You are on page 1of 14

Numer Algor (2007) 45:61–74

DOI 10.1007/s11075-007-9110-6
ORIGINAL PAPER

A comparison of numerical integration rules


for the meshless local Petrov–Galerkin method

Annamaria Mazzia · Massimiliano Ferronato ·


Giorgio Pini · Giuseppe Gambolati

Received: 30 October 2006 / Accepted: 22 May 2007 /


Published online: 20 July 2007
© Springer Science + Business Media B.V. 2007

Abstract The meshless local Petrov–Galerkin (MLPG) method is a mesh-


free procedure for solving partial differential equations. However, the benefit
in avoiding the mesh construction and refinement is counterbalanced by the
use of complicated non polynomial shape functions with subsequent difficul-
ties, and a potentially large cost, when implementing numerical integration
schemes. In this paper we describe and compare some numerical quadrature
rules with the aim at preserving the MLPG solution accuracy and at the same
time reducing its computational cost.

Keywords Meshless methods · Meshless local Petrov–Galerkin method ·


Numerical integration rules

1 Introduction

Mesh-free techniques for solving partial differential equations (PDEs) in


physics and engineering have the attractive advantage of eliminating the need
of building geometric meshes, as is required by any mesh-based technique,
at the cost, however, of using complicated shape functions with subsequent
difficulties in implementing accurate and inexpensive numerical integration
rules [2, 3, 11, 12, 17].

A. Mazzia (B) · M. Ferronato · G. Pini · G. Gambolati


Dipartimento di Metodi e Modelli Matematici per le Scienze Applicate,
Università degli Studi di Padova, via Trieste 63, 35121 Padova, Italy
e-mail: mazzia@dmsa.unipd.it
62 Numer Algor (2007) 45:61–74

In this paper, we describe and compare some numerical integration schemes


specifically developed for circular sectors with the aim at obtaining an efficient
integration rule for the meshless local Petrov–Galerkin (MLPG) [4–7]. This is
a “truly” meshless scheme since it does not require any background integration
cells, with the local symmetric weak form (LSWF) yielding the computation of
integrals carried out over sub-domains of regular shape. The great flexibility
in the choice of the trial and test functions leads to the development of
different MLPG variants, called MLPG1..., MLPG6 [5, 7]. Among them,
MLPG1 and MLPG5 appear to be the most promising formulations, with
the latter involving only boundary integrals over each sub-domain and over-
coming the problem of integrating rational functions. However, the MLPG5
implementation, besides suffering from a few theoretical drawbacks, in some
engineering applications, e.g. axi-symmetric problems, does not lead to such a
simplification [13], thus reducing the attractiveness of this method. Therefore,
in this study we investigate the role of numerical integration in MLPG1.
The numerical integration to approximate the integrals in the LSWF is
actually a crucial issue in any MLPG since the integrands are non polynomial
functions and conventional schemes, such as the Gaussian quadrature rule,
require several points to provide results of acceptable accuracy [3]. Moreover,
as observed in [2], an accurate integration is particularly difficult because the
Gaussian quadrature rule is based on an infinitely differentiable interpolation
function, whereas the functions involved in the MLPG integrals are not so.
A way to overcome such a difficulty may be the partition of the integration
domain into many small volumes or areas or segments in order to get a higher
accuracy [1, 3]. Quadrature formulae based on few integration points in MLPG
are given in [17], with their use restricted, however, to the local domains fully
embedded in the global domain.
In this paper we compare the conventional Gaussian quadrature formulae
with a less known formula introduced in [18] and developed in [11–13] and
with a piecewise midpoint quadrature rule [12]. The MLPG1 scheme is briefly
reviewed and the integrals arising in the solution process defined. Different
numerical quadrature formulae are then addressed and some numerical results
discussed. A few remarks close the paper.

