Comput. Methods Appl. Mech. Engrg.

201204 (2012) 91111

Contents lists available at SciVerse ScienceDirect

Comput. Methods Appl. Mech. Engrg.

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

Stable Generalized Finite Element Method (SGFEM) q

I. Babuka a,1, U. Banerjee b,
a
ICES, University of Texas at Austin, Austin, TX 78712, United States
b
Department of Mathematics, 215 Carnegie, Syracuse University, Syracuse, NY 13244, United States

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

Article history: The Generalized Finite Element Method (GFEM) is a Partition of Unity Method (PUM), where the trial
Received 8 April 2011 space of standard Finite Element Method (FEM) is augmented with non-polynomial shape functions with
Received in revised form 22 August 2011 compact support. These shape functions, which are also known as the enrichments, mimic the local
Accepted 23 September 2011
behavior of the unknown solution of the underlying variational problem. GFEM has been successfully
Available online 1 October 2011
used to solve a variety of problems with complicated features and microstructure. However, the stiffness
matrix of GFEM is badly conditioned (much worse compared to the standard FEM) and there could be a
Keywords:
severe loss of accuracy in the computed solution of the associated linear system. In this paper, we address
Generalized Finite Element Method (GFEM)
Partition of Unity (PU)
this issue and propose a modication of the GFEM, referred to as the Stable GFEM (SGFEM). We show that
Extended Finite Element Method (XFEM) SGFEM retains the excellent convergence properties of GFEM, does not require a ramp-function in the
Approximation presence of blending elements, and the conditioning of the associated stiffness matrix is not worse than
Condition number that of the standard FEM. Moreover, SGFEM is very robust with respect to the parameters of the enrich-
Validation and verication ments. We show these features of SGFEM on several examples.
2011 Elsevier B.V. All rights reserved.

1. Introduction
During the last decade, the Generalized Finite Element Method
(GFEM) and the Extended Finite Element Method (XFEM) two ap-
proaches based on the Partition of Unity Method (PUM) were
developed independently and have been widely used to solve var-
ious types of problems. Only recently, it was clearly recognized
that these two methods are same and were referred to as XFEM/
GFEM [23]. Hence we believe that it is interesting to briey de-
scribe the early development of these methods. It was also recog-
nized that, though these methods have excellent convergence
properties, the stiffness matrices associated with these methods
could be ill-conditioned. In this paper, we especially address this
issue and propose an easy modication, which we call the Stable
Generalized Finite Element Method (SGFEM), that does not have
the above mentioned conditioning problem and is very robust.
We start with a brief history of the early development of the
methods based on PUM. Since this is history and not a survey, it
is important to provide not only the publication date, but in addi-
q
This research was partially supported by IMA, University of Minnesota, Minneap-
olis, MN and J. T. Oden Faculty Fellowship, ICES, University of Texas at Austin, Austin,
TX.
Corresponding author. Tel.: +1 315 443 1460; fax: +1 315 443 1475.
E-mail addresses: babuska@ices.utexas.edu (I. Babuka), banerjee@syr.edu
(U. Banerjee).
1
Tel.: +1 512 471 2156.
tion, the submission dates for various papers, and pay careful
attention to nomenclature for the various methods.
Brief early history: The idea of adding non-polynomial basis
functions into the trial space of the FEM started in 1970s
[10,12,21]. However, these basis functions had global support
and the associated stiffness and mass matrices lost their local
structure.
Three Special FEMs, which used non-polynomial shape func-
tions, were proposed in [5] (1994, sub: Mar.1992) to solve second
order problems with rough coefcients. In particular, the shape
functions used in the Special FEM #3 have compact supports and
are products of piecewise linear FE hat-functions and a non-poly-
nomial function that mimic the special features of the unknown
solution. This idea was further generalized with detailed mathe-
matical theory and applications in the Ph.D. dissertation of Melenk
[32] (1995), where it was shown that the hat-functions could be
replaced by any PU (with compact support). This method was re-
ferred to as PUM and PUFEM in [33] (1996, sub: Apr.1996) and
[6] (1997, sub: Jul.1995), respectively; these papers contain major
results on the method and its application to the problems with
highly oscillatory solutions, problems with solutions with bound-
ary layers, differential equations with rough coefcients, etc.
The PUM was referred to as the GFEM in [48] (2000, sub:
Jul.1998), [49] (2000, sub: Nov.1998), [50] (2001, sub: Jul.2000),
where the hat-functions were used as the PU (similar to the Special
FEM #3 in [5]). In these papers, GFEM is interpreted as an FEM aug-
mented with non-polynomial shape functions with compact sup-
port, and it is shown that the use of only a few of these shape

