Solution of partial differential equations by a global radial basis

function-based differential quadrature method

C. Shu
*
, H. Ding, K.S. Yeo
Department of Mechanical Engineering, National University of Singapore, 10 Kent Ridge Crescent, Singapore, Singapore 117576
Received 8 September 2002; revised 20 December 2002; accepted 26 February 2003
Available online 17 April 2004
Abstract
The conventional differential quadrature (DQ) method is limited in its application to regular regions by using functional values along a
mesh line to approximate derivatives. In this work, we extend the idea of DQ method to a general case. In other words, any spatial derivative
is approximated by a linear weighted sum of all the functional values in the whole physical domain. The weighting coefcients in the new
approach are determined by the radial basis functions (RBFs). The proposed method combines the advantages of the conventional DQ
method and the RBFs. Since the method directly approximates the derivative, it can be consistently well applied to linear and nonlinear
problems. It also remains mesh free feature of RBFs. Numerical examples of linear and nonlinear cases showed that RBF-based DQ method
has a potential to become an efcient approach for solving partial differential equations.
q 2004 Elsevier Ltd. All rights reserved.
Keywords: Differential quadrature method; Radial basis function; Partial differential equation
1. Introduction
The differential quadrature (DQ) method was introduced
by Richard Bellman and his associates in the early of 1970s
[1,2], following the idea of integral quadrature. The basic
idea of the DQ method is that any derivative at a mesh point
can be approximated by a weighted linear sum of all the
functional values along a mesh line. The key procedure in
the DQ method is the determination of weighting coef-
cients. As shown by Shu and Richards [3], when the solution
of a partial differential equation (PDE) is approximated by a
high order polynomial, the weighting coefcients can be
computed by a simple algebraic formulation or by a
recurrence relationship. Later, Shu and Chew [4] also
showed that when the solution of the PDE is approximated
by a Fourier series expansion, the weighting coefcients of
the rst- and second-order derivatives can be computed
explicitly by algebraic formulations. The details of the
polynomial-based and Fourier series expansion-based DQ
methods can be found in the book of Shu [5]. Currently, the
DQ method has been extensively applied in engineering for
the rapid and accurate solution of various linear and
nonlinear differential equations [311].
On the other hand, it is noted that the function
approximation (polynomial or Fourier series expansion) in
the DQ method is along a straight line. This means that
numerical discretization of derivatives by the DQ method is
also along a straight line. Due to this feature, the DQ method
can be directly applied to regular regions, such as
rectangular and circular domains. For complex geometry,
the DQ method cannot be directly applied. One has to rely
on the coordinate transformation technique [911]. In this
technique, the irregular domain in the physical space is rst
mapped to a regular domain in the computational space.
Then the differential equations and their associated
boundary conditions are transformed into relevant forms
in the computational space. The numerical discretization is
only made in the computational space by the DQ method.
Although this technique can obtain very good results for
problems with complex geometry, we have to admit that the
process is very complicated, and the approach is not as
exible as the nite element method. Practically, there is a
demanding to develop a more efcient method for solving
complex problems.
It was found that the need of coordinate transformation
by the DQ method for complex problems is actually due to
its discretization along a straight line. Furthermore, we
found that the DQ discretization along a straight line is due
0955-7997/$ - see front matter q 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enganabound.2003.02.001
Engineering Analysis with Boundary Elements 28 (2004) 12171226
www.elsevier.com/locate/enganabound
* Corresponding author.
to one-dimensional function approximation used [5]. It is
expected that if a two-dimensional polynomial is used to
approximate a function, the DQ approximation of a
derivative can involve any point on the two-dimensional
plane. And as a consequence, no coordinate transformation
is needed. This is the idea of differential cubature (DC)
method [12]. Unfortunately, due to oscillatory feature of
high order polynomials, the DC method can only obtain
stable solution of a PDE by using a limited number of mesh
points. It seems that the multi-dimensional polynomial
approximation as the test function may not be a good choice
in the DQ approximation. As will be shown in this paper, the
radial basis functions (RBFs), which have truly meshless
