(Heinz Antes) A Short Course On Boundary Element Methods

A Short Course on Boundary Element Methods
Prof. Dr. rer. nat. Heinz Antes
Institut für Angewandte Mechanik

Spielmannstraße 11
38106 Braunschweig
March 29, 2005

Contents
1 Introduction 4
2 Mathematical Preliminaries 7
2.1 Some notations and de…nitions . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Indicial and symbolic notation . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Contraction and di¤erent products of tensors . . . . . . . . . . . . . 8
2.1.3 The Euclidian distance r and its derivatives . . . . . . . . . . . . . 9
2.2 The Gauss theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 The gradient theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 The divergence theorem . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Generalized Gauss theorems . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Integration by parts - Green’s identities . . . . . . . . . . . . . . . . . . . . 12
2.4 Fundamental solutions of di¤erential equations . . . . . . . . . . . . . . . . 14
2.4.1 Adjoint and self-adjoint operators . . . . . . . . . . . . . . . . . . . 14
2.4.2 The Dirac -function . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.3 Green’s functions of boundary value problems . . . . . . . . . . . . 20
2.4.4 Ordinary di¤erential equations with constant coe¢ cients . . . . . . 21
2.4.5 Scalar partial di¤erential equations with constant coe¢ cients . . . . 23
2.5 Singular integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5.1 Weak singularities - improper integrals . . . . . . . . . . . . . . . . 26
2.5.2 The Cauchy Principal Value of strongly singular integrals . . . . . . 27
2.5.3 Cauchy Principal Value integrals in boundary integral equations . . 28
3 Transformation of Di¤erential Equations to Integral Equations 32

3.1 Introductary 1-d problems: Transformation of ordinary di¤erential equations 32
3.1.1 Integral equations by direct integration . . . . . . . . . . . . . . . . 32
3.1.2 Direct integral equations by the method of weighted residuals . . . 35
3.1.3 Integral formulation with Green’s functions . . . . . . . . . . . . . . 50
3.1.4 Indirect integral formulations: the singularity method . . . . . . . . 56
3.2 2-d and 3-d problems: Transformation of partial di¤erential equations . . . 60
3.2.1 Direct integral equations by the method of weighted residuals . . . 60
3.2.2 Indirect integral formulations: the singularity method . . . . . . . . 70
3.2.3 Integral formulation with Green’s functions . . . . . . . . . . . . . . 76
2
3
4 Numerical solution of boundary integral equations: The boundary ele-

ment method 80
4.1 Approximation of the boundary and of boundary states . . . . . . . . . . . 80
4.1.1 On boundary curves in R2 . . . . . . . . . . . . . . . . . . . . . . . 80
4.1.2 On boundary surfaces in R3 . . . . . . . . . . . . . . . . . . . . . . 82
4.2 Integration over boundary elements . . . . . . . . . . . . . . . . . . . . . . 84
4.2.1 Elements on boundary curves . . . . . . . . . . . . . . . . . . . . . 85
4.2.2 Elements on boundary surfaces . . . . . . . . . . . . . . . . . . . . 87
4.3 Boundary element equations by point collocation . . . . . . . . . . . . . . 89
4.3.1 Approximation by constant shape functions . . . . . . . . . . . . . 90
4.3.2 Approximation by linear and higher grade shape functions . . . . . 93
5 Appendices 99
5.1 A: Exercise Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2 B: Analytic integration of singular boundary integrals . . . . . . . . . . . . 113
5.2.1 Analytic integration in the case of logarithmic kernels . . . . . . . . 113
5.2.2 Analytic integration in the case of (1/r)- kernels . . . . . . . . . . . 115
1 Introduction
Engineers who are familiar with …nite elements very often ask why it is necessary to
develop yet another computational technique. The answer is that …nite elements have
been proved to be inadequate or ine¢ cient in many engineering applications and what
is perhaps more important is in many cases cumbersome to use and hence di¢ cult to
implement in Computer Aided Engineering systems. Finite Element (FE) analysis is still
a comparatively slow process due to the need to de…ne and rede…ne meshes in the piece
or domain under study.
Boundary elements (BE) have emerged as a powerful alternative to …nite elements
particularly in cases where better accuracy is required due to problems such as stress
concentration or where the domain extends to in…nity. The most important feature of
boundary elements, however, is that di¤erent to the …nite domain methods as, e.g., the
…nite di¤erence method or the …nite element method, the methodology of formulating
boundary value problems as boundary integral equations describes problems only by equa-
tions with known and unknown boundary states.Hence, it only requires discretization of
the surface rather than the volume, i.e., the dimension of problems is reduced by one.
Consequently, the necessary discretization e¤ort is mostly much smaller and , moreover,
meshes can easily be generated and design changes do not require a complete remeshing.
The BE method is especially advantageous in the case of problems with in…nite or
semi-in…nite domains, e.g., so-called exterior domain problems: there, although only the
…nite surface of the in…nite domain has to be discretized, the solution at any arbitrary
point of the domain can be found after determining the unknown boundary data.
To be objective, the features of the BE method should be compared to its main rival,
the FE method. Its advantages and disadvantages can be summarized as follows
Advantages of the Boundary Element Method

1. Less data preparation time: This is adirect result of the ’surface-only’modelling. Thus,
the analyst’s time required for data preparation and data checking for a given problem
should be greatly reduced. Furthermore, subsequent changes in meshes are made easier.
2. High resolution of stress: Stresses are accurate because no further approximation
is imposed on the solution at interior points, i.e., solution is exact and fully continuous
inside the domain.
3. Less computer time and storage: For the same level of accuracy, the BE method
uses a lesser number of nodes and elements (but a fully populated matrix), i.e., to achieve
comparable accuracy in stress values, FE meshes would need more boundary divisions
than the equivalent BE meshes.
4. Less unwanted information: In most engineering problems, the ’worst’ situation
(such as fracture, stress concentration, thermal shocks a.s.o.) usually occur on the surface.
4
5
Thus, modelling an entire three-dimensional body with …nite elements and calculating
stress (or other states) at every nodal point is very ine¢ cient because only a few of these
values will be incorporated in the design analysis. Therefore, using boundary elements is
a very e¤ective use of computing resources, and, furthermore, since internal points in BE
solutions are optional, the user can focus on a particular interior region rather than the
whole interior.
Disadvantages of the BE method

1. Unfamiliar mathematics: The mathematics used in BE formulations may seem unfa-
miliar to engineers (but not di¢ cult to learn). However, many FE numerical procedures
are directly applicable to BE solutions (such as numerical integration, surface approxima-
tion, treatment of boundary conditions).
2. In non-linear problems, the interior must be modelled: Interior modelling is un-
avoidable in non-linear material problems. However, in many non-linear cases (such as
elastoplasticity) interior modelling can be restricted to selected areas such as the region
around a crack tip.
3. Fully populated and unsymmetric solution matrix : The solution matrix resulting
from the BE formulation is unsymmetric and fully populated with non-zero coe¢ cients,
whereas the FE solution matrices are usually much larger but sparsely populated. Thsi
means that the entire BE solution matrix must be saved in the computer core memory.
However, this is not a serious disadvantage because to obtain the same level of accuracy
as the FE solution, the BE method needs only a relatively modest number of nodes and
elements.
4. Poor for thin structures (shell) three-dimensional analyses: This is because of the
large surface/volume ratio and the close proximity of nodal points on either side of the
structure thickness. This causes inaccuracies in the numerical integrations.
Choosing BE or FE?
To decide whether BE or FE solutions are more suitable for a particular problem, three
factors must be taken into consideration:
1. The type of problem (linear, non-linear, shell-like analysis, etc.)
2. The degree of accuracy required
3. The amount ot time to be spent in preparing and interpreting data.
Both techniques should be made available to engineers, because in certain types of
applications one of them may display a distinct advantage over the other. Considering
the advantages and disadvantages of the BE method listed above, the following points
may help in deciding which technique to use:
a) The BE method is very suitabable and more accurate for linear problems, particu-
larly for three-dimensional problems with rapidly changing variables such as fracture or
contact problems
b) Because of the much reduced time needed to model a particular problem, the BE
method is very suitable for preliminary design analyses where geometry and loads can
6 1 Introduction
be subsequently modi…ed with minimal e¤ort. This gives designers more freedom in
experimenting with new shapes and geometries.
c) The FE method is more established and more commercially developed, particularly
for complex non-linear problems where thorough tests to establish its reliability have been
performed. The temptatipn for engineers is to use a well-established computer program
rather than venture into new methods.
d) Mesh generators and plotting routines developed for FE applications are directly
applicable to BE problems. It should not be a di¢ cult task to write ’translator’programs
to interface with commercial FE packages. Furthermore, many load incrementation and
iterative routines developed for FE applications in non-linear problems are also directly
applicable in BE algorithms.
2 Mathematical Preliminaries
For an easy understanding of the boundary integral equation derivation, some mathemat-
ical techniques are very important. They will be used time and time again to transform
the di¤erential equations governing continuum mechanic problems into equivalent bound-
ary integral equations. Moreover, some notations, de…nitions and useful formulas should
be familiar to the reader in order to feel con…dent about their subsequent use. Proofs for
these formulas and results can be found in textbooks on calculus and analysis.
2.1 Some notations and de…nitions

Here, the notations used in the following text are introduced and some de…nitions are
given.
2.1.1 Indicial and symbolic notation

The components of a tensor of any order may be represented clearly by the use of the
indicial notation, i.e., letter indices as subscripts are appended to the generic letter repre-
senting the tensor quantity of interest. Dependent on the number these indices, a tensor
of …rst order, mostly called vector, bears one free index, a second-order tensor, sometimes
called dyadic, has two free indices, a.s.o. Hence, a symbol such as which has no indices
attached, represents a scalar or tensor of zero order.
When an index appears twice in a term, that index is understood to take on all the
values of its range, and the resulting terms summed. In this so-called Einstein summation
convention, repeated indices are often referred to as dummy indices, since their replace-
ment by any other letter not appearing as a free index does not change the meaning of the
term in which they occur. In ordinary physical space, the range of the indices is 1; 2; 3.
The representation of a vector and a tensor in the symbolic notation is designated by
bold-faced letters, e.g., a and D, respectively, where unit vectors êi are further distin-
guished by a caret placed over the bold-faced letter. There, the summation convention is
often also employed in connection with indexed base vectors ^ ei , i.e., for a vector
v = vi ^
ei = v1^
e1 + v2^
e2 + v3^
e3 (2.1)
and similarly for an arbitrary dyadic
D = Dij ^
ei^
ej (2.2)
A special operational vector is r, the so-called Nabla vector, containing di¤erentiations
with respect to all coordinate axis, e.g., for Cartesian coordinates x1 ; x2 ; x3
@ @ @ @
r= e1 +
^ e2 +
^ e3 =
^ ei , @i
^ (2.3)
@x1 @x2 @x3 @xi
7
8 2 Mathematical Preliminaries
Exercise 1: Nabla vector
Derive the Nabla vector for the Polar coordinates r and ' where the relations between
the unit vectors ^
er and ^
e' of the Polar coordinate system and the unit vecors ^
e1 and ^
e2
of theCartesian coordinate system are
e1 = ^
^ er cos ' ^ e' sin '
e2 = ^
^ er sin ' + ^
e' cos '
2.1.2 Contraction and di¤erent products of tensors

The outer product of two tensors is the tensor whose components are formed by multiply-
ing each component of one of the tensors by every component of the other, i.e., a dyad is
formed from two vectors by this very product
Indicial Notation Symbolic Notation

(2.4)
ai bj = Tij ab = T
where the symbols ai and bj can be in any order. Also, one obtains, e.g.,
ij nk = sijk ; ij "kl = Eijkl
Contraction of a tensor with respect to two free indices is the operation of assigning to
both indices the same letter subscript, i.e., changing them to dummy indices, and, hence,
performing the summation convention, e.g.,
Tii = T11 + T22 + T33 = a1 b1 + a2 b2 + a3 b3 = ai bi
ij nj = pi
An inner product or scalar product of two tensors of arbitrary order is the result of a
contraction, involving one index from each tensor, performed on the outer product of the
two tensors, e.g.,
Indicial Notation Symbolic Notation
ai b i = a b=
(2.5)
Dij nj = pi D n=p
Dij ni = fj n D=f
Example: The directional derivative of a scalar function f (x1 ; x2 ; x3 ) in the direction of

a unit vector, e.g., the unit normal vector n is de…ned by the scalar product of this unit
vector n with gradf = rf :
@f @f
= rf n = ni (2.6)
@n @xi
2.1 Some notations and de…nitions 9
Exercise 2: Laplace operator

Determine the Laplace operator, i.e., the scalar product of the nabla vector with itself in
Polar coordinates.
In order to express the cross products in the indicial notation, the third order tensor
ijk , known as the permutation symbol, must be introduced:
if the values of i; j; k are an even permutation of 1; 2; 3

1
(i.e. if they appear in sequence as in the arrangement 12312...)
if the values of i; j; k are an odd permutation of 1; 2; 3
ijk = 1 (2.7)
(i.e. if they appear in sequence as in the arrangement 32132...)
if the values of i; j; k are not a permutation of 1; 2; 3
0
(i.e. if two or all three of the indices have the same value)
From this de…nition, the indicial notation of cross products is written by, e.g.
a b = c; ijk aj bk = ci (2.8)
r E = D; ijk @j Ekl = Dil (2.9)
Example: The coss product of of two vectors a and b may also be expanded as
^
e1 ^
e2 ^
e3
a b = a1 a2 a3
b1 b2 b3
a2 a3 a1 a3 a1 a2
= ^
e1 ^
e2 +^
e3 (2.10)
b2 b3 b1 b3 b1 b2
2.1.3 The Euclidian distance r and its derivatives

Considering two points x and with its Cartesian coordinates (x1 ; x2 ; x3 ) and ( 1 ; 2 ; 3 ),
respectively, their Euclidian distance r is de…ned as
p
r =j x j = (x1 2 2 2 (2.11)
1 ) + (x2 2 ) + (x3 3)
Using the summation convention, r may be written in the indicial notation form as
q p 1=2
r= ij (xi i )(x j j ) = (xi i )(xi i ) = [(xi i )(xi i )] (2.12)
where ij is the so-called Kronecker delta meaning

n
0 for i6=j
ij = 1 for i=j (2.13)
The …rst derivative of r with respect to xj follows from (2.12) to be
@r 1 1=2 (xj j)
= r;j = [(xi i )(xi i )] 2(xj j) = (2.14)
@xj 2 r
where, here, the comma is used to denote partial derivatives with respect to the coordi-
nates of the point x. The …rst derivative of r in special directions, e.g., in the direction
of a normal vector ni or a tangential vector ti is simply
r;n = r;i ni and r;t = r;i ti (2.15)
In R1 , this …rst derivatives of r may also be expressed as
(x1 1)
r;1 = = sign(x1 1) = 2H(x1 1) 1 (2.16)
j x1 1 j
where sign(x1 1 ) gives the sign of (x1 1 ) and H(x1 1 ) means the Heaviside function.
Hence, as shown above , in R1 , the second derivative of r is
r;11 = 2 (x1 1) (2.17)
This is di¤erent in R2 and R3 . There, since @(xj j )=@xk = jk , the second derivative
is obtained to be (j; k = 1; 2 and j; k = 1; 2; 3 in R2 and R3 , respectively)
@2r jk (xj j) @r jk r;j r;k
r;jk = = = (2.18)
@xj @xk r r2 @xk r
Examples: If the summation rule is applied, one obtains
n
2 in R2
ii = 3 in R3
and, in R1 ; R2 , and in R3
r;j r;j = 1
whereas
r;11 = 2 (x1 in R1
1)
1
r;jj = r;11 +r;22 = in R2
r
2
r;jj = r;11 +r;22 +r;33 = in R3
r
Remark: In general, i.e., when the normal vector ni is de…ned at a curved boundary, it
holds
r;in = r;ij nj 6= r;ni = (r;j nj );i = r;ji nj + r;j nj;i
Since with the curvature radius of the boundary
1
nj;t = tj and nj;n = 0
one …nds
nj;i = nj;t ti + nj;n ni
1
= tj ti
it holds
1
r;ni = (r;j nj );i = r;ji nj + r;j nj;i = r;ji nj + tj ti
2.2 The Gauss theorems 11
2.2 The Gauss theorems

The Gauss-Green theorem is a very fundamental identity that relates a domain integral
of a derivative of a tensorial function to an integral of that function around the boundary
of that domain. The only requirement is that the integrand of the domain integral is a
derivative, i.e., may be expressed as a product (inner, outer or cross) of the Nabla-vector
r with a tensorial function.
Let be a …nite domain in Rn bounded by a piecewise smooth orientable surface
with the outward normal vector n at a point on . Dependent on the order of the tensorial
function, following special theorems are known.
2.2.1 The gradient theorem

When F is a scalar function, the following identity holds
Z I
rF d = nF d (2.19)
where, obviously, the integrand of the surface integral is obtained by simply exchanging
the Nabla-vector r by the normal vector n.
2.2.2 The divergence theorem

When a is a vector function and the Nabla-vector r is multiplied via the inner product
with this function, the divergence theorem holds
R H R H H
r ad = n ad or @i ai d = ni ai d = an d (2.20)
which is also known as Gauss theorem. Similarly, when S is a dyadic function, one obtains
R H R H H
r Sd = n Sd or @i Sij d = ni Sij d = tj d (2.21)
2.2.3 Generalized Gauss theorems

Not often used, but nevertheless valid are generalizations where the cross product or the
outer product is applied for multiplication:
R H R H
r ad = n ad or ijk @j ak d = ijk nj ak d (2.22)
R H R H
ru d = nu d or @i uj d = ni uj d (2.23)
Example 1.1
The realization of formula (2.19) in R1 gives a well known result, when one notices that
in this special case the domain is simply an intervall [a; b], the boundary means here
only the two points x1 = a and x1 = b and, hence, boundary integration is summation at
these two points. Since, moreover, the outward normal vector at these points is n1 (a) = 1
and n1 (b) = 1,respectively, one obtains
Z I
rF d = nF d
Zb
d
F (x1 )dx1 = n1 (a)F (a) + n1 (b)F (b) = [F (x1 )]xx11 =b
=a (2.24)
dx1
a
exactly as one knows it from basic calculus.
2.3 Integration by parts - Green’s identities

By applying the above Gauss theorems, one can easily perform integrations by part. When
the Laplace operator = r2 = r r acts on the scalar function v and this result is
multiplied with the scalar function u; one obtains in symbolic notation
Z Z
(r rv) u d = [r ((rv)u) (rv) (ru)] d
I Z
= n (rv) ud (rv) (ru) d (2.25)
or in indicial notation
Z Z
@2v @ @v @v @u
ud = u d
@xi @xi @xi @xi @xi @xi
I Z
@v @v @u
= ni ud d (2.26)
@xi @xi @xi
Changing the sequence of these terms, this is the so-called Green’s …rst identity
Z I
[(r rv) u + (rv) (ru)] d = n (rv) ud (2.27)
When the remaining domain integral in (2.25) is integrated by parts a second time
Z Z
(rv) (ru) d = [r (v (ru)) v (r ru)] d
I Z
= n (v (ru)) d v (r ru) d (2.28)
the …nal result is an additional boundary integral and a domain integral where all di¤en-
tiations are shifted from v to u:
Z I Z
(r rv) u d = [n (rv) u v n (ru)] d + v (r ru) d
Z I Z
@2v @v @u @2u
ud = u v d + v d (2.29)
@xi @xi @n @n @xi @xi
2.3 Integration by parts - Green’s identities 13
This is exactly the transformation which one needs for deriving an integral representation
of the Laplace equation. In the form
Z I
2 2
(r v) u v (r u) d = [n (rv) u v n (ru)] d (2.30)
the relation (2.29) is known as Green’s second identity.

The above demonstrated steps of integration by parts are naturally possible also for
vectorial and tensorial states, e.g. with ru , @i uj = uj;i and , ij
Z Z
(ru) d , uj;i ij d
Z
= [(uj ij );i uj ij;i ] d
I Z
= uj ij ni d uj ij;i d
I Z
= (n ) ud (r ) u d (2.31)
Example 1.2
Again, the realization in R1 , here of formula (2.26), gives a well known result, when
one notices that in this special case the domain is simply an intervall [a; b], the boundary
means here only the two points x1 = a and x1 = b and, hence, boundary integration
is summation at these two points. Since, moreover, the outward normal vector at these
df
points is n1 (a) = 1 and n1 (b) = 1, respectively, one obtains from (2.26) ( dx = f 0)
Zb Zb
00 0 0
f (x1 )g(x1 )dx1 = [n1 (a)f (a)g(a) + n1 (b)f (b)g(b)] f 0 (x1 )g 0 (x1 )dx1
a a
Zb
x =b
= [f 0 (x1 )g(x1 )]x11 =a f 0 (x1 )g 0 (x1 )dx1 (2.32)
a
exactly what one knows from basic calculus.
Exercise 3: Integration by parts

Evaluate by integrations by parts the following integrals
a) in R1 on the intervall a < x < b :
Zb
xn ln(x)dx
a
n p o
b) in R2 on the circular domain = (x1 ; x2 ) j r = x21 + x22 R :
ZR
rn ln(r)d
0
2.4 Fundamental solutions of di¤erential equations

Let Rn denote the the Euclidian n-space and the ordered n-tuple k = (k1 ; k2 ; :::; kn ) with
non-negative integers k1 ; :::; kn be a multi-index of dimension n. If we let j k j= k1 +::::+kn ,
then the k-th (partial) di¤erential operator is de…ned by
@ jkj @ k1 +::::+kn
Dk = = ; x = (x1 ; :::; xn ) 2 Rn (2.33)
@xk11 @xknn k1
@x1 @xn k n
such that if any component of k is zero, the partial derivative with respect to that variable
is omitted. Moreover,
@ jkj u(x1 ; :::; xn )

Dk u(x) = ; D0 u(x) =u(x) (2.34)
@xk11 @xknn
@
D = (D1 ; D2 ; :::; Dn ) with Di = @i = ; i = 1; 2; :::; n: (2.35)
@xi
An arbitrary linear di¤erential operator L of order p in n independent variables x1 ; :::; xn
is denoted by X
L L(D) = ak (x)Dk (2.36)
jkj p
where the coe¢ cients ak (x) = a(k1 ;k2 ;:::;kn ) (x1 ; :::; xn ) are arbitrary functions.
2.4.1 Adjoint and self-adjoint operators

For the arbitrary linear di¤erential operator L (2.36), the so-called adjoint operator L is
formally de…ned by X
Lv= ( 1)k Dk (ak v): (2.37)
jkj p
If ak (x) = ak are constant, then L (D) = L( D). An operator is said to be self-adjoint

if L = L .
Considering the boundary value problem
L(D)u(x) = f (x) in Rn ; (2.38)
B(u) = 0 on @ = (2.39)
2.4 Fundamental solutions of di¤erential equations 15
where the Eq (2.39) represents linear boundary conditions, and introducing the inner
product hf1 ; f2 i of two functions f1 and f2 in the euclidean space Rn as
Z
hf1 ; f2 i = f1 (x)f1 (x)d x (2.40)
are more speci…c de…nition of L can be formulated:

Through integrations by part, i.e., by shifting all di¤erentiations acting in on the
functions u to the functions w, one obtains
Z
hLu; wi = hu; L wi + fE(w)N (u) N (u)E(w)g d (2.41)
Here, again, the operator L is adjoint to L, and, if L = L, then L is said to be self-

adjoint.
The formula (2.41) represents the variational formulation for the equation (2.38),
where E(u) generates essential boundary conditions ( which must be enforced at some
points so as to have a unique solution), while N (u) generates non-essential boundary con-
ditions, also called the natural boundary conditions, depending on the degree of derivatives
that appear in the operators E and N that respectively de…ne them.
Then, the fundamental solution u (x; ) is the solution of (2.38) for the special case
when f (x) is replace by (x; ), the so-called Dirac -function (whose exact mathematical
de…nition is only possible considering the theory of distributions, see, e.g., [4])
L(D)u (x; ) = (x; ) (2.42)
The function u (x; ) is unique only up to a function w (x; ) which is the solution of
the homogeneous equation L(D)w = 0, i.e., the function u + w is also a fundamental
solution for the operator L(D):
L(D) (u (x; ) + w (x; )) = L(D)u (x; ) + L(D)w (x; ) = (x; ) (2.43)
2.4.2 The Dirac -function

This symbolic function or distribution has the following basic properties:
(x; ) = 0 f or x 6= (2.44)
where, in general, it is only a function of the distance between the two points x and ,
i.e.,
(x; ) = (x ) = (x1 1 ) (x2 2) (xn n) (2.45)
Z
(x) (x; ) d x = ( ) for 2 (2.46)
for all su¢ ciently ’smooth’, i.e., continous functions (x). (2.46) is the selection property
which means that the -function, when involved in an integration process with another
function, selects the value of the other function at the point where the -function (x )
has a zero argument, i.e., at x = . For the special case (x) 1, (2.46) gives
Z Z
(x; ) d x = (x ) d x = 1 for 2 (2.47)
Example 1.3
For a rectangular plane domain , i.e., x1 2 [a1 ; b1 ]; x2 2 [a2 ; b2 ], the postulate (2.46)
gives with (2.45)
Z Z
(x )d x = (x1 1 ) (x2 2) d x
Zb1 Zb2
= (x1 1 )dx1 (x2 2 )dx2 = 1 for 2
a1 a2
Zbi
or (xi i )dxi = 1 for i 2 [ai ; bi ]; i = 1; 2 (2.48)
ai
The expression for the -function becomes more complicated when one introduces curvilin-
ear co-ordinates, i.e., the considered di¤erential equations are formulated using curvilinear
co-ordinates.
Example 1.4
The transformation from rectangular Cartesian x1 ; x2 to plane polar co-ordinates r; '.
The transformation is given by
x1 = u(r; ') = r cos '; x2 = v(r; ') = r sin '
and the Jacobian J of the transformation is J = r which yields
d x , r drd'
Then, the integral statement (2.47), e.g., for a circular domain with radius R around
= 0, becomes due to the fact that the -function is only a function of the distance
r =j x j and not of the angle '
Z Z ZR Z2
(x) d x = (x1 ) (x2 ) d x = (r; ') r drd'
0 0
ZR Z2 ZR
= (r; ') r dr d' = (r; ') r dr2 = 1
0 0 0
and postulates
(r)
(r; ') = (2.49)
2 r
Hence, if a partial di¤erential equation is formulated in polar co-ordinates, the adequate

fundamental solution has to satisfy
(r)
L(D)u (r; ') =
2 r
3
and, correspondingly, using in R spherical co-ordinates r; ' and #
(r)
L(D)u (r; '; #) = (2.50)
4 r2
The -function in R1 and the Heaviside function

In R1 , it can be shown that the -function can be handled algebraically as if it were an
ordinary function, but one must always interpret any equation involving (x) as follows:
if the equation is multiplied throughout by an arbitrary continuous function f (x), and
integrated over an intervall [a; b] by using the -function’s selection property, i.e.
Zb
(x )f (x)dx = f ( ) for 2 [a; b] (2.51)
a
= 0 for 2
= [a; b]
then the resulting equation is correct and involves only ordinary functions.
For example
(x ) (x )=0 (2.52)
because for any arbitrary continuous function f (x) one obtains with g(x) = (x )f (x)
Zb
(x )g(x)dx = g( ) = 0 for 2 [a; b]
a
Similarly, one can state

w(x) (x ) = 0 if w(x = ) = 0 (2.53)
Moreover, the familiar techniques of integration, such as integration by parts and substi-
tution, can be shown to apply to integrals involving -functions. As an example, consider
the integral
Z1
I= (g(x)) f (x)dx
1
where f (x) is an arbitrary continuous function and g(x) is a monotonic function of x

which vanishes when x = . Write y = g(x) and it follows that dy = g 0 (x)dx. The
integral then becomes
Z1 Z1
f( )
I= (g(x)) f (x)dx = (y) (y)dy = (0) = 0
jg( )j
1 1
with (y) = f (x)= j g 0 (x) j, where the modulus sign is to ensure that the integration is
always from 1 to 1. Consequently, it follows if g(x = ) = 0 that
(x )
(g(x)) = 0
(2.54)
jg( )j
As a special case of (2.54) one obtains

Z1
f ( ab )
(ax b) f (x)dx = (2.55)
jaj
1
And, again by integration by parts, one obtains
Zb Zb
@ (x )
f (x)dx = (x ) f 0 (x)dx = f 0 ( ) for 2 [a; b] (2.56)
@x
a a
Besides, in R1 , the Dirac function can be considered to be the derivative of the Heaviside
unit function H de…ned as (see, e.g., [3], p. 147)
H(x ) = 1 for x > (2.57)

= 0 for x <
To see this, we integrate by parts (a < < b)
Zb Zb
@H(x )
f (x)dx = [H(x )f (x)]x=b
x=a H(x )f 0 (x)dx
@x
a a
Zb
= f (b) f 0 (x)dx
= f (b) (f (b) f ( )) = f ( )
which yields by comparison with (2.51)
@H(x )
= (x ) (2.58)
@x
The -function in Rn with n 2

In Cartesian co-ordinates, the -function in a n-dimensional geometric space is the product
of n one-dimensional -functions, e.g., in R3
(x ) = (x1 1) (x2 2) (x3 3 ): (2.59)

The expressions for the -functions become much more complicated when one introduces
curvilinear coordinates. To …nd corresponding forms, we will, for simplicity, con…ne our-
selves to two-dimensional space. Suppose that we transform from Cartesian co-ordinates
x1 ; x2 to curvilinear co-ordinates 1 ; 2 by means of the relations
x1 = u( 1 ; 2 ); x2 = v( 1 ; 2) (2.60)
where u and v are single-valued, continously di¤erentable functions of their arguments.