0045-7825/$ - see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2011.09.012
92 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111
functions is enough to address the problems with singular solu-
tions. Moreover, the idea of obtaining the non-polynomial shape
functions by solving certain local problems is also introduced in
these papers in the context of the analysis of a perforated plate.
In a parallel development, PUM with hat-functions serving as
the PU was also investigated in [9] (1999, sub: Jul.1998) and
[36], (1999, sub: Feb.1999) in the context of crack propagation
problems. This method is similar to GFEM as it also uses the stan-
dard FE trial space augmented with non-polynomial shape func-
tions. However, the major contributions of these papers is to
show that the method does not need remeshing as the crack prop-
agates. Also shape functions with jump discontinuities were used
in these papers. The method was rst referred to as the XFEM in
the Ph.D. thesis of Dolbow [16] (1999) and almost simultaneously
in [51] (2000, sub: Sep. 1999), [13] (2000, sub: Sep.1999], and [15]
(2000, sub: Sep.1999). We mention that XFEM was employed in
[51,13,15] to address crack propagation problems in 3-d, problems
with branched cracks, and fracture in ReissnerMindlin plates.
Another idea similar to the PUM was used in the h-p Cloud
method in [17] (1996, sub: Jun.1995), [18] (1996, sub: Apr.
1996), where the shape functions were the products of a PU and
polynomials. The goal of this method is to obtain h-p FEM like
approximation without using a FE mesh, in the spirit of meshless
methods. The use of the customized function (which mimicked
the exact solution) for crack problems was also suggested in [39]
(1997, sub: Dec.1996), under this framework. Later, the hat-func-
tions were also used as the PU in the h-p Cloud method in [40]
(1998, sub: Dec. 1996).
Lot of work has been done in the area of GFEM and XFEM since
these early work, described above. We will comment on some of
the recent developments near the end of this section.
GFEM and the problem with conditioning: PUM is a exible
framework to design Galerkin methods that accurately approxi-
mate solutions of variational problems. The framework involves
(a) accurately approximating the solution, locally, using functions
in a local approximation space, and (b) gluing the local approxima-
tions, using a PU, to construct a globally conforming approximate
solution. The GFEM, which is a PUM with FE hat functions serving
as the PU, retains the important exibility of choosing the local
approximation space. The efciency of GFEM lies in the fact that
it requires only modifying an existing FE code to incorporate spe-
cial shape functions with compact support. The GFEM, with appro-
priate choice of special shape functions, leads to excellent
convergence properties. However, the use of hat-functions as PU
may result into almost linearly dependent shape functions in
GFEM, and the stiffness matrix could be severely ill-conditioned;
the ill-conditioning could be much worse than the conditioning
of the stiffness matrix of the FEM. This results into the loss of accu-
racy in the solution of the linear system associated with the GFEM.
In fact, the shape functions could be linearly dependent yielding a
singular stiffness matrix.
Various ad hoc approaches have been developed in the litera-
ture to address this issue. For example, the stiffness matrix of
GFEM was perturbed by an identity matrix of size  (small) in
[48,50] and an iterative method was used to solve the perturbed
linear system. Preconditioning of the stiffness matrix, based on do-
main decomposition, have been recently suggested in [34] to ad-
dress the conditioning problem. In [25,44], a at-hat PU
(modied FE hat functions with attened top) was used in the
PUM instead of hat-functions. The use of at-hat PU certainly
avoids the problem of loss of accuracy in the linear system, but it
requires developing a code from the scratch.
Naturally, it is pertinent to ask if GFEM could be modied so
that it retains the excellent convergence properties of the GFEM,
and the loss of accuracy in the computed solution of the linear
system of the modied GFEM is of the same order as that of the
standard FEM. In this paper, we will show that the SGFEM has both
of these features and it does not use a ramp-function for optimal
convergence in the presence of so called blending elements, as re-
quired in the Corrected XFEM for certain problems [22]. We have
chosen a 1-d problem to present the idea of the SGFEM primarily
for the clarity of exposition and not to obscure the analysis with
details that are not directly related to the SGFEM. However, the
ideas and the associated analysis (including the notational machin-
ery) could be easily generalized to higher dimensions and will be
reported in a future publication.
Indicator of the loss of accuracy in computed solution of the
linear system: Consider the linear system Ax b, associated with
FEM, GFEM, or SGFEM, where A is an n  n sparse symmetric posi-
tive denite matrix. Let ^ x be the computed solution of the linear
system, obtained from an elimination method encoded in a linear
algebra package and the computations follow the IEEE standard
for oating point arithmetic (with guard digits). Set
g : kx  ^xk2 =kxk2 the relative error that measures the loss of
accuracy in the computed solution. g depends on the round-off,
but in general, it also depends on the elimination algorithm and
its implementation in the package, the compiler, the processor,
and the computing platform with single or multiple processors. g
is related to the relative error in the approximate solution due to
round-off.
We seek an indicator that reliably indicates the loss of accuracy
in the computed solution, characterized by g, and is practically
independent of other factors mentioned above. Let H DAD,
1=2
where D is a diagonal matrix with Dii Aii . Dene the scaled
condition number KA of A by KA : j2 H, where
j2 H kHk2 kH1 k2 is the condition number of H based on the
k  k2 vector norm. We hypothesize that KA is the indicator, which
we formalize as follows:
Hypothesis H.
g  Cnb KA; b  0; 1:1
where  is the machine precision.
We will elaborate on the precise meaning of the hypothesis and
validate it in the Appendix, borrowing the ideas from the area of
Validation and Verication. The indicator KA will be used to com-
pare various GFEMs with respect to the loss of accuracy in the com-
puted solution, which will allow us to choose a preferable GFEM. In
particular, we will show in this paper that KASGFEM 6 KAGFEM ,
where ASGFEM and AGFEM are the stiffness matrices of SGFEM and
GFEM, respectively, and therefore the SGFEM is preferable over
the GFEM.
Some current work in GFEM/XFEM: These methods have been
used in a variety of applications. For example, XFEM has been used
recently to address two-phase uid ow problems [19], mechanical
behavior of nano-structures [20], and heterogeneous material with
random interfaces [37]); GFEM has been used to address heat trans-
fer problems with sharp thermal gradient [41], grain boundary in
polycrystals [45], and electromagnetic problems [30]. Special shape
functions for problems with locally periodic coefcients are con-
structed in [31] that yield exponential order of convergence. Also
local problems to compute the shape functions for problems with
rough coefcients are constructed in [3], and it has been proved
that GFEM yields exponential order of convergence. For an exten-
sive collection of references in XFEM/GFEM, we refer to [23].
Organization of the paper: In Section 2, we give the model
problem in 1-d. We intentionally chose the problem in 1-d so that
we could communicate the main ideas of SGFEM, when applied to
this problem, in a fairly general fashion, without the notational and
other technical complexity associated with higher dimensions. We
describe the PUM and GFEM, together with the approximation
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 93

results, in Section 3 and show the conditioning problem in GFEM For a parameter 0 < h < 1, let Ih : fi 2Z : 0 6  i 6 Ng, where
h
on an example. In Section 4, we rst describe the SGFEM in a sim- N Nh is an integer. For i 2 Ih , let xhi : ahi ; bi  X such that
pler setting, show that SGFEM retains the convergence properties (i) X [i2Ih xhi , and (ii) any x 2 X belongs to at most j of the open
of GFEM, and establish that the scaled condition numbers of the intervals xhi ; j is independent of i; h. The open interval h
n o xi is called
 
stiffness matrices of the SGFEM and FEM are of the same order. a patch. Subordinate to the cover xhi i2Ih , let N hi be a C 0 PU
i2Ih
We chose the simpler setting primarily to communicate the main satisfying
idea of the method and the associated analysis. We then describe X  
 h
the SGFEM and provide the analysis in full generality. We note that Nhi x 1; 8x 2 X; Ni  6 C;
L1 X
some of the ideas presented here could have been presented in a i2Ih
  
 0 

simpler fashion by using 1-d arguments. However we did not take diam xhi  Nhi  6 C;
this approach; the notations and framework of the analysis, devel- L1 X

oped in this section, could be easily generalized to higher dimen- where C > 0 is independent of i (for details, see [6,33,4]).
sions. In Section 5, we applied SGFEM to three specic examples, On each patch xhi , i 2 Ih , we consider an ni 1-dimensional
namely, interface problems, problems with singular solutions, space V hi the local approximating space, namely
and problems with discontinuous solutions. In the Appendix, we n oni
i;h
discuss the validation of Hypothesis H and present many validation V hi span uj ; ui;h
j
i;h
2 H1 xi and u0 1; 3:2
j0
experiments. We note that the Appendix is a very important part of
i;h
this paper. where ni s are non-negative integers. The functions uj , j > 0, are
carefully chosen such the functions in V hi mimic the exact solution
2. Model problem u, locally in xhi . We will further comment on this issue later. In the
i
rest of the paper, we will write I; xi ; Ni ; V i ; uj in place of
h h h h i;h
Let X 0; 1 and, for an integer k P 0, we denote the standard I ; xi ; N i ; V i , uj , respectively, with an understanding that they de-
Sobolev spaces by Hk X with the norm k  kHk X and seminorm pend on the parameter h. The PUM is precisely (3.1), with the nite
dimensional space S is given by
j  jHk X ; for k 0, H0 X L2 X. We would also use the spaces X n o
i
Hk A, where A is a sub-domain of X. Consider the variational S Ni V i span Ni uj ; 0 6 j 6 ni ; i 2 I : S 1 S 2 ; 3:3
i2I
problem
where
u 2 H1 X; Bu; v Fv ; 8v 2 H1 X; 2:1 ( ) ( )
X ni
XX
i i i i
where S1 f:f y0 u 0 Ni ; S2 f:f yj u j Ni ; 3:4
Z Z i2I i2I j1
Bu; v : au v dx and Fv :
0 0
f v dx 2:2 i i i
X X and y0 ; yj 2 R. The functions uj , j P 1, and the associated spaces V i
R are sometimes referred to as enrichments and enrichment spaces
such that F1 X fdx 0. We assume that the function ax is
respectively in the literature. We will refer to S 1 as the basic part
bounded, i.e., there are constants a; b such that
of S; S 2 will be referred to as the enrichment part of S. Moreover,
we will refer to the Galerkin method with S S 1 as the basic part
0 < a 6 ax 6 b; 8x 2 X 2:3
of PUM. Thus every PUM has a basic part based only on the PU.
We note that ax could be smooth, but it also could be rough. It is We now present the main approximation result of PUM in the
well known that the problem (2.1) has a unique solution, up to an Energy norm (see [6,33,4]).
additive constant.
We dene the Energy norm, kv kEA , of v 2 H1 A, where A is a Theorem 3.1. Suppose u 2 H1 X. Suppose for i 2 I, there exists
sub-domain of X, by ni 2 V i and C 1 > 0, independent of i, such that
Z
ku  ni kL2 xi 6 C 1 diam xi ku  ni kExi and ku  ni kExi 6 i :
kv k2EA : BA v ; v ; where BA w; z : aw0 z0 dx:
A

It is well known that the solution u of (2.1) is also the solution of a

Then there exists v 2 S such that
" #1=2
boundary value problem (BVP), posed in the strong form as X
ku  v kEX 6 C  2
; 3:5
axu0 0 f ; au0 0 au0 1 0 2:4 i2I
i

provided au0 is differentiable. where the positive constant C depends on j; C 1 ; b=a.