property and insensitivity to high dimension, could be a
good choice in the DQ approximation.
RBFs have been under intensive research as a technique
for multivariate data and function interpolation in the past
decades, especially in multi-dimensional applications. Their
performance demonstrates that RBFs constitute a powerful
framework for interpolating or approximating data on
nonuniform grids. Furthermore, Buhmann and Micchelli
[13] showed that RBFs are attractive for pre-wavelet
construction due to their exceptional rates of convergence
and innite differentiability. Since RBFs have excellent
performance for function approximation, many researchers
turn to explore their ability for solving PDEs. The rst trial
of such exploration was made by Kansa [14,15]. As shown
in Ref. [14], using RBFs as a meshless collocation method
to solve PDEs possesses the following advantages: (1) rst
of all, it is a truly mesh-free method, and is independent of
spatial dimension in the sense that the convergence order is
of O(h
d-1
); where h is the density of the collocation points
and d is the spatial dimension; (2) furthermore, in the
context of scattered data interpolation, it is known that some
RBFs have spectral convergence. In other words, as the
spatial dimension of the problem increases, the convergence
order also increases, and hence, much fewer scattered
collocation points will be needed to maintain the same
accuracy as compared with conventional nite difference,
nite element and nite volume methods. This shows the
applicability of the RBFs for solving high-dimensional
problems. Apart from the work of Kansa [14,15], other
researchers, such as Fornberg et al. [16], Hon and Wu [17]
also made a great contribution in this area. It should be
indicated that although some excellent results were
obtained, all previous works related to the application of
RBFs for the numerical solution of PDEs are actually based
on the function approximation instead of derivative
approximation. In other words, these works directly
substitute the expression of function approximation by
RBFs into a PDE, and then change the dependent variables
into the coefcients of function approximation. The process
is very complicated, especially for nonlinear problems. For
the nonlinear case, some special techniques such as
numerical continuation and bifurcation approach proposed
by Fedoseyev et al. [18] have to be used to solve the resultant
nonlinear equations. Since the techniques are very compli-
cated, it is not easy to apply them for solving practical
problems such as uid dynamics, which usually require a
large number of mesh points for accurate solution.
As will be shown in this paper, the advantages of the DQ
approximation and RBFs can be combined to provide an
efcient discretization method, which is a derivative
approximation approach and is mesh-free. In our proposed
method, the RBFs are taken as the test functions in the DQ
approximation to compute the weighting coefcients. Once
the weighting coefcients are computed, the solution
process for a PDE by the new method is exactly the same
as the conventional DQ method and nite difference
schemes. Moreover, the new method can be consistently
well applied to linear and nonlinear problems. Our
numerical experiments demonstrate that this new method
not only inherits the advantages of the DQ method such as
high accuracy and efcient computation, but also owns the
merits of RBFs such as mesh-free feature and easy extension
to high dimension. This article is the rst of a series works.
We hope to present a new framework in applying the DQ
method to practical problems.
2. Development of RBF-based DQ method
The DQ method is a numerical discretization technique
for approximation of derivatives. It was initiated from the
idea of conventional integral quadrature. The conventional
DQ method is applied along a mesh line. As shown by Shu
and co-workers [35], the conventional DQ method is
actually based on the one-dimensional function approxi-
mation. The weighting coefcients in the DQ approximation
can be computed under the analysis of function approxi-
mation and linear vector space. In this work, we extend the
idea of DQ approximation to a general case. The limitation
on using functional values along a mesh line in the
conventional DQ method is released. In our approach, any
spatial derivative is approximated by a linear weighted sum
of all the functional values in the whole physical domain.
The weighting coefcients in our method are determined by
RBF approximation and linear vector space analysis. In the
following, we will show the details of this approach.
2.1. Derivative approximation by differential quadrature
method
It is well known that any integral over a closed domain
can be approximated by a linear weighted sum of all the
functional values in the integral domain. Following this
idea, Bellman et al. [1,2] suggested that the partial
