1

Comparison of Kansas method versus Method

of Fundamental Solution (MFS) and Dual Reciprocity Method of Fundamental Solution (MFSDRM)

1.1

Introduction

In this work, performances of two most widely used radial basis function based
meshless methods (Kansas and MFS/MFS-DRM) are presented and evaluated.
Both were tested on two problems, the first assuming the potential field, and the
second with internal source, using three different boundary conditions-Dirichlet,
Neumann, and Robin boundary conditions.

1.2

Benchmark test 1

For comparison, NAFEMS benchmark test has been chosen. This test describes
2D thermal problem without internal heat generation and with the following
boundary conditions:
on boundary AB: T = 1000 C
on boundary DA: zero flux

T
n

=0

on boundaries BC,CD: heat transfer with surrounding temperature 00 C

Material properties for this problem are set as follows:

thermal conductivity k = 52 W/mK
heat transfer coefficient h = 750 W/mK

Combining all listed properties we have:

52 2 u
u
52
+ 750 u
n
u
u
n

in

on BC,CD

100

on AB

on AD

(1)

The results are compared in the reference point at the coordinate (0.6,0.2),
where analytical value is calculated to be 18.25375654.
1
0.9
T
100
92.8571
85.7143
78.5714
71.4286
64.2857
57.1429
50
42.8571
35.7143
28.5714
21.4286
14.2857
7.14286
0

0.8
0.7

0.6
0.5
0.4
0.3
0.2
0.1
0

0.1

0.2

0.3

0.4

0.5

0.6

Figure 1: Results of NAFEMS thermal problem

1.3

Benchmark test 2

For this comparison, benchmark test describing 2D thermal problem with internal heat generation and with the following boundary conditions was chosen:
on boundary AB, DA T = ((1 x)2 + (1 y)2 ) 0 C
on boundary BC, CD zero flux

T
n

=0

The problem description is

2 u = 4
in
2
u = (1 x) + (1 y)2
u
= 0
on BC,CD
n

(2)
on AB,DA

The results are compared across the whole area. The analytical solution for this
problem is
u = (1 x)2 + (1 y)2
(3)

1
0.9
T
2
1.85714
1.71429
1.57143
1.42857
1.28571
1.14286
1
0.857143
0.714286
0.571429
0.428571
0.285714
0.142857
0

0.8
0.7

0.6
0.5
0.4
0.3
0.2
0.1
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure 2: Results of the second NAFEMS thermal problem

1.4

Kansas method

Kansas method or Radial basis functions collocation method uses global interpolation direct collocation with radial basis functions (RBS) in order to obtain
the solution of partial differential equation. For the problem described above,
we have the following formulation: For the elliptic problem, the approximate

solution for u is given by

u
(p) =

N
X

j j (p),

(4)

j=1

where j are coefficients to be determined. Similarly we can define the first and
the second derivatives of the function u
in cartesian coordinates
N

u
(p) X
j (p)
j
=
,

j=1
and

x, y

(5)

2u
(p) X
2 j (p)
=
,
j

j=1

x, y,

x, y

(6)

Substituting Eq. (4),(5) and (6) into Eq. (1), we get the formulation for
Dirichlet boundary condition
N
X

j (p) j = T D

j=1

p D ,

(7)

Neumann boundary condition

N
X
j
j=1

(p) j = T N ,

p N

(8)

Robin boundary condition


N 
X
j
R
k
+ h j (p) j = h Tref
n
j=1

p R ,

(9)

where p
Domain interpolation
N  2
X
j
j=1

x2

2 j
+
y 2

(p) j = f,

(10)

which is a N N linear system from where we can calculate the unknowns

{j }N
j=1 . The source f is 0 in the case of Laplace equation.

1.5

MFS

The basic idea of the Method of Fundamental Solution is to solve potential field
problem i.e. Laplace equation only by the interpolation over boundary. This
is possible by representing solution u as a linear combination of fundamental
solutions. Fundamental solution satisfies the solution of Laplace equation
2 G(r) = 0,

(11)

from which

1
log r,
(12)
2
where r is an euclidian distance between two collocation points. Since fundamental solution has singularity when r = 0, the approximate solution u
can
be represented by a linear combination of fundamental solutions Eq.(13) of the
governing equation with the singularities placed outside the physical domain,
is the fictitious boundary.
where
G(r) =

u
(p) =

N
X
j=1

x j , yj

j G(x xj , y yj ),

(13)

Similarly to Eq.(5) and Eq.(6) we can write

N

u
(p) X
Gj (p)
j
=
,

j=1
and

(14)

2u
(p) X
2 Gj (p)
j
=
,

j=1

1.6

x, y

x, y,

x, y

(15)

