Journal of Computational and Applied Mathematics 448 (2024) 115896

Contents lists available at ScienceDirect

Journal of Computational and Applied Mathematics

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

Multinode Shepard method for two-dimensional elliptic boundary

problems on different shaped domains
Francesco Dell’Accio a ,∗, Filomena Di Tommaso a , Elisa Francomano b
a Department of Mathematics and Computer Science, University of Calabria, Via P. Bucci, Rende (CS), Italy
b Department of Engineering, University of Palermo, Viale delle Scienze, Palermo, Italy

ARTICLE INFO ABSTRACT

Keywords: In this paper, we continue the study on the application of multinode Shepard method to
Approximation by rational functions numerically solve elliptic Partial Differential Equations (PDEs) equipped with various conditions
Multinode Shepard method at the boundary of domains of different shapes. In particular, for the first time, the multinode
Collocation method
Shepard method is proposed to solve elliptic PDEs with Dirichlet and/or Neumann boundary
Elliptic PDEs
conditions. The method has been opportunely handled to efficiently work dealing with scattered
distribution of points and, to this aim, several experiments in different 2d domains have been
performed. Comparisons with the analytic solution and the results generated by the Kansa’s
RBF solvers have been reported referring to Halton points. The results are very promising and
should be of interest for applications in the real world.

1. Introduction

As well known, every field in science and engineering uses models based on differential equations: from biology to medicine,
from physics to chemistry, and all engineering sciences. Despite the huge number of known Partial Differential Equations (PDEs),
continuously discovered in modeling new problems, only a few possess closed-form solutions and numerical methods allow to
represent such solutions with the help of adequate designed software [1].
Elliptic PDEs are special kinds of second-order equations, which find applications in almost all areas of mathematics, from
harmonic analysis to geometry to Lie theory, as well as numerous applications in physics, and they have a well-developed theory. The
Laplacian operator, in particular, is a fundamental component in a large number of multi-dimensional linear models of mathematical
physics: to name a few, the Laplacian arises in descriptions of various classical and quantum wave phenomena, as well as models
involving diffusion and viscosity effects [2].
Often the elliptic PDEs are coupled with boundary conditions, to define a boundary value problem in which the solution of the
given equation is required to satisfy conditions, such as Dirichlet or Neumann conditions or mixed ones, on the boundary of the
problem domain.
The most popular numerical methods for solving PDEs are finite difference methods (FDMs), spectral methods, finite element
methods (FEMs), boundary element methods (BEMs), finite volume methods (FVMs). Despite their popularity, classical methods
based on local polynomial interpolation, have low algebraic convergence rates. On the other hand, global polynomial methods, have
exponential convergence rates, but in any case a structured grid is required. Radial basis functions (RBFs) [3–5] have been adopted
into the PDEs framework too [6–8]. The direct use of RBFs to solve PDEs via collocation was first attempted by Kansa in 1990 [6]

∗ Corresponding author.
E-mail addresses: [email protected] (F. Dell’Accio), [email protected] (F. Di Tommaso), [email protected]
(E. Francomano).

https://doi.org/10.1016/j.cam.2024.115896
Received 30 October 2023; Received in revised form 6 February 2024
Available online 2 April 2024
0377-0427/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

and widely used in the applications. In this method, multiquadric RBFs were used to approximate the solution, and point collocation
was considered to enforce the governing equation and boundary conditions [9]. In the last years, it has been also investigated a
generalization of the Kansa method, called polynomial-augmented RBF, which consists of adding multivariate polynomials to the RBF
basis, and then imposing matching constraints [10,11]. This approach increases the accuracy of the approximation by high-degree
polynomials and stabilizes the system through the RBFs [12]. The RBFs collocation methods do not require a grid and are powerful
when one has to work with irregular domains [13–17]. The disadvantages are that often the condition number of the collocation
matrix is too large and severe ill-conditioning results [18]. As shown in recent works [19,20], these drawbacks do not occur in using
the multinode Shepard (MS) method, instead of the RBFs, in collocation.
In this paper, we continue our studies on the possible application of the MS method to solve elliptic PDEs with Dirichlet or
mixed boundary conditions by considering domains of different shapes in 2d. In order to maintain the paper self-contained, we
briefly outline the main features of the MS method in Section 2. In Section 3, the MS method is used to efficiently solve different
elliptic PDEs, with Dirichlet boundary conditions, dealing with scattered distribution of points, and comparisons with the analytic
solution and the results provided by Kansa’s RBFs solver are reported. Similar experiments, using Halton points, have been conducted
in Section 5 for the case of elliptic PDEs with mixed boundary conditions. Finally, Section 6 is devoted to conclusions and further
perspective of work.

2. A brief overview on the multinode Shepard method

To maintain the paper self-contained, in this section, we shortly describe the MS method, which is the main ingredient of the
adopted numerical approach. The MS method is an accurate procedure for reconstructing functions from scattered data, which arises
by combining local interpolation polynomials of a fixed degree with inverse distance weighted basis functions [19].
Let 𝛺 ⊂ R𝑑 , 𝑑 ≥ 2, be a region (i.e. a non-empty connected open set), 𝜕𝛺 its boundary, 𝛯 = {𝝃 𝑖 }𝑛𝑖=1 ⊂ 𝛺 ∪ 𝜕𝛺 a finite set of
( ) ( )
pairwise distinct scattered nodes and 𝑓 = {𝑓𝑖 }𝑛𝑖=1 the associated function values. Let 𝑟 ∈ N and 𝑚 = 𝑟+𝑑 𝑑
= dim 𝑟 (R𝑑 ) , where
𝑑
𝑟 (R ) denotes the space of polynomials of 𝑑 variables of total degree ≤ 𝑟.
We assume that a set {𝜎𝑗 }𝑠𝑗=1 is given, such that for each 𝑗 = 1, … , 𝑠, 𝜎𝑗 = {𝝃 𝑗𝜄 }𝑚
𝜄=1
⊂ 𝛯 is unisolvent for the polynomial space
𝑟 (R𝑑 ) and
⋃
𝑠
𝜎𝑗 = 𝛯 (1)
𝑗=1

(to shorten the notation, for each 𝑗 = 1, … , 𝑠, we are denoting with 𝑗𝜄 = 𝜑𝑗 (𝜄) the image of 𝜄 ∈ {1, … , 𝑚} by an injective map 𝜑𝑗 from
{1, … , 𝑚} into {1, … , 𝑛}).
A convenient way to represent the unique polynomial 𝑃𝑗 ∈ 𝑟 (R𝑑 ), 𝑗 = 1, … , 𝑠 interpolating on 𝜎𝑗 = {𝝃 𝑗1 , … , 𝝃 𝑗𝑚 } the data
{𝑓𝑗1 , … , 𝑓𝑗𝑚 } is as follows
∑
𝑚
𝑃𝑗 (𝒙) = 𝓁𝑗,𝜄 (𝒙)𝑓𝑗𝜄 , 𝒙 ∈ R𝑑 ,
𝜄=1

where
∑ ( )𝛼
𝓁𝑗,𝜄 (𝒙) = 𝑎(𝑗,𝜄)
𝛼 𝒙 − 𝝃 (𝑏)
𝑗 ,
|𝛼|≤𝑟

are the Lagrange fundamental polynomials written in the Taylor basis centered at the barycenter 𝝃 (𝑏) 𝑚
𝑗 of 𝜎𝑗 and 𝛼 ∈ N ∪ {(0, … , 0)}
is a multi-index (for more details see [21]). It is a well-known fact that
( )
𝓁𝑗,𝜄 𝝃 𝑗𝜅 = 𝛿𝜄𝜅 , 𝑗 = 1, … , 𝑠, 𝜄, 𝜅 = 1, … , 𝑚. (2)
{ }𝑠
The existence of a covering 𝜎𝑗 𝑗=1 of 𝛯 is almost surely guaranteed [22]. In [19] we detailed a procedure to determine such a
covering which consists of considering, for each scattered point 𝝃 𝑖 , the set of 𝑚 + 𝑞, 𝑞 > 0, nearest points and in choosing, among
them, the subset of 𝑚 discrete Leja points computed by the algorithm presented in [23]. In terms of computational cost { and
}𝑠 accuracy
of the results, such a procedure can be much improved by minimizing the number 𝑠 of elements of the covering 𝜎𝑗 𝑗=1 . In the
numerical experiments, we make use of the improved procedure.
The multinode inverse distance weighted basis functions based on the covering {𝜎𝑗 }𝑠𝑗=1 are defined as follows [24]
∏𝑚 ‖ ‖−𝜇
𝜄=1 ‖
‖
𝒙 − 𝝃 𝑗𝜄 ‖
‖
𝑊𝜇,𝑗 (𝒙) = ∑ ∏ , 𝑗 = 1, … , 𝑠,
𝑠 𝑚 ‖ ‖−𝜇
𝑙=1 𝜆=1 ‖
‖
𝒙 − 𝝃 𝑙𝜆 ‖
‖
where 𝜇 > 0 is fixed. They are a partition of unity
∑
𝑠
𝑊𝜇,𝑗 (𝒙) = 1, 𝒙 ∈ R𝑑 , (3)
𝑗=1

and satisfy the following interpolation properties

( ) ∑ ( )
𝑊𝜇,𝑗 𝝃 𝑖 = 0 for all 𝝃 𝑖 ∉ 𝜎𝑗 , 𝑊𝜇,𝑗 𝝃 𝑖 = 1, (4)
𝑗∈𝑖

2
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

where we set

𝑖 = {𝑗 ∈ {1, … , 𝑠} ∶ 𝝃 𝑖 ∈ 𝜎𝑗 }, 𝑖 = 1, … , 𝑛.

