You are on page 1of 22

Applied Mathematics and Computation 245 (2014) 321–342

Contents lists available at ScienceDirect

Applied Mathematics and Computation


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

Error estimates for the interpolating moving least-squares


method
J.F. Wang a,b, F.X. Sun a,c, Y.M. Cheng a,⇑, A.X. Huang d
a
Shanghai Institute of Applied Mathematics and Mechanics, Shanghai University, Shanghai 200072, China
b
Ningbo Institute of Technology, Zhejiang University, Ningbo 315100, China
c
Faculty of Science, Ningbo University of Technology, Ningbo 315016, China
d
School of Mathematics and Statistics, Xi’an Jiaotong University, Xi’an 710049, China

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

Keywords: In this paper, the interpolating moving least-squares (IMLS) method presented by Lancas-
Meshless method ter and Salkauskas is discussed in details. The advantage of the IMLS method is that the
Interpolating moving least-squares (IMLS) meshless method which is constructed based on the IMLS method can apply the essential
method boundary conditions directly and easily. A simpler expression of the approximation
Error estimate
function of the IMLS method is obtained. Then the error estimate of the approximation
Rate of convergence
function and its first and second order derivatives of the IMLS method are presented in
one-dimensional case in this paper. The theoretical results show that if the order of the
polynomial basis functions is big enough and the original function is sufficiently smooth,
then the approximation function and its partial derivatives are convergent to the exact val-
ues in terms of the maximum radius of the domains of influence of nodes. For the purpose
of demonstration, some selected numerical examples are given to prove the theories in this
paper.
Ó 2014 Elsevier Inc. All rights reserved.

1. Introduction

The meshless method has some advantages over the traditional numerical methods, such as finite element method (FEM)
and boundary element method (BEM) [1]. In meshless methods, the approximations entirely base on a set of scattered nodes,
then some complex problems, such as the large deformation and crack growth problems in mechanics, can be simulated with
the method without the re-meshing technique.
The smoothed particle (SPH) method is the oldest meshless method [2,3]. In the SPH method, the approximation is con-
structed from the kernel estimation. Based on the moving least square (MLS) approximations [4], Nayroles et al. presented
the diffuse element method (DEM) [5]. By combining the MLS approximation and the global Galerkin weak form, Belytschko
et al. proposed the element-free Galerkin (EFG) method [6]. The EFG method has high computational precision and compu-
tational stability, and can avoid volumetric locking. The EFG method has been one of the most important meshless methods,
and has been widely applied to solve many science and engineering problems [1,7,8]. After the EFG method, many kinds of
meshless methods have been developed, such as radial basis functions (RBF) method [9,10], finite point method (FPM)
[11–13], meshless local Petrov–Galerkin (MLPG) method [14–18], element-free kp-Ritz method [19–25], meshless manifold

⇑ Corresponding author.
E-mail address: ymcheng@shu.edu.cn (Y.M. Cheng).

http://dx.doi.org/10.1016/j.amc.2014.07.072
0096-3003/Ó 2014 Elsevier Inc. All rights reserved.
322 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

method [26,27], complex variable meshless method [28–46], boundary element-free method (BEFM) [47–54], local bound-
ary integral equation (LBIE) method [55,56], and boundary node method [57].
The MLS approximation is one of the most important methods to form an approximation function in meshless methods.
The MLS approximation was introduced by Lancaster and Salkauskas for surface generation problems [4]. The MLS approx-
imation can be used to form a lot of meshless methods, such as element-free Galerkin method, finite point method, meshless
local Petrov–Galerkin method, complex variable element-free Galerkin method [28–40], boundary element-free method,
boundary node method, etc. The meshless methods based on the MLS approximation have high computational precision
and computational stability.
A disadvantage of the MLS approximation is that the final algebra equations system is sometimes ill-conditioned, and
sometimes the meshless methods based on the MLS approximation cannot obtain a correct solution. Then the improved
moving least-squares approximation was proposed by Cheng and Peng [47]. In the improved MLS approximation, the algebra
equation system is not ill-conditioned, and can be solved directly without obtaining the inverse matrix, and then the mesh-
less methods based on the MLS approximation can always obtain a correct solution. The number of coefficients in the
improved MLS approximation is fewer than in the MLS approximation. Hence, this improved MLS approximation has greater
computational efficiency. Based on the improved MLS approximation, the BEFM [47–53] and the improved EFG method are
proposed to solve many science and engineering problems [58–64].
Based on the MLS approximation and the complex theory, Cheng et al. proposed a new approximation technique, which is
the complex variable moving least-squares (CVMLS) approximation [28–30]. In the CVMLS approximation, the trial function
of a two-dimensional problem is formed with a one-dimensional basis function. Then the number of unknown coefficients in
the trial function of the CVMLS approximation is less than that in the trial function of the MLS approximation. Therefore
nodes needed in the meshless method based on the CVMLS approximation is also less than in the meshless method based
on the MLS approximation with no loss of precision. The meshless method derived from the CVMLS approximation has a
greater computational efficiency. Based on the CVMLS approximation, the complex variable element-free Galerkin
[28–41], complex variable reproducing kernel particle method [42–46], complex variable boundary element-free method
[54], and complex variable meshless manifold method [65] have been proposed to solve many problems.
In recent year, meshless method has been developed greatly in many fields of science and engineering. However, only a
few papers have been published for considering the corresponding mathematical theory. In order to obtain the error esti-
mates of meshless methods based on the MLS approximation, it is fundamental to analyze error estimates of the MLS
approximation function and its derivatives. For a particular weight function, Levin analyzed the MLS approximation and
obtained error estimates in the uniform norm of a regular function in high dimensions [66]. Under appropriate hypotheses
on the weight function and the distribution of nodes, Armentano et al. obtained the error estimates of the MLS approxima-
tion in one-dimensional case [67]. Zuppa also obtained error estimates for the MLS approximation and its derivatives by
introducing condition numbers of the star of nodes in the normal equation [68]. Cheng et al. studied the error estimates
of the MLS approximation in multiple dimensions of Sobolev space [69]. And Li and Zhu also studied the error estimates
of the MLS approximation for one-dimensional problems in Sobolev space [70].
The shape function of the MLS approximation does not satisfy the property of Kronecker d function, and then when a
problem is solved with the meshless method based on the MLS approximation, the essential boundary conditions cannot
be applied directly.
Lancaster and Salkauskas [4] presented the interpolating moving least-squares (IMLS) method by using a singular weight
function. The shape function of the IMLS method satisfies the property of Kronecker d function, and then the meshless
method based on the IMLS method can apply the essential boundary condition directly without any additional numerical
effort. Kaljevic and Saigal presented an improved formulation of the EFG method based on the IMLS method [71]. Maisuradze
et al. applied the IMLS method for fitting potential energy surfaces in one dimensional chemical application [72]. Netuzhylov
used the meshfree collocation method based on the IMLS method to solve boundary value problems [73]. By revising the
formulae of the IMLS method, Ren et al. presented the interpolating boundary element-free (IBEF) method [74,75] and inter-
polating element-free Galerkin (IEFG) method [76,77] for two-dimensional potential and elasticity problems. Based on the
IMLS method and the CVMLS approximation, Ren et al. also presented the complex variable interpolating moving least-
squares (CVIMLS) method [78,79]. To overcome the singularity of weight function in the IMLS method, Wang et al. presented
the improved interpolating moving least-squares (IIMLS) method with the nonsingular weight function [80–83].
In order to study the mathematical theory of meshless methods based on the IMLS method, it is certainly also important
to analyze error estimates of the approximation function of the IMLS method and its derivatives. In the MLS approximation,
the weight function is bounded. But in the IMLS method, the weight function is singular. Thus the error estimates of the
approximation function of the IMLS method and its derivatives are more difficult than the ones of the MLS approximation.
In this paper, the IMLS method presented by Lancaster and Salkauskas is discussed in details. Some inner products pre-
sented by Lancaster and Salkauskas are omitted, thus the corresponding simpler expression of the approximation function of
the IMLS method is obtained. The advantage of the IMLS method is that the meshless method which is constructed based on
the IMLS method can apply the essential boundary conditions directly and easily. Then the error estimate of the approxima-
tion function of the IMLS method is presented in one-dimensional case, and the error estimates of the first and second order
derivatives are also given. The theoretical results show that if the order of the polynomial basis functions is big enough, the
approximation function and its partial derivatives are convergent to the exact values in terms of the maximum radius of the
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 323

domains of influence of nodes. For the purpose of demonstration, some selected numerical examples are given to confirm the
theories in this paper.

2. Interpolating moving least-squares method

Let X ¼ fx1 ; x2 ; . . . ; xM g be a set of all nodes in the bounded domain X  Rn , where M is the number of nodes. The param-
eter qI denotes the radius of the domain of influence of node xI , and jj  jj denotes the Euclidean norm. The domain of influ-
ence of xI is defined by XI ¼ fxjjjx  xI jj 6 qI ; x 2 Xg. Let e ¼ min fjjxI  xJ jjg and q ¼ maxfqI g. For a given point x 2 X,
xI ;xJ 2X;I – J xI 2X
define the index set sx ¼ fIjjjxI  xjj < qI ; xI 2 Xg.
Let uðxÞ be the function of the field variable defined in X. The approximation function of uðxÞ is denoted by uh ðxÞ. In order
to let the approximation function uh ðxÞ in the IMLS method satisfy the interpolating property, Lancaster and Salkauskas [4]
defined a singular weight function, i.e.
(
jjx  xI jja ; jjx  xI jj 6 qI
wðx; xI Þ ¼ wðx  xI Þ ¼ ; ð1Þ
0; others

where the parameter a is an even positive integer.


