Meshless and Generalized Finite Element

Methods: A Survey of Some Major Results

I. Babuska1 ? , U. Banerjee2 ??
, and J. E. Osborn3 ???

1
Texas Institute for Computational and Applied Mathematics, ACE 6.412,
University of Texas at Austin, Austin, TX 78712.
2
Department of Mathematics, 215 Carnegie, Syracuse University, Syracuse, NY
13244. Email address: [email protected].
3
Department of Mathematics, University of Maryland, College Park, MD 20742.
Email address: [email protected].

Abstract. In this lecture, we discuss Meshless and Generalized Finite Element

Methods. We survey major results in this area with a unified approach.

1 Introduction
For concreteness and simplicity we will address the weak solution of the model
problem
u + u = f (x), on Rn , (1)
u
= 0 on (2)
n
or
u = 0 on , (3)
for f L2 () given. We will assume that is a Lipschitz domain. Ad-
ditional assumptions on will given as needed. The weak solution u0
H 1 () (H01 (), respectively) satisfies

B(u0 , v) = F (v), for all v H 1 () (v H01 (), respectively), (4)

where
Z Z
B(u, v) (u v + uv) dx and F (v) f v dx. (5)

The energy norm of u0 is defined by

ku0 kE B(u0 , u0 )1/2 = ku0 kH 1 () . (6)

?
The work of this author was partially supported by Office of Naval Research
Grant N00014-99-1-0724.
??
The work of this author was partially supported by the Texas Institute for Com-
putational and Applied Mathematics, University of Texas at Austin.
???
The work of this author was partially supported by the Texas Institute for Com-
putational and Applied Mathematics, University of Texas at Austin.
2 I. Babuska, U. Banerjee, and J. E. Osborn

We will write H instead of H 1 () or H01 () if no misunderstanding can

occur.
Let S H be a finite dimensional subspace, called the approximation
space. Then the Galerkin approximation, uS S, to u0 is determined by

B(uS , v) = F (v), for all v S. (7)

It is immediate that

ku0 uS kH 1 () = inf ku0 kH 1 () . (8)

S

Hence, the main problem is the approximation of u0 by functions in S.

Remark 1. The Finite Element Method (FEM) is the Galerkin Method where
S is the span of functions with small supports. For the history of the FEM,
see [5] and the reference therein.

Remark 2. The classical Ritz method uses spaces of polynomials on for

the approximation spaces; see, e.g., [31].

As mentioned above, the Finite Element Method uses basis functions

with small supports, e.g., hill functions. The theory of approximation with
general hill functions with translation invariant supports was developed in
1970 in [1] using the Fourier Transform. The results in [1] were applied to
the numercial solution of PDE in [2]. A very similar theory, also based on the
Fourier Transform, was later developed in [35], [36]; see also [26]. Later, hill
functions were called Particle Functions (see [22]). In the 1990s, hill functions
began to be used in the framework of meshless methods. For a broad survey
of meshless methods see [27]. A survey of the approximation properties of
radial hill functions is given in [14].
In this paper we will survey basic meshless approximation results and
their use in the framework of Galerkin Methods.

2 Approximation by Particle Shape Functions

Associated with Uniformly Distributed Particles in
Rn . The h-version
Let
Z n {j = (j1 , j2 , . . . , jn ) : j1 , . . . , jn integers}
be the integer lattice, and let

xhj {(j1 h, . . . , jn h) = hj, where j = (j1 , . . . , jn ) Z n }.

The xhj s are called (uniformly distributed) particles. When considering such
a family of particles, we often construct associated shape functions as follows:
 Meshless and Generalized Finite Element Methods: 3

Let (x) H q (Rn ), for some 0 q, be a function with compact support. Let
supp and, for the sake of simplicity, assume that 0 (interior of ).
is called the basic shape function. Then, for 0 < h and j Z n , define
x
hj (x) = ( j). (9)
h
We will be interested in the approximation properties of the space

X
V h () = v = wjh hj (x) : wjh R , as h 0. (10)
n

jZ

Here is the family hj ; hj s are called particle shape functions associated
with particles xhj s and wjh s are called weights. More specifically, given u
H k (Rn ), we will be interested in estimating
inf ku vkH 1 (Rn ) ,
vV h ()

where S in (8) is replaced by V h (). This problem was discussed in [1] and
later in [36]. Essentially the same results were proved in these two papers.
Since has compact support, its Fourier Transform, (), has derivatives of
all orders with respect to , (in fact, it is an entire function). We now cite
the main theorem from [36].
Theorem 1 Suppose H q (Rn ) has compact support. Then the following
three conditions are are equivalent:
1.
(0) 6= 0 (11)
and
D (2j) = 0, for 0 6= j Z n and || p. (12)
Here we use the usual multi-index notation for partial derivatives ( =
(1 , . . . , n ), with i 0, is a multi-index, || 1 + . . . + n , and

||
D = ).
11 nn
2. For || p,
X
j (x j) = d x + q||1 (x), where d 6= 0, (13)
jZ n
and degree q||1 < ||.
The equality in (13) is equality in L2 (Rn ), i.e., equality for almost all x
Rn . The function of the right-hand side of (13) is, of course, continuous.
If the function on the left-hand side is continuous, which will be the case
if q > n/2, then (13) will hold for all x Rn .
4 I. Babuska, U. Banerjee, and J. E. Osborn

3. For each u H p+1 (Rn ), there are weights wjh R, for j Z n and 0 < h,
such that
X
ku wjh hj kH s (Rn )
jZ n

Chp+1s kukH p+1 (Rn ) , for 0 s min{q, p + 1}, (14)

and X
hn (wjh )2 K 2 kuk2H 0 (Rn ) . (15)
jZ n

Here C and K may depend on q, p, and s, but are independent of u and h.

