Engineering Analysis with Boundary Elements 57 (2015) 66–71

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements

journal homepage: www.elsevier.com/locate/enganabound

A regularized multi-level technique for solving potential problems

by the method of fundamental solutions
Csaba Gáspár
Széchenyi István University, Egyetem tér 1, H-9026 Győr, Hungary

art ic l e i nf o a b s t r a c t

Article history: The method of fundamental solutions is investigated in the case when the source points are located
Received 6 October 2013 along the boundary of the domain of the original problem and coincide with the collocation points. The
Accepted 7 May 2014 appearing singularities are eliminated by several techniques: by using approximate but continuous
Available online 6 June 2014
fundamental solutions (regularization) and via auxiliary subproblems to avoid the stronger singularities
Keywords: that appear in the normal derivatives of the fundamental solution (desingularization). Both monopole
Meshless method and dipole formulations are investigated. A special iterative solution algorithm is presented, which
Method of fundamental solutions converts the original (mixed) problem to a sequence of pure Dirichlet and pure Neumann subproblems.
Regularization The pure subproblems can be handled efficiently by using conjugate gradients. The efficiency is
Desingularization
significantly increased by embedding the resulting method in a natural multi-level context. At the same
Multi-level method
time, the problem of the use of highly ill-conditioned matrices is also avoided.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction defined in a sufficiently smooth domain Ω supplied with mixed

boundary conditions:
Elliptic partial differential equations play an essential role in a

∂u


lot of fields of application. Modelling stationary phenomena such ujΓ D ¼ u0 ; ¼ v0 ; ð2Þ
∂n
ΓN
as diffusion in fluids or in gases, heat transfer in machines e.g. in
parts of traditional or hybrid cars, seepage through porous media where Γ≔∂Ω denotes the boundary of Ω, which has a disjoint
lead to solving elliptic problems. The usual implicit time discreti- decomposition into a Dirichlet part Γ D and a Neumann part Γ N .
zation techniques of time-dependent problems result in elliptic In its traditional form, the MFS produces an approximate
problems as well, at every time step. solution of the problem (1) and (2) in the following form:
To handle elliptic problems in a meshless way, the popular N
Boundary Element Method is not suitable, since it requires a uN ðxÞ ¼ ∑ αj Φðx  x~ j Þ; ð3Þ
j¼1
boundary mesh structure. The strength of the meshless methods is
to circumvent the generation of both domain and boundary mesh. So where Φ is a fundamental solution of the operator L, i.e. ΔΦ ¼ δ
far, a number of boundary meshless methods have been developed e. (here δ denotes the Dirac distribution concentrated at the origin).
g. the boundary knot method [3,4] which uses nonsingular general The predefined points x~ 1 ; …; x~ N (the source points) are located
solutions, or the method of fundamental solutions (MFS, see e.g. [1]), outside of the domain Ω. The a priori unknown coefficients
which is based on the fundamental solution of the applied partial α1 ; …; αN can be computed by enforcing the boundary conditions:
differential operator, i.e. on solutions with singularities. In this paper,
N
we restrict ourselves to the MFS applied to homogeneous problem. ∑ αj Φðxk  x~ j Þ ¼ u0 ðxk Þ ðxk A Γ D Þ;
(For non-homogeneous problems, the approach can be combined j¼1
ð4Þ
with the well-known principle of the Method of Particular Solutions). N ∂Φ
∑ αj ðx  x~ j Þ ¼ v0 ðxk Þ ðxk A Γ N Þ;
Consider a second-order elliptic homogeneous linear partial j¼1 ∂nk k
differential equation:
where x1 ; …; xN A Γ are predefined boundary collocation points.
Lu ¼ 0 in Ω ð1Þ Then the function uN exactly satisfies Eq. (1), and exhibits
singularities at the source points.
Instead of the formulation (3) (called monopole formulation
E-mail address: [email protected] hereafter), it is often more advantageous to use the dipole

http://dx.doi.org/10.1016/j.enganabound.2014.05.002
0955-7997/& 2015 Elsevier Ltd. All rights reserved.
 C. Gáspár / Engineering Analysis with Boundary Elements 57 (2015) 66–71 67

