PONTIFICIA UNIVERSIDAD CATOLICA DE CHILE

SCHOOL OF ENGINEERING

AN ACCURATE
NUMERICAL-ANALYTICAL METHOD FOR
COMPUTING STRESSES IN ROCK MASS
AROUND MINING EXCAVATIONS

VALERIA BOCCARDO SALVO

Thesis submitted to the Office of Graduate Studies

in partial fulfillment of the requirements for the Degree of
Doctor in Engineering Sciences

Advisor:
MARIO DURAN

Santiago de Chile, Dic 2017

c MMXVII, VALERIA B OCCARDO S ALVO

PONTIFICIA UNIVERSIDAD CATOLICA DE CHILE
SCHOOL OF ENGINEERING

AN ACCURATE
NUMERICAL-ANALYTICAL METHOD FOR
COMPUTING STRESSES IN ROCK MASS
AROUND MINING EXCAVATIONS

VALERIA BOCCARDO SALVO

Members of the Committee:

MARIO DURAN
EDUARDO GODOY
GONZALO YÁÑEZ
RAFAEL BENGURIA
JEAN-CLAUDE NÉDÉLEC
JORGE VÁSQUEZ

Thesis submitted to the Office of Graduate Studies

in partial fulfillment of the requirements for the Degree of
Doctor in Engineering Sciences

Santiago de Chile, Dic 2017

c MMXVII, VALERIA B OCCARDO S ALVO

 A mi familia, especialmente a
Germán y Antonio, gracias por
darle sentido a mis dı́as.
 ACKNOWLEDGEMENTS

Quisiera agradecer al proyecto MECE Educación superior PUC0710 por la ayuda

recibida para realizar mis estudios.

A mi supervisor Mario por su tiempo y dedicación, y al equipo de jovenes investi-

gadores que lidera, Ricardo Hein, Carlos Pérez, Ignacio Vargas y Juan La Rivera, pues
sin ellos nada de esto habrı́a sido posible. Especialmente mi reconocimiento a Eduardo
Godoy por su constante apoyo en el desarrollo de esta tesis. Al equipo de INGMAT por
sus consejos y ayuda.

A Rafael Benguria, quien desde que era una estudiante de pregrado me ha apoyado,
aconsejado e inspirado. A los profesores Jean-Claude Nédélec, Gonzalo Yáñez y Jorge
Vásquez les agradezco por participar en mi comisión examinadora y revisar esta tesis.

A Aldo Valcarce, Sanzia, Rafael González, Mabel Vega, Mauricio Ipinza, Daniela
Cerón, Sebastián Sarmiento y Cristian Gutierrez les agradezco por su paciencia y compañia
en estos años.

A mi familia, mi madre Patricia, mis hermanos Andrea, Tamara y Mauro, y mis sobri-
nos Franco, Stefano y Rafaella, por su amor y apoyo.

Finalmente a Germán, fuimos muy afortunados de encontrarnos, llegaste a cambiar mi

vida, llenaste de luz mis ojos y de risas mis dı́as. Gracias por ser mi compañero, y el mejor
padre que podrı́a existir para nuestro hijo Antonio.

iv
 TABLE OF CONTENTS

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

RESUMEN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii

1. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2. GENERAL MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1. Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1. Papkovich-Neuber’s decomposition . . . . . . . . . . . . . . . . . . 13

3. A SIMPLE SEMI-INFINITE GEOMETRY . . . . . . . . . . . . . . . . . . . 20

3.1. Series’s solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2. First set of boundary conditions: free surface boundary condition on the
surface of the plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3. Second set of boundary conditions: boundary conditions on the surface of the
semiesphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1. Traction-free boundary . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.2. Loaded boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.3. Zero displacement . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.4. Prescribed displacements . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4. Accelerating the convergence of series . . . . . . . . . . . . . . . . . . . 40
3.4.1. Asymptotic behaviour of Legendre’s polynomials . . . . . . . . . . . 42
3.4.2. Traction-free boundary and loaded boundary . . . . . . . . . . . . . . 45
3.4.3. Zero displacement and prescribed displacements . . . . . . . . . . . . 49
3.5. Some clues about the programming . . . . . . . . . . . . . . . . . . . . . 50

v
 3.6. Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.1. Traction-free boundary . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6.2. Loaded boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6.3. Zero displacement . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6.4. Prescribed displacements . . . . . . . . . . . . . . . . . . . . . . . . 55

4. AN EFFICIENT SEMI-ANALYTICAL METHOD TO COMPUTE DISPLACEMENTS 61

4.1. Series solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1.1. Traction-free boundary on Γ∞ . . . . . . . . . . . . . . . . . . . . . 64
4.1.2. Axisymmetric boundary conditions . . . . . . . . . . . . . . . . . . 68
4.2. Numerical enforcement of boundary conditions on the hemispherical pit . . 69
4.2.1. Truncation of the series . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.2. Quadratic functional and its matrix form . . . . . . . . . . . . . . . . 70
4.2.3. Linear system and method of inversion . . . . . . . . . . . . . . . . . 74
4.3. Numerical results and validation . . . . . . . . . . . . . . . . . . . . . . 75
4.3.1. Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.2. Validation of the procedure . . . . . . . . . . . . . . . . . . . . . . . 79

5. A DIRICHLET-TO-NEUMANN FINITE ELEMENT METHOD FOR AXISYMMETRIC

ELASTOSTATICS IN A SEMI-INFINITE DOMAIN . . . . . . . . . . . . . . 83
5.1. Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.1. Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.2. Axisymmetric elastostatic model . . . . . . . . . . . . . . . . . . . . 83
5.2. FEM formulation in the computational domain . . . . . . . . . . . . . . . 84
5.2.1. Equivalent bounded boundary-value problem . . . . . . . . . . . . . 84
5.2.2. Weak formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2.3. FEM discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3. Approximation of the exact artificial boundary conditions . . . . . . . . . 89
5.3.1. Definition of the DtN map . . . . . . . . . . . . . . . . . . . . . . . 89

vi
 5.3.2. Numerical enforcement of exact boundary conditions on the artificial
boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3.3. Numerical approximation of integral terms involving the DtN map in the
FEM formulation . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4. Numerical experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4.1. Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4.2. Implementation aspects . . . . . . . . . . . . . . . . . . . . . . . . 101
5.4.3. Results and accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 102

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

6. Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.1. Some properties of Legendre polynomials . . . . . . . . . . . . . . . . . 113
6.1.1. Expansion of the odd terms in pair terms from the Legendre polynomials 114

vii
 LIST OF FIGURES

2.1 Geometrical model of the elastoestatic problem. . . . . . . . . . . . . . . . . 10

2.2 Stress components in spherical coordinates. . . . . . . . . . . . . . . . . . . 11

3.1 Boundaries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.2 Displacement relative error between our method and the results yielded by
COMSOL in a square of size 13000 m. . . . . . . . . . . . . . . . . . . . . 53

3.3 Displacement for traction-free boundary condition with ν = 0.3, λ = 40.5 GPa,
µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.4 Stresses for traction-free boundary condition with ν = 0.3, λ = 40.5 GPa,
µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.4 Stresses for traction-free boundary condition with ν = 0.3, λ = 40.5 GPa,
µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.5 Displacement for zero displacement boundary condition with ν = 0.3, λ = 40.5
GPa, µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . 58

3.6 Stresses for zero displacement boundary condition with ν = 0.3, λ = 40.5 GPa,
µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.6 Stresses for zero displacement boundary condition with ν = 0.3, λ = 40.5 GPa,
µ = 27 GPa, % = 2725 kg/m3 . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.1 Structure of matrices Q(AA) ,Q(AB) and Q(BB) . . . . . . . . . . . . . . . . . 74

4.2 Relative error of the solution between two successive iterations in N . . . . . . 77

4.3 Plots of displacement and stress components obtained in Ω. . . . . . . . . . . 78

4.4 Schematic representation of the domain Ω truncated by a square box of length L. 79

4.5 Comparison of displacement components on Γh . . . . . . . . . . . . . . . . . 81

4.6 Comparison of stress components on Γh . . . . . . . . . . . . . . . . . . . . . 81

viii
4.7 Relative errors associated with the displacement vector and the stress tensor on
Γh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5.1 Axisymmetric semi-infinite domain. . . . . . . . . . . . . . . . . . . . . . . 84

5.2 Axisymmetric computational domain. . . . . . . . . . . . . . . . . . . . . . 85

5.3 Axisymmetric semi-infinite residual domain. . . . . . . . . . . . . . . . . . 90

5.4 Two of the considered structured triangular meshes. (a) I = 5. (b) I = 10. . . 103
5.5 Computed displacement components. (a) uhρ (ρ, z). (b) uhz (ρ, z). . . . . . . . . 104

5.6 Computed stress components. (a) σρh (ρ, z). (b) σθh (ρ, z). (c) σzh (ρ, z). (d)
h
σρz (ρ, z). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.7 Log-log plot of Eu in function of h. . . . . . . . . . . . . . . . . . . . . . . 107

ix
 LIST OF TABLES

2.1 Relationship between the different constants that characterise a material. . . . 14

3.1 Comparison of the absolute error between the semi-analytical solution and the
model implemented in COMSOL in a square of side 10000 m. . . . . . . . . 55
3.2 Comparison of the absolute error between the semi-analytical solution and the
model implemented in COMSOL in a square of side 10000 m. . . . . . . . . 55

4.1 Numerical values of the physical parameters. . . . . . . . . . . . . . . . . . 76

5.1 Parameters of the structured triangular meshes considered. . . . . . . . . . . 103

5.2 Some components of the solution evaluated at points (0, R) and (−R, 0). . . . 106

x
 ABSTRACT

The Earth’s crust is subjected to a state of stress, as a result of gravitational, tectonic,

and other forces. When removing material from rock masses, stresses are redistributed,
which is especially important in mining, since high stress concentration areas could arise
in mining excavations, resulting in undesired events such as rock-falls or slides. Therefore,
there is a need for determining, in a reliable way, the state of stress generated in rock masses
by mining activities, which constitutes a complex problem. The aim of this work is thus
to propose an accurate mathematical methodology to compute such stresses and show its
performance in a simplified case. The main strength of the proposed methodology is the
effective and efficient numerical treatment of the unbounded medium that surrounds the
excavation, when compared with the customary approach used by standard commercial
software for the same purpose, which results in a considerable reduction of the required
computational resources. The methodology is developed in two stages. In the first stage,
we consider a simplified problem where the excavation is assumed to be a hemispherical
pit, placed on the surface of an elastic semi-infinite domain subjected to gravity. This
problem is solved by using two semi-analytical methods, where the second method is an
improvement of the first one. In the second stage, we deal with an excavation of arbitrary
axisymmetric geometry by means of a coupling technique, which employs the second semi-
analytical method to compute an approximation of the Dirichlet-to-Neumann map (DtN) on
an artificial semi-spherical boundary. The DtN map provides the exact boundary conditions
that complete the mathematical problem posed in the computational domain lying between
the excavation and the artificial boundary, which is then solved by the finite element method
(FEM). Numerical results are provided to demonstrate the effectiveness and accuracy of the
proposed methodology.

Keywords: Numerical-analytical method, mining excavation, stress, elasticity, semi-

infinite domain, finite elements, Dirichlet-to-Neumann map

xi
 RESUMEN

La corteza terrestre se encuentra sometida a un estado de esfuerzos o tensiones debido a

fuerzas de origen gravitacional y tectónico, entre otros. Al extraer material desde un macizo
rocoso, los esfuerzos se redistribuyen, lo cual es especialmente importante en minerı́a, de-
bido a que pueden aparecer áreas de alta concentración de esfuerzos en excavaciones min-
eras, produciendo eventos no deseados como caı́das de roca o deslizamientos. Por lo tanto,
hay una necesidad por determinar, de manera fehaciente, el estado de esfuerzos inducido en
un macizo rocoso por la actividad minera, lo cual constituye un problema complejo. El ob-
jetivo de este trabajo es entonces proponer una metodologı́a matemática certera para calcu-
lar estos esfuerzos y mostrar su desempeño en un caso simplificado. La principal fortaleza
de esta metodologı́a es el tratamiento numérico eficiente y eficaz del medio no acotado
que rodea la excavación, comparado con el enfoque habitual usado por software comercial
estándar para el mismo propósito, lo que conlleva una considerable reducción del recurso
computacional requerido. El desarrollo de la metodologı́a se divide en dos etapas. En la
primera etapa se considera un problema simplificado donde se asume que la excavación
es un pit semiesférico situado sobre la superficie de un dominio elástico semi-infinito y
sometido a gravedad. Este problema es resuelto usando dos métodos semianalı́ticos, donde
el segundo método es una mejora del primero. En la segunda etapa, se trata el caso de una
excavación de geometrı́a axisimétrica arbitraria mediante una técnica de acoplamiento, la
cual emplea el segundo método semi-analı́tico para calcular una aproximación del operador
Dirichlet-to-Neumann (DtN) sobre una frontera artificial semi-esférica. El operador DtN
provee las condiciones de borde exactas que completan el problema matemático planteado
en el dominio computacional existente entre la excavación y la frontera artificial, el cual es
entonces resuelto mediante el método de los elementos finitos (FEM). Se presentan resul-
tados numéricos para demostrar la eficiencia y precisión de la metodologı́a propuesta.

xii
Palabras Claves: Método numérico-analı́tico, excavación minera, esfuerzos, elastici-
dad, dominio semi-infinito, elementos finitos, operador Dirichlet-to-
Neumann

xiii
 1

1. INTRODUCTION

Many productive activities involve large scale excavation works. Typical examples
are the exploitation of mineral resources, hydrocarbons, or the construction of civil works.
From an operational point of view, the stability of excavations represents a crucial issue.
If an excavation has not been properly designed, particular zones may become unstable,
resulting in undesired events such as rock-falls or slides, which could put at risk the safety
of workers, equipment, infrastructure, or cause economic losses. Hence the need for pre-
dictive tools capable of determining those areas of the excavation that may be in risk of
collapse, in order to take the necessary measures in advance. Previous to any excavation
work, the geological structure of the rock mass is subjected to an initial state of stress of
gravitational, tectonic, hydraulic or residual origin. When a large scale excavation takes
place, the initial state of stress is modified, both on the new surfaces created by the exca-
vation and in the surrounding area. The new state of stress may give rise to instabilities,
particularly if certain areas of the excavation undergo high-stress concentration. A dis-
placement field is also generated in the excavation as a consequence of the modified state
of stress. Therefore, reliable and opportune information about stresses and displacements
in an excavation would make possible to detect beforehand the areas in potential risk of
failure.

Numerical methods are a common and powerful tool to compute the stresses and dis-
placements around structures in a rock mass, such an excavation. Among them, one of
the most widely used is the finite element method (FEM), mainly due to its advantages
in treating complex geometries and material properties. Nevertheless, the FEM applied to
problems in geomechanics presents a computational drawback that needs to be overcome.
As the whole rock mass around an excavation is of infinitely large size, it is not possible
in practice to cover the entire region with a finite element mesh. This drawback is often
remedied by performing the finite element analysis only within a finite region of rock mass
near the structure under study. This region is bounded by artificial boundaries, which nor-
mally consist of line segments or planes, depending on whether the study is two or three
 2

dimensional, respectively. Some authors that have adopted this approach in underground
mining are (Kulhawy, 1974), (Huttelmaier & Glockner, 1985) and (Ou, Chiou, & Wu,
1996). It has also been applied by (Griffiths & Lane, 1999) to slope stability in open pits.
However, this approach is not exempt from limitations. As it was established in (Kulhawy,
1974), the artificial boundaries need to be located far away enough from the structure in
order to achieve an acceptable accuracy. Hence, it is necessary to consider a sufficiently
large computational domain, which will require a large number of points in order to be
discretized, leading to an increasing amount in the required computational resources. This
is a essential point, because the chosen boundary conditions strongly affect the numerical
solution obtained. Nevertheless, it often receives little attention in related literature. What
is usually done is to assume that either all displacements are zero, or tangent stresses and
normal displacements are zero (cf. (Ou et al., 1996; Griffiths & Lane, 1999)). Other ap-
proximations suppose spring-like behaviours at boundaries (cf. (Huttelmaier & Glockner,
1985)). However, no mathematical justification of these boundary conditions is given.

In this thesis, we propose an accurate and efficient methodology to compute stresses

and displacements around an excavation, whose main strength is the effective and efficient
numerical treatment of the unbounded surrounding medium, when compared with the usual
approach considered by standard commercial software to perform the same calculations.
The methodology uses exact artificial boundary conditions in semi-infinite domains, which
allow us to considerably minimize the size of the computational domain to be meshed
and at the same time, to avoid the need of using spurious boundary conditions, resulting
in an important reduction of the required computational resources. In order to demon-
strate the efficiency and effectiveness of the proposed methodology, a simplified problem
is considered, where the unbounded rock mass is assumed to be an elastic, isotropic and
homogeneous semi-infinite medium, and the excavation shape is supposed axisymmetric.
By using spherical coordinates, the partial differential equations of elastostatics are ana-
lytically solved and exact boundary conditions are determined in order to be prescribed
on an artificial half-spherical boundary. Then a finite element analysis of stresses can be
performed inside the computational domain lying between the structure and the artificial
 3

boundary, where more complex characterizations of the rock mass could be assumed. As a
first stage of this study, we consider a hemispherical pit in the surface of an elastic, homo-
geneous and isotropic half-space subject to gravitational stress. Both the infinite flat surface
of the half-space and the pit are assumed to be traction-free boundaries. The complete so-
lution of this problem is obtained in a semi-analytical way. Analogous problems have been
approached for several authors in contexts different from geomechanics or mining. The first
of them was (R. Eubanks, 1953), who studied analytically the stresses and displacements
of an elastic half-space with a free boundary and a hemispherical cavity, induced due to a
constant pressure, parallel to the flat surface of the solid. Subsequently, (Fujita, Sadayasu,
Tsuchida, & Nakahara, 1978) improved some of the expressions given by Eubanks and pro-
vided a complete analytical solution. In a later work, (Fujita, Tsuchida, & Nakahara, 1982)
treated the case when the external force is no longer a homogeneous pressure, but a di-
rected force (simulating the traction or compression to which a metallic piece is subjected).
The same problem was solved by (Ovsyannikov & Starikov, 1987), who proposed some
variations based on singular solutions, in order to increase the efficiency of the solution
method. In more recent works, most of the attention has been concentrated on stresses on
hemispherical cavities when in addition there is corrosion in the metal (see (Cerit, Genel,
& Eksi, 2009; Turnbull, Horner, & Connolly, 2009)). All of these articles do not deal with
the volume forces due to the weight of the solid medium, which are essential in an elastic
model of the infinite rock mass. Eubanks employs the so-called Boussinesq potentials to
obtain an analytical solution as infinite series satisfying traction-free boundary conditions
on the plane surface and decaying conditions at infinity. However, imposing boundary con-
ditions on the surface of the pit leads to an infinite set of simultaneous linear equations for
the coefficients of the series, which cannot be solved exactly. After some unwieldy alge-
braic manipulation, these equations are solved numerically, yielding approximate values
for a finite number of coefficients. Hence, the solution given in (R. A. Eubanks, 1954) is
actually not fully analytical but semi-analytical, and similar phenomena occur in (Fujita et
al., 1978, 1982) and (Givoli & Vigdergauz, 1993). In addition, the numerical evaluation
of the solution in (R. A. Eubanks, 1954) involves computing the sum of double series that
 4

show a slow convergence at the surface of the pit, which is computationally expensive and
further complicates obtaining explicitly the solution.

In the second stage of this study we deal with an excavation of arbitrary axisymmetric
geometry placed on the surface of a half-space, assumed again elastic, isotropic and homo-
geneous. The boundary-value problem (BVP) that describes the stresses and displacements
around the excavation is formulated on an unbounded domain. Different mathematical and
numerical approaches have been devised to solve BVPs in unbounded domains. According
to (Givoli, 1992, 1999b), they are classified into four main categories: boundary integral
methods, infinite element methods, absorbing layer methods, and artificial boundary con-
dition (ABC) methods. The advantages and limitations of each one of them are discussed
in (Givoli, 1992). The present work is concerned with the latter category, also referred to
as artificial boundary method (H. D. Han & Wu, 2013b, 2013a).

In the standard ABC method, the original unbounded domain is truncated by introduc-
ing an artificial boundary enclosing the particular area of interest, thus defining a bounded
computational domain, where standard numerical methods may be used to solve the BVP.
This is possible on the condition that suitable boundary conditions are set on the artificial
boundary (the ABCs), which are supposed to properly represent the unbounded residual
domain that was eliminated. In general, many different choices of ABCs are possible.
We focus herein on ABCs based on the Dirichlet-to-Neumann (DtN) map and their use in
combination with finite elements, which results in the Dirichlet-to-Neumann finite element
method (DtN FEM) (Givoli, 1999b, 1999a). The main advantage of this approach is that the
DtN map provides exact ABCs, in such a way that the resulting BVP in the computational
domain is mathematically equivalent to the original unbounded BVP. Therefore, the use of
the FEM to solve the former results in highly accurate and robust numerical schemes.

The DtN FEM method has been mainly developed for BVPs formulated in exterior do-
mains, i.e. the complement of a compact set. It is customary to assume a circular or spher-
ical artificial boundary, in order to be able to apply the method of separation of variables to
solve the resulting BVP in the residual domain. Thus, an analytical solution in series form
 5

is obtained, with coefficients that are computed exactly. From it, an explicit closed-form
expression for the DtN map is deduced, given as an exact nonlocal relation between the
solution and its normal derivative on the artificial boundary, which is used as a boundary
condition that completes the mathematical formulation of the BVP in the computational
domain, making it available for numerical solution by finite elements. This procedure
has been extensively used to solve wave BVPs, namely the Helmholtz equation (Keller &
Givoli, 1989), (Grote & Keller, 1995b), (Deakin & Rasmussen, 2001), the time-dependent
wave equation (Grote & Keller, 1995a, 1996; Aladl, Deakin, & Rasmussen, 2002), and
the elastodynamic equation, both in frequency domain (Givoli & Keller, 1990; Harari &
Shohet, 1998) and time domain (Grote & Keller, 2000; Grote, 2000; Gächter & Grote,
2003). In this context, ABCs are also called non-reflecting boundary conditions, since they
are aimed at preventing any spurious reflection of waves from the artificial boundary. The
DtN FEM method has also been successfully applied to linear elliptic BVPs in exterior do-
mains, mainly the Laplace equation and linear elastostatics (H. D. Han & Wu, 1985; Givoli
& Keller, 1989; H. D. Han & Wu, 1992; H. D. Han & Zheng, 2005).

A particularly important field of application of unbounded BVPs is geophysics, where

the ground is usually modelled as an unbounded elastic domain. A typical example is
the problem of determining deformations and stresses around structures of interest in the
ground, such as an excavation in mining or civil engineering. Even though an exterior
domain may be used as a model of the ground in a first approximation, a more realistic
model needs to take into account the ground surface, which is customarily assumed, for
simplicity, to be a plane boundary that extends to infinity, where a traction-free (Neumann)
boundary condition holds. The resulting unbounded domain is called semi-infinite. To
apply a DtN FEM approach to such a domain, a natural approach would be to consider
as artificial boundary a semi-circle or semi-sphere surrounding the structure of interest, in
analogy to the exterior case. This procedure works properly for some standard 2D scalar
BVPs, such as the Laplace equation (Givoli, 1992) and the Helmholtz equation (Givoli &
Vigdergauz, 1993) with Dirichlet or Neumann boundary conditions holding on the infinite
plane boundary, since in these cases the BVP in the residual semi-infinite domain can be
 6

solved by separation of variables. However, the same procedure cannot be directly applied
to linear elastostatics in semi-infinite domains, since in this case the method of separation
of variables fails in solving the BVP in the residual domain in a fully analytical way, and
thus it is not possible to get an explicit closed-form expression for the DtN map. Givoli and
Vigdergauz (Givoli & Vigdergauz, 1993) proposed an alternative to deal with this draw-
back in the 2D case. They considered a semi-circular artificial boundary and solved the
BVP in the semi-infinite residual domain using complex analysis techniques, resulting in a
solution in series form, whose coefficients can only be computed in an approximated way.
A different approach was used by Han, Bao and Wang (H. Han, Bao, & Wang, 1997), who
employed the direct method of lines with semi-discretisation to solve the BVP in the resid-
ual domain, obtaining an approximation of the DtN map. No previous work on DtN FEM
procedures for 3D elastostatics in semi-infinite domains was found in the literature.

In the last chapter, we present a DtN FEM for a semi-infinite elastic domain in 3D. A
geometrical perturbation on the plane surface representing the structure of interest is con-
sidered, which is assumed to be axisymmetric about the vertical axis. Assuming further
that the elastic domain is only under axisymmetric loading, the whole problem is treated
as a BVP of axisymmetric elastostatics. The semi-infinite domain is truncated by intro-
ducing a semi-spherical artificial boundary surrounding the perturbation. Then a standard
FEM discretisation in the resulting computational domain with exact ABCs in the artificial
boundary is established at a theoretical level. As it is not possible to obtain a fully explicit
closed-form expression for the DtN map, we proceed in a similar way to (Givoli & Vigder-
gauz, 1993), solving the BVP in the residual domain just for particular Dirichlet data on
the artificial boundary, in order to approximate those boundary integral terms involving the
DtN map that occur in the finite element formulation. The BVP in the residual domain for
each required case is solved by a semi-analytical method, analogous to that presented in
Chapter IV. By applying separation of variables, a general analytical solution is calculated,
expressed as a series with unknown coefficients, which are approximated by minimising
a quadratic energy functional appropriately chosen. The minimisation yields a symmetric
and positive definite linear system of equations for a finite number of coefficients, which is
 7

efficiently solved by exploiting its particular block structure, in such a way that the coeffi-
cients are computed by mere forward and backward substitutions. This procedure allows
an approximate but effective coupling of the exact nonlocal ABCs provided by the DtN
map with the FEM scheme. The method is validated by solving a model problem whose
exact solution is available. The relative error between the numerical and the exact solution
is analysed for different mesh sizes.
 8

2. GENERAL MODEL

In this chapter, the mathematical model describing the elastic deformations and stresses
in a half-space with a hemispherical pit is formulated.