For the weight function in Eq. (1), when jjx  xI jj 6 qI and x – xI , we have
lim wðx; xI Þ ¼ þ1 ð2Þ
qI !0

Then if the maximum radius of the domains of influence is very small, the values of the weight function would be very
large, which is disadvantageous for a numerical method. However, from the MLS approximation, we know that, in order to
guarantee the precision of the numerical solutions, the fine distribution of nodes in the domain is used, which makes the
radius of influence small. Then a proper value of the radius of influence must be selected.
To overcome the disadvantage of the weight function of Eq. (1), in this paper, we use the following weight function
8  a
< xxI  ; jjx  x jj 6 q
 qI  I I
wðx; xI Þ ¼ wðx  xI Þ ¼ ð3Þ
:
0; others

Define the inner product


X
ðf ; gÞx ¼ wðx; xI Þf ðxI ÞgðxI Þ; 8f ; g 2 C 0 ðXÞ ð4Þ
I2sx

where the subscript x denotes a point in X. Then the corresponding norm at x is


" #1=2
X 2
jjf jjx ¼ wðx; xI Þf ðxI Þ ð5Þ
I2sx

 þ 1 denotes the number of the basis functions. We will


Let p0 ðxÞ  1; p1 ðxÞ; . . . ; pm ðxÞ be given basis functions, where m
generate a new set of basis functions from these given basis. First normalizing p0 ðxÞ and let
1
p  Þ ¼ P
~0 ðx; x 1=2 ð6Þ
I2sx wðx; xI Þ

Þ as
~0 ðx; x
Then we can generate new basis functions orthogonal to p
p  Þ ¼ p i ðx
~i ðx; x Þ  Spi ðxÞ; 
i ¼ 1; 2; . . . ; m; ð7Þ
where S is a linear operator defined as
X
Spi ðxÞ ¼ v ðx; xI Þpi ðxI Þ ð8Þ
I2sx

and
wðx; xI Þ
v ðx; xI Þ ¼ P ð9Þ
J2sx wðx; xJ Þ

The function v ðx; xI Þ has the following properties [4].

 Þ, and
Lemma 2.1. If the weight function of Eq. (3) is used, then v ðx; xI Þ 2 C 1 ðX

(a) v ðxI ; xJ Þ ¼ dI J , 8I; J 2 sx ;


P
(b) I2sx v ðx; xI Þ ¼ 1; 8x 2 X;
324 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

(c) 0 6 v ðx; xI Þ 6 1; 8x 2 X, and v ðx; xI Þ ¼ 0 if and only if x ¼ xJ and J – I;


@ v ðx ;x Þ
(d) @xI J ¼ 0, 8I; J 2 sx

To obtain the expression of the approximation function uh ðxÞ which satisfies the interpolating property, Lancaster and
Salkauskas [4] defined a local approximation, i.e.,
X

m
uh ðx; x
Þ ¼ p Þa0 ðxÞ þ
~0 ðx; x Þai ðxÞ;
~i ðx; x
p ð10Þ
i¼1

where x is the point in the domain of influence of x, and ai ðxÞ, i ¼ 0; 1; . . . ; m,


 are the unknown coefficients of basis functions.
For a given x, the difference between the local approximation function uh ðx; x Þ and the function uðx
Þ will be minimized by
a weighted least-squares method. Then define the weighted discrete L2 norm as
X 2
J¼ wðx; xI Þ½uh ðx; xI Þ  uI  ; ð11Þ
I2sx

where wðx; xI Þ, as shown in Eq. (3), is a weight function with compact support, xI ðI 2 sx Þ are the nodes with domains of influ-
ence that cover the point x, and uI ¼ uðxI Þ.
By minimizing the weighted discrete L2 norm of Eq. (11) we have

ðuðÞ  uh ðx; Þ; p


~0 Þx ¼ 0; ð12Þ

ðuðÞ  uh ðx; Þ; p


~i Þx ¼ 0; 
i ¼ 1; 2;    ; m ð13Þ
In terms of the orthogonality, Eqs. (12) and (13) can be rewritten as
~ 0 Þx ;
a0 ðxÞ ¼ ðu; p ð14Þ

X

m
~0 ; p
a0 ðxÞðp ~ j Þx þ ~i ; p
ai ðxÞðp ~j Þx ¼ ðu; p
~ j Þx ; 
j ¼ 1; 2; . . . ; m ð15Þ
i¼1

According to Eq. (6) and the definition of inner product, we have


1 X
p Þa0 ðxÞ ¼ P
~0 ðx; x ~0 Þx ¼
1=2 ðu; p v ðx; xI ÞuI ¼ Su ð16Þ
I2sx wðx; xI Þ I2sx

Then Eq. (15) reduces to


X

m
~i ; p
ai ðxÞðp ~j Þx ¼ ðu  Su; p
~ j Þx ; 
j ¼ 1; 2; . . . ; m ð17Þ
i¼1

 are solved from Eq. (17). In fact, by the following lemma, Eq. (17)
In Ref. [4], the unknown parameters ai ðxÞði ¼ 1; 2; . . . ; mÞ
can be simplified.

Lemma 2.2. If the weight function of Eq. (3) is used, for 8x 2 X, there exists
~i Þx ¼ 0; i ¼ 1; 2; . . . ; m
ðSu; p  ð18Þ

P
Proof. Since Su ¼ I2sx v ðx; xI ÞuI , we only need to show that
~i Þx ¼ 0;
ð1; p 
i ¼ 1; 2; . . . ; m ð19Þ
In terms of Eqs. (6), (7), (12) and (13), if x R X and x 2 X, this lemma is certainly satisfied. Then we only need to prove
~i ; 1Þx ¼ 0;
limðp 8xI 2 X ð20Þ
x!xI

Since
X X
~i ; 1Þx ¼ lim
limðp ~i ðx; xJ Þ ¼ lim
wðx; xJ Þp wðx; xJ Þ½pi ðxJ Þ  Spi ðxÞ þ limwðx; xI Þ½pi ðxI Þ  Spi ðxÞ
x!xI x!xI x!xI x!xI
J2sx J2sx ;J – I
X wðx; xI Þ X
¼ wðxI ; xJ Þ½pi ðxJ Þ  pi ðxI Þ þ lim P wðx; xJ Þ½pi ðxI Þ  pi ðxJ Þ
J2sx wðx; xJ Þ J2sx ;J – I
x!xI
J2sx ;J – I
X X
¼ wðxI ; xJ Þ½pi ðxJ Þ  pi ðxI Þ þ wðxI ; xJ Þ½pi ðxI Þ  pi ðxJ Þ ¼ 0; ð21Þ
J2sx ;J – I J2sx ;J – I

then the lemma is proved. h


J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 325

According to Lemma 2.2, Eq. (17) can be simplified as


X

m
~i ; p
ai ðxÞðp ~j Þx ¼ ðu; p
~j Þx ; 
j ¼ 1; 2; . . . ; m ð22Þ
i¼1

Eq. (22) is simpler than the corresponding expression in Ref. [4], and can be rewritten as
AðxÞaðxÞ ¼ F W ðxÞu; ð23Þ
where

aT ðxÞ ¼ ða1 ðxÞ; a2 ðxÞ; . . . ; am ðxÞÞ; ð24Þ

uT ¼ ðu1 ; u2 ; . . . ; uN Þ; ð25Þ

AðxÞ ¼ F W ðxÞF T ðxÞ; ð26Þ


2~ ~1 ðx; x2 Þ . . . p~1 ðx; xN Þ 3
p1 ðx; x1 Þ p
6p ~2 ðx; xN Þ 7
6 ~2 ðx; x1 Þ p~2 ðx; x2 Þ . . . p 7
FðxÞ ¼ 6
6 .. .. .. ..
7
7 ð27Þ
4 . . . . 5
p ~m ðx; xN Þ
~m ðx; x2 Þ . . . p
~m ðx; x1 Þ p
and F W ðxÞ ¼ ð-k J ðxÞÞmN

  N matrix, and
is a m
8
< wðx; ~k ðx; xJ Þ;
xJ Þp x – xJ ;
X
-k J ðxÞ ¼ wðx ; x Þ½p ðx Þ  p ðx Þ; x ¼ xJ : ð28Þ
: J I k J k I
I2sx ;I ne J

