Engineering Analysis with Boundary Elements 50 (2015) 341–351

Contents lists available at ScienceDirect

Engineering Analysis with Boundary Elements

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

Higher order multipoint method – from Collatz to meshless FDM

Irena Jaworska n, Janusz Orkisz
Institute for Computational Civil Engineering, Cracow University of Technology, ul.Warszawska 24, 31-155 Cracow, Poland

art ic l e i nf o a b s t r a c t

Article history: The higher order multipoint meshless method for boundary value problems is considered in this paper.
Received 30 March 2014 The new method relies on the Collatz multipoint concept and the meshless FDM.
Received in revised form The main idea of multipoint technique is based on raising the local approximation order of a
5 August 2014
searched function by assuming additional degrees of freedom at each node of stencil, e.g. by including a
Accepted 20 September 2014
combination of nodal values of the right-hand side of considered differential equation. Thus, the FD
formula takes into account a combination of unknown function values which is equal a combination of
Keywords: additional d.o.f., e.g. right-hand side of PDE. In this way one may generate higher order FD operators
Multipoint finite difference method without any additional unknowns using the same set of nodes in stencil as in the non-multipoint case.
Meshless method
Essential modifications and extensions of the old classical multipoint formulation have been
Moving least squares
introduced for the purpose of this research. New fully automatic multipoint meshless FDM uses the
Higher order approximation
moving weighted least squares approximation instead of the interpolation proposed by Collatz, and is
based on arbitrarily distributed cloud of nodes. Moreover, besides the local formulation, also various
global formulations of b.v. problems are possible.
Several numerical benchmark problems analyzed illustrate the effectiveness of the proposed
approach.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction additional degrees of freedom at the star nodes, taking into

Various ways exist in modern numerical analysis to improve
the solution precision of boundary value (b.v.) problems. There are
two main directions of such improvement. The first one is based
on the mesh density increase, preferably using an adaptive (h-
type) solution approach. The second option is provided by rising
the approximation order (p-type). It may be raised in several ways.
In the finite difference method (FDM) the oldest one is called the
deferred correction technique [1]. It uses FD stars (stencils) with
the increased number of nodes. Another way uses the same simple
stars, but with additional generalized degrees of freedom (d.o.f.)
introduced there [2]. The most recent concepts [3] are based either
on evaluation of the higher order derivatives and using them as
“correction terms” in the FDM equations [4], or on the modified
multipoint approach [5] considered in this paper.
The new approach is based on the meshless finite difference
method (MFDM) [2] using arbitrarily irregular grids as well as
various formulations of b.v. problem. The multipoint MFDM
provides raising the order of approximation by introducing
n
Corresponding author.
E-mail addresses: irena@L5.pk.edu.pl (I. Jaworska),
plorkisz@cyf-kr.edu.pl (J. Orkisz).
account a combination of searched function values together with
combination of the right hand side values of the considered
equation (specific approach) or other chosen operator (general
approach). In this way one may generate higher order FD operators
using the same set of nodes in a meshless FD (MFD) star (stencil,
Fig. 1a) as in the non multipoint case. This increase the computa-
tional effectiveness in the specific multipoint version (Fig. 1b),
where the higher order (HO) approach is obtained without any
additional unknowns. The general version (Fig. 1c) of multipoint
method instead provides general HO solution approach for all
types of b.v. problems and various physical processes being
modeled on the same geometric domain due to relations between
unknown function and its derivatives.
The basic concept of the multipoint solution approach was
introduced long time ago by Lothar Collatz [6] as improvement of
the FDM. In Section 2 of this paper the main idea of the Collatz
multipoint method is presented, whereas in Section 3 the features of
multipoint concept are summarized. In his book [6] Collatz has
described the essence of HO multipoint concept and has calculated
multipoint FD formulas for several differential operators only due to
computational difficulties with hand calculations. This basic multi-
point approach has been reformulated by the authors and extended to
the fully automatic multipoint meshless finite difference method
which is discussed in Section 4. In comparison with the Collatz

http://dx.doi.org/10.1016/j.enganabound.2014.09.007
0955-7997/& 2014 Elsevier Ltd. All rights reserved.
342 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351

Fig. 1. MFD star used in (a) meshless FDM; (b) specific multipoint MFDM and (c) general multipoint MFDM.

approach, the following basic modifications have been included in the an auxiliary function Φ summed over all nodes j (j¼1,2,…m ) of
new method: the FD star with the central node i, the basic multipoint formula is
assumed as follows:
– the various global (weak) and global-local formulations of b.v.
problems may be applied besides the local (strong) one; Φi ¼ ∑cj uj  ∑αj f j ; ð2Þ
jðiÞ jðiÞ
– cloud of arbitrarily distributed nodes may be used instead of
regular meshes; and where cj ¼cj(i), α j ¼ α j(i), and
– the moving weighted least squares (MWLS) approximation
technique [2,7] is used instead of the interpolation one. ∑cj uj ¼ ∑αj f j for Φi ¼ 0: ð3Þ
jðiÞ jðiÞ

The paper is illustrated with the selected results of benchmark

To obtain this formula, one expands uj and the right hand side
tests (Section 5). Eventually, benefits of the proposed approach are
of the considered equation f j ¼ L uj into the truncated (by includ-
highlighted in the final remarks (Section 6).
ing higher order terms and neglecting Δ0j and Δj ) Taylor series
Only the main concept of the new multipoint MFDM approach
in respect of the stencil central node i
is described here, having the details of particular problems to our
other papers ([8] and articles that follows). uj uj fj fj j m ð4Þ
j j

