See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327993261

Trefftz methods with cracklets and their relation to BEM and MFS

Article  in  Engineering Analysis with Boundary Elements · October 2018

DOI: 10.1016/j.enganabound.2018.06.021

CITATION READS

1 48

3 authors, including:

Carlos J S Alves Svilen S. Valtchev

University of Lisbon University of Lisbon
77 PUBLICATIONS   926 CITATIONS    31 PUBLICATIONS   164 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Inverse Source Problems View project

Inverse Problems View project

All content following this page was uploaded by Svilen S. Valtchev on 03 October 2018.

The user has requested enhancement of the downloaded file.

 Trefftz methods with cracklets and their relation to
BEM and MFS

Carlos J. S. Alvesa , Nuno F. M. Martinsb,∗, Svilen S. Valtcheva,c

a CEMAT–IST, University of Lisbon, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal
b Department of Mathematics, Faculdade de Ciências e Tecnologia, Univ. Nova de Lisboa, Quinta da
Torre, 2829-516 Caparica, Portugal
c ESTG - Polytechnic Institute of Leiria, 2411-901 Leiria, Portugal

Abstract
In this paper we consider Trefftz methods which are based on functions defined by single
layer or double layer potentials, integrals of the fundamental solution, or their normal
derivative, on cracks. These functions are called cracklets, and satisfy the partial differ-
ential equation, as long as the crack support is not placed inside the domain. A boundary
element method (BEM) interpretation is to consider these cracks as elements of the orig-
inal boundary, in a direct BEM approach, or elements of an artificial boundary, in an
indirect BEM approach. In this paper we consider the cracklets just as basis functions
in Trefftz methods, as the method of fundamental solutions (MFS). We focus on the
2D Laplace equation, and establish some comparisons and connections between these
methods with cracklets and standard approaches like the BEM, indirect BEM, and the
MFS. Namely, we propose the enrichment of the MFS basis with the cracklets. Several
numerical simulations are presented to test the performance of the methods, in particular
comparing the results with the MFS and the BEM.
Keywords: Trefftz method, Cracklets, Boundary element method, Method of
fundamental solutions
2010 MSC: 65N38, 65N80

1. Introduction

The solution of boundary value problems (BVP) for partial differential equations
(PDEs) benefits from the fact that for homogeneous linear PDEs, it is possible to write
them as a linear combination of basis functions that satisfy the PDE, reducing the BVP
to the fitting of the boundary conditions. This is a general context of the Trefftz methods
in its different variants (e.g. [21]), and it has the advantage that no meshing procedure
is needed for the domain.

∗ Corresponding
author
Email addresses: carlos.alves@math.tecnico.ulisboa.pt (Carlos J. S. Alves), nfm@fct.unl.pt
(Nuno F. M. Martins), ssv@math.ist.utl.pt (Svilen S. Valtchev)

Preprint submitted to Engineering Analysis with Boundary Elements October 3, 2018

 The Boundary Element Method (BEM) has also a long history of application (e.g.
[12]) and it may be included in the class of Trefftz methods, as it relies on boundary
potentials that are solutions of the PDE inside the domain. The indirect BEM (IBEM) is
another Trefftz variant, where the boundary potentials are considered on some artificial
boundary surrounding the original one. The use of an artificial boundary is also the
context of application of another Trefftz method - the method of fundamental solutions
(MFS), no longer with boundary potentials, but simply as the location of point sources,
i.e. fundamental solutions centered on source points, that are located on some artificial
boundary (e.g. [14, 13, 2]).
A connection between some of these approaches was already pointed out in [10], and
in this paper we explore this known connection further, using the concept of cracklets
(e.g. [1]). Since the single and double layer potentials are defined on a boundary, which
is decomposed in multiple boundary elements, we consider the elements of the boundary
themselves to be cracks (or screens), and the cracklets are therefore defined as single or
double layers over the support of a crack. If the boundary elements are seen as the union
of cracks, then the BEM formulation may be understood with the use of cracklets on the
boundary, and the IBEM formulation with the use of cracklets on some artificial domain.
However since these cracklets are themselves solutions of the PDE, with an appro-
priate behavior at infinity, they might be understood as a Trefftz basis, and we propose
that they may be used with no direct connection to a boundary (artificial or original). In
section 3.5 we prove a density theorem and its corollary that justifies the completeness
of this set of basis functions.
As in the MFS, a good location for the support of the cracklets, is an issue of current
research (cf. [2, 8, 23]), and we either considered a standard approach using boundary
dilation, like in the IBEM, or used a MFS choice as proposed in [2], or even considered
cracklets on the original boundary as in the BEM, as was proposed in [5] to tackle
difficulties in the approximation of discontinuous functions. In particular, to avoid the
Gibbs oscillations when approximating discontinuous boundary conditions (e.g. [16]).
Several different methods have been considered to avoid the singularities associated
with the fundamental solutions and allowing direct collocation on the boundary, such as
the the boundary knot method [15], the regularized meshless method [24], the modified
method of fundamental solutions [22], or the singular boundary method [11]. In section
3.7 we emphasize that when the solution does not have an analytic extension that goes
beyond the analytic support of the approximation, the results will become worse. In
fact it is known [18, 7] that we can get exponential convergence of the MFS, as long
as the solutions are entire, but the rate of convergence decreases fast, if the solutions
do not extend analitically beyond the analytic support of the approximation, which is
determined by the artificial boundary. In this sense the Trefftz method with cracklets
proposed here is not attached to a specific artificial boundary as in an indirect boundary
integral equation method. It inherits the ill conditioning of the inversion of compact
operators (e.g. [19]), and these methods have been used to solve inverse problems, sharing
the need of some regularization techniques (e.g. [17, 4]).
The cracklets are here considered not only as a Trefftz method, but also as an en-
richment technique for the MFS, for instance as proposed in [5, 3, 6], or to serve as a
link between the MFS and the BEM. Recently, the method of angular basis functions
(MABF) was proposed in [25, 20] which considers double layer solutions, that are linked
to angular measurements in the case of the Laplace equation. These functions are also
2
considered in the fast implementation of BEM (e.g. [9]) and correspond to double layer
cracklets.
In Section 2 we briefly recall the notions of layer potentials, and in Section 3 we
consider the cracklet Trefftz method, as proposed for single and double layer potentials,
but reduced to constant densities, and establish the main mathematical result that proves
the completeness of the method. Finally in Section 4 we present numerical simulations
that illustrate the performance of the different approaches.