Then we can obtain

aðxÞ ¼ A1 ðxÞF W ðxÞu ð29Þ


Then the local approximation function can be obtained as
X

m
uh ðx; x
Þ ¼ Su þ Þ
~i ðx; x
ai ðxÞp ð30Þ
i¼1

Thus the global interpolating approximation function of uðxÞ can be obtained as


X

m X
N
uh ðxÞ ¼ Su þ ai ðxÞg i ðxÞ  UðxÞu ¼ UI ðxÞuðxI Þ; ð31Þ
i¼1 I¼1

where UðxÞ is a matrix of shape functions,

UðxÞ ¼ ðU1 ðxÞ; U2 ðxÞ; . . . ; UN ðxÞÞ ¼ v T þ pT ðxÞA1 ðxÞF W ðxÞ; ð32Þ


and

v T ¼ ðv ðx; x1 Þ; v ðx; x2 Þ; . . . ; v ðx; xN ÞÞ; ð33Þ

pT ðxÞ ¼ ðg 1 ðxÞ; g 2 ðxÞ; . . . ; g m ðxÞÞ; ð34Þ

g i ðxÞ ¼ pi ðxÞ  Spi ðxÞ ð35Þ


Then the first partial derivatives of the shape functions of the IMLS method can be obtained as

U;i ðxÞ ¼ v T;i ðxÞ þ pT;i ðxÞA1 ðxÞF W ðxÞ þ pT ðxÞA1 ðxÞF W;i ðxÞ þ pT ðxÞA1
;i ðxÞF W ðxÞ; ð36Þ

where
F W;i ðxÞ ¼ ð-k J;i ðxÞÞmN
 ; ð37Þ
8
~k ðx; xJ Þ þ wðx; xJ Þp
;i ðx; xJ Þp
< wX ~k;i ðx; xJ Þ; x – xJ
-k J;i ðxÞ ¼ w;i ðx; xI Þ½pk ðxJ Þ  pk ðxI Þ; x ¼ xJ ; ð38Þ
:
I2sx ;I – J

A1 1 1
;i ðxÞ ¼ A ðxÞA;i ðxÞA ðxÞ ð39Þ
326 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

Eq. (32) is the shape function of the IMLS method, and Eqs. (17) and (22) show that the computation of the shape function
of this paper is simpler than the corresponding expression in Ref. [4]. Then the IMLS method is presented.

3. Error estimates of the IMLS method in one-dimensional space

Let Bx ðyÞ ¼ fyjjy  xj 6 lq; y 2 X  Rg; where 1 < l < 2 is a constant. Define the index set H ¼ fIjxI 2 X; xI 2 Bx ðyÞg. Sup-
pose h is a sufficient small number such that sx  H and sxþh  H.
When x R X and wðx; xI Þ – 0; there certainly exist constants cb such that
 
db wðx; x Þ c
 I  b
 6 ð40Þ
 dxb  qbI

In our subsequent discussion of this section, the basis functions are selected as
p0 ðxÞ ¼ 1; p1 ðxÞ ¼ x; . . . ; pm ðxÞ ¼ xm ð41Þ
To present the theorems of the error estimates of the IMLS method in one-dimensional case, we firstly give the following
hypotheses.

Hypothesis 3.1. There exist constants ce and cI such that q 6 ce e and q 6 cI qI respectively.
This hypothesis shows that, for 8x 2 X; the number of nodes in the domain of influence of x is finite.

Hypothesis 3.2. For 8x 2 X; there at least exist m þ 1 elements in the set sx \ sxþh .
Suppose d be a positive integer. Let C d ðXÞ be the set of all the functions whose derivative of order d exists and is
continuous in X. Then we can prove the following lemma.

Lemma 3.1. If pi ðxÞ 2 C d ðXÞ; where d is a positive integer, and uh ðx; yÞ and uh ðxÞ are determined respectively by Eqs. (30) and
(31), then the following results hold:

(a) For a given x, uh ðx; yÞ2 C d ðXÞ;


(b) uh ðxÞ2 C d ðXÞ; and uh ðxI Þ ¼ uðxI Þ, 8xI 2 X;
(c) For 8qðyÞ 2 P m ; where P m denotes the set of all polynomials of degree not greater than m; we have

ðuh ðx; yÞ  uðyÞ; qðyÞÞx ¼ 0; 8x 2 X ð42Þ

Proof. (a) and (b) certainly hold [4]. Then we need only prove that (c) is valid.
From Eqs. (6), (7), (12) and (13), if x R X; (c) certainly holds. Then the remainder to be proved is that, for 8x ¼ xI 2 X; there
exists

limðuh ðx; yÞ  uðyÞ; qðyÞÞx ¼ 0 ð43Þ


x!xI

In fact, if limx!xI ðuh ðx; yÞ  uðyÞ; 1Þx ¼ 0; Eq. (43) holds.


From Eq. (30), we have
X X
limðuh ðx; yÞ  uðyÞ; 1Þx ¼ lim wðx; xJ Þ½uh ðx; xJ Þ  uðxJ Þ ¼ lim wðx; xJ Þ½uh ðx; xJ Þ  uðxJ Þ
x!xI x!xI x!xI
J2sx J2sx ;J–I
2
X X
þ limwðx; xI Þ½uh ðx; xI Þ  uðxI Þ ¼ wðxI ; xJ Þ4 uðxk Þv ðxI ; xk Þ
x!xI
J2sxI ;J–I k2sxI
# (
Xm
wðx; xI Þ X
þ ai ðxI Þp~i ðxI ; xJ Þ  uðxJ Þ þ lim P wðx; xJ Þ½uðxJ Þ  uðxI Þ
J2sx wðx; xJ Þ J2sx ;J–I
x!xI
i¼1
) (
Xm X X X m
þ ai ðxI Þ wðx; xJ Þ½pi ðxI Þ  pi ðxJ Þ ¼ wðxI ; xJ Þ uðxI Þ þ ai ðxI Þ½pi ðxJ Þ
i¼1 J2sx ;J–I J2sxI ;J–I i¼1
 X
pi ðxI Þ  uðxJ Þ þ wðxI ; xJ Þ½uðxJ Þ  uðxI Þ
J2sxI ;J–I

X
m X
þ ai ðxI Þ wðxI ; xJ Þ½pi ðxI Þ  pi ðxJ Þ ¼ 0: ð44Þ
i¼1 J2sxI ;J–I
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 327

Then this lemma is proved. h

The convergence order of the local approximation uh ðx; yÞ can be obtained from the known results on the best least-
squares approximation. Following Motzkin’s definitions in Ref. [84], we introduce the concept of ‘weak oscillation’ as follows.

Definition 3.1 [84]]. A function uðxÞ is said to be weakly oscillatory on the set E ¼ fx1 ; x2 ; . . . ; xk g; where x1 < x2 < . . . < xk ;
if we have either (a) or (b):

(a) uðx1 Þ P 0; /ðx2 Þ 6 0; uðx3 Þ P 0; . . .;


(b) uðx1 Þ 6 0; uðx2 Þ P 0; uðx3 Þ 6 0; . . ..

A pair of adjacent inequalities on /ðxÞ is said to define a weak sign change of /ðxÞ on E; and then we have the following
lemma.

Lemma 3.2 [84]]. If the function /ðxÞ oscillates weakly on the set E ¼ fx1 ; x2 ; . . . ; xk g and is continuous for x1 6 x 6 xk ; there exist
zero points yk of /ðxÞ such that

x1 6 y1 6 x2 6 y2 6 x3 6 . . . 6 yk1 6 xk

Then we can prove the following lemma.

Lemma 3.3. If Hypotheses 3.1 and 3.2 are satisfied, and uh ðx; yÞ is determined by Eq. (30), then for 8u 2 C mþ1 ðX
 Þ; uh ðx; yÞ  uðyÞ
have at least m þ 1 zero points on y in Bx ðyÞ.

Proof. This lemma will be proved by the method of reduction to absurdity.


Let E ¼ fx1 ; x2 ; . . . ; xN g be the set of nodes whose domain of influence cover x; and satisfy x1 < x2 < . . . < xN . For a given
point x; let ux ðyÞ ¼ uh ðx; yÞ  uðyÞ; and suppose the function ux ðyÞ has no more than m zero points on y in Bx ðyÞ.
Then from Lemma 3.2, the sequence ux ðx1 Þ; ux ðx2 Þ; . . . ; ux ðxN Þ exhibits no more than m weak sign changes. Assume the
number of the weak sign changes is m ^ (m
^ 6 m). Then there exists a polynomial qx ðyÞ of degree m
^ such that qx ðxI Þ ð1 6 I 6 NÞ
has the same algebraic sign as ux ðxI Þ. The specific form of the polynomial qx ðyÞ will be constructed as follows.
If x is not a node, let

X
^
m
qx ðyÞ ¼ p i ðxÞ;
~i ðx; yÞa ð45Þ
i¼0

where ai ðxÞ; i ¼ 0; 1; . . . ; m;


 are the unknown coefficients and qx ðyÞ cannot vanish identically.
If we can find out m ^ þ 1 different coordinates xi1 ; xi2 ; . . . ; ximþ1
^
2 ½x1 ; xN  and the corresponding values of qx ðyÞ; the
coefficients a i ðxÞ in Eq. (45) can be obtained.
Without loss of generality, suppose ux ðx1 Þ P 0. Firstly let xi1 ¼ x1 and qx ðxi1 Þ ¼ ux ðxi1 Þ. Then let qx ðxi2 Þ ¼ ux ðxi2 Þ; where xi2
satisfies the following (a) and (b):

(a) xi2 2 E; xi2 > xi1 ; and ux ðxi2 Þ 6 0;


(b) for 8xk 2 E; if xk > xi1 and ux ðxk Þ 6 0; we have xk P xi2 .

