Applied Mathematical Modelling 39 (2015) 4979–4994

Contents lists available at ScienceDirect

Applied Mathematical Modelling

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

Finite integration method for solving multi-dimensional partial

differential equations
M. Li b,⇑, C.S. Chen b,c, Y.C. Hon b,d, P.H. Wen a,b
a
School of Engineering and Materials Science, Queen Mary, University of London, London E1 4NS, UK
b
College of Mathematics, Taiyuan University of Technology, Taiyuan, China
c
Department of Mathematics, University of Southern Mississippi, Hattiesburg, MS 39406, USA
d
Department of Mathematics, City University of Hong Kong, Hong Kong, China

a r t i c l e i n f o a b s t r a c t

Article history: Based on the recently developed Finite Integration Method (FIM) for solving
Received 16 November 2013 one-dimensional ordinary and partial differential equations, this paper extends the tech-
Received in revised form 9 February 2015 nique to higher dimensional partial differential equations. The main idea is to extend the
Accepted 18 March 2015
first order finite integration matrices constructed by using either Ordinary Linear
Available online 24 April 2015
Approach (OLA) (uniform distribution of nodes) or Radial Basis Function (RBF) interpola-
tion (uniform/random distributions of nodes) to higher order integration matrices. Using
Keywords:
standard time integration techniques, such as Laplace transform, we have shown that
Finite integration method
Radial basis functions
the FIM is capable for solving time-dependent partial differential equations. Illustrative
Partial differential equation numerical examples are given in two-dimension to compare the FIM (FIM-OLA and
Wave equation FIM-RBF) with the finite difference method and point collocation method to demonstrate
Laplace transformation its superior accuracy and efficiency.
Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction

Mathematical models in terms of partial differential equations (PDEs) have commonly been used to describe a wide vari-
ety of physical phenomena such as sound, heat, electrostatics, electrodynamics, fluid flow, and elasticity. Under various
boundary conditions, it is very rare that these models can be solved in closed form solutions. Numerical methods are
unavoidable for seeking approximate solutions to simulate the dynamics and characteristics of the models. Due to the
advance of computational methods, these kinds of numerical approximation can usually be achieved inexpensively to high
accuracy together with a reliable bound on the error between the analytical solution and its numerical approximation. There
are many numerical techniques available for solving differential equations [1–5] including the Finite Element Method (FEM)
and Boundary Element Method (BEM). In the last decade, the development of the Radial Basis Functions (RBFs) as a truly
meshless method has drawn attention from many researchers. In particular, the use of multiquadric radial basis function
(MQ-RBF) in [6–8] has shown the superior convergence of the method in comparing with FEM and BEM. Li et al. [9] later
compared the accuracy of the RBFs with another two meshless methods, the Method of Fundamental Solutions and Dual
Reciprocity Method. Numerical results indicated that these meshless methods provide a similar optimal accuracy for solving
both elliptic and parabolic equations in 2D.

⇑ Corresponding author.
E-mail address: liming01@tyut.edu.cn (M. Li).

http://dx.doi.org/10.1016/j.apm.2015.03.049
0307-904X/Ó 2015 Elsevier Inc. All rights reserved.
4980 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

Recently, Wen et al. [10] developed a Finite Integration Method (FIM) for solving differential equation in 1D and demon-
strated its applications to nonlocal elasticity problems [11]. It has been shown that the FIM gives higher degree of accuracy
than the Finite Difference Method (FDM) and Point Collocation Method (PCM). In this paper, the FIM is further extended to
solve multi-dimensional partial differential equations. In this paper, two-dimensional partial differential equations are given
for static and dynamic problems as illustrative examples. Similar to the FDM and the PCM, a finite number of points, known
as field points, are distributed in the computational domain. The field points are generated either uniformly (grid) along the
independent coordinate or randomly in the domain. The integration matrix of the first order is obtained by the direct inte-
gration with either OLA approximation or RBFs interpolation. Based on these first order integration matrices, any finite inte-
gration matrix with multi-layer integration can easily be obtained. To compare with other numerical methods, the PCM and
analytical solutions are used. For two-dimensional time-dependent problems, Laplace transform has been applied to remove
the time-dependent variable. The differential equation is then solved in the Laplace space with given initial and boundary
conditions. To demonstrate the accuracy and efficiency of the FIM, several numerical examples are given.

2. The FIM for one-dimensional problems

Numerical quadrature rule based on Ordinary Linear Approach (OLA) is the simplest computational scheme for integra-
tion [10]. Starting from one-dimension problem, an integral of a given function UðxÞ can be written as
Z x
UðxÞ ¼ uðnÞdn: ð1Þ
0

Applying the linear interpolation technique to (1), we have

Z xk X
k
Uðxk Þ ¼ uðnÞdn ¼ aki uðxi Þ; ð2Þ
0 i¼1

where, using trapezoidal rule,

a1i ¼ 0; ð3Þ
8
> 0:5D; i ¼ 1;
>
>
< D; i ¼ 2; 3; :::; k  1;
aki ¼ ð4Þ
> 0:5D;
> i ¼ k;
>
:
0; i > k;
and xi ¼ D  ði  1Þ; D ¼ b=ðN  1Þ; i ¼ 1; 2; ::; N are nodal points in [0, b], and x1 ¼ 0; xN ¼ b. Note that (2) can be written in a
matrix form as
U ¼ Au; ð5Þ
T T
where U ¼ ½U 1 ; U 2 ; :::; U N  , u ¼ ½u1 ; u2 ; :::; uN  , the first order integration matrix
0 1
0 0 0 0 0 0
B 1=2 1=2 0 0 0 0 C
B C
B C
B 1=2 1 1=2 0 0 0 C
A ¼ ðaki Þ ¼ DB
B 1=2
C
C ; ð6Þ
B 1 1 1=2 0 0 C
B C
@ ::: ::: ::: ::: ::: ::: A
1=2 1 1 1 1 1=2 NN

and U i ¼ Uðxi Þ; ui ¼ uðxi Þ are the values of integration and the integral function respectively at each nodes.
Thereafter, consider a multi-integral for one-dimensional problem
Z x Z f
U ð2Þ ðxÞ ¼ uðnÞdndf; x 2 ½0; b: ð7Þ
0 0

Applying the above OLA technique again for integral function U ð2Þ ðxÞ, we have
Z xk Z f X
k X
i X
k
ð2Þ
U ð2Þ ðxk Þ ¼ uðnÞdndf ¼ aki aij uðxi Þ ¼ aki uðxi Þ: ð8Þ
0 0 i¼1 j¼1 i¼1