Assuming the matrix notation, one may write relation (2) as

2. Interpolation based multipoint approach

Φi ¼ Ci ui , where Ci ¼ cjðiÞ ; αjðiÞ are unknown coefficients and
n o
As mentioned before, the main idea of the Collatz multipoint ui ¼ ujðiÞ ; f jðiÞ – degrees of freedom, and the truncated Taylor
approach [6] relies on an improvement of the standard FDM due to series (4) with respect to the central node Pi as follows ui  Ψi Dui ,
introducing into the FD operator values of the right hand side of  
where Dui ¼ ui ; u0i ; :::; ui ðpmax Þ is the vector of all local type
the considered differential equation (the specific approach) or the
derivatives up to pmax order, while the matrix Ψ is the Taylor
unknown function derivatives (the general approach) defined at
series coefficients matrix.
nodes of the FD star (Fig. 1).
Unknown coefficients C of the multipoint finite difference
Thus, in the multipoint method two approaches are distin-
formula discretizing the problem (1) may be found from the
guished namely the specific and the general ones. Both of them
provide higher order FD operators. The specific approach is more requirement Φi  Ci Ψi Dui ¼ 0 and Ci Ψi ¼ 0, taking into
simple, but it may be applied only to linear boundary value account their obvious linear dependence. Here Φi is a part of Φi
problems. The general multipoint approach is more complex, but involving only the terms up to the order pmax , resulting from the
it may be used in all types of b.v. problem formulation, linear and truncated Taylor series.
non-linear ones. Therefore, the coefficients cj ¼cj(i) and αj ¼aj(i) are computed by
solving the following system of equations:
2.1. Specific case 8 p
>
>
h
∑ cj j ¼ 0; p ¼ 0; 1; :::; r  1
>
> for derivative orders
< j p!
The specific approach is simpler than the general one, allows to ! :
provide HO solution without any additional unknowns, but its >
>
p ðp  qÞ þ
> ∑ cj hj  αj ∑ lq hj
> ¼ 0; for p ¼ r; :::; p ;
application is more restricted. Besides the d.o.f. assigned to : p!
q
ðp  qÞ þ ! max
j
unknown function values it also uses known values of a given ð5Þ
differential operator (e.g. right hand side of ODE or PDE).
Consider the locally formulated boundary value problem where r is the lowest derivative order in f, pmax is the highest
n1 derivative order with imposed zero coefficient value.
Lu ¼ uðnÞ þ ∑ lq uðqÞ ¼ f ðPÞ; u ¼ uðPÞ; P A Ω; ð1Þ In this way e.g. for the 1D b.v. problem
q¼0
a U u0 þ u″ ¼ f ; u0 ¼ uN ¼ 0 ð6Þ
with relevant boundary conditions
Gu ¼ gðPÞ; P A ∂Ω: and tri-nodal stencil, the FD multipoint operator
When the nodes in the domain Ω are introduced, one may
Au Bu Cu f f f
select appropriate FD stars and discretize the given operator. Using
 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351 343

is calculated as follows: The multipoint FD discretization is applied assuming FD stars

 and d.o.f. as well as the function
 2 3 4 5
A  u0 ¼ u1  hu01 þ h2 u″1  h6 u″0
1 þ 24u1  120u1 þ:::
h IV h V

 k
Φi ¼ ∑k cj uj  ∑k αj uðkÞ
j ; ð9Þ
B  u1
 2 3 4 5
C  u2 ¼ u1 þ hu01 þ h2 u″1 þ h6 u″0
1 þ 24u1 þ 120u1 þ:::
h IV h V jðiÞ jðiÞ

 hence
α  f 0 ¼ au01 þ u″1  hðau″1 þ u″01 Þ þ h2 ðau″01 þ uIV
2 3 4

1 Þ  6 ðau1 þ u1 Þ þ 24ðau1 þ u1 Þ þ:::

h IV V h V VI

β  f 1 ¼ au01 þ u″1 ∑k cj uj ¼ ∑k αj uðkÞ for k
Φi ¼ 0; ð10Þ
γ  f 2 ¼ au01 þ u″1 þ hðau″1 þ u″01 Þ þ h2 ðau″01 þ uIV 3 4 j
1 Þ þ ðau1 þ u1 Þ þ ðau1 þ u1 Þ þ:::
h IV V h V VI jðiÞ jðiÞ
2 6 24

The following system of equations in respect of unknown where k Φi corresponds to Φi for k-th derivative uðkÞ
j replacing f j .
coefficients A, B, C, α, β, γ is obtained by using summation of all Values uj and uðkÞ
j are expanded into the truncated Taylor series,
coefficients at the function value u1 and its derivatives up to 5th including higher order terms, with respect to a chosen central
order: node i
( 2
! 2
!)
u B A C k
Φi ¼ ∑ k
cjðiÞ ui þ hj u01 þ
hj
u″1 þ ::: :  k αjðiÞ ui ðkÞ þhj ui ðk þ 1Þ þ
hj
ui ðk þ 2Þ þ ::: :
2 2
hC A a j
u
h ð11Þ
u A C ha
As in the specific case, the unknown coefficients c and α are
h h computed by setting to zero relevant coefficients at subsequent
u C A h a
derivatives uðqÞ
i , q ¼ 0; 1; 2; :::; p, and solving resulting system of
u IV h A C
h h
a
equations
8
> hj p
uV h h h > ∑ cj p! ¼ 0;
>
<
for derivative orders p ¼ 0; 1; …; k  1;
C A a j
  :
>
> hp h ðp  kÞ þ
>
: ∑ cj p!j  αj ðpj  kÞ ! ¼ 0; for p ¼ k; …; pmax :
These equations are linearly dependent. Only five of them are j
þ

solved in terms of the parameter A. Assuming a ¼1 and h ¼1, the

ð12Þ
multipoint FD operator
39ui  1  144ui þ 105ui þ 1 ¼ 2f i  1 þ 56f i þ 8f i þ 1 ð7Þ Thus the relation

with O(h5) truncation error is obtained.

k
CU ¼ k α UðkÞ -UðkÞ ¼ k α  1 k CU ð13Þ

is found in the whole domain Ω. Here U ¼ fui g is the vector of

2.2. General case searched unknowns ui in the whole domain, while
C ¼ ½cjðiÞ  and α ¼ ½αjðiÞ  are the global matrices of coefficients
When the specific formulation cannot be applied (e.g. for non- for all nodes i¼1,2,…N of the considered mesh.
linear b.v. problems), the general one may be used. The combina- Finally, in this way, all required FD operators of derivatives uðkÞ
j
tion of the u(k) derivatives is used in order to obtain the higher for each k emerging from the considered differential equation are
order multipoint finite difference operators instead of the known generated in the whole domain Ω. Such operators are used to
right hand side values of a considered equation, as it is in the discretize the given b.v. problem (8) and solve the resulting
specific case. simultaneous equations in respect of U.
Let us consider a local (strong) formulation of a given b.v. In Fig. 2 convergence of the specific and the general multipoint
problem for the n-th order ODE (in a similar way also the PDE may approaches, as well as the standard FD solution is compared for
be considered) the simple 1D test problem:
uðnÞ ¼ FðP; u; u0 ; :::; uðn  1Þ Þ for P AΩ ð8Þ u″ þ u0 ¼ f ðxÞ; uð0Þ ¼ 0; uð4Þ ¼ 0; x A ½0; 4;
and relevant b.c. for P A ∂Ω. π π x
π2 π x