The exponent p + 1 s is the best possible if p is the largest integer for which
(13) holds.

Definition: If (13) holds, the basic shape function is called Quasi-

Reproducible of Order p. If (13) holds with d = 1 and q||1 (x) = 0,
is called Reproducible of Order p.
If is quasi-reproducible of order p (respectively, reproducible of order
p), then the corresponding particle shape functions ji are also called quasi-
reproducible of order p (respectively, reproducible of order p).
Uniformly distributed particles and associated particle shape functions
are translation invariant in the sense that

xhj+l = xhj + xhl and hj+l (x) = hj (x xhl ),

and will sometimes be referred to as translation invariant. They are a special

case of general (nonuniformly distributed) particles, which will be addressed
in the next section. We have, however, more detailed results for translation
invariant particles.
In one dimension we can prove more.

Theorem 2 [36] Let satisfy the conditions of Theorem 1 (with n = 1).

Then  p+1
sin(/2)
() = Z() ,
/2
where Z() is an entire function.

This result means the is a convolution of B-splines with functions of

compact support.

Theorem 3 [1] The interval ( p+1 p+1

2 + , 2 ), for any  > 0, cannot be
the support of , where satisfies the conditions of Theorem 1.

This result means that the B-spline has minimal support.

 Meshless and Generalized Finite Element Methods: 5

3 Approximation by Particle Shape Functions

Associated with Arbitrary (Non-Uniformly
Distributed) Particles in Rn . The h-version
In this section we will generalize the major part of Theorem 1
Suppose {X }N is a family of countable subsets of points in Rn ; the
family is indexed by the parameter , which varies over the index set N . The
points in X are called particles, and will be denoted by x, to distinguish
them from general points in Rn . If it is necessary to underline that x X ,
we will write x = x . To each x X we associate
hx = hx = a positive number;
x = x = a bounded domain in Rn ;
x = x = a function in H q (Rn ), with x = x supp x assumed
compact.

Regarding {X }, hx , x , and x , we make several assumptions:

1. For each , [
x = Rn ,
xX

i.e., for each , {x }xX is an open cover of Rn .

2. For x X , let
Sx {y X : x y 6= }.

There is a constant < depending on {X }N , but neither on , nor

on x X , such that

card Sx , for all x X and all N.

3. Let
Bx = {x Rn : kx xk }
denote the ball of radius centered at x. There is a 0 < < 1 such that
h h
x Bx , for all x X , for all N.
x x
Bx

4. For all x X , for all N , x x and x x .

5. For x X , let
Qx {y X : x y 6= }.
It follows from Items 2 and 4 that

card Qx , for all x X , for all N.

6 I. Babuska, U. Banerjee, and J. E. Osborn

Let [
x = y .
yQ
x

There is a 0 < < , which may depend on {X }, but is independent

of x and , so that
h
x Bx , for all x X , for all N.
x

6. There is a set Q
x Qx such that given data v(y), for y Qx , there is
a unique polynomial Px,p (x) of deg p that minimizes
X
|v(y) Px,p (y)|2 .
yQ
x

Px,p (x) satisfies

kPx,p kL2 (x ) C1 max

|v(y)|,
yQx

where C1 is independent of x and .

7. There is a one-to-one mapping Ax : Pp Pp such that

X  1 
Ax Pp (y)y (x) = Pp (x), for x Rn , and any Pp Pp ,
yQ
x

(16)
and  
1
Ax C2 . (17)
)L ( )
L2 (x 2 x

C2 may depend on {X }N , but neither on x X , nor on .

Remark 3. It is necessary for Item 7 of these assumptions to be valid, that

for each , [
x = Rn ,
xX

which in turn implies that card Qx , card Sx 1, in Items 5 and 2 respectively.

Also, Item 7 of these assumptions plays the role of quasi-reproducibility

of order p of the shape functions x , for x X . The shape functions x are
said to be quasi-reproducible of order p if,
X
x x (x) = d x + q||1 (x),
xX

where d 6= 0 and degree q||1 (x) < ||.