2.1. Mathematical Model

Let us consider the lower half-space with a geometrical perturbation on its plane sur-
face which is local, i.e. it is restricted to a bounded region. This type of domain is often-
times referred to as a locally perturbed half-space. In addition, the perturbation is assumed
to be rotationally symmetric about a vertical axis, in such a way that the whole semi-infinite
domain is actually an axisymmetric solid (or solid of revolution). Therefore, its geometry
is completely described by its generating cross section, which consists in a 2D domain
that we denote by Ω. The boundary of Ω consists of three parts: A vertical boundary of
axisymmetry coinciding with the axis of revolution, denoted by Γs , a horizontal unper-
turbed boundary coinciding with the infinite plane surface of the half-space, denoted by
Γ∞ , and a perturbed bounded boundary which is assumed piecewise smooth, denoted by
Γh (see Figure 2.1). The domain Ω will be described with the aid of axisymmetric cylindri-
cal coordinates (ρ, z) or spherical coordinates (r, φ) as appropriate, with the origin placed
at the point of intersection between the axis of revolution and the plane boundary of the
unperturbed half-space. Variables ρ, z, r and φ are linked by the relations
z
r 2 = ρ2 + z 2 , φ = arctan , ρ = r sin φ, z = r cos φ.
ρ
The unit vectors associated with variables ρ, z, r and φ are denoted respectively by ρ̂, ẑ, r̂
and φ̂. They are linked by the relations

ρ̂ = r̂ sin φ + φ̂ cos φ, ẑ = r̂ cos φ − φ̂ sin φ, (2.1a)

r̂ = ρ̂ sin φ + ẑ cos φ, φ̂ = ρ̂ cos φ − ẑ sin φ. (2.1b)

Thus an arbitrary position in Ω is expressed either as (ρ, z) or (r, φ), depending on the
chosen system of coordinates. Moreover, we denote by θ the azimuthal angle, which is not
 9

necessary to describe geometrically the 2D domain Ω, but it is required in the axisymmetric

elastostatic model.

We assume that Ω is occupied by an isotropic, homogeneous, linear elastic medium.

Under the condition that the 3D axisymmetric domain is only subject to axisymmetric load-
ing, the resulting elastic deformations will keep the axisymmetric nature of the problem.
A generic displacement field defined in Ω is denoted by u and its associated stress tensor
is denoted by σ = σ(u). By virtue of the axisymmetry, u has only components uρ and
uz (resp. ur and uφ ), whereas σ has normal components σρ and σz (resp. σr and σφ ), a
shear component σρz (resp. σrφ ), and an additional non-vanishing normal component σθ
(see (Sadd, 2005) for details). It is assumed that σ is given in terms of u by the isotropic
Hooke’s law, that is,
σ(u) = λ (∇ · u) I + µ(∇u + ∇uT ), (2.2)

where λ, µ > 0 are the Lamé constants of the elastic solid and I is the 3×3 identity matrix.
We assume the downward gravitational force to be the only body force acting on the solid
medium. The elastic equilibrium is thus governed by Navier’s equation:

− ∇ · σ(u) = %gẑ, (2.3)

or written in terms of the displacements

(λ + µ)∇(∇ · u) + µ∆u = −%gẑ, (2.4)

where g denotes the acceleration of gravity and ẑ stands for the unit vector in the direction
of positive z-axis. The right-hand side of (2.3) takes into account the effect of the gravity
force per unit volume of solid. It has constant magnitude %g and acts in the direction of
−ẑ everywhere in R3− . As the half-space is unbounded in both horizontal directions x
and y, the displacement field due to the gravity force, which we shall call ug , must be in
the direction of −ẑ. On the other hand, it is assumed that the infinite plane surface of
the half-space, defined by z = 0, is traction-free, which is mathematically expressed by
 10

homogeneous Neumann boundary conditions:

σ(u)ẑ = 0. (2.5)

r≥h z
z π
2 ≤φ≤π
0 ≤ θ ≤ 2π φ
Γ∞
Γ∞ x

Γh
h r
y
Γh
Ω
x Γs Ω

Figure 2.1. Geometrical model of the elastoestatic problem.

hemispherical boundary, by Γ∞ its infinite horizontal boundary, and by Γs its vertical

boundary. These sets are defined in terms of (r, φ) as follows:

Ω = {(r, φ) : h < r < ∞, π/2 < φ < π}, (2.6a)

Γh = {(r, φ) : r = h, π/2 < φ < π}, (2.6b)

Γ∞ = {(r, φ) : r ≥ h, φ = π/2}, (2.6c)

Γs = {(r, φ) : r ≥ h, φ = π}. (2.6d)

The stress (see Figure 2.2) and the displacements are related through the Hooke’s law,
which in this isotropic case is σij = λ∇ · uδij + 2µεij .The Hooke’s law (2.2) is written by
components as
∂uρ uρ ∂uz
σρ (u) = (λ + 2µ) +λ +λ , (2.7a)
∂ρ ρ ∂z
∂uρ uρ ∂uz
σθ (u) = λ + (λ + 2µ) + λ , (2.7b)
∂ρ ρ ∂z
∂uρ uρ ∂uz
σz (u) = λ + λ + (λ + 2µ) , (2.7c)
∂ρ ρ ∂z
 11

σrθ
σr

σrφ

σθ

φ r
σφ
σφθ

y
θ

Figure 2.2. Stress components in spherical coordinates.

 ∂u ∂uz 
ρ
σρz (u) = µ + , (2.7d)
∂z ∂ρ
in cylindrical coordinates and as
∂ur ur λ  ∂uφ 
σr (u) = (λ + 2µ) + 2λ + + cot φ uφ , (2.8a)
∂r r r ∂φ
∂ur ur 1  ∂uφ 
σφ (u) = λ + 2(λ + µ) + (λ + 2µ) + λ cot φ uφ , (2.8b)
∂r r r ∂φ
∂ur ur 1  ∂uφ 
σθ (u) = λ + 2(λ + µ) + λ + (λ + 2µ) cot φ uφ , (2.8c)
∂r r r ∂φ
 1 ∂u ∂uφ uφ 
r
σrφ (u) = µ + − , (2.8d)
r ∂φ ∂r r
in spherical coordinates (see (Sadd, 2005) for details).
 12

The elastic equilibrium of the perturbed domain Ω subject to gravity is governed by

Navier’s equation, given in (2.3). On the other hand, Γ∞ and Γh are both assumed to be
traction-free, which means that homogeneous Neumann boundary conditions as those given
in (2.5) hold on both boundaries, with ẑ substituted by −r̂ in the latter case. In addition,
the vertical boundary Γs is assumed to be free of shear traction and constrained against
normal displacement (otherwise the axisymmetry would be destroyed). Furthermore, from
physical intuition it is reasonable to assume that the effect of the hemispherical pit on
the displacement field in Ω is essentially local, that is, at large distances from the origin
the elastic half-space deforms as if there is no geometrical perturbation. We thus assume
that, as the distance from the pit tends to infinity, the displacement field u approaches
asymptotically ug , i.e., the displacement field in absence of perturbation. An important
related issue is the rate at which this asymptotic approach at infinity takes place, which
is not a simple matter. It rather concerns questions of existence and uniqueness, which
are beyond the scope of this work. We simply assume that when r tends to infinity, the
difference in norm between u and ug decays to zero as O(1/r), which is sufficient for our
purposes. Taking into account all these assumptions, u is obtained as a solution of the
boundary-value problem: Find u : Ω → R2 such that

∇ · σ(u) = −ρgẑ in Ω, (2.9a)

σ(u)ẑ = 0 on Γ∞ , (2.9b)

σ(u)r̂ = 0 on Γh , (2.9c)

σ(u)φ̂ · r̂ = u · φ̂ = 0 on Γs , (2.9d)
1
|u − ug | = O as r → ∞, (2.9e)
r
%gz 2
where |·| stands for the Euclidean norm. We make the change of variables v = u− 2(λ+2µ) ẑ
in order to achieve a homogeneous equation. The displacement field u is then

%gz 2
u = ug + v, ug (x, y, z) = − ẑ. (2.10)
2(λ + 2µ)
 13

The Navier equation in this new displacement field v is

(λ + µ)∇(∇ · v) + µ∆v = 0. (2.11)

The problem is then: Find v : Ω → R2 such that

∇ · σ(v) = 0 in Ω, (2.12a)

σ(v)ẑ = 0 on Γ∞ , (2.12b)

σ(v)r̂ = f on Γh , (2.12c)

σ(v)φ̂ · r̂ = v · φ̂ = 0 on Γs , (2.12d)
1
|v| = O as r → ∞, (2.12e)
r
where f : Γh → R2 is a vector function defined by its components in r and φ as
π
f (φ) = fr (φ) r̂ + fφ (φ) φ̂, ≤ φ ≤ π, (2.13)
2
with
ν + (1 − 2ν) cos2 φ cos φ


fr (φ) = %gh , (2.14a)
1−ν
(1 − 2ν) cos2 φ sin φ
fφ (φ) = −%gh , (2.14b)
1−ν
The boundary-value problem (2.12) has certain advantages over the original problem, such
as a homogeneous Navier’s equation (2.12a) and a solution decaying to zero at infinity
(2.12e). Hence, in order to determine the displacement field u, we first solve (2.12) to
obtain the displacement field v and then we calculate u from (2.10), given that an explicit
expression for ug is available.

2.1.1. Papkovich-Neuber’s decomposition

It should be noted that only two independent elastic constants are needed to describe
the behaviour of isotropic materials. The relationships between elastic constants E, λ, µ
and ν are provided in Table 2.1.
 14

Table 2.1. Relationship between the different constants that characterise a

material.

(λ, µ) (E, µ) (λ, ν) (µ, ν) (E, ν)

3λ + 2µ λ(1 + ν)(1 − 2ν)

E µ E 2µ(1 + ν) E
λ+µ ν

E − 2µ 2µν Eν
λ λ µ λ
3µ − E 1 − 2ν (1 + ν)(1 − 2ν)

1 − 2ν E
µ µ µ λ µ
2ν 2(1 + ν)

λ E
ν −1 ν ν ν
2(λ + µ) 2µ

We can use the Helmholtz decomposition, which states that any sufficiently continuous
vector field can be represented as the sum of the gradient of a scalar potential ϕ and the curl
of a vector potential φ, that is, u = ∇ϕ + ∇ × φ. The gradient term in the decomposition
has a zero curl and is referred to as the irrotational part, while the curl term has a zero
divergence and is called solenoidal. Note that this representation specifies three displace-
ment components in terms of four potential components, and furthermore the divergence
of φ is arbitrary. In order to address these problems, it is common to choose φ with zero
divergence, i.e., ∇ · φ = 0.

Based on the Helmholtz’s decomposition, we can define ϕ such that ∇ · u = ∆ϕ, so

1 F
∆u + ∇(∆ϕ) = − , (2.15)
(1 − 2ν) µ
 15

and arranging in a suitable way,

 
1 F
∆ u+ ∇ϕ = − . (2.16)
(1 − 2ν) µ
 
2µ
We define the vectorial potential such that A = 2µu + ∇ϕ , so for A =
(1 − 2ν)
[Ar Aθ Az ]T we have the following relations

4µ(1 − ν)
∆A = −2F and ∇·A= ∆ϕ. (2.17)
(1 − 2ν)
The second relation follows from ∇ · ∇ϕ = ∆ϕ, and by definition ∇ · u = ∆ϕ.

From vectorial calculus we know that for any vectorial field F (·) and a constant vector
R the following relation holds,

∆(R · F ) = R · ∆F + 2(∇ · F ).

In particular we have,
∆(R · A) = R · ∆A + 2(∇ · A), (2.18)

which is the same as,

1
∇·A= (∆(R · A) + 2(R · f )) , (2.19)
2
and this is equivalent to
4µ(1 − ν) 1
∆ϕ = (∆(R · A) + 2(R · F )) , (2.20)
(1 − 2ν) 2
which implies
4µ(1 − ν)
 
1
∆ ϕ − (R · A) = R · F . (2.21)
(1 − 2ν) 2
(R · A)
 
2µ
We introduce the scalar potential B = ϕ− in sustitution of ϕ,
(1 − 2ν) 4(1 − ν)
which satisfies
R·F
∆B = ,
2(1 − ν)
in our case we have F = 0, which means the potentials satisfy

∆A = 0, ∆B = 0. (2.22)
 16

Moreover, in this axisymmetric case Ar = Aθ = 0, Az = Az (r, z) and B = B(r, z), where

Az , B are commonly called the Boussinesq potentials. If we take the definitions of A y B,
it holds that
R·A
 
2µu = A − ∇ B + . (2.23)
4(1 − ν)
If we use the decomposition of Papkovich-Neuber, we can find the solution of (2.4) through
the scalar potentials B(r, φ) and Az (r, φ), which are related with v by means of the expres-
sion
2µv = ∇(B + zAz ) − 4(1 − ν)Az ẑ, (2.24)

where both potentials are harmonic functions, i.e., ∆B = ∆Az = 0, and v must satisfy the
boundary conditions. The completeness of this representation was shown by (R. Eubanks
& Sternberg, 1956), and thus all elasticity solutions are representable by this scheme. The
solution to Navier’s equation in Ω is sought with the aid of the so-called Boussinesq po-
tentials, which are a particular case of the more general Papkovich-Neuber (or Boussinesq-
Papkovich) representation (cf. (Amenzade, 1979; Sadd, 2005)). According to (R. A. Eu-
banks, 1954), v is defined through the following relation

2µv = ∇(Φ + zΨ) − 4(1 − ν)Ψ k̂, (2.25)

where Φ and Ψ are the Boussinesq potentials, which are harmonic functions thanks to the
fact that Navier’s equation (2.12a) is homogeneous. Otherwise, Φ and Ψ would satisfy the
Poisson equation, with nonzero right-hand sides. Expressing (2.25) by its components in r
and φ yields
∂Φ ∂Ψ
2µvr (r, φ) = (r, φ) + r cos φ (r, φ) − (3 − 4ν) cos φ Ψ(r, φ), (2.26a)
∂r ∂r
1 ∂Φ ∂Ψ
2µvφ (r, φ) = (r, φ) + cos φ (r, φ) + (3 − 4ν) sin φ Ψ(r, φ). (2.26b)
r ∂φ ∂φ
In order to calculate a solution of (2.12a), it becomes necessary to solve the Laplace equa-
tion in Ω for Φ and Ψ. Prior to this, we analyse in more detail the decaying condition at
infinity for v. By replacing (2.25) in (2.12e), we infer that Φ and Ψ have to satisfy certain
asymptotic behaviours when r tends to infinity. Specifically, Ψ has to decay to zero as
 17

O(1/r). It is also required that the term ∇(zΨ) appearing in (2.25) decreases to zero at the
same rate. We also need that ∇Φ decays to zero as O(1/r), which is a weaker condition
than the one required for Ψ. Therefore, Ψ is sought as a solution of

∆Ψ = 0 in Ω, (2.27a)
1
Ψ=O as r → ∞, (2.27b)
r
and Φ is sought as a solution of

∆Φ = 0 in Ω, (2.28a)
1
|∇Φ| = O as r → ∞, (2.28b)
r
where ∆ stands for the Laplacian. Let us first determine Ψ. Expressing Ψ = Ψ(r, φ), the
Laplace equation in axisymmetric spherical coordinates is given by
   
1 ∂ 2 ∂Ψ(r, φ) 1 ∂ ∂Ψ(r, φ)
∆Ψ(r, φ) = 2 r + 2 sin φ = 0. (2.29)
r ∂r ∂r r sin φ ∂φ ∂φ
If we apply standard separation of variables in r and φ to (2.29) (cf. (Arfken & Weber,
2001)), and we discard those solutions that are unbounded in Ω, we obtain that for each
integer n ≥ 0, the function Ψn defined as
Pn (cos φ)
Ψn (r, φ) = , (2.30)
rn+1
is a solution of (2.27a), where Pn (·) denotes the Legendre polynomial of order n (see
Appendix 6.1). The general solution of (2.27a) is then expressed as an infinite linear com-
bination of functions Ψn . By virtue of (2.30) it is immediate that
 1 
Ψn (r, φ) = O as r → ∞,
rn+1
which means that when r increases, the first term of the infinite linear combination (the one
for n = 0) dominates over the rest of the terms, and as this term behaves asymptotically as
O(1/r), the linear combination of all terms behaves asymptotically the same way. There-
fore, the obtained general solution of (2.27a) also fulfils the decaying condition (2.27b). In
 18

addition, if we compute the terms ∇(zΨn ) for each n ≥ 0, we verify that

 1 
|∇(zΨn (r, φ))| = O as r → ∞,
rn+1
and an analogous reasoning allows us to state that when r tends to infinity, the term ∇(zΨ)
decreases asymptotically in norm as O(1/r). Hence, this term does not affect the fulfilment
of (2.12e). Let us now determine Φ. Proceeding analogously as for Ψ, we obtain that the
functions Φn defined by
Pn (cos φ)
Φn (r, φ) = , (2.31)
rn+1
satisfy (2.28a) for each integer n ≥ 0. As done in (R. A. Eubanks, 1954), it seems rea-
sonable to propose an infinite linear combination of functions Φn as the general solution of
(2.28a), in analogy to the solution established for (2.27a). Nevertheless, according to the
authors’ opinion such a solution is not general enough, owing to the decaying condition at
infinity (2.28b) imposed on Φ, which differs from that imposed on Ψ. If we calculate the
gradients of functions Φn defined in (2.31) and we take their norms, it is easy to see that
 1 
|∇Φn (r, φ)| = O n+2 as r → ∞.
r
Consequently, if we express Φ as a linear combination of functions Φn defined in (2.31),
and if we study the asymptotic behaviour of |∇Φ| as r increases, we obtain that the first
term, which is again the dominating one, behaves asymptotically as O(1/r2 ). Thus, when
r tends to infinity |∇Φ| decreases to zero as O(1/r2 ). This means that the solution Φ, just
expressed as a linear combination of functions Φn , satisfies a more restrictive condition
than (2.28b), so it is not general enough to be the sought solution of (2.28). In order to
achieve the required generality in Φ, we add a new component to the set of functions Φn ,
which gives rise to an asymptotic behaviour of order O(1/r) in |∇Φ| as r tends to infinity.
This component, associated for convenience with the index n = −1, corresponds to the
following logarithmic potential

Φ−1 (r, φ) = ln(r − r cos φ). (2.32)

 19

This potential arises when solving the Boussinesq’s problem, i.e., the one of a concentrated
force acting normal to the free surface of an elastic half-space (cf. (Sadd, 2005)). It is also
known as the Boussinesq’s elementary solution of the second kind (cf. (Amenzade, 1979)).
It is straightforward to verify that Φ−1 is a solution of (2.29). In addition, the singularity of
the logarithm at zero in (2.32) does not cause problems of unboundedness, since all those
points such that cos φ = 1 (i.e. φ = 0) are not contained in the domain Ω. Moreover, if we
compute the gradient of Φ−1 and we take its norm, we arrive at
1
|∇Φ−1 (r, φ)| = O as r → ∞.
r
Hence, the sought solution to (2.28) corresponds to an infinite linear combination of func-
tions Φn including the case n = −1, since such a solution satisfies the right decaying
condition at infinity (2.28b).
 20

3. A SIMPLE SEMI-INFINITE GEOMETRY

In this chapter we want to find the displacements and the stresses in a semi-infinite
region. In order to do so, we use the Papkovich-Neuber decomposition technique and by
imposing different boundary conditions in the perturbation of the domain, we can obtain
an expression that allows us to evaluate the displacements and the stresses in any point of
our domain.

3.1. Series’s solution

In order to find the solution to (2.25) we note that Laplace’s equation in spherical
coordinates is
∂ 2f
   
1 ∂ 2 ∂f 1 1 ∂ ∂f
∆f = 2 r + 2 2 2
+ 2 sin φ . (3.1)
r ∂r ∂r r sin φ ∂θ r sin φ ∂φ ∂φ

Let us consider the problem of finding solutions of the form f (r, φ) = g(r)h(φ). By
separation of variables, two differential equations result by imposing Laplace’s equation.
We use then g(r) = rn . Furthermore, the change of variables x = cos φ transforms this
equation into the Legendre differential equation

∂ 2 h(x) ∂h(x)
(1 − x2 ) − 2x + n(n + 1)h(x) = 0, (3.2)
∂x2 ∂x
whose bounded solutions are the Legendre polynomials. We assume a bounded perturba-
tion, so at infinity, the expression for Az must decay, while the gradient for B must tend to
zero. Hence, we write in general the solutions in the form
∞
X
Az (r, φ) = an r−(n+1) Pn (cos φ),
n=0
∞
X
B(r, φ) = b̃ ln(r − r cos φ) + bn r−(n+1) Pn (cos φ). (3.3)
n=0
 21

In order to solve the problem, we replace (3.3) in (2.25). For the sake of convenience,
we omit henceforth the argument of the Legendre polynomials, since all of them are eval-
uated at cos φ unless otherwise is indicated. We thus obtain the following expressions for
the displacements
∞ 
(n + 4 − 4ν) cos φPn

1 X (n + 1)
2µvr = b̃ + − n+1
an − n+2 Pn bn , (3.4)
r n=0
r r
∞ 
sin φ((3 − 4ν)Pn − cos φPn0 ) sin φPn0

sin φ X
2µvφ = b̃ + n+1
an − n+2 bn . (3.5)
r(1 − cos φ) n=0
r r

We make the linear combination An = (2n + 1)an − (n + 4 − 4ν)bn−1 valid for

n = 0, 1, . . ., in order to simplify the calculation. In this axisymmetric model, it holds that
σφθ = σrθ = 0. The remaining stresses are

2(2(1 − ν) + (2 − ν) cos φ) 1
σr = 2
A0 − 2 b̃
r r
∞
X (n + 1)((n + 1)(n + 4) − 2ν)Pn+1
 
(n + 1)(n + 2)
+ An + Pn b n , (3.6)
n=0
rn+2 rn+3
(3 − 2ν + (1 − 2ν) cos φ) cos φ cos φ
σφ = 2
A0 − 2 b̃
(1 − cos φ)r r (1 − cos φ)
∞ 
X (n + 1)(n2 − n + 1 − 2ν)Pn+1 − (n − 3 + 4ν)Pn0
+ − An
n=0
rn+2
0
Pn+1 − (n + 1)(n + 2)Pn

+ bn , (3.7)
rn+3
4(1 − ν) + (1 − 2ν)(1 − cos φ) cos φ 1
σθ = − 2
A0 + 2 b̃
(1 − cos φ)r r (1 − cos φ)
∞ 
−(1 + 2ν)(n + 1)(2n + 1)Pn+1 − (n − 3 + 4ν)Pn0

X Pn+1
+ n+2
An − n+3 bn , (3.8)
n=0
r r
(3 − 2ν + (1 − 2ν) cos φ) sin φ
σrφ = 2
sin φA0 − 2 b̃
(1 − cos φ)r r (1 − cos φ)
∞  0
(n2 + 2n − 1 + 2ν)Pn+1 (n + 2) sin φPn0
X 
+ sin φAn + bn . (3.9)
n=0
rn+2 rn+3
 22

3.2. First set of boundary conditions: free surface boundary condition on the surface
of the plane.

We know that any picewise continuous function can be written in the base of {P2n }∞
n=0
0 ∞
in the range [−1, 0], and also we can make a descomposition in terms of the basis {sin φP2n }n=0
in the same range (see (Brown & Churchill, 2006)), where the orthonormalization constants
are
∞
X
F (φ) = F2n P2n (cos φ), (3.10)
n=0
Z π
F2n = (4n + 1) F (φ) sin φP2n (cos φ)dφ, (3.11)
π/2
∞
X
0
G(φ) = G2n sin φP2n (cos φ), (3.12)
n=0
π
(4n + 1)
Z
0
G2n = G(φ) sin2 φP2n (cos φ)dφ. (3.13)
2n(2n + 1) π/2

In particular, Legendre’s polynomials of odd degree are expressed in terms of this basis,
for −1 ≤ x < 0 and k = 0, 1, . . ., as follows (see Appendix A)
∞
(n)
X
P2k+1 (x) = −(2k + 1)P2k (0) ωk P2n (x), (3.14)
n=0
∞
(n)
X
0 0
P2k+1 (x) = −(2k + 1)P2k (0) ωk P2n (x), (3.15)
n=0

(n) (4n + 1)P2n (0)

ωk = . (3.16)
(2k + 1 − 2n)(2k + 2 + 2n)

In addition, by using the recurrence relations for Legendre polynomials, we find that

0
P2k+1 (0) = (2k + 1)P2k (0), (3.17)
(2k + 1)
P2k+2 (0) = − P2k (0), (3.18)
(2k + 2)
(2k)!
P2k (0) = (−1)k . (3.19)
22k (k!)2
 23

P2n (0)
In the specific case of F (x) = ln(1 − x) we find that F0 = 2 ln 2 − 1 and Fn = − 2n(2n+1) .

Next, we impose the free surface boundary condition on the surface of the plane Γ∞
(r > h, φ = π/2), whose components are σφ (v) = σrφ (v) = 0. We use for this purpose
0
the relations P2n+1 (0) = P2n (0) = 0 for all n, and define the coefficients αn = (n + 1)2 −
2 + 2ν and βn = (n + 2)(n + 5) − 2ν. This yields the following contributions, which are
arranged according to the parity of the coefficients

3−2ν
A0 ⇒ σφ = 0, σrφ = r2
,
(2n+1)α2n
A2n+1 ⇒ σφ = r2n+3
P2n (0), σrφ = 0,
A2n+2 ⇒ σφ = 0, σrφ = − (2n+1)(2n+3)α2n+2
(2n+2)r2n+4
P2n (0),
(3.20)
b̃ ⇒ σφ = 0, σrφ = − r12 ,
2
b2n ⇒ σφ = − (2n+1)
r2n+3
P2n (0), σrφ = 0,
(2n+1)(2n+3)
b2n+1 ⇒ σφ = 0, σrφ = r2n+4
P2n (0).

