
also be applied to obtain results in frequency-domain, for all frequencies in one
computation, with the help of the Fast Fourier Transform to translate between time
and frequency. This approach proves competitive if a broad band of frequencies is
of interest.
As an alternative to time-domain boundary elements, the past few years have
seen rapid progress for convolution quadrature methods [30, 31, 361]. These
exploit the convolution structure in time for integral equations to approximate them
through the frequency domain by an inverse Laplace transform. Given a frequency
domain solver, their implementation does not struggle with the careful, accurate
computation of distributional integrals like time domain boundary elements do.
However, for long time simulations and certain nonlinear problems with constraints,
such as dynamic contact and friction problems, the variational nature of Galerkin
time domain methods maybe advantageous.

538 13 Time-Domain BEM
13.1 Integral Equations and Anisotropic Space-Time Sobolev

Spaces
Let d = 2 or 3 and Ω i ⊂ Rd be a bounded polygonal domain. For simplicity, we

assume that the exterior domain Ω e = Rd \Ω i is connected and that the boundary
Γ = ∂Ω is a Lipschitz manifold. Our emphasis will be on the case d = 3.
We aim to find a weak solution to an initial-boundary problem for the wave
equation in Ω e :
∂ 2u
− Δu = 0 in R+ × Ω e (13.1)
∂t 2
∂u
u(0, x) = (0, x) = 0 in Ω e ,
∂t
with either Dirichlet boundary conditions u = g, Neumann boundary conditions
∂n = g or more generally acoustic boundary conditions
∂u
∂u ∂u
−α =g on R+ × Γ .
∂n ∂t
Here n denotes the inward unit normal vector to ∂Ω e , g lies in a suitable Sobolev
space, α ∈ L∞ (Γ ). In the case of an incoming wave uinc scattered by Ω i , the right
inc inc
hand side is g = − ∂u∂n +α ∂u∂t . In order for (13.1) to be well-posed, α should have
nonnegative real part, so that waves are not amplified at reflection. We also consider
the simpler Dirichlet problem on Γ , for which instead of the absorbing boundary
condition, u|R+ ×Γ is given.
This section reduces the acoustic and Dirichlet boundary problems to time-
dependent integral equations on R+ × Γ and studies a Galerkin time-domain
boundary element method for their approximation. It presents from [193, 194] an a
priori and an a posteriori error analysis for methods based on integral formulations
of the first kind.
Time-dependent Galerkin boundary element methods for wave problems were
introduced by Bamberger and Ha-Duong [29]. Some relevant works on the numer-
ical implementation of the resulting marching-in-on-time scheme include the
Ph.D. thesis of Terrasse and [224]. For a survey see Costabel’s article [117]
In the special case of the half-space, our work is motivated by the recent explicit
formulas for the fundamental solutions obtained by Ochmann [329], which include
acoustic boundary conditions on the surface of the street.
Similar to elliptic problems, the initial-boundary value problem (13.1) for the
wave equation can be formulated as an integral equation of either the first or second
kind on the boundary. Using an appropriate Green’s function for the absorbing half-
space an equation on the subset Γ of the boundary is obtained.
13.1 Integral Equations, Space-Time Sobolev Spaces 539
We introduce the single layer potential in time domain as

Sϕ(t, x) = G(t − τ, x, y) ϕ(τ, y) dτ dsy ,
R+ ×Γ
where
1 H (t − τ − |x − y|)
G(t − τ, x, y) = M (2D)
2π (t − τ )2 − |x − y|2
1 δ(t − τ − |x − y|)
G(t − τ, x, y) = 3D
4π |x − y|
is a fundamental solution to the wave equation with the Heaviside function H and
the delta-distribution δ. Specifically in 3 dimensions, it is given by

1 ϕ(t − |x − y|, y)
Sϕ(t, x) = dsy .
4π Γ |x − y|
We similarly define the double layer potential in time domain as

∂G
Dϕ(t, x) = (t − τ, x, y) ϕ(τ, y) dτ dsy .
R+ ×Γ ∂ny
For acoustic boundary conditions we require the single-layer operator V , its

normal derivative K , the double-layer operator K and hypersingular operator W
for x ∈ Γ , t > 0:

V ϕ(t, x) = 2 G(t − τ, x, y) ϕ(τ, y) dτ dsy ,
R+ ×Γ

∂G
Kϕ(t, x) = 2 (t − τ, x, y) ϕ(τ, y) dτ dsy ,
R+ ×Γ ∂ny

∂G
K ϕ(t, x) = 2 (t − τ, x, y) ϕ(τ, y) dτ dsy ,
R+ ×Γ ∂nx

∂ 2G
W ϕ(t, x) = −2 (t − τ, x, y) ϕ(τ, y) dτ dsy .
R+ ×Γ ∂nx ∂ny
The boundary integral operators are considered between space-time anisotropic

Sobolev spaces Hσs (R+ , H=r (Γ )). To define them, if ∂Γ = ∅, first extend Γ to a
closed, orientable Lipschitz manifold Γ=.
On Γ one defines the usual Sobolev spaces of supported distributions:
=r (Γ ) = {u ∈ H r (Γ=) : supp u ⊂ Γ } ,
H r ∈R.
Furthermore, H r (Γ ) is the quotient space H r (Γ=)/H=r (Γ= \ Γ ).

To write down an explicit family of Sobolev norms, introduce a partition of unity
αi subordinate to a covering of Γ= by open sets Bi . For diffeomorphisms ϕi mapping
each Bi into the unit cube ⊂ Rd , a family of Sobolev norms is induced from Rd :
( p ) 12
& '
||u||r,ω,Γ= = (|ω|2 + |ξ |2 )r |F (αi u) ◦ ϕi−1 (ξ )|2 dξ .
i=1 Rd
The norms for different ω ∈ C \ {0} are equivalent, and F denotes the Fourier
transform. They induce norms on H r (Γ ), ||u||r,ω,Γ = infv∈H=r (Γ=\Γ ) ||u + v||r,ω,Γ=,
and on H=r (Γ ), ||u||r,ω,Γ,∗ = ||e+ u||r,ω,Γ=. e+ extends the distribution u by 0 from
=
Γ to Γ . It is stronger than ||u||r,ω,Γ whenever r ∈ 12 + Z.
We now define a class of space-time anisotropic Sobolev spaces:
Definition 13.1 For s, r ∈ R define

Hσs (R+ , H r (Γ )) = {u ∈ D+ (H r (Γ )) : e−σ t u ∈ S+ (H r (Γ )) and ||u||s,r,Γ < ∞} ,

=r (Γ )) = {u ∈ D+ (H
Hσs (R+ , H =r (Γ )) : e−σ t u ∈ S+ (H
=r (Γ )) and ||u||s,r,Γ,∗ < ∞} .

D+ (E) resp. S+ (E) denote the spaces of distributions, resp. tempered distributions,
on R with support in [0, ∞), taking values in E = H r (Γ ), H =r (Γ ). The relevant
norms are given by
+∞+iσ
1
2
us,r,σ := us,r,Γ = |ω| 2s
û(ω)2r,ω,Γ dω ,
−∞+iσ
+∞+iσ
1
2
2s
us,r,σ,∗ := us,r,Γ,∗ = |ω| û(ω)2r,ω,Γ,∗ dω .
−∞+iσ
For |r| ≤ 1 the spaces are independent of the choice of αi and ϕi . See [193, 223]
for a more detailed discussion.
The representation formula uses S and D to express a solution to the wave
equation in terms of its Dirichlet and Neumann data on Γ :
Theorem 13.1 Let u ∈ L2 (R+ , H 1 (Ω)) ∩ H01 (R+ , L2 (Ω)) be the solution
of (13.1) for a Lipschitz boundary Γ . Then
u(t, x) = Sϕ(t, x) − Dp(t, x) ,
where ϕ = [u] is the jump of u across Γ and p = [ ∂u

∂n ] is the jump of the normal
flux.
13.1 Integral Equations, Space-Time Sobolev Spaces 541
The initial boundary value problem (13.1) with acoustic boundary conditions is
then equivalent to a system of integral equations of the first kind,
%
K p − W ϕ + α ∂ϕ
∂t = F (13.2)
p + α(V ∂t p + K∂t ϕ) = G .
Here, ϕ = [u] and p = [ ∂u

∂n ] as above, and for an incoming wave u
inc scattered
by Ω i , we have F = −2 ∂u∂n and G = −2α ∂u∂t . If α −1 ∈ L∞ (Γ ), pairing these

inc inc
equations with test functions ∂t ψ respectively αq , we obtain the following space-time

variational formulation:
Find Φ = (ϕ, p) ∈ Hσ1 (R+ , H = 12 (Γ )) × Hσ1 (R+ , L2 (Γ )) such that for all Ψ =
(ψ, q) ∈ Hσ1 (R+ , H= 12 (Γ )) × Hσ1 (R+ , L2 (Γ )):
a(Φ, Ψ ) = l(Ψ ) . (13.3)
where
∞ ∞ Gq
l(Ψ ) = F ∂t ψ dsx dσ t + dsx dσ t (13.4)
0 Γ 0 Γ α
and a(Φ, Ψ ) is given by
∞
1
α(∂t ϕ)(∂t ψ)+ pq + K p(∂t ψ)−W ϕ(∂t ψ)+V (∂t p)q +K(∂t ϕ)q dsx dσ t.
α
0Γ
(13.5)
∞
Here dσ t = e−2σ t dt, σ > 0, and u, v := 0 Γ uv̄dsx dσ t.
The variational formulation of the Dirichlet problem, V ∂t φ = ∂t f , similarly
=− 21 (Γ )) such that
reads: Find φ ∈ Hσ1 (R+ , H
=− 2 (Γ )) ,
1
b(φ, ψ) = ∂t f, ψ ∀ψ ∈ Hσ1 (R+ , H (13.6)
where
∞
b(φ, ψ) = (V ∂t φ(t, x))ψ(t, x)dsx dσ t ,
0 Γ
∞
∂t f, ψ = (∂t f (t, x))ψ(t, x)dsx dσ t .
0 Γ
Adapting fundamental observations in [29] and [223] to our situation, the bilinear
forms a(Φ, Ψ ) and b(φ, ψ) are continous and, in a weak sense, coercive. They are
related to the physical energy of the system. As a consequence, both the acoustic
and the Dirichlet problem admit unique solutions for sufficiently smooth data. See
[193] for details.
The Neumann problem, corresponding to α = 0, may be discussed similarly
[34, 195]. In addition to the variational formulations as integral equations of the
first kind, related to the energy, for computations an integral equation of the
second kind will prove useful. We will only state the Neumann case, α = 0:
1
Find ϕ(t, x) ∈ Hσ2 ([0, ∞), H − 2 (Γ )) such that for all test functions ψ(t, x) ∈
1
1
Hσ2 ([0, ∞), H − 2 (Γ )) there holds:
1
∞ ∞

−I + K ϕ(t, x)ψ(t, x) dsx dσ t = 2 g(t, x)ψ(t, x) dsx dσ t.
0 Γ 0 Γ
(13.7)
As it is equivalent to the original initial boundary value problem, also this

formulation admits a unique solution for smooth right hand sides, though it is not
known to be coercive.
For applications to traffic noise, also the wave equation in the half-space R3+ is
of interest [34]. By choosing an appropriate, modified Green’s function G which
satisfies the boundary conditions, the formulation may be reduced to the surface Γ
of the scatterer. Partially absorbing, acoustic boundary conditions on the road ∂R3+ ,
using a modification of K have been discussed in [193, 329].
We now discuss the discrete spaces used for the numerical approximation
of the weak formulations (13.3), (13.6) and (13.7). If Γ is not polygonal we
approximate it by a piecewise polygonal curve resp. surface and write Γ again
for the approximation. For simplicity, when d = 3 we will use here a surface
composed of N triangular facets Γi such that Γ = ∪N i=1 Γi . When d = 2, we
assume Γ = ∪N i=1 Γ i is composed of line segments Γ i . In each case, the elements
Γi are closed with int (Γi ) = ∅, and for distinct Γi , Γj ⊂ Γ the intersection
int (Γi ) ∩ int (Γj ) = ∅.
For the time discretisation we consider a uniform decomposition of the time
interval [0, ∞) into subintervals In = [tn−1 , tn ) with time step |In | = Δt, such
that tn = nΔt (n = 0, 1, . . . ).
p p p
We choose a basis ϕ1 , · · · , ϕNs of the space Vh of piecewise polynomial
functions of degree p in space (continuous and vanishing at ∂Γ if p ≥ 1) and a
q
basis β 1,q , · · · , β Nt ,q of the space VΔt of piecewise polynomial functions of degree
of q in time (continuous and vanishing at t = 0 if q ≥ 1).
Let TS = {T1 , · · · , TNs } be the spatial mesh for Γ and TT = {[0, t1 ), [t1 , t2 ), · · · ,
[tNt −1 , T )} the time mesh for a finite subinterval [0, T ).
We consider the tensor product of the approximation spaces in space and time,
p q
Vh and VΔt , associated to the space-time mesh TS,T = TS × TT , and we write
p,q p q
VΔt,h = Vh ⊗ VΔt .
13.2 A Priori and A Posteriori Error Estimates 543
13.2 A Priori and A Posteriori Error Estimates
The approximation spaces lead to Galerkin formulations for the acoustic and
Dirichlet problems (13.3), (13.6) and (13.7). E.g. the Galerkin formulation of (13.6)
p,q
reads: Find φΔt,h ∈ VΔt,h such that
p,q
b(φΔt,h , ψΔt,h ) = (∂t f )Δt,h , ψΔt,h ∀ψΔt,h ∈ VΔt,h . (13.8)
The well-posedness of the continuous and discretized problems is a basic

consequence of the continuity and weak coercivity of the bilinear form b:
1
Corollary 13.1 Let f ∈ Hσ1 (R+ , H 2 (Γ )). Then the Dirichlet problem (13.6) and
its discretization (13.8) admit unique solutions φ, φΔt,h ∈ Hσ0 (R+ , H − 2 (Γ )) and
1
φ0,− 1 ,Γ,∗ , φΔt,h 0,− 1 ,Γ,∗ f 1, 1 ,Γ .

2 2 2
In [193], we discuss a priori error estimates and the convergence of Galerkin

approximations for (13.3) and (13.6). For the Dirichlet problem the basic estimate
is the following:
=− 12 (Γ )) of (13.6), φΔt,h ∈
Theorem 13.2 ([193]) For the solutions φ ∈ Hσ1 (R+ , H
= p,q
V Δt,h of (13.8) there holds:
φ − φΔt,h 0,− 1 ,Γ,∗ ||(∂t f )Δt,h − ∂t f ||0, 1 ,Γ

2 2
"
1 1
+ inf p,q (1 + )φ − ψΔt,h ||0,− 1 ,Γ,∗ + ∂t φ − ∂t ψΔt,h ||0,− 1 ,Γ,∗ .
=
ψΔt,h ∈V Δt 2 Δt 2
Δt,h
If in addition φ ∈ Hσs (R+ , H m (Γ )) for s > 1 and m > − 12 , one obtains

convergence rates.
For the proof, we assume for simplicity that ∂t f = (∂t f )Δt,h . Then from the
weak coercivtiy of the bilinear form b, and adding a 0, we have
φ − φΔt,h 20,− 1 ,Γ,∗ b(φΔt,h − φ, φΔt,h − ψΔt,h ) + b(φ − ψΔt,h , φΔt,h − ψΔt,h )
2
for all test functions ψΔt,h . Using the Galerkin orthogonality
b(φΔt,h − φ, φΔt,h − ψΔt,h ) = 0 ,
the first term vanishes. For the second, we use the continuity of the duality pairing,
the mapping properties of V and an inverse estimate in t:
∂
b(φ − ψΔt,h , φΔt,h − ψΔt,h ) ≤ V (φ − ψΔt,h )−1, 1 ,Γ φΔt,h − ψΔt,h 1,− 1 ,Γ,∗
∂t 2 2
φ − ψΔt,h 1,− 1 ,Γ,∗ φΔt,h − ψΔt,h 1,− 1 ,Γ,∗

2 2
1
φΔt,h − ψΔt,h 0,− 1 ,Γ,∗ φ − ψΔt,h 1,− 1 ,Γ,∗
Δt 2 2
Combining this with a triangle inequality, we obtain the claimed a priori bound:
φ − φΔt,h 0,− 1 ,Γ,∗ φ − ψΔt,h 0,− 1 ,Γ,∗ + φΔt,h − ψΔt,h 0,− 1 ,Γ,∗
2 2 2

1
1+ φ − ψΔt,h 1,− 1 ,Γ,∗ .
Δt 2
For the acoustic problem, we introduce the norm

1
2
|||p, ϕ||| = ||p||20,0,Γ + ||ϕ||20, 1 ,Γ + ||∂t ϕ||20,0,Γ .
2
Theorem 13.3 ([193]) Assume (for simplicity) that α1 ∈ L∞ (Γ ). For the solutions
Φ = (p, ϕ) ∈ Hσ1 (R+ , H = 12 (Γ )) × Hσ1 (R+ , L2 (Γ )) of (13.3) and ΦΔt,h =
p̃,q̃ p,q
(pΔt,h , ϕΔt,h) ∈ VΔt,h × VΔt,h of its discretisation there holds:
|||p − pΔt,h , ϕ − ϕΔt,h ||| ||FΔt,h − F ||0,0,Γ + ||GΔt,h − G||0,0,Γ

1 1
+ max ,√ inf ||p − qΔt,h ||1,0,Γ + ||ϕ − ψΔt,h ||1, 1 ,Γ .
Δt h (qΔt,h ,ψΔt,h )∈VΔt,h
p̃,q̃ p,q
×VΔt,h
2
As for the Dirichlet problem, better estimates are obtained under smoothness
assumptions, ϕ ∈ Hσs1 (R+ , H m1 (Γ )), p ∈ Hσs2 (R+ , H m2 (Γ )), [193].
Computable error indicators are a key ingredient to design adaptive mesh refine-
ments. For the time-dependent boundary element methods efficient and reliable such
estimates of residual type have been obtained in [194], see also [197] and [358] for
alternative error indicators.
Using ideas going back to Carstensen [74] and Carstensen and Stephan [92] for
the boundary element method for elliptic problems (see Section 10.1 and 10.2), we
obtain an a posteriori error estimate with residual error estimator for the Galerkin
solution to the Dirichlet problem in [194].
Theorem 13.4 Let φ, φΔt,h ∈ Hσ1 (R+ , H − 2 (Γ )) the solutions to (13.6)

1
resp. (13.8). Assume that R = ∂t f − V ∂t φΔt,h ∈ Hσ0 (R+ , H 1 (Γ )). Then

φ − φΔt,h 20,− 1 ,Γ,∗ R0,1,Γ Δt∂t R0,0,Γ + h · ∇R0,0,Γ
2
max{Δt, h}(∂t R0,0,Γ + ∇R0,0,Γ )2

p,q
Proof We first note that for all ψΔt,h ∈ VΔt,h
φ − φΔt,h 20,− 1 ,Γ,∗ b(φ − φΔt,h , φ − φΔt,h )

2

= ∂t f (φ − φΔt,h ) dsx dσ t − b(φΔt,h , φ − φΔt,h )
R+ Γ

= ∂t f (φ − ψΔt,h ) dsx dσ t − b(φΔt,h , φ − ψΔt,h )
R+ Γ

= (∂t f − V ∂t φΔt,h )(φ − ψΔt,h ) dsx dσ t .
R+ Γ
The last term may be estimated by:

(∂t f − V φ̇Δt,h )(φ − ψΔt,h ) dsx dσ t
R+ Γ
≤ R0, 1 ,Γ φ − ψΔt,h 0,− 1 ,Γ,∗ .

2 2
We use ψΔt,h = φΔt,h together with the interpolation inequality
R20, 1 ,Γ ≤ R0,0,Γ R0,1,Γ .

2
p,q
As the residual is perpendicular to VΔt,h ,
=Δt,h
R20,0,Γ = R, R = R, R − ψ
=Δt,h 0,0,Γ
≤ R0,0,Γ R − ψ
=Δt,h ∈ V p,q , we obtain

for all ψ Δt,h
=Δt,h 0,0,Γ : ψ
R0,0,Γ ≤ inf{R − ψ =Δt,h ∈ V p,q
Δt,h } .
=Δt,h = Π
Choosing ψ =Δt,h R, based on the interpolation operator defined earlier, we
obtain
R0,0,Γ Δt∂t R0,0,Γ + h · ∇R0,0,Γ .
The theorem follows.

The result for the single layer potential generalizes to a theorem without any
assumptions on the underlying meshes.
Theorem 13.5 Let φ ∈ Hσ0 (R+ , H − 2 (Γ )) be the solution to (13.6), and let
1
1
φh,Δt ∈ Hσ0 (R+ , H − 2 (Γ )) such that R = ∂t f − V ∂t φh,Δt ∈ Hσ0 (R+ , H 1 (Γ )).
Then

φ − φh,Δt 20,− 1 ,Γ,∗ max{Δt, hΔ } R20,1,[ti ,ti+1 )×Δ .
2
i,Δ
Because of the different norms in the upper and lower bounds for b, the a
posteriori estimate only satisfies a weak variant of efficiency: For ε ∈ (0, 1):
max{Δt, h}−
1−ε
2 φ − φΔt,h 0,− 1 ,Γ R0,1−ε,Γ = V (φ̇ − φ̇Δt,h )0,1−ε,Γ
2
φ − φΔt,h 2,−ε,Γ ≤ φ − φΔt,h 2,0,Γ .
A proof of the sharp estimate, ε = 0, would require sharp mapping properties of the
layer potentials outside the energy spaces.
One then uses the mapping properties of V together with approximation proper-
ties of the finite element spaces to recover the same spatial Sobolev index − 12 in the
upper and lower estimates.
Theorem 13.6 ([194]) Assume that the R ∈ H 0 ([0, T ], H 1 (Γ )) and that the
p,q
ansatz functions VΔt,h ⊆ H 2 ([0, T ], H 0 (Γ )) satisfy
inf p,q φ − ψhΔt 2,0,Γ,∗ 0 max{Δt, h}β (13.9)

ψhΔt ∈VΔt,h
for some β > 0. Then for all ε ∈ (0, 1)
R0,1−ε,Γ max{h− 2 , (Δt)− 2 }φ − φhΔt 2,−1/2,Γ,∗.

1 1
Remark 13.1 The hypothesis (13.9) can be verified using the singular expansion of
the solution φ at the edges and corners [194].
For the acoustic problem, a simple error estimate reads as follows:
1
Theorem 13.7 ([194]) Let (ϕ, p), (ϕΔt,h , pΔt,h ) ∈ H01 ([0, T ], H 2 (Γ ))
×H 1 ([0, T ], L2 (Γ )) be the solutions to (13.3) and its discretized variant, and
assume that
R1 = F − α ϕ̇Δt,h + 2K pΔt,h − 2W ϕΔt,h ∈ L2 ([0, T ], L2 (Γ )) ,

R2 = G + α −1 pΔt,h + 2S ṗΔt,h − 2K ϕ̇Δt,h ∈ L2 ([0, T ], L2 (Γ )) .
Then
|p − pΔt,h , ϕ − ϕΔt,h | R1 0,0,Γ + R2 0,0,Γ .
In [195] the Neumann problem is solved with a double layer potential ansatz
leading to the hyper singular integral equation and corresponding a priori and a
posteriori error estimates are given.
13.2.1 Adaptive Mesh Refinements
Space-time adaptive methods are still in their infancy. As a test case in [194] we
concentrate on time-independent geometric singularities of the solution, e.g. in the
horn geometry between the tyre and the street. In this case we expect to have time-
independent meshes, refined near the singularities, which do not require an update
of the Galerkin matrices in every time step.
From the discrete solution ϕ̇Δt,h of the Dirichlet problem (13.8) and f˙ we
determine in every triangle the time integrated local error indicator
T
2
η = [h∇Γ (f˙ − V ϕ̇Δt,h )]2 ,
0
where the time integral is approximated by a Riemann sum.

The error indicators η lead to an adaptive algorithm, based on the 4 steps
SOLVE −→ ESTIMATE −→ MARK −→ REFINE.
Adaptive Algorithm [194]:

Input: Mesh T = T0 , refinement parameter θ ∈ (0, 1), tolerance > 0, data f .
(i) Solve V ϕ̇Δt,h = f˙ on T .
(ii) Compute the error indicators η() in each triangle ∈ T .
(iii) Find ηmax = max η().
(iv) Stop if i η2 ( i ) < 2 .
(v) Mark all ∈ T with η( i ) > θ ηmax .
(vi) Refine each marked triangle into 4 new triangles to obtain a new mesh T
Δt
(and project the new nodes onto the sphere). Choose Δt such that Δx ≤ 1 for
all traingles.
(vii) Go to 1.
Output: Approximation of ϕ̇.
Example We consider the single layer integral equation V φ = f on the square
screen Γ = [−0.5, 0.5]2 × {0} with right hand side f (t, x, y, z) = sin(t)5 x 2 for
times [0, 2.5]. Using a discretization by linear ansatz and test functions in space and
time, we compare the error of a uniform discetization to the error of an adaptive
series of meshes, steered by the residual error estimate. The time step is Δt = 0.1,
and the uniform meshes consist of 18, 288, 648, 1352, and 6050 triangles, while
the adaptive refinements correspond to 36, 74, 164, 370, 784, 1676, 3485, and 7432
triangles. Figure 13.1 shows the convergence in of the indicators and the energy
error, for both the uniform and adaptive series of meshes.
Figure 13.2 shows some representative adaptive meshes, where the color scale
highlights the residual-based indicator values for each element. Mesh refinements
0
10 E − unif
E − adap
IND − unif
IND − adap
−2
10
−4
10
−6
10 2 3 4 5 6
10 10 10 10 10
DOF
Fig. 13.1 Energy error and residual error indicators for Dirichlet problem on Γ = [−0.5, 0.5]2 ×
{0} [194]
Fig. 13.2 Meshes 1, 2, 3 and 6 generated by adaptive refinements [194]

13.3 Time Domain BEM for Contact Problems 549
concentrate at the left and right edges, where the right hand side is steep, and to a
lesser extent also at the top and bottom edges.
13.3 Time Domain BEM for Contact Problems
The previous section considered the use of time-independent adapted (graded)

meshes to resolve the geometric singularities of the solution to the wave equation
near edges and corners. In this section we discuss a class of problems where the
singularities are moving in space and time, motivating future developments of
space-time adaptive boundary element procedures. For the contact problems under
consideration, the free contact boundary is changing with time, and so is the location
of its associated edge and corner singularities.
In [192] we propose a time domain boundary element method for a dynamical
contact problem for the scalar wave equation and provide a priori error estimates
in the case of a flat contact area. For the (scalar) displacement w : Ω → R
contact conditions along a crack G with a non-penetrable material in a Lipschitz
domain Ω ⊂ Rn are described in terms of the traction −μ ∂w ∂ν |G (with normal ν)
and prescribed forces h:
%
w|G ≥ 0 , −μ ∂w ∂ν |G ≥ h
w|G > 0 ⇒ μ ∂w
∂ν = h.
The full system of equations for the dynamical contact problem is given by:
⎧ 2
⎪
⎪ ∂ w
= cs2 Δw for (t, x) ∈ R × Ω
⎪
⎪ ∂t 2
⎪
⎪
⎨w = 0
⎪ on Γ \G (Γ = ∂Ω)
w ≥ 0 , −mu ∂ν ≥ 0 on G
∂w (13.10)
⎪
⎪
⎪
⎪
⎪
⎪ (−μ ∂w∂ν − h)w = 0 on G
⎪
⎩ w=0 for (t, x) ∈ (−∞, 0) × Ω .
In [192] we reduce (13.10) to an equivalent variational inequality (in a weak

form) for the trace uσ = wσ |Γ , wσ := e−σ t w, hσ := e−σ t h (for some σ > 0)
on the boundary: Find uσ ∈ Hσ (R+ , H̃ 1/2 (G)) such that uσ ≥ 0 and for all
1/2
1/2 +
v ∈ Hσ (R , H̃ 1/2 (G)) with v ≥ 0:
pQ Sσ uσ , v − uσ ≥ hσ , v − uσ . (13.11)
Here pQ is the restriction to Q = R × G and Sσ is the Dirichlet-to-Neumann

operator, defined as
∂w
Sσ wσ |Γ := −μ |Γ .
∂ν
The discretized variational inequality then reads:

+
Find uΔt,h ∈ KΔt,h such that
pQ Sσ uΔt,h , vΔt,h − uΔt,h ≥ h, vΔt,h − uΔt,h (13.12)
+ + p,q
for all vΔt,h ∈ K̃Δt,h . Here, K̃Δt,h ⊂ ṼΔt,h is the convex subset of nonnegative
piecewise polynomials.
There holds the following a priori error estimate:
3 1 1 1
Theorem 13.8 Let h ∈ Hσ2 (R+ , H − 2 (G)) and let u ∈ Hσ2 (R+ , H̃ 2 (G))+ ,
1
+ 1
uΔt,h ∈ K̃Δt,h ⊂ Hσ (R+ , H̃ 2 (G))+ be the solutions of (13.11), resp. (13.12). Then
2
the following estimate holds:
uσ − uΔt,h 2− 1 , 1 ,σ,∗ σ (13.13)

2 2
inf (h − pQ Sσ uσ 1 ,− 1 ,σ uσ − φΔt,h − 1 , 1 ,σ,∗ + uσ − φΔt,h 21 , 1 ,σ,∗ ).

+
φΔt,h ∈K̃Δt,h 2 2 2 2 2 2
To assure conservation of energy in the numerical approximation, it proves useful to

impose the constraints on the displacement only indirectly. We therefore reformulate
the variational inequality as an equivalent mixed system. The Lagrange multiplier λ
in this formulation also provides a measure to which extent the variational inequality
is not an equality; physically, it indicates the contact area and the contact forces
within the computational domain.
The variational inequality (13.11) is equivalent to the mixed formulation:
Find (u, λ) ∈ X = Hσ (R+ , H̃ 1/2(G)) × Hσ (R+ , H −1/2 (G)), λ ≥ 0,
1/2 1/2
such that
%
Sσ u, v − λ, v = h, v
(13.14)
u, μ − λ ≥ 0
holds for all (v, μ) ∈ X, μ ≥ 0.