Now we get,
 Meshless and Generalized Finite Element Methods: 7

Theorem 4 Suppose {X }, hx , x , and x satisfy Assumptions 17. Sup-

pose
X
kuk2 r +1 hx < , where rx p, for all x X , for all N,
x
H (Bx )
xX

with > 1 independent of x and . Then there are weights wx R, for

x X , for all N , such that
X
ku wx x kH l (Rn )
xX
1/2

(rx +1l)2
X
C (hx ) kuk2 +1
rx h
x
, (18)
H (Bx )
xX

for 0 l l min{q, rx + 1 : x X } for all N . C here depends on

the constants in assumptions 17, but neither on u, nor on .

The proof of this theorem is given in [12].

Remark 4. For uniformly distributed particles (and associated particle shape

functions) Assumption 17 are satisfied with hx = h and x = xhj .

Remark 5. In [12], we will address some ways to construct particle shape

functions that satisfy Assumption 17. In general, the verification of these
assumption is not easy for higher p; for p = 0 the construction and verification
is usually quite easy.

Remark 6. Theorem 4 is a generalization of the h-version of the FEM. It per-

mits approximation of functions that are characterized by weighted Sobolev
spaces.

So far we have addressed the case when one particle function is associated
with each particle. Let us now consider the case when more shape functions
are used.

Theorem 5 Suppose {X }, hx , x , and x satisfy Assumptions 17. In

terms of x , x X , for all N , define the shape functions

x,x = x (x)(x x) x ,

where |x | tx . Suppose
X
kuk2 +1
rx h
x
< , where rx p + tx ,
H (Bx )
xX
8 I. Babuska, U. Banerjee, and J. E. Osborn

for all x X , for all N . Then there are weights wx,

R such that
x

X X

ku wx, x, kH l (Rn )
x x
xX |t
|x x

1/2

(rx +1l)2
X
C (hx ) kuk2 +1
rx h
x
, (19)
H (Bx )
xX

for 0 l l min{q, rx + tx : x X )}, for all . C here depends on the

constants in assumptions 17, but neither on u, nor on .

The proof of this result is given in [12].

Remark 7. Theorem 5 avoids the difficulties connected with the construction
of particle shape function for higher p (see Remark 5) because we construct
the functions x only for small values of p.
Remark 8. Although Theorem 5 is formulated in the framework of the h
version of the FEM, it is essentially a generalization of the h p version. See
Section 4.
Remark 9. It is also possible to prove an inverse theorem that characterizes
the smoothness of the approximated function in terms of the convergence
rate. This analysis is similar to that in [8].
In section 2 we defined the space V h for the case of uniformly distributed
particle and associated shape functions. We will also use the same symbol, V h
(instead of V ), for the general case of non-uniformly distributed particles,
which was addressed in Theorems 4 and 5.

4 The Generalized Finite Element Method

In the previous section we addressed the approximation by particle functions

of the form x (x)(x x ) x . We now further generalize the character of
these shape functions. We let {x }xX , for N , be shape functions with
compact supports x , and assume they are a partition of unity, i.e.,
X
x = 1, for all N. (20)
xX

We note that we do not assume Items 17, given in Section

S 3, for the particles
considered in this section. However, (20) implies that xX x = Rn , for
each . We also assume that the pointwise overlap

M sup card {x X : y x } < .

yRn , N
 Meshless and Generalized Finite Element Methods: 9

We further assume that

kx kL (Rn ) C1 (x), for all x X , for all , (21)

C2 (x)
kx kL (Rn ) , for all x X , for all , (22)
diam x
and
x H 1 (x ) is an N (x) dimensional space defined on x , (23)
for x X , for all . Then we have
Theorem 6 Let u H 1 (Rn ) and suppose for every x X , for all , there
exists vx x such that

ku vx kL2 (x ) 1 (x), for all x X , for all (24)

and
ku vx kL2 (x ) 2 (x), for all x X , for all . (25)
Then

x vx H 1 (Rn ), with compact support, for all x X , for all (26)

and
X
ku x vx kH 1 (Rn ) (27)
xX
! 1/2
X C2 (x)2 X
2M 1 (x)2 C1 (x)2 + + 2 (x)2 C1 (x)2
( diam x )2
xX xX

for all .

Theorem 7 Suppose 1 x and that

inf ku kL2 (x ) CP (diam x ) |u|H 1 (x ) , for all x X , for all , (28)

R

with CP < independent of x and , and where | |H 1 (x ) is the semi-norm.

Also assume that C1 , C2 in (21) and (22) respectively, are independent of x
and . Then there exists vx x , satisfying (25) such that
1/2
X X
ku x vx kH 1 (Rn ) C 2 (x)2 , for all , (29)
xX xX

where C depends on C1 , C2 , M , and CP , but is independent of u and 2 .

10 I. Babuska, U. Banerjee, and J. E. Osborn

Remark 10. Inequality (28) is the Poincare inequality. It holds if there are
balls Bx and Bx Bx , with diameters x and x , respectively, such
that x x , with independent of x and , and Bx x Bx .
Remark 11. If x is a space of polynomials, then Theorem 6 leads to Theo-
rem 5. Moreover, if the assumptions of Theorem 7 hold, then Theorem 7 also
leads to Theorem 5.
Remark 12. Theorem 6 and Theorem 7 are generalizations of the p-version
of the FEM.
Remark 13. Theorem 6 allows the use of additional information about the
approximated function. For example, if u is harmonic, then x can be the
space of harmonic polynomials; if u has various singularities, as in the neigh-
borhood of the corners or has boundary layer behavior, x can be chosen
accordingly. For proofs of Theorem 6 and Theorem 7, we refer to [7], [9], [12],
[30].
Remark 14. If u = 0 on part of x , we do not include the constant function
in the space x .
Remark 15. The GFEM is obviously a generalization of the meshless method.