It should be observed that the contributions to σφ are given by the odd terms of An
and the even terms of bn , while for σrφ the opposite happens. Motivated by this fact, we
take linear combinations of the coefficients, in order to continue fulfilling the boundary
conditions and to simplify further calculations. In particular, the displacements are
−1 + 2ν 4(1 − ν)
 
2µvr (r, φ) = P0 − P1 B̃
r r
∞ 
X (2n + 1)(2n + 2)(2n + 5 − 4ν)P2n+2 (2n + 1)α2n P2n
− 2n+2
Ãn + 2n+2
Ãn
n=0
r r

(2n + 2)α2n+2 P2n+1 (2n+2)(2n+3)(2n +6−4ν)P2n+3
+ B̃n + B̃n , (3.21)
r2n+3 r2n+3
2(1 − ν) (3 − 4ν) cos φ
 
sin φ
2µvφ (r, φ) = − B̃
r (1 − cos φ) (1 − cos φ)
∞  0 0
X (2n + 1)(2n − 2 + 4ν)P2n+2 α2n P2n
− sin φ Ã n + Ãn
n=0
r2n+2 r2n+2
0 0
α2n+2 P2n+1 (2n + 2)(2n − 1 + 4ν)P2n+3

+ B̃n + B̃n . (3.22)
r2n+3 r2n+3
 24

The stresses are

1 − 2ν 2(2 − ν)
 
σr = P0 + P1 B̃
r2 r2
∞ 
X (2n + 1)(2n + 2)
+ 2n+3
(α2n P2n + β2n P2n+2 ) Ãn (3.23)
n=0
r

(2n + 2)(2n + 3)
+ (α2n+2 P2n+1 + β2n+1 P2n+3 ) B̃n ,
r2n+4
(1 − 2ν) cos2 φ
σφ = B̃
(1 − cos φ)r2
∞
"
(n + 1)(4n2 + 2n + 1 − 2ν)

X (n + 1)α2n
+ − 2(2n + 1) 2n+3
P 2n + 2n+3
P2n+2
n=0
r r
(4n + 3)(2n + 2ν − 1) 0

+ P2n+1 Ãn
r2n+3

(5 + 4n)(2n + 1 + 2ν) 0
+ P2n+2
r2n+4
 #
α2n+2 P2n+1 + (4n2 + 6n + 3 − 2ν)P2n+3
−(2n + 3)(2n + 2) B̃n , (3.24)
r2n+4

(1 − 2ν)(cos2 φ − cos φ − 1)
σθ = B̃
(1 − cos φ)r2
∞
"
0
(1 − 2ν)(2n + 1)(2n + 2)P2n+2 (2n − 1 + 2ν)P2n+1
X  
+ − (4n + 3) + Ãn
n=0
r2n+3 r2n+3
0
 #
(1 − 2ν)(2n + 2)(2n + 3)P2n+3 (2n + 1 + 2ν)P2n+2

− (4n + 5) + B̃n , (3.25)
r2n+4 r2n+4
(1 − 2ν) sin φ cos φ
σrφ = B̃
(1 − cos φ)r2
∞
"
X sin φ 0 0
+ 2n+3
((2n + 2)α2n P2n + (2n + 1)α2n+1 P2n+2 )Ãn (3.26)
n=0
r
#
sin φ 0 0
+ 2n+4 α2n+2 ((2n + 2)P2n+3 + (2n + 3)P2n+1 )B̃n .
r
 25

3.3. Second set of boundary conditions: boundary conditions on the surface of the
semiesphere

Until now we have the solution for the following model

  

 Find v(r, φ) =  r
 v (r, φ)

  such that:
v (r, φ)



 φ



(P )

 (λ + µ)∇(∇ · v) + µ∆v = 0 in Ω, (3.27)


σφ (v) = σrφ (v) = 0 on Γ∞ , (3.28)






 v→0

r → ∞. (3.29)

We still have to impose the second set of boundary conditions on the boundary Γh (the
surface of the semiesphere). To achieve this we use various options (with respect to the
boundary conditions) in the following sections.

3.3.1. Traction-free boundary

In this case we set σr (u) = σrφ (u) = 0. We make a change of variables (which we
already use in (2.10)), and combining with (3.14) we get

ν + (1 − 2ν) cos2 φ
 
σr (v) = %gh cos φ
1−ν
 
%gh 1 2
= (3 − ν)P1 + (1 − 2ν)P3
1−ν 5 5
∞
%gh X  (n) (n)

=− (3 − ν)ω0 − 3(1 − 2ν)ω1 P2n , (3.30)
5(1 − ν) n=0
1 − 2ν
 
2
σrφ (v) = −%gh cos φ sin φ
1−ν
1 − 2ν 1
   
= −%gh sin φ 3P10 + 2P30
1 − ν 15
 ∞
1 − 2ν X  (n)

1 (n)

0
= − %gh ω1 − ω0 sin φP2n . (3.31)
5 1 − ν n=0
 26

We replace (3.23) in (3.30). Rearranging the terms appropiately and introducing a

Kronecker delta gives
∞ 
X %gh  (n) (n)
 (2n + 1)(2n + 2)
(3 − ν)ω0 − 3(1 − 2ν)ω1 + α2n 2n+3
Ãn
n=0
5(1 − ν) h
1 − 2ν (n) 2(2 − ν) (n)
 
+ δ − ω0 B̃
h2 0 h2
∞
#
X (2k+1)(2k+3)P2k (0)  (n) (n)

− 2k+4
(2k+2)α2k+2 ωk − (2k+3)β2k+1 ωk+1 B̃k P2n
k=0
h
!
(2n+1)(2n+2)
+β2n Ãn P2n+2 = 0. (3.32)
h2n+3

Defining the coefficients C̃ = B̃ and Ck = (2k + 1)(2k + 3)P2k (0)B̃k , the first boundary
condition at Γh becomes
2n(2n − 1) 1 − 2ν (n) 2(2 − ν) (n)
 
(2n + 1)(2n + 2)
α2n Ãn + β2n−2 Ãn−1 + δ − ω0 C̃
h2n+3 h2n+1 h2 0 h2
∞
X Ck  (n) (n)

+ −(2k + 2)α ω
2k+2 k + (2k + 3)β ω
2k+1 k+1
k=0
h2k+4
%gh  (n) (n)

+ (3 − ν)ω0 − 3(1 − 2ν)ω1 = 0. (3.33)
5(1 − ν)

In the same way as we proceeded before, we substitute (3.26) in (3.31)

∞
"
1 − 2ν  (n)  (1− 2ν)  
X (n) (n) (4n + 1)
%gh ω − ω0 + ω0 + P2n (0) B̃
n=0
5(1 − ν) 1 h2 2n(2n + 1)
α2n (2n + 2) (2n + 1)α2n+1 0
+ Ãn + P2n+2 Ãn
h2n+3 h2n+3
∞
# !
X (2k + 1)(2k + 3)α2k+2 P2k (0)  (n) (n)

0
+ 2k+4
ωk+1 − ωk B̃k P2n = 0. (3.34)
k=0
h
 27

The second boundary condition is

2n − 1 1 − 2ν
 
2n + 2 (n) (4n + 1)
α2n 2n+3 Ãn + α2n−1 2n+1 Ãn−1 + ω0 + P2n (0) C̃
h h h2 2n(2n + 1)
∞
X α2k+2  (n) (n)
 1 − 2ν  (n) (n)

+ 2k+4
ωk+1 − ω k C k + %gh ω 1 − ω 0 = 0. (3.35)
k=0
h 5(1 − ν)

Ãn
Resolution by a system of equations: Defining the coefficients A0n = %gh2n+4
and
Cn
Cn0 = %gh2n+5
, we can express the problem as the following system of equations

 α2n (2n + 1)(2n + 2)A0n + β2n−2 2n(2n − 1)A0n−1 = Dn1 ,
(3.36)
α2n (2n + 2)A0n + α2n−1 (2n − 1)A0n−1 = Dn2 ,


where
  ∞ 
(n) (n) (n)
X
Dn1 = C̃ 2(2 − ν)ω0 − (1 − 2ν)δ0 + Ck0 (2k + 2)α2k+2 ωk
k=0
 (3 − ν)ω (n) − 3(1 − 2ν)ω (n)
(n) 0 1
−(2k + 3)β2k+1 ωk+1 − , (3.37)
5(1 − ν)
  X ∞
2 (4n + 1) (n) 0

(n) (n)

Dn = C̃(1 − 2ν) − P2n (0) − ω0 + Ck α2k+2 ωk − ωk+1
2n(2n + 1) k=0

(1 − 2ν)  (n) (n)


− ω1 − ω0 . (3.38)
5(1 − ν)

This is equivalent to the matrix system

    
0 1
2n(2n − 1)β2n−2 (2n + 1)(2n + 2)α2n A D
   n−1  =  n  , (3.39)
(2n − 1)α2n−1 (2n + 2)α2n A0n Dn2
 28

whose determinant is ∆n = (2n − 1)(2n + 2)α2n (2nβ2n−2 − (2n + 1)α2n−1 ). The solutions
(A0n and A0n−1 ) for this system of equations are
(n) (n)
(2n + 2)α2n (2n + 4 − (4n + 3)ν)ω0 (1 − 2ν)(2n + 4)ω1
A0n−1 = − +
∆n 5(1 − ν) 5(1 − ν)
 
(n) (4n + 1)
+ (2n + 5 − 4ν(n + 1))ω0 + (1 − 2ν) P2n (0) C̃
2n
∞ h
!
i
(n) (n)
X
α2k+2 (2k−2n+1)ωk + α2k+2 (2n+1) − β2k+1 (2k + 3) ωk+1 Ck0 ,


+ (3.40)
k=0

(n) (n)
(2n−1) (1−2ν)2nβ2n−2 + (3 − ν)α2n−1 ω0 (1 − 2ν)(2nβ2n−2 + 3α2n−1 )ω1
A0n = −
∆n 5(1 − ν) 5(1 − ν)
  

(n) (n)

(n) (4n + 1)
+ α2n−1 (1 − 2ν)δ0 − 2(2 − ν)ω0 −2nβ2n−2 (1−2ν) ω0 + P2n (0) C̃ 0
2n(2n+1)
∞ h
X
(n)
+ α2k+2 2nβ2n−2 − (2k + 2)α2n−1 ωk
k=0
!
i
(n)
Ck0


+ (2k + 3)α2n−1 β2k+1 − 2nβ2n−2 α2k+2 ωk+1 . (3.41)
 29

By evaluating (3.40) at (n + 1) and combining with (3.41), we obtain

  
(n+1) (4n + 5)
∆n (2n + 4)α2n+2 (2n + 7 − 4ν(n + 2))ω0 + (1 − 2ν) P2n+2 (0)
(2n + 2)
  
(n) (n)
− ∆n+1 (2n − 1) α2n−1 (1 − 2ν)δ0 − 2(2 − ν)ω0
 
(n) (4n + 1)
−2nβ2n−2 (1 − 2ν) ω0 + P2n (0) C̃ 0
2n(2n + 1)
∞
"
h
(n+1)
X
+ ∆n (2n + 4)α2n+2 α2k+2 (2k − 2n − 1)ωk
k=0

(n+1) i
+ (2n + 3)α2k+2 − (2k + 3)β2k+1 ωk+1
h
(n)
− ∆n+1 (2n − 1) α2k+2 2nβ2n−2 − (2k + 2)α2n−1 ωk
#

(n) i
+ (2k + 3)α2n−1 β2k+1 − 2nα2k+2 β2n−2 ωk+1 Ck0
"
−1 
(n+1) (n+1)

= (2n + 4)α2n+2 ∆n (1 − 2ν)(2n + 6)ω1 − (2n + 6 − (4n + 7)ν)ω0
5(1−ν)

(n)
− (2n − 1)∆n+1 (1 − 2ν)(2nβ2n−2 + 3α2n−1 )ω1
#

(n)
−((1 − 2ν)2nβ2n−2 + (3 − ν)α2n−1 )ω0 . (3.42)

By truncating the infinite series in (3.42) at a finite order N , we get a linear system of
equations for coefficients C̃ 0 , C00 , C10 , . . . , CN0 . Once we have solved this system, we can
obtain the solution to the displacements and stresses with the expressions (3.21) to (3.26).

3.3.2. Loaded boundary

In this case we take σr (v) = g1 (cos φ) and σrφ (v) = g2 (cos φ). We decompose these
functions in terms of the basis of Legendre polynomials, and call the associated coefficients
l2n and m2n , in the case that the basis is {P2n }∞ 0 ∞
n=0 or {sin φP2n }n=0 .
 30

In the case of the first boundary condition (which involves g1 ), we have that
∞ 
1 − 2ν 2(2 − ν)
  X (2n + 1)(2n + 2)
2
P0 + 2
P1 B̃ + [α2n P2n + β2n P2n+2 ]Ãn
h h n=0
h2n+3
∞
)
X (2k + 1)(2k + 3)P2k (0) h (n) (n)
i
+ −α2k+2 (2k + 2)ωk + β2k+1 (2k + 3)ωk+1 P2n B̃k
k=0
h2k+4
∞
X
= l2n P2n . (3.43)
n=0

Rearranging the terms appropiately, introducing a Kronecker delta, and defining the coef-
ficients C̃ = B̃ and Ck = (2k + 1)(2k + 3)P2k (0)B̃k , gives us

2n(2n − 1) (2n + 1)(2n + 2) 1 − 2ν (n) 2(2 − ν)ω0 (n)

2n+1
β2n−2 Ãn−1 + 2n+3
α2n Ãn + 2
δ0 C̃ − C̃
h h h h2
∞
X 1 h (n) (n)
i
+ 2k+4
−α 2k+2 (2k + 2)ω k + β2k+1 (2k + 3)ωk+1 Ck − l2n = 0 (3.44)
k=0
h

If we proceed in the same way as before, then the second boundary condition is
 
(2n−1)α2n−1 (2n + 2)α2n (1−2ν) (n) (4n + 1)
Ãn−1 + Ãn + ω0 + P2n (0) C̃
h2n+1 h2n+3 h2 2n(2n + 1)
∞
X α2k+2  (n) (n)

+ 2k+4
ωk+1 −ω k Ck − m2n = 0. (3.45)
k=0
h

Ãn
Resolution by a system of equations We make the change of variables A0n = h2n+3
,
Cn C̃
Cn0 = h2n+4
and C̃ 0 = h2
. In order to find the stresses, we must solve the system of
equations generated by (3.44) and (3.45)

 α2n (2n + 1)(2n + 2)A0n + β2n−2 2n(2n − 1)A0n−1 = Dn1 ,
(3.46)
α2n (2n + 2)A0n + α2n−1 (2n − 1)A0n−1 = Dn2 ,

 31

where

(n) (n)
Dn1 = l2n − (1 − 2ν)δ0 C̃ 0 + 2(2 − ν)ω0 C̃ 0
∞ h i
(n) (n)
X
+ α2k+2 (2k + 2)ωk − β2k+1 (2k + 3)ωk+1 Ck0 , (3.47)
k=0
  ∞
(n) (4n + 1) X 
(n) (n)

Dn2 = m2n − (1 − 2ν) ω0 + P2n C̃ 0 − α2k+2 ωk+1 − ωk Ck0 .
2n(2n + 1) k=0
(3.48)

Using matrix notation, this becomes

    
2n(2n − 1)β2n−2 (2n + 1)(2n + 2)α2n A0n−1 Dn1
  =  (3.49)
(2n − 1)α2n−1 (2n + 2)α2n A0n Dn2

where the determinant is ∆n = (2n − 1)(2n + 2)α2n (2nβ2n−2 − (2n + 1)α2n−1 ). Using
this we can find the solutions A0n and A0n−1 , which are given by

(2n + 2)α2n
Dn1 − (2n + 1)Dn2 ,


An−1 =
∆n
(2n − 1)
−α2n−1 Dn1 + 2nβ2n−2 Dn2 .


An = (3.50)
∆n
 32

This is equivalent to

(2n + 2)α2n
A0n−1 = l2n − (2n + 1)m2n
∆n
 
(n) (4n + 1)
+ (2n + 5 − 4ν(n + 1))ω0 + (1 − 2ν) P2n (0) C̃ 0
2n
∞ h 
X (n)
(n) i 0
+ α2k+2 (2k− 2n+1)ωk + α2k+2 (2n+1)−β2k+1 (2k+3) ωk+1 Ck ,
k=0
(3.51)
(2n − 1)

A0n = − l2n α2n−1 + 2nm2n β2n−2
∆n
h  
(n) (n)
+ α2n−1 (1 − 2ν)δ0 − 2(2 − ν)ω0
 
(n) (4n + 1)
−2nβ2n−2 (1 − 2ν) ω0 + P2n (0) C̃ 0
2n(2n + 1)
∞
Xh
(n)
+ α2k+2 2nβ2n−2 − (2k + 2)α2n−1 ωk
k=0


(n) i 0
+ (2k + 3)α2n−1 β2k+1 − 2nα2k+2 β2n−2 ωk+1 Ck . (3.52)

Evaluating (3.51) at (n + 1), we find that

  
∆n (n+1) (4n + 5)
α2n+2 (2n + 7 − 4ν(n + 2))ω0 + (1 − 2ν) P2n+2 (0)
2n − 1 (2n + 2)
  
∆n+1 
(n) (n)

(n) (4n + 1)
− α2n−1 (1−2ν)δ0 −2(2−ν)ω0 −2nβ2n−2 (1−2ν) ω0 + P2n C̃ 0
2n + 4 2n(2n+1)
∞
"
X ∆n h
(n+1)
(n+1) i
+ α2n+2 α2k+2 (2k− 2n −1)ωk + (2n +3)α2k+2 −(2k+3)β2k+1 ωk+1
2n − 1
k=0
#
∆n+1 h
(n)
(n) i 0
− α2k+2 2nβ2n−2 −(2k+2)α2n−1 ωk + (2k+3)α2n−1 β2k+1 −2nα2k+2 β2n−2 ωk+1 Ck
2n+4
   
∆n ∆n+1
=− α2n+2 l2n+2 − (2n + 3)m2n+2 + l2n α2n−1 − 2nm2n β2n−2 . (3.53)
2n − 1 2n + 4

By truncating the infinite series in (3.53) at a finite order N , we get a linear system of
equations for coefficients C̃ 0 , C00 , C10 , . . . , CN0 . Once we have solved this system, we can
obtain the solution to the displacements and stresses with the expressions (3.21) to (3.26).
 33

3.3.3. Zero displacement

If we take u = 0, then the change of variables given in (2.10), becomes

1 − 2ν
 
2µvr (r, φ) = − %gh2 cos3 φ,
2 − 2ν
1 − 2ν
 
2µvφ (r, φ) = %gh2 sin φ cos2 φ. (3.54)
2 − 2ν

Due to linearity, it is natural to replace the boundary conditions on Γh by the simpler ones
2µvr = − cos3 φ and 2µvφ = sin φ cos2 φ, and once the problem is solved, the resulting
1−2ν
%gh2 . If we write these conditions in


solution must be multiplied by the constant 2−2ν
terms of the Legendre basis, we obtain that
1
2µvr (r, φ) = − (3P1 + 2P3 )
5
∞ 
3 X (n) (n)

= ω0 − ω1 P2n , (3.55)
5 n=0
d cos3 φ
 
d 1
2µvφ (r, φ) = sin φ = sin φ (3P1 + 2P3 )
dφ 3 dφ 5
∞
sin φ X  (n) (n)

0
=− ω0 − ω1 P2n . (3.56)
5 n=0

Substituting (3.21) in (3.55),exchanging the summations, collecting the terms accord-

ing to the order of Legendre polynomials and introducing a Kronecker delta, we arrive
at
∞
1 − 2ν (n) 4(1 − ν) (n)
   
X 3 (n) (n)
 (2n + 1)
ω − ω1 + α2n 2n+2 Ãn + δ0 − ω0 B̃
n=0
5 0 h h h
∞
#
X (2k + 1)P2k (0)  (n) (n)

− (2k + 2)α2k+2 ωk − (2k + 3)2 (2k + 6 − 4ν)ωk+1 B̃k P2n
k=0
h2k+3
!
(2n + 1)(2n + 2)(2n + 5 − 4ν)
+ Ãn P2n+2 = 0. (3.57)
h2n+2
 34

By using the orthogonality of the basis {P2n }∞

n=0 in the range [−1, 0], this condition be-

comes
(2n + 1) 2n(2n − 1)(2n + 3− 4ν)
2n+2
α2n
Ãn + Ãn−1
h h2n
∞
X (2k + 1)P2k (0)  (n) 2 (n)

− (2k + 2)α2k+2 ωk − (2k + 3) (2k + 6 − 4ν)ωk+1 B̃k
k=0
h2k+3
 
1− 2ν (n) 4(1− ν) (n) 3  (n) (n)

+ δ0 − ω0 B̃ + ω0 − ω1 = 0. (3.58)
h h 5
In the same way as we did before, we substitute (3.56) in (3.22), yielding
∞
"
(3 − 4ν) (n) (1 − 2ν)(4n + 1)P2n (0)
 
X α2n
à +
2n+2 n
ω0 − B̃
n=1
h h 2n(2n + 1)h
∞
X (2k + 1)P2k (0)  (n) (n)

+ (2k + 3)(2k − 1 + 4ν)ωk+1 − α2k+2 ωk B̃k
k=0
h2k+3
# !
1  (n) (n)

0 (2n + 1)(2n − 2 + 4ν) 0
− ω0 − ω1 sin φP2n + Ãn sin φP2n+2 = 0. (3.59)
5 h2n+2
0 ∞
By using the orthogonality of the basis {sin φP2n }n=1 in the range [−1, 0], the second
boundary condition becomes
(3 − 4ν) (n) (1 − 2ν)(4n + 1)P2n (0)
 
α2n
Ãn + ω0 − B̃
h2n+2 h 2n(2n + 1)h
∞
X (2k + 1)P2k (0)  (n) (n)

+ (2k + 3)(2k − 1 + 4ν)ωk+1 − α2k+2 ωk B̃k
k=0
h2k+3
1  (n)  (2n − 1)(2n − 4 + 4ν)
(n)
− ω0 − ω1 + Ãn−1 = 0. (3.60)
5 h2n

Ãn
Resolution by a system of equations: If we make the change of variables A0n = h2n+2
,
(2n+1)P2n (0)B̃n B̃
Cn0 = h2n+3
and C̃ 0 = h
, then the system we must solve is

 α2n (2n + 1)Ã0n + 2n(2n − 1)(2n + 3 − 4ν)Ã0n−1 = Dn1 ,
(3.61)
α2n Ã0n + (2n − 1)(2n − 4 + 4ν)Ã0n−1 = Dn2 ,

 35

where
  ∞ 
(n) (n) (n)
X
0
Dn1 = 4(1 − ν)ω0 − (1 − 2ν)δ0 C̃ + (2k + 2)α2k+2 ωk
k=0

(n) 3  (n)

(n)

−(2k + 3)2 (2k + 6 − 4ν)ωk+1 Ck0 − ω0 − ω1 , (3.62)
5
  ∞ 
2 (n) (4n + 1) 0
X (n)
Dn = − (3 − 4ν)ω0 − (1 − 2ν) P2n (0) C̃ + α2k+2 ωk
2n(2n + 1) k=0

(n)
 1 
(n) (n)

−(2k + 3)(2k − 1 + 4ν)ωk+1 Ck0 + ω − ω1 . (3.63)
5 0
Using matrix notation, this becomes
    
0 1
2n(2n − 1)(2n + 3 − 4ν) (2n + 1)α2n A D
   n−1  =  n  , (3.64)
(2n − 1)(2n − 4 + 4ν) α2n A0n Dn2

where the determinant is ∆n = 4α2n (2n − 1)((3 − 4ν)n + 1 − ν). The solutions A0n and
A0n−1 are

α2n
2(n + 2)  (n) (n)
 h
(n)
A0n−1 = ω1 − ω0 + (2(3 − 4ν)n + 7 − 8ν)ω0
∆n 5
 ∞ h
(4n + 1) X (n)
+(1 − 2ν) P2n (0) C̃ 0 + (2k − 2n + 1)α2k+2 ωk
2n k=0
!

(n) i 0
+(2k+3) (2n + 1)(2k − 1 + 4ν) − (2k + 3)(2k + 6 − 4ν) ωk+1 Ck , (3.65)
 36

(2n − 1) 4 
(n) (n)

A0n =− (n(n + 3 − 2ν) − 3(1 − ν)) ω1 − ω0
∆n 5
  
(n) (4n + 1)
− (1 − 2ν) (2n − 4 + 4ν)δ0 + (2n + 3 − 4ν) P2n (0)
(2n + 1)
i
(n)
−2(2n − 1) (3 − 4ν)n + 8(1 − ν)2 )ω0 C̃ 0
∞ h
(n)
X

+ − 2n(2n − 2k + 1 − 4ν) + 8(k + 1)(1 − ν) α2k+2 ωk
k=0
!

(n) i 0
+2(2n−2k−3)(2k+3) n(2k−1+4ν) − 4(k+ 3 − 2ν)(1− ν) ωk+1 Ck . (3.66)

We evaluate (3.65) at (n + 1) and obtain

2(n + 3)  (n+1) (n+1)

 h
(n+1)
α2n+2 ∆n ω1 − ω0 + (2(3 − 4ν)(n + 1) + 7 − 8ν)ω0
5
 ∞ h
(4n + 5) 0
X (n+1)
+(1 − 2ν) P2n+2 (0) C̃ + (2k − 2n − 1)α2k+2 ωk
(2n + 2) k=0
!

