You are on page 1of 15

Applied Mathematical Modelling 37 (2013) 10092–10106

Contents lists available at SciVerse ScienceDirect

Applied Mathematical Modelling


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

Finite integration method for partial differential equations


P.H. Wen a,⇑, Y.C. Hon b,c, M. Li c, T. Korakianitis d
a
School of Engineering and Materials Science, Queen Mary, University of London, London E1 4NS, UK
b
Department of Mathematics, City University of Hong Kong, Hong Kong Special Administrative Region
c
College of Mathematics, Taiyuan University of Technology, Taiyuan, China
d
Parks College of Engineering, Aviation and Technology, Saint Louis University, St. Louis, MO 63103, USA

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

Article history: A finite integration method is proposed in this paper to deal with partial differential equa-
Received 2 October 2012 tions in which the finite integration matrices of the first order are constructed by using
Received in revised form 6 March 2013 both standard integral algorithm and radial basis functions interpolation respectively.
Accepted 31 May 2013
These matrices of first order can directly be used to obtain finite integration matrices of
Available online 22 June 2013
higher order. Combining with the Laplace transform technique, the finite integration
method is extended to solve time dependent partial differential equations. The accuracy
Keywords:
of both the finite integration method and finite difference method are demonstrated with
Finite integral method
Radial basis functions
several examples. It has been observed that the finite integration method using either
Partial differential equation radial basis function or simple linear approximation gives a much higher degree of accu-
Partial differential equation with fractional racy than the traditional finite difference method.
order Ó 2013 Elsevier Inc. All rights reserved.
Elasto-dynamics
Laplace transformation

1. Introduction

Ordinary and partial differential equations occur as mathematical models in many branches of science, engineering and
economy. Since it is rare that these equations have solutions in closed form, it is common to develop numerical methods to
seek approximate solutions. Due to the advance of computational methods, these kinds of numerical approximation can usu-
ally 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 to deal with differential equations [1–5]. In the last
decade, developments of the Radial Basis Functions (RBF) as a truly meshless method has drawn the attention of many inves-
tigators (see Golberg et al. [6]). Hardy [7] and Hon et al. [8] used the multi-quadric interpolation method for solving linear
differential equations. Li et al. [9] compared two meshless methods, method of fundamental solutions and dual reciprocity
method, by the use of radial basis functions. Numerical results indicate that these two methods provide a similar optimal
accuracy in solving both 2D Poisson’s and parabolic equations.
In this paper, we present a finite integration method for solving systems of linear boundary value problems. Like the finite
difference method, we replace the solution domain with a finite number of points, known as grid points, and obtain the solu-
tion at these points. The grids are generally equally spaced along the independent coordinate. The integration matrix of the
first order is obtained by the direct integration with either linear approximation or radial basis functions. Based on this ma-
trix, any finite integration matrix with multi-layer integration can be obtained directly by the use of the matrix of the first
order. To compare with other numerical methods, the finite difference method and point collocation method are used. For

⇑ Corresponding author. Tel.: +44 20 7882 5371; fax: +44 20 8983 1007.
E-mail address: p.h.wen@qmul.ac.uk (P.H. Wen).

0307-904X/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.apm.2013.05.054
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10093

one dimensional time dependent problem, it is studied in the Laplace transformed domain. In this case, the continuous var-
iable t is replaced by the discrete variables s and the differential equation is solved in the Laplace space with known initial
conditions. The solution is obtained by suitable selection of the free parameters. Several numerical examples demonstrate
the accuracy and efficiency of the proposed method.

2. Finite integration with ordinary linear approximation

Considering an integral starting from point a in the region x 2 ½a; b


Z x
F ð1Þ ðxÞ ¼ f ðxÞdx; ð1Þ
a

and applying linear interpolation technique for integral function f ðxÞ, we have
Z xk X
k
ð1Þ
F ð1Þ ðxk Þ ¼ f ðxÞdx ¼ aki f ðxi Þ; ð2Þ
a i¼0

ð1Þ
a0i ¼ 0; ð3Þ
8
> 0:5D i¼0
>
>
< D i ¼ 1; 2; . . . ; k  1
ð1Þ
aki ¼ ; ð4Þ
>
> 0:5D i¼k
>
:
0 i>k
where xi ¼ a þ Di; D ¼ ðb  aÞ=N; i ¼ 0; 1; 2; . . . ; N are collocation points in the region of x 2 ½a; b, x0 ¼ a; xL ¼ b. We also can
write integrations in a matrix form as

Fð1Þ ¼ Að1Þ f; ð5Þ


ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ T T
where F ¼ fF 0 ; F 1 ; F 2 ; . . . ; F N g , f ¼ ff0 ; f1 ; f2 ; . . . ; fN g and 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
¼ ðaki Þ ¼ DB C
ð1Þ
Að1Þ B 1=2 ; ð6Þ
B 1 1 1=2 0 0 CC
B C
@ ... ... ... ... ... ... A
1=2 1 1 1 1 1=2 ðNþ1ÞðNþ1Þ

ð1Þ ð1Þ
in which Fi
¼ F ðxi Þ; fi ¼ f ðxi Þ are the values of integration and the integral function respectively at each nodes. Thereafter,
consider a double-layer integral
Z x Z x
F ð2Þ ðxÞ ¼ f ðxÞdxdx; x 2 ½a; b; ð7Þ
a a

and apply linear interpolation technique again for integral function F ð2Þ ðxÞ, we have
Z xk Z x X
k X
i X
k
ð1Þ ð1Þ ð2Þ
F ð2Þ ðxk Þ ¼ f ðxÞdxdx ¼ aki aij f ðxi Þ ¼ aki f ðxi Þ: ð8Þ
a a i¼0 j¼0 i¼0