In addition, if 𝜇 > 2 the multinode basis functions 𝑊𝜇,𝑗 (𝒙) satisfy the following differential properties
( ) ∑ ( )
∇𝑊𝜇,𝑗 𝝃 𝑖 = 𝟎 for all 𝝃 𝑖 ∉ 𝜎𝑗 , ∇𝑊𝜇,𝑗 𝝃 𝑖 = 𝟎, (5)
𝑗∈𝑖

and
( ) ∑ ( )
𝐻𝑊𝜇,𝑗 𝝃 𝑖 = 𝟎 for all 𝝃 𝑖 ∉ 𝜎𝑗 , 𝐻𝑊𝜇,𝑗 𝝃 𝑖 = 𝟎, (6)
𝑗∈𝑖

where, as usual, ∇𝑊𝜇,𝑗 (𝒙) and 𝐻𝑊𝜇,𝑗 (𝒙) denote the gradient and the hessian matrix of 𝑊𝜇,𝑗 (𝒙), respectively. Finally, they are
rational functions without singularities if 𝜇 is an even integer (for more details, see [24]).
The MS operator is a blend of local interpolation polynomials realized by using multinode basis functions as follows
∑
𝑠
 𝜇 [𝑓 ] (𝒙) = 𝑊𝜇,𝑗 (𝒙) 𝑃𝑗 (𝒙) .
𝑗=1

Since the property (3)  𝜇 [⋅] reproduces polynomials of 𝑑 variables of total degree ≤ 𝑟, while (4) imply that  𝜇 [𝑓 ] interpolates
data 𝑓𝑖 at 𝝃 𝑖 , 𝑖 = 1, … , 𝑛. Moreover, by taking into account that for each 𝑗 ∈ 𝑖 there exists a unique 𝜄 ∈ {1, … , 𝑚} such that 𝑗𝜄 = 𝑖,
with a little abuse of notation we denote the polynomial 𝓁𝑗,𝜄 by 𝓁𝑗,𝑖 and rewrite the MS operator as follows

∑
𝑛 ∑ ∑
𝑛
 𝜇 [𝑓 ] (𝒙) = 𝑊𝜇,𝑗 (𝒙) 𝓁𝑗,𝑖 (𝒙) 𝑓𝑖 = 𝐵𝜇,𝑖 (𝒙) 𝑓𝑖 (7)
𝑖=1 𝑗∈𝑖 𝑖=1

where, for each 𝑖 = 1, … , 𝑛 we set

∑
𝐵𝜇,𝑖 (𝒙) = 𝑊𝜇,𝑗 (𝒙) 𝓁𝑗,𝑖 (𝒙) . (8)
𝑗∈𝑖
{ }
Functions 𝐵𝜇,𝑖 (𝒙) are linearly independent on the set 𝝃 1 , … , 𝝃 𝑛 since they satisfy the Kronecker delta property

𝐵𝜇,𝑖 (𝝃 𝑘 ) = 𝛿𝑘𝑖 , 𝑘, 𝑖 = 1, … , 𝑛. (9)

In the following, it is useful to split the sum over 𝑖 in (7) in two parts, the first one related to the interior points of the integration
domain 𝛺 and the second one related to the boundary points of 𝛺. Moreover, without loss of generality, we can assume that the
nodes in 𝜩 are ordered in such a way that 𝝃 1 , … , 𝝃 𝑛𝐼 are the interior points and 𝝃 𝑛𝐼 +1 , … , 𝝃 𝑛 are the boundary points. Therefore we
have
𝑛𝐼
∑ ∑
𝑛
 𝜇 [𝑓 ] (𝒙) = 𝐵𝜇,𝑖 (𝒙) 𝑓𝑖 + 𝐵𝜇,𝑖 (𝒙) 𝑓𝑖 . (10)
𝑖=1 𝑖=𝑛𝐼 +1

3. Poisson problem with Dirichlet boundary conditions

In this section, we focus on the Poisson problem with continuous (or piecewise continuous) Dirichlet boundary conditions
{
𝛥𝑢 (𝒙) = 𝑓 (𝒙) , 𝒙 ∈ 𝛺,
(11)
𝑢 (𝒙) = 𝑔 (𝒙) , 𝒙 ∈ 𝜕𝛺,
and we assume, for 𝜇 > 2 even integer, that its approximate solution 𝑢̃ is represented by (10), that is
𝑛𝐼
∑ ∑
𝑛
𝑢̃ (𝒙) = 𝐵𝜇,𝑖 (𝒙) 𝑢̃ 𝑖 + 𝐵𝜇,𝑖 (𝒙) 𝑢̃ 𝑖 , (12)
𝑖=1 𝑖=𝑛𝐼 +1

where 𝑢̃ 𝑖 = 𝑢(𝝃
̃ 𝑖 ), 𝑖 = 1, … , 𝑛, are the unknown coefficients. To determine these coefficients, we impose differential and boundary
conditions to the approximate solution 𝑢̃ on the interior nodes 𝝃 1 , … , 𝝃 𝑛𝐼 and boundary nodes 𝝃 𝑛𝐼 +1 , … , 𝝃 𝑛 , respectively. that is
{ ( )
̃ 𝑘)
𝛥𝑢(𝝃 = 𝑓 𝝃 𝑘 , 𝑘 = 1, … , 𝑛𝐼 ,
( ) ( ) (13)
𝑢̃ 𝝃 𝑘 = 𝑔 𝝃 𝑘 , 𝑘 = 𝑛𝐼 + 1, … , 𝑛.

By taking into account the representation of the candidate solution (12) we can write (13) in matrix form as follows

𝐴̃𝐮 = 𝐲 (14)

where we set 𝐴 = [𝐴𝑘𝑖 ], 𝑘, 𝑖 = 1, … , 𝑛, 𝐮̃ = [𝑢̃ 1 , … , 𝑢̃ 𝑛 ]𝑇 , 𝐲 = [𝑦1 , … , 𝑦𝑛 ]𝑇 , with 𝑦𝑘 = 𝑓 (𝝃 𝑘 ), 𝑘 = 1, … , 𝑛𝐼 and 𝑦𝑘 = 𝑔(𝝃 𝑘 ), 𝑘 = 𝑛𝐼 + 1, … , 𝑛.

3
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

By the setting (8), by the properties (4)–(6) satisfied by the basis functions 𝑊𝜇,𝑗 and the Kronecker delta property (2) of the
fundamental Lagrange polynomials 𝓁𝑗,𝑖 , simple algebraic computations show that the entries of the collocation matrix are

𝐴𝑘𝑖 = 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘

∑ ( )|
= 𝛥 𝑊𝜇,𝑗 (𝒙) 𝓁𝑗,𝑖 (𝒙) |
|𝒙=𝝃 𝑘
∑𝑖 (
𝑗∈
( ) ( ) ( ) ( ))
= 𝑊𝜇,𝑗 𝝃 𝑘 𝛥𝓁𝑗,𝑖 𝝃 𝑘 + 2∇𝑊𝜇,𝑗 𝝃 𝑘 ⋅ ∇𝓁𝑗,𝑖 𝝃 𝑘 , 𝑘 = 1, … , 𝑛𝐼 , 𝑖 = 1, … , 𝑛,
𝑗∈𝑖

𝐴𝑘𝑖 = 0, 𝑘 = 𝑛𝐼 + 1, … , 𝑛 𝑖 = 1, … , 𝑛𝐼

and finally

𝐴𝑘𝑖 = 𝛿𝑘𝑖 , 𝑘, 𝑖 = 𝑛𝐼 + 1, … , 𝑛.