Supposing that under this transformation 1 = 1 and 2 = 2 correspond to x1 = 1
and x2 = 2 , respectively.
If one changes the co-ordinates according to (2.60), the equation
Z Z
(x1 ; x2 ) (x1 1 ) (x2 2 )dx1 dx2 = ( 1 ; 2 )
becomes
Z Z
(u; v) [u( 1 ; 2) 1 )] [v( 1 ; 2) 2 )] jJj d 1 d 2 = ( 1; 2)
where J = @(u; v)=@( 1 ; 2 ) is the Jacobian of (2.60).

Correspondingly, we may write
[u( 1 ; 2) 1 )] [v( 1 ; 2) 2 )] jJj = ( 1 1) ( 2 2)
or, provided J 6= 0
( 1 1) ( 2 2)
(x1 1) (x2 2) = (2.61)
jJj
Example: Transformation from rectangular Cartesian co-ordinates x; y to plane polar

co-ordinates r; :
x = r cos ; y = r sin
Since the Jacobian of this transformation is
@x @x
cos r sin
J= @r
@y
@
@y = = r cos2 + sin2 =r
@r @
sin r cos
one obtains the relation

(r r0 ) ( 0)
(x x0 ) (y y0 ) =
r
if x0 = r0 cos 0 and y0 = r0 sin 0 .
For the event that J = 0 at some so-called singular point, the considered transfor-
mation is no longer one-to-one and, moreover, some co-ordinate , then called ignorable
co-ordinate, is either many-valued or has no determinate value at such a singular point of
the transformation. When, for example, the co-ordinate 2 is ignorable, the Jacobian has
to be integrated with respect to this co-ordinate (see, [3], p.219)
Z
J1 = jJj d 2
and, consequently, in this case when J = 0 for x1 = 1, we have the relation, e.g., in R2
( 1 1)
(x1 1) (x2 2) =
jJ1 j
Example (see also Example 1.4): In the case of the transformation x = r cos ; y =
r sin , the Jacobian J = r vanishes at x = 0; y = 0 or r = 0 which means that may
take on any value at this point, i.e., is ignorable. It follows that
Z2
J1 = rd = 2 r
0
and, hence,
(r)
(x) (y) =
2 r
2.4.3 Green’s functions of boundary value problems

As explained above, a boundary value problem is described by a di¤erential equation
L(D)u(x) = f (x) in Rn ; (2.62)
and associated prescribed boundary conditions, e.g., on = 1 [ 2
E(u(x)) = u(x) for x 2 1 (2.63)

N (u(x)) = q(x) for x 2 2 (2.64)
Consequently, besides the unknown solution u(x) in the interior , there are also unknown
boundary reactions E(u(x)) = u(x) on 2 and N (u(x)) = q(x) on 1 which are, in general,
not zero along the respective boundary.
The so-called Green’s function G (x; ) of such a boundary value problem is de…ned
as being a special fundamental solution, i.e.,
L(D)G (x; ) = (x; ) (2.65)
which satis…es homogeneous conditions for those boundary states which are prescribed in
the actual problem, i.e.,
E(G (x; )) = 0 for x 2 1 (2.66)
N (G (x; )) = 0 for x 2 2 (2.67)
When such Green’s function of a boundary value problem is available - in general, it is
very di¢ cult to construct it -the problem’s solution can be formulated in an integral form
which contains no unknown terms (see, Chapter ?).
2.4.4 Ordinary di¤erential equations with constant coe¢ cients

The fundamental solution of an ordinary di¤erential equation of n-th order with constant
coe¢ cients satis…es
dn u dn 1 u du
L(D)u (x; ) = + a1 n 1 + + an 1 + an u = (x; ) (2.68)
dxn dx dx
and is given by
u (x; ) = H(x )w(x; ) (2.69)
where w(x; ) 2 C n (R1 ) satis…es the homogeneous equation L(D)w(x; ) = 0 with the
conditions
dw dn 2 w dn 1 w
w(x = ) = = = = 0; =1 (2.70)
dx x= dxn 2 x= dxn 1 x=
Since, in view of (2.58) and (2.53),
du (x; ) dn 1 u (x; )
= H(x )w0 (x; ); ::::; = H(x )w(n 1)
(x; )
dx dxn 1
and
dn u (x; )
= (x ) + H(x )w(n) (x; );
dxn
one …nds with L(D)w(x; ) = 0 that
L(D)u (x; ) = (x )
+H(x ) w(n) (x; ) + a1 w(n 1) (x; ) + + an w(x; )
= (x ) + H(x )L(D)w(x; )
= (x )
Example 1.5
From the above rule (2.69) with the conditions (2.70) follows that the fundamental
d d2 2 d2
solutions for the operators L1 = dx + a, L2 = dx 2 + a , and L3 = dx2 a2 are given by
(see, [4], p.40)
u1 (x; ) = H(x )e a(x ) (2.71)
where, here for n = 1, w1 (x; ) = e a(x )
and satis…es the homogeneous di¤erential
d0 w
equation dw
dx
+ aw = 0 and the condition dx0
= w(x = ) = 1, while for n = 2 the
x=
fundamental solutions are
sin a(x )
u2 (x; ) = H(x ) (2.72)
a
sinh a(x )
u3 (x; ) = H(x ) (2.73)
a
where the function w2 (x; ) = a1 sin a(x ) and w3 (x; ) = a1 sinh a(x ), respectively,
satis…es again the related homogeneous di¤erential equation and the conditions w(x =
) = 0 and dw
dx x=
= 1.
When the constant a in the operator L2 and L3 tends to zero, one obtains the operator
Lbar of the bar equation for constant sti¤ness EA (which also represents the potential
equation in R1 )
d2 u(x) p(x)
Lbar u(x) = 2
= = f (x) (2.74)
dx EA
whose fundamental solution, i.e., the solution for f (x) = (x ) is found by considering
the limit of (2.72) for a ! 0:
sin a(x )
u (x; ) = H(x ) lim = (x )H(x ) (2.75)
a!0 a
But, as expressed in (2.43), this function u (x; ) is unique only up to a function w (x; )
which is the solution of the homogeneous equation Lbar w = 0, i.e., the function u + w
with w = 21 (x ) is also a fundamental solution for the operator LBar
1 1
ubar (x; ) = (x )H(x ) (x ) = (x ) [2H(x ) 1]
2 2
1 jx j r
= (x )sign(x )= = (2.76)
2 2 2
where sign(x ) = 1 for x > and sign(x )= 1 for x < . This is obviously correct
since (x ) (x ) = 0 and
@ 1 @r 1 1
ubar (x; ) = = H(x ) + (x ) [2 (x )]
@x 2 @x 2 2
1 1
= H(x ) = sign(x ) (2.77)
2 2
@2
u (x; ) = (x ) (2.78)
@x2
Correspondingly, a fundamental solution of the basic di¤erential equation for an elastic

beam with constant ‡exural rigidity EI
d4 u(x) q(x)
Lbeam u(x) = 4
= = f (x) (2.79)
dx EI
can easily found by integrating the solution (2.76)of (2.74) twice:
r3
ubeam (x; ) = (2.80)
12
2.4.5 Scalar partial di¤erential equations with constant coe¢ -

cients
The fundamental solution for the three-dimensional Laplace equation
u (x; ) = (x; ) (2.81)
can be obtained directly as follows ([4], page 79):
Since the operator = r r is invariant under a rotation of coordinate axes, we shall
seek a solution that depends only on the distance r =j x j. For r > 0, u (x; ) = u (r)
will satisfy the homogeneous equation u = 0, i.e., in spherical coordinates
1 @ @u
r2 = 0; (2.82)
r @r @r
which has a solution u = Ar + B. If one requires the solution u to vanish at in…nity, then
B = 0. In order to determine A, one has to take into account the magnitude of the source
at x = . Integrating (2.81) over a small sphere " of radius " and center at x = , one
obtains Z Z
r ru (x; )d x = (x; )d x = 1 (2.83)
" "
which, by using the divergence theorem (2.20), gives

Z Z Z
@u (x; ) @u (x; )
n ru (x; )d x = d x= d x =1 (2.84)
@n @r
" " "
where " =@ " is the surface of the sphere " and d x = r2 sin d d' with 0
and ' . Now, substituting u = A=r, i.e., @u =@r = A=r2 in (2.84) yields
Z Z Z Z
A 2
r sin d d' = A d' sin d = A2 [ cos ]0 = 4 A = 1
r2
0 0
i.e., A = 1=(4 ). Hence, the fundamental solution for the three-dimensional Laplace
equation is
1
u (x; ) = u (r) = (2.85)
4 r
Correspondingly, the homogeneous two-dimensional Laplace equation in polar coordinates
1 @ @u
r = 0; (2.86)
r @r @r
has for r > 0 the solution u = C ln(r=a) + D. Arbitrarily setting D = 0 and applying
the same steps as above in (2.83) and (2.84), determines C = 1=(2 ), and hence the
fundamental solution of the two-dimensional Laplace equation is
1 r 1 a
u (x; ) = u (r) = ln( ) = ln( ) (2.87)
2 a 2 r
where a > 0 is an arbitrary real constant making the ratio r=a dimensionless.
Similarly (see, e.g., [5] or [4]), one can determine the fundamental solutions for the
Helmholtz equations in Rn ; n = 1; 2; 3
( + k 2 )u (x; ) = (x; ) (2.88)
where k 6= 0, real, is the so-called wave number. One gets

1 ikr
for n = 3: u (x; ) = u (r) = e (2.89)
4 r
i (2)
for n = 2: u (x; ) = u (r) = H0 (kr) (2.90)
4
1
for n = 1: u (x; ) = u (r) = sin(kr) (2.91)
2k
(2)
where H0 (kr) is a Hankel function. With h = ik or k = ih, the operator (2.88) changes
to
( h2 )u (x; ) = (x; ) (2.92)
and the fundamental solutions become
1
for n = 3: u (x; ) = u (r) = eihr (2.93)
4 r
1
for n = 2: u (x; ) = u (r) = K0 (hr) (2.94)
2
1
for n = 1: u (x; ) = u (r) = sinh(hr) (2.95)
2h
where K0 (hr) is a modi…ed Besselfunction. Note that here lim u (r) = 1, i.e., these
r!1
fundamental solutions represent wave forms that diverge to in…nity.
Remark 1: The R1 fundamental solutions given in (2.91) and (2.72) for the operator
(2.88) as well as in (2.95) and (2.73) for the operator (2.92) are di¤erent, but, as one can
easily check, both versions are correct, i.e., give the typical …ltering e¤ect of the Dirac -
1
function. For example, gives the fundamental solution (2.91) u (x; ) = u (r) = 2k sin(kr)
2
of the equation (2.88)(note that r;x = @r=@x = 2H(x ) 1 and (r;x ) = 1)
Zb Zb
2
( + k )u (x; )dx = cos(kr) (x; )dx = 1 for 2 [a; b]
a a
while the fundamental solution (2.73) u2 (x; ) = H(x ) sin k(x

k
)
for the same equation
results in
Zb Zb
2
( + k )u2 (x; )dx = cos(k(x )) (x; )dx = 1 for 2 [a; b]
a a
The essential di¤erence of both versions is that the forms (2.72) and (2.72) contain the
Heaviside function H(x ) and, therefore, are cut o¤ (zero) for x < , while the other
versions (2.91) and (2.95) are due to their dependence on r = jx j symmetric to x = .
1
Remark 2: In R , it is also possible to use instead of the divergent form (2.95)
1 hr
u (x; ) = u (r) = e (2.96)
2h
This form is convergent, i.e., lim u (r) = 0, and has, as easily can be checked, the essential
r!1
behaviour of a fundamental solution:
Zb Zb
( h2 )u (x; )dx = e hr (x; )dx = 1 for 2 [a; b] (2.97)
a a
Remark 3: When taking into account that u (r) = 1=4 r is the fundamental solution
of the Laplace operator in R3 , i.e.,
1 1
( ) = r2 ( ) = 4 (x; )
r r
1 ikr
the check that u = 4 r e is really the fundamental solution of the Helmholtz equation
(2.88) in R is straight-forward by applying the Leibniz formula r2 (af ) = f r2 (a) +
3
2r(a) r(f ) + ar2 (f ), i.e. here,

1 1 1 1
( + k 2 )u = e ikr
r2 ( ) + 2r( ) r(e ikr
) + r2 (e ikr
) + k2u
4 r r r
Since di¤erentiations give ( @x@ i =;i )
1 @ 1 1 (xi i)
r( ) = ( )= 2
r;i = 3
r @xi r r r
@ (xi i)
r(e ikr ) = (e ikr ) = ikr;i e ikr = ik e ikr
@xi r
@ (xi i) 2ik
r2 (e ikr ) = ( ik e ikr ) = ( + k 2 )e ikr
@xi r r
one obtains …nally
2 1 e ikr
4 (x; ) + 2 (xir3 i ) ik (xi r i)
e ikr k2 ikr
( + k )u = 1 2ik e
4 ( + k 2 )e ikr )
r r 4 r
ikr
= e (x; ) = (x; )
Certainly, one should know that
@2 ikr @
(e ) = ( ikr;i e ikr ) = ik r;ij e ikr
+ r;i ( ikr;j )e ikr
@xi @xj @xj
= ik (r;ij ikr;i r;j )) e ikr
with
@ (xi i) ij (xi i )r;j 1
r;ij = = = ( ij r;i r;j )
@xj r r r2 r
2.5 Singular integrals

Singular integrals are those whose integrands reach an in…nite value at some points on
the integration domain . They are, in general, de…ned by eliminating a small space
(initially arbitrary) including the singular point, and obtaining the limit when this small
space tends to disappear
Z Z
f (x; )d x = lim f (x; )d x with 2 " (2.98)
"!0
"
" can be a ball of radius " in 3D, a circle of radius " in 2D, and in 1D, i.e. on a line a
segment of dimension " at each side of the point where the singularity is located.
2.5.1 Weak singularities - improper integrals

If the limit of (2.98) exists independently of how " tends to zero, it is said that this integral
exists as improper and the singularity is said to be weak.
In R1 , a representative example is the integral
Zb
ln j x j dx; a < <b (2.99)
a
which can evaluated as

0 1
Zb Z "1 Zb
ln jx j dx = lim @ ln( x)dx + ln(x )dxA
"1 !0;"2 !0
a a +"2
"1
= lim [ ( x)[ln( x) 1]]a + lim [(x )[ln(x ) 1]]b +"2
"1 !0 "2 !0
= lim ( "1 [ln("1 ) 1]) + ( a)[ln( a) 1]
"1 !0
+(b )[ln(b ) 1] lim ("2 [ln("2 ) 1])
"2 !0
= (a b) + ( a) ln( a) + (b ) ln(b ) (2.100)
obviously exists, since from the rule of Bernoulli-de l’Hospital
ln(") " 1
lim (" ln(")) = lim = lim = lim( ") = 0 (2.101)
"!0 "!0 " 1 "!0 " 2 "!0
follows that all "-terms disappear independently from each other.

Physically, this fact implies that the area under the function at any side if the singular
point has a …nite value.
Remark: The following integrals also exist as improper integrals
Zb
dx
I= for 0 < k < 1
jx jk
a
2.5 Singular integrals 27
2.5.2 The Cauchy Principal Value of strongly singular integrals

When one evaluates the integral
Zb
dx
; a< <b (2.102)
x
a
as an improper integral
0 1
Zb Z "1 Zb
dx dx dx A b "1
= lim @ + = ln + lim ln (2.103)
x "1 !0;"2 !0 x x a "1 !0;"2 !0 "2
a a +"2
the limit of the last expression obviously depends on the way in which "1 and "2 tend
to zero. Hence, the improper integral does not exist. This integral is called a strongly
singular integral. However, this integral can be assigned a meaning if we assume that
there is some relationship between "1 and "2 , e.g., if the deleted intervall is symmetric
with respect to the point , i.e., "1 = "2 = ". Then, one obtains
0 1
Zb Z " Zb
dx dx dx A b
= lim @ + = ln (2.104)
x "!0 x x a
a a +"
the so-called Cauchy principal value (CPV) of a singular integral.

Remark: Evaluating weakly singular integrals as Cauchy principal value gives obvi-
ously the same result.
Now, consider the more general integral
Zb
'(x)
dx; a < <b (2.105)
x
a
where '(x); x 2 [a; b] is a function satisfying the Hölder condition, i.e., for any two points
t1 and t2 on a smooth curve L and for positive constants A and with 0 < 1 holds
j '(t2 ) '(t1 ) j< A j t2 t1 j (2.106)
which means that '(t) is di¤erentiable and has a bounded derivative. The Hölder condi-
tion is sometimes referred to as an intermediate situation between continuity and deriv-
ability, establishing in fact a division in the set of continuous non-derivable functions (see,
Paris and Canas [6], p. 14)
Let us understand this integral (2.105) in the sense of the Cauchy principal value
(CPV) as 0 " 1
Zb Z Zb
'(x) '(x) '(x) A
dx = lim @ dx + dx (2.107)
x "!0 x x
a a +"
and take the identity

Zb Zb Zb
'(x) '(x) '( ) dx
dx = dx + '( ) (2.108)
x x x
a a a
Now, one can see that the …rst integral on the right-hand side of (2.108) is convergent as
an improper integral, because it follows from the Hölder condition that
'(x) '( ) A
j j< ; 0< 1 (2.109)
x jx j1
and the second integral coincides with (2.104). Thus, if '(x) satis…es the Hölder condition,
the singular integral (2.105) exists in the sense of the Cauchy principle value (CVP)and
is equal to
Zb Zb
'(x) '(x) '( ) b
dx = dx + '( ) ln (2.110)
x x a
a a
Remark: The following integral does not exist, neither as improper integral nor as Cauchy
Principal Value:
Zb
dx
I=
jx j
a
Example: Determine the CVP of the following integral (a < < b):
Zb Z " Zb
dx dx dx
= +
sin(x ) sin( x) sin(x )
a a +"
Z" Zb
dy dy
= +
sin(y) sin(y)
y= a "
h y i" h y ib
= ln tan( ) + ln tan( )
2 a 2 "
!
tan( b 2 )
= ln
tan( 2 a )
2.5.3 Cauchy Principal Value integrals in boundary integral equa-

tions
In boundary integral equations, the integrands contain often singularities of inverse powers
of r =j x j multiplied by certain bounded functions, i.e., have the general structure
Z
'(x; )
I= u(x)d x
jx jk
where the so-called characteristic '(x; ) does not include any singularity, the singular
kernel '(x; )=rk results from the fundamental solution while the so-called density u(x)
usually represents boundary values of the considered problem.
Conditions for the existence of a CPV of a singular curvilinear line integral
Let be a smooth contour, x and be coordinates of its points, and xa and xb be the
endpoints of . Consider the singular curvilinear integral where the contour points are
expressed in terms of a parameter, e.g., the arc length s, so that x = x(s) and sa , sb are
the values of the parameter s corresponding to the endpoints of :
Z Zsb Zsb
'(x; ) '(x(s); ) d x '(x(s); )
u(x)d x = u(x(s)) ds = u(x(s))J(s)ds
jx j j x(s) j ds j x(s) j
sa sa
(2.111)
where J(s) is a Jacobian associated with the change of variable from x = (x1 ; x2 ) to s
s
2 2
d x dx1 dx2
J(s) = = + (2.112)
ds ds ds
Let us take a circle of some radius " centered at the point on the contour, + = + x"
and = x" be the points of intersection of this circle with the curve, and assume
that the radius is so small that the circle has no other points of intersection with . Let
" be the part of the contour cut out by the circle and consider the integral over the
remaining arc. Then, its limit for " ! 0 is the principal value of the singular integral
(2.111).
Z Z
'(x; ) '(x; )
u(x)d x = lim u(x)d x
jx j "!0 jx j
"
Z Z
'(x; ) '(x; )
= [u(x) u( )] d x + u( ) d x(2.113)
jx j jx j
where, if u(x) satis…es the Hölder condition for points x being placed in the neighbourhood
of , the …rst integral exists. The second integral is, e.g., along a straight line contour
of length s = sb sa where with x = x(s) and = x(s) the distance r =j x j is
r =j s s j and J(s) = 1
2s " 3
Zxb Z Zsb
'(x; ) '(x(s); x(s)) '(x(s); x(s)) 5
d x = lim 4 ds + ds (2.114)
jx j "!0 js s j js s j
xa s s+"
2 sa " 3
Z Zsb
'(x(s); x(s)) '(x(s); x(s)) 5
= lim 4 ds + ds
"!0 s s s s
s s+"
2 as 3
Za Zsb
'(x(s); x(s)) '(x(s); x(s)) 5
= lim 4 ds + ds
"!0 s s s s
s " s+"
= I1 + I2
Remark: In many discretizations, the boundary line of a domain is approximated by a
polygon, i.e., by elementwise straight lines.
Evaluating I1 by integration by parts gives
Zsa
'(x(s); x(s))
I1 = lim ds
"!0 s s
s "
2 3
Zsa
d'(x(s); x(s)) 5
= lim 4 ln(js s j)'(x(s); x(s))jssa " ln(js s j) ds
"!0 ds
s "
= lim [ln(jsa s j)'(x(sa ); x(s)) ln(")'(x(s "); x(s)) I12 ]
"!0
The integral I12 and the corresponding integral from I2 exists if d'=dx takes …nite values
along the integration zone outside r = 0 while the …rst evaluated terms of I1 and the
corresponding ones from I2 lead to
ln(jsa s j)'(x(sa ); x(s)) ln(")'(x(s "); x(s))
lim
"!0 + ln(jsb s j)'(x(sb ); x(s)) ln(")'(x(s + "); x(s))
Hence, if
lim fln(") ['(x(s "); x(s)) + '(x(s + "); x(s))]g = 0
"!0
and d'=dx takes …nite values along the integration zone outside r = 0, the integral I1 + I2
would have a …nite value:
2s " 3
Z Zsb
'(x(s); x(s)) '(x(s); x(s)) 5
lim 4 ds + ds = ln(jsa s j)'(x(sa ); x(s))
"!0 js s j js s j
sa s+"
+ ln( j sb s j)'(x(sb ); x(s))
2 3
Zsa Zsb
4 5 ln( d'(x(s); x(s))
+ j s s j) ds(2.115)
ds
s " s+"
This condition is satis…ed by a function ' if in the neighbourhood of = x(s) it holds

that
j'(x(s "); x(s)) + '(x(s + "); x(s))j A" ; A > 0; 0 < 1; " > 0 (2.116)
Conditions for the existence of a CPV of a singular surface integral

For a study of singular integrals in R2 where singularities of order two (1=r2 ) can have
a Cauchy Principal Value, let us consider the integral on a domain 2 in R2 , here, in
general, a surface domain of three-dimensional areas
Z
'( ; )
I( ) = u(x)d x (2.117)
r2
2
where r =j x j and represents the angle formed by r with respect to the coordinate
axes. This integral can be divided in the following manner:
Z Z
'( ; ) '( ; )
2
u(x)d x = u(x)d x +
r r2
2 2 \(r )
Z Z
'( ; ) '( ; )
[u(x) u( )] d x + u( ) d x (2.118)
r2 r2
r< r<
where is a …xed distance su¢ ciently small to guarantee that the points x : jx j<
belong to 2 .
The …rst integral of (2.118) is de…ned and the second exists if the Hölder condition is
satis…ed:
ju(x) u( )j Ar (x; ); A > 0; 0 < 1: (2.119)
The third integral of (2.118) is speci…ed as
Z Z
'( ; ) '( ; )
d x = lim d x (2.120)
r2 "!0 r2
r< "<r<
and polar coordinates, i.e., d x = rd dr are introduced which yields

Z Z Z2 Z2
'( ; ) 1
lim rd dr = lim dr '( ; )d = lim ln( ) '( ; )d
"!0 r2 "!0 r "!0 "
"<r< " 0 0
The condition for this integral to exist is then:

Z2
'( ; )d = 0 (2.121)
0
Hence, if the integral of the characteristic '( ; ) on the surface around the pole x =
is zero and the density u(x) satis…es the Hölder condition, the integral (2.117) with a
singularity of order two in the two-dimensional space has a Cauchy Principal value.
3 Transformation of Di¤erential Equations
to Integral Equations
There exist several methods for transforming di¤erential equations decribing a boundary
value problem or an initial/boundary value problem to an equivalent representation by
integral equations. Two are essentially di¤erent when regarding their basic ideas and
result also in quite di¤erent formulations: the so-called direct boundary integral equation
method and the indirect boundary integral equation method. Both derivations will be
described by some representative examples.
3.1 Introductary 1-d problems: Transformation of

ordinary di¤erential equations
In R1 , not only the transformation of the considered di¤erential equations to an integral
equation representation, but also their solution can be found in many cases analytically
exact. Hence, it makes sense to consider some typical examples in order to clarify some es-
sentials, e.g., the importance of the fundamental solution of the analysed basic di¤erential
equations.
3.1.1 Integral equations by direct integration

Here, some simple integral equations are derived by performing straightforward integra-
tions in order to show the equivalence of the formulations and to point-out the importance
of the respective fundamental solution for these integral equation representations.
First order ordinary di¤erential equation.

The most simple di¤erential equation is certainly
d
y(x) = f (x) (3.1)
dx
where f (x) is given, and the solution is unique by the initial condition y(a) = y0 in an
arbitrary point x = a and shall be considered to be de…ned in a certain intervall [a; b].
Integrating of both sides of the equation (3.1) gives
Z
y(x) = f (x)dx +c (3.2)
x=x
32
When one satis…es the prescribed initial condition, i.e.,

Z
y(a) = y0 = f (x)dx +c (3.3)
x=a
one …nds the constant c as Z

c = y0 f (x)dx (3.4)
x=a
and the ’solution’ of the …rst order ordinary di¤erential equation (3.1) with the initial
condition y(a) = y0 to be
x
y(x) = s f (x)dx + y0 (3.5)
a
This example looks trivial. But, taking additionally into account that x is restricted to
a one-dimensional domain whose boundary consists of the two endpoints of the closed
intervall [a; b], and introducing the Heaviside function H(x x) (see (2.57))
1 x>x
H(x )= für (3.6)
0 x<x
the solution (3.5) may also be written as an integral over the whole de…nition domain
= [a; b]:
Zb
y(x) = H(x x)f (x)dx + y0 (3.7)
a
The kernel H(x x) of the above integral operator is obviously (see (2.58)) the fundamental
solution y (x; x) of the considered di¤erential equation (3.1).
Second order ordinary di¤erential equations.

Transfering the idea of the integral equation formulation (3.7) to the di¤erential equation
of second order
d2
y(x) = f (x) (3.8)
dx2
with the two initial conditions
d 0 0
y(a) = y0 and y(x) = y (a) = y0 (3.9)
dx x=a
and de…ned on the domain = [a; b], its solution should be given with its fundamental
solution y (x; x) and the initial conditions (3.9) as
Zb
0
y(x) = y (x; x)f (x)dx + y0 + y0 (x a) (3.10)
a
34 3 Transformation of Di¤erential Equations to Integral Equations
Certainly, one has to know the adequate fundamental solution of the di¤erential equation
(3.8). Since the fundamental solution of the …rst order di¤erential equation (3.1) is the
Heaviside function H(x x), one has only to know what is the result of integrating
H(x x). For this purpose, it advantageous to represent the Heaviside function as ’cut
polynomial’of zero degree:
H(x x) = [x x]0+ (3.11)
In this form, it is easy to integrate the Heaviside function and to …nd the fundamental
solution of (3.8):
y (x; x) = [x x]1+ = (x x)H(x x) (3.12)
This is known as fundamental solution of the bar equation (2.74). Another possible
fundamental solution is (2.76)
r jx xj 1
y (x; x) = = = (x x)sign(x x) (3.13)
2 2 2
1 1
= (x x) [2H(x x) 1] = (x x)H(x x) (x x)
2 2
because the additional linear term 21 (x x) is only a trivial solution of the homogeneous
di¤erential equation (3.8).
It is not easy to recognize that the expression (3.10) is really the solution of the
di¤erential equation (3.8), but this can be shown by two straightforward integrations and
a little more trickery transformation of the double integral into a single integral.
One integration of both sides of the di¤erential equation (3.8) gives
Zx
dy(x) 0
= y0 + f (x)dx (3.14)
dx
a
0 0
satisfying the initial condition y (a) = y0 , and a second produces
2 3
Zx Zs
0
y(x) = y0 + (x a)y0 + 4 f (x)dx5 ds (3.15)
a a
the further constant of integration having been taken so that y(a) = y0 .

Simpli…cation of the double integral in (3.15) follows on using the result of the more
general formula 2 3 2 3
Zx Zs Zx Zx
4 G(x; s)dx5 ds = 4 G(x; s)ds5 dx (3.16)
a a a x
for which it is su¢ cient that G(x; s) be a continous function of both variables. To establish
(3.16) note that the repeated integral on the left hand side is evaluated over a triangular
region of the x s plane, …rst the inner integral at a …xed s from x = a to x = s and
the outer integral then runs from s = a to s = x. On reversing the integration order, the
same triangular region must be covered. This is achieved by integrating from s = x to
s = x at a …xed x, followed by integration with respect to x from x = a to x = x.
Exercise 4: Reversed order integrations

a) Check the formula (3.16) for the integrand G(x; s) = x3 s.
b) Derive a reversed order integration form for the double integral
x b
s s G(x; s)dx ds; x 2 [a; b];
a s
and check it for the integrand G(x; s) = x3 s.