MFS-DRM

In Dual Reciprocity Method of Fundamental Solution the solution u

is casted
into homogenous uh and particular up solution. First, particular solution is
solved regardless of boundary conditions using global interpolation of the field.
If f is the source, then we can write poisson equation
N
X
j=1

2 j (p)
= f,
2

x, y

(16)

where j are coefficients to be determined, and f are known values. Once j

coefficients are known, the particular solution over boundary and domain
is constructed as:
up (p) =

N
X

j (p)

(17)

j=1

When the particular solution up is known, the homogenous solution can be

calculated as elaborated in section 1.5, only with changed boundary conditions.
Dirichlet boundary conditions
5

uh (p) = T D (p) up (p)

(18)

Neumann boundary conditions

uh
up
(x, y) = T N (p)
(p)
n
n

(19)

Robin boundary conditions

uh (p) = h T R (p) k

1.7

up
(p) h up (p)
n

(20)

Comparison

Both methods were tested on the two NAFEMS thermal problems described in
Section 1.2 and 1.3. Comparison has been made using several different mesh
discretizations, RBSs in Kansas method and different distances of the fictitious
boundary.
1.7.1

Comparison results for the first problem

For Kansas method the following RBSs has been tested:

= r3
= r5
= r7

= r2 + c2 with c = 0.0025, 0.01, and 0.025

Fictitious boundary distances for MFS are 2, 4, and 6 times of typical mesh
distance. Errors for Kansas method are given in Table 2, errors for MFS are
given in Table 1.
In Figure 3 and 4 the influence of boundary points (with constant number of
internal points in the case of Kansas method) on accuracy is shown.

ku u

Figure 3: Relative error ( = Kurefref 100[%]) for different RBSs. (Number marked as No. of boundary points is number of boundary points on each
boundary)

ku

Figure 4: Relative error ( = M FuSref ref 100[%]) for different distances of

fictitious boundary from real boundary. (Number marked as No. of boundary
points is number of boundary points on each boundary)

N,x
10
20
40
60
10
20
40
60
10
20
40
60
10
20
40
60
10
20
40
60
10
20
40
60

N,y
10
20
30
60
15
30
70
100
10
20
30
60
15
30
70
100
10
20
30
60
15
30
70
100

Ntot
40
80
160
240
50
100
220
320
40
80
160
240
50
100
220
320
40
80
160
240
50
100
220
320

factor
2
2
2
2
2
2
2
2
4
4
4
4
4
4
4
4
6
6
6
6
6
6
6
6

value
19.0395
18.2723
18.2612
18.2649
17.9843
18.2421
18.2488
18.2515
19.7658
18.4503
18.2543
18.2505
17.9828
18.2634
18.2497
18.2518
19.7796
18.5197
18.2723
18.2532
18.0605
18.2584
18.2501
18.2520

[%]
4.3046
0.1016
0.0409
0.0611
-1.4762
-0.0639
-0.2693
-0.0122
8.2834
1.0769
0.0030
-0.0181
-1.4844
0.0527
-0.0221
-0.0106
8.3592
1.4568
0.1018
-0.0030
-1.0585
0.0254
-0.0201
-0.0094

Table 1: List of errors for different boundary discretizations. N,x is the number
of boundary points in x direction, N,y is the number of boundary points in y
direction,

N
10
20
8
10
20
30
10
20
8
10
20
30
10
20
8
10
20
30
20
10
20
30
20
10
20
30
20
10
20
30

N
9
19
29
29
29
29
9
19
29
29
29
29
9
19
29
29
29
29
19
29
29
29
19
29
29
29
19
29
29
29

Ntot
121
441
873
881
921
961
121
441
873
881
921
961
121
441
873
881
921
961
441
881
921
961
441
881
921
961
441
881
921
961

RBS
R3
R3
R3
R3
R3
R3
R5
R5
R5
R5
R5
R5
R7
R7
R7
R7
R7
R7
MQ
MQ
MQ
MQ
MQ
MQ
MQ
MQ
MQ
MQ
MQ
MQ

c2
0.0025
0.0025
0.0025
0.0025
0.01
0.01
0.01
0.01
0.025
0.025
0.025
0.025

value
21.6837
19.3996
18.5316
18.7596
18.8846
19.0223
23.4962
20.7280
18.5550
19.2374
19.1456
19.8282
20.0371
18.7426
18.4751
19.3584
18.8787
18.4766
16.6076
17.8163
17.4038
17.4700
18.3234
18.6241
18.3089
17.5271
17.3716
19.0602
18.3891
16.3260

[%]
18.7904
6.2771
1.5221
2.7713
3.4561
4.2102
28.7199
13.5548
1.6506
5.3885
4.8856
8.6251
9.7699
2.6780
1.2126
6.0516
3.4236
1.2207
-9.0182
-2.3965
-4.6563
-4.2936
0.3815
2.0289
0.3020
-3.9809
-4.8327
4.4179
0.7416
-10.5609

Table 2: List of errors for different mesh definitions and RBSs.