The corresponding discrete formulation with different meshes for the displace-
ment and the Lagrange multiplier reads as follows:
1,1
Find (uΔt,h , λΔt2 ,h2 ) ∈ Ṽt,h × (Vt0,0
2 ,h2
)+ such that
%
Sσ uΔt,h , vΔt,h − λΔt2 ,h2 ,vΔt,h = h, vΔt,h
(13.15)
uΔt,h , μΔt2 ,h2 − λΔt2 ,h2 ≥ 0
1,1
holds for all (vΔt,h , μΔt2 ,h2 ) ∈ Ṽt,h × (Vt0,0
2 ,h2
)+ .
1 ,Δt1 }
Theorem 13.9 ([192]) Let C > 0 sufficiently small, and max{h
min{h2 ,Δt2 } < C. Then the
discrete mixed formulation (13.15) admits a unique solution, and the following a
13.3 Time Domain BEM for Contact Problems 551
priori estimates hold:
λ−λΔt2 ,h2 0,− 1 ,σ inf λ− λ̃Δt2 ,h2 0,− 1 ,σ +(Δt1 )− 2 u − uΔt1 ,h1 − 1 , 1 ,σ,∗ ,
1
2 λ̃Δt2 ,h2 2 2 2
(13.16)
u − uΔt1 ,h1 − 1 , 1 ,σ,∗ σ inf u − vΔt1 ,h1 1 , 1 ,σ,∗
2 2 vΔt1 ,h1 2 2
& '
+ inf λ̃Δt2 ,h2 − λ 1 ,− 1 ,σ +λ̃Δt2 ,h2 −λΔt2 ,h2 1 ,− 1 ,σ .
λ̃Δt2 ,h2 2 2 2 2
(13.17)
A crucial ingredient in the proof is the inf-sup condition in space-time, which holds
1 ,Δt1 }
for max{h
min{h2 ,Δt2 } sufficiently small [192]: There exists α > 0 such that for all λΔt2 ,h2 :
μΔt1 ,h1 , λΔt2 ,h2

sup ≥ αλΔt2 ,h2 0,− 1 ,σ .
μΔt1 ,h1 μΔt1 ,h1 0, 1 ,σ,∗ 2
2
Numerically, system (13.15) is solved with an Uzawa algorithm, and the Dirichlet-
to-Neumann operator is computed in terms of the retarded boundary layer potentials
as W + (K − 1/2)V −1 (K − 1/2). See [192] for details of the discretization.
We now consider the discretization of the dynamical contact problem (13.10),
(13.11) for a flat contact area. No exact solutions are known, so that we compare the
numerical approximations to a reference solutions on an appropriate finer space-
time mesh.
Example We choose Γ = [−2, 2]2 × {0} with contact area G = [−1, 1]2 × {0}
for times [0, 5], with the CFL-ratio Δt
h = 0.7. The right hand side of the contact
problem (13.15) is given by
h(t, x) = e−2t t 4 cos(2πx) cos(2πy)χ[−0.25,0.25](x)χ[−0.25,0.25](y) .
The numerical solutions are compared to a reference solution on a mesh with 12800
triangles, and we use Δt = 0.075.
In this example, contact takes place from time t = 4.25 on. Figure 13.3 considers
the relative error to the reference solution in L2 ([0, T ] × G). The numerical
approximations converge at a rate of α = 0.8 with increasing degrees of freedom.
Algorithmically, the computational cost of the nonlinear solver is dominated by the
cost of computing the matrix entries.
0
10
Relative L2 error in space-time
-1
10
-2
10
10 -3
10 1 10 2 10 3 10 4 10 5
Degrees of Freedom
Fig. 13.3 Relative L2 ([0, T ] × Γ )-error vs. degrees of freedom of the solutions to the contact
problem for fixed Δt
h [192]
13.4 Algorithmic Aspects of Time Domain BEM
13.4.1 MOT Algorithm
The Galerkin discretization in space and time leads to a block-lower-triangular sys-

tem of equations, which can be solved by blockwise forward substitution. For ease
of presentation we consider the Galerkin scheme for the Dirichlet problem (13.6)
which can be rewritten for piecewise constant test and trial functions in space and
time as:
0,0 0,0
Find φt,h ∈ Vt,h such that for all ψ ∈ Vt,h
∞ ∞
(V φt,h (t, x))∂t ψ(t, x)dsx dσ t = (ft,h (t, x))∂t ψ(t, x)dsx dσ t .
0 Γ 0 Γ
(13.18)
This yields an algebraic system of the form

n
V n−m bm = 2(f n−1 − f n )
m=1
in time step n = 1, 2, 3, . . . . It can be solved by forward substitution, giving rise to

the marching in on time (MOT) scheme
n−1
V 0 bn = 2(f n−1 − f n ) − V n−m bm .
m=1
13.4 Algorithmic Aspects of Time Domain BEM 553
This can be seen as follows: Setting

Nt Ns m m,0
ϕΔt,h (x, t) = m=1 i=1 bi β (t)ϕi0 (x)
and
ψ(x, t) = β n,0 (t)ϕj0 (x)
one computes
∞
β m,0 (t − |x − y|)β̇ n,0 (t)dt = H (tn−m − |x − y|) − H (tn−m−1 − |x − y|)
0
− H (tn−m+1 − |x − y|) + H (tn−m − |x − y|)
with the Heaviside funcion

%
1 |x − y| ≤ tl
H (tl − |x − y|) = .
0 elsewhere
Therefore the left hand side in (13.18) becomes

Nt
Ns - ϕi0 (y)ϕj0 (x) ϕi0 (y)ϕj0 (x) .
bim dsy dsx − dsy dsx
4π|x − y| 4π|x − y|
m=1 i=1 Γ ×Γ ∩En−m−1 Γ ×Γ ∩En−m
with
El := {(x, y) ∈ Γ × Γ : tl ≤ |x − y| ≤ tl+1 }.
Similarly, setting

Nt
Ns
ft,h (x, t) = fim β m,0 (t)ϕi0 (x)
m=1 i=1
yields for the right hand side in (13.18)

Ns
n−1
[fi − fi ]
n
ϕi0 (x)ϕj0 (x)dsx ,
i=1 Γ
where

Nt
fin−1 = fim β m,0 (tn−1 ),
m=1
because
∞
β m,0 (t)β̇ n,0 (t)dt = β m,0 (tn−1 ) − β m,0 (tn ).
0
The above fully discrete systems involve the computation of a series of matrices, that
(if α∞ = 0) are sparsely populated, because the Dirac-delta fundamental solution
restricts the number of interacting elements per time step. Note that the computation
of each matrix only depends on the time difference. Furthermore, for bounded
surfaces Γ the .
- matrices V n−m vanish whenever the time difference l := n − m
diam Γ
satisfies l > Δt , i.e. the light cone has traveled through the entire surface Γ .
The most time consuming part in the MOT algorithm is the matrix computation,
even though the resulting matrices are sparse. An efficient hp-composite Gauss-
quadrature allows to compute the entries in V l [402, 403].
13.4.2 An hp-Composite Quadrature of Matrix Elements
The most time consuming part in the MOT algorithm is the matrix computation,
even though the resulting matrices are sparse in each time step. An efficient hp-
composite Gauss-quadrature allows to compute the entries in V l , and similarly for
the other layer operators [331, 402, 403].
Recall the form of the matrix entries of V l in R3 as an example:
p
1 ϕi (y)∂t β n,q (t − |x − y|) p
ϕj (x)β m,q (t) dsy dsx dσ t .
2π R+ ×Γ ×Γ |x − y|
First, the time integrals are evaluated analytically and result in an integration domain
E = {(x, y) ∈ Γ × Γ : rmin ≤ |x − y| ≤ rmax }
of the form of a light cone, rmin and rmax depending on tm and tn . It remains to
evaluate terms like

p p
Gνij = kν (x − y)ϕi (y)ϕj (x) dsy dsx , (13.19)
E
where kν (x−y) = |x−y|ν denotes a weakly singular kernel function. Our numerical
quadrature separates the outer spatial integration from the singular inner one. Define
the domain of influence of x ∈ R3 by
& '
E(x) := Brmax (x) \ Brmin (x) = y ∈ R3 : rmin ≤ |x − y| ≤ rmax
13.4 Algorithmic Aspects of Time Domain BEM 555
(a) (b)
E(T )
rmin rmax
E(T ) ∩ T̂
x
T
E(x) ∩ T
Fig. 13.4 Domains of influence and the illumination of test and trial element T> and T during the
evaluation of the inner and outer integral [190]. (a) Outer integral: Domain of influence of triangle
T̂ intersected with triangle T . (b) Inner integral: Domain of influence E(x) of point x ∈ E(T ) ∩ T̂
as in Fig. 13.4b. Figure 13.4a similarly sketches the domain of influence of a

triangle T ,

E(T ) := E(x) = {y ∈ R3 : rmin ≤ |x − y| ≤ rmax , x ∈ T } .
x∈T
Defining E(Tj , Ti ) := E ∩ (Tj × Ti ), we rewrite (13.19) as

p p
Gνij = kν (x − y)ϕi (y)ϕj (x) dsy dsx
Ti ⊂suppϕi E(T ,T )
Tj ⊂suppϕj j i

p p
= ϕj (x)Pi,i (x) dsx ,
Ti ⊂suppϕi T ∩E(T )
Tj ⊂suppϕj j i
with a retarded potential Pi,i given by

p
Pi,i (x) := kν (x − y)ϕi (y) dsy .
E(x)∩Ti
To simplify notation, we explain the quadrature for a simplified integral. Given

triangles T , T̂ and basis functions ϕ, > >, respectively, a typical
ϕ defined on T and T
entry in the Galerkin matrix reads

ϕ (x) dsx , P ϕ(x) :=
P ϕ(x)> kν (x − y)ϕ(y) dsy . (13.20)
E(T )∩T> E(x)∩T
We evaluate the outer and the inner integral step by step decomposing the integration
domain and using a grading strategy for the different singularities. It is crucial to
take into account the cut-off behavior due to the different domains of influence.
As explained in [190] we obtain from (13.20)
nd
ν
P ϕ(x) = (d 2 + r 2 ) 2 ϕ(r, θ )r dr dθ ,
l=1 >
Dl
where d > 0 and ϕ is sufficiently regular. For each of the domains D̂l (see Fig. 13.5),
we can write the integral as
θ2 r2 (θ) ν
I (D̂l ) f := f (r, θ ) dr dθ , f (r, θ ) := (d 2 + r 2 ) 2 ϕ(r, θ )r. (13.21)
θ1 r1 (θ)
To introduce our quadrature method, denote by Q[a,b]

n f := n
i=1 wi f (xi ) the
b
Gauss-Legendre quadrature rule with n quadrature points to evaluate f dx. Given
a
a subdivision of [a, b] into m subintervals Ij , a variable order composite Gauss
I
rule with degree vector n = (n1 , . . . , nm ) is defined by Qn,m,σ f := m j
j =1 Qnj f .
We use a geometric subdivision
* of [a, b] with m levels and grading parameter
σ ∈ (0, 1): [a, b] = m j =1 Ij , where for j = 1, . . . , m we let Ij := [xj −1 , xj ],
(r) (r)
x0 := a, xj := a + (b − a)σ m−j . For nr = (n1 , . . . , nm ), mr ≥ 1 and σr ∈ (0, 1],
the integral (13.21) is then computed as
QD̂l f := Qn[θθ1 ,θ2 ] (Q[r 1 (θ),r2 (θ)] f ).

nr ,mr ,σr
D̂1 D̂2 D̂3 D̂4
Fig. 13.5 Generic integration domains [190]

13.5 Screen Problems and Graded Meshes 557
An error analysis for the evaluation of (13.21) is given in [331] by showing that
the integrand belongs to the countably normed, weighted space Bβ0 (T ) of Babuska
[19].
Definition 13.2 (Countably normed space Bβl (T )) We say u ∈ Bβl (T ) with
respect to a weight function Φβ,α,l , if u ∈ H l−1 (T ) and if
Φβ,α,l D α uL2 (Ω) ≤ Cd |α|−l (|α| − l)!
for |α| = l, l + 1, . . .. Here the constants C > 0 and d ≥ 1 are independent of |α|.
If the number of angular quadrature points, nθ , is chosen proportional to mr , we
obtain the following theorem on the accuracy of the quadrature in our TDBEM:
Theorem 13.10 ([331]) Given a function f ∈ Bβ0 (T) with a weight function
Φβ,α,0 (r) = r |α|+β , then there holds for D̂l :
√
3
|I (D̂l ) f − Q(D̂l ) f | ≤ Ce−b N
for l = 1, . . . , 4. Here N denotes the total number of quadrature points and C and
b are positive constants independently of N, but depending on the grading factor
σr , the number of levels mr and on f .
13.5 Screen Problems and Graded Meshes
For solutions to elliptic equations in a polyhedral domain, the asymptotic behavior

near the edges and corners, as well as its numerical approximation has been studied
in Sect. 4.3 and Chap. 7.
In the case of the wave equation in domains with conical or wedge singularities,
a similar asymptotic behavior has been obtained by Kokotov, Neittaanmäki and
Plamenevskii since the late 1990’s [268, 269]. Their results imply that at a fixed
time t, the solution to the wave equation admits an explicit singular expansion with
the same exponents as for elliptic equations.
The realistic scattering and diffraction of waves in R3 is crucially affected by
geometric singularities of the scatterer, with significant new challenges for both the
singular and numerical analysis. The article [191] studies the solution of the wave
equation in the most singular case, outside a screen Γ in R3 or, equivalently, for an
opening crack. From the singular expansion one obtains optimal convergence rates
for piecewise polynomial approximations on graded meshes.
The computations below are conducted on gradedM meshes on the square [−1, 1]2,
respectively on the circular screen {(x, y, 0) : x + y 2 ≤ 1}. To define β-graded
2
meshes on the square, due to symmetry, it suffices to consider a β-graded mesh on
[−1, 0]. We define yk = xk = −1 + ( Nkl )β for k = 1, . . . , Nl and for a constant
Fig. 13.6 β-graded meshes for square and circular screens, with β = 2 [191]
β ≥ 0. The nodes of the β-graded mesh on the square are therefore (xk , yl ), k, l =
1, . . . , Nl . We note that for β = 1 we would have a uniform mesh.
In a general convex, polyhedral geometry graded meshes are locally modeled
on this example. In particular, on the circular screen of radius 1, for β = 1 we
take a uniform mesh with nodes on concentric circles of radius rk = 1 − Nkl for
k = 0, . . . , Nl −1. For the β-graded mesh, the radii are moved to rk = 1−( Nkl )β for
k = 0, . . . , Nl − 1. While the triangles become increasingly flat near the boundary,
their total number remains proportional to Nl2 .
Examples of the resulting 2-graded meshes on the square and the circular screens
are depicted in Fig. 13.6.
For these meshes one shows:
Theorem 13.11 ([191]) Let ε > 0.
β
a) Let ψ be the solution to the hypersingular integral equation W ψ = g and ψh,Δt
the best approximation in the norm of Hσr (R+ , H = 12 −s (Γ )) to ψ in V
=p,1 on a
Δt,h
β-graded spatial mesh with Δt hβ . Then
1 3
ψ − ψh,Δt r, 1 −s,Γ,∗ ≤ Cβ,ε hmin{β( 2 +s), 2 +s}−ε , where s ∈ [0, 12 ] and r ∈
β
2
[0, p).
b) Let φ be the solution to the single layer integral equation V φ = f and
φh,Δt the best approximation in the norm of Hσr (R+ , H
β =− 21 (Γ )) to φ in V p,0
Δt,h
β
on a β-graded spatial mesh with Δt hβ . Then φ − φh,Δt r,− 1 ,Γ,∗ ≤
2
β 3
Cβ,ε hmin{ 2 , 2 }−ε , where r ∈ [0, p + 1).
Note that the energy norm associated to the weak form of the single layer integral
1
equation is weaker than the norm of Hσ1 (R+ , H − 2 (Γ )) and stronger than the norm
of Hσ0 (R+ , H − 2 (Γ )), according to the coercivity and continuity properties of V

1
on screen. Similarly, for the weak form of the hypersingular integral equation, the
1
energy norm is weaker than the norm of Hσ1 (R+ , H 2 (Γ )) and stronger than the
1
norm of Hσ0 (R+ , H 2 (Γ )).
Together with the a priori estimates for the time domain boundary element
methods on screens in this chapter, the theorem implies convergence rates for
the Galerkin approximations, which recover those for smooth solutions (up to an
arbitrarily small ε > 0) provided the grading parameter β is chosen sufficiently
large.
The crucial ingredient in the proof of Theorem 13.11 is a precise description
of the corner and edge singularities of the solution. In analogy with the work of
Plamenevskii and coauthors, the asymptotic expansion of the solution u to the wave
equation, respectively its normal derivative ∂ν u, near the corner of a polygonal
screen in R3 in the time domain is as follows:
1
u(t, x)|+ = ψ0 (t, r, θ ) + χ(r)r γ α(t, θ ) + χ̃ (θ )b1 (t, r)(sin(θ )) 2
1
+ χ̃( π2 − θ )b2 (t, r)(cos(θ )) 2 ,
∂ν u(t, x)|+ = φ0 (t, r, θ ) + χ(r)r γ −1 α(t, θ ) + χ̃ (θ )b1 (t, r)r −1 (sin(θ ))− 2
1
1
+ χ̃( π2 − θ )b2 (t, r)r −1 (cos(θ ))− 2 .
The remainders ψ0 and φ0 are less singular, and γ is the singular exponent of the
corner singularity known from the elliptic case. In particular, for the square screen
γ 0 0.2966. For a circular screen, only the edge singularity with singular exponent
2 (u), respectively − 2 (∂ν u), is present.
1 1
For the algorithmic details of the numerical experiments we refer to [191].

Example 13.1 We compute the solution to theM integral equation V φ = f on R+ t ×Γ
with the circular screen Γ = {(x, y, 0) : 0 ≤ x 2 + y 2 ≤ 1} depicted in Fig. 13.6.
We use constant test and ansatz functions in space and time. The right hand side is
given by f (t, x) = cos(|k|t − k · x) exp −1/(10t 2 )), where k = (0.2, 0.2, 0.2). The
time discretization errors are negligibly small in this numerical experiment, when
the time step is chosen to be Δt = 0.005. We compute the solution up to T = 1.
The finest graded mesh consists of 2662 triangles, and we use the solution on this
mesh as reference solution using the same Δt = 0.005.
We consider the error compared to the benchmark solution on the 2-graded mesh.
For the error as a function of the degrees of freedom, Fig. 13.7 shows convergence in
the energy norm with a rate −0.52 on the 2-graded mesh, respectively −0.26 on the
uniform mesh. The error therefore behaves in agreement with the approximation
properties proportional to ∼ h (equivalently, ∼ DOF − 2 ) on the 2-graded mesh,
1
while the convergence is ∼ h1/2 (∼ DOF −1/4 ) on a uniform mesh.

10 0
uniform; = -0.26
-graded, =2; = -0.52
-0.25
O(DOF )
-0.5
O(DOF )
relative energy error
10 -1
10 -2
10 4 10 5 10 6
DOF
Fig. 13.7 Energy error for single layer equation on circular screen, Example 13.1 [191]
Example 13.2 We compute the solution to the integral equation W φ = g on R+ t ×Γ

with the square screen Γ = [−1, 1]2 × {0}. We use linear ansatz and test functions
in space, linear ansatz and constant test functions in time. The right hand side is
given by
3 π π π 1
g(t, x) =(− +cos( (4−t))+ sin( (4−t))− (cos(π(4−t))+π sin(π(4−t))))
4 2 2 2 4
× [H (4 − t) − H (−t)],
where H is the Heaviside function, and Δt = 0.01, T = 4. The finest graded mesh
consists of 2312 triangles, and we use the solution on this mesh as reference solution
using the same Δt = 0.01.
Figure 13.8 shows the error in both the energy and L2 ([0, T ], L2 (Γ )) norms with
respect to the benchmark solution. The convergence rate in terms of the degrees
of freedom on the 2-graded mesh is −0.51 in energy and −1.05 in L2 . On the
uniform mesh the rate is −0.26 in energy and −0.50 in L2 . The rates on the 2-graded
meshes are in close agreement with a convergence proportional to ∼ h (equivalently,
∼ DOF −1/2 ) predicted by the approximation properties in the energy norm, and
∼ h1/2 (∼ DOF −1/4 ) on uniform meshes. Also in L2 norm, the convergence
corresponds to the expected rates: Approximately ∼ h2 (equivalently, ∼ DOF −1 )
on 2-graded meshes, ∼ h (equivalently, ∼ DOF −1/2 ) on uniform meshes. In all
cases the convergence is twice as fast on the 2-graded compared to the uniform
meshes.
0
10
relative L 2 and energy errors
-1
10
-2 energy error, uniform; = -0.26

10
energy error, -graded, =2; = -0.51
L error, uniform; = -0.50
2
L2 error, -graded, =2; = -1.05
-0.25
O(DOF )
10 -3 -0.5
O(DOF )
-1
O(DOF )
5 6
10 10
DOF
Fig. 13.8 L2 ([0, T ], L2 (Γ )) and energy error for hypersingular equation on square screen,
Example 13.2 [191]
Appendix A
Linear Operator Theory
Here we recall some concepts from linear operator theory, in particular compact
operators and Fredholm operators. It is assumed that the reader is familiar with
basics of linear functional analysis like norm, metric, completeness, Banach space,
Hilbert space etc. Some theorems are proved here, some are only stated.
Definition A.1 Let E, F be vector spaces (linear spaces) over a field Δ, a linear
operator from E to F is a mapping A : E → F such that
A(λx + μy) = λAx + μAy ∀λ, μ ∈ Δ; x, y ∈ E .
In the following, E, F are linear normed spaces.

Theorem A.1 Let A be a linear operator from E into F . Then A is bounded, if and
only if A is continuous, written A ∈ B(E, F ).
Definition A.2 The operator norm A = AE→F of A ∈ B(E, F ) is defined by
AxF
A := sup AxF = sup
x∈E x∈E\{0} xE
x≤1
Definition A.3 An operator A is said to be continuously invertible if A−1 exists

and is continuous.
Theorem A.2 (Banach’s Theorem) Let A ∈ B(X, Y ) and X, Y be Banach
spaces. Assume that A is injective and surjective (onto). Then A is continuously
invertible.

564 A Linear Operator Theory
Let E denote the dual space of E. Let M be an arbitrary nonvoid subset of E.

Then the set
{x ∈ E : x , x = 0 ∀x ∈ M}
is the annihilator of M, written M ⊥ . Similarly, N ⊥ ⊂ E for N ⊂ E .

Let A ∈ B(E, F ). Then the adjoint A is in B(F , E ). Moreover immediately
ker A = (im A )⊥ , ker A = (im A)⊥ . (A.1)
Also im A ⊂ (ker A )⊥ is obvious; however, equality holds only with closedness.

More precisely, due to Banach:
Theorem A.3 (Closed Range Theorem) Let A ∈ B(X, Y ) and X, Y be Banach
spaces. Then the following assertions are equivalent:
(i) im A is closed in Y ,
(ii) im A is closed in X ,
(iii) im A = (ker A )⊥ ,
(iv) im A = (ker A)⊥ .
A class of continuous linear operators that have closed range are Fredholm
operators, see below Definition A.8.
Definition A.4 Let X, Y be Banach spaces. A sequence of linear operators
{An : X → Y } is called strongly convergent (or pointwise convergent) to an
operator A : X → Y , written “An → A”, if for all x ∈ X, limn→∞ An x −AxY =
0.
Definition A.5 Let X, Y be Banach spaces. A sequence of linear operators
{An : X → Y } is called convergent in the norm, written “An ⇒ A”, if lim An −
n→∞
A = 0.
Theorem A.4 (Theorem of Banach-Steinhaus) The sequence {An : X → Y }
of continuous linear operators converges pointwise to a continuous linear operator
A : X → Y if and only if
(i) The sequence of operator norms An is bounded.
(ii) The sequence {An x} converges for all x ∈ M where M is a dense subset of X.
Definition A.6 Let X, Y be Banach spaces and A ∈ B(X, Y ) . Then A : X → Y
is said to be compact if the image of the closed unit ball B1X = {x ∈ X : x ≤ 1}
under A is relatively compact in Y .
Remark A.1 A set K ⊂ Y is relatively compact in Y if for every ε > 0 there is a
finite number of elements in K, say y1 , . . . , ym ∈ K, such that the ε-balls around
yj (j = 1, . . . , m) cover K, K ⊂ ∪m Y
j =1 Bε (yj ). Thus the yj are a finite ε-net in K.
Furthermore A is compact if and only if, for every bounded sequence {xn } in X, the
sequence {Axn } has a convergent subsequence in Y .
A Linear Operator Theory 565
Lemma A.1
(i) Let A : X −→ Y be compact and B : Y −→ Z bounded. Then, the operator
AB : X −→ Z is compact, too.
(ii) Let again A : X −→ Y be compact and let Bn −→ 0 be strongly convergent.
Then, Bn A −→ 0 is strongly convergent, too.
Proof (i) Exercise. (ii) Suppose the assertion does not hold. Then there exists a
sequence {xn } ⊂ X satisfying xn X = 1 ∀ n and Bn Axn Z ≥ α > 0 ∀ n . With
A compact there further exists a subsequence {xn } with Axn −→ y in Y . Thus,
0 < α ≤ Bn Axn ≤ Bn y + Bn (Axn − y)
But Bn y → 0 and Bn (Axn − y) ≤ MAxn − y → 0 which is a contradiction.

Definition A.7 Let X, Y be Banach spaces and {An } a sequence of linear operators.
Then {An : X → Y } is called collectively compact, if K := ∪n∈N An (B1X (0)) is
relatively compact in Y .
Lemma A.2 Assume that the sequence of the linear operators An : X → Y
converges pointwise to A. Then there holds:
(i) If K ⊂ X is relatively compact in X, then An ⇒ A uniformly in K.
(ii) Let B : W → X (with another Banach space W ) be a compact linear operator
and assume that Bn ⇒ B. Then An Bn ⇒ AB.
(iii) Let {Bn : W → X} be collectively compact and A = 0. Then An Bn ⇒ 0.
Proof
1. Assume that {x1 , . . . , xm } ⊂ K is a ε-net for a given (but arbitrary) ε > 0,
i.e. K ⊂ ∪m j =1 Bε (xj ). Choose a n0 (ε) ∈ N such that for all n > n0 (ε), j =
X
1, . . . , m, An xj − Axj Y < ε. Let x ∈ K arbitrary. Then there exists a j ∈

{1, . . . , m} with x − xj X ≤ ε and thus:
An x − Ax ≤ (An − A)(x − xj ) + (An − A)xj

≤ (An − A) (x − xj ) + ε
≤ (1 + An − A) · ε
By the Theorem of Banach-Steinhaus (see A.4) we have An − A ≤ C ∀n ∈

N and therefore An x − Ax ≤ (1 + C) · ε → 0 uniformly for all x ∈ K.
2. Since B is compact, {Bw|w ∈ W, w ≤ 1} ⊂ X is relatively compact in X.
Therefore by 1., for any ε > 0 there exists n0 ∈ N such that (An − A)Bw ≤ ε
for n ≥ n0 and w ≤ 1. Moreover since Bn ⇒ B, (Bn −B)w ≤ ε for n ≥ n0
and w ≤ 1. Therefore for w ≤ 1,
An Bn w − ABw ≤ An (Bn − B)w + (An − A)Bw

≤ An · (Bn − B)w + ε
≤ (1 + C)ε for w ≤ 1.
where An ≤ C by Banach-Steinhaus Theorem. Hence An Bn ⇒ AB.

3. From 1. we have that An converges uniformly to 0 = A on K := ∪n Bn (B1W (0)),
i.e. ∀ε > 0 ∃n0 : An x ≤ ε ∀n ≥ n0 ∀x ∈ K. Therefore An Bn w ≤ ε for
∀n ≥ n0 , and w ≤ 1 and so
An Bn W →Y := sup An Bn w −→ 0 for n → ∞.

w≤1

Definition A.8 Let X, Y be Banach spaces and let A ∈ B(X, Y ). Then A is called
a Fredholm operator, written A ∈ F (X, Y ), if A enjoys the following properties:
(I) The kernel ker A has finite dimension,
(II) im A is closed in Y ,
(III) the range im A has finite codimension: codim im A = dim (Y/im A) < ∞.
The number
ind (A) := dim ker A − codim im A
is called the (Fredholm) index of A.

Remark A.2 By the closed range theorem, here Theorem A.3, the operator equation
Au = f with A ∈ F (X, Y ) is solvable, if and only if, f ∈ (ker A )⊥ .
Theorem A.5 Let A ∈ F (X, Y ) and X, Y be Banach spaces. Then the adjoint A
is in F (X, Y ), too. Moreover,
dim ker A = codim im A and codim im A = dim ker A ,
hence
ind A = −ind A . (A.2)
Proof By A.1, ker A = (im A)⊥ ∼ = (Y/im A) = (coker A) . This shows the
first formula in the theorem. To prove the second formula, by the closedness of im
A, we can apply Theorem A.3, hence im A = (ker A)⊥ . This gives (ker A) ∼ =
X /(ker A)⊥ = X /im A = coker A , hence the second formula, (A.2), and also
A ∈ F (X, Y ).

A Linear Operator Theory 567
Without proof we list from [364, 435] the following.

Theorem A.6 Let A ∈ B(X, Y ) and X, Y be Banach spaces. Then A is a Fredholm
operator, if and only if, there exist Q1 , Q2 ∈ B(Y, X) such that
Q1 A = I − C1 in X and AQ2 = I − C2 in Y (A.3)
with compact operators C1 , C2 .

Let A ∈ F (X, Y ), B ∈ F (Y, Z) and X, Y, Z Banach spaces. Then B ◦ A ∈
F (X, Z) and
ind (B ◦ A) = ind B + ind A . (A.4)
For a Fredholm operator A and a compact operator C, the sum A+C is a Fredholm
operator and
ind (A + C) = ind A . (A.5)
The set of Fredholm operators is an open subset in the space of bounded linear
operators and the index is a continuous function.
To conclude Appendix A we recall from [259] Fredholm’s alternative for a
sesquilinear form a in a Hilbert space H under a Gårding inequality,
& '
% a(v, v) + (Cv, v)H ≥ α0 v2H , v∈H (A.6)
for a constant α0 > 0 and a compact operator C from H into H .

Theorem A.7 Suppose the continuous sesquilinear form a : H × H → C satisfies
Gårding inequality (A.6). Then for the variational equation
Find u ∈ H such that a(u, v) = (v) , ∀v ∈ H (A.7)
there holds the alternative:

Either
(A.7) has exactly one solution u ∈ H for every given ∈ H ∗
or
The homogeneous problem,
Find u0 ∈ H such that a(u0, v) = 0 , ∀v ∈ H (A.8)
and its adjoint problem,
Find v0 ∈ H such that a(u, v0 ) = 0 , ∀u ∈ H (A.9)

have finite dimensional kernels of the same dimension k > 0. The nonhomogeneous
problem (A.7) and its adjoint,
Find w ∈ H such that a(v, w) = ∗ (v) , ∀v ∈ H
have solutions iff the orthogonality conditions
(v0(j ) ) = 0, respectively, ∗ (u0(j ) ) = 0 for j = 1, . . . , k
hold where {u0(j ) }kj =1 spans the eigenspace of (A.8) and {v0(j ) }kj =1 spans the
eigenspace of (A.9), respectively.
Rewrite the variational equation (A.7) as
a(u, v) = (j Au, v)H = (f, v)H = (v) ,
where f ∈ H represents ∈ H ∗ due to the Riesz representation theorem. Then

the maps j A : H → H, A : H → H ∗ are linear and bounded. Moreover from
Fredholm’s alternative above, Theorem A.7 we derive the following
Remark A.3 Suppose the Gårding inequality (A.6). Then the linear operator A has
closed range, moreover A is a Fredholm operator with index zero.
Appendix B
Pseudodifferential Operators
Here we recall some concepts from Fourier transform and the theory of pseudodif-
ferential operators. These are especially used in Chap. 4. For further reading see e.g.
[253, 376, 415].
For 1 ≤ p < ∞ the space Lp (E) consists of all measurable functions f : E →
R (C) with norm
⎛ ⎞1/p

f Lp := ⎝ |f (x)|p dx ⎠ < ∞,
E
whereas L∞ (E) has norm f L∞ := ess supx∈E |f (x) | < ∞ .

Definition B.1 Let u ∈ L1 (Rn ). Then the Fourier transform of u is given by

û (ξ ) = (F u) (ξ ) = e−ixξ u (x) dx (B.1)
Rn
with its inverse

=u (ξ ) = (2π)−n
ũ (ξ ) = F eixξ u (x) dx. (B.2)
Rn
Clearly û ∈ L∞ (Rn ) , û L∞ ≤ u L1 .

Definition B.2 (Rapidly Decreasing Functions) ϕ ∈ S (Rn ) ⇐⇒ ϕ ∈
C ∞ (Rn , C) , supx∈Rn | < x >α D β ϕ (x) | < ∞ ∀α, β ∈ Nn0 , with
< x >:= 1 + |x|.
Proposition B.1 u ∈ S (Rn ) ⇒ û, ũ ∈ S (Rn ) .

570 B Pseudodifferential Operators
Proof u ∈ S (Rn ) ⇒ x α u ∈ S (Rn ) ∀α ∈ Nn0 ; D β u ∈ S (Rn ) ∀β ∈ Nn0 .

Therefore we have D β (x α u) ∈ S (Rn ) ⊂ L1 (Rn ) and consequently F D β (x α u)
is absolutely convergent. So we can interchange differentiation and integration in
(B.1). With ∂j := ∂x∂ j and Dj := 1i ∂j we have
Dξα e−ixξ = (−i)|α| (−ix)α e−ixξ = (−1)|α| x α e−ixξ
and thus ξ β e−ixξ = (−1)β Dx e−ix . It follows that

β

Dξα û (ξ ) = (−x)α u (x) e−ixξ dx = F (−x)α u (ξ ) (B.3)
Rn
and

ξ β û (ξ ) = (−i)β u (x) Dxβ e−ixξ dx (B.4)
Rn

= Dxβ u (x) e−ixξ dx = F D β u (ξ ) .
Rn
Applying (B.3) and (B.4) yields

ξ β D α û (ξ ) = F D β (−x)α u (ξ ) ∈ L∞ Rn ∀α, β ∈ Nn0 .
Hence û ∈ S (Rn ).

==
Proposition B.2 The map F : S (Rn ) → S (Rn ) is an isomorphism with F F
=
F F = 1. Moreover, for u ∈ S (R ) there holds:
n

−n
u(x) = (2π) û (ξ ) eixξ dξ. (B.5)
Rn
Remark B.1 By Proposition B.1 we have u = F = (F (u)) for all u ∈ S (Rn ) .

Moreover, F and F = are related to each other via (F u) (x) = (2π)n F =u (−x) .
It follows that
F F u (x) = (2π)n u (−x) . (B.6)
Remark B.2 With equation (B.5) we write the differential operator

N
p (x, D) = aα (x) Dxα as
|α|=0

−n
p (x, D) u (x) = (2π) eixξ p (x, ξ ) û (ξ ) dξ.
Rn
B Pseudodifferential Operators 571
N
Definition B.3 p (x, ξ ) = |α|=0 aα (x) ξ α is called the symbol of p (x, D) .
Definition B.4 A tempered distribution T on Rn is a continuous linear functional
T : S (Rn ) → C. The linear space of tempered distributions is denoted by
S (Rn ) .
Next we extend F to a map on tempered distributions:
Find F + : S (Rn ) → S (Rn ) such that

< F f, ϕ > = < f, F + ϕ > ∀f ∈ S Rn , ϕ ∈ S Rn ,
< f, ϕ >:= f (x)ϕ(x) dx ∀ϕ ∈ S (Rn ) and |< f, ϕ >| ≤ c bounded with

Rn
c ∈ R.
Then we define
F |S (Rn ) := (F + |S (Rn ) ) .
If f, ϕ ∈ S (Rn ) ⊂ S (Rn ) , then by Fubini’s theorem

< F f, ϕ > = f (x) e−ixξ dx ϕ (ξ ) dξ
Rn Rn

= f (x) e−ixξ ϕ (ξ ) dξ dx = < f, F ϕ > .
Rn Rn
So we get that F + = F on S (Rn ) .

Definition B.5 If f ∈ S (Rn ) , then F f = fˆ ∈ S (Rn ) is defined by

< F f, ϕ > = < f, F ϕ > ∀ϕ ∈ S Rn . (B.7)
(B.7) is equivalent to

=ϕ ∀ϕ ∈ S Rn
(F f, ϕ) = (2π)n f, F (B.8)
with (f, ϕ) :=< f, ϕ̄ > .