Consequently

𝑢̃ 𝑘 = 𝑔(𝝃 𝑘 ), 𝑘 = 𝑛𝐼 + 1, … , 𝑛

and the linear system (14) reduces to order 𝑛𝐼 with

𝐴𝑘𝑖 = 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘 , 𝑘, 𝑖 = 1, … , 𝑛𝐼 , (15)

and
∑
𝑛
𝑦𝑘 = 𝑓 (𝝃 𝑘 ) − 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘 𝑔(𝝃 𝑖 ), 𝑘 = 1, … , 𝑛𝐼 . (16)
𝑖=𝑛𝐼 +1

We emphasize that the { unknown} vector is now related only to the Laplacian of the functions 𝐵𝜇,𝑖 (𝒙), 𝑖 = 1, … , 𝑛𝐼 which are
linearly independent on 𝝃 1 , … , 𝝃 𝑛𝐼 by (9). However, the non singularity of the matrix 𝐴 for general configuration of collocation
{ }
nodes 𝝃 1 , … , 𝝃 𝑛 is still an open problem.
As a consequence of the polynomial reproduction property of the operator  𝜇 , the following result can be stated.

Theorem 3.1. Let 𝐴 be a non-singular matrix and assume that the problem (11) admits a unique polynomial solution 𝑝 ∈ 𝑟 (R𝑑 ). Then
𝑢̃ = 𝑝.

Proof. By the polynomial reproduction property of the operator  𝜇 we have  𝜇 [𝑝] = 𝑝 and then 𝛥 𝜇 [𝑝] = 𝛥𝑝. Being 𝛥𝑓 = 𝛥𝑝
and 𝑔 = 𝑝 in (11) the vector 𝐮̃ = [𝑝(𝝃 1 ), … , 𝑝(𝝃 𝑛 )] is the unique solution of system (14) and the thesis follows since, by (12)
𝑛𝐼
∑ ∑
𝑛
𝑢̃ (𝒙) = 𝐵𝜇,𝑖 (𝒙) 𝑝(𝝃 𝑖 ) + 𝐵𝜇,𝑖 (𝒙) 𝑝(𝝃 𝑖 ) = 𝑝(𝒙). ■
𝑖=1 𝑖=𝑛𝐼 +1

{ }
Remark 3.2. Theorem 3.1 holds in the case of more general BVPs for elliptic PDEs, when a covering 𝐶1 (𝜕𝛺), … , 𝐶𝑡 (𝜕𝛺) of 𝜕𝛺
by its pairwise disjoint connected subsets and differential operators , 1 , … , 𝑡 are given, such that problem

⎧
⎪𝑢 (𝒙) = 𝑓 (𝒙) , 𝒙 ∈ 𝛺,
⎪1 𝑢 (𝒙) = 𝑔1 (𝒙) , 𝒙 ∈ 𝐶1 (𝜕𝛺),
⎨⋮ ⋮
⎪
⎪𝑡 𝑢 (𝒙) = 𝑔𝑡 (𝒙) , 𝒙 ∈ 𝐶𝑡 (𝜕𝛺),
⎩

admits a unique polynomial solution 𝑝 ∈ 𝑟 (R𝑑 ).

In the following, we assume 𝑑 = 2 and set 𝒙 = (𝑥, 𝑦). Moreover, we fix the value of 𝜇 to 4 in order to work with rational
approximants satisfying both properties (5)–(6). In Algorithm 1 we report the pseudocode for the Multinode Shepard Collocation
(MSC) method.
In the numerical experiments, provided in the following sections, we compare the results obtained by using the MSC method
with those obtained by the Kansa method [6] with the following RBFs:
( )1
Hardy’s Multiquadric 𝐶 ∞ (MQ) 𝜑 (𝜌) = 1 + 𝑐 2 𝜌2 2 ;
( )− 1 (17)
Inverse Multiquadric 𝐶 ∞ (IMQ) 𝜑 (𝜌) = 1 + 𝑐 2 𝜌2 2 ;
( )
Matérn 𝐶 6 (M6) 𝜑 (𝜌) = 𝑒−𝑐𝜌 15 + 15𝑐𝜌 + 6𝑐 2 𝜌2 + 𝑐 3 𝜌3 ,

4
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Table 1
Parameter values for the experiment in Fig. 2.
𝑟 𝑚 𝑞 𝑠
3 10 10 16
4 15 1 9
5 21 4 9
6 28 5 7
7 3 8 6
8 45 3 4
9 55 3 2

√
where 𝜌 = ‖𝒙‖ = 𝑥2 + 𝑦2 and 𝑐 > 0 is the shape parameter, which plays an important role in the accuracy of the method and is
set by means of trial and error procedure [25,26]. Moreover, we consider a comparison with the polynomial-augmented RBF Kansa
method [10], by adding a polynomial term to get the same reproduction degree 𝑟 as the MSC method.

Algorithm 1 Algorithm for the MSC method

{ }𝑛 ( )
Require: A set of collocation points 𝛯 = 𝝃 𝑖 𝑖=1 ; 𝑟 degree of the local polynomial interpolants; 𝑚 = 𝑟+2 ; 𝑚 + 𝑞, 𝑞 > 0, number of
{ }𝑠 2
nearest neighbor points for the selection of the covering 𝜎𝑗 𝑗=1
Ensure: Approximation of the solution 𝑢(𝒙)
{ }
Step 1: Determine the minimal set 𝜎𝑗 𝑠𝑗=1 covering 𝛯
Step 2: Compute the entries of the collocation matrix by the formula (15)
Step 3: Compute the entries of the known term by the formula (16)
Step 4: Solve the linear system 𝐴𝒖̃ = 𝒚 to compute the approximation 𝒖̃ 𝑖 of the solution 𝑢(𝒙) at the collocation points 𝝃 𝑖 ,
𝑖 = 1, … , 𝑛𝐼
Step 5: Reconstruct the approximation of 𝑢(𝒙) through the MS approximant in (12)

3.1. Elliptic problems on the unit disk with Dirichlet boundary conditions

{ }
We focus on the elliptic problem (11), with 𝛺 = (𝑥, 𝑦) ∈ R2 ∶ 𝑥2 + 𝑦2 ≤ 1 being the unit disk centered at the origin, and
65
𝑢 (𝑥, 𝑦) = .
65 + (𝑥 − 0.2)2 + (𝑦 + 0.1)2
This problem has been approached in [13], in order to study the conditioning of the collocation matrix, when the shape parameter
𝑐 tends to 0, by means of three slightly different variations of the Kansa pure collocation method. For the aim of the paper [13], a
small number of collocation points is sufficient. In our experiment, we consider the set of 50 Halton points in [0, 1] × [0, 1], moved
into [−1, 1] × [−1, 1] by means of the affine map
( )
1
𝒙↦2 𝒙− , 𝒙 = (𝑥, 𝑦) ∈ R2 ,
2
from which we extract the subset of points that lie in the unit disk 𝛺. As a result, we get 𝑛𝐼 = 41 interior points which we couple with
𝑛𝐵 = 22 uniformly distributed boundary points (see Fig. 1 (left)). We compute the mean absolute error and the condition number of
the collocation matrix obtained by using the Kansa method and the polynomial-augmented Kansa method with RBFs in (17) with
the best shape parameters (𝑐 = 0.11 for the MQ, 𝑐 = 0.21 for the IMQ and 𝑐 = 0.31 for the M6) and the MSC method varying the
degree 𝑟 of the local polynomial interpolants from 3 to 9. We stress that the polynomial term in the polynomial-augmented Kansa
method is chosen to get the same reproduction degree as the MSC method for each 𝑟. The shape parameter 𝑐 used in the simulations
is chosen using a trial and error procedure. The minimum degree 𝑟 = 3 of the polynomial interpolants is necessary to avoid local
interpolants with constant Laplacian. In its turn, the maximum degree 𝑟 = 9 is determined by the total number of collocation points,
which does not allow the use of local polynomial interpolants of greater degree. The pointwise errors are computed at a set of 801
points distributed on concentric circles of radii 𝑖∕8, 𝑖 = 0, … , 8 (see Fig. 1 (right)). The numerical results are displayed in Fig. 2.
{ }𝑠
For each value of 𝑟, the covering 𝜎𝑗 𝑗=1 satisfying condition (1) is realized by minimizing the number 𝑠 of subsets. The values
of 𝑞 and 𝑠 are reported in Table 1. As one expects, by increasing the degree of the local polynomial interpolants, and therefore
the polynomial reproduction degree of the approximate solution (12), the accuracy of the MSC method increases. In particular,
the numerical results, summarized in Fig. 2 and in Fig. 3, show an improvement in the accuracy of the approximation of the MSC
method from a value of 2.63𝑒−5 for 𝑟 = 3 up to a value of 7.44𝑒−12 for 𝑟 = 9 and a low condition number which does not vary as
the polynomial degree increases.
Finally, in Fig. 4 we plot the error on the surface of the approximate solutions obtained by the MSC method with 𝑟 = 9, the
Kansa M6 method, and the polynomial-augmented Kansa IMQ method with 𝑟 = 9 by using 𝑛 = 63 collocation points.