f ðxÞ ¼ cos  sin : ð14Þ

4 4 16 4

2.3. Boundary equations

Taking into account a combination of different order deriva-

tives, e.g. u0 and u” as additional d.o.f. (in the general case) or the
right hand side of the given differential equation together with its
derivatives (in the specific case), one may obtain the difference
operator of the considered approximation order by using FD star
with number of nodes relatively smaller than in the standard case.
Such approach may be specially useful in order to obtain the finite
difference operator on the boundary of the domain. Assume e.g.
multipoint relation based on two nodes stencil only

Au0 þ Bu1 ¼ α0 u00 þ α1 u01 þ β 0 u″0 þ β 1 u″1 ; ð15Þ

Fig. 2. Comparison of solution convergence of the interpolation (Collatz) and
MWLS based general and specific multipoint approaches with the standard low and provide the FD discretization, as well as the expansion into the
order FD solution. Taylor series. Finally, the resulting higher order boundary
344 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351

difference operator is as follows: (ii) Lui ?=Mf i – task (i) reversed, M is known, L is searched; and
u1  u0 u00 þ u01 h  ″  (iii) Lui ?=Mf i ? – both L and M are sought to achieve the most
¼ þ u u″1 ; ð16Þ precise FD formula.
h 2 12 0
5
where truncation error term O(h ) is obtained.
As mentioned before, two basic versions of the multipoint
method namely the general and the specific ones are distin-
3. The main idea of the multipoint approach guished. The specific approach, based on multipoint formula type

Let us consider the local (strong) formulation of boundary value Cu ¼ αf ; ð24Þ

problem for the n-th order differential equation with appropriate
is simpler and easier in implementation than the general one, but
b.c.
( it deals with the linear b.v. problems only. In the specific
Lu ¼ f ; for P A Ω formulation the additional degrees of freedom are known, e.g.
u ¼ uðPÞ; ; ð17Þ
Gu ¼ g; for P A ∂Ω right hand side of ODE or PDE. In this way, the higher order
approximation solution can be obtained without any additional
where L ; G are differential operators, or an equivalent global unknowns and with the same computational cost as in the
(weak) formulation involving integral of the type standard MFDM.
Z In the general approach, where the unknown derivatives u(k)
FðP; u; u0 ; :::; uðnÞ ÞdΩ: ð18Þ are used besides the searched values instead of the right hand side
Ω
function
Using the FDM or MFDM solution approach, the classical
difference operator Lu is presented in the following form: Cu ¼ αuðkÞ ; ð25Þ
Lui  Lui  ∑cj uj ¼ f i ) Lui ¼ f i : ð19Þ
jðiÞ
the additional d.o.f. are sought. The advantage is wide applicability
of this case – the general HO solution approach for all types of b.v.
For example, well-known FD formulas problem may be obtained in this way. Solving a problem in general
   
ui þ 1  ui  1 ui þ 1  2ui þ ui  1 formulation, all derivatives up to order p are defined by relations
 u0i and 2
 ″i ð20Þ type u(k)C u, which are depend on discretization of the problem
2h h
domain only.
with error bounds O(h2) and O(h2) respectively. Each of these multipoint FD cases provides high order approx-
In the multipoint approach (both in the classical Collatz and the imation FD operators, using mostly the same number of nodes in
meshless ones), a combination of the right hand side values fj (j¼1, the stencil (or even smaller one taking into account two d.o.f. at a
…m) of the given differential equation (or other chosen operator) node), as needed to generate difference operator L in the non-
taken at all nodes of considered stencil is used rather than central multipoint FDM approach. This fact is advantage in comparison
nodal value fi only with other HO meshless method, since less calculations are
Lui  Lui  ∑cj uj ¼ ∑αj f j ) Lui ¼ Mf i : ð21Þ required – excessive number of nodes in stencil, especially for
jðiÞ jðiÞ arbitrarily irregular grid, leads to a large size of MFD star seriously
For example, using first derivatives as d.o.f., the multipoint affecting the computational efficiency and worth quality of FD
operator based on 3-nodes stencil is as follows: operators.

The multipoint approach may be combined with almost any
ui þ 1  ui  1 1 0 
 ui  1 þ 4u0i þ u0i þ 1 ; with error bound O h
4
ð22Þ other method, presenting an efficiency solution tool of various b.v.
2h 6
problems that improves the numerical precision.
and using second derivatives
ui  1  2ui þ ui þ 1 1 ″ 

þ 10u″i þ u″i þ 1 ;
4
 u with error bound O h :
h
2 12 i  1
4. Multipoint meshless FDM formulation
ð23Þ
The first truncated error term in the case of Eq. (22) is h4  uV/180, The most commonly applied method of computer analysis is
and in the case of Eq. (23) is h4  uVI/240. the Finite Element Method (FEM, [9]). Using FEM implies that the
Here, L is a differential and L is a difference operator respec- required discretization must be obtained by the mesh generation.
tively; j is a number of node in the selected FD star, where i is the This procedure remains challenging for complex geometries, mesh
central node, Mfi is e.g. a combination of the equations right hand distortions in large deformation problems, when remeshing dom-
side values, fj may presents value of the whole operator L or its inates computer solution time. Such difficulties are the reason of
part only, e.g. a specific derivative uðkÞ
j . meshless methods rising. Meshless methods become more impor-
In the multipoint formulation, the difference operators L and M tant for solving b.v. problems [10–15], so developing extended HO
are obtained by means of the Taylor series expansion of unknown meshless multipoint method seems to be a step in the right
function u and additional degrees of freedom, e.g. right hand side direction. For this purpose the multipoint approach based upon
part f, including higher order derivatives. the meshless discretization and approximation – meshless finite
In general, the FD operator L may be referred to: differential difference method [2] – is especially investigated.
eqs, boundary conditions, and integrand in the global formulation The main concept of the new meshless multipoint method
of b.v. problem. developed and presented here is based on the following
Several task types may appear in the multipoint solution foundations:
approach:
– the old original Collatz multipoint (interpolation) approach [6],
(i) Lui =Mf i ? – operator L is known from the standard equation discussed above;
and FD discretization, while M is sought (upgrade of the – arbitrarily irregular cloud of nodes [2,16] rather than a regular
classical FDM formulas); meshes;
 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351 345