Remark B.3 Also on S (Rn ) there holds F F ==F =F = 1.
Furthermore , F |S (Rn ) is continuous extension of F |S (Rn ) .
For u ∈ S (Rn ) there holds

D α u = F −1 ξ α û , (B.9)

(−x)β u = F −1 D β û . (B.10)
For all u ∈ S (Rn ) there holds Parseval’s equality:

⎛ ⎞
û 2L2 (Rn ) = ⎝F u, =ϕ) = (2π)n u 2 2 n .
F u ⎠ = (2π)n (u, F L (R )
=:ϕ
−n
Thus u → (2π) 2 û is an isometry on L2 (Rn ) . There holds
−n −1 ∗
(2π) 2 F
n
= = (2π) −n
= (2π) 2 F 2 F ,
where the adjoint A∗ of a given operator A is defined by

(Af, ϕ) = f, A∗ ϕ ∀ϕ ∈ S Rn , f ∈ S Rn .
Definition B.6 For all u, v in S (Rn ) the convolution is defined by (u ∗ v) (x) :=

Rn u (x − y) v (y) dy.
The Fourier transform satisfies:

û ∗ v̂ (ξ ) = û (ξ − η) v̂ (η) dη = < û (ξ − ·) , v̂ >
Rn
=< F û (ξ − ·) , v > = (2π)n u?

· v (ξ ) .
The last equation follows from

û (ξ − η) = =x→η e−ixξ u (x) (η)
u (x) e−ix(ξ −η) dx = (2π)n F
Rn
Thus

Fη→x û (ξ − η) (x) = (2π)n e−ixξ u (x)
and

< F û (ξ − ·) , v > = (2π) n
e−ixξ u (x) v (x) dx = (2π)n u?
· v (ξ ) .
Rn
The following formulae are valid for u, v ∈ S (Rn ):
û ∗ v̂ = (2π)n u?
·v (B.11)
u
∗ v = û · v̂. (B.12)
We prove (B.12): ũ?

· ṽ = ũˆ ∗ ṽˆ (2π)−n = u ∗ v (2π)−n .
Hence by Remark B.1
∗ v (x) = (2π)n F ũ?

u · ṽ (x) = (2π)2n (ũ · ṽ) (−x)

= (2π)2n ũ (−x) ṽ (−x) = (2π)2n (2π)−n û (x) (2π)−n v̂ (x) = û · v̂ (x)
which completes the proof.

Next we give a brief introduction in pseudodifferential operators and symbol-
classes. We define for s ∈ R

2
||φ||H s (Rn ) := (2π) −n
(1 + |ξ |2 )s |>
φ(ξ )|2 dξ φ ∈ S (Rn ) (B.13)
Rn
In the following we also use the abbreviation < ξ >2s := (1 + |ξ |2 )s .

We remark
||φ||2H 0 (Rn ) ≡ ||φ||2L2 (Rn )
and introduce the Sobolev space
H s (Rn ) := {ϕ ∈ S (Rn )|>

ϕ ∈ L2loc (Rn ), ||ϕ||H s (Rn ) < ∞}
= completion of S (Rn ) with || · ||H s (Rn )
This gives rise to the Bessel potential operator Λs : H t → H t −s ∀t, s ∈ R.

The differentiation D α : H s −→ H s−|α| is continuous and for all 0 ≤ s ≤ t,
S ⊂ H t ⊂ H s ⊂ H 0 = L2 ⊂ H −s ⊂ S .
Now we introduce symbol classes. Let p(x, D) = |α|≤k aα (x)D α . Then we

have

−n
p(x, D)u(x) = (2π) u(ξ )dξ , ∀u ∈ S (Rn ) .
eixξ p(x, ξ )> (B.14)
Rn
By the property of the Fourier transform,
?
F −1 D α u = F −1 (ξ α>
u) ,
hence

p(x, D)u(x) = ?
aα (x)F −1 D αu = aα (x)F −1 (ξ α>
u) = F −1 p(x, ξ )>
u.
Example B.1
s
@ =
(i) p(ξ ) =< ξ >s (= (1 + |ξ |2 ) 2 ). This gives p(D) = (1 − Δ)s , since −Δu
2
− dx
d
2 u = −(i ξ ) >
2 u(ξ ) = ξ 2>
u(ξ ).
(ii) p(ξ ) = e ia·ξ with a ∈ R fixed gives p(D)u(x) = u(x + a).
(iii) For n = 1, define using the Poisson kernel,
∞
1 tf (y)
Kt f (x) := dy .
π t2 + (x − y)2
−∞
Then Kt = p(x, D) with p(x, ξ ) = e−t |ξ | .

Next we write (B.14) for some special functions p(x, ξ ).
p(x, ξ ) = p(x) gives a multiplication operator that maps C0∞ → C ∞
p(x, ξ ) = p(ξ ) gives all convolution operators S → S , provided p ∈ S ,
(respectively L2 → L2 ⇔ p ∈ L∞ ).
In the general case, formal calculus gives

p(x, D)u(x) = (2π)−n eix·ξ p(x, ξ )( e−iy·ξ u(y) dy) dξ

= (2π)−n e−i(x−y)·ξ p(x, ξ )u(y) dy dξ

= p̌(x, x − y)u(y) dy ,
where p̌(x, z) = (2π)−n eiz·ξ p(x, ξ ) dξ or equivalently,
p(x, ξ ) = Fz→ξ (p̌(x, z))(ξ ) (B.15)
Hence p(x, D) has a kernel k(x, y) = p̌(x, x − y) or equivalently, p̌(x, z) =

k(x, x − z), and thus

p(x, ξ ) = e−iz·ξ k(x, x − z) dz
= F k(x, x − ·)(ξ ) if k(x, x − ·) ∈ S .
Let us introduce the following notions.

(i) Let Ω open in Rn , fix m, ρ, δ ∈ R with ρ ≤ 1, δ ≥ 0. Then define
m
Sρ,δ (Ω) :={p ∈ C ∞ (Ω × Rn )|∀K ⊂⊂ Ω ∀α, β ∈ Nn0 ∃ CK,α,β :
(B.16)
β
|Dxα Dξ p(x, ξ )| ≤ CK,α,β < ξ >n−ρ|β|+δ|α| , ∀x ∈ K, ξ ∈ Rn }
In the following we often have ρ = 1, δ = 0.

(ii) Classical symbols:
p ∈ S m (Ω) :⇔ p ∈ S1,0
m and
∃ sequence (pm−j )j ∈N0 ⊂ C ∞ (Ω × Rn ) with
pm−j (x, rξ ) = r m−j pm−j (x, ξ ) , ∀ |ξ | ≥ 1, r ≥ 1
(positively homogenous of degree m − j for |ξ | ≥ 1, r ≥ 1 ) that decays faster

than any power,

N
m−N−1
p− pm−j ∈ S1,0 (Ω) (⇔ p ∼ pm−j )
j =0 j ≥0
(iii)
L L
S −∞ := Sm = m
Sρ,δ (independent of ρ, δ)
m m
Note that when v(x) −→ > v (ξ ), the asymptotic behaviour of v for small x
corresponds to the asymptotic behaviour of >v for large ξ .
Remember

(v ∗ u)(x) = v(x − y)u(y) dy v∗ u(ξ ) = >
v (ξ )>
u(ξ )
and consider the following

Example B.2
Δu = 0 in Ω = R+
2 , u = g on ∂Ω = R1
Then

1
u(x) = − ln |x − y|φ(y) dy
2π R
satisfies Δu = 0 in Ω. We have the convolution u(x) = v(x − y) φ(y) dy with

the simple kernel v(x) = ln |x|. Hence > v (ξ )>
g (ξ ) = > φ (ξ ) and thus
−1 >g
φ(x) = Fξ−1 >
→x φ (ξ ) = Fξ →x (ξ ) .
>
v
More generally let

u = Aφ , Aφ(x) = (2π)−n eix·ξ a(x, ξ )>
φ(ξ ) dξ
Here the singularity of the kernel k(x, x − y) of A when x → y is determined by

the behaviour of a(x, ξ ) for |ξ | → ∞.
Example B.3
(i) Let p(x, ξ ) = |α|≤k aα (x)ξ , then p(x, ξ ) ∈
α S k . Here we have
β
pk−j (x, ξ ) = α Dξ ξ α = Cα β ξ α−β ,
|α|=k−j aα (x)ξ . Note that we can write
∀ β ≤ α with some constants Cα β .
(ii) Let q(ξ ) ∈ C ∞ (Rn \ {0}) be homogeneous of degree m,
q(rξ ) = r m q(ξ ) , ∀r > 0, ξ = 0 .
Let χ ∈ C0∞ (Rn ) with χ ≡ 1 in a neighborhood of 0 and supp χ ⊂⊂ B1 (0).

Then put p(x, ξ ) = p(ξ ) := (1 − χ(ξ ))q(ξ ). Hence there holds p ∈ S1,0m
(Rn )
(even ∈ S (R )). Note that we can modify symbols for small |ξ |, since we
m n
are interested in the behaviour for large |ξ |.

−2
(iii) Let p(ξ ) =< ξ >−2 = (1 + |ξ |2 )−1 . Then p ∈ S −2 ⊂ S1,0 .
∞
Indeed, p ∈ C , D p(ξ ) = (1 + |ξ | )
α 2 −1−|α| |α|
· h (ξ ) with an appropriate
−2
polynomial h. Hence |D α p(ξ )| ≤ C < ξ >−2−|α| and p ∈ S1,0 follows.
Moreover, we use the asymptotic expansion
∞
1 1
= |ξ |−2 −2
=− (−1)k |ξ |−2k (|ξ | > 1)
1 + |ξ | 2 1 + |ξ |
k=1
and put
p−2k (ξ ) := (1 − χ(ξ ))(−1)k+1 |ξ |−2k

−2k
Then as seen above in (2.) p−2k ∈ S1,0 and hence

N
−2N−2 −2N−1
p(ξ ) − p−2k (ξ ) ∈ S1,0 ⊂ S1,0 (N ≥ 1)
k=1
(iv) Let p(ξ ) =< ξ >s . Then p ∈ S s ⊂ S1,0

s
, ∀s ∈ R
D α1 u D αk u
Lemma B.1 D α ( u1 ) = 1
u k≤|α| Cα1 ,...,αk u · ...· u
Theorem B.1 Let p ∈ S1,0
m (Ω) and
1
| | ≤ c < ξ >−m for |ξ | ≥ 1 (⇐⇒ elliptic)
p(x, ξ )
1−χ −m
Then p ∈ S1,0 (Ω).
Proof We have
β
Dxα1 Dξ λ p(x, ξ )
| | ≤ C < ξ >m−|βλ | < ξ >−m = C < ξ >−|βλ |
p(x, ξ )
and hence by the lemma above
β 1 B β
Dxα1 Dξ λ p
|Dxα Dξ | ≤ | p(x,ξ
1
)| Cαβ | p |
p(x, ξ ) ακ =α
βλ =β
≤ C < ξ >−m < ξ >−|β| .

m (Ω), q ∈ S (Ω) . Then
Theorem B.2 Let p ∈ S1,0 m
1,0
β m−|β| m+m
Dxα Dξ p ∈ S1,0 (Ω) and p · q ∈ S1,0 (Ω) .
Proof Use Leibniz rule

β α β α β β
Dxα Dξ (p · q)(x, ξ ) = D D p(x, ξ ) D α
Dξ q(x, ξ )

α β x ξ x
α +α =α

β +β =β

and estimate the partial derivatives of p by < ξ >m−|β | , respectively the partial

derivatives of q by < ξ >m −|β | modulo some positive constant factor, what leads

to the upper bound < ξ >m+m −|β| modulo a positive constant.

m
We define a pseudodifferential operator of class S1,0
p(x, D) ∈ OP S1,0
m
(Ω) :⇔ p(x, ξ ) ∈ S1,0
m
(Ω)
with

−n
p(x, D)u(x) = (2π) eix·ξ p(x, ξ )>
u(ξ ) dξ , ∀u ∈ C0∞ (Rn ) (B.17)
Rn
Theorem B.3 Let m ∈ R, , p(x, D) ∈ OP S1,0

m (Ω). There holds
p(x, D) : C0∞ (Ω) −→ C ∞ (Ω) continuous, linear.
Proof Now u ∈ C0∞ (Rn ) implies > u ∈ S (Rn ). Hence p(x, ξ )> u(ξ ) still decays
fast and the integral converges absolutely. Hence interchanging differentiation and
integration yields with (B.17)

|Dxα (p(x, D)u(x))| ≤ c < ξ >m < ξ >|α | |>
u(ξ )| dξ < ∞ ∀α

α +α =α
Thus p(x, D)u ∈ C ∞ .

Exercise: Show the mapping p(x, D) ∈ OP S1,0 m : H s (Rm ) −→
H s−m (Rm ) is continuous, that is, there exists C > 0 such that p(x, D)uH s−m ≤
C uH s .
Next we consider the relation between strong ellipticity of a pseudodifferential
operator and Gårding’s inequality. As shown in Sect. 4.2 with the example of the
single layer operator, considering integral operators as pseudodifferential operators
allows to deduced the mapping properties of boundary integral operators by exam-
ining the symbols of the pseudodifferential operators. On the other hand, Garding’s
inequality for integal equations is the key property to guarantee convergence of
Galerkin’s method, see Theorem 6.1, Theorem 6.11. Now, Garding’s inequality
follows from the definition of uniform strong ellipticity of pseudodifferential
operators, see Theorem 6.2.7 in [259].
s +t
Definition B.7 A system of pseudodifferential operators Aj k ∈ OP S1,0 j k
(Ω) is
called uniformly strongly elliptic if for the principal part matrix a ( x ; ξ ) =
0
((asj +tk (x; ξ )))p×p there exist a C ∞ -matrix valued function Θ(x) =
j k0
((Θj k (x)))p×p and a constant γ0 > 0 such that
%ζ T Θ(x)a 0(x, ξ )ζ ≥ γ0 |ζ |2
for all x ∈ Ω, ζ ∈ Cp and ξ ∈ Rn with |ξ | = 1.

A uniformly strongly elliptic system of pseudodifferential operators satisfies a
Gårding inequality, see [259, Theorem 6.2.7.] . In the following we present and
prove the corresponding result for a single pseudodifferential operator:
Theorem B.4 (Gårding Inequality) Let p(x, ξ ) ∈ OP S1,0 m (Ω) be strongly
elliptic , i.e. ∀K ⊂⊂ Ω let there exist positive constants CK , RK such that there
holds
%p(x, ξ ) ≥ CK < ξ >m ∀x ∈ K, |ξ | ≥ RK
Then ∀K ⊂⊂ Ω and ∀s ∈ R there exist constants γK , CK,s such that
%(p(x, D)u, u) ≥ γK ||u||2H m/2 (Ω) − CK,s ||u||sH s (Ω)
∀u ∈ C0∞ (Ω).
Lemma B.2 Let p ∈ S1,0 0

(Ω), %p(x, ξ ) ≥ C > 0 ∀x, ξ ( |ξ | sufficiently large)
then there exist B ∈ OP S1,0 (Ω), K ∈ OP S −∞ (Ω) such that
m
%p(x, D) = B ∗ B + K.
Proof Setting q(x, D) := Λ− 2 pΛ− 2 we have for u ∈ C0∞ (pu, u) =

m m
m m m
(qΛ 2 u, Λ 2 u) and ||u||2H 1/2 ∼ ||Λ 2 u||2L2 . Therefore it suffices to show for m = 0.
Now we use the above lemma for p0 (x, ξ ) := %p(x, ξ ) − c , with 0<c’<c.
Hence p0 (x, ξ ) ≥ c − c > 0. Then there exists b ∈ OP S1,0 0 such that
∗ ∞
%p0 (x, D) − B B =: S ∈ OP S such that
%(p(x, D)u, u) − c ||u||2L2 = ||Bu||2L2 + %(Su, u)
and this yields
%(p(x, D)u, u) ≥ c ||u||2L2 + %(Su, u).

As a consequence, any strongly elliptic pseudodifferential operator defines a
Fredholm operator of index zero since for the corresponding bilinear form one may
apply the classical Fredholm alternative (A.3).
Example B.4 Writing the single layer potential as

u(ξ )
>
V ψ(x) = (2π)1−n eix ξ dξ
Rn−1 |ξ |
gives

? 1
(V ψ(x )ψ(x )dx = %(V > )) = %(2π)1−n
ψ(ξ ), ψ(ξ ψ(ξ > )dξ
> )ψ(ξ
Γ Rn−1 |ξ |
≥ γ ||ψ||2H −1/2 (Γ ) − compact perturbation .
Example B.5 The single layer potential in linear elasticity with fundamental solu-
tion
E(x, y) = 1/|x − y|I + κ(x − y)(x − y)T

λ+μ
where κ = λ+3μ has principal symbol
⎛ 2 ⎞
|ξ | + κξ22 −κξ1 ξ2 0
λ + 3μ 1 ⎝
σ0 (V )(ξ ) = −κξ1 ξ2 |ξ |2 + κξ12 0 ⎠
2μ(λ + 2μ) |ξ |3
0 0 |ξ |2
The corresponding hypersingular operator has principal symbol

⎛ 2 ⎞
|ξ | + ξ12 ξ1 ξ2 0
−μ2 ⎝ ⎠
σ0 (W )(ξ ) = ξ1 ξ2 |ξ |2 + ξ22 0
|ξ |
0 0 (1 + )|ξ |2
with −1/2 < := λ(λ + 2μ)−1 < 1 see [130, 396].

Appendix C
Convex and Nonsmooth Analysis,
Variational Inequalities
C.1 Convex Optimization, Lagrange Multipliers
By this section of Appendix C we invite the reader to get acquainted with

some fundamental concepts, methods, and results of convex optimization that are
necessary for the proper understanding of the mathematical and numerical treatment
of inequality constrained problems that occur in the Signorini boundary value
problem and in further nonsmooth boundary value problems, see Chap. 5, and in
contact problems, see Chap. 11 and also Sect. 12.5
Based on the monograph [55] of Blum and Oettli, we start with convex quadratic
optimization in finite dimensions. Already at this level we encounter different
formulations, namely a “primal” and a “mixed” formulation with signed Lagrange
multipliers that are associated to inequality constraints. In fact, the existence of such
Lagrange multipliers can be derived from the celebrated duality theory of linear
optimization (“linear programming”) without any further assumptions. Moreover, a
solution in convex quadratic optimization is characterized by a “linear complemen-
tarity problem” and by a variational inequality (VI) of a special structure.
Then we proceed to convex variational problems in Hilbert space. As a straight-
forward extension of the finite dimensional case, we characterize solutions by vari-
ational inequalities with symmetric bilinear forms. Also guided by the finite dimen-
sional case, we readily introduce the Lagrange function for convex cone constraints.
However, the existence of Lagrange multipliers is more involved than in the finite
dimensional case. First we construct the Lagrange multiplier in the space dual to the
solution space of the primal variable, which is a Sobolev space of negative order in
application to contact problems. Then in the subsequent subsection we follow [219]
and present mixed formulations with Lagrange multipliers that live in the Hilbert
space of constraints, which is the more regular L2 function space on the contact
boundary part in the application to unilateral contact problems. To this end we pro-
vide an extension of the famous Brezzi splitting theorem that originally covers sad-

582 C Convex and Nonsmooth Analysis
dle point problems with equality constraints, only, to a class of nonsmooth inequality
constrained variational problems. Under the celebrated Babuška-Brezzi condition
we obtain independent Lagrange multipliers in the ordering cone of the inequality
constraints and in the subdifferential of the convex nonsmooth sublinear functional.
C.1.1 Convex Quadratic Optimization in Finite Dimensions
For given data b ∈ Rn, d ∈ Rm, C ∈ Rm×n, A ∈ Rn×n , where A is a symmetric

positive semidefinite matrix, shortly A = AT ≥ 0, we consider the convex
finite dimensional quadratic optimization problem with linear inequality constraints
(“quadratic program”)

minimize f (x) = 12 x T Ax − bT x
(QP)
subject to x ≥ 0, Cx ≤ d .
Put in another way, among all feasible solutions x to (QP), that is, x ∈ Rn+ , shortly
x ≥ 0, that satisfy the constraints (Cx)j ≤ dj (∀j = 1, . . . , m) we are looking for
that feasible x̂ that minimizes the objective function f .
The symmetry requirement A = AT is not essential, since we can replace the
1
matrix A by its symmetric part (A + AT ) in the objective function f . In the formal
2
discussion to follow, considering only signed variables xi for i = 1, . . . , n does
not lead to a loss of generality either, since for a free variable xi we can use its
decomposition xi = xi+ − xi−, xi+ ≥ 0, xi− ≥ 0. Also an equality constraint cjT x =
dj can be rewritten as
% ,
cjT x ≤ dj
.
−cjT x ≤ −dj
Of course, these two latter trivial reformulations are not appropriate in numerical
computation, but are convenient here to reduce the discussion of constrained
optimization problems to the standard form (QP) given above.
Now we take (QP) as primal optimization problem (“primal program”) and
proceed to its mixed formulation via the Lagrange function
L(x, y) = f (x) + y T (Cx − d) .
In view of the sign conditions and the inequality constraints, the Lagrange function
is considered only for x ≥ 0, y ≥ 0, since we have for any x ≥ 0,

f (x) if x is feasible;
sup L(x, y) =
y≥0 +∞ otherwise.
C Convex and Nonsmooth Analysis 583
This gives
inf sup L(x, y) = inf (QP) ,

x≥0 y≥0
where inf (QP) denotes the optimal value of (QP). Therefore in the sense of convex
duality theory, the dual optimization problem (“dual quadratic program”) to (QP)
reads
%
maximize inf L(x, y)
(DQP) x≥0
subject to y ≥ 0 .
Obviously, inf sup L ≥ sup inf L is trivial. But in finite dimensions, without further
assumptions, we have even the “duality equality” inf (QP) = sup (DQP); moreover,
the dual problem attains an optimal solution, what is nothing else than a Lagrange
multiplier to the inequality constrained optimization problem (QP):
Theorem C.1 If (QP) has an optimal solution x̂, then there exists a Lagrange
multiplier ŷ ≥ 0 such that (x̂, ŷ) is a saddle point of L on Rn+ × Rm
+ , that is,
we have
(SP) L(x̂, y) ≤ L(x̂, ŷ) ≤ L(x, ŷ); ∀x ∈ Rn+ , y ∈ Rm

+.
Proof To prove (SP) it is enough to establish the Karush–Kuhn–Tucker conditions,

which read for the linear constraints in (QP) here
(KKT) f (x̂) ≤ f (x) + ŷ T (Cx − d) ∀x ∈ Rn+ .
Indeed, in view of feasibility, C x̂ − d ≤ 0, (KKT) implies the equality
(*) ŷ T (C x̂ − d) = 0 .
Hence, the right hand side of (SP) follows from (KKT) directly, whereas the left
hand side of (SP) is equivalent to (*) and the feasibility of x̂.
Therefore it remains to show the existence of ŷ ∈ Rm + that satisfies (KKT). Here
we rely on the duality theorem of finite dimensional linear optimization (“linear
programming”) and first show the following
Proposition C.1 Let x̂ be an optimal solution to (QP). Then x̂ is an optimal
solution to the linear program

minimize (Ax̂ − b)T x =: cT x
(LP )
subject x ≥ 0, Cx ≤ d .
Proof of the Proposition Since the constraints of (QP) and (LP) are the same, it is
enough to give the following contradiction argument. Suppose there exists x̃ ∈ Rn+
such that cT x̃ < cT x̂ and C x̃ ≤ d. Then consider xt = x̂ + t (x̃ − x̂), where

0 < t < 1; xt is feasible for (QP). By ∇f (x̂)T (x̃ − x̂) = cT (x̃ − x̂) < 0, for small
enough t > 0, we arrive at f (xt ) < f (x̂) contradicting the optimality of x̂.

Proof of the theorem continued. The dual linear optimization problem (“dual pro-
gram”) to (LP) reads

maximize − d T y
(DLP )
subject y ≥ 0, C T y + c ≥ 0 ;
this can be seen by means of the associated Lagrange function
l(x, y) = cT x + y T (Cx − d) = −d T y + x T (c + C T y)
on Rn+ × Rm
+ and by the relation

−d T y if C T y + c ≥ 0 ;
inf l(x, y) =
x≥0 −∞ otherwise.
In virtue of the duality theorem of linear programming, see [55, 140], there exists
ŷ ∈ Rm+ such that
(i) C T ŷ ≥ −c ,
(ii) cT x̂ = −d T ŷ .
Then multiplying (i) by arbitrary x ≥ 0 gives
x T C T ŷ ≥ −x T Ax̂ + bT x ,
hence by (ii)
(Cx − d)T ŷ ≥ x̂ T A(x̂ − x) + b T x − b T x̂ .
Thus we obtain
f (x) + (Cx − d)T ŷ ≥ x̂ T A(x̂ − x) + 12 x T Ax − bT x̂

= 12 x̂ T Ax̂ + 12 [(x̂ − x)T A(x̂ − x)] − bT x̂ ,
and since A is positive semidefinite,
1 T
f (x) + (Cx − d)T ŷ ≥ x̂ Ax̂ − bT x̂ ,
2
what is the claimed (KKT) inequality.

We remark that the saddle point inequalities (SP) are clearly also sufficient for
the optimality of x̂.
We can characterize the optimality of x̂ in another way using slack variables.
Define the primal slack variable
v = d − Cx ∈ Rm ,
then feasibility is equivalent to v ≥ 0 and (*) reads v̂ T ŷ = 0 with v̂ = d − C x̂.

Likewise define the dual slack variable
u = c + C T y = Ax̂ − b + C T y ∈ Rn .
Then for y ≥ 0, feasibility in (DLP) is equivalent to u ≥ 0 and with û = c + C T ŷ =

Ax̂ − b + C T ŷ, we conclude from (ii) and (*) that
(**) x̂ T û = 0 .
Since v̂j ≥ 0, ŷi ≥ 0, ûi ≥ 0, x̂i ≥ 0, (*) means v̂j = 0 or ŷj = 0 and (**) means
ûi = 0 or x̂i = 0. In this sense v̂ and ŷ, respectively û and x̂ are “complementary
variables”.
Altogether we obtain the following
Corollary C.1 x̂ is an optimal solution to (QP), if and only if (x̂, ŷ, û, v̂) ∈ Rn+m ×
Rn+m satisfies

u A CT x −b
= +
v −C 0 y d
T
u x u x
≥0 ≥0 =0.
v y v y
The above system of linear equations and sign inequalities can be considered as
a “mixed formulation” of the convex quadratic optimization problem (QP). It leads
to the
Definition Let F : RN → RN be given. Then the “complementarity problem”
consists in finding ẑ ∈ RN
+ such that F (ẑ) ∈ R+ and ẑ F (ẑ) = 0 hold. We have
N T
a “linear complementarity problem” (LCP), if F (z) = Bz − a is affine-linear for

some B ∈ RN×N , a ∈ RN .
Thus the solution of (QP) can be characterized as the solution of a special (LCP),
where the matrix B has the special saddle point structure

A CT
B= .
−C 0
Furthermore a solution ẑ to the complementarity problem can be characterized by

the following “variational inequality”:
ẑ ∈ RN
+ , F (ẑ)T (z − ẑ) ≥ 0 ∀z ∈ RN
+ .
Indeed, the direct implication being obvious, only the reverse implication needs
an argument; for that choose z = 12 ẑ and z = 2ẑ. In the case of a linear
complementarity problem, the variational inequality reads
ẑ ∈ RN
+ , (B ẑ)T (z − ẑ) ≥ a T (z − ẑ) ∀z ∈ RN
+ .
Remark Also in the case of a linear complementarity problem, the solution

generally depends nonlinear on the data, e.g. on the datum a!
An unessential extension of the problem is obtained by a simple translation: Let
c ∈ RN be given; find z ∈ RN , such that z ≥ c, F (z) ≥ 0, (z − c)T F (z) = 0.
An essential extension of the problem is obtained as follows. Instead of R+ N,
consider an arbitrary convex cone K (that is, K + K ⊆ K, R+ K ⊆ K) in RN ,

not necessarily polyhedric, define the positive polar cone K + = {u ∈ RN |uT x ≥
0, ∀x ∈ K}. Then the complementarity problem consists in finding ẑ ∈ RN such that
ẑ ∈ K, F (ẑ) ∈ K + , ẑT F (ẑ) = 0. Again, this can be characterized by a variational
inequality. In the case of a linear complementarity problem with F affine-linear as
above, this variational inequality reads
ẑ ∈ K , (B ẑ)T (z − ẑ) ≥ a T (z − ẑ) ∀z ∈ K .
For more information on linear complementarity problems and variational

inequalities in finite dimensions we refer to the monographs of Cottle, Pang, and
Stone [139] and of Facchinei and Pang [168, 169], respectively.
C.1.2 Convex Quadratic Optimization in Hilbert Spaces
Let V be a real Hilbert space (may be also a reflexive Banach space) and Z another
real Hilbert space with its dual Z . Let A ∈ L (V , V ) with A = A , A ≥ 0 (i.e.
Av, v ≥ 0, ∀v ∈ V ), further B ∈ L (V , Z ) and f ∈ V , g ∈ Z fixed elements.
We also need the adjoint B ∈ L (Z, V ). Moreover let an order ≤ be defined in Z
via a convex closed cone P ⊂ Z via z ≥ 0 iff z ∈ P . Also ζ ∈ Z ≤ 0 iff ζ lies in
the negative dual cone P − = { ζ ∈ Z : ζ (p) ≤ 0, ∀p ∈ P }. With these given data,
similar to (QP ) in C.1.1, we consider the convex quadratic optimization problem

minimize f (v) = 12 Av, v − f, v
(CP )
subject to Bv ≤ g .
This gives rise to the bilinear form a(u, v) := Au, v and the convex closed
sets,
K(g) := {v ∈ V | Bv ≤ g}
which is a translate of the convex closed cone (with vertex at zero)
K := {v ∈ V | Bv ≤ 0}.
As in C.1.1 a solution u of (CP ) is characterized by a variational inequality, here
u ∈ K(g), a(u, v − u) ≥ f, v − u , ∀v ∈ K(g) . (C.1)
Analogously to C.1.1 , we introduce the Lagrangian
L(v, p) := f (v) + p, Bv − g Z×Z = f (v) + B p, v V ×V − g(p), v ∈ V, p ∈ P ,
to arrive at saddle points and to mixed formulations.

We can drop the requirement that A = A and now start from the primal
VI (C.1). However, the existence of Lagrange multipliers in the cone P in the
infinite dimensional space Z is more involved than in finite dimensions. As the
recent paper [219] shows, this can be accomplished by an extension of the Brezzi
splitting theorem under the Babuška-Brezzi condition. We postpone a sketch of this
approach to Lagrange multipliers to the next subsection.
Before that we describe here first an easier approach under the assumption that
there exists a preimage w of g under B, thus Bw = g. This allows to to work with
the duality on V × V and to obtain the following characterization via multipliers in
the negative dual cone K − to K.
Proposition C.2 u ∈ K(g) solves the VI (C.1), iff there exists λ ∈ V such that
(u, λ) ∈ K(g) × K − solves the mixed system

a(u, v) + λ, v = f, v
(MP )
κ − λ, u − w ≤ 0 ,
for all v ∈ V , κ ∈ K − . Then there holds the complementarity condition
(∗) λ, u − w = 0 .
Proof Let u ∈ K(g) solve the VI (C.1). Define λ ∈ V by λ(v) = f (v)−a(u, v).
Then (MP )1 holds. Further, for any v ∈ K, ṽ := v + u lies in K(g) and hence
−λ(v) = a(u, ṽ − u) − f (ṽ − u) ≥ 0 .

Thus λ ∈ K − . Since w ∈ K(g), u − w ∈ K,
κ − λ, u − w = κ, u − w + [a(u, u − w) − f (u − w)] ≤ 0
for any κ ∈ K − and therefore (MP ) holds.

The complementarity condition (∗) follows from (MP )2 by the choice μ =
2λ, μ = 0.
Vice versa, let v ∈ K(g), hence v −w ∈ K. This implies by the complementarity
condition (∗)
λ, v − u = λ, v − w − λ, u − w ≤ 0 .
Hence we arrive at a(u, v − u) = (f − λ)(v − u) ≥ f (v − u) .