It is immediate from Theorem 3.1 that the PUM solution
3. Generalized Finite Element Method (GFEM) uh 2 S S 1 S 2 of (3.1) satises
" #1=2
Let S be a nite dimensional subspace of H1 X. The Ritz-Galer- X
ku  uh kEX 6 inf ku  v kEX 6 C  2
i ; 3:6
kin method to approximate the solution u of (2.1) is given by v 2S
i2I
uh 2 S; Buh ; v Fv ; 8v 2 S: 3:1
where u is the solution of (2.1). It is clear from above that the global
The solution uh is unique up to an additive constant. We can obtain accuracy of the PUM solution uh depends on how accurately the
a unique solution by imposing a natural constraint on uh , namely, solution u of (2.1) can be approximated by the functions in V i , lo-
uh 0 0. cally on the patches xi .
A Partition of Unity method (PUM) is a Ritz-Galerkin method, We mention that in higher dimensions, the patches xi are sub-
where S is constructed employing a (a) Partition of Unity (PU) domains, which can have quite general shape. Theorem 3.1, as pre-
and (b) Local approximating spaces. A Generalized Finite Element sented above, is also true is higher dimensions.
method (GFEM) is a PUM with special PU. We rst briey described We now describe the GFEM. Recall that the choice of PU in PUM
the PUM. is arbitrary. The GFEM is a PUM, where (a) the patches xi are FE
94 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

stars relative to a nite element (FE) triangulation of X, and (b) (see [47,21,11,12,10,42]), where the standard nite element trial
the piecewise linear FE hat-functions N i , associated with the verti- space is augmented by the global function sx. Moreover, we note
ces of FE triangulation, serve as the PU. that ni in (3.2) could be different for different values of i. In fact, one
Let N 1=h and recalling I fi : 0 6 i 6 Ng, let T : may use ni > 0 only for some of the patches xi , as needed for
fxi ih : i 2 Ig. Let fsk gk2Inf0g be the uniform mesh on X, where accuracy, based on the available information; for other patches,
o ni 0, i.e., V i spanf1g. This idea was also discussed and imple-
sk : xk1 ; xk  are the elements; sk : xk1 ; xk is the interior of sk .
mented in the original GFEM papers [48,49].
The points xi are called the vertices of the mesh. The patches
fxi gi2I are dened as xi : xi1 ; xi1 , i 1; 2; . . . ; N  1; also 3.1. Scaled condition number of the stiffness matrix of GFEM
x0 : x0 ; x1 and xN : xN1 ; xN . For i 2 I, let Ni be the standard
hat-functions associated with the vertex xi ; the support of N i is The stiffness matrix A of the GFEM is positive semi-denite.
x i . Note that x  0 s1 ; x
 N sN and x  i si [ si1 for Even when the GFEM solution is naturally constrained with
i 1; 2; . . . ; N  1. x
 i is the FE star associated with the vertex xi . uh 0 0, i.e., when A is positive denite, the condition number
Clearly, fN i gi2I form a PU subordinate to the patches fxi gi2I . The j2 A can be extremely large, specically larger than the condition
2
associated GFEM is the Galerkin method (3.1) with S S 1 S 2 number of standard FE stiffness matrix, which is Oh for second
(see (3.3)). Clearly S 1 is the standard FE space of piecewise linear order problems. However, according to Hypothesis H, the scaled
functions, and consequently, the basic part of the GFEM is the stan- condition number KA j2 H is a reliable indicator of the loss
dard nite element method (FEM). Thus the trial space S of the of accuracy in the computed solution of Ax b. Recall H DAD,
GFEM is precisely the standard FE trial space, augmented with where D is a diagonal matrix with Dii A1=2 ii . We now present
the space S 2 . Thus GFEM could be implemented by incorporating an example where KA is much larger than the scaled condition
2
enrichments into an existing FE code. The name GFEM was rst number of the standard FE stiffness matrix, which is again Oh .
a
Suppose ax 1 in (2.2) and let u x with 1=2 < a < 3=2,
used in [48,49] to highlight exactly this point. The description of
GFEM is exactly same in higher dimensions; it is based on the stan- a 1. Note that xa 2 H1 X but xa R H2 X. We consider a GFEM
i
dard FE triangulation of X. with ni 1 and u1 : xa jxi ; i 2 I. From the denition of S, any
v 2 S is of the form
X X
Remark 3.2. We note that we have considered a uniform mesh v x ai Ni x bi Ni xxa ; ai ; bi 2 R: 3:7
i2Inf0g i2I
only for the simplicity of exposition; in fact, the ideas and theory in
this paper could also be presented for locally quasi-uniform meshes, We have set a0 0 to impose the constraint uh 0 0 on the GFEM
i.e., when C 1 6 jsk1 j=jsk j 6 C 2 for k 1; . . . ; N  1, with C 1 ; C 2 > 0 solution uh . It can be easily shown that uh u, i.e. there is no approx-
independent of k. h imation error.
We let g : a1 ; . . . ; aN ; b0 ; . . . ; bN T 2 R2N1 . Then Bv ; v gT Ag,
The accuracy of the GFEM (also PUM) solution depends on the where A is the 2N 1  2N 1 positive denite stiffness matrix
choice of V i , as mentioned before (see Theorem 3.1). The func- of the GFEM. We note that Aii jN i j2H1 xi for 1 6 i 6 N and
i
tions uj 2 V i (see (3.2), (3.3)) are carefully chosen based on the AN1j;N1j jN j xa j2H1 xj for 0 6 j 6 N. Therefore by considering
available information on the unknown solution u of (2.1) to mi- v 2 S of the form
mic the unknown solution locally in xi . Examples of V i , suitable X X
Ni x Ni xxa
for specic applications are available in the literature (e.g., see v x ai bi ; ai ; bi 2 R; 3:8
jNi jH1 X jNi xa jH1 X
[4]). We briey mention some of the examples that we will con- i2Inf0g i2I

sider in this paper:

it is easy to see that Bv ; v gT Hg, where H is as mentioned before.
i We consider a v 2 S of the form (3.8) with ai 0 for
 If the unknown solution u is smooth in xi , then the uj s are usu-
1 6 i 6 N  1, aN 1, and bi 0 for i 2 I. Then
ally chosen to be polynomials in P j xi and the associated Z
spaces V i are spaces of polynomials of degree ni . We note that
X X 2
Bv ; v v 0 2 dx 1 and kgk2 : a2i bi 1:
ni could be different for different i, based on the available infor- X i2Inf0g i2I
mation on u.
 When ax is a piecewise smooth and discontinuous h i0 function Therefore,
i i
(interface problems), uj s are chosen such that a uj is smooth Bv ; v
i
on xi . Clearly, uj are continuous piecewise smooth functions 1 6 kM ; 3:9
kgk2
with derivatives that are discontinuous at the discontinuities
of ax. where kM is the largest eigenvalue of H.
i
 If the unknown solution u is singular, then uj should be chosen Let gx 2 H2 X be a non-decreasing function with gx 0 for
as singular functions, mimicking the singularity of u. 0 6 x 6 1=4 and 0 < C 6 gxi 6 1 for i P dN=2e. For h small en-
i
 If u is discontinuous at x c in the domain, then uj s are chosen ough, let 1=8 < xk 6 1=4 be the vertex closest to x 1=4. Clearly
to be discontinuous functions on those xi s that contain x c. xa and gxa are in H2 X b , where X
b : 1=8; 1. We now consider a
We note however that problems with discontinuous solutions v 2 S of the form (3.8) with ai gxi xai jNi jH1 X and
cannot be cast as (2.1); we will address these problems in Sec- bi gxi jN i xa jH1 X . Then
tion 5.3 of this paper.
X
N X
N
v x  gxi xai Ni x gxi N i xxa :
ik ik
Remark 3.3. GFEM provides a exible framework to obtain
various Galerkin methods. Many classical methods could be cast Thus v 0 on 0; xk . Moreover, on si , i P k 1, we have
in this framework. For example, with ni 0 for i 2 I in (3.2), GFEM
v js I ih gxa xa I ih g;
(with S S 1 ) yields the classical FEM. Moreover, let sx be a func- i