5
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 1. Point distributions for the example in Section 3.1. Left: the set of 𝑛𝐼 = 41 Halton interior points in red with 𝑛𝐵 = 22 uniformly distributed boundary
points in blue. Right: the set of 801 evaluation points arranged on concentric circles of radii 𝑖∕8, 𝑖 = 0, … , 8.

Fig. 2. Semilog plots of the mean absolute error for the example presented in Section 3.1. The MSC method is with 𝑟 from 3 to 9. The same values of 𝑟 are for
the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 0.11 for the MQ, 𝑐 = 0.21 for the IMQ and 𝑐 = 0.31 for the M6.

Fig. 3. Semilog plots of the condition number of the collocation matrices for the example presented in Section 3.1. The MSC method is with 𝑟 from 3 to 9. The
same values of 𝑟 are for the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 0.11 for the MQ, 𝑐 = 0.21 for the IMQ
and 𝑐 = 0.31 for the M6.

6
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 4. Error on the surface of the approximated solution for the example presented in Section 3.1. The MSC method, on the top left, is with 𝑟 = 9. The Kansa
M6 method is on the top right and the polynomial-augmented Kansa IMQ method, on the bottom, is with 𝑟 = 9. The simulations use 𝑛𝐼 = 41 Halton interior
points and 𝑛𝐵 = 22 uniformly distributed boundary points.

3.2. Elliptic problem on L-shaped domain with Dirichlet boundary conditions

Let us consider the L-shaped domain

𝛺̄ = [0, 0.5] × [0, 1] ∪ [0.5, 1] × [0, 0.5]

and the function [27]

( 𝜋𝑦 )
𝑢(𝑥, 𝑦) = sin(𝜋𝑥) cos
2
̄ of the corresponding Poisson problem with Dirichlet boundary conditions (11). To test how the MSC method performs
solution, on 𝛺,
in numerically solving this problem, we consider the set of 500 Halton points in [0, 1] × [0, 1] from which we extract the maximal
subset of points that lie in the L-shaped domain 𝛺. As a result, we get a set of 𝑛𝐼 = 376 interior points which we couple with 𝑛𝐵 = 82
uniformly distributed boundary points, respectively (see Fig. 5). Moreover, the local polynomial interpolant degree 𝑟 varies from
{ }𝑠
4 to 11. For each value of 𝑟, the covering 𝜎𝑗 𝑗=1 satisfying condition (1) is realized by minimizing the number 𝑠 of subsets. The
values of 𝑞 and 𝑠 are reported in Table 2. The Kansa method and the polynomial-augmented Kansa method adopt the best shape
parameters for the employed RBFs (𝑐 = 1.9 for the MQ, 𝑐 = 1.8 for the IMQ and 𝑐 = 0.8 for the M6). The shape parameters 𝑐 used
in the simulations are chosen using a trial and error procedure. We stress that the polynomial term in the polynomial-augmented
Kansa method is chosen to get the same polynomial reproduction degree as the MSC method for each 𝑟. The pointwise errors are
evaluated at a set of 7500 points distributed on a regular grid on the L-shaped domain and the mean absolute error and the condition
number of the collocation matrices are computed for overall methods. The numerical results are displayed in Fig. 6. In Fig. 7 we
represent the error on the surface of the approximate solution for the case 𝑟 = 11. Finally, in Fig. 8 we display the semilog plot of
the condition number of the various collocation matrices by varying the polynomial reproduction degree 𝑟.

7
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 5. L-shaped domain, for the example in Section 3.2, with the set of interior Halton points in red and uniformly distributed boundary points in blue.

Table 2
Parameter values for the experiment in Fig. 6.
𝑟 𝑚 𝑞 𝑠
4 15 10 87
5 21 14 70
6 28 9 52
7 36 13 44
8 45 4 32
9 55 9 29
10 66 15 29
11 78 18 24

Fig. 6. Semilog plots of the mean absolute error for the example presented in Section 3.2. The MSC method is with 𝑟 from 4 to 11. The same values of 𝑟 are for
the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 1.8 for the MQ, 𝑐 = 1.9 for the IMQ and 𝑐 = 1.4 for the M6.

4. Seepage flow in a heterogeneous and anisotropic medium

In hydrology, seepage flow refers to the flow of a fluid (water) in permeable soil layers such as sand. From the mathematical point
of view, such a problem is formulated by a PDE problem based on an extension of Darcy’s law [28]. More precisely, the mathematical

8
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 7. Error on the surface of the approximated solution for the example presented in Section 3.2. The MSC method, on the top left, is with 𝑟 = 11. The Kansa
IMQ method with 𝑐 = 1.8 is on the top right and the polynomial-augmented Kansa IMQ method, on the bottom, is with 𝑐 = 1.8 and 𝑟 = 11. The simulations use
𝑛𝐼 = 376 interior points and 𝑛𝐵 = 82 uniformly distributed boundary points.

Fig. 8. Semilog plots of the condition number of the collocation matrices for the example presented in Section 3.2. The MSC method is with 𝑟 from 4 to 11.
The same values of 𝑟 are for the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 1.8 for both RBF methods.

9
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

model for the seepage flow in a heterogeneous and anisotropic medium corresponds to an elliptic operator with variable coefficients
coupled with Dirichlet boundary conditions
⎧ 𝜕 [ ] [ ]
𝜕𝑢 𝜕 𝜕𝑢
⎪ 𝑎 (𝑥, 𝑦) + 𝑏 (𝑥, 𝑦) = 𝑓 (𝑥, 𝑦) , (𝑥, 𝑦) ∈ 𝛺,
⎨ 𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 (18)
⎪𝑢 (𝑥, 𝑦) = 𝑔 (𝑥, 𝑦) , (𝑥, 𝑦) ∈ 𝜕𝛺,
⎩
where the derivatives
𝜕𝑢 𝜕𝑢
𝑣(𝑥, 𝑦) = −𝑎(𝑥, 𝑦) , 𝑤(𝑥, 𝑦) = −𝑏(𝑥, 𝑦) ,
𝜕𝑥 𝜕𝑦
represent the Darcy’s velocity [27]. If 𝑎 (𝑥, 𝑦) = 𝑏 (𝑥, 𝑦) = 𝑘, 𝑘 ∈ R∖ {0}, the problem (18) becomes a Poisson problem with Dirichlet
boundary conditions. In applying the MSC method to approximate the solution 𝑢 of the problem (18), as before we distinguish
between interior and boundary collocation points. By the setting (8), by the properties (4)–(6) satisfied by the basis functions 𝑊𝜇,𝑗
and the Kronecker delta property (2) of the fundamental Lagrange polynomials 𝓁𝑗,𝑖 , simple algebraic computations show that the
entries of the collocation matrix are
( [ ] [ ])|
𝜕 𝜕𝐵𝜇,𝑖 𝜕 𝜕𝐵𝜇,𝑖 |
𝐴𝑘𝑖 = 𝑎 (𝒙) (𝒙) + 𝑏 (𝒙) (𝒙) | 𝑘 = 1, … , 𝑛𝐼 , 𝑖 = 1, … , 𝑛,
𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 |
|𝒙=𝝃 𝑘
( )
𝜕𝑎 ( ) ∑ ( ) 𝜕𝓁𝑗,𝑖 ( ) ∑ 𝜕𝑊𝜇,𝑗 ( ) 𝜕𝓁𝑗,𝑖 ( ) ( ) 𝜕 2 𝓁𝑗,𝑖 ( )
= 𝝃 𝑊 𝝃 𝝃 𝑘 + 𝑎(𝝃 𝑘 ) 𝝃𝑘 𝝃 𝑘 + 𝑊𝜇,𝑗 𝝃 𝑘 𝝃𝑘
𝜕𝑥 𝑘 𝑗∈ 𝜇,𝑗 𝑘 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥2
𝑖 𝑗∈𝑖 ( )
𝜕𝑏 ∑ ( ) 𝜕𝓁𝑗,𝑖 ( ) ∑ 𝜕𝑊𝜇,𝑗 ( ) 𝜕𝓁𝑗,𝑖 ( ) ( ) 𝜕 2 𝓁𝑗,𝑖 ( )
+ (𝝃 𝑘 ) 𝑊𝜇,𝑗 𝝃 𝑘 𝝃 𝑘 + 𝑏(𝝃 𝑘 ) 𝝃𝑘 𝝃 𝑘 + 𝑊𝜇,𝑗 𝝃 𝑘 𝝃𝑘 ,
𝜕𝑦 𝑗∈
𝜕𝑦 𝑗∈
𝜕𝑦 𝜕𝑦 𝜕𝑦2
𝑖 𝑖