From the proof above, it follows that u ∈ K(g) solves the VI (C.1), iff there
exists λ such that [u, λ] ∈ K(g) × K − solves (MP )1 and (∗) holds. Therefore the
above mixed form does not depend on the chosen preimage w. Indeed, let Bwi =
g (i = 1, 2). Then u±(w1 −w2 ) ∈ K(g) and thus by the VI (C.1), λ(w1 −w2 ) = 0.
The mixed formulation above applies to unilateral contact problems with Sig-
norini condition on some boundary part Γc in appropriate function spaces, where
for a boundary variable u the linear map u → Bu is the restriction to the boundary
part Γc ; see Sect. 11.4.1.
With friction problems we encounter nonsmooth optimization problems of the
form
1
(NOP ) minimize f (v) = Av, v − f, v + ϕ(v), v ∈ V ,
2
where ϕ is convex, even positively homogeneous, hence sublinear on V , but not
differentiable in the classic sense. A prominent example is

ϕ(v) = g|v| ds (g ∈ L∞ (Γc ), g > 0) .
Γc
An optimal solution of (NOP ) is characterized as solution to the so-called

variational inequality of the second kind:
u ∈ V , Au, v − u + ϕ(v) − ϕ(u) ≥ f (v − u), ∀v ∈ V .
Here one can obtain by (L1 , L∞ ) duality and density the useful duality formula

ϕ(v) = g|v| ds = sup{ gvμ ds | μ ∈ C(Γ ), |μ| ≤ 1} .
Γc Γc
C.1.3 Lagrange Multipliers for Some Inequality Constrained

Variational Inequalities
In this subsection we deal with a canonical class of inequality constrained vari-

ational inequalities of the second kind, where the sum of a bilinear form and a
sublinear functional and further a linear functional as right hand side occur and
where the constraints are defined by linear inequalties with respect to a closed
convex ordering cone. More precisely, let V , Z be real reflexive Banach spaces with
(topological) dual spaces V , Z . Let P ⊂ Z be a closed convex cone with vertex
at zero. Let A ∈ L (V , V ), B ∈ L (V , Z ) be continuous linear operators that
give rise to the continuous bilinear forms a : V × V → R, b : V × Z → R
via a(v, w) = Av, w V ×V , b(v, z) = Bv, z Z ×Z . We use the null space
W := ker B of B and its polar W ◦ contained in V . Further let ϕ : V → R be
sublinear, thus there holds the representation formula
ϕ(v) = maxσ, v , ∀v ∈ V , (C.2)

σ ∈S
where S ⊂ V is weak∗ compact and coincides with the convex subdifferential

∂ϕ(0) = {ξ ∈ V |ξ, · ≤ ϕ}. In other words, ϕ is the support function [252] of S.
Finally let f ∈ V , g ∈ Z be fixed. Then introduce the feasible set
K(g) = {v ∈ V : b(v, p) ≤ g, p , ∀p ∈ P }
and pose the variational inequality in its primal form: Find u ∈ V that satisfies
(V I ) u ∈ K(g), a(u, v − u) + ϕ(v) − ϕ(u) ≥ f, v − u , ∀v ∈ K(g) .
Our goal in this subsection is to arrive at the following mixed form with Lagrange
multipliers q ∈ P and τ ∈ S :
%
(MF − 1) a(u, v) + b(v, q) + τ, v = f, v , ∀v ∈ V ,
(MF )
(MF − 2) b(u, p − q) + u, σ − τ ≤ g, p − q , ∀[p, σ ] ∈ P × S .
To achieve this goal we use the famous Brezzi lemma which characterizes that
B , the adjoint operator of B, is isomorph, i.e. is bijective with continuous inverse,
by the celebrated Babuška-Brezzi condition (BB). More precisely, there holds
Lemma C.1 The following assertions are equivalent.
(i) There exists a number β > 0 such that
b(v, z)
(BB) sup ≥ β zZ , ∀z ∈ Z ,
v∈V ,v=0 vV
(ii) B : Z → W ◦ is isomorph with
B zV ≥ β zZ , ∀z ∈ Z (C.3)
for some β > 0.

For the proof of the Brezzi lemma we can e.g. refer to [60, Theorem 3.6,
Lemma 4.2].
Now we focus to the homogeneous case, where g = 0 with feasible set K =:
K(0), since the proof of this case is simpler and nearer to the linear functional
analytical proof of the classic case of equality constrained variational problems than
the proof for general g.
Theorem C.2 The two problems (V I ) and (MF ) are related as follows. If
[u, q, λ] ∈ V × Z × V solves (MF ) (with g = 0), then u lies in K and solves
(V I ). Vice versa, let u ∈ K solve (V I ), then there exist q ∈ P and τ ∈ S such that
[u, q, τ ] solves (MF ) (with g = 0), provided (BB) holds for some β > 0.
Proof We give a sketch of the proof divided in several steps.
I. Since P is a cone, we can choose p = 1/2 q and p = 2q in (MF − 2).
Moreover we use (C.2). Thus we first observe that (MF − 2) with g = 0 splits
equivalently into the statements
⎧
⎨ b(u, p) ≤ 0 , ∀p ∈ P ,
(MF − 3) b(u, q) = 0 ,
⎩
ϕ(u) = τ, u .
II. Let [u, q, τ ] ∈ V × Z × V solve (MF ) with g = 0. Then from (MF − 3)1 it
is immediate that u ∈ K.
To show that u solves (V I ), let v ∈ K be arbitrary. Then b(v, q) ≤ 0 and
from (MF −3)2 , b(v−u, q) ≤ 0. Hence from (MF −3)3 , (C.2), and (MF −1),
a(u, v − u) + ϕ(v) − ϕ(u)

≥ a(u, v − u) + τ, v − u
= −b(v − u, q) + f, v − u
≥ f, v − u .
III. The proof of the second part of the theorem runs in 5 steps.
1. Let u ∈ K solve (V I ). Since K is a cone, we can choose v = 2u and v = 1/2 u.
This gives
a(u, u) + ϕ(u) = f, u , (C.4)

hence by addition,
a(u, v) + ϕ(v) ≥ f, v , ∀v ∈ K . (C.5)
Note that (C.4) and (C.5) are equivalent to (V I ).

2. By (C.2), (C.5) means: ∀v ∈ K ∃σ ∈ S such that a(u, v) + σ, v ≥ f, v .
Since S is convex and weak∗ compact, it can be shown that there exists some
τ ∈ S such that
a(u, v) + τ, v ≥ f, v , ∀v ∈ K . (C.6)
3. By construction, W = ker B ⊂ K. Hence (C.6) implies
a(u, w) + τ, w = f, w , ∀w ∈ W ,
or f − Au − τ ∈ W ◦ . In virtue of the (BB) condition, Lemma C.1 applies and

entails the existence of q ∈ Z such that B q = f − Au − τ or
a(u, v) + b(v, q) + τ, v = f, v , ∀v ∈ V .
Thus we obtain (MF − 1).

4. We claim that q ∈ P . Indeed, (C.6) gives by definition of q,
B q, v = Bv, q ≤ 0 , ∀v ∈ K .
This means Bv ∈ P − ⇒ Bv ∈ Q− , where P − = {ζ ∈ Z |ζ, p ≤ 0, ∀p ∈ P }

is the negative dual cone to P and Q := R+ q ⊂ Z. In virtue of the (BB)
condition, Lemma C.1 applies and hence B : (W ◦ ) → Z is isomorph, in
particular is onto. Therefore the implication above gives P − ⊂ Q− , what results
by the bipolar theorem in P −− = P ⊃ Q−− = Q. This proves the claim.
5. To prove (MF − 2), we show (MF − 3). By feasibility of u ∈ K, (MF − 31 ) is
obvious. Since τ ∈ S = ∂ϕ(0), ϕ(u) ≥ τ, u . From (C.4) and (C.6), we get
f, v = a(u, u) + ϕ(u) ≥ a(u, u) + τ, u ≥ f, v ,
hence (MF − 3)3 , and also by definition of q, B q, u = b(u, q) = 0, thus

finally (MF − 3)2 .

For a more detailed proof and for the proof of the general case of arbitrary g,
moreover for further references see [219].
Here let us first consider the special case ϕ = 0, S = {0}. Our aim is to
derive from the present mixed form (MF ) the mixed form (MP ) of the previous
subsection.
The present mixed form becomes then with some preimage w = B −1 g

%
a(u, v) + B q, v = f, v , ∀v ∈ V ,
(MF )0
u, B p − B q ≤ w, B p − B q , ∀p ∈ P ,
where the multiplier q exists in P . Note that λ := B q ∈ K − , the latter inequality

(MF )0−2 extends to the closure of B P , what coincides with [B −1 (P − )]− = K − .
Hence we arrive at the mixed form (MP ).
To conclude this subsection, we want to bring the present mixed form (MF )
in relation to the mixed form used in BEM solution of frictional unilateral contact
problems in [33, 37], see Sect. 11.4.1. To this end, we proceed as in the special case
above and obtain from (MF ) with again Bw = g the pair [λ, τ ] ∈ K − × S that
together with u ∈ V solves the mixed system
%
a(u, v) + λ, v + τ, v = f, v , ∀v ∈ V ,
κ − λ, u + σ − τ, u ≤ κ − λ, w , ∀[κ, σ ] ∈ K − × S .
Note that K − + S is convex and closed in V . Thus using the indicator function
χK of K (χK (v) = 0 iff v ∈ K, = +∞ elsewhere),
K − + S = ∂χK (0) + ∂ϕ(0) = ∂(χK + ϕ)(0)

= {μ|μ, · ≤ χK + ϕ}
= {μ|μ, v ≤ ϕ(v), ∀v ∈ V with Bv ∈ P − } =: M
what is the analog to the set of multipliers in [33, 37].

On the other hand, for any μ ∈ M - in the case of a general reflexive Banach
space V in virtue of Troyanski’s renorming theorem an equivalent norm can be
introduced so that V and V are locally uniformly convex, and thus also strictly
convex - the constrained best approximation problem
minimize κ2 + σ 2 , κ ∈ K − , σ ∈ S
subject to κ + σ = μ
admits unique solutions μ− ∈ K − , μS ∈ S with μ = μ− + μS .

Therefore we arrive at the multiplier ν := λ + τ ∈ M that together with u ∈ V
solves the somewhat condensed mixed system
%
a(u, v) + ν, v = f, v , ∀v ∈ V ,
(MF )c
μ − ν, u ≤ μ− − ν− , w , ∀μ ∈ M ,
what corresponds to the mixed form in [33, 37].

C.2 Nonsmooth Analysis
With nonmonotone contact problems we encounter locally Lipschitz functions that

are not necessarily convex or smooth in the sense of classical differentiability.
Therefore in this section we draw some basics from Clarke’s monograph [109]
on nonsmooth analysis. We collect some fundamental concepts of the Clarke
generalized differential calculus, in particular introduce his generalized directional
derivative along with its basic properties. Following [332, 335] we also provide reg-
ularization techniques of nondifferentiable optimization to smooth locally Lipschitz
functions that are minima or maxima of smooth functions. These regularization
techniques are needed in addition for the numerical treatment of nonmonotone
contact problems, see Sect. 11.5.
C.2.1 Nonsmooth Analysis of Locally Lipschitz Functions
Throughout this subsection, let X denote a (real) Banach space. Let f : X → R be

Lipschitz of rank K near a given point x ∈ X; that is, for some ε > 0, we have
|f (y) − f (z)| ≤ K y − z; ∀y, z ∈ B(x, ε) .
Definition C.1
f (y + tv) − f (y)
f 0 (x; v) := lim sup{ | y ∈ X, y → x; t > 0, t → 0}
t
is called the generalized directional derivative of f in the direction v.
Note that this definition does not presuppose the existence of a limit and that
it differs from the common definition of the directional derivative (or Gâteaux
derivative, which is continuous in v) in that the base point y in the difference
quotient varies. Also note that in general f 0 (x; ·) is not linear. The utility of this
definition is seen from the properties listed below.
Proposition C.3 Let f be Lipschitz of rank K near x. Then:
(i) The function v → f 0 (x; v) ∈ R is sublinear, hence convex, and satisfies
|f 0 (x;v)| ≤ K v for all v ∈ X;
(ii) The function (z, w) → f 0 (z; w) is upper semicontinuous at (x, v); the
function w → f 0 (x; w) is Lipschitz of rank K on X;
(iii) There holds f 0 (x; −v) = (−f )0 (x; v) for v ∈ X.
Definition C.2 The generalized gradient of the function f at x, denoted by
(simply) ∂f (x), is the unique nonempty weak∗ compact convex subset of the dual
space X , whose support function is f 0 (x; .).
Thus
ξ ∈ ∂f (x) ⇔ f 0 (x, v) ≥ ξ, v , ∀v ∈ X ,

f 0 (x; v) = max{ξ, v : ξ ∈ ∂f (x)}, ∀v ∈ X .
A function f : X → R which is continously differentiable near a point x is locally

Lipschitz near x by the mean value theorem. Also a function f : X → R which is
convex and lower semicontinuous is locally Lipschitz on all of X. In either case,
∂f reduces to the familiar concept of the derivative, respectively of that of the
subdifferential of convex analysis:
Theorem C.3 If f : X → R is continously differentiable near x, then ∂f (x) =
{f (x)}. If f : X → R is convex and lower semicontinuous on X, then for any
x ∈ X,
∂f (x) = {ξ ∈ X∗ : ξ, y − x ≤ f (y) − f (x), ∀y ∈ X} .
On the other hand, let f be Lipschitz near x and suppose that ∂f (x) is a singleton
{ξ }, then f is Gâteaux differentiable with f (x) = ξ .
Definition C.3 Let f : X → R be locally Lipschitz near x. Then f is called
regular at x, if f 0 (x; v) coincides with the classical directional derivative f (x, v)
for all v ∈ X.
There is a calculus of generalized gradiens including a sum rule, mean value
theorem, and chain rule; see [109] for details. Here we only provide an important
formula of nonsmooth analysis (’Danskin’s formula’, see [109, (2.3.12)]) that
chararacterizes the generalized directional derivative of max functions.
Let I be a finite index set and let {fi : i ∈ I } be a finite collection of functions
that are Lipschitz near x. Then the function f defined by
f (x) := max fi (x)

i∈I
is Lipschitz near x as well. Let I (x) := {i ∈ I : fi (x) = f (x)} and “co” denote
the convex hull.
Theorem C.4 There holds
∂f (x) ⊂ co {∂fi (x) : i ∈ I (x)}.
If fi is regular at x for each i ∈ I (x), then equality holds and f is regular at x.

C.2.2 Regularization of Nonsmooth Functions
In this subsection we follow [332, 335] and present a unified approach to regular-
ization of nonsmooth functions with focus to locally Lipschitz functions that are
minima or maxima of smooth functions.
According to Bertsekas [49] the maximum function f : Rn → R,
f (x) = max{g1 (x), g2 (x), . . . , gm (x)} (C.7)
of m continuously differentiable functions gi can be expressed by means of the plus

function p(x) = x + = max(x, 0), x ∈ R as
# $
f (x) = g1 (x) + p g2 (x) − g1 (x) + . . . + p [gm (x) − gm−1 (x)] . (C.8)
Replacing now the plus function by an approximation P (ε, )˙ , the smoothing

function S : Rn × R++ → R is given by
S(x, ε) := g1 (x)+P [ε, g2 (x) − g1 (x) + . . . + P [ε, gm (x) − gm−1 (x)]] (C.9)
as suggested by Chen et al. in [100]. The advantage of this procedure lies in the use
of one single regularization parameter ε to smooth an eventually larger number of
kinks. Here, P : R++ × R → R is the smoothing function via convolution for the
plus function p defined by
t
ε
P (ε, t) = (t − εs)ρ(s) ds.
−∞
We restrict ρ : R → R+ to be a density function of finite absolute mean; that is

k := |s|ρ(s) ds < ∞.
R
The major properties of S, see [346], that follow from the properties of the
function P , see [169, section 11.8.2], are collected in the following lemma.
Lemma C.2
(i) For any ε > 0 and for all x ∈ Rn ,
|S(x, ε) − f (x)| ≤ (m − 1)kε. (C.10)

(ii) The function S is continuously differentiable on Rn × R++ and for any x ∈ Rn

m
and ε > 0 there exist Λi ≥ 0 such that Λi = 1 and
i=1

m
∇x S(x, ε) = Λi ∇gi (x). (C.11)
i=1
Moreover,
co{ξ ∈ Rn : ξ = lim ∇x S(xk , εk ), xk → x, εk → 0+ } ⊆ ∂f (x),

k→∞
(C.12)
where “co” denotes the convex hull and ∂f (x) is the Clarke subdifferential.
We recall that the Clarke subdifferential of a locally Lipschitz function f at a point
x ∈ Rn can be characterized by
∂f (x) = co {ξ ∈ Rn : ξ = lim ∇f (xk ), xk → x, f is differentiable at xk },

k→∞
since in finite dimensional case, according to Rademacher’s theorem, f is differen-

tiable almost everywhere.
The maximum function given by (C.7) is clearly locally Lipschitz continuous
and by Theorem C.4, the Clarke subdifferential can be written as
∂f (x) = co{∇gi (x) : i ∈ I (x)}
with
I (x) := {i : f (x) = gi (x)}.
In particular, if x ∈ Rn is a point such that f (x) = gi (x) then ∂f (x) = {∇gi (x)}.
For such a point x ∈ Rn we show later on that
lim ∇x S(z, ε) = ∇gi (x).

z→x,ε→0+
Note that the set on the left-hand side in (C.12) goes back to [353]. In [99], this
set is denoted by GS (x) and is called there the subdifferential associated with
the smoothing function. The inclusion (C.12) shows in fact that GS (x) ⊆ ∂f (x).
Moreover, according to the part (b) of Corollary 8.47 in [353], ∂f (x) ⊆ GS (x).
Thus, ∂f (x) = GS (x).
Remark C.1 Note also that S is a smoothing approximation of f in the sense that
lim S(z, ε) = f (x) ∀ x ∈ Rn . (C.13)

z→x,ε→0+
This is immediate from (C.10).

Remark C.2 The regularization procedure (C.9) can be also applied to a minimum
function by
min{g1 (x), g2 (x), . . . , gm (x)}=−max{−g1 (x), −g2 (x), . . . , −gm (x)}≈−S(x, ε).
Denote now
Si = gi − gi−1 + P [ε, gi+1 − gi + P [ε, gi+2 − gi+1 + . . . + P [ε, gm − gm−1 ]]] .
This function should approximate

# # $$
gi − gi−1 + p gi+1 − gi + p gi+2 − gi+1 + . . . + p [gm − gm−1 ]
(C.8)
= max{gi − gi−1 , gi+1 − gi−1 , . . . , gm − gi−1 } =: Ti−1 .
Lemma C.3 It holds
lim P (ε, Si (z, ε)) = p(Ti−1 (x)). (C.14)

z→x,ε→0+
Proof First, for any ε0 > 0 there exists δ0 > 0 such that
|P (ε, z) − p(Ti−1 (x))| < ε0 (C.15)
for any z ∈ Bδ0 (Ti−1 (x)) and ε ∈ (0, δ0 ). Next, since Si is a smoothing
approximation of Ti−1 in the sense of (C.13), there exists δ̄0 > 0 such that
|Si (z, ε) − Ti−1 (x)| < δ0 (C.16)
for any z ∈ Bδ̄0 (x) and ε ∈ (0, δ̄0 ). Combining (C.15) and (C.16), it follows that
|P (ε, Si (z, ε)) − p(Ti−1 (x))| < ε0
holds for any ε < min{δ0 , δ̄0 } and any z ∈ Bδ̄0 (x). Thus, the assertion of the lemma
is proved.

Since the nonsmooth functions that occur in the nonmonotone contact problems
can be reformulated by using the plus function, all our regularizations are based in
fact on a class of smoothing approximations for the plus function. Some examples
from [168] and the references therein are in order:
t
ε
(t − εs) ρ1 (s) ds = t + ε ln(1 + e− ε ) = ε ln(1 + e ε ),
t t
P1 (ε, t) = (C.17)
−∞
e−s
where ρ1 (s) =
(1 + e−s )2
t √
ε t 2 + 4ε2 + t
P2 (ε, t) = (t − εs) ρ2 (s) ds = , (C.18)
−∞ 2
2
where ρ2 (s) =
(s 2 + 4)3/2
⎧
t ⎪
⎨0 if t < − 2ε
ε
P3 (ε, t) = (t − εs) ρ3 (s) ds = 1
(t + 2ε )2 if − ε
≤t ≤ ε (C.19)
−∞ ⎪
⎩
2ε 2 2
ε
t if t > 2,
%
1 if − 1
≤s≤ 1
where ρ3 (s) = 2 2
0 otherwise.
⎧
t ⎪
⎨0 if t < 0
ε
P4 (ε, t) = (t − εs) ρ4 (s) ds = t2
if 0 ≤ t ≤ ε (C.20)
−∞ ⎪
⎩ 2ε
t− ε
2 if t > ε,
%
1 if 0 ≤ s ≤ 1
where ρ4 (s) =
0 otherwise.
In the following we need
Ai = {x ∈ Rn : gi (x) > gj (x), ∀j = 1, . . . , m, j = i} for all i = 1, . . . , m
and compute
t
ε
Pt (ε, t) = ρ(s) ds. (C.21)
−∞
Lemma C.4 The following properties hold:

a) If x ∈ Ai , i = 1, 2, . . . , m − 1, then
lim Pt (ε, Si+1 (z, ε)) = 0. (C.22)

z→x,ε→0+
b) if x ∈ Ai , i = 2, 3, . . . , m, then
lim Pt (ε, Sj (z, ε)) = 1 for all j = 2, 3, . . . , i. (C.23)

z→x,ε→0+
Proof
a) Let i ∈ {1, 2, . . . , m − 1} and x ∈ Ai , i.e., gi (x) > gj (x) for all j = 1, . . . , m,
j = i, and Si+1 be a smoothing approximation of Ti defined as above by
Ti = max{gi+1 − gi , gi+2 − gi , . . . , gm − gi }.
Clearly, Ti (x) < 0. Since by (C.16) in the proof of Lemma C.3
Si+1 (z, ε) → Ti (x) as z → x and ε → 0+ (C.24)
and due to Ti (x) < 0, it follows from (C.21) that

Si+1 (z,ε)
ε
Pt (ε, Si+1 (z, ε)) = ρ(s) ds → 0 as z → x, ε → 0+
−∞
and (C.22) is verified.

b) Let now x ∈ Ai , i ∈ {2, 3, . . . , m}. We first prove the statement of the lemma
for j = i. By the representation
Si (z, ε) = gi (z) − gi−1 (z) + P (ε, Si+1 (z, ε))
and using (C.14) from Lemma C.3, it follows that
Si (z, ε) → gi (x) − gi−1 (x) as z → x and ε → 0+ . (C.25)
Hence, since gi (x) − gi−1 (x) > 0, we have
Si (z,ε)
ε
Pt (ε, Si (z, ε)) = ρ(s) ds → 1 as z → x, ε → 0+ (C.26)
−∞
and therefore (C.23) is verified for j = i. Thus, we completely proved the lemma
in the case m = 2. The remaining case can be based on an induction argument,
see [335].

Now we are ready to show that the gradient of the given function gi on Ai can be
approximated by the gradients of the smoothing function.
Theorem C.5 For any x ∈ Ai , i = 1, 2, . . . , m,
lim ∇x S(z, ε) = ∇gi (x).

z→x,ε→0+
Proof From (C.9), by direct differentiation with respect to x, it follows that
∇x S(z, ε) = 1 − Pt (ε, S2 (z, ε)) ∇g1 (z)

m−1 R
i
+ 1 − Pt (ε, Si+1 (z, ε)) Pt (ε, Sj (z, ε))∇gi (z)
i=2 j =2
R
m
+ Pt (ε, Si (z, ε))∇gm (z).
i=2
We shall distinguish the following three cases.

1) First, we take x ∈ A1 . From Lemma C.4 a)
lim Pt (ε, S2 (z, ε)) = 0

z→x,ε→0+
and, consequently, the following relations hold as z → x and ε → 0+ :
Λ1 := 1 − Pt (ε, S2 (z, ε)) → 1,
R
i
Λi := 1−Pt (ε, Si+1 (z, ε)) Pt (ε, Sj (z, ε)) → 0, i=2, . . . , m−1 (m≥3)
j =2
and
R
m
Λm := Pt (ε, Sj (z, ε)) → 0.
j =2
Hence, if x ∈ A1 then lim ∇x S(z, ε) = ∇g1 (x).

z→x,ε→0+
2) Let now x ∈ Ai for some i ∈ {2, 3, . . . , m − 1}, m ≥ 3. By (C.22) and (C.23), it
follows immediately that
Λi → 1 as z → x, ε → 0+ .
Further, we shall show that for any k, k ∈ {1, 2, . . . , m}, k = i, it holds for any
z → x and ε → 0+ that Λk → 0.
Indeed, the relation (C.23) implies
lim Pt (ε, Sk+1 (z, ε)) = 1 ∀k = 1, . . . , i − 1.

z→x,ε→0+
Therefore,
Λ1 = 1 − Pt (ε, S2 (z, ε)) → 0
and
R
k
Λk = (1 − Pt (ε, Sk+1 (z, ε))) Pt (ε, Sj (z, ε)) → 0 ∀k = 2, . . . , i − 1
j =2
(C.27)
as z → x and ε → 0+ . Altogether, Λk → 0 for all k = 1, . . . , i − 1.
Let now k ∈ {i +1, i +2, . . . , m−1}. According to (C.22), the (i +1)− multiplier
Pt (ε, Si+1 (z, ε)) in (C.27) goes to zero and consequently, Λk → 0.
Further, since (i+1) ∈ {3, 4, . . . , m} and Pt (ε, Si+1 (z, ε)) goes to zero, it follows
that
R
m
Λm = Pt (ε, Sj (z, ε)) → 0 as z → x, ε → 0+ .
j =2
In this way, we have proved that Λk → 0 for every k = 1, . . . , m, k = i, and

therefore, if x ∈ Ai then lim ∇x S(z, ε) → ∇gi (x).
z→x,ε→0+
3) Finally, let x ∈ Am . From Lemma C.4 b),
lim Pt (ε, Si (z, ε)) = 1 i = 2, . . . , m.

z→x,ε→0+
Hence,
Λ1 = 1 − Pt (ε, S2 (z, ε)) → 0 and Λm → 1.
Clearly, we can also write
Pt (ε, Si+1 (z, ε)) → 1 ∀i = 2, . . . , m − 1
and consequently,
Λi → 0 for all i = 2, . . . , m − 1.
Therefore, we have proved that if x ∈ Am then lim ∇x S(z, ε) → ∇gm (x).

z→x,ε→0+
Collecting all cases, the proof of the theorem is complete.

Remark C.3 Note that if x ∈ Rn is a point such that gi (x) = gj (x) for some i and
j , i = j , then for any sequences {xk } ⊂ Rn , {εk } ⊂ R++ such that xk → x and
εk → 0+ we have
lim ∇x S(xk , εk ) ∈ ∂f (x).

k→∞
C.3 Existence and Approximation Results for Variational

Inequalities
C.3.1 Existence Results for Linear VIs
Let (V , ., ., . ) be a real Hilbert space. Let λ ∈ V ∗ be a continuous linear form,
K ⊂ V a nonvoid closed, convex set, and β : V × V → R be a continuous bilinear
form, not necessarily symmetric. With these data given we consider the subsequent
variational inequality (P): Find û ∈ K such that
β(û, v − û) ≥ λ(v − û) ∀v ∈ K .
We require that β is positive semidefinite, i.e. β(v, v) ≥ 0 for all v ∈ V . Hence the
closed set
N := {u ∈ V : β(u, u) = 0}
is a (generally nontrivial) subspace, as it is seen as follows. Clearly, RN ⊆ N .

The symmetric bilinear form
1
β symm (u, v) := {β(u, v) + β(v, u)}
2
satisfies the Schwarz inequality. Therefore for any u, v ∈ N
0 ≤ β(u + v, u + v) = 2 β symm (u, v) ≤ 0 ,
hence u + v ∈ N . This also shows that
N = {u ∈ V : β symm (u, .) ≡ 0} .
Although the solution of (P) generally depends nonlinearly on the datum λ, the
solution set of (P) is convex. This is an easy consequence of the following useful
characterization.
Lemma C.5 Let û ∈ K. Then û solves (P), if and only if
β(v, û − v) ≤ λ(û − v) ∀v ∈ K .
Proof To show the “≤” inequality, use positive semidefiniteness of β and obtain
β(v, û − v) ≤ −β(û, v − û) ≤ −λ(v − û) ∀v ∈ K .
To show conversely (P), for any v ∈ K take wt := û + t (v − û) , t ∈ (0, 1). Then
wt ∈ K and
β(wt , û − wt ) ≤ λ(û − wt ) ,
hence
β(wt , v − û) ≥ λ(v − û) .
Letting t → 0, the inequality of (P) follows.

To obtain existence results one needs further assumptions. In accordance to the

Signorini problem in Sect. 5.1 , we assume that β is semicoercive in the sense that
β should satisfy a Gårding inequality:
(G) β(v, v) + Cv, v ≥ c v2 for all v ∈ V
with some real number c > 0 and a compact linear operator C : V → V ∗ . If (G)
holds with C = 0, then β is usually termed coercive or elliptic. In the coercive case,
the Lions - Stampacchia theorem that extends the Lax - Milgram lemma guarantees
unique solvability of (P) for each λ ∈ V ∗ :
Theorem C.6 (Lions - Stampacchia Theorem) Let β : V × V → R be a
continuous elliptic bilinear form on the Hilbert space V . Moreover, let K = ∅,
convex, closed ⊂ V , λ ∈ V ∗ . Then the variational inequality (P) has a unique
solution û. Moreover, the mapping λ → û is Lipschitz continuous.
Proof We give a sketch of the proof divided in three steps.
1. Let ui be solutions to the data λi . Then choose v = u2 , respectively v = u1 in
(P), sum up and obtain β(u1 − u2 , u1 − u2 ) ≤ (λ1 − λ2 )(u1 − u2 ). Since β is
elliptic, c u1 − u2 2 ≤ λ1 − λ2 V ∗ u1 − u2 , what shows Lipschitz continuity
and uniqueness.
2. Existence in the case of symmetric β
1
Method: Minimize “energy” J (v) = β(v, v) − λ(v), since minimization
2
problem on K is equivalent to (P ) in the symmetric case.
Consider minimizing sequence {un }; this is a Cauchy sequence, what can be
seen by the parallelogram rule. Then un → û ∈ K, J (un ) → J (û), since J is
continuous.
3. Existence in the general case.
Let in addition σ a symmetric elliptic bilinear form, e. g. σ (v, w) = v, w or
σ (v, w) = β symm (v, w) = 12 β(v, w) + β(w, v) .
For fixed u ∈ K, ρ > 0 ∃1 w ∈ K (according to the symmetric case above) such
that
σ (w, v − w) ≥ σ (u, v − w) − ρ[β(u, v − w) − λ(v − w)] ∀v ∈ K .
Hence u → w = S(u) gives a mapping S: K → K. Clearly û solves (P ), if

and only if û = S(û). Choose now ρ > 0 sufficiently small, such that S is a
contraction that gives the fixed point û.
For details see e.g. the monograph of Kinderlehrer and Stampacchia [267,
theorem II.2.19].

Coercivity is also necessary for well-posedness; this is clarified in the following

Proposition C.4 Let A : H → H be a linear continuous operator. Suppose that
the bilinear form
α(x, y) = < Ax, y >
is symmetric and positive semidefinite. If A is bijective, then α is coercive.

Proof By Banach’s inverse mapping theorem, A−1 is continuous. Then for any fixed
x ∈ H with x = 1 we have
0 0 5 6
0 −1 Ax 0
0 Ax, Ax = x · Ax = 1
2
A−1 sup |Ax, y | ≥ 0
0A 0
y=1 Ax Ax Ax Ax
and hence
1
inf sup |Ax, y | ≥ .
||x||=1 ||y||=1 ||A−1 ||
Thus by the Cauchy-Schwarz inequality applied to the positive semidefinite and

symmetric bilinear form a,
|α(x, y)|
α(x, x) ≥ supy=0
α(y, y)
1 |α(x, y)| 2
≥ [ supy=0 ]
A y
1
≥ x2
A A−1 2
what proves the asssertion.