derivative of a function with respect to an independent
variable can be approximated by a linear weighted sum of
functional values at all mesh points in that direction. As
shown in Fig. 1, DQ approximates the derivative of a
function with respect to x at a mesh point (x
i
; y
j
)
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1218
(represented by the symbol X) by all the functional values
along the mesh line of y = y
j
(represented by the symbol W),
and the derivative of the function with respect to y by all the
functional values along the mesh line of x = x
i
(represented
by the symbol A). Mathematically, the DQ approximation
of the nth order derivative with respect to x; f
(n)
x
; and the mth
order derivative with respect to y; f
(m)
y
; at (x
i
; y
j
) can be
written as
f
(n)
x
(x
i
; y
j
) =
X
N
k=1
w
(n)
i;k
f (x
k
; y
j
) (1a)
f
(m)
y
(x
i
; y
j
) =
X
M
k=1
w
(m)
j;k
f (x
i
; y
k
) (1b)
where N; M are, respectively, the number of mesh points in
the x and y direction, w
(n)
i;k
; w
(m)
j;k
are the DQ weighting
coefcients in the x and y directions. As shown in Ref. [5],
w
(n)
i;k
depends on the approximation of the one-dimensional
function f (x; y
j
) (x is the variable), while w
(m)
j;k
depends on the
approximation of the one-dimensional function f (x
i
; y) (y is
the variable). When f (x; y
j
) or f (x
i
; y) is approximated by a
high order polynomial, Shu and Richards [3] derived a
simple algebraic formulation and a recurrence relationship
to compute w
(n)
i;k
and w
(m)
j;k
: When the function is approxi-
mated by a Fourier series expansion, Shu and Chew [4] also
derived simple algebraic formulations to compute the
weighting coefcients of the rst- and second-order
derivatives. For simple geometry, the above DQ approach
can obtain very accurate results by using a considerably
small number of mesh points. However, as mentioned in
Section 1, for complex geometry, the above scheme
cannot be applied directly. The coordinate transformation
technique must be introduced. To remove this drawback, we
need to develop a more efcient approach.
It is noted that the basic idea of the DQ method is that any
derivative can be approximated by a linear weighted sum of
functional values at some mesh points. We can keep this
idea but release the choice of functional values along a mesh
line in the conventional DQ approximation. In other words,
for a two-dimensional problem shown in Fig. 1, any spatial
derivative is approximated by a linear weighted sum of all
the functional values in the whole two-dimensional domain.
In this approximation, a mesh point in the two-dimensional
domain is represented by one index, k; while in the
conventional DQ approximation like Eq. (1), the mesh
point is represented by two indexes i; j: If the mesh is
structured, it is easy to establish the relationship between i; j
and k: For the example shown in Fig. 1, k can be written as
k = (i 21)M - j; i = 1; 2; ; N; j = 1; 2; ; M: Clearly,
when i is changed from 1 to N and j is changed from 1 to
M; k is changed from 1 to NM = N M: The new DQ
approximation for the nth order derivative with respect to
x; f
(n)
x
; and the mth order derivative with respect to y; f
(m)
y
; at
(x
k
; y
k
) can be written as
f
(n)
x
(x
k
; y
k
) =
X
NM
k1=1
w
(n)
k;k1
f (x
k1
; y
k1
) (2a)
f
(m)
y
(x
k
; y
k
) =
X
NM
k1=1
w
(m)
k;k1
f (x
k1
; y
k1
) (2b)
In the following, we will show that the weighting
coefcients in Eq. (2) can be determined by the function
approximation of RBFs and the analysis of a linear vector
space.
2.2. Function approximation by radial basis functions
For scattered points x [ V , R
d
; RBF interpolation
methods use a radial function w : R
d
!R to construct an
interpolant g satisfying g(x
i
) = f
i
as
g(x)
X
N
j=1
l
j
w(x; x
j
) -c(x) (3)
where N is the number of data points, x = (x
1
; x
2
; ; x
d
); d is
the dimension of the problem, ls are coefcients to be
determined and w is the RBF. Eq. (3) can be written without
the additional polynomial c: In that case w must be an
unconditionally positive-denite function to guarantee the
solvability of the resulting system (e.g. Gaussian, inverse
multiquadrics, or compactly supported functions). How-
ever, c is usually required when w is conditionally positive
denite, i.e. when w has a polynomial growth towards
innity. Examples are thin plate splines and multiquadrics.
Moreover, the polynomial in Eq. (3) is added for a special
proof technique of nonsingularity of the extended inter-
polation system. In the comprehensive study by Franke [19],
Fig. 1. A structured mesh for a two-dimensional problem.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1219
it was shown that multiquadrics have the excellent proper-
ties for the interpolation of 2D scattered data. Therefore,
they are highly recommended as test functions for the
numerical schemes proposed in Section 3. Multiquadric
RBFs (for generality, it is written with a variable shape
parameter c
j
) can be written as
w(x; x
j
) = w(r
j
) =