formulation, where the approximate solution of (1) and (2) is where the coefficients α1 ; …; αN solve the algebraic system:
sought in the following form: N
∑ αj C kj ¼ u0 ðxk Þ ðxk A Γ D Þ;
N∂Φ
uN ðxÞ ¼ ∑ αj ðx  x~ j Þ: ð5Þ j¼1
ð10Þ
j¼1 ∂n j N
∑ αj Q kj ¼ v0 ðxk Þ ðxk A Γ N Þ:
Again, Φ is the fundamental solution of the operator L, and the j¼1

source points x~ 1 ; …; x~ N are located outside of Ω. Now the coeffi-

cients α1 ; …; αN can be computed by solving the linear system: Here the entries of the matrices A, B, C, Q are defined as follows:
∂Φ
N ∂Φ Akj ¼ Φðxk xj Þ; Bkj ¼ ðx xj Þ;
∑ αj ðx  x~ j Þ ¼ u0 ðxk Þ ðxk A Γ D Þ; ∂nk k
j¼1 ∂nj k
ð6Þ ∂Φ ∂2 Φ
N ∂2 Φ C kj ¼ ðx  xj Þ; Q kj ¼ ðx  xj Þ ð11Þ
∑ αj ðxk  x~ j Þ ¼ v0 ðxk Þ ðxk A Γ N Þ: ∂nj k ∂nk ∂nj k
j ¼ 1 ∂nk ∂nj
for j a k. Due to the singularity of the fundamental solution at the
Note that the monopole and the dipole formulations can be origin, the diagonal entries cannot be computed by the above
considered as meshless discretizations of the indirect BEM based definition.
on single layer and double layer potentials, respectively. For the proper definition of Akk, one should replace the funda-
A common disadvantage of the traditional forms (3) and (5) is mental solution Φ with an approximate fundamental solution Φ,
the use of external source points, the location of which can be which has no singularity at the origin. Such an approximate funda-
hardly automatized. mental solution can be defined e.g. by truncation. In polar coordi-
Though the MFS has excellent accuracy properties (see e.g. [10] nates:
and references therein), the systems (4) and (6) are highly ill- 8
>
> 1
conditioned in general, which is a severe drawback of the method. >
< ΦðrÞ if r Z
  c
This is the case especially when the source points are located ΦðrÞ≔ 1 1 ð12Þ
>
> Φ o ;
far from the boundary. On the other hand, if they are close to the >
: if r
c c
boundary, the systems (4) and (6) become much better condi-
tioned, however, the accuracy goes wrong due to the appearance provided that Φ is a radial function i.e. it depends only on r ¼ jjxjj,
of numerical singularities at the boundary collocation points. which is often the case. Here c denotes a carefully chosen scaling
A usual technique is to define the source and the boundary constant which should remain inversely proportional to the char-
collocation points to coincide. Special techniques are required to acteristic distance of the boundary collocation points, when N varies.
avoid the problem of singularity (regularization and desingular- Another regularization technique is to replace Φ with the funda-
ization, see e.g. [14,7,12,8]). mental solution of the singularly perturbed fourth-order operator
In this paper, we investigate some regularized versions of both LðI  ð1=c2 ÞLÞ, where I denotes the identity operator and c is again
the monopole and the dipole formulations. It turns out that, from a a scaling constant, see [8] for details. Thus, the diagonal terms Akk
computational point of view, the dipole formulation is much more can be computed without difficulty. Using the simplest truncation,
advantageous for handling pure Dirichlet problems, while for pure Akk ¼ Φð0Þ, while for ja k, Akj ¼ Φðxk  xj Þ.
Neumann problems, the monopole formulation overperforms The proper definition of the diagonal terms Bkk is somewhat
the dipole formulation. For mixed boundary conditions, a special more difficult since the derivatives of Φ have stronger singularities
iterative technique is proposed which converts the original mixed at the origin than Φ itself. Let w be a smooth, easily computable
problem to a convergent sequence of pure Dirichlet and pure particular solution of Eq. (1), then w can be approximated by the
Neumann subproblems. This results in a computationally efficient monopole formulation:
method, which avoids also the problem of highly ill-conditioned
N
linear systems. The efficiency can be increased further by embed- wN ðxÞ≔ ∑ βj Φðx  xj Þ;
ding the method in a natural multi-level context. j¼1

from where the coefficients β1 ; …; βN can be computed by solving (1)

supplied with a pure Dirichlet condition. Computing the normal
2. Regularization and desingularization
derivative of the particular solution w, we have the following:

In the rest of the paper, suppose that the source points and the ∂wN N ∂Φ
ðxÞ ¼ ∑ βj ðx  xj Þ:
boundary collocation points x1 ; …; xN A Γ coincide. Then the mono- ∂n j¼1 ∂n
pole formulation has the form
Hence
N
uN ðxÞ ¼ ∑ αj Φðx  xj Þ; ð7Þ ∂w
∑ βj Bkj þ βk Bkk ¼ ðx Þ ðk ¼ 1; 2; …; NÞ:
j¼1
jak ∂nk k
where the coefficients α1 ; …; αN can be computed by solving the
Thus, the diagonal terms Bkk can be defined as
algebraic system: !
1 ∂w
N Bkk ≔  ∑ βj Bkj þ ðxk Þ :
∑ αj Akj ¼ u0 ðxk Þ ðxk A Γ D Þ; βk jak ∂nk
j¼1
ð8Þ
N This is the desingularization idea, see e.g. [14,11,5] for details. Note
∑ αj Bkj ¼ v0 ðxk Þ ðxk A Γ N Þ: that in the simplest case of the Laplace equation w can often be
j¼1
chosen e.g. by w :  1; in the case of the modified 2D Helmholtz
Similarly, the dipole formulation has the form equation Δu  λ2 u ¼ 0, a particular solution wðxÞ≔I 0 ðλ  jjxjjÞ can be
used, where I0 denotes the familiar modified Bessel function of the
N ∂Φ first kind, and so on. By a proper choice of the particular solution w,
uN ðxÞ ¼ ∑ αj ðx  xj Þ; ð9Þ
j¼1 ∂nj one can ensure that the coefficients βk do not vanish, so that the
68 C. Gáspár / Engineering Analysis with Boundary Elements 57 (2015) 66–71

definition of Bkk is correct (using possibly several particular solu- layer potential at the boundary, we have
tions). The desingularization can be improved by redefining not only Z
1 〈xk  y; ny 〉 1
the diagonal terms but also the neighbouring entries [9]. lim uðξÞ ¼  wðyÞ dΓ y  wðxk Þ ¼ uk :
ξ-xk 2π Γ jjxk  yjj2 2
In the dipole formulation, a straightforward desingularization
is C kk ≔Bkk (while for ja k, C kj ¼  Bjk ). Finally, the diagonal terms A procedure similar to the case of the single layer potential yields
Qkk can be defined by a desingularization procedure similar to the that these equations can be approximated by the discrete forms:
monopole formulation, by solving an auxiliary pure Dirichlet 1
problem. See [8] for details. ∑ C kj  αj  α ¼ uk ;
jak 2hk k
Remark. The definition of the diagonal entries C kk ≔Bkk is straight- where
forward but not trivial, so that it needs some explanation. For
simplicity, assume that Ω is a polygon with sides Γj (j ¼ 1; 2; …; N) 1 〈xk  xj ; nj 〉
C kj ≔  ðj a kÞ
and the boundary collocation point xj is the midpoint of Γj (more 2π jjxk  xj jj2
precisely let us approximate Ω in this way provided that N is large and the diagonal entries are C kk ¼  1=ð2hk Þ. Thus, C kk ¼ Bkk , and
enough). Let us express the solution of the original problem in the for ja k:
form of a single layer potential:
Z 1 〈xj  xk ; nj 〉
1