For the convenience of following analysis, the first order integration matrix is written as A ¼ Að1Þ and the double-layer
integral can be also written again in a matrix form as

Fð2Þ ¼ Að2Þ f ¼ A2 f; ð9Þ


where second order integration matrix
0 1
0 0 0 0 0 0
B 1=4 1=4 0 0 0 0 C
B C
B C
B 3=4 1 1=4 0 0 0 C
¼ ðaki Þ ¼ D2 B C;
ð2Þ
Að2Þ B ð10Þ
B 5=4 2 1 1=4 0 0 CC
B C
@ ... ... ... ... ... ... A
½1 þ 2ðN  1Þ=4 N  1 N  2 ... 1 1=4
where the elements in double-layer integration matrix Að2Þ are
10094 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

ð2Þ
a0i ¼ 0
8
>
> ½1 þ 2ðk  1ÞD2 =4 i¼0
>
>
< ðk  iÞD 2
i ¼ 1; 2; . . . ; k  1 ð11Þ
ð2Þ
aki ¼ :
>
> D 2
=4 i¼k
>
>
:
0 i>k
In the same way, for the higher order integration matrix, we have
Z x Z x
F ðmÞ ðxÞ ¼ ... f ðxÞdx . . . dx x 2 ½a; b ð12Þ
a a
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
m layers

Applying linear interpolation technique again for integral function F ðmÞ ðxÞ, we have
Z xk Z x X
k X
i X
k
ð1Þ ð1Þ ðmÞ
F ðmÞ ðxk Þ ¼ ... f ðxÞdx . . . dx ¼ ... aki . . . aij f ðxi Þ ¼ aki f ðxi Þ: ð13Þ
a a i¼0 j¼0 i¼0

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

FðmÞ ¼ AðmÞ f ¼ Am f: ð14Þ


ðmÞ
It is worth to point out that the integral matrix with any order A is triangular in which half elements in the matrix are
zeroes.

3. Finite integration with radial basis functions

In this section, high accurate approximation technique is to be introduced to construct finite integration matrix. The mul-
ti-quadric RBF was introduced by Hardy [7] for the interpolation of topographical surfaces. The field variable u(x) in the
interval [a, b] can be interpolated over a number of randomly distributed nodes x ¼ fxi gT ; i ¼ 0; 1; 2; . . . ; N, and nodal values
u ¼ fui gT with x0 ¼ a and xN ¼ b; as
X
N X
Q
uðxÞ ¼ Ri ðx; xi Þai þ P q ðxÞbq ¼ RðxÞa þ PðxÞb; ð15Þ
i¼0 q¼0

where RðxÞ ¼ fR0 ðx; x0 Þ; R1 ðx; x1 Þ; . . . ; RN ðx; xN Þg is a set of radial basis functions centred at x, fai gNi¼0 are the unknown coeffi-
cients and fbq gQq¼0 are coefficients for the polynomial basis fP q ðxÞgQq¼0 . The polynomial term has to satisfy an extra require-
ment that guarantees unique approximation of a function as follows
X
N
Pq ðxi Þai ¼ 0; q ¼ 0; 1; 2; . . . ; Q; ð16Þ
i¼0

which can be written in matrix form

PT a ¼ 0: ð17Þ
A set of linear equations can be written in the matrix form as

R0 a þ Pb ¼ u
^; PT a ¼ 0; ð18Þ
where the coefficient matrices are defined as
2 3 2 3
R0 ðx0 ; x0 Þ R1 ðx0 ; x1 Þ ... RN ðx0 ; xN Þ P0 ðx0 Þ P1 ðx0 Þ ... PQ ðx0 Þ
6 R ðx ; x Þ R1 ðx1 ; x1 Þ ... RN ðx1 ; xN Þ 7 6 P0 ðx1 Þ P1 ðx1 Þ ... PQ ðx1 Þ 7
6 0 1 0 7 6 7
6 7 6 7
6 : : ... : 7 6 : : ... : 7
R0 ¼ 6
6
7; P ¼ 6
7 6 :
7; ð19Þ
6 : : ... : 7 6 : ... : 7 7
6 7 6 7
4 : : ... : 5 4 : : ... : 5
R0 ðxN ; x0 Þ R1 ðxN ; x1 Þ . . . RN ðxN ; xN Þ P0 ðxN Þ P1 ðxN Þ . . . P Q ðxN Þ
^ denotes the vector of nodal values. Solving Eq. (18) gives
and u
1 1
b ¼ ðPT R1 T 1
^
0 PÞ P R 0 u; a ¼ R1 T 1 T 1
^
0 ½I  PðP R 0 PÞ P R 0 u; ð20Þ
where u is the vector containing all the field nodal values at the L local nodes, I denotes the diagonal unit matrix. Substituting
the coefficients a and b from (20) into (15), we can obtain the approximation of the field function, in terms of the nodal
values
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10095

D 1 T 1 1 T 1
E X
N
T 1 T 1
uðxÞ ¼ RðxÞR1
0 ½I  PðP R 0 PÞ P R 0  þ PðxÞðP R 0 PÞ P R 0

u ^i :
/i ðxÞu ð21Þ
i¼0

As the inverse matrix of the coefficients a depends only on the distribution of nodes, it is easy to evaluate the derivatives
of the shape function with respect to the coordinate. From (15) and (21), we have

@u X N XQ XN
¼ R0i ðxÞai þ P0q ðxÞbq ¼ R0 ðxÞa þ P0 ðxÞb ¼ /0i ðxÞu
^i : ð22Þ
@x i¼0 q¼0 i¼1

It is worth noticing that the shape function depends uniquely on the distribution of scattered nodes within the domain
and has the Kronecker delta property. The multi-quadric RBF [7,8] is written as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Rl ðxÞ ¼ c2 þ ðx  xl Þ2 and Pq ðxÞ ¼ xq ; ð23Þ