Thus for the general semicoercive bilinear form β under study, we need extra
conditions for the specific λ ∈ V ∗ to yield existence of solutions to (P). Referring to
[28, 201] a sufficient condition for solvability is the recession condition C = −C ,
where the convex cone C is given by
C := {w ∈ ac K ∩ N : β(v, w) ≤ λ(w) ∀v ∈ K}
and with some fixed k0 ∈ K

L
ac K := t (K − k0 )
t >0
denotes the asymptotic cone or the recession cone of K. A stronger condition is that
there exists some v0 ∈ K such that
λ(w) < β(v0 , w) ∀w ∈ ac K ∩ N \ {0} ,
since this latter condition obviously implies that C = {0}. In the case 0 ∈ K, this
latter condition simplifies to
λ(w) < 0 ∀w ∈ ac K ∩ N \ {0} ,
which can already be found with Fichera [179] and Stampacchia [388].
In the Signorini problem discussed in Sect. 5.1, see (5.6), K is already a convex
cone (with vertex at zero) and the set K ∩ N coincides with the set of constant
functions that are nonpositive on ΓS , thus nonpositive throughout Rd . Therefore the
recession condition of Fichera–Stampacchia is here simply

(1) = g ds + h ds > 0 . (C.28)
ΓN ΓS
This latter condition also guarantees the uniqueness of the solution of the Neumann–
Signorini problem, where ΓD may be empty.
C.3.2 Approximation of Linear VIs
In this subsection we present an approximation result, which is based on [213], for

linear variational inequalities in Hilbert space. So we have the same setting as in the
previous section and consider the problem (P), but now for simplicity K is assumed
to be a nonvoid closed, convex cone (with vertex at zero).
To describe the approximation of our variational problem (P) we suppose that we
are given a positive parameter h converging to 0 and a family {V h }h>0 of closed
finite dimensional subspaces contained in V . In addition we have a family {K h }h>0
of closed convex nonempty cones of V h . These sets K h should approximate the
given set K. However, piecewise polynomial interpolation - except piecewise linear
interpolation - does not preserve order, thus generally K h cannot assumed to be
contained in K. To cope with this difficulty of nonconforming approximation we
follow the discretization theory of Glowinski [199, Chapter 1], which refines the
set convergence notion due to Mosco [309] (see also [6] for definition and further
study) and independently to Stummel, see [414] and introduce the following two
hypotheses (H1) and (H2):
(H1) If for some sequence {hj }j ∈N with hj → 0, v hj ∈ K hj (j ∈ N) and v hj
converges weakly to v ∈ V (j → ∞), then v ∈ K.
(H2) There exist a subset M ⊂ V such that M = K and mappings r h : M → V h

with the property that, for each v ∈ M, r h v → v (h → 0) and r h v ∈ K h for
all h ≤ h0 (v) for some h0 (v) > 0 .
Thus we approximate the problem (P ) by the following variational inequality (P h ):
Find uh ∈ K h such that
β(uh , v h − uh ) ≥ λ(v h − uh ) ∀v h ∈ K h .
By the existence theory in the infinite dimensional case, also solutions uh to these
finite dimensional problems exist.
Note that in most computations, however, it will be necessary to replace also β
and λ by some approximations β h and λh , defined by a numerical integration rule
which is used in the finite element, respectively boundary element discretization.
Since there is nothing new compared to the case of linear elliptic boundary value
problems and variational equalities, we do not discuss this aspect here.
Now we can state and prove our basic convergence result.
Theorem C.7 Let β , λ, K, and {K h }h satisfy the conditions (G), (H1) and (H2). If
the solution û of (P ) is unique, then limh→0 uh − û = 0 holds.
Proof We divide the proof in five parts. We first show a priori estimates for {uh }h ,
before we can establish the convergence results.
1) | . | – estimate for {uh } .
Fix w0 ∈ M, let wh := r h w0 ∈ K h for 0 < h = h0 := h0 (w0 ). Then we have
lim wh − w0 = 0, and with uh , a solution of (P h )
|uh |2 = β(uh , uh ) ≤ c0 + c1 uh + λ(uh ) (C.29)

≤ c0 + c2 uh . (C.30)
Here and in the following c0 , c1 , c2 , . . . are generic positive constants. Moreover,

by positive semidefiniteness,
β(wh , uh ) − λ(uh ) ≤ β(wh , wh ) − λ(wh ) ≤ c3 . (C.31)
2) Norm-boundedness of {uh } .
Here we modify a contradiction argument, which in the existence theory of
semicoercive variational inequalities goes back to Fichera [179] and Stampac-
chia [388]. We assume there exists a subsequence {u }∈N := {uh } such that
u → +∞ ( → ∞) . With y := u −1 u in the Hilbert space V , we can
extract a subsequence, again denoted by {y } , that converges weakly to some
y ∈ V . In virtue of (C.30) , we get
|y |2 u ≤ c4 .
Thus we have |y | → 0. [ Assume not. Then for a subsequence |yk | ≥ c5 > 0
and hence
c4
|yk | ≤ ,
c5 uk
what by uk → +∞ leads to a contradiction. ]

Since | . | is continuous and sublinear, hence weakly sequentially lower semicon-
tinuous, we obtain y ∈ N . Since {u } belongs to the cone K h , (H1) implies
that y ∈ K, too.
We claim that y = 0. From (C.31) we obtain
c3
β(w , y ) − λ(y ) ≤ ,
u
hence
β(w0 , y) ≤ λ(y) ∀w0 ∈ M , (C.32)
which extends to M = K by continuity. Moreover, for the solution û we have

by the characterization lemma C.5
β(u, û) − β(u, u) ≤ −λ(u − û) ∀u ∈ K . (C.33)
From (C.32) and (C.33) it follows for any t > 0
β(u, û + tu) − β(u, u) ≤ λ(û + tu) − λ(u) ∀u ∈ K .
Hence by the characterization lemma C.5, û + ty solves (P), and by uniqueness,

y = 0 follows.
Now we use (G). By compactness of C, for some subsequence
limk→∞ Cyk = 0 and
cyk 2 ≤ β(yk , yk ) + Cyk , yk ,
hence yk → 0. However, yk = 1, and a contradiction is reached proving the

boundedness of {uh } .
3) Any weak limit point u∗ of {uh } solves (P ) .
By the preceding step, there exists a subsequence, again denoted by {u } such
that u → u∗ . By (H1) , u∗ belongs to K. To show that u∗ solves (P ), take
v ∈ M arbitrarily. Then v := r h v converges strongly to v , and for h ≤ h0 (v)
we have
β(u , v − u ) ≥ λ(v − u ) .
Since β is positive semidefinite,
β(v , u − v ) ≤ λ(u − v ) .
Hence in the limit
β(v, u∗ − v) ≤ λ(u∗ − v) .
This inequality extends by continuity to M = K . Finally by the characterization

lemma C.5, we conclude for any v ∈ K
β(u∗ , v − u∗ ) ≥ λ(v − u∗ ) .
4) Convergence with respect to | . | .

Here we use an argument due to Glowinski [199, Chapter 1]. Since the solution
û of (P) is unique, the entire family {uh } converges weakly to û. Now take v ∈ M
arbitrarily. Then v h := r h v converges strongly to v , and for h ≤ h0 (v) we have
β(uh − û, uh − û) = β(uh , v h − û) − β(uh , v h − uh ) − β(û, uh − û)

≤ c6 v h − û + λ(uh − v h ) − β(û, uh − û) .
Hence in the limit, for any v ∈ M ,
0 ≤ lim sup |uh − û|2 ≤ c6 v − û + λ(û − v) .

h→0
The obtained inequality extends to K by density and continuity. Finally, the

choice v = û leads to the desired |.| – convergence.
5) Convergence with respect to . .
Assume there exists a sequence {u } such that u is a solution to (P h ) and
u − û ≥ δ > 0. By part (2), u − û is bounded and therefore we can
extract a subsequence, again denoted by {u } such that u − û converges weakly
to some w ∈ V . By part (3), û + w solves (P ), hence by uniqueness w = 0V .
Now we again use (G). By compactness of C, we can extract a subsequence,
again denoted by {u } such that C u converges strongly to C û and moreover by
part (4), |u − û| → 0 ( → ∞). Therefore by (G), u − û → 0 ( → ∞),
and a contradiction is reached.

For the more general approximation of general convex closed sets (instead of cones)
we refer to [213].
C.3.3 Pseudomonotone VIs—Existence Result
The Lions-Stampacchia theorem was substantially extended by Brézis to a very

large class of (non-linear) operators, called pseudomonotone operators in [64,
Theorem 24], see also [438, section 27.2]. With the symbol ! denoting weak
convergence on V , T : V → V ∗ is called pseudomonotone, if it is bounded
and if for any sequence {un }n∈N in V ,
un ! u and lim inf T (un ), u − un ≥ 0,

n→∞
imply
T (u), v − u ≥ lim sup T (un ), v − un , ∀v ∈ V .

n→∞
Such a pseudomonotone operator T : K ⊂ V → V ∗ as defined above gives

rise to the bifunction ψ : K × K → R via ψ(u, v) := T (u), v − u . Then ψ is
pseudomonotone (PM) in the sense that for any sequence {un } in K,
un ! u and lim inf ψ(un , u) ≥ 0

n→∞
imply that for any v ∈ K there holds
ψ(u, v) ≥ lim sup ψ(un , v) .

n→∞
A simple example of a pseudomonotone bifunction (not represented by an operator)

is ψ(u, v) = g(v) − g(u), where g is a weakly lower semicontinuous function.
Let T be weakly continuous on subsets F ∩ K of K, where F is a finite dimen-
sional subspace of V . Then the function ψ(·, v) becomes upper semicontinuous on
each finite dimensional part F ∩ K of K. Here, we assume only that ψ(u, u) ≥ 0
and ψ(u, ·) is convex for any u ∈ K; thus we do not require that ψ(u, ·) is linear-
affine. This is a suitable extension for the treatment of hemivariational inequalities to
follow. In this setting we have the following existence result from [212, Theorem 3].
Theorem C.8 Let K be a closed convex nonvoid subset of a reflexive Banach space
V . Let the bifunction ψ : K × K → R be pseudomonotone with ψ(·, v) upper
semicontinuous on each finite dimensional part of K, ψ(u, u) ≥ 0 and ψ(u, ·)
convex for any u ∈ K. Suppose that for some u0 ∈ K, ψ satisfies the coercivity
condition
ψ(u, u0 )
(CC) → −∞ as u ∈ K, u → ∞ .
u − u0
Then for any f ∈ V ∗ the variational inequality V I (ψ, f, K) admits a solution, i.e.
there exists u ∈ K such that
ψ(u, v) ≥ f, v − u , ∀v ∈ K . (C.34)

C.3.4 Mosco Convergence, Approximation of Pseudomonotone

VIs
In this subsection we present an approximation procedure for pseudomonotone

variational inequalities, where the given data (ψ, f, K) of the variational inequality
are approximated by bifunctions ψt , linear continuous functionals ft and closed
convex sets Kt , respectively, indexed by a directed set T . While K is contained in
a general reflexive Banach space V , Kt is a subset of a subspace Vt of V . For the
approximation of K by Kt we employ Mosco convergence, since we do not assume
that Kt is a subset of K. We provide a general approximation result, which with
finite-dimensional subspaces Vt of V can be considered as an abstract convergence
result for the Galerkin method for the solution of V I (ψ, f, K). Our approximation
result includes also the existence of solutions to the approximate V I (ψt , ft , Kt )
under an appropriate coerciveness condition.
We assume the following hypotheses:
(H1) If {vt }t ∈T weakly converges to v in V , vt ∈ Kt (t ∈ T ) for a subnet
{Kt }t ∈T of the net {Kt }t ∈T , then v ∈ K.
(H2) For any v ∈ K and any t ∈ T there exists vt ∈ Kt such that vt → v (strongly)
in V .
(H3) ψt is pseudomonotone for any t ∈ T .
(H4) ft → f in V ∗ .
(H5) For any nets {ut } and {vt } such that ut ∈ Kt , vt ∈ Kt , ut ! u, and vt → v in
V it follows that
lim inf ψt (ut , vt ) ≤ ψ(u, v) .

t ∈T
(H6) The family {−ψt } is uniformly bounded from below in the sense that there
exist constants c > 0, d, d0 ∈ R and α > 1 (independent of t ∈ T ) such that
for some wt ∈ Kt with wt → w there holds
−ψt (ut , wt ) ≥ cut αV + dut V + d0 , ∀ut ∈ Kt , ∀t ∈ T .
Remark C.4 The hypotheses (H1) and (H2) describe the Mosco convergence [6] of
the family {Kt } to K.
Remark C.5 Without loss of generality we can assume that 0 ∈ K ∩ {∩t ∈T Kt }.
Indeed, since K is nonvoid, by (H2) for any w ∈ K there exist wt ∈ Kt such
that wt → w. Then we consider the transformations v ∈ K → v − w ∈ K̃ :=
K − w; vt ∈ Kt → vt − wt ∈ K̃t := Kt − wt . Thus, K̃t Mosco converges to K̃ and
the hypotheses (H3), (H5), (H6) hold for the transformed bifunctions ψ̃, ψ̃t as well.
Under these hypotheses we have the following basic convergence result.
Theorem C.9 (General Approximation Result) Under conditions (H1)–[(H6),
there exist solutions ut to the approximate problem V I (ψt , ft , Kt ) and the family
{ut } is uniformly bounded in V . Moreover, there exists a subnet of {ut } that
converges weakly in V to a solution of the problem V I (ψ, f, K). Furthermore,

any weak accumulation point of {ut } is a solution to the problem V I (ψ, f, K).
Proof Using (H3) and (H6), the existence of a solution ut to V I (ψt , ft , Kt ) follows
from Theorem C.8. Inserting vt = 0 in V I (ψt , ft , Kt ) and using (H6) and (H4) we
obtain
cut αV + dut V + d0 ≤ −ψt (ut , 0) ≤ ft V ∗ ut ≤ C f V ∗ ut V ,
what proves the norm boundedness of {ut }. So we can extract a subnet of {ut }
denoted by {ut }t ∈T such that ut converges weakly to u in V . By (H1), u ∈ K.
Now, take an arbitrary v ∈ K. By (H2), there exists a net {vt } such that vt ∈ Kt and
vt → v in V . By (H4) and (H5), we get from V I (ψt , ft , Kt ) that for any v ∈ K
ψ(u, v) ≥ lim

inf

ψt (ut , vt ) ≥ lim

ft , vt − ut = f, v − u
t ∈T t ∈T
and consequently u is a solution to V I (ψ, f, K). At the same time we have proved
that any weak accumulation point of {ut } is a solution to V I (ψ, f, K). This should
be understood in the sense that every weak limit of any subnet of {ut } is a solution
to V I (ψ, f, K).

Remark C.6 Without the coercivity hypothesis (H6) we get a stability result in the
sense of Painlevé-Kuratowski set convergence that guarantees the inclusion
lim sup S (ψt , ft , Kt ) ⊂ S (ψ, f, K) .

t ∈T
Here, the set S (ψ, f, K), depending on ψ, f and K, consists of all functions u ∈ K
satisfying the variational inequality V I (ψ; f ; K).
C.3.5 A Hemivariational Inequality as a Pseudomonotone VI
Let V be the classical Sobolev space H 1 (Ω; Rd ), where Ω ⊂ Rd with d = 2, 3

is a bounded domain with Lipschitz boundary ∂Ω, and let K ⊆ V be a nonempty
closed, convex set specified later. Further, let the boundary ∂Ω = Γ¯D ∪ Γ¯c ∪ Γ¯F
be composed of three mutually disjoint parts: a Dirichlet boundary ΓD , a contact
boundary Γc and a part ΓF , where given external forces are applied. We also assume
that the measure of ΓD and Γc is strictly positive.
With γ we denote the trace operator from V into L2 (Γc ; Rd ), which is a linear
continuous mapping. Hence, there exists a constant c0 depending on Ω and Γc such
that
γ vL2 (Γc ;Rd ) ≤ c0 vV , ∀v ∈ V . (C.35)
Moreover, by the trace theorem [277, Theorem 6.10.5], γ is compact.

We introduce the linear elastic operator A : V → V ∗ ,

Au, v = ε(u) : σ (v) dx, (C.36)
Ω
where ε(u) = 12 (∇u + (∇u)T ) is the linearized strain tensor and σ (v) = C :
ε(v) is the stress tensor. Here, C is the elasticity tensor with symmetric positive
L∞ coefficients. Hence, the linear elastic operator A : V → V ∗ is continuous,
symmetric and due to the Korn’s inequality coercive, i.e. there exists a constant
cK > 0 such that
Av, v ≥ cK v2V , ∀v ∈ V . (C.37)
We define the linear form f : V → R by

f, v = fT0 v dx + fT1 v ds,
Ω ΓF
where f0 ∈ L2 (Ω; Rd ) are the prescribed body forces and f1 ∈ L2 (ΓF ; Rd ) are the
prescribed surface tractions on ΓF .
In what follows we consider a function j : Γc × Rd → R such that
j (·, ξ ) : Γc → R is measurable on Γc for all ξ ∈ Rd and j (s, ·) : Rd → R
is locally Lipschitz on Rd for almost all (a.a.) s ∈ Γc . Moreover, j 0 (s, · ; ·)
stands for the generalized Clarke directional derivative [109] of j (s, ·), as used in
Sect. 5.3 and analyzed in Sect. C.2.1 above. With this data we consider the following
hemivariational inequality: Find u ∈ K such that

Au, v−u + j 0 (s, γ u(s); γ v(s)−γ u(s)) ds ≥ f, v−u , ∀v ∈ K. (C.38)
Γc
We denote by ∂j (s, ξ ) := ∂j (s, ·)(ξ ) the Clarke generalized subdifferential of

j (s, ·) at the point ξ . We assume that there exist positive constants c1 and c2 such
that for a.a. s ∈ Γc , all ξ ∈ Rd and for all η ∈ ∂j (s, ξ ) the following inequalities
hold
(i) |η| ≤ c1 (1 + |ξ |);
(ii) ηT ξ ≥ −c2 |ξ |.
This growth condition assures that the integral in (C.38) is well defined. Indeed, it
follows from (i) and (ii) that for a.a. s ∈ Γc

0 T
j (s, ξ ; ς ) = max η ς ≤ max |η| |ς | ≤ c1 (1 + |ξ |)|ς |, ∀ ξ, ς ∈ Rd
η∈∂j (s,ξ ) η∈∂j (s,ξ )
(C.39)
and
j 0 (s, ξ ; −ξ ) = max ηT (−ξ ) ≤ c2 |ξ |, ∀ξ ∈ Rd . (C.40)

η∈∂j (s,ξ )
The existence of a solution u to problem (P) can be derived from Theorem C.8. To
this end we define the functional ϕ : V × V → R by

ϕ(u, v) = j 0 (s, γ u(s); γ v(s) − γ u(s)) ds, ∀u, v ∈ V . (C.41)
Γc
Lemma C.6 ([220]) The functional ϕ is pseudomonotone and satisfies
ϕ(u, 0) ≤ c3 uV , ∀u ∈ V (C.42)
for some positive constant c3 .

Proof Let {um } be a sequence in V such that um ! u in V as m → ∞. Since γ is
compact, it follows for a subsequence of {γ um }, which we denote again by {γ um },
that
γ um → γ u in L2 (Γc ; Rd ) as m → ∞. (C.43)
Now, we fix v ∈ V and show that
lim sup ϕ(um , v) ≤ ϕ(u, v). (C.44)

m→∞
We first observe that by (C.43) there exists a subsequence of {γ um }, which we

denote again by {γ um }, such that
γ um (s) → γ u(s) for a.a. s ∈ Γc (C.45)
and
|γ um (s)| ≤ κ0 (s) for some nonnegative function κ0 ∈ L2 (Γc ). (C.46)
Using (C.39) and (C.46), it follows that
j 0 (s, γ um (s); γ v(s) − γ um (s)) ≤ c1 (1 + |γ um (s)|)|γ v(s) − γ um (s)|

≤ c1 (1 + κ0 (s)) |γ v(s)| + κ0 (s) ∈ L1 (Γc ).
From (C.45) and the upper semicontinuity of j 0 (s; ·, ·), we conclude by applying
the Fatou lemma that

lim sup ϕ(um , v) = lim sup j 0 (s, γ um (s); γ v(s) − γ um (s)) ds
m→∞ m→∞ Γc

≤ lim sup j 0 (s, γ um (s); γ v(s) − γ um (s)) ds
Γc m→∞

≤ j 0 (s, γ u(s); γ v(s) − γ u(s)) ds = ϕ(u, v) (C.47)
Γc
and thus, (C.44) is shown. Hence, the functional ϕ is pseudomonotone.

Furthermore, by (C.40) for any u ∈ V we can estimate

ϕ(u, 0) = j 0 (s, γ u(s); −γ u(s)) ds ≤ c2 |γ u(s)| ds
Γc Γc
(C.35)
≤ c2 ((meas (Γc ))1/2 γ uL2 (Γc ;Rd ) ≤ c2 ((meas (Γc ))1/2 c0 uV ,
which implies (C.42). The proof of the lemma is thus complete.

Since summation preserves pseudomonotonicity, see [216], the bifunction
ψ(u, v) := Au, v − u + ϕ(u, v)
is pseudomonotone and satisfies the assumptions of Theorem C.8; in particular the

coercivity condition (CC) holds, since by (C.37) and (C.42),
−ψ(u, 0) ≥ cK u2V − c3 uV .
We point out that uniqueness of the solution u to the hemivariational inequal-

ity (C.38) can be ensured for a large enough Korn constant cK , see [332] for a
proof of such an uniqueness result. We also refer to [326] for a similar uniqueness
result to related nonconvex nonsmooth optimization problems.
Appendix D
Some Implementations for BEM
D.1 Symm’s Equation on an Interval

1
V u(x) := − ln |x − y|u(y)dsy = f (x) for x ∈ Γ = (−1, 1). (D.1)
π Γ
On a uniform mesh with meshsize h on Γ , xj = −1+j h, h = n2 , j = 0, . . . , n,

we take the space V h of piecewise constant functions and perform the h-version of
the Galerkin scheme for (D.1):
Find uh ∈ V h , such that
a(uh , vh ) := V uh , vh = f, vh ∀vh ∈ V h . (D.2)
With the auxiliary function F (x) = x 2 ln |x| the Galerkin element aij becomes:
xi+1 xj+1
1 1
− ln |x − y| dy dx = F (xi+1 − xj +1 ) − F (xi+1 − xj )
π xi xj 2π
−F (xi − xj +1 ) + F (xi − xj ) (D.3)
+3(xi+1 − xi )(xj +1 − xj ) .
Exercise: Write a program,which implements the Galerkin scheme (D.2).

(i) Compute (D.2) for f = 1 and f = x. Note u = 1
ln 2
√1 for f = 1.
1−x 2
(ii) Plot the solution for n = 4, n = 8 and n = 16.
(iii) Write for the program unilap2.f90 a subroutine, which computes the energy
norm of the solution of the Galerkin equations. Compute for f = 1 the error

616 D Some Implementations for BEM
in the energy norm:

/ A
π
u − uh V = u2V − uh 2V = − uh 2V .
ln 2
Compute for different n the errors in the energy norm and plot them in a double
logarithmic scale.
D.2 The Dirichlet Problem in 2D
Now we consider the integral equation
V u(x) = (I + K)g(x), for x ∈ Γ (D.4)

1
V u(x) := − ln |x − y|u(y) dsy
π Γ

1 ∂ 1 ny (y − x)
Ku(x) := − ln |x − y|u(y) dsy = − u(y) dsy .
π Γ ∂ny π Γ |x − y|2
As geometry we take the L-shape Γ with vertices (0, 0), (0, 0.5), (−0.5, 0.5),
(−0.5, −0.5), (0.5, −0.5), (0.5, 0). We use a uniform mesh with length h and
define there the space V h of piecewise constant functions. Then the h-Version of
the Galerkin BEM for the equation (D.4) reads: Find uh ∈ V h , such that
a(uh , vh ) := V uh , vh = (I + K)g, vh ∀vh ∈ V h .
Write a program, which implements this method. Use routines of maiprogs.

To describe the mesh use the following data structure
integer:: ng
integer,parameter :: ngmax=2048
real(kind=dp):: rx(0:1,0:ngmax-1)
real(kind=dp):: rh(0:1,0:ngmax-1)
real(kind=dp):: rn(0:1,0:ngmax-1)
Here ngmax denotes the maximal number of elements and ng denotes the actual
amount. rx(0,i) and rx(1,i) are the x- and y- components of a vertex of the
element with the number i. rh(.,i) points from a vertex to the next vertex and
rn(.,i) is the direction of the exterior normal of this element.
(i) Create a mesh generator, which creates for an arbitrary number of elements a
uniform mesh.
(ii) Compute the Galerkin matrix for this data structure. Use the routine lapintegmd
of liblap2.f90.
D Some Implementations for BEM 617
(iii) Compute the right hand side with Gaussian quadrature using lapid und
lapkspot.
(iv) Test the last routine with g ≡ 1. (There holds (I + K)1 ≡ 0, why?)
(v) Solve the linear system with Gauss elimination.
The above mentioned subroutines can be downloaded from the home page of M.
Maischak, http://people.brunel.ac.uk/~mastmmm/. Use
⎡ ⎤

1 ny · (y − x)
Kg, vh =− vh (x) ⎣ g(y) dsy ⎦ dsx
π |x − y|2
Γ Γ
⎡ ⎤

1 ny · (y − x)
g(y) ⎣ vh (x)dsx ⎦ dsy = g, K vh .
Fubini
= −
π |x − y|2
Γ Γ
Let x ∈ Γi , y ∈ Γj . If Γi and Γj are on the same edge then nx · (y − x) = 0 and

g, K vh = 0. Otherwise we compute with the parametrisation of the vector y − x
1
1 n · (y − x) 1 n · (at + b)
dsy = dt
π |x − y|2 π |at + b|2
Γi −1
⎛ ⎞
1 1
1 t 1
= ⎝n · a dt + n · b dt ⎠ .
π a t + 2a · bt + b2
2 2 a2 t 2 + 2a · bt + b2
−1 −1
1
For the determination of these integrals, let gkn (α, β, γ ) := t k (αt 2 + βt + γ )n dt
−1
with α = a2 , β = 2a · b, γ = b2 . Now with = 4αγ − β 2:
1
1 2 2α + β −2α + β
g0−1 = dt = √ arctan √ − arctan √
αt + βt + γ
2
−1
and
1 1
t 2αt + β 2α dt
g1−1 = dt = + .
αt + βt + γ
2 (αt + βt + γ )
2 αt 2 + βt + γ
−1 −1
D.3 Symm’s Equation on a Surface Piece
We can consider on a plane surface piece Γ

1 ψ(y)
dσy = f (x) , x∈Γ . (D.5)
4π Γ |x − y|
Decomposing Γ into rectangles Ri ∈ Th , we can choose our basis functions ϕj (x)

to be 1 only on one element,

N
1, x ∈ Rj ,
ψh (x) = ψj ϕj (x) , ϕj (x) =
0, else.
i=1
1
We get the Galerkin scheme: Find ψh ∈ H̃ − 2 (Γ1 ) such that, for k = 1, . . . , N,
N
1 ψh (y) 1 ϕj (y)ϕk (x)
dσy ϕk (x) dσx = ψj dσy dσx
4π |x − y| 4π |y − x|
Γ Γ j =1 Γ Γ

=:I

= f (x)ϕk (x) dσx (D.6)
Γ
By definition of ϕj and ϕk , I can be written as

1 1 1 1
I = dσy dσx = M dσy dσx .
4π Rk Rj |x − y| 4π Rk Rj (x1 −y1 )2 + (x2 −y2 )2
The inner integral can be calculated by transforming Rj to the reference square

R̃ = [0, 1]2 :
3 4
Rj := (ξ, η) : xj ≤ ξ ≤ xj + hx , yj ≤ η ≤ yj + hy
ξ = xj + hx u , 0≤u≤1
η = yj + hy v , 0≤v≤1.
With fixed points x = (a, b) and y = (ξ, η), we get for the inner integral

1
dσy =
Rj |x − y|

dξ dη hx du hy dv
= M = 1
Rj (ξ − a) + (η − b)
2 2 R̃ (x + h u − a)2 + (y + h v − b)2 2
j x j y

−1 hy − β −1 β −1 hx − α −1 α
=α sinh + sinh +β sinh + sinh
|α| |α| |β| |β|

hy − β β
+(hx − α) sinh−1 + sinh−1
|hx − α| |hx − α|

hx −α α
+(hy − β) sinh−1 + sinh−1 , α = a −xj , β = b−yj .
|hy −β| |hy −β|
The outer integral can be approximated, e.g. by a 4-point quadrature formula that
is exact for polynomials√of degree √≤ 2: Let the quadrature nodes x̂1 , x̂2 , x̂3 , x̂4 be
given by x̂i = (0.5 ± 63 , 0.5 ± 63 ), i = 1, 2, 3, 4 on R̃, then on the reference
square, for the polynomial P there holds

1
4
P (u, v) du dv = P (x̂i ) .
R̃ 4
i=1
Note that the outer integration can also be performed analytically. This is imple-
mented in the software package maiprogs, see [291].
For decomposing Γ we consider 4 different methods, firstly a uniform mesh
with axis-parallel rectangles, secondly a graded mesh described by a tensor product
mesh based on a 1-d graded mesh with grading constant β. And finally two adaptive
strategies based on a two-level error estimator, one where we split each appropriate
element into four equal sized elements and another one where we split the element
horizontally or vertically into two parts or into four equal parts depending on the
composition of the local error indicator (see Fig. D.1).
For the hierarchical error estimator we decompose every brick function φih
associated with the element i and element size h into a set of three jump functions
βi,j by uniformly refining the element i into four equal sized sub elements. Then
there holds for 1 ≤ j ≤ 3

4
h/2

4
h/2
βi,j = cl,j φi,l = φih + c̃l,j φi,l ,
l=1 l=2
h/2
where φi,l is the brick function on the sub element l to the element i. Further we
define with the energynorm ·V
7 8
9 : 4 h/2
V ψN − f, βi,j V ψN − f, φih + l=2 c̃l,j φi,l
ϑi,j := 0 0 = 0 0
0βi,j 0 0βi,j 0
V V
7 8
4 h/2
V ψN − f, l=2 c̃l,j φi,l
= 0 0
0βi,j 0
V
Fig. D.1 Different boundary decomposition techniques . (a) Uniform mesh decomposition. (b)
Graded mesh decomposition. (c) Adaptive mesh decomposition strategies 1. (d) Adaptive mesh
decomposition strategies 2
7 8 7 8
N 4 h/2 4 h/2
V h
i=1 αi φi (x), l=2 c̃l,j φi,l − f, l=2 c̃l,j φi,l
= 0 0
0βi,j 0
V
Note that ϑi,j can be implemented efficiently when using the linearity of the scalar
product and the reuse of old values. When making use of the Galerkin orthogonality
as above we are able to reduce the computation time for ϑi,j by 14 . The local error
indicator is now defined by
/
ϑi := 2 + ϑ2 + ϑ2
ϑi,1 i,2 i,3
For the adaptive strategy 2 we save in an additional vector if there holds ϑi,2 ≥
1.5ϑi,1 , ϑi,1 ≥ 1.5ϑi,2 or neither. If the element has been marked for refinement and
the first condition is true split the element vertically into 2 equal sized rectangles, if
the second condition is true then split horizontally into 2 equal sized rectangles and
else into 4 equal sized rectangles. If the saturation assumption holds one can prove
the efficiency and reliability of the error indicator η = ϑ2 .
The numerical experiments were carried out by Lothar Banz on the Laptop
Fujitsu Siemens Amilo M1439G with MatLab R2007. For solving the discrete linear
system a CG algorithm is applied.
As we can see from Fig. D.3 the condition √ number of Galerkin matrix with
an underlying uniform mesh behaves like O( N ).The condition numbers for the
different mesh strategies are growing much faster than for the uniform mesh. For the
graded meshes there holds the greater β is the faster the condition number grows.
The solution is obviously singular at the boundary of the boundary-domain Γ
with strong singularities in the edges. The uniform mesh does not take the singular
behavior into account which yields a lower convergence rate than for the different
strategies. If we apply a graded mesh as described earlier we can improve the
convergence rate in the energy norm of 0.25 for the uniform mesh to 0.73 for the
graded mesh with β = 4 (see Fig. D.2). The more we take the singularity into
account the better is the convergence rate. The local error is a product of the local
1
energy error
0.1
uniform mesh
Graded mesh beta=2
0.01 Graded mesh beta=3
Graded mesh beta=4
adaptiv strategie 1
adaptiv strategie 2
1 10 100 1000
dof
Fig. D.2 Energy error for different meshes

1e+014
uniform mesh
Graded mesh beta=2
Graded mesh beta=3
Graded mesh beta=4
1e+012 adaptiv strategie 1
adaptiv strategie 2
1e+010
Condition number
1e+008
1e+006
10000
100
1
1 10 100 1000
dof
Fig. D.3 Condition number for different meshes
element size with the local error-behavior of the solution ψ. Therefore a reduction
of the local element size will reduce the local error and thus the global error. As the
reduction strategy of the first adaptive strategy is limited by no reduction or splitting
into four equally sized elements we expect a convergence rate which is greater than
of the uniform mesh and less than of a graded mesh. The second adaptive strategy
has a broader reduction strategy and can therefore take the singular behavior better
into account as the first strategy. However it is still worse than the graded strategy as
it has no continuously, systematic, slow reduction of the elements close to the center
of the boundary-domain.
D.3.1 Implementation of hp-BEM on Surfaces
In the following we report from [310]. The combination of geometric mesh

refinement and h-p approximation with boundary element techniques gives a
powerful tool for the approximate solution of boundary integral equations. In [235]
an h-p Galerkin scheme for weakly singular and hypersingular integral equations
on plane screens in R3 was analyzed and in [251] exponential convergence could be
proved.
Although the singular integrals for plane surfaces in [235] can be evaluated
analytically the assembly of the Galerkin matrix is extremely expensive. This
becomes even worse if curved surfaces are considered and the entries of the Galerkin
matrix have to be computed by a numerical quadrature rule. Here, an application of
the h-p boundary element method has the advantage that the Galerkin error decays
exponentially fast with the size of the Galerkin matrix, i.e. the number of Galerkin
entries is kept low.
In this subsection we focus on the weakly singular integral equation on an open
surface Γ , which corresponds to the direct single layer potential formulation of the
Dirichlet problem for the homogeneous Laplace equation in R3 \ Γ . Our aim is now
to define a quadrature rule which approximates the Galerkin entries exponentially
fast with the number n of kernel evaluations. By increasing n at each h-p refinement
step we may, hence, expect to preserve the exponential convergence of the Galerkin
scheme while keeping the computational costs low.
The quadrature rules which we use are basically applications of Schwab’s [372]
graded quadrature rules for singular integrals to the inner and outer integrals in
our Galerkin matrix. Schwab’s rule can be applied directly to assemble collocation
matrices or the inner integrals of the Galerkin entries. Based on the h-p approx-
imation results in [17] and the interpolation property of e.g. Gaussian quadrature
formulae exponential convergence could be proved [372]. For the outer integrals we
need a similar rule which is designed to approximate the singularities of the single
layer potential.
Let G ⊂ R3 be an open curved surface with parameter region Γ = [−1, 1]2 and
parameter function γ : Γ → G. We assume that G satisfies a Lipschitz condition
and that γ (∂Γ ) = ∂G. Let V : H =−1/2(G) → H 1/2(G) be the single layer potential
operator defined as

1 ψ(y)
V ψ(x) = dσy .
4π Γ x − y
Let X= be a finite dimensional space of piecewise polynomial functions over Γ and

let X = {φ ◦ γ −1 ; φ ∈ X}= ⊂H =−1/2(G). Let ψi , ψj be two basis functions in X=
and let Γi , Γj be elements in Γ with supp(ψi ) ⊂ Γ i and supp(ψj ) ⊂ Γ j . Now, the
entries of the Galerkin matrix and the right hand side vector are
( )
−1 −1 1 ψj (x)
V ψj ◦γ , ψi ◦γ = ψi (x) J (y)dy J (x)dx (D.7)
4π Γi Γj γ (x)−γ (y)

f, ψi ◦ γ −1 = ψi (x)f (γ (x))J (x)dx (D.8)
Γi
where J (x) = ∂γ∂x(x)

1
× ∂γ∂x(x)
2
and . denotes the Euklidean norm in R3 .
It could be shown in [251] for plane surfaces Γ that the h-p version of the
boundary element method converges exponentially fast, see also Sect. 8.2. The
h-p meshes are geometrically graded towards ∂Γ and the polynomial degrees
of the test and trial functions in x ∈ Γ are small if x is close to ∂Γ and are
increased perpendicular to ∂Γ (for details see [235]). For plane surfaces, however,
the integrals (D.7) can be evaluated analytically and the computational cost for the
assembly of the Galerkin matrix grows only algebraically with the number N of h-
p refinement steps, i.e. like O(N α ). We show that there is a quadrature rule which
approximates the singular integrals (D.7) exponentially fast (with N) and which
needs O(N α ) kernel evaluations (α ∈ N fixed). Furthermore, we give (numerical)
evidence that the h-p Galerkin method applied to (D.5) in combination with this
quadrature rule leads to exponential convergence of the approximate solutions. To
approximate both integrals we have to deal with point and edge singularites. The
kernel |γ (x) − γ (y)|−1 of the inner integral has obviously a point singularity at
y = x whereas the single layer potential has singular behaviour at γ (∂ supp(ψ)).
For point singularities Schwab suggested the following rule for the approxima-
tion of the integral Γ0 ψ0 (x) dx where Γ0 = (0, 1)2 and ψ0 is singular at the origin:
Given a fixed parameter σ1 ∈ (0, 1) and an integer n one considers geometric
subdivisions of Γ0 into smaller rectangles Rl,k . We define z0 = 0, zk = σ1n−k ,
1 ≤ k ≤ n, and
R1,k = (zk−1 , zk ) × (0, zk−1 ) for 2 ≤ k ≤ n ,

R2,k = (zk−1 , zk ) × (zk−1 , zk ) for 1 ≤ k ≤ n ,
R3,k = (0, zk−1 ) × (zk−1 , zk ) for 2 ≤ k ≤ n .
For fixed 1 , 2 ∈ N0 let Ql,k denote the tensor product of the (k + 1 )-point
Gaussian quadrature rule in x1 -direction and the (k + 2 )-point Gaussian quadrature
rule in x2 -direction, scaled to Rl,k . Hence

Ql,k ψ0 ≈ ψ0 (x) dx .
Rl,k
(1)
The composite quadrature rule Qn is now defined as

n
3
Qn(1) ψ0 = Q2,1 ψ0 + Ql,k ψ0 ≈ ψ0 (x) dx .
k=2 l=1 Γ0
The quadrature points and the subdivision of Γ0 for σ1 = 0.4, n = 4 and 1 = 2 =

0 are shown in Fig. D.4a.
Remark D.1 When approximating the Galerkin entries (D.7) we will choose k =
pk where pk is the polynomial degree of ψj (or ψi ) in xk -direction (k = 1, 2).
For corner-edge singularities we consider again the reference element Γ0 . Let φ0
have corner singularities at the origin (0, 0) and at the point (x1 , x2 ) = (0, 1) and
an edge singularity at x1 ≡ 0. We use a geometric subdivision of Γ0 towards the
corners (0, 0) and (0, 1) and towards the corresponding edge with grading parameter
Fig. D.4 Subdivision of [0, 1]2 and quadrature points for (a) and (b), where σ1 = σ2 = 0.4 and
1 = 2 = 0 [310]
σ2 ∈ (0, 1/2). This defines a quadrature rule Q(2)

n . Hence, we have

Q(2)
n φ0 ≈ φ0 (x) dx .
Γ0
For an example see Fig. D.4b.