The above multi-integral can also be written in a matrix form as

Uð2Þ ¼ Að2Þ u ¼ A2 u; ð9Þ

where
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4981

0 1
0 0 0 0 0 0
B 1=4 1=4 0 0 0 0 C
B C
B C
B
2 B 3=4 1 1=4 0 0 0 C
Að2Þ ¼ AA ¼ D B C ; ð10Þ
C
B 5=4 2 1 1=4 0 0 C
B C
@ ::: ::: ::: ::: ::: ::: A
½1 þ 2ðN  2Þ=4 N  2 N  3 ::: 1 1=4 NN

ð2Þ
and the elements of matrix A are [10,11]
ð2Þ
a1i ¼ 0;
8
>
> ½1 þ 2ðk  2ÞD2 =4; i ¼ 1;
>
>
< ðk  iÞD2 ; i ¼ 2; 3; :::; k  1; ð11Þ
ð2Þ
aki ¼
> D2 =4;
> i ¼ k;
>
>
:
0; i > k:
For two-dimensional problems, let us consider a uniform distribution of collocation points as shown in Fig. 1. Similar to
(1), we define
Z x Z xk
U x ðx; yÞ ¼ uðn; yÞdn; U x ðxk ; yk Þ ¼ uðn; yk Þdn; ð12Þ
0 0

and the total number of point is k ¼ N 1 ðj  1Þ þ i, where i and j denote the number of column and the number of row respec-
tively. This numbering system is called the global number system. We can also express each nodal value of integration in
(12) in a matrix form as
Ux ¼ Ax u; ð13Þ
T T
where integral nodal value U x ¼ ½U x1 ; U x2 ; :::; U xM  , nodal value u ¼ ½u1 ; u2 ; . . . ; uM  and M is the total number of collocation
points (M ¼ N 1  N 2 for grid shown in Fig. 1). For a rectangular domain, the first order integration matrix
0 1
A 0 ::: 0
B0 A 0 0 C
B C
Ax ¼ B C;
@ ::: ::: ::: ::: A ð14Þ
0 0 0 A
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
N1 N2

in which, A is integration matrix for one-dimension given in (6) with dimension N 1  N 1 . Similarly, the integration along y
axis is

y
N2

column i
……

1 2 ( xk , y k ) N1

row j
.......
.......

2
x

Fig. 1. Uniform distribution of collocation points and the number of nodes.

4982 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

Z y Z yk
U y ðx; yÞ ¼ uðx; gÞdg; U y ðxk ; yk Þ ¼ uðxk ; yÞdy; ð15Þ
0 0

which can be written in the matrix form as

Uy ¼ A0 u; ð16Þ
in the local system for the collocation points, where k ¼ N 2 ði  1Þ þ j. The first order integration matrix in the local system is
0 1
A 0 ::: 0
B0 A 0 0 C
B C
A0 ¼ B C;
@ ::: ::: ::: ::: A ð17Þ
0 0 0 A
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
N1 N 2

in which A is the integration matrix for one-dimension integral given in (6) with dimension N 2  N 2 . By a simple
re-arrangement of the number of the nodes, (16) can be rewritten, in the global system, as
Uy ¼ Ay u: ð18Þ
For the multi-integration in two-dimensional problem in a rectangular domain, we consider the following integral with
respect to coordinate x
Z xi Z f
U ð2Þ
x ðxi ; yÞ ¼ uðn; yÞdndf; xi 2 ½0; b1 ; y 2 ½0; b2 ; ð19Þ
0 0

and use the same procedure for one-dimension, one has

Z xk Z f X
k X
i
U ð2Þ ðxk ; yk Þ ¼ uðn; yk Þdndf ¼ ðakj Þx ðaji Þx ui ; ð20Þ
0 0 i¼1 j¼1

or in a matrix form
2
Uð2Þ
x ¼ Ax u; ð21Þ
where
0 1
A2 0::: 0
B C
B0 A2
0 0 C
A2x ¼ Ax Ax ¼ B
B
C;
C
@ ::: ::: ::: ::: A ð22Þ
2
0 0 0 A
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
N1 N 2

Similarly, one has multi-integration U ð2Þ

y ðx; yÞ with respect to coordinate y

Z yk Z s X
k X
i
U ð2Þ
y ðxk ; yk Þ ¼ uðxk ; gÞdgds ¼ ðakj Þy ðaji Þy ui ; ð23Þ
0 0 i¼1 j¼1

and
2
Uð2Þ
y ¼ Ay u: ð24Þ

This method can be extended to the higher order integrations, i.e.

Z xk Z ::: Z yk Z :::
U ðmÞ
x ðxk ; yk Þ ¼ ::: uðn1 ; yk Þdn1 :::dnm ; U yðmÞ ðxk ; yk Þ ¼ ::: uðxk ; g1 Þdg1 :::dgm
0 0 0 0
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} ð25Þ
m layers m layers

xk 2 ½0; b1 ; yk 2 ½0; b2 :

Applying ordinary linear interpolation technique again for integral function U ðmÞ ðx; yÞ, we have
Z xk Z ::: X
M X
M X
M
ðmÞ
U xðmÞ ðxk ; yk Þ ¼ ::: uðn1 ; yk Þdn1 :::dn1 ¼ ::: ðakj Þx :::ðani Þx uðxi ; yi Þ ¼ ðaki Þx ui ; ð26Þ
0 0 i¼1 j¼1 i¼1
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4983

Z yk Z ::: X
M X
M X
M
ðmÞ
U yðmÞ ðxk ; yk Þ ¼ ::: uðxk ; g1 Þdg1 :::dgm ¼ ::: ðakj Þy :::ðani Þy uðxi ; yi Þ ¼ ðaki Þy ui ; ð27Þ
0 0 i¼1 j¼1 i¼1