2 MLPG1 scheme

For the sake of simplicity, let us consider as a test problem the linear 2-D
Poisson’s equation
∇ 2 u(x) = p(x) on  (1)
where p is a given source function and  is the domain enclosed by  = u ∪
q , with boundary conditions
u=u on u
∂u
≡q=q on q (2)
∂n
Numer Algor (2007) 45:61–74 63

where u and q are the prescribed potential and normal flux, respectively, on the
Dirichlet boundary u and on the Neumann boundary q , and n is the outward
normal direction to .
A LSWF over a local sub-domain (i) s may be written as

     
∂u ∂v ∂u ∂v
qv d + qv d + qv d − + + pv d = 0
Ls su sq (i)
s
∂x ∂x ∂y ∂y
(3)

with v the test function. The local sub-domain s(i) can have an arbitrary shape,
but it is typically set as the intersection of a circle centered at node xi with fixed
radius r0(i) with  [4]. The boundary ∂(i) s generally consists of a portion Ls
located inside the global domain, where no boundary conditions are specified,
and two portions su and sq located on the global domain boundary  where
Dirichlet and Neumann boundary conditions are given, respectively. For (i) s
located entirely within , there is no intersection between ∂s(i) and , hence
the integrals over su and sq vanish.
The LSWF is prescribed over any node xi . Generally n randomly distributed
nodes are chosen thus n LSWFs have to be considered over the local sub-
domains (i) s . In MLPG1, the trial function u is defined by the moving least
square (MLS) approximation [7, 8, 15] as u(x) = nj=1 j(x)û j where j are
the MLS shape functions and û j are fictitious nodal values. Moreover, a test
function v vanishing over Ls is selected, thus eliminating the corresponding
integral in (3). Typically v is the same weight function as in the MLS approx-
imation, i.e. Gaussian or quartic spline [7, 8, 16], with compact support over
(i)
s . For more details, see for instance [14].
The final system of equations is

Kû = f (4)

with û the vector of fictitious nodal values and K the stiffness matrix. Setting
v = v(x, xi ), for i, j = 1, . . . , n, the K and f coefficients read:

  
∂ j ∂v(x, xi ) ∂ j ∂v(x, xi )
Kij = + d
(i)

s ∂ x ∂x ∂y ∂y
∂j ∂j
− nx + n y v(x, xi ) d
 su ∂ x ∂y
fi = qv(x, xi ) d − pv(x, xi ) d
sq (i)
s


If xl is a Dirichlet node, the lth equation in (4) is replaced by nj=1 j(xl )û j =
ul , thus exactly enforcing the Dirichlet conditions over u [3].
An accurate and efficient evaluation of the integrals over the sub-domain
(i)
s in Kij and fi is a most crucial issue to obtain a numerical solution of (1).
64 Numer Algor (2007) 45:61–74

3 Numerical integration schemes

The local sub-domains (i) s can be: (a) circles inside  for interior nodes
sufficiently far from the boundary, (b) circular sectors for boundary nodes,
(c) sub-domains formed by circular sectors plus one or more triangles for
interior nodes for which (i) s intersects . We focus on numerical integration
rules developed for sub-domains of type (a) and (b) only. Integration over
regions of type (c) is implemented with the same procedure as type (a)
neglecting the quadrature points lying outside (i) s . Such a procedure is chosen
for the sake of its simplicity, recalling that, by the way r0(i) is selected, typically
in a real problem there are very few (i)s regions of type (c).
Let us consider the integral  f (x, y) d, where f (x, y) is an integrable
function and the domain  is a circular sector with center (x0 , y0 ) and radius r0 ,
bounded by the angles θ1 and θ2 (for θ1 = 0 and θ2 = 2π we obtain a circle). The
simplest idea is to use the Gauss–Legendre product rule with the integration
points chosen along the x- and y-directions. Unfortunately, this rule is not
as accurate as one should expect because it is not able to exactly evaluate
the area of a unit circle, i.e. the integral with f (x, y) = 1. Indeed, a 20 × 20
Gauss–Legendre formula is necessary to achieve an absolute error of about
2 × 10−4 [12].
Polar coordinates are more appropriate for integration on circular domains.
Applying an affine transformation from the unitary square to the circular
sector we obtain the formula:

  
nρ nθ

f (x, y) d ≈ wi w j G ξi , η j , (5)
 i=1 j=1

where nρ and nθ are the number of quadrature points along ξ - and η-axis;
ξi , i = 1, . . . , nρ , and η j, j = 1, . . . , nθ , are the abscissae of the Gauss–
Legendre formula with corresponding weights wi and w j over the inter-
val [−1, 1]; the function G(ξ, η) = F(ρ, θ)ρ [r0 (θ2 − θ1 )] /4, F(ρ, θ ) = f (x0 +
ρ cos (θ), y0 + ρ sin (θ)), and the polar coordinates ρ and θ are related to ξ and
η by the affine transformation

r0 r0
ρ= ξ+
2 2
θ2 − θ1 θ2 + θ1
θ= η+ .
2 2

We denote this formula as Rule 1.


Generally, it is advisable to develop a rule taking into account the geometry
of the integration domain. We follow the guidelines suggested by Peirce [18]
to integrate over the planar annulus, and by De and Bathe [11] to integrate
over an interior disk and a boundary sector. We adopt a different approach
using a Gauss–Legendre formula for the ρ integral and a midpoint rule for the
Numer Algor (2007) 45:61–74 65

θ integral, instead of the Gauss–Chebyshev rule [18] that is easy to apply only
for θ2 = 2π , and the “engineering solution” proposed in [11]:
  r0  θ2  
nρ nθ
f (x, y) d = F(ρ, θ)ρ dρ dθ ≈ ai b j F(ρ j, θi ) (6)
 0 θ1 i=1 j=1

where F(ρ, θ) = f (x0 + ρ cos (θ), y0 + ρ sin (θ)), ρi is the square root of the
ith zero of the nρ -degree Legendre polynomial with ai the corresponding
weight, and θ j denotes the midpoint of the jth circular sector with angular size
b j. Since ai = (r02 wi )/4 and θ j
= θ1 + ( j − 1/2)b
j , b j = (θ2 − θ1 )/nθ it follows
straightforwardly that ai b j = wi r02 (θ2 − θ1 ) /(4nθ ) = wi A/(2nθ ) with A the
circular sector area. The quadrature formula (6) will be denoted as Rule 2.
Another formula has proved computationally efficient in the method of
finite spheres [12]. A piecewise midpoint rule is implemented by subdividing
the domain in concentric circular sectors and radial lines, and computing
the integral over each subdomain as the subdomain area multiplied by the
integrand calculated at the centroid [12]. By considering nθ subdivisions of the
sector, nρ subdivisions along each radius, and approximating sin (x/2) with x/2
we obtain the following polar coordinates for the centroid of each subdomain:
j 2 − j + 1/3
ρj = ρ j = 1, . . . , nρ
j − 1/2
θi = (i − 1/2) θ i = 1, . . . , nθ

where ρ = r0 /nρ and θ = (θ2 − θ1 )/nθ . With the above transformations, the
midpoint formula gives the following approximation:
  
nρ nθ
f (x, y) d ≈ Dij f (ρ j cos (θi ), ρ j sin (θi )) (7)
 j=1 i=1

where Dij is the area of the subdomain, i.e. Dij = ( j − 1/2) θ ( ρ)2 . We
denote this formula as Rule 3.
Several other integration rules for circles have been suggested in the spe-
cialist literature (e.g. [9, 10] and references therein). Implementation of these
cubature formulae in the MLPG seems to yield some promising advances with
respect to traditional rules and requires more investigation. This is currently
under study and will be discussed in a future work.