2. Single and double layer potentials

This work will focus on harmonic boundary value problems, but it may be extended
to problems such as 
Du = 0, in Ω
(P ) (1)
Bu = g, on Γ = ∂Ω
where a fundamental solution Φ of the differential operator D is available.
In this paper, for simplicity, we assume D = ∆ to be the 2D Laplace operator, with
fundamental solution given by
1
Φ(x) = log ||x||, (2)
2π
and when B is the identity operator, (P ) is a Dirichlet boundary value problem, with
unique solution u ∈ H 1 (Ω), for a given g ∈ H 1/2 (Γ).
The corresponding single layer or double layer potentials (e.g. [19]) are respectively
represented by Lγ and Mγ ,
Z
Lγ α(x) = Φ(x − y)α(y)dsy (3)
γ
Z
Mγ β(x) = ∂ny Φ(x − y)β(y)dsy (4)
γ

where γ is a boundary that may coincide with Γ, α ∈ H −1/2 (γ) and β ∈ H 1/2 (γ) are
unknown densities. Moreover, ∂n = ∇ · n represents the normal derivative (we write
∂ny to be clear that the gradient is with respect to the y variable).
Remark 2.1. These layer potentials are defined for x ∈
/ γ, and trace formulas may be
obtained when x ∈ Ω → xγ ∈ γ (along the normal direction),

Lγ α(x) → Sγ α(xγ ), Mγ β(x) → (τ I + Kγ )β(xγ ) (5)

where the expressions for Sγ and Kγ are the same as for Lγ and Mγ , respectively, but
now with xγ ∈ Γ, thus implying the definition of a (weakly) singular integral operator.
The parameter τ (xγ ) = 12 for all regular boundary points, but if xγ is a corner point,
θ
then τ (xγ ) = 2π , where θ represents the internal angle (from 0 to 2π). Taking the trace
from the exterior domain, with x ∈ / Ω̄, then τ has negative sign, corresponding to the
external angle.

3
2.1. Integral equations
To find the unknown densities we consider the first kind integral equations when
x ∈ Γ, and Γ 6= γ.
(i) Using the single layer potential, Lγ α(x) = g(x), for x ∈ Γ, i.e.
Z
Φ(x − y)α(y)dsy = g(x). (6)
γ

(ii) Using the double layer potential, Mγ β(x) = g(x), for x ∈ Γ, i.e.
Z
∂ny Φ(x − y)β(y)dsy = g(x). (7)
γ

Remark 2.2. When Γ = γ, we obtain

Sα(x) = g(x), (8)

which is also a first kind integral equation on Γ, but a second kind integral equation is
obtained for the double layer potential, since (τ I + K)β(x) = g(x), which means
Z
τ (x)β(x) + ∂ny Φ(x − y)β(y)dsy = g(x). (9)
Γ

Remark 2.3. In the case of Neumann boundary conditions, the inner normal trace of
the single and double layer potentials gives, when x ∈ Ω → xγ ∈ γ (along the normal
direction),
n(xγ ) · ∇Lγ α(x) → (−τ I + Kγ0 )α(xγ ) (10)

n(xγ ) · ∇Mγ β(x) → Tγ β(xγ ) (11)

with
Z Z
Kγ0 α(xγ ) = ∂nx Φ(xγ − y)α(y)dsy and Tγ β(xγ ) = ∂nx ∂ny Φ(xγ − y)β(y)dsy .
γ γ

The operator Tγ presents then a singular integration which is understood in the sense of
the Cauchy principal value.

2.2. Direct and Indirect Boundary Element approach

Consider the boundary element method in its two variants - the direct the and indirect
approach.
(i) In the BEM, equations (8) or (9) are considered, taking γ = Γ.
(ii) In the IBEM, equations (6) or (7) are considered, with γ being an artificial
boundary surrounding the original Γ. That is, we may take ω ⊃ Ω̄ and γ = ∂ω.
The integral equation (9) is of the second kind, it is better conditioned and usually
preferred to the first kind integral equation (8).

4
 The discretization of the integral operators on γ, equal or not to Γ, usualy consists
in splitting that boundary into boundary elements

γ = γ1 ∪ ... ∪ γN (12)

which may reproduce or approximate the full boundary. Thus,

N Z
X
Mβ(x) = ∂ny Φ(x − y)βn (y)dsy (13)
n=1 γn

using local densities βn = β|Γn and, for example, a trigonometric or a power series
expansion, βn (s) = βn,0 + βn,1 s + βn,2 s2 + ...
In the simplest situation βn ∼= βn,0 are constant terms, and the calculation resumes
to the evaluation of Z
Kn (x) = ∂ny Φ(x − y)dsy (14)
γn

leading to the approximation

N
X
Mβ(x) ≈ βn Kn (x). (15)
n=1

Using the boundary condition (τ I + K)β(xk ) = g(xk ) for points xk ∈ Γ, we get the
following linear system
N
X
(τ (xk )In (xk ) + Kn (xk )) βn = g(xk ), (16)
n=1
(
1, if xk ∈ γn ,
where In (xk ) = δnk = . In this case, the discretized solution is given
0, if xk ∈
/ γn ,
by
N
X
u(x) = Mβ(x) ≈ Kn (x)βn (17)
n=1

for x ∈ Ω.
The difference between the direct and indirect approach consists in taking γn on Γ
(direct BEM) or on the artificial boundary γ (indirect BEM), taking into account that in
the indirect approach, the discretized boundary equation (16) is considered with τ ≡ 0.

Advantages and disadvantages. A clear advantage of the BEM with the double layer
approach is that it leads to a second kind integral equation, avoiding ill conditioning
problems. But, on the other hand, the approximation on the boundary reflects the
approximation being taken on β.
In the simplest situation, while considering an approximation by piecewise constant
densities, the approximation on the boundary will be affected by the discontinuous ap-
proximation of the density, and this is not suited for more regular boundary conditions,
as a better approximation requires more basis functions and a larger system.
5
Figure 1: Cracklets for the double layer potential (K(γε ) on the left picture), and for the single layer
potential (S(γε ) on the right picture), with ε = 1.

On the other hand, the indirect approach may lead to an analytic approximation
of the boundary condition, but this might be inconvenient for discontinuous boundary
conditions. Moreover it leads to ill conditioned systems, and the major problem is that
it defines an extension of the solution outside Ω.
This is the same problem that may occur with the MFS, while taking an artificial
boundary γ = ∂ω. Thus, we will proceed in the same direction, proposing a Trefftz
method with cracklets placed outside Ω, as is done with point sources in the MFS.

3. A Trefftz method with cracklet functions

We will consider as basis functions for the Trefftz method, the cracklet functions
as they will be defined entirely by the crack location γn outside Ω. In the limit, if we
consider these cracks to be located on the boundary Γ, we get the same approach as in
the boundary element method.

3.1. Double layer cracklets

We start with the double layer cracklets, because their expression is simpler. In fact,
we define the cracklets K by
Z
K(γn )(x) = ∂ny Φ(x − y)dsy , (18)
γn

and this calculation may be reduced to a reference linear crack γε = [0, ε] × {0}, by a
rigid transformation γn = an + Qγε , when γn is a linear crack, with Q being a rotation
matrix, and ε = |γn | being the length of the cracklet.
For the Laplace equation in 2D,
    
1 ε − x1 x1
K(γε )(x) = arctan + arctan (19)
2π x2 x2

6
Remark 3.1. Considering the complex Argand plane, K may be written as (eg. [9])
1
K(γ)(z) = (arg(b − z) − arg(a − z)), (20)
2π
for a line crack γ = [a, b], where a, b, z ∈ C.
In Fig.1 we present the cracklet function u = K(γε ) with ε = 1, which presents a
discontinuity with a jump [u] = u− − u+ = 1. On the other hand, the normal derivative
is continuous across the crack, that is [∂n u] = 0.