Again, it can also be written, in a matrix form, as

UxðmÞ ¼ Am
x u; UðmÞ
y ¼ Am
y u: ð28Þ

In addition, this method can be extended to multi-layers integration with two coordinates x and y as follow:
Z xk Z ::: Z yk Z :::
U ðmnÞ ðxk ; yk Þ ¼ ::: ::: uðn; gÞdn1 :::dnn dg1 :::dgm xk 2 ½0; b1 ; yk 2 ½0; b2 ; ð29Þ
0 0 0 0 |fflfflfflfflffl{zfflfflfflfflffl} |fflfflfflfflfflffl{zfflfflfflfflfflffl}
nlayer mlayer

and the nodal values of the above integration are obtained in the matrix form as

UðmnÞ ¼ Am n
x Ay u: ð30Þ

3. The FIM with radial basis functions

For uniform distribution of nodes (grid), the multi-layer integrations at each node can be obtained quite easily in a matrix
form. However, in general case, if the nodes distribution is random, the algorithm OLA discussed in the Section 2 is not valid.
In this case, interpolation schemes have to be introduced. Recently, the radial basis functions interpolation schemes and
moving least squares method are very popular meshless methods. For example, the MQ-RBF was introduced by Hardy [7]
for the interpolation of topographical surfaces in the early stage of radial bases function application. Note that uðxÞ in the
domain X can be interpolated over a number of randomly distributed nodes xi ¼ ðxi ; yi Þ; i ¼ 1; 2; . . . ; M, as
X
M X
Q
uðxÞ ¼ Ri ðx; xi Þai þ Pq ðxÞbq ¼ RðxÞa þ PðxÞb; x 2 X; ð31Þ
i¼1 q¼1

where RðxÞ ¼ ½R1 ðx; x1 Þ; R2 ðx; x2 Þ; :::; RM ðx; xM Þ is a set of radial basis functions centred at x ¼ ðx; yÞ, a ¼ ½a1 ; a2 ; . . . ; aM T and
b ¼ ½b1 ; b2 ; . . . ; bQ T are the coefficients to be determined. For uniqueness, the polynomial terms Pq(x) satisfies the following
additional requirement
X
M
Pq ðxi Þai ¼ 0; q ¼ 1; 2; . . . ; Q : ð32Þ
i¼1

From (31) and (32), we have

R0 a þ P0 b ¼ u; PT0 a ¼ 0; ð33Þ
where
2 3 2 3
R1 ðx1 ; x1 Þ R2 ðx1 ; x2 Þ ::: RM ðx1 ; xM Þ P1 ðx1 Þ P 2 ðx1 Þ ::: PQ ðx1 Þ
6 R ðx ; x Þ R2 ðx2 ; x2 Þ ::: RM ðx2 ; xM Þ 7 6 P1 ðx2 Þ P 2 ðx2 Þ ::: PQ ðx2 Þ 7
6 1 2 1 7 6 7
6 7 6 7
6: : ::: : 7 6: : ::: : 7
R0 ¼ 6
6:
7;
7 P0 ¼ 6
6:
7;
7 ð34Þ
6 : ::: : 7 6 : ::: : 7
6 7 6 7
4: : ::: : 5 4: : ::: : 5
R1 ðxM ; x1 Þ R2 ðxM ; x2 Þ ::: RM ðxM ; xM Þ P1 ðxM Þ P 2 ðxM Þ ::: PQ ðxM Þ
T
and u ¼ ½uðx1 Þ; uðx2 Þ; :::; uðxM Þ . Eq. (33) gives
1 1
b ¼ ðPT0 R1 T 1
0 P0 Þ P0 R 0 u; a ¼ R1 T 1 T 1
0 ½I  P0 ðP0 R 0 P0 Þ P0 R 0  u: ð35Þ
Substituting the coefficients a and b from (35) into (31), we have
h i X
M
T 1 T 1 T 1 T 1 1 T 1
uðxÞ ¼ RðxÞR1
0 ðI  P0 ðP0 R 0 P0 Þ P0 R 0 Þ þ PðxÞðP0 R 0 P0 Þ P0 R 0 u¼ /i ðxÞui ; ð36Þ
i¼1

where ui ðxÞ is the shape function. From (36), we have

@u X M XQ XN
¼ Ri;x ðxÞai þ Pq;x ðxÞbq ¼ /i;x ðxÞui or u;x ¼ Dx u; ð37Þ
@x i¼1 q¼1 i¼1
4984 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

@u X M XQ XN
¼ Ri;y ðxÞai þ Pq;y ðxÞbq ¼ /i;y ðxÞui or u;y ¼ Dy u: ð38Þ
@y i¼1 q¼1 i¼1

Furthermore,
Z X
M X
Q X
M
uðxÞdx ¼  xi ðxÞai þ
R  xq ðxÞb þ f ðyÞ ¼
P  xi ðxÞui þ f ðyÞ;
/ ð39Þ
q 0 0
i¼1 q¼1 i¼1

and
Z X
M X
Q X
M
uðxÞdy ¼  yi ðxÞai þ
R P yq ðxÞbq þ g 0 ðxÞ ¼  yi ðxÞui þ g ðxÞ:
/ ð40Þ
0
i¼1 q¼1 i¼1

where
Z Z Z Z
 xi ðxÞ ¼
R Ri ðxÞdx;  xq ðxÞ ¼
P Pq ðxÞdx;  yi ðxÞ ¼
R Ri ðxÞdy; Pyq ðxÞ ¼ Pq ðxÞdy; ð41Þ

Z Z
 xi ¼
/ /i ðxÞdx;  yi ¼
/ /i ðxÞdy; ð42Þ

and f 0 ðyÞ and g 0 ðxÞ are arbitrary functions. Therefore, the integration matrices of the first order are
0  x2 ðx1 Þ  xM ðx1 Þ 1
/x1 ðx1 Þ / ::: /
B/  xM ðx2 Þ C
B  x1 ðx2 Þ  x2 ðx2 Þ
/ ::: / C
Ax ¼ B C ; ð43Þ
@ ::: ::: ::: ::: A
 x1 ðxM Þ
/  x2 ðxM Þ ::: /
/  xM ðxM Þ
MM

and
0  y2 ðx1 Þ  yM ðx1 Þ 1
/y1 ðx1 Þ / ::: /
B/  yM ðx2 Þ C
B  y1 ðx2 Þ  y2 ðx2 Þ
/ ::: / C
Ay ¼ B C : ð44Þ
@ ::: ::: ::: ::: A
 y1 ðxM Þ
/  y2 ðxM Þ ::: /
/  yM ðxM Þ
MM