– application of the moving weighted least squares approxima- The MFD formulas for all derivatives Du up to order p
tion [2,7,16] instead of the interpolation technique used by 8 0 ðpÞ
Collatz; < fu; u ; u″; …u g
> in 1D;
– any type formulations (global, local, mixed [15]) of b.v. Du ¼ fu; ux ; uy ; uxx ; …uðpÞ g in 2D; ð31Þ
>
: fu; u ; u ; u ; u ; …uðpÞ g in 3D;
problem. x y z xx

may be found then by solving the set of equations

The last three constitute the bases of the meshless finite
Ψ Du ¼ Cu u  Cf f; ð32Þ
difference method and are for the first time introduced into the
concept of the multipoint method. In other words, an extension of where
the MFDM onto the multipoint solution approach is
presented here. Ψ ¼ PT W2 P þPf T Wf 2 Pf ; Cu ¼ PT W2 ; Cf ¼ P f T W f 2 ; ð33Þ
 
Multipoint meshless FDM is specially dedicated to higher order ui ¼ ujðiÞ is the vector of function values at all nodes belonging to
approximation spanned over domains covered by arbitrarily dis- the considered MFD star with central node P i .
tributed cloud of nodes. Like in the original Collatz approach, two 2 3
1 2 1 2 1 p
cases namely the specific and the general ones, will be 1 h1 k1 2h1 h1 k 1 2k1 ::: p!k1
6 7
discussed below. 61 1 2 1 2
::: 1 p 7 hj ¼ hij ¼ xj  xi ;
6 h2 k2 2h2 h2 k 2 2k2 p!k27
Here P¼6 7;
6: : : : : : ::: : 7 kj ¼ kij ¼ yj  yi ;
4 5
1 p
1 hm : : : : ::: p!km
4.1. Specific case
ð34Þ

In the specific case the known values of a given differential is the matrix built upon the Taylor series expansion of all nodal
operator (e.g. right hand side of ODE or PDE) are used as additional values uj ; Pf is the similar type matrix obtained by the expansion
degrees of freedom for raising the approximation order. of the right hand side values f into the truncated Taylor series: if e.
Consider the locally formulated linear boundary value problem g. f ¼ uxx þ uyy , then Pf ¼Pxx þPyy and
(1). The multipoint MFD discretization based on the selected MFD 2 p2
3
0 0 0 1 0 0 h1 k1 0 0 ::: ðp 1 2Þ!k1
stars and expansion (4) a nodal function value uj and right hand 6 7
side value f j in respect of chosen central point Pi into the truncated 60 0 0 1 0 0 h k2 0 0 ::: ðp 1 2Þ!k2 7
p2
6 2 7
Pf ¼ 6 7
Taylor series (dashed terms uj and f j ) is introduced here. Due to 6: : : : : : : : : : ::: : 7
4 5
using additional d.o.f. fj, the Taylor expansions include the higher 0 0 1 0 0 hm km 0 0 ::: ðp  2Þ!km 1 p2
order terms up to the order p
2 p2
3
r rq 0 0 0 0 0 1 0 0 h1 k1 ::: 1
ðp  2Þ!k1
p
h ðrÞ p n hij 6 7
uj ¼ ∑ ui and LujðiÞ ¼ f j ¼ ∑ ∑ lq uðr þ qÞ : ð26Þ 60
6 0 0 0 0 1 0 0 h2 k2 ::: 1 p2 7
7
r ¼ 0 r! ðr qÞ! i þ6 ðp  2Þ!k2
r ¼0q¼0 7
6: : : : : : : : : : ::: 7
4 5
The purpose of the above manipulations is to form the multi- p2
0 0 0 0 0 1 0 0 hm km ::: 1
ðp  2Þ!km
point finite difference operator similarly to the specific case of the
Collatz multipoint approach as follows: ð35Þ

W, Wf are corresponding diagonal matrices consisting of the

∑ c j uj ¼ ∑ α j f j ; ð27Þ
jðiÞ jðiÞ squares of the weighting factors wij at these nodes, and m is a
number of nodes in the selected FD star at the central node i.
or the extended version, when the derivatives of the right hand In general, in the weighted least squares methods [17], the
side of considered PDE are used as additional d.o.f. also matrix Ψ will be non-singular for nodes of each MFD star,
ðsÞ provided there is a set on which the rows of matrix P are linearly
∑cj uj ¼ ∑αj f j þ ∑∑ s αj f j : ð28Þ independent. Therefore, at each node P i (i ¼ 1; …; N, where N is the
jðiÞ jðiÞ jðiÞ s
number of nodes in the domain), the local multipoint MFD
The MWLS approach will be applied here instead of interpolation formula, as well as all derivatives up to order p
one used by Collatz. All nodes of a MFD star should be
Du ¼ c u  α f ð36Þ
involved here.
The locally (stencil) defined weighted error functional may be are obtained by solving the simultaneous Eq. (32). Substituting
assumed in the form given below. Values of function u, right hand this result into Eq. (1) one obtains
side f, and optional its derivatives (in extended version) are n n  
included as follows ∑ lq uðqÞ  ∑ lq cq u  αq f ¼ f ; ð37Þ
q¼0 q¼0
 2
2
J i ¼ ∑ uj  uj 0 w2ij þ ∑ f j  Luj n w2ij : ð29Þ where u, c, and α are relevant vectors defined for each MFD star.
jðiÞ jðiÞ
Application of the formula (37) at each node in the whole domain
Here, the weighting factor n wij ¼ ρij p þ n  1 is a function of the
Ω provides a simultaneous algebraic equations in respect of the
unknown U¼{ui}, i¼1,…N in Ω.
distances ρij ¼ jP i P j j between central node Pi and node Pj
Considering an example of Poisson's equation uxx þ uyy ¼f and
belonging to FD star. The weight depends on the parameters p,
extracting the third (for uxx) and fifth (for uyy) equations from the
n. Moreover p is the local approximation order, n is the differential
system (36) one may obtain as follows:
equation order.