3.2. Single layer cracklets

In this case, we define the cracklets S by
Z
S(γn )(x) = Φ(x − y)dsy , (21)
γn

and this calculation may again be reduced to a reference linear crack γε = [0, ε] × {0},
by the same rigid transformation γn = an + Qγε .
For the Laplace equation in 2D,
ε x1 − ε x1
S(γε )(x) = − x2 K(γε )(x) + log(||x − (ε, 0)||) − log(||x||). (22)
2π 2π 2π
In Fig.1 we present the cracklet function v = S(γε ) with ε = 1. In this case the
function is continuous across the crack [v] = 0, and if fact when x ∈ γε
ε x1 − ε x1
v(x) = S(γε )(x) = + log |x1 − ε| − log |x1 |
2π 2π 2π
ε
and v(x) = 2π on both tips of the crack. A jump appears only on the normal derivative
through γε .
Remark 3.2. We are only considering here the simplest cracklets, with constant density,
but it is clear that we could also consider more complex cracklets (cf. [5]) using, for
instance, polynomial densities up to degree p, with
Z Z
p
Sp (γn )(x) = Φ(x − y) y dsy and Kp (γn )(x) = ∂ny Φ(x − y) y p dsy . (23)
γn γn

3.3. Trefftz methods with cracklets

We do not require that the union of the cracks γ1 ∪ ... ∪ γN forms a closed boundary.
Taking γn outside Ω, each cracklet K(γn ) or S(γn ) is a solution of the PDE in Ω.
We do not exclude that γn ⊂ Γ, and in that case, the cracklet may be understood as a
boundary element.
Therefore a linear combination of these particular solutions gives also a solution of
the PDE.
(i) Using cracklets K, the approximation will be
N
X
uN (x) = βn K(γn )(x), (24)
n=1
7
calculating βn from the system

K(γn )(x−
 
m ) [βn ] = [g(xm )] , (25)

where xm ∈ Γ, are collocation points.

Remark 3.3. It is important to notice that we write x− m = xm − εn(xm ), with ε → 0 ,
+

to include the case when γn ⊂ Γ. Taking into account the trace formula (9), system (25)
resumes to  
1
δnm + K(γn )(xm ) [βn ] = [g(xm )] . (26)
2

(ii) Using cracklets S, it is similar, as the approximation will be

N
X
vN (x) = αn S(γn )(x), (27)
n=1

calculating αn from the system

[S(γn )(xm )] [αn ] = [g(xm )] . (28)

Example 3.4. In the simplest situation, we may consider the domain to be a square Ω =
(−1, 1)2 , and double layer cracklets on the four edges of the square, ∂Ω = γ1 ∪γ2 ∪γ3 ∪γ4 .
In this example we are using a direct BEM method, and we may use the middle points
on each edge to collocate and get the system. Using the complex notation we get
1
K(γn )(z) = (arg(qn+1 − z) − arg(qn − z)),
2π
using q1 , ..., q4 as the complex coordinates for the vertices, and assuming q5 = q1 . This
leads to the matrix
1
Mmn = K(γn )(zm ) = (arg (qn+1 − zm ) − arg (qn − zm )) .
2π
Taking as collocation points zm = 12 (qm + qm+1 ), the system resumes to be
      1    
0 c t c β1 2 c t c β1 g1
1 1
 I +  c 0 c t   β2  =  c c t   β2  =  g2 
       
2 
2  t c 0 c   β3   t c 21 c   β3   g3 
c t c 0 β4 c t c 12 β4 g4
arc tan(2)
where c = 1−2t 1
4 , t= 2 − π = 0.147584.., and gm = g(zm ).
When g(z) = 1, and since 2c + t = 21 , the solution is β1 = ... = β4 = 1, therefore
4 4
X X 1
u(z) = K(γn )(z) = (arg(qn+1 − z) − arg(qn − z))
n=1 n=1
2π
1
= (arg(q5 − z) − arg(q1 − z)) .
2π
8
Max.Error
Log(Max.Error)
0.6
dist
2 4 6 8 10
0.5

0.4 -2

0.3
-4
0.2

-6
0.1

dist
2 4 6 8 10 -8

Figure 2: Example 1. On the left, approximation of the maximum error with N = 4 cracklets for
g(z) = z, decreasing with larger distance to the boundary. On the right, the same experiment with
N = 32, for g(z) = z p , taking the average of the maximum errors for p = 0, ..., 8, here presented in log10
scale.

Since q5 = q1 , for points z outside the square, the argument does not jump and u(z) = 0,
but for points z inside the square the argument rotates by 2π, and u(z) = 1, giving the
exact solution inside the square. In fact, while using the argument function or arctangent,
some attention must be given to the implementation.
A similar result can be obtained for a larger square Ω̂ = (−a, a)2 with a > 1, giving the
same exact solution for the indirect BEM, although for a different matrix with diagonal
entries Mjj = π1 arctan( a−1a
).
However, for g(z) = z it is clear that the approximation does no longer yield an exact
result. Moreover, the best result is not given with the cracklets on the original boundary.
The error converges to zero as the distance of the artificial boundary goes from zero to
infinity. When the distance is zero, and the boundaries coincide, the maximum error is
0.65, and when the distance is 10 it is only 5.81 × 10−3 (see Fig.2, on the left). Moreover,
using N = 8 cracklets (two on each edge), the maximum error decreases from 0.357.. to
1.84 × 10−5 , and with N = 16 the maximum error decreases from 0.21 to 8.51 × 10−10 .
Even if we consider monomials g(z) = z p up to degree p = 8 (or more), the results are
extremely good with N = 32 cracklets at a distance d = 10, giving an average maximum
error of 10−8 (see Fig.2, right).
Thus, we may conclude that a few cracklets allow to approximate harmonic polyno-
mials up to a considerable degree, with extremely small errors.
In fact, since we observe that ||z p − u32,p (z)||∞ ≤ 10−8 , we have for polynomials
P8 P8
P8 (z) = p=0 ap z p that u32 (z) = p=0 ap u32,p (z) satisfies
8
X 8
X
||P8 (z) − u32 (z)||∞ ≤ |ap | ||z p − u32,p (z)||∞ ≤ 10−8 |ap |.
p=0 p=0

Remark 3.5. Since any analytic function in C has a power series expansion, in Laurent
series, we can capture a considerable number of functions with good precision, using just
32 cracklets. This is not only true for the square, but also for any other shape inside the
square.
Example 3.6. The excellent results obtained for traces of harmonic polynomials, do not
occur for non harmonic polynomials. Consider polynomials of the form g(z) = Re(z)p
9
which do not satisfy the Laplace equation, for p ≥ 2, and vary N from 32 to 128 elements.
The smallest maximum error is achieved at a variable distance, which is about the length
of the distance between the collocation points (in this case, d = 8/N ). These results are
shown in Fig. 3, left (due to some numerical ill conditioning, the linear system was solved
in the least squares sense).
We must note that when the cracklets are close to the boundary, the system matrix
presents diagonal elements that are close to 1, and the other entries are comparatively
very small, but not small enough to give a diagonally dominant matrix.
In Fig 3, right, we plot a similar result for boundary functions g(z) = |z|p , which also
do not satisfy the Laplace equation. We took the distance to be d = 8/N, and considered
N = 2m cracklets, taking the average errors up to p = 4. In this case we considered not
only the maximum norm, but also a discrete `2 norm, using P test points z1 , ..., zP ∈ ∂Ω
(with P  N ). This norm is given by
P
1 X
||g − uN ||2`2 = |g(zp ) − uN (zp )|2 , (29)
P p=1

and this may be seen to approximate the L2 (Γ) norm, because

Z
||g − uN ||2L2 (Γ) = |g(z) − uN (z)|2 dz
Γ
P Z
X
≈ |g(z) − uN (z)|2 dz
p=1 [zp ,zp+1 ]

P
X
≈ |g(zp ) − uN (zp )|2 |zp+1 − zp |
p=1
P
|Γ| X
≈ |g(zp ) − uN (zp )|2
P p=1

assuming that the points are almost equally spaced and |zp+1 − zp | ≈ |Γ| P . The only
difference is that we neglect the length of the boundary |Γ|, like we do in the `1 average
error
P
1 X
||g − uN ||`1 = |g(zp ) − uN (zp )|, (30)
P p=1
1
that approximates |Γ| ||g − uN ||L1 (Γ) , the averaged L1 (Γ) norm.