4 Numerical tests

4.1 Computation of a single integral

We compare the above three numerical integration rules using the absolute
error made in the calculation of known integrals. We assume a circular
integration domain or a circular sector with center in (0, 0).
66 Numer Algor (2007) 45:61–74

Table 1 Absolute error for the integral of cos (x2 + y2 ) on the circle

nρ × nθ r0 = 1 r0 = 4

Rule 1 Rule 2 Rule 3 Rule 1 Rule 2 Rule 3

2×2 3.332e–2 6.319e–4 7.505e–2 2.533e+1 1.588e+0 3.816e+1


3×3 1.054e–3 1.357e–6 3.473e–2 1.415e+1 6.394e+0 9.376e+0
4×4 1.518e–5 1.541e–9 1.976e–2 1.846e+1 3.165e+0 3.398e+1
5×5 4.351e–7 1.079e–12 1.271e–2 2.710e+1 7.826e–1 1.182e+1
6×6 1.626e–9 4.441e–16 8.848e–3 6.942e+1 1.185e–1 1.732e+0
7×7 6.698e–11 2.664e–15 6.510e–3 1.736e+0 1.225e–2 4.042e–1
8×8 8.882e–15 7.994e–15 4.999e–3 9.965e–1 9.238e–4 6.744e–1
9×9 6.217e–15 3.553e–15 3.944e–3 8.497e–2 5.323e–5 6.126e–1
10 × 10 1.776e–15 1.554e–14 3.196e–3 2.687e–2 2.425e–6 5.125e–1

Let f (x, y) = cos (x2 + y2 ), i.e. an axi-symmetric non-polynomial oscillatory


function. The exact integral is:
  θ2  r0
sin (r02 )
cos (x + y ) d =
2 2
cos (ρ 2 )ρ dρ dθ = (θ2 − θ1 )
 θ1 0 2
Table 1 reports the absolute errors from the three integration rules on the
circle (θ1 = 0 and θ2 = 2π ) with radius r0 = 1 and r0 = 4, respectively. With
r0 = 1, 5 × 5 points are enough for Rule 2 to achieve accuracy of order
O(10−12 ), while Rule 1 requires 7 × 7 points to obtain a similar result. On the
contrary, the absolute error from Rule 3 is no less than 10−3 .
Increasing the radius from r0 = 1 to r0 = 4, the integrand function displays
oscillations that Rule 1 and Rule 3 are unable to capture, while Rule 2
requires at least 8 × 8 points to approximate the solution with an error of order
O(10−4 ). Rule 2 is about 4 and 5 orders of magnitude more accurate than Rule
1 and Rule 3, respectively. If the integration domain is a circular sector the
results do not change.
Now let f (x, y) = x2 , i.e. a non axi-symmetric non-oscillatory function. The
exact integral over the domain  is:
  θ2  r0  
r04 sin (2θ2 ) − sin (2θ1 )
x d =
2
ρ cos θρ dρ dθ =
2 2
θ2 − θ1 +
 θ1 0 8 2
Here a different behavior can be observed according to whether the inte-
gration domain is a circle or a circular sector. As shown in Table 2, Rule 1
gives good results on the circle, with errors of order of O(10−5 ) with 7 × 7
points and up to O(10−9 ) with 10 × 10 points. On the circular sector (with
θ1 = 0 and θ2 = 2.5) the error decreases from an initial value of order O(10−2 )
with 2 × 2 points down to O(10−16 ) with 10 × 10 points. By distinction, Rule 2
gives very accurate results on the circle, where the error is of order O(10−16 )
with any distribution of integration points, and a quite unsatisfactory outcome
(O(10−5 )) as well on the circular sector. Rule 3 displays a similar behavior
on the circle and on the circular sector with a relatively large error of order
O(10−3 ). The distribution of the integration points over a circle (Fig. 1) and
Numer Algor (2007) 45:61–74 67