tion (could be singular) dened on X. Consider ni 1 and where I ih f is the linear interpolant of f on si , interpolating at xi1
ui1 sxjxi in the denition of S in (3.4). Then GFEM, with and xi . Therefore,
i
y1 b (a constant) for i 2 I, yields the classical singular FEM
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 95

 
jv jH1 so gxa  I ih gxa  gxa xa I ih gH1 so i
I xi uj is the piecewise linear interpolant of uj on the patch xi .
i
i i
h i We adjust the operators I x0 and I x1 ; they interpolate at fx0 ; x1 g
a a
6 Ch jgx jH2 so kx kW 1;1 so jgjH2 so ;
i i i and fxN1 ; xN g respectively. We dene a modied local approxima-
n o2
where we have used standard interpolation estimates. Thus recall-  i  i
tion space V i span u j , associated with V i . Clearly, u j 0
ing that v 0 on 0; xk , we have j0
n o
X
N  for j 0; 1 and thus V i span u
i
2 .
2
Bv ; v jv j2H1 X jv j2H1 so 6 Ch jgxa j2 2 b kxa k2 1;1 b kgk2 2 b
i H X W X H X It is well known (see [33,50]) that the scaled condition number
ik1
2 a 2
of the stiffness matrix of the GFEM, with V i as the local approxima-
: Ch jjjgx jjj : 3:10 tion spaces, could be extremely large or even unbounded. We will
Also, use the GFEM with V i precisely to address this issue, and show that
X
N X
N the GFEM based on the approximation space
C
kgk2 P gxi 2 jN i xa j2H1 X P gxi 2 jNi xa j2H1 X P 2
; X X
ik idN=2e h S S1 S2 ; with S 1 ai Ni and S 2 Ni V i
i2Inf0g i2I
where we have used that jN i xa j2H1 X P C=h for i P dN=2e. Thus using
(3.10), we have is an SGFEM. Note that v 0 0 for all v 2 S. We have chosen a0 0
in the denition of S 1 to impose the constraint uh 0 0 to obtain a
Bv ; v 4
6 Ch jjjgxa jjj2 ; unique GFEM solution uh 2 S.
kgk2 It is easy to check that the assumptions in Theorem 3.1 hold; in
and hence, 2
fact, there exists ni 2 V i such that ku  ni kExi 6 Ch jujH3 xi . There-
4 a 2
km 6 Ch jjjgx jjj 2
fore it is clear from (3.6) that ku  uh kEX Oh , where uh is the
P
where km is the smallest eigenvalue of H. Finally, from (3.9), we get GFEM solution, based on S S 1 S 2 (recall S 2 i2I N i V i ). We
kM Ch
4 rst show that the GFEM based on S S 1 S 2 also yields the same
KA j2 H P ; optimal order of convergence.
km jjjgxa jjj2
which is much bigger than the scaled condition number of the stiff- Proposition 4.1. There exists a v 2 S S 1 S 2 , such that
ness matrix of the standard FEM; we recall that the standard FEM is
2
basic part of the GFEM. Thus from Hypothesis H, there will be se- ku  v kEX 6 Ch jujH3 X ;
vere loss of accuracy in the computed solution of Ax b. We will
where the positive constant C independent of h.
show this feature in the Appendix.
It is interesting to note that using v 2 S of the form (3.7) and
following the same arguments as before, we can also show that Proof. Since u 2 H3 xi for 0 6 i 6 N, it is well known that there
4
the condition number j2 A P Ch =jjjgxa jjj2 . We stated this prop- exists ni 2 V i P 2 xi such that
erty at the beginning of this subsection. 2
ku  ni kExi 6 Ch jujH3 xi : 4:1
4. Stable Generalized Finite Element Method (SGFEM) P
Let I h u i2I uxi N i . It is clear that I h u I xi u on xi . Now from the
Poincar inequality, we have
A GFEM will be referred to as an SGFEM if the GFEM satises the
following property: the scaled condition number KA of the asso- ku  I h u  ni  I xi ni kL2 xi 6 Cdiamxi ku  I h u
ciated stiffness matrix A is of the same order with respect to h as of  ni  I xi ni kExi :
the stiffness matrix of the basic part of the GFEM. Since the basic
part of any GFEM is the standard FEM, therefore a GFEM is an Also using the boundedness of the interpolation operator and (4.1),
2
SGFEM provided KA Oh for second order problems. As we get
mentioned before, we will present the analysis for uniform
2
meshes. However, the analysis is valid for locally quasi-uniform ku  I h u  ni  I xi ni kExi 6 Cku  ni kExi 6 Ch jujH3 xi :
meshes.
We rst present a particular example highlighting the ideas and Let w : u  I h u; clearly w 2 H1 X. From above, ni  I xi ni 2 V i
results related to SGFEM in a simpler setting. approximates w locally in xi . Therefore, from the Theorem 3.1,
there is v 2 S 2 such that
4.1. An example of the SGFEM X 4 4
kw  v k2EX 6 C 2 h juj2H3 xi 6 C 2 h juj2H3 X : 4:2
i2I
Let ax 1 in (2.1) and suppose the solution of (2.1) is smooth,
in particular let u 2 H3 X. Since the solution is unique up to an Let v I h u  ux0 v . Since fN i gi2I is a PU, we have
P
additive constant, we seek u with u0 0. It is well known that I h u  ux0 i2Inf0g uxi  ux0 N i 2 S 1 . Thus v 2 S and using
a function in H3 X could be accurately approximated, locally in (4.2), we get
xi , by polynomials of degree 2; recall that the patches xi have 2
been dened in Section 3. Based on this information, we consider ku  v kEX kw  v kEX 6 Ch jujH3 X ;
n o2
i i
V i span uj (i.e., ni 2), where u1 x  xi and which is the desired result. h
j0
i 2 i 2
u x  xi , for 0 6 i 6 N. Recall that u 1. Thus V i P xi .
2 0
Using Proposition 4.1, we immediately get that
We let 2
 ku  uh kEX Oh , where uh is the GFEM solution based on
    X  S S 1 S 2 . We also note that we approximated u  I h u by the
 i
u i i
j : uj  I xi uj
i
; where I xi uj : uij xk Nk  ; functions in V i in the proof of Proposition 4.1 this is the main idea
i16k6i1

xi of SGFEM. Later, we will further comment on this issue.
96 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