𝐴𝑘𝑖 = 0, 𝑘 = 𝑛𝐼 + 1, … , 𝑛, 𝑖 = 1, … , 𝑛𝐼
and finally
𝐴𝑘𝑖 = 𝛿𝑘𝑖 , 𝑘, 𝑖 = 𝑛𝐼 + 1, … , 𝑛.
Consequently
𝑢̃ 𝑘 = 𝑔(𝝃 𝑘 ), 𝑘 = 𝑛𝐼 + 1, … , 𝑛
and the collocation linear system reduces to order 𝑛𝐼 with
( [ ] [ ])|
𝜕 𝜕𝐵𝜇,𝑖 𝜕 𝜕𝐵𝜇,𝑖 |
𝐴𝑘𝑖 = 𝑎 (𝒙) (𝒙) + 𝑏 (𝒙) (𝒙) | , 𝑘, 𝑖 = 1, … , 𝑛𝐼 ,
𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 |
|𝒙=𝝃 𝑘
and
( [ ] [ ])|
∑
𝑛
𝜕 𝜕𝐵𝜇,𝑖 𝜕 𝜕𝐵𝜇,𝑖 |
𝑦𝑘 = 𝑓 (𝝃 𝑘 ) − 𝑎 (𝒙) (𝒙) + 𝑏 (𝒙) (𝒙) | 𝑔(𝝃 𝑖 ), 𝑘 = 1, … , 𝑛𝐼 .
𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 |
𝑖=𝑛𝐼 +1 |𝒙=𝝃 𝑘

4.1. A case study

In line with [27], we set

𝑎 (𝑥, 𝑦) = 2 − 𝑥2 − 𝑦2 , 𝑏 (𝑥, 𝑦) = 𝑒𝑥−𝑦 ,
and, on 𝛺̄ = [0, 1] × [0, 1], we consider the function
𝑢 (𝑥, 𝑦) = 𝑥𝑦(1 − 𝑥)(1 − 𝑦),
solution of the corresponding seepage flow problem (18) with
𝑓 (𝑥, 𝑦) = −𝑥𝑒𝑥−𝑦 (1 − 𝑥)(3 − 2𝑦) + 2𝑦(1 − 𝑦)(3𝑥2 + 𝑦2 − 𝑥 − 2).
In line with [27], we conduct the experiments by using 𝑛𝐼 = 113 interior points and 𝑛𝐵 = 32 boundary points. In Fig. 9 we display
the error on the surface of the approximated solution obtained by the MSC method with local interpolation polynomial of degree
𝑟 = 4 and those obtained by using the Kansa MQ method with shape parameter 𝑐 = 1.2 and the polynomial-augmented Kansa MQ
method with 𝑟 = { 4 and
} shape parameter 𝑐 = 1.2. The error is computed on a regular grid of 100 × 100 points on the solution domain
̄ The covering 𝜎𝑗 𝑠 satisfying condition (1) is realized by minimizing the number 𝑠 of subsets. By setting 𝑚 = 15 and 𝑞 = 6 we
𝛺. 𝑗=1
get 𝑠 = 30 subsets. Since the solution is a polynomial of degree 3, from Theorem 3.1 we expect 𝑢̃ = 𝑢 and the obtained numerical
results related to the absolute error of 6.03𝑒−16 for the MSC method confirm the thesis of the Theorem (see Fig. 9). By applying
the Kansa MQ method we get a maximum absolute error of 2.24𝑒−5 while the polynomial-augmented Kansa MQ method gives a
maximum absolute error of 9.07𝑒−16.
Moreover, the condition number of the MSC method is 1.60𝑒+3 against a condition number of 8.12𝑒+13 for the Kansa MQ method
and of 7.80𝑒+13 for the polynomial-augmented Kansa MQ.

10
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 9. Error on the surface of the approximated solution for the example presented in Section 4.1. The MSC method, on the top left, is with 𝑟 = 4. The Kansa
MQ method with 𝑐 = 1.2 is on the top right and the polynomial-augmented Kansa MQ method, on the bottom, is with 𝑐 = 1.2 and 𝑟 = 4. The simulations use
𝑛𝐼 = 113 interior points and 𝑛𝐵 = 32 uniformly distributed boundary points.

5. Poisson problem with mixed boundary conditions

In this Section, we consider the Poisson problem with mixed boundary conditions on different shaped domains. In particular, we
analyze the case in which the boundary conditions are constituted by both Dirichlet and Neumann conditions. Due to the particular
nature of such problems, the entries of the collocation matrix depend on the Neumann conditions of the specific problem and
therefore we cannot give a general representation for it as we do for the case of Dirichlet boundary conditions.

5.1. Poisson problem with mixed boundary conditions on the unit square domain

Let 𝛺̄ be the unit square [0, 1] × [0, 1]. In line with [5], we consider the Poisson problem

𝛥𝑢(𝑥, 𝑦) = −5.4𝑥 in 𝛺 (19)

with the mixed Dirichlet–Neumann conditions

⎧ 𝜕 𝑢(𝑥, 𝑦) = 0, 𝑦 = 0, 0 ≤ 𝑥 ≤ 1, 𝛤1 ,
⎪ 𝜕𝑦
⎪𝑢(𝑥, 𝑦) = 0.1, 𝑥 = 1, 0 ≤ 𝑦 ≤ 1, 𝛤2 ,
⎨ 𝜕 (20)
⎪− 𝑢(𝑥, 𝑦) = 0, 𝑦 = 1, 0≤𝑥<1 𝛤3 ,
⎪ 𝜕𝑦
⎩𝑢(𝑥, 𝑦) = 1, 𝑥 = 0, 0 ≤ 𝑦 ≤ 1, 𝛤4 ,
whose exact solution is the cubic polynomial

𝑢 (𝑥, 𝑦) = 1 − 0.9𝑥3 .

11
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

In Fig. 10 we display the geometry domain of this problem. In applying the MSC method to approximate the solution 𝑢 of the
problem (19) with boundary conditions (20), we have to distinguish between interior and boundary collocation points. Without loss
of generality, we assume that the collocation points are ordered as follows:
𝝃 𝑘 ∈ 𝛺, 𝑘 = 1, … , 𝑛𝐼 ;
𝝃 𝑘 ∈ 𝛤1 ∪ 𝛤3 , 𝑘 = 𝑛𝐼 + 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 ;
𝝃 𝑘 ∈ 𝛤2 ∪ 𝛤4 , 𝑘 = 𝑛𝐼 + 𝑛𝐵𝑁 + 1, … , 𝑛;
where 𝑛𝐵𝑁 is the number of collocation points on 𝛤1 ∪𝛤3 . By the setting (8), by the properties (4)–(6) satisfied by the basis functions
𝑊𝜇,𝑗 and the Kronecker delta property (2) of the fundamental Lagrange polynomials 𝓁𝑗,𝑖 , simple algebraic computations show that
the entries of the collocation matrix are
𝐴𝑘𝑖 = 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘
∑ ( )|
= 𝛥 𝑊𝜇,𝑗 (𝒙) 𝓁𝑗,𝑖 (𝒙) |
|𝒙=𝝃 𝑘
∑𝑖 (
𝑗∈
( ) ( ) ( ) ( ))
= 𝑊𝜇,𝑗 𝝃 𝑘 𝛥𝓁𝑗,𝑖 𝝃 𝑘 + 2∇𝑊𝜇,𝑗 𝝃 𝑘 ⋅ ∇𝓁𝑗,𝑖 𝝃 𝑘 , 𝑘 = 1, … , 𝑛𝐼 , 𝑖 = 1, … , 𝑛,
𝑗∈𝑖

𝜕𝐵𝜇,𝑖 ( )
𝐴𝑘𝑖 = 𝝃𝑘
𝜕𝑦
∑ ( ) 𝜕𝓁𝑗,𝑖 ( )
= 𝑊𝜇,𝑗 𝝃 𝑘 𝝃𝑘 , 𝑘 = 𝑛𝐼 + 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 , 𝑖 = 1, … , 𝑛,
𝑗∈
𝜕𝑦
𝑖

𝐴𝑘𝑖 = 0, 𝑘 = 𝑛𝐼 + 𝑛𝐵𝑁 + 1, … , 𝑛, 𝑖 = 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 + 1,

and finally

𝐴𝑘𝑖 = 𝛿𝑘𝑖 , 𝑘, 𝑖 = 𝑛𝐼 + 𝑛𝐵𝑁 + 2, … , 𝑛.