Minimizing the functional J is done by taking advantage of the uxx ji ¼ ∑C 3;j uj ∑α3;j f j ; uyy i ¼ ∑C 5;j uj  ∑α5;j f j ; ð38Þ
condition jðiÞ jðiÞ jðiÞ jðiÞ

    
∂J uxx þ uyy i  f i ¼ ∑ C 3;j þ C 5;j uj  ∑ α3;j þ α5;j f j : ð39Þ
¼ 0: ð30Þ
∂fDug jðiÞ jðiÞ
346 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351

Finally, for each FD star (with central node i ) the relation involving simultaneous algebraic equations:
only unknowns u is obtained
   
Ψ U Du ¼ Cu U u  C k U uðkÞ ; ð47Þ
∑ C 3;j þ C 5;j uj ¼ f i þ ∑ α3;j þ α5;j f j ð40Þ
jðiÞ jðiÞ
where Ψ, Cu and Ck are formed in the same way as (33). Hence
unknown derivative values Du of a searched function u are
and generally in the whole domain Ω calculated. Consequently, all local multipoint MFD formulas for
C U U ¼ F þA UF: ð41Þ derivatives of order q ¼ 0; 1; …; p are obtained as follows:

Here C and A are banded matrices of multipoint coefficients, F Du ¼ c u  α uðkÞ : ð48Þ

is right hand side values vector. It is worth noticing, that besides However, it is impossible to obtain relation between u and u(k)
the right hand side function f values (there are two d.o.f. at each from formula (48) straightforward, because two of these equations are
node in this case), its derivatives may be also taken into account linearly dependent and present the identity (ui ¼ ui ; uðkÞ ðkÞ
i ¼ ui ):
likewise (28) in order to obtain the raised order MFD operators 0 1 2 3
ui 1 0 ::: 0
(due to more than two d.o.f. at a node), where the final set of B u0 C 6 c 7 0 1
simultaneous equations is as follows: B i C 6 1;0 c1;1 ::: c1;m  1 7 ui
B C 6 7 B
B ::: C 6 : : ::: : 7 B ui þ 1 C C
B C 6 7 B C
C U U ¼ F þA UF þ ∑ Bs UFðsÞ : ð42Þ B uðkÞ C ¼ 6
:::
7U
0 7 B ::: C
B i C 6 0 0
s B C 6 7 @ A
B ::: C 6 : : ::: : 7 ui þ m  1
@ A 4 5
uðpÞ
i cp;0 cp;1 ::: cp;m  1
4.2. General case 2 3
0 0 ::: 0 0 1
6α 7 uðkÞ
6 1;0 α1;1 ::: α1;m  1 7 i
Let us consider the b.v. problem (17), or an equivalent global 6 7 B C
6 : : ::: : 7 B uðkÞ C
(weak) formulation involving integral (18). The general multipoint 6 7 B iþ1 C
þ6 7UB C: ð49Þ
case may be used here and in all situations, when the specific case 6 1 0 ::: 0 7 B ::: C
6 7 @ A
cannot be applied, e.g. for non-linear b.v. problems. 6 : : ::: : 7 uðkÞ
4 5 iþm1
In order to obtain any required multipoint finite difference αp;0 αp;1 ::: αp;m  1
formula considered is the combination
Therefore, required are additional equations with additional
∑cj uj ¼ ∑αj uðkÞ
j ð43Þ degrees of freedom, e.g. derivative u(l) instead of u(k) (l op, p –
jðiÞ jðiÞ
approximation order) assumed at nodes of the FD star. Minimiza-
of unknown function values uj and its k-th derivative values uðkÞ j ,
tion of the functional of the same type as (45) yields similarly to
assumed as unknown additional degrees of freedom. These deri- (48)
vatives are used instead of known values of the right hand side of
Du ¼ c~ u  α
~ uðlÞ : ð50Þ
the given differential equation emerging in the specific case
formulation. Using set of Eq. (49), the derivative uðlÞ
i may be obtained as
All nodes of each FD star are involved, similarly as in the follows:
specific approach. Several different k values are chosen to match
uðlÞ
i ¼ cl u  αl u
ðkÞ
ð51Þ
all derivative orders appearing in the differential operator L. The
main feature of the multipoint MFDM general concept is that the Moreover simultaneous Eq. (50) give the derivative uðkÞ
i
partial problems should be solved first. The relations between
u and uðkÞ (between u and ux , u and uy ,… in 2D or 3D) uðkÞ
i ¼ ck u  αk u :
~ ~ ðlÞ
ð52Þ
inside the whole domain Ω should be provided in this case, Here cl ; c~ k ; αl ; α ~ k are vectors of coefficients related to the
assuming each time required k-th derivative uðkÞ instead of f. For corresponding derivatives (l-th and k-th) of the matrices c, c; ~ α~
this purpose, likewise the specific case, the function value uj and and α. In this way, using Eqs. (51, 52) the relation between u and
its derivative uðkÞ
j are developed into the truncated Taylor series u(k), as well as relation between u and u (l) may be finally obtained.
The required number of such relations may be reduced to M for
uj uj j u jk ujk kj ð44Þ
the M-dimensional b.v. problem (e.g. uxCu and uyCu in 2D).
Afterwards the locally defined weighted error functional Higher order derivative relations could be obtained by multiplying