We now address the scaled condition number of the stiffness Proposition 4.2. Suppose KA be the scaled condition number of A
matrix A associated with the GFEM based on S S 1 S 2 . With a b 11 , kmax A
and let kmin A b 11 be the smallest and largest eigenvalue of
suitable ordering of the shape function of S, the matrix A is of b 11 , respectively. Then
A
the form b 11
maxf1; C 2 =kmax A
 KA11 6 KA 6 KA11 ;
A11 A12 b 11 g
minf1; C 1 =kmin A
A ; 4:3
A21 A22
where C 1 3=4 and C 2 5=4.
where Aij are block matrices. The matrix A11 fBN i ; Nj gi;j2Inf0g ,
which is the stiffness matrix of the basic part of GFEM, is the stan- Proof. Let z z1 ; z2 T 2 R2N1 , where z1 2 RN and z2 2 RN1 . Then
dard N  N FE stiffness matrix. The N 1  N 1 matrix A22 is of
n o b D1 z1 T A11 D1 z1 D2 z2 T A22 D2 z2 zT A
b Tb
zT Az 1 11 z1 z2 A 22 z2 : 4:8
the form A22 BN i u i  j
2 ; N j u2 . Also A21 AT12 . For the clarity of 
i;j2I 12
T 12
Let z2 y1 ; y2 ; . . . ; yN1 , then D2 z2 m2 y1 ; 2 y2 ; . . . ;
notation, we will write A22 fA22 ij gM i;j1 , where M N 1. Note 1 3
22 yN ; yN1 T , where m2 2h =15. We dene z2;k : yk ; yk1 T , and
that A22 jj are associated with the vertices xj1 , respectively, and
the GFEM solution uh is computed by postprocessing. We remark z2;1 : m21=2 y1 ; 21=2 y2 T ; z2;N : m21=2 21=2 yN ; yN1 T ;
that, in general, M will vary based on the application and A22 jj will z2;k : 2m2 1=2 yk ; yk1 T ; for k 2; . . . ; N  1:
be associated with some vertex xij .
i
We rst note that u  i Recalling that A22 could be obtained from the element matrices A22
2 xj 0 for j i  1; i; i 1. Therefore it is
easy to show that S 1 and S 2 are orthogonal in the inner product through the assembly process, using (4.7) we get,
B; , i.e., X 3 N
b 22 z2 D2 z2 T A22 D2 z2
N
h X
kz2;k k2
k
zT2 A zT2;k A22 z2;k 6
Bv 1 ; v 2 0; 8 v 1 2 S1 ; v 2 2 S2 : 4:4 k1
6 k1
3 N1
Thus it is immediate that A12 and A21 in (4.3) are zero-matrices. h X 5
m1 y2 kz2 k2 :
The matrix A11 is tridiagonal and is constructed by the assembly 6 i1 2 i 4
k
process from the element stiffness matrices A11 , for the element
k
sk xk1 ; xk ; k 1; 2; . . . ; N. The matrices A11 are given by Similarly from (4.7), we also get
 3
k 1 b k b k : 1 1 b 1 : 1: b 22 z2 ;
kz2 k2 6 zT2 A
A11 A ; A ; 2 6 k 6 N; and A 4
h 11 11
1 1 11

4:5 and therefore from (4.8),

b 11 z1 C 1 kz2 k2 6 zT Az
zT1 A b 6 zT A
b 2
Similarly, the matrix A22 , which is also tridiagonal, is constructed by 1 11 z1 C 2 kz2 k : 4:9
the assembly process from the element matrices It is clear from above that
2    3
u k1
u  k1  k
Bsk Nk u  k1 b P zT A
zT Az b 2 b 2 2
k
B
6 sk
Nk1  2 ; Nk1 2 2 ; N k1 u2 7 1 11 z1 C 1 kz2 k P kmin A 11 kz1 k C 1 kz2 k
A22 4     5;
 k1  k k k b 11 gkzk2 ;
P minfC 1 ; kmin A
Bsk Nk1 u 2 ; Nk u 2 Bsk Nk u 2 ; Nk u
2

4:6 where C 1 : 34. Therefore,

R b P minfC 1 ; kmin A
b 11 g kmin A
b 11 minf1; C 1 =kmin A
b 11 g:
for the element sk , k 1; 2; . . . ; N, and Bsk w; v : sk au0 v 0 dx. A di- kmin A
rect computation yields
Similarly from the upper bound of (4.9), we can show that
" #
2 1
k 3 b k b k 15 30 b 6 kmax A
kmax A b 11 maxf1; C 2 =kmax A
b 11 g;
A22 h A 22 ; A 22 1 2
:
30 15
where C 2 54. Thus
b k is positive denite (the eigen-
It is easy to check that the matrix A 22 b b b
1 1
values are 10 and 6) and thus b kmax A 6 kmax A 11 maxf1; C 2 =kmax A 11 g
KA j2 A
b
kmin A b b
kmin A 11 minf1; C 1 =kmin A 11 g
3 3
h h b 11 g
kyk2 6 yT A22 y 6 kyk2 ;
k
8 y y1 ; y2 2 R2 : 4:7 maxf1; C 2 =kmax A
10 6 KA11 ;
b 11 g
minf1; C 1 =kmin A
We now consider the diagonal matrix D diagD1 ; D2 with
b 11 . Thus we have the required upper bound of
where KA11 j2 A
D1 diagd1 ; d1 m1=2
1 21=2 ; . . . ; 21=2 ; 1T 2 RN ; KA.
Now let z1 be an eigenvector of A b 11 associated with kmax A
b 11 .
D2 diagd2 ; d2 m1=2
2 1; 21=2 ; . . . ; 21=2 ; 1T 2 RN1 ;
Also let z2 0. Then from (4.8), we have
3
where m1 1=h and m2 2h =15.
zT Az b 11 kz1 k2 kmax A
b kmax A b 11 kzk2 ;
We next dene
 " #
b 11 and therefore,
b : DAD D1 A11 D1
A
0

A 0
;
0 D2 A22 D2 0 b 22
A b P kmax A
kmax A b 11 :

b 11 D1 A11 D1 and A b 22 D2 A22 D2 . Clearly A

b 11 and A
b 22 are b 11 associated with
Similarly, considering z1 to be an eigenvector of A
where A
b
kmin A 11 and z2 0, we have
N  N and N 1  N 1 tri-diagonal matrices, respectively.
The diagonal elements of A b 11 and Ab 22 are equal to 1. Consequently,
zT Az b 11 kz1 k2 kmin A
b kmin A b 11 kzk2 ;
diagonal elements of A b are equal to 1 and the scaled condition num-
b
ber KA of A is j2 A. and therefore,
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 97

b 6 kmin A
kmin A b 11 : vertices. In Section 4.1, we chose T 2 T . We will present examples
with f2 N 1 (i.e., only few patches enriched) later in the paper.
Now,
b
kmax A b 11
kmax A Remark 4.6. The sets T 1 ; T 2  T provide a framework to address
KA P KA11 ; numerical treatment of many applications. Selection of both sets
b
kmin A b
kmin A 11 depends on a priori information on the problem and its solution.
which is the required lower bound of KA. h Selection of T 2 will be apparent from the examples in Section 5.
Suppose T 0  T contains all the vertices xj 2 T , where it is known
The Proposition 4.2 establishes that KA  KA11 , i.e., the a priori that uxj 0. We choose T 1 T n T 0 . Typically, T 1 will
scaled condition numbers of the stiffness matrices for the GFEM not contain any boundary vertex with homogeneous Dirichlet
and the basic part of the GFEM are of same order. Thus the GFEM condition. However T 1 may exclude other vertices in T based on a
P
with S S 1 S 2 is indeed an SGFEM. priori information. For example, let f x 1k0 c k cos2p2k 1x,
ax 1, and suppose it it known that u0 0. Then
Remark 4.3. We note that the orthogonality of the spaces S 1 and u1=4 u3=4 0, and the vertices xj R T 1 if xj 1=4 or
S 2 was essential in proving Proposition 4.2. This property does not xj 3=4. Thus we can accommodate many a priori information in
hold in general. Later we will dene a notion of almost orthog- this framework. Only for simplicity, we have considered T 0 fx0 g
onality of S 1 and S 2 , which will address this issue. h in this section. h