Applying the formula (3.16) to the double integral in (3.15) gives
x s x x x x x
s s f (x)dx ds = s s f (x)ds dx = s f (x) s ds dx = s f (x)(x x)dx
a a a x a x a
and the integral representation (3.15) is simpli…ed to

0 x
y(x) = y0 + (x a)y0 + s (x x)f (x)dx (3.17)
a
or, by introducing the Heaviside function to correctly perform the integration over the
whole intervall
0 b
y(x) = y0 + (x a)y0 + s (x x)H(x x)f (x)dx (3.18)
a
This is obviously the same solution as given in (3.10).
Exercise 5: Integral equation by straightforward integrations

The second order di¤erential equation (3.8) has to satisfy the two boundary conditions
d 0 0
y(a) = y0 and y(x) = y (b) = y1
dx x=b
Derive the solution of the boundary value problem by straightforward integrations and
transform the resulting double integral into single integrals applying the formula derived
in Exercise 2.
3.1.2 Direct integral equations by the method of weighted resid-

uals
In general, the above described proceeding of transforming di¤erential equations into
integral equations by straightforward integrations is not applicable. Therefore, a very
general methodology - the method of weighted residuals - is now introduced where, for a
better understanding, the same simple second order ordinary di¤erential equation (3.8) is
considered again …rst.
The idea of this method is as follows: One considers the residual which
remains when an approximative solution is inserted in the di¤erential equa-
tion, multiplies this residual with certain (known) weighting functions, and
demands that the integral of this product over the problem domain disap-
pears, i.e., is zero.
Then, all di¤erentiations acting on the unknown states of the di¤erential
equation are shifted through integration by parts to act on the known weight-
ing functions. If the chosen weighting function is the fundamental solution of
the actually considered di¤erential equation, one obtains an equivalent inte-
gral equation formulation of the boundary value problem.
It can be used to determine the unknown boundary reactions of the problem, and,
when these are found, also the sought solution of the considered di¤erential equation at
any arbitrary interior point.
Transformation of Poisson or Laplace equations
The second order ordinary di¤erential equation (3.8) is the one-dimensional representation
of the so-called Poisson’s or Laplace equation which has, as shown above (see (3.12) and
(3.13)), the fundamental solution ( x)H( x) or equivalently 12 r = 12 j x j.
As physically meaningful examples, the equations from Euler-Bernoulli’theory of elastic
beams are considered.
Bending de‡ection of elastic beams The de‡ection w(x) of an elastic beam un-
der a prescribed bending moment distribution M (x) has to satisfy the inhomogeneous
di¤erential equation of second order:
d2 M (x)
w(x) = ; = f x jx 2 [a; b]; a b = lg (3.19)
dx2 EI
Following the above advices, this equation is multiplied with the fundamental solution
w (x; ) = 21 r as adequate weighting function, integrated over the problem domain, i.e.,
over the beam length l from x = a to x = b, and the result is demanded to be zero:
Zb
d2 w(x) M (x)
+ w (x; )dx = 0 (3.20)
dx2 EI
a
or
Zb Zb
d2 w(x) M (x)
w (x; )dx = w (x; )dx (3.21)
dx2 EI
a a
Now, the left hand side of (3.21) has to be integrated by parts twice in order to shift
the two di¤erentiation in the ’domain’ integral from the unknown w(x) to the known
fundamental solution w (x; ):

Zb b Zb
d2 w(x) dw(x) dw(x) @w (x; )
w (x; )dx = w (x; ) dx
dx2 dx a dx @x
a a
b
dw(x) @w (x; )
= w (x; ) w(x)
dx @x a
Zb
@ 2 w (x; )
+ w(x) dx (3.22)
@x2
a
Since the second derivative of the fundamental solution (2.76) w (x; ) = 12 r = 12 j x j

gives the Dirac function (x; ), the result of the respective domain integral is with 2
[a; b] simply the value of w at (see (2.51)):
Zb Zb
@ 2 w (x; )
w(x) dx = w(x) (x; )dx = w( ) (3.23)
@x2
a a
Hence, with (3.23) and (3.22), the equation (3.21) is transformed into the boundary
integral representation
b Zb
dw(x) @w (x; ) M (x)
w( ) = w (x; ) w(x) w (x; )dx (3.24)
dx @x a EI
a
Explicitly, with (2.77), i.e.,

@w (x; ) 1 1
= H(x ) = sign(x ) (3.25)
@x 2 2
the evaluation of (3.22) gives:
Zl
d2 w(x) 0 jb j 1
w (x; )dx = w (b) w(b) sign(b )
dx2 2 2
0
0 ja j 1
w (a) + w(a) sign(a ) + w( )
2 2
0 (b ) 1 0 a 1
= w (b) w(b) w (a) w(a) + w( )(3.26)
2 2 2 2
Hence, the equation (3.24), the integral representation of the di¤erential equation (3.19),
can also be expressed as:
Zl
1 a 0 1 (b ) 0 M (x) 1
w( ) = w(a) + w (a) + w(b) w (b) jx j dx (3.27)
2 2 2 2 EI 2
0
It is valid for all interior points 2 [a; b] and all combinations of boundary conditions,
but, before it is possible to evaluate this expression, all unknown boundary reactions must
be determined.
For this purpose, the point has to be shifted on the boundary, i.e. here, at the two
boundary points = a and = b, to obtain two equations for the two unknown boundary
values. This gives the following equation system (b a = l):
2 3
w(a) Zb
1 1 0 1 l 6 w0 (a) 7 1 (x a)M (x)
6 7 (3.28)
2 1 l 1 0 4 w(b) 5 = 2EI (b x)M (x)
dx
0 a
w (b)
Dependent on the actually prescribed boundary conditions, the corresponding colums

have to be multiplied with the respective known values and transfered to the right hand
side. The solution of the resulting system delivers the unknown boundary reactions which
are necessary for the evaluation of the integral representation (3.27).
In order to compare with the already solved problem, the initial value problem with
the conditions (3.9)
0 0
w(a) = w0 and w (a) = w0 (3.29)
and f (x) = M (x)=EI is considered here again. Then, the actual algebraic equation
system is
Zl
1 l w(b) (x a)f (x) 1 0 w0
0 = dx 0 (3.30)
1 0 w (b) (b x)f (x) 1 l w0
0
from which one obtains
Zb
0
w(b) = (b x)f (x)dx + w0 + lw0 (3.31)
a
0 1
Zb
0 1@
w (b) = (x a)f (x)dx w0 + w(b)A
l
a
Zb
0
= f (x)dx + w0 (3.32)
a
Inserting these boundary reactions in the integral representation (3.27) of the solution
Zl
1 a 10 (b ) 0 jx j
w( ) = w0 + w0 + w(b) w (b) + f (x) dx
2 2 2 2 2
0
0 l 1
Z
1 a 0 1@ 0
= w0 + w0 + (b x)f (x)dx + w0 + lw0 A
2 2 2
0
0 l 1
Z Zl
(b )@ 0 jx j
f (x)dx + w0 A + f (x) dx
2 2
0 0
Zl
0 1
= w0 + ( a)w0 + f (x) ( x+ j x j) dx
2
0
Zl
0
= w0 + ( a)w0 + f (x)( x)H( x)dx (3.33)
0
where the equality of 12 ( x+ j x j)and ( x)H( x) has been taken into account.
The …nal line in (3.33) is exactly the same expression for the solution of the initial value
problem as given in (3.18) when one recognizes that and x are there x and x, respectively.
Remark: An essential di¤erence of the direct boundary integral formulation obtain
via the method of weighted residuals in comparison to the integral solution by straight-
forward integrations is that one has to determine …rst the unknown boundary reactions
before one can evaluate the expression for the solution at arbitrary interior points.
Exercise 6: Beam de‡ection under prescribed moments Use the above system
(3.28) to solve the boundary value problem with the prescribed conditions w(a) = w0
and w (b) = w1 which corresponds with y , w and f (x) = M (x)=EI to the problem of
0 0
Exercise 3.
Exercise 7: Bending moment of an elastic beams under transversal loading

Transfer the above solution (3.27) and the system (3.28) for determining the boundary
reactions to the di¤erential equation for the bending moment M (x) of an elastic beam
under the prescribed transversal loading q(x):
d2 M (x)
= q(x) (3.34)
dx2
Note that the shear force Q(x)is the …rst derivative of the bending moment, i.e., Q(x) =
0
M (x) = dM (x)=dx.
Exercise 8: Axial displacement of an elastic bar Transfer the above solution (3.27)
and the system (3.28) for determining the boundary reactions to the di¤erential equation
for the axial displacement u(x) of an elastic bar of length l with sectional area A and
Young’s modulus E under the prescribed axial loading p(x):
d2 u(x) p(x)
= (3.35)
dx2 EA
Note that the resultant axial force N (x) is related to the axial displacement u(x) via
0
N (x) = EAu (x).
Exercise 9: Bar stretching under axial loadings Solve the stretching problem of
an elastic bar, which is …xed at x = a = 0, i.e, u(0) = 0 and has a free ending at x = b = l,
i.e., N (l) = 0, with the integral equation system and the solution expression determined
in Exercise 8. The prescribed axial loading is p(x) = p0 xl .
Transformation of Helmholtz equations

The till now considered di¤erential equations contained only …rst or second order deriva-
tives of the sought solution. When the sought solution function multiplied by a constant
factor is added to or substracted from the second derivative term, e.g.,
d2 w(x) 2 d2 w(x)
+ k w(x) = f (x) or h2 w(x) = f (x) (3.36)
dx2 dx2
this di¤erential equation is called of Helmholtz type. When one prefers symmetric forms,
the respective fundamental solutions are (see, (2.88) and (2.92), respectively, and corre-
sponding Remark 1 there)
1
w (x; ) = w (r) = sin(kr) with k 6= 0, real; r = jx j (3.37)
2k
1
w (x; ) = w (r) = sinh(hr) with h 6= 0, real; r = jx j (3.38)
2h
or, instead of the for r ! 1 divergent form (3.38), the convergent one (2.96)
1 hr
u (x; ) = u (r) = e (3.39)
2h
Stationary longitudinal waves in an elastic bar The dynamic equilibrium of a bar

element with the cross section A and material density under an axial loading p(x; t) and
a londitudinal accelleration u•(x) = @ 2 u(x; t)=@t2 is described by
@N (x; t) @ 2 u(x; t)
= p(x; t) + A (3.40)
@x @t2
where the axial resultant force N (x; t) is related to the longitudinal displacement u(x; t)
by
@u(x; t)
N (x; t) = EA = EAu0 (x; t) (3.41)
@x
Connecting both equations and assuming constant cross section A and modulus of elas-
ticity E yields the basic di¤erential equation
@ 2 u(x; t) @ 2 u(x; t)
EA = p(x; t) + A (3.42)
@x2 @t2
p
or, introducing the longitudinal wave speed cL = E= ,
@ 2 u(x; t) 1 @ 2 u(x; t) p(x; t)

= (3.43)
@x2 c2L @t2 EA
For time-harmonic loadings with the excitation frequency !
p(x; t) = p(x)ei!t (3.44)
the response can also be assumed to be time-harmonic:
u(x;t) = u(x)ei!t (3.45)
The result is an equation which is not longer time-dependent and is of Helmholtz type
d2 u(x) 2 p(x)
+ u(x) = (3.46)
dx2 EA
The ratio = !=cL is the so-called wave number.
As we already know, the transformation of this di¤erential equation (3.46) to an
equivalent integral equation may be performed by integrations by parts of the integral of
the weighted residual over the problem domain, i.e. here, over the bar length l:
Zl
d2 u(x) 2 p(x)
+ u(x) + u (x; )dx = 0 (3.47)
dx2 EA
0
where the fundamental solution u (x; ) of the di¤erential equation is taken as special
weigthing function.
The two integrations by parts of the …rst term in (3.47) gives
Zl Zl
d2 u(x) l @u (x; )
u (x; )dx = [u0 (x)u (x; )]0 u0 (x) dx
dx2 @x
0 0
l Zl
0 @u (x; ) @ 2 u (x; )
= u (x)u (x; ) u(x) + u(x) (3.48)
dx
@x 0 @x2
0
such that the complete transformation of (3.47) becomes
l Zl Zl
0 @u (x; ) @ 2 u (x; ) 2 p(x)
u (x)u (x; ) u(x) + + u (x; ) u(x)dx = u (x; )dx
@x 0 @x2 EA
0 0
(3.49)
Here, the adequate fundamental solution is (see (3.38)
1
u (x; ) = u (r) = sin(kr) (3.50)
2k
Its …rst and second derivative, respectively, is (2H(x ) 1 =sign(x ))
@u (x; ) 1 @r 1
= cos(kr) = cos(kr)[2H(x ) 1] (3.51)
@x 2 @x 2
@ 2 u (x; ) k
= sin(kr) + cos(kr) (x ) (3.52)
@x2 2
such that
Zl
l @ 2 u (x; ) 2
s + u (x; ) u(x)dx = cos(kr) (x )u(x)dx = u( ) for 2 [0; l]
0 @x2
0
Thus, the …nal result of (3.49) is the following integral expression for the axial displacement
at an arbitrary point 2 [0; l]:
l Zl
@u (x; ) p(x)
u( ) = u0 (x)u (x; ) u(x) u (x; )dx
@x 0 EA
0
1 1
= sink(l )u0 (l) + cosk(l )sign(l )u(l)
2k 2
1 1 l p(x)
+ sin(k )u0 (0) cos(k )sign( )u(0) s u (x; )dx
2k 2 0 EA
1 1
= sink(l )N (l) + cosk(l )u(l)
2EAk 2
Zl
1 1 p(x) 1
+ sin(k )N (0) + cos(k )u(0) sin(kr)dx (3.53)
2EAk 2 EA 2k
0
where the relation N (x) = EAu0 (x) was applied in order to introduce the adequate
boundary state N (x).
Since two of the four boundary values are unknown, one needs two equation to deter-
mine their values. They are obtain by collocation, i.e., evaluation of the equation (3.53)
at the two boundary points (both equations have been multiplied by 2):
Zl
1 p(x)
= 0 : u(0) cos(kl) u(l) + sin(kl) N (l) = sin(kx)dx
EAk EAk
0
Zl
1 p(x)
= l: cos(kl) u(0) + u(l) sin(kl) N (0) = sink(l x)dx(3.54)
EAk EAk
0
or in matrix notation
2 3
u(0)
Zl
1 0 cos(kl) sin(kl) 6 N (0) 7 1 p(x)sin(kx)
6 EAk 7 = dx
cos(kl) sin(kl) 1 0 4 u(l) 5 EAk p(x)sink(l x)
N (l) 0
EAk
(3.55)
Exercise 10: Torsional twist of an elastic bar Transfer the above integral form
(3.53) of the solution and the equation system (3.54) for determining the boundary reac-
tions to the di¤erential equation for the angular twist change #(x) = d =dx of an elastic
bar of length l under a torsional moment MT (x):
d2 #(x) MT (x)
h2 #(x) = (3.56)
dx2 ECT
GIT
The constant factor h2 = ECT
is the ratio of the torsional sti¤ness GIT and the warping
resistance ECT .
Transformation of a Bilaplacian equation
When the 1-d form of the Laplace operator =d2 =dx2 is applied twice to a sought
function w(x), the so-called Bilaplacian or Biharmonic equation is obtained:
d2 d2
L(w) = ( w(x)) = f (x) (3.57)
dx2 dx2
Its fundamental solution is found by integrating two times the fundamental solution
w (x; ) = 12 r of the Poisson equation (??) which gives
1 3
w (x; ) = r (3.58)
12
This can easily be checked by straight-forward di¤erentiations
@w (x; ) 1 2 @r 1
= r = r2 (2H(x ) 1) (3.59)
@x 4 @x 4
2 2
@ w (x; ) 1 @r 1 1
= r + r2 (x )= r (3.60)
@x2 2 @x 2 2
@ 3 w (x; ) 1 @r 1
= = H(x ) (3.61)
@x3 2 @x 2
4
@ w (x; )
= (x ) (3.62)
@x4
Remark: The term 21 r2 (x ) in the second derivative can be neglected since it is zero
due to r = 0 for x = .
Now, this most simple form of a 4th order di¤erential equation will be tranformed into
0
an integral formulation for its solution w(x) and for other related states, e.g., w (x).
The Euler-Bernoulli beam In the Euler-Bernoulli theory for the bending of elastic
beams, the de‡ection w(x) is described by the 4th order di¤erential equation
d4 w(x)
EI = q(x) (3.63)
dx4
where EI means its bending sti¤ness. For a unique solution, four boundary conditions
have to be prescribed where at each boundary point two boundary values are known
corresponding to the actual support while the other two are unknown reactions, e.g.,
0
for clamped endings w = 0 w = 0
for a free ending M =0 Q=0 (3.64)
for a simple support w = 0 M = 0
Integral equation for the beam de‡ection The method of weighted residual
postulates
Zl
d4 w(x)
EI q(x) w (x; )dx = 0 (3.65)
dx4
0
or
Zl Zl
d4 w(x)
EI w (x; )dx = q(x) w (x; )dx (3.66)
dx4
0 0
where the weighting function w (x; ) has to be the fundamental solution of the di¤erential
equation (3.63). This is obviously obtained from (3.58) by simply dividing by EI
r3
w (x; ) = (3.67)
12EI
Its derivatives di¤er from (3.59) to (3.62) only by the factor 1=EI, e.g.,
@ 3 w (x; ) 1 @r 1
3
= = [2H(x ) 1] (3.68)
@x 2EI @x 2EI
The procedure for deriving the boundary integral form is analogous to that in the case
of the di¤erential equations of second order: one has only to integrate by parts four times
instead of two times. Having in mind that
d2 w(x) d3 w(x)
EI = M (x); EI = Q(x) (3.69)
dx2 dx3
one obtains by the …rst integration by parts
Zl l Zl
d4 w(x) d3 w(x) d3 w(x) @w (x; )
EI w (x; )dx = EI w (x; ) EI dx
dx4 dx3 0 dx3 @x
0 0
by two integrations by parts

Zl l
d4 w(x) d2 w(x) @w (x; )
EI w (x; )dx = Q(x)w (x; ) EI
dx4 dx2 @x 0
0
Zl
d2 w(x) @ 2 w (x; )
+ EI dx
dx2 @x2
0
by three integrations by parts

Zl h il
d4 w(x) 0 0
EI 4
w (x; )dx = Q(x)w (x; ) + M (x)w (x; ) w (x)M (x; )
dx 0
0
Zl
dw(x) @ 3 w (x; )
EI dx
dx @x3
0
and by the last forth integration by parts

Zl h il
d4 w(x) 0 0
EI 4
w (x; )dx = Q(x)w (x; ) + M (x)w (x; ) w (x)M (x; )
dx 0
0
l Zl
@ 3 w (x; ) @ 4 w (x; )
EIw(x) + EIw(x) dx
@x3 0 @x4
0
0 l
Q(x)w (x; ) + M (x)w (x; )
= 0
w (x)M (x; ) + w(x)Q (x; ) 0
Zl
(x )
+ EIw(x) dx
EI
0
Hence, taking the ’…ltering’e¤ect of the Dirac function into account, one gets
h 0 0
il
w( ) = Q(x)w (x; ) M (x)w (x; ) + w (x)M (x; ) w(x)Q (x; )
0
l
+s q(x) w (x; )dx (3.70)
0
or explicitly
w( ) = Q(l)w (l; ) Q(0)w (0; ) w(l)Q (l; ) + w(0)Q (0; )

0 0 0 0
M (l)w (l; ) + w (l)M (l; ) + M (0)w (0; ) w (0)M (0; )
Zl
+ q(x) w (x; )dx (3.71)
0
In this equation, four of the eight boundary values are known and the other four are
unknown reactions. Evaluating this equation at the two boundary points, i.e., at = 0+"
and = l " gives the two boundary ’integral’equations
Zl
1 1 l3 l2 0 l x3
w(0) w(l) Q(l) + M (l) + w (l) = q(x) dx (3.72)
2 2 12EI 4EI 2 12EI
0
Zl
1 1 l3 l2 0 l (x l)3
w(0) + w(l) + Q(0) + M (0) w (0) = q(x) dx (3.73)
2 2 12EI 4EI 2 12EI
0
where the evaluation of the fundamental solution has given (" ! 0)
l3 l3
w (l; 0) = , w (0; 0) = 0, w (l; l) = 0, w (0; l) = (3.74)
12EI 12EI
and for the related fundamental states
0 l2 0 0 0 l2
w (l; 0) = , w (0; 0) = 0, w (l; l) = 0, w (0; l) = (3.75)
4EI 4EI
00 l l
M (l; 0) = EIw (l; 0) = , M (0; 0) = 0, M (l; l) = 0, M (0; l) = (3.76)
2 2
1 1 1 1
Q (l; 0) = , Q (0; 0) = , Q (l; l) = , Q (0; l) = (3.77)
2 2 2 2
Now, we obtained two boundary integral equations but one needs two equations more.
Integral equation for the beam slope Obviously, one gets a new integral equation
0
for the beam slope w ( ), which means a rotation about the y-axis, when one di¤erentiates
the integral equation (3.70) with respect to the variable . Since the equation (3.70) was
obtained by integration by parts of the weighted residual (3.66) of the di¤erential equation
(3.63), it is clear that one can perform the di¤erentiation with respect to also directly
to the weighted residual:
Zl Zl
d4 w(x) @w (x; ) @w (x; )
EI dx = q(x) dx (3.78)
dx4 @ @
0 0
Hence, the relation (3.78) is also a weighted residual form of the beam equation (3.63)
where only another weighting function is applied, namely
@w (x; ) @w (x; ) 1
w2 (x; ) = = = (x )2 sgn(x ) (3.79)
@ @x 4EI
with
@w2 (x; ) @ 2 w (x; ) r
= 2
=
@x @x 2EI
@2 1 @r 1
w (x; ) = = [2H(x ) 1]
@x2 2 2EI @x 2EI
@3 1
3
w2 (x; ) = (x )
@x EI
Then, integration by parts yields at …rst
Zl l Zl
d4 w(x) d3 w(x) d3 w(x) @w2 (x; )
EI w2 (x; )dx = EI w2 (x; ) EI dx
dx4 dx3 0 dx3 @x
0 0
Zl
d3 w(x) @w2 (x; )
= [ Q(x)w2 (x; )]l0 EI dx (3.80)
dx3 @x
0
by the second integration

Zl l
d4 w(x) d2 w(x) @w2 (x; )
EI w2 (x; )dx = Q(x)w2 (x; ) EI
dx4 dx2 @x 0
0
Zl
d2 w(x) @ 2 w2 (x; )
+ EI dx
dx2 @x2
0
l Zl
@w (x; ) d2 w(x) @ 2 w2 (x; )
= Q(x)w2 (x; ) + M (x) 2 + EI dx
@x 0 dx2 @x2
0
and …nally already by the third integration by parts (M2 (x; ) = EI@ 2 w2 (x; )=@x2 )
Zl l
d4 w(x) @w (x; ) dw(x) @ 2 w2 (x; )
EI w2 (x; )dx = Q(x)w2 (x; ) + M (x) 2 + EI
dx4 @x dx @x2 0
0
Zl
dw(x) @ 3 w2 (x; )
EI dx
dx @x3
0
l
@w (x; ) dw(x)
= Q(x)w2 (x; ) + M (x) 2 M2 (x; )
@x dx 0
Z l
dw(x)
+ (x; )dx
dx
0
l
@w (x; ) dw(x)
= Q(x)w2 (x; ) + M (x) 2 M2 (x; )
@x dx 0
dw(x)
+
dx x=
0
Consequentyl, the integral equation for w ( ) at interior points 2 (0; l) reads as
l Zl
0 @w (x; ) dw(x)
w ( ) = Q(x)w2 (x; ) M (x) 2 + M2 (x; ) + q(x) w2 (x; )dx (3.81)
@x dx 0
0
The evaluation of this equation at the two boundary points gives the two extra equations
for the determination of the four unknown boundary reactions:
for = 0 + " (" ! 0):
l Zl
0 @w (x; ") dw(x)
w (0) Q(x)w2 (x; ") M (x) 2 + M2 (x; ") = q(x) w2 (x; 0)dx (3.82)
@x dx 0
0
for =l " (" ! 0):
l Zl
0 @w (x; l ") dw(x)
w (l) Q(x)w2 (x; l ") M (x) 2 + M2 (x; l ") = q(x) w2 (x; l)dx
@x dx 0
0
(3.83)
When the respective function values (with " ! 0) of this weighting function and of its
derivatives, respectively,
l2 l2
w2 (l; ") = , w (0; ") = 0, w2 (l; l ") = 0, w2 (0; l ") = (3.84)
4EI 2 4EI
0 l 0 0 0 l
w2 (l; ") = , w2 (0; ") = 0, w2 (l; l ") = 0, w2 (0; l ") = (3.85)
2EI 2EI
1 1 1 1
M2 (l; ") = , M2 (0; ") = , M2 (l; l ") = , M2 (0; l ") = (3.86)
2 2 2 2
are introduced, one obtains the two equations:
for = 0:
Zl
1 0 0 l2 l x2
[w (0) w (l)] + Q(l) M (l) = q(x) dx (3.87)
2 4EI 2EI 4EI
0
for = l:
Zl
1 0 0 l2 l (x l)2
[w (l) w (0)] + Q(0) + M (0) = q(x) dx (3.88)
2 4EI 2EI 4EI
0
The complete system of integral equations for the de‡ection and the slope
Althogether, the equations (3.72), (3.73), (3.87), and (3.88) result the following system
(in matrix-vector notation)
2 3
w(0)
36 7
0
2 6 w (0) 7 2 3
3
1 1 l l2 l3 6 7 x
q(x) 12EI
0 0 0 M (0)
6 2 2 2 4EI 12EI
76 7 Zl 6 l)3 7
6 1 l l2 l3 1
0 0 0 76 Q(0) 7 6 q(x) (x12EI 7
6 2 2 4EI 12EI 2 76 7= 6 7 dx
4 0 1
2
0 0 0 2
1 l
2EI
l2
4EI
56
6 w(l) 7
7 4 x2
q(x) 4EI 5
1 l l2 1 6 0
w (l) 7 0 2
0 2 2EI 4EI
0 2
0 0 6 7 q(x) (x4EIl)
4 M (l) 5
Q(l)
(3.89)
A rearrangement of these equations, i.e., an interchanging of the …rst and forth line, gives
a more systematic order of the coe¢ cient matrix::
2 3
w(0)
36 7
0
2 6 w (0) 7 2 2
3
l2
0 1 l
0 1
0 0 6 M (0) 7 q(x) (x4EIl)
6 1 2 2EI 4EI 2
76 7 Zl 6 7
6 2 l l2 l3 1
0 0 0 76 Q(0) 7 6 q(x) (x12EIl)3
7
6 2 4EI 12EI 2 76 7= 6 7 dx
4 0 1
2
0 0 0 2
1
2EI
l l2
4EI
56
6 w(l) 7
7 4 x2
q(x) 4EI 5
1
0 0 0 1 l l 2 l3 6 0
w (l) 7 0 x3
2 2 2 4EI 12EI 6 7 q(x) 12EI
4 M (l) 5
Q(l)
(3.90)
3.1.3 Integral formulation with Green’s functions

As described in Chapter 2.?, Green’s functions are special fundamental solutions which
additionally satisfy certain homogeneous boundary conditions, or more exactly, those
boundary states of the Green’s function have to be zero which are prescribed in the
actual problem.
Stretching of bars
As derived in Chapter ? for general Laplace equations and explicitly given for bars in the
solution of Exercise 7, the direct form of the integral equation for the axial displacement
of an elastic bar of length l = b a is
b Zb
N (x) N (x; ) p(x)
u( ) = u (x; ) u(x) u (x; )dx (3.91)
EA EA a EA
a
There, the fundamental solution is (see, (2.76))

1 1
u (x; ) = r = j x j (3.92)
2 2
and satis…es the 1-dimensional Laplace equation
d2 u (x; )
= (x; ) (3.93)
dx2
When an actual boundary value problem is, e.g.,
d2 u(x) p(x)
= (3.94)
dx2 EA
with the boundary conditions
du(x)
u(x = a) = u0 and N (x = b) = EA = Nl (3.95)
dx x=b
the adequate Green’s function GuN (x; ) of this problem has to satisfy the equation (3.93),
i.e., has to be a fundamental solution, but has additionally to ful…ll the homogeneous
conditions
@GuN (x; )
GuN (x = a; ) = 0 and EA =0 (3.96)
@x x=b
Hence, as easily can be found, this special Green’s function is given as
@u (x; )
GuN (x; ) = u (x; ) u (a; ) (x a) (3.97)
@x x=b
Now, applying in the integral equation (3.91) instead of u (x; ) the adequate Green’s
function (3.97) and taking the actual boundary conditions (3.95) and the corresponding
homogeneous boundary conditions of the Green’s function (3.96) into account, the solution
of the boundary value problem, the axial displacement at any interior point (a b)
is directly found to be
Nl uN @GuN (x; ) b p(x) uN

u( ) = G (b; ) u0 s G (x; )dx (3.98)
EA @x x=a a EA
Since with a b
@u (x; )
GuN (b; ) = u (b; ) u (a; ) (b a)
@x x=b
1 1 1
= jb j ja j sign(b )l
2 2 2
1
= (b +a l)
2
= a (3.99)
and
@GuN (x; ) @u (x; ) @u (x; )