r
2
j
- c
2
j
q
(4)
where r
j
= kx 2x
j
k is the Euclidian norm. Since w given by
Eq. (4) is C
1
continuous, the multiquadric RBFs can be
used directly. If C
d
q
denotes the space of d-variate
polynomials of order not exceeding q; and letting the
polynomials P
1
; ; P
m
be the basis of C
d
q
in R
d
; then
the polynomial c(x); in Eq. (3), is usually written in the
following form
c(x) =
X
m
i=1
z
i
P
i
(x) (5)
where m = (q 21 - d)!=(d!(q 21)!): To determine the
coefcients (l
1
; ; l
N
) and (z
1
; ; z
m
); extra m equations
are required in addition to the N equations resulting from the
collocating Eq. (3) at the N points. This is ensured by the m
conditions for Eq. (3), viz.
X
N
j=1
l
j
P
i
(x
j
) = 0; i = 1; ; m (6)
2.3. Determination of weighting coefcients in RBF-based
DQ approximation
In this section, we will use the multiquadric RBFs as test
functions to determine the weighting coefcients in the new
DQ approximation of derivatives for a two-dimensional
problem. However, the method can be easily extended to the
three-dimensional case as it is dimension-independent, and
other RBFs can also be used as test functions.
Suppose that the solution of a PDE is continuous, which
can be approximated by multiquadric RBFs, and only a
constant is included in the polynomial term Y (x): So, for a
two-dimensional case, Eq. (3) can be reduced to
f (x; y) =
X
N
j=1
l
j
w(kx 2x
j
k) -l
N-1
(7)
To make the problem be well-posed, one more equation
is required. From Eq. (6), we have
X
N
j=1
l
j
= 0 = l
i
= 2
X
N
j=1;ji
l
j
(8)
Substituting Eq. (8) into Eq. (7) gives
f (x; y) =
X
N
j=1;ji
l
j
(w(kx 2x
j
k) 2w(kx 2x
i
k)) -l
N-1
(9)
The number of unknowns in Eq. (9) is N: As no
confusion rises, l
N-1
can be replaced by l
i
; and Eq. (9) can
be written as
f (x; y) =
X
N
j=1;ji
l
j
(w(kx 2x
j
k) 2w(kx 2x
i
k)) -l
i
(10)
It is easy to see that f (x; y) in Eq. (10) constitutes
N-dimensional linear vector space V
N
with respect to the
operation of addition and multiplication. From the concept
of linear independence, the bases of a vector space can be
considered as linearly independent subset that spans the
entire space. In the space V
N
; one set of base vectors is 1,
g
j
(x; y) = w(kx 2x
j
k) 2w(kx 2x
i
k); j = 1; ; N but j i:
From the property of a linear vector space, if all the base
functions satisfy the linear Eq. (2), so does any function in
the space V
N
represented by Eq. (10). There is an interesting
feature. From Eq. (10), while all the base functions are
given, the function f (x; y) is still unknown since the
coefcients l
i
are unknown. However, when all the base
functions satisfy Eq. (2), we can guarantee that f (x; y) also
satises Eq. (2). In other words, we can guarantee that the
solution of a PDE approximated by the RBF satises
Eq. (2). Thus, when the weighting coefcients of DQ
approximation are determined by all the base functions, they
can be used to discretize the derivatives in a PDE. That is the
essence of the RBF-based DQ method.
Substituting all the base functions 1, g
j
(x; y) = w(kx 2x
j
k)
2w(kx 2x
i
k); j = 1; ; N; but j i into Eq. (2a), we can
obtain
0 =
X
N
k=1
w
(n)
i;k
(11a)