5 Approximation in Bounded Domains

In the previous sections we have addressed the approximation of functions
in Rn by linear combinations of particle shape functions. These results lead
immediately to approximation results in Lipschitz domains. We need only to
use an extension theorem.

Theorem 8 [33] Suppose is a Lipschitz domain and suppose k 0. Then

there is a bounded extension operator E : H k () H k (Rn ), i.e., an operator
E such that for all u H k (), v = E(u) has compact support in Rn ,

kvkH k (Rn ) CkukH k () , for all u H k (),

and
v(x) = u(x), for all x .
Here C is independent of u (it does depend on and k).

Theorem 8 yields the desired result when usual Sobolev spaces are used.
Theorems 4 and 5 address the approximation by particle shape functions with
refinements. As stated in Remark 6, we can approximate functions that lie
in the weighted Sobolev spaces. Here we need extension of weighted Sobolev
spaces when the weight is a power of the distance from the vertex of the
domain where the solution of a PDE has singular behavior. Such theorems
for Rn with piecewise smooth boundary were proved in [8].
 Meshless and Generalized Finite Element Methods: 11

Remark 16. We addressed only Lipschitz domains, which exclude slit do-
mains. These domains require special treatment, which, because of the length
restriction of this paper, we will not discuss.
Let us now introduce the notion of a (t, k)-regular system of functions.
For 0 k t, supposes {ht,k ()}0<h1 is a-one parameter family of linear
spaces. We say {ht,k ()}0<h1 is a (t, k)-regular system if
1. ht,k () H k (), for 0 < h 1, and
2. For every w H l (), with 0 l, there is a function g ht,k such that

kw gkH s () Ch kwkH l () , for all 0 s min{l, k}, (30)

where = min{t s, l s}. The constant C is independent of g and h.

Remark 17. In the definition of a (t, k)regular system, we can have = Rn .

We now introduce two additional notions. A (t, k)-regular system {ht,k ()}
is said to satisfy a local assumption if for w H l () with compact support
, the function g ht,k () can be chosen so that the support h of g has
the property that

h h {x : d(x, ) h},

where d(x, ) is the distance form x to , and is independent of , s, k. We

say that {ht,k ()} satisfies an inverse assumption if

kgkH k () Ch(ks) kgkH s () , for all 0 s k and all g H k ().

In [6] many important properties of (t, k)regular systems are proved. Some
of these properties will be used in Section 7.
A linear system of particle shape functions is a (t, k)regular system; more
precisely we have,

Theorem 9 Let {x } and {x }, for x X and all N , be a system of

particles and associated particle shape functions that are quasi-reproducible
of order p. Suppose that the particles are quasi-uniform, i.e.,
hx
1 2 , for all x, y X , for all N,
hy

where 0 < 1 < 2 < are independent of x, y and . Then the associated
linear system of particle shapes functions is a (p + 1, q)-regular system.

The proof of this theorem is given in [12].

Remark 18. With some mild assumptions on the distribution of particle near
the boundary of , the particle shape functions are a (p+1, q)-regular system
that satisfies an inverse assumption [12].
12 I. Babuska, U. Banerjee, and J. E. Osborn

6 Construction of Particle Shape Functions and Some

of Their Properties
Various particle shape functions are used today in meshless methods. Some
major representatives are
1. Reproducing Kernel Particle (RKP) shape functions. These are concep-
tually based on smooth particle hydrodynamics (SPH) approximations
[22]. For RKP, see e.g., [12], [28].
2. Moving Least Squares (MLS) particle shape functions [25]. This idea was
further generalized into Moving Least Squares Kernel (MLSK). See e.g.,
[27], [29].
3. In the GFEM often the classical finite element shape functions are used
as partition of unity functions [9], [11], [30].
One of the major problems in this area is the assessment of the perfor-
mance of the shape functions, respectively, their selection. The shape func-
tions, constructed in Items 1 and 2 above, are reproducible of order p. Error
estimates for the meshless approximation, using such shape functions, are
obtained via an interpolation error estimate. For translation invariant shape
functions (addressed in Theorem 1) associated with uniformly distributed
particles, the error can be analyzed in detail. Let u be sufficiently smooth,
and consider the interpolant I h u of u defined as
X
I h u(x) = u(xhj )hj (x), xhj = hj,
j

where {hj } are reproducible of order p. The accuracy of Galerkin approxi-

mation (8) of a smooth function u (using u instead of u0 ), by these shape
functions, is indicated by the estimate,

inf ku vkH 1 (Rn ) ku I h ukH 1 (Rn ) Chp , (31)

vS

where S = span {hj }. The following result is proved in [13] for n = 1.