For multi-layer integration, the integration matrix can be obtained from (30). Note that the double layer integration can
be obtained analytically in the following algorithm as
Z Z X  ðxÞa þ XP ðxÞb þ xf ðyÞ þ f ðyÞ ¼ X/
M Q M
uðxÞdxdx ¼ R  ðxÞu þ xf ðyÞ þ f ðyÞ; ð45Þ
xi i xq q 0 1 xi i 0 1
i¼1 q¼1 i¼1

and
Z Z X  ðxÞa þ XP ðxÞb þ yg ðxÞ þ g ðxÞ ¼ X/
M Q M
uðxÞdydy ¼ R  ðxÞu þ yg ðxÞ þ g ðxÞ; ð46Þ
yi i yq q 0 1 yi 0 1
i¼1 q¼1 i¼1

where
 ðxÞ ¼ R R R ðxÞdxdx;
R P xq ðxÞ ¼
RR
Pq ðxÞdx;
xi i
ð47Þ
 ðxÞ ¼ R R R ðxÞdydy;
R 
Pyq ðxÞ ¼
RR
Pq ðxÞdydy;
yi i

and
Z Z Z Z
 ¼
/ /i ðxÞdxdx;  yi ¼
u /i ðxÞdydy; ð48Þ
xi

and f i ðyÞ and g i ðxÞ; i ¼ 0; 1; are arbitrary functions. Therefore, the integration matrices of multi- integration matrix are as
follows:
0 1
 ðx Þ
/  ðx Þ
/  ðx Þ
::: /
x1 1 x2 1 xM 1
B  ðx Þ  ðx Þ C
B /x1 ðx2 Þ / ::: / C
Bx ¼ B
B
x2 2 xM 2 C
C ; ð49Þ
@ ::: ::: ::: ::: A
 ðx Þ
/  ðx Þ ::: /
/  ðx Þ
x1 M x2 M xM M MM

and
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4985

0 1
 ðx Þ
/  ðx Þ
/  ðx Þ
::: /
y1 1 y2 1 yM 1
B  ðx Þ  ðx Þ C
B /y1 ðx2 Þ / ::: / C
By ¼ B
B
y2 2 yM 2 C
C : ð50Þ
@ ::: ::: ::: ::: A
 ðx Þ
/  ðx Þ ::: /
/  ðx Þ
y1 M y2 M yM M MM

It is worth noting that these two integration matrices can replace integration matrix Ax and Ay in (43) and (44) respec-
tively. Three types of radial bases functions, i.e. MQ, linear, and Thin-Plate Splines (TPS) are observed in this paper. Their
derivatives and integrals in equations above are presented in the Appendix. For the mth order derivative, we have the rela-
tionship between higher order derivative matrix and first order derivative matrix as
@m @nu
¼ Dm n
x Dy u; ð51Þ
@xm @yn
which can be used by the Point Collocation Method (PCM) directly. Unlike the OLA in Section 2, the integral matrices and the
derivative matrices with any order using radial basis functions, AðmÞ , B and DðmÞ , are full ranked matrices.

4. The FIM for multi-dimensional problems

The FIM described in Sections 2 and 3 is readily extendable to solving higher dimensional problems. For illustration, con-
sider the following two-dimensional partial differential equation

@2u @2u
a1 ðx; yÞ þ a2 ðx; yÞ 2 þ a3 ðx; yÞu ¼ bðx; yÞ; x 2 X;
@x 2 @y ð52Þ
K½uðx; yÞ ¼ hðx; yÞ; x 2 @ X;
where K is a boundary operator, a1 ðx; yÞ; a2 ðx; yÞ; a3 ðx; yÞ; bðx; yÞ and hðx; yÞ are given functions. u is generally referred as
potential, which represents the transversal displacement of a membrane. X and @ X are simple connected domain and its
boundary respectively. Integrating twice in (52) with respect to coordinates x and y respectively, one has
ZZ ZZ " #
@2u @2u
a1 ðx; yÞ 2 þ a2 ðx; yÞ 2 þ a3 ðx; yÞu dxdxdydy
@x @y
ZZ ZZ
¼ bðx; yÞdxdxdydy þ xf0 ðyÞ þ f 1 ðyÞ þ yg 0 ðxÞ þ g 1 ðxÞ; ð53Þ

where f 0 ðyÞ; f 1 ðyÞ; g 0 ðxÞ and g 1 ðxÞ are unknown one-dimensional functions. Using the technique of integration by part, we
have
ZZ " Z #
ZZ ZZ " Z ZZ #
@ 2 a1
@ a1 @ a2 @ 2 a2
a1 u  2 u dxdx
dx þ u
dydy þ a 2 u  2 u dx þ u dydy dxdx
@x2
@x @y @y2
ZZ ZZ ZZ ZZ
þ a3 udxdxdydy ¼ bðx; yÞdxdxdydy þ xf0 ðyÞ þ f 1 ðyÞ þ yg 0 ðxÞ þ g 1 ðxÞ: ð54Þ

By using integration matrix mentioned in the previous sections, we have

h i
A2y a1 þ A2x a2  2Ax Ay ðAy a1;x þ Ax a2;y Þ þ A2x A2y ða1;xx þ a2;yy þ a3 Þ u ¼ A2y A2x b þ XWy f 0 þ Wy f 1 þ YWx g0 þ Wx g1 ; ð55Þ

1 2 r T 1 2 r T T T
where X ¼ fx1 ; x2 ; :::; xM g; Y ¼ fy1 ; y2 ; :::; yM g; f 0 ¼ ½f 0 ; f 0 ; :::; f 0  ; f 1 ¼ ½f 1 ; f 1 ; :::; f 1  ; g0 ¼ ½g 10 ; g 20 ; :::; g p0  ; g1 ¼ ½g 11 ; g 21 ; :::; g p1  , p
and r are numbers of point to be used for interpolation of functions f(y) and g(x) respectively, Wx and Wy are matrices of
one-dimensional shape functions with respect to coordinates x and y respectively, and
0 1
al ðx1 Þ 0 ::: 0
B 0 al ðx2 Þ ::: 0 C
al ¼ B
B
C
C l ¼ 1; 2; 3: ð56Þ
@ ::: ::: ::: ::: A
0 0 ::: al ðxM Þ
Integral functions f 0 ðyÞ; f 1 ðyÞ; g 0 ðxÞ and g 1 ðxÞ can be interpolated in terms of the nodal values in the following procedure:
(1) Determine the regions of functions f(y) and g(x), i.e. ½y 1 ; y
r , ½
x1 ; 
xp , and uniformly distributed points in these regions as
shown in Fig. 2;
(2) Determine one-dimensional shape function matrices Wx and Wy
(a) Ordinary Linear Approximation (OLA)
By using linear interpolation, one has
4986 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

Fig. 2. Interpolations for one dimensional function f(y) and g(x).

f ðyÞ ¼ 0 if 1 < y < y

y m1 or ym < y < y r ;
ð57Þ
gðxÞ ¼ 0 if x1 < x < xn1 or xn < x < xp ;

m  y
y yy m1
f ðyÞ ¼ f þ f m1 < y < y
if y m ;
m  y
y m1 m1 y m  y m1 m
ð58Þ
xn  x x  xn1
gðxÞ ¼ g þ g if xn1 < x < xn :
xn  xn1 n1 xn  xn1 n
Therefore, the matrices of shape function are