=
@x x=a @x x=a @x x=b
1 1
= sign(a ) sign(b )
2 2
= 1 (3.100)
the solution (3.98) is explicitly
Zb
Nl p(x) 1
u( ) = (a ) + u0 (j x j ja j sign(b )(x a)) dx
EA EA 2
a
Zb
Nl p(x) 1
= ( a) + u0 (j x j ( a) (x a)) dx
EA EA 2
a
Z
Nl p(x) 1
= ( a) + u0 ( x + 2a x) dx
EA EA 2
a
Zb
p(x) 1
(x + 2a x) dx
EA 2
Z Zb
Nl p(x) p(x)
= ( a) + u0 ( x + a) dx ( + a) dx (3.101)
EA EA EA
a
e.g., for p(x) = p0 =const. and with a = 0 and b = l

0 1
Z Zl
Nl p0 @
u( ) = + u0 ( x) dx + ( )dxA
EA EA
0
!
Nl p0 x2
= + u0 (l )
EA EA 2 0
2
Nl p0
= + u0 (l )
EA EA 2
2
Nl p0
= + u0 + + l (3.102)
EA EA 2
As easily can be checked, this is the exact solution of the problem.
Remark: The Green’s functions for bar problems where
a) at x = a a prescribed force N0 is acting and at x = b the axial displacement is
prescribed, i.e.,
du(x)
N (x = a) = EA = N0 and u(x = b) = ul (3.103)
dx x=a
is given by
@u (x; )
GN u (x; ) = u (x; ) u (b; ) (x b) (3.104)
@x x=a
b) at x = a and at x = b displacements are prescribed, i.e.,
u(x = a) = u0 and u(x = b) = ul (3.105)
is given by
b x x a
Guu (x; ) = u (x; ) u (a; ) u (b; ) (3.106)
b a b a
Bending of beams
As given in (3.71), the direct form of the integral equation for the de‡ection of an elastic
beam of length l is
w( ) = Q(l)w (l; ) Q(0)w (0; ) w(l)Q (l; ) + w(0)Q (0; )
0 0 0 0
M (l)w (l; ) + w (l)M (l; ) + M (0)w (0; ) w (0)M (0; )
Zl
+ q(x) w (x; )dx (3.107)
0
where, the fundamental solution (see, (3.67))

r3
w (x; ) = (3.108)
12EI
satis…es the 1-dimensional Bi-potential equation

@ 4 w (x; ) (x; )
4
= (3.109)
@x EI
Considering as an actual problem
d4 w(x) q(x)
4
= (3.110)
dx EI
with the boundary conditions
0
w(x = 0) = 0 and w (x = 0) = 0 (3.111)
w(x = l) = wl and M (x = l) = 0 (3.112)
i.e., a beam with a clamped boundary at x = 0 and with a pinned support at x = l which
has su¤ered a vertical settlement of w(x = l) = wl , the above integral equation reads as
0 0
w( ) = Q(l)w (l; ) Q(0)w (0; ) + w (l)M (l; ) + M (0)w (0; )
Zl
wl Q (l; ) + q(x) w (x; )dx (3.113)
0
The adequate Green’s function Gcs (x; ) of this problem has to satisfy the equation (3.109),
i.e., has to be a fundamental solution, but has additionally to ful…ll the homogeneous
conditions
Gcs (l; ) = 0 and Gcs (0; ) = 0 (3.114)

@ 2 Gcs (x; )
M (Gcs (x; ))jx=l = EI =0 (3.115)
@x2 x=l
@Gcs (x; )
= 0 (3.116)
@x x=0
Then, the de‡ection w( ) at an arbitrary point solution is simply
Zl
cs
w( ) = wl Q (G (x; ))jx=l + q(x) Gcs (x; )dx (3.117)
0
The derivation of the Green’s function Gcs (x; ) of the above de…ned problem can start
with an ’ansatz’ which combines the fundamental solution w (x; ) and its derivatives,
respectively, with unknown polynomials h1 (x); h2 (x); h3 (x); and h4 (x) adequately
Gcs (x; ) = w (x; ) w (0; )h1 (x) w (l; )h2 (x)

@w (x; ) @ 2 w (x; )
h3 (x) h4 (x) (3.118)
@x x=0 @x2 x=l
These polynomials have to be cubic and, as follows from the conditions (3.114) to (3.116)
must satisfy the conditions
0 00
h1 (0) = 1; h1 (l) = 0; h1 (0) = 0; h1 (l) = 0
0 00
h2 (0) = 0; h2 (l) = 1; h2 (0) = 0; h2 (l) = 0
0 00
h3 (0) = 0; h3 (l) = 0; h3 (0) = 1; h3 (l) = 0
0 00
h4 (0) = 0; h4 (l) = 0; h4 (0) = 0; h4 (l) = 1 (3.119)
Some simple analysis gives
3 x 2 1 x 3
h1 (x) = 1 + (3.120)
2 l 2 l
3 x 2 1 x 3
h2 (x) = (3.121)
2 l 2 l
x 3 x 2 1 x 3
h3 (x) = l + (3.122)
l 2 l 2 l
2
l x 2 x 3
h3 (x) = + (3.123)
4 l l
Then, introducing for 0 l with @r=@x =sign(x ) and r2 (x )=0
3
(l )3
w (0; ) = ; w (l; ) = ;
12EI 12EI
2
@w (x; ) @ 2 w (x; ) (l )
= ; 2
= (3.124)
@x x=0 4EI @x x=l 2EI
one obtains after some re-arrangements the Green’s function explicitly as (r3 = jx j3 =
(x )3 sign(x ))
( )
1 3 2 3 3 2 2 x 2
r + 3x + 3( 3 l + l )
Gcs (x; ) = 3
l (3.125)
12EI ( 3 3 2 l + l3 ) xl
Remark: It should be mentioned that the same result can be found by starting with the
general polynomial ’ansatz’
1
Gcs (x; ) = r 3 + c1 x3 + c2 x2 + c3 x + c4 (3.126)
12EI
and determing the four constants c1 ; c2 ; c3 ; and c4 via the four homogeneous boundary
conditions (3.114) to (3.116).
Example: .By using the above Green’s function, the de‡ection function w( ) of a
beam with a clamped boundary at x = 0 and with a pinned support at x = l which has
a vertical settlement at x = l of w(x = l) = wl and is continuously loaded by q(x) = q0 is
given as (see, (3.117))
Zl
cs
w( ) = wl Q (G (x; ))jx=l + q0 Gcs (x; )dx
0
i.e., one has to evaluate at x = l (r;x =sign(x ))

@ 3 cs
Q (Gcs (x; )) = EI G (x; )
x3
1 3 6
= 6r;x ( 3 2 l + l3 ) ;
12 l3
i.e.
( )
3 2
1
Q (Gcs (x; ))jx=l = 1 +3 1
2 l l
2
1
= 3
2 l l
and to integrate Gcs (x; ) along the beam
Zl Zl ( 3 x 2
)
1 r + 3x 2 3
+ 3( 3 3 2 l + l ) 2
Gcs (x; )dx = 3
l dx
12EI ( 3 3 2 l + l3 ) xl
0 0
8 h 2 il 9
>
> 1 4
+ (l ) 4
+ 3 x 2
x 3 >
>
>
> 4 2 >
>
1 < 3 2 2
h
l x
i 0
3 l
=
= +3( 3 l+ l ) 3 l
12EI >
> h il 0 >
>
>
> l x 4
>
>
: ( 3 2 3
3 l+l ) 4 l ;
0
4 3 2 2
1 + (l 1
) +
4
4
2
l l 3
= 2 3 3 2
12EI +( 2
3 l + l )l ( 3 l + l3 ) 4l
1
= 2 4 5l 3 + 3l2 2
48EI
Remark: For 0 l, one has to integrate r3 = jx j3 as follows
Zl Zl
r3 dx = jx j3 dx
0 0
Z Zl
3
= ( x) dx + (x )3 dx
0
l
( x)4 (x )4
= +
4 0 4
1 4
= + (l )4
4
Finally, one …nds the de‡ection for any position 2 [0; l] to be
2
1 q0 4 3 2
w( ) = wl 3 + 2 5l + 3l2
2 l l 48EI
which is, as easily can be checked, the exact solution for these boundary conditions and
the constant loading.
Exercise 11: Green’functions for beam problems

Determine the adequate Green’s function for a beam which has
a) a clamped support at both endings
b) a clamped support at x = 0 while the other ending is free:
3.1.4 Indirect integral formulations: the singularity method

One important feature of the indirect method is that the physical variables of the boundary
value problems, the unknown boundary reactions, do not remain the unknown quantities
of the integral equation: Intermediary unknowns - unknown intensities of certain
singularity layers - are introduced instead. For these singularities, so-called
in‡uence functions, i.e., the complete response to the action of a singularity
(e.g., a unit point force) must be known everywhere in the considered material.
Then, these singularity layers are distributed on a ’…ctitious’boundary + in
a certain small distance d" from the real boundary outside the domain
and their intensities have to be determined such that the integrated response
is equal to the prescribed boundary values on the real boundary .
Since it is di¢ cult to choose an optimal size of this small distance d" , it is the best
choice to use d" = 0, i.e., let the …ctitious boundary + coincide (in the limit from outside)
with the real boundary .
Representation of Poisson equation problems: Stretching of bars

The di¤erential equation for the axial displacement u(x) of an elastic bar with sectional
area A and modulus of elasticity E under the prescribed axial loading p(x) is by (see
(3.35), Example 6):
d2 u(x) p(x)
2
= (3.127)
dx EA
and we know (see (2.78)) the fundamental solution u (x; ) = r=2 of the Poisson equation
@ 2 u (x; )
= (x; )
@x2
Comparing both equations, it is obvious that from its physical meaning
u (x; ) r
(up) (x; ) = = (3.128)
EA 2EA
gives the axial displacement at the point x due to a axial unit point force at point , i.e.,
is the in‡uence function of a point force with intensity 1 for the axial displacement in the
bar stretching problem.
Since the resultant axial force N (x) is related to the axial displacement u(x) via
0
N (x) = EAu (x), the corresponding in‡uence function for this state is obtained by ap-
plying this de…nition to (3.128) as
@ (up) (x; ) 1 @r 1 1
(N p) (x; ) = EA = = [2H(x ) 1] = sgn(x ) (3.129)
@x 2 @x 2 2
Following the above described idea of the indirect method, one has to in-
troduce at the points on the …ctitious boundary + (which is either enclosing
the real boundary with a certain distance d" or both boundaries coincide
with each other) the intensity p ( ) of an adequate singularity, here, of a point
forces, as new unknown function. Then, this intensity p ( ) of the singularity
layers must take such a distribution that all prescribed boundary conditions
on the real boundary will be satis…ed, either pointwise or in some other
sense (certain norm).
For demonstrating this idea, the stretching of an elastic bar of length l = b a with
sectional area A and Young’s modulus E under the prescribed axial loading p(x) (see,
Exercise 6) is considered with mixed boundary conditions: the bar shall be …xed at the
boundary point x = a, i.e., u(x = a) = 0, and shall have a free ending at the other
boundary point x = b, i.e., N (x = b) = 0.
The prescribed axial loading p has to be considered as a point force singu-
larity layer with prescribed intensity p( ) in the interior of the bar’s domain
= (a; b).
The axial displacement u at the point x caused by the point force distribution with
unknown intensity p ( ) on the …ctitiuos boundary + with an arbitrary small distance d"
from the boundary points x = a and x = b, i.e., at = a d" and at = b + d" , and
by the axial loading in the bar’s interior (a; b) with the prescribed intensity p( ) may be
expressed applying the in‡uence function (up) (x; ) as
b
=b+d"
u(x) = [(up) (x; )p ( )] =a d" + s (up) (x; ) p( )d (3.130)
a
Correspondingly, using the in‡uence function (N p) (x; ), the resultant axial force N (x)
may be expressed as:
b
=b+d"
N (x) = [(N p) (x; )p ( )] =a d" + s (N p) (x; )p( )d (3.131)
a
These two boundary ’integral’equations give the boundary value problem solution at all
points x in the closed domain a x b if the intensities p (a d" ) and p (b + d" ) are
determined so that the prescribed boundary conditions are satis…ed, i.e.:
b
=b+d"
u(a) = 0 = [(up) (a; )p ( )] =a d" + s (up) (a; )p( )d (3.132)
a
b
=b+d"
N (b) = 0 = [(N p) (b; )p ( )] =a d" + s (N p) (b; )p( )d (3.133)
a
or, more detailed, by inserting the above de…ned in‡uence functions (r = jx j, l = b a,

d" > 0)
ja b d" j ja a + d" j bja j

p (b + d" ) + p (a d" ) = s p( )d
2EA 2EA a 2EA
b
(l + d" ) p (b + d" ) + d" p (a d" ) = s( a)p( )d (3.134)
a
sgn(b b d" ) sgn(b a + d" ) b sgn(b )

p (b + d" ) + p (a d" ) = s p( )d
2 2 a 2
b
p (b + d" ) + p (a d" ) = s p( )d (3.135)
a
For demonstrating the correctness of these two indirect boundary integral forms, an ex-
ample shall be solved explicitly and compared with the exact analytical solution.
Example: Fixed-free bar under linear axial loading The above equations (3.134)
and (3.135) are for a prescribed axial loading p(x) = p0 x l a :
b a p0 l 2
(l + d" ) p (b + d" ) + d" p (a d" ) = s ( a)p0 d = (3.136)
a l 3
and
b a l
p (b + d" ) + p (a d" ) = s p0 d = p0 (3.137)
a l 2
These two equations give the sought intensities
p0 l 5l + 3d" p0 l 2l 3d"
p (a d" ) = and p (b + d" ) = (3.138)
6 l + 2d" 6 l + 2d"
For the special case d" = 0, i.e., for choosing the …ctitious boundary to be identic with the real
boundary, i.e., for + = , the result is simpli…ed to
5 1
p (a) = p0 l and p (b) = p0 l (3.139)
6 3
With the intensities (3.138) the indirect integral equations (3.130) can be evaluated for any
arbitrary interior point x as
=b+d"
jx j bjx j a
u(x) = p( ) s p0 d
2EA =a d" a 2EA l
p0 l 2l 3d" 5l + 3d"
= (b + d" x) + (x a + d" )
12EA l + 2d" l + 2d"
1 p0 x b
s s (x )( a)d
2EA l a x
p0 l
= (3xl + 2bl 5al + d" (6x 3a 3b + 7l))
12EA(l + 2d" )
p0 (x a)3 (x a)l2 l3
+ (3.140)
2EAl 3 2 3
At …rst, this solution does not look like the exact solution which easily can be determined by
direct integrations to be
p0 (x a)
uexact (x) = (x a)2 3l2 (3.141)
EA 6l
but eliminating b in (3.140) by b = a + l shows that one can cancel the term (l + 2d" ) in
p0 l (3xl + 2bl 5al + d" (6x 3a 3b + 7l)) p0 l (3l(x a) + 2l2 + d" (6(x a) + 4l))
=
12EA(l + 2d" ) 12EA(l + 2d" )
p0 l
= (3(x a) + 2l) (l + 2d" )
12EA(l + 2d" )
p0 l
= (3(x a) + 2l) (3.142)
12EA
Obviously , this term, i.e., the result of the boundary ’integral’on the …ctitious boundary + , is
independent on the distance d" . Now, subtracting from (3.142) the result of the domain integral
p0 l p0 (x a)3 (x a)l2 l3
u(x) = (3(x a) + 2l) +
12EA 2EAl 3 2 3
p0 2 3 3 2 3
= 3l (x a) + 2l 2(x a) + 3(x a)l 2l
12EAl
p0 (x a)
= 3l2 (x a)2
6EAl
the exact solution (3.141) is obtained. Evaluating in the same way the indirect integral equation
(3.131) for the resultant axial force
=b+d"
1 b1 a
N (x) = sgn(x )p ( ) s sgn(x )p0 d
2 =a d" a2 l
1 p0 l 2l 3d" 1 p0 l 5l + 3d" p0 x b
= + s s ( a)d
2 6 l + 2d" 2 6 l + 2d" 2l a x
p0 l 3l + 6d" p0 l2 p0 l p0 l2
= (x a)2 = 3 (x a)2
12 l + 2d" 2l 2 12 2l 2
p0 2
= l (x a)2 (3.143)
2l
gives again the exact solution.

Remark: Obviously, for one-dimensional problems where no approximation errors (no
discretization of the boundary, no ansatz functions for the state functions, no numerical
integration, no point collocation, and no numerical equation solver) occur, the exact
solution is always obtained and the distance d" of the …ctitious boundary + does not
in‡uence the solution.
3.2 2-d and 3-d problems: Transformation of partial

di¤erential equations
In the case of two- and three-dimensional problems, a direct analytical integration of the
partial di¤erential equations is not possible. It is necessary to transform the di¤erential
equation either by the method of weighted residuals or by the so-called singularity method
in integral equations. Their solution is generally possible only by using discretization
techniques.
3.2.1 Direct integral equations by the method of weighted resid-

uals
The Poisson or Laplace equation is a partial di¤erential equation for a scalar function
and can, therefore, be handled relatively simple. Hence, the transformation of a partial
di¤erential equation to an integral equation shall be demonstrated for this case …rst.
Problems governed by this equation appear in di¤erent …elds. Without aiming to be
exhaustive a certain number of them have been summarized in the following table.
3.2 2-d and 3-d problems: Transformation of partial di¤erential equations 61
Problems Scalar function Boundary conditions Constants

=0 = q = K @@n (K)
Thermal
Temperature Heat ‡ow
Heat transfer T =T @T conductivity
(T Deg:) q= @n ( )
Stream function @
Ideal ‡uid ‡ow = q=
( m2 s 1 ) @n
Groundwater Hydraulic head Permeability

= q = K @@n
‡ow ( m) (K)
Hydrodynamic P =P
Pressure q= an Density
pressure on free surface:
(P N m 2 ) (an normal accel.) ( )
moving surfaces P =0
Field potential Permittivity
Electrostatic V =V q= " @V
@n
(V volt) (")
Electric Electropotential 1 @E Resistivity
E=E q= k @n
conduction (E volt) (k)
Magnetic
Magnetic Potential @M
Magnetostatic M =M B= @n
Permeability
(M amp)
( )
Stationary heat conduction - a Poisson equation

In the case of a homogeneous isotropic body with constant thermal conductivity 0 , the
stationary (not longer changing with time) temperature …eld (x) is governed by the
scalar Poisson equation
1
(x) = Wq (x) (3.144)
0
where Wq (x) represents the heat source generation rate. Associated boundary conditions
may involve either a …xed given temperature (a so-called Dirichlet boundary condition)
at a part of the boundary
(x) = (x) for x 2 (3.145)
or a prescribed heat ‡ux through a boundary part (a so-called Neumann boundary con-
dition):
@ (x)
qn (x) = 0 = qn (x) für x 2 (3.146)
@n(x)
where n(x) is the outward normal unit vector. This heat ‡ux can be zero when the
boundary is insulated.
Almost no real problems have purely temperature or ‡ux speci…ed boundary condi-
tions, so it is necessary to consider mixed boundary conditions from the beginning, i.e.,
one has on one boundary part 1 the condition (3.145) and on the remaining part 2 the
condition (3.146).
Now, as in the above discussed one-dimensional problems, the weighted residual of the
considered di¤erential equation (3.144) is integrated over the domain of the problem
and set to be zero:
Wq (x)
s (x) + (x; )d x = 0 (3.147)
0
where the fundamental solution (x; ) of the respective di¤erential equation is taken as
special weighting function.
The respective fundamental solutions are known (see, (2.87) and (2.85)) to be (r =j
x j):
1 r 1
(x; ) = ln( ) = [ln(r) ln(c)] in R2 (3.148)
2 c 2
1
(x; ) = (r) = in R3 (3.149)
4 r
where c > 0 is an arbitrary real constant making the ratio r=a dimensionless, e.g., in the
case of a numerical solution procedure, taken as the smallest geometrical dimension of
the discretization.
Now, following the above introduced rules for deriving an equivalent integral equation
representation by the method of weighted residuals, the …rst di¤erential operator term in
(3.147) has to integrated by parts till all di¤erentiations are transfered from the unknown
function (x) to the known weighting function, the fundamental solution (x; ). This
gives ( @x@ i =;i ) for 2 :
Z Z
(x) (x; )d x = [ (x)];ii (x; )d x
Z
= f[ (x)];i (x; )g;i [ (x)];i [ (x; )];i d x
Z
= s [ (x)];i (x; )ni (x)d x [ (x)];i [ (x; )];i d x
1
= s qn (x) (x; )d x
0
Z
f (x)[ (x; )];i g;i (x)[ (x; )];ii d x
1
= s qn (x) (x; )d x s (x)[ (x; )];i ni (x)d x
0
Z
+ (x) (x; )d x
1
= s [qn (x) (x; ) (x)qn (x; )] d x ( ) (3.150)
0
where, as de…ned in (3.146), the heat ‡ux qn (x) was introduced as second boundary
state. Combining this transformed expression with the other terms of (3.147) delivers the
following integral equation for the temperature at an arbitrary interior point 2 :
1 Wq (x)
( )=s (qn (x) (x; ) (x)qn (x; )) d x +s (x; )d x (3.151)
0 0
where
qn (x; ) = 0[ (x; )];i ni (x)

0 1 @r 0 (xi i)
= ni (x) = ni (x) in R2 (3.152)
2 r @xi 2 r2
0 @r 0 (xi i)
= 2
ni (x) = 3
ni (x) in R3 (3.153)
4 r @xi 4 r
This integral equation (3.151) contains unknown boundary terms: the heat ‡ux, where
the temperature is prescribed, and the temperature, where the ‡ux is prescribed. In
order to obtain equations which are only dependent on boundary values and can be used
to determine the unknown boundary reactions, the point has to be shifted from the
interior to the boundary .
Then, the kernel (x; ) and qn (x; ) in the boundary integral becomes weakly and
strongly singular, respectively, when integration points x coincide with , and, hence, it
is necessary to avoid this. For this purpose, in the two-(three-)dimensional case, a small
"-intervall " ahead and behind (circular region of radius " around) is cut out on the
boundary line (surface) , and the integration around is performed along a circle line
(on a spherical surface) " with radius " (with lim " ! 0):
Z Z
1 Wq (x)
( )= [qn (x) (x; ) (x)qn (x; )] d x + (x; )d x (3.154)
0 0
0
"+ "
0
where = " with " = fx 2 : j x j "g.
The integrals with the weakly singular kernel (x; ) can be evaluated as improper
integrals, while those with the strongly singular kernel qn (x; ) can be determined on
" as Cauchy principal values (see, above the respective section), and on " , one has
to consider two integrals
Z Z Z
1 1 1
(x)qn (x; )d x = (x)qn (x; )d x + ( ) qn (x; )d x
0 0 0
"+ " " "
Z
1
+ [ (x) ( )]qn (x; )d x (3.155)
0
"
where the second exists as an improper integral delivering zero since the temperature
…eld is continous while the …rst one can directly be evaluated. One obtains in R2 with
@r=@n = 1 and d x = rd' and in R3 with ni (x)d x = r;' r; d'd (see [1], p.37)
Z
1 1 '2 1 @r '2 '1
qn (x; )d x = s rd' = in R2 (3.156)
0 2 '1 r @n 2
Z" Z Z Z
1 1 (xi i) 1
qn (x; )d x = ni (x)d x = sin 'd'd in R3(3.157)
0 4 r3 4
" "
where, in R2 , '2 '1 means the ’external angle’of the boundary at the point , i.e., the
di¤erence of the outer normal direction at the beginning and at the end of " . Finally,
one obtains from (3.154) the following boundary integral equation:
Z Z
( ) 1 Wq (x)
( )= [qn (x) (x; ) (x)qn (x; )] d x + (x; )d x
2 0 0
0
"
(3.158)
2
where for in R , ( )=2 ('2 '1 ) means the internal angle of at the point ,
i.e., ( ) = for all points besides for corner points while for in R3 , ( ) means
the inner solid angle, i.e., ( ) = 2 for all points besides for points at corners and
edges.
Stationary sound radiation - a Helmholtz equation

Adding to (or subtracting from) the Laplace operator a constant factor 2 gives the
so-called Helmholtz operator, e.g., for the stationary sound radiation problem
2
p(x) + p(x) = b(x) = +ic 0 a(x) (3.159)
where p(x) is the sound pressure distribution when considering time-harmonic processes,
the so-called wave number = !=c with the excitation frequency ! and the sound speed
c in the considered medium (air, water, a.s.o) with the density 0 , and a(x) is the sound
source intensity distribution.
Remark: In general, the Helmholtz equation is the result of a resolution in the Fourier
domain or in the Laplace domain of a transient dynamical problem, or is decribing the
response to steady-state excitations assuming that a permanent regime has been reached.
In any case, the …eld variables are time-harmonic with a …xed angular frequency !, i.e.,
of the form
p(x)e i!t ]
p(x; t) = R[^ (3.160)
where p^(x) is a complex-valued function which encodes amplitude and phase information.
In the sequel, following the traditional convention, the factor ei!t is systematically omitted
and the notation p(x) is used instead of p^(x).
Associated boundary conditions may involve either a …xed given sound pressure (a
so-called Dirichlet boundary condition) at a part of the boundary
p(x) = p(x) for x 2 1 (3.161)

or a prescribed sound ‡ux through a boundary part (a so-called Neumann boundary

condition):
@p(x)
q(x) = = q(x) für x 2 2 (3.162)
@n(x)
where n(x) is the outward normal unit vector, or a certain connexion between both can
be described.
The boundary condition p = 0 models a ’free’surface, e.g., the free surface of a water
domain when gravity waves are neglected, while q = 0 describes the complete re‡exion
of an incoming pressure wave. A boundary condition q(x) = q(x) 6= 0 means a ’sound
production’with a prescribed spatial change of intensity.
A partial re‡exion, i.e., a damped re‡exion of waves may be described by (i2 = 1)
1 1
q(x) = i!Ap(x) = i! p(x) (3.163)
1+ c
where the damping coe¢ cient A depends on the re‡exion coe¢ cient 1 1, who
describes the ratio of re‡ected to incoming pressure wave, i.e., = 0 means no re‡exion
and j j = 1 a complete re‡exion, either symmetric or antimetric corresponding to the
sign of .
The weighed residual of the Helmholtz equation (3.159) is very similar to that of the
Poisson equation (3.147)
2
s p(x) + p(x) + b(x) p (x; )d x =0 (3.164)
The adequate fundamental soution p (x; ) is for a real (see (2.89) and (2.90))
1
p (x; ) = p (r) = e i r in R3 (3.165)
4 r
i (2) 1
p (x; ) = p (r) = H0 ( r) = K0 (i r) in R2 (3.166)
4 2
(2)
where H0 (kr) is a Hankel function of second kind and order zero, while K0 (i r) is a
modi…ed Besselfunction of order zero (Macdonald function).
Remarks: The derivatives of this Besselfunction are obtained by the following rules (r;k =
@r=@xk ):
@K0 (i r) @r
= i K1 (i r) (3.167)
@xk @xk
@ 2 K0 (i r) 2 i
= K0 (i r)r;k r;j + K1 (i r) f2r;k r;j kj g (3.168)
@xk @xj r
such that with (r;1 )2 + (r;2 )2 = 1 and 11 + 22 = 2 the homogeneous Helmholtz equation is
shown to be satis…ed for x 6= :
@ 2 K0 (i r) @ 2 K0 (i r) 2
K0 (i r) = + = K0 (i r)
@x21 @x22
For (3.168), the following recursion formula have been used:
d
2 Kn (x) = Kn 1 (x) + Kn (x)
dx
2n
Kn (x) = Kn 1 (x) Kn+1 (x)
x
The integration by parts of the …rst integral term in (3.164) is formally identic to that
in the case of the Poisson equation in (3.150) and can, therefore, be transfered (only
the constant factor 0 has to be taken as 1). Hence, one obtains the following integral
transformation of (3.164)
Z
2
s p(x) + p(x) + b(x) p (x; )d x = [qn (x)p (x; ) p(x)qn (x; )] d x
Z
2
+ p(x) p (x; ) + p (x; ) d x
Z
+ b(x)p (x; )d x (3.169)
or, since p (x; ) + 2 p (x; ) = (x; ) due to the …ltering e¤ect of the -function
Z Z
p( ) = [qn (x)p (x; ) p(x)qn (x; )] d x + b(x) p (x; )d x (3.170)
where the normal derivative of the fundamental solution qn (x; ) is (see, for R2 , (3.167))
@p (x; ) r;k
qn (x; ) = nk (x) = [1 + i r] e i r in R3 (3.171)
@xk 4 r2
@p (x; ) i
qn (x; ) = nk (x) = K1 (i r)r;k nk (x) in R2 (3.172)
@xk 2
Since the integral equation (3.170) for the sound pressure at interior points 2
contains unknown boundary reaction terms, one needs a boundary integral equation for
their determination. Hence, as for the Poisson equation analysis, has to be shifted on
the boundary , whereby the integral kernels p (x; ) and qn (x; ) become weakly and
strongly singular, respectively, for x ! as well in R3 due to the 1=r and 1=r2 behaviour
as in R2 since K0 (z) ! ln(z) and K1 (z) ! 1=z.
As already explained in detail in the derivation of (3.158), the weakly singular integral
in (3.170) exists as improper integral while the integral with the strongly singular kernel
qn (x; )
s p(x)qn (x; )d x = s p(x)qn (x; )d x + p( ) s qn (x; )d x (3.173)

"+ " " "
exists on " as Cauchy principal value, and can be evaluated on " explicitly giving
the same factors as in the Poisson equation case, e.g.:
i @r i '2 1 '2 '1

s qn (x; )d x = s K1 (i r) d x = s "d' = in R2
" "
2 @n 2 '1 i " 2
Finally, one obtains from (3.170) with b(x) = ic 0 a(x) the boundary integral equation
c( )p( ) = s [qn (x)p (x; ) p(x)qn (x; )] d x s ic 0 a(x) p (x; )d x (3.174)

0
"
with
'( ) '2 '1
c( ) = =1 in R2
2 2
which is, besides the use of di¤erent fundamental solutions, formally almost identic
to(3.158). The main di¤erence to the scalar integral equation for the Poisson equation is
the necessity of calculating with complex numbers.
Linear elastostatics - the Navier equations

When the state variables are vectorial states, boundary value problems are described
by systems of partial di¤erential equations. As a representative example, the Navier
equations describing the deformation displacements u(x) of an linearly elastic body under
body forces b(x) are considered here:
u(x)+( + )rr u(x) = b(x) (3.175)
where the Lamé constants ; are related to the Young’s modulus E and the Poisson’s
E E
ratio by = G = 2(1+ )
and = (1 2 )(1+ )
for three-dimensional and plane strain
E
states and = 1 2 for plane stress states, respectively.
The weighted residual for (3.175) is in indicial notation
Z
@ 2 ui (x) @ 2 uj (x) (k)
+( + ) + bi (x) ui (x; )d x = 0 (3.176)
@xj @xj @xj @xi
(k)
where the adequate fundamental solution, the so-called Kelvin solution ui (x; ) repre-
(k)
senting the response to a unit point force bi (x) = (x )ei applied at a given …xed
point 2 along the k-direction, is given by (note the di¤erent de…nitions of in R2
and in R3 )
(k) 1 1
ui (x; ) = (3 + ) ik ln r + (1 + )r;i r;k in R2 (3.177)
4 2 +
1 1 1
= (3 + ) ik + (1 + )r;i r;k in R3 (3.178)
8 2 + r
For the integration by parts, it is helpful to substitute the Navier equations, which are the
displacement representation of the interior equilibrium to the body forces, by its original
stress-based form
@ 2 ui (x) @ 2 uj (x) @ ik (x)
+( + ) = = bi (x) (3.179)
@xj @xj @xj @xi @xk
since, then, it is easy to perform the …rst integration by parts:
Z Z h i
@ ij (x) (k) (k) (k)
ui (x; )d x = [ ij (x)ui (x; )];j ij (x)[u i (x; )];j d x
@xj
Z Z (k)
(k) @ui (x; )
= ij (x)ui (x; )nj (x)d x ij (x) d x
@xj
Z Z (k)
(k) @ui (x; )
= Ti (x)ui (x; )d x ij (x) d x (3.180)
@xj
where Ti = ij nj is the so-called traction vector.