Next let qx ðxi3 Þ ¼ ux ðxi3 Þ; where xi3 satisfies the following (a) and (b):

(a) xi3 2 E; xi3 > xi2 ; and ux ðxi3 Þ P 0;


(b) for 8xk 2 E; if xk > xi2 and ux ðxk Þ P 0; we have xk P xi3 .

Since ux ðyÞ has m ^ weak sign changes on E; the above process can be alternately repeated m ^ times. Then the coordinates of
xi1 ; xi2 ; . . . ; ximþ1
^
can be obtained, and satisfy qx ðxik Þ ¼ ux ðxik Þ 6 0 if k is even, and qx ðxik Þ ¼ ux ðxik Þ P 0 if k is odd. Now the
coordinates of m ^ þ 1 different points and the corresponding values of qx ðyÞ have been given. Then qx ðyÞ can be obtained.
If qx ðyÞ has greater than m ^ weak sign changes on E; from Lemma 3.2, there would exist at least m ^ þ 1 zero points of qx ðyÞ;
and then qx ðyÞ would vanish identically. Then qx ðyÞ only exists m ^ weak sign changes on E, and then it has the same algebraic
sign as ux ðyÞ on E.
If x is a node such as x ¼ xI 2 E; let

X
^
m
qx ðyÞ ¼ p i ðxÞ;
~i ðx; yÞa ð46Þ
i¼1

i ðxÞ; i ¼ 1; 2; . . . ; m;
where a ^ are the unknown coefficients and qx ðyÞ cannot vanish identically.
328 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

By the same method as used when this lemma is proved in the case of x R E; we can determinate the coordinates of m ^ þ1
points xi1 ; xi2 ; . . . ; ximþ1
^
and the corresponding values of the polynomial (46). Since x I 2 fxi1 ; x i2 ; . . . ; x imþ1
^
g; q ðx
x I Þ ¼ 0 and
~i ðxI ; xI Þ ¼ 0; then by substituting the coordinates and corresponding values into Eq. (46), the coefficients can be obtained,
p
and Eq. (46) has the same algebraic sign as ux ðyÞ on E too.
Then for each x 2 X, when eð> 0Þ is chosen appropriately, we have
jux ðyÞ  e qx ðyÞj 6 jux ðyÞj; 8y 2 E
Then for the weighted discrete L2 norm, we have
X 2 X 2
wðx; xI Þ½uh ðx; xI Þ  e qx ðxI Þ  uðxI Þ 6 wðx; xI Þ½uh ðx; xI Þ  uðxI Þ
I2sx I2sx

h
Thus it is contrary to the hypothesis that u ðx; yÞ is the best least-squares approximation to uðyÞ on E. And then this lemma
is proved. h

Thus, from the error estimates of Lagrange interpolation polynomial we have the following theorem.

Theorem 3.1. If u 2 C mþ1 ðX Þ and Hypotheses 3.1 and 3.2 are satisfied, for each x 2 X, there exists a constant C, which is
independent of q, such that

juh ðx; yÞ  uðyÞj 6 Cjjuðmþ1Þ jjL1 ðXÞ qmþ1 ; 8y 2 Bx ðyÞ ð47Þ

In particular, when y ¼ x; we have

jjuh ðxÞ  uðxÞjjL1 ðXÞ 6 Cjjuðmþ1Þ jjL1 ðXÞ qmþ1 ð48Þ


Now we will discuss the error analysis of the first and second order derivatives of the approximation function.

Lemma 3.4. If xI 2 X; we have


X X
m X X
wðxI ; xJ Þ½uðxJ Þ  uðxI Þ  ai ðxI Þ wðxI ; xJ Þ½pi ðxJ Þ  pi ðxI Þ ¼ wðxI ; xJ Þ½uðxJ Þ  uh ðxI ; xJ Þ
J2sxI ;J – I i¼1 J2sxI ;J–I J2sxI ;J–I

Proof. From the local approximation function of Eq. (30) we have


X X
m X
m
uh ðxI ; xJ Þ ¼ uðxJ Þv ðxI ; xJ Þ þ ~i ðxI ; xJ Þ ¼ uðxI Þ þ
ai ðxI Þp ai ðxI Þ½pi ðxJ Þ  pi ðxI Þ ð49Þ
J2sxI i¼1 i¼1

Then
X
m
ai ðxI Þ½pi ðxJ Þ  pi ðxI Þ ¼ uh ðxI ; xJ Þ  uðxI Þ ð50Þ
i¼1

It follows that
X
m X X X
m X
ai ðxI Þ wðxI ; xJ Þ½pi ðxJ Þ  pi ðxI Þ ¼ wðxI ; xJ Þ ai ðxI Þ½pi ðxJ Þ  pi ðxI Þ ¼ wðxI ; xJ Þ½uh ðxI ; xJ Þ  uðxI Þ ð51Þ
i¼1 J2sxI ;J–I J2sxI ;J–I i¼1 J2sxI ;J–I

Then
X X
m X X
wðxI ; xJ Þ½uðxJ Þ  uðxI Þ  ai ðxI Þ wðxI ; xJ Þ½pi ðxJ Þ  pi ðxI Þ ¼ wðxI ; xJ Þ½uðxJ Þ  uðxI Þ
J2sxI ;J–I i¼1 J2sxI ;J–I J2sxI ;J–I
X
 wðxI ; xJ Þ½uh ðxI ; xJ Þ  uðxI Þ
J2sxI ;J–I
X
¼ wðxI ; xJ Þ½uðxJ Þ  uh ðxI ; xJ Þ ð52Þ
J2sxI ;J–I

Then this lemma is proved. h

@uh ðx;yÞ  Þ; there exists a constant C, which is independent of q,


Lemma 3.5. For each x 2 X; if the derivative @x
exists and u 2 C mþ1 ðX
such that
 h 
@u ðx; yÞ
  ðmþ1Þ
jjL1 ðXÞ qm ; 8y 2 Bx ðyÞ
 @x  6 Cjju ð53Þ
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 329

Proof. Let
X 2
fðxÞ ¼ ½uh ðx þ h; xI Þ  uh ðx; xI Þ ð54Þ
I2K

and

Q ðyÞ ¼ uh ðx þ h; yÞ  uh ðx; yÞ; ð55Þ


where K is an index set of only m þ 1 elements in sx \ sxþh .
Firstly, we will prove this lemma when x R X.
From Eqs. (54) and (3), we have
X 2
fðxÞ 6 wðx; xI Þ½uh ðx þ h; xI Þ  uh ðx; xI Þ
I2K
X X
6 wðx; xI ÞQ ðxI Þ½uh ðx þ h; xI Þ  uðxI Þ þ wðx; xI ÞQ ðxI Þ½uðxI Þ  uh ðx; xI Þ ð56Þ
I2H I2H

Q ðyÞ is a polynomial of degree m; and then from Eq. (42) we have


X
ðuh ðx; yÞ  uðyÞ; Q ðyÞÞx ¼ wðx; xI Þ½uh ðx þ h; xI Þ  uh ðx; xI Þ½uðxI Þ  uh ðx; xI Þ ¼ 0 ð57Þ
I2H

Then
X
fðxÞ 6 wðx; xI ÞQ ðxI Þ½uh ðx þ h; xI Þ  uðxI Þ ð58Þ
I2H

When x – xI ; for small enough h; there exists #I such that


wðx; xI Þ ¼ wðx þ h; xI Þ  w0 ð#I ; xI Þh ð59Þ
From Eq. (42) we have
X
ðuh ðx þ h; yÞ  uðyÞ; QðyÞÞx ¼ wðx þ h; xI ÞQ ðxI Þ½uh ðx þ h; xI Þ  uðxI Þ ¼ 0 ð60Þ
I2H

Then we have
X X
fðxÞ 6 ½wðx þ h; xI Þ  w0 ð#I ; xI Þh½uh ðx þ h; xI Þ  uðxI ÞQ ðxI Þ ¼ h w0 ð#I ; xI Þ½uh ðx þ h; xI Þ  uðxI ÞQðxI Þ ð61Þ
I2H I2H

For x R X; from the inequality (40), Hypothesis 3.1 and Theorem 3.1, we have
 
 X  c hX X
 h  w
0
h w ð#I ; xI ÞQ ðxI Þ½u ðx þ h; xI Þ  uðxI Þ 6 jQðxI Þj  juh ðx þ h; xI Þ  uðxI Þj 6 Chjjuðmþ1Þ jjL1 ðXÞ qm jQ ðxI Þj ð62Þ
 I2H  q I2H I2H

Thus
X
fðxÞ 6 Chjjuðmþ1Þ jjL1 ðXÞ qm jQ ðxI Þj; x R X ð63Þ
I2H

Using the Holder’s inequality, we have


" #2
X X X
jQ ðxI Þj 6 ðm þ 1Þ jQ ðxI Þj2 ¼ ðm þ 1ÞfðxÞ 6 ðm þ 1ÞChjjuðmþ1Þ jjL1 ðXÞ qm jQðxI Þj; x R X ð64Þ
I2K I2K I2H

Because Q ðyÞ is a polynomial of degree m; it can be written as


X
Q ðyÞ ¼ QðxI ÞlI ðyÞ; ð65Þ
I2K

where lI ðyÞ is a Lagrange interpolation polynomial of degree m.