Bjk ¼ ¼  C kj :
uðξÞ ¼ ðlog

ξ  y

Þ  wðyÞ dΓ y ðξ A ΩÞ; 2π jjxj  xk jj2
2π Γ
Note that the above definitions of Bkk and Ckk mimic the jump
with some continuous density function w. It is well known that
of the normal derivative of the single layer potential at the
the normal derivative of a single layer potential has a jump at the
boundary and the jump of the double layer potential, respectively.
boundary i.e.
Z Both the monopole and the dipole formulations are suitable for
∂u 1 〈xk  y; nk 〉 1 solving mixed boundary problems using a regularized fundamen-
lim ðξÞ ¼  wðyÞ dΓ y  wðxk Þ:
ξ-xk ∂nk 2π Γ jjxk  yjj2 2 tal solution; the singularities of the diagonal entries of the systems
(8) and (10) are avoided by regularization and desingularization.
(Here 〈; 〉 denotes the Euclidean scalar product.) The integral in
To illustrate the above techniques, consider the simplest 2D
the right-hand side can be expressed as a sum of integrals taken
Laplace equation
along the sides Γj, whence
Z Δu ¼ 0 ð13Þ
∂u 1 〈xk  y; nk 〉
lim ðξÞ ¼ ∑  wðyÞ dΓ y
ξ-xk ∂nk j a k2π Γ j jjxk  yjj
2
in the square Ω≔½  1; 1  ½ 1; 1 with the test solution
Z
1 〈xk  y; nk 〉 1 uðxÞ≔log ððxð1Þ þ 2Þ2 þ ðxð2Þ þ 3Þ2 Þ; ð14Þ
þ  wðyÞ dΓ y  wðxk Þ:
2π Γk jjxk  yjj2 2
ð1Þ ð2Þ
where x , x denote the components of the vector variable x. The
The last integral (taken along Γk) vanishes since xk  y is orthogo- boundary was discretized by N collocation points in an equidistant
nal to the normal vector nk. Approximating the preceding integrals way. The scaling parameter was set as follows: c≔N, which is a quasi-
by the midpoint rule, we obtain the following discrete form of the optimal choice (more precisely, define c≔const:=h, where h denotes
equations: the characteristic distance of the boundary collocation points, and
1 const: means a dimensionless constant which is independent of
∑ Bkj  hj wj  wk ¼ vk ; the discretization of the boundary). The problem was supplied with
jak 2
mixed boundary conditions; along the half of the boundary, a
where wj ≔wðxj Þ, hj ≔jΓ j j (the length of Γj) and Dirichlet condition was prescribed, while along the remaining part
1 〈xk  xj ; nk 〉 ∂Φ of the boundary, a Neumann condition was imposed. Table 1 shows
Bkj ¼  ¼ ðx  xj Þ: the relative L2-errors of the approximate solution computed on the
2π jjxk  xj jj2 ∂nk k
boundary Γ and also the condition numbers of the corresponding
Denoting by αj ≔hj wj , we have algebraic systems. The method applied was the desingularized MFS
based on monopoles (7) and dipoles (9). This seems an acceptable
1
∑ Bkj αj  α ¼ vk ; compromise: the accuracy is still acceptable and the condition
jak 2hk k
numbers are still moderate; they are in fact much less than in the
whence Bkk ¼  1=ð2hk Þ. case of a traditional MFS with external source points located from the
boundary at a distance which is comparable with the size of the
On the other hand, expressing the solution of the original domain.
problem in the form of a double layer potential:
Z Remark. For Neumann and mixed problems, it is sufficient to
1 〈ξ  y; ny 〉
uðξÞ ¼  wðyÞ dΓ y ðξ A ΩÞ; compute the error of the approximate solution along the boundary
2π Γ jjξ  yjj2
only. Once the boundary values of the solution have been
(with a density function w which is different from the previous (approximately) computed, the solution inside the domain can
case). Using the well-known theorem of the jump of the double be reconstructed by solving an additional pure Dirichlet problem.

Table 1
MFS with truncation. Mixed boundary conditions. Relative L2-errors on the boundary and condition numbers.

N 8 16 32 64 128 256 512

Monopoles: relative L2-error (%) 0.4541 0.2009 0.1058 0.0552 0.0282 0.0142 0.0071
Condition number 8 21 53 127 294 670 1.5E þ3
Dipoles: relative L2-error (%) 0.4982 0.1516 0.0458 0.0177 0.0081 0.0040 0.0020
Condition number 3 6 13 29 63 135 287
 C. Gáspár / Engineering Analysis with Boundary Elements 57 (2015) 66–71 69

Table 2
MFS with truncation and desingularization. Pure Neumann boundary condition (using monopoles) and pure Dirichlet boundary condition (using dipoles). Relative L2-errors
on the boundary and condition numbers.