We now consider a GFEM with

X X
Remark 4.4. The inequality (4.7) played an important role in S S1 S2 ai Ni Ni V i : 4:10
obtaining Proposition 4.2. This property depends on the functions xi 2T 1 xi 2T 2
in V i . For general approximation spaces V i , we need an assumption
Note that v 0 0 for all v 2 S. We will show that this GFEM is an
that will be presented later. h
SGFEM, under certain assumptions on the space S 2 , which we will
present later. We mention that T 1 and T 2 are called S 1 and S 2 rel-
Remark 4.5. SGFEM uses V i as the enrichment space, which is a evant vertices, respectively, since the degrees of freedom associated
modication of V i . Other modications of V i have been reported only with these vertices appear in the GFEM.
in different contexts. For example, the shifting modication, We rst present an approximation result for the GFEM with
namely, u i i i S S1 S2 .
j x uj x  uj xi ; j > 0, is used in XFEM in the con-
text of approximation error as well as enforcement the Kronecker
delta property (see [23]). h Theorem 4.7. Let u 2 H1 X be the solution of (2.1). Suppose for each
xi 2 T 2 , there exists 
ni 2 V i and C 1 > 0, independent of i, such that
ku  I xi u  ni kL2 xi 6 C 1 diamxi ku  I xi u  ni kExi ;
4.2. SGFEM and its analysis
and ku  I xi u  
ni kExi 6 i :Then there exists v 2 S S1 S2 such
1 that
We now present the SGFEM for (2.1), with a 2 L X and
0 < a 6 ax 6 b. Moreover, suppose it is a priori known that
( )1=2
X X
u0 0 (we will further comment on a priori information later). ku  v kEX 6 C ku  I xi uk2Exi 2i : 4:11
We consider the uniform mesh fsk gk2Inf0g with the set of vertices xi 2T nT 2 xi 2T 2

T , as described in Section 3. Recall that the hat function N i and

the patch xi is associated with each xi 2 T . We will refer to P
Proof. Let I h u xi 2T uxi N i be the piecewise linear interpolant
fxi1 ; xi ; xi1 g as the vertices of xi ; the vertices of x0 ; xN are
of u. We note that I h u I xi u on xi . Dene w : u  I h u and let
fx0 ; x1 g; fxN1 ; xN g, respectively. Let P
v : xi 2T 2 Ni ni 2 S 2 . Then recalling that fNi gxi 2T is a PU, we have
T 1; T 2  T ; f1 : cardT 1 ; f2 : cardT 2 ; f1 ; f2 6 N 1: X X X X
P w  v Ni w  Ni ni Ni w Ni w  ni :
We dene S 1 xi 2T 1 ai N i ; ai 2 R; T 1 will be referred to as S 1 -relevant xi 2T xi 2T 2 xi 2T 2
xi 2T nT 2
set of vertices. We consider T 1 fxi 2 T : 1 6 i 6 Ng as in the exam-
ple in Section 4.1. For othernchoices Therefore 2
oni of T 1 , we refer to Remark 4.6. 2  2
i  X  X 
For xi 2 T , let V i span uj  H1 xi such that there exists    i 
j0 kw  v k2EX 4
6C  Ni w  
Ni w  n  : 4:12
i
n 2 V i satisfying ku  n k i
o 6 i for all sk  x  i . Clearly, x 2T nT  x 2T 
Esk i 2 EX i 2 EX
ku  ni kExi 6 2i . We consider the modied space
n oni We rst address the last term of (4.12). Using the fact that x 2 X is
i
V i span u j , where in at most two patches xi , xi1 , we see that the sum
j1 P i 0
 xi 2T 2 N i w  n  has at most two terms
  for any x 2 X. Using this
X  observation, the assumption that N 0i L1 X 6 Cdiamfxi g1 , and
 i
u i i
j uj  I xi uj ;
i
I xi uj : uij xk Nk  : the hypothesis of the Theorem, we can show that
i16k6i1

xi  2 " #
X  X kw  ni k2L2 x X
 
kw  ni k2Exi
1
I xi v is the piecewise linear interpolant of v 2 H xi on the patch  Ni w  ni  6C i

xi based on the vertices of xi ; we adjust I x0 and I xN accord-
x 2T
i 2

EX xi 2T 2 diamfxi g2 xi 2T 2
X X
ingly as before. It is important to note that if for some 6 kw  ni k2Exi 6 2i : 4:13
xi 2 T ; V i fn 2 H1 xi : njs 2 P 1 sk for all sk  x
k
 i g, then V i f0g. xi 2T 2 xi 2T 2

Also ni xk 0 with k i  1; i; i 1 for all 

ni 2 V i . We refer to a patch (We refer to the proof of Theorem 3.2 in [4] for details of the argument
xi as enriched if V i f0g. Let T 2 : fxi 2 T : xi is enrichedg and de- leading to (4.13)). Using exactly same argument and the interpola-
P
ne S 2 xi 2T 2 N i V i ; T 2 will be referred to as the S 2 -relevant set of tion estimate kwkL2 xi ku  I xi ukL2 xi 6 Chku  I xi ukExi , we get
98 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

 2
 X  X are equal to 1.
 
 Ni w 6C ku  I xi uk2Exi : The matrix A22 plays a central role in our analysis and depends
x 2T nT  xi 2T nT 2
i 2 EX on elements that have been enriched. We will refer to an element
Therefore, from (4.12) and (4.13), we have  k1
sk xk1 ; xk  as enriched if (a) xk1 2 T 2 and u 1 js X 0, or (b) k

" #  k
xk 2 T 2 and u 1 jsk X 0. Let
X X
kw  v k2EX 6 C ku  I xi uk2Exi 2i : Kenr : fsk : sk is enrichedg:
xi 2T nT 2 xi 2T 2
The matrix A22 is constructed by the assembly process using the ele-
Finally, writing w u  I h u and setting v I h u v 2 S 1 S 2 , we k
ment stiffness matrices A22 dened only on sk 2 Kenr .
get the desired result. h k
We now address the structure of the element matrices A22 in
detail and set up some notions and notations that will be used in
We mention that unlike in Proposition 4.1, we did not assume the analysis. We denote the vertices of the element sk as
T 2 T in Theorem 4.7. We further note that I h u for u 2 H1 X is k k
x1 : xk1 and x2 : xk ; we consider only sk 2 Kenr . The element
not dened in higher dimensions, since the point values of u, in k
stiffness matrix A22 is of the form
general, do not exist in higher dimensions (in contrast to 1-d). " #
k k
However, using a generalized interpolant based on the average of k b11 b12
A22 ; 4:16
u in a ball around the vertices xi , the proof of the above result k
b12 b22
k

can be easily generalized to higher dimensions.

where
Remark 4.8. From the proof of Theorem 4.7, it is clear that  
k
 k2i
bij Bsk Nk2i u 1
 k2j
; Nk2j u 1 ; 1 6 i; j 6 2:
accurate local approximation of u  I h u by functions in V i is crucial
to obtain the desired result. This is the main idea of SGFEM the k k k
If b11 ; b22 > 0, then A22 is 2  2 and we say that the local stiffness
spaces V i are constructed such that the functions in V i accurately k k
matrix A22 is associated with the vertices x1 xn
k
k1 and o x2 xk .
approximate u  I h u in xi . This is in contrast to the standard We dene a diagonal matrix Dk diag d1 ; d2
k k
with
GFEM, where the functions in local approximating spaces V i k k
d1 ; d2 > 0, such that the diagonal elements of
accurately approximate u in xi . h
b k : Dk Ak Dk
A 22 22