Table 2 Absolute errors for the integral of x2

nρ × nθ r0 = 1, θ1 = 0, θ2 = 2π r0 = 1, θ1 = 0, θ2 = 2.5

Rule 1 Rule 2 Rule 3 Rule 1 Rule 2 Rule 3

2×2 6.944e–1 2.220e–16 7.854e–1 2.812e–2 9.867e–4 2.766e–2


3×3 4.162e–1 2.220e–16 1.372e–2 1.600e–3 4.357e–4 1.182e–2
4×4 9.875e–2 3.331e–16 7.895e–3 4.709e–5 2.446e–4 6.585e–3
5×5 1.309e–2 4.441e–16 5.111e–3 8.471e–7 1.564e–4 4.200e–3
6×6 1.117e–3 7.771e–16 3.573e–3 1.028e–8 1.085e–4 2.911e–3
7×7 6.679e–5 2.220e–16 2.636e–3 8.999e–11 7.970e–5 2.137e–3
8×8 2.961e–6 3.331e–16 2.024e–3 5.945e–13 6.101e–5 1.635e–3
9×9 1.013e–7 2.220e–16 1.600e–3 3.108e–15 4.820e–5 1.291e–3
10 × 10 2.763e–9 2.220e–16 1.300e–3 1.110e–16 3.904e–5 1.046e–3

the circular sector (Fig. 2), along with the behavior of the integrand function,
helps explain the different behavior of the three Rules.
From these numerical experiments we expect Rule 2 to be more appropriate
and robust for the integral of oscillatory functions, while Rule 1 is perhaps
superior whenever the integrand requires more integration points close to the
circular sector center. Rule 3 seems not to be as accurate in both cases.

4.2 Implementation in MLPG1

Implementation of Rules 1, 2 and 3 in MLPG1 does not yield such a pro-


nounced difference as in the approximation of a single integral and, above all,
increasing the number of integration points does not guarantee a correspond-
ing improvement of accuracy.
Indeed in MLPG1 two other factors have to be considered:
(a) The integrand may be zero over portions of (i)
s with no intersection with
the local supports of adjacent nodes, hence some quadrature points are
not actually used;
(b) The error on the numerical solution is due to both the inexact integration
and the MLPG approximation.
Moreover the typical integrand function has a quite complicated behavior and
the above integration rules may be relatively inaccurate even with several
integration points.
As it is not possible to calculate analytically the MLPG1 integrals, we
compare the global errors from elliptic problems with a prescribed solution.
The error is now the outcome of both integral and MLPG approximations. It is
estimated as the L2 norm of the difference between the vector of the numerical
solution u and the vector of the analytical solution uexact :

n n
  exact 2
|e| =  (ui − ui )
exact 2  (ui ) .
i=1 i=1
68 Numer Algor (2007) 45:61–74

1 1

0.5 0.5

0 0

–0.5 –0.5

–1 –1
–1 –0.5 0 0.5 1 –1 –0.5 0 0.5 1
1.0

0.2 0.1
0.5
0.5
0.5 0.6 0.5 .6
0
0.2
0.7
0.4
0.4 0.8
0.1 0.3
0.3 0.7 0.9
0.0
0 0.9
0.8 0.2 0.1
0.9
0.3 0.8
0.5 0.7
0.6
–0.5 0.1 0.6
–0.5 0.5
0.4
0.4
0.3
0.2

–1.0 –1
–1.0 –0.5 0.0 0.5 1.0 –1 –0.5 0 0.5 1

Fig. 1 9 × 9 integration points on the circle by applying Rule 1 (top left), Rule 2 (top right) and
Rule 3 (bottom left), respectively. Bottom right: contour lines of x2 on the circle

4.2.1 Patch tests


Linear and quadratic patch tests are performed with Dirichlet problems
whose exact solutions are given, respectively, by uexact = x + y and uexact =
x2 − y2 + xy, over a domain of size 2 × 2. Boundary conditions are prescribed
in agreement with the exact solution. To pass the patch tests, the MLPG1
algorithm with linear (quadratic) basis functions for the MLS approximation
must provide the exact solution for the linear (quadratic) patch test over all
the interior nodes of the domain to machine accuracy.
We consider nine nodes regularly distributed on the 2 × 2 domain, with
eight nodes on the boundary and one internal. We use a Gaussian weight
function in the MLS approximation with the parameters described in [14].
Eight integration points are selected for quadrature along the boundary.
MLPG1 with Rules 2 and 3 passes the patch tests whatever couple nρ × nθ
of integration points is selected. On the contrary Rule 1 requires at least nθ = 7
in the linear case or nθ = 9 in the quadratic test, to get the correct value over
the internal point.
Numer Algor (2007) 45:61–74 69