Theorem 10 Suppose u is a sufficiently smooth function and consider the

shape functions that are reproducible of order p. Then

ku I h uk2H 1 (R) |p+1 |2H 1 (0,1)

sup lim = , (32)
uH p+2 (R) h0 h2p [|u|2H p+1 (R) + |u|2H p+2 (R) ] (p + 1)!
2

where X
p+1 (x) = xp+1 j p+1 (x j).
jZ

p+1 (x) is a periodic function with period 1, and (x) is the basic shape
function.
 Meshless and Generalized Finite Element Methods: 13

A similar result also holds in Rn for bounded domains.

Theorem 11 Consider the shape functions that are reproducible of order p.

Let u be sufficiently smooth and suppose

ku I h ukL2 (R) Chp+1+ ,

where C is independent of h. Then u is a polynomial of degree p and

I h u = u.

Remark 19. In [13], we computed |p+1 |H 1 (0,1) for several particle shape func-
tions as a comparison of the performance of various particle shape functions.

Remark 20. Theorem 11 is a saturation theorem, and is proved in [13]. It

can serve as a basis for the verification of the implementation code, as h
approaches 0.

Remark 21. In [10], [11] we introduced the theory of the robustness of the
selection of the spaces xhh . We note that xhh is x (defined in Section 4)
j j
for uniformly distributed particles.

We note that the use of reproducibility of order p of shape functions

and the associated analysis in terms of interpolation may give a pessimistic
order of convergence of the usual Sobolev norm error in the approximation
of smooth functions. Thus the notion of quasi-reproducibility of order p of
the shape functions is an appropriate framework for approximation analysis,
but the notion of reproducibility of order p is not appropriate.
Let us consider the situation where the particles are uniformly distributed.
From the definition of reproducibility and quasi-reproducibility of shape func-
tions, it is clear that if a basic shape function (x) is reproducible of order p,
then it is also quasi-reproducible of order p. But a basic shape function may be
reproducible of order p, and quasi-reproducible of higher order p + k, k 1,
in which case the approximation error inf vS ku vkH 1 (R) O(hp+k ) for
smooth u by Theorem 1. But note that (31), which is based on interpolation,
will only yield O(hp ). The basic shape function (x), with Fourier Transform
() = (sin(/2)/(/2))4 (B-spline of order 4), is reproducible of order 1, but
quasi-reproducible of order 3.
We further note that a basic shape function (x) may be reproducible of
order p, and not quasi-reproducible of any higher order. In that case, Theorem
1 and (31) yield the same order of convergence. But even in this case, (x)
may be almost quasi-reproducible of order (p + 1), i.e.,

(a) (x) is reproducible of order p

X
(b) p+1 (x) xp+1 ip+1 (x i) Constant.
iZ
14 I. Babuska, U. Banerjee, and J. E. Osborn

For such basic shape functions, the approximation error (for smooth func-
tions) may decrease at a higher rate, in the pre-asymptotic range (i.e., for
larger h), than is predicted. We note, however, that the pre-asymptotic can
be so large that for practical accuracy, the asymptotic range is not visible. It
can be shown using Theorem 10 that,

|p+1 |H 1 (0,1) hp
 
ku I h ukH 1 (R) + O(hp+ ) [|u|2H p+1 (R) + |u|2H p+2 (R) ]1/2 .
(p + 1)!

For an almost quasi-reproducible basic shape function of the order (p + 1)

(which is also reproducible of order p), we have |p+1 |H 1 (0,1) 0, and ku
I h ukH 1 (R) will exhibit higher order of convergence than the expected order,
which is O(hp ), in the pre-asymptotic range. An example of a RKP-shape
function, which is almost quasi-reproducible of order 2, is given in [13].

7 Solution of Elliptic Boundary Value Problems

In Section 1 we introduced the boundary value problem (1), (2) ((3), respec-
tively), its variational formulation, and its approximate solution. Equation
(8) shows that the approximation theory introduced in the previous sections,
with S = V h , can be used directly for the approximate solution and for the
error estimation when purely natural boundary conditions (2) are prescribed.
Essential (Dirichlet) boundary conditions (3) require the construction of ap-
proximation functions that vanish on , or the approximate satisfaction of
the boundary condition. The GFEM naturally allows the implementation of
Dirichlet boundary conditions by the appropriate choice of the space x . See
also Remark 14.
Let us now consider several methods for boundary value problems with
Dirichlet boundary conditions that use the same approximation space V h
used for Neumann boundary conditions. We will consider three methods:

1. The Penalty Method

2. Nitsches Method
3. The Characteristic Function Method.

The aim is to obtain as much as possible the same rate of convergence as

for the Neumann boundary conditions, i.e., that the Dirichlet conditions will
not adversely influence the rate of convergence. For simplicity we will assume
here that the boundary, , is sufficiently smooth and that the distribution
of particles is quasi-uniform. We will assume the space V h is a (t, k)regular
family of spaces (introduced in Section 5, see Theorem 9) satisfying a local
assumption.