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


x  xl
R0l ðxÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi and P0q ðxÞ ¼ qxq1 : ð24Þ
c2 þ ðx  xl Þ2

Also, the first order integration


Z x X
N X
Q X
N
uðxÞdx ¼ Ri ðxÞai þ Pq ðxÞbq ¼ RðxÞa þ PðxÞb ¼ ^i ;
/i ðxÞu ð25Þ
a i¼0 q¼0 i¼0

where the matrices R and P are defined as


qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 x  x þ c2 þ ðx  xi Þ2
x  xi x i c i
Ri ðxÞ ¼ c2 þ ðx  xi Þ2 þ c2 þ x2i þ ln qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð26Þ
2 2 2 c 2 þ x2  x i i

and

Pq ðxÞ ¼ xqþ1 =ðq þ 1Þ: ð27Þ


Hence, the finite integration matrix of the first order is
 ki Þ;
A ¼ ð/  ki ¼ /
/  i ðxk Þ: ð28Þ
For multi-layer integration matrix, we have the same property as that using the ordinary linear approximation as

AðmÞ ¼ Am : ð29Þ
To compare with different numerical methods, the point collocation method [10] (solutions in strong form) is used for
comparison. It is known that the solution is in strong form using the point collocation method. Therefore, higher order deriv-
atives for variables are needed. For the first order derivative, we have the finite derivation matrix [10] as

df Xk
G1 ðxk Þ ¼ ¼ /0 fi ; ð30Þ
dx i¼0 ki

and

Gð1Þ ¼ Dð1Þ f; Dð1Þ ¼ ð/0ki Þ; /0ki ¼ /0i ðxk Þ: ð31Þ


For the mth order derivative, we have the relationship between higher order derivative matrix and first order derivative
matrix as
m
d f
m ¼ D
ðmÞ
f; DðmÞ ¼ Dm : ð32Þ
dx
In which D ¼ Dð1Þ . It is worth pointing out that, unlike the ordinary linear approximation, the integral matrix and the
derivative matrix of any order in this case AðmÞ and DðmÞ are full.

4. Accuracy of different methods

We firstly consider the following simple ordinary differential equation


2
d u
2
 u ¼ 0 x 2 ½0; 1; ð33Þ
dx
10096 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

with boundary conditions uð0Þ ¼ 0 and u0 ð1Þ ¼ 1. Applying integration operation twice on both sides of the equation, we
have

u  A2 u ¼ c0 x þ c1 i; ð34Þ
T
where c0 and c1 are two integral constants and i ¼ f1; 1; . . . ; 1g . Two extra equations related to the boundary conditions are
X
N
ð1Þ
u0 ¼ 0; aNi ui þ c0 ¼ 1; ð35Þ
i¼0

for the ordinary linear approximation, or


X
N
u0 ¼ 0;  ui þ c0 ¼ 1;
/ ð36Þ
Ni
i¼0

for the radial basis function. There are ðN þ 1Þ linear algebraic equations in (34) and two extra conditions. By solving these
linear equations, ðN þ 3Þ unknowns including uðN þ 1Þ, c0 and c1 can be obtained consequently. However, for the point col-
location method, the differential equation becomes

D2 u  u ¼ 0; ð37Þ
and the first and last equations should be replaced with
X
N
u0 ¼ 0; /0Ni ui ¼ 1: ð38Þ
i¼0

There are ðN þ 1Þ linear algebraic equations to be used for solving the unknown u. The analytical solution of (33) is ob-
tained as
ex  ex
u ðxÞ ¼ : ð39Þ
e þ e1
To demonstrate the accuracy of the finite integration method, the typical finite difference method is compared. The error
is defined as

1 X N
Er ¼ jui  ui j: ð40Þ
N þ 1 i¼0

The variations of average error with the number of collocation point are shown in Table 1. In radial basis function, two
free parameter are selected c ¼ 1=N and Q ¼ 7 for all examples in this paper if they are not specified. It is very clear that even
for a few number of collocation points for the finite integration method, the average errors are less than 1011 and 104 for
the radial basis function and the ordinary linear approximation respectively. However, the average error by the typical finite
difference method (FDM) is around 102 . The increase of accuracy by the finite integration method with the increase of col-
location point number is shown to be faster than that by the finite difference method as well. To investigate the sensitivity of
the free parameter c, it is necessary to observe the influence by the selection of parameter c. The average error variation is
shown in Table 2 by using the finite integration method and the point collocation method with different parameters c. In this
case, the number of collocation point is chosen as N ¼ 10; Q ¼ 7. When c < 1:6, the average errors are less than 1%. However,
the accuracy degree of the finite integration method is much higher than with the point collocation method, both using ra-
dial basis function.
Secondly, consider a differential equation with variable coefficient as below
2
d u
2
 2ð1 þ 2x2 Þu ¼ 0 x 2 ½0; 1; ð41Þ
dx
under boundary conditions uð0Þ ¼ 1 and u0 ð0Þ ¼ 0. The analytical solution is given by

Table 1
Average errors (Er) for different methods.