1.0 1.0

0.5 0.5

0.0 0.0

–0.5 –0.5

–1.0 –1.0
–1.0 –0.5 0.0 0.5 1.0 –1.0 –0.5 0.0 0.5 1.0
1.0

0.1
0.2 0.1

0.5 0.3
0.4 0.3
0.5
0.5 0.4 0.6
0.1 0.2 0.7
0.5
0.1
0.8
0.3
0.6 0.9
0.4
0.2 0.5 0.7
0.0 0.1
0

–0.5 –0.5

–1.0 –1
–1.0 –0.5 0.0 0.5 1.0 –1 –0.5 0 0.5 1

Fig. 2 9 × 9 integration points on the circular sector (θ1 = 0, θ2 = 2.5) by applying Rule 1
(top left), Rule 2 (top right) and Rule 3 (bottom left), respectively. Bottom right: contour lines
of x2 on the circular sector

4.2.2 Laplace equation


Over the same 2 × 2 domain we solve the Laplace equation with exact solution
uexact = −x3 − y3 + 3x2 y + 3xy2 . Three regular grids of 25, 81 and 289 nodes
are used starting from the nine point pattern of the previous example and
progressively refining it with a node addition at the midpoint of each virtual
edge connecting two adjacent nodes. Essential (Dirichlet) boundary conditions
are prescribed on all sides.
The influence of the integration error on the MLPG1 accuracy is investi-
gated by comparing the theoretical convergence to the numerical one as the
node pattern is regularly refined. Figure 3 shows the convergence profiles for
Rule 1, 2 and 3 obtained with 7 × 7 up to 10 × 10 integration points using
both a linear and a quadratic basis for the MLS approximation. The so-called
“exact” profile in Fig. 3 is computed using 64 × 64 integration points, with
which every rule turns out to provide the same outcome. Increasing the
number of integration points generally improves the MLPG1 accuracy, but
with Rule 1 the convergence rate of the numerical solution to the analytical
one is often lost. Rule 1 appears to be particularly inappropriate when using a
70 Numer Algor (2007) 45:61–74

1e-04 1e-04
|e|

|e|
7x7 points 7x7 points
9x9 points 9x9 points
10x10 points 10x10 points
1e-05 MLS linear basis exact 1e-05 MLS linear basis exact

1e-04 1e-04
|e|

|e|
1e-05 MLS quadratic basis 1e-05 MLS quadratic basis

0.1 0.1
h h

1e-04
|e|

7x7 points
9x9 points
10x10 points
1e-05 MLS linear basis exact

1e-04
|e|

1e-05 MLS quadratic basis

0.1
h

Fig. 3 Error |e| vs. the node spacing h solving the Laplace equation with MLPG1 and: Rule 1
(top left) ; Rule 2 (top right); Rule 3 (bottom)

quadratic basis for the MLS approximation. In fact, the inaccurate integration
yields the paradoxical result that the global MLPG1 error with a quadratic
basis is larger than that with a linear basis. By contrast, Rule 2 and 3 exhibit
a much better behavior. With a linear basis, the numerical and theoretical
convergence profiles are almost indistinguishable even with 7 × 7 points, while
with a quadratic basis the theoretical convergence is practically obtained using
10 × 10 points. Note that, anyway, even with Rule 2 and 3 the integration
errors are larger for the quadratic basis than for the linear one.
Ferronato et al. [13] have shown that the MLPG1 accuracy can be improved
by properly scaling the radius of the local sub-domains (i)s by a dimensionless
factor β. However, the possible advantage of using β > 1, i.e. a larger local
sub-domain, can be concealed by larger integration errors associated to the
increased number of domains of type (c). This is investigated by setting β = 1.2
in the current Laplace test case. The comparison between the theoretical and
the numerical convergence profiles, obtained using 10 × 10 integration points,
is shown in Fig. 4. The error induced by the approximated integrals is larger
Numer Algor (2007) 45:61–74 71