2 the first ones. For example, if u x ¼A  u, then u xx ¼A  (A  u)¼A2  u.
 2
J ik ¼ ∑ uj  uj 0 w2ij þ ∑ uðkÞ ðkÞ
j  uj
k 2
wij ð45Þ This alternative approach, in which only the first order relations
jðiÞ jðiÞ are generated, is called AA instead of the standard approach AB,
is formed. Assumed is one of the following weights: where each relation for higher order derivative is calculated
independently – u x ¼A  u and u xx ¼B  u.
!pþk1
g4 Variety of tests have been done for versions AA and AB. The
wij ¼ ρ2ijð  p þ k  1Þ ; or k w2ij ¼ ρ2ij þ 2
k 2
; ð46Þ very first results show that AA procedure is more effective, saving
ρij þg 2
computer time and storage space. Sometimes, however, the AB
where ρij ¼ jP j  P i j, and gZ0 is a smoothing parameter. Both types approach exhibits a better solution convergence for a series of
of weights depend on ρij and have the property that the weight meshes. In Fig. 3, the convergence of the AA and AB general
function magnitude decreases as the distance between nodes approaches is compared for regular grids.
increases. Using the weight of the first type, numerically accurate Having found in the whole domain Ω the multipoint FD
interpolation at the central node is achieves. In the case of relations u(k)Cu for all k-th order derivatives emerging in the
smoothing parameter g ¼0, the second type of weight turns out considered b.v. problem, one substitutes these relations to the
to be the same as the first one. given linear or non-linear differential equation. After such dis-
Using arbitrarily distributed nodes and the MWLS approxima- cretization the ODE (PDE) depends on the primary unknowns U
tion at each node P i considered as the central one, minimization of only. It is worth noticing, that the d.o.f. like u(k), u(l) (e.g. ux, uy ) are
the functional ∂J=∂fDug ¼ 0 yields the following type of used in the multipoint general version only additionally, in order
 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351 347

to obtain the HO approximation approach. The final set of simulta-
neous algebraic equations depends only on the basic d.o.f. the
unknown values of u at each node.
Numerical implementation of the above general approach
algorithm for PDE is much more complicated than it is in the
ODE case, because the problem (47) may be ill-conditioned. Such
situation happens e.g. when high order approximation and nearly
regular symmetric arrangement of star nodes are used. Success is
possible when the d.o.f. assumed together with the stencil con-
figuration allow for evaluation of all partial derivatives emerging
in the Du vector up to the order p. To overcome the possibility of
ill-conditioning, several different versions of multipoint algo-
rithms have been proposed and tested [8].
It should be emphasized, that the multipoint general formula-
tion provides HO general solution approach for all types of
boundary value problem and physical process being modeled on
the same mesh, due to independence of relations u(k)Cu on
considered differential equations.
method the relations between u and its derivatives are applied to
obtain equation type Au ¼ g(Pi).
Without introduction of additional points outside the domain,
the stencil central node located on the boundary is not at the
center of gravity, where the approximation is expected to be the
most accurate. However, the size of such MFD star may be
reduced, providing the approximation of better quality then. Both
cases of multipoint MFD method use at least two d.o.f. at each
node. Due to the abundance of stencils d.o.f. the same approxima-
tion order of the boundary conditions as inside the domain may be
obtained by using a MFD star with smaller number of nodes
(Fig. 4a–b). The extended variant of multipoint approach, where
more than two d.o.f. are used at each node (Fig. 4c – even four
nodes would be enough here to obtain the full third order of
approximation) can also be applied. In the specific case the
derivatives of right hand of considered equation may be used as
additional d.o.f. in the multipoint operator like (28) to obtain

4.3. Boundary conditions

Quality of the b.v. problems solution essentially depends on the

approximation precision in the boundary neighborhood, which
should be at least the same as inside the domain.
For imposing the essential (Dirichlet) boundary conditions, the
first type of weight functions (46) may be used. Singular weight
guarantee numerically accurate interpolation at the central node
of each MFD star and satisfy boundary conditions ui ¼g(Pi) at
nodes Pi A ∂Ω. In case of natural (Neumann) or combined bound-
ary conditions – discretization of b.c. differential operator is
applied in the same way as for the operator inside the domain.
When the specific form is used, the Eq. similar to (36) cu – αf ¼g
(Pi) are generated on the boundary. In general form of multipoint

Fig. 3. Comparison of AA–AB solution convergence on regular grid; 5-129 nodes, Fig. 5. Solution convergence for various approximation orders, general approach:
benchmark test (14). (a) regular and (b) adaptive irregular meshes presented in Fig. 6.

u u
u, f u, f u, f u, f, f ’ u, f, f’
u
u u, f u, f, f’
u u, f u, f, f’
u

Fig. 4. MFD approximation on the boundary: (a) 2nd order, MFDM; (b) 2nd order, multipoint MFDM and (c) 3rd order, extended multipoint MFDM.
348 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351

better quality solution on the boundary and in the general case – where for the 4th approximation order with O(h5) error term, the
the derivatives u(k) together with u(l), as follows: coefficients are as follows:

∑cj uj ¼ ∑αj uðkÞ ðlÞ

j þ ∑β j uj : ð53Þ a ¼ ½  0:067 1:000 0:000  1:000 0:067;
jðiÞ jðiÞ jðiÞ
α ¼ ½0:024 0:061 1:565 0:061 0:024;
Collocation approach carried out in the whole domain Ω [ ∂Ω
b ¼ ½  0:043  0:457 1:000  0:457  0:043;
may be used then in order to generate the simultaneous algebraic
equations and provide the final high order FD solution. β ¼ ½0:023 0:005 0:575 0:005 0:023;
Precision of the multipoint FD solutions essentially depends on
the order of Taylor series expansion. If truncated Taylor series in
the multipoint approaches includes terms up to the order p (one
– 5th approximation order:
may use a high order), the error bound of solution is expected
[5,6,18,22] to be an O(hp þ 1), where h is an average distance a ¼ ½  0:330 1:000 0:000  1:000 0:330;
between the nodes. Analysis of analytical as well as a posteriori α ¼ ½  0:114  0:092 1:091  0:092  0:114;
error estimate is presented in the paper that follows. b ¼ ½0:026  0:526 1:000 0:526 0:026;
The same multipoint MFD approach may be applied in the case β ¼ ½0:002 0:002 0:414 0:002 0:002;
of the global and global-local [15] formulations of b.v. problem
given instead of PDE (ODE). The integral is discretized using
– 6th approximation order:
formulas for derivatives u(k)Cu, found in the way mentioned
above. The detailed description of global and global-local multi- a ¼ ½  0:262 1:000 0:000  1:000 0:262;
point solution approach [19,20] will be presented in the α ¼ ½  0:093 0:005 1:127 0:005  0:093;
next paper. b ¼ ½1:000 2:917  7:835 2:917 1:000;
β ¼ ½  0:055  1:358  4:092  1:358  0:055;