3.4. Connection with the MFS

An approximation with the method of fundamental solutions can be considered as a
combination
N
X N
X
vN (x) = ak Φ(x − yk ) + bk · ∇Φ(x − yk ), (31)
k=1 k=1
10
Log(Max.Error) Log(Error)
dist
0.05 0.10 0.15 0.20 0.25 N
50 100 150 200 250
-0.5
-1
-1.0
N=32 Max|Error|
-1.5
N=64 -2 L2(Error)
-2.0 N=128 L1(Error)
-3
-2.5

-3.0 -4

-3.5

Figure 3: Example 2. On the left, approximation maximum error for g(z) = Re(z)p , with N = 32, 64
and 128 cracklets decreasing near the boundary. On the right, a similar experiment with g(z) = |z|p ,
taking the average errors for p = 1, ..., 4, here presented in log10 scale for three different norms.

using only monopolar sources (with bk ≡ 0, which is the classical case) or also using
dipolar sources, that might follow a prescribed direction bk = n(yk )bk , corresponding to
a discretization of the double layer potential on the artificial boundary. The connection
of the MFS with the boundary layers is well known, and we provide here some error
estimates considering cracklets.

Taking a cracklet defined by the single layer potential on a small γ, with a middle
point ym , the mid-point integration rule gives for x not close to the boundary
Z
S(γ)(x) = Φ(x − y)dsy = Φ(x − ym )|γ| + O(|γ|3 ) (32)
γ

with |γ| being the length of the cracklet support.

Thus, this small cracklet is expected to have a similar effect as considering a point
source at ym as a pole. A similar reasoning occurs for the double layer potential,
Z
K(γ)(x) = ∂ny Φ(x − y)dsy ≈ n(ym ) · ∇Φ(x − ym )|γ| (33)
γ

relating it to the dipole centered in ym with direction given by the normal.

Therefore, the use of very small single layer cracklets, at some distance from the
boundary, will have a similar effect as the use of the standard MFS (and the use of small
double layer cracklets will be similar to a dipole MFS). We resume this in the following
theorem (that may also be deduced for the double layer cracklets).

Theorem 3.7. Consider line cracks such that |γn | ≤ ε and dist(Ω, γn ) > 0. Then an
approximation using cracklets S, given by
N
X
uN (x) = αn S(γn )(x) (34)
n=1

with |αn | ≤ ||α||∞ , can be reproduced by a standard MFS approximation

N
X
vN (x) = αn |γn |Φ(x − yn ) (35)
k=1
11
where yn is the middle point of γn , with the following error estimate for x ∈ Ω̄,

N ||α||∞ ε3
|uN (x) − vN (x)| ≤ . (36)
48πdist(Ω, γ)2

Proof. It is an immediate consequence of (32), more precisely for x ∈ Ω̄

|γn |3
Z 
 

 Φ(x − y)dsy − Φ(x − yn )|γ n | ≤
 48πdist(Ω, γn )2
γn

2
which is a consequence of the middle point integration error estimate, since |∂ij Φ(x)| ≤
1
2π||x||2 . Thus,

N N
X X |αn ||γn |3
|uN (x) − vN (x)| ≤ |αn | |S(γn )(x) − |γn |Φ(x − yn )| ≤ .
n=1 n=1
48πdist(Ω, γn )2

3.5. Completeness of the Cracklets

Here we present a theorem for the completeness of external cracklets in L2 (Γ).
Theorem 3.8. Let ω ⊃ Ω̄, and consider the spaces of cracklets on γ = ∂ω, defined by

GS = span S(γ[y,ε] ) : y ∈ γ, ε > 0 ∪ {1}, (37)


GK = span K(γ[y,ε] ) : y ∈ γ, ε > 0 , (38)
where γ[y,ε] is a part of the boundary, connecting y ∈ γ to ỹ ∈ γ such that ||y − ỹ|| = ε.
Then GS and GK are dense in L2 (Γ).
Proof. We prove the result for GS (the proof for GK is similar).
Consider any discrete number of points y1 , ..., yN ∈ γ, such that when N → ∞, this
set of points is dense in γ. We assume that ||yk+1 − yk || = ε > 0, such that yN +1 = y1 .
Given any function ψ ∈ L2 (Γ), we want to prove that the orthogonality of the cracklets
implies ψ to be null. That is, if
Z Z Z


S(γ[y,ε] ), ψ L2 (Γ) = S(γ[y,ε] )(x)ψ(x)dx = Φ(x − s)ds ψ(x)dx = 0
Γ Γ γ[y,ε]

then ψ ≡ 0. Using this equality for each yn we define any linear combination of cracklets
N
X Z
v(x) = αn S(γ[yn ,ε] )(x) = α(s)Φ(x − s)ds
n=1 γ

while considering α(s) = αn , when s ∈ γ[yn ,ε] , and we have

N
X Z Z


hv, ψiL2 (Γ) = αn S(γ[y,ε] ), ψ L2 (Γ)
= α(s)Φ(x − s)ds ψ(x)dx = 0.
n=1 Γ γ
12
 This means that for any piecewise constant density α, we get
Z Z
α(s) Φ(x − s)ψ(x)dx ds = 0.
γ Γ

As this is true for any ε > 0, from the previous weak equality, we conclude the strong
equality Z
u(s) = Φ(x − s)ψ(x)dx = 0,
Γ
when s ∈ γ.
Since u satisfies the null exterior Dirichlet problem in ω c = R2 \ω̄ and it is ortogonal
to constants, this implies u ≡ 0 in R2 \Ω̄, by analytic extension. As u is defined by a
continuous single layer potential on Γ, then [u] = 0, and u− = u+ = 0. The uniqueness
of the interior Dirichlet problem also implies that u ≡ 0 in Ω, and this forces ψ ≡ 0.
Corollary 3.9. In Theorem 3.8 we may consider any set γ ⊂ Ωc as long as there is a
boundary γ ? ⊆ γ, with γ ? = ∂ω ? and ω ? ⊃ Ω̄.
Proof. This is a consequence of the proof of Theorem 3.8, since u(s) = 0 with s ∈ γ
implies, in particular, that u(s) = 0 with s ∈ γ ? , and the other steps remain true taking
γ ? instead of γ.
Remark 3.10. Theorem 3.8 does not specify which kind of artificial ω to consider, and
we may assume that it is a polygon, such that each crack γ[yn ,ε] is a straight segment.
Therefore, due to the corollary, we may consider intersecting crack supports that in the
limit contain an exterior artificial boundary.