ð59Þ

and

: ð60Þ

(b) Radial basis functions

In this case, two one-dimensional shape function matrices with the radial bases function interpolation Ux and Uy are

ð61Þ

and

; ð62Þ

column 1

in which wi ðxÞ and wi ðyÞ are shape functions in one dimensional case as shown in Fig. 2. In (55), we have M nodal unknowns
of u, 2q unknowns of f 0 , f 1 and 2r unknowns of g0 , g1 . For a rectangular plate with uniform distribution of nodes ðN 1  N 2 Þ,
obviously one has 2p þ 2r nodes located on the boundary. By selecting p ¼ N 1  1 and r ¼ N 2  1 for uniform distribution of
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4987

node, there are N 1  N 2 þ 2ðN 1 þ N 2  2Þ linear system of equations to determine all unknowns, i.e. u, f 0 , f 1 , g0 , and g1 . In fact,
the number of boundary points to determine four one-dimensional integral functions is arbitrary. The number of points (L)
on the boundary should be greater than or equal to 2ðp þ rÞ. If L ¼ 2ðp þ rÞ, the standard Gaussian solver can be used directly.
Otherwise, the Singular Value Decomposition (SVD) [12] scheme should be introduced.
For partial differential equation with constants coefficients in (52), i.e. a1 ¼ a2 ¼ 1, using double layer integrations in (49)
and (50), we have
½By þ Bx þ Bx By a3 u ¼ Bx By b þ XWy f 0 þ Wy f 1 þ YWx g0 þ Wx g1 : ð63Þ
Under this circumstance, the first order integration matrices Ax and Ay are not necessary.

5. The FIM for Time-dependent problems

For two-dimensional dynamics, the governing equation is written, in time domain, as

@2u @2u @u @2u

a1 ðxÞ 2
þ a2 ðxÞ 2 þ AðxÞ þ BðxÞ 2 þ a3 ðxÞu ¼ bðx; tÞ; t > 0; x 2 X; ð64Þ
@x @y @t @t
where a1 ðxÞ; a2 ðxÞ; a3 ðxÞ; AðxÞ; BðxÞ; and bðx; tÞ are given functions. In addition, the boundary conditions and initial conditions
are specified as
_
uðx; 0Þ ¼ d1 ðxÞ; uðx; 0Þ ¼ d2 ðxÞ; x 2 X;
ð65Þ
K½uðx; tÞ ¼ eðx; tÞ; x 2 @ X;
where d1 ðxÞ, d2 ðxÞ; and eðx; tÞ are given functions. By applying the Laplace transform to (64) with consideration of the initial
conditions, one obtains

@2u
~ @2u
~ ~
a1 ~  d1  þ B½s2 u
þ a2 2 þ A½su ~  sd1  d2  þ a3 u
~ ¼ b; ð66Þ
@x 2 @y
where the Laplace transform is defined as
Z 1
~f ¼ f ðtÞest dt: ð67Þ
0

From (66), we have

@2u
~ @2u
~ ~ ~
a1 2
þ a2 2 þ E ~ ¼ h;
u ð68Þ
@x @y
where
~ sÞ ¼ ½AðxÞs þ BðxÞs2 þ a3 ðxÞ;
Eðx;
ð69Þ
~ sÞ ¼ bðx;
hðx; ~ sÞ þ AðxÞd1 ðxÞ þ BðxÞ½sd1 ðxÞ þ d2 ðxÞ:

The numerical procedure of solving (68) is the same as solving (52) in Section 4. Applying the integration operation twice
on (68), we have
h i
A2y a1 þ A2x a2  2Ax Ay ðAy a1;x þ Ax a2;y Þ þ A2x A2y ða1;xx þ a2;yy þ EÞ
~ u ~ þ XUy f 0 þ Uy f 1 þ YUx g þ Ux g ;
~ ¼ A2y A2x h ð70Þ
0 1

where 0 1 0 1
~ 1 ; sÞ 0
Eðx ::: 0 ~ 1 ; sÞ
hðx 0 ::: 0
B C B C
B ~ 2 ; sÞ ::: 0 C ~ sÞ ¼ B
~ 2 ; sÞ ::: 0 C
~ sÞ ¼ B 0
Eðx;
Eðx C; hðx; B0 hðx C: ð71Þ
B C B C
@ ::: ::: ::: ::: A @ ::: ::: ::: ::: A
0 0 ~ M ; sÞ
::: Eðx 0 0 ~ M ; sÞ
::: hðx
If a1 ¼ a2 ¼ B ¼ 1; A ¼ a; d1 ¼ d2 ¼ b ¼ 0, Eq. (70) becomes

½A2y þ A2x  ða þ sÞsA2x A2y u

~ ¼ XUy f 0 þ Uy f 1 þ YUx g0 þ Ux g1 : ð72Þ

By using double layer integration matrix, (72) becomes

~ ¼ XUy f 0 þ Uy f 1 þ YUx g0 þ Ux g1 :
ðBy þ Bx  ða þ sÞsBx By Þu ð73Þ
All unknowns including nodal values of potential u and four one-dimensional functions f 0 ; f 1 and g 0 ; g 1 can be obtained
numerically in the Laplace transform space for each specified Laplace parameter sk . If (K + 1) samples in the transformation
~ ðxi ; sk Þ, are evaluated by
space sk ; k ¼ 0; 1; . . . ; K, are selected in the Laplace transform domain, the transformed variables i.e. u
the numerical procedure above. Then u(x, t) in the time domain can be determined by Laplace inversion techniques. A simple
and accurate method proposed by Durbin [13] is adopted in this paper. The formula of Durbin inversion scheme is given by
4988 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

2egt 1 ~ XK
f ðtÞ ¼ ½ f ðgÞ þ Ref~f ðg þ 2kpi=TÞe2kpti=T g ð74Þ
T 2 k¼0

where ~f ðsk Þ denotes the transformed variable in the Laplace domain, the parameter of the Laplace transform is chosen as:
pffiffiffiffiffiffiffi
sk ¼ g þ 2kp i=T ð i ¼ 1Þ. There are two free normalised parameters in sk : g and T. The selection of parameters T depends
on the observing period in the time domain.