5. Preliminary numerical results

– 7th approximation order:
The multipoint approach described here is used to analyze a ¼ ½0:156 1:000 0:000  1:000  0:156;
boundary value problems. A variety of 1D and 2D tests done show α ¼ ½0:038 0:600 1:350 0:600 0:038;
that the new multipoint meshless FDM may be a useful solution b ¼ ½ 0:309  0:191 1:000  0:191  0:309;
tool for analysis of b.v. problems locally or globally formulated. In β ¼ ½0:020 0:347 0:690 0:347 0:020:
the benchmark tests examined was precision of the results, their
convergence and efficiency obtained for various formulations of
– and 8th approximation order:
the multipoint method.
Examples of the following comparisons have been investigated: a ¼ ½0:156 1:000 0:000  1:000  0:156;
α ¼ ½0:038 0:600 1:350 0:600 0:038;
– interpolation based (Collatz type) and MWLS based multipoint b ¼ ½ 0:098  0:403 1:000  0:403  0:098;
approach (Fig. 2); β ¼ ½0:005 0:144 0:494 0:144 0:005:
– specific and general type analysis (Fig. 2);
– version AA and AB convergences (Fig. 3);
– results obtained on regular and irregular meshes (Fig. 5); and
– solution and its derivative convergences (Fig. 7).
5.2. 2D benchmark test
Using preliminary simple examples, the influences of the
approximation order (Figs. 3,5), weighting factor type [21] and In Figs. 7 and 8 the two-dimensional Poisson's b.v. problem
discretization of boundary conditions on solution quality and ∇2 u ¼ f ðx; yÞ together with the Dirichlet boundary conditions given
effectiveness were analyzed. in the domain for 0 r x r 1; 0 r y r 1 was analyzed and com-
pared with the FEM solution, where
5.1. 1D benchmark test "  2  2 #
4 xβ yβ
f ¼ 6x  6y  2  4 4 U
Among other one-dimensional problems, the benchmark test α 2 α 2 α
described above in chapter 2.2, Eq. (14), was solved. Figs. 2,3,5 "  2  2 #
xβ yβ
present effectiveness of the proposed technique in terms of exp   ; α ¼ 0:2; β ¼ 0:5;
convergence rates of this problem. The finite difference operators α α
used in Fig. 2 (based on the tri-nodal stencil and regular distances
h between nodes) are as follows: corresponding to the exact solution [8]
– in the case of low order FD this is the sum of both FD  2  2 !
xβ yβ
formulas (20) with appropriate error term; uðx; yÞ ¼  x3  y3 þ exp   : ð55Þ
α α
– in the case of Collatz (interpolation) and the MWLS (mesh-
less) based multipoint methods, the obtained operators for chosen The general multipoint formulas, e.g. used to calculate solution
stencil are the same. The specific approach uses formula (7), and convergence presented in Fig. 7, are based on the nine-nodes
the general one uses the sum of Eqs. (22) and (23). regular stencil and relations between uCux and uCuy, similarly as
In Fig. 5a, for example, the general multipoint formulas of five- in the 1D case Eq. (54), with the coefficiens as follows:
nodes stencil are based on the relations
½  0:032 0 0:032 1:0 0 1:0  0:032 0 0:032
5 5 5 5
∑ aj uj ¼ ∑ αj u0j and ∑ bj uj ¼ ∑ βj u″j ; ð54Þ  u ¼ ½0:039  0:045 0:039 0:1 0:8 0:1 0:039  0:045 0:039
j¼1 j¼1 j¼1 j¼1  ux
 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351 349

Fig. 6. An example of the set of adaptive irregular meshes (8, 18, 63 and 119 nodes) used for the benchmark test (14).

solution, 3 app.order
0 X derivative, 3 app.order
Y derivative, 3 app.order
FEM, triangular elements
1
log(error)

1.4 1.2 1 0.8 0.6

log(h)

Fig. 7. Convergence rates of FEM (triangular elements) and multipoint solution and
its derivatives.

Fig. 9. Irregular mesh (971 nodes) for railroad shape.

The local formulation in the form of the Poisson equation with

the essential boundary conditions
(
∇2 Φ ¼  2Gθ; in Ω
; ð56Þ
Φ ¼ 0; on ∂Ω

is considered, where Φ is the Prandtl stress function, Gθ ¼1 is the

torsional stiffness, Ω – domain of the bar cross-section.
The total shear stress for prismatic bar is
  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
τ ¼ grad Φ ¼ τ2zx þ τ2zy ; ð57Þ