n
g
j
(x
i
; y
i
)
x
n
=
X
N
k=1
w
(n)
i;k
g
j
(x
k
; y
k
);
j = 1; 2; N; but j i
(11b)
For the given i; equation system(11) has N unknowns with
N equations. So, solving this equation system can obtain the
weighting coefcients w
(n)
i;k
: In a similar manner, when all the
base functions are substituted into Eq. (2b), the following
equation system can be obtained:
0 =
X
N
k=1
w
(m)
i;k
(12a)

m
g
j
(x
i
; y
i
)
y
m
=
X
N
k=1
w
(m)
i;k
g
j
(x
k
; y
k
);
j = 1; 2; ; N; but j i
(12b)
Equation system (12) is used to compute the weighting
coefcients w
(m)
i;k
of y-derivatives. For multiquadric RBFs,
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1220
g
j
(x; y) can be written as
g
j
(x; y) =

(x 2x
j
)
2
- (y 2y
j
)
2
- c
2
j
q
2

(x 2x
i
)
2
- (y 2y
i
)
2
- c
2
i
q
(13)
From Eq. (13), one can easily obtain the rst-order
derivatives of g
j
(x; y) as
g
j
(x; y)
x
=
x 2x
j

(x 2x
j
)
2
- (y 2y
j
)
2
- c
2
j
q
2
x 2x
i

(x 2x
i
)
2
- (y 2y
i
)
2
-c
2
i
q (14a)
g
j
(x; y)
y
=
y 2y
j

(x 2x
j
)
2
- (y 2y
j
)
2
- c
2
j
q
2
y 2y
i

(x 2x
i
)
2
- (y 2y
i
)
2
-c
2
i
q (14b)
The second- and higher-order derivatives of g
j
(x; y) can
also be obtained by differentiating Eq. (13).
One of the most attractive properties in the above method
is that the weighting coefcients are only related to the test
functions and the position of the collocation points. That
character is very appealing when we deal with the nonlinear
problems. Since the derivatives are directly discretized, the
method can be consistently well applied to linear and
nonlinear problems. Another attractive property of RBF-
based DQ method is that it is naturally mesh free, i.e. all the
information required about the nodes in the domain is
nothing but their positions.
2.4. Shape parameter c
It is well known that the multiquadrics suffer from a
trade-off principle, i.e. one can adjust the parameter c to
increase the approximation power of the basis, but pays for
this by making the problem more and more ill-conditioned.
These effects were initially observed for scattered data
interpolation, but we will see that they also occur in the
numerical solution of PDEs. Thus, the problem of how to
select a good value for the parameter c appears in front of us.
Several methods for selecting c for the multiquadric
interpolants in the two-dimensional case were suggested
in the literature.
Hardy [20] used c = 0:815d; where d = (1=N)
P
N
i=1
d
i
and d
i
is the distance between the ith node and its
neighbouring node. Franke [19] replaced d by D=

N
_
;
where D is the diameter of the minimal circle enclosing all
supporting points and suggested to use c = 1:25(D=

N
_
): Up
to now, optimization of shape parameter and its distribution
are still under research. In Section 3, we will show the
optimal shape parameter c through the trial-and-error
process for the solution of PDEs by RBF-based DQ method.
To study whether the optimal shape parameter c is varied
with different problems, the optimal c achieved in the linear
PDE case is also used in the nonlinear case for the same
point distribution.
3. Numerical examples
By using the presently developed RBF-based DQ
method, several numerical examples are worked out to
investigate the numerical characteristics of the proposed
method.
3.1. Linear PDE case
Consider the problem of a two-dimensional Poisson
equation in a square domain (0 # x # 1; 0 # y # 1): The
governing equation and boundary condition are dened as

2
u
x
2
-

2
u
y
2
= 22p
2
sin px sin py;
0 # x # 1; 0 # y # 1
(15)
Boundary condition:
u = 0 on V (16)
The analytical solution for this problem is u(x; y) =
sin px sin py: Two kinds of mesh point distributions are
considered for this problem. One is the uniform point
distribution and the other is the random point distri-
bution. A typical plot of uniform and random point
distributions is shown in Fig. 2. After numerical
discretization of Eq. (15) by the RBF-based DQ method,
the resultant algebraic equations are solved by the SOR
iteration method. The convergence criteria are chosen as
(lu
n-1
2u
n
l=lu
n
l - 10
28
)
max
, 10
25
: When the criteria
are satised, the convergence results are subsequently
obtained.
To study the convergence behavior of the method, 52, 65,
84, 100, 138 and 217 points are used for the random point
distribution, while mesh sizes of 7 7, 8 8, 9 9,
10 10, 12 12, 15 15 are used for the uniform point
distribution. The numerical results are compared with
analytical solution. To observe the convergence, three
Fig. 2. The uniform and random point distributions.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1221
relative error norms are measured. They are dened as
follows:
Relative L
2
error norm:

X
N
i=1
lu
num
2u
exact
l
2
v
u
u
t

X
N
i=1
lu
exact
l
2
v
u
u
t
(17)
Relative H
1
error norm:

X
N
i=1

lu
num
2u
exact
l
2
- lu
/
num
2u
/
exact
l
2

v
u
u
t

X
N
i=1

lu
exact
l
2
- lu
/
exact
l
2

v
u
u
t
(18)
Relative H
2
error norm:

X
N
i=1

lu
num
2u
exact
l
2
-lu
/
num
2u
/
exact
l
2
-lu
//
num
2u
//
exact
l
2

v
u
u
t

X
N
i=1
lu
exact
l
2
-lu
/
exact
l
2
-lu
//
exact
l
2

v
u
u
t
(19)
where the superscripts ()
/
and ()
//
denote the rst- and
second-order derivatives of the dependent variable,
respectively.
Table 1 presents the relative errors of L
2
; H
1
; and H
2
in
the interior region and on the boundary obtained for the
random point distributions. Table 2 shows the relative errors
for the uniform point distributions. The optimal parameter c
shown in Tables 1 and 2 is obtained through the trial-and-
error process. It can be seen from the two tables that within
the certain number of mesh points, the accuracy of
numerical results can be improved by increasing the number
of mesh points. However, when the number of mesh points
is further increased after a critical value, the accuracy of
numerical results is decreased. The reason may be due
to the fact that, when the number of mesh points is
increased, the condition number of the matrix becomes very
large and the matrix tends to be ill-conditioned.
From Tables 1 and 2, we can see that accurate numerical
results can be obtained by the RBF-based DQ method using
very few mesh points. We also notice that the relative errors
for the derivative approximation at the boundary points
are relatively larger than those in the interior region.
This phenomenon has also been observed by Zhang [21].
So, when a practical problem with derivative boundary
Table 1
Relative errors obtained for a linear problem on random point distributions
Shape parameter, c Total node number L
2
error (10
23
) H
1
error (10
23
) H
2
error (10
23
) Boundary H
1
error (10
23
) Boundary H
2
error (10
23
)
1.272 52 0.0594 0.0628 0.115 0.657 6.861
1.191 65 0.0325 0.0597 0.0782 0.415 4.047
1.037 84 0.00740 0.0126 0.0264 0.127 1.654
0.629 100 0.116 0.0893 0.156 0.639 6.211
0.502 138 0.0579 0.0428 0.102 0.447 4.928
0.156 217 0.316 0.312 1.068 2.258 33.89
Table 2
Relative errors obtained for a linear problem on uniform point distributions
Shape parameter, c Grid L
2
error (10
23
) H
1
error (10
23
) H
2
error (10
23
) Boundary H
1
error (10
23
) Boundary H
2
error (10
23
)
1.471 7 7 0.0335 0.0652 0.0790 0.503 5.219
1.392 8 8 0.0209 0.0492 0.0510 0.375 3.966
1.151 9 9 0.00551 0.0146 0.0181 0.0993 1.289
1.113 10 10 0.00380 0.0107 0.0128 0.0721 0.975
0.675 12 12 0.0571 0.0303 0.0187 0.343 3.929
0.311 15 15 0.0865 0.0790 0.119 0.621 9.816
Fig. 3. Convergence trend for a linear problem on random point
distributions.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1222
condition is considered, we may need to use a more efcient
method to implement the boundary condition.
To clearly show the convergence behavior of the RBF-
based DQ method for the random and uniform point
distributions, the relative errors are plotted against the mesh
spacing h: For the uniform point distribution, h is the
distance of two nearest mesh points. For the random point
distribution, h is dened as
h =
Length

N
_
21
(20)
Using this denition, the numerical results obtained on
the uniform point distribution are a little bit more accurate
than those on the random point distribution. This can be
observed from Figs. 3 and 4, which show the convergence
of relative errors on the random and uniform point
distributions. It can be seen from these two gures that
the convergence trends for the solution of the PDE and
approximation of the rst- and second-order derivatives
are similar. Figs. 3 and 4 also reveal that the convergence
trends of random and uniform point distributions are
similar.
3.2. Nonlinear PDE case
To further validate the performance of RBF-based DQ
method, we will apply it to solve nonlinear PDEs in this
section. Like the linear case, the resultant algebraic
equations from numerical discretization are solved by the
SOR iteration method. We will consider two nonlinear
examples. The rst example is