Remark 3.11. As previously mentioned when γ ⊂ Ωc , the inversion of Sγ or Kγ leads

to ill-posed problems, because they imply the inversion of compact operators. Therefore
we can not have Sγ α = g or Kγ β = g with ||α||, ||β|| < ∞. On the other hand, this
theorem implies that we may find a sequence of αn or of βn such that ||Sγ αn − g|| → 0
or ||Kγ βn − g|| → 0. The main problem is that, in that case, ||αn || → ∞ and ||βn || → ∞.
Thus, like in the MFS, we expect that the coefficients in the solution of the linear
system will tend to infinity, unless some regularization technique (like Tikhonov regu-
larization) is used. The ill conditioning problems of the MFS will appear also using
cracklets, or using a similar indirect BEM.

3.6. Neumann Boundary Conditions

In the case of Neumann boundary conditions, the procedure is basically the same.
Moreover, using the single layer potential cracklets, the gradient of the reference
cracklet is given by the vector
 
Φ(x − (ε, 0)) − Φ(x)
∇S(γε )(x) = . (39)
−K(γε )(x)

In the reference crack we may consider the normal vector to be (0, 1), giving exactly

∂nx S(γε ) = −K(γε ),

13
and this means that we may use the double layer cracklets directly for the Neumann
problem, as long as we understand that the density being recovered in the system is α,
to be used in the single layer representation (and not β which was used for the double
layer representation).
This results from the duality between the expression for the trace of the double layer
(9), and the normal trace of the single layer given in (10), and is well known in the
BEM. It does not change much if we take the normal derivative on a boundary Γ which
is different from γ, as long as the normal vector is the same. For instance, considering
line cracklets, parallel to a polygonal boundary.
In any case, with the expression of the gradient being available, for any cracklet (by
rigid transformation), we just have to solve the linear system

n(xm ) · ∇S(γn )(x−

 
m ) [αn ] = [gn (xm )] (40)

(here gn (xm ) represents the Neumann data on xm ∈ γ), and the solution will be given
by the discrete single layer representation (27).
Another possibility is to consider the gradient of the double layer representation,
giving  
−ε (ε − 2x1 )x2
∇K(γε )(x) =
2π||x − (ε, 0)||2 ||x||2 (ε − x1 )x1 + x22
with singularities on the tips, and this is associated with the singular operator Tγ as
defined in (11), when Γ and γ coincide.

3.7. Analytic support

The use of cracklets as basis functions to solve boundary value problems can be seen,
in the limiting case as a BEM, when the union of γn coincides with Γ, or the indirect
BEM, when that union γ is a boundary that surrounds Γ.
With this flexibility it can be used to either capture the advantages of one or the
other.
Moreover, since the cracklets are used as basis functions, we do not require them to
replicate a full boundary, we can place some cracklets near the boundary (for instance,
when discontinuities are present or expected), and some other cracklets away from the
boundary.
The main goal is just to diminish the boundary error, and approximate g as well as
possible, since all the functions satisfy the PDE.
Unlike the BEM, we do not propose to increase the polynomial degree of the density
approximation, using functions Sp or Kp (as it was mentioned in Remark 3.2), with
higher p, unless this is strictly necessary - and in that case it is clear that the standard
BEM will perform better.
We propose to place the cracklets at a reasonable distance from the original boundary,
as their trace leads to analytical functions on the boundary. The compromise is that the
system will become more ill conditioned, as occurs in the MFS.
In fact, the use of cracklets may be combined with the MFS, taking the advantages
of one and the other. The use of cracklets may be seen as an enrichment of the MFS
basis, as it was already done in [5] to solve boundary value problems with discontinuities
on the Dirichlet data.

14
 The best performance of a method is related to the type of boundary functions being
tested, and if a method suits better to some of them, it may fail considerably for others.
The main difficulties for the Trefftz methods appear in approximating solutions with
smaller analytical support than the basis, and in increasing the ill conditioning of the
linear systems to be solved.
Remark 3.12. Consider a Trefftz basis φ1 , ..., φn , ... that has analytical support in a set
Ω̂ and a solution u with analytic support in Ω̃ ⊇ Ω. If Ω̃ ⊂ Ω̂ then it is impossible to
write u as a linear combination of φn . This does not imply that it is not possible to
approximateP u with a linear combination of φn , in appropriate norms. In fact, assume
that v = n αn φn satisfies v ≡ u in Ω ⊆ Ω̃. As the analytical support of v is Ω̂, by
the analytical continuation theorem, it is possible to extend analytically v from Ω̃ to Ω̂.
However this is not true for u as its analytical support is reduced to Ω̃ and these sets are
different.
In the case of the MFS, it is known that the method performs better when the point
sources are placed far from the boundary [18], if the solution extends as an analytic
function to infinity, meaning that Ω̃ = R2 . The same performance was seen here with
cracklets taken far away from the boundary. In these situations there is no problem as
Ω̂ ⊂ R2 . But in these situations also polynomials, or other analytic functions, satisfying
the PDE, could be used in a Trefftz method.
However, if the solution does not extend analytically outside the domain, i.e.Ω̃ = Ω
the best possibility would be to consider the BEM, even if the approximation on the
boundary is not so great.
Frequently there is an analytic extension of the solution outside Ω but it may be
constrained in some parts of the boundary, for instance, corners, or discontinuities on
the boundary condition. In those situations, the best strategy is to combine methods. On
the part of the boundary where no analytic extension is expected, we may use cracklets
very close to the boundary, like in the BEM, and on the part of the boundary where no
problems are expected, the cracklets may be placed away from the boundary, like in the
MFS.