Fig. 4 Error |e| vs. the node


spacing h solving the Laplace
equation with MLPG1 and 1e-03
β = 1.2. The profiles for Rule
1, 2 and 3 are obtained using

|e|
10 × 10 integration points
1e-04 Rule 1
Rule 2
Rule 3
MLS linear basis exact
1e-05

|e| 1e-03

1e-04

MLS quadratic basis


1e-05
0.1
h

than with β = 1 (Fig. 3) with the theoretically higher MLPG1 accuracy lost
in practice. In particular, Rule 1 appears to be the most inaccurate technique,
while Rule 2 allows for the best outcome.
The influence of the integration rule on the MLPG1 accuracy is finally
investigated in the same Laplace test case discretized with a random node
pattern. The two patterns shown in Fig. 5, consisting of 81 and 289 nodes,
respectively, are used. This numerical experiment resembles a most typical
practical situation where the node distribution and its possible refinement are
generally non uniform. The global MLPG1 numerical error obtained with Rule
1, 2 and 3 and 10 × 10 points is compared to the theoretical one obtained

2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2

Fig. 5 Non uniform node patterns with: 81 nodes (left); 289 nodes (right)
72 Numer Algor (2007) 45:61–74

1e-03 1e-03
|e|

|e|
Rule 1 Rule 1
Rule 2 Rule 2
Rule 3 Rule 3
1e-04 MLS linear basis exact 1e-04 MLS linear basis exact

1e-03 1e-03
|e|

|e|
1e-04 MLS quadratic basis 1e-04 MLS quadratic basis

0.5 1 0.5 1
β β

Fig. 6 Error |e| vs. the scaling factor β solving the Laplace equation with MLPG1 and the node
pattern: (left) in Fig. 5–left; (right) in Fig. 5–right

with an “exact” integration for a variable size of (i) s (Fig. 6). The “exact”
integration profile is practically computed using again 64 × 64 points, with
all Rules giving the same outcome. The integration error and the choice of
the Rule have practically no influence on the MLPG1 accuracy when using
a linear basis for the MLS approximation. By contrast, the role of numerical
integration is quite important when using a quadratic basis. With β = 0.4 every
sub-domain (i)s is a circle inside the global domain and the error introduced by
the numerical integration is negligible. Increasing β, however, the global error
increases too, almost independently of the integration Rule, while theoretically
it should be expected to decrease.
On summary, we can conclude that the integration error turns out to be
more important with a quadratic basis than with a linear basis for the MLS
approximation. Rule 1 is generally the less accurate technique, while Rule 2
and 3 behave similarly with a slight advantage for Rule 2 in most of the test
case considered herein.

5 Conclusions

The accurate and efficient computation of integrals in MLPG is a crucial issue


for the success of the method. The present contribution is devoted to the
comparison of three integration rules in the MLPG1 solution of a Laplace
problem with Dirichlet boundary condition, with the aim at emphasizing the
role of the numerical integral accuracy in the global MLPG1 approximation.
Potentially accurate quadrature rules, such as Rules 1 and 2, prove less effec-
tive when used with the MLPG1 scheme because of the generally complicated
non-polynomial behavior of the integrand functions. A simple rule, such as
Rule 3, provides a similar MLPG1 outcome although it does not appear to
Numer Algor (2007) 45:61–74 73

be as accurate in the evaluation of a single integral. The main results can be


summarized as follows:

– Inaccurate integral computations in MLPG1 can yield significant devia-


tions from theoretical convergence. For example, contrary to what one
would expect the use of a quadratic basis for MLS can provide larger errors
than a linear basis because of integral inaccuracy;
– Rule 2 and 3 generally behave similarly, while Rule 1, which is the most
frequently used technique in MLPG, needs more integration points to be
as accurate;
– The presence of a large number of non-circular domains, e.g. when a
scaling factor β > 1 for the radius of (i)s is used, generally worsens the
global MLPG accuracy because of a larger approximation in the numerical
integral computation. Among the investigated formulas, Rule 2 appears to
provide in this case the most accurate results.