1. The Penalty Method

 Meshless and Generalized Finite Element Methods: 15

Here, instead of the bilinear form (5), we will use

Z Z
B (u, v) = (u v + uv) dx + h uv ds, (33)

with > 0, and as before

Z
L(v) = f v dx. (34)

The penalty method approximation w,h V h satisfies

B (w,h , v) = L(v), for all v V h . (35)

Then we have
Theorem 12 [3] [4] Let u0 be the solution of the problem (1), (3) (H =
H01 ()) with f H l (), l 0. Let w,h V h be the approximate
solution, defined in (35), with k 1, and t l + 2. Then

ku0 w,h kH 1 () C()h kf kH l () , (36)

for  > 0 arbitrary, where C() is independent of f and h, and

+1 t
= min(, + l, (l + 1)), (37)
2 t1
where
1
= max(1, ( + 1)). (38)
2
Remark 22. For the Neumann boundary condition we have

ku0 w,h kH 1 () Ch kf kH l () ,

where = min(l + 1, t 1), which is the maximal rate of convergence.

We see a loss in the rate of convergence, but not a significant one. For
example, using l = 2, t = 4, the maximal rate of convergnce is 3, while
with a certain value of , which minimizes (37), we get the rate 2.74.
2. Nitsches Method
Here we use the bilinear form BN, defined on V h V h by
Z
BN, (u, v) = (u v + uv) dx
Z Z Z
u v 1
v ds u ds + h uv ds, (39)
n n

with > 0, and as before

Z
L(v) = f v dx. (40)

16 I. Babuska, U. Banerjee, and J. E. Osborn

The Nitsches method approximation u,h V h satisfies

BN, (u,h , v) = L(v), for all v V h . (41)

Let us assume that

Z
v
h ( )2 ds C1 kvk2H 0 () , for all v V h , with C1 < . (42)
n
Then we have from [32] (see also [34]),
Theorem 13 Let u0 be the solution of the problem (1), (3) (H = H01 ())
with f H l (), l 0. Let u,h V h be the approximate solution, defined
by (41), with t 1 and k 1.5. Then

ku0 u,h kH 1 () C()h kf kH l () ,  > 0 arbitrary, (43)

where C() is independent of f and h, and

= min(l + 1, t 1). (44)

Remark 23. Condition (42) can be guaranteed by the proper selection of

the particles, e.g., |x | |x |, with > 0 independent of x and
.

3. The Characteristic Function Method

Let be a smooth function on such that

> 0 on , (45)

= 0 on , (46)
and
|| > 0 on . (47)
Let Vh h
= {u : u = v, v V }. Then obviously Vh H01 (), and we
have (see [12]).
Theorem 14 Let uh Vh be the solution of (7) with S = Vh . Then

ku0 uh kH 1 () Ch kukH l () , (48)

where C is independent of u and h, and

= min(l + 1, t 1). (49)

Remark 24. (48), (49) gives the optimal rate of convergence.

Remark 25. We have assumed homogeneous boundary conditions (2),

quasi-uniform particles, and a smooth boundary. This method can be
generalized for non-smooth boundaries.
 Meshless and Generalized Finite Element Methods: 17

8 Certain Implementational Aspects

The implementation of meshless and GFE methods involve certain basic
steps, which are similar to standard FEM. These steps are as follows:
(a) Construction of the stiffness matrix and the load vector
(b) Solving the system of linear equations, and computation of data of inter-
est.
Let us discuss these steps separately:
(a) Computation of the elements of the stiffness matrix and the load
vector for meshless methods, which must be done numerically, is much more
expensive than with standard FEM, because numerical integration cannot be
performed on the master element. If the supports of the shape functions
corresponding to the particles are circles (or spheres), then the underlying
code becomes complex, and the computation of these elements become ex-
pensive (see e.g.,[18]). Also, the quadrature error can have a severe negative
influence on the accuracy of the computed solution (see e.g.,[15], [18]). The
quadrature error analysis, in the context of FEM, has been discussed in de-
tail in Chapter 4 of [16]. This analysis is essentially based on the fact that
D O(1)h|| for || p, and D = 0 for p + 1, where is a
straight element of degree p. For curvilinear elements, the additional fact
that the mapping from master element to physical element converges to a
linear mapping as h 0, is used in the analysis. For meshless methods, the
shape functions satisfy D O(1)h|| for all || p, but we do not
have D = 0 for || p + 1. Moreover, to maintain the optimal order of
convergence, the numerical quadrature rule must approximate the elements
of the stiffness matrix with a relative accuracy of O(hp+2 ). In practice, the
use of adaptive integration to compute the diagonal elements of the stiff-
ness matrix with relative error 0.01hp+2 is quite sufficient. The quadrature
rule, which is used to compute the diagonal elements, is then used for the
off-diagonal elements of the same row and the corresponding element of the
load vector, i.e., the elements are computed simultaneously. In GFEM, the
use of partition of unity function , based of finite elements, simplifies numer-
ical quadrature. Nevertheless, the adaptive quadrature is essential if special
functions, e.g., functions with corner singularities, are used to construct the
shape functions in GFEM. In most cases, numerical integration is the most
expensive part of the computation.
(b) The stiffness matrix in meshless methods may be ill-conditioned for
various distribution of particle points. The GFEM may give rise to a singular
stiffness matrix, because of the linear dependency of the basis functions. For
example, if the space x (see Section 4) is the space of polynomials, and
the partition of unity functions are piecewise polynomial finite elements
functions, then the basis functions are linearly dependent. In such situations,
for each row of the stiffness matrix, every element of that row and the cor-
responding element of the load vector have to be computed simultaneously
18 I. Babuska, U. Banerjee, and J. E. Osborn