4. Numerical simulations

We consider some experiments that show the performance of the proposed methods,
for non trivial domains, and compare them with the MFS and the BEM.
There are several known factors that may change the quality of the approximation,
concerning the position of the point sources in the MFS, and in this case the position,
orientation and length of the cracklets. We will focus on some specific approximations
that worked well, but better results may be obtained for each of the methods, depending
on the geometry of the domain and on the boundary functions.
Sinc Trefftz methods yield in general ill conditioned systems, the linear systems were
solved in the least squares sense (a Tikhonov regularization factor was not needed in
these examples). This would not be needed for cracklets very close to the boundary, as
this would lead a to well conditioned system, especially for cracklets K, nevertheless, as
this will not change the N × N system solution, as we are dealing with square matrices.
The matrices are rectangular only for the MFS which is considered with 2N collocation
points.
15
4.1. Simulation 1 (polygonal domain)
Consider a polygonal domain Ω1 defined by the six vertices with complex coordinates
{1, 43 + i, i − 12 , −1, −1 − i, −i}.
We divided each segment in the same number of points, that acted as collocation
− +
points, x1 , ..., xM . From these set points we defined the cracklets, as segments ]ym , ym [
with tips given by
ym = xm + δm t⊥
m
±
and ym = ym ± θ m t m , (41)
where tm = 1
2 (xm+1 −xm−1 ) and t⊥
mis orthogonal to tm , pointing in the normal direction
(or its average). The choice of ym has already been used for the MFS (e.g. [2]).
The parameter δm reflects the distance to the boundary with respect to the distance
between points, and the parameter θm reflects half the length of the crack with a middle
point given by ym . Keeping the size of the crack with respect to the distance between
collocation points, this means θm = 1, which was the best choice that we considered.
It should be noticed that the choice given by (41) is thought for convex shapes,
because the edges of the cracklet may fall inside the domain, and θm must be considered
small enough.
In Fig. 4 (left) we present the distribution of the cracklets, given 120 collocation
points (20 each side), for δm = 5, θm = 1.
With these cracks we tested the Trefftz basis for cracklets K (double layer) and
cracklets S (single layer), and compared the results with the results for the MFS, with
the double number of collocation points, and using the same ym as sources.
• (a) We considered the Dirichlet data given by the complex function (z = x1 + x2 i)
(
0,
ga (z) = Im(z)   if Im(z) ≤ 0, (42)
sin(2z)
10 Re 1+i−z if Im(z) > 0,
which is discontinuous in the derivatives, and presents a singularity at 1+i, meaning
at the point (1, 1) which is close to the boundary point ( 34 , 1).
In Fig. 4 (right), we plot the function in 1002 points on the boundary, starting with
point (1, 0) going anti-clockwise, and finishing on the same point.
In Fig. 5 (left) we plot the results using 120 cracklets S or K, or using 120 source
points for the MFS. We may see that there is not much difference taking cracklets in the
Trefftz method, or source points in the MFS, the results were similar in this case. The
graph of the solution is given in Fig. 5 (right).
The highest values of the absolute errors occur in the corner points, especially in the
corner point nearest to the singularity, as expected. Considering different norms, as in
(29), (30), we obtained the errors on the boundary (due to the maximum principle for
Laplace equation the errors in the domain are smaller than the errors on the boundary),
presented in Table 1:
||ga − u120 ||∞ ||ga − u120 ||`2 ||ga − u120 ||`1
BEM S 0.078 5.8 × 10−3 1.4 × 10−3
BEM K 0.12 1.5 × 10−2 5.6 × 10−3
MFS 3.9 × 10−4 3.4 × 10−5 1.4 × 10−5
Cracklet S 7.3 × 10−4 5.9 × 10−5 2.2 × 10−5
Cracklet K 8.0 × 10−4 5.7 × 10−5 1.2 × 10−5
16
 1.0

0.5
1.5

0.0
1.0

-0.5
0.5

-1.0

200 400 600 800 1000

-1.0 -0.5 0.0 0.5 1.0

Figure 4: Simulation 1. On the left, the domain of approximation with the collocation points (120 blue
points) on the boundary ∂Ω1 , and the cracklets (in red) using the procedure defined in (41). On the
right, the values of the function ga along 1002 test points on the polygonal boundary.

Table 1: Errors in different norms for the different methods, considered for Simulation 1 with ga .

It should be observed that the approximation given by the piecewise constant density
in the BEM, either in the single layer formulation (BEM S), or in the double layer
formulation (BEM K), is not good. For instance, with BEM K, the approximation with
a piecewise discontinuous function performs poorly in the maximum norm, but is not so
bad in the `2 or `1 norms. The `2 error for BEM K performed consistently like O(h1.3 ),
being h the average element length, and even with N = 960 elements (h = 0.0071), the
`2 error was 1.9 × 10−3 .
On the other hand, by doing a simple dilation of the artificial boundary to γ = 1.5Γ,
both indirect BEM, S or K, with N = 120, gave maximum errors about 10−3 , and `2 or
`1 errors below 10−4 , being close to the errors obtained with the MFS or cracklets.
• (b) With exactly the same geometry, boundary points, cracklets, and point sources,
we now considered the boundary function
 
1 sin(2z)
gb (z) = Re , (43)
10 1+i−z
which is almost the same function as considered in (42), but now it is only the real
part of an analytic function on Ω and therefore it is the exact solution. Except
for the BEM, the errors are comparatively smaller than before, due to the higher
regularity, and the results are presented in a similar table (Table 2):
||gb − u120 ||∞ ||gb − u120 ||`2 ||gb − u120 ||`1
BEM S 0.077 4.8 × 10−3 1.1 × 10−3
BEM K 0.12 1.5 × 10−2 6.0 × 10−3
MFS 3.7 × 10−5 4.8 × 10−6 2.0 × 10−6
Cracklet S 9.9 × 10−5 9.5 × 10−6 3.3 × 10−6
Cracklet K 1.6 × 10−5 1.5 × 10−6 4.0 × 10−7
17
 0.0008

0.0006
Absolute Error

0.0004
Cracklet K
Cracklet S
MFS
0.0002

0.0000
0 200 400 600 800 1000
Points

Figure 5: Simulation 1. On the left, absolute errors for 3 methods - using Cracklets K or S, and using
the MFS, with the boundary function ga . On the right, plot of the solution in the polygonal domain.

0.0001

0.000025
0.00008
0.000020
Absolute Error

0.00006
MFS 0.000015

0.00004 Cracklet S 0.000010

Cracklet K 0.5 × 10-5

0.00002
0

0.0000
0 200 400 600 800 1000
Points

Figure 6: Simulation 1. On the left, absolute error plots for the three methods, considering the bound-
ary function gb . On the right, density plot of the absolute error in the domain for the Cracklet K
approximation.

Table 2: Errors in different norms for the methods considered for Simulation 1, now with the analytic
gb .

In Fig. 6 (left) we plot a similar figure to Fig.5 (left), now with gb as boundary
function, and in Fig. 6 (right) we present the error in the whole domain, which is in
average below 10−6 , except around the corner point, close to the singularity, where it is
slightly higher.
Increasing the number of points near the corner singularity improves the results. For
instance taking the same 120 points, but with a higher concentration at the corner, and
at a distance δm = 10, decreased the maximum absolute error of the MFS to 7 × 10−7
(50 times less than before), with a similar effect on the cracklet methods (4.2 × 10−6 for
cracklets S and 1.4 × 10−6 for cracklets K).

4.2. Simulation 2 (mixed round shapes)

In this simulation we considered the same domain for Im(z) > 0 together with the
semicircle |z| < 1, for Im(z) < 0. This defines Ω2 . The collocation points on the boundary
18
 1.0

0.5
0.00012

0.00010
0.0

Absolute Error
0.00008
Cracklet S
0.00006
-0.5
MFS
0.00004 Cracklet K
0.00002
-1.0
0.00000
0 200 400 600 800 1000 1200

-1.0 -0.5 0.0 0.5 1.0 Points

Figure 7: Simulation 2. On the left, the collocation points (120 blue points) on the boundary ∂Ω2 , and
the cracklets (in red) using the procedure defined in (41). On the right, absolute error plots for the three
methods, considering the boundary function gb .

∂Ω2 and the cracklets are shown in Fig. 7 (left), at a similar distance δm = 5. Again
we used 20 points in each segment and 60 points on the semicircle, in a total of 120
collocation points, as before.
The results obtained for this shape and the data given by ga or gb are very similar
to the previous case. In Fig. 7 (right) we present the variation of the absolute error for
the three methods, considering the analytic boundary function gb . The error results are
presented in Table 3.
||gb − u120 ||∞ ||gb − u120 ||`2 ||gb − u120 ||`1
MFS 3.4 × 10−5 4.8 × 10−6 2.3 × 10−6
Cracklet S 1.3 × 10−4 1.3 × 10−5 5.8 × 10−6
Cracklet K 1.6 × 10−5 1.4 × 10−6 3.4 × 10−7
Table 3: Errors in different norms for the 3 methods, considered for Simulation 2 with gb .