From the de…nition of the strain tensor "ij = 0:5(ui;j + uj;i ), one obtains due to
the symmetry of the stress tensor ij , the symmetry of the elastic constitutive law for
isotropic material ij = 2 "ij + ij "ll , and the reciprocity with its inverse form "ij = ( ij
ij ll =(2 + 3 ))=2 and "ij = ( ij ij ll =(2 + 2 ))=2 in 3D and 2D, respectively,
i.e., ij "ij = "ij ij
(k)
!
(k) (k)
@ui (x; ) 1 @ui (x; ) @uj (x; ) (k)
ij (x) = ij (x) + = ij (x)"ij (x; )
@xj 2 @xj @xi
(k) @ui (x) (k)
= "ij (x) ij (x; )= ij (x; ) (3.181)
@xj
which allows the second integration by parts of the remaining domain integral
Z Z h i
@ui (x) (k) (k) (k)
(x; ) d x = [u i (x) (x; )];j u i (x)[ (x; )];j d x
@xj ij ij ij
Z Z
(k) (k)
= ui (x) ij (x; )nj (x)d x ui (x)[ ij (x; )];j d x
Z Z
(k) (k)
= ui (x)Ti (x; )d x + ui (x) (x )ei d x (3.182)
Taking (3.181) into account when substituting (3.182) in (3.180) and in the equivalent
weighted residual form (3.176), respectively, yields with the …ltering e¤ect of the -function
the integral equation for the displacements at arbitrary interior points 2 :
Z h i Z
(k) (k) (k)
uk ( ) = Ti (x)ui (x; ) ui (x)Ti (x; ) d x + bi (x) ui (x; )d x (3.183)
(k)
The boundary traction vector Ti (x; ) of the fundamental solution may be determined
by di¤erentiating the fundamental solution (3.177) and (3.178), respectively, via the def-
inition of the strain tensor and the constitutive relations as:
(k) 1 1 nj
Ti = ( (r;k ij r;i jk r;j ik ) 2( + )r;i r;j r;k ) in R2 (3.184)
2 2 + r
1 1 nj
= ( (r;k ij r;i jk r;j ik ) 3( + )r;i r;j r;k ) in R3 (3.185)
4 2 + r2
Exercise 12: Strain tensor of the elastostatic fundamental solution

(k) (k)
Determine the strain tensor "ij (x; ) of the fundamental solution ui (x; ) (3.178) in R3 .
Since the displacement integral equations (3.183) contain unknown boundary reactions
(at each boundary point two and three in R2 and R3 , respectively), the arbitrary source
point is placed on the boundary to obtain equations which connects, besides the
known body force domain integral, only boundary terms with each other. Then, the
(k) (k)
integral kernels ui (x; ) and Ti (x; ) become weakly and stronly singular for x = ,
respectively, and, similarly, but not identically (see [1], page 71) to the above considered
scalar problems, the weakly singular integrals exists as improper integrals, while the those
(k)
with the strongly singular kernel Ti (x; ) produces two contributions: the …rst are their
Cauchy principal values on " (the remaining part of the boundary contour (surface)
from which the part " was cut out by a circle (sphere) of some radius ", centered at
(k)
) and the second is an integral over the singular traction Ti (x; ) itself on " , i.e., on
the outer (outside the domain ) part of the contour (surface) of that circle (sphere) with
radius " which was used for the Cauchy principal value cut-out:
(k) (k) (k)

s ui (x)Ti (x; )d x = s ui (x)Ti (x; )d x + ui ( ) s Ti (x; )d x (3.186)
"+ " " "
Since on " the components ni of the outward normal unit vector and those of the deriv-
ative of r are identic, i.e., ni = r;i and, therefore, @r=@n = 1, one obtains in R2 where
(r;1 ; r;2 ) = (cos '; sin ') and d x = rd'
(k) 11 1
cki ( ) = s Ti (x; )d x = s [ ik + 2( + )r;i r;k ] d x
"
2 2 + "
r
'2
11 + 2( + ) cos2 ' 2( + ) sin ' cos '
= s d'
2 2 + '1 2( + ) sin ' cos ' + 2( + ) sin2 '
8 9
1<
('2 '1 ) ki =
= + 1 sin 2'2 sin 2'1 cos 2'2 + cos 2'1 (3.187)
2 : +2 + 2 ;
cos 2'2 + cos 2'1 sin 2'2 + sin 2'1
while in R3 with (r;1 ; r;2 ; r;3 ) = (cos ' sin #; sin ' sin #; cos #) and d x = r2 sin #d#d'
(k) 1 1 1
cki ( ) = s Ti (x; )d x = s [ ik + 3( + )r;i r;k ] d x
"
2 2 + "
r2
Z'2 Z#2
11
= [ ik + 3( + )r;i r;k ] sin #d#d' (3.188)
2 2 +
' 1 #1
In general, the actual values of these factors dependent on the shape of the actually
considered boundary. But, when the position of is on a smooth boundary region, i.e.,
'2 '1 = in R2 and '2 '1 = 2 with #2 #1 = 2 in R3 , both above integrals (3.187)
and (3.188) become simply
1
cki ( ) = ki (3.189)
2
Finally, the boundary integral equations for determining unknown boundary reactions
read as (k = 1; 2 in R2 and k = 1; 2; 3 in R3 )
h i
(k) (k) (k)
[ ki + cki ( )]ui ( ) = s Ti (x)ui (x; ) ui (x)Ti (x; ) d x + s bi (x) ui (x; )d x
0
"
(3.190)
3.2.2 Indirect integral formulations: the singularity method

As already explain for one-dimensional problems, the basic idea of this indirect method is
to introduce intermediary unknowns - unknown intensities of certain singularity
layers. For these singularities, so-called in‡uence functions, i.e., the complete
response to the action of a singularity (e.g., a unit point force) must be known
everywhere in the considered material. Then, these singularity layers are dis-
tributed on a ’…ctitious’boundary + in a certain small distance d" from the
real boundary outside the domain and their intensities have to be deter-
mined such that the integrated response is equal to the prescribed boundary
values on the real boundary .
In a second step, when these intensities of the singularities are determined,
the unknown physical boundary reaction as well as corresponding interior
states can easily be found by the same integral equation used in the …rst step
to …nd the singularity intensities.
Sound pressure and sound ‡ux in stationary acoustics (Helmholtz equation)

As known, the stationary sound radiation problem is described by the Helmholtz equation
(see (3.159))
p(x) + 2 p(x) = b(x) = +ic 0 a(x) (3.191)
for the sound pressure distribution p(x) and the fundamental solution solving the equation
2
p (x; ) + p (x; ) = (x; )
is in R2 (siehe (3.166))
1
p (x; ) =
K0 (i r) (3.192)
2
Hence, the in‡uence function decribing the sound pressure at a point x due to a unit point
sound source a(x) = (x; ) at point is found by comparison to be
1
(pa) (x; ) = ic 0 K0 (i r) (3.193)
2
Since the sound ‡ux qn (x) in the direction of the normal vector nk (x) is de…ned as the
normal derivative of the sound pressure, the corresponding in‡uence function for the sound
‡ux is determined by (r = jx j)
@ (pa) (x; ) 2 1 @r
(qn a) (x; ) = nk (x) = c 0 K1 (i r) nk (x) (3.194)
@xk 2 @xk
Following the above described idea of the indirect integral equation method, at points
on a …ctitious boundary + , which is either enclosing with a certain distance d" the
real boundary or is coincident to , layers of point sources are introduced with such
an intensity a ( ) that the prescribed boundary conditions on the real boundary are
satis…ed.
For a sound radiation problem where on one part of the boundary 1 the sound pressure
p(x) = p(x) and on the remaining part 2 the sound ‡ux qn (x) = qn (x) is prescribed, and,
moreover, a sound source density a(x) is acting in the interior of the considered domain
, the two indirect integral equations for determining the unknown point source layer
intensity a ( ) are simply given by
s (pa) (x; )a ( )d + s (pa) (x; )a( )d = p(x) for x 2 1

+
s (qn a) (x; )a ( )d + s (qn a) (x; )a( )d = qn (x) for x 2 2

+
which is explicitly
2 i p(x)
s K0 (i r)a ( )d + s K0 (i r)a( )d = for x 2(3.195)
1
+ c 0
@r @r 2 i qn (x)
nk (x) s i K1 (i r) a ( )d + s i K1 (i r) a( )d = for x
(3.196)
2 2
+ @xk @xk c 0
Since the size of the distance d" of the …ctitious boundary + from the real boundary
has a large e¤ect on the solution if these integral equations have to be solved numerically,
it is mostly better to transfer the …ctitious boundary into the real boundary, i.e., d" ! 0.
In this case, the in‡uence function (pb) (x; ) becomes with K0 (i r) ! ln(i r) weakly
singular and (qn b) (x; ) with K1 (i r) ! 1=(i r) strongly singular for ! x.
Hence, similarly to direct integral equation method, the singular point x = has to
be avoided in the integration on + ! , but di¤erent to the handling there, the integral
on + is only to split into one on + + +

" and one on " (while the integration on the
+
"-circular arc " is dropped here), and, for both contributions, the limit towards "
and " is performed.
In the case of the weakly singular integral equation (3.195) for the sound pressure,
the integral on + " ! " gives no contribution such that the equation remains formally
unchanged:
2 i p(x)
s K0 (i r)a ( )d + s K0 (i r)a( )d = for x 2 1 (3.197)
" c 0
The integral with the strongly singular kernel in the equation (3.196) for the sound ‡ux
@r @r
nk (x) s i K1 (i r) a ( )d = nk (x) s i K1 (i r) a ( )d
+ @xk + +
"
@xk
@r
+nk (x)a (x) s i K1 (i r) d
+
"
@xk
@r
+nk (x) s [a ( ) a (x)]i K1 (i r) d (3.198)
+
"
@xk
is split in three parts where that on + +
" ! " exists as Cauchy principal value,
that on " can be analytically integrated while the third one on +
+
" ! " gives zero
assuming at x = a continous intensity a ( ).
Since here, di¤erent to the direct integral equations, the normal vector nk (x) ,
(cos '; sin ') is not that of the integration point , and is only on a smooth bound-
ary uniquely de…ned, for the evaluation on + " , a smooth boundary
+
is assumed which
is in a distance d" parallel to the real boundary . Denoting the distance of the projection
of x on p+ from with s, one can express the distance r between x 2 and 2 +
by r = s2 + d2" and r;i = r;n ni + r;s ti can be described by (' is the angle between the
normal direction and the x1 -axis)
r;1 1 x1 1 1 d" cos ' s sin ' d" n1 (x)t1 (x) s
= = = +
r;2 r x2 2 r d" sin ' + s cos ' r n2 (x)t2 (x) r
(3.199)
Applying this and the approximation of K1 (i r) 1=(i r) for small distances r, gives,
since nk (x)nk (x) = 1 and nk (x)tk (x) = 0
@r " 1 d"
r k
n (x)
nk (x)a (x) lim s i K1 (i r) d = a (x)nk (x) lim s i s ds
d" !0 +
"
@xk d" !0 " i r + r tk (x)
d" "
= a (x) lim s ds
d" !0 " + d2" s2
"
s
= a (x) lim arctan
d" !0 d" "
" "
= a (x) lim arctan arctan
d" !0 d" d"
= a (x) (3.200)
Finally, the following singular version of (3.196) for satisfying prescribed sound ‡ux con-
ditions is obtained
a (x) nk (x) @r iqn (x)
@r
s i K1 (i r) a ( )d + s i K1 (i r) for x 2 2
a( )d =
2 2 " @xk @xk
c 0
(3.201)
When by solving the two equations (3.197) and (3.201) the adequate intensities a ( ) are
found to represent the prescribed boundary values, the same two equations can be used to
…nd the unknown boundary reactions p(x) on 2 and qn (x) on 1 , and the two equations
(3.195) and (3.196) (due to + ! by integration on ) for …nding the sound pressure
and the sound ‡ux at any interior point x.
Displacements and stresses in elastostatics (Navier equations)

The Navier equations (3.175)
@ 2 ui (x) @ 2 uj (x)
+( + ) = bi (x) (3.202)
@xj @xj @xj @xi
describe the displacements of an elastic body under its dead weight bi (x) and as a result of
prescribed displacements ui (x) =ui (x) on a part of the boundary 1 and/or of prescribed
boundary tractions Ti (x) = Ti (x) on the remaining boundary part 2 .
Besides these, other dynamic or geometric states, e.g., point forces, single moments,
or dislocations, may occur and produce displacements and deformations.
(k)
The solutions ui (x; ) of the equations
(k) (k)
@ 2 ui (x; ) @ 2 uj (x; ) (k)
+( + ) = bi (x; ) = (x )ei (3.203)
@xj @xj @xj @xi
i.e., the displacements (see (3.177))
(k) 1 1
ui (x; ) = (3 + ) ik ln r + (1 + )r;i r;k in R2 (3.204)
4 2 +
(k)
are the reactions at a point x on a unit point force bi (x; ) = Fi (x; ) = (x )ei
applied at a given …xed point 2 along the k-direction. Hence, this fundamental
solution is also a in‡uence function representing displacements due to unit point forces
(k)
(uF )i:k (x; ) = ui (x; ) (3.205)
Then, following the idea of the indirect integral equation method, point force layers with
unknown intensity Fk ( ) along the k-direction at points on a …ctitious boundary +
enclosing with a distance d" the real boundary are introduced which produce together
with the known prescribed dead weight loading bk ( ), 2 , displacements at points x
ui (x) = s (uF )i:k (x; )Fk ( )d + s (uF )i:k (x; )bk ( )d (3.206)
+
For representing prescribed displacements ui (x) on a part 1 of the real boundary , the
intensities Fk ( ) have to satisfy the indirect boundary integral equation:
s (uF )i:k (x; )Fk ( )d + s (uF )i:k (x; )bk ( )d = ui (x) for x 2 1 (3.207)
+
When on the remaining part 2 = 1 boundary tractions Ti (x) are prescribed, in‡u-
ence functions representing boundary tractions are needed which can be determined by
di¤erentiating the displacement in‡uence function (3.205), respectively (3.204), via the
de…nition of the strain tensor and the constitutive relations (see (3.184))
(k) (k) (k)

(T F )i:k (x; ) = ij (x; )nj (x) = 2 "ij (x; ) + ij "ll (x; ) nj (x)
1 1 r;k ij r;i jk r;j ik
= nj (x) (3.208)
2 2 + r 2(1 + )r;i r;j r;k
This gives the indirect boundary integral equations for representing boundary tractions
as
s (T F )i:k (x; )Fk ( )d + s (T F )i:k (x; )bk ( )d = Ti (x) for x 2 2 (3.209)

+
In this regular version of the indirect boundary equations (3.207) und (3.209) for a
mixed elastostatic boundary value problem, the distance d" between the …ctitious bound-
ary + and the real boundary has a large e¤ect on a numerically determined solution.
Therefore, it is advantageous to shift with d" ! 0 the …ctitious boundary + towards
the real boundary . Then, the integral kernels (uF )i:k (x; ) and (T F )i:k (x; ) become
for ! x weakly singular and strongly singular, as ln r and 1=r, respectively.
Hence, similarly to direct integral equation method, the singular point x = has to
be avoided in the integration on + ! , but di¤erent to the handling there, the integral
on + is only to split into one on + + +
" and one on " (while the integration on a
+
"-circular arc " , equivalent to " in the diect method, is dropped here), and, for both
contributions, the limit towards " and " is performed.
In the case of the weakly singular integral equation (3.207) for the displacements,
the integral on + " ! " gives no contribution such that the equation remains formally
unchanged:
s (uF )i:k (x; )Fk ( )d + s (uF )i:k (x; )bk ( )d = ui (x) for x 2 1 (3.210)
"
The integral with the strongly singular kernel in the equation (3.209) for the boundary
tractions
nj (x) 1 r;k ij r;i jk r;j ik

s (T F )i:k (x; )Fk ( )d = s (3.211)
Fk ( )d
+ 2 2 + +r 2(1 + )r;i r;j r;k
= s (T F )i:k (x; )Fk ( )d
+ +
"
+Fk (x) s (T F )i:k (x; )d

+
"
+ s [Fk ( ) Fk (x)] (T F )i:k (x; )d (3.212)

+
"
is split in three parts where that on + +

" ! " exists as Cauchy principal value,
that on " can be analytically integrated while the third one on +
+
" ! " gives zero
assuming at x = a continous intensity Fk ( ).
Since here, di¤erent to the direct integral equations, the normal vector nk (x) ,
(cos '; sin ') is not that of the integration point , and is only on a smooth bound-
ary uniquely de…ned, for the evaluation on + " , a smooth boundary
+
is assumed which
is in a distance d" parallel to the real boundary . Denoting the distance of the projection
of x pon + from with s, one can express the distance r between x 2 and 2 + by
r = s2 + d2" and r;i = r;n ni + r;s ti can be described by (see (3.199))
d" s
r;i = ni (x) ti (x) (3.213)
r r
Following the evaluation in (3.200), one obtains for the essential singular part
1 @r " 1 d" s
lim s d = lim s nk (x) + tk (x) ds = nk (x) (3.214)
d" !0 + r @xk d" !0 " r r r
"
and with limd" !0 for
1 @r @r @r " 1 d" d" d"

r i
n (x) r j
n (x) r k
n (x)
s d = s ds
+ r @xi @xj @xk
"
" r + rs ti (x) + rs tj (x) + rs tk (x)
" d3" ni nj nk + d2" s (ni nj tk + ni tj nk + ti nj nk )
1
= s ds
" r4 +d" s2 (ni tj tk + ti nj tk + ti tj nk ) + s3 ti tj tk
8 h i" 9
>
> sd" 1
+ 2 arctan d" s
ni nj nk >
>
>
> 2(d2" + 2 ) >
>
>
> h i "
" >
>
>
< + s2 +d2 d2
"
(ni nj tk + ni tj nk + ti nj nk ) >
=
= h " " i"
>
> + 2(s2sd "
+ 12 arctan ds" (ni tj tk + ti nj tk + ti tj nk ) >
>
>
> +d2" )
h i" >
>
>
>
" >
>
>
:
2
d" 1
+ s2 +d2 + 2 ln(s + d" ) 2 2
ti tj tk >
;
" "
= fni nj nk + ni tj tk + ti nj tk + ti tj nk g for d" ! 0 (3.215)

2
This yields with nj nj = 1, tj nj = 0, and ni nk + ti tk = ik for
1 (nk ni ni nk nj ik nj )
lim s (T F )i:k (x; )d =
d" !0 + 2 2 + +2(1 + ) 2 (ni nj nk + ni tj tk + ti nj tk + ti tj nk )nj
"
1 1
= ik + (1 + )(ni nk + ti tk ) = ik (3.216)
22 + 2
Finally, the singular version of the indirect boundary integral equation for the boundary
tractions is obtained to be
1
F (x) + s (T F )i:k (x; )Fk ( )d + s (T F )i:k (x; )bk ( )d = Ti (x) for x 2 2
2 i "
(3.217)
Are the singularity layer intensities Fk ( ) determined by solving the integral equations
(3.210) and (3.217), in a second step, the unknown boundary reactions, i.e., ui (x) for
x 2 2 and Ti (x) for x 2 1 can easily found be evaluating the integral equations (3.210)
and (3.217), respectively.
Then, it is also possible to analyse the stresses at interior points by evaluating with
the determined intensities Fk ( ) the integral relation
ij (x) = s ( F )ij:k (x; )Fk ( )d + s ( F )ij:k (x; )bk ( )d for x 2 (3.218)
where the in‡uence function ( F )ij:k (x; ) for the stresses is easily found from that of the
boundary tractions (3.208) to be
1 1
( F )ij:k (x; ) = r;k ij r;i jk r;j ik 2(1 + )r;i r;j r;k : (3.219)
2 2 + r
3.2.3 Integral formulation with Green’s functions

As de…ned above, the so-called Green’s function G (x; ) of a boundary value problem is
a special fundamental solution, i.e.,
L(D)G (x; ) = (x; )
which satis…es homogeneous conditions for those boundary states which are prescribed in
the actual problem, i.e.,
E(G (x; )) = 0 for x 2 1

N (G (x; )) = 0 for x 2 2
The meaning of this de…nition shall now be demonstrated for some explicit problems.
Temperature distribution in stationary heat conduction

As described above in equation (3.158), the direct form of the integral equation for the
temperature ( ) on the boundary, i.e., for 2 ; reads
Z Z
( ) 1 Wq (x)
( )= [qn (x) (x; ) (x)qn (x; )] d x (x; )d x
2 0 0
"
(3.220)
while it is given in the interior, i.e., at points 2 (see equation (3.151)) by
Z Z
1 Wq (x)
( )= (qn (x) (x; ) (x)qn (x; )) d x (x; )d x (3.221)
0 0
When the actual boundary value problem is de…ned by a prescribed temperature (x) = (x)
on a part of the boundary 1 and/or of prescribed temperature ‡ux qn (x) = qn (x) on the
remaining boundary part 2 , the more detailed form of (3.226)
Z Z
1 1
( ) = qn (x) (x; ) (x)qn (x; ) d x [qn (x) (x; ) (x)qn (x; )] d x
0 0
Z1 2
Wq (x)
(x; )d x (3.222)
0
shows explicitly the unknown boundary reactions qn (x) on 1 and (x) on 2 . Hence,
it is obvious that the temperature ( ) could directly be determined by this integral if
the respective integrands containing unknowns would be zero, i.e., if one …nds a special
fundamental solution G(x; ) which satis…es additionally the conditions
G(x; ) = 0 for x 2 1 and qn (G(x; )) = 0 for x 2 2 (3.223)
Then, the temperature at any interior point 2 is expressed by

Z Z Z
1 1 Wq (x)
( )= (x)qn (G(x; ))d x qn (x)G(x; )d x G(x; )d x (3.224)
0 0 0
1 2
without determining …rst the unknown boundary reactions by solving the boundary inte-
gral equation (3.225).
Sound pressure in stationary acoustics

As described above in equation (3.174), the direct form of the integral equation for the
sound pressure p( ) on the boundary, i.e., for 2 ; reads
Z
'( )
p( ) = [qn (x)p (x; ) p(x)qn (x; )] d x + s b(x) p (x; )d x (3.225)
2
"
while it is given in the interior, i.e., at points 2 (see equation (3.170)) by

Z Z
p( ) = [qn (x)p (x; ) p(x)qn (x; )] d x+ b(x) p (x; )d x (3.226)
When the actual boundary value problem is de…ned by a prescribed pressure p(x) =p(x)
on a part of the boundary 1 and/or of prescribed sound ‡ux qn (x) = qn (x) on the
remaining boundary part 2 , the more detailed form of (3.226)
Z Z
p( ) = [qn (x)p (x; ) p(x)qn (x; )] d x + [qn (x)p (x; ) p(x)qn (x; )] d x
1 2
Z
+ b(x) p (x; )d x (3.227)
shows explicitly the unknown boundary reactions qn (x) on 1 and p(x) on 2 . Hence, it
is obvious that the sound pressure p( ) could directly be determined by this integral if
the respective integrands containing unknowns would be zero, i.e., if one …nds a special
fundamental solution G(x; ) which satis…es additionally the conditions
G(x; ) = 0 for x 2 1 and qn (G(x; )) = 0 for x 2 2 (3.228)
Then, the sound pressure at any interior point 2 is expressed by
p( ) = s p(x)qn (G(x; ))d x + s qn (x)G(x; )d x + s b(x) G(x; )d x (3.229)
1 2
without determining …rst the unknown boundary reactions by solving the boundary inte-
gral equation (3.225).
Displacements in elastic bodies

As described above in equation (3.190), the direct form of the integral equation for dis-
placements ui ( ) on the boundary, i.e., for 2 ; reads
h i
(k) (k) (k)
[ ki + cki ( )]ui ( ) = s Ti (x)ui (x; ) ui (x)Ti (x; ) d x + s bi (x) ui (x; )d x
0
"
(3.230)
while for interior displacements ui ( ) with 2 holds (see equation (3.183))
h i
(k) (k) (k)
uk ( ) =s Ti (x)ui (x; ) ui (x)Ti (x; ) d x + s bi (x) ui (x; )d x (3.231)
When the boundary value problem is de…ned by prescribed displacements ui (x) =ui (x)
on a part of the boundary 1 and/or of prescribed boundary tractions Ti (x) = Ti (x) on
the remaining boundary part 2 , the more detailed form of (3.183)
h i h i
(k) (k) (k) (k)
uk ( ) = s Ti (x)ui (x; ) ui (x)Ti (x; ) d x + s Ti (x)ui (x; ) ui (x)Ti (x; ) d x
1 2
(k)
+s bi (x) ui (x; )d x (3.232)
shows the unknown boundary reactions Ti (x) on 1 and ui (x) on 2 . Hence, it is obvious
that the displacements uk ( ) could directly be determined by these integrals if the respec-
tive integrands containing unknowns would be zero, i.e., if one …nds a special fundamental
(k)
solution Gi (x; ) which satis…es additionally the conditions
(k) (k)
Gi (x; ) = 0 for x 2 1 and Ti (Gj (x; )) = 0 for x 2 2 (3.233)
Then, the displacement at any position 2 is determined by

(k) (k)
uk ( ) = s ui (x)Ti (Gj (x; ))d x + s Ti (x)Gi (x; )d x
1 2
(k)
+s bi (x) Gi (x; )d x (3.234)
4 Numerical solution of boundary integral
equations: The boundary element method
In the case of two- and three-dimensional problems, in general, boundary integral equa-
tions can only approximatively be solved, i.e., approximations have to be introduced for
the boundary and/or the state functions, integrations are performed numerically and not
analytically, the integral equations are satis…ed only pointwise (i.e., point collocation ),
and, …nally, the resulting system of algebraic equations is solved also only numerically
either by a direct solver (Gauss elimination) or iteratively.
These steps are explained in the following sections.
4.1 Approximation of the boundary and of boundary

states
The …rst discretization step is the partition of the boundary curve in non-intersecting
so-called boundary elements e (e = 1; ::; m) by introducing on the boundary so-called
node points xle (l = 1; ::; n) where the nodes on the interface between two neighbour
elements e and e+1 have the same coordinates but di¤erent node indices, e.g., xne =
x1e+1 in R2 .
Inside these elements, both the boundary curve (if it is not straight) and the state
functions of the boundary value problem have to be approximated usually by polynoms
as exact as wanted or needed. The polynomial order for approximating the geometry and
the boundary states may be di¤erent; if these orders are chosen to be equal, one uses a
so-called isoparametric concept.
Remark: Of course, exactly the same considerations led in past to the concepts and tech-
niques now commonly used in …nite element methods. The usual concept of boundary elements
is a mere transposition of that of a …nite element and, hence, based on the use of nodes and
shape functions.
4.1.1 On boundary curves in R2

For the geometric approximation of the boundary, usually a mapping of each physical
boundary element e onto a parent element e in a parameter space is introduced where
the parent element assumes a simple shape, i.e., a line segment 2 [0; 1] or 2 [ 1; 1]:
X
n
2 e ! xe ( ) = xle Nln ( ) with 0 1 (4.1)
l=1
80
4.1 Approximation of the boundary and of boundary states 81
where the index l (1 l n) de…nes the local numbering of the nodes on element e.
The n shape functions, usually of polynomial type and at least linear, since the boundary
approximation should be continous, are subjected to the following restrictions:
X
n
Npn ( q ) = pq and Nln ( ) = 1 8 2 e (4.2)
l=1
where q 2 e is the antecedent of the physical node xqe . Of course, the speci…c choice
of shape functions and the number n of nodes de…ning the element are related, e.g.,
for a linear approximation is n = 2 with initial and end node x1e and x2e , respectively,
the shape functions are de…ned as
N12 ( ) = 1 ; N22 ( ) = for 2 e = [0; 1] (4.3)
and the boundary element is approximated by
xe ( ) = x1e (1 ) + x2e (4.4)
for a quadratic approximation is n = 3 with initial, middle, and end node x1e 2e
i , xi
3e
and xi , respectively, the shape functions are
N13 ( ) = (1 )(1 2 );
3
N2 ( ) = 4 (1 );
3
N3 ( ) = (2 1) for 2 e = [0; 1] (4.5)
and the boundary element is approximated by
xe ( ) = x1e (1 )(1 2 ) + x2e 4 (1 ) + x3e (2 1) (4.6)
Remark: It should be mentioned that on a straight boundary element when x2e i is taken
to be the central node, i.e., is taken as the arithmetic mean of the initial and the end
node coordinates, x2e = 0:5(x1e + x3e ), the quadratic approximation (4.6) is reduced to
xe ( ) = x1e (1 )(1 2 ) + 0:5(x1e + x3e )4 (1 ) + x3e (2 1)

= x1e (1 ) + x3e
i.e., to the linear approximation (4.4)
Exercises 13: Shape functions with the local coordinate 1 1

Determine the linear and the quadratic shape functions (corresponding to (4.3) and (4.5),
respectively) when the local coordinate is de…ned in the range 1 1.
In the same way, the boundary states of the problem can be approximated by
shape functions in the local coordinate , but, here, it is not necessary to guarantee a
continuous approximation across the elements; sometimes it is even necessary to simulate
82 4 Numerical solution of boundary integral equations: The boundary element method
discontinuities, e.g., when one element is loaded by constant tractions and the next is
unloaded. Hence, besides the above introduced linear and quadratic shape functions (4.3)
and (4.5), respectively, a single middle node x1e
i and the constant shape function (n = 1)
N11 ( ) = 1 (4.7)
can be used to approximate a boundary state (x).