1.7.2

Comparison results for the second problem

For both methods the following RBSs has been tested:

= r3
= r5
= r7
All distances for fictitious boundary for this problem are set to 3 typical mesh
distances in the domain. For the comparison we used different kinds of meshes.
Errors for Kansas method are given in Table 3, errors for MFS are given in
Table 4.
N
10
20
8
10
20
30
10
20
8
10
20
30
10
20
8
10
20
30

N
9
19
29
29
29
29
9
19
29
29
29
29
9
19
29
29
29
29

Ntot
121
441
873
881
921
961
121
441
873
881
921
961
121
441
873
881
921
961

RBS
R3
R3
R3
R3
R3
R3
R5
R5
R5
R5
R5
R5
R7
R7
R7
R7
R7
R7

Abs. error
0.0451
0.0167
0.0164
0.0167
0.0140
0.0095
0.0167
0.0029
0.0014
0.0016
0.0015
0.0010
0.0087
0.0008
0.0014
0.0001
0.0003
0.0002

Table 3: List of errors for Kansas method for different mesh definitions and
RBSs.

10

Figure 5: Relative error (error = kuK uref k) for different RBSs. (Number
marked as No. of points is the total number of points)

Figure 6: Relative error (error = kuM F S uref k) for different distances of

fictitious boundary from real boundary. (Number marked as No. of points is
the total number of points)

11

N
10
20
8
10
20
30
10
20
8
10
20
30
10
20
8
10
20
30

N
9
19
29
29
29
29
9
19
29
29
29
29
9
19
29
29
29
29

Ntot
121
441
873
881
921
961
121
441
873
881
921
961
121
441
873
881
921
961

RBS
R3
R3
R3
R3
R3
R3
R5
R5
R5
R5
R5
R5
R7
R7
R7
R7
R7
R7

Abs. error
0.9025
0.9506
0.8805
0.9025
0.9506
0.0009
0.0017
0.0011
0.0429
0.0237
0.0038
0.0009
0.0032
0.0008
0.0447
0.0247
0.0037
0.0009

Table 4: List of errors for MFS-DRM for different mesh definitions and RBSs.

1.8

conclusions

The first comparison showed, that MFS is far more precise than Kansas method.
For the case of Laplace equation this is quite understandable, since in MFS interpolation is made using analytical solution of Laplace equation, whereas in
Kansas method RBS functions dont have physical background. Another observation is, that in the case of Kansas method it is very hard to predict what
ratio between boundary and domain points is preferable. The reason for this
is, that the definition of the boundary in Kansas method does not give real
information about the problem involved i.e. the definition of the boundary conditions is the same in the case of for example Laplace, Poisson or Navier-Stokes
equations. Due to that fact it can happen, that if there is more boundary points
than domain points, method does not give us the physically sound solution. In
the case of MFS, tests were carried out also for augmented and non augmented
formulation. It turns out, that there is difference only on the 6th digit of the
calculated value.
Probably more comparable results we get in case of Poisson equation where
MFS also needs internal interpolation for particular solution. In the second
comparison we showed, that errors are much more comparable. It should be
stressed out that in cases with 873,881, and 921 number of points, the typical
mesh distance on the boundary is larger than in the domain. That is why the
errors are growing in this area. Again we can observe better consistency of the
results in the case of MFS-DRM, although the minimum error is lower for some
mesh definitions in Kansas method.
The general conclusion is, that in MFS-DRM the results are far more predictable in terms of grid size.

12

References
[1] Li J., Hon, Y.C., Chen C.S.: Numerical comparison of two meshless methods using radial basis functions, Eng. Anal. Bound. Elements, 26, (2002),
pp. 205-225.
[2] Fairweather G., Karageorghis A.: The method of fundamental solution for
elliptic boundary value problems, Advances in Computational Mathematics,
9, (1998), pp. 69-95.
[3] Golberg A. M.: The method of fundamental solutions for Poissons equation, Eng. Anal. Bound. Elements, 16, (1995), pp. 205-213.
[4] Chen C.S., Golberg M.A. and Rashed Y.F.: A mesh free method for linear
diffusion equations, Numerical Heat Transfer, Part B, 33, (1998), 469-486.
[5] Chen C.S., Golberg M.A. , Schaback R.S.: Recent developments of the
dual reciprocity method using compactly supported radial basis functions, to
appear in: Transformation of Domain Effects to the Boundary, ed. Y.F.
Rashed, (2002), WIT Press, Southampton, Boston.

13