(3)
If the function φ0 is singular at all four edges of Γ0 we use a quadrature rule Qn
with geometrical grading towards all the edges of Γ0 and with grading parameter
σ3 ∈ (0, 1/2).
To approximate the Galerkin entries (1.3) we have to deal with three critical cases
where the kernel becomes singular:
i. Γi and Γj have a common node
ii. Γi and Γj have a common edge
iii. Γi = Γj
(1)
In the first case we use affine images of the quadrature rule Qn on Γi and Γj with
grading towards the common node.
We define the integer m = 1n4/3 2.
(2)
In the second case we use the affine image of Qm on Γi with grading towards the
common edge and the common nodes. Let xk denote the quadrature points of Q(2) m
on Γi . For any of these points xk we consider the straight line Lk which contains xk
and which is orthogonal to the common edge E of Γi and Γj . The line Lk divides
Γj into two rectangles Γj,k 1 and Γ 2 . On each of these rectangles we apply the
j,k
quadrature rule Q(1)
n with grading towards the point Lk ∩ E (see Fig. D.5).
Γj,k
1
xk
Γj,k
2
Γi E Γj
Fig. D.5 Composite quadrature for two elements Γi and Γj with common edge E. The arrows
indicate direction of grading on Γj . The grading on Γj varies with the location of the quadrature
points xk in Γi [310]
Γi
xk xk
Γj
Fig. D.6 Composite quadrature for Γi = Γj . The grading for the inner quadrature (over Γj ) varies
with the location of the outer quadrature points xk ∈ Γi [310]
(3)
In the third case we use the affine image of Qm on Γi . For any quadrature point
(3)
xk which belongs to Qm we divide Γj into four rectangles with common node xk
(1)
and use the affine image of Qn on each of these rectangles (with grading towards
xk ). See Fig. (D.6).
Next we prove exponential convergence of the quadrature rule introduced above
for a simple example of two square elements in the (x1 , x2 )-plane with a common
node. In the proof we restrict ourselves to piecewise constant test and trial functions,
i.e. we have 1 = 2 = 0. Numerical results for higher polynomial degrees, for the
case of two elements Γi and Γj with common edge and for the case Γi = Γj are
included in [310]: There the experimental results indicate exponential convergence
in this case.
As a simple example we consider the elements Γ1 = (0, 1)2 and Γ2 = (−1, 0)2
and the parameter function γ : [−1, 1]2 → G defined as γ (x1 , x2 ) = (x1 , x2 , 0).
Let ψ1 , ψ2 ∈ H =−1/2 (G) be defined as

1 if x ∈ Γj × {0}
ψj (x) = (j = 1, 2) .
0 if x ∈ G \ (Γj × {0})
Hence,

1 1
V (ψ1 , ψ2 ) := dy dx . (D.9)
4π Γ2 Γ1 |x − y|
(1) (1)
For 0 < σ < 1 and n ∈ N define Q1 = Qn and let Q2 be the affine image of Qn
(j ) (j )
on Γ2 with grading towards the origin. For i ∈ {1, 2} let xk and wk be the knots
and weights of the rule Qj , i.e.

M
(j ) (j )
Qj g = wk g(xk )
k=1
where M = 1 + 3 ni=2 i 2 = n3 + 32 n2 + 12 n − 2.
The following result states exponential convergence of the composite quadrature
rule Q2 Q1 applied to |x − y|−1 .
Theorem D.1 ([310]) For any 0 < σ < 1 there are constants c1 , c2 > 0 such that

M M
(2) −1
V (ψ1 , ψ2 ) − (1) (2) (1)
|x − | −c2 n
wk wj k x j ≤ c1 e (D.10)
k=1 j =1
for all n ∈ N.
To prove the theorem we need the theory of countably normed spaces and the result
from [372] on the exponential convergence of Qj :
For a domain A ⊂ R2 \ {0} and a parameter 0 < β < 1 let Hβk (A) be the closure
of C ∞ (A) with respect to the weighted Sobolev norm

k
g2H k (A) = |D α g(x)|2 φβ+j
2
(x) dx
β A
j =0 |α|=j
where the weight function φs is defined as φs (x) = |x|s (s ∈ R).

The countably
T normed space Bβ (A) is defined as the subspace of all functions g
in L1 (A) ∩ ∞ k
k=0 Hβ (A) whose derivatives satisfy the growth condition
1/2
|D α g(x)|2 φβ+j
2
(x) dx ≤ Dg (dg )j j ! (D.11)
A
for all α ∈ N20 with |α| = j . The constants dg ≥ 1 and Dg > 0 depend on A and g
but not on j .
As an example consider the family of functions
1
kx (y) = ∈ Bβ (Γ1 ) for 0 < β < 1 uniformly for x ∈ Γ2 .
|x − y|
The following result was proved in [372]:

Lemma D.1 Let g ∈ Bβ (Γ1 ) with β > 0 sufficiently small. Then, for any 0 <
σ1 < 1 there exist constants b1 , b2 > 0 independent of n such that

g(y) dy − Q(1) ≤ b1 e−b2 n (D.12)
n g
Γ1
where the constants b1 and b2 depend only on σ1 , β, dg , Dg and Γ1 .

We are now in the position to prove Theorem D.1.
Proof The triangle inequality yields

M
M

4π V (ψ1 , ψ2 ) − wk(1) wj(2) |xk(1) − xj(2)|−1 ≤
k=1 j =1

1 1
≤ dy dx − Q(x) dy
Γ Γ |x − y| Γ1 |x − y|
2
1 2
= 1

1 1
+ Q(x) dy − Q (x) (y)
Q 1 |x − y| .
Γ1 |x − y|
2 2

= 2
We will estimate 2 and 1 seperately. By definition of Qi and kx we have

M

=
(2)
wj kx (2) (y) dy − Q(1)
2 k
n x .
(2)
Γ1 j j
j =1
From Lemma D.1 it follows that

M
2 ≤ wj(2) b1 e−b2 n
j =1
where b1 and b2 are independent of j . Hence,
2 ≤ |Γ2 | b1 e−b2 n ≤ b1 e−b2 n . (D.13)

To estimate 1 we have to show that V ψ1 (x) ∈ Bβ (Γ2 ) for all 0 < β < 1. For
plane rectangular elements the single layer potential can be calculated analytically
[235]. For x ∈ Γ2 we have:
1
y2 − x2 y1 − x1 1
V ψ1 (x) = (y1 − x1 ) arsinh + (y2 − x2 ) arsinh
|y1 − x1 | |y2 − x2 | y1 =0
y2 =0
= −x1 ln(−x2 + |x|) − x2 ln(−x1 + |x|) + g1 (x)

= −(cos θ + sin θ ) r ln r + g2 (r, θ )
where (r, θ ) are the usual polar co-ordinates and g1 , g2 are analytic in Γ2 .
With an alternative formulation of the growth condition in polar co-ordinates
[19] it can be shown easily that V ψ1 ∈ Bβ (Γ2 ) for all 0 < β < 1. Hence, from
Lemma D.1 it follows that
1 ≤ b3 e−b4 n . (D.14)
From (D.13) and (D.14) we conclude (D.10) with c1 = (b1 + b3 )/(4π) and c2 =
min{b2 , b4 }.

For further reading see [104].
A comprehensive list of numerical experiments can be found e.g. in the Book of
Numerical Experiments – BONE which can be downloaded from the home page of
M. Maischak, http://people.brunel.ac.uk/mastmmm/.
References
1. H. Ammari, A. Buffa, J.-C. Nédélec, A justification of eddy currents model for the Maxwell
equations. SIAM J. Appl. Math. 60, 1805–1823 (2000)
2. P.M. Anselone, Collectively Compact Operator Approximation Theory and Applications to
Integral Equations (Prentice-Hall, Englewood Cliffs, NJ, 1971)
3. D.N. Arnold, W.L. Wendland, On the asymptotic convergence of collocation methods. Math.
Comput. 41, 349–381 (1983)
4. K.E. Atkinson, A discrete Galerkin method for first kind integral equations with a logarithmic
kernel. J. Integr. Equ. Appl. 1, 343–363 (1988)
5. K.E. Atkinson, I.H. Sloan, The numerical solution of first-kind logarithmic-kernel integral
equations on smooth open arcs. Math. Comput. 56, 119–139 (1991)
6. H. Attouch, Variational Convergence for Functions and Operators. Applicable Mathematics
Series (Pitman (Advanced Publishing Program), Boston, MA, 1984)
7. J.-P. Aubin, Approximation of Elliptic Boundary-Value Problems. Pure and Applied Mathe-
matics, Vol. XXVI (Wiley-Interscience, New York-London-Sydney, 1972)
8. M. Aurada, M. Feischl, Th. Führer, M. Karkulik, J.M. Melenk, D. Praetorius, Classical FEM-
BEM coupling methods: nonlinearities, well-posedness, and adaptivity. Comput. Mech. 51,
399–419 (2014)
9. M. Aurada, M. Feischl, Th. Führer, M. Karkulik, J.M. Melenk, D. Praetorius, Local inverse
estimates for non-local boundary integral operators. Math. Comput. 86, 2651–2686 (2017)
10. M. Aurada, M. Feischl, Th. Führer, M. Karkulik, D. Praetorius, Efficiency and optimality of
some weighted-residual error estimator for adaptive 2D boundary element methods. Comput.
Methods Appl. Math. 13, 305–332 (2013)
11. M. Aurada, M. Feischl, Th. Führer, M. Karkulik, D. Praetorius, Energy norm based error
estimators for adaptive BEM for hypersingular integral equations. Appl. Numer. Math. 95,
15–35 (2015)
12. M. Aurada, M. Feischl, M. Karkulik, D. Praetorius, A posteriori error estimates for the
Johnson-Nédélec FEM-BEM coupling. Eng. Anal. Bound. Elem., 255–266 (2012)
13. M. Aurada, M. Feischl, D. Praetorius, Convergence of some adaptive FEM-BEM coupling
for elliptic but possibly nonlinear interface problems. ESAIM Math. Model. Numer. Anal.
46, 1147–1173 (2012)
14. I. Babuška, A.K. Aziz, Survey Lectures on the Mathematical Foundations of the Finite Ele-
ment Method. The Mathematical Foundations of the Finite Element Method with Applications