where shear stresses are

Fig. 8. Solution convergence results for both the general multipoint method and
the standard FDM.
and
½  0:032  1:0  0:032 0 0 0 0:032 1:0 0:032
 u ¼ ½0:039 0:1 0:039  0:045 0:8 0:045 0:039 0:1 0:039
 uy :
An example of multipoint FD operator derivation has been
presented in detail in [8].
5.3. Rail shape prismatic bar twisting
The regular and arbitrarily irregular meshes applied to analysis
of various b.v. problems, including de Saint-Venants one, were also
examined. Among others, the shear stress in prismatic bar of
railroad rail shape cross-section has been solved.
∂Φ ∂Φ
τzx ¼ ; τzy ¼ : ð58Þ
∂y ∂x
The results of shear stress analysis based on irregular mesh (see
Fig. 9) are presented in Fig. 10.
Variety of tests done confirm the observation that the higher
order approximation multipoint approaches ensure improvement
of solution quality (see Fig. 5).
Using the multipoint approach, the good solution precision can
be obtained with reducing number of nodes (Fig. 8). The advan-
tage in comparison with the finite element method is higher
convergence of derivatives in the multipoint method. Likewise
MFDM, the convergence rates of derivatives are very close to
convergence rate of unknown function, which is important to
solve engineering problems (Fig. 7).
However, besides convergence improvement in comparison
with the standard MFDM, highest approximation order results
may be not stable enough and may have worse convergence rate
due to truncation error (Fig. 5, orders 7 and 8). Also, the weighting
function employed, non-sufficient number and inappropriate dis-
tribution of nodes in MFD stars may influence stability of the FD
350 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351
Fig. 10. Prandtl stress function and total shear stress in railroad rail shape bar.
methods based on arbitrarily irregular meshes [23,24,8]. On the
other hand, the following security measures against inappropriate
MFD stars are assumed in the multipoint MFDM approach: each
stencil consists of a surplus of a number of d.o.f., and certain
geometrical criteria of node locations are satisfied [8,25]. Never-
theless, detailed analysis of stability of FD methods at arbitrary
irregular meshes still constitute an open research problem, espe-
cially of mathematical nature, and are beyond the scope of
this paper.
Results of tests for both multipoint versions, the general and
the specific ones, are encouraging and further tests are planned.
More details of the numerical analysis will be presented in the
following papers, dealing with particular features of the
multipoint MFDM.
6. Final remarks
The new multipoint technique based on the old Collatz higher
order concept and meshless FDM is being developed. This method
uses various formulations of b.v. problems and arbitrarily distrib-
uted cloud of nodes in the domain, as well as the MWLS
approximation. The main idea of the method has been presented
in this paper.
Two basic cases have been considered namely the general and
the specific ones. The specific version allows for more effective
analysis of linear b.v. problems. The higher order approximation
with corresponding high quality of the solution is obtained using
the same set of nodes in a stencil as in the standard MFDM and
without any additional unknowns. This increase the computa-
tional effectiveness of the multipoint approach. One of the great
attractions of the general case is its potential wide range of
applicability. This multipoint approach allows for analysis of all
types of b.v. problems, including non-linear ones. Solving a b.v.
problem in general formulation, all required derivatives are
defined as relations type u(k)Cu, which are depend on discretiza-
tion of the problem domain only. Therefore, the solution of almost
all problems in computational mechanics may be obtained
this way.
The main features of the proposed here generalized multipoint
method are as follows:
– the possibility of using various formulations of b.v. problems
(global, local, mixed);
– higher order approximation and improved precision of the
solution obtained on the arbitrarily irregular meshes;
– unknown function derivatives are obtained without any addi-
tional computational cost besides solution;
– convergence rates of derivatives are very close to those for the
unknown function itself.
This paper presents the main idea of the new multipoint
meshless FDM. There are, however, many open issues that need
further investigation. Among them, we have
– Further development of error analysis.
– Efficiency of various versions of general multipoint approach
[8].
– Further development of the multipoint approach for analysis of
b.v. problems given in the various global formulations, espe-
cially global-local (MLPG, [15,20]) ones.
– Comparison and combination of the multipoint MFDM with the
other discrete methods.
– Various tests and analysis of engineering applications.
References
[1] Hackbush B. Multi-grid methods and applications. Berlin: Springer-Verlag;
1985.
[2] Orkisz J. Finite difference method (Part III). In: Kleiber M, editor. Handbook of
Computational Solid Mechanics. Berlin: Springer-Verlag; 1998. p. 336–432.
[3] Orkisz J, Jaworska I, Milewski S. Meshless finite difference method for higher
order approximation. In: Proceedings of the third international workshop
meshfree methods for partial differential equations. Bonn, Germany; 2005.
[4] Milewski S. Meshless finite difference method with higher order approxima-
tion—applications in mechanics. Arch Comput Methods Eng 2012;19(1):1–49.
[5] Jaworska I. The multipoint meshless finite difference method for analysis of
boundary value problems of mechanics. (Ph.D. thesis). Cracow: Cracow
University of Technology; 2009.
[6] Collatz L. Numerische Behandlung von Differentialgleichungen. Berlin-
Göttingen-Heidelberg: Springer-Verlag; 1955.
[7] Lancaster P, Salkauskas K. Surfaces generated by moving least-squares
method. Math Comp 1981;155(37):141–58.
[8] Jaworska I. On the ill-conditioning in the new higher order multipoint
method. Comput Math Appl 2013;66(3):238–49.
[9] Zienkiewicz OC, Taylor RL. Finite element method its basis and fundamentals.
Elsevier; 2005.
[10] Belytschko T, Krongauz Y, Organ D, Fleming M, Krysl P. Meshless methods: an
overview and recent developments. Comput Methods Appl Mech Eng
1996;139:3–47.
[11] Fries TP, Matthies HG. Classification and overview of meshless methods.
Brunswick, Germany: Technical University Braunschweig; 2004.
 I. Jaworska, J. Orkisz / Engineering Analysis with Boundary Elements 50 (2015) 341–351
[12] Leitao VMA, Alves CJS, Duarte CA. Advances in meshless techniques. Springer;
2007.
[13] Griebel M, Schweitzer MA. Meshfree methods for partial differential equations
VI. Springer; 2013.
[14] Liu GR. Mesh free methods: moving beyond the finite element method. Boca
Raton: CRC Press LLC; 2003.
[15] Atluri SN, Shen S. The meshless local Petrov–Galerkin (MLPG) method.
Forsyth, GA: Tech Science Press; 2002.
[16] Liszka T, Orkisz J. The finite difference method at arbitrary irregular grids and
its applications in applied mechanics. Comput Struct 1980;11:83–95.
[17] Lancaster P, Salkauskas K. Curve and surface fitting. Academic Press Inc.; 1990.
[18] Jaworska I, Orkisz J. On some aspects of a posteriori error estimation in the
multipoint meshless FDM. Barcelona: WCCM2014; 2014; 2737–43.
[19] Jaworska I, Orkisz J. Multipoint meshless finite difference method – recent
development. Venice: WCCM8, ECCOMAS; 2008.
351
[20] Jaworska I, Orkisz J. On the multipoint meshless FD method using the local
Petrov–Galerkin approach. In: Eberhardsteiner J, Böhm HJ, Rammerstorfer FG,
editors. ECCOMAS, 2012. Vienna; 2012. p. 6582–91.
[21] Most T, Bucher C. New concepts for moving least squares: an interpolating
non-singular weighting function and weighted nodal least squares. Eng Anal
Bound Elem 2008;32:461–70.
[22] Wendland H. Scattered data approximation. Cambridge University Press; 2005.
[23] Demkowicz L, Karafiat A, Liszka T. On some convergence results for FDM with
irregular mesh. Comp Methods Appl Mech Eng 1984;42:343–55.
[24] Benito JJ, Urena F, Gavete L. Influence of several factors in the generalized
finite difference method. Appl Math Model 2001;25:1039–53.
[25] Kączkowski Z, Tribiłło R. A generalization of the finite difference formulas. Int J
Numer Methods Eng 1975;3(9):581–99.