(n+1) i 0
+(2k + 3) (2n + 3)(2k − 1 + 4ν) − (2k + 3)(2k + 6 − 4ν) ωk+1 Ck

4 
(n) (n)

+ (2n − 1)∆n+1 (n(n + 3 − 2ν) − 3(1 − ν)) ω1 − ω0
5
  
(n) (4n + 1)
− (1 − 2ν) (2n − 4 + 4ν)δ0 + (2n + 3 − 4ν) P2n (0)
(2n + 1)
i
(n)
−2(2n − 1) (3 − 4ν)n + 8(1 − ν)2 )ω0 C̃ 0
∞ h
X
(n)
+ 2(2n − 2k − 3)(2k + 3) n(2k−1 + 4ν) − 4(k+ 3 − 2ν)(1− ν) ωk+1
k=0
!
i
(n)
− 2n(2n − 2k + 1 − 4ν) + 8(k + 1)(1 − ν) α2k+2 ωk Ck0


= 0. (3.67)

By truncating the infinite series in (3.67) at a finite order N , we get a linear system of
equations for coefficients C̃ 0 , C00 , C10 , . . . , CN0 . Once we have solved this system, we can
obtain the solution to the displacements and stresses with the expressions (3.21) to (3.26).
 37

3.3.4. Prescribed displacements

In this case we use as boundary conditions 2µv r = g1 (cos φ) and 2µv φ = g2 (cos φ).
We decompose these functions in terms of the basis of Legendre polynomials as we did
before in Section 3.3.2. Then, the first boundary condition is given by
∞
1 − 2ν (n) 4(1 − ν) (n)
  
X (2n + 1)
l2n + α2n 2n+2 Ãn + δ0 − ω0 B̃
n=0
h h h
∞
#
X (2k + 1)P2k (0)  (n) 2 (n)

− (2k + 2)α2k+2 ωk − (2k + 3) (2k + 6 − 4ν)ωk+1 B̃k P2n
k=0
h2k+3
!
(2n + 1)(2n + 2)(2n + 5 − 4ν)
+ Ãn P2n+2 = 0. (3.68)
h2n+2

The use of the orthogonality of the basis {P2n }∞

n=0 in the range [−1, 0], turns this condition

in
1 − 2ν (n) 4(1 − ν) (n) 2n(2n− 1)(2n+ 3 − 4ν)
 
(2n+1)α2n
l2n + δ0 − ω0 B̃ + 2n
Ãn−1 + Ãn
h h h h2n+2
∞ 
(n) (2k+1)

(n)
X
+ (2k+ 3)2 (2k+6−4ν)ωk+1 −(2k+2)α2k+2 ωk 2k+3
P2k (0)B̃k = 0. (3.69)
k=0
h

Now we use the condition for vφ , which gives us

∞
"
(3 − 4ν) (n) (1 − 2ν)(4n + 1)P2n (0)
 
X α2n
m2n + 2n+2 Ãn + ω0 − B̃
n=1
h h 2n(2n + 1)h
∞ 
#
(2k + 1)P2k (0)   
(n) (n)
X
0
+ 2k+3
(2k + 3)(2k − 1 + 4ν)ωk+1 − α2k+2 ωk B̃k sin φP2n
k=0
h
!
(2n + 1)(2n − 2 + 4ν) 0
+ Ãn sin φP2n+2 = 0. (3.70)
h2n+2
 38

0
Using the orthogonality of the basis {sin φP2n (cos φ)}∞n=0 in [−1, 0], gives us

(2n−1)(2n − 4 + 4ν) (3 − 4ν) (n) (1 − 2ν)(4n + 1)P2n (0)

 
α2n
Ãn + Ãn−1 + ω0 − B̃
h2n+2 h2n h 2n(2n + 1)h
∞  
X (2k + 1)P2k (0)  (n) (n)

+ 2k+3
(2k + 3)(2k − 1 + 4ν)ωk+1 − α2k+2 ωk B̃k + m2n = 0.
k=0
h
(3.71)

Ãn
Resolution by a system of equations: We make the change of variables A0n = h2n+2
,
(2n+1)P2n (0)B̃n B̃
Cn0 = h2n+3
and C̃ 0 = h
, which gives us

 α2n (2n + 1)Ã0n + 2n(2n − 1)(2n + 3 − 4ν)Ã0n−1 = Dn1 ,
(3.72)
α2n Ã0n + (2n − 1)(2n − 4 + 4ν)Ã0n−1 = Dn2 ,


where
 
(n) (n)
Dn1 = 4(1 − ν)ω0 − (1 − 2ν)δ0 C̃ 0 − l2n
∞  
(n) (n)
X
− (2k+ 3)2 (2k+6−4ν)ωk+1 −(2k+2)α2k+2 ωk C̃k0 , (3.73)
k=0
 
(n) (4n + 1)
Dn2 = − (3 − 4ν)ω0 − (1 − 2ν) P2n (0) C̃ 0 − m2n
2n(2n + 1)
∞  
(n) (n)
X
+ α2k+2 ωk − (2k+3)(2k−1+4ν)ωk+1 C̃k0 . (3.74)
k=0

By writing this in matrix form, we have

    
0 1
2n(2n − 1)(2n + 3 − 4ν) (2n + 1)α2n A D
   n−1  =  n  , (3.75)
(2n − 1)(2n − 4 + 4ν) α2n A0n Dn2
 39

where the determinant is ∆n = 4α2n (2n − 1)((3 − 4ν)n + 1 − ν). Using this we find the
solutions A0n and A0n−1 , which are given by

α2n
A0n−1 = (2n + 1)m2n − l2n
∆n
 
(n) (4n + 1)
+ (2(3 − 4ν)n + 7 − 8ν)ω0 − (1 − 2ν) P2n (0) C̃ 0
2n
∞ h
(n)
X
+ (2k−2n+1)α2k+2 ωk
k=0
!

(n) i 0
+ (2k+3) (2n + 1)(2k−1+4ν)− (2k+3)(2k + 6 − 4ν) ωk+1 Ck , (3.76)

(2n − 1)
A0n = − 2n(2n + 3 − 4ν)m2n − 2(n + 2ν − 2)l2n
∆n
  
(n) (4n + 1)
− (1 − 2ν) (2n − 4 + 4ν)δ0 + (2n + 3 − 4ν) P2n (0)
(2n + 1)
i
(n)
−2(2n − 1) (3 − 4ν)n + 8(1 − ν)2 )ω0 C̃ 0
∞ h
(n)
X

+ − 2n(2n − 2k + 1 − 4ν) + 8(k + 1)(1 − ν) α2k+2 ωk
k=0
!

(n) i 0
+2(2n− 2k− 3)(2k+3) n(2k−1 + 4ν) − 4(k+ 3 − 2ν)(1− ν) ωk+1 Ck . (3.77)
 40

Finally, we evaluate (3.76) at (n + 1) and obtain

h
(n+1)
α2n+2 ∆n (2n + 3)m2n+2 − l2n+2 + (2(3 − 4ν)(n + 1) + 7 − 8ν)ω0
 ∞ h
(4n + 5) 0
X (n+1)
−(1 − 2ν) P2n+2 (0) C̃ + (2k − 2n − 1)α2k+2 ωk
(2n + 2) k=0
!

(n+1) i 0
+(2k + 3) (2n + 3)(2k − 1 + 4ν) − (2k + 3)(2k + 6 − 4ν) ωk+1 Ck

+ (2n − 1)∆n+1 − 2(n + 2ν − 2)l2n + 2n(2n + 3 − 4ν)m2n

  
(n) (4n + 1)
− (1 − 2ν) (2n − 4 + 4ν)δ0 + (2n + 3 − 4ν) P2n (0)
(2n + 1)
i
(n)
−2(2n − 1) (3 − 4ν)n + 8(1 − ν) )ω0 C̃ 0
2

∞ h
X
(n)
+ 2(2n − 2k − 3)(2k + 3) n(2k−1 + 4ν) − 4(k+ 3 − 2ν)(1− ν) ωk+1
k=0
!
i
(n)
− 2n(2n − 2k + 1 − 4ν) + 8(k + 1)(1 − ν) α2k+2 ωk Ck0


= 0. (3.78)

By truncating the infinite series in (3.78) at a finite order N , we get a linear system of
equations for coefficients C̃ 0 , C00 , C10 , . . . , CN0 . Once we have solved this system, we can
obtain the solution to the displacements and stresses with the expressions (3.21) to (3.26).

3.4. Accelerating the convergence of series

In each of the above cases (for the stresses see (3.42), (3.53) and for the displacements
see (3.67), (3.78)), after we have truncated the series at a finite order N , we get a system
of N + 2 equations and N + 2 unknows for the coefficients Ck0 , which can be solved by
some standard procedure to that purpose. To determine the stress values (and the displace-
ment values), it is necessary to find the values of coefficients A0n , but these series have
a slow convergence. To overcome this difficulty, we use the technique of transformation
of Kummer for series (see (Abramowitz & Stegun, 1965)), i.e., we obtain and calculate
 41

numerically the dominant asymptotic term for (3.40) (respectively for (3.51), (3.65) and
(3.76)). The values of these key terms are used to estimate the value of the series for A0n ,
by subtracting the asymptotic term in order to accelerate convergence, and once we have
the value of the series, we must add the value of the asymptotic term. To do this, we write
the term An−1 in the following way
∞
X
An−1 = Fn + G(k)
n Ck , (3.79)
k=−1

(k)
where Fn and Gn will depend on the case under consideration. We define s to be the
contribution to the general solution given by the terms that multiply the coefficients An ,
that is,
∞
X
s= An−1 sn−1 . (3.80)
n=1

Using the decomposition of An−1 , we can isolate the three contributions

∞ ∞
" #
X X
s= Fn + Gn(k) Ck sn−1
n=1 k=−1
∞
X ∞ X
X ∞
 (k) 
= Fn sn−1 + Gn sn−1 Ck
n=1 k=−1 n=1
∞
X ∞
X ∞
∞ X
 (−1)  0 X  (k) 
= Fn sn−1 + Gn sn−1 C̃ + Gn sn−1 Ck
n=1 n=1 k=0 n=1
∞
X
= F + Q−1 + Qk , (3.81)
k=0

where the terms sn−1 depend on the stresses (or displacements) that we are considering.
We review the terms of the stresses given in (3.21)-(3.26) and evaluated in n = n − 1.
For the sake of convenience we use the dimensionless parameter h/r instead 1/r in these
equations. Thus we have
 2n
h
− ((2n − 1)2 − 2 + 2ν)P2n−2 − 2n(2n + 3 − 4ν)P2n , (3.82)


sn−1 (vr ) = (2n − 1)
r
 42

 2n
h 0 0
((2n − 1)2 − 2 + 2ν)P2n−2


sn−1 (vφ ) = − sin φ + (2n − 1)(2n − 4 + 4ν)P2n ,
r
(3.83)
 2n+1
h
((2n − 1)2 − 2 + 2ν)P2n−2 + (2n(2n + 3) − 2ν)P2n ,


sn−1 (σr ) = 2n(2n − 1)
r
(3.84)
 2n+1 
h 0
sn−1 (σφ ) = (4n − 1)(2n + 2ν − 3)P2n−1
r

− 2n(2n − 1)(((2n − 1)2 − 2 + 2ν)P2n−2 + (4n2 − 6n + 3 − 2ν)P2n ) , (3.85)
 2n+1
h 0


sn−1 (σθ ) = −(4n − 1) 2n(2n − 1)(1 − 2ν)P2n + (2n − 3 + 2ν)P2n−1 , (3.86)
r
 2n+1
h 0 0
2n((2n − 1)2 − 2 + 2ν)P2n−2


sn−1 (σrφ ) = sin φ + (2n − 1)(2n(2n + 3) − 2ν)P2n .
r
(3.87)

3.4.1. Asymptotic behaviour of Legendre’s polynomials

For large n, the following asymptotic behaviour holds (see (Abramowitz & Stegun,
1965))
 − 12   
Γ(n + 1) 1 1 π
Pn (cos φ) = π sin φ cos n + φ− . (3.88)
Γ n + 23


2 2 4

We use one of the classical properties for the gamma function, namely
           √
3 1 1 1 1 π (2n)!
Γ n+ =Γ n+ +1 = n+ Γ n+ = n+ 2n−1
.
2 2 2 2 2 2 n!
(3.89)

With this, (3.88) becomes

− 12
22n−1 (n!)2
   
1 1 π
Pn (cos φ) =
√ π sin φ cos n + φ− . (3.90)
n + 12 π(2n)! 2 2 4

We use the Stirling formula, in order to approximate the factorials of high order, this gives
us
√ n n 2


1 2


(n!)2 2πn 1+ √
 
e 12n −2n 1 1
= √ = πn2 1+ + − ... . (3.91)
(2n)! 2n 2n 1 8n 576n2



4πn e
1+ 24n
 43

If we only keep the dominant term, then we have for a larger n the following expression
r   
2 1 π
Pn (cos φ) ≈ cos n + φ− . (3.92)
nπ sin φ 2 4

If we calculate dPn (x)/dx in (3.92), then we obtain for the derivatives of the Legendre
polynomials the following expression
  
cos φ 1 π
Pn0 (cos φ) = √ cos n + φ−
sin2 φ 2nπ sin φ 2 4
r     
2 1 1 1 π
+ n+ sin n + φ− . (3.93)
nπ sin φ 2 sin φ 2 4

Then, for large n we arrive at

r   
2n 1 1 3π
Pn0 (cos φ) ≈ cos n+ φ− . (3.94)
π sin φ sin φ 2 4

Using again Stirling’s formula, it is easy to see that for larger n we have
1
P2n (0) = (−1)n √ . (3.95)
πn

In the case we need to consider the approximation to second order for the Legendre poly-
nomials, we use

r     
2 3 1 π
Pn (cos φ) ≈ 1− cos n + φ−
nπ sin φ 8n 2 4
  
1 1 3 3π
+ cos n+ φ− , (3.96)
8n sin φ 2 4
 44

s "      
0 2 1 π 1 1 π
Pn (cos φ) ≈ − n cos n + φ+ − cos n + φ+
nπ sin3 φ 2 4 8 2 4
     
1 3 π 1 cos φ 1 π
+ cos n + φ− + cos n + φ−
8 sin φ 2 4 2 sin φ 2 4
     
1 3 1 π 1 cos φ 3 π
+ cos n + φ+ − cos n + φ+
n 16 2 4 8 sin2 φ 2 4
  !
3 1 3 π
− cos n + φ−
16 sin φ 2 4
      #
1 cos φ 1 π 1 3 π
− 3 cos n+ φ − + cos n+ φ + . (3.97)
16 n sin φ 2 4 sin φ 2 4

This second order approximation in the case φ = π/2 gives us

(−1)n
 
1
P2n (0) ≈ √ 1− . (3.98)
nπ 8n
These approximations are valid for φ 6= π. In the case φ = π, we have that
(−1)n
Pn (cos π) = (−1)n , y Pn0 (cos π) = − n(n + 1). (3.99)
2

Using (3.92) and (3.94), we obtain the following asymptotic behaviour of the terms sn−1 in
each case of interest
 2n  
(2n)3
    
h 3 π 1 π
sn−1 (vr ) = − √ cos 2n− φ− +cos 2n+ φ−
nπ sin φ r 2 4 2 4
2n
(2n)3 i − 1  − 3 iφ
   
h i

=√ Re √ e 2 + e 2 φ e2inφ , (3.100)
nπ sin φ r 2
 2n  
(2n)3
    
h 3 π 1 π
sn−1 (vφ ) = √ cos 2n− φ+ +cos 2n+ φ+
nπ sin φ r 2 4 2 4
2n
(2n)3
   
h 1 + i  − 3 iφ i

=√ Re √ e 2 + e 2 φ e2inφ , (3.101)
nπ sin φ r 2
 2n+1  
(2n)4
    
h 3 π 1 π
sn−1 (σr ) = √ cos 2n− φ− +cos 2n+ φ−
nπ sin φ r 2 4 2 4
2n+1
(2n)4 1 − i  − 3 iφ
   
h i

=√ Re √ e 2 + e 2 φ e2inφ , (3.102)
nπ sin φ r 2
 45

 2n+1  
−(2n)4
    
h 3 π 1 π
sn−1 (σφ ) = √ cos 2n− φ− +cos 2n+ φ−
nπ sin φ r 2 4 2 4
2n+1
(2n)4 i − 1  − 3 iφ
   
h i
φ

=√ Re √ e 2 +e 2 e2inφ , (3.103)
nπ sin φ r 2
 2n+1 
−2(2n)3 h
  
1 π
sn−1 (σθ ) = √ (1 − 2ν) cos 2n+ φ−
nπ sin φ r 2 4
  
1 1 3π
+ cos 2n+ φ− , (3.104)
sin φ 2 4
 2n+1 
(2n)4
    
h 3 3π 1 3π
sn−1 (σrφ ) = √ cos 2n− φ− +cos 2n+ φ−
nπ sin φ r 2 4 2 4
2n+1
(2n)4
   
h 1 + i  − 3 iφ i

=√ Re − √ e 2 + e 2 φ e2inφ . (3.105)
nπ sin φ r 2

In the case of φ = π, this becomes

 2n
h 3
sn−1 (vr ) = −16n , (3.106)
r
sn−1 (vφ ) = 0, (3.107)
 2n+1
h
4
sn−1 (σr ) = 32n , (3.108)
r
 2n+1
4 h
sn−1 (σφ ) = −(2n) , (3.109)
r
 2n+1
4 h
sn−1 (σθ ) = −(2n) , (3.110)
r
sn−1 (σrφ ) = 0. (3.111)

(k)
We just need to find the value of Fn and Gn to calculate the value of the coefficients.

3.4.2. Traction-free boundary and loaded boundary

We write the term An−1 from the equation (3.40) and (3.51) in the following way:
∞
X
An−1 = Fn + G(k)
n Ck , (3.112)
k=−1
 46

where we have two possible expressions for Fn depending on whether it is free or loaded
boundary
" #
(n) (n)
(2n + 2)α2n (2n + 4 − (4n + 3)ν)ω0 (1 − 2ν)(2n + 4)ω1
Fn = − , (3.113)
∆n 5(1 − ν) 5(1 − ν)
(2n + 2)α2n
or Fn = [l2n − (2n + 1)m2n ] , (3.114)
∆n
 
(2n+2)α2n (n) (4n + 1)
G(−1)
n = (2n + 5 − 4ν(n + 1))ω0 + (1 − 2ν) P2n (0) , (3.115)
∆n 2n
(2n+2)α2n h (n)
G(k)
n = (2k−2n+1)α2k+2 ωk
∆n

(n) i
+ (2n + 1)α2k+2 − (2k + 3)β2k+1 ωk+1 . (3.116)

Then we have to calculate three different sums in the case of traction-free boundary, the
first one involves the free term (Fn ) and the other two (which we also use in the case of
loaded boundary) are
∞
X ∞
X
 (−1)
G(k)

Q−1 = Gn sn−1 C̃ and Qk = n sn−1 Ck k = 0, 1, . . . (3.117)
n=1 n=1

We need to find the asymptotic behaviour of the series, so we must find the dominant term
in each case, i.e.,
−ν 1 (−1)n
Fn ≈ √ , (3.118)
(1 − ν) 16n4 πn
1 (2 − ν) (−1)n
G(−1)
n ≈− √ , (3.119)
8 n4 πn
1 (4k + 5)(2k + 4 − ν) (−1)n
G(k)
n ≈ √ . (3.120)
8 n4 πn
 47

We use the following sums

∞  2n+1  3
X
n h 2inφ h 1
(−1) n e =− e2iφ  2 , (3.121)
r r h 2


n=1 1+ r
e2iφ
∞  2n+1  3
X
n h 2inφ h 1
(−1) e =− e2iφ
2 , (3.122)
n=1
r r h
1 + r e2iφ
∞  2n+1  2 !
(−1)n h
 
X h h
e2inφ = − ln 1 + e2iφ . (3.123)
n=1
n r r r

Then, the terms that we need include in the sums φ < π are
 2 !!
(1 − i) h  − 3 iφ
 
ν 1 i
 h 2iφ
F (σr ) = Re √ √ e 2 + e 2 φ ln 1 + e C̃ 0 ,
(1 − ν) π sin φ 2 r r
(3.124)
 2 !!
(i − 1) h  − 3 iφ
 
ν 1 i
 h 2iφ
F (σφ ) = Re √ √ e 2 +e 2
φ
ln 1 + e C̃ 0 ,
(1 − ν) π sin φ 2 r r
(3.125)
 2 !!
−ν
 
1 (i + 1) h  − 3 iφ i
 h 2iφ
F (σrφ ) = Re √ √ φ
e 2 +e 2 ln 1 + e C̃ 0 ,
(1 − ν) π sin φ 2 r r
(3.126)

 2 !!
(2 − ν) √ h  − 3 iφ
 
i
 h
Q−1 (σr ) = Re √ (1 − i) 2 e 2 + e 2 φ ln 1 + e2iφ C̃ 0 ,
π sin φ r r
(3.127)
 2 !!
(2 − ν) √ h  − 3 iφ
 
i
 h
Q−1 (σφ ) = Re √ (i − 1) 2 φ
e 2 + e 2 ln 1 + e2iφ C̃ 0 ,
π sin φ r r
(3.128)
 2 !!
(2 − ν) √ h  − 3 iφ
 
i
 h
Q−1 (σrφ ) = Re − √ (i + 1) 2 e 2 +e 2
φ
ln 1 + e2iφ C̃ 0 ,
π sin φ r r
(3.129)
 48

Qk (σr )
!
(4k+5)(2k−ν +4) √ h  − 3 iφ
 
i
  h2 
= Re √ (i−1) 2 e 2 + e 2 φ ln 1+ e2iφ Ck0 , (3.130)
π sin φ r r

Qk (σφ )
!
(4k+5)(2k−ν +4) √ h  − 3 iφ
 
i
  h2 
= Re √ (1−i) 2 e 2 + e 2 φ ln 1+ e2iφ Ck0 , (3.131)
π sin φ r r

Qk (σrφ )
!
(4k+5)(2k−ν +4) √ h  − 3 iφ
 
i
  h2 
= Re √ (i + 1) 2 e 2 + e 2 φ ln 1+ e2iφ Ck0 .
π sin φ r r
(3.132)

We note that the terms related to σθ , vr and vφ grow slowly, therefore, they are not included
in our acceleration technique. In the case of φ = π we have that
  s 
 2
h  0
Q−1 (vr ) = 4(2 − ν) ln(2) − ln1 + 1 + C̃ , (3.133)
r
 3
h 1
Q−1 (σr ) = 4(2 − ν) q C̃ 0 , (3.134)
r 2 2
1 + hr + 1 + hr

 3
h 1 0
Q−1 (σφ ) = −2(2 − ν) q
C̃ , (3.135)
r h 2

h 2
1+ r +1+ r
 3
h 1
Q−1 (σθ ) = −2(2 − ν) q C̃ 0 , (3.136)
r 2 2
1 + hr + 1 + hr

  s 
 2
h  0
Qk (vr ) = −4(4k + 5)(2k − ν + 4) ln(2) − ln1 + 1+ Ck , (3.137)
r
 3
h 1 0
Qk (σr ) = −4(4k + 5)(2k − ν + 4) q
Ck , (3.138)
r h

2 h 2
1+ r
+1+ r
 49

 3
h 1 0
Qk (σφ ) = 2(4k + 5)(2k − ν + 4) q
Ck , (3.139)
r h 2

h 2
1+ r
+1+ r
 3
h 1
Qk (σθ ) = 2(4k + 5)(2k − ν + 4) q Ck0 . (3.140)
r h 2

h 2


1+ r
+1+ r

3.4.3. Zero displacement and prescribed displacements

The term An−1 from the equation (3.65) gives us the contributions
 
α2n 2(n + 2)  (n) (n)
Fn = ω1 − ω0 , (3.141)
∆n 5
(4 n + 1) P2n (0)
or Fn = 1/4 , (3.142)
(1 + n) (2 n − 3) (−3 n + 4 ν n − 1 + ν) (2 n − 1)2
 
α2n (n) (4n + 1)
G(−1)
n = (2(3 − 4ν)n + 7 − 8ν)ω0 + (1 − 2ν) P2n , (3.143)
∆n 2n
α2n h
(n)
G(k)
n = (2k + 3) (2n + 1)(2k − 1 + 4ν) − (2k + 3)(2k + 6 − 4ν) ωk+1
∆n
i
(n)
+(2k − 2n + 1)α2k+2 ωk , (3.144)

whose dominant terms are

1
Fn ≈ − P2n (0), (3.145)
8(3 − 4ν)n4
1 (1 − ν)
G(−1)
n ≈− P2n (0), (3.146)
2 (3 − 4ν)n2
1 (4k + 5)(1 − ν)
G(k)
n ≈ P2n (0). (3.147)
2 (3 − 4ν)n2

Using these expressions we can determine the dominant term. For the sake of convenience,
we work in the complex plane, and subsequently we consider just the real part. For exam-
ple, for σr we have
√ s  2n+1
2 1 3 h

sn−1 (σr ) = − 3 (1 + i)n e 2inφ
−16ne−5/2iφ + 3e−5/2iφ
4 nπ sin φ r

+2e−1/2iφ + 16ne3/2iφ − e3/2iφ . (3.148)
 50

The asymptotic term are

√  2n+1
(−1)n
r  
2(1 + i) 1 h 1  5
Fn sn−1 = 1 − e 2inφ
e− 2 iφ (16n − 3)
32π(−3 + 4ν) sin φr n2 8n
1 3

−2e− 2 iφ − e 2 iφ (16n − 1) , (3.149)
√ r  2n+1  
2(1 + i) 1 h 1  5
G(−1)
n sn−1 =− n
(−1) 1 − e2inφ e− 2 iφ (16n − 3)
8π(−3 + 4ν) sin φ r 8n
1 3

−2e− 2 iφ − e 2 iφ (16n − 1) , (3.150)
√ r  2n+1  
2(1 + i) 1 h 1  5
Gn(k) sn−1 = (4k + 5) n
(−1) 1 − e2inφ e− 2 iφ (16n − 3)
8π(−3 + 4ν) sin φ r 8n
1 3

−2e− 2 iφ − e 2 iφ (16n − 1) . (3.151)

The following sums hold

∞  2n+1  3
X h n 2inφ h 1
(−1) e n = − e2iφ  2 , (3.152)
r r h 2


n=1 1+ r
e2iφ
∞  2n+1  3
X h n 2inφ h 1
(−1) e =− e2iφ , (3.153)
r r h 2 2iφ


n=1 1 + r
e
∞
!
X h   2n+1    2
1 h h
(−1)n e2inφ = − ln 1 + e2iφ , (3.154)
n=1
r n r r
   
2iφ h 2 h 2 2iφ h 2 2iφ




∞
X h  2n+1 2inφ e − 1 + e ln 1 + e
n e r r r
(−1) = h

. (3.155)
n=1
r n(n+1) r

3.5. Some clues about the programming

Once we have computed all the expressions for the displacements and the stresses, we
face the challenge of putting them in a computer code, in order to evaluate each expression
and plot the results. The way to do so is tricky, so we present next some of the solutions
we have used to solve those problems that every implementation usually has.

One of our biggest concerns was the implementation of the coefficients l2n , m2n from
(3.3.2), (3.3.4), because we need to calculate the integrals for the Legendre polynomials of
 51

degree from 0 to N , where N could take values up to 500, so we need a suitable method to
compute these integrals.

The first step is to determine the most general form that these integrals could have. We
use the following recurrence relation for the Legendre polynomials,

(1 − x2 )Pn0 (x) = n(Pn−1 (x) − xPn (x)), (3.156)

so we can write any of the integrals that will arise in the following way
Z 0
g(x)P2n (x)dx. (3.157)
−1

We use a partition of the interval [−1, 0] defined as x1 = −1 < x2 < . . . < xN +1 = 0, and
we take the Lagrange interpolation functions, that is

 x2 − φ

[x1 , x2 ]
φ1 (x) = x2 − x1 (3.158)
0 [x3 , xN +1 ]


φ − xj−1
[xj−1 , xj ]


xj − xj−1

φj (x) = xj+1 − φ j = 2, . . . , N (3.159)

 [xj , xj+1 ]
xj+1 − xj



 0 [x1 , xN −1 ]
φN +1 (x) = φ − x N (3.160)
 [xN , xN +1 ]
xN +1 − xN


With this particular partition, we use the approximation

N
X +1
g(x) ≈ g(xj )φj (x), (3.161)
j=1
N +1
!
Z 0 Z 0 X
g(x)P2n (x)dx ≈ g(xj )φj (x)P2n (x) dx, (3.162)
−1 −1 j=1
N
X +1 Z 0

≈ g(xj ) φj (x)P2n (x) dx. (3.163)
j=1 −1
 52

So, we express the integrals in the following ways

Z xj+1
xj+1 P2n+1 (xj+1 ) − xj P2n+1 (xj ) xj+1 P2n−1 (xj+1 ) − xj P2n−1 (xj )
xP2n (x)dx = −
xj 4n + 1 4n + 1
P2n+2 (xj+1 ) − P2n+2 (xj ) P2n (xj+1 ) − P2n (xj )
− +2
(4n + 1)(4n + 3) (4n − 1)(4n + 3)
P2n−2 (xj+1 ) − P2n−2 (xj )
− , (3.164)
(4n − 1)(4n + 1)
xj+1
P2n+1 (xj+1 ) − P2n+1 (xj ) P2n−1 (xj+1 ) − P2n−1 (xj )
Z
P2n (x)dx = − . (3.165)
xj 4n + 1 4n + 1

3.6. Numerical results

Γ∞
Γh

ΩR
Γs

ΓR

Figure 3.1. Boundaries.