N 8 16 32 64 128 256 512

Monopoles: relative L2-error (%) 0.5580 0.3566 0.2057 0.1111 0.0568 0.0286 0.0143
Condition number 15 17 18 19 21 30 44
Dipoles: relative L2-error (%) 4.8602 1.3709 0.4715 0.1869 0.0821 0.0385 0.0186
Condition number 3 3 3 4 4 4 4

This latter problem can be treated by a regularized MFS and needs Consider the mixed problem:
no desingularization. ∂u


Δu ¼ 0 in Ω; ujΓD ¼ u0 ;
¼ v0
The condition numbers dramatically decrease if the above methods ∂n ΓN
are applied to pure problems. More precisely, the regularized- in the Sobolev space H 1 ðΩÞ. Let P be a (not necessarily orthogonal)
desingularized MFS based on monopoles (7) is quite efficient for projector of the closed subspace of the functions of H 1=2 ðΓÞ
solving pure Neumann problems, while the dipole formulation (9) vanishing along Γ D . Then the operators P 1 ≔I P, P 2 ≔P n (the
overperforms the regularized MFS when solving pure Dirichlet adjoint of P) are also projectors in the spaces H 1=2 ðΓÞ and
problems. This is in accordance with the indirect boundary element H  1=2 ðΓÞ, respectively, and can be interpreted as certain extensions
methods: pure Dirichlet (resp. Neumann) problems should be treated from Γ D to Γ and from Γ N to Γ, respectively. Define the following
by double layer (resp. single layer) potentials. iteration:
To illustrate this, consider again the problem (13) in the square
ΔU n þ 1=2 ¼ 0; U n þ 1=2 jΓ ¼ un þP 1 ðu0  un Þ; ð15Þ
Ω≔½  1; 1  ½  1; 1 with the test solution (14). Set c≔N, as earlier.
Table 2 shows the relative L2-errors of the approximate solution ∂U n þ 1


computed on the boundary Γ and the condition numbers of the ΔU n þ 1 ¼ 0;
¼ vn þ 1=2 þ P 2 ðv0  vn þ 1=2 Þ; ð16Þ
∂n Γ
corresponding algebraic system (8), when a pure Neumann condition
(n ¼0, 1, 2,…), where
was prescribed. Here the monopole formulation was applied. Table 2
shows also the relative L2-errors of the normal derivative of ∂U n þ 1=2


un ≔U n jΓ ; vn þ 1=2 ≔
:
the approximate solution computed on the boundary Γ and also ∂n Γ
the condition numbers of the corresponding algebraic system (10), The first equation is a pure Dirichlet, while the second one is a
when a pure Dirichlet condition was prescribed. Here the dipole pure Neumann problem. The iteration can be interpreted as
formulation was applied. Since the algebraic system is always follows: in the first half-step, the Dirichlet boundary condition
singular (due to the fact that the solution of the pure Neumann along Γ D is exactly satisfied, while in the second half-step, it is the
problem is unique up to an additive constant only), the first row of Neumann condition along Γ N that is exactly fulfilled.
the matrix was substituted by ones, which make the matrix non- Introducing the Dirichlet-to-Neumann operator (also referred
singular (in fact, this prescription defines well the above additive to as Poincaré–Steklov operator) J by Ju≔∂U=∂n, where ΔU ¼ 0 in Ω
constant). and UjΓ ≔u, (note that J is an invertible mapping which maps
Comparing with the previous examples presented in Table 1, the H 1=2 ðΓÞ onto a one-codimensional subspace of H  1=2 ðΓÞ) the above
accuracy of the approximate solutions remain more or less the same, iteration can be written in the form
but the condition numbers are much lower, so that the numerical
properties are much more advantageous. This is not surprising; in un þ 1 ≔J  1 ðI  P 2 ÞJðI  P 1 Þun þ b;
both cases, the discretization mimics a boundary integral equation, where b≔J  1 ðP 2 v0 þ ðI P 2 ÞJP 1 u0 Þ. If the operator
1
which is a Fredholm equation of the second kind; the operators are A≔J ðI  P 2 ÞJðI  P 1 Þ is a contraction in the space H 1=2 ðΓÞ, the
compact perturbations of the identity, which implies e.g. that the iteration is convergent.
conjugate gradient method converges quite rapidly: the speed of The ‘ideal’ choice of the extension operators is as follows. Let
convergence is superlinear in fact (cf. [2]). This shows that in the case P 1 u≔wjΓ , where w is the (unique) solution of the special mixed
of a pure Dirichlet and a pure Neumann condition, the computational problem:
∂w


cost of the solution of the discrete systems (10) and (8) can be
significantly reduced by applying e.g. the simple conjugate gradient Δw ¼ 0 in Ω; wjΓD ¼ ujΓD ;
¼ 0:
∂n ΓN
method (or another Krylov subspace method). In addition to this, the
problems of dense and ill-conditioned matrices are also avoided. Similarly, let P 2 u≔∂w=∂n, where now w is the (unique) solution of
In the case of mixed problems, the above advantages can be the mixed problem:
preserved if the original mixed problem is converted to ‘pure’ ∂w