From Hypothesis 3.1, we have
 m
2l q
jlI ðyÞj 6 6 ð2lÞm cm
e ð66Þ
e
Since the number of elements in set H is finite, then there exists a constant c2 such that
 
X XX  XX
 X X X
jQðxI Þj ¼  Q ðxJ ÞlJ ðxI Þ 6 jQ ðxJ Þj  jlJ ðxI Þj ¼ ½jQ ðxJ Þj  jlJ ðxI Þj 6 ð2lÞm cm
e c2 jQ ðxJ Þj ð67Þ
I2H

I2H J2K
 I2H J2K J2K I2H J2K
330 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

It follows from the inequalities (64) and (67) that


" #2
X X
jQðxI Þj 6 ð2lÞm c2 cm
e ðm þ 1ÞChjju
ðmþ1Þ
jjL1 ðXÞ qm jQ ðxI Þj ð68Þ
I2K I2K

Then we have
X
jQðxI Þj 6 ð2lÞm c2 cm
e ðm þ 1ÞChjju
ðmþ1Þ
jjL1 ðXÞ qm ð69Þ
I2K

Thus
 
jQ ðyÞj 1 X  1X
 1 X
¼  QðxI ÞlI ðyÞ 6 jQ ðxI Þj  jlI ðyÞj 6 ð2lÞm cm
e jQ ðxI Þj 6 Cjjuðmþ1Þ jjL1 ðXÞ qm ; x R X ð70Þ
h h  I2K  h I2K h I2K

If the first order derivative of uh ðx; yÞ at x exists, we have


 h   
@u ðx; yÞ  
  ¼ lim QðyÞ 6 Cjjuðmþ1Þ jj 1 qm ; xRX ð71Þ
 @x   h!0 h  L ðXÞ

Then this lemma is proved when x R X.


Secondly, we will prove this lemma when x ¼ xI 2 X.
From Eqs. (54), (57), (59) and (60), we have
X
fðxÞ 6 h w0 ð#J ; xJ Þ½uh ðx þ h; xJ Þ  uðxJ ÞQ ðxJ Þ þ cðx; hÞ; ð72Þ
J2H;J – I

where
2
cðx; hÞ ¼ wðx þ h; xI Þ½uh ðx þ h; xI Þ  uðxI ÞQ ðxI Þ  wðx; xI ÞQ ðxI Þ½uðxI Þ  uh ðx; xI Þ þ ½uh ðx þ h; xI Þ  uh ðx; xI Þ ð73Þ
Then from the inequalities (67), (72), and Theorem 3.1, we have
X
fðxÞ 6 Chjjuðmþ1Þ jjL1 ðXÞ qm jQ ðxJ Þj þ cðx; hÞ ð74Þ
J2K

Let Q I ðyÞ ¼ uh ðxI þ h; yÞ  uh ðxI ; yÞ; then


" #2
X X X
jQ I ðxJ Þj 6 ðm þ 1Þ jQ I ðxJ Þj2 ¼ ðm þ 1ÞfðxI Þ 6 ðm þ 1ÞChjjuðmþ1Þ jjL1 ðXÞ qm jQ I ðxJ Þj þ ðm þ 1ÞcðxI ; hÞ ð75Þ
J2K J2K J2K

Solving the inequality (75), we have


X qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 1 2
jQ I ðxJ Þj 6 ðm þ 1ÞChjjuðmþ1Þ jjL1 ðXÞ qm þ ½ðm þ 1ÞChjjuðmþ1Þ jjL1 ðXÞ qm  þ 4ðm þ 1ÞcðxI ; hÞ ð76Þ
J2K
2 2

From Eq. (73) and Lemma 3.4 we have


8 92
cðxI ; hÞ < X =
lim 2
¼ 2ca q2 wðxI ; xJ Þ½uðxJ Þ  uh ðxI ; xJ Þ ; ð77Þ
h!0 h :J2s ;
xI ;J–I

where

1 a¼2
ca ¼ ð78Þ
0 a–2
Then from Theorem 3.1 we have
cðxI ; hÞ 2
lim 2
6 C jjuðmþ1Þ jjL1 ðXÞ qm ð79Þ
h!0 h
From the inequalities (76) and (79), there exists a constant C such that
1X
lim jQ I ðxJ Þj 6 Cjjuðmþ1Þ jjL1 ðXÞ qm ð80Þ
h!0 h
J2K

Thus we have
 h    X
@u ðxI ; yÞ 


 ¼ lim Q I ðyÞ 6 lim 1 1X
 @x   h!0 h  h!0 h jQ I ðxJ Þj  jlJ ðyÞj 6 ð2lÞm cm
e lim jQ I ðxJ Þj 6 Cjjuðmþ1Þ jjL1 ðXÞ qm ð81Þ
h!0 h
J2K J2K
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 331

Then this lemma is proved. h

Theorem 3.2. If u 2 C mþ1 ðX


 Þ and Hypotheses 3.1 and 3.2 hold, there exists a constant C; which is independent on q; such that
 h 
du ðxÞ duðxÞ
   6 Cjjuðmþ1Þ jjL1 ðXÞ qm ð82Þ
 dx dx L1 ðXÞ

Proof. Since

 h   
du ðxÞ duðxÞ duh ðx; xÞ duðxÞ
  ¼   ð83Þ
 dx dx   dx dx 

and
 h 
duh ðx; xÞ @u ðx; yÞ @uh ðx; yÞ @y
¼ þ j ; ð84Þ
dx @x @y @x y¼x

then we only need to estimate the upper limit of


 h 
@u ðx; yÞ @uh ðx; yÞ @uðyÞ
 þ  ; 8y 2 Bx ðyÞ ð85Þ
 @x @y @y 
Since uh ðx; yÞ  uðyÞ has at least m þ 1 zero points on y in Bx ðyÞ; @y
@
½uh ðx; yÞ  uðyÞ has at least m zero points on y. Hence
from the error estimates of Lagrange interpolation polynomial we have
 h 
@u ðx; yÞ @uðyÞ
   6 Cjjuðmþ1Þ jj 1 qm ; 8y 2 Bx ðyÞ: ð86Þ
 @y @y  L ðXÞ

Then it follows from Lemma 3.5 and the inequality (86) that
 h     
@u ðx; yÞ @uh ðx; yÞ @uðyÞ @uh ðx; yÞ @uðyÞ @uh ðx; yÞ
 þ  6  þ  6 Cjjuðmþ1Þ jj 1 qm ð87Þ
 @x @y @y   @y @y   @x  L ðXÞ

In particular, by taking y ¼ x; this theorem is proved. h

2 h
Lemma 3.6. For each x 2 X, if the derivative @ u ðx;yÞ
@x@y
exists and u 2 C mþ1 ðX
 Þ, then there exists a constant C, which is independent on
q, such that
 
@ 2 uh ðx; yÞ
 
  6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 : ð88Þ
 @x@y 

Proof. For a given x, uh ðx; yÞ is a polynomial of degree m on y; it can be written as

X
uh ðx; yÞ ¼ uh ðx; xI ÞlI ðyÞ; 8x 2 X: ð89Þ
I2K

By using Lemma 3.5, we have


 
@ 2 uh ðx; yÞ X@uh ðx; x Þ @l ðyÞ X 




6  I   I
  6 Cjjuðmþ1Þ jj 1 qm @lI ðyÞ: ð90Þ
 @x@y  I2K  @x   @y  L ðXÞ  @y 
I2K

From Hypothesis 3.1, it follows that


 
@lI ðyÞ mð2lÞm1 cm
  e
: ð91Þ
 @y  6 q
Thus it follows from the inequality (90) that
 
@ 2 uh ðx; yÞ
  m1
  6 ð2lÞ Cmðm þ 1Þcm
e jju
ðmþ1Þ
jjL1 ðXÞ qm1 : ð92Þ
 @x@y 

Then this lemma holds. h


332 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

2 h
Lemma 3.7. For each x R X, if the derivative @ u ðx;yÞ  Þ, then there exists a constant C, which is independent on
exists and u 2 C mþ1 ðX
@x2
q, such that
 
@ 2 uh ðx; yÞ
 
  6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 ; 8y 2 Bx ðyÞ: ð93Þ
 @x2 

Since the weight function is bounded when x R X, then in terms of Ref. [67], this lemma can be proved similarly to Lemma 3.5.

 Þ and Hypotheses 3.1 and 3.2 hold, then there exists a constant C, which is independent on q, such
Theorem 3.3. If u 2 C mþ1 ðX
that
 
d2 uh ðxÞ d2 uðxÞ
 
   6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 : ð94Þ
 dx2 dx2  1
L ðXÞ

1.0
0.8
0.6
0.4
0.2
0.0
error

-0.2
-0.4
-0.6
-0.8
-1.0
-1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 1. The L1 error norms of uðxÞ.