The choice of the most efficient quadrature rule along with the number of
integration points must be generally a trade-off between the accuracy and the
cost of the global MLPG solution. The results of the present contribution seem
to indicate that Rule 1, which is suggested in classical MLPG formulations,
can often be expensive and inaccurate. Simpler rules, such as Rule 2 and 3,
are slightly cheaper, as a smaller number of operations are required for each
function evaluation, and often allow for a more accurate global MLPG result.
Research is currently underway to implementing in MLPG more advanced
integration rules which may allow for promising results.

Acknowledgement This study has been supported by the Italian MIUR project “Numerical
models for multiphase flow and deformation in porous media.”

References

1. Atluri, S.N., Han, Z.D., Rajendran, A.M.: A new implementation of the meshless finite volume
method, through the MLPG “mixed” approach. Comput. Model. Eng. Sci. 6(6), 491–513
(2004)
2. Atluri, S.N., J, Y.C., Kim, H.G.: Analysis of thin beams, using the meshless local
Petrov–Galerkin method, with generalized moving least squares interpolations. Comput.
Mech. 24, 334–347 (1999)
3. Atluri, S.N., Kim, H.G., Cho, J.: A critical assessment of the truly meshless local
Petrov–Galerkin (MLPG) and local boundary integral equation (LBIE) methods. Comput.
Mech. 24, 348–372 (1999)
4. Atluri, S.N., Shen, S.: The Meshless Local Petrov–Galerkin (MLPG) Method. Tech Science,
Forsyth, GA (2002)
5. Atluri, S.N., Shen, S.: The Meshless Method (MLPG) for Domain and BIE Discretizations.
Tech Science, Forsyth, GA (2004)
6. Atluri, S.N., Zhu, T.: A new meshless local Petrov–Galerkin (MLPG) approach in computa-
tional mechanics. Comput. Mech. 22, 117–127 (1998)
7. Atluri, S.N., Zhu, T.: The meshless local Petrov–Galerkin (MLPG) method: a simple and less-
costly alternative to the finite element methods. Comput. Model. Eng. Sci. 3(1), 11–51 (2002)
8. Belytschko, T., Krongauz, Y., Organ, D., Fleming, M., Krysl, P.: Meshless methods: an
overview and recent developments. Comput. Methods Appl. Mech. Eng. 139, 3–47 (1996)
74 Numer Algor (2007) 45:61–74

9. Cools, R., Kim, K.: A survey of known and new cubature formulas for the unit disk. Report
TW 300, Katholieke Universiteit Leuven, Department of Computer Science (2000)
10. Cools, R.: An encyclopaedia of cubature formulas. J. Complex. 19(3), 445–453 (2003)
11. De, S., Bathe, K.J.: The method of finite spheres. Comput. Mech. 25, 329–345 (2000)
12. De, S., Bathe, K.J.: The method of finite spheres with improved numerical integration.
Comput. Struct. 79, 2183–2196 (2001)
13. Ferronato, M., Mazzia, A., Pini, G., Gambolati, G.: A meshless method for axi-symmetric
poroelastic simulations: numerical study. Int. J. Numer. Methods Eng. 70(12), 1346–1365
(2007) (doi:http://dx.doi.org/10.1002/nme.1931)
14. Ferronato, M., Mazzia, A., Pini, G., Gambolati, G.: Accuracy and performance of meshless
local Petrov–Galerkin methods in 2-D elastostatic problems. JP J. Solids Struct. 1, 35–59 (2007)
15. Lancaster, P., S̆alkauskas, K.: Surfaces generated by moving least squares methods. Math.
Comput. 37(155), 141–158 (1981)
16. Lu, Y.Y., Belytschko, T., Gu, L.: A new implementation of the element free Galerkin method.
Comput. Methods Appl. Mech. Eng. 113, 397–414 (1994)
17. Pecher, R.: Efficient cubature formulae for MLPG and related methods. Int. J. Numer.
Methods Eng. 65, 566–593 (2006)
18. Peirce, W.H.: Numerical integration over the planar annulus. J. Soc. Ind. Appl. Math. 5(2),
66–73 (1957)

You might also like