N FIM(RBF) PCM(RBF) FIM(OLA) FDM


10 2.8127  1011 2.7894  108 1.3615  104 1.4851  102
20 3.3665  1012 9.5536  109 3.1729  105 7.0650  103
30 1.1737  1012 6.3275  109 1.3768  105 4.6293  103
40 5.2770  1013 5.0772  109 7.6513  106 3.4417  103
50 2.3958  1013 4.4270  109 4.8613  106 2.7388  103
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10097

Table 2
Average errors (Er) for different methods with different free parameter c.

c FIM(RBF) PCM(RBF)
0.0 3.3308  1011 3.0695  107
0.4 1.6019  1011 1.8940  108
0.8 1.8630  108 3.9502  108
1.2 4.2030  106 1.6585  105
1.6 7.5021  104 1.3198  103
2.0 3.5350  101 6.5444  101

2
uðxÞ ¼ ex : ð42Þ
Applying integration operator on both sides of the equation, we have

u  A2 Bu ¼ c0 x þ c1 i; ð43Þ
where
0 1
2ð1 þ 2x20 Þ 0 ... 0
B 0 2ð1 þ 2x21 Þ . . . 0 C
B C
B¼B C: ð44Þ
@ ... ... ... ... A
0 0 . . . 2ð1 þ 2x2N Þ
For the point collocation method, the differential equation becomes

D2 u  Bu ¼ 0: ð45Þ
If the relative error is defined as

1 X N
REr ¼ jui  ui j=jui j: ð46Þ
N þ 1 i¼0

The numerical results that REr various with the number N of collocation point are shown in Table 3. It is apparent that
both the finite integration method using the radial basis function FIM (RBF) and the ordinary linear approximation FIM (OLA)
proposed in this paper have higher degree of accuracy than the typical finite difference method. The accuracy by PCM with
radial basis function lies between FIM (RBF) and FIM (OLA).
Finally, consider a beam rested on an elastic foundation and subjected to a unit uniformly distributed load as shown in
Fig. 1. The normalised fourth order differential equation of deflection can be written as

Table 3
Average relative errors (REr) for different methods.

N FIM(RBF) PCM(RBF) FIM(OLA) FDM


6 5 3
10 6.0950  10 7.7037  10 3.9962  10 4.5134  102
20 7.6558  107 1.6548  105 9.2555  104 2.0775  102
30 2.5724  107 7.7647  106 4.0096  104 1.3469  102
40 1.2414  107 4.8063  106 2.2266  104 9.9615  103
50 7.2096  108 3.4076  106 1.4141  104 7.9024  103

q=1

O x

u 1

Fig. 1. A straight bar rested on an elastic foundation.


10098 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

Table 4
Average errors (Er) for different methods ðk ¼ 1Þ.

N FIM(RBF) PCM(RBF) FIM(OLA)


10 5.3652  1014 1.7164  108 2.6501  107
20 1.9795  1014 3.4587  108 1.3437  107
30 7.8857  1015 6.1391  108 8.9815  108
40 1.8704  1014 9.6627  108 6.7422  108
50 2.7109  1014 1.4014  107 5.3960  108

4
d u
4
þ ku ¼ 1 x 2 ½0; 1; ð47Þ
dx
where k is stiffness coefficient of foundation and the boundary conditions (simply supported) are
2
d u
u¼ 2
¼ 0; x ¼ 0 and x ¼ 1: ð48Þ
dx
Applying the finite integration method and the point collocation method, the system equation can be written as
x4 x3 x
u þ kA4 u ¼ þ c0 þ c1 þ c2 x þ c3 i; ð49Þ
24 6 2
and

D4 u þ ku ¼ i; ð50Þ
i T
respectively, where x ¼ fxi0 ; xi1 ; . . . xiN g . Four extra equations related to the boundary conditions for the finite integration
method can be introduced by
u0 ¼ uN ¼ 0
XN
ð2Þ
X
N
ð2Þ ð51Þ
k a0i ui ¼ c1 ; k aNi ui ¼ 12 þ c0 þ c1
i¼0 i¼0

The analytical solution for Eq. (47) can be obtained in Fourier’s series form as
4X
1
sinð2m  1Þpx
u ðxÞ ¼ h i: ð52Þ
p m¼1 ð2m  1Þ ð2m  1Þ4 p4 þ k

The average errors for the different methods varies with the number of collocation are presented in Table 4, where the
free parameters c ¼ 1=N and Q ¼ 7. Again, the finite integration method using radial basis function FIM (RBF) has much high-
er accuracy than the ordinary linear approximation FIM (OLA). In addition, the point collocation method always has much
less accuracy than the finite integration method. For the fourth order differential equation, it is not so easy to obtain the fi-
nite difference equations which satisfy all boundary conditions at two boundary ends.

5. Time dependent problems

For one dimensional elasto-dynamics, the normalised equilibrium equation is written, in time domain, as

@2u @u @2u
2
þ AðxÞ þ BðxÞ 2 þ CðxÞu ¼ bðx; tÞ; 0 < t; 0 6 x 6 1 ð53Þ
@x @t @t
where the coefficients AðxÞ, BðxÞ and the source team bðx; tÞ are known. In addition, the boundary conditions and initial con-
ditions are given by
_ 0Þ ¼ g 2 ðxÞ
uðx; 0Þ ¼ g 1 ðxÞ; uðx;
uð0; tÞ ¼ qðtÞ ð54Þ
u;x ð1; tÞ ¼ pðtÞ
By applying the Laplace transformation on both sides in Eqs. (53) and (54) with consideration of the initial condition, one
obtains
2
d u~ ~ sÞ;
2
~  g 1 ðxÞ þ BðxÞ½s2 u
þ AðxÞ½su ~  sg 1 ðxÞ  g 2 ðxÞ þ CðxÞu
~ ¼ bðx; 0 6 x 6 1: ð55Þ
dx
The Laplace transformation is defined as
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10099

Z 1
~f ¼ f ðtÞest dt: ð56Þ
0

Finally we can re-arrange Eq. (55) as


2
d u~
2
~ ¼ hðx; sÞ;
þ Eðx; sÞu ð57Þ
dx
where

Eðx; sÞ ¼ ½AðxÞs þ BðxÞs2 þ CðxÞ


ð58Þ
~ sÞ þ AðxÞg ðxÞ þ BðxÞ½sg ðxÞ þ g ðxÞ:
hðx; sÞ ¼ bðx; 1 1 2

The numerical procedure is the same as the first example in Section 4. Applying the integration operation twice on both
sides of the equation, we have