6. Numerical examples

Example 6.1. We firstly consider the following partial differential equation

@2u @2u
þ þ ku ¼ sin px sin py; ðx; yÞ 2 X;
@x2 @y2 ð75Þ
uðx; yÞ ¼ 0; ðx; yÞ 2 @ X;
where X [ @ X ¼ ½0; 1  ½0; 1.
For the OLA, uniform distribution of nodes is selected (N 1 ¼ N 2 ¼ 21, M ¼ 441, q ¼ r ¼ N 1 ). For radial basis functions
approach in Eqs. (61–63), three radial basis functions are considered, i.e.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(1) MQ: RðrÞ ¼ c2 þ r2 ;
(2) Linear (R): RðrÞ ¼ r;
(3) Thin-Plate Splines (TPS): RðrÞ ¼ r2 ln r.

Double layer integration scheme given in (73) is used. For the choice of the free parameter c in [10], it was observed that a
stable and high accurate solution can be obtained when c  1:8 for one-dimensional problems. It is remarked here that a
theoretical justification on the choice of this free parameter c in MQ was recently given by Luh [14]. In this paper, we simply
select this parameter as c ¼ 1=N 1 and the number of polynomial terms is Q ¼ 6 for all examples in this section. The analytical
solution of (75) is given by
sin px sin py
u ðx; yÞ ¼ : ð76Þ
k  2p 2
The average relative error is defined as

1X M
jui  ui j 1
e¼ ; umax ¼ : ð77Þ
M i¼1 jumax j k  2p2

To demonstrate the accuracy of the proposed method, we compare the numerical results with the Point Collocation
Method (PCM) and the Finite Difference Method (FDM). The logarithms of the average error e to base 10 varied with two
kinds of node distributions as shown in Fig. 3, i.e. uniform and randomly distributions (441 nodes for uniform grid and
102 nodes for random grid). Numerical results of average relative error are shown in Fig. 4 and Fig. 5 using these two dis-
tributions of node respectively. Due to the singularity at k ¼ 2p2 , the relative errors are increased significantly near this point
for each case. In addition, the accuracy of the FIM is higher than that of the PCM and FDM in most cases. However, the accu-
racy is also different for using other interpolation schemes in the FIM. The OLA is the simplest method, but suffers the low
accuracy using the FIM as shown in Fig. 4. Also in the selection of radial basis functions, the MQ and the TPS are apparently
superior to others. However, there is a free parameter in the MQ which has significant effect on the accuracy and has to be
selected carefully. Fig. 5 shows the logarithms of the average error e with irregular distribution of nodes in the domain. The
conclusion is similar to that for the uniform node distribution. The accuracy of the PCM is the lowest and very close to that by
the FIM with linear RBF. For the irregular node distribution, the method of OLA is not available. To observe the effect of the
selection of shape parameter c in MQ, the results using two solvers, i.e. SVD and Gaussian algorithms, are presented in Fig. 6
for the case k = 10. It is shown that the differences between these two solvers are small. In addition, we can conclude that the
reasonable degree of accuracy can be obtained for c < 0:35 for two-dimensional problems using FIM. Comparing the com-
putational effort, the CPU times used by PCM, FDM and FIM are very close as there are no any special function need to be
calculated in the system matrix and also the Gaussian linear algebraic equation solver is used.

Example 6.2. In this example, consider the following partial differential equation

@2u @2u
xð1  xÞ þ yð1  yÞ 2 ¼ 4xyð1  xÞð1  yÞ; ðx; yÞ 2 X;
@x 2 @y ð78Þ
uðx; yÞ ¼ 0; ðx; yÞ 2 @ X;
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4989

Fig. 3. Distributions of nodes: (a) uniform; (b) random.

k
0.0
0 5 10 15 20 25 30

-1.0 Series1
OLA Series2
MQ Series3
R

Series4
TPS Series5
PCM Series6
FDM

-2.0

-3.0
Log ( ε)

-4.0

-5.0

-6.0 2π 2

-7.0

Fig. 4. The logarithms of the average errors versus parameter k for different method (regular distribution).

k
0.0
0 5 10 15 20 25 30
-0.5
MQF
MQ L
RF

-1.0
PSF
TPS PC
PCMM

-1.5
Log(ε)

-2.0

-2.5

-3.0 2π 2
-3.5

-4.0

Fig. 5. The The average errors versus parameter k for different methods (irregular distribution of nodes).
4990 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

where X [ @ X ¼ ½0; 1  ½0; 1. The analytical solution is given by u ðx; yÞ ¼ xyð1  xÞð1  yÞ . The average relative error is
defined as

1X M
jui  ui j 1
e¼ ; umax ¼ u ð0:5; 0:5Þ ¼ : ð79Þ
M i¼1 jumax j 16

In this example, we chose N 1 ¼ N 2 ¼ N and p ¼ r ¼ N. Similar to the previous example, the shape parameter c of MQ is
selected as c = 1 / N. For the RBF approach, single integration matrix in (70) is used. The average errors e for various number
of collocation point are shown in Table 1. Among these algorithms, the accuracy of OLA is the lowest and PSF of radial basis
function is the highest.

Example 6.3. Consider the following partial differential equation with a unit circle domain

@2u @2u
þ ¼ 12xy; ðx; yÞ 2 X;
@x2 @y2 ð80Þ
uðx; yÞ ¼ 0; ðx; yÞ 2 @ X;
where X [ @ X ¼ fðx; yÞ : x2 þ y2  1g:
The analytical solution is given by u ðx; yÞ ¼ xyð1  x2  y2 Þ. For implementation, three nodal densities with node
numbers M = 93, 343 and 747 and the node numbers for determining four one-dimensional functions p = r = 11, 21 and 31
are given respectively. The distribution of node in a unit circle domain is shown in Fig. 7 for M = 343. To solve the linear
system of equations, the SVD is used [12]. In this example, TPS is employed. It is because TPS has no shape parameter and its
degree of accuracy is very close to that by MQ. The numerical solutions are plotted in Fig. 8 and the exact solution is given for
comparison. The average error is defined as

1X m
e¼ jui ðxi ; yi Þ  u ðxi ; yi Þj; ð81Þ
m i¼1
where the test points are fðxi ; yi Þ : xi ¼ yi ¼ ði  1Þ=ðm  1Þ cos p=4; i ¼ 1; 2; :::; m; m ¼ 21g. The average errors e using various
number of collocation points M are shown in Table 2.