2
u
x
2
-

2
u
y
2
- u
u
x
-
u
y

22(x - y)u = 4 (21)
The exact solution of Eq. (21) in a square domain is
u(x; y) = x
2
- y
2
(22)
Dirichlet boundary condition for the four edges of the
square domain can be obtained from Eq. (22), that is
u(x = 0) = y
2
; u(x = 1) = 1 - y
2
;
u(y = 0) = x
2
; u(y = 1) = 1 - x
2
(23)
Numerical results are obtained by employing the same
point distributions used in the linear PDE case. The shape
parameters c and the criteria for the convergence are also
chosenthesameasthoseinthelinear case. Fromthenumerical
Table 3
Relative errors obtained for a nonlinear problem on random point distributions
Shape parameter, c Total node number L
2
error (10
23
) H
1
error (10
23
) H
2
error (10
23
) Boundary H
1
error (10
23
) Boundary H
2
error (10
23
)
1.272 52 0.0676 0.112 0.310 0.845 12.32
1.191 65 0.0390 0.0600 0.251 0.540 8.891
1.037 84 0.00458 0.0127 0.0687 0.121 3.063
0.629 100 0.0285 0.0560 0.240 0.317 7.355
0.502 138 0.00615 0.0142 0.0702 0.147 4.222
0.156 217 0.0376 0.108 0.712 0.787 24.57
Fig. 4. Convergence trend for a linear problem on uniform point
distributions.
Table 4
Relative errors obtained for a nonlinear problem on uniform point distributions
Shape parameter, c Grid L
2
error (10
23
) H
1
error (10
23
) H
2
error (10
23
) Boundary H
1
error (10
23
) Boundary H
2
error (10
23
)
1.471 7 7 0.0985 0.151 0.249 0.917 11.60
1.392 8 8 0.0489 0.0477 0.0557 0.477 6.970
1.151 9 9 0.0162 0.0259 0.0508 0.190 3.295
1.113 10 10 0.0107 0.0109 0.0158 0.122 2.361
0.675 12 12 0.0111 0.0145 0.0283 0.151 3.359
0.311 15 15 0.0110 0.0212 0.0589 0.191 5.775
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1223
results shown in Tables 3 and 4, we notice that the optimal
shape parameter c achieved in the linear case works very well
in the nonlinear case. It can also provide very accurate results.
From this study, we may expect that, for a xed mesh point
distribution, the optimal shape parameter c in the RBF-based
DQ method remains the same for various problems. The
convergence trends of this nonlinear case on the random and
uniform point distributions are shown in Figs. 5 and 6. As
compared with Figs. 3and 4, we can ndthat the convergence
trend of the nonlinear case is similar to the linear case.
The second example is chosen from Ref. [22]. The
governing equation and boundary condition of this problem
are
21
2
7
2
u 2u - u
3
= f ; in V = [0; 1] [0; 1] (24)
u = 0; on V (25)
The source function f on the right-hand side is chosen so
that Eq. (24) has an analytic solution of the form
u(x; y) = c(x)c(y) (26)
with
c(t) = 1 - e
21=1
2e
2t=1
2e
(t21)=1
(27)
Here, the parameter 1 determines the size of the
boundary layers near the edges of the domain V: We use
a value of 1 = 0:1 in our numerical experiments. In this
case, we solve the PDE on the random point distribution.
The optimal shape parameter c is also taken the same as that
we achieved in the linear PDE case. The L
2
relative error of
this problem is shown in Table 5. From the relative error in
the L
2
norm, we can see that the RBF-based DQ method also
works well for this nonlinear problem. To visualize
the numerical solution on the random point distribution,
we need to interpolate the results on a uniform grid
Fig. 5. Convergence trend for a nonlinear problem on random point
distributions.
Fig. 6. Convergence trend for a nonlinear problem on uniform point
distributions.
Table 5
The L
2
error obtained for a nonlinear problemon random point distributions
Shape parameter c Total node number L
2
error
1.272 52 1.4474 10
22
1.191 65 8.4122 10
23
1.037 84 7.3988 10
23
0.629 100 1.1071 10
22
0.502 138 9.8744 10
23
0.156 217 9.4433 10
23
Fig. 7. Initial guess for the second nonlinear PDE case.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1224
of 101 101. The interpolation scheme introduced
in Eq. (10) is used to do this job. The initial guess,
numerical and exact solutions of the problem are shown,
respectively, in Figs. 79. As compared with the exact
solution (shown in Fig. 9), the numerical solution shows no
visual difference.
In general, we can observe that for the RBF-based DQ
method, even with a small number of mesh points, as
shown in Tables 14, good accuracy of numerical results
can be achieved for both the linear and nonlinear
problems.
4. Conclusions
A mesh free radial basis function-based differential
quadrature (RBF-based DQ) method is presented in this
paper. The proposed method is similar to nite
difference schemes in the sense of derivative approxi-