1.5

1.0

0.5
u(x)

0.0
Exact
Numerical
-0.5

-1.0

-0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5


x
Fig. 2. The numerical and exact values of uðxÞ.
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 333

Proof. Since
!
2 2 2
d uh ðxÞ d uðxÞ d uh ðx; xÞ d uðxÞ
2
@ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uðyÞ 

 ¼  ¼ þ 2 þ   ; ð95Þ
dx2 dx2 dx2 dx2 @x2 @x@y @y2 @y2 
y¼x

then we only need to estimate the upper limit of


 
@ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uðyÞ

 þ2 þ  ; 8y 2 Bx ðyÞ: ð96Þ
 @x2 @x@y @y2 @y2 
2 h 2
In terms of Lemma 3.3, @ u@yðx;yÞ
2  @ @yuðyÞ
2 have at least m  1 zero points on y in Bx ðyÞ. Hence from the error estimates of
Lagrange interpolation polynomial we have
 
@ 2 uh ðx; yÞ @ 2 uðyÞ
 
   6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 : ð97Þ
 @y2 @y2 

0.5

0.0

-0.5
du/dx

-1.0 Exact
Numerical
-1.5

-2.0
-0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
x
duðxÞ
Fig. 3. The numerical and exact values of dx
.

2
Exact
1 Numerical
2
d u/dx

0
2

-1

-2

-0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5


x
d2 uðxÞ
Fig. 4. The numerical and exact values of dx2
.
334 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

If x R X, From the inequality (97) and Lemmas 3.6 and 3.7 we have
       
@ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uðyÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uðyÞ
    
 þ 2 þ   
6  þ 2 þ
    6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 :
 @x2 @x@y @y2 @y2   @x2   @x@y   @y2 @y2 
ð98Þ
Then this theorem is proved when x R X.
For any xI 2 X, we have
!
2 2
d uh ðxI Þ d uðxI Þ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uh ðx; yÞ @ 2 uðyÞ 

 ¼ þ2 þ   : ð99Þ
dx2 dx2 @x 2 @x@y @y2 @y2 
y¼x¼xI

It follows from Eq. (30) and Lemma 3.4 that



 2 )
2 X
@ 2 uh ðx;xI Þ dwðx;xI Þ 2 1 d wðx;xI Þ
lim ¼ lim h i  h i  wðxI ; xJ Þ½uðxJ Þ  uh ðxI ; xJ Þ: ð100Þ
x!xI @x 2 x!x I dx P 3 P 2 dx2
J2sx wðx; xJ Þ J2sx wðx;xJ Þ
J2 sx I
;J–I

-1.0

-1.2 L(1)

-1.4 L (2)

-1.6
lg(error)

-1.8

-2.0

-2.2

-2.4
-1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 5. The L1 error norms of the first and second derivatives.

0.5
0.0
-0.5 L(2)
2
-1.0
-1.5
-2.0
lg(error)

-2.5 L(1)
2
-3.0
-3.5
-4.0 L(0)
2
-4.5
-5.0
-1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 6. The L2 error norms of uðxÞ and its first and second derivatives.
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 335

From Eq. (3), there exists a constant c3 such that



 2 )
2
dwðx; xI Þ 2 1 d wðx; xI Þ
lim hP i3  hP i2 ¼ c3 q2 : ð101Þ
x!xI dx dx2
J2sx wðx; xJ Þ J2sx wðx; xI Þ

From Theorem 3.1 and Eq. (100), it easily follows that


 
@ 2 uh ðx ; x Þ
 I I 
  6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 : ð102Þ
 @x2 

From the inequalities (97) and (102), Eq. (99) and Lemma 3.6, we have
 
d2 uh ðx Þ d2 uðx Þ
 I I 
   6 Cjjuðmþ1Þ jjL1 ðXÞ qm1 : ð103Þ
 dx2 dx2 

Then this theorem is proved. h

1.0
0.8
0.6
0.4
0.2
0.0
error

-0.2
-0.4
-0.6
-0.8
-1.0
-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 7. The L1 error norms of uðxÞ.

1.0

0.8
Exact
0.6 Numerical
u(x)

0.4

0.2

0.0

0.0 0.5 1.0 1.5 2.0


x
Fig. 8. The numerical and exact values of uðxÞ.
336 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

4. Numerical examples

In this section, we will present some numerical examples to verify the theoretical error estimates of the IMLS method. In
our three examples, the approximation function uh ðxÞ is constructed from a given function uðxÞ by the IMLS method, and
these examples are used to verify the error estimates of the IMLS method. The errors are measured by L1 and L2 error norms
defined respectively as
ðkÞ
L1  maxjDðkÞ uh ðxI Þ  DðkÞ uðxI Þj; k ¼ 0; 1; 2; ð104Þ
xI 2X

Z 1=2
ðkÞ 2
L2  jjDðkÞ uh  DðkÞ ujjL2 ðXÞ ¼ ðDðkÞ uh  DðkÞ uÞ dX ; k ¼ 0; 1; 2; ð105Þ
X
k
where X  R is a bounded domain, and DðkÞ is a derivative operator defined as DðkÞ uðxÞ ¼ d dxuðxÞ k .

In our numerical computation of this section, the nodes are arranged regularly, and the radius of the domain of influence
of node xI is determined by qI ¼ dmax  jxI  xI1 j, where dmax is a positive scalar. The value of dmax must be chosen so that the

0.0

-0.5
du/dx

-1.0 Exact
Numerical
-1.5

-2.0

0.0 0.5 1.0 1.5 2.0


x
duðxÞ
Fig. 9. The numerical and exact values of dx
.

Exact
3 Numerical
2
d u/dx

2
2

0.0 0.5 1.0 1.5 2.0


x
d2 uðxÞ
Fig. 10. The numerical and exact values of dx2
.
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 337

matrix of Eq. (26) is invertible. In our examples, dmax ¼ 2:5, and the quadric basis functions are used. The integration in Eq.
(105) is obtained numerically by the fourth-order Gaussian quadrature.

Example 5.1. The given function uðxÞ is selected to be a trigonometric function, and uðxÞ ¼ sin2 x þ cos x, x 2 ½0; p.

The L1 error norms of the approximation function of the IMLS method at the nodes are shown in Fig. 1 for different radii,
where the errors are equal to zeros. It can be seen that the approximation function of the IMLS method satisfies the
interpolating property at nodes. The numerical and exact values of the given function and its first and second derivatives are
shown in Figs. 2–4 respectively with 41 nodes, where the numerical values are in excellent agreement with the exact ones.
The L1 error norms of the first and second derivatives are shown in Fig. 5 in log-log scale. And the L2 error norms of the
approximation function and its first and second derivatives are shown in Fig. 6 in log-log scale. Figs. 5 and 6 show that the
error norms of the approximation function and its first and second order derivatives decrease as the radii of the influence
domains decrease.

Example 5.2. The function uðxÞ is chosen to be an exponential function, and uðxÞ ¼ e2x , x 2 ½0; 2.

0.0
-0.3
-0.6
-0.9
-1.2
L(1)
lg(error)


-1.5
-1.8 L(2)

-2.1
-2.4
-2.7
-3.0
-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 11. The L1 error norms of the first and second derivatives.

-1 L(2)
2

-2
lg(error)

-3 L(1)
2

-4

-5 L(0)
2

-6
-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 12. The L2 error norms of uðxÞ and its first and second derivatives.
338 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

1.0
0.8
0.6
0.4
0.2
0.0
error
-0.2
-0.4
-0.6
-0.8
-1.0
-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 13. The L1 error norms of uðxÞ.

1.4

Exact
1.2 Numerical
u(x)

1.0

0.8

0.6
0.0 0.5 1.0 1.5 2.0
x
Fig. 14. The numerical and exact values of uðxÞ.

0.50

Exact
0.45
Numerical

0.40
du/dx

0.35

0.30

0.25

0.0 0.5 1.0 1.5 2.0


x
duðxÞ
Fig. 15. The numerical and exact values of dx
.
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 339

The L1 error norms of the approximation function of the IMLS method at the nodes is shown in Fig. 7. It can also be seen
that the approximation function of the IMLS method satisfies the interpolating property at nodes. The numerical and exact
values of the given function and its first and second derivatives are respectively shown in Figs. 8–10 with 51 nodes, where
the numerical values are also in excellent agreement with the exact ones. And the L1 and L2 error norms of the first and
second derivatives for different radii are shown in Figs. 11 and 12, respectively, in log-log scale, and it is also shown that the
error norms decrease as the radii of the influence domains decrease.

Example 5.3. The logarithmic function is chosen to construct the approximation function, and uðxÞ ¼ lnðx þ 2Þ, x 2 ½0; 2.
The L1 error norms of the approximation function of the IMLS method at nodes are shown in Fig. 13. It can also be seen
that the approximation function of the IMLS method is interpolating at nodes. The numerical and exact values of the given
function and its first and second derivatives are respectively shown in Figs. 14–16 with 51 nodes, where the numerical
values are also in good agreement with the exact ones. The L1 and L2 error norms of the first and second derivatives are
shown respectively in Figs. 17 and 18 in log-log scale, and it is shown again that the error norms decrease as the radii of the
influence domains decrease.

-0.05
Exact
Numerical
-0.10
2

-0.15
d u/dx
2

-0.20

-0.25

0.0 0.5 1.0 1.5 2.0


x
d2 uðxÞ
Fig. 16. The numerical and exact values of dx2
.