We need a proper benchmark to test our results. For this purpose we choose the recog-
nized comercial software, COMSOL (see (Multiphysics, 1994)). We first need to evaluate
wheter the results given by COMSOL for our semi-infinite problem are trustable, because
it is necessary to use artificial boundaries in order to limit the domain. We use an axisym-
metric domain, with a hemispherical perturbation of radius h in the presence of the gravity
force. In order to determine a suitable size for the domain, we make the following exper-
iment: We consider domains consisting of squares of different sizes with a hemispherical
 53

perturbation of radius h = 600 m, and we solve our semi-infinite problem using COMSOL
in each one of these domains. The radii vary from a minimum value of 1000 m to a maxi-
mum value of 13000 m, in increments of 1000 m. The results obtained with the maximum
radius are taken as a reference. Let us call XY the results yielded by Comsol, and xy the
results yielded by our methodology. We use an interpolation between this sets of points in
order to measure the point by point error, this is
pP
2
i (Yi − yi )
Eboundary = pP . (3.166)
2
i (Yi )

The results in the important boundaries (see Figure 3.1) for both displacements are shown

0.35 0.5
Perturbation Perturbation
Transparency 0.45 Transparency
0.3
Semiplane Semiplane
Symmetry 0.4

0.25 0.35

0.3
0.2
Error

Error

0.25
0.15
0.2

0.1 0.15

0.1
0.05
0.05

0 0
2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000
Size m Size m

(a) Displacement ur (b) Displacement uφ

Figure 3.2. Displacement relative error between our method and the re-
sults yielded by COMSOL in a square of size 13000 m.

in 3.2a and 3.2b. According to these results, we choose a square of size 10000 m as bench-
mark.

3.6.1. Traction-free boundary

The Figures 3.3a to 3.4h show the results provided by our method and the benchmark
for the case of traction-free boundary condition on the hemisphere.

The numerical error associated with the traction-free boundary condition on each boundary
(see Figure 3.1) are shown in Table 3.1. These results allow us to validate our methodology.
 54

(a) Displacement ur m, with traction- (b) Displacement ur m, cal-

free boundary condition. culated with COMSOL

(c) Displacement uφ m, with traction- (d) Displacement uφ m, cal-

free boundary condition. culated with COMSOL

Figure 3.3. Displacement for traction-free boundary condition with ν = 0.3,

λ = 40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .

3.6.2. Loaded boundary

In this case we can obtain again the results presented in Figures 3.3a to 3.6h if we use
(3.30) and (3.31) as our boundary condition.
 55

Table 3.1. Comparison of the absolute error between the semi-analytical

solution and the model implemented in COMSOL in a square of side 10000
m.

Boundarie ur uφ
Plane Γ∞ 0.0092 0.0924
Perturbation Γh 0.0314 0.0789
Transparency ΓR 0.0348 0.1485
Axisymmetric Γs 0.0311 0

3.6.3. Zero displacement

The Figures 3.5a to 3.6h show the results provided by our method and the benchmark
in the case of zero displacement on the hemisphere.
Table 3.2. Comparison of the absolute error between the semi-analytical
solution and the model implemented in COMSOL in a square of side 10000
m.

ur uφ
Plane Γ∞ 0.3290 0.1080
Transparency ΓR 0.0114 0.1823
Axisymmetry Γs 0.0065 0

The relative error associated with the zero displacement boundary condition in each
of the boundaries (see 3.1) are shown in Table 3.2. These results allow us to validate our
methodology.

3.6.4. Prescribed displacements

In this case we obtain again the results of Figures 3.5a to 3.6g if we use (3.55) and
(3.56) as our boundary condition.
 56

(a) Stress σr Pa, with traction-free (b) Stress σr Pa, calcu-

boundary condition lated with COMSOL

(c) Stress σφ Pa, with traction-free (d) Stress σφ Pa, calcu-

boundary condition lated with COMSOL

Figure 3.4. Stresses for traction-free boundary condition with ν = 0.3, λ =

40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .
 57

(e) Stress σθ Pa, with traction-free (f) Stress σθ Pa, calculated

boundary condition with COMSOL

(g) Stress σrφ Pa, with traction-free (h) Stress σrφ Pa, calcu-
boundary condition lated with COMSOL

Figure 3.4. Stresses for traction-free boundary condition with ν = 0.3, λ =

40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .
 58

(a) Displacement ur m, with zero dis- (b) Displacement ur m, cal-

placement boundary condition. culated with COMSOL

(c) Displacement uφ m, with zero dis- (d) Displacement uφ m, cal-

placement boundary condition. culated with COMSOL

Figure 3.5. Displacement for zero displacement boundary condition with

ν = 0.3, λ = 40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .
 59

(a) Stress σr Pa, with zero displace- (b) Stress σr Pa, calcu-
ment boundary condition lated with COMSOL

(c) Stress σφ Pa, with zero displace- (d) Stress σφ Pa, calcu-
ment boundary condition lated with COMSOL

Figure 3.6. Stresses for zero displacement boundary condition with ν =

0.3, λ = 40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .
 60

(e) Stress σθ Pa, with zero displace- (f) Stress σθ Pa, calculated
ment boundary condition with COMSOL

(g) Stress σrφ Pa, with zero displace- (h) Stress σrφ Pa, calcu-
ment boundary condition lated with COMSOL

Figure 3.6. Stresses for zero displacement boundary condition with ν =

0.3, λ = 40.5 GPa, µ = 27 GPa, % = 2725 kg/m3 .
 61

4. AN EFFICIENT SEMI-ANALYTICAL METHOD TO COMPUTE DISPLACE-

MENTS

Throughout this chapter, we consider the same half-space with a hemispherical pit used
in the previous chapter. In what follows, we are going to impose the boundary conditions
on the pit by minimising a quadratic functional which represents the surface elastic energy.
Even though this approach does not eliminate the problem of an infinite number of equa-
tions and coefficients, it provides a systematic method to deal with it. The minimisation
of the functional leads to a linear system of equations for a finite number of coefficients of
the series. The associated matrix is symmetric, positive definite and possesses a particular
block structure that allows us to solve it in an efficient and robust way, thus avoiding to
have to deal with cumbersome equations and slowly convergent double series.

Furthermore, in order to simplify the calculations, we re-define the notation for the
displacements and stresses.

4.1. Series solution

Having already solved (2.27) and (2.28), we now resume the calculation of a solution
to (2.12) in series form. We have obtained potentials Φn , given in (2.31) for n ≥ 0 and in
(2.32) for n = −1, and potentials Ψn , given in (2.30) for n ≥ 0. Both kinds of potentials
may be combined in different ways in (2.25) or (2.26), giving rise to displacement fields v
satisfying (2.12a) and (2.12e). We shall consider two sets of such displacement fields. The
(1)
first set will consist of displacement fields v n , obtained by setting Φ = Φn and Ψ = 0 in
(2.25) for n ≥ −1, that is,
2µv (1)
n = ∇Φn , (4.1)
(2)
whereas the second set will consist of displacement fields v n , obtained by setting Ψ =
(2n + 1)Ψn and Φ = −(n − 4 + 4ν)Φn−1 in (2.25) for n ≥ 0, that is,

2µv (2)
 
n = −(n − 4 + 4ν)∇Φn−1 + (2n + 1) ∇(zΨn ) − 4(1 − ν)Ψn k̂ . (4.2)
 62

The reason for considering this particular combination of potentials Φn and Ψn is that, as
stated in (R. A. Eubanks, 1954), the forthcoming calculations are considerably simplified.
The general solution to Navier’s equation (2.11) corresponds to an infinite linear combina-
(1) (2)
tion of displacement fields v n and v n , that is,
∞
X ∞
X
v(r, φ) = a(1) (1)
n v n (r, φ) + a(2) (2)
n v n (r, φ), (4.3)
n=−1 n=0

(1) (1) (1) (2) (2) (2)

where a−1 , a0 , a1 , . . . and a0 , a1 , a2 , . . . are unknown real coefficients. Displacement
(1) (2)
fields v n and v n are also expressed in terms of their components in r and φ by employing
(2.26). Substituting these components in the Hooke’s law yields the components of the
(1) (2) (1) (2)
stress fields associated with v n and v n , which we denote respectively by σ n and σ n .
The stress tensor σ associated with v in (4.3) is thus expressed as
∞
X ∞
X
σ(r, φ) = a(1) (1)
n σ n (r, φ) + a(2) (2)
n σ n (r, φ). (4.4)
n=−1 n=0

Making all the necessary substitutions and rearranging the resulting expressions in
(1) (2)
each case, we obtain that the displacement fields v n and v n , together with their associated
(1) (2)
respective stress fields σ n and σ n , can be expressed as
1 1
v (1)
n (r, φ) = w(1)
n (φ), σ (1)
n (r, φ) = τ (1) (φ), n = −1, 0, 1, . . . (4.5a)
rn+2 rn+3 n
1 1
v (2) (2)
n (r, φ) = n+1 w n (φ), σ (2) (2)
n (r, φ) = n+2 τ n (φ), n = 0, 1, 2, . . . (4.5b)
r r
(1) (2) (1) (2)
where wn , wn are vector functions, and τ n , τ n are tensor functions, all of them de-
pending only on the angle φ. This is a convenient form of expressing the displacement and
stress fields, since it allows us to make explicit the dependence on r, separating it from the
(1) (1)
dependence on φ. The components of functions wn and τ n for n = −1 are

(1)
2µ[w−1 ]r (φ) = 1, (4.6a)
(1)
2µ[w−1 ]φ (φ) = q(φ) sin φ, (4.6b)
(1)
[τ−1 ]r (φ) = −1, (4.6c)
 63

(1)
[τ−1 ]φ (φ) = −q(φ) cos φ, (4.6d)
(1)
[τ−1 ]θ (φ) = q(φ), (4.6e)
(1)
[τ−1 ]rφ (φ) = −q(φ) sin φ, (4.6f)

where the function q(·) is defined as

1 π
q(φ) = , ≤ φ ≤ π. (4.7)
1 − cos φ 2
(1) (1)
The components of functions wn and τ n for n ≥ 0 are

2µ[wn(1) ]r (φ) = −(n + 1)Pn (cos φ), (4.8a)

2µ[wn(1) ]φ (φ) = − sin φ Pn0 (cos φ), (4.8b)

[τn(1) ]r (φ) = (n + 1)(n + 2)Pn (cos φ), (4.8c)

0
[τn(1) ]φ (φ) = Pn+1 (cos φ) − (n + 1)(n + 2)Pn (cos φ), (4.8d)
0
[τn(1) ]θ (φ) = −Pn+1 (cos φ), (4.8e)

[τn(1) ]rφ (φ) = (n + 2) sin φ Pn0 (cos φ). (4.8f)

(2) (2)
The components of functions wn and τ n for n = 0 are

(2)
2µ[w0 ]r (φ) = −4(1 − ν)(1 + cos φ), (4.9a)
(2)

2µ[w0 ]φ (φ) = − 4(1 − ν)q(φ) − 3 + 4ν sin φ, (4.9b)
(2)

[τ0 ]r (φ) = 2 2 − 2ν + (2 − ν) cos φ , (4.9c)
(2)
[τ0 ]φ (φ) = (4(1 − ν)q(φ) − 1 + 2ν) cos φ, (4.9d)
(2)
[τ0 ]θ (φ) = −4(1 − ν)q(φ) − (1 − 2ν) cos φ, (4.9e)
(2)

[τ0 ]rφ (φ) = 4(1 − ν)q(φ) − 1 + 2ν sin φ, (4.9f)

and for n ≥ 1 are

2µ[wn(2) ]r (φ) = −(n + 1)(n + 4 − 4ν)Pn+1 (cos φ), (4.10a)

 64

0
2µ[wn(2) ]φ (φ) = −(n − 3 + 4ν) sin φ Pn+1 (cos φ), (4.10b)

[τn(2) ]r (φ) = (n + 1) (n + 1)(n + 4) − 2ν Pn+1 (cos φ),

(4.10c)

[τn(2) ]φ (φ) = −(n + 1)(n2 − n + 1 − 2ν)Pn+1 (cos φ)

(4.10d)
+ (n − 3 + 4ν)Pn0 (cos φ),

[τn(2) ]θ (φ) = −(1 − 2ν)(n + 1)(2n + 1)Pn+1 (cos φ)

(4.10e)
− (n − 3 + 4ν)Pn0 (cos φ),
0
[τn(2) ]rφ (φ) = (n2 + 2n − 1 + 2ν) sin φ Pn+1 (cos φ). (4.10f)

Substituting (4.5) in (4.3) and (4.4), and grouping terms with the same power of r, we
obtain that v and σ are also expressed as
∞
X 1  (2) (2)

v(r, φ) = a(1) (1)
n w n (φ) + an+1 wn+1 (φ) , (4.11)
n=−1
rn+2
∞
X 1  (2) (2)

σ(r, φ) = a(1)
n τ (1)
n (φ) + a τ
n+1 n+1 (φ) . (4.12)
n=−1
rn+3

4.1.1. Traction-free boundary on Γ∞

As done in the previous chapter, we need to impose the traction-free boundary condi-
tion on the plane boundary (which is σφ (v) = σrφ (v) = 0), that is
∞
X 1  π 
(2) (2)
 π 
σφ (v) = a(1) (1)
n [τn ]φ + an+1 [τn+1 ]φ = 0, (4.13a)
n=−1
rn+3 2 2
∞
X 1  π 
(2) (2)
 π 
σrφ (v) = a(1) (1)
n [τn ]rφ + an+1 [τn+1 ]rφ = 0. (4.13b)
n=−1
rn+3 2 2

In order to find out under which conditions (4.13a) and (4.13b) hold, we evaluate (4.6d),
(4.6f), (4.8d), (4.8f), (4.9d), (4.9f), (4.10d) and (4.10f) at φ = π/2. The result varies
depending on whether n is even or odd, so we discriminate between these two cases before
 65

evaluating the expressions. In the even case, we arrive at

π 
(1)
[τ2n ]φ = −(2n + 1)2 P2n (0) n ≥ 0, (4.14a)
2
π 
(1)
[τ2n ]rφ = 0 n ≥ 0, (4.14b)
2
π 
(2)
[τ2n ]φ = 0 n ≥ 0, (4.14c)
2

π   3 − 2ν n = 0,
(2)
[τ2n ]rφ = (4.14d)
2 (2n + 1)α2n P2n (0) n ≥ 1,


and in the odd case, we obtain

π 
(1)
[τ2n+1 ]φ = 0 n ≥ −1, (4.15a)
2

π   −1 n = −1,
(1)
[τ2n+1 ]rφ = (4.15b)
2 (2n + 1)(2n + 3)P2n (0) n ≥ 0,

π 
(2)
[τ2n+1 ]φ = (2n + 1)α2n P2n (0) n ≥ 0, (4.15c)
2
π 
(2)
[τ2n+1 ]rφ = 0 n ≥ 0, (4.15d)
2
where
α2n = (2n + 1)2 − 2(1 − ν),

and having combined with the recurrence relations (6.17) as appropriate. It is easy to
verify that by virtue of (4.14) and (4.15), identities (4.13a) and (4.13b) hold, provided that
(1) (2)
the coefficients an and an satisfy the relations

(1) (2)
a−1 = (3 − 2ν)a0 , (4.16a)
(1) (2)
(2n + 1)a2n = α2n a2n+1 n ≥ 0, (4.16b)
(1) (2)
(2n + 2)a2n+1 = α2n+2 a2n+2 n ≥ 0. (4.16c)
 66

Thus, separating the infinite sums in (4.11) and (4.12) into even and odd terms, and com-
bining with (4.16), we arrive at the following expressions for v and σ:
(2)
a0  (1) (2)

v(r, φ) = (3 − 2ν)w−1 (φ) + w0 (φ)
r
∞ (2)
X a2n+2 
(1) (2)

+ α2n+2 w2n+1 (φ) + (2n + 2)w2n+2 (φ) (4.17a)
n=0
(2n + 2)r2n+3
∞ (2)
X a2n+1 
(1) (2)

+ α w
2n 2n (φ) + (2n + 1)w 2n+1 (φ) ,
n=0
(2n + 1)r2n+2
(2)
a0  (1) (2)

σ(r, φ) = (3 − 2ν)τ −1 (φ) + τ 0 (φ)
r2
∞ (2)
X a2n+2 
(1) (2)

+ α τ
2n+2 2n+1 (φ) + (2n + 2)τ 2n+2 (φ) (4.17b)
n=0
(2n + 2)r2n+4
∞ (2)
X a2n+1 
(1) (2)

+ α2n τ 2n (φ) + (2n + 1)τ 2n+1 (φ) .
n=0
(2n + 1)r2n+3

This solution satisfies (2.12b), as well as (2.12a) and (2.12e). Defining the vector functions

(1) (2)
w(A)
n (φ) = α2n w 2n (φ) + (2n + 1)w 2n+1 (φ) n ≥ 0, (4.18a)

(3 − 2ν)w(1)
 (2)
−1 (φ) + w 0 (φ) n = −1,
(B)
wn (φ) = (4.18b)
(1) (2)
n ≥ 0,

α
2n+2 w 2n+1 (φ) + (2n + 2)w 2n+2 (φ)

and the tensor functions

(1) (2)
τ (A)
n (φ) = α2n τ 2n (φ) + (2n + 1)τ 2n+1 (φ) n ≥ 0, (4.19a)

(3 − 2ν)τ (1)
 (2)
−1 (φ) + τ 0 (φ) n = −1,
τ (B)
n (φ) = (4.19b)
(1) (2)
n ≥ 0,

α
2n+2 τ 2n+1 (φ) + (2n + 2)τ 2n+2 (φ)
 67

(2)
and using the fact that the coefficients an are arbitrary, it is possible to reexpress (4.17a)
and (4.17b) as
∞
X  h 2n+2 ∞
X  h 2n+3
v(r, φ) = An w(A)
n (φ) + Bn w(B)
n (φ), (4.20a)
n=0
r n=−1
r
∞
X ∞ 
1  h 2n+3 X  h 2n+4
σ(r, φ) = An τ (A)
n (φ) + Bn τ n(B) (φ) , (4.20b)
h n=0
r n=−1
r

(A) (A)
where An and Bn are unknown real coefficients. The components of functions wn , τ n
for n = 0, 1, 2, . . . are

2µ[wn(A) ]r (φ) = −(2n + 1) α2n P2n (cos φ) + γ2n P2n+2 (cos φ) ,

(4.21a)
0 0
2µ[wn(A) ]φ (φ) = − sin φ α2n P2n


(cos φ) + 2n P2n+2 (cos φ) , (4.21b)

[τn(A) ]r (φ) = (2n + 1)(2n + 2) α2n P2n (cos φ) + β2n P2n+2 (cos φ) ,


(4.21c)
0
[τn(A) ]φ (φ) = (α2n + 2n )P2n+1 (cos φ) − (2n + 1)(2n + 2)

(4.21d)
× α2n P2n (cos φ) + (α2n − 2n + 2 − 4ν)P2n+2 (cos φ) ,

[τn(A) ]θ (φ) = −(4n + 3) (2n + 1)(2n + 2)(1 − 2ν)P2n+2 (cos φ)

(4.21e)
0


+ (2n − 1 + 2ν)P2n+1 (cos φ) ,
0 0
[τn(A) ]rφ (φ) = sin φ (2n + 2)α2n P2n


(cos φ) + (2n + 1)α2n+1 P2n+2 (cos φ) . (4.21f)

(B) (B)
In the case n = −1, the components of functions wn , τ n are

(B)

2µ[w−1 ]r (φ) = − 1 − 2ν + 4(1 − ν) cos φ , (4.22a)
(B)

2µ[w−1 ]φ (φ) = sin φ 3 − 4ν − (1 − 2ν)q(φ) , (4.22b)
(B)
[τ−1 ]r (φ) = 1 − 2ν + 2(2 − ν) cos φ, (4.22c)
(B)

[τ−1 ]φ (φ) = −(1 − 2ν) 1 + cos φ − q(φ) , (4.22d)
(B)

[τ−1 ]θ (φ) = −(1 − 2ν) cos φ + q(φ) , (4.22e)
(B)

[τ−1 ]rφ (φ) = −(1 − 2ν) sin φ 1 − q(φ) , (4.22f)
 68

whereas in the case n = 0, 1, 2, . . ., they are

2µ[wn(B) ]r (φ) = −(2n + 2) α2n+2 P2n+1 (cos φ) + γ2n+1 P2n+3 (cos φ) ,

(4.23a)
0 0
2µ[wn(B) ]φ (φ) = − sin φ α2n+2 P2n+1


(cos φ) + 2n+1 P2n+3 (cos φ) , (4.23b)

[τn(B) ]r (φ) = (2n + 2)(2n + 3) α2n+2 P2n+1 (cos φ) + β2n+1 P2n+3 (cos φ) ,


(4.23c)

0
[τn(B) ]φ (φ) = α2n+2 + 2n+1 P2n+2 (cos φ) − (2n + 2)(2n + 3)

(4.23d)
× α2n+2 P2n+1 (cos φ) + (α2n+1 − 2n + 1 − 4ν)P2n+3 (cos φ) ,

[τn(B) ]θ (φ) = −(4n + 5) (2n + 2)(2n + 3)(1 − 2ν)P2n+3 (cos φ)

(4.23e)
0


+ (2n + 1 + 2ν)P2n+2 (cos φ) ,
0 0
[τn(B) ]rφ (φ) = sin φ α2n+2 (2n + 3)P2n+1


(cos φ) + (2n + 2)P2n+3 (cos φ) , (4.23f)

where the coefficients β2n , γ2n and 2n are defined as

β2n = (2n + 2)(2n + 5) − 2ν,

γ2n = (2n + 2)(2n + 5 − 4ν),

2n = (2n + 1)(2n − 2 + 4ν).

4.1.2. Axisymmetric boundary conditions

Let us briefly analyse the fulfilment of the axisymmetric boundary conditions on Γs ,

given in (2.12d). We have that

σ(v)φ̂ · r̂ = (σrφ (v)r̂ + σφ (v)φ̂) · r̂ = σrφ (v), (4.24)

and
v · φ̂ = vφ , (4.25)

so we need that the component vφ of (4.20a) and the component σrφ of (4.20b) vanish on
Γ∞ , that is, for φ = π (cf. (2.6b)). Actually, these conditions are already fulfilled due to the
factor sin φ existing in (4.21b), (4.21f), (4.22b), (4.22f), (4.23b) and (4.23f). Consequently,
 69

the analytical solution given in (4.20a)-(4.20b) satisfies also (2.12d). In the next section,
we impose the boundary conditions on Γh , given in (2.12c).

4.2. Numerical enforcement of boundary conditions on the hemispherical pit

Up to now, we have obtained a solution in series form, given in (4.20a)-(4.20b), which

satisfies the elasticity equation (2.12a), the traction-free boundary conditions on the in-
finite plane surface (2.12b), the axisymmetric boundary conditions on the vertical surface
(2.12d), and the decaying condition at infinity (2.12e). It is a fully analytical solution, since
no numerical approximation has been introduced yet. In the present section, we enforce this
solution to satisfy the boundary conditions on the hemispherical surface (2.12c), which is
only possible in numerical form, so the sought solution becomes semi-analytical. This nu-
merical enforcement is done by means of minimising a quadratic functional, as described
below.

4.2.1. Truncation of the series

The fulfilment of the boundary conditions on the pit (2.12c) is directly related to the
coefficients An and Bn in (4.20a)-(4.20b), which have to be chosen in such a way that
(2.12c) is satisfied. However, there is an infinite number of coefficients An and Bn , which
are determined by an infinite set of simultaneous linear equations. Hence, it is not possible
to calculate them exactly, keeping the analytical nature of the solution. This drawback
is overcome by truncating the infinite series in (4.20a)-(4.20b) at a finite order N , which
introduces the first numerical approximation to our procedure and gives rise to a semi-
analytical solution of (2.12). The truncated solution is given by
N
X  h 2n+2 N
X  h 2n+3
v N (r, φ) = An w(A)
n (φ) + Bn w(B)
n (φ), (4.26a)
n=0
r n=−1
r
N
X N 
1  h 2n+3 X  h 2n+4
σN (r, φ) = An τ (A)
n (φ) + Bn τ (B)
n (φ) . (4.26b)
h n=0
r n=−1
r
 70

This approximation fixes a finite number of coefficients An and Bn that are to be calculated.
This is done by solving a finite linear system of equations for the coefficients An and Bn ,
which is obtained next.