632 References
to Partial Differential Equations (Proc. Sympos., Univ. Maryland, Baltimore, MD, 1972).
Academic Press, New York, 1972, With the collaboration of G. Fix and R. B. Kellogg,
pp. 1–359
15. I. Babuška, A. Craig, J. Mandel, J. Pitkäranta, Efficient preconditioning for the p-version finite
element method in two dimensions. SIAM J. Numer. Anal. 28, 624–661 (1991)
16. I. Babuška, M.R. Dorr, Error estimates for the combined h and p versions of the finite element
method. Numer. Math. 37, 257–277 (1981)
17. I. Babuška, B.Q. Guo, The h-p version of the finite element method, part 1: The basic
approximation results. Comput. Mech. 1, 21–41 (1986)
18. I. Babuška, B.Q. Guo, The h-p version of the finite element method, part 2: General results
and applications. Comput. Mech. 1, 203–220 (1986)
19. I. Babuška, B.Q. Guo, Regularity of the solution of elliptic problems with piecewise analytic
data. I. Boundary value problems for linear elliptic equation of second order. SIAM J. Math.
Anal. 19, 172–203 (1988)
20. I. Babuška, B.Q. Guo, E.P. Stephan, The h-p version of the boundary element method with
geometric mesh on polygonal domains. Comput. Methods Appl. Mech. Eng. 80, 319–325
(1990)
21. I. Babuška, B.Q. Guo, E.P. Stephan, On the exponential convergence of the h-p version for
boundary element Galerkin methods on polygons. Math. Methods Appl. Sci. 12, 413–427
(1990)
22. I. Babuška, B.Q. Guo, M. Suri, Implementation of nonhomogeneous Dirichlet boundary
conditions in the p-version of the finite element method. Impact Comput. Sci. Eng. 1, 36–
63 (1989)
23. I. Babuška, M. Suri, The h-p version of the finite element method with quasi-uniform meshes.
RAIRO Modél. Math. Anal. Numér. 21, 199–238 (1987)
24. I. Babuška, M. Suri, The optimal convergence rate of the p-version of the finite element
method. SIAM J. Numer. Anal. 24, 750–776 (1987)
25. I. Babuška, M. Suri, The treatment of nonhomogeneous Dirichlet boundary conditions by the
p-version of the finite element method. Numer. Math. 55, 97–121 (1989)
26. I. Babuška, M. Vogelius, Feedback and adaptive finite element solution of one-dimensional
boundary value problems. Numer. Math. 44, 75–102 (1984)
27. M. Bach, S.A. Nazarov, W.L. Wendland, Stable Propagation of a Mode-1 Planar Crack in
an Isotropic Elastic Space. Comparison of the Irwin and the Griffith Approaches. Current
Problems of Analysis and Mathematical Physics (Taormina, Italy, 1998), Aracne, Rome,
2000, pp. 167–189
28. C. Baiocchi, A. Capelo, Variational and Quasivariational Inequalities - Applications to Free
Boundary Problems (Wiley, New York, 1984)
29. A. Bamberger, T. Ha-Duong, Formulation variationnelle espace-temps pour le calcul par
potentiel retardé de la diffraction d’une onde acoustique. I. Math. Methods Appl. Sci. 8, 405–
435 (1986)
30. L. Banjai, S. Sauter, Rapid solution of the wave equation in unbounded domains. SIAM J.
Numer. Anal. 47, 227–249 (2008/09)
31. L. Banjai, M. Schanz, Wave Propagation Problems Treated with Convolution Quadrature
and BEM. Fast Boundary Element Methods in Engineering and Industrial Applications. Lect.
Notes Appl. Comput. Mech., vol. 63 (Springer, Heidelberg, 2012), pp. 145–184
32. R.E. Bank, R.K. Smith, A posteriori error estimates based on hierarchical bases. SIAM J.
Numer. Anal. 30, 921–935 (1993)
33. L. Banz, H. Gimperlein, A. Issaoui, E.P. Stephan, Stabilized mixed hp-BEM for frictional
contact problems in linear elasticity. Numer. Math. 135, 217–263 (2017)
34. L. Banz, H. Gimperlein, Z. Nezhi, E.P. Stephan, Time domain BEM for sound radiation of
tires. Comput. Mech. 58, 45–57 (2016)
35. L. Banz, G. Milicic, N. Ovcharova, Improved stabilization technique for frictional contact
problems solved with hp-bem. Comput. Math. Appl. Mech. Eng. (2018, to appear)
References 633
36. L. Banz, A. Schröder, Biorthogonal Basis Functions in hp-adaptive fem for Elliptic Obstacle
Problems, 2015, pp. 1721–1742
37. L. Banz, E.P. Stephan, On hp-adaptive BEM for frictional contact problems in linear elasticity.
Comput. Math. Appl. 69, 559–581 (2015)
38. L. Banz, E.P. Stephan, Comparison of mixed hp-bem (stabilized and non-stabilized) for
frictional contact problems. J. Comput. Appl. Math. 295, 92–102 (2016)
39. H.J.C. Barbosa, T.J.R. Hughes, Circumventing the Babuška-Brezzi condition in mixed finite
element approximations of elliptic variational inequalities. Comput. Methods Appl. Mech.
Eng. 97, 193–210 (1992)
40. R. Beck, P. Deuflhard, R. Hiptmair, R.H.W. Hoppe, B. Wohlmuth, Adaptive multilevel
methods for edge element discretizations of Maxwell’s equations. Surveys Math. Indust. 8,
271–312 (1999)
41. A. Bendali, Numerical analysis of the exterior boundary value problem for the time-harmonic
Maxwell equations by a boundary finite element method. I. The continuous problem. Math.
Comput. 43, 29–46 (1984)
42. A. Bendali, Numerical analysis of the exterior boundary value problem for the time-harmonic
Maxwell equations by a boundary finite element method. II. The discrete problem. Math.
Comput. 43, 47–68 (1984)
43. C.A. Berenstein, R. Gay, Complex Variables. Graduate Texts in Mathematics, vol. 125
(Springer, New York, 1991)
44. J. Bergh, J. Löfström, Interpolation Spaces. An Introduction. Grundlehren der Mathematis-
chen Wissenschaften, No. 223 (Springer, Berlin-New York, 1976)
45. C. Bernardi, Y. Maday, A.T. Patera, Domain Decomposition by the Mortar Element Method.
Asymptotic and Numerical Methods for Partial Differential Equations with Critical Parame-
ters (Beaune, 1992), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 384 (Kluwer Acad.
Publ., Dordrecht, 1993), pp. 269–286
46. Ch. Bernardi, Y. Maday, Polynomial interpolation results in Sobolev spaces. J. Comput. Appl.
Math. 43, 53–80 (1992)
47. Ch. Bernardi, Y. Maday, Spectral Methods. Handbook of Numerical Analysis, Vol. V. Handb.
Numer. Anal. (North-Holland, Amsterdam, 1997), pp. 209–485
48. Ch. Bernardi, Y. Maday, Spectral, Spectral Element and Mortar Element Methods. Theory
and Numerics of Differential Equations (Durham, 2000) (Springer, Berlin, 2001), pp. 1–57
49. D.P. Bertsekas, Nondifferentiable optimization via approximation. Math. Program. Stud. 3,
1–25 (1975)
50. A. Bespalov, N. Heuer, The p-version of the boundary element method for a three-dimensional
crack problem. J. Integr. Equ. Appl. 17, 243–258 (2005)
51. A. Bespalov, N. Heuer, The p-version of the boundary element method for weakly singular
operators on piecewise plane open surfaces. Numer. Math. 106, 69–97 (2007)
52. A. Bespalov, N. Heuer, The hp-version of the boundary element method with quasi-uniform
meshes in three dimensions. M2AN Math. Model. Numer. Anal. 42, 821–849 (2008)
53. A. Bespalov, S. Nicaise, A priori error analysis of the BEM with graded meshes for the electric
field integral equation on polyhedral surfaces. Comput. Math. Appl. 71, 1636–1644 (2016)
54. J. Bielak, R.C. MacCamy, An exterior interface problem in two-dimensional elastodynamics.
Quart. Appl. Math. 41, 143–159 (1983/84)
55. E. Blum, W. Oettli, Mathematische Optimierung (Springer, Berlin-New York, 1975)
56. D. Boffi, F. Brezzi, M. Fortin, Mixed Finite Element Methods and Applications. Springer
Series in Computational Mathematics, vol. 44 (Springer, Heidelberg, 2013)
57. D. Boffi, M. Costabel, M. Dauge, L. Demkowicz, R. Hiptmair, Discrete compactness for the
p-version of discrete differential forms. SIAM J. Numer. Anal. 49, 135–158 (2011)
58. A. Bossavit, The computation of eddy-currents, in dimension 3, by using mixed finite
elements and boundary elements in association. Math. Comput. Modell. 15, 33–42 (1991)
59. D. Braess, A posteriori error estimators for obstacle problems—another look. Numer. Math.
101, 415–421 (2005)
634 References
60. D. Braess, Finite Elements, 3rd edn. (Cambridge University Press, Cambridge, 2007). Theory,
Fast Solvers, and Applications in Elasticity Theory. Translated from the German by L.L.
Schumaker
61. J.H. Bramble, A.H. Schatz, Higher order local accuracy by averaging in the finite element
method. Math. Comput. 31, 94–111 (1977)
62. J.H. Bramble, R.D. Lazarov, J.E. Pasciak, A least-squares approach based on a discrete minus
one inner product for first order systems. Math. Comput. 66, 935–955 (1997)
63. S.C. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods, 3rd edn. Texts
in Applied Mathematics, vol. 15 (Springer, New York, 2008)
64. H. Brezis, Èquations et inéquations non linéaires dans les espaces vectoriels en dualité. Ann.
Inst. Fourier (Grenoble) 18, 115–175 (1968)
65. F. Brezzi, M. Fortin, Mixed and Hybrid Finite Element Methods. Springer Series in Compu-
tational Mathematics, vol. 15 (Springer, New York, 1991)
66. U. Brink, E.P. Stephan, Adaptive coupling of boundary elements and mixed finite elements
for incompressible elasticity. Numer. Methods Partial Differ. Equ. 17, 79–92 (2001)
67. A. Buffa, Trace Theorems on Non-smooth Boundaries for Functional Spaces Related to
Maxwell Equations: An Overview. Computational Electromagnetics (Kiel, 2001). Lect. Notes
Comput. Sci. Eng., vol. 28 (Springer, Berlin, 2003), pp. 23–34
68. A. Buffa, S.H. Christiansen, The electric field integral equation on Lipschitz screens:
definitions and numerical approximation. Numer. Math. 94, 229–267 (2003)
69. A. Buffa, P. Ciarlet, Jr., On traces for functional spaces related to Maxwell’s equations. I.
An integration by parts formula in Lipschitz polyhedra. Math. Methods Appl. Sci. 24, 9–30
(2001)
70. A. Buffa, P. Ciarlet, Jr., On traces for functional spaces related to Maxwell’s equations.
II. Hodge decompositions on the boundary of Lipschitz polyhedra and applications. Math.
Methods Appl. Sci. 24, 31–48 (2001)
71. A. Buffa, M. Costabel, D. Sheen, On traces for H(curl, Ω) in Lipschitz domains. J. Math.
Anal. Appl. 276, 845–867 (2002)
72. A. Buffa, R. Hiptmair, Galerkin Boundary Element Methods for Electromagnetic Scattering.
Topics in Computational Wave Propagation. Lect. Notes Comput. Sci. Eng., vol. 31 (Springer,
Berlin, 2003), pp. 83–124
73. C. Carstensen, Interface problem in holonomic elastoplasticity. Math. Methods Appl. Sci. 16,
819–835 (1993)
74. C. Carstensen, Efficiency of a posteriori BEM-error estimates for first-kind integral equations
on quasi-uniform meshes. Math. Comput. 65, 69–84 (1996)
75. C. Carstensen, A posteriori error estimate for the symmetric coupling of finite elements and
boundary elements. Computing 57, 301–322 (1996)
76. C. Carstensen, An a posteriori error estimate for a first-kind integral equation. Math. Comput.
66, 139–155 (1997)
77. C. Carstensen, S. Bartels, Each averaging technique yields reliable a posteriori error control
in FEM on unstructured grids. I. Low order conforming, nonconforming, and mixed FEM.
Math. Comput. 71, 945–969 (2002)
78. C. Carstensen, B. Faermann, Mathematical foundation of a posteriori error estimates and
adaptive mesh-refining algorithms for boundary integral equations of the first kind. Eng. Anal.
Bound. Elem. 25, 497–509 (2001)
79. C. Carstensen, M. Feischl, M. Page, D. Praetorius, Axioms of adaptivity. Comput. Math.
Appl. 67, 1195–1253 (2016)
80. C. Carstensen, S.A. Funken, Coupling of nonconforming finite elements and boundary
elements. I. A priori estimates. Computing 62, 229–241 (1999)
81. C. Carstensen, S.A. Funken, Coupling of mixed finite elements and boundary elements. IMA
J. Numer. Anal. 20, 461–480 (2000)
82. C. Carstensen, S.A. Funken, E.P. Stephan, A posteriori error estimates for hp-boundary
element methods. Appl. Anal. 61, 233–253 (1996)
References 635
83. C. Carstensen, S.A. Funken, E.P. Stephan, On the adaptive coupling of FEM and BEM in
2-d-elasticity. Numer. Math. 77, 187–221 (1997)
84. C. Carstensen, D. Gallistl, J. Gedicke, Justification of the saturation assumption. Numer.
Math. 134, 1–25 (2016)
85. C. Carstensen, J. Gwinner, FEM and BEM coupling for a nonlinear transmission problem
with Signorini contact. SIAM J. Numer. Anal. 34, 1845–1864 (1997)
86. C. Carstensen, M. Maischak, D. Praetorius, E.P. Stephan, Residual-based a posteriori error
estimate for hypersingular equation on surfaces. Numer. Math. 97, 397–425 (2004)
87. C. Carstensen, M. Maischak, E.P. Stephan, A posteriori error estimate and h-adaptive
algorithm on surfaces for Symm’s integral equation. Numer. Math. 90, 197–213 (2001)
88. C. Carstensen, D. Praetorius, Averaging techniques for the effective numerical solution of
Symm’s integral equation of the first kind. SIAM J. Sci. Comput. 27, 1226–1260 (2006)
89. C. Carstensen, D. Praetorius, Convergence of adaptive boundary element methods. J. Integr.
Equ. Appl. 24, 1–23 (2012)
90. C. Carstensen, O. Scherf, P. Wriggers, Adaptive finite elements for elastic bodies in contact.
SIAM J. Sci. Comput. 20, 1605–1626 (1999)
91. C. Carstensen, E.P. Stephan, Adaptive coupling of boundary elements and finite elements.
RAIRO Modél. Math. Anal. Numér. 29, 779–817 (1995)
92. C. Carstensen, E.P. Stephan, A posteriori error estimates for boundary element methods.
Math. Comput. 64, 483–500 (1995)
93. C. Carstensen, E.P. Stephan, Adaptive boundary element methods for some first kind integral
equations. SIAM J. Numer. Anal. 33, 2166–2183 (1996)
94. J. Chabassier, A. Chaigne, P. Joly, Time domain simulation of a piano. Part 1: Model
description. ESAIM Math. Model. Numer. Anal. 48, 1241–1278 (2014)
95. G.A. Chandler, I.G. Graham, Product integration-collocation methods for noncompact inte-
gral operator equations. Math. Comput. 50(181), 125–138 (1988)
96. G.A. Chandler, I.H. Sloan, Spline qualocation methods for boundary integral equations.
Numer. Math. 58, 537–567 (1990)
97. D. Chen, V.A. Menegatto, X. Sun, A necessary and sufficient condition for strictly positive
definite functions on spheres. Proc. Am. Math. Soc. 131, 2733–2740 (2003)
98. G. Chen, J. Zhou, Boundary Element Methods with Applications to Nonlinear Problems, 2nd
edn. Atlantis Studies in Mathematics for Engineering and Science, vol. 7b (Atlantis Press,
Paris; World Scientific Publishing, Hackensack, NJ, 2010)
99. X. Chen, Smoothing methods for nonsmooth, nonconvex minimization. Math. Program. 134,
71–99 (2012)
100. X. Chen, L. Qi, D. Sun, Global and superlinear convergence of the smoothing Newton method
and its application to general box constrained variational inequalities. Math. Comput. 67,
519–540 (1998)
101. A. Chernov, Nonconforming boundary elements and finite elements for interface and contact
problems with friction – hp-version for mortar, penalty and Nitsche’s methods, Ph.D. thesis,
Leibniz Universität Hannover, 2006
102. A. Chernov, M. Maischak, E.P. Stephan, A priori error estimates for hp penalty BEM for
contact problems in elasticity. Comput. Methods Appl. Mech. Eng. 196, 3871–3880 (2007)
103. A. Chernov, M. Maischak, E.P. Stephan, hp-mortar boundary element method for two-body
contact problems with friction. Math. Methods Appl. Sci. 31, 2029–2054 (2008)
104. A. Chernov, T. von Petersdorff, Ch. Schwab, Exponential convergence of hp quadrature for
integral operators with Gevrey kernels. ESAIM Math. Model. Numer. Anal. 45, 387–422
(2011)
105. P.G. Ciarlet, Basic Error Estimates for Elliptic Problems. Handbook of Numerical Analysis,
Vol. II (North-Holland, Amsterdam, 1991), pp. 17–351
106. P.G. Ciarlet, The Finite Element Method for Elliptic Problems. Classics in Applied Mathe-
matics, vol. 40 (Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,
2002)
636 References
107. X. Claeys, R. Hiptmair, Integral equations on multi-screens. Integr. Equ. Oper. Theory 77,
167–197 (2013)
108. X. Claeys, R. Hiptmair, Integral equations for electromagnetic scattering at multi-screens.
Integr. Equ. Oper. Theory 84, 33–68 (2016)
109. F.H. Clarke, Optimization and Nonsmooth Analysis, 2nd edn. Classics in Applied Mathemat-
ics, vol. 5 (Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1990)
110. Ph. Clément, Approximation by finite element functions using local regularization. RAIRO
Analyse Numérique 9, 77–84 (1975)
111. A.R. Conn, N.I.M. Gould, P.L. Toint, Trust-Region Methods. MPS/SIAM Series on Optimiza-
tion (Society for Industrial and Applied Mathematics (SIAM); Mathematical Programming
Society (MPS), Philadelphia, PA, 2000)
112. M. Costabel, Principles of Boundary Element Methods. Finite Elements in Physics (Lausanne,
1986) (North-Holland, Amsterdam, 1987), pp. 243–274
113. M. Costabel, Symmetric Methods for the Coupling of Finite Elements and Boundary Elements
(Invited Contribution). Boundary Elements IX, Vol. 1 (Stuttgart, 1987) (Comput. Mech.,
Southampton, 1987), pp. 411–420
114. M. Costabel, Boundary integral operators on Lipschitz domains: elementary results. SIAM J.
Math. Anal. 19, 613–626 (1988)
115. M. Costabel, A Symmetric Method for the Coupling of Finite Elements and Boundary
Elements. The Mathematics of Finite Elements and Applications, VI (Uxbridge, 1987)
(Academic Press, London, 1988), pp. 281–288
116. M. Costabel, Randelemente (Vorlesungsmanuskript, TU Darmstadt, 1989)
117. M. Costabel, Time-Dependent Problems with a Boundary Integral Equation Method. Ency-
clopedia of Computational Mechanics, 2004
118. M. Costabel, Some Historical Remarks on the Positivity of Boundary Integral Operators.
Boundary Element Analysis. Lect. Notes Appl. Comput. Mech., vol. 29 (Springer, Berlin,
2007), pp. 1–27
119. M. Costabel, M. Dauge, S. Nicaise, Analytic regularity for linear elliptic systems in polygons
and polyhedra. Math. Models Methods Appl. Sci. 22, 1250015, 63 (2012)
120. M. Costabel, V.J. Ervin, E.P. Stephan, Symmetric coupling of finite elements and boundary
elements for a parabolic-elliptic interface problem. Quart. Appl. Math. 48, 265–279 (1990)
121. M. Costabel, V.J. Ervin, E.P. Stephan, Experimental convergence rates for various couplings
of boundary and finite elements. Math. Comput. Model. 15, 93–102 (1991)
122. M. Costabel, V.J. Ervin, E.P. Stephan, Quadrature and collocation methods for the double
layer potential on polygons. Z. Anal. Anwendungen 12, 699–707 (1993)
123. M. Costabel, A. McIntosh, On Bogovskiı̆ and regularized Poincaré integral operators for de
Rham complexes on Lipschitz domains. Math. Z. 265, 297–320 (2010)
124. M. Costabel, F. Penzel, R. Schneider, Convergence of boundary element collocation methods
for Dirichlet and Neumann screen problems in R3 . Appl. Anal. 49, 101–117 (1993)
125. M. Costabel, E. Stephan, W.L. Wendland, Zur Randintegralmethode für das erste Fundamen-
talproblem der ebenen Elastizitätstheorie auf Polygongebieten. Recent Trends in Mathematics
(Reinhardsbrunn, 1982). Teubner-Texte zur Math., vol. 50 (Teubner, Leipzig, 1982), pp. 56–
68
126. M. Costabel, E.P. Stephan, Curvature terms in the asymptotic expansions for solutions of
boundary integral equations on curved polygons. J. Integr. Equ. 5, 353–371 (1983)
127. M. Costabel, E.P. Stephan, The normal derivative of the double layer potential on polygons
and Galerkin approximation. Appl. Anal. 16, 205–228 (1983)
128. M. Costabel, E.P. Stephan, Boundary Integral Equation for Mixed Boundary Value Problem
in Polygonal Domains and Galerkin Approximation. Mathematical Models and Methods in
Mechanics. Banach Center Publications, vol. 15 (PWN - Polish Scientific Publishers, 1985),
pp. 175–251
129. M. Costabel, E.P. Stephan, A direct boundary integral equation method for transmission
problems. J. Math. Anal. Appl. 106, 367–413 (1985)
References 637
130. M. Costabel, E.P. Stephan, An improved boundary element Galerkin method for three-
dimensional crack problems. Integr. Equ. Oper. Theory 10, 467–504 (1987)
131. M. Costabel, E.P. Stephan, On the convergence of collocation methods for boundary integral
equations on polygons. Math. Comput. 49, 461–478 (1987)
132. M. Costabel, E.P. Stephan, Coupling of Finite Elements and Boundary Elements for Inhomo-
geneous Transmission Problems in R3 . The Mathematics of Finite Elements and Applications,
VI (Uxbridge, 1987) (Academic Press, London, 1988), pp. 289–296
133. M. Costabel, E.P. Stephan, Coupling of Finite Elements and Boundary Elements for Transmis-
sion Problems of Elastic Waves in R3 . Advanced Boundary Element Methods (San Antonio,
TX, 1987) (Springer, Berlin, 1988), pp. 117–124
134. M. Costabel, E.P. Stephan, Duality estimates for the numerical solution of integral equations.
Numer. Math. 54, 339–353 (1988)
135. M. Costabel, E.P. Stephan, Strongly elliptic boundary integral equations for electromagnetic
transmission problems. Proc. Roy. Soc. Edinburgh Sect. A 109, 271–296 (1988)
136. M. Costabel, E.P. Stephan, Coupling of finite and boundary element methods for an
elastoplastic interface problem. SIAM J. Numer. Anal. 27, 1212–1226 (1990)
137. M. Costabel, E.P. Stephan, Integral equations for transmission problems in linear elasticity. J.
Integr. Equ. Appl. 2, 211–223 (1990)
138. M. Costabel, E.P. Stephan, W.L. Wendland, On boundary integral equations of the first kind
for the bi-Laplacian in a polygonal plane domain. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4)
10, 197–241 (1983)
139. R.W. Cottle, J.S. Pang, R.E. Stone, The Linear Complementarity Problem (Academic Press,
Boston, 1992)
140. G.B. Dantzig, Linear Programming and Extensions, corrected edn. Princeton Landmarks in
Mathematics (Princeton University Press, Princeton, NJ, 1998)
141. M. Dauge, Elliptic Boundary Value Problems on Corner Domains. Smoothness and Asymp-
totics of Solutions. Lecture Notes in Mathematics, vol. 1341 (Springer, Berlin, 1988)
142. R.A. DeVore, G.G. Lorentz, Constructive Approximation. Grundlehren der Mathematischen
Wissenschaften, vol. 303 (Springer, Berlin, 1993)
143. W. Dörfler, A convergent adaptive algorithm for Poisson’s equation. SIAM J. Numer. Anal.
33, 1106–1124 (1996)
144. W. Dörfler, R.H. Nochetto, Small data oscillation implies the saturation assumption. Numer.
Math. 91, 1–12 (2002)
145. C. Eck, Existence and regularity of solutions to contact problems with friction (german),
Ph.D. thesis, Universität Stuttgart, 1996
146. C. Eck, J. Jarušek, M. Krbec, Unilateral Contact Problems. Pure and Applied Mathematics,
vol. 270 (Chapman & Hall/CRC, Boca Raton, FL, 2005)
147. C. Eck, S.A. Nazarov, W.L. Wendland, Asymptotic analysis for a mixed boundary-value
contact problem. Arch. Ration. Mech. Anal. 156, 275–316 (2001)
148. C. Eck, H. Schulz, O. Steinbach, W.L. Wendland, An Adaptive Boundary Element Method
for Contact Problems. Error Controlled Adaptive Finite Elements in Solid Mechanics (Wiley,
2003), pp. 181–209
149. C. Eck, O. Steinbach, W.L. Wendland, A symmetric boundary element method for contact
problems with friction. Math. Comput. Simul. 50, 43–61 (1999)
150. C. Eck, W.L. Wendland, A residual-based error estimator for BEM-discretizations of contact
problems. Numer. Math. 95, 253–282 (2003)
151. I. Ekeland, R. Témam, Convex Analysis and Variational Problems. Classics in Applied
Mathematics, vol. 28 (Society for Industrial and Applied Mathematics (SIAM), Philadelphia,
PA, 1999)
152. J. Elschner, The double layer potential operator over polyhedral domains. I. Solvability in
weighted Sobolev spaces. Appl. Anal. 45, 117–134 (1992)
153. J. Elschner, The double layer potential operator over polyhedral domains. II. Spline Galerkin
methods. Math. Methods Appl. Sci. 15, 23–37 (1992)
638 References
154. J. Elschner, The h-p-version of spline approximation methods for Mellin convolution
equations. J. Integr. Equ. Appl. 5, 47–73 (1993)
155. J. Elschner, I.G. Graham, An optimal order collocation method for first kind boundary integral
equations on polygons. Numer. Math. 70, 1–31 (1995)
156. J. Elschner, I.G. Graham, Quadrature methods for Symm’s integral equation on polygons.
IMA J. Numer. Anal. 17, 643–664 (1997)
157. J. Elschner, I.G. Graham, Numerical methods for integral equations of Mellin type. J. Comput.
Appl. Math. 125(1–2), 423–437 (2000)
158. J. Elschner, Y. Jeon, I.H. Sloan, E.P. Stephan, The collocation method for mixed boundary
value problems on domains with curved polygonal boundaries. Numer. Math. 76, 355–381
(1997)
159. J. Elschner, E.P. Stephan, A discrete collocation method for Symm’s integral equation on
curves with corners. J. Comput. Appl. Math. 75, 131–146 (1996)
160. H. Engels, Numerical Quadrature and Cubature. Computational Mathematics and Applica-
tions (Academic Press, London-New York, 1980)
161. C. Erath, S. Ferraz-Leite, S. Funken, D. Praetorius, Energy norm based a posteriori error
estimation for boundary element methods in two dimensions. Appl. Numer. Math. 59, 2713–
2734 (2009)
162. C. Erath, S. Funken, P. Goldenits, D. Praetorius, Simple error estimators for the Galerkin
BEM for some hypersingular integral equation in 2D. Appl. Anal. 92, 1194–1216 (2013)
163. E. Ernst, M. Théra, A converse to the Lions-Stampacchia theorem. ESAIM Control Optim.
Calc. Var. 15, 810–817 (2009)
164. V.J. Ervin, N. Heuer, An adaptive boundary element method for the exterior Stokes problem
in three dimensions. IMA J. Numer. Anal. 26, 297–325 (2006)
165. V.J. Ervin, N. Heuer, E.P. Stephan, On the h-p version of the boundary element method
for Symm’s integral equation on polygons. Comput. Methods Appl. Mech. Eng. 110, 25–38
(1993)
166. V.J. Ervin, E.P. Stephan, Collocation with Chebyshev polynomials for a hypersingular integral
equation on an interval. J. Comput. Appl. Math. 43, 221–229 (1992)
167. G.I. Eskin, Boundary Value Problems for Elliptic Pseudodifferential Equations. Translations
of Mathematical Monographs, vol. 52 (American Mathematical Society, Providence, RI,
1981)
168. F. Facchinei, J.-S. Pang, Finite-Dimensional Variational Inequalities and Complementarity
Problems, Vol. I. Springer Series in Operations Research (Springer, New York, 2003)
169. F. Facchinei, J.-S. Pang, Finite-Dimensional Variational Inequalities and Complementarity
Problems, Vol. II. Springer Series in Operations Research (Springer, New York, 2003)
170. B. Faermann, Local a-posteriori error indicators for the Galerkin discretization of boundary
integral equations. Numer. Math. 79, 43–76 (1998)
171. B. Faermann, Localization of the Aronszajn-Slobodeckij norm and application to adaptive
boundary element methods. I. The two-dimensional case. IMA J. Numer. Anal. 20, 203–234
(2000)
172. B. Faermann, Localization of the Aronszajn-Slobodeckij norm and application to adaptive
boundary element methods. II. The three-dimensional case. Numer. Math. 92, 467–499 (2002)
173. R.S. Falk, Error estimates for the approximation of a class of variational inequalities. Math.
Comput. 28, 963–971 (1974)
174. M. Feischl, Th. Führer, N. Heuer, M. Karkulik, D. Praetorius, Adaptive boundary element
methods. Arch. Comput. Methods Eng. 22, 309–389 (2015)
175. M. Feischl, Th. Führer, M. Karkulik, J.M. Melenk, D. Praetorius, Quasi-optimal convergence
rates for adaptive boundary element methods with data approximation, part I: Weakly-
singular integral equation. Calcolo 51, 531–562 (2014)
176. M. Feischl, Th. Führer, G. Mitscha-Eibl, D. Praetorius, E.P. Stephan, Convergence of adaptive
BEM and adaptive FEM-BEM coupling for estimators without h-weighting factor. Comput.
Methods Appl. Math. 14, 485–508 (2014)
References 639
177. M. Feischl, M. Karkulik, J.M. Melenk, D. Praetorius, Quasi-optimal convergence rate for an
adaptive boundary element method. SIAM J. Numer. Anal. 51, 1327–1348 (2013)
178. S. Ferraz-Leite, D. Praetorius, Simple a posteriori error estimators for the h-version of the
boundary element method. Computing 83, 135–162 (2008)
179. G. Fichera, Boundary Value Problems of Elasticity with Unilateral Constraints. Handbuch der
Physik, Festkörpermechanik, vol. VIa/2, ed. by S. Flügge (Springer, Berlin, 1972), pp. 391–
424
180. G.J. Fix, S. Gulati, G.I. Wakoff, On the use of singular functions with finite element
approximations. J. Comput. Phys. 13, 209–228 (1973)
181. Th. Führer, Zur Kopplung von Finiten Elementen und Randelementen, Ph.D. thesis, TU Wien,
2014
182. A. Gachechiladze, R. Gachechiladze, J. Gwinner, D. Natroshvili, A boundary variational
inequality approach to unilateral contact problems with friction for micropolar hemitropic
solids. Math. Methods Appl. Sci. 33, 2145–2161 (2010)
183. A. Gachechiladze, R. Gachechiladze, J. Gwinner, D. Natroshvili, Contact problems with
friction for hemitropic solids: boundary variational inequality approach. Appl. Anal. 90, 279–
303 (2011)
184. R. Gachechiladze, J. Gwinner, D. Natroshvili, A boundary variational inequality approach
to unilateral contact with hemitropic materials. Mem. Differ. Equ. Math. Phys. 39, 69–103
(2006)
185. D. Gaier, Integralgleichungen erster Art und konforme Abbildung. Math. Z. 147, 113–129
(1976)
186. G.N. Gatica, G.C. Hsiao, F.-J. Sayas, Relaxing the hypotheses of Bielak-MacCamy’s BEM-
FEM coupling. Numer. Math. 120, 465–487 (2012)
187. G.N. Gatica, N. Heuer, E.P. Stephan, An implicit-explicit residual error estimator for the
coupling of dual-mixed finite elements and boundary elements in elastostatics. Math. Methods
Appl. Sci. 24, 179–191 (2001)
188. G.N. Gatica, M. Maischak, E.P. Stephan, Numerical analysis of a transmission problem with
Signorini contact using mixed-FEM and BEM. ESAIM Math. Model. Numer. Anal. 45, 779–
802 (2011)
189. H. Gimperlein, M. Maischak, E. Schrohe, E.P. Stephan, Adaptive FE-BE coupling for strongly
nonlinear transmission problems with Coulomb friction. Numer. Math. 117, 307–332 (2011)
190. H. Gimperlein, M. Maischak, E.P. Stephan, Adaptive time domain boundary element methods
with engineering applications. J. Integr. Equ. Appl. 29, 75–105 (2017)
191. H. Gimperlein, F. Meyer, C. Özdemir, D. Stark, E.P. Stephan, Boundary elements with mesh
refinements for the wave equation. Numer. Math. (2018)
192. H. Gimperlein, F. Meyer, C. Özdemir, D. Stark, E.P. Stephan, Time domain boundary
elements for dynamic contact problems. Comput. Methods Appl. Mech. Eng. 333, 147–175
(2018)
193. H. Gimperlein, Z. Nezhi, E.P. Stephan, A priori error estimates for a time-dependent boundary
element method for the acoustic wave equation in a half-space. Math. Methods Appl. Sci. 40,
448–462 (2017)
194. H. Gimperlein, C. Özdemir, D. Stark, E.P. Stephan, A residual a posteriori error estimate for
the time-domain boundary element method. Preprint 2017
195. H. Gimperlein, C. Özdemir, E.P. Stephan, Time domain boundary element methods for the
Neumann problem: error estimates and acoustics problems. J. Comput. Math. 36, 70–89
(2018)
196. V. Girault, P.-A. Raviart, Finite Element Methods for Navier-Stokes Equations. Springer
Series in Computational Mathematics, vol. 5 (Springer, Berlin, 1986). Theory and Algorithms
197. M. Gläfke, Adaptive methods for time domain boundary integral equations, Ph.D. thesis,
Brunel University, 2012
198. M. Gläfke, M. Maischak, E.P. Stephan, Coupling of FEM and BEM for a transmission
problem with nonlinear interface conditions. Hierarchical and residual error indicators. Appl.
Numer. Math. 62, 736–753 (2012)
640 References
199. R. Glowinski, Numerical Methods for Nonlinear Variational Problems. Scientific Computa-
tion (Springer, Berlin, 2008)
200. R. Glowinski, J.-L. Lions, R. Trémolières, Numerical Analysis of Variational Inequalities.
Studies in Mathematics and Its Applications, vol. 8 (North-Holland Publishing, Amsterdam-
New York, 1981)
201. D. Goeleven, Noncoercive Variational Problems and Related Results. Pitman Research Notes
in Mathematics Series, vol. 357 (Longman, Harlow, 1996)
202. I.C. Gohberg, I.A. Feldman, Convolution Equations and Projection Methods for Their Solu-
tion (American Mathematical Society, Providence, RI, 1974). Translations of Mathematical
Monographs, Vol. 41
203. M.O. González, Classical Complex Analysis. Monographs and Textbooks in Pure and Applied
Mathematics, vol. 151 (Marcel Dekker, New York, 1992)
204. P. Grisvard, Elliptic Problems in Nonsmooth Domains. Monographs and Studies in Mathe-
matics, vol. 24 (Pitman (Advanced Publishing Program), Boston, MA, 1985). Reprint 2011
205. H. Guediri, On a boundary variational inequality of the second kind modelling a friction
problem. Math. Methods Appl. Sci. 25, 93–114 (2002)
206. N.M. Günter, Potential Theory and Its Applications to Basic Problems of Mathematical
Physics (Frederick Ungar Publishing, New York, 1967)
207. B.Q. Guo, The h-p Version of the Finite Element Method for Solving Boundary Value
Problems in Polyhedral Domains. Boundary Value Problems and Integral Equations in
Nonsmooth Domains (Luminy, 1993). Lecture Notes in Pure and Appl. Math., vol. 167
(Dekker, New York, 1995), pp. 101–120
208. B.Q. Guo, N. Heuer, The optimal rate of convergence of the p-version of the boundary element
method in two dimensions. Numer. Math. 98, 499–538 (2004)
209. B.Q. Guo, N. Heuer, The optimal convergence of the h-p version of the boundary element
method with quasiuniform meshes for elliptic problems on polygonal domains. Adv. Comput.
Math. 24, 353–374 (2006)
210. B.Q. Guo, N. Heuer, E.P. Stephan, The h-p version of the boundary element method for
transmission problems with piecewise analytic data. SIAM J. Numer. Anal. 33, 789–808
(1996)
211. B.Q. Guo, T. von Petersdorff, E.P. Stephan, An hp Version of the Boundary Element Method
for Plane Mixed Boundary Value Problems. Advances in Boundary Elements. Comput.
Mech., vol. 1 (Springer, Berlin, 1989), pp. 95–103
212. J. Gwinner, On fixed points and variational inequalities—a circular tour. Nonlinear Anal. 5,
565–583 (1981)
213. J. Gwinner, Discretization of semicoercive variational inequalities. Aequationes Math. 42,
72–79 (1991)
214. J. Gwinner, Boundary Element Convergence for Unilateral Harmonic Problems. Progress in
Partial Differential Equations: Calculus of Variations, Applications (Pont-à-Mousson, 1991).
Pitman Res. Notes Math. Ser., vol. 267 (Longman Sci. Tech., Harlow, 1992), pp. 200–213
215. J. Gwinner, A discretization theory for monotone semicoercive problems and finite element
convergence for p-harmonic Signorini problems. Z. Angew. Math. Mech. 74, 417–427 (1994)
216. J. Gwinner, A note on pseudomonotone functions, regularization, and relaxed coerciveness.
Nonlinear Anal. 30, 4217–4227 (1997)
217. J. Gwinner, On the p-version approximation in the boundary element method for a variational
inequality of the second kind modelling unilateral contact and given friction. Appl. Numer.
Math. 59, 2774–2784 (2009)
218. J. Gwinner, hp-FEM convergence for unilateral contact problems with Tresca friction in plane
linear elastostatics. J. Comput. Appl. Math. 254, 175–184 (2013)
219. J. Gwinner, Lagrange multipliers and mixed formulations for some inequality constrained
variational inequalities and some nonsmooth unilateral problems. Optimization 66, 1323–
1336 (2017)
220. J. Gwinner, N. Ovcharova, From solvability and approximation of variational inequalities to
solution of nondifferentiable optimization problems in contact mechanics. Optimization 64,
1683–1702 (2015)
References 641
221. J. Gwinner, E.P. Stephan, Boundary Element Convergence for a Variational Inequality of
the Second Kind. Parametric Optimization and Related Topics, III (Güstrow, 1991). Approx.
Optim., vol. 3 (Lang, Frankfurt am Main, 1993), pp. 227–241
222. J. Gwinner, E.P. Stephan, A boundary element procedure for contact problems in plane linear
elastostatics. RAIRO M2AN 27, 457–480 (1993)
223. T. Ha-Duong, On Retarded Potential Boundary Integral Equations and Their Discretisation.
Topics in Computational Wave Propagation. Lect. Notes Comput. Sci. Eng., vol. 31 (Springer,
Berlin, 2003), pp. 301–336
224. T. Ha-Duong, B. Ludwig, I. Terrasse, A Galerkin BEM for transient acoustic scattering by an
absorbing obstacle. Int. J. Numer. Methods Eng. 57, 1845–1882 (2003)
225. W. Hackbusch, Integral Equations. International Series of Numerical Mathematics, vol. 120
(Birkhäuser Verlag, Basel, 1995)
226. G. Hämmerlin, K.-H. Hoffmann, Numerische Mathematik, 4th edn. (Springer-Lehrbuch,
Springer, Berlin, 1994)
227. H. Han, A direct boundary element method for Signorini problems. Math. Comput. 55, 115–
128 (1990)
228. H. Han, A new class of variational formulations for the coupling of finite and boundary
element methods. J. Comput. Math. 8, 223–232 (1990)
229. F. Hartmann, The Somigliana identity on piecewise smooth surfaces. J. Elasticity 11, 403–423
(1981)
230. T. Hartmann, E.P. Stephan, Rates of convergence for collocation with Jacobi polynomials for
the airfoil equation. J. Comput. Appl. Math. 51, 179–191 (1994)
231. T. Hartmann, E.P. Stephan, A Discrete Collocation Method for a Hypersingular Integral
Equation on Curves with Corners. Contemporary Computational Mathematics - a Celebration
of the 80th Birthday of Ian Sloan (Springer, 2018)
232. J. Haslinger, C.C. Baniotopoulos, P.D. Panagiotopoulos, A boundary multivalued integral
“equation” approach to the semipermeability problem. Appl. Math. 38, 39–60 (1993)
233. N. Heuer, Additive Schwarz method for the p-version of the boundary element method for the
single layer potential operator on a plane screen. Numer. Math. 88, 485–511 (2001)
234. N. Heuer, An hp-adaptive refinement strategy for hypersingular operators on surfaces. Numer.
Methods Partial Differ. Equ. 18, 396–419 (2002)
235. N. Heuer, M. Maischak, E.P. Stephan, Exponential convergence of the hp-version for the
boundary element method on open surfaces. Numer. Math. 83, 641–666 (1999)
236. N. Heuer, M. Maischak, E.P. Stephan, Preconditioned minimum residual iteration for the h-p
version of the coupled FEM/BEM with quasi-uniform meshes. Numer. Linear Algebra Appl.
6, 435–456 (1999)
237. N. Heuer, S. Meddahi, F.-J. Sayas, Symmetric coupling of LDG-FEM and DG-BEM. J. Sci.
Comput. 68, 303–325 (2016)
238. N. Heuer, M.E. Mellado, E.P. Stephan, hp-adaptive two-level methods for boundary integral
equations on curves. Computing 67, 305–334 (2001)
239. N. Heuer, M.E. Mellado, E.P. Stephan, A p-adaptive algorithm for the BEM with the
hypersingular operator on the plane screen. Int. J. Numer. Methods Eng. 53, 85–104 (2002)
240. N. Heuer, E.P. Stephan, The hp-version of the boundary element method on polygons. J.
Integr. Equ. Appl. 8, 173–212 (1996)
241. N. Heuer, E.P. Stephan, Boundary integral operators in countably normed spaces. Math.
Nachr. 191, 123–151 (1998)
242. N. Heuer, E.P. Stephan, The Poincaré-Steklov Operator Within Countably Normed Spaces.
Mathematical Aspects of Boundary Element Methods (Palaiseau, 1998). Res. Notes Math.,
vol. 414 (Chapman & Hall/CRC, Boca Raton, FL, 2000), pp. 152–164
243. N. Heuer, T. Tran, Radial basis functions for the solution of hypersingular operators on open
surfaces. Comput. Math. Appl. 63, 1504–1518 (2012)
244. S. Hildebrandt, E. Wienholtz, Constructive proofs of representation theorems in separable
Hilbert space. Commun. Pure Appl. Math. 17, 369–373 (1964)
245. E. Hille, Analytic Function Theory, Vol. II (Ginn and Co., Boston-New York-Toronto, 1962)
642 References
246. R. Hiptmair, Symmetric coupling for eddy current problems. SIAM J. Numer. Anal. 40, 41–
65 (2002)
247. R. Hiptmair, Coupling of finite elements and boundary elements in electromagnetic scattering.
SIAM J. Numer. Anal. 41, 919–944 (2003)
248. R. Hiptmair, C. Schwab, Natural boundary element methods for the electric field integral
equation on polyhedra. SIAM J. Numer. Anal. 40, 66–86 (2002)
249. I. Hlaváček, J. Haslinger, J. Nečas, J. Lovíšek, Solution of Variational Inequalities in
Mechanics. Applied Mathematical Sciences, vol. 66 (Springer, New York, 1988)
250. H. Holm, M. Maischak, E.P. Stephan, The hp-version of the boundary element method for
Helmholtz screen problems. Computing 57, 105–134 (1996)
251. H. Holm, M. Maischak, E.P. Stephan, Exponential convergence of the h-p version BEM for
mixed boundary value problems on polyhedrons. Math. Methods Appl. Sci. 31, 2069–2093
(2008)
252. L. Hörmander, Sur la fonction d’appui des ensembles convexes dans un espace localement
convexe. Ark. Mat. 3, 181–186 (1955)
253. L. Hörmander, Linear Partial Differential Operators (Springer, Berlin-New York, 1976)
254. P. Houston, E. Süli, A note on the design of hp-adaptive finite element methods for elliptic
partial differential equations. Comput. Methods Appl. Mech. Eng. 194, 229–243 (2005)
255. G.C. Hsiao, O. Steinbach, W.L. Wendland, Domain decomposition methods via boundary
integral equations. J. Comput. Appl. Math. 125, 521–537 (2000)
256. G.C. Hsiao, E.P. Stephan, W.L. Wendland, On the Dirichlet problem in elasticity for a domain
exterior to an arc. J. Comput. Appl. Math. 34, 1–19 (1991)
257. G.C. Hsiao, W.L. Wendland, A finite element method for some integral equations of the first
kind. J. Math. Anal. Appl. 58, 449–481 (1977)
258. G.C. Hsiao, W.L. Wendland, The Aubin-Nitsche lemma for integral equations. J. Integr. Equ.
3, 299–315 (1981)
259. G.C. Hsiao, W.L. Wendland, Boundary Integral Equations. Applied Mathematical Sciences,
vol. 164 (Springer, Berlin, 2008)
260. G.C. Hsiao, W.L. Wendland, Boundary Element Methods: Foundation and Error Analysis.
Encyclopedia of Computational Mechanics, vol. 1, ed. by E. Stein et al. (Wiley, Chichester,
2004), pp. 339–373
261. J. Jarušek, Contact problems with bounded friction coercive case. Czechoslov. Math. J.
33(108), 237–261 (1983)
262. Y. Jeon, I.H. Sloan, E.P. Stephan, J. Elschner, Discrete qualocation methods for logarithmic-
kernel integral equations on a piecewise smooth boundary. Adv. Comput. Math. 7, 547–571
(1997)
263. C. Johnson, J.C. Nédélec, On the coupling of boundary integral and finite element methods.
Math. Comput. 35, 1063–1079 (1980)
264. M. Karkulik, Zur Konvergenz und Quasioptimalität adaptiver Randelementmethoden, Ph.D.
thesis, TU Wien, 2012
265. M. Karkulik, J.M. Melenk, Local high-order regularization and applications to hp-methods.
Comput. Math. Appl. 70, 1606–1639 (2015)
266. N. Kikuchi, J.T. Oden, Contact Problems in Elasticity: A Study of Variational Inequalities and
Finite Element Methods. SIAM Studies in Applied Mathematics, vol. 8 (Society for Industrial
and Applied Mathematics (SIAM), Philadelphia, PA, 1988)
267. D. Kinderlehrer, G. Stampacchia, An Introduction to Variational Inequalities and Their
Applications. Classics in Applied Mathematics, vol. 31 (Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA, 2000)
268. A.Yu. Kokotov, P. Neittaanmäki, B.A. Plamenevskii, The Neumann problem for the wave
equation in a cone. J. Math. Sci. (New York) 102, 4400–4428 (2000)
269. A.Yu. Kokotov, P. Neittaanmäki, B.A. Plamenevskii, Diffraction on a cone: the asymptotic of
solutions near the vertex.. J. Math. Sci. (New York) 109, 1894–1910 (2002)
270. V.A. Kondratiev, Boundary value problems for elliptic equations in domains with conical and
angular points. Trans. Moscow Math. Soc. 16, 227–313 (1967)
References 643
271. V.A. Kozlov, V.G. Maz’ya, J. Rossmann, Elliptic Boundary Value Problems in Domains with
Point Singularities. Mathematical Surveys and Monographs, vol. 52 (American Mathematical
Society, Providence, RI, 1997)
272. J. Král, Integral Operators in Potential Theory. Lecture Notes in Mathematics, vol. 823
(Springer, Berlin, 1980)
273. M.A. Krasnosel’skiı̆, G.M. Vaı̆nikko, P.P. Zabreı̆ko, Ya.B. Rutitskii, V.Ya. Stetsenko, Approx-
imate Solution of Operator Equations (Wolters-Noordhoff Publishing, Groningen, 1972)
274. A. Krebs, M. Maischak, E.P. Stephan, Adaptive FEM-BEM coupling with a Schur comple-
ment error indicator. Appl. Numer. Math. 60, 798–808 (2010)
275. A. Krebs, E.P. Stephan, A p-version finite element method for nonlinear elliptic variational
inequalities in 2D. Numer. Math. 105, 457–480 (2007)
276. R. Kress, Linear Integral Equations, 3rd edn. Applied Mathematical Sciences, vol. 82
(Springer, New York, 2014)
277. A. Kufner, O. John, S. Fučík, Function Spaces (Noordhoff International Publishing, Leyden;
Academia, Prague, 1977)
278. V.D. Kupradze, T.G. Gegelia, M.O. Basheleĭshvili, T.V. Burchuladze, Three-Dimensional
Problems of the Mathematical Theory of Elasticity and Thermoelasticity (North-Holland,
Amsterdam-New York, 1979)
279. U. Lamp, T. Schleicher, E.P. Stephan, W.L. Wendland, Galerkin collocation for an improved
boundary element method for a plane mixed boundary value problem. Computing 33, 269–
296 (1984)
280. U. Langer, O. Steinbach, Boundary element tearing and interconnecting methods. Computing
71, 205–228 (2003)
281. F. Leydecker, hp-version of the boundary element method for electromagnetic problems :
error analysis, adaptivity, preconditioners, Ph.D. thesis, Leibniz Universität Hannover, 2006
282. F. Leydecker, M. Maischak, E.P. Stephan, M. Teltscher, Adaptive FE-BE coupling for an
electromagnetic problem in R3 —a residual error estimator. Math. Methods Appl. Sci. 33,
2162–2186 (2010)
283. F. Leydecker, M. Maischak, E.P. Stephan, M. Teltscher, A p-hierarchical error estimator for a
fe-be coupling formulation applied to electromagnetic scattering problems in R3 . Appl. Anal.
91, 277–293 (2012)
284. J.-L. Lions, E. Magenes, Non-homogeneous Boundary Value Problems and Applications.
Vol. I. Grundlehren der mathematischen Wissenschaften, vol. 181 (Springer, New York-
Heidelberg, 1972). Translated from the French by P. Kenneth
285. R.C. MacCamy, E.P. Stephan, A Simple Layer Potential Method for Three-Dimensional Eddy
Current Problems. Ordinary and Partial Differential Equations. Lecture Notes in Math.,
vol. 1964 (Springer, Berlin, 1982), pp. 477–484
286. R.C. MacCamy, E.P. Stephan, A boundary element method for an exterior problem for three-
dimensional Max-well’s equations. Appl. Anal. 16, 141–163 (1983)
287. R.C. MacCamy, E.P. Stephan, Solution procedures for three-dimensional eddy current
problems. J. Math. Anal. Appl. 101, 348–379 (1984)
288. R.C. MacCamy, E.P. Stephan, A skin effect approximation for eddy current problems. Arch.
Ration. Mech. Anal. 90, 87–98 (1985)
289. R.C. MacCamy, M. Suri, A time-dependent interface problem for two-dimensional eddy
currents. Quart. Appl. Math. 44, 675–690 (1987)
290. M. Maischak, hp-methoden für randintegralgleichungen bei 3d-problemen, theorie und
implementierung, Ph.D. thesis, Leibniz Universität Hannover, 1996
291. M. Maischak, The analytical computation of the Galerkin elements for the Laplace, Lamé
and Helmholtz equation in 3d-BEM, MAIPROGS documentation, 2000
292. M. Maischak, Fem/bem Methods for Signorini-Type Problems: Error Analysis, Adaptivity,
Preconditioners. Habilitationsschrift (Leibniz Universität Hannover, 2004)
293. M. Maischak, Manual of the Software Package Maiprogs, version 3.7.1 edn., 2012
294. M. Maischak, P. Mund, E.P. Stephan, Adaptive multilevel BEM for acoustic scattering.
Comput. Methods Appl. Mech. Eng. 150, 351–367 (1997)
644 References
295. M. Maischak, E.P. Stephan, The hp-version of the boundary element method in R 3 : the basic
approximation results. Math. Methods Appl. Sci. 20, 461–476 (1997)
296. M. Maischak, E.P. Stephan, A least squares coupling method with finite elements and
boundary elements for transmission problems. Comput. Math. Appl. 48, 995–1016 (2004)
297. M. Maischak, E.P. Stephan, Adaptive hp-versions of BEM for Signorini problems. Appl.
Numer. Math. 54, 425–449 (2005)
298. M. Maischak, E.P. Stephan, A FEM-BEM coupling method for a nonlinear transmission
problem modelling Coulomb friction contact. Comput. Methods Appl. Mech. Eng. 194, 453–
466 (2005)
299. M. Maischak, E.P. Stephan, Adaptive hp-versions of boundary element methods for elastic
contact problems. Comput. Mech. 39, 597–607 (2007)
300. M. Maischak, E.P. Stephan, The hp-Version of the Boundary Element Method for the Lamé
Equation in 3D. Boundary Element Analysis. Lect. Notes Appl. Comput. Mech., vol. 29
(Springer, Berlin, 2007), pp. 97–112
301. V.G. Maz’ya, Boundary Integral Equations. Analysis, IV, Encyclopaedia Math. Sci., vol. 27
302. V.G. Maz’ya, B.A. Plamenevskiı̆, The coefficients in the asymptotic expansion of the
solutions of elliptic boundary value problems near an edge. Dokl. Akad. Nauk SSSR 229,
33–36 (1976)
303. V.G. Maz’ya, B.A. Plamenevskiı̆, The coefficients in the asymptotics of solutions of elliptic
boundary value problems with conical points. Math. Nachr. 76, 29–60 (1977)
304. W. McLean, Strongly Elliptic Systems and Boundary Integral Equations (Cambridge Univer-
sity Press, Cambridge, 2000)
305. W. McLean, S. Prössdorf, W.L. Wendland, A fully-discrete trigonometric collocation method.
J. Integr. Equ. Appl. 5, 103–129 (1993)
306. S. Meddahi, J. Valdés, O. Menéndez, P. Pérez, On the coupling of boundary integral and
mixed finite element methods. J. Comput. Appl. Math. 69, 113–124 (1996)
307. M. Mitrea, M. Wright, Boundary value problems for the Stokes system in arbitrary Lipschitz
domains. Astérisque 344, viii+241 (2012)
308. P. Morin, K.G. Siebert, A. Veeser, A basic convergence result for conforming adaptive finite
elements. Math. Models Methods Appl. Sci. 18, 707–737 (2008)
309. U. Mosco, Convergence of convex sets and of solutions of variational inequalities. Adv. Math.
3, 510–585 (1969)
310. P. Mund, On the Implementation of the h-p Boundary Element Method on Curved Surfaces.
Boundary Elements: Implementation and the Analysis of Advanced Algorithms. Notes on
Numerical Fluid Mechanics, vol. 54 (Vieweg, Braunschweig, 1996), pp. 182–193
311. P. Mund, E.P. Stephan, Adaptive coupling and fast solution of FEM-BEM equations for
parabolic-elliptic interface problems. Math. Methods Appl. Sci. 20, 403–423 (1997)
312. P. Mund, E.P. Stephan, An adaptive two-level method for the coupling of nonlinear FEM-
BEM equations. SIAM J. Numer. Anal. 36, 1001–1021 (1999)
313. P. Mund, E.P. Stephan, An adaptive two-level method for hypersingular integral equations in
R3 . In: Proceedings of the 1999 International Conference on Computational Techniques and
Applications (Canberra), vol. 42, 2000, pp. C1019–C1033
314. P. Mund, E.P. Stephan, J. Weisse, Two-level methods for the single layer potential in R3 .
Computing 60, 243–266 (1998)
315. F.J. Narcowich, J.D. Ward, Scattered data interpolation on spheres: error estimates and locally
supported basis functions. SIAM J. Math. Anal. 33, 1393–1410 (2002)
316. S.A. Nazarov, Asymptotic behavior of the solution of the Dirichlet problem in an angular
domain with a periodically changing boundary Mat. Zametki 49, 86–96 (1991)
317. S.A. Nazarov, B.A. Plamenevskiı̆, Elliptic Problems in Domains with Piecewise Smooth
Boundaries. De Gruyter Expositions in Mathematics, vol. 13 (Walter de Gruyter & Co.,
Berlin, 1994)
318. M. Neamtu, L.L. Schumaker, On the approximation order of splines on spherical triangula-
tions. Adv. Comput. Math. 21, 3–20 (2004)
References 645
319. J. Nečas, J. Jarušek, J. Haslinger, On the solution of the variational inequality to the Signorini
problem with small friction. Boll. Un. Mat. Ital. B (5) 17, 796–811 (1980)
320. J.C. Nédélec, Approximation des équations intégrales en mécanique et en physique (Ecole
Polytechnique, Palaiseau, 1977)
321. J.C. Nédélec, Mixed finite elements in R3 . Numer. Math. 35, 315–341 (1980)
322. J.C. Nédélec, Integral equations with nonintegrable kernels. Integr. Equ. Oper. Theory 5, 562–
572 (1982)
323. J.C. Nédélec, Acoustic and Electromagnetic Equations. Integral Representations for Har-
monic Problems. Applied Mathematical Sciences, vol. 144 (Springer, New York, 2001)
324. J.C. Nédélec, J. Planchard, Une méthode variationnelle d’ éléments finis pour la résolution
numérique d’un problème extérieur dans R 3 . Rev. Franc. Automat. Informat. Rech. Operat.
Sér. Rouge 7, 105–129 (1973)
325. J.C. Nédélec, Computation of eddy currents on a surface in R3 by finite element methods.
SIAM J. Numer. Anal. 15, 580–594 (1978)
326. L. Nesemann, E.P. Stephan, Numerical solution of an adhesion problem with FEM and BEM.
Appl. Numer. Math. 62, 606–619 (2012)
327. J. Nečas, Direct Methods in the Theory of Elliptic Equations. Springer Monographs in
Mathematics (Springer, Heidelberg, 2012). Translated from the 1967 French original by G.
Tronel and A. Kufner, and a contribution by Ch. G. Simader
328. J. Nitsche, Ein Kriterium für die Quasi-Optimalität des Ritzschen Verfahrens. Numer. Math.
11, 346–348 (1968)
329. M. Ochmann, Closed form solutions for the acoustical impulse response over a masslike or
an absorbing plane. J. Acoust. Soc. Am. 129, 3502–3512 (2011)
330. G. Of, O. Steinbach, Is the one-equation coupling of finite and boundary element methods
always stable?. ZAMM Z. Angew. Math. Mech. 93, 476–484 (2013)
331. E. Ostermann, Numerical methods for space-time variational formulations of retarded
potential boundary integral equations, Ph.D. thesis, Leibniz Universität Hannover, 2009
332. N. Ovcharova, Regularization Methods and Finite Element Approximation of Hemivari-
ational Inequalities with Applications to Nonmonotone Contact Problems, Ph.D. thesis,
Universität der Bundeswehr München, 2012
333. N. Ovcharova, On the coupling of regularization techniques and the boundary element method
for a hemivariational inequality modelling a delamination problem. Math. Methods Appl. Sci.
40, 60–77 (2017)
334. N. Ovcharova, L. Banz, Coupling regularization and adaptive hp-BEM for the solution of a
delamination problem. Numer. Math. 137, 303–337 (2017)
335. N. Ovcharova, J. Gwinner, A study of regularization techniques of nondifferentiable opti-
mization in view of application to hemivariational inequalities. J. Optim. Theory Appl. 162,
754–778 (2014)
336. P.D. Panagiotopoulos, J. Haslinger, On the dual reciprocal variational approach to the
Signorini-Fichera problem. Convex and nonconvex generalization. Z. Angew. Math. Mech.
72, 497–506 (1992)
337. L.F. Pavarino, Additive Schwarz methods for the p-version finite element method. Numer.
Math. 66, 493–515 (1994)
338. T.D. Pham, T. Tran, Strongly elliptic pseudodifferential equations on the sphere with radial
basis functions. Numer. Math. 128, 589–614 (2014)
339. T.D. Pham, T. Tran, A. Chernov, Pseudodifferential equations on the sphere with spherical
splines. Math. Models Methods Appl. Sci. 21, 1933–1959 (2011)
340. F.V. Postell, E.P. Stephan, On the h-, p- and h-p versions of the boundary element method—
numerical results. Comput. Methods Appl. Mech. Eng. 83, 69–89 (1990)
341. R.A. Prato Torres, E.P. Stephan, F. Leydecker, A FE/BE coupling for the 3D time-dependent
eddy current problem. Part I: a priori error estimates. Computing 88, 131–154 (2010)
342. R.A. Prato Torres, E.P. Stephan, F. Leydecker, A FE/BE coupling for the 3D time-dependent
eddy current problem. Part II: a posteriori error estimates and adaptive computations.
Computing 88, 155–172 (2010)
646 References
343. S. Prössdorf, Linear Integral Equations. Analysis, IV, Encyclopaedia Math. Sci., vol. 27
344. S. Prössdorf, A. Rathsfeld, Quadrature and collocation methods for singular integral equations
on curves with corners. Z. Anal. Anwendungen 8, 197–220 (1989)
345. S. Prössdorf, B. Silbermann, Numerical Analysis for Integral and Related Operator Equa-
tions. Operator Theory: Advances and Applications, vol. 52 (Birkhäuser Verlag, Basel, 1991)
346. L. Qi, D. Sun, Smoothing functions and smoothing Newton method for complementarity and
variational inequality problems. J. Optim. Theory Appl. 113, 121–147 (2002)
347. J. Radon, Über die Randwertaufgabe beim logarithmischen Potential. Math.–Nat. Kl. Abt. IIa
128, 1123–1167 (1919)
348. E. Rank, A posteriori error estimates and adaptive refinement for some boundary integral
element method. In: Proceedings Int. Conf. on Accuracy Estimates and Adaptive Refinements
in FE Computations (ARFEC, Lisbon, 1984), pp. 55–64
349. R. Rannacher, W.L. Wendland, On the order of pointwise convergence of some boundary
element methods. I. Operators of negative and zero order. RAIRO Modél. Math. Anal. Numér.
19, 65–87 (1985)
350. R. Rannacher, W.L. Wendland, Pointwise convergence of some boundary element methods.
II. RAIRO Modél. Math. Anal. Numér. 22, 343–362 (1988)
351. S. Rempel, G. Schmidt, Eigenvalues for spherical domains with corners via boundary integral
equations. Integr. Equ. Oper. Theory 14, 229–250 (1991)
352. J.E. Roberts, J.-M. Thomas, Mixed and Hybrid Methods. Handbook of Numerical Analysis,
Vol. II. Handb. Numer. Anal., II (North-Holland, Amsterdam, 1991), pp. 523–639
353. R.T. Rockafellar, R.J.-B. Wets, Variational Analysis. Grundlehren der Mathematischen
Wissenschaften, vol. 317 (Springer, Berlin, 1998)
354. E.B. Saff, A.B.J. Kuijlaars, Distributing many points on a sphere. Math. Intell. 19, 5–11
(1997)
355. J. Saranen, Local error estimates for some Petrov-Galerkin methods applied to strongly
elliptic equations on curves. Math. Comput. 48, 485–502 (1987)
356. J. Saranen, G. Vainikko, Periodic Integral and Pseudodifferential Equations with Numerical
Approximation. Springer Monographs in Mathematics (Springer, Berlin, 2002)
357. J. Saranen, W.L. Wendland, On the asymptotic convergence of collocation methods with
spline functions of even degree. Math. Comput. 45, 91–108 (1985)
358. S. Sauter, A. Veit, Retarded boundary integral equations on the sphere: exact and numerical
solution. IMA J. Numer. Anal. 34, 675–699 (2014)
359. S.A. Sauter, Ch. Schwab, Boundary Element Methods. Springer Series in Computational
Mathematics, vol. 39 (Springer, Berlin, 2011)
360. F.-J. Sayas, The validity of Johnson-Nédélec’s BEM-FEM coupling on polygonal interfaces.
SIAM J. Numer. Anal. 47, 3451–3463 (2009)
361. F.-J. Sayas, Retarded Potentials and Time Domain Boundary Integral Equations. A Road Map.
Springer Series in Computational Mathematics, vol. 50 (Springer, Cham, 2016)
362. A.H. Schatz, V. Thomée, W.L. Wendland, Mathematical Theory of Finite and Boundary
Element Methods. DMV Seminar, vol. 15 (Birkhäuser Verlag, Basel, 1990)
363. A.H. Schatz, L.B. Wahlbin, Interior maximum norm estimates for finite element methods.
Math. Comput. 31, 414–442 (1977)
364. M. Schechter, Principles of Functional Analysis (Academic Press, New York-London,
1971,1973)
365. G. Schmidt, On spline collocation methods for boundary integral equations in the plane. Math.
Methods Appl. Sci. 7, 74–89 (1985)
366. G. Schmidt, Boundary element discretization of Poincaré-Steklov operators. Numer. Math.
69, 83–101 (1994)
367. H. Schmitz, K. Volk, W.L. Wendland, Three-dimensional singularities of elastic fields near
vertices. Numer. Methods Partial Differ. Equ. 9, 323–337 (1993)
368. D. Schötzau, Ch. Schwab, Exponential convergence for hp-version and spectral finite element
methods for elliptic problems in polyhedra. Math. Models Methods Appl. Sci. 25, 1617–1661
(2015)
References 647
369. H. Schulz, O. Steinbach, A new a posteriori error estimator in adaptive direct boundary
element methods: the Dirichlet problem. Calcolo 37, 79–96 (2000)
370. H. Schulz, O. Steinbach, A New a Posteriori Error Estimator in Adaptive Direct Boundary
Element Methods. The Neumann Problem. Multifield Problems (Springer, Berlin, 2000),
pp. 201–208
371. H. Schulz, O. Steinbach, W.L. Wendland, On Adaptivity in Boundary Element Methods.
Aspects of the Boundary Element Method (Kluwer, 2001). In: IUTAM/IACM/IABEM
Symposium on Advanced Mathematical and Computational Mechanics, pp. 315–325
372. Ch. Schwab, Variable order composite quadrature of singular and nearly singular integrals.
Computing 53, 173–194 (1994)
373. Ch. Schwab, p- and hp-Finite Element Methods.Theory and Applications in Solid and
Fluid Mechanics. Numerical Mathematics and Scientific Computation (The Clarendon Press,
Oxford University Press, New York, 1998)
374. Ch. Schwab, M. Suri, The optimal p-version approximation of singularities on polyhedra in
the boundary element method. SIAM J. Numer. Anal. 33, 729–759 (1996)
375. Ch. Schwab, W.L. Wendland, On the extraction technique in boundary integral equations.
Math. Comput. 68, 91–122 (1999)
376. R. Seeley, Topics in Pseudo-Differential Operators. Pseudo-Diff. Operators (C.I.M.E., Stresa,
1968; Edizioni Cremonese, Rome, 1969), pp. 167–305
377. K.G. Siebert, A convergence proof for adaptive finite elements without lower bound. IMA J.
Numer. Anal. 31, 947–970 (2011)
378. I.H. Sloan, A quadrature-based approach to improving the collocation method. Numer. Math.
54, 41–56 (1988)
379. I.H. Sloan, Error Analysis of Boundary Integral Methods. Acta Numerica, 1992 (Cambridge
Univ. Press, Cambridge, 1992), pp. 287–339
380. I.H. Sloan, A. Spence, The Galerkin method for integral equations of the first kind with
logarithmic kernel: theory. IMA J. Numer. Anal. 8, 105–122 (1988)
381. I.H. Sloan, E.P. Stephan, Collocation with Chebyshev polynomials for Symm’s integral
equation on an interval. J. Aust. Math. Soc. Ser. B 34, 199–211 (1992)
382. I.H. Sloan, T. Tran, The tolerant qualocation method for variable-coefficient elliptic equations
on curves. J. Integr. Equ. Appl. 13, 73–98 (2001)
383. I.H. Sloan, W.L. Wendland, Commutator properties for periodic splines. J. Approx. Theory
97, 254–281 (1999)
384. I.H. Sloan, W.L. Wendland, Spline qualocation methods for variable-coefficient elliptic
equations on curves. Numer. Math. 83, 497–533 (1999)
385. I.H. Sloan, W.L. Wendland, Qualocation methods for elliptic boundary integral equations.
Numer. Math. 79, 451–483 (1998)
386. W. Spann, On the boundary element method for the Signorini problem of the Laplacian.
Numer. Math. 65, 337–356 (1993)
387. W. Spann, Error estimates for the boundary element approximation of a contact problem in
elasticity. Math. Methods Appl. Sci. 20, 205–217 (1997)
388. G. Stampacchia, Variational Inequalities. Theory and Applications of Monotone Operators,
ed. by A. Ghizetti (Edizione Odersi, Gubbio, 1969), pp. 101–192
389. O. Steinbach, Adaptive finite element-boundary element solution of boundary value prob-
lems. J. Comput. Appl. Math. 106, 307–316 (1999)
390. O. Steinbach, Adaptive boundary element methods based on computational schemes for
Sobolev norms. SIAM J. Sci. Comput. 22, 604–616 (2000)
391. O. Steinbach, Numerische Näherungsverfahren für elliptische Randwertprobleme (Springer,
2003)
392. O. Steinbach, A note on the stable one-equation coupling of finite and boundary elements.
SIAM J. Numer. Anal. 49, 1521–1531 (2011)
393. O. Steinbach, Boundary element methods for variational inequalities. Numer. Math. 126, 173–
197 (2014)
648 References
394. O. Steinbach, W.L. Wendland, On C. Neumann’s method for second-order elliptic systems in
domains with non-smooth boundaries. J. Math. Anal. Appl. 262, 733–748 (2001)
395. E.P. Stephan, Solution Procedures for Interface Problems in Acoustics and Electromagnetics.
Theoretical Acoustics and Numerical Techniques. CISM Courses and Lectures, vol. 277
(Springer, Vienna, 1983), pp. 291–348
396. E.P. Stephan, A boundary integral equation method for three-dimensional crack problems in
elasticity. Math. Methods Appl. Sci. 8, 609–623 (1986)
397. E.P. Stephan, Boundary integral equations for mixed boundary value problems in R3 . Math.
Nachr. 134, 21–53 (1987)
398. E.P. Stephan, Boundary integral equations for screen problems in R3 . Integr. Equ. Oper.
Theory 10, 236–257 (1987)
399. E.P. Stephan, Coupling of finite elements and boundary elements for some nonlinear interface
problems. Comput. Methods Appl. Mech. Eng. 101, 61–72 (1992)
400. E.P. Stephan, The h-p boundary element method for solving 2- and 3-dimensional problems.
Comput. Methods Appl. Mech. Eng. 133, 183–208 (1996)
401. E.P. Stephan, M. Maischak, A posteriori error estimates for fem-bem couplings of three-
dimensional electromagnetic problems. Comput. Methods Appl. Mech. Eng. 194, 441–452
(2005)
402. E.P. Stephan, M. Maischak, E. Ostermann, Transient boundary element method and numerical
evaluation of retarded potentials. In: Computational Science - ICCS 2008, 8th International
Conference, Kraków, Poland, June 23–25, 2008, Proceedings, Part II, 2008, pp. 321–330
403. E.P. Stephan, M. Maischak, E. Ostermann, TD-BEM for sound radiation in three dimensions
and the numerical evaluation of retarded potentials. In: International Conference on Acous-
tics, NAG/DAGA, 2009
404. E.P. Stephan, M. Suri, On the convergence of the p-version of the boundary element Galerkin
method. Math. Comput. 52, 31–48 (1989)
405. E.P. Stephan, M. Suri, The h-p version of the boundary element method on polygonal domains
with quasiuniform meshes. RAIRO Modél. Math. Anal. Numér. 25, 783–807 (1991)
406. E.P. Stephan, M.T. Teltscher, Collocation with trigonometric polynomials for integral equa-
tions to the mixed boundary value problem. Numerische Mathematik (2018, to appear)
407. E.P. Stephan, T. Tran, Localization and post processing for the Galerkin boundary element
method applied to three-dimensional screen problems. J. Integr. Equ. Appl. 8, 457–481 (1996)
408. E.P. Stephan, W.L. Wendland, Remarks to Galerkin and Least Squares Methods with Finite
Elements for General Elliptic Problems. Ordinary and Partial Differential Equations (Proc.
Fourth Conf., Univ. Dundee, Dundee, 1976). Lecture Notes in Math., vol. 564 (Springer,
Berlin, 1976), pp. 461–471
409. E.P. Stephan, W.L. Wendland, Mathematische Grundlagen der finiten Element-Methoden.
Methoden und Verfahren der Mathematischen Physik [Methods and Procedures in Mathe-
matical Physics], vol. 23 (Verlag Peter D. Lang, Frankfurt, 1982)
410. E.P. Stephan, W.L. Wendland, An augmented Galerkin procedure for the boundary integral
method applied to two-dimensional screen and crack problems. Appl. Anal. 18, 183–219
(1984)
411. E.P. Stephan, W.L. Wendland, An augmented Galerkin procedure for the boundary integral
method applied to mixed boundary value problems. Appl. Numer. Math. 1, 121–143 (1985)
412. G. Strang, G.J. Fix, An Analysis of the Finite Element Method (Prentice-Hall, Englewood
Cliffs, NJ, 1973)
413. J.A. Stratton, Electromagnetic Rheory (Wiley, 2007)
414. F. Stummel, Perturbation theory for Sobolev spaces. Proc. Roy. Soc. Edinburgh Sect. A 73,
5–49 (1975)
415. M.E. Taylor, Pseudodifferential Operators. Princeton Mathematical Series, vol. 34 (Princeton
University Press, Princeton, NJ, 1981)
416. T. Tran, The K-operator and the Galerkin method for strongly elliptic equations on smooth
curves: local estimates. Math. Comput. 64, 501–513 (1995)
References 649
417. T. Tran, Local error estimates for the Galerkin method applied to strongly elliptic integral
equations on open curves. SIAM J. Numer. Anal. 33, 1484–1493 (1996)
418. T. Tran, Q.T. Le Gia, I.H. Sloan, E.P. Stephan, Boundary integral equations on the sphere with
radial basis functions: error analysis. Appl. Numer. Math. 59, 2857–2871 (2009)
419. T. Tran, I.H. Sloan, Tolerant qualocation—a qualocation method for boundary integral
equations with reduced regularity requirement. J. Integr. Equ. Appl. 10, 85–115 (1998)
420. T. Tran, E.P. Stephan, Additive Schwarz methods for the h-version boundary element method.
Appl. Anal. 60, 63–84 (1996)
421. G. Verchota, Layer potentials and regularity for the Dirichlet problem for Laplace’s equation
in Lipschitz domains. J. Funct. Anal. 59, 572–611 (1984)
422. T. von Petersdorff, Boundary integral equations for mixed Dirichlet, Neumann and transmis-
sion problems. Math. Methods Appl. Sci. 11, 185–213 (1989)
423. T. von Petersdorff, Randwertprobleme der Elastizitätstheorie für Polyeder - Singularitäten
und Approximation mit Randelementmethoden, Ph.D. thesis, TU Darmstadt, 1989
424. T. von Petersdorff, E.P. Stephan, A Direct Boundary Element Method for Interface Crack
Problems. Computational Mechanics ‘88 (Springer, 1988). In: Proceedings Atlanta, Vol 1,
pp. 11.vii.1–11.vii.5
425. T. von Petersdorff, E.P. Stephan, Decompositions in edge and corner singularities for the
solution of the Dirichlet problem of the Laplacian in a polyhedron. Math. Nachr. 149, 71–103
(1990)
426. T. von Petersdorff, E.P. Stephan, Regularity of mixed boundary value problems in R3 and
boundary element methods on graded meshes. Math. Methods Appl. Sci. 12, 229–249 (1990)
427. H. Wendland, Scattered Data Approximation. Cambridge Monographs on Applied and
Computational Mathematics, vol. 17 (Cambridge University Press, Cambridge, 2005)
428. W.L. Wendland, Asymptotic Convergence of Boundary Element Methods; Integral Equation
Methods for Mixed bvp’s. Preprint 611 (TU Darmstadt, Fachbereich Mathematik, 1981)
429. W.L. Wendland, Boundary Element Methods and Their Asymptotic Convergence. Theoretical
Acoustics and Numerical Techniques. CISM Courses and Lectures, vol. 277 (Springer,
Vienna, 1983), pp. 135–216
430. W.L. Wendland, Randelementmethoden – eine Einführung (Vorlesungsmanuskript, TU Darm-
stadt, 1985)
431. W.L. Wendland, Qualocation, the new variety of boundary element methods. Wiss. Z. Tech.
Univ. Karl-Marx-Stadt 31, 276–284 (1989)
432. W.L. Wendland, E.P. Stephan, A hypersingular boundary integral method for two-dimensional
screen and crack problems. Arch. Ration. Mech. Anal. 112, 363–390 (1990)
433. W.L. Wendland, E.P. Stephan, G.C. Hsiao, On the integral equation method for the plane
mixed boundary value problem of the Laplacian. Math. Methods Appl. Sci. 1, 265–321 (1979)
434. W.L. Wendland, D.H. Yu, Adaptive boundary element methods for strongly elliptic integral
equations. Numer. Math. 53, 539–558 (1988)
435. J. Wloka, Partial Differential Equations (Cambridge University Press, Cambridge, 1987).
Translated from the German by C. B. Thomas and M. J. Thomas
436. H. Yserentant, On the multilevel splitting of finite element spaces. Numer. Math. 49, 379–412
(1986)
437. E. Zeidler, Nonlinear Functional Analysis and Its Applications. IV (Springer, New York,
1988)
438. E. Zeidler, Nonlinear Functional Analysis and Its Applications. II/B (Springer, New York,
1990)
439. O.C. Zienkiewicz, D.W. Kelly, P. Bettess, Marriage à la Mode—The Best of Both Worlds
(Finite Elements and Boundary Integrals). Energy Methods in Finite Element Analysis
(Wiley, Chichester, 1979), pp. 81–107
Index
Aubin-Nitsche duality estimate, 153, 186 Fredholm operator, 566