This means that a boundary state (xi ) may be approximated in a boundary element
e as
X
n
(xei ( )) = e ( ) = le n
Nl ( ) with 0 1: (4.8)
l=1
where n indicates the number of applied nodes per element.
4.1.2 On boundary surfaces in R3

Again, the …rst step is to divide the boundary surface into m, in general, curved either
quadrilateral or triangular surface elements e by introducing on the boundary surface
node points xle (l = 1; :::; n) where l de…nes the local numbering of the nodes on the
element. The position of a point on e is expressed by
X
n
xe ( 1 ; 2) = xle Nln ( 1 ; 2) (4.9)
l=1
with the given nodal position vectors xle multiplied by appropriate shape functions Nln ( 1 ; 2 )
with local coordinates ( 1 ; 2 ) lying in the range ( 1; 1) or (0; 1). Every shape function
has unit value at its associated node and zero value at all other nodes.
Remark: These shape functions have been developed in the Finite Element Method
and taken over into the Boundary Element Method.
The parameters ( 1 ; 2 ) de…ne a plane and the curved element is thus mapped, for
quadrilateral elements onto a square in this plane. An element represented linearly in each
of the local coordinates 1 and 2 is speci…ed by four given nodal values xle (l = 1; :::; 4)
each of which has an associated shape function Nl4 ( 1 ; 2 ). The following Table shows the
shape functions for the four node so-called serendipity element in the range ( 1; 1):
xle ( 1; 2) Nl4 ( 1 ; 2 )
x1e (1; 1) (1 + 1 )(1 + 2 )=4
x2e ( 1; 1) (1 1 )(1 + 2 )=4 (4.10)
x3e ( 1; 1) (1 1 )(1 2 )=4
x4e (1; 1) (1 + 1 )(1 2 )=4
It can easily be seen that, e.g., N14 ( 1 ; 2 ) is 1 when 1 = 1 and 2 = 1, that is at node
x1e , while N14 ( 1 ; 2 ) = 0 when either 1 = 1 or 2 = 1, and hence it is zero at nodes
x2e , x3e , and x4e .
4.1 Approximation of the boundary and of boundary states 83
When the variation with respect to each of the local coordinates shall be quadratic,
one needs eight nodes, i.e.,
X
8
e
x ( 1; 2) = xle Nl8 ( 1 ; 2) (4.11)
l=1
and the following shape functions in the range ( 1; 1)
xle ( 1; 2) Nl8 ( 1 ; 2 )
x1e (1; 1) (1 + 1 )(1 + 2 )( 1 + 21)=4
x2e ( 1; 1) (1 1 )(1 + 2 )( 1 2 1)=4
x3e ( 1; 1) (1 1 )(1 2 )( 1 + 2 + 1)=4
x4e (1; 1) (1 + 1 )(1 2 )( 1 2 + 1)=4 (4.12)
x5e (1; 0) (1 + 1 )(1 2
2 )=2
x6e (0; 1) (1 2
1 )(1 + 2 )=2
x7e ( 1; 0) (1 1 )(1
2
2 )=2
x8e (0; 1) (1 2
1 )(1 2 )=2
Checks con…rm the wanted shape function properties, that is of having unit value at their
’own’node and zero at other nodes. e.g., the shape function N38 ( 1 ; 2 ) associated with
the node x3e gives at ( 1; 1)
N38 ( 1; 1) = (1 + 1)(1 + 1)( 1 1 + 1)=4 = 1
and evaluated at (1; 1)
N38 (1; 1) = (1 1)(1 1)(1 + 1 + 1)=4 = 0
Similarly, the values of N38 at the remaining nodes are all zero.
An approximation using triangular elements and being linear in each of the local
coordinates 1 and 2 needs only three nodes, the three corner nodes of the triangle
X
3
e
x ( 1; 2) = xle Nl3 ( 1 ; 2) (4.13)
l=1
where the shape functions associated with these corner nodes are dependent on the range
( 1; +1) or (0; +1) of the local coordinates ( 1 ; 2 ) de…ned as
xle ( 1; 2) Nl3 ( 1 ; 2 ) xle ( 1 ; 2 ) Nl3 ( 1 ; 2)

x1e (1; 1) (1 + 1 )=2 x1e (1; 0) 1
or 2e (4.14)
x2e ( 1; 1) (1 + 2 )=2 x (0; 1) 2
x3e ( 1; 1) ( 1 + 2 )=2 x3e (0; 0) 1 1 2
Obviously, this simple three node triangular element produces planar approximations.
For curved surfaces, six node triangular elements with the associated quadratic shape
functions Nl6 ( 1 ; 2 ), (l = 1; 2; :::; 6) on the range ( 1; +1)
xle ( 1; 2) Nl6 ( 1 ; 2 )
x1e (1; 1) 1 (1 + 1 )=2
x2e ( 1; 1) 2 (1 + 2 )=2
x3e ( 1; 1) ( 1 + 2 )( 1 + 2 + 1)=2 (4.15)
x4e (0; 0) (1 + 1 )(1 + 2 )
x5e ( 1; 0) (1 + 2 )( 1 + 2 )
x6e (0; 1) ( 1 + 1)( 1 + 2 )
are more appropriate.

In the same way as the element surface, the boundary states of the problem can be
approximated by shape functions in the local coordinates 1 and 2 , but, here, it is not
necessary to guarantee a continuous approximation across the elements; sometimes it is
even necessary to simulate discontinuities, e.g., when one element is loaded by constant
tractions and the next is unloaded. Hence, besides the above introduced bilinear and
biquadratic shape functions (4.10) and (4.12), respectively, for quadrilateral elements and
(4.14) and (4.15), respectively, for triangular elements, a single middle node x1ei and the
constant shape function (n = 1)
N11 ( 1 ; 2 ) = 1 (4.16)
can be used to approximate a boundary state (x).
Remark: It should be mentioned that, as in the Finite Element Method, also higher
order discontinuous elements are possible, e.g., a four-node discontinuous quadrilateral
element with the shape functions (for more details, see [9])
xle ( 1; 2) Nl4 ( 1 ; 2 )
x1e ( 21 ; 12 ) (1 + 2 1 )(1 + 2 2 )=4
x2e ( 21 ; 12 ) (1 2 1 )(1 + 2 2 )=4 (4.17)
x3e ( 21 ; 12 ) (1 2 1 )(1 2 2 )=4
x4e ( 12 ; 12 ) (1 + 2 1 )(1 2 2 )=4
e
Hence, a boundary state (xi ) may be approximated in a boundary element as
X
n
(xei ( 1 ; 2 )) = e
( 1; 2) = le
Nln ( 1 ; 2) (4.18)
l=1
where n indicates the number of applied nodes.
4.2 Integration over boundary elements

After the …rst step, the discretization of the boundary in boundary elements, the integrals
over the whole boundary are decomposed into a sum of m integrals over single elements
4.2 Integration over boundary elements 85
e
. As a consequence of the next steps, the elementwise approximation of the boundary
geometry and of the boundary states, the integration over the boundary elements has to
take into account the actually applied approximation schemes. This will be described in
the following sections.
4.2.1 Elements on boundary curves

When integrating along the boundary curve , in each boundary element e , the mapping
between the global physical coordinates x and the local coordinate has to be taken into
account by s
dxe dxe
d x = ( 1 )2 + ( 2 )2 d = j J( ) j d (4.19)
d d
where j J( ) j is the so-called Jacobian. For a straight boundary element with linear
approximation is
dxei
= x1e 2e 2e
i ( 1) + xi (+1) = xi x1e
i (4.20)
d
which results with the element length le
q
j J( ) j = (x2e 1 x1e 2 2e
1 ) + (x2 x1e 2
2 ) = le
so that in this case

d x = le d (4.21)
For a parabolic boundary element with quadratic boundary approximation (4.5) is
dxei
= x1e 2e
i ( 3 + 4 ) + xi (4 8 ) + x3e
i ( 1+4 ) (4.22)
d
and, hence, the Jacobian j J( ) j (see, (4.19)) becomes rather complicated and analytical
integration are rarely possible.
When a boundary element e of a two-dimensional domain is circular with a curvature
radius Re and the coordinates’center of the circumference xei0 , the coordinates of a generic
point along the element can be expressed by means of these 3 parameters as a function
of the angle ' (between the x1 -axis and the radial vector xe xe0 ) in the form
xe1 = Re cos ' + xe10 (4.23)
xe2 = Re sin ' + xe20 (4.24)
Representing the angle ' by quadratic shape functions in the local coordinate
'( ) = N13 ( )'e1 + N23 ( )'e2 + N33 ( )'e3 (4.25)
where '1 ; '2 ;and '3 is the angle of the initial node, of an inner node, and of the end node,
respectively, on the circular element, gives the quadratic expression
'( ) = D3e 2
+ E3e + F3e (4.26)
Dependent on the de…nition range of , these coe¢ cients are
D3e = 2('e1 2'e2 + 'e3 ); E3e = 3'e1 + 4'e2 'e3 ; F3e = 'e1 for 0 1 (4.27)
D3e = 0:5('e1 2'e2 + 'e3 ); E3e = 0:5('e3 'e1 ); F3e = 'e2 for 1 1 (4.28)
The quadratic relation (4.26) between ' and allows also an explicit de…nition of as a
function of '
( p
1
2D3
E 3 + E32 4D3 (F3 ') for D3 6= 0
= 1
(4.29)
E3
(' F3 ) for D3 = 0
where D3 = 0 means '2 = 0:5('1 + '3 ), i.e., the node with the angle '2 has the same
distance to the initial node of the circular element with the angle '1 and to the end node
with the angle '3 .
The Jacobian of this transformation is calculated via
dxe1 dxe1 d'
= = Re sin ' (2D3e + E3e )
d d' d
dxe2 dxe2 d'
= = Re cos ' (2D3e + E3e )
d d' d
to be
s
dxe1 2 dxe
J( ) = ( ) + ( 2 )2 = Re (2D3e + E3e ) for D3 6= 0 (4.30)
d d
= Re E3e for D3 = 0 (4.31)
Hence, when D3 = 0, the Jacobian is simply equal to the length of the circular element:
J( ) = se = Re ('e3 'e1 ) for 0 1, or to the half of it: J( ) = 21 se = Re 12 ('e3 'e1 )
for 1 1.
When the angle ' is represented by linear shape functions in the local coordinate
'( ) = N12 ( )'e1 + N22 ( )'e2 (4.32)
where '1 and '2 is the angle between the outer normal vector at the initial node, at an
inner node, and at the end node, respectively, on the circular element and the x1 -axis,
this linear expression is
'( ) = E2e + F2e (4.33)
with
E2e = 'e2 'e1 ; F2e = 'e1 for 0 1

e e e e e e
E2 = 0:5('2 '1 ); F2 = 0:5('2 + '1 ) for 1 1
From (4.33) one obtains

' F2e
=
E2e
4.2 Integration over boundary elements 87
and the Jacobian of this linear transformation with

dxe1 dxe d' dxe2 dxe d'
= 1 = Re sin 'E2e ; = 2 = Re cos 'E2e
d d' d d d' d
as
J( ) = Re E2e (4.34)
Hence, as in the case of the quadratic approximation for the choice D3 = 0, the Jacobian
is simply equal to the length of the circular element: J( ) = se = Re ('e2 'e1 ) for
0 1, or to the half of it: J( ) = 12 se = Re 12 ('e2 'e1 ) for 1 1.
A boundary integral with the kernel g(x; ) and the state variable (x) over a bound-
ary element e is then transformed into an integral over the local coordinate as:
1 X
n
e le
s g(x; ) (x)d x = s g(x ( ); ) Nln ( ) j J( ) j d
e =0
l=1
X
n
1
le
= s g(xe ( ); )Nln ( ) j J( ) j d (4.35)
=0
l=1
0 0
For points in a di¤erent boundary element e , e 6= e, the integrand of (4.35) is regular
for all possible kernels g(x; ) so that those integrals can be numerically evaluated by a
Gaussian quadrature formula
1 1 X
s g(xe ( ); )Nln ( ) j J( ) j d = s flne ( )d = flne ( p )wp (4.36)
=0 =0
p
The number of Gaussian points with the abscissae p and the weights wp is dependent on
the distance of the point from the integration element e , generally, between 4 and 10
Gaussian points are taken.
Coincides the point with one of the nodes xle of the considered boundary element e ,
most of the integral kernels become either weakly singular or strongly singular, some even
hypersingular. Such integrals should be integrated analytically if possible (see Appendix
B).
4.2.2 Elements on boundary surfaces

When integrating over a boundary surface , in each boundary element e , the mapping
between the global physical coordinates x and the local coordinates 1 and 2 has to be
taken into account by the surface Jacobian which relates the element of area d x on the
surface to the element of area d 1 d 2 in the local parameter domain
d x =j J( 1 ; 2) j d 1d 2 (4.37)
The surface will contain lines corresponding to constant values of the local coordinates
1 and 2 . Moving along the constant 1 -coordinate line from ( 1 ; 2 ) to ( 1 ; 2 + d 2 )
results in a point x( 1 ; 2 + d 2 ) with the approximative position vector

@x
x( 1 ; 2 + d 2 ) = x( 1 ; 2) + d 2 + O((d 2 )2 )
@ 2
and similarly, moving along a constant 2 -coordinate line gives
@x
x( 1 + d 1; 2) = x( 1 ; 2) + d 1 + O((d 1 )2 )
@ 1
Hence, the vectors along the sides of the surface element of area d x are
@x
x( 1 ; 2 + d 2) x( 1 ; 2) d = 2 + O((d 2 )2 )
@ 2
@x
x( 1 + d 1; 2) x( 1 ; 2 ) = d 1 + O((d 1 )2 )
@ 1
To the …rst order, the surface area will be plane and d x is thus given by the formula from
vector calculus relating the area of a plane rectangle to the cross product of the vectors
on its sides:
@x @x @x @x
d x= j d 1 d 2j = j j d 1d 2 (4.38)
@ 1 @ 2 @ 1 @ 2
By comparison with (4.37), the surface Jacobian is
@x @x
j J( 1 ; 2) j = j j (4.39)
@ 1 @ 2
Since the cross product may be expanded as (see (2.10)
@x2 @x3 @x1 @x3 @x1 @x2
@x @x @ 1 @ 1 @ 1 @ 1 @ 1 @ 1
= ^
e1 @x2 @x3 ^
e2 @x1 @x3 +^
e3 @x1 @x2
@ 1 @ 2 @ 2 @ 2 @ 2 @ 2 @ 2 @ 2
= ^
e1 m11 + ^
e2 m12 + ^
e3 m13
@x @x
where mij are the minors of @ 1 @ 2
. Thus
q
@x @x
j J( 1 ; 2 ) j = j j = m211 + m212 + m213 (4.40)
@ 1 @ 2
A boundary integral with the kernel g(x; ) and the state variable (x) over a bound-
ary element e is then transformed into an integral over the local coordinates 1 and 2 ,
e.g., with the range 1 1; 2 1 as:
1 1 X
n
e le
s g(x; ) (x)d x = s s g(x ( 1 ; 2 ); ) Nln ( 1 ; 2) j J( 1 ; 2) j d 1d 2
e
1= 1 2= 1
l=1
X
n
1 1
le
= s s g(xe ( 1 ; 2 ); )Nln ( 1 ; 2) j J( 1 ; 2) j d 1(4.41)
d 2
l=1 1= 1 2= 1
4.3 Boundary element equations by point collocation 89
0 0
For points in a di¤erent boundary element e , e 6= e, the integrand of (4.41) is regular
for all possible kernels g(x; ) so that those integrals can be numerically evaluated by a
Gaussian quadrature formula
1 1 1 1
s s g(xe ( 1 ; 2 ); )Nln ( 1 ; 2) j J( 1 ; 2 ) j d 1 d 2 = s s flne ( 1 ; 2 )d 1 d 2
1= 1 2= 1 1= 1 2= 1
XX
= flne (ap ; aq )wp wq (4.42)
p q
The number of Gaussian points with the abscissae ap and the weights wp as well as with
the abscissae aq and the weights wq is dependent on the distance of the point from the
integration element e , generally, between 4 and 10 Gaussian points are taken.
Coincides the point with one of the nodes xle of the considered boundary element e ,
most of the integral kernels become either weakly singular or strongly singular, some even
hypersingular. Such integrals should be integrated analytically if possible (see Appendix
B).
4.3 Boundary element equations by point collocation

A ’direct’integral equation, e.g., in the case of a scalar boundary value problem (see, e.g.,
(3.158))
c( ) ( ) + s g(x; ) (x)d x = s h(x; ) (x)d x (4.43)
with a weakly singular and a strongly singular kernel h(x; ) and g(x; ), respectively,
will be transformed by partitioning the boundary into elements e (e = 1; :::; m), i.e.,
decomposing the integrals over the whole boundary into a sum of m integrals over single
elements e
X
m
s [::::] d x = s [::::] d x (4.44)
e
e=1
and by the approximation of the boundary, i.e., of the boundary coordinates xei (see,
(4.1)), and of the boundary states (x) and (x) (see (4.8)), as it is shown in (4.35)),
into the following approximate equation
X
m X
n
1
le
c( ) ( ) = s g(xe ( ); )Nln ( ) j J( ) j d
=0
e=1 l=1
Xm X n
1
le
+ s h(xe ( ); )Nln ( ) j J( ) j d (4.45)
=0
e=1 l=1
Remark: In many physical boundary value problems, the state variable (x) is continous
along the whole boundary (e.g., the temperature in the case of heat conduction problems) while
the state (x) is discontinuous since it is often related to the normal vector which jumps at
corner points. Such discontinuities have to be taken into account in approximating this state.
The number of unknown node values le or le of the shape functions of (x) and
(x), respectively, in this single equation (4.7) is dependent on the grade (n 1) of the
applied shape functions: - for constant shape functions, i.e., n = 1, as well as for linear
shape functions, i.e., n = 2, one has m unknowns in the case of a closed smooth contour
(i.e., in the linear case assuming along the whole boundary a continous approximation)
and 2m unknowns for quadratic shape functions, i.e., n = 3, ( again assuming a continous
approximation). Hence, one need m and 2m, respectively, equations for their unique
determination.
The most simple and mostly used way is to evaluate this equation (4.45) at as many
points = xj as one needs where it is advisable and useful to choose for this purpose,
i.e., as these collocation points, the node points xle of the shape functions:
Xm X n
1
j j le
c(x ) (x ) = s g(xe ( ); xj )Nln ( ) j J( ) j d
=0
e=1 l=1
Xm X n
1
le
+ s h(xe ( ); xj )Nln ( ) j J( ) j d (4.46)
=0
e=1 l=1
Since for linear and all higher grade approximations (n = 2) the end node of an element
ne
e is also the initial node of the sequent element e+1 , i.e., x = x1e+1 , it is helpful to
j
introduce a global node numbering x ; j = 1; 2; :::; N , as follows
for n = 2, i.e. N = m : x1 = x11 = x2m ; x2 = x21 = x12 ; x3 = x22 = x13 ; :::
for n = 3, i.e. N = 2m : x1 = x11 = x3m ; x2 = x21 ; x3 = x31 = x12 ; :::
4.3.1 Approximation by constant shape functions

Physically not very realistic but most simple is an approximation of the state functions
by constant shape functions (n = 1) and collocation at the middle node x1e = xe of each
element e .
With N11 ( ) = 1 one has only one representative value 1e = e and 1e = e for
the boundary states in each element and (4.46) yields with a linear approximation of the
boundary (i.e., j J( ) j= le )
1 j 1 Xm
1 Xm
1
j e e j e
= (x ) = s g(x ( ); x )le d + s h(xe ( ); xj )le d (4.47)
2 2 e=1
=0
e=1
=0
where, due to the collocation at the middle nodes xj of the straight and, therefore, smooth
boundary elements, the factor c(xj ) is always c(xj ) = 1=2. Finally, one obains with the
matrices
1 1
Gje = s g(xe ( ); xj )le d and Hje = s h(xe ( ); xj )le d (4.48)
=0 =0
the following system of m equations for m unknown boundary values
Xm
1 Xm
e
je + Gje = Hje e for j = 1; 2; :::; m (4.49)
e=1
2 e=1
Example: Stationary heat conduction

In the case of a stationary heat conduction problem, the boundary states in (4.49) are
the temperature and the heat ‡ux qn , i.e., e , e and e , qne , and the kernels
g(x ( ); ) , qn (x; )= 0 = 2 1 r;n =r and h(xe ( ); xj ) ,
e
(x; )= 0 = 2 10 ln(r=c)
(see, (3.152) and (3.148)). Hence, one has to evaluate the following expressions for the
matrices Gje and Hje , respectively, (r1ej = x1e xj ):
1 1 1 @rje ( ) e le nei 1 x1e (1 ) + x2e xji

Gje = s ni le d = s i 2
i
d
2 =0 rje ( ) @xi 2 =0 rje ( )
le 1 nei (x1e
i xji ) + (x2e i x1e e
i )ni
= s 2
d (4.50)
2 =0 rje ( )
le 1 d
= ne r1ej s 2 2
(4.51)
2 =0 le + 2r1ej te le + j r1ej j2
2
le 1 rje ( )
Hje = s ln d
4 0 =0 l02
le 1 l2 2 + 2r1ej te le + j r1ej j2
= s ln e d
4 0 =0 l02
le 1 2 j r1ej j2 le2
= s ln 2 + r1ej te + + ln d (4.52)
4 0 =0 le le2 l02
where the constant reference value c in has been taken as the length of the shortest
element e , i.e., l0 = min fle j e = 1; :::; m g and, moreover, the following relations hold for
a linear boundary approximation
2 2
2
rje ( ) = x1e
1 (1 ) + x2e
1 xj1 + x1e
2 (1 ) + x2e
2 xj2
= le2 2
+ 2(x1e
i xji )le tei + (x1ei xji )(x1e
i xji )
= le2 2
+ 2r 1ej e
t le + j r j 1ej 2
(4.53)
(x2e
i x1e e 2e
i ) = le ti ! (xi x1e e e e
i )ni = le ti ni = 0 (4.54)
For the singular elements, i.e., for a collocation at the middle node xji = 0; 5(x2e 1e
i + xi )
of the considered element e , one obtains
r1ej , x1e
i xji = 0; 5(x2e
i x1e
i ) = 0; 5le tei
so that
ne r1ej , nei (x1e
i xji ) = 0; 5le nei tei = 0
and, hence, one obtains from (4.50) regarding also (4.54)
Gee = 0:
The reason is @r=@n = 0 for this middle node collocation point.