By taking into account the Dirichlet boundary conditions (20) on 𝛤2 and 𝛤3 , we get

𝑢̃ 𝑘 = 𝑢(𝝃 𝑘 ), 𝑘 = 𝑛𝐼 + 𝑛𝐵𝑁 + 1, … , 𝑛,

and the collocation linear system reduces to order 𝑛𝐼 + 𝑛𝐵𝑁 with

𝐴𝑘𝑖 = 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘 , 𝑘 = 1, … , 𝑛𝐼 , 𝑖 = 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 ,
𝜕𝐵𝜇,𝑖 ( )
𝐴𝑘𝑖 = 𝝃𝑘 , 𝑘 = 𝑛𝐼 + 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 , 𝑖 = 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 ,
𝜕𝑦
and
∑
𝑛
𝑦𝑘 = −5.4𝝃 𝑥𝑘 − 𝛥𝐵𝜇,𝑖 (𝒙) |𝒙=𝝃 𝑘 𝑢(𝝃 𝑖 ), 𝑘 = 1, … , 𝑛𝐼 ,
𝑖=𝑛𝐼 +𝑛𝐵𝑁 +1
∑
𝑛
𝜕𝐵𝜇,𝑖 ( )
𝑦𝑘 = − 𝝃 𝑘 𝑢(𝝃 𝑖 ) 𝑘 = 𝑛𝐼 + 1, … , 𝑛𝐼 + 𝑛𝐵𝑁 ,
𝑖=𝑛𝐼 +𝑛𝐵𝑁 +1
𝜕𝑦

where we assume 𝝃 𝑘 = (𝝃 𝑥𝑘 , 𝝃 𝑦𝑘 ). In line with [5], we conduct the experiments by using 𝑛𝐼 = 289 Halton points as interior nodes
and 𝑛𝐵 = 64 uniformly distributed boundary points. In Table 3 we report the maximum absolute error 𝑒𝑚𝑎𝑥 , the mean absolute
error 𝑒𝑚𝑒𝑎𝑛 and the root mean square error 𝑒𝑅𝑀𝑆 evaluated at a grid of 40 × 40 points in [0, 1] × [0, 1] obtained by the MSC method
with local interpolation polynomial of degree 𝑟 = 4, by the Kansa MQ method with 𝑐 = 1.1, the Kansa IMQ and M6 methods with
{ }𝑠
𝑐 = 1.2 and by the polynomial-augmented Kansa method with the same shape parameter and 𝑟 = 4. The covering 𝜎𝑗 𝑗=1 satisfying
conditions (1) is realized by minimizing the number 𝑠 of subsets. By setting 𝑚 = 15 and 𝑞 = 19 we get 𝑠 = 69 subsets. Since the
solution is a polynomial of degree 3, from Theorem 3.1 we expect 𝑢̃ = 𝑢 and the obtained numerical results related to the maximum
absolute error of 9.57𝑒−14 for the MSC method confirm the thesis of the Theorem (see Fig. 11). In Table 4 we report the condition
number of the methods adopted in the simulations. We note that the MSC method gives rise to a collocation matrix with the lowest
condition number.

5.2. Poisson problem with mixed boundary conditions on an elliptical domain

Let us consider the elliptical domain

{ }
𝑥2 𝑦2
𝛺̄ = (𝑥, 𝑦) ∈ R2 ∶ + ≤1
𝑎2 𝑏2
with major and minor semiaxes 𝑎 and 𝑏 and, on it, the function [29,30]
( 2 2 )( 2 )
𝑎 𝑏 𝑥 𝑦2
𝑢(𝑥, 𝑦) = − + −1 ,
2
𝑎 +𝑏 2 𝑎 2 𝑏 2

12
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 10. The geometry domain of the problem in Section 5.1.

Table 3
Maximum absolute error 𝑒𝑚𝑎𝑥 , mean absolute error 𝑒𝑚𝑒𝑎𝑛 and root mean square error 𝑒𝑅𝑀𝑆 for
the example presented in Section 5.1. The MSC method and the polynomial-augmented Kansa
methods (pol) are with 𝑟 = 4, 𝑐 = 1.2 for the Kansa’s IMQ, M6, polynomial-augmented IMQ and
M6 methods, while 𝑐 = 1.1 for the Kansa’s MQ and polynomial-augmented MQ methods.
𝑒𝑚𝑎𝑥 𝑒𝑚𝑒𝑎𝑛 𝑒𝑅𝑀𝑆
MSC 9.57𝑒−14 2.70𝑒−14 3.53𝑒−14
Kansa MQ 3.19𝑒−06 6.33𝑒−07 7.91𝑒−07
Kansa IMQ 2.51𝑒−06 4.60𝑒−07 5.96𝑒−07
Kansa M6 6.57𝑒−05 2.31𝑒−06 5.27𝑒−06
Kansa MQ (pol) 9.08𝑒−11 2.68𝑒−11 3.23𝑒−11
Kansa IMQ (pol) 1.84𝑒−09 5.98𝑒−10 7.10𝑒−10
Kansa M6 (pol) 3.53𝑒−13 7.53𝑒−14 1.06𝑒−13

Table 4
Condition number for the example presented in Section 5.1. The MSC method and the
polynomial-augmented Kansa methods (pol) are with 𝑟 = 4, 𝑐 = 1.2 for the Kansa’s IMQ,
M6, polynomial-augmented IMQ and M6 methods, while 𝑐 = 1.1 for the Kansa’s MQ and
polynomial-augmented MQ methods.
𝐶𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 𝑛𝑢𝑚𝑏𝑒𝑟
MSC 6.41𝑒+05
Kansa MQ 1.55𝑒+19
Kansa IMQ 2.28𝑒+19
Kansa M6 1.11𝑒+14
Kansa MQ (pol) 5.16𝑒+21
Kansa IMQ (pol) 1.47𝑒+19
Kansa M6 (pol) 1.09𝑒+14

solution of the Poisson problem with mixed boundary conditions

⎧𝛥𝑢(𝒙) = −2, 𝒙 ∈ 𝛺,
⎪ 𝜕𝑢
⎪ 𝜕𝑥 = 0, 𝑥 = 0,
⎨ 𝜕𝑢 (21)
⎪ 𝜕𝑦 = 0, 𝑦 = 0,
⎪
⎩𝑢(𝒙) = 0, 𝒙 ∈ 𝜕𝛺 ∖ (0, 𝑦) ∪ (𝑥, 0) .
As specified in [31], the problem (21) provides a function 𝑢 from which the angle of twist of a cylindrical shaft of an elliptical
cross-section under torsion can be calculated. Moreover, since the solution of (21) is expected to be symmetric with respect to 𝑥
and 𝑦, it is sufficient to solve it in the first quadrant of the ellipse. To this aim, in the experiments we consider the set of 100 Halton
points in [0, 1] × [0, 1], moved into [0, 𝑎] × [0, 𝑎] by means of the affine map
𝒙 ↦ 𝑎𝒙 𝒙 = (𝑥, 𝑦) ∈ R2 ,
from which we extract the subset of points that lie in the first quadrant of the ellipse of semiaxes 𝑎 and 𝑏. The collocation matrix
can be obtained similarly to the one discussed in Section 5.1. By following [29], in the first experiment we set 𝑎 = 10, 𝑏 = 8 and we
get 𝑛𝐼 = 67 interior points which we couple with 𝑛𝐵 = 44 boundary points (see Fig. 12 (left)). We compute the maximum absolute
error 𝑒𝑚𝑎𝑥 , the mean absolute error 𝑒𝑚𝑒𝑎𝑛 and the root mean square error 𝑒𝑅𝑀𝑆 by evaluating the pointwise errors at a set of 161

13
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 11. Error on the surface of the approximated solution for the example presented in Section 5.1. The MSC method, on the top left, is with 𝑟 = 4. The Kansa
IMQ method with 𝑐 = 1.2 is on the top right and the polynomial-augmented Kansa M6 method, on the bottom, is with 𝑐 = 1.2 and 𝑟 = 4. The simulations use
𝑛𝐼 = 289 interior points and 𝑛𝐵 = 64 uniformly distributed boundary points.