Δw ¼ 0 in Ω; wjΓD ¼ 0;
¼ vjΓ N :
problems, as shown in the next section. ∂n ΓN
Then, as can be easily seen, the operator A ¼ J  1 ðI  P 2 ÞJðI  P 1 Þ
equals to the zero operator. That is, the iteration (15) and (16)
3. Converting mixed problems to pure problems results in the exact solution after a single iteration step. Unfortu-
nately, to numerically realize these extension operators, mixed
The following technique was originally proposed in connection problems should be solved, i.e. from a computational point of
with the boundary element method, see [6]. Here we briefly recall view, the implementation is at least as costly as that of the
the main idea but it should be pointed out that the appearing solution of the original problem itself. However, in practice, P1,
subproblems can be treated in principle by any other method; here P2 can be defined by a coarse-grid approximation of the above
the regularized and desingularized MFS are used. mixed problems, so that the computation of the projections can be
For the sake of simplicity, we restrict ourselves to the 2D embedded in the multi-level context without difficulty. In this
Laplace equation; note, however, that the technique essentially case, the operator A is not zero any more, but might remain
remains the same for more general and/or 3D problems as well. contractive, so the iteration is still convergent.
70 C. Gáspár / Engineering Analysis with Boundary Elements 57 (2015) 66–71

Once the extension operators have been defined, the steps of point xðkÞ ðk þ 1Þ
j A Sk the ‘unification’ of some points xj , … , xðk
j
þ 1Þ
of
1 r
the iteration can be realized by solving pure Dirichlet and pure the finer level Sk þ 1 (e.g. xðkÞ
j can be defined as the barycentre of the
Neumann subproblems using the dipole and the monopole for-
mulations as discussed earlier. points xðk
j
þ 1Þ
, … , xðk
j
þ 1Þ
which are located in a small neighbour-
1 r

The extension operators P1, P2 can also be defined in many hood of the point xðkÞ
j ). Moreover, between the consecutive levels,
other ways and also independently of each other. Our experience ‘inter-grid’ transfer operators (restrictions and prolongations)
was that even the extension operators P1, P2 defined by a simple should be defined. This task is also independent of the concrete
Shepard interpolation technique from Γ D to Γ and a constant partial differential equation itself. Suppose that on the level Sk, the
extension from Γ N to Γ, respectively, result in excellent conver- components of the vector ðα1 ; …; αNk Þ denote the coefficients of the
gence properties.
monopoles (or dipoles) associated with the points xðkÞ ðkÞ ðkÞ
1 ; x2 ; …; xN k .
To illustrate the method, consider again the problem (13) in the
square Ω≔½ 1; 1  ½  1; 1 with the test solution (14). Set c≔N, as The simplest projection is defined by averaging
earlier. The problem (13) was supplied with mixed boundary 1 r
αðkÞ ðk þ 1Þ
j ≔ r  ∑ αjp ðj ¼ 1; 2; …; N k Þ;
conditions. Along half of the boundary, a Dirichlet condition p¼1
was prescribed, while along the remaining part of the boundary,
a Neumann condition was imposed. We have implemented the while the simplest prolongation is the uniform (piecewise con-
above conversion to a sequence of ‘pure’ subproblems (15) and stant) prolongation:
(16). Table 3 shows the relative L2-errors of the approximate þ 1Þ 1
αðk ≔  αðkÞ ðp ¼ 1; 2; …; rÞ;
solution computed on the boundary Γ with different numbers of jp r j
boundary collocation points (N). The parameter iteration indicates As a next step, the problem has to be discretized at each level, i.
the applied number of iterations (15) and (16). The extension e. the matrices A, B, C, Q as well as (8) and (10) should be defined
operator P1 was defined by Shepard interpolation: for each collocation point set S0, …, SL.
∑x A ΓD uj  wkj In the simplest cascade technique, only prolongations are
uk ≔ j ðxk A Γ N Þ; needed. The algorithm is as follows:
∑xj A ΓD wkj