For a collocation at this middle node is due to
te r1ej , tei (x1e

i xji ) = 0; 5le tei tei = 0; 5le
r1ej r1ej , (x1e i xji )(x1e
i xji ) = 0:25le2
the respective singular element in the H-matrix (for details about the evaluation of this
improper integral, see Appendix B, 4.5.1)
( )
2
le 1 1 le2
Hee = s ln + ln d
4 0 =0 2 l02
"1 #
" 2 2
le 2 1 1 1 le2
= lim s ln d + s ln d + ln
4 0 "!0 =0 2 = 12 +" 2 l02
le
= [1 + ln(2) ln (le =l0 )] (4.55)
2 0
For all other collocation points (where nei (x1e

i xji ) = ne r1ej 6= 0) holds
=1
le e 1ej 2 le + te r1ej
Gje = n r arctan
2 2le j ne r1ej j j ne r1ej j =0
e 1ej
1 le + t r e
t r1ej
= sign ne r1ej arctan arctan (4.56)
2 j ne r1ej j j ne r1ej j
8 9
> e 1ej 2 e 1ej jr1ej j2 >
>
> (le + t r ) ln 1 + le
t r + le2 >
>
1 < jr 1ej j2 2
=
Hje = te r1ej ln l 2 2le + le ln lle2 (4.57)
4 0>> h e 0 i >
>
>
: +2 j ne r1ej j arctan le +te e 1ej r1ej te r1ej >
;
jn r j
arctan jne r1ej j
Remarks:
a) There holds the following separation of components for length of the distance vector
j
between the collocation point xi and the initial node x1e
i of the element e
2 2 2
r1ej = (x1e
i xji )(x1e
i xji ) = nei (x1e
i xji ) + tei (x1e
i xji )
where nei and tei means the unit normal vector and unit tangential vecor, respectively, in the
straight element e .
b) In the above evaluation of Gje , the following integral was applied (for 4b a2 > 0):
2 a 2
p 2 +a
s ln( + a + b) d = ( + ) ln( + a + b) 2 + 4b a2 arctan p
2 4b a2
Exercise 14: Stationary heat conduction in a rectangular domain- constant

shape functions
In a rectangular domain = f(x1 ; x2 ) j 0 x1 l1 = 1m; 0 x2 l2 = 2mg, the tem-
perature is prescribed at two opposite sides x2 = 0 and x2 = l2 to be (x1 ; 0) = 0 and
(x1 ; l2 ) = > 0 while at the other two sides x1 = 0 and x1 = l1 the heat ‡ux is stopped,
i.e., qn = 0. For this problem (see [2]), the exact temperature distribution is known to
be (x1 ; x2 ) = x2 =l2 from which one can determine via (3.146) the heat ‡ux across the
boundary to be as qn (x1 ; x2 ) = 0 ni (x1 ; x2 )@ (x1 ; x2 )=@xi = 0 n2 (x1 ; x2 ) /l2 .
Solve this problem applying the above given boundary element approximation by taken
each of the four sides as a boundary element and compare with the known exact solution.
4.3.2 Approximation by linear and higher grade shape functions

Continous state functions (x) and (x), which are approximated by shape functions
with n (= 2) nodes per element, have in successive boundary elements e und e+1 the
same node value, i.e., ne = 1e+1 and ne = 1e+1 , respectively, and, hence, an equal
global node value numbering i and i , respectively. Therefore, for each collcation point
xj , the elements in one line of the matrices G and H related to the boundary state node
values i and i , respectively, read as:
8 1 1
>
> s g(x e
( ); xj
)N n
( ) j J( ) j d + s g(xe+1 ( ); xj )N1n ( ) j J( ) j d
>
> n
< =0 =0
Gji = for i = ne = 1e+1 (4.58)
>
> 1
>
>
: s g(xe ( ); xj )Nln ( ) j J( ) j d for i = le if l 6= 1; n (n > 2)
=0
8 1 1
>
> s h(xe
( ); xj
)N n
( ) j J( ) j d + s h(xe+1 ( ); xj )N1n ( ) j J( ) j d
>
> n
< =0 =0
Hji = for i = ne = 1e+1 (4.59)
>
> 1
>
>
: s h(xe ( ); xj )Nln ( ) j J( ) j d for i = le if l 6= 1; n (n > 2)
=0
With that, for continous state functions on a smooth boundary of a 2d-body, the following
system of equations for the N (for n = 2 is N = m and for n = 3 is N = 2m) unknown
node values is given (cj = c(xj ) and j = (xj ))
X
N X
N
j i i
c ji + Gji = Hji for j = 1; 2; ::; N (4.60)
i=1 i=1
Remark: For a linear approximation of the boundary and linear shape functions in each
boundary element e for a continous state function the elements (4.58) are explicitly:
1 1
for i = 2 : Gji = s g(xe ( ); xj ) le d + s g(xe+1 ( ); xj )(1 )le+1 d with i = e + 1
=0 =0
1 1
for i = 1 : Gj1 = s g(xm ( ); xj ) lm d + s g(x1 ( ); xj )(1 )l1 d at x1 = xm+1
=0 =0
At corners of the boundary approximation, the direction of the normal vector jumps
and, hence, a boundary state, e.g., (x), which is dependent on the normal vector, is
also discontinous there, i.e., ne 6= 1e+1 when the boundary elements e and e+1 form
a corner. But also at a point on a smooth boundary, the prescribed boundary state
function (x) itself can be discontinous, e.g., in a heat conduction problem when the
permeability and consequently the heat ‡ux changes suddenly. For the case that such
discontinuities ne 6= 1e+1 are assumed at the transition points between all elements
0
increase the number of node values i , e.g., for n = 2 from N = m to N = 2m and for
0
n = 3 from N = 2m to N = 3m:
X
m m X
X n
j i l le
c ji + Gji = Hje für j = 1; 2; ::; m (4.61)
i=1 e=1 l=1
le
where, di¤erent to (4.59), the coe¢ cients of , i.e., the elements of H are de…ned as
1
l
Hje = s h(xe ( ); xj )Nln ( ) j J( ) j d
=0
with j = 1(1)m and l = 1(1)n; e = 1(1)m (n > 2) (4.62)
It should be mentioned that in spite of those discontinuities, ne 6= 1e+1 , for the most
types of boundary conditions, the system (4.61) obtained by collocation at the nodes of
the shape functions allows to determine all unknown nodal values uniquely since:
at a node on a smooth boundary, a discontinuity of (x) appears only when it is
prescibed, i.e., only the nodal value of the continous state (x) is unknown at those nodes,
at a corner node, almost all combinations of boundary conditions ’ahead’and ’be-
hind’a corner produce only one unknown nodal value, e.g.,
-when ahead ne and behind 1e+1 is prescribed: ne = 1e+1 is the single unknown,
- when ahead ne and behind 1e+1 (= ne ) is prescribed: 1e+1 is the single unknown,
- when ahead ne (= 1e+1 ) and behind 1e+1 is prescribed: ne is the single unknown,
Only for the case
- when ahead ne and behind 1e+1 = ne is prescribed, the two values ne and 1e+1
are both unknown and, since the collocation at the corner node delivers only one equation,
one equation is missing.
Often, these needed two equations (instead of only one) are produced by using the
’double node concept’(for more details see [6], pp. 87), i.e., two collocation points (instead
of the one at the corner) located ’near’ the corner in the element ahead and behind,
respectively, in the case of linear shape functions often in a distance of le =3 from the
corner.
If the discontinuity at a corner x = xE is only produced by the jump of the normal
vector ahead and behind the corner, i.e., n (xE ) 6= n+ (xE ), sometimes a physically based
extra condition can be found and the ’double nodes’at the corner can be avoided (see the
following example).
Example: Stationary heat conduction

In the stationary heat conduction problem, the state (x) means the heat ‡ux qn (x) which
is de…ned on the boundary as qn (x) = 0 ni (x)@ (x)=@xi and, hence, is dependent on
the normal vector ni (x). Since the gradient of the temperature is continous also at a
corner xE of the boundary, the following relations hold for the normal and tangential
components of this temperature gradient ahead and behind a corner
@ (x) @ (x) @ (x) @ (x) + @ (x) +
ni + ti = = n + t (4.63)
@n @t @xi @n+ i @t+ i
where n = (cos ' ; sin ' ); t = ( sin ' ; cos ' ) and n+ = (cos '+ ; sin '+ ); t+ =
( sin '+ ; cos '+ ) are the normal and tangential unit vectors, respectively, ahead and be-
hind the corner. With that, the above mentioned extra equation (which is only necessary
when ahead and behind a corner the temperature is prescribed and, consequently, the
tangential derivatives ahead @ (x)=@t and behind @ (x)=@t+ are known) at a corner
with internal angle ' = '+ + ' can be given as the following relation between
+
the heat ‡ux qn (x) and qn (x) ahead and behind the corner, respectively:
@ (x) @ (x)
qn+ (x) = 0 = 0 sin ' cos 'qn (x) (4.64)
@n+ @t
or equivalently
@ (x) @ (x)
qn (x) = 0= cos 'qn+ (x) 0 sin ' (4.65)
@n @t+
Applying a linear approximation for the boundary and linear shape functions for (x)
and qn (x); one obtains for heat conduction problems without interior heat sources, i.e.,
with Wq (x) = 0, from (3.158) the following integral equation for the temperature at a
point xj
8 1
9
m >
> 1e e j
)le d > >
X < qn s (x ( ); x )(1 =
j j =0
c(x ) (x ) = 1
> >
e=1 >
: +qn2e s (xe ( ); xj ) le d >
;
=0
8 1
9
m <
>
> 1e e
s qn (x ( ); x )(1 j
)le d >>
X =
=0
1 (4.66)
> >
e=1 > >
2e e j
: + s q n (x ( ); x ) le d ;
=0
with
'(xj )
j for boundary points xj
c(x ) = 2 (4.67)
1 for interior points xj
e 1e 2e 1e+1
or corresponding (4.61) in a matrix notation ( = with = )
X
m X
m
c(xj ) (xj ) = 1 1e
Hje 2 2e
qn + Hje qn Gje e
für j = 1; 2; ::; m (4.68)
e=1 e=1
where '(xj ) = 2 ('2 '1 ) means the internal angle at a boundary point xj .
Related to the everywhere continous temperture nodal value k = 2e = 1e+1 ,
k = e + 1; e = 1(1)m ( m+1 = 1 = 1m+1 means that the boundary contour is closed)
one obtains from (3.152) explicitly the following expressions for the matrix coe¢ cients
Gjk (ri1ej = x1e
i xji ):
j j
for xi = x1e+1
i = x2e i and each k = 1(1)m (even when the collocation point xi is at
a corner of the linearly approximated boundary):
Gjk = 0;
for xji
6= x1e+1
i = x2e e 1e
i and k = e + 1 > 1 and ni (xi xji ) = ne r1ej 6= 0 or ne+1 r1e+1j
6= 0 (for ne r1ej = 0 or ne+1 r1e+1j = 0, the respective integrals with those factors
disappear):
le 1 d
Gjk = ne r1ej s 2 2
2 =0 le + 2te r1ej le
+ j r1ej j2
le+1 e+1 1e+1j 1 (1 )d
n r s 2 2 e+1 1e+1j
2 =0 le+1 + 2t r le+1 + j r1e+1j j2
8 9
le2
e 1ej < 1 le te rej
ln 1 + 2 jr1ej j2 + jr1ej j2 =
n r 2
=
2 le : te rej e r1ej
arctan lejn+te r1ej arctan jnt e rr1ej j ;
e 1ej
jne rej j j
8 2
9
> 1 le+1 te+1 r1e+1j le+1
>
>
< 2
ln 1 + 2 jr 1e+1j j2 + jr 1e+1j j2 >
=
ne+1 r1e+1j !
+ le+1 +te+1 r1e+1j (4.69)
> le+1 +t e+1 r 1e+1j arctan >
2 le+1 > jne+1 r1e+1j j
>
: jne+1 r1e+1j j
arctan t r
e+1 1e+1j ;
jne+1 r1e+1j j
for xji 6= x11

i = xi
1m+1
= x2m
i and k = 1 as well as nm i (xi
1m
xji ) 6= 0 or n1i (x11i xji )
6= 0 (for nmi (xi
1m
xji ) = 0 or n1i (x11 i xji ) = 0, the respective integrals with those factors
disappear):
8 9
j < 1 lm (x1m xji )tm 2
lm =
m 1m
ni (xi xi ) 2
ln 1 + 2 i
jx1m xj j2
i
+ jx1m xj j2
Gj1 = 1m j m 1m j m 1m j m
2 lm : (xi xi )ti l +(x
arctan mjnm (xi 1m xi j )ji
x )t
arctan jnmi (x1m i xij )j ;
(x x )t
jnm (x 1m xj )j
i i i i i i i i i
8 9
l1 (x11 xji )t1i l 2
>
>
1
ln 1 + 2 jx11 xj j2 + jx11 xj j2
i 1 >
>
1 11 j <> 2
0 1 >
=
ni (xi xi ) l1 +(x11 x j
)t 1
+ l1 +(x11 xji )t1i @
arctan jn1 (x11 xj )ji i i
(4.70)
2 l1 >
> i i i i A > >
>
: 1 11
jni (xi xi )j j
(x 11 j
x )t 1 >
;
arctan i i ji jn1i (x11 xi )j
i
Related to the heat ‡uxes qn1e and qn2e respectively, at xji one obtains from (3.148) the both
contributions from the adjacent elements e and e+1
2 le 1 2 e 1ej
2 j r1ej j2 le2
Hje = s ln + t r + + ln 2 d (4.71)
4 0 =0 le le2 l0
2
1 le+1 1 2 e+1 1e+1j j r1e+1j j2 le+1
Hje+1 = s (1 ) ln 2 + t r + 2
+ ln (4.72)
d
4 0 =0 le+1 le+1 l02
2 1
where these two contributions are only added to Hjk = Hje + Hje+1 , k = e + 1, if qnk is
continous at this point. This gives explicitly:
for xji = x1e+1
i = x2e
i
j
i.e., x1e
i xi = x1ei x2e
i = le tei ! te r1ej = tei ( le tei ) = le and j r1ej j= le
1e+1 j 1e+1
and xi xi = xi xi = 0 ! te+1 r1e+1j = 0 and j r1e+1j j= 0
2e
2 le 1 "
2 le2 le le2
Hje = lim s ln 2 + 1 + ln d = 3 ln (4.73)
4 0 "!0 =0 l02 8 0 l02
2 2
1 le+1 1
2 le+1 le+1 le+1
Hje+1 = lim s (1 ) ln + ln d = 3 ln (4.74)
4 0 "!0 =" l02 8 0 l02
for xji = x1e

i
j
i.e., x1e
i x i = 0
2 le 1
2 le2 le le2
Hje = lim s ln + ln d = 1 ln (4.75)
4 0 "!0 =" l02 8 0 l02
for xji = x2e+1

i
i.e., xi1e+1 xji = x1e+1
i x2e+1
i = le+1 te+1
i ! te+1 r1e+1j = te+1
i ( le+1 te+1
i )= le+1
and j r1e+1j j= le+1
2 2
1 le+1 1 "
2 le+1 le+1 le+1
Hje+1 = lim s (1 ) ln 2 + 1 + ln d = 1 ln
4 0 "!0 =0 l02 8 0 l02
(4.76)
for xji 6= x1e j
i and xi =6 x2ei = xi
1e+1
and nel (x1e
l xjl ) = ne r1ej 6= 0 :
8 ! 9
jne r1ej j2
>
> le2 2 >
>
>
>
2
le e 1ej
ln 1 + 2 lejrt1ejr j2 + jr1ej + ln lle2 >
>
>
> jte r1ej j2 j 2 >
>
>
< le2
0
>
=
le 1ej 2
2
Hje = + ln 1 + l2e te r1ej + jr l2 j + l2e te r1ej 1 (4.77)
8 0> > e ! >
>
>
> arctan le +te r1ej >
>
>
> 4 e e
jn r j1ej >
>
>
: le2
t r1ej j ne r1ej j e 1ej >
;
arctan t r jne r1ej j
for xji 6= x1e+1

i and ne+1
l (x1e+1
l xjl ) = ne+1 r1e+1j 6= 0 with e + 1 = k
8 0 1 9
2 k 1kj
>
> t r >
>
>
> B
lk
k r1kj j2 C >
>
>
> B jn
C l t k r 1kj l 2 l 2 >
>
>
> @ lk2
A ln 1 + 2 k
jr 1kj j2 + jr
k
1kj j2 + ln k
2 >
>
>
> k r1kj j2
l 0 >
>
l < + jt
2
=
1 k lk
Hjk = 1kj 2 (4.78)
8 0> > + ln 1 + l2k tk r1kj + jr l2 j 2 k
t r1kj 3 >
>
>
> lk
! >
>
>
>
k
l +t k r 1kj >
>
>
> 1+ arctan k >
>
>
> + 4
j n k
r 1kj
j jn k r1kj j
>
>
: lk 1 k
t r 1kj
arctan k 1kj t k r1kj ;
lk jn r j
Remark:
a) Integration by parts yields the following result which was important for the integration
of Gje (for 4b a2 > 0):
2 ( +pa2 ) ln( 2 + a + b) 2
s ln( + a + b) d =
+ 4b a2 arctan p24b+aa2
" #
2 a2 2 2
1 ( +b p2 ) ln( + a + b) +a
s ln( 2 + a + b) d = 2 2 +a
2 a 4b a arctan p4b a2
b) In evaluating the integrals, one obtains the following di¤erence of two ln-terms which can
be assembled as follows:
2 e 1ej j r1ej j2 j r1ej j2

ln 1 + t r + ln
le le2 le2
le te r1ej le2
= ln 1 + 2 1ej 2 + 1ej
jr j j r j2
c) It is well known that the scalar product of a directional unit vector with an arbitrary
vector gives the vector’s component in the direction of the unit vector. Hence, for the unit
vectors te and ne , which are orthogonal to each other, holds (see Fig. )
j r1ej j2 =j te r1ej j2 + j ne r1ej j2
Exercise 15: Stationary heat conduction in a rectangular domain- linear shape

functions
In a rectangular domain = f(x1 ; x2 ) j 0 x1 l1 = 1m; 0 x2 l2 = 2mg, the tem-
perature is prescribed at two opposite sides x2 = 0 and x2 = l2 to be (x1 ; 0) = 0 and
(x1 ; l2 ) = > 0 while at the other two sides x1 = 0 and x1 = l1 the heat ‡ux is stopped,
i.e., qn = 0. For this problem (see [2]), the exact temperature distribution is known to
be (x1 ; x2 ) = x2 =l2 from which one can determine via (3.146) the heat ‡ux across the
boundary to be as qn (x1 ; x2 ) = 0 ni (x1 ; x2 )@ (x1 ; x2 )=@xi = 0 n2 (x1 ; x2 ) /l2 .
Solve this problem applying the above given linear boundary element approximation
by taken each of the four sides as a boundary element and compare with the known exact
solution
5 Appendices
5.1 A: Exercise Solutions

Exercise 1: Nabla Vector
@U @U
gradU = e1 +
^ ^
e2
@x1 @x2
@U @r @U @'
= + (^
er cos ' ^ e' sin ')
@r @x1 @' @x1
@U @r @U @'
+ + (^
er sin ' + ^
e' cos ')
@r @x2 @' @x2
Since
q
@r x1 @r x2
r = x21 + x21 !
= = cos '; = = sin '
@x1 r @x2 r
x2 @' x2 sin ' @' x1 cos '
' = arctan( ) ! = 2 = ; = 2 =
x1 @x1 r r @x2 r r
the above expression for gradU may be expressed as
@U @U sin '
gradU = cos ' + ( ) (êr cos ' ^ e' sin ')
@r @' r
@U @U cos '
+ sin ' + (^
er sin ' + ^
e' cos ')
@r @' r
@U 1 @U
= er +
^ ^
e'
@r r @'
@ 1 @
= (^
er +ê' )U
@r r @'
Exercise 2: Laplace operator

@ 1 @ @ 1 @
U = (^
er +ê' ) (^
er +^
e' )U
@r r @' @r r @'
The scalar product is only ’working’between unit vectors, i.e., ^ er = ^
er ^ e' ê' = 1 and
^ e' = 0, but the di¤erentiations have to be performed following the chain rule, i.e.,
er ^
@ @ 1 @ @^
er @ @2 @ê' 1 @ 1 @ 1 @2
(^
er +^
e' ) = +^
er 2 + +^
e' ( 2 + )
@r @r r @' @r @r @r @r r @' r @' r @'@r
@ @ 1 @ @^
er @ @2 @ê' 1 @ 1 @2
(^
er +^
e' ) = +^
er + +ê'
@' @r r @' @' @r @r@' @' r @' r @'2
99
100 5 Appendices
For the di¤erentiations of the unit vectors of the Polar coordinate system holds
@^
er @ê' @^
er @^
e'
= = 0; =^
e' ; = ^
er
@r @r @' @'
which reduces the above results to

@ @ 1 @ @2 1 @ 1 @2
(^
er +^
e' ) = ^
er 2 + ê' ( 2 + )
@r @r r @' @r r @' r @'@r
@ @ 1 @ @ @2 1 @ 1 @2
(^
er +^
e' ) = ^
e' +^
er ^
er +ê'
@' @r r @' @r @r@' r @' r @'2
Now, performing the scalar product of the …rst line result with ^
er and of the second one
e' 1r gives …nally
with ^
@2 1 @ 1 @2
U =( 2 + + 2 2 )U
@r r @r r @'
Exercise 3: Integration by parts

a) in R1 on the intervall a < x < b :
Zb x=b Zb
n xn+1 xn+1 1
x ln(x)dx = ln(x) dx
n+1 x=a n+1x
a a
n+1 n+1 x=b
x x
= ln(x)
n+1 (n + 1)2 x=a
n p o
b) in R2 on the circular domain = (x1 ; x2 ) j r = x21 + x22 R :
Z Z2 ZR ZR
n n
r ln(r)d = r ln(r)rdrd' = 2 rn+1 ln(r)dr
0 0 0
R ZR
rn+2 rn+2 1
= 2 ln(r) 2 dr
n+2 0 n+2r
0
n+2 n+2 R
r r
= 2 ln(r)
n+2 (n + 2)2 0
2 1
= Rn+2 ln(R)
n+2 n+2
Exercise 4: Reversed order integrations

a)
5.1 A: Exercise Solutions 101
x=s s=x
x s x4 x x s5 a4 s6 a4 2
s s x3 s dx ds = s s ds = s s ds = s
a a a 4 x=a a 4 4 24 8 s=a
x 6 1 4 2 a6 a6 1
= ax + = (x6 3a4 x2 + 2a6 )
24 8 24 8 24
2 s=x 2 x=x
x x
3
x
3s
x
3x x5 1 x6
s s x s ds dx = s x dx = s x dx = x4 x2
a x a 2 s=x a 2 2 8 12 x=a
6
x x6 1 4 2 a 1 6
= ax + = (x6 3a4 x2 + 2a6 )
8 12 8 12 24
b)
x b x x x b
s s G(x; s)dx ds = s s G(x; s)dx ds + s s G(x; s)dx ds
a s a s a x
x x b x
= s s G(x; s)ds dx + s s G(x; s)ds dx (5.1)
a a x a
x=b s=x
x b
3x4 x b4 s s5
x b 4 s2 s6
s s x s dx ds = s s ds = s ds =
a s a 4 x=s a 4 4 8 24 s=a
4 2 6 4 2 6
bx x ba a 1
= + = (3b4 x2 x6 3b4 a2 + a6 )
8 24 8 24 24
2 s=x 2 s=x
3s 3s
x x b x x b
3 3
s s x s ds dx + s s x s ds dx = s x dx + s x dx
a a x a a 2 s=a x 2 s=a
x x5 a2 b x 2
a2
= s x3 dx + s x3 x3 dx
a 2 2 x 2 2
x=x x=b
x 6 x 4 a2 x 4 x 2 a2
= + ( )
12 4 2 x=a 4 2 2 x=x
x 6 x 4 a2 a6 a6 b 4 x 2 a2 x4 x2 a2
= [ + ]+[ ( ) ( )]
12 8 12 8 4 2 2 4 2 2
1
= (3b4 x2 x6 3b4 a2 + a6 )
24
Exercise 5: Integral equation by straightforward integrations

One integration of both sides of the di¤erential equation (3.8) gives
Z
dy(x)
= f (x)dx + c0
dx x=x
102 5 Appendices
0 0
and satisfying the boundary condition y (b) = y1
Zb
dy(x) 0
= y1 f (x)dx
dx
x
Then, a second produces

2 0 b 1 3
Z Z
0
y(x) = y1 x 4 @ f (x)dxA ds5 + c1
s s=x
and …nally
0 x b
y(x) = y0 + (x a)y1 s s f (x)dx ds
a s
when the further constant of integration c1 having been taken so that y(a) = y0 . Applying
the formula (5.1), derived in Exercise 2, to the above double integral gives
x b x x b x
s s f (x)dx ds = s s f (x)ds dx + s s f (x)ds dx
a s a a x a
x b
= s (x a)f (x)dx + s (x a)f (x)dx
a x
and the …nal simpli…ed expression for the solution of the boundary value problem:
0 x b
y(x) = y0 + (x a)y1 s (x a)f (x)dx (x a)s f (x)dx: (5.2)
a x
Exercise 6: Beam de‡ection under prescribed moments

When in (3.28) the prescribed boundary conditions are taken into account and the system
is reordered, one obtains
0
0 1 w (a) 1 b (x a)M (x) 1 l w0
= s dx 0
l 1 w(b) EI a (b x)M (x) 1 0 w1
which has the solutions (b a = l)
1 b 0
w(b) = s (x a)M (x)dx + w0 + lw1
EI a
0 1 1 b
w (a) = s (b x)M (x)dx w0 + w(b)
l EI a
1 1 b 1 b 0
= s (b x)M (x)dx w0 + s (x a)M (x)dx + w0 + lw1
l EI a EI a
1 1 b 0
= s (b a)M (x)dx + lw1
l EI a
1 b 0
= s M (x)dx + w1
EI a
With these boundary reactions and the prescribed boundary values, the solution (3.27)
becomes
1 a 0 1 (b ) 0 b M (x) 1
w( ) = w(a) + w (a) + w(b) w (b) s jx j dx
2 2 2 2 a EI 2
1 a 1 b 0 1 1 b 0
= w0 + s M (x)dx + w1 + s (x a)M (x)dx + w0 + lw1
2 2 EI a 2 EI a
(b ) 0 b M (x) 1
w1 s jx j dx
2 a EI 2
0 1 1 b
= w0 + ( a)w1 + s( 2a + x jx j) M (x)dx
2 EI a
0 1 1 1 1 b
= w0 + ( a)w1 + s (2x 2a) M (x)dx + s (2 2a) M (x)dx
2 EI a 2 EI
0 1 1 b
= w0 + ( a)w1 + s (x a) M (x)dx + ( a)s M (x)dx
EI a EI
This is exactly the solution (5.2) when one recognizes that and x are there x and x,
respectively, and f = M=EI.
Exercise 7: Bending moment of an elastic beams under transversal loading

0
By formally setting M , Q, and q instead of w, w , and M=EI, respectively, the solution
is directly obtained from (3.27) as
1 a 1 (b ) b 1
M ( ) = M (a) + Q(a) + M (b) Q(b) s q(x) j x j dx (5.3)
2 2 2 2 a 2
while the algebraic system (3.28) to determine the unknown boundary reactions becomes
(b a = l):
2 3
M (a)
1 0 1 l 6 7 b
6 Q(a) 7 = s (x a)q(x) dx (5.4)
1 l 1 0 4 M (b) 5 a (b x)q(x)
Q(b)
Exercise 8: Axial displacement of elastic bars

0
By formally setting u, N=EA, and p=EA instead of w, w , and M=EI, respectively, the
solution is directly obtained from (3.24)
b
du(x) @u (x; ) b p(x)
u( ) = u (x; ) u(x) s u (x; )dx (5.5)
dx @x a a EA
104 5 Appendices
and with N (x) = EAu0 (x)

b
N (x) N (x; ) b p(x)
u( ) = u (x; ) u(x) s u (x; )dx (5.6)
EA EA a a EA
or in detail with (3.27) as
1 a N (a) 1 (b ) N (b) b p(x) 1

u( ) = u(a) + + u(b) s jx j dx (5.7)
2 2 EA 2 2 EA a EA 2
Then, the algebraic system (3.28) to determine the unknown boundary reactions becomes
(b a = l):
2 3
u(a)
1 0 1 l 6 N (a) 7 1 b (x a)p(x)
6 EA 7 = s dx (5.8)
1 l 1 0 4 u(b) 5 EA a (b x)p(x)
N (b)
EA
Exercise 9: Bar stretching under axial loadings

The system (5.8) becomes for this case
N (0) l2
0 1 1 l x p0 xl p0 3
EA = s dx = l2
l 1 u(l) EA 0 (l x)p0 xl EA 6
and its solutions are

p0 l 2
u(l) =
EA 3
l EA l l l
N (0) = p0 + u(l) = p0 + = p0
6 l 6 3 2
Inserting the prescribed boundary conditions and these boundary reaction in the solution
(5.7) gives
N (0) 1 l p x1
0
u( ) = + u(l) s jx j dx
2 EA 2 0 EA l 2
p0 l 1 p0 l 2 p0 x 1 l p0 x 1
= + s ( x)dx s (x )dx
2 2EA 2 EA 3 0 EA l 2 EA l 2
3
p0 l l2 1 3
= + (2l3 3 l2 + )
EA 4 6 12l 12l
3
p0
= 3 l
6EA l
Exercise 10: Torsional twist of an elastic bar

The adequate fundamental solution of this Helmholtz type equation is (3.39)
1 hr
# (x; ) = # (r) = e
2h
which has the …rst derivative
@# (x; ) 1 @r 1
= e hr = e hr sign(x )
@x 2 @x 2
Hence, substituting the corresponding fundamental solution terms and introducing the
0
actual state variables #(x) and # (x) in the integral form of the solution (3.53) yields
1 ) 0 1 1 1 l MT (x)
#( ) = e h(l
# (l) + e h(l )
#(l) e h
#0 (0) + e h
#(0) + s e hjx j
dx
2h 2 2h 2 0 ECT 2h
and, correspondingly, one obtains from (3.55)

2 3
#(0)
1 1 e hl e hl 6 6
1 0
h
# (0) 7
7= 1 l MT (x)e hx
hl hl 4 5 s hjx j dx
e e 1 1 #(l) ECT h 0 MT (x)e
1 0
h
# (l)
Exercise 11: Green’functions for beam problems

a) When the beam has a clamped support at both endings, the boundary conditions for
the Green’s function of this boundary value problem are
Gcc (0; ) = 0 and Gcc (l; ) = 0
@Gcc (x; ) @Gcc (x; )
= 0 and =0
@x x=0 @x x=l
With the polynomial ’ansatz’

1
Gcc (x; ) =
r 3 + c1 x3 + c2 x2 + c3 x + c4
12EI
one easily …nds the following conditions for the four free constants c1 ; c2 , c3 , and c4
(r = jx j, 0 l)
Gcc (0; ) = 0 = 3 + c4
Gcc (l; ) = 0 = (l )3 + c1 l3 + c2 l2 + c3 l + c4
@Gcc (x; )
= 0 = 3r2 sign(x ) x=0 + c3
@x x=0
= 3 2 + c3
@Gcc (x; )
= 0 = 3r2 sign(x ) x=l + 3c1 l2 + 2c2 l + c3
@x x=l
= 3(l )2 + 3c1 l2 + 2c2 l + c3
106 5 Appendices
Solving these four equations gives

" #
2 3 2
2 3
c1 = 1+6 4 ; c2 = 3 1 4 +2 c3 = 3 c4 =
l l l l
and, …nally, the following Green’s function

8 h i 9
2 3
1 < r 3
+ 1+6 4 x3 =
l l
Gcc (x; ) = h 2
i
12EI : + 1 4 +2 3 x + 3 2x
2 3 ;
l l
( )
2 3 2 3 x 3
1 r 3 x3 + 3 x +3 x 2
+ 6l 4 l
= 2
12EI + 2 2 4l 3 xl
b) For the Green’s function of a beam with a clamped support at x = 0 while the
other ending is free, one starts again with the polynomial ’ansatz’
1
Gcf (x; ) = r 3 + c1 x3 + c2 x2 + c3 x + c4
12EI
There, the following conditions hold
Gcf (0; ) = 0 = 3 + c4
@Gcf (x; )
= 0 = 3r2 sign(x ) x=0
+ c3
@x x=0
= 3 2 + c3
M (Gcf (x; )) x=l = 0 !
@ 2 Gcf (x; )
= 0 = 6rjx=l + 6c1 l + 2c2
@x2 x=l
= 6(l ) + 6c1 l + 2c2
cf
Q(G (x; )) x=l
= 0!
3 cf
@ G (x; )
= 0 = 6sign(x )jx=l + 6c1
@x3 x=l
= 6 + 6c1
One easily gets

c1 = 1; c2 = 3 ; c3 = 3 2 ; c4 = 3
and hence, the following Green’s function
1
Gcf (x; ) = r3 x3 + 3 x2 + 3 2 x 3
12EI
Exercise 12: Strain tensor of the elastostatic fundamental solution