4.2.2. Quadratic functional and its matrix form

The method to determine a linear system of equations satisfied by the coefficients An

and Bn is based upon the minimisation of a quadratic functional, which we define as
1 1
Z Z
J (v N ) = σN r̂ · v N ds − f · v N ds, (4.27)
2h Γh h Γh

where v N and σN are given in (4.26a) and (4.26b), respectively, and f is the vector func-
tion that arises at the right-hand side of (2.12c), whose components in r and φ were de-
fined in (2.14a) and (2.14b), respectively. The first term in the right-hand side of (4.27)
is quadratic in v N and represents the surface elastic potential energy on Γh . The minus
sign has been chosen in order to obtain a strictly convex functional. The second term is
linear in v N and is related to the right-hand side vector function f . Although we have
assumed a particular function f , the subsequent analysis is valid for any piecewise con-
tinuous function f : Γh → R2 satisfying fφ (π) = 0. Moreover, v N and σN are regarded
in (4.27) as generic functions depending on the arbitrary coefficients A0 , A1 , A2 , . . . , AN
and B−1 , B0 , B1 , . . . , BN , respectively, which are to be determined in order to minimise
the functional J. Expressing v N , σN f in terms of their respective components in r and φ,
making explicit the integrals and rearranging terms, (4.27) is rewritten as
h π
Z


J (v N ) = [σN ]r (h, φ)[vN ]r (h, φ) + [σN ]rφ (h, φ)[vN ]φ (h, φ) sin φ dφ
2 π/2
Z π (4.28)


−h fr (h, φ)[vN ]r (h, φ) + fφ (h, φ)[vN ]φ (h, φ) sin φ dφ,
π/2

and next, the components of v N and σN in (4.26a) and (4.26b) are evaluated at r = h
and substituted in (4.28). Expanding the resulting terms and collecting the products An Ak ,
 71

An Bk , Bn Ak and Bn Bk , the functional J is reexpressed as

N N N N
1 X X (AA) 1 X X (AB)
J (v N ) = Q An Ak + Q An Bk
2 n=0 k=0 nk 2 n=0 k=−1 nk
N N N N
1 X X (BA) 1 X X (BB)
+ Q Bn Ak + Q Bn Bk (4.29)
2 n=−1 k=0 nk 2 n=−1 k=−1 nk
N
X N
X
− yn(A) An − yn(B) Bn ,
n=0 n=−1

(AA) (AB) (BA) (BB) (A) (B)

where the terms Qnk , Qnk , Qnk , Qnk , yn and yn correspond to the entries of
matrices Q(AA) ∈ MN +1 (R), Q(AB) ∈ M(N +1)×(N +2) (R), Q(BA) ∈ M(N +2)×(N +1) (R),
Q(BB) ∈ MN +2 (R) and vectors y (A) ∈ RN +1 and y (B) ∈ RN +2 , respectively. These
entries are defined as the following integrals:
Z π  
(αβ) (α) (β) (α) (β)
Qnk = − [wn ]r (φ)[τk ]r (φ) + [wn ]φ (φ)[τk ]rφ (φ) sin φ dφ, (4.30a)
π/2
Z π  
yn(α) = −h [wn(α) ]r (φ)fr (φ) + [wn(α) ]φ (φ)fφ (φ) sin φ dφ, (4.30b)
π/2

where α, β = A, B. Substituting (2.14a), (2.14b), (4.21a), (4.21b), (4.21c), (4.21f), (4.22a),

(4.22b), (4.22c), (4.22f), (4.23a), (4.23b), (4.23c), (4.23f) in (4.30a) and (4.30b) as appro-
(αβ) (α)
priate leads us to obtain expressions for the quantities Qnk and yn in terms of explicit
integrals. These expressions are too cumbersome to be reproduced here. However, with
the aid of the integral formulae provided in the Appendix 6.1, all the involved integrals
are calculated exactly, yielding explicit expressions for the entries of matrices Q(αβ) and
vectors y (α) which we reproduce next for α, β = A, B. The entries of the matrix Q(AA)
are computed by using (6.4a) and (6.4d), corresponding to a tridiagonal symmetric matrix
with main diagonal entries
(2n + 1)(2n + 2)(4n + 3)
2µQ(AA)
nn =
(4n + 5) (4.31a)
3 2 2 2


× (2n + 3)(16n + 32n + 22n + 8ν − 12ν + 9) + 4(1 − ν ) ,
 72

where 0 ≤ n ≤ N , and sub-diagonal and super-diagonal entries

(AA) (AA) (2n + 1)(2n + 2)(2n + 3)(2n + 4)(4n + 3)

2µQn,n+1 = Qn+1,n =
(4n + 5) (4.31b)


× (2n + 2)(2n + 4) − 1 + 2ν ,

where 0 ≤ n ≤ N − 1. The entries of matrices Q(AB) and Q(BA) are calculated by

using (6.4a), (6.4c), (6.4f) and (6.5a). It is obtained that they correspond to full matrices
satisfying Q(BA) = [Q(AB) ]T . The entries of Q(AB) in the case k = −1 are given by

(AB) 4(4n + 3)(2n(2n + 3)ν(ν − 2) − 3 + 5ν − 4ν 2 )P2n (0)

2µQn,−1 = − , (4.32a)
(2n − 1)(2n + 2)(2n + 4)
and in the case 0 ≤ k ≤ N + 1 are given by

(AB) 4(2k + 1)(2k + 3)(4k + 5)(2n + 1)(4n + 3)ηnk P2n (0)P2k (0)
2µQnk = , (4.32b)
(2k+6+2n)(2k+3−2n)(2k+4+2n)(2k−1−2n)(2k+1−2n)
where

ηnk = 51 + 58k + 32k 2 n2 + 188kn + 138n + 56k 2 n + 72n2 + 16k 2 + 104kn2

− (2k − 1 − 2n) 4n2 (2k + 4) − 4k((2k + 5)n + 2k + 6) − 21 − 8k(2k + 2) ν

− (2k + 3 − 2n)(2k + 6 + 2n)(2k − 1 − 2n)ν 2 ,

and 0 ≤ n ≤ N . The entries of matrix Q(BB) are calculated by using (6.4b), (6.4c),
(6.4e), and (6.5b), together with some elementary primitives of trigonometric functions. It
is obtained that Q(BB) corresponds to a symmetric matrix that is almost tridiagonal, except
for its first row and its first column, which are full. The entries associated with the first row
and column of Q(BB) are

(BB) 2
2µQ−1,−1 = 2(1 − 2ν)2 ln 2 + (2 + 5ν − 6ν 2 ), (4.33a)
3


(BB) (BB) 2(1 − 2ν)(4n + 5)(2n + 1) (2n + 4)ν − 1 P2n (0)
2µQn,−1 = Q−1,n = −
(2n + 2)(2n + 4) (4.33b)
+ 2(7 + 2ν)δn0 ,
 73

where 0 ≤ n ≤ N and δnk stands for the Kronecker delta. The entries associated with the
tridiagonal part of Q(BB) are the main diagonal entries
(4n + 5)(2n + 2)(2n + 3)
2µQ(BB)
nn =
(4n + 7) (4.33c)
× (32n4 +208n3 +492n2 +4(125−2ν +4ν 2 )n+187−20ν +28ν 2 ),

where 0 ≤ n ≤ N , and the and sub-diagonal and super-diagonal entries

(BB) (BB) (4n + 5)(2n + 2)(2n + 3)(2n + 4)(2n + 5)

2µQn+1,n = Qn,n+1 =
(4n + 7) (4.33d)
× ((2n + 4)(2n + 6) − 1 + 2ν),

where 0 ≤ n ≤ N − 1. The entries of vector y (A) are computed by using (6.4c) and (6.4f),
yielding

2%gh2 (4n + 3)(2n + 1)P2n (0)

yn(A) =
(1 − ν)(2n − 1)(2n + 2)(2n + 4)(2n + 6) (4.34)
2
× (4(2n + 3) − 6ν(3n + 5) + ν (2n + 4)(2n + 6)),

where 0 ≤ n ≤ N . Fig. 4.1 shows schematically the structure of matrices Q. The entries
of vector y (B) are computed by using (6.4b), (6.4c), (6.4e) and (6.5b). It is obtained that
only the first two entries of y (B) are different from zero. These entries are given by

(B) %gh2 (26 − 47ν + 30ν 2 )

y−1 = , (4.35a)
30(1 − ν)
(B) 2%gh2 (46 − 71ν + 70ν 2 )
y0 = . (4.35b)
21(1 − ν)

We define also the vectors A ∈ RN +1 and B ∈ RN +2 as those whose entries correspond

to the coefficients An and Bn to be computed. From the above matrices and vectors, we
also define the square matrix Q ∈ M2N +3 (R) and the vectors x, y ∈ R2N +3 by blocks as
     
(AA) (AB) (A)
Q Q A y
Q= , x= , y= , (4.36)
(AB) T (BB) (B)
[Q ] Q B y
 74

(AA) (AB)
Q(BB)
Q Q

N +1 N +2

N +1 N +2 N +2

Figure 4.1. Structure of matrices Q(AA) ,Q(AB) and Q(BB) .

which allow us to reexpress the functional J given in (4.29) as the following quadratic
form in x:
1
J (x) = xT Qx − xT y. (4.37)
2

4.2.3. Linear system and method of inversion

A linear system for the coefficients An and Bn is obtained by minimisation of the

functional J. It holds that Q is a symmetric matrix, due to the properties of its blocks Q(ij) ,
and it is also positive definite, thanks to the elliptic nature of the elastostatic boundary-value
problem. Therefore, J has a global minimum, which is reached when ∇J (x) = Qx−y =
0, or equivalently, when
Qx = y. (4.38)

Eq. (4.38) corresponds to the sought linear system of equations for the coefficients An and
Bn . We take advantage of the symmetry and positive definiteness of Q in solving (4.38).
Expressing Q, x and y by blocks in (4.38) yields
    
Q(AA) Q(AB) A y (A)
  = , (4.39)
(AB) T (BB) (B)
[Q ] Q B y
 75

and this system is inverted with the aid of the so-called Schur-Banachiewicz blockwise
inversion formula (cf. (Björck, 1996)), which is given by

A = [Q(AA) ]−1 + [Q(AA) ]−1 Q(AB) [Q

e (BB) ]−1 [Q(AB) ]T [Q(AA) ]−1 y (A)


(4.40a)
(AA) −1 (AB) e (BB) ]−1 y (B) ,
− [Q ] Q [Q
e (BB) ]−1 [Q(AB) ]T [Q(AA) ]−1 y (A) + [Q
B = −[Q e (BB) ]−1 y (B) , (4.40b)

e (BB) denotes the Schur complement of Q(BB) in Q, defined as

where Q

e (BB) = Q(BB) − [Q(AB) ]T [Q(AA) ]−1 Q(AB) .

Q (4.41)

To evaluate (4.40a) and (4.40b), it suffices to inverse the symmetric and positive definite
matrices Q(AA) and Q e (BB) . As Q(AA) is in addition a tridiagonal matrix, it is efficiently

inverted by using the Thomas algorithm for tridiagonal systems (cf. (Golub & Loan, 1996),
e (BB) , we employ its Cholesky factorisation, which is quickly
Algorithm 4.3.6). To invert Q
obtained with the aid of a suitable algorithm (cf. (Golub & Loan, 1996), Algorithm 4.2.1).
We thus establish the following algorithm to compute the coefficients An and Bn :

(i) Compute the coefficients for tridiagonal inversion of Q(AA) .

(ii) Use the coefficients to evaluate [Q(AA) ]−1 y (A) and [Q(AA) ]−1 Q(AB) .
(iii) Calculate the Schur complement Q e (BB) using (4.41).

(iv) Obtain the Cholesky factorisation of Qe (BB) .

e (BB) ]−1 [Q(AB) ]T .
e (BB) ]−1 y (B) and [Q
(v) Use the Cholesky factorisation to evaluate [Q
(vi) Assemble A and B as indicated in (4.40a) and (4.40b).

This algorithm yields approximate values for the two series of coefficients A0 , A1 , A2 , . . . , AN
and B−1 , B0 , B1 , . . . , BN . By substituting these values in (4.26a) and (4.26b), we obtain
the desired semi-analytical solution of the boundary-value problem (2.12) in explicit way.

4.3. Numerical results and validation

In this section, we present numerical results obtained with the semi-analytical method
described throughout Chapter 4. Prior to this, the convergence of the series is numerically
 76

verified and a convergence criterion is established in order to determine a suitable value

for the truncation parameter N . Furthermore, the procedure is validated by comparing our
semi-analytical solution with results obtained numerically using the commercial software
COMSOL Multiphysics. The unbounded domain Ω is truncated by means of an artificial
square box.

4.3.1. Numerical results

The semi-analytical method described in Chapter 3 and Section 4.2 was implemented
in its entirety, and the obtained solution was numerically evaluated for different values of
N . The considered numerical values of the gravity acceleration g, the radius of the pit h,
the density %, the Young’s modulus E and the Poisson’s ratio ν are shown in Table ??.To
Table 4.1. Numerical values of the physical parameters.

Parameter h g % E ν
Value 600 m 9.81 m/s2 2725 kg/m3 70.2 GPa 0.3

evaluate numerically the semi-analytical solution of (2.12) and analyse its convergence as
the truncation parameter N tends to infinity, we consider a bounded subregion of Ω, defined
in axisymmetric spherical coordinates by the points (r, φ) satisfying h ≤ r ≤ 2000 m and
π/2 ≤ φ ≤ π. The evaluations are performed at the points (ri , φj ) of an equispaced
rectangular grid in the (r, φ)-plane, defined as ri = h + i∆r and φj = π/2 + j∆φ, where
i = 0, 1, 2, . . . , 140, ∆r = 10 m, j = 0, 1, 2, . . . , 500, and ∆φ = π/1000. In order to
test numerically the convergence of the series in terms of N , we compute the relative error
between two successive solutions, considering separately the displacement field v and the
stress tensor σ(v). The respective relative errors are defined as

|v N +1 − v N | |σ N +1 (v N +1 ) − σ N (v N )|
eN (v) = , eN (σ) = , (4.42)
|v N | |σ N (v N )|

where in the latter case, | · | corresponds to the matrix Euclidean norm (or Frobenius norm).
Successive values of the truncation parameter up to N = 100 were considered in the anal-
ysis. A semi-logarithmic plot with both relative errors in function of N is presented in
Fig. 4.2, where it is observed that both of them decrease as N increases, exhibiting the
 77

error associated with the stress tensor a slower decrease. We conclude from this analysis
that the semi-analytical method shows an acceptable convergence. In order to set a value of
N for evaluation of the semi-analytical solution, we have used as a convergence criterion
that both relative errors have to be smaller than a tolerance of 0.001%, yielding the value
N = 40 (see Fig. 4.2), which we assume from now on. Having already calculated the solu-
tion v, the physical displacement field u is obtained by adding the lithostatic displacement
field ug (evaluated at the same rectangular grid), as indicated in (2.10). The physical stress
tensor σ(u) is obtained in a similar way, by adding to σ(v) the lithostatic stress tensor
σ(ug ), whose components are explicitly obtained by substituting (2.10) in (2.7). The val-
ues of the displacement components ur and uφ , and the stress components σr (u), σφ (u),
σθ (u) and σrφ (u) are depicted in Fig. 4.3. It is observed from the plots of stress compo-
nents that obtained semi-analytical solution fulfils the boundary conditions (2.9b), (2.9c)
and (2.9d).

Relative errors
1
10
eN(v)
0
10 eN(σ)

−1
10

−2
10
Error (%)

−3
10

−4
10

−5
10

−6
10

−7
10
0 20 40 60 80 100
Truncation parameter N

Figure 4.2. Relative error of the solution between two successive iterations in N .
 78

Figure 4.3. Plots of displacement and stress components obtained in Ω.

 79

4.3.2. Validation of the procedure

In order to confirm the correctness of the semi-analytical solution, a validation test

was carried out. For this, the axisymmetric boundary-value problem in u (2.9) was numer-
ically solved using the commercial finite element software COMSOL Multiphysics, with
the physical parameters of the problem fixed to the same numerical values indicated in
Table ??. The purpose of this numerical simulation is to compare the results calculated
by such means with those obtained by using the semi-analytical method proposed herein.
As this type of software is only able to deal with bounded regions, the unbounded domain
Ω was truncated by means of an artificial square box of length L, as indicated schemat-
ically in Fig. 4.4. In order to be able to solve (2.9) in the truncated domain (denoted by
Ωtr ), the decaying condition at infinity (2.9e) was replaced by artificial Dirichlet boundary
conditions on the right and bottom boundaries, with the displacements set to the lithostatic
displacement ug field given in (2.10). Different values of the length L were considered,
starting from a minimum value of L = 1000 m up to a maximum value of L = 13000 m,
with successive increases of ∆L = 1000 m. For each one of these values, non-uniform tri-

Ωtr L

Figure 4.4. Schematic representation of the domain Ω truncated by a

square box of length L.

angular meshes of the domain Ωtr were generated, with the element size varying between a
 80

minimum value of 2 m and a maximum value of 50 m, and the maximum element growth
rate set to 1.3. To solve numerically the boundary-value problem, standard conforming P1
finite elements in each mesh were used. Common sense indicates that as the domain size
grows, the approximation of the unbounded boundary-value problem (2.9) by a bounded
one becomes more accurate. Therefore, it is expected that the solution calculated numeri-
cally by finite elements approaches the semi-analytical solution as the length L increases.
In order to find out whether this occurs, the displacement components ur and uφ , and the
stress components σφ and σθ obtained by both means were compared on the surface of the
pit Γh . The remaining stress components σr and σrφ were not considered in the compar-
ison, since they are set to zero by the boundary condition prescribed on Γh (2.9c). The
displacement curves obtained by the semi-analytical method, and by finite elements for
some values of L, are presented separately for each displacement component in Fig. 4.5,
where the horizontal axis corresponds to the angle φ. The stress curves for σφ and σθ are
given in analogous way in Fig. 4.6. It is noticed from Figs. 4.5 and 4.6 that, as expected,
the solution obtained by the semi-analytical method is better approximated by the solution
computed numerically as the value of L increases. This behaviour is seen in the displace-
ment as well as the stress curves, being more evident in the latter case. It is also observed
from Fig. 4.5 that the component ur obtained semi-analytically is better approximated than
the component uφ obtained in the same way as the value of L increases. This difference
between both displacement components can be explained by the error introduced in the
numerical solution by the artificial boundary conditions, particularly on the right boundary
of Ωtr , which affects more strongly the displacement uφ near the infinite plane surface. In
order to further verify the already observed tendency as L increases, the relative error of
the semi-analytical solution with respect to the numerical one was evaluated on Γh , for all
the assumed values of L, considering separately the errors associated with the displace-
ment vector u and the stress tensor σ(u). Both relative error curves as functions of L are
presented in a semilogarithmic plot in Fig. 4.7. These curves confirm the behaviour seen
in Figs. 4.5 and 4.6: The relative error between the semi-analytical and the numerical so-
lution decreases as L increases, both for the displacement vector and the stress tensor. In
 81

Displacement component ur Displacement component uφ

0.12
Semi−analytical
Numerical L=2000 m
0.2 Numerical L=5000 m 0.1
Numerical L=10000 m

0.15 0.08
Distance (m)

Distance (m)
0.1 0.06

0.05 0.04

Semi−analytical
0 0.02 Numerical L=2000 m
Numerical L=5000 m
Numerical L=10000 m
−0.05 0
1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2
Angle (rad) Angle (rad)

Figure 4.5. Comparison of displacement components on Γh .

Stress component σφ Stress component σθ

12 5

4
10

8
2
Stress (MPa)

Stress (MPa)

6 1

0
4

−1
Semi−analytical Semi−analytical
2 Numerical L=2000 m Numerical L=2000 m
Numerical L=5000 m −2 Numerical L=5000 m
Numerical L=10000 m Numerical L=10000 m
0 −3
1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2
Angle (rad) Angle (rad)

Figure 4.6. Comparison of stress components on Γh .

fact, both relative errors show a similar tendency, being the error in stress smaller than the
error in displacement for all values of L. We confirm from this analysis the validity of the
proposed semi-analytical solution.
 82

Relative errors
2
10
err(u)
err(σ)

1
10
Error (%)

0
10

−1
10
0 2000 4000 6000 8000 10000 12000 14000
Distance L (m)

Figure 4.7. Relative errors associated with the displacement vector and the
stress tensor on Γh .
 83

5. A DIRICHLET-TO-NEUMANN FINITE ELEMENT METHOD FOR AXISYM-

METRIC ELASTOSTATICS IN A SEMI-INFINITE DOMAIN

We present a DtN FEM procedure for boundary-value problems of elastostatics in

semi-infinite domains with axisymmetry about the vertical axis. A semi-spherical artificial
boundary is used to truncate the semi-infinite domain and to obtain a bounded computa-
tional domain, where a FEM scheme is employed. By using a semi-analytical procedure of
solution in the unbounded residual domain lying outside the artificial boundary, the exact
nonlocal boundary conditions provided by the DtN map are numerically approximated and
efficiently coupled with the FEM scheme. Numerical results are provided to demonstrate
the effectiveness and accuracy of the proposed method.

5.1. Mathematical formulation

5.1.1. Generalities

In this Chapter the boundary of Ω consists of three parts: A vertical boundary of ax-
isymmetry coinciding with the axis of revolution, denoted by Γs , a horizontal unperturbed
boundary coinciding with the infinite plane surface of the half-space, denoted by Γ∞ , and
a perturbed bounded boundary which is assumed piecewise smooth, denoted by Γp (see
Figure 5.1).

5.1.2. Axisymmetric elastostatic model

We consider the following elastostatic BVP in Ω: Find u : Ω → R2 such that

∇ · σ(u) = 0 in Ω, (5.1a)

σ(u)ẑ = 0 on Γ∞ , (5.1b)

σ(u)n̂ = f on Γp , (5.1c)

σ(u)ρ̂ · ẑ = u · ρ̂ = 0 on Γs , (5.1d)

|u| = O r−1


as r → ∞, (5.1e)
 84

φ
z r Γ∞
ρ

Γp

Γs

Figure 5.1. Axisymmetric semi-infinite domain.

where n̂ denotes the unit outward normal vector on Γp and f : Γp → R2 is a given

piecewise continuous vector function.

5.2. FEM formulation in the computational domain

5.2.1. Equivalent bounded boundary-value problem

In order to apply the DtN FEM to solve (5.1), an artificial boundary is introduced to
truncate Ω, which is chosen to be a quarter circle of radius R, denoted by ΓR (cf. Fig-
ure 5.2). Notice that if Ω is rotated about the axis of revolution to generate the 3D semi-
infinite solid, ΓR gives rise to a hemispherical surface of radius R. The bounded domain
lying inside ΓR is denoted by Ωi , which corresponds to the computational domain. The re-
sulting bounded parts of boundaries Γ∞ and Γs are respectively denoted by Γi∞ and Γis (cf.
Figure 5.2). In order to solve (5.1), we formulate the following mathematically equivalent
BVP in Ωi : Find u : Ωi → R2 such that

∇ · σ(u) = 0 in Ωi , (5.2a)

σ(u)ẑ = 0 on Γi∞ , (5.2b)

 85

σ(u)n̂ = f on Γp , (5.2c)

σ(u)ρ̂ · ẑ = u · ρ̂ = 0 on Γis , (5.2d)

σ(u)r̂ = −Mu on ΓR , (5.2e)

where (5.2e) is the exact ABC on ΓR , expressed in terms of the DtN map M, which
we assume for the moment to be known. The definition of M, as well as its numerical
approximation, will be discussed in detail in the next section. We have assumed that the

z
Γi∞
ρ

Γp

R Ωi
ΓR

Γis

Figure 5.2. Axisymmetric computational domain.

medium occupying Ω is isotropic, homogeneous, and linear elastic. Nevertheless, the DtN
FEM also applies if more general hypotheses in this regard are assumed, but restricted to
the the computational domain, e.g. the medium in Ωi could be inhomogeneous, anisotropic,
nonlinear, and even inelastic. Moreover, nonzero right-hand sides in (5.2a) or (5.2b) are
allowed as well.

5.2.2. Weak formulation

Throughout this section, we work in axisymmetric cylindrical coordinates (ρ, z). To

state a weak (or variational) formulation of (5.2), we consider a Hilbert space containing
 86

physically admissible displacement fields in Ωi , defined as

V = v = (vρ , vz ) ∈ [H 1 (Ωi )]2 : vρ = 0 on Γis ,



where H 1 (Ωi ) is the usual Sobolev space of L2 functions with L2 first derivatives. Notice
that the boundary condition of zero normal displacement in Γs is included in V. The weak
formulation of (5.2) is expressed as: Find u ∈ V such that

a(u, v) + b(u, v) = (f , v)Γp ∀ v ∈ V, (5.3)

where a corresponds to the bilinear form associated with the weak formulation of the elas-
tostatic equation, expressed in axisymmetric cylindrical coordinates (cf. (Szabo & Babuška,
1991)) as
Z   
∂vρ vρ ∂vz ∂vρ ∂vz
a(u, v) = σρ (u) + σθ (u) + σz (u) + σρz (u) + ρ dρ dz, (5.4)
Ωi ∂ρ ρ ∂z ∂z ∂ρ
whereas b is a bilinear form involving the DtN map, defined as
Z
b(u, v) = Mu · v dΓR , (5.5)
ΓR

and the right-hand side of (5.3) is

Z
(f , v)Γp = f · v dΓp .
Γp
 87

By substituting (2.7) in (5.4), expanding terms and rearranging appropriately, a is reex-

pressed as
Z  
∂uρ ∂vρ ∂uz ∂vz
a(u, v) = (λ + 2µ) + ρ dρ dz
Ωi ∂ρ ∂ρ ∂z ∂z
Z  
∂uρ ∂vz ∂uz ∂vρ
+λ + ρ dρ dz
Ωi ∂ρ ∂z ∂z ∂ρ
Z   
∂uρ ∂uz ∂vρ ∂vz
+µ + + ρ dρ dz (5.6)
Ωi ∂z ∂ρ ∂z ∂ρ
Z    
∂uρ ∂uz ∂vρ ∂vz
+λ + vρ + uρ + dρ dz
Ωi ∂ρ ∂z ∂ρ ∂z
uρ vρ
Z
+ (λ + 2µ) dρ dz.
Ωi ρ
The first three integrals in (5.6) are terms analogous to the 2D elastostatic case (except for
the differential element), whereas the last two integrals are additional terms that arise due
to the axisymmetry.

5.2.3. FEM discretisation

In what follows, the weak formulation (5.3) is discretised by using standard conform-
ing P 1 finite elements. Let us assume for simplicity that Γp is a polygonal line. We denote
by Th a family of regular triangular meshes of Ωi such that Ωi = T ∈Th T (the quarter
S

circle ΓR is approximated by a polygonal line). The indexing parameter h measures the

size of each triangular element T and is defined as h = max{diam T : T ∈ Th }, with
p
diam T = sup{ (ρ2 − ρ1 )2 + (z2 − z1 )2 : (ρ1 , z1 ), (ρ2 , z2 ) ∈ T }. For each triangular
mesh T h , let V h be a finite-dimensional vector subspace of V consisting of continuous
piecewise linear vector functions, defined as

V h = v h ∈ V : v h ∈ [C 0 (Ωi )]2 , v h ∈ [P1 (T )]2 ∀T ∈ T h ,


T

where C 0 (Ωi ) denotes the space of continuous functions in Ωi and P1 (T ) stands for the
space of polynomials of degree at most one defined in T . The discrete version of (5.3) is
 88

expressed as: Find uh ∈ V h such that

a(uh , v h ) + b(uh , v h ) = (f , v h )Γp ∀ vh ∈ V h. (5.7)

Let I be the set of all nodes of the triangular mesh T h . For each i ∈ I, we introduce a
nodal shape (or basis) function ψi ∈ V h which has unit value at node i and zero value at all
other nodes. Let also Is ⊂ I be the set of nodes lying on Γis . A nodal basis of V h is thus
built by considering vector functions ψi ρ̂ for nodes i ∈ I \ Is , to allow for the condition
of zero normal displacement given by (5.2d), and vector functions ψi ẑ for all nodes i ∈ I.
Consequently, the solution of (5.7) is expressed as a linear combination of the nodal basis
functions as
X X
uh (ρ, z) = dρi ψi (ρ, z)ρ̂ + dzi ψi (ρ, z)ẑ, (5.8)
i∈I\Is i∈I

where dρi and dzi are respectively the unknown nodal values of components uhρ and uhz
of the solution uh . Replacing (5.8) in (5.7) and substituting v h by each one of the nodal
basis functions ψi ρ̂ and ψi ẑ, we arrive at the finite element matrix form of the problem,
expressed as
Kd = F , (5.9)

with
   
K aρρ K aρz K bρρ K bρz
K = K a + K b, Ka =  , Kb =  ,
K azρ K azz K bzρ K bzz

and    
dρ Fρ
d= , F = ,
dz Fz
where

[K aρρ ]ij = a(ψi ρ̂, ψj ρ̂), [K aρz ]ij = a(ψi ρ̂, ψj ẑ),
(5.10)
[K azρ ]ij = a(ψi ẑ, ψj ρ̂), [K azz ]ij = a(ψi ẑ, ψj ẑ),
 89

[K bρρ ]ij = b(ψi ρ̂, ψj ρ̂), [K bρz ]ij = b(ψi ρ̂, ψj ẑ),
(5.11)
[K bzρ ]ij = b(ψi ẑ, ψj ρ̂), [K bzz ]ij = b(ψi ẑ, ψj ẑ),
[dρ ]j = dρj , [dz ]j = dzj , (5.12)

[F ρ ]i = (f , ψi ρ̂)Γp , [F z ]i = (f , ψi ẑ)Γp . (5.13)

The entries of matrix K a and vector F are computed by standard numerical integration
techniques. However, this is not possible in the case of matrix K b . Substituting (5.5) in
(5.11), we reexpress the entries of K b as
Z Z
b
[K ρρ ]ij = ψi ρ̂ · Mψj ρ̂ dΓR , [K bρz ]ij = ψi ρ̂ · Mψj ẑ dΓR , (5.14a)
ΓR ΓR
Z Z
[K bzρ ]ij = ψi ẑ · Mψj ρ̂ dΓR , [K bzz ]ij = ψi ẑ · Mψj ẑ dΓR . (5.14b)
ΓR ΓR

Computing these terms is far from being straightforward, since they involve the DtN map
M, which has to be calculated in an accurate way. Notice that these terms are different
from zero only if i, j ∈ IR , where IR ⊂ I denotes the set of mesh nodes lying on ΓR .
A correct calculation of the nonzero entries of matrix K b is critical for obtaining an ac-
curate solution of (5.2), since K b accounts for the contribution from the exact ABCs in
the standard finite element scheme used for numerical discretisation. This crucial issue is
thoroughly treated in the next section.