In Fig. 8 we present the solution for the problem in Simulation 2 (on the left), and
the absolute error for the approximation using cracklets K. As previously mentioned, for
Simulation 1, the highest errors occur near the singularity, as expected.

4.3. Simulation 3 (round shapes)

We also considered shapes defined by parts of radial star-shape functions, in this case
Ω3 is defined by the boundary
(
 it
1 − 13 sin(3t) t ∈ [0, π]
∂Ω3 = r(t)e : t ∈ [0, 2π[ , with r(t) = (44)
1 t ∈]π, 2π[

The domain is presented in Fig. 9 (left) together with the associated cracklets, as defined
in (41) at distance δm = 6, and it presents two discontinuities (at t = 0, t = π).

19
Figure 8: Simulation 2. On the left, plot of the solution in the domain and, on the right, the error for
the approximation obtained with 120 cracklets K.

As in the IBEM we could have used an artificial boundary to place the cracklets, but
here we used the technique that is being considered for the MFS, in the distribution of
source points (see [2]). Note that unlike the BEM, where the use of straight elements
compromises the geometry of the boundary, here we may use collocation on the original
boundary as long as these cracklets are placed outside the domain.
Using the boundary function gb , in Fig. 9 (right) we present the variation of the
absolute error for the three methods. The error results are presented in Table 4.
||gb − u120 ||∞ ||gb − u120 ||`2 ||gb − u120 ||`1
MFS 9.7 × 10−8 4.0 × 10−8 3.4 × 10−8
Cracklet S 5.2 × 10−7 9.0 × 10−8 6.5 × 10−8
Cracklet K 8.0 × 10−7 8.9 × 10−8 3.0 × 10−8
Table 4: Errors in different norms for the 3 methods, considered for Simulation 3 with gb .

A similar table could be obtained for the function ga , but with higher errors, never-
theless smaller than the ones obtained in Table 1.
Remark 4.1. It should be noted that in all these simulations we considered θm = 1 in
(41), but taking a different θm would slightly change the results. In this case, we also
used θm = 14 , 12 , 2, and the results were basically the same (only a slightly worse). This
may be justified by the fact that, at a large distance from the boundary, the use of
cracklet with half size is compensated by a coefficient taken as double the size.

4.4. Simulation 4 (non convex polygon)

In the previous cases the MFS provided good results, and now we will consider a
case with discontinuous data, using a similar function to ga but being 1 instead of 0, for
Im(z) ≤ 0. More precisely:
(
1, if Im(z) ≤ 0,
gc (z) = (45)
ga (z) if Im(z) > 0,
20
 1.5

1.0

0.5 8.× 10-7

6.× 10-7
0.0

Absolute Error
Cracklet S
4.× 10-7
-0.5 MFS
Cracklet K
2.× 10-7

-1.0
0
0 200 400 600 800 1000

-1.0 -0.5 0.0 0.5 1.0 Points

Figure 9: Simulation 3. On the left, the collocation points (120 blue points) on the boundary ∂Ω3 , and
the cracklets (in red) using the procedure defined in (41). On the right, absolute error plots for the three
methods, considering the boundary function gb .

Furthermore, we will consider a polygon with the same vertices as in Simulation 1, but
with a new corner point (− 41 , 0), with an angle greater than 32 π. We will call this domain
Ω4 , which is represented in Fig. 10 (left) with the 280 collocation points on the boundary,
together with cracklets placed on a surrounding curve with tip points ym as given in (41),
which is closer to the boundary for the first 4 segments, at a distance δm = 1, and with
δm = 5, for the other 3 segments.
We also compared with the BEM S and BEM K (see the plots in Fig. 10, on the
right), as the results were not so good in this case, and are given in Table 5.
||gc − u280 ||`2 ||gc − u280 ||`1
BEM S 7.7 × 10−2 1.2 × 10−2
BEM K 1.3 × 10−2 2.8 × 10−3
MFS 4.3 × 10−2 7.1 × 10−3
Cracklet S 3.0 × 10−2 4.2 × 10−3
Cracklet K 2.4 × 10−2 2.5 × 10−3
Table 5: Errors in different norms for the methods considered for Simulation 4, with the discontinuous
function gc .

We did not consider the maximum norm, as this is not appropriate for discontinuous
functions, but we may see in Fig. 10 (right) that the BEM K was the most appropriate
method in this case, since it is the one that best captures the discontinuities. On the
other hand, BEM S, fitted the function, but presented high instabilities in the corner
points. The MFS and the cracklet S or K methods also presented difficulties to fit the
discontinuities, due to the Gibbs effect. In fact, since the cracklets were placed outside
the boundary all functions are analytic there, and the Gibbs effect occurs.

We also considered the simulation for the same geometry, but now for the analytic
21
 1.0

0.5

1.5

0.0 1.0

Function gc

Value
BEM S
0.5
BEM K
-0.5
MFS
0.0

-1.0 0 500 1000 1500 2000

-1.0 -0.5 0.0 0.5 1.0 Points

Figure 10: Simulation 3. On the left, the collocation points (280 blue points) on the boundary ∂Ω4 , and
the used cracklets (in red). On the right, comparison of the plots for different methods, with N = 280
and the boundary function gc .

function gb , and as expected, the results are much better, especially for the Trefftz
methods (MFS, cracklets S and K), as we may see in the error Table 6.

||gb − u280 ||∞ ||gb − u280 ||`2 ||gb − u280 ||`1

BEM S 0.22 7.9 × 10−3 1.0 × 10−3
BEM K 6.5 × 10−2 5.6 × 10−3 2.2 × 10−3
MFS 2.0 × 10−3 1.3 × 10−4 4.7 × 10−5
Cracklet S 8.0 × 10−3 2.5 × 10−4 3.0 × 10−5
Cracklet K 1.1 × 10−2 3.6 × 10−4 2.5 × 10−5
Table 6: Errors in different norms for the methods considered for Simulation 4, but now with the
analytic function gb .

4.5. MFS Enrichment

Finally we considered the possibility of combining methods to improve the results
obtained in Table 5.
We took the MFS enriched by cracklets K.
This meant to consider 7 cracklets K, one at each edge of the polygon, to avoid the
Gibbs effect on discontinuities and corner problems, and this also meant to place the
cracks near the boundary, at a very close distance(= 10−4 ).
To keep the same number of basis functions as before, we considered 273 source points
for the MFS, using the 7 cracklets as an enrichment of the MFS and kept the distance
δm = 1 for all source points ym as defined in (41).
The results of this method (abbreviated MFS+K) were much better than the ones
of the MFS, and also better than BEM K, as we may see in Table 7 (we used 4008 test
points on the boundary):

The biggest improvement was cancelling the Gibbs effect in the MFS, but difficulties
remained in the corners. This lead to a considerable decrease in the average error, i.e. in
22
 ||gc − u280 ||`2 ||gc − u280 ||`1
BEM K 1.3 × 10−2 2.8 × 10−3
MFS 4.3 × 10−2 7.1 × 10−3
MFS + K 1.1 × 10−2 4.7 × 10−4
Table 7: Improving the MFS with 7 edge cracklets K, for the discontinuous function gc .

the `1 norm, as the error at each point was below 10−3 except around 4 corner points, as
we may see in Fig.11 (left), compared to the worse performance of the BEM with double
layer formulation using 280 elements, not affected by the corners, but with difficulties
matching the peak of the boundary function gc .
We present the solution given by this MFS+K method in Fig. 11 (right).