-1.5
-1.8
-2.1
-2.4
-2.7
lg(error)

-3.0 L(1)

-3.3 L(2)

-3.6
-3.9
-4.2
-4.5
-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 17. The L1 error norms of the first and second derivatives.
340 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

-1

-2
L(2)
2

-3

lg(error)
-4
L(1)
2

-5

-6
L(0)
2
-7

-1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5
lg(ρ)
Fig. 18. The L2 error norms of uðxÞ and its first and second derivatives.

5. Conclusions

The shape function of the MLS approximation does not satisfy the property of Kronecker d function, and then when a
problem is solved with the meshless method based on the MLS approximation, the essential boundary conditions cannot
be applied directly, which always affects the accuracy of the final solution. In this paper, the IMLS method presented by Lan-
caster and Salkauskas is discussed in details. The advantage of the IMLS method is that the meshless method which is con-
structed based on the IMLS method can apply the essential boundary conditions directly and easily. Some inner products
presented by Lancaster and Salkauskas are omitted, then the corresponding simpler expression of the approximation func-
tion of the IMLS method is obtained. The simpler expression of the approximation function of the IMLS method makes the
method more convenient to be used.
In order to obtain the error estimates of meshless methods based on the IMLS method, it is certainly fundamental to ana-
lyze error estimates of the IMLS approximation function and its derivatives. In this paper, the error estimates of the approx-
imation function of the IMLS method and its first and second order derivatives are presented in one-dimensional case. The
theoretical results show that if the order of the polynomial basis functions is big enough and the original function is suffi-
ciently smooth, then the approximation function and its partial derivatives are convergent to the exact values in terms of the
maximum radius of the domains of influence of nodes.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (No. 11171208), Shanghai Leading Aca-
demic Discipline Project (No. S30106) and the Natural Science Foundation of Ningbo City (No. 2013A610103).

References

[1] T. Belytschko, Y. Krongauz, D. Organ, M. Fleming, P. Krysl, Meshless methods: an overview and recent developments, Comput. Meth. Appl. Mech. Eng.
139 (1996) 3–47.
[2] L.B. Lucy, A numerical approach to the testing of the fission hypothesis, Astron. J. 82 (1977) 1013–1024.
[3] R.A. Gingold, J.J. Monaghan, Smoothed particle hydrodynamics: theory and allocation to non-spherical stars, Mon. Not. R. Astron. Soc. 181 (1977) 375–
389.
[4] P. Lancaster, K. Salkauskas, Surfaces generated by moving least squares methods, Math. Comput. 37 (1981) 141–158.
[5] B. Nayroles, G. Touzot, P. Villon, Generalizing the finite element method: diffuse approximation and diffuse elements, Comput. Mech. 10 (1992) 307–
318.
[6] T. Belytschko, Y.Y. Lu, L. Gu, Element-free Galerkin methods, Int. J. Numer. Meth. Eng. 37 (1994) 229–256.
[7] J.F. Wang, F.X. Sun, R.J. Cheng, Element-free Galerkin method for a kind of KdV equation, Chin. Phys. B 19 (2010) 060201.
[8] R.J. Cheng, Y.M. Cheng, A meshless method for the compound KdV-Burges equation, Chin. Phys. B 20 (2011) 070206.
[9] H. Wendland, Meshless Galerkin method using radial basis functions, Math. Comput. 68 (1999) 1521–1531.
[10] B.D. Dai, Y.M. Cheng, Local boundary integral equation method based on radial basis functions for potential problems, Acta Phys. Sin. 56 (2007) 597–
603.
[11] R.J. Cheng, Y.M. Cheng, The meshless method for solving the inverse heat conduction problem with a source parameter, Acta Phys. Sin. 56 (2007) 5569–
5574.
[12] R.J. Cheng, Y.M. Cheng, The meshless method for a two-dimensional inverse heat conduction problem with a source parameter, Acta Mech. Sin. 39
(2007) 843–847.
[13] J.F. Wang, F.N. Bai, Y.M. Cheng, A meshless method for the nonlinear generalized regularized long wave equation, Chin. Phys. B 20 (2011) 030206.
J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342 341

[14] S.N. Atluri, T. Zhu, A new meshless local Petrov–Galerkin (MLPG) approach in computational mechanics, Comput. Mech. 22 (1998) 117–127.
[15] B.D. Dai, B.J. Zheng, Numerical solution of transient heat conduction problems using improved meshless local Petrov–Galerkin method, Appl. Math.
Comput. 219 (2013) 10044–10052.
[16] B.D. Dai, J. Cheng, B.J. Zheng, A moving Kriging interpolation-based meshless local Petrov–Galerkin method for elastodynamic analysis, Int. J. Appl.
Mech. 5 (1) (2013) 1350011.
[17] L. Chen, C. Liu, H.P. Ma, Y.M. Cheng, An interpolating local Petrov–Galerkin method for potential problems, Int. J. Appl. Mech. 6 (2014) 1450009.
[18] P. Zhu, L.W. Zhang, K.M. Liew, Geometrically nonlinear thermomechanical analysis of moderately thick functionally graded plates using a local Petrov–
Galerkin approach with moving Kriging interpolation, Compos. Struct. 107 (2014) 298–314.
[19] R.J. Cheng, K.M. Liew, Analyzing two-dimensional sine-Gordon equation with the mesh-free reproducing kernel particle Ritz method, Comput. Meth.
Appl. Mech. Eng. 245–246 (2012) 132–143.
[20] K.M. Liew, Z.X. Lei, J.L. Yu, L.W. Zhang, Postbuckling of carbon nanotube-reinforced functionally graded cylindrical panels under axial compression
using a meshless approach, Comput. Meth. Appl. Mech. Eng. 268 (2014) 1–17.
[21] L.W. Zhang, Z.X. Lei, K.M. Liew, J.L. Yu, Static and dynamic of carbon nanotube reinforced functionally graded cylindrical panels, Compos. Struct. 111
(2014) 205–212.
[22] L.W. Zhang, Z.X. Lei, K.M. Liew, J.L. Yu, Large deflection geometrically nonlinear analysis of carbon nanotube-reinforced functionally graded cylindrical
panels, Comput. Meth. Appl. Mech. Eng. 273 (2014) 1–18.
[23] R.J. Cheng, L.W. Zhang, K.M. Liew, Modeling of biological population problems using the element-free kp-Ritz method, Appl. Math. Comput. 227 (2014)
274–290.
[24] R.J. Cheng, K.M. Liew, Modeling of biological population problems using the element-free Kp-Ritz method, Appl. Math. Comput. 227 (2014) 274–290.
[25] Q. Wei, R.J. Cheng, The improved moving least-square ritz method for the one-dimensional Sine-Gordon equation, Math. Prob. Eng. 12 (2014) 1–10.
[26] S. Li, Y.M. Cheng, Y.F. Wu, Numerical manifold method based on the method of weighted residuals, Comput. Mech. 35 (2005) 470–480.
[27] S.C. Li, S.C. Li, Y.M. Cheng, Enriched meshless manifold method for two-dimensional crack modeling, Theor. Appl. Fract. Mech. 44 (2005) 234–248.
[28] Y.M. Cheng, M.J. Peng, J. Li, The complex variable moving least-squares approximation and its application, Chin. J. Theor. Appl. Mech. 37 (2005) 719–
723.
[29] Y.M. Cheng, J.H. Li, A meshless method with complex variables for elasticity, Acta Phys. Sin. 4 (2005) 4463–4471.
[30] Y.M. Cheng, J.H. Li, A complex variable meshless method for fracture problems, Sci. China Ser. G 49 (2006) 46–59.
[31] K.M. Liew, C. Feng, Y.M. Cheng, S. Kitipornchai, Complex variable moving least-squares method: a meshless approximation technique, Int. J. Numer.
Meth. Eng. 70 (2007) 46–70.
[32] M. Peng, P. Liu, Y.M. Cheng, The complex variable element-free Galerkin (CVEFG) method for two-dimensional elasticity problems, Int. J. Appl. Mech. 1
(2009) 367–385.
[33] M. Peng, D. Li, Y.M. Cheng, The complex variable element-free Galerkin (CVEFG) method for elasto-plasticity problems, Eng. Struct. 33 (2011) 127–135.
[34] F.N. Bai, D.M. Li, J.F. Wang, Y.M. Cheng, An improved complex variable element-free Galerkin (ICVEFG) method for two-dimensional elasticity
problems, Chin. Phys. B 21 (2012) 020204.
[35] D.M. Li, F.N. Bai, Y.M. Cheng, K.M. Liew, A novel complex variable element-free Galerkin method for two-dimensional large deformation problems,
Comput. Meth. Appl. Mech. Eng. 233–236 (2012) 1–10.
[36] Y.M. Cheng, J.F. Wang, F.N. Bai, A new complex variable element-free Galerkin method for two-dimensional potential problems, Chin. Phys. B 21 (2012)
090203.
[37] Y.M. Cheng, R. Li, M. Peng, Complex variable element-free Galerkin (CVEFG) method for viscoelasticity problems, Chin. Phys. B 21 (2012) 090205.
[38] J. Wang, Y.M. Cheng, A new complex variable meshless method for transient heat conduction problems, Chin. Phys. B 21 (2012) 120206.
[39] Y.M. Cheng, J. Wang, R. Li, The complex variable element-free Galerkin (CVEFG) method for two-dimensional elastodynamics problems, Int. J. Appl.
Mech. 4 (2012) 1250042.
[40] D.M. Li, K.M. Liew, Y.M. Cheng, An improved complex variable element-free Galerkin method for two-dimensional large deformation elastoplasticity
problems, Comput. Meth. Appl. Mech. Eng. 269 (2014) 72–86.
[41] J. Wang, Y.M. Cheng, New complex variable meshless method for advection–diffusion problems, Chin. Phys. B 22 (2013) 030208.
[42] L. Chen, Y.M. Cheng, The complex variable reproducing kernel particle method for elasto-plasticity problems, Sci. China Phys. Mech. Astron. 53 (2010)
954–965.
[43] Li Chen, K.M. Liew, Y.M. Cheng, The coupling of complex variable-reproducing kernel particle method and finite element method for two-dimensional
potential problems, Interact. Multiscale Mech. 3 (2010) 277.
[44] L. Chen, Y.M. Cheng, The complex variable reproducing kernel particle method for two-dimensional elastodynamics, Chin. Phys. B 19 (2010) 090204.
[45] L. Chen, H. Ma, Y.M. Cheng, Combining the complex variable reproducing kernel particle method and the finite element method for solving transient
heat conduction problems, Chin. Phys. B 22 (2013) 050202.
[46] Y. Weng, Y.M. Cheng, Analyzing variable coefficient advection–diffusion problems via complex variable reproducing kernel particle method, Chin.
Phys. B 22 (2013) 090204.
[47] Y.M. Cheng, M.J. Peng, Boundary element-free method for elastodynamics, Sci. China Ser. G 48 (2005) 641–657.
[48] K.M. Liew, Y.M. Cheng, S. Kitipornchai, Boundary element-free method (BEFM) for two-dimensional elastodynamic analysis using Laplace transform,
Int. J. Numer. Meth. Eng. 64 (2005) 1610–1627.
[49] S. Kitipornchai, K.M. Liew, Y.M. Cheng, A boundary element-free method (BEFM) for three-dimensional elasticity problems, Comput. Mech. 36 (2005)
13–20.
[50] K.M. Liew, Y.M. Cheng, S. Kitipornchai, Boundary element-free method (BEFM) and its application to two-dimensional elasticity problems, Int. J.
Numer. Meth. Eng. 65 (2006) 1310–1332.
[51] K.M. Liew, Y.M. Cheng, S. Kitipornchai, Analyzing the 2D fracture problems via the enriched boundary element-free method, Int. J. Solids Struct. 44
(2007) 4220–4233.
[52] M. Peng, Y.M. Cheng, A boundary element-free method (BEFM) for two-dimensional potential problems, Eng. Anal. Boundary Elem. 33 (2009) 77–82.
[53] Y.M. Cheng, K.M. Liew, S. Kitipornchai, Reply to Comments on Boundary element-free method (BEFM) and its application to two-dimensional elasticity
problems, Int. J. Numer. Meth. Eng. 78 (2009) 1258–1260.
[54] K.M. Liew, Y.M. Cheng, Complex variable boundary element-free method for two-dimensional elastodynamic problems, Comput. Meth. Appl. Mech.
Eng. 198 (2009) 3925–3933.
[55] S.N. Atluri, J. Sladek, V. Sladek, T. Zhu, The local boundary integral equation (LBIE) and its meshless implementation for linear elasticity, Comput. Mech.
25 (2000) 180–198.
[56] B.D. Dai, Y.M. Cheng, An improved local boundary integral equation method for two-dimensional potential problems, Int. J. Appl. Mech. 2 (2010) 421–
436.
[57] V.S. Kothnur, S. Mukherjee, Y.X. Mukherjee, Two dimensional linear elasticity by the boundary node method, Int. J. Solids Struct. 36 (1999) 1129–1147.
[58] Z. Zhang, K.M. Liew, Y.M. Cheng, Y.Y. Li, Analyzing 2D fracture problems with the improved element-free Galerkin method, Eng. Anal. Boundary Elem.
32 (2008) 241–250.
[59] Z. Zhang, K.M. Liew, Y.M. Cheng, Coupling of improved element-free Galerkin and boundary element methods for the 2D elasticity problems, Eng. Anal.
Boundary Elem. 32 (2008) 100–107.
[60] Z. Zhang, D. Li, Y.M. Cheng, K.M. Liew, The improved element-free Galerkin method for three-dimensional wave equation, Acta Mech. Sin. 28 (2012)
808–818.
342 J.F. Wang et al. / Applied Mathematics and Computation 245 (2014) 321–342