where wkj ≔1=jjxk  xj jj2 . The extension operator P2 was defined to  On the coarsest level (S0), solve the discrete problem ((8) or
be a constant extension: (10)) exactly.
vk ≔const: ðxk A Γ D Þ;  Prolongate the solution to the next finer level.
 Improve the solution by applying e.g. several conjugate gradi-
where the const: is defined in such a way that the integral of the
ent iterations to the normal equation.
extended function v over the boundary vanishes (which is the case  Repeat the procedure until the finest level (SL ) is achieved.
when v is the normal derivative of a harmonic function).
Comparing with the results of Table 1, it can be seen that the
The number of conjugate gradient iterations at each level can be
relative errors of the latter methods are achieved after 4 and
kept moderate, which may significantly reduce the computational
5 iteration steps of the solution of ‘pure’ problems (15) and (16).
complexity.
The above ‘pure’ problems can be treated efficiently from a
Note that, once the ‘inter-grid’ transfer operators have been
computational point of view, as pointed out in the previous
constructed, the usual items of the multi-level methods (coarse
section. Thus, a remarkably efficient method has been obtained.
grid correction, multigrid cycle, etc., see e.g. [13] for details) can be
The efficiency can be increased by embedding this iteration into a
defined in a straightforward way.
multi-level technique.
As an example, consider again the problem (13) in the square
Ω≔½  1; 1  ½ 1; 1 with the test solution (14). Set c≔N, as before.
3.1. Multi-level solution The problem (13) was supplied with mixed boundary conditions.
Along the half of the boundary, a Dirichlet condition was pre-
To handle the problem in a multi-level way, first of all, a scribed, while along the remaining part of the boundary, a
sequence of finite sets of collocation points Neumann condition was imposed. We have applied the above
Sk ≔fxðkÞ ðkÞ ðkÞ
1 ; x2 ; …; xN k g ðk ¼ 0; 1; …; LÞ; cascade type multi-level technique based on monopoles (see (8)).
The next finer ‘grid’ was always constructed by doubling the
is needed. Here S0 is considered the ‘coarsest grid’ and SL is the number of boundary collocation points in an equidistant manner.
‘finest grid’ (note that, in the meshless context they need not have At each level, the approximate solution was defined by applying
any grid or mesh structure). The inclusions Sk  Sk þ 1 are not 4 iteration steps of ‘pure’ problems (15) and (16), where the initial
required; however, it is often more comfortable to consider a guess was transferred from the previous coarser level. At each
‘pure’ problem, 5 conjugate gradient iteration steps were applied
Table 3 to the corresponding normal equation. Table 4 shows the relative
MFS with truncation. Mixed boundary condition. Iteration by converting to pure
L2-errors of the approximate solution computed on the boundary
subproblems. Relative L2-errors on the boundary.
Γ at the different levels (l) and the corresponding numbers of
N boundary collocation points (N). The results indicate that the same
accuracy can be achieved as in the iterative solution technique (cf.
Iteration 8 16 32 64 128 256 512 Table 3), but the number of necessary arithmetic operations is
significantly reduced.
1 2.9570 6.0692 8.6811 11.069 13.374 15.635 17.904
2 0.4598 0.2344 0.3843 0.5148 0.6287 0.7378 0.8021
3 0.5038 0.1933 0.1042 0.0739 0.0668 0.0681 0.0691
4 0.5018 0.1849 0.0862 0.0421 0.0201 0.0095 0.0053 4. Summary and conclusions
5 0.5019 0.1855 0.0874 0.0435 0.0217 0.0108 0.0054
6 0.5019 0.1856 0.0874 0.0435 0.0216 0.0107 0.0053 Some versions of the method of fundamental solutions have
7 0.5019 0.1856 0.0874 0.0435 0.0216 0.0107 0.0053
been developed. The source and boundary collocation points were
8 0.5019 0.1856 0.0874 0.0435 0.0216 0.0107 0.0053
assumed to coincide. This causes singularities, which have been
 C. Gáspár / Engineering Analysis with Boundary Elements 57 (2015) 66–71 71

Table 4
MFS with truncation using monopoles and multi-level method. Mixed boundary condition. Relative L2-errors on the boundary.