0.05
Error

0.00
BEM K
MFS+ K

-0.05

0 1000 2000 3000 4000

Points

Figure 11: Simulation 4. Comparison of the errors between MFS+K and BEM (on the left), and the
solution for gc given with the MFS+K (on the right).

We could also consider more than 7 cracklets K, taking several more on each side, but
this also generates further discontinuities at the end of each cracklet, as the functions
K(γn ) are discontinuous at the tips. To avoid this we could also add K1 (γn ) cracklets
as mentioned in Remark 3.2, but this goes beyond the scope of this paper, as this would
also mean a comparison with BEM K formulated with piecewise linear, and not piecewise
constant, densities.

5. Conclusions

In this work we presented two Trefftz methods with cracklets, based on the single layer
potential (cracklet S) or the double layer potential (cracklet K), that in some particular
situations may be viewed as a direct or as an indirect BEM, or may also be viewed
as equivalent to the MFS with monopole or dipole source points. This depends if the
cracklets are placed on the boundary (direct BEM), form a boundary outside (indirect
BEM), or may be considered with small crack support (MFS). However, understanding
these cracklets as the basis functions of a Trefftz method allows all these possibilities,
and allows to establish a more flexible link between the methods.
For smooth boundary conditions the proposed Trefftz methods with cracklets perform
in a similar way as the MFS and significantly better than the direct BEM. On the other
23
hand, in the presence of discontinuities, the cracklets improve the accuracy of the MFS,
if their supports are chosen on the boundary. In particular, we considered the cracklets
as an enrichment of the MFS that proved to be helpful to avoid the Gibbs effect on
discontinuous boundary conditions (as was already suggested in [5]).
The simulations have been carried for the Dirichlet problem of the Laplace equation
in 2D, but may be extended with no great difficulty to Neumann or Robin boundary
conditions, with the same performance of the method. It is also possible to consider
other differential operators with a known fundamental solution, and an extension to 3D,
but this would lead to new cracklet functions (some of which are already known from
BEM research), and some extra difficulties with the 3D geometries are expected. This
work is currently under investigation.

Acknowledgements

The financial support received from the Fundação para a Ciência e a Tecnologia
(FCT) project UID/Multi/ 04621/2013 is gratefully acknowledged.

References
[1] C.J. Alves, J.B. Abdallah, M. Jaoua, Recovery of cracks using a point-source reciprocity gap func-
tion, Inverse Problems in Science and Engineering 12 (2004) 519–534.
[2] C.J.S. Alves, On the choice of source points in the method of fundamental solutions, Engineering
Analysis with Boundary Elements 33 (2009) 1348–1361.
[3] C.J.S. Alves, V.M.A. Leitão, Crack analysis using an enriched MFS domain decomposition tech-
nique, Engineering Analysis with Boundary Elements 30 (2006) 160–166.
[4] C.J.S. Alves, N.F.M. Martins, The direct method of fundamental solutions and the inverse kirsch-
kress method for the reconstruction of elastic inclusions or cavities, Journal of Integral Equations
and Applications 21 (2009) 153–178.
[5] C.J.S. Alves, S.S. Valtchev, On the application of the method of fundamental solutions to boundary
value problems with jump discontinuities, Applied Mathematics and Computation 320 (2018) 61–74.
[6] P.R.S. Antunes, S.S. Valtchev, A meshfree numerical method for acoustic wave propagation prob-
lems in planar domains with corners and cracks, Journal of Computational and Applied Mathemat-
ics 234 (2010) 2646–2662.
[7] A.H. Barnett, T. Betcke, Stability and convergence of the method of fundamental solutions for
Helmholtz problems on analytic domains, Journal of Computational Physics 227 (2008) 7003–7026.
[8] C.S. Chen, A. Karageorghis, Y. Li, On choosing the location of the sources in the MFS, Numerical
Algorithms 72 (2015) 107–130.
[9] G. Chen, J. Zhou, Boundary Element Methods with Applications to Nonlinear Problems, Atlantis
Press, 2010.
[10] J.T. Chen, J.J. Tsai, Y.T. Lee, J.W. Lee, Study on connections of the MFS, trefftz method, indirect
BIEM and invariant MFS in the three-dimensional laplace problems containing spherical boundaries,
Applied Mathematics and Computation 218 (2011) 4056–4074.
[11] W. Chen, F. Wang, A method of fundamental solutions without fictitious boundary, Engineering
Analysis with Boundary Elements 34 (2010) 530–532.
[12] A.H.D. Cheng, D.T. Cheng, Heritage and early history of the boundary element method, Engineer-
ing Analysis with Boundary Elements 29 (2005) 268–302.
[13] G. Fairweather, A. Karageorghis, The method of fundamental solutions for elliptic boundary value
problems, Advances in Computational Mathematics 9 (1998) 69–95.
[14] M.A. Golberg, C.S. Chen, The method of fundamental solutions for potential, Helmholtz and
diffusion problems, in: Boundary Integral Methods: Numerical and Mathematical Aspects, WIT
Press, Computational Mechanics Publications, Boston, Southampton, 1999, pp. 103–176.
[15] Y.C. Hon, W. Chen, Boundary knot method for 2d and 3d helmholtz and convection-diffusion
problems under complicated geometry, International Journal for Numerical Methods in Engineering
56 (2003) 1931–1948.
24
 [16] J.H. Jung, A note on the Gibbs phenomenon with multiquadric radial basis functions, Applied
Numerical Mathematics 57 (2007) 213–229.
[17] A. Karageorghis, D. Lesnic, L. Marin, A survey of applications of the MFS to inverse problems,
Inverse Problems in Science and Engineering 19 (2011) 309–336.
[18] M. Katsurada, H. Okamoto, A mathematical study for the charge simulation method I, J Fac Sci
Univ of Tokyo Sect 1A Math 35 (1988) 507–518.
[19] R. Kress, Linear Integral Equations, volume 82 of Applied Mathematical Sciences, 2nd ed., Springer,
1999.
[20] X. Li, J. Oh, Y. Wang, H. Zhu, The method of transformed angular basis function for solving the
Laplace equation, Engineering Analysis with Boundary Elements 93 (2018) 72–82.
[21] Z.C. Li, T.T. Lu, H.T. Huang, A.H.D. Cheng, Trefftz, collocation, and other boundary methods—a
comparison, Numerical Methods for Partial Differential Equations 23 (2006) 93–144.
[22] B. Šarler, Solution of potential flow problems by the modified method of fundamental solutions:
Formulations with the single layer and the double layer fundamental solutions, Engineering Analysis
with Boundary Elements 33 (2009) 1374–1382.
[23] R. Schaback, An adaptive numerical solution of mfs systems, in: The Method of Fundamental
Solutions - A Meshless Method, Dynamic Publishers, 2008, pp. 1–27.
[24] D. Young, K. Chen, C. Lee, Novel meshless method for solving the potential problems with arbitrary
domain, Journal of Computational Physics 209 (2005) 290–321.
[25] D.L. Young, Y.J. Huang, C.S. Wu, V. Sladek, J. Sladek, Angular basis functions formulation for
2D potential flows with non-smooth boundaries, Engineering Analysis with Boundary Elements 61
(2015) 1–15.

25

View publication stats