[61] Z. Zhang, Y.M. Cheng, K.M. Liew, The improved element-free Galerkin method for two-dimensional elastodynamics problems, Eng. Anal. Boundary
Elem. 37 (2013) 1576–1584.
[62] Z. Zhang, J. Wang, Y.M. Cheng, K.M. Liew, The improved element-free Galerkin method for three-dimensional transient heat conduction problems, Sci.
China Phys. Mech. Astron. 56 (2013) 1568–1580.
[63] M.J. Peng, R.X. Li, Y.M. Cheng, Analyzing three-dimensional viscoelasticity problems via the improved element-free Galerkin (IEFG) method, Eng. Anal.
Boundary Elem. 40 (2014) 104–113.
[64] L.W. Zhang, Y.J. Deng, K.M. Liew, An improved element-free Galerkin method for numerical modeling of the biological population problems, Eng. Anal.
Boundary Elem. 40 (2014) 181–188.
[65] H. Gao, Y. Cheng, A complex variable meshless manifold method for fracture problems, Int. J. Comput. Meth. 7 (2010) 55–81.
[66] D. Levin, The approximation power of moving least-squares, Math. Comput. 67 (1998) 1335–1754.
[67] M.G. Armentano, R.G. Durán, Error estimates for moving least square approximations, Appl. Numer. Math. 37 (2001) 397–416.
[68] C. Zuppa, Error estimates for moving least square approximations, Bull. Braz. Math. Soc. 34 (2003) 231–249.
[69] R. Cheng, Y. Cheng, Error estimates for the finite point method, Appl. Numer. Math. 58 (2008) 884–898.
[70] X. Li, J. Zhu, A Galerkin boundary node method and its convergence analysis, J. Comput. Appl. Math. 230 (2009) 314–328.
[71] I. Kaljevic, S. Saigal, An improved element free Galerkin formulation, Int. J. Numer. Meth. Eng. 40 (1997) 2953–2974.
[72] G. Maisuradze, D. Thompson, A. Wagner, M. Minkoff, Interpolating moving least-squares methods for fitting potential energy surfaces: detailed
analysis of one-dimensional applications, J. Chem. Phys. 119 (2003) 10002–10014.
[73] H. Netuzhylov, Enforcement of boundary conditions in meshfree methods using interpolating moving least squares, Eng. Anal. Boundary Elem. 32
(2008) 512–516.
[74] H. Ren, Y.M. Cheng, W. Zhang, An improved boundary element-free method (IBEFM) for two-dimensional potential problems, Chin. Phys. B 18 (2009)
4065–4073.
[75] H. Ren, Y.M. Cheng, W. Zhang, An interpolating boundary element-free method (IBEFM) for elasticity problems, Sci. China Phys.: Mech. Astron. 53
(2010) 758–766.
[76] H. Ren, Y.M. Cheng, The interpolating element-free Galerkin (IEFG) method for two-dimensional elasticity problems, Int. J. Appl. Mech. 3 (2011) 735–
758.
[77] H. Ren, Y.M. Cheng, The interpolating element-free Galerkin (IEFG) method for two-dimensional potential problems, Eng. Anal. Boundary Elem. 36
(2012) 873–880.
[78] H. Ren, J. Cheng, A. Huang, The complex variable interpolating moving least-squares method, Appl. Math. Comput. 219 (2012) 1724–1736.
[79] H. Ren, Y.M. Cheng, A new element-free Galerkin method based on improved complex variable moving least-squares approximation for elasticity, Int. J.
Comput. Mater. Sci. Eng. 1 (2012) 1250011.
[80] J.F. Wang, F.X. Sun, Y.M. Cheng, An improved interpolating element-free Galerkin method with nonsingular weight function for two-dimensional
potential problems, Chin. Phys. B 21 (2012) 090204.
[81] J.F. Wang, J.F. Wang, F.X. Sun, Y.M. Cheng, An interpolating boundary element-free method with nonsingular weight function for two-dimensional
potential problems, Int. J. Comput. Meth. 10 (2013) 1350043.
[82] F.X. Sun, J.F. Wang, Y.M. Cheng, An improved interpolating element-free Galerkin method for elasticity, Chin. Phys. B 22 (2013) 120203.
[83] F.X. Sun, C. Liu, Y.M. Cheng, An improved interpolating element-free Galerkin method based on nonsingular weight functions, Math. Prob. Eng. 2014
(2014) 323945.
[84] M. Bitaraf, S. Mohammadi, Large deflection analysis of flexible plates by the meshless finite point method, Thin-Walled Struct. 48 (2010) 200–214.

You might also like