~ þ A 2 Eu
u ~ ¼ A2 h þ c0 x þ c1 i; ð59Þ
where c0 and c1 are two integral constants. Two extra equations related to the boundary conditions are
X
N
ð1Þ
~ 0 ¼ 0; p
u ~þ ~ i  c0 ¼ 0;
aNi Ei u ð60Þ
i¼0

for the ordinary linear approximation, or


X
N
~ 0 ¼ 0; p
u ~þ  Ni u
/ ~ i  c0 ¼ 0; ð61Þ
i¼0

for the radial basis function. However, for the point collocation method, the differential equation becomes

D2 u
~ þ Eu
~ ¼ h; ð62Þ
and the first and last equations should be replaced with
X
N
~ 0 ¼ 0;
u /0Ni u
~i ¼ p
~; ð63Þ
i¼0

All unknowns including nodal values of displacement can be obtained numerically in the Laplace transform space for each
specified Laplace parameter sk . If (K + 1) samples in the transformation space sk ; k ¼ 0; 1; . . . ; K, are selected in the Laplace
transform domain, the transformed variables i.e. u ~ ðsk ; xi Þ, are evaluated by the numerical procedure above. Then each vari-
able in the time domain can be determined by any Laplace inversion technique. Here, a method proposed by Durbin [11] is
adopted. The formula of inversion used is written as
" #
2egt 1 XK n o
f ðtÞ ¼  ~f ðgÞ þ Re ~f ðg þ 2kpi=TÞe2kpti=T ; ð64Þ
T 2 k¼0

where ~f ðsk Þ denotes the


pffiffiffiffiffiffiffitransformed variable in the Laplace domain, the parameter of the Laplace transform is chosen as:
sk ¼ g þ 2kp i=T ði 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.
Firstly, consider a classic elastic straight bar with constant section of area and with the following boundary and initial
conditions: AðxÞ ¼ CðxÞ ¼ g 1 ðxÞ ¼ g 2 ðxÞ ¼ qðtÞ ¼ 0, BðxÞ ¼ 1, pðtÞ ¼ HðtÞ and the source team bðx; tÞ ¼ 0, where HðtÞ is the
Heaviside function. The governing equation, in the Laplace domain, becomes
2
d u
2
 s2 u ¼ 0 0 6 x 6 1: ð65Þ
dx
Therefore, we have boundary condition u ~ ;x ðs; 1Þ ¼ p
~ ¼ 1=s. The number of collocation point N is chosen as 20 and the
number of samples in the Laplace space K is taken as 100. In this example, two free parameters in (64) are selected as
g ¼ 5=t0 and T=t0 ¼ 20. The displacements given by the finite integration method FIM (RBF) at the different points
ðx ¼ 1 and x ¼ 0:5Þ are presented in Fig. 2. Good agreement has been achieved for the time dependent displacement
and stress with classic elastic analytical solutions. The average error for time dependent displacement at the end (x = 1) is
calculated by

1 X L
Erð%Þ ¼ juð1; ti Þ  ui ð1; ti Þj  100; ti ¼ 16i=L; L ¼ 160; ð66aÞ
L þ 1 i¼0

where the exact solution is


10100 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

2.5
u (t ,1)
Series1 u(t ,0.5)
Series2
2.0

1.5
u(t)

1.0

0.5

0.0
0 2 4 6 8 10 12 14 16
t
-0.5

Fig. 2. Normalized displacement at top and middle of the bar: classic elasticity.

Table 5
Average errors (Er%) in percentage for different methods.

N FIM(RBF) PCM(RBF) FIM(OLA) FDM


10 0.7017 1.9426 2.6314 16.6523
20 0.4464 1.0403 1.2693 9.0574
30 0.3771 0.8750 0.8627 6.2302
40 0.3601 0.7990 0.6987 4.7696
50 0.3562 0.7534 0.5852 3.8654
100 0.3553 0.6989 0.4239 1.9884

2r1

O x
2r0

p = H (t )
1

Fig. 3. A cylindrical bar with variable section area subjected to impact load at the end.

8
>
< t 06t62
u ð1; tÞ ¼ 4t 2 6 t 6 4; ð66bÞ
>
:
... ...
and the average errors are shown in Table 5 for different number of collocation point with different methods. It is clear that
the finite integration method is more accurate than the finite difference method. With increasing the number of collocation
point, the error decreases very slightly. Compared with analytical solutions in the transformed domain, the numerical solu-
tion has very high accuracy similar to that shown in Table 1. It can be concluded that the error in the time domain is due to
the noise from the inverse of the Laplace transform method.
Second, consider straight cylinder with linear variation section as shown in Fig. 3. In this case, the area of section
   2
r0 r0
SðxÞ ¼ pr 21 þ 1 x ; r1  1; r 0  1; ð67Þ
r1 r1
and the normalised equilibrium equation is obtained, in time domain, by
    
@2u @u @ 2 u r0 r0 r0
þ EðxÞ  ¼ 0; EðxÞ ¼ 2 1  þ 1 x ; 0 6 x 6 1; ð68Þ
@x2 @x @t2 r1 r1 r1
with displacement and traction boundary conditions
uð0; tÞ ¼ 0 and u;x ð1; tÞ ¼ pðtÞ: ð69Þ
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10101

Then, it can be written in the Laplace domain, with zero initial conditions, as
2
d u~ ~
du
2
þ EðxÞ  s2 u
~ ¼ 0 0 6 x 6 1: ð70Þ
dx dx
Applying integration operation twice on both sides of the equation, we have