Since di¤erentiating the distance r = jx j with respect to the coordinates xi gives in
2 3
R and in R
@r xi i @2r 1
= r;i = ; = r;ij = ( ij r;i r;j )
@xi r @xi @xj r
it is easy to determine
(k)
@ui (x; ) 1 1 1
= (3 + ) ik r;j +(1 + )( ij r;k + kj r;i 3r;i r;j r;k ) in R3
@xj 8 2 + r2
and then via the de…nition of the strain tensor
(k)
!
(k)
(k) 1 @u i (x; ) @u j (x; )
"ij (x; ) = +
2 @xj @xi
1 1 1
= ( ik r;j + jk r;i ) + (1 + )( ij r;k 3r;i r;j r;k ) in R3
8 2 + r2
Exercise 13: Shape functions with the local coordinate 1 1

Since the restrictions (4.2) hold, the linear shape functions
N12 ( ) = a1 + b1 ; N22 ( ) = a2 + b2
are de…ned by the two conditions for each of the shape functions
N12 ( 1) = a1 + b1 = 1; N12 (+1) = a1 + b1 = 0

N22 ( 1) = a2 + b2 = 0; N22 (+1) = a2 + b2 = 1
This gives easily for 1 1

1 1
N12 ( ) = (1 ); N22 ( ) = (1 + )
2 2
For the three quadratic shape functions
N13 ( ) = a1 2
+ b1 + c1 ; N23 ( ) = a2 2
+ b2 + c2 ; N33 ( ) = a3 2
+ b3 + c 3
the restrictions (4.2) give three conditions for each of them
N13 ( 1) = a1 b1 + c1 = 1; N13 (0) = c1 = 0; N13 (1) = a1 + b1 + c1 = 0

N23 ( 1) = a2 b2 + c2 = 0; N23 (0) = c2 = 1; N23 (1) = a2 + b2 + c2 = 0
N33 ( 1) = a3 b3 + c3 = 0; N33 (0) = c3 = 0; N33 (1) = a3 + b3 + c3 = 1
and one obtains for 1 1

1 1
N13 ( ) = ( 1); N23 ( ) = 1 2
; N33 ( ) = ( + 1)
2 2
108 5 Appendices
Exercise 14: Stationary heat conduction in a rectangular domain- constant

shape functions
For the simplest discretization of the boundary, i.e., each side of the quadrilateral domain
as an element, collocations at the middle nodes of these four elements, at x1 = (l1 =2; 0),
x2 = (l1 ; l2 =2), x3 = (l1 =2; l2 ), and x4 = (0; l2 =2) yield from (4.55), (4.56), and (4.57),
respectively, the following matrix elements of the algebraic systemrixelemente (4.49):
G11 = G22 = G33 = G44 = 0

1
G12 = G14 = G32 = G34 = arctan(4) 0:2110
2
1 1
G13 = G31 = arctan( ) 0:07798
4
G21 = G23 = G41 = G43 = 0:125 und G24 = G42 = 0:25
1 1 1 0:269473
H11 = H33 = H22 = H44 = (1 + ln(2))
2 2 2 0 0
1 17 0:203156
H12 = H14 = H32 = H34 = 4 2 ln( ) arctan(4)
4 0 16 0
1 17 1 0:1119454
H13 = H31 = 2 ln( ) 8 arctan( )
4 0 4 4 0
1 n o 0:0210
H21 = H23 = H41 = H43 = 2 ln(2)
4 0 2 0
1 0:178627
H24 = H42 = f4 + 2 ln(2) g
4 0 0
Taking the boundary conditions 1 = 0, 3 = ; and qn2 = qn4 = 0 into account, and
reordering the system with respect to known and unknown node values, gives the system
2 32 2 3 2 3
G12 G12 H11 H13 G13
6 0:5 G24 H21 H21 7 6 4 7 6 7
6 7 6 1 7 = 6 G23 7
4 G12 G12 H13 H11 5 4 qn 5 4 0:5 5
3
G24 0:5 H21 H21 qn G23
2 4
Subtraction of the forth from the second equationt gives = and then by adding
these two equations
2 4 G23 H21 1
= = + (qn1 + qn3 ) = + 4H21 (qn1 + qn3 )
0:5 + G24 0:5 + G24 2
Finally, subtracting and adding the …rst and third equation and taking the above result
for 2 = 4 into account gives
0:5 G13 0:5 + G13 + 2G12
qn1 qn3 = and qn1 + qn3 =
H13 H11 H11 + H13 16G12 H12
Since G13 + 2G12 = 1 arctan( 14 ) 2 21 arctan(4) = 1 (arctan( 14 ) + arctan(4)) = 1

2
=
0:5, the nominator 0:5 + G13 + 2G12 = 0 and, hence, qn1 = qn3 . This yields
2 4 1
= =
2
0:5 G13 0:5 G13
qn1 qn3 = 2qn1 = =) qn1 = qn3 = 0:757 0
H13 H11 2(H13 H11 )
The exact solution gives for these node points the same value for the temperature (l1 ; l2 =2)
= (0; l2 =2) = 0:5 , while with n2 (l1 =2; 0) = 1, n2 (l1 =2; 0) = 1, and l2 = 2m the heat
‡ux is there exactly qn1 = qn (l1 =2; 0) = 0 =l2 = 0:5 0 = qn3 = qn (l1 =2; l2 ) which
means about 50% error. This not at all astonishing since physically wrong elementwise
constant approximations have been used to model the correctly continous temperature
…eld.
Exercise 15: Stationary heat conduction in a rectangular domain- linear shape

functions
When applying the simplest possible discretization of the boundary, i.e., taking each of the
4 sides of the rectangular domain as one boundary element, using linear shape functions
for the temperature and the heat ‡ux qn , and performing collocation at the nodes
x = (0; 0), x = (l1 ; 0), x3 = (l1 ; l2 ), and x4 = (0; l2 ), (here, the 4 corners of the domain),
1 2
one obtains the system
X
4
1 X
4
e 2 2e 1 1e
je + Gje = Hje qn + Hje qn (5.9)
e=1
4 e=1
where here, since all node points are corner points, qn2e 6= qn1e+1 for all 4 elemente e
.
From (4.69) and (4.70), respectively, result the following matrix elements
G11 = G22 = G33 = G44 = 0

1
G21 = G12 = G34 = G43 = [ln(5) 4 arctan(2)]
8
1
G31 = G13 = G24 = G42 = [4 ln(4) 5 ln(5)]
8
1 5 1
G41 = G14 = G23 = G32 = ln( ) arctan( )
2 4 2
1 2
while from (4.71) and (4.72), respectively, the elements of Hje and Hje can be evaluated
j
corresponding to the location of the collocation point xi from (4.73) to (4.76) and (4.78)
110 5 Appendices
to (4.77),respectively: (l1 = l3 = l = 1m; l2 = l4 = 2l = 2m)
1 1 3l1 1 1 3l2
H11 = H33 = ; H22 = H44 =
8 0 4 0
1 1 l1 1 1 l2
H21 = H43 = ; H12 = H34 = [3 ln(5) 4 ln(4) 12 + 8 arctan(2)]
8 0 32 0
1 1 l1 1 1 l2
H31 = H13 = [5 ln(5) 4 ln(4) 1] ; H24 = H42 = [5 ln(5) 4 ln(4) 4]
8 0 32 0
1 1 l1 1 1 1 l2
H41 = H23 = 4 ln(4) 3 ln(5) 3 + 8 arctan( ) ; H14 = H32 =
8 0 2 8 0
2 2 1 2 2 1
H11 = H33 = H21 ; H22 = H44 = H14
2 2 1 2 2 1
H21 = H43 = H11 ; H12 = H34 = H24
2 2 1 2 2 1
H31 = H13 = H41 ; H24 = H42 = H34
2 2 1 2 2 1
H41 = H23 = H13 ; H14 = H32 = H22
With that, and using the abbreviations h1 = 2 ln(4) + 2 52 ln(5); h2 = 4 ln(4) + 3 ln(5) +
3 8 arctan( 21 ), h3 = 4 ln(4) + 1 5 ln(5); h4 = 32 ln(5) + 2 ln(4) + 6 4 arctan(2) and
g1 = 14 ln(5) arctan(2), g2 = ln(4) 54 ln(5); g3 = ln( 54 ) arctan( 21 ) the system (5.9)
becomes
02 3 2 21 3 2 3 2 11 31
1 h1 h2 6 qn 3 h4 h3 2 qn
l B6 3 2 h3 h4 7 6 qn 7 6 1 6 h2 h1 7 6 qn12 7
B 6 7 6 22 7 6 7 6 C
7C =
+
8 0 @4 h2 6 1 h1 5 4 qn 5 4 h3 2 3 h4 5 4 qn 5A
23 13
h3 h4 3 2 qn24 h2 h1 1 6 qn14
2 32 1
3
2
g1 g2 g3
1 6
6 g1 2
g3 g2 7
76
6 2 7
7
=
2 4 g2 g3 2 g1 5 4 3 5
4
g3 g2 g1 2
Now, incorporating the actual boundary conditions of the problem, i.e.,
1 2 3 4
= = 0 and = =
and
qn22 = qn24 = qn12 = qn14 = 0
reduces the system to:
2 3 2 11 3 2 3
3 1 h3 h2 qn g2 + g3
l 6 1 3 h2 7 6
h3 7 6 qn21 7 6 g2 + g3 7
6 7= 6 7
8 0
4 h3 h2 3 1 5 4 qn13 5 2 4 + g1 5
2
h2 h3 1 3 qn23 2
+ g1
Subtraction of the second from the …rst and of the forth from the third equation yield
qn11 = qn21 and qn13 = qn23 , so that only two equations remain
1 1 1
qn11 + 1 ln(5) 2 arctan( ) qn13 = 0 ln(5) + 4 arctan( )
2 2 4l 2
1 1
1 ln(5) 2 arctan( ) qn11 + qn13 = 0 [2 + ln(5) 4 arctan(2)]
2 2 4l
>From that, one obtains with arctan( 12 ) + arctan(2) = 2

the exact boundary values for
the heat ‡ux
qn13 = qn23 = 11 21
0 and qn = qn = 0
2l 2l
The temperature distribution is by the linear shape function on the element 2 ,i.e., on x1
= x12 22 12
1 = x1 = l1 , 0 = x2 x2 x22 2 = l2 , with = x2 =l2 and the boundary conditions
12 22
= 0 and = exactly described
x2
(x2i ( )) = 12
(1 )+ 22
=
l2
x2 l 2
This is also true on 4 , i.e., x1 = x14 24 24
1 = x1 = 0, 0 = x2 x2 x14
2 = l2 , with = l2
and the boundary conditions 24 = 0 and 14 = by
x2 l2 x2
(x4i ( )) = 14
(1 )+ 24
= (1 )= :
l2 l2
From (4.66), the temperature at internal points xj is in this case , i.e., with the prescribed
boundary values 1 = 11 = 24 = 0; 2 = 12 = 21 = 0 , 3 = 22 = 13 = ;
4 23 14 22 24 12 14
= = = , and qn = qn = qn = qn = 0 and the just determined boundary
reactions qn = qn = 2l 0 and qn11 = qn21 = 2l 0 given by
13 23
1 (x3 ( ); xj )(1 )l3 (x1 ( ); xj )(1 )l1

(xj ) = 0 s 3 j 1 j d
2l =0 + (x ( ); x ) l3 (x ( ); x ) l1
1 [qn (x3 ( ); xj )l3 + qn (x4 ( ); xj )l4 ] (1 )
s d
=0 + [qn (x2 ( ); xj )l2 + qn (x3 ( ); xj )l3 ]
1
= 0 s (x3 ( ); xj )l3 (x1 ( ); xj )l1 d
2l =0
1 [qn (x3 ( ); xj )l3 + qn (x4 ( ); xj )l4 ]
s d (5.10)
=0 + [qn (x2 ( ); xj )l2 qn (x4 ( ); xj )l4 ]
where as in (4.57)
" jx1e xj j2
#1
1 1 (le + rej te ) ln 2
+ l2e rej te + le2
s (xe ( ); xj )le d = le +rej te
=0 4 0 2le + 2 j rej ne j arctan jrej ne j =0
(5.11)
112 5 Appendices
and rej =j x1e xj j means the distance between the observation point xj and the
initial point x1e of the element e . With the angle ej = ^(rej ; te ), between rej and the
tangential unit vector te along this element holds
(x1e
i xji )tei = rej cos ej ; nei (x1e
i xji ) = rej sin ej
le2 + 2
rej 2
2le rej cos ej = re+1;j (5.12)
where for rectangular domains, as here, is especially
le rej cos ej = re+1;j sin e+1;j

re+1;j cos e+1;j = rej sin ej für e = 1; 2; 3; 4 mit e = 5 , 1 (5.13)
With that, the following integral can be evaluated as

2 2 2
3
re+1;j re;j
1 1 r e+1;j sin e+1;j ln le2
2le + rej cos ej ln le2
s (xe ( ); xj )le d = 4 5
4 0 +2rej sin ej arctan e+1;j r sin e+1;j
=0
rej sin ej
+ arctan(cot ej )
" r2 2
re;j
#
1 re+1;j sin e+1;j ln e+1;j
le2
2le + rej cos ej ln le2
=
4 0 +2rej sin ej e+1;j + ej 2
From (4.56) is
1 1 le + te rej te rej
s qn (xe ( ); xj )le d = sign ne rej arctan arctan
=0 2 j ne rej j j ne rej j
1 re+1;j sin e+1;j
= sign ne rej arctan arctan(cot ej )
2 rej sin ej
and, especially for the actual rectangular domain with ej 0:5
1
e j 1n o
s qn (x ( ); x )le d = e+1;j + ej
=0 2 2
and, moreover, from (4.69)
8 2
9
n r < =
e ej
1 e ej 1
2
ln 1 + 2 jxle1et rxj j2 + jx1e le xj j2
s qn (xe ( ); xj ) le d =
=0 2 le : te rej e ej
arctan lejn+te rejr j
e ej
arctan jnt e rrej j ;
jne rej j
8 9
r2
>
> 1
r sin ej ln e+1;j + >
>
1 < 2 ej 2
re;j
! =
= r
arctan e+1;j
sin e+1;j
2 le > e ej
> sign (n r ) rej cos ej r ej sin ej >
>
: + ;
2 ej
and especially for rectangular domains

2
1 1 1 re+1;j
s qn (xe ( ); xj ) le d = rej sin ej ln 2
+ rej cos ej e+1;j + ej
=0 2 le 2 re;j 2
5.2 B: Analytic integration of singular boundary integrals 113
Finally, one obtains for (5.10) in the actual rectangular domain with l1 = l3 = l and
l2 = l4 = 2l
8 " 2 2
# 9
r4j r3j
>
> r sin ln 2l + r cos ln >
>
>
>
4j 4j l 2 3j 3j l 2 >
>
< =
j " +2r 3j sin 3j 4j + 2 3j #
(x ) = r 2 r 2
8 l >> r2j sin 2j ln l2j2 2l + r1j cos 1j ln l1j2 >
>
>
> >
>
: ;
+2r1j sin 1j 2j + 2 1j
8 9
>
> 4j + 2 3j + 1j + 2 >
>
< 1 h1 i
4j
2
r3j =
+ + 2l 2 r2j sin 2j ln r2 + r2j cos 2j 3j + 2 2j
2 > h 2j i >
>
: 1 1 r4j sin 4j ln 1j r 2
>
;
2l 2 2 +
r4j
r 4j cos 4j 1j + 2 4j
Taking into account that xj2 = r1j sin 1j and 2l xj2 = r3j sin 3j as well as (5.13) hold,
this is reduced to
(xj ) = xj2
2l
the exact solution for internal points.
5.2 B: Analytic integration of singular boundary in-

tegrals
5.2.1 Analytic integration in the case of logarithmic kernels
In R2 , kernels containing ln(r=c) are weakly singular and can be represented on a straight
boundary element modelled by the linear approximation (4.1) xei ( ) = x1e i (1 ) + x2e
i
as
r 1 h 2 2
i
ln( ) = ln x1e 1 (1 ) + x 2e
1 1 + x 1e
2 (1 ) + x 2e
2 2 ln(c) (5.14)
c 2
For an approximation of the boundary state (x) in e by a constant value (xei ( )) = e
and a collocation at its respective node, i.e., at the middle of the element i = 21 (x1e 2e
i +xi ),
an integral with a logarithmically singular kernel over this element can be evaluated
exactly
r(x; ) e le 1 1
s ln( ) (x)d x = s ln ( )2 (x2e
1 x1e 2
1 ) + (x2
2e
x1e
2 )
2
d le ln(c)
e c 2 =0 2
e le 1
1
= s ln( )2 + ln le2 d le ln(c)
2 =0 2
e le 1
= 2(ln 1) + 2 ln le le ln(c)
2 2
e le
= le ln( ) 1 (5.15)
2c
e
= le (ln 2 + 1) for c = le (see the Remark below) (5.16)
114 5 Appendices
Remark: The argument of the ln-function has to be dimensionless and, hence, the constant
factor c in the above integral kernel has to represent a distance, but can be arbitrarily chosen.
le
In order to avoid that for smaller elements, i.e., le ! 0, the term ln( 2c ) becomes dominant due
le
to limle !0 ln( 2c ) ! 1, the factor c should be chosen problem orientated, e,g., as c = le .
When the state function is linearly approximated
(xei ( )) = 1e
(1 )+ 2e
and the collocation point is placed at the initial node of the element, i.e., i = x1e i ,
the square of the distance between this collocation point and the integration points is
r2 = le2 2 . Then, the integral with a logarithmic kernel becomes (c is an arbitrary constant
distance, e.g., c = le )
r(x; ) 1 le 1e 2e
s ln( ) (x)d x = le s ln + ln( ) (1 )+ d
e c =0 c
le 1e le 2e le
= (2 ln( ) 3) + (2 ln( ) 1)
4 c c
le 1e 2e
= 3 + for c = le (5.17)
4
while for a collocation at the end node of the element, i.e., at i = x2e
i , the square of the
distance between this collocation point and the integration points is r2 = le2 (1 )2 and
the integral becomes
r(x; ) 1 le 1e 2e
s ln( ) (x)d x = le s ln(1 ) + ln( ) (1 )+ d
e c =0 c
le 1e le 2e le
= 2 ln( ) 1 + 2 ln( ) 3
4 c c
le 1e 2e
= +3 for c = le (5.18)
4
In the case of a quadratic approximation of the state function
(xei ( )) = 1e
(1 3 + 2 2) + 2e
(4 4 2) + 3e
(2 2
)
and collocation at the initial node of the element, i.e., at i = x1e i , the square of the
distance between this collocation point and the integration points is r2 = le2 2 and the
integral becomes
8 2 1e 39
< (1 3 + 2 2 )+ =
r(x; ) 1 le
s ln( ) (x)d x = le s ln + ln( ) 4 2e
(4 4 2) 5 d
c =0 : c 3e 2 ;
e
+ (2 )
le 1e le 17 le 10 le 1
= (ln( ) ) + 2e (4 ln( ) ) + 3e (ln( ) + )
6 c 6 c 3 c 6
le
= 17 1e + 20 2e 3e
for c = le (5.19)
36
2 1
while for collocation at the middle point, i.e., at i = x2e 2
i gives r = le ( 2 )2 and
8 2 1e 39
< (1 3 + 2 2 )+ =
r(x; ) le 1 1 le
s ln( ) (x)d x = s ln( )2 + 2 ln( ) 4 2e
(4 4 2) 5 d
c 2 =0 : 2 c ;
e
+ 3e (2 2 )
le 1e le 1 le 4
= + 3e (ln( ) ) + 4 2e (ln ) (5.20)
6 2c 3 2c 3
le 1e 1 4
= + 3e (ln 2 + ) + 4 2e (ln 2 + ) for c = le
6 3 3
and for collokation at the end point of the element, i.e., at i = x3e 2 2
i gives r = le (1 )2
and results in
8 2 1e 2
39
r(x; ) 1 < le
(1 3 + 2 )+ =
s ln( ) (x)d x = le s ln(1 ) + ln( ) 4 2e
(4 4 2) 5 d
c =0 : c ;
e
+ 3e (2 2 )
le 1e le 1 le 10 le 17
= ln( ) + + 2e 4 ln( ) + 3e ln( )
6 c 6 c 3 c 6
le 1e
= + 20 2e + 17 3e for c = le (5.21)
36
5.2.2 Analytic integration in the case of (1/r)- kernels

In many cases, integral kernels contain r;s =r which are strongly singular and, hence, the
respective boundary integral has to be evaluated in the Cauchy principal value sense.
For a linear geometric approximation (4.1), i.e., for a straight boundary element with
e
xi ( ) = x1e
i (1 ) + x2e
i , the tangential unit vector is given as
dxei dxe dxe dxe x2e x1e

ti = = j i j= i = j i j= i i
(5.22)
ds ds d d le
so that the distance between a collocation point and the integration point on the element
q
2 2
r = [x1e1 (1 ) + x2e
1
1e
1 ] + [x2 (1 ) + x2e
2 2] (5.23)
is found to have as tangential derivative
[x1e
1 (1 ) + x2e
1
1e
1 ; x2 (1 ) + x2e
2 2] 1 x2e
1 x1e
1
r;s = r;i ti = q
2 2 le x2e
2 x1e
2
[x1e
1 (1 ) + x2e
1 1] + [x1e
2 (1 ) + x2e
2 2]
(5.24)
This allows to represent the singular kernel as
r;s [x1e (1 ) + x2e 1e

1 ; x2 (1 ) + x2e 2] 1 x2e x1e
= 1e 1 1 2 1 1
(5.25)
r [x1 (1 ) + x2e
1
2
1 ] + [x2 (1
1e
) + x2e
2 2]
2
le x2e
2 x1e
2
116 5 Appendices
and gives with a constant shape function for the state function and, accordingly, a col-
location at the middle node of the element, i.e., at i = 21 (x1e 2e
i + xi ), as distance (5.23)
between the collocation point and the integration points r = le j 12 j. Finally, the
strongly singular boundary integral can be evaluated to
1
!
1
r;s e
2
" 1 [x1e
1 x2e
1 ; x2
1e
x2e
2 ] (2 ) 1 x2e 1 x1e
1
s (x)d x = le lim s + s 2 j 1 2e 1e d
e r "!0 =0 1
= 2 +" le 2
j 2 le x 2 x 2
1
!
"
e
2 1 1
= lim s + s 1d
"!0 =0 = 12 +" 2
e 1 1
= lim ln " ln + ln ln " =0 (5.26)
"!0 2 2
For a linear approximation of the state function
(xei ( )) = 1e
(1 )+ 2e
and collocation at the initial node of the element, i.e., at i = x1ei one obtains for the
distance between the integration points on the element e and that collocation point
re = le . At the same time, that collocation point also coincides with the end point
of the neighbour element e 1 , i.e., i = x2ei
1
, which gives for the respective distance
re 1 = le 1 (1 ). In this case, the Cauchy principal value has to be evaluated with
lim"!0 by considering both neighbour elements e 1 and e (regarding that 2e 1 = 1e ):
1 " x1e
[1 1
x12e 1
;x1e
2
1
2 ]
x2e x2e
1
1
x1e
1
1 1e 1
(1 )
s d
r;s =0
le2 1 (1 ) x2e
2
1
x1e
2
1
+ 2e 1
s +s (x)d x =
r 1
[x2e
1 x1e 2e
1 ;x2 2 ]
x1e x2e
1 x1e
1
1e
(1 )
e 1 e
+s d
="
2
le 2
x2e
2 x1e
2 + 2e
1 " 1 1e 1 2e 1
= s (1 )+ d
=0 1
1 1 1e 2e
+s (1 )+ d
="
1e 1 2e 1
1 "
1e
1 (1 ) 2e
= s d + s d +
=0 1 ="
0
2e 1e 1 2e 1
" (1 ) 1 (1 )
= + s 0
d 0+ 1e
s d
0 =1 ="
2e 1e 1
= (5.27)
The unit normal vector is for straight boundary elements
1 x2e
2 + x2
1e
n= (5.28)
le x2e
1 x1e
1
from which due to the linear approximation of this element at all collocation points, e.g.,
at the initial point i = x1e
i
[x1e
1 x2e 1e
1 ; x2 x2e
2 ] 1 x2e
2 + x2
1e
r;n = r;i ni = =0 (5.29)
le le x2e
1
1e
x1
at the middle point i = 21 (x1e 2e

i + xi )
1
[x1e
1 x2e 1e
1 ; x2 x2e
2 ] (2 ) 1 x2e
2 + x2
1e
r;n = r;i ni = =0 (5.30)
le j 12 j le x2e
1 x1e
1
and also at the end point i = x2e

i
[x1e
1 x2e 1e
1 ; x2 x2e
2 ] (1 ) 1 x2e
2 + x2
1e
r;n = r;i ni = =0 (5.31)
le j 1 j le x2e
1 x1e
1
follows that the normal derivative r;n = 0 on straight elements for all types of shape
functions, i.e., an integral containing r;n gives zero on straight boundary parts.
Bibliography
[1] Bonnet, M.: Boundary Integral Equation Methods for Solids and Fluids, John Wiley
& Sons, Chicester 1995.
[2] Gaul, L. and Fiedler, Ch.: Methode der Randelemente in Statik und Dynamik, Vieweg,
Braunscheig 1997.
[3] Roach, G.F.: Green’s Functions, Cambridge University Press, New York 1995.
[4] Prem K. Kythe: Fundamental solutions for di¤erential operators and applications,
Birkhäuser, Boston, Basel, Berlin 1996.
[5] Prem K. Kythe: An Introduction to Boundary Element Methods, CRC Press, Boca
Raton, London 1995.
[6] Paris, F. and Canas, J.: Boundary Element Method - Fundamentals and Applications,
Oxford Science Publications, Oxford 1997
[7] Hörmander, L.: Linear Partial Di¤erential Operators, Springer, Berlin 1963.
[8] Gipson, G. St: Boundary Element Fundamentals, Topics in Engineering Vol. 2, Com-
putational Mechanics Publ., Southampton
[9] Kane, J.H.: Boundary Element Analysis in Engineering Continuum Mechanics, Pren-
tice Hall, Englewood Cli¤s, 1994.
118

(Heinz Antes) A Short Course On Boundary Element Methods

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Heinz Antes) A Short Course On Boundary Element Methods

Uploaded by

Copyright:

Available Formats

A Short Course on Boundary Element Methods

Prof. Dr. rer. nat. Heinz Antes

Institut für Angewandte Mechanik

March 29, 2005

3 Transformation of Di¤erential Equations to Integral Equations 32

4 Numerical solution of boundary integral equations: The boundary ele-

Advantages of the Boundary Element Method

Disadvantages of the BE method

2.1 Some notations and de…nitions

2.1.1 Indicial and symbolic notation

Exercise 1: Nabla vector

2.1.2 Contraction and di¤erent products of tensors

Indicial Notation Symbolic Notation

ij nk = sijk ; ij "kl = Eijkl

Tii = T11 + T22 + T33 = a1 b1 + a2 b2 + a3 b3 = ai bi

Example: The directional derivative of a scalar function f (x1 ; x2 ; x3 ) in the direction of

Exercise 2: Laplace operator

if the values of i; j; k are an even permutation of 1; 2; 3

2.1.3 The Euclidian distance r and its derivatives

where ij is the so-called Kronecker delta meaning

The …rst derivative of r with respect to xj follows from (2.12) to be

2.2 The Gauss theorems

2.2.1 The gradient theorem

2.2.2 The divergence theorem

2.2.3 Generalized Gauss theorems

exactly as one knows it from basic calculus.

2.3 Integration by parts - Green’s identities

the relation (2.29) is known as Green’s second identity.

exactly what one knows from basic calculus.

Exercise 3: Integration by parts

2.4 Fundamental solutions of di¤erential equations

@ jkj u(x1 ; :::; xn )

2.4.1 Adjoint and self-adjoint operators

If ak (x) = ak are constant, then L (D) = L( D). An operator is said to be self-adjoint

L(D)u(x) = f (x) in Rn ; (2.38)

are more speci…c de…nition of L can be formulated:

Here, again, the operator L is adjoint to L, and, if L = L, then L is said to be self-

L(D)u (x; ) = (x; ) (2.42)

L(D) (u (x; ) + w (x; )) = L(D)u (x; ) + L(D)w (x; ) = (x; ) (2.43)

2.4.2 The Dirac -function

x1 = u(r; ') = r cos '; x2 = v(r; ') = r sin '

and the Jacobian J of the transformation is J = r which yields

Hence, if a partial di¤erential equation is formulated in polar co-ordinates, the adequate

The -function in R1 and the Heaviside function

Similarly, one can state

where f (x) is an arbitrary continuous function and g(x) is a monotonic function of x

As a special case of (2.54) one obtains

And, again by integration by parts, one obtains

H(x ) = 1 for x > (2.57)

To see this, we integrate by parts (a < < b)

which yields by comparison with (2.51)

The -function in Rn with n 2

(x ) = (x1 1) (x2 2) (x3 3 ): (2.59)

where u and v are single-valued, continously di¤erentable functions of their arguments.

where J = @(u; v)=@( 1 ; 2 ) is the Jacobian of (2.60).

[u( 1 ; 2) 1 )] [v( 1 ; 2) 2 )] jJj = ( 1 1) ( 2 2)

Example: Transformation from rectangular Cartesian co-ordinates x; y to plane polar

one obtains the relation

2.4.3 Green’s functions of boundary value problems

E(u(x)) = u(x) for x 2 1 (2.63)

2.4.4 Ordinary di¤erential equations with constant coe¢ cients

Since, in view of (2.58) and (2.53),