Table 5
Maximum absolute error 𝑒𝑚𝑎𝑥 , mean absolute error 𝑒𝑚𝑒𝑎𝑛 and root mean square error 𝑒𝑅𝑀𝑆 for
the example presented in Section 5.2. The MSC method and the polynomial-augmented Kansa
methods (pol) are with 𝑟 = 3, 𝑐 = 0.1 for the Kansa and polynomial-augmented Kansa methods.
The collocation points lie in the first quadrant of the ellipse of semiaxes 𝑎 = 10 and 𝑏 = 8.
𝑒𝑚𝑎𝑥 𝑒𝑚𝑒𝑎𝑛 𝑒𝑅𝑀𝑆
MSC 1.25𝑒−12 1.59𝑒−13 3.14𝑒−13
Kansa MQ 1.49𝑒−04 3.91𝑒−05 6.83𝑒−05
Kansa IMQ 1.03𝑒−03 3.45𝑒−04 5.49𝑒−04
Kansa M6 3.94𝑒−03 8.57𝑒−04 1.66𝑒−03
Kansa MQ (pol) 6.68𝑒−08 3.28𝑒−08 4.75𝑒−08
Kansa IMQ (pol) 4.56𝑒−10 2.29𝑒−10 3.31𝑒−10
Kansa M6 (pol) 6.31𝑒−09 3.06𝑒−09 4.52𝑒−09

points in the first quadrant distributed on concentric ellipses of semiaxes 𝑎𝑖∕16 and 𝑏𝑖∕16, 𝑖 = 0, … , 16 (see Fig. 12 (right)). The
numerical results, obtained by the MSC method with local interpolation polynomials of degree 𝑟 = 3, the Kansa method with the
RBFs in (17) and the polynomial-augmented Kansa method with 𝑟 = 3, are reported in Table 5. The best shape parameter is 𝑐 = 0.1
{ }𝑠
for all the RBFs. The covering 𝜎𝑗 𝑗=1 satisfying condition (1) is realized by minimizing the number 𝑠 of subsets. By setting 𝑚 = 10
and 𝑞 = 10 we get 𝑠 = 26 subsets. In this case the values for the condition number are reported in Table 6, which shows again the
best conditioning for the MSC method.

Finally, in Fig. 13 we display the error on the surface of the approximated solution obtained by the MSC method, the Kansa MQ
method and the polynomial-augmented Kansa IMQ method. In the second experiment, we set 𝑎 = 2 and 𝑏 = 1 and we get 𝑛𝐼 = 40
interior points which we couple with 𝑛𝐵 = 24 boundary points. The related numerical results are reported in Table 7, where the
best shape parameter for the Kansa MQ is 𝑐 = 0.3, for the Kansa IMQ is 𝑐 = 0.2 and for the Kansa M6 is 𝑐 = 0.1.

14
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 12. Set of 𝑛𝐼 = 67 Halton interior points in red with 𝑛𝐵 = 44 boundary points in blue (left) and set of 161 evaluation points in the first quadrant distributed
on concentric ellipses of semiaxes 10𝑖∕16 and 8𝑖∕16, 𝑖 = 0, … , 16 used for the example presented in Section 5.2.

Table 6
Condition number for the example presented in Section 5.2. The MSC method and the
polynomial-augmented Kansa methods (pol) are with 𝑟 = 4, 𝑐 = 1.2 for the Kansa’s IMQ,
M6, polynomial-augmented IMQ and M6 methods, while 𝑐 = 1.1 for the Kansa’s MQ and
polynomial-augmented MQ methods.
𝐶𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 𝑛𝑢𝑚𝑏𝑒𝑟
MSC 2.16𝑒+04
Kansa MQ 9.34𝑒+17
Kansa IMQ 8.20𝑒+17
Kansa M6 1.55𝑒+13
Kansa MQ (pol) 1.55𝑒+19
Kansa IMQ (pol) 1.35𝑒+20
Kansa M6 (pol) 2.38𝑒+17

Table 7
Maximum absolute error 𝑒𝑚𝑎𝑥 , mean absolute error 𝑒𝑚𝑒𝑎𝑛 and root mean square error 𝑒𝑅𝑀𝑆 for
the example presented in Section 5.2. The MSC method and the polynomial-augmented Kansa
methods (pol) are with 𝑟 = 3, 𝑐 = 0.3 for the Kansa MQ and polynomial-augmented MQ, 𝑐 = 0.2
for the Kansa IMQ and polynomial-augmented IMQ and 𝑐 = 0.1 for the Kansa M6 and polynomial-
augmented M6. The collocation points lie in the first quadrant of the ellipse of semiaxes 𝑎 = 2
and 𝑏 = 1.
𝑒𝑚𝑎𝑥 𝑒𝑚𝑒𝑎𝑛 𝑒𝑅𝑀𝑆
MSC 5.77𝑒−15 2.85𝑒−15 4.15𝑒−15
Kansa MQ 1.90𝑒−05 5.67𝑒−06 9.44𝑒−06
Kansa IMQ 1.53𝑒−04 4.59𝑒−05 7.16𝑒−05
Kansa M6 3.70𝑒−05 1.34𝑒−05 2.07𝑒−05
Kansa MQ (pol) 2.02𝑒−11 1.32𝑒−11 1.74𝑒−11
Kansa IMQ (pol) 7.47𝑒−11 4.85𝑒−11 6.41𝑒−11
Kansa M6 (pol) 2.30𝑒−12 1.51𝑒−12 2.01𝑒−12

5.3. Poisson problem with mixed boundary conditions on the unit square domain with no polynomial exact solution

Let 𝛺̄ = [0, 1] × [0, 1]. In line with [30], we consider the Poisson problem

𝛥𝑢(𝑥, 𝑦) = (𝜆2 + 𝜈 2 )𝑒𝜆𝑥+𝜈𝑦 , (𝑥, 𝑦) ∈ 𝛺,

with the mixed Dirichlet–Neumann boundary conditions

⎧𝑢(𝑥, 𝑦) = 𝑒𝜆𝑥+𝜈𝑦 , 𝑦 = 0, 0 ≤ 𝑥 ≤ 1,
⎪
⎪𝑢(𝑥, 𝑦) = 𝑒𝜆𝑥+𝜈𝑦 , 𝑦 = 1, 0 ≤ 𝑥 ≤ 1,
⎪
⎨ 𝜕𝑢(𝑥, 𝑦) = 𝜆𝑒𝜆𝑥+𝜈𝑦 , 𝑥 = 0, 0 < 𝑦 < 1,
⎪ 𝜕𝑥
⎪ 𝜕𝑢(𝑥, 𝑦)
⎪ = 𝜆𝑒𝜆𝑥+𝜈𝑦 , 𝑥 = 1, 0 < 𝑦 < 1,
⎩ 𝜕𝑥

15
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 13. Error on the surface of the approximated solution for the example presented in Section 5.2. The MSC method, on the top left, is with 𝑟 = 3. The Kansa
MQ method with 𝑐 = 0.1 is on the top right and the polynomial-augmented Kansa IMQ method, on the bottom, is with 𝑐 = 0.1 and 𝑟 = 3. The simulations use
𝑛𝐼 = 67 interior points and 𝑛𝐵 = 44 uniformly distributed boundary points. The collocation points lie in the first quadrant of the ellipse of semiaxes 𝑎 = 10 and
𝑏 = 8.

whose exact solution is

𝑢(𝑥, 𝑦) = 𝑒𝜆𝑥+𝜈𝑦 .

By following [6,30], we set 𝜆 = 2 and 𝜈 = 3. In this case, the collocation matrix is the same as the one discussed in Section 5.1 with
𝜕 𝜕
𝜕𝑥
in place of 𝜕𝑦 . To test how the MSC method performs in numerically solving this problem, we consider the set of 𝑛𝐼 = 500 Halton
points in [0, 1] × [0, 1] coupled with 𝑛𝐵 = 84 uniformly distributed boundary points, respectively. Moreover, the local polynomial
{ }𝑠
interpolant degree 𝑟 varies from 4 to 11. For each value of 𝑟, the covering 𝜎𝑗 𝑗=1 satisfying condition (1) is realized by minimizing
the number 𝑠 of subsets. The values of 𝑞 and 𝑠 are reported in Table 8. The Kansa method and the polynomial-augmented Kansa
method adopt the best shape parameters for the employed RBFs (𝑐 = 2.1 for the MQ, 𝑐 = 1.6 for the IMQ and 𝑐 = 0.4 for the M6).
The shape parameters 𝑐 used in the simulations are chosen using a trial and error procedure. We stress that the polynomial term in
the polynomial-augmented Kansa method is chosen to get the same polynomial reproduction degree as the MSC method for each 𝑟.
The pointwise errors are evaluated at a set of 40 × 40 points distributed on a regular grid on [0, 1] × [0, 1] and the mean absolute
error and the condition number of the collocation matrices are computed for overall methods. The numerical results are displayed
in Fig. 14. In Fig. 16 we represent the error on the surface of the approximated solution for the case 𝑟 = 11. Finally, in Fig. 15 we
display the semilog plot of the condition number of the various collocation matrices by varying the polynomial reproduction degree
𝑟.