Example 6.4. Finally, we consider the following scale wave equation with initial and boundary conditions

@2u @2u @u @ 2 u
þ a  ¼ 0; t > 0; ðx; yÞ 2 X;
@x2 @y2 @t @t 2
_ y; 0Þ ¼ 0;
uðx; y; 0Þ ¼ uðx; ðx; yÞ 2 X [ @ X; ð82Þ
u;x ð0; y; tÞ ¼ u;x ð1; y; tÞ ¼ uðx; 0; tÞ ¼ 0; 0  x  1; 0  y  1;
u;y ðx; 1; tÞ ¼ HðtÞ; 0  x  1;
where X [ @ X ¼ ½0; 1  ½0; 1; HðtÞ is Heaviside function, a is the damping factor. Therefore, the boundary condition of the
last equation of Eq. (82) in the Laplace transform domain is u ~ ;y ðx; 1; sÞ ¼ 1=s. Uniform distribution of node is considered using
N 1 ¼ N 2 ¼ 20 and p ¼ r ¼ 20 for one-dimensional functions. The number of samples in the Laplace space is K = 100. In this

1.0

SSVD
eries1
0.0
SGaussian
eries2
-1.0
Series3
-2.0
Log(ε)

<1%
-3.0

-4.0

-5.0

-6.0

-7.0
0.00 0.10 0.20 0.30 0.40 0.50
c

Fig. 6. The logarithms of the average error versus shape parameter c for two solvers for the case k = 10.
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4991

Table 1
Average errors (e) for different interpolation methods.

N OLA MQF LF PSF

10 0.019110 0.013707 0.017554 0.012985
20 0.005689 0.005365 0.006462 0.004496
30 0.013820 0.003083 0.003671 0.002373

θ
x
-1 +1

Fig. 7. Nodal distribution (M = 343) in a unit circle domain.

example, two free parameters in (74) are selected as g ¼ 5 and T = 20. The displacements using MQ and TPS at two locations
(x = 1 and x = 0.5 if y = 0.5) are presented in Fig. 9 without damping (a = 0). Double layer integration scheme in (73) is used.
Good agreement has been observed for the time-dependent displacement with analytical solution. In this case, the analytical
solution can be presented by a periodic function, at the top of the plate, as

t; 0  t  2;
u ð1; tÞ ¼ u ð1; t þ 4Þ ¼ u ð1; tÞ; ð83Þ
4  t; 2  t  4;
and
8
>
> 0; 0  t  0:5;
>
>
>
< t  0:5;
> 0:5  t  1:5;
u ð0:5; tÞ ¼ 1; 1:5  t  2:5; u ð0:5; t þ 4Þ ¼ u ð0:5; tÞ: ð84Þ
>
>
>
> 3:5  t; 2:5  t  3:5;
>
>
:
0; 3:5  t  4;

0.16

0.14

0.12

0.10

0.08
u(r, π/4)

Series1
M=93

0.06
Series2
M=343

0.04
Series3
M=747

0.02
Series4
Exact
0.00

-0.02
0.0 0.2 0.4 0.6 0.8 1.0
r

Fig. 8. Numerical solutions using various numbers of nodes.

4992 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

Table 2
Average errors for various node numbers.

M 93 343 747
e 0.009361 0.003909 0.000467

2.5
Series1
Top (MQ) Series2
Mid (MQ) Series3
Top (TPS)
Series4
Mid (TPS) Series5
Top (exact) Series6
Mid (exact)
2.0

1.5
u(t)

1.0

0.5

0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
t

Fig. 9. Variations of displacement on the top and middle of the plate versus time without damping.

1.8

Series1
Top (MQ) Series2
Mid (MQ)
1.6
Series3
Top (TPS) Series4
Mid (TPS)
1.4
Series5
Static limit Series6
Static limit
1.2

1.0
u (t)

0.8

0.6

0.4

0.2

0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
t

Fig. 10. Variations of displacement on the top and middle of the plate versus time with damping effect.

Considering the damping effect a = 0.5, the displacements using MQ and TPS for two points are presented in Fig. 10. In this
case, there is no close form solution available. However, the results using these two RBFs are too close to see the difference. It
is apparent that the solutions at these two locations tend to the values of static value due to the effect of damping when
t > 11.

7. Conclusion

In this paper, the Finite Integration Method (FIM) with ordinary linear approach and radial basis functions interpolation
was extended to solve multi-dimensional differential equations. Coupled with the technique of Laplace transform, we
demonstrated that the method can be applied to solve time-dependent partial differential equation. Compared with the
 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994 4993

Point Collocation Method (PCM) and the finite difference method, the proposed FIM performs much superior in accuracy and
stability. For the FIM with radial basis functions interpolation, the use of randomly distributed nodes in the domain allows
solving problems under irregular domains. In addition. FIM should be workable for arbitrary shape of the boundary.
However, there are still some issues to deal with arbitrary shape boundary such as the determination of functions f(y)
and g(x), particularly for the problem with domain containing holes. Obviously the application of the FIM for solving two
and three dimensional more complicated elasticity problems and time fractional diffusion equations [15] will be investi-
gated in our future work.

Appendix A.

(I) MQ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ri ðxÞ ¼ c2 þ r 2 ; r ¼ ðx  xi Þ2 þ ðy  yi Þ2 ; ðA1Þ

where c is a free parameter. It is easy to obtain its first derivative as

x  xi y  yi
Ri;x ðx; yÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; Ri;y ðx; yÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi; ðA2Þ
c2 þ r 2 c2 þ r 2

2 2  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 xi ðx; yÞ ¼ x  xi r þ c þ ðy  yi Þ ln x  xi þ c2 þ r2 ;
R ðA3Þ
2 2

2 2  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 yi ðx; yÞ ¼ y  yi r þ c þ ðx  xi Þ ln y  y þ c2 þ r 2 ;
R ðA4Þ
i
2 2

3 2 2  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2
 ðx; yÞ ¼ r þ c þ ðy  yi Þ ðx  x Þ ln x  x þ c2 þ r 2  c þ ðy  yi Þ r;
R ðA5Þ
xi i i
6 2 2

3 2 2  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2
 ðx; yÞ ¼ r þ c þ ðx  xi Þ ðy  y Þ ln y  y þ c2 þ r 2  c þ ðx  xi Þ r;
R ðA6Þ
yi i i
6 2 2
(II) Linear function
x  xi y  yi
Ri ðxÞ ¼ r; Ri;x ðx; yÞ ¼ ; Ri;y ðx; yÞ ¼ ; ðA7Þ
r r
2
 xi ðx; yÞ ¼ x  xi r þ ðy  yi Þ lnðx  xi þ rÞ;
R ðA8Þ
2 2
2
 yi ðx; yÞ ¼ y  yi r þ ðx  xi Þ lnðy  y þ rÞ;
R ðA9Þ
i
2 2