5.3. Approximation of the exact artificial boundary conditions

5.3.1. Definition of the DtN map

Next, we provide the mathematical definition of the DtN map M introduced in the pre-
vious section. Let us consider the residual semi-infinite domain lying outside the artificial
boundary ΓR , denoted by Ωe . The resulting unbounded parts of boundaries Γ∞ and Γs are
respectively denoted by Γe∞ and Γes (cf. Figure 5.3). Let H s (ΓR ) be the standard Sobolev
space on ΓR of real order s. Given any v ∈ [H 1/2 (ΓR )]2 , we obtain Mv ∈ [H −1/2 (ΓR )]2
as
Mv = −σ(u)r̂ ΓR
, (5.15)
 90

where u is the solution of the following BVP in Ωe : Find u : Ωe → R2 such that

∇ · σ(u) = 0 in Ωe , (5.16a)

σ(u)ẑ = 0 on Γe∞ , (5.16b)

u=v on ΓR , (5.16c)

σ(u)ρ̂ · ẑ = u · ρ̂ = 0 on Γes , (5.16d)

|u| = O r−1


as r → ∞. (5.16e)

As already mentioned, unlike most exterior BVPs, in this case the method of separation
of variables fails in obtaining a full analytical closed-form expression for the solution u
of (5.16) in function of a generic Dirichlet datum v on ΓR . This drawback is overcome by
solving (5.16) approximately, just for those v required to compute the nonzero entries of
matrix K b . By virtue of (5.14), the required right-hand sides in (5.16c) are v = ψj ρ̂ and
v = ψj ẑ, for every j ∈ IR .

φ
r Γe∞

R
ΓR

Ωe

Γes

Figure 5.3. Axisymmetric semi-infinite residual domain.

 91

5.3.2. Numerical enforcement of exact boundary conditions on the artificial boundary

In order to solve (5.16), we apply the semi-analytical method presented in Chapter 4

with some suitable modifications. Proceeding analogously as in Section 4.1, we obtain
an analytical solution in series form satisfying (5.16a), (5.16b), (5.16b) and (5.16e). This
solution reads exactly as (4.20a)-(4.20b), but with h substituted by R. In the same way
as we did in Section 4.2, this solution is numerically forced to satisfy (5.16c), leading to
determine the coefficients An and Bn associated with a particular right-hand side v. In the
next subsection, this procedure will be applied to the particular right-hand sides v = ψj ρ̂
and v = ψj ẑ. First, we truncate the series in the aforementioned analytical solution at a
finite order N . The resulting expressions for the displacement vector uN and the stress
tensor σ N are
N
X  R 2n+2 N
X  R 2n+3
uN (r, φ) = An w(A)
n (φ) + Bn wn(B) (φ), (5.17a)
n=0
r n=−1
r
 N N 
1 X  R 2n+3 (A) X  R 2n+4
(B)
σ N (r, φ) = An τ n (φ) + Bn τ n (φ) , (5.17b)
R n=0 r n=−1
r

and this time we consider the quadratic energy functional J defined as

1 1
Z Z
J (uN ) = σ N n̂ · uN dΓR − σ N n̂ · v dΓR , (5.18)
2R ΓR R ΓR

where n̂ = −r̂ is the unit normal vector on ΓR , pointing outwards Ωe . The first term in
(5.18) is quadratic in uN and represents the surface elastic potential energy on ΓR , whereas
the second term is linear in uN and is related to the Dirichlet boundary condition assumed
on ΓR . The functional J has been defined in such a way that its minimisation provides nu-
merical values for coefficients A0 , A1 , A2 , . . . , AN and B−1 , B0 , B1 , . . . , BN so that (5.16c)
holds approximately. By expanding terms in (5.18), replacing n̂ = −r̂ and making explicit
 92

the integrals (notice that dΓR = R2 sin φ dφ), we arrive at

R π
Z


J (uN ) = − [σN ]r (R, φ) [uN ]r (R, φ) + [σN ]rφ (R, φ) [uN ]φ (R, φ) sin φ dφ
2 π/2
Z π


+R [σN ]r (R, φ) vr (R, φ) + [σN ]rφ (R, φ) vφ (R, φ) sin φ dφ.
π/2

(5.19)

As we did in Subsection 4.2.2, we define the column vectors A ∈ RN +1 and B ∈ RN +2 as

those whose entries are the coefficients A0 , A1 , A2 , . . . , AN and B−1 , B0 , B1 , . . . , BN , re-
spectively. By evaluating the corresponding components of uN and σ N at r = R in (5.17a)
and (5.17b), substituting in (5.19) and expanding terms appropriately, J is reexpressed as
an explicit function of A and B as
N N N N N N
1 X X (AA) 1 X X (AB) 1 X X (BA)
J (A, B) = Q An Ak + Q An Bk + Q Bn Ak
2 n=0 k=0 nk 2 n=0 k=−1 nk 2 n=−1 k=0 nk
N N N N
1 X X (BB) X
(A)
X
+ Qnk Bn Bk − yn A n − yn(B) Bn ,
2 n=−1 k=−1 n=0 n=−1

(5.20)

where the coefficients Qαβ

nk are given by (4.31a)-(4.33d) and
Z π  
(α) (α) (α)
yn = − [τn ]r (φ)vr (R, φ) + [τn ]rφ (φ)vφ (R, φ) sin φ dφ. (5.21)
π/2

(α)
where α = A, B. Coefficients yn depend on the particular right-hand side v assumed
in (5.16c), so their calculation is discussed in the next subsection. Defining matrix Q ∈
M2N +3 (R) and vectors x, y ∈ R2N +3 exactly as in Subsection 4.2.2, it is possible to reex-
press J in (5.20) as a quadratic form in x
1
J (x) = xT Qx − xT y. (5.22)
2
whose minimization yields again a linear sistem of equations for the coefficients A0 , A1 , . . . , AN
and B−1 , B0 , B1 , . . . , BN stored in the vector x. The procedure and associated algorithm
to solve the system are completely analogous to those presented in Section 4.2.3. This
 93

algorithm has to be applied repeatedly for each pair of vectors y (A) and y (B) necessary in
the computations. Nevertheless, as matrices Q(αβ) remain unchanged, each one of them,
together with the coefficients for inversion of Q(AA) and Q
e (BB) , are computed and stored

only once.

5.3.3. Numerical approximation of integral terms involving the DtN map in the FEM
formulation

In what follows, the numerical procedure described in the previous subsection is ap-
(α)
plied to approximate the terms in (5.14). First, coefficients yn in (5.21) are to be calculated
for v = ψj ρ̂ and v = ψj ẑ for every j ∈ IR . Combining with (2.1a), we obtain that these
two cases are respectively equivalent to consider

y (A) = C jρ , y (B) = D jρ ,

and
y (A) = C jz , y (B) = D jz ,

where C jρ , C jz ∈ RN +1 and D jρ , D jz ∈ RN +2 are the vectors of components

Z π
j
[τn(A) ]r (φ) sin φ + [τn(A) ]rφ (φ) cos φ ψj (R, φ) sin φ dφ


Cρ,n = −
π/2

n = 0, . . . , N, (5.23a)
Z π
j
[τn(A) ]r (φ) cos φ − [τn(A) ]rφ (φ) sin φ ψj (R, φ) sin φ dφ


Cz,n =−
π/2

n = 0, . . . , N, (5.23b)
Z π
j
[τn(B) ]r (φ) sin φ + [τn(B) ]rφ (φ) cos φ ψj (R, φ) sin φ dφ


Dρ,n =−
π/2

n = 0, . . . , N, (5.23c)
Z π
j
[τn(B) ]r (φ) cos φ − [τn(B) ]rφ (φ) sin φ ψj (R, φ) sin φ dφ


Dz,n =−
π/2

n = 0, . . . , N, (5.23d)
 94

respectively. By substituting (4.21c), (4.21f), (4.22c), (4.22f), (4.23c) and (4.23f) in (5.23)
and expanding, components of vectors C jρ , C jz , D jρ , D jz are rewritten as

j j j


Cs,n = −(2n + 1)(2n + 2) α2n ls,2n + β2n ls,2n+2

− (2n + 2)α2n mjs,2n − (2n + 1)α2n+1 mjs,2n+2 n = 0, . . . , N, (5.24a)

j j
Ds,−1 = −3 ls,1 − (1 − 2ν)qsj , (5.24b)
j j j


Ds,n = −(2n + 2)(2n + 3) α2n+2 ls,2n+1 + β2n+1 ls,2n+3

− α2n+2 (2n + 3)mjs,2n+1 + (2n + 2)mjs,2n+3

n = 0, . . . , N, (5.24c)

where s = ρ, z and
Z π
j
lρ,n = sin2 φPn (cos φ)ψj (R, φ)dφ, (5.25a)
π/2
Z π
j
lz,n = sin φ cos φPn (cos φ)ψj (R, φ)dφ, (5.25b)
π/2
Z π
mjρ,n = sin2 φ cos φPn0 (cos φ)ψj (R, φ)dφ, (5.25c)
π/2
Z π
mjz,n =− sin3 φPn0 (cos φ)ψj (R, φ)dφ, (5.25d)
π/2
Z π
qρj = (1 + cos φ)ψj (R, φ)dφ, (5.25e)
π/2

qzj = 0. (5.25f)

To compute these integrals, we observe that the mesh Th restricted to ΓR gives rise to a
partition of the quarter circle into segments (usually equispaced). Let J be the number
of these segments. Then the set IR necessarily consists of J + 1 mesh nodes, which we
assume for simplicity to be correlatively numbered, that is, IR = {1, 2, . . . , J, J + 1}. We
associate with each node j ∈ IR an angle φj , in such a way that π/2 = φ1 < φ2 < . . . <
φJ < φJ+1 = π. The P 1-nodal shape functions ψj , restricted to ΓR , do not depend on the
 95

radius R, and have the following explicit expressions

φ2 − φ
 
ψ1 (φ) = 1[φ1 ,φ2 ] (φ) , (5.26a)
φ2 − φ1
φ − φj−1 φj+1 − φ
   
ψj (φ) = 1[φj−1 ,φj ] (φ) + 1[φj ,φj+1 ] (φ) j = 2, . . . , J,
φj − φj−1 φj+1 − φj
(5.26b)
φ − φJ
 
ψJ+1 (φ) = 1[φJ ,φJ+1 ] (φ) , (5.26c)
φJ+1 − φJ

where 1[a,b] (·) stands for the indicator function of the interval [a, b]. Substituting (5.26) in
s
(5.25), coefficients ln,j , msn,j and qjρ are decomposed as

1 1,+
ls,n = ls,n , m1s,n = ms,n
1,+
, 1
qρ,n 1,+
= qρ,n ,
j j,− j,+
ls,n = ls,n + ls,n , mjs,n = mj,− j,+
s,n + ms,n ,
j
qρ,n j,−
= qρ,n j,+
+ qρ,n , j = 2, . . . , J,
J+1 J+1,−
ls,n = ls,n , mJ+1 J+1,−
s,n = ms,n , J+1
qρ,n J+1,−
= qρ,n ,

where for j = 2, . . . , J + 1,
φj
φ − φj−1
Z  
j,− 2
lρ,n = sin φPn (cos φ) dφ, (5.27a)
φj−1 φj − φj−1
Z φj
φ − φj−1
 
0
mj,−
ρ,n = 2
sin φ cos φPn (cos φ) dφ, (5.27b)
φj−1 φj − φj−1
Z φj
φ − φj−1
 
qρj,− = (1 + cos φ) dφ, (5.27c)
φj−1 φj − φj−1
Z φj
φ − φj−1
 
j,−
lz,n = sin φ cos φPn (cos φ) dφ, (5.27d)
φj−1 φj − φj−1
Z φj
φ − φj−1
 
0
mj,−
z,n =− 3
sin φPn (cos φ) dφ, (5.27e)
φj−1 φj − φj−1

and for j = 1, . . . , J,
φj+1
φj+1 − φ
Z  
j,+ 2
lρ,n = sin φPn (cos φ) dφ, (5.28a)
φj φj+1 − φj
Z φj+1
φj+1 − φ
 
0
mj,+
ρ,n = 2
sin φ cos φPn (cos φ) dφ, (5.28b)
φj φj+1 − φj
 96

φj+1
φj+1 − φ
Z  
qρj,+ = (1 + cos φ) dφ, (5.28c)
φj φj+1 − φj
Z φj+1
φj+1 − φ
 
j,+
lz,n = sin φ cos φPn (cos φ) dφ, (5.28d)
φj φj+1 − φj
Z φj+1
φj+1 − φ
 
0
mj,+
z,n =− 3
sin φPn (cos φ) dφ. (5.28e)
φj φj+1 − φj

j,±
Explicit expressions for a few coefficients ls,n , mj,± j,±
s,n and qs are given next

j,− φj − φj−1 1 cos2 φj − cos2 φj−1

lρ,0 = − cos φj sin φj − ,
4 2 4(φj − φj−1 )
j,+ φj+1 − φj 1 cos2 φj+1 − cos2 φj
lρ,0 = + cos φj sin φj + ,
4 2 4(φj+1 − φj )
j,− 1 3 (2 + sin2 φj ) cos φj − (2 + sin2 φj−1 ) cos φj−1
lρ,1 = mj,−
ρ,1 = sin φj + ,
3 9(φj − φj−1 )
j,+ 1 3 (2 + sin2 φj+1 ) cos φj+1 − (2 + sin2 φj ) cos φj
lρ,1 = mj,+
ρ,1 = − sin φj − ,
3 9(φj+1 − φj )
j,− φj − φj−1 cos φj sin φj (6 cos2 φj − 7)
lρ,2 =− −
32 16
(cos φj − cos φj−1 ) 3(cos2 φj + cos2 φj−1 ) − 7
2 2


−
32(φj − φj−1 )
j,+ φj+1 − φj cos φj sin φj (6 cos2 φj − 7)
lρ,2 =− +
32 16
(cos φj+1 − cos φj ) 3(cos2 φj + cos2 φj+1 ) − 7
2 2


+
32(φj+1 − φj )
φj − φj−1 cos φj − cos φj−1
qρj,− = + sin φj + ,
2 φj − φj−1
φj+1 − φj cos φj+1 − cos φj
qρj,+ = − sin φj − ,
2 φj+1 − φj
j,− cos2 φj − sin2 φj sin φj cos φj − sin φj−1 cos φj−1
lz,0 =− + ,
4 4(φj − φj−1 )
j,+ cos2 φj − sin2 φj sin φj+1 cos φj+1 − sin φj cos φj
lz,0 = − ,
4 4(φj+1 − φj )
j,− 1 3 (2 + cos2 φj ) sin φj − (2 + cos2 φj−1 ) sin φj−1
lz,1 = − cos φj + ,
3 9(φj − φj−1 )
 97

j,+ 1 (2 + cos2 φj+1 ) sin φj+1 − (2 + cos2 φj ) sin φj

lz,1 = cos3 φj − ,
3 9(φj+1 − φj )
(2 + sin2 φj ) cos φj (6 + sin2 φj ) sin φj − (6 + sin2 φj−1 ) sin φj−1
mj,−
z,1 = − ,
3 9(φj − φj−1 )
(2 + sin2 φj ) cos φj (6 + sin2 φj+1 ) sin φj+1 − (6 + sin2 φj ) sin φj
mj,+
z,1 = − + ,
3 9(φj+1 − φj )
j,− 8(cos2 φj − 1)(3 cos2 φj + 1) + 7
lz,2 =−
64
cos φj sin φj (6 cos2 φj + 1) − cos φj−1 sin φj−1 (6 cos2 φj−1 + 1)
+
64(φj − φj−1 )
j,+ 8(cos2 φj − 1)(3 cos2 φj + 1) + 7
lz,2 =
64
cos φj+1 sin φj+1 (6 cos2 φj+1 + 1) − cos φj sin φj (6 cos2 φj + 1)
−
64(φj+1 − φj )

Integrals in (5.27) and (5.28) are computed analytically with the aid of symbolic com-
putation techniques. The resulting expressions become more and more unwieldy as n
increases, so only a few of them are provided by the way of example. We denote by
j
Ajs,n and Bs,n the coefficients An and Bn obtained by applying the algorithm of the pre-
vious subsection with y (A) = C js and y (B) = D js , respectively. Notice that as for each
j = 1, 2, . . . , J, J + 1 two cases need to be considered (s = ρ and s = z), the algorithm is
to be applied 2(J + 1) times in total. Additionally, let us define the vectors Ajs ∈ RN +1 and
B js ∈ RN +2 as those whose components are the coefficients Ajs,n and Bs,n
j
, respectively,
which by virtue of (4.40a) and (4.40b) are given by

Ajs = [Q(AA) ]−1 + [Q(AA) ]−1 Q(AB) [Q

e (BB) ]−1 [Q(AB) ]T [Q(AA) ]−1 C j


s
(5.29a)
− [Q(AA) ]−1 Q(AB) [Q
e (BB) ]−1 D j ,
s

e (BB) ]−1 [Q(AB) ]T [Q(AA) ]−1 C j + [Q

B js = −[Q e (BB) ]−1 D j . (5.29b)
s s
 98

Thus, using the definition of the DtN map (5.15) and combining with (4.26a) and (5.17b)
yields the approximations
 N
1 X j  (A) 
Mψj ρ̂(φ) ≈ − Aρ,n [τn ]r (φ)r̂ + [τn(A) ]rφ (φ)φ̂
R n=0
N
X  
j (B) (B)
+ Bρ,n [τn ]r (φ)r̂ + [τn ]rφ (φ)φ̂ ,
n=−1
 N
1 X j  (A) 
Mψj ẑ(φ) ≈ − Az,n [τn ]r (φ)r̂ + [τn(A) ]rφ (φ)φ̂
R n=0
N
X  
j (B) (B)
+ Bz,n [τn ]r (φ)r̂ + [τn ]rφ (φ)φ̂ .
n=−1

By replacing these expressions in (5.14), combining with (2.1a) and expanding, we arrive
at
N
X Z π  
[K bρρ ]ij ≈−R Ajρ,n [τn(A) ]r (φ) sin φ + [τn(A) ]rφ (φ) cos φ ψi (R, φ) sin φ dφ
n=0 π/2

N
X Z π  
j
−R Bρ,n [τn(B) ]r (φ) sin φ + [τn(B) ]rφ (φ) cos φ ψi (R, φ) sin φ dφ,
n=−1 π/2

N
X Z π  
[K bρz ]ij ≈−R Ajz,n [τn(A) ]r (φ) sin φ + [τn(A) ]rφ (φ) cos φ ψi (R, φ) sin φ dφ
n=0 π/2

N
X Z π  
j
−R Bz,n [τn(B) ]r (φ) sin φ + [τn(B) ]rφ (φ) cos φ ψi (R, φ) sin φ dφ,
n=−1 π/2

N
X Z π  
[K bzρ ]ij ≈−R Ajρ,n [τn(A) ]r (φ) cos φ − [τn(A) ]rφ (φ) sin φ ψi (R, φ) sin φ dφ
n=0 π/2

N
X Z π  
j
−R Bρ,n [τn(B) ]r (φ) cos φ − [τn(B) ]rφ (φ) sin φ ψi (R, φ) sin φ dφ,
n=−1 π/2
 99

N
X Z π  
[K bzz ]ij ≈−R Ajz,n [τn(A) ]r (φ) cos φ − [τn(A) ]rφ (φ) sin φ ψi (R, φ) sin φ dφ
n=0 π/2

N
X Z π  
j
−R Bz,n [τn(B) ]r (φ) cos φ − [τn(B) ]rφ (φ) sin φ ψi (R, φ) sin φ dφ,
n=−1 π/2

and combining with (5.23), these terms are conveniently reexpressed as

N
X N
X 
[K bρρ ]ij i
Ajρ,n i j
= R C iρ · Ajρ + D iρ · B jρ ,


≈R Cρ,n + Dρ,n Bρ,n (5.30a)
n=0 n=−1
N
X N
X 
[K bρz ]ij i
Ajz,n i j
= R C iρ · Ajz + D iρ · B jz ,


≈R Cρ,n + Dρ,n Bz,n (5.30b)
n=0 n=−1
N
X N
X 
[K bzρ ]ij i
Ajρ,n i j
= R C iz · Ajρ + D iz · B jρ ,


≈R Cz,n + Dz,n Bρ,n (5.30c)
n=0 n=−1
N
X N
X 
[K bzz ]ij i
Ajz,n i j
= R C iz · Ajz + D iz · B jz .


≈R Cz,n + Dz,n Bz,n (5.30d)
n=0 n=−1

It should be observed that vectors C iρ , C iz , D iρ , D iz actually do not depend on the radius R,

and so do vectors Aiρ , Aiz , B iρ , B iz . Hence, by virtue of (5.30), the entries of matrix K b are
linear in R.

5.4. Numerical experiments

5.4.1. Model problem

Throughout this section we consider a model problem whose exact solution is avail-
able. Let us assume the perturbed boundary Γp to be a quarter circle of radius a > 0. In
this case, it holds in axisymmetric spherical coordinates that

Ω = {(r, φ) : a < r < ∞, π/2 < φ < π},

Γp = {(r, φ) : r = a, π/2 < φ < π}.

 100

We define the function f = fr r̂ + fφ φ̂ : Γp → R2 as

fr (φ) = 2µc 1 − 2ν + 2(2 − ν) cos φ , (5.31a)
 
cos φ sin φ
fφ (φ) = 2µc (1 − 2ν) , (5.31b)
1 − cos φ

where c is a scale factor. Then the displacement u = ur r̂ + uφ φ̂ : Ω → R2 , defined as

c a2

ur (r, φ) = − 1 − 2ν + 4(1 − ν) cos φ , (5.32a)
r
2
1 − 2ν
 
ca
uφ (r, φ) = 3 − 4ν − sin φ, (5.32b)
r 1 − cos φ
is an exact solution of (5.1) with f given by (5.31). This solution is obtained from (4.20a)
with R replaced by a and setting An = Bn = 0 for all n = 0, 1, 2, . . . and B−1 = 2µc a.
The stress components associated with (5.32) are obtained by setting these same values of
coefficients An and Bn in (4.20b) with R replaced by a, or simply by substituting (5.32) in
(2.8), yielding

2µc a2

σr (r, φ) = 1 − 2ν + 2(2 − ν) cos φ , (5.33a)
r2
2µc a2 (1 − 2ν) cos2 φ
 
σφ (r, φ) = , (5.33b)
r2 1 − cos φ
2µc a2 (1 − 2ν) 1 + cos φ − cos2 φ
 
σθ (r, φ) = − , (5.33c)
r2 1 − cos φ
2µc a2 (1 − 2ν) cos φ sin φ
 
σrφ (r, φ) = . (5.33d)
r2 1 − cos φ
In axisymmetric cylindrical coordinates, (5.32) reads u = uρ ρ̂ + uz ẑ, with

c a2 ρ 1 − 2ν
 
z
uρ (ρ, z) = − 2 + , (5.34a)
(ρ + z 2 )1/2 (ρ2 + z 2 )1/2 − z ρ2 + z 2
c a2 z2
 
uz (ρ, z) = − 2 2(1 − ν) + 2 , (5.34b)
(ρ + z 2 )1/2 ρ + z2
and associated stress components given by

2µc a2 1 − 2ν 3ρ2 z
 
σρ (ρ, z) = 2 + , (5.35a)
(ρ + z 2 )1/2 (ρ2 + z 2 )1/2 − z (ρ2 + z 2 )2
 101

2µc a2 (1 − 2ν)
 
1 z
σθ (ρ, z) = − + , (5.35b)
(ρ2 + z 2 )1/2 (ρ2 + z 2 )1/2 − z ρ2 + z 2
6µc a2 z 3
σz (ρ, z) = , (5.35c)
(ρ2 + z 2 )5/2
6µc a2 ρz 2
σρz (ρ, z) = , (5.35d)
(ρ2 + z 2 )5/2
which follow from substituting (5.34) in (2.7). This exact solution will be used as a bench-
mark to test the performance of the DtN FEM approach.

5.4.2. Implementation aspects

To apply the DtN FEM to solve the model problem, the radius of the artificial boundary
R must be, of course, greater than a. The resulting computational domain Ωi is described
in axisymmetric spherical coordinates (r, φ) as

Ωi = {(r, φ) : a < r < R, π/2 < φ < π}.

In general, in any DtN FEM approach, the radius R is to be chosen large enough so that any
possible irregularity (anisotropy, inhomogeneity) is enclosed by ΓR , but at the same time
small enough to minimise the size of Ωi and thus the number of finite elements considered
(cf. (Givoli & Vigdergauz, 1993)). In our case, for the purpose of solving the model prob-
lem in order to test the accuracy of the DtN FEM, it is reasonable to assume, for instance,
R = (3/2)a. In addition, as we have deduced a linear dependence of the matrix K b on R,
the accuracy of the method should not be strongly affected by the location of ΓR . Thus,
from now on, R is assumed to be fixed at the same value.