6. Conclusion and future perspectives of work

In this paper, we investigated on the multinode Shepard method for solving elliptic boundary value problems equipped with
Dirichlet and/or Neumann boundary conditions on two-dimensional domains of different shapes. The method performs very well
by reaching good approximation accuracy comparable with or better than the one reached by the RBF Kansa methods with or
without polynomial precision. We highlight the particular feature of the MSC method, which gives rise to a collocation matrix with
a low condition number.

16
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Table 8
Parameter values for the experiment in Fig. 14.
𝑟 𝑚 𝑞 𝑠
4 15 12 112
5 21 11 83
6 28 4 58
7 36 10 57
8 45 18 47
9 55 11 43
10 66 5 30
11 78 9 29

Fig. 14. Semilog plots of the mean absolute error for the example presented in Section 5.3. The MSC method is with 𝑟 from 4 to 11. The same values of 𝑟 are
for the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 2.1 for the MQ, 𝑐 = 1.6 for the IMQ and 𝑐 = 0.4 for the
M6.

Fig. 15. Semilog plots of the condition number of the collocation matrices for the example presented in Section 5.3. The MSC method is with 𝑟 from 4 to 11.
The same values of 𝑟 are for the polynomial-augmented Kansa method (pol). The shape parameter for the RBFs in (17) are 𝑐 = 2.1 for the MQ, 𝑐 = 1.6 for the
IMQ and 𝑐 = 0.4 for the M6.

17
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

Fig. 16. Error on the surface of the approximated solution for the example presented in Section 5.3. The MSC method, on the top left, is with 𝑟 = 11. The
Kansa MQ method with 𝑐 = 2.1 is on the top right and the polynomial-augmented Kansa MQ method, on the bottom, is with 𝑐 = 2.1 and 𝑟 = 3. The simulations
use 𝑛𝐼 = 500 interior points and 𝑛𝐵 = 84 uniformly distributed boundary points.

Data availability

No data was used for the research described in the article.

Acknowledgments

This research has been achieved as part of RITA ‘‘Research ITalian network on Approximation’’, as part of the UMI group ‘‘Teoria
dell’Approssimazione e Applicazioni’’ and was supported by INDAM-GNCS project 2023, UNIPA D26-PREMIO-GRUPPI-RIC-2023,
FFR2023. The authors are members of the INdAM research group GNCS.

References

[1] M.J. Gander, F. Kwok, Numerical Analysis of Partial Differential Equations using Maple and MATLAB, Society for Industrial and Applied Mathematics,
Philadelphia, PA, 2018.
[2] A.S. Reimer, A.F. Cheviakov, A matlab-based finite-difference solver for the Poisson problem with mixed Dirichlet–Neumann boundary conditions, Comput.
Phys. Comm. 184 (3) (2013) 783–798.
[3] H. Wendland, Scattered Data Approximation, in: Cambridge Monographs on Applied and Computational Mathematics, Cambridge University Press, 2004.
[4] M.D. Buhmann, Radial Basis Functions: Theory and Implementations, in: Cambridge Monographs on Applied and Computational Mathematics, Cambridge
University Press, 2003.
[5] G.E. Fasshauer, Meshfree Approximation Methods with MATLAB, vol. 6, World Scientific, 2007.
[6] E. Kansa, Multiquadrics—A scattered data approximation scheme with applications to computational fluid-dynamics—II solutions to parabolic, hyperbolic
and elliptic partial differential equations, Comput. Math. Appl. 19 (8) (1990) 147–161.
[7] G.E. Fasshauer, M.J. McCourt, Kernel-Based Approximation Methods Using Matlab, vol. 19, World Scientific Publishing Company, 2015.
[8] E. Francomano, M. Paliaga, Highlighting numerical insights of an efficient SPH method, Appl. Math. Comput. 339 (2018) 899–915.
[9] S.A. Sarra, E.J. Kansa, Multiquadric radial basis function approximation methods for the numerical solution of partial differential equations, Adv. Comput.
Mech. 2 (2) (2009).
[10] B. Fornberg, N. Flyer, Solving PDEs with radial basis functions, Acta Numer. 24 (2015) 215–258.
[11] M.A. Jankowska, A. Karageorghis, C. Chen, Improved Kansa RBF method for the solution of nonlinear boundary value problems, Eng. Anal. Bound. Elem.
87 (2018) 173–183.

18
F. Dell’Accio et al. Journal of Computational and Applied Mathematics 448 (2024) 115896

[12] D. Cao, X. Li, H. Zhu, A polynomial-augmented RBF collocation method using fictitious centres for solving the Cahn–Hilliard equation, Eng. Anal. Bound.
Elem. 137 (2022) 41–55.
[13] E. Larsson, B. Fornberg, A numerical study of some radial basis function based solution methods for elliptic PDEs, Comput. Math. Appl. 46 (5) (2003)
891–902.
[14] A. Karageorghis, M.A. Jankowska, C.-S. Chen, Kansa-RBF algorithms for elliptic problems in regular polygonal domains, Numer. Algorithms 79 (2018)
399–421.
[15] S. Li, L. Ling, Weighted least-squares collocation methods for elliptic PDEs with mixed boundary conditions, Eng. Anal. Bound. Elem. 105 (2019) 146–154.
[16] A. Karageorghis, C. Chen, X.-Y. Liu, Kansa-RBF algorithms for elliptic problems in axisymmetric domains, SIAM J. Sci. Comput. 38 (1) (2016) A435–A470.
[17] S. Reutskiy, A Novel Semi-Analytic Meshless Method for Solving Two- and Three-Dimensional Elliptic Equations of General Form with Variable Coefficients
in Irregular Domains, CMES Comput. Model. Eng. Sci. 99 (4) (2014) 327–349.
[18] E. Kansa, P. Holoborodko, On the ill-conditioned nature of 𝐶 ∞ RBF strong collocation, Eng. Anal. Bound. Elem. 78 (2017) 26–30.
[19] F. Dell’Accio, F. Di Tommaso, O. Nouisser, N. Siar, Solving Poisson equation with Dirichlet conditions through multinode shepard operators, Comput.
Math. Appl. 98 (2021) 254–260.
[20] F. Dell’Accio, F. Di Tommaso, G. Ala, E. Francomano, Electric scalar potential estimations for non-invasive brain activity detection through multinode
Shepard method, in: 2022 IEEE 21st Mediterranean Electrotechnical Conference, MELECON, 2022, pp. 1264–1268.
[21] F. Dell’Accio, F. Di Tommaso, N. Siar, On the numerical computation of bivariate Lagrange polynomials, Appl. Math. Lett. 112 (2021) 106845.
[22] F. Dell’Accio, A. Sommariva, M. Vianello, Random sampling and unisolvent interpolation by almost everywhere analytic functions, Appl. Math. Lett. (2023)
108734.
[23] L. Bos, S. De Marchi, A. Sommariva, M. Vianello, Computing multivariate Fekete and Leja points by numerical linear algebra, SIAM J. Numer. Anal. 48
(5) (2010) 1984–1999.
[24] F. Dell’Accio, F. Di Tommaso, Rate of convergence of multinode Shepard operators, Dolomites Res. Notes Approx. 12 (1) (2019) 1–6.
[25] G.E. Fasshauer, J.G. Zhang, On choosing ‘‘optimal’’ shape parameters for RBF approximation, Numer. Algorithms 45 (2007) 345–368.
[26] R. Cavoretto, A. De Rossi, M.S. Mukhametzhanov, Y.D. Sergeyev, On the search of the shape parameter in radial basis functions using univariate global
optimization methods, J. Global Optim. 79 (2021) 305–327.
[27] J. Li, A.H.-D. Cheng, C.-S. Chen, A comparison of efficiency and error convergence of multiquadric collocation method and finite element method, Eng.
Anal. Bound. Elem. 27 (3) (2003) 251–257.
[28] Q. Jiang, C. Yao, Z. Ye, C. Zhou, Seepage flow with free surface in fracture networks, Water Resour. Res. 49 (1) (2013) 176–186.
[29] M. Sharan, E. Kansa, S. Gupta, Application of the multiquadric method for numerical solution of elliptic partial differential equations, Appl. Math. Comput.
84 (2) (1997) 275–302.
[30] N. Mai-Duy, T. Tran-Cong, Numerical solution of differential equations using multiquadric radial basis function networks, Neural Netw. 14 (2) (2001)
185–199.
[31] W.F. Hughes, E.W. Gaylord, Basic Equations of Engineering Science, Schaum Publishing Company, New York, 1964.

19