using the same adaptive quadrature procedure. Numerical integration is per-

formed only locally over the elements, which have non-empty intersection
with the support of . This is essential to keep the linear system consistent.
The computed stiffness matrix is positive semi-definite, and elimination with
pivoting, or band elimination with perturbation and iteration, can be imple-
mented successfully. Certainly, such a linear system has many solutions, and
this leads to erratic computed solution. Nevertheless, the FEM solution is
unique. This erratic behaviour disappears when data of interest are com-
puted. In fact, the effect of round-off errors and global accuracy are similar
as in the classical FEM (see [37]).
We further note that, the character of the stiffness matrix (i.e., the dis-
tribution of non-zero entries) obtained from GFEM is different from that
obtained from classical FEM [20]. This influences the performance of sparse
elimination solvers. However, one may still use solvers based on the conju-
gate gradient method. We note that the multigrid method is not applicable
to these problems, because the eigenfunctions associated with the zero eigen-
value are oscillatory. Nevertheless, changing the partition of unity function
(at the expense of decreased rate of convergence) gives rise to a non-singular
stiffness matrix, and some form of multigrid method can be used to solve the
linear system (see [23], [24]).

9 Applications
The meshless method is applied in various fields. For a survey, we refer to [27].
In [38], a survey of GFEM with emphasis on complex geometries is given.

10 Additional Comments
Meshless methods (also referred to as Meshfree methods) and GFEM appear
in the literature under different names, although their content is essentially
same. For example, the partition of unity method, the GFEM, the extented
FEM, and the method of clouds are essentially identical methods. See e.g.,
[17], [19], [21]
Meshless methods are based on meshless approximation and discretiza-
tion. The discretization could be based on a variational formulation or on
collocations, and can be used for linear and non-linear problems.

References
1. Babuska, I. (1970): Approximation by Hill Functions, Comment Math. Univ.
Carolinae, 11, pp. 787-811.
2. Babuska, I. (1971): The Finite Element Method for Elliptic Equations, in
Numerical Solution of Partial Differential Equations II, SYNSPADE 1970, B.
Hubbard, ed., Academic Press, London, pp. 69-106.
 Meshless and Generalized Finite Element Methods: 19

3. Babuska, I. (1972): Approximation by Hill Functions II, Comment Math.

Univ. Carolinae, 13, pp. 1-22.
4. Babuska, I. (1973): The Finite Element Method with Penalty, Math. Comp,
27, pp. 221-228.
5. Babuska, I. (1994): Courant Element: Before and After, in Finite Element
Methods: Fifty Years of Courant Element, Lecture Notes in Pure and Applied
Mathematics, Vol. 164, Marcel Dekker, pp. 37-51.
6. Babuska, I. and Aziz, K. A. (1972): Survey Lectures on the Mathematical
Foundations of the Finite Element Method, in The Mathematical Foundations
of the Finite Element Method with Applications to Partial Differential, A. K.
Aziz, ed., Academic Press, pp. 3-345.
7. Babuska, I., Caloz, G. and Osborn, J. (1994): Special Finite Element
Methods for a class of second order elliptic problems with rough coefficients,
SIAM J. Numer. Anal., 31, pp. 945-981.
8. Babuska, I., Kellogg, R. B. and Pitkaranta, J. (1979): Direct and In-
verse Error Estimates for Finite Elements with Mesh Refinements, Numer.
Math., 33, pp. 447-471.
9. Babuska, I. and Melenk, J. M. (1997): The Partition of Unity Finite Ele-
ment Method, Int. J. Numer. Meth. Engng., 40, pp. 727-758.
10. Babuska, I., Banerjee, U. and Osborn, J.: On Principles for the Selection
of Shape Functions for the Generalized Finite Element Method, to appear.
11. Babuska, I., Banerjee, U. and Osborn, J., (2001) : On Principles for
the Selection of Shape Functions for the Generalized Finite Element Method,
Technical Report #01-16, TICAM, University of Texas at Austin.
12. Babuska, I., Banerjee, U. and Osborn, J.: Survey of Meshless and Gen-
eralized Finite Element Method: A Unified Approach, in preparation.
13. Babuska, I., Banerjee, U. and Osborn, J.: On the Approximability and the
Selection of Particle Shape Functions, in preparation.
14. Buhmann, M. D. (2000): Radial Basis Functions, Acta Numerica, 9, pp. 1-38.
15. Chen, J. S., Wu, C. T. and You, Y. (2001): A Stabilized Conforming Nodal
Integration for Galerkin Meshfree Methods, Int. J. Numer. Meth. Engng., 50,
pp. 435-466.
16. Ciarlet, P. G. (1991): Basic Error Estimates for Elliptic Problems, in Hand-
book of Numerical analysis, Vol. II, Part 1, P. G. Ciarlet and J. L. Lions eds.,
Elsivier Science Publ., pp. 19-351.
17. Daux, C., Moes, N., Dolbow, J., Sukumar, N. and Belytschko, T.
(2001): Arbitrary Branched and Intersecting Cracks with the Extended Finite
Element Method, Int. J. Numer. Meth. Engng., 48, pp. 1741-1760.
18. Dolbow, J. and Belytschko, T. (1999): Numerical Integration of the
Galerkin Weak Form in Meshfree Methods, Comp. Mech., 23, pp. 219-230.
19. Duarte, C. A. and Babuska, I. (2001): Mesh Independent p-Orthotropic En-
richment using Generalized Finite Element Method, Technical Report, TICAM,
University of Texas at Austin.
20. Duarte, C. A., Babuska I., and Oden, J. T. (2000): Generalized Finite El-
ement Methods for three dimensional structural mechanics problems, Computer
and Stuctures, 77, pp. 215-232.
21. Duarte, C. A., Hamzeh O. H.,Liszka, T. J. and Tworzydlo, W. W.
(2001): A Generalized Finite Element Method for the Simulation of Three-
Dimensional Crack Propagation, Comput. Methods Appl. Mech. Engrg., 190,
pp. 2227-2262.
20 I. Babuska, U. Banerjee, and J. E. Osborn