3 2 2
 ðx; yÞ ¼ r þ ðy  yi Þ ðx  x Þ lnðx  x þ rÞ  ðy  yi Þ r;
R ðA10Þ
xi i i
6 2 2

3 2 2
 ðx; yÞ ¼ r þ ðx  xi Þ ðy  y Þ lnðy  y þ rÞ  ðx  xi Þ r;
R ðA11Þ
yi i i
6 2 2
(III) Thin-Plate Splines

Ri ðxÞ ¼ r 2 ln r; Ri;x ðx; yÞ ¼ ðx  xi Þð1 þ 2 ln rÞ; Ri;y ðx; yÞ ¼ ðy  yi Þð1 þ 2 ln rÞ; ðA12Þ
! !
ðx  xi Þ2 2 ðx  xi Þ2 2 x  xi
 xi ðx; yÞ ¼
R ðy  yi Þ2 þ ðx  xi Þ ln r  ðy  yi Þ2 þ ðx  xi Þ þ ðy  yi Þ3 tan1 ; ðA13Þ
3 3 6 3 y  yi

! !
ðy  yi Þ2 2 ðy  yi Þ2 2 y  yi
 yi ðx; yÞ ¼
R ðx  xi Þ2 þ ðy  yi Þ ln r  ðx  xi Þ2 þ ðy  yi Þ þ ðx  xi Þ3 tan1 ; ðA14Þ
3 3 6 3 x  xi

 ðx; yÞ ¼ 1 ððx  x Þ4 þ 6ðx  x Þ2 ðy  y Þ2  3ðy  y Þ4 Þ ln r  13 ðx  x Þ2 ðy  y Þ2 þ 2 ðx  x Þðy  y Þ3 tan1 x  xi ;

Rxi i i i i i i i i
12 24 3 y  yi
ðA15Þ
4994 M. Li et al. / Applied Mathematical Modelling 39 (2015) 4979–4994

 ðx; yÞ ¼ 1 ððy  y Þ4 þ 6ðy  y Þ2 ðx  x Þ2  3ðx  x Þ4 Þ ln r  13 ðy  y Þ2 ðx  x Þ2 þ 2 ðy  y Þðx  x Þ3 tan1 y  yi ;

Rxi i i i i i i i i
12 24 3 x  xi
ðA16Þ
(IV) Polynomial Basis Functions
We select polynomial basis function as (Q ¼ 6)

P1 ¼ 1; P 2 ¼ x; P 3 ¼ y; P4 ¼ x2 ; P5 ¼ xy; P6 ¼ y2 : ðA17Þ
Then, we have
P1;x ¼ 0; P 2;x ¼ 1; P3;x ¼ 0; P4;x ¼ 2x; P5;x ¼ y; P 6;x ¼ 0; ðA18Þ

P1;y ¼ 0; P 2;y ¼ 0; P3;y ¼ 1; P4;y ¼ 0; P5;y ¼ x; P6;y ¼ 2y; ðA19Þ

2 3 2
 x1 ¼ x; P x2 ¼ x ; Px3 ¼ xy; P
P  x4 ¼ x ; Px5 ¼ x y ; P
 x6 ¼ xy2 ; ðA20Þ
2 3 2
2 2 3
 y2 ¼ xy; Py3 ¼ y ; P
 y1 ¼ y; P
P  y4 ¼ x2 y; P y5 ¼ xy ; P
 y6 ¼ y ; ðA21Þ
2 2 3
2 3 2 4 3 2 2
 ¼ x ; P ¼ x ; P ¼ x y ; P
P  ¼ x ; P ¼ x y ; P
 ¼x y ; ðA22Þ
x1 x2 x3 x4 x5 x6
2 6 2 12 6 2
2 2 3 2 2 3 4
 ¼ y ;P
P  ¼ xy ; P ¼ y ; P
 ¼ x y ;P
 ¼ xy ; P
 ¼y : ðA23Þ
y1 y2 y3 y4 y5 y6
2 2 6 2 6 12

References

[1] J.D. Lambert, Numerical Methods for Ordinary Differential Systems: The Initial Value Problem, John Wiley & Sons, New York, 1991.
[2] E. Hairer, Solving Ordinary Differential Equations, vol. 1 and 2, Springer-Verlag, Berlin-New York, 1993.
[3] W.E. Boyce, R.C. DiPrima, Numerical Methods for Ordinary Differential Systems: The Initial Value Problem, John Wiley & Sons, New York, 2001.
[4] S. Roberts, J. Shipman, Two-Point Boundary Value Problems: Shooting Methods, Elsevier, New York, 1972.
[5] M. Scott, H. Watts, Computational solution of linear two point boundary value problems via orthonormalization, SIAM J. Numer. Anal. 14 (1977) 40–70.
[6] M.A. Goldberg, C.S. Chen, S.R. Karur, Improved multiquadric approximation for partial differential equations, Eng. Anal. Boundary Elem. 18 (1996) 9–
17.
[7] R.L. Hardy, Multiquadric equations of topography and other irregular surface, J. Geophys. Res. 76 (1971) 1905–1915.
[8] Y.C. Hon, X.Z. Mao, A multiquadric interpolation method for solving initial value problems, J. Sci. Comput. 12 (1997) 51–55.
[9] J. Li, A.H.-D. Cheng, C.-S. Chen, A comparing of efficiency and error convergence of multiquadric collocation method and finite element method, Eng.
Anal. Boundary Elem. 27 (2003) 251–257.
[10] P.H. Wen, Y.C. Hon, M. Li, T. Korakianitis, Finite integration method for partial differential equations, Appl. Math. Model. 37 (2013) 10092–10106.
[11] M. Li, Y.C. Hon, T. Korakianitis, P.H. Wen, Finite integration method for nonlocal elastic bar under static and dynamic loads, Eng. Anal. Boundary Elem.
37 (5) (2013) 842–849.
[12] W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Numerical Recipes in Fortran, Cambridge University Press, Cambridge, UK, 1992.
[13] F. Durbin, Numerical inversion of Laplace transforms: an efficient improvement to Dubner and Abate’s method, Comput. J. 17 (1975) 371–376.
[14] L.T. Luh, The shape parameter in the shifted surface sphere, Eng. Anal. Boundary Elem. 36 (2012) 1604–1617.
[15] Fu. Zhuo-Jia, Wen. Chen, Qing.-Hua. Qin, Three Boundary meshless methods for heat conduction analysis in nonlinear FGMs with Kirchhoff and Laplace
transformation, Adv. Appl. Math. Mech. 4 (2012) 519–542.