~ þ A2 ED1 u
u ~  s2 A2 u
~ ¼ c0 x þ c1 i ð71Þ
where c0 and c1 are two integral constants and
0 1
E0 0 . . . 0
B 0 E ... 0 C
B 1 C
E¼B C and Ei ¼ Eðxi Þ: ð72Þ
@... ... ... ...A
0 0 ... EN
For ordinary linear approximation, the derivative matrix is approximated
0 1
1 1 0 0 0 0
B 0 1 1 0 0 0 C
B C

1B B... ... ...
C
... ... ... C
ð1Þ
D ¼ dki ¼ B C; ð73Þ
DB
B 0 0 0 1 1 0 C C
B C
@ 0 0 0 0 1 1 A
0 0 0 0 1 1
and D for radial basis function is given by (31). Let
W ¼ AED ¼ ðwij Þ: ð74Þ
We have two extra equations related to the boundary conditions for the finite integration method as
X
N X
N
~ 0 ¼ 0;
u ~þ
p ~ i  s2 að1Þ
wNi u ~
Ni ui  c 0 ¼ 0; ð75Þ
i¼0 i¼0

for ordinary linear approximation, or


X
N X
N
~ 0 ¼ 0; p
u ~þ ~i  s2 /Ni u
wNi u ~ i  c0 ¼ 0; ð76Þ
i¼0 i¼0

for radial basis function. For the point collocation method, the differential equation simply becomes

D2 u ~  s2 u
~ þ ED u ~ ¼ 0; ð77Þ
and the first and the last equations should be replaced with
X
N
~ 0 ¼ 0;
u /0Ni u
~i ¼ 0 ð78Þ
i¼0

Eq. (70) can also be written as


2
d u~ d dEðxÞ
2
þ ~  u
½EðxÞu ~  s2 u
~ ¼ 0 0 6 x 6 1: ð79Þ
dx dx dx
Applying integration operation twice over (79), we obtain

u ~  A 2 E0 u
~ þ AEu ~  s2 A2 u
~ ¼ c0 x þ c1 i; ð80Þ
where
0 1
E00 0 . . . 0
B 0 E0 . . . 0 C
B C
E0 ¼ B 1 0
C and Ei ¼ @Eðxi Þ=@x: ð81Þ
@... ... ... ...A
0 0 ... E0N
In this case, the derivative matrix disappears. Eqs. (70) and (79) give almost the same results. If the traction applied at the
end x ¼ 1 is considered as pðtÞ ¼ HðtÞ; and u ~ ;x ðs; 1Þ ¼ p
~ ¼ 1=s. The number of collocation point N is chosen to be 100 and the
number of sample in the Laplace space K is taken as 200. The displacements given by the finite integration method FIM (RBF)
at the end of bar ðx ¼ 1Þ are presented in Fig. 4 for different ratios r0 =r1 . It can be observed from Fig. 2 that the effect of the
10102 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

3.0

2.5

2.0
u(t)

1.5

1.0

0.5

0.0
0 2 4 6 8 10 12 14 16
t

Fig. 4. Normalized displacement at top of bar with variable section area.

variation of the section diameter is significant. It is obvious that the slimmer of the straight bar at the fixed end ðx ¼ 0Þ is, the
larger of the displacement at the end will be resulted on the right hand side. In addition, there are some kinks for each curves
of displacement at the arrival time of the longitudinal waves, i.e. t ¼ 2; 4; 6 . . .. However, these kinks become weaker grad-
ually as time increases. Furthermore, it should be mentioned here that the numerical results are almost the same for differ-
ent methods, i.e. FIM (RBF), PCM (RBF), FIM (OLA) and FDM and one can hardly see the difference between them.

6. Partial differential equations of fractional order

6.1. Partial differential equations of fractional order

Fractional advection–dispersion equations are used in groundwater hydrology to model the transport of passive tracers
carried by fluid flow in a porous medium. Analysis of the diffusion-wave equation in mathematical physics has been of con-
siderable interest in the literature. Ordinary and partial differential equations of fractional order have been the focus of many
studies because of their frequent appearance in various applications in fluid mechanics, viscoelasticity, biology, physics, and
engineering [12–19]. Here we consider a simple form for such transport equations

@u @2u
Da ðuÞ þ AðxÞ þ BðxÞ 2 þ CðxÞu ¼ f ðx; tÞ; 0 < t; 0 6 x 6 1; ð82Þ
@t @t
where the coefficients AðxÞ, BðxÞ and the source team f ðx; tÞ are given. In (82), Da is a fractional derivative operator and is
defined as
Z
1 @m x
uðsÞds
Da ðuÞ ¼ ; ð83Þ
Cðm  aÞ @xm 0 ðx  sÞamþ1
where a > 0 and m is an integer such that m > a > m  1. The boundary conditions and initial conditions are given by

uðx; 0Þ ¼ g 1 ðxÞ; @uðx;0Þ


@t
¼ g 2 ðxÞ
uð0; tÞ ¼ 0 ; ð84Þ
@uð0;tÞ
@x
¼0
or

uðx; 0Þ ¼ g 1 ðxÞ; @uðx;0Þ


@t
¼ g 2 ðxÞ
uð0; tÞ ¼ 0 ð85Þ
uð1; tÞ ¼ hðtÞ:
In this paper, we only consider fractional derivatives in space and the index a should be in the range 1 6 a 6 2 and there-
fore m ¼ 2: By applying the Laplace transformation on Eqs. (82), (84), and (85) with consideration of initial condition, one
obtains

Da ðu ~  g 1 ðxÞ þ BðxÞ½p2 u
~ Þ þ AðxÞ½pu ~ ¼ ~f ðxÞ; 0 < t;
~  pg 1 ðxÞ  g 2 ðxÞ þ CðxÞu 0 6 x 6 1; ð86Þ
with the boundary condition
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10103

~ ð0; pÞ
@u ~
~ ð0; pÞ ¼ 0;
u ~ ð0; pÞ ¼ 0;
¼ 0 or u ~ ð1; pÞ ¼ h:
u ð87Þ
@x
Although there are several numerical methods developed to solve differential equations of fractional order, there is little
information on the numerical solution of partial differential equations of fractional order. The rest of this paper is organised
as follows.