Remark 4.9. We note that V i f0g for xi 2 T n T 2 . If u 2 H1 X is are of equal to 1 or O1, independent of h. Clearly, d1 ; d2 are asso-
k k

locally smooth, namely, u 2 H2 xi for xi 2 T n T 2 , then ciated with vertices xk1 ; xk respectively.
k
ku  I xi ukExi 6 ChjujH2 xi for xi 2 T n T 2 and (4.11) could be writ- On the other hand, if b22 0 (i.e., u  k
1 jsk
0 and consequently
ten as k k
b12 b21 0) in (4.16), then the local stiffness matrix
( )1=2 h i
k k
2
X X A22 b11 is of size 1  1 and is associated only with the vertex
ku  v kEX 6 C h juj2H2 xi 2i : 4:14 h i n o1=2
k k k k k
xi 2T nT 2 xi 2T 2 x1 xk1 . We dene Dk d1 , where d1 b11 ; d1 is
k k
By incorporating the available information on the solution u in V i , associated with the vertex x1 xk1 . Similarly, if b11 0 in
for xi 2 T 2 , we can have i Oh, and consequently, (4.16), then the local stiffness matrix
k
A22
k
is associated only
b22 
ku  v kEX Oh. We note that blending elements (an element h i n o1=2
k k k k k
with at least one vertex in T 2 and at least one vertex in T n T 2 ) with the vertex x2 xk . Also D d2 with d1 b22
are present in this situation, and the optimal convergence of SGFEM associated with the vertex
k
x2 xk . Let 1k be the number of verti-
(i.e., the GFEM with S S 1 S 2 ) is not affected by the presence of ces associated with the local stiffness matrix A22 . Thus the size of
k

these elements. The set T 2 can be chosen adaptively with respect to k

A22 is 1k  1k ; note that 1k is either 1 or 2 with our assumption
a prescribed tolerance, which we do not elaborate in this paper. h
ni 1.
Recall that A22 is obtained by the assembly process using the
k
Remark 4.10. A rate of convergence of Oh for various problems element stiffness matrices A22 ; the size of A22 is f2  f2 . Let
have been reported for the XFEM and the Corrected XFEM (which c c1 ; c2 ; . . . ; cf2 , then
is also a GFEM); see e.g., [22]. However, for the interface problems, X
ck T A22 ck ;
k
the enrichment spaces V i in XFEM requires the use of a ramp-func- cT A22 c 4:17
sk 2Kenr
tion to obtain the Oh rate of convergence. In contrast, the SGFEM
does not require the use of a ramp-function to obtain the rate of k
where ck 2 R1 . Moreover, the components of ck are also the
convergence of Oh in such situations. components of c that correspond to those vertices of sk that are
We now address the scaled condition number of the stiffness k k k k
associated with A22 . For example, if b11 ; b22 > 0 in A22 , then as men-
matrix of the GFEM with S S 1 S 2 . For clarity of the exposition, k k
tioned before, the vertices x1 xk1 ; x2 xk are associated with
we will present the analysis for the case when ni 1 i.e.,
k
V i spanfu i A22 . Suppose the components cjk1 ; cjk of c are associated with
1 g . The analysis for general ni is similar.
As in the example presented in Section 4.1, the stiffness matrix the vertices xk1 ; xk , respectively, of sk . Then ck cjk1 ; cjk T . Sim-
 h i
k k k k
A11 A12 ilarly, if A22 b11 , then A22 is associated with x1 and ck cjk1 
A is of the form A , where A11 fBN i ; N j gxi ;xj 2T 1 is
A21 A22
a vector with one component. Later in our analysis, we will use
the f1  f1 stiffness matrix of the basic part of the GFEM. Let D1
(4.17) with a particular vector c and ck as dened above.
be a diagonal matrix with D1 ii A11 1=2
ii . Clearly, the diagonal Next we note that each vertex xi of the FE mesh is associated
elements of with a FE star union of all elements sk  x  i (equivalently, union
of all elements sk with common vertex xi ). For xi 2 T 2 , we dene
b 11 : D1 A11 D1
A 4:15 Ki : fsk 2 Kenr : sk  x
 i g. For xi 2 T 2 and sk 2 Ki , we set the index
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 99

1 6 li; k 6 2 as follows. We rst note that k 2 fi; i 1g. For k i, where A, A11 and A22 are matrices dened before.
we set li; k li; i 2 and for k i 1, we set li; k li; i 1
k k
1. Thus li; k is the index such that xli;k xi ; note xli;k may P
k Proof. Let n ni xi 2T 1 and g gi xi 2T 2 . Consider v 1 xi 2T 1 ni N i
not be associated with A22 . We dene P i
n o 2 S 1 and v 2 xi 2T 2 gi N i u  1 2 S 2 . Then Bv 1 v 2 ; v 1 v 2
Ki :
k
sk 2 Ki : xli;k k
is associated with A22 : xT Ax, Bv 1 ; v 1 nT A11 n, and Bv 2 ; v 2 gT A22 g. The desired result
 is now immediate from Assumption 1. h
 i 
Thus Ki is the set of sk 2 Ki such that u 1 X 0. For xi 2 T 2 , we Theorem 4.12. Suppose the Assumptions 13 are satised. Let A be
sk
dene the stiffness matrix of the GFEM with S S 1 S 2 . Then

Xh i2 L1 b 11 g
U 1 maxf1; U 2 U 3 =kmax A
Di : k
dli;k ; 4:18 KA11 6 KA 6 KA11 ;
U1 L1 minf1; L2 L3 =kmin Ab 11 g
sk 2Ki
b 11 , kmax A
where kmin A b are the smallest and largest eigenvalues,
which will be used later in our analysis. b 11 dened before.
respectively, of the matrix A
Each diagonal element of A22 is associated with a vertex in T 2 .
Let A22 ji ji be associated with xi 2 T 2 . Moreover, we note that
P k k Remark 4.13. This result shows that under the Assumptions 13,
A22 ji ji sk 2K bli;k;li;k , where bpq was dened in (4.16). Thus
i
the scaled condition numbers of the stiffness matrices of the GFEM
A22 ji ji > 0 for all xi 2 T 2 (i.e., all the diagonal elements of A22
with S S 1 S 2 and the basic part of the GFEM are of the same
are positive). We now dene the diagonal matrix D2 diagfd1 ; d2 order. Thus the GFEM with S S 1 S 2 is indeed an SGFEM.
1=2 1=2
; . . . ; df2 g with dj A22 jj ; 1 6 j 6 f2 . Note that dji A22 ji ji is
associated with xi 2 T 2 . Clearly, the diagonal elements of
Proof. Let z z1 ; z2 T 2 Rf1 f2 , where z1 2 Rf1 and z2 2 Rf2 . Then
b 22 : D2 A22 D2 from the denition of b
A (see (4.20)), we have
A 4:19
b zT DADz DzT ADz, and since Dz D1 z1 T ; D2 z2 T T ,
zT Az
are equal to 1. Dene the diagonal matrix D : diagfD1 ; D2 g. Since from Lemma 4.11 we get
the diagonal elements of Ab 11 ; A
b 22 (see (4.15) and (4.19)) are equal
to 1, the diagonal elements of
b
L1 D1 z1 T A11 D1 z1 D2 z2 T A22 D2 z2  6 zT Az
" #
b 11
A b 12
A 6 U 1 D1 z1 T A11 D1 z1 D2 z2 T A22 D2 z2 : 4:21
b : DAD
A 4:20
b 21
A b 22
A Let z2 f1 ; f2 ; . . . ; ff2 T and consider D2 diagd1 ; d2 ; . . . ; df2 with
b 12 D1 A12 D2 and A
are also equal to 1. Also A b 21 A
bT . di A22 1=2
ii as dened before. Then D2 z2 d1 f1 ; d2 f2 ; . . . ; df2 ff2 T .
12
We will show that the GFEM with S S 1 S 2 is an SGFEM, un- Recall that dji is associated with xi 2 T 2 . Consequently, dji fji is asso-
der the following assumptions on the local approximation spaces ciated with xi 2 T 2 .
V i and the enrichment part of S, namely S2 . Consider an element sk 2 Kenr . Following the notation given after
k k k
z2 : D2 z2 k 2 R1 such that the components of 
(4.17), let  z2 are
Assumption 1. The spaces S 1 and S 2 are almost orthogonal with the components of D2 z2 corresponding to the vertices of sk associ-
k
respect to the inner product B:, i.e., there exist constants ated with A22 . Now from (4.17) and using Assumption 2, we have
0 < L1 ; U 1 < 1, independent of h, such that X X  
T  k 1 k 2
D2 z2 T A22 D2 z2
k k
n o zk2 A22 z2 P L2 D  z2  : 4:22
L1 kv 1 k2EX kv 2 k2EX 6 jBv 1 v 2 ; v 1 v 2 j sk 2Kenr sk 2Kenr
n o n o
k k
6 U 1 kv 1 k2EX kv 2 k2EX ; We note that if Dk diag d1 ; d2 , then
  h i2 h i2
 k 1 k 2
djk1 2 fjk1 2 d2 djk 2 fjk 2 ;
k k
for all v 1 2 S 1 and v 2 2 S 2 . D  z2  d1