augmented finite element space, 183
Gårding inequality, 101, 118, 194, 322

biorthogonal basis function, 422 Galerkin method, 4, 116
boundary hemivariational inequality, 110 Galerkin-Petrov method, 185
boundary variational inequality, 99, 105, 106 Gauss-Lobatto quadrature, 396
Gauss-Lobatto-Lagrange basis function, 422
geometric mesh, 230, 271, 284
Cauchy data, 65 graded mesh, 228, 234
collectively compact, 132, 133, 565
collocation method, 127, 185, 189
collocation methods, 125
corner singularity, 231 hemivariational inequality, 109
Coulomb friction, 104 hypersingular integral equation, 339, 340
countably normed space, 273, 285, 324 hypersingular operator, 358
crack problem, 267
interface problem, 71, 79

decaying condition, 57
delamination, 106
discrete Poincaré–Steklov operator, 398
Lagrange multiplier, 432
discrete Poincaré-Steklov operator, 470
Lions - Stampacchia theorem, 603
discrete Poincaré-Steklov operator, 436
double layer potential in time domain, 539
Mellin convolution operator, 176, 178

eddy current problem, 89, 522 Mellin symbol, 297, 319
edge singularity, 231 Mellin transform, 296
elliptic, 74 mesh grading transformation, 173
exponential convergence, 271, 285 mixed boundary value problem, 63
MOT scheme, 552
Fredholm integral equation of the second kind,

77, 153 Navier–Lamé equation, 287

652 Index
penalty formulation, 432 spline, 128, 140

Poincaré–Steklov operator, 106, 325, 396 Stratton-Chu representation formulas, 90
principal symbol, 67, 74, 75 strongly elliptic, 118
projection method, 131 strongly elliptic system of pseudodifferential
pseudodifferential operator, 66, 337, 577 equations, 69
pseudodifferential operator of order r, 74 superapproximation property, 153
pseudomonotone, 111, 608 symbol, 74
Symm’s integral equation, 24, 153, 338, 340,
343, 349
qualocation method, 201, 203
quasiuniform mesh, 151
tempered distribution, 129
time harmonic Maxwell’s equations, 90
Radial basis, 205
transmission problem, 79, 279
representation formula, 10
transmission problem in steady state
residual error estimator, 336
elastodynamics, 84
Tresca friction, 103
trigonometric polynomial, 44, 119
saddle point, 432, 455 two-level additive Schwarz operator, 374, 400
screen problem, 81, 265 two-level error estimator, 360, 388
Second Strang Lemma, 157
Signorini boundary condition, 96
single layer potential in time domain, 539
Sobolev space, 64 uniform strong ellipticity, 578
Somigliana representation formula, 85, 109
Sommerfeld radiation condition, 71
spherical splines, 219 wave equation, 538

Advanced Boundary Element Methods: Joachim Gwinner Ernst Peter Stephan

Uploaded by

Copyright:

Available Formats

Advanced Boundary Element Methods: Joachim Gwinner Ernst Peter Stephan

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Boundary Element Methods: Joachim Gwinner Ernst Peter Stephan

Uploaded by

Copyright:

Available Formats

Springer Series in Computational Mathematics 52

Joachim Gwinner · Ernst Peter Stephan

Advanced Boundary Element

ISSN 0179-3632 ISSN 2198-3712 (electronic)

Library of Congress Control Number: 2018946112

Mathematics Subject Classification (2010): 35-XX, 45-XX, 49-XX, 65-XX

© Springer International Publishing AG, part of Springer Nature 2018

Printed on acid-free paper

Chapter 1 introduces to the theory of approximation methods for the solution of

München, Germany Joachim Gwinner

first kind BIE’s for general projection methods

standard BEM advanced BEM

FE/BE coupling TD-BEM

4.3 Screen Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 81

7 Advanced BEM for BVPs in Polygonal/Polyhedral Domains:

10.3 The Weakly Singular Integral Equation in 3D . .. . . . . . . . . . . . . . . . . . . . 349

12.3.3 Adaptive FE/BE Coupling with a Schur

A Linear Operator Theory .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 563

B Pseudodifferential Operators . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 569

C Convex and Nonsmooth Analysis, Variational Inequalities .. . . . . . . . . . . 581

C.3 Existence and Approximation Results for Variational

D Some Implementations for BEM . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 615

1.1 The Basic Approximation Problems

Problem 1.1 Let A : X → Y be a continuous linear operator between two

© Springer International Publishing AG, part of Springer Nature 2018 1

above equation in certain subspaces XN ⊂ X, YN ⊂ Y :

where AN is an approximation to A and to hope that uN ∈ XN is a “reasonable”

where Γ is the boundary of a bounded domain in R2 , such that (1.1) corresponds to

Note for Symm’s integral equation the quadratical functional becomes

Here Γc is a part of Γ and S : H 1/2(Γ ) → H −1/2 (Γ ) is the Dirichlet-to-Neumann

Here F (x), z denotes the directional derivative of F at x in direction z.

where C := {v ∈ H 1/2 (Γ )|v|Γc ≤ g}. Now we may choose XN as continuous

1.2 Convergence of Projection Methods

t, AuN = t, f ∀ t ∈ TN . (1.5)

Theorem 1.1 (Galerkin Method) Let X = Y , Y = X and TN = XN for the

3. First of all, the triangle inequality yields for all v ∈ XN

4. By 3., lim inf u − vX = 0 ⇒ lim u − uN X = 0. 

∃ α > 0(independent of N) ∀ v ∈ XN ∃ t ∈ TN \{0} :

∃ QN : XN → TN , ∃ M > 0 ∃ α > 0(independent of N) :

By (1.7) ∃ t ∈ TN \{0} : 0 = t, AuN ≥ αuN X tY . Hence uN = 0. Thus

3. The Galerkin projector GN is well-defined by 1. Furthermore, 2. yields |||GN ||| ≤

u − uN X ≤ u − vX + GN (u − v)X

Dwh , v = Dw, v for all v ∈ Sh (1.10)

define an operator GhD : w ∈ X → wh ∈ Sh with the property

GhD w − w → 0 as h → 0 for all w ∈ X. (1.11)

Then for small h > 0 the equations

Auh , v = Au, v for all v ∈ Sh (1.12)

define an operator GhA : u ∈ X → uh ∈ Sh such that

In this chapter we collect well-known concepts and results of classical potential

2.1 Representation Formulas

© Springer International Publishing AG, part of Springer Nature 2018 9

Proof Consider the following Fig. 2.1.

Fig. 2.1 Geometrical setting

Thus, the assertion follows with ln |x − y| Δu(y) dy = lim J . 

Remark 2.1 Setting G(x, y) = 2π 1

supp (f )) the function

satisfies the Poisson equation Δu = f in R2 .

Then we have for u ∈ C02 (R2 )

u = GΔu = ΔGu. (2.3)

Here F (x), z denotes the directional derivative of F at x in direction z.

t, AuN = t, f ∀ t ∈ TN . (1.5)

4. By 3., lim inf u − vX = 0 ⇒ lim u − uN X = 0.

By (1.7) ∃ t ∈ TN \{0} : 0 = t, AuN ≥ αuN X tY . Hence uN = 0. Thus

u − uN X ≤ u − vX + GN (u − v)X

Dwh , v = Dw, v for all v ∈ Sh (1.10)

define an operator GhD : w ∈ X → wh ∈ Sh with the property

GhD w − w → 0 as h → 0 for all w ∈ X. (1.11)

Auh , v = Au, v for all v ∈ Sh (1.12)

define an operator GhA : u ∈ X → uh ∈ Sh such that

Thus, the assertion follows with ln |x − y| Δu(y) dy = lim J .

as seen as follows: Introduce polar coordinatesz − x = ρ(cos θ, sin θ )T ,

Once more we apply the divergence theorem to − log(|x − z|) (z = x, x fixed on

δ0 , ϕ = ϕ(0), ϕ ∈ D(Ω) .

∂j t, ϕ := − t, ∂j ϕ , ∀ ϕ ∈ C0∞ (R2 ).

Gt, ϕ := t, χG∗ ϕ

Gδ, ϕ = δ, χG∗ ϕ = (χG∗ ϕ)(0) = χ(0) · (G∗ ϕ)(0)

ΔGψ, ϕ = Gψ, Δϕ = ψ, G∗ Δϕ = ψ, GΔϕ = ψ, ϕ ,

Δu, ϕ = u, Δϕ = uΔϕ dx = u− Δϕ dx + u+ Δϕ dx

Δu, ϕ = f, ϕ + [∂n u]ϕ ds − [u]∂n ϕ ds

using γ1∗ ψ, ϕ = ψ∂n ϕ ds and γ0∗ ψ, ϕ = ψϕ ds . Thus, we finally obtain

2ψ = [γ0 (Dψ)], 0 = [γ1 (Dψ)] .

2ϕ, V ϕ = γ1 u− , γ0 u− − γ1 u+ , γ0 u+

ϕ2V := ϕ0 , V ϕ0 + m2 (ϕ) .

hence by Lemma 2.4 ψ = 0, since V ψ, ψ = 0

t, ϕ = (ψ, ϕ)V = ψ0 , V ϕ0 + m(ϕ)m(ψ)