6.2. Difference integration method

The spatial variable x is defined as a unit length interval in (86). By using the method of integration by part, the fractional
derivative can be written as
Z x Z x Z x Z x Z x
1 ~ ðs; pÞds
u
þ ~ ðs; pÞdxdx ¼
Hðx; pÞu Q ðx; pÞdxdx þ c0 x þ c1 ; ð88Þ
Cð2  aÞ 0 ðx  sÞa1 0 0 0 0

where Hðx; pÞ ¼ pAðxÞ þ p2 BðxÞ þ CðxÞ, Q ðxÞ ¼ ~f ðxÞ þ AðxÞg 1 ðxÞ þ BðxÞ½pg 1 ðxÞ þ g 2 ðxÞ and c0 and c1 are constants to be obtained
from the boundary and initial conditions:
~ ð0Þ ¼ 0;
u ~;x ð0Þ ¼ 0;
u ð89Þ
and hence we have c0 ¼ c1 ¼ 0. As there is a weak singularity in the first integral in (82), the algorithm in (2) could not be
used directly. In this case, the coefficient matrix can be written, for linear variation of integral function f ðxÞ, as

R xk f ðsÞ
X
k
ðsÞ
Fðxk Þ ¼ 0 ðxsÞb
ds ¼ aki fi
i¼0
ðsÞ ;
a0i ¼ 0 for 0 6 i 6 N
ðsÞ x1b x2b ðxk DÞ2b
ak0 ¼ 1b
k
 k
ð1bÞð2bÞD

ðsÞ 2ðxk  xi Þ2b  ðxk  xiþ1 Þ2b  ðxk  xi1 Þ2b


aki ¼ for 0 < i < k; ð90Þ
ð1  bÞð2  bÞD

ðsÞ 1b
D
akk ¼ ð1bÞð2bÞ
ðsÞ
aki ¼ 0 for i > k; k ¼ 0; 1; 2; . . . ; N;
ðsÞ ð1Þ
where b ¼ a  1. When b ¼ 0, aki ¼ aki . Then Eq. (88) can be arranged in matrix form as
1
~ þ A 2 Hu
AðsÞ u ~ ¼ A2 q; ð91Þ
Cð2  aÞ
in which
0 1 0 1
Hðx0 ; pÞ 0 ... 0 Q ðx0 ; pÞ
B 0 Hðx1 ; pÞ . . . 0 C B Q ðx ; pÞ C
B C B 1 C
H¼B C; q¼B C: ð92Þ
@ ... ... ... ... A @ ... A
0 0 . . . HðxN ; pÞ Q ðxN ; pÞ
Therefore, the system equation in matrix form can be obtained as
~ ¼ b;
Au ð93Þ
where
1
A¼ AðsÞ þ A2 H; b ¼ A2 q: ð94Þ
Cð2  aÞ
In addition, the first equation ðk ¼ 0Þ and the last equation ðk ¼ NÞ in (94) must be modified with the following boundary
conditions

~ ð0; pÞ ¼ u
u ~ 0 ¼ 0; ~ ð1; pÞ ¼ u
u ~
~N ¼ h; ð95Þ
which leads to
a0i ¼ d0i ; b0 ¼ 0
~ i ¼ 0; 1; 2; . . . ; N: ð96Þ
aNi ¼ dNi ; bN ¼ h
10104 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

~ in
Thus, we have obtained the complete system of algebraic equations for the computation of (N + 1) nodal unknowns u
the Laplace transform domain.

6.3. Numerical examples

In the following test cases, we assume that 1 6 a 6 2; m ¼ 2. To test the numerical schemes, it is important to use simple
analytical models. For that purpose, in the first example we consider the case of evolution of a density profile to an equilib-
rium solution [20]. The equilibrium density profile is the solution of the equation,
Da ðuÞ ¼ f ðxÞ; 0 6 x 6 1: ð97Þ
In order to have a simple solution, we use a simple form for the source function

f ðxÞ ¼ 1  x2a : ð98Þ


Along with the boundary condition uð0Þ ¼ 0; u ð0Þ ¼ 0: In this case, the analytical solution for any value of a > 1 can be
0

obtained as
xa Cð3  aÞx2
u ðxÞ ¼  : ð99Þ
Cð1 þ aÞ 2
The numerical results uð1Þ that various with the number of collocation point (N) are listed in Table 6 when fractional or-
der a ¼ 1:5. It is clear that even for a small number of nodes, the relative error is less than 2%. In the second example, we
consider the following differential equation

@2u
Da ðuÞ þ uðxÞ þ ¼ f ðxÞ; 0 6 x 6 1; ð100Þ
@x2
along with the boundary condition uð0Þ ¼ 0, uð1Þ ¼ 1 and source term
2x2a
f ðxÞ ¼ þ x2 þ 2: ð101Þ
Cð3  aÞ
The analytical solution is given as uðxÞ ¼ x2 . For numerical error estimation, we define the relative average error as

1 XN
REr ¼ jui  u ðxi Þj: ð102Þ
ðN þ 1Þumax i¼0


Numerical results are presented in Table 7 for different fractional order a.