22. Gingold R. A. and Monaghan J. J. (1977): Smoothed Particle Hydrody-

namics: Theory and Application to Non Spherical Stars, Mon. Not. R. astr.
Soc., 181, pp. 375-389.
23. Griebel, M. and Schweitzer, M. A. (2001): A Particle-Partition of
Unity Method, Part II, Efficient Cover Construction and Reliable Integration,
Preprint, University of Bonn.
24. Griebel, M. and Schweitzer, M. A. (2001): A Particle-Partition of Unity
Method, Part III, A Multilevel Solver, Preprint, University of Bonn.
25. Lancaster, P. and Salkauskas, K. (1981): Surfaces Generated by Moving
Least Squares Method, Math. Comp, 37, pp. 141-158.
26. Li, S. and Liu, W. K. (1996): Moving Least Squares Reproducing Kernal
Particle Method, Part II, Fourier Analysis, Comput. Methods Appl. Mech.
Engrg., 139, pp. 159-194.
27. Li, S. and Liu, W. K., (2001): Meshfree and Particle Methods and Their
Application, to appear in Applied Mechanics Review.
28. W. K. Liu, S. Jun, and Y. F. Zhang (1995): Reproducing Kernel Particle
Methods, Int. J. Numer. Meth. Fluids, 20, pp. 1081-1106.
29. Liu, W. K., Li, S. and Belytschko, T. (1997): Moving Least Square Repro-
ducing Kernel Particle Method, Methodology and Convergence, Comput. Meth-
ods Appl. Mech. Engrg., 143, pp. 422-453.
30. Melenk, J. M. and Babuska, I., (1996): The Partition of Unity Finite Ele-
ment Method: Theory and Application, Comput. Methods Appl. Mech. Engrg.,
139, pp. 289-314.
31. Mikhlin, S. G., (1971): The Numerical Performance of Variational Methods,
Walkers-Noordhoff.
32. Nitsche, J., (1970/1971): Uber ein Variationaprinzip zur Losung von
Dirichlet-Problemen bei Verwendung von Teilraumen, die keinen Randbedin-
gungen unterworfen sind, Abh. Math. Univ. Hamburg, 36, pp. 9-15.
33. Stein, E. M., (1970): Singular Integrals and Differentiability Properties of
Functions, Princeton Univ. Press.
34. Stenberg, R., (1995): On Some Techniques for approximating Boundary Con-
ditions in the Finite Element Method, Journal of Computational and Applied
Mathematics, 63, pp. 139-148.
35. Strang, G., (1971): The Finite Element Method and Approximation Theory,
in Numerical Solution of Partial Differential Equations II, SYNSPADE 1970,
B. Hubbard eds., Academic Press, London, pp. 547-584.
36. Strang, G. and Fix, G., (1973): A Fourier Analysis of Finite Element Vari-
ational Method, in Constructive Aspects of Functional analysis, Edizioni Cre-
monese, pp. 795-840.
37. Stroubolis, T., Babuska, I. and Copps, K., (2000): The Design and Anal-
ysis of the Generalized Finite Element Method, Comput. Methods Appl. Mech.
Engrg., 181,pp. 43-69.
38. Stroubolis, T., Copps, K. and Babuska, I., (2001): The Generalized Finite
Element Method, Comput. Methods Appl. Mech. Engrg., 190, pp. 4081-4193.