mation. It is emphasized that the method permits the
preservation of the simplicity of the nite
differences, while eliminating their drawbacks when it
applies to arbitrary nodal distribution. Furthermore,
it also makes no distinction regarding the dimension of
the problem. Numerical results showed that our RBF-
based DQ scheme is an efcient approach for solution
of PDEs.
On the other hand, it should be indicated that when the
number of nodes is increased to a certain value, the
accuracy of RBF-DQ approximation for derivative will
decrease with the increase of nodes. This is because the
matrix for computing the weighting coefcients in the
DQ approximation would be highly ill-conditioned
when the number of nodes is large. To remove this
difculty, a local RBF-DQ approximation for derivative is
recommended, which will be discussed in detail in
subsequent papers.
References
[1] Bellman RE, Casti J. Differential quadrature and long-term inte-
gration. J Math Anal Appl 1971;34:2358.
[2] Bellman RE, Kashef BG, Casti J. Differential quadrature: a technique
for the rapid solution of nonlinear partial differential equations.
J Comput Phys 1972;10:4052.
[3] Shu C, Richards BE. Application of generalized differential quad-
rature to solve two-dimensional incompressible Navier Stokes
equations. Int J Numer Meth Fluids 1992;15:7918.
[4] Shu C, Chew YT. Fourier expansion-based differential quadrature and
its application to Helmholtz eigenvalue problems. Commun Numer
Meth Engng 1997;13:64353.
[5] Shu C. Differential quadrature and its application in engineering.
London: Springer; 2000.
[6] Bert CW, Jang SK, Striz AG. Two new methods for analyzing free
vibration of structure components. AIAA J 1988;26:6128.
[7] Sherbourne AN, Pandy MD. Differential quadrature method in the
bucking analysis of beams and composite plates. Comput Struct 1991;
40:90313.
[8] Laura PAA, Rossi RE. The method of differential quadrature and its
application to the approximate solution of ocean engineering
problems. Ocean Engng 1994;21:5766.
[9] Bert CW, Malik M. The differential quadrature method for irregular
domains and application to plate vibration. Int J Mech Sci 1996;38(6):
589606.
[10] Han JB, Liew KM. An eight-node curvilinear differential quadrature
formulation for Reissner/Mindlin plates. Comput Meth Appl Mech
Engng 1997;141(3/4):26580.
[11] Shu C, Chen W, Du H. Free vibration analysis of curvilinear
quadrilateral plates by the differential quadrature method. J Comput
Phys 2000;163:45266.
Fig. 8. Numerical solution for the second nonlinear PDE case. Fig. 9. Exact solution for the second nonlinear PDE case.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1225
[12] Liew KM, Liu FL. Differential cubature method: a solution technique
for Kirchhoff plates of arbitrary shape. Comput Meth Appl Mech
1997;145:110.
[13] Buhmann MD, Micchelli CA. Multiquadric interpolation improved.
Computers Math Appl 1992;24:215.
[14] Kansa EJ. Multiquadrics: a scattered data approximation scheme with
applications to computational uid-dynamics. I. Surface approxi-
mations and partial derivative estimates. Computers Math Appl 1990;
19(68):12745.
[15] Kansa EJ. Multiquadrics: a scattered data approximation scheme with
applications to computational uid-dynamics. II. Solutions to
parabolic, hyperbolic, and elliptic partial differential equations.
Computers Math Appl 1990;19(68):14761.
[16] Fornberg B, Driscoll TA, Wright G, Charles R. Observations on
the behavior of radial basis function approximations near boundaries.
Computers Math Appl 2002;43:47390.
[17] Hon YC, Wu ZM. A quasi-interpolation method for solving stiff
ordinary differential equations. Int J Numer Meth Engng 2000;48:
118797.
[18] Fedoseyev AI, Friedman MJ, Kansa EJ. Continuation for nonlinear
elliptic partial differential equation discretized by the multiquadric
method. Int J Bifurcat Chaos 2000;10:48192.
[19] Franke R. Scattered data interpolation: tests of some methods. Math
Comput 1982;38:18199.
[20] Hardy RL. Multiquadric equations of topography and other irregular
surfaces. J Geophys Res 1971;76:190515.
[21] Zhang X, Song KZ, Liu MW. Meshless methods based on
collocation with radial basis functions. Comput Mech 2000;26:
33343.
[22] Fasshauer GE, Gartland EC, Jerome JW. Newton iteration for partial
differential equations and the approximation of the identity. Numer
Algorithms 2000;25:18195.
C. Shu et al. / Engineering Analysis with Boundary Elements 28 (2004) 12171226 1226