In the case of a ¼ 2, the partial differential equation becomes an equation with strong singularity and the solution does
not exist. It is understandable that the degree of accuracy of the solution depends on the fractional order. Surprisingly, this
numerical scheme with integral operator algorithm is of high accuracy to solve partial differential equation of fractional or-
der even with small number of segment N.
Next, we consider following differential equation of fractional order with time dependency as
!
@u @ 2 u
Da ðuÞ  ex þ 2 ¼ f ðx; tÞ; f ðx; tÞ ¼ tet þ xa ext =Cð1 þ aÞ; 0 < t; 0 6 x 6 1; ð103Þ
@t @t

@uðx; 0Þ
uðx; 0Þ ¼ 0; ¼ xa =Cð1 þ aÞ; ð104aÞ
@t

Table 6
Numerical results with different number of segment uð1Þ.

N ¼ 10 N ¼ 20 N ¼ 50 N ¼ 100 Exact in (28)


0.3151356 0.3114126 0.3097477 0.3093602 0.3091393

Table 7
Relative error REr, where umax ¼ 1:.

a N ¼ 10 N ¼ 20 N ¼ 50 N ¼ 100
1.001 2.5492  106 8.4525  107 1.7419  107 5.0267  108
1.1 3.1509  104 1.0870  104 2.3591  105 7.0784  106
1.5 3.6165  103 1.5302  103 4.3656  104 1.6187  104
1.9 1.3637  102 7.4245  103 2.9700  103 1.4288  103
1.999 1.7736  102 1.0311  102 4.5038  103 2.3174  103
P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106 10105

0.30

N=2
0.25
N=4

0.20 N=10

u(1,t)
Exact
0.15

0.10

0.05

0.00
0 1 2 3 4 5 6 7 8 9 10
t

Fig. 5. Variation of uð1; tÞ against time t.

Table 8
Average errors (Er%) in percentage for different methods.

N N¼2 N¼4 N ¼ 10
Er% 2.3143 0.15206 0.0105

@uð0; tÞ
uð0; tÞ ¼ 0; ¼ 0: ð104bÞ
@x
The analytical solution is determined by u ðx; tÞ ¼ xa tet =Cð1 þ aÞ. The numerical solutions of proposed scheme are plot-
ted in Fig. 5 for 0 6 t 6 10 and fractional order a ¼ 1:5. The two free parameters are g ¼ 5, T ¼ 20 and the number of samples
in the Laplace space is L ¼ 200. To show the accuracy of the integral operator algorithm, the analytical solutions are plotted
in this figure for comparison. Three nodal numbers are selected and the results are shown in Fig. 5 and the average errors
(Er%) defined in (66) are presented in Table 8. In this case, the numerical solution with high degree can be obtained with
very small number of nodes.

7. Conclusion

In this paper, the finite integration method with two interpolation method, namely, ordinary linear approximation and
radial basis function, was developed to solve differential equations numerically with high accuracy. The integration matrix
with multi-layers was obtained by the integration matrix of the first order. This method is also extended to the time depen-
dent differential equation and variable section bar under dynamic load was analysed. Compared with the point collocation
method and finite difference method, the finite integration method has much higher accuracy as shown in the several exam-
ples. This method is readily extendable to solve two-dimensional partial differential equation such as potential and elasticity
problems with regular shape of the boundary (grid). The application of FIM for two-dimensional problems both in potential
and elasticity will be demonstrated in future papers.

Acknowledgements

The work described in this paper was supported by a grant from the ImpactQM, Queen Mary, University of London, UK,
and partially supported by a grant from Research Grant Council of the Hong Kong Special Administrative Region, China (Pro-
ject No. CityU 101310) and the Shanxi Hundred Talents Programme through the Taiyuan University of Technology.

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.
10106 P.H. Wen et al. / Applied Mathematical Modelling 37 (2013) 10092–10106

[6] M.A. Golgerg, 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, M.H. Aliabadi, An improved meshless collocation method for elastostatic and elastodynamic problems, J. Commun. Numer. Methods Eng. 24
(8) (2008) 635–651.
[11] F. Durbin, Numerical inversion of Laplace transforms: an efficient improvement to Dubner and Abate’s method, Comput. J. 17 (1975) 371–376.
[12] E. Barkai, R. Metzler, J. Klafter, From continuous time random walks to the fractional Fokker–Planck equation, Phys. Rev. E 61 (2000) 132–138.
[13] J.P. Bouchaud, A. Georges, Anomalous diffusion in disordered media-statistical mechanisms, models and physical applications, Phys. Rep. 195 (1990)
127–293.
[14] M. Meerschaert, D. Benson, H.P. ScheXer, B. Baeumer, Stochastic solution of space-time fractional diffusion equations, Phys. Rev. E 65 (2002) 1103–
1106.
[15] G. Zaslavsky, Fractional kinetic equation for Hamiltonian chaos. Chaotic advection, tracer dynamics and turbulent dispersion, Phys. D 76 (1994) 110–
122.
[16] M. Raberto, E. Scalas, F. Mainardi, Waiting-times and returns in high-frequency financial data: an empirical study, Phys. A 314 (2002) 749–755.
[17] L. Sabatelli, S. Keating, J. Dudley, P. Richmond, Waiting time distributions in financial markets, Eur. Phys. J. B 27 (2002) 273–275.
[18] B. Baeumer, M.M. Meerschaert, D.A. Benson, S.W. Wheatcraft, Subordinated advection–dispersion equation for contaminant transport, Water Resour.
Res. 37 (2001) 1543–1550.
[19] D. Benson, S. Wheatcraft, M. Meerschaert, Application of a fractional advection–dispersion equation, Water Resour. Res. 36 (2000) 1403–1412.
[20] V.E. Lynch, B.A. Carreras, D. del-Castillo-Negrete, K.M. Ferreira-Mejias, H.R. Hicks, Numerical methods for the solution of partial differential equations
of fractional order, J. Comput. Phys. 192 (2003) 406–421.

You might also like