l 0 1 2 3 4 5 6
N 8 16 32 64 128 256 512
Relative L2-error (%) 0.5018 0.1856 0.0875 0.0436 0.0216 0.0109 0.0057

boundary elements. Proceedings of the 24th international conference on the

eliminated by several techniques. The singularities caused by the
boundary element method incorporating meshless solution seminar, vol. 13.
singularity of the original fundamental solution have been avoided Southampton, Boston: WitPress; 2002. p. 67–76.
by using a truncated (or other approximate) fundamental solution [2] Axelsson O, Karátson J. Superlinearly convergent CG methods via equivalent
controlled by a scaling constant. The stronger singularities caused preconditioning for nonsymmetric elliptic operators. Numer Math 2004;99:
197–223.
by the Neumann boundary condition i.e. the normal derivatives
[3] Chen W. Symmetric boundary knot method. Eng Anal Bound Elem 2002;26:
of the fundamental solution have been eliminated by a special 489–94.
desingularization technique. The same procedure was performed [4] Chen W, Shen LJ, Shen ZJ, Yuan GW. Boundary knot method for Poisson
also for the dipole formulation. From a computational point of equations. Eng Anal Bound Elem 2005;29:756–60.
[5] Chen W, Wang FZ. A method of fundamental solutions without fictitious
view, the monopole formulation has proven more advantageous boundary. Eng Anal Bound Elem 2010;34:530–2.
for pure Neumann problems, while the dipole formulation over- [6] Gáspár C. Multigrid and multipole techniques in the boundary integral
performs the traditional monopole formulation for pure Dirichlet equation method. In: Hackbusch W, Wittum G, editors. Notes on numerical
problems. For a mixed boundary condition, a special iterative fluid mechanics, vol. 54. Braunschweig, Wiesbaden: Vieweg-Verlag; 1996.
p. 102–14.
solution algorithm has been presented, converting the original [7] Gáspár C. A multi-level regularized version of the method of fundamental
(mixed) problem to a sequence of pure Dirichlet and pure solutions. In: Chen CS, Karageorghis A, Smyrlis YS, editors. The method of
Neumann subproblems, the solutions of which converge rapidly fundamental solutions – a meshless method. Atlanta, USA: Dynamic Publish-
to the solution of the original mixed problem. These pure sub- ers, Inc.; 2008. p. 145–64.
[8] Gáspár C. Regularization techniques for the method of fundamental solutions.
problems can be handled efficiently by using conjugate gradients. Int J Comput Methods 2013;10(2) p. 1341004-1–21.
The efficiency can be increased further by embedding the resulting [9] Gáspár C. Combining regularization and desingularization techniques in the
method in a multi-level context. In addition to it, the problem method of fundamental solutions. In: Sellier A, Aliabadi MH, editors. Advances
in boundary element and meshless techniques XIV. UK: EC, Ltd.; 2013.
of highly ill-conditioned matrices is also avoided. The use of the
p. 219–24.
fast multipole summation technique may speed up the method [10] Li X. On convergence of the method of fundamental solutions for solving the
further. Dirichlet problem of Poisson's equation. Adv Comput Math 2005;23:265–77.
[11] Sarler B. Desingularised method of double layer fundamental solutions for
potential flow problems. In: Skerget L, Brebbia CA, editors. Boundary elements
Acknowledgments and other mesh reduction methods XXX. Southampton, Boston: WitPress;
2008. p. 159–68.
[12] Sarler B. Solution of potential flow problems by the modified method of
The research was partly supported by the European Union fundamental solutions: formulations with the single layer and the double
(co-financed by the European Social Fund) under the project layer fundamental solutions. Eng Anal Bound Elem 2009;33:1374–82.
[13] Stüben K, Trottenberg U. Multigrid methods: fundamental algorithms, model
TÁMOP-4.2.2.A-11/1/KONV-2012-0012.
problem analysis and applications. GDM-Studien, No. 96, Birlinghoven, Germany;
1984.
References [14] Young DL, Chen KH, Lee CW. Novel meshless method for solving the potential
problems with arbitrary domain. J Comput Phys 2005;209:290–321.

[1] Alves CJS, Chen CS, Sarler B. The method of fundamental solutions for solving
Poisson problems. In: Brebbia CA, Tadeu A, Popov V. Int. series on advances in