In order to maintain a constant mesh quality, we consider structured triangular meshes

of Ωi generated as follows: We assume equispaced partitions of intervals [a, R] and [π/2, π]
consisting respectively of I and J segments, delimited by nodes satisfying a = r1 < r2 <
. . . < rI < rI+1 = R and π/2 = φ1 < φ2 < . . . < φJ < φJ+1 = π (the latter partition
already introduced in Section 5.3), in such a way that ri = a + (i − 1)(R − a)/I and
φj = (π/2)(1+(j−1)/J). The cartesian product of both partitions gives rise to a structured
quadrilateral mesh of Ωi , which consists of (I + 1)(J + 1) nodes and IJ elements. If each
 102

quadrilateral element is split by its diagonal, two triangles are created. Then the resulting
triangular mesh consists of (I + 1)(J + 1) nodes and 2IJ elements in total. Given that
we have assumed R = (3/2)a, and in order to ensure relatively good quality triangles, it is
reasonable to set J = 4I. The mesh size h is calculated using relatively simple geometrical
considerations, arriving at the formula

R
   π 1/2
h= 1 + 6I(3I − 1) 1 − cos . (5.36)
3I 8I
To choose an optimal value of the truncation order N for fixed R and h, we use the simple
crude formula first proposed in (Keller & Givoli, 1989) for the Laplace equation, and sub-
sequently modified and used in (Givoli & Keller, 1989) and (Givoli & Vigdergauz, 1993)
for the 2D elasticity equation. According to this formula, the optimal N is estimated as
 
ln (h/a)
Nopt = − (p + 1) , (5.37)
ln (R/a)

where d·e stands for the ceiling function and p is the highest degree of complete polynomial
in V h (in this case p = 1 since we are considering P 1 finite elements). Substituting R =
(3/2)a and approximating ln(3/2) ≈ 2/5, (5.37) becomes
   h 
Nopt = − 2 + 5 ln . (5.38)
R
This formula provides a good estimation of N , as we shall see below.

5.4.3. Results and accuracy

Next, we present numerical results obtained by solving the model problem using our
DtN FEM procedure. We assume a = 600 m and R = (3/2)a = 900 m. Six meshes of
Ωi for different values of I are considered, whose parameters are summarised in Table 5.1.
The optimal series truncation order Nopt is computed from (5.38) and is shown in the last
column of the table. The meshes corresponding to I = 5 and I = 10 are depicted in
Figure 5.4 by the way of example.

We assume an elastic solid with Young’s modulus E = 70 GPa and Poisson’s ratio
ν = 0.3. The Lamé constants are immediate from E and ν through the standard formulae
 103

Table 5.1. Parameters of the structured triangular meshes considered.

I J No. nodes No. elements hm Nopt

3 12 52 72 149.39679 7
5 20 126 200 90.89000 10
10 40 451 800 45.90566 13
18 72 1387 2592 25.61552 16
32 128 4257 8192 14.44309 19
60 240 14701 28800 7.71394 22

λ = Eν/(1 + ν)/(1 − 2ν) and µ = E/2/(1 + ν). In order to get displacements and stresses
of physically realistic magnitudes, we set c = 10−4 . Then the model problem is solved
by the DtN FEM in the six triangular meshes described in Table 5.1. For each mesh, the
optimal series truncation order Nopt indicated in the same table is assumed in all the de-
velopment of Section 5.3. The results calculated in the finest mesh (I = 60) are presented
next. Figures 5.5 shows the components of the computed displacement vector uh . The

(a) (b)

Figure 5.4. Two of the considered structured triangular meshes. (a) I = 5.

(b) I = 10.
 104

associated stress tensor σ h is numerically computed from uh by substituting the displace-

ment components in (2.7), with the derivatives calculated numerically in the mesh. The
components of σ h are presented in Figure 5.6.

(a) (b)

Figure 5.5. Computed displacement components. (a) uhρ (ρ, z). (b) uhz (ρ, z).

In order to test preliminarily the effectiveness of our DtN FEM, the numerical solu-
tions calculated in each mesh are compared against the exact solution given by (5.34) and
(5.35) at some specific points of the domain. Table 5.2 presents numerical values of some
displacement and stress components evaluated at the extreme points of ΓR . From this table
we see that, as the mesh size decreases, the values of the numerical solution get closer to the
values of the exact one at the considered points. Moreover, in the case of the displacement,
this approximation is faster.

To study more rigorously the accuracy and convergence of the DtN FEM procedure,
we employ the relative error between the numerical and the exact solution, defined as

kuh − Πh uk0,Ωi
Eu (h) = ,
kΠh uk0,Ωi
 105

(a) (b)

(c) (d)

Figure 5.6. Computed stress components. (a) σρh (ρ, z). (b) σθh (ρ, z). (c)
σzh (ρ, z). (d) σρz
h (ρ, z).

where k · k0,Ωi is the usual L2 -norm in Ωi and Πh u denotes the Lagrange interpolation
of the exact solution u over the mesh of size h. Figure 5.7 presents a log-log plot of Eu
in function of h, where we observe that the relative error decays as mesh size diminishes.
Therefore, the whole numerical solution computed by the DtN FEM converges to the exact
solution as the mesh is refined, which confirms the effectiveness of the method. In fact, the
 106

Table 5.2. Some components of the solution evaluated at points (0, R) and
(−R, 0).

Solution uρ (0, R) m uz (−R, 0) m σρ (0, R) MPa σρ (−R, 0) MPa σz (−R, 0) MPa

I=3 -0.01709 -0.09664 0.29365 0.08399 -7.58848
I=5 -0.01648 -0.09633 0.59049 0.20644 -7.49848
I = 10 -0.01615 -0.09613 0.78277 0.37847 -7.33654
I = 18 -0.01606 -0.09605 0.85841 0.42163 -7.27821
I = 32 -0.01602 -0.09602 0.89876 0.44859 -7.23198
I = 60 -0.01601 -0.09601 0.92393 0.46549 -7.20743
Exact -0.01600 -0.09600 0.95726 0.47863 -7.17949

relative error in the finest mesh is approximately 0.003%. This good agreement between
the numerical and the exact solution is achieved with a relatively low number of terms in
the series of the DtN map (N = 22). In addition, the graph of Eu in the log-log plot is
nearly a straight line, which suggests that the DtN FEM has a constant rate of convergence.
To estimate it, we compute the slope of this line, using for this the relative errors associated
with the two finest meshes, since in these cases the method yields the best approximations.
The computed slope is 2.00257. A slope nearly 2 in the relative error of the solution predicts
that the rate of convergence of the method is 2. Hence, the numerical evidence indicates
that the DtN FEM presented throughout this work has second-order accuracy.
 107

1
10

0
10
Relative error [%]

−1
10

−2
10

−3
10 0 1 2 3
10 10 10 10
h [m]

Figure 5.7. Log-log plot of Eu in function of h.

 108

References

Abramowitz, M., & Stegun, I. (1965). Handbook of mathematical functions with

formulas, graphs, and mathematical tables. New York: Dover Publications.

Aladl, U. E., Deakin, A. S., & Rasmussen, H. (2002). Nonreflecting boundary con-
dition for the wave equation. J. Comput. Appl. Math., 138, 309–323.

Amenzade, Y. A. (1979). Theory of elasticity. Moscow: Mir Publishers.

Andrews, L. C. (1985). Special functions for engineers and applied mathematicians.

New York: Macmillan Publishing Co.

Arfken, G. B., & Weber, H. J. (2001). Mathematical methods for physicists. Lon-
don: Harcourt/Academic Press.

Bell, W. W. (1968). Special functions for scientists and engineers. London: Van
Nostrand.

Björck, A. (1996). Numerical methods for least square problems. Philadelphia:

SIAM.

Brown, J. W., & Churchill, R. (2006). Fourier series and boundary value problems.
New York: McGraw-Hill.

Cerit, M., Genel, K., & Eksi, S. (2009, October). Numerical investigation on stress
concentration of corrosion pit. Engineering Failure Analysis, 16(7), 2467–2472. doi:
10.1016/j.engfailanal.2009.04.004

Deakin, A. S., & Rasmussen, H. (2001). Nonreflecting boundary condition for the
Helmholtz equation. Comput. Math. Appl., 41, 307–318.
 109

Eubanks, R. (1953). Stress concentration due to a hemispherical pit at a free surface.

Journal of Applied Mechanics, 21, 57.

Eubanks, R., & Sternberg, E. (1956). On the completeness of the boussinesq-

papkovich stress functions. Jour. Rational Mech. Anal., 5, 735 - 746.

Eubanks, R. A. (1954). Stress concentration due to a hemispherical pit at a free

surface. J. Appl. Mech., 21, 57–62.

Fujita, T., Sadayasu, T., Tsuchida, E., & Nakahara, I. (1978). Stress concentration
due to a hemispherical pit at a free surface of a thick plate under all-around tension.
Bulletin of JSME, 21(154), 561-565.

Fujita, T., Tsuchida, E., & Nakahara, I. (1982, April). Asymmetric problem of a
semi-infinite body having a hemispherical pit under uniaxial tension. Journal of Elas-
ticity, 12(2), 177–192. doi: 10.1007/BF00042214

Gächter, G., & Grote, M. J. (2003). Dirichlet-to-Neumann map for three-

dimensional elastic waves. Wave Motion, 37, 293–311.

Givoli, D. (1992). Numerical methods for problems in infinite domains. Amster-

dam: Elsevier.

Givoli, D. (1999a). Exact representations on artificial interfaces and applications in

mechanics. Appl. Mech. Rev., 52(11), 333–349.

Givoli, D. (1999b). Recent advances in the DtN FE method. Arch. Comput. Method
E., 6(2), 71–116.

Givoli, D., & Keller, J. B. (1989). A finite element method for large domains. Com-
put. Methods Appl. Mech. Engrg., 76, 41-66.

Givoli, D., & Keller, J. B. (1990). Non-reflecting boundary conditions for elastic
waves. Wave Motion, 12, 261–279.
 110

Givoli, D., & Vigdergauz, S. (1993). Artificial boundary conditions for 2D problems
in geophysics. Comput. Methods Appl. Mech. Engrg., 110, 87-101.

Golub, G. H., & Loan, C. F. V. (1996). Matrix computations. Baltimore: Johns

Hopkins University Press.

Gradshteyn, I. S., & Ryzhik, I. M. (2007). Table of integrals, series and products.
Amsterdam: Elsevier/Academic Press.

Griffiths, D., & Lane, P. (1999). Slope stability analysis by finite elements. Geotech-
nique, 49(3), 387–403.

Grote, M. J. (2000). Nonreflecting boundary conditions for elastodynamic scatter-

ing. J. Comput. Phys., 161, 331–353.

Grote, M. J., & Keller, J. B. (1995a). Exact nonreflecting boundary condition for the
time dependent wave equation. SIAM J. Appl. Math., 55, 280–297.

Grote, M. J., & Keller, J. B. (1995b). On nonreflecting boundary conditions. J.

Comput. Phys., 122, 231–243.

Grote, M. J., & Keller, J. B. (1996). Nonreflecting boundary conditions for time-
dependent scattering. J. Comput. Phys., 127, 52–65.

Grote, M. J., & Keller, J. B. (2000). Exact nonreflecting boundary condition for
elastic waves. SIAM J. Appl. Math., 60, 803–819.

Han, H., Bao, W., & Wang, T. (1997). Numerical simulation for the problem of
infinite elastic foundation. Comput. Methods Appl. Mech. Engrg., 147, 369–385.

Han, H. D., & Wu, X. N. (1985). Approximations of infinite boundary conditions

and its application to finite elements methods. J. Comput. Math., 3, 179–192.
 111

Han, H. D., & Wu, X. N. (1992). The approximation of the exact boundary condi-
tions at an artificial boundary for linear elastic equations and its application. Math.
Comput., 59(199), 21–37.

Han, H. D., & Wu, X. N. (2013a). Artificial boundary method. Beijing-Berlin-

Heidelberg: Tsinghua University Press and Springer-Verlag.

Han, H. D., & Wu, X. N. (2013b). A survey on artificial boundary method. Sci.
China Math., 56, 2439–2488.

Han, H. D., & Zheng, C. X. (2005). Artificial boundary method for the three-
dimensional exterior problem of elasticity. J. Comput. Math., 23, 603-–618.

Harari, I., & Shohet, Z. (1998). On non-reflecting boundary conditions in unbounded

elastic solids. Comput. Methods Appl. Mech. Engrg., 163, 123—139.

Huttelmaier, H. P., & Glockner, P. G. (1985). Stresses and displacements due to

underground mining using a finite element procedure. Geotechnical and Geological
Engineering, 3(1), 49 - 63. doi: DOI:10.1007/BF00881341

Keller, J. B., & Givoli, D. (1989). Exact non-reflecting boundary conditions. J.

Comput. Phys., 82, 172–192.

Kulhawy, F. H. (1974). Finite element modeling criteria for underground openings

in rock. International Journal of Rock Mechanics and Mining Sciences and Geome-
chanics Abstracts, 11(12), 465 - 472. doi: DOI:10.1016/0148-9062(74)91996-2

Multiphysics, C. (1994). Comsol. Inc., Burlington, MA, www. comsol. com.

Ou, C.-Y., Chiou, D.-C., & Wu, T.-S. (1996). Three-Dimensional Finite Ele-
ment Analysis of Deep Excavations (Vol. 122) (No. 5). doi: 10.1061/(ASCE)0733
-9410(1996)122:5(337)
 112

Ovsyannikov, A. S., & Starikov, V. A. (1987, September). Numerical solution of an

axisymmetric problem of the theory of elasticity for a half-space with a notch. Soviet
Applied Mechanics, 23(9), 849–852. doi: 10.1007/BF00887788

Sadd, M. H. (2005). Elasticity: Theory, applications, and numerics. Boston: Else-

vier Butterworth-Heinemann.

Szabo, B., & Babuška, I. (1991). Finite element analysis. New York: John Wiley &
Sons.

Turnbull, A., Horner, D., & Connolly, B. (2009, March). Challenges in modelling
the evolution of stress corrosion cracks from pits. Engineering Fracture Mechanics,
76(5), 633–640. doi: 10.1016/j.engfracmech.2008.09.004
 113

6. APPENDIX

6.1. Some properties of Legendre polynomials

The Legendre polynomial Pn (cos φ) as a function of φ is a solution of the Legendre’s

differential equation, given by
 
1 d d
sin φ Pn (cos φ) + n(n + 1)Pn (cos φ) = 0. (6.1)
sin φ dφ dφ
Some useful recurrence relations fulfilled by the Legendre polynomials and their deriva-
tives are (cf. (Andrews, 1985; Bell, 1968)):

(n + 1)Pn+1 (cos φ) − (2n + 1) cos φ Pn (cos φ) + nPn−1 (cos φ) = 0, (6.2a)

0 0
Pn+1 (cos φ) − Pn−1 (cos φ) = (2n + 1)Pn (cos φ), (6.2b)

sin2 φ Pn0 (cos φ) = (n + 1) cos φ Pn (cos φ) − Pn+1 (cos φ) ,

(6.2c)

sin2 φ Pn0 (cos φ) = n Pn−1 (cos φ) − cos φ Pn (cos φ) ,

(6.2d)

where n ≥ 1. The Legendre polynomials also satisfy the following properties for n ≥ 0:
(−1)n (2n)!
P2n (0) = , (6.3a)
22n (n!)2
P2n+1 (0) = 0, (6.3b)
2n + 1
P2n+2 (0) = − P2n (0), (6.3c)
2n + 2
0
P2n (0) = 0, (6.3d)
0
P2n+1 (0) = (2n + 1) P2n (0). (6.3e)

Properties (6.3a) and (6.3b) are standard and can be found, for instance, in (Andrews, 1985)
or (Bell, 1968). Property (6.3c) is a consequence of formula (6.2a). Property (6.3d) follows
from (6.2c) and (6.3b). Property (6.3e) is obtained from (6.2d). Moreover, the Legendre
polynomials fulfil the following integral formulae for n ≥ 0 and k ≥ 0:
Z π
δn,k
P2n (cos φ) P2k (cos φ) sin φ dφ = , (6.4a)
π/2 4n + 1
 114

π
δn,k
Z
P2n+1 (cos φ) P2k+1 (cos φ) sin φ dφ = , (6.4b)
π/2 4n + 3
π
(2k + 1)P2n (0)P2k (0)
Z
P2n (cos φ) P2k+1 (cos φ) sin φ dφ = − , (6.4c)
π/2 (2k + 1 − 2n)(2k + 2 + 2n)
π
2n(2n + 1)δn,k
Z
0 0
P2n (cos φ) P2k (cos φ) sin3 φ dφ = , (6.4d)
π/2 4n + 1
π
(2n + 1)(2n + 2)δn,k
Z
0 0
P2n+1 (cos φ) P2k+1 (cos φ) sin3 φ dφ = , (6.4e)
π/2 4n + 3
π
2n(2n + 1)(2k + 1)P2n (0)P2k (0)
Z
0 0
P2n (cos φ) P2k+1 (cos φ) sin3 φ dφ = − . (6.4f)
π/2 (2k + 1 − 2n)(2k + 2 + 2n)

Formulae (6.4a), (6.4b), and (6.4c) can be deduced from (Gradshteyn & Ryzhik, 2007),
Subsection 7.221. To obtain (6.4c) it is, in addition, necessary to combine with (6.3a).
Formulae (6.4d), (6.4e), and (6.4f) follow from formulae (6.4a), (6.4b), and (6.4c), re-
spectively, together with the Legendre’s differential equation (6.1). The following two
additional integral formulae are also valid
Z π
0
P2n (cos φ)(1 + cos φ) sin φ dφ = P2n (0), (6.5a)
π/2
π
P2n (0)
Z
0
P2n+1 (cos φ)(1 + cos φ) sin φ dφ = , (6.5b)
π/2 2n + 2

which are obtained by integrating by parts and combining with (6.4a) and (6.4c), respec-
tively.

6.1.1. Expansion of the odd terms in pair terms from the Legendre polynomials

We need to show that ∀k = 0, 1, . . . and −1 ≤ x ≤ 0

∞
(n)
X
P2k+1 (x) = −(2k + 1)P2k (0) ωk P2n (x), (6.6)
n=0

where
(n) (4n + 1)P2n (0)
ωk = . (6.7)
(2k + 1 − 2n)(2k + 2 + 2n)
 115

The set {P2n (x)}∞

n=0 is an orthogonal basis (Brown & Churchill, 2006) in [−1, 0] (and

in [0, 1]) such that

0
δnm
Z
P2n (x)P2m (x)dx = , (6.8)
−1 4n + 1
then we can expand P2k+1 (x) in a Legendre series
∞
X Z 0
P2k+1 (x) = (4n + 1) P2k+1 (t)P2k (t)dtP2n (x). (6.9)
n=0 −1

Then, our problem becomes

Z 0
(2k + 1)P2k (0)P2n (0)
P2k+1 (x)P2n (x)dx = − ∀n, k = 0, 1 . . . (6.10)
−1 (2k + 1 − 2n)(2k + 2 + 2n)
which can also be written as
Z 1
(2k + 1)P2k (0)P2n (0)
P2n (x)P2k+1 (x)dx = ∀n, k = 0, 1 . . . (6.11)
0 (2k + 1 − 2n)(2k + 2 + 2n)
We define
Z 1
An,k = P2n (x)P2k+1 (x)dx. (6.12)
0

This will be proved by induction. We have the following hypothesis

(i) The statement (6.12) is valid for n = k = 0.

(ii) If the statement holds for some N > 0 for any n, k with n + k = N , then the
statement also holds for any n, k when n + k = N + 1.

Basis: Show that the statement holds for n = k = 0.

Z 1 Z 1 1
x2 1
A0,0 = P0 (x)P1 (x)dx = xdx = = .
0 0 2 0 2

On the other hand if n = k = 0

(2k + 1)P2k (0)P2n (0) 1·1·1 1
= = . (6.13)
(2k + 1 − 2n)(2k + 2 + 2n) 1·2 2

The two sides are equal, so the statement is true for n = k = 0. Thus it has been shown
that A0,0 holds.
 116

Inductive step: Show that if An,k with n + k = N holds, then also An,k with n + k =
N + 1 holds, where
(2k + 1)P2k (0)P2n (0)
An,k = . (6.14)
(2k + 1 − 2n)(2k + 2 + 2n)
This only can occur in two cases, (n + 1, k) and (n, k + 1). Then we need to prove two
statements
1
(2k + 1)P2k (0)P2n+2 (0)
Z
An+1,k = P2n+2 (x)P2k+1 (x)dx = , (6.15)
0 (2k − 1 − 2n)(2k + 4 + 2n)
and
1
(2k + 3)P2k+2 (0)P2n (0)
Z
An,k+1 = P2n (x)P2k+3 (x)dx = . (6.16)
0 (2k + 3 − 2n)(2k + 4 + 2n)
The Legendre polynomials satisfy the recurrence relations

(n + 1)Pn+1 (x) − (2n − 1)xPn (x) + nPn−1 (x) = 0,

(1 − x2 )Pn0 (x) = −nxPn (x) + nPn−1 (x) = (n + 1) xPn (x) − Pn+1 .

(6.17)

Combining this we obtain

(4n + 3) 0
P2n+2 (x) = P2n (x) − (1 − x2 )P2n+1 (x), (6.18)
(2n + 1)(2n + 2)

multiplying by P2k+1 (x) and integrating between 0 and 1 we have

Z 1
(4n + 3) 0
An+1,k = An,k − (1 − x2 )P2n+1 (x)P2k+1 (x)dx. (6.19)
(2n + 1)(2n + 2) 0

Using integration by parts we obtain

Z 1 Z 1
2 0 0
P2n+1 (x) (1 − x2 )P2k+1


(1 − x )P2n+1 (x)P2k+1 (x)dx = − (x) − 2xP2k+1 (x) dx.
0 0
(6.20)

Using this we obtain

1
(4n + 3)
Z
0
(1 − x2 )P2n+1


An+1,k = An,k + (x) − 2P2n+1 (x) P2k+1 (x)dx.
(2n + 1)(2n + 2) 0
(6.21)
 117

With (6.17), this becomes

 Z 1
(4n + 3)
An+1,k = An,k + (2k + 1) P2n+1 (x)P2k (x)dx
(2n + 1)(2n + 2) 0
Z 1 
−(2k + 3) xP2n+1 (x)P2k+1 (x)dx (6.22)
0

Again using the recurrence relations (6.17) multiplying by P2k+1 (x) and integration
between 0 and 1, and using the definition of Ak,n we obtain
Z 1
(2n + 1) (2n + 2)
xP2n+1 (x)P2k+1 (x)dx = An,k + An+1,k . (6.23)
0 (4n + 3) (4n + 3)

This yields

(4n + 3) (2k + 3)(2n + 1)
An+1,k = An,k + (2k + 1)Ak,n − An,k
(2n + 1)(2n + 2) (4n + 3)

(2k + 3)(2n + 2)
− An+1,k . (6.24)
(4n + 3)

Rearranging in a suitable way gives us

(2n + 1)(2k + 1 − 2n) (4n + 3)(2k + 1)
An+1,k = − An,k + Ak,n . (6.25)
(2n + 2)(2k + 4 + 2n) (2n + 2)(2k + 4 + 2n)

Applying (6.14) in this expression and using that (2n+2)P2n+2 (0) = −(2n+1)P2n (0)
yields
(2k + 1) (2n + 1)
An+1,k = P2k (0) − P2n (0)
(2k − 1 − 2n)(2k + 4 + 2n) (2n + 2)
(2k + 1)P2k (0)P2n+2 (0)
= , (6.26)
(2k − 1 − 2n)(2k + 4 + 2n)

thereby showing that indeed An+1,k holds. Next we need to show te same for An,k+1 , in
order to do that we use (6.17)
(4k + 5) 0
P2k+3 (x) = P2k+1 − (1 − x2 )P2k+2 (x). (6.27)
(2k + 2)(2k + 3)
 118

Multiplying by P2n (x) and integration between 0 and 1 gives us

Z 1
(4n + 5) 0
An,k+1 = An,k − (1 − x2 )P2k+2 (x)P2n (x)dx. (6.28)
(2k + 2)(2k + 3) 0

Since
Z 1
0
(1 − x2 )P2k+2 (x)P2n (x)dx = −P2n (0)P2k+2 (x)
0
Z 1
0
(1 − x2 )P2n


− (x) − 2xP2n (x) P2k+2 (x)dx. (6.29)
0

Hence

(4k + 5)
An,k+1 = An,k + P2n (0)P2k+2 (0)
(2k + 2)(2k + 3)
Z 1 Z 1 
2 0
+ (1 − x )P2n (x)P2k+2 (x)dx − 2 xP2n (x)P2k+2 (x)dx . (6.30)
0 0

Using (6.17) yields


(4k + 5)
An,k+1 = An,k + P2n (0)P2k+2 (0)
(2k + 2)(2k + 3)
Z 1 Z 1 
+ (2n − 1) xP2n (x)P2k+2 (x)dx − (2n + 1) P2n+1 (x)P2k+2 (x)dx .
0 0
(6.31)

Again using the recurrence relations (6.17) multiplying by P2n (x) and integration be-
tween 0 and 1, and using the definition of Ak+1,n we obtain
Z 1
(2k + 3) (2k + 2)
xP2n (x)P2k+2 (x)dx = An,k+1 + An,k , (6.32)
0 (4k + 5) (4k + 5)
hence

(4k + 5)
An,k+1 = An,k + P2n (0)P2k+2 (0)
(2k + 2)(2k + 3)
(2n − 1)(2k + 3) (2n − 1)(2k + 2)

+ An,k+1 + An,k − (2n + 1)Ak+1,n . (6.33)
(4k + 5) (4k + 5)
 119

Rearranging in a suitable way gives us

(2k + 2)(2k + 2 + 2n) (4k + 5)(2n + 1)
An,k+1 = An,k − Ak+1,n
(2k + 3)(2k + 3 − 2n) (2k + 3)(2k + 3 − 2n)
(4k + 5)
+ P2n (0)P2k+2 (0). (6.34)
(2k + 3)(2k + 3 − 2n)

If we interchange k by n in (6.25) we obtain

(2k + 2)(2k + 3)(2k + 3 − 2n)(2k + 4 + 2n)An,k+1

+ (2k + 1)(4k + 5)(2k − 1 − 2n)(2n + 1)Ak,n

+ (4k + 3)(4k + 5)(2n + 1)2 − (2k + 2)2 (2k + 2 + 2n)(2k + 4 + 2n) An,k

= −(2k + 1)(4k + 5)(2k + 4 + 2n)P2n (0)P2k (0). (6.35)

Using (6.14) yields

(2k + 3)P2k+2 (0)P2n (0)
An,k+1 = . (6.36)
(2k + 3 − 2n)(2k + 4 + 2n)

End of proof.