where z k2 djk1 fjk1 ; djk fjk T following the notation given after
Assumption 2. For sk 2 Kenr , there exist constants 0 < L2 ; U 2 < 1, h i  2 h i2
 
djk1 2
k k
independent of k and h such that (4.17). Similarly, if Dk d1 , then Dk 1  zk2  d1
h i  2 h i2
L2 kDk 1 xk2 6 xT A22 x 6 U 2 kDk 1 xk2 ;
k
8x 2 R 1 ;
k  
fjk1 2 , and if Dk d2 , then Dk 1  djk 2 fjk 2 .
k k
zk2  d2
where the diagonal matrices Dk have been dened before. Now, it is important to note that if J 1 : fdjk1 fjk1 ;
djk fjk gsk 2Kenr (where the repeated elements appear only once)
and J 2 : fdji fji gxi 2T 2 , then J 1 J 2 . Thus from (4.22), we have
Assumption 3. For xi 2 T 2 , there exist constants 0 < L3 ; U 3 < 1,
independent of i and h such that X 2
D2 z2 T A22 D2 z2 P L2 Di dji fj2i P L2 L3 kz2 k2 ; 4:23
xi 2T 2
L3 6 A22 1
ji ji Di 6 U3 ; where we used Assumption 3 to get the last inequality. Similarly,
where A22 ji ji is the diagonal element of A22 associated with xi , and we can show that
Di is as dened in (4.18). D2 z2 T A22 D2 z2 6 U 2 U 3 kz2 k2 :
The following result is an easy consequence of Assumption 1. b 11 , we get
Therefore from (4.21) and using the denition of A
h i h i
b 11 z1 L2 L3 kz2 k2 6 zT Az
L1 zT1 A b 6 U 1 zT A
b 2
Lemma 4.11. Let x nT ; gT T 2 Rf1 f2 where n 2 Rf1 and g 2 Rf2 . 1 11 z1 U 2 U 3 kz2 k : 4:24

Then there exist positive constants L1 and U 1 , independent of h, such b in (4.24), we have
Now from the lower bound of zT Az
that b 11 kz1 k2 L2 L3 kz2 k2 ;
b P L1 kmin A
zT Az
T T T T T
L1 n A11 n g A22 g 6 x Ax 6 U 1 n A11 n g A22 g;
and therefore,
100 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

b P L1 kmin A
kmin A b 11 minf1; L2 L3 =kmin A
b 11 g: 4:25 Remark 4.16. As shown in the Appendix, the implementation of
the SGFEM does not require scaling the stiffness matrix, i.e., the lin-
b in (4.24), we can show
Similarly, using the upper bound of zT Az ear system involving the stiffness matrix A, and not scaled version
b 6 U 1 kmax A
b 11 maxf1; U 2 U 3 =kmax A
b 11 g: b is solved. The scaling was used only to dene KA and to study
A,
kmax A 4:26
its order through Theorem 4.12. We will show in the Appendix that
Thus from (4.25) and (4.26), we have KA is an indicator of the loss of accuracy in the computed solution
b b 11 g of the linear system associated with FEM, GFEM, and SGFEM. h
kmax A U 1 maxf1; U 2 U 3 =kmax A
KA 6 KA11 ; 4:27
b
kmin A L1 minf1; L2 L3 =kmin Ab 11 g
5. Applications
which the required upper bound. The required lower bound could
be obtained by following the exact arguments in Proposition 4.2 In this section we will present the SGFEM, when applied to
and (4.24). Thus we get the desired result. h three specic applications. We will primarily address in detail
the scaled condition number of the stiffness matrix of the method
We mention that the notions and notations developed leading and show that the assumptions presented in the last section hold.
to Theorem 4.12 can also be extended to higher dimensions. An The SGFEM, applied to each of these applications, will based on the
element will have ne vertices, e.g., ne could be 3 or 4 in 2-d. And uniform mesh fsk gk2Inf0g with the set of vertices T , dened before.
k
the element stiffness matrices A22 could be at most ne  ne . The
assembly argument (4.17) could be easily generalized to higher 5.1. Interface problems
dimensions. For a given vertex xi and an enriched element sk in
the FE star associated with xi , the index li; k will again represent Let ax in (2.1) be a piecewise constant function and let f be
k
the local index of the vertex xli;k of sk that coincides with xi , i.e., smooth. We will consider two situations, namely, ax a1 x
k
xli;k xi . The expressions for Di , A22 ii and the Assumptions 13, and ax a2 x, where
are exactly same in higher dimensions. Using these notions, the 8 
(
proof of Theorem 4.12 can be easily extended to higher dimen- 1
; 06x<b
 < 1; 0 6 x < b1
>
2  
sions. The approach presented here can also be extended for elas- a1 x  and a2 x 12 ; b1 6 x < b2
1; b 6 x 6 1 >
: 
ticity equations etc. We note however, the notations become a 1; b2 6 x 6 1
little more involved if ni > 1.
We note that the solution u of (2.1) does not belong to H2 X.
Remark 4.14. We now make comments on the assumptions. The We rst consider ax a1 x. We consider the set T 2  T as
R o
Assumption 1 is always satised in 1-d. Let B0 u; v : X u0 v 0 dx. before. There exists an m such that b 2 sm1 xm ; xm1 and

i
Since u  j xk 0 for k i  1; i; i 1, it can be easily shown that therefore,

b 2 xm \ xm1 . For xi 2 T , we consider
n Rx o
B0 v 1 ; v 2 0 for all v 1 2 S 1 and v 2 2 S 2 . Therefore, i
V i span 1; u1 xi1 1=a1 tdt . Clearly, for i m; m 1, we
Bv 1 v 2 ; v 1 v 2 P aB0 v 1 v 2 ; v 1 v 2 aB0 v 1 ; v 1 B0 v 2 ; v 2  have V i spanf1; x  xi1 g. Therefore recalling that
a n o
P kv 1 k2EX kv 2 k2EX : V i span u  i
1 , where u
 i
1 u i
1  I xi
u i
1 , we get V i f0g for
b
i m; m 1. We set T 2 fxm ; xm1 g  T and from the denition
Similarly, we can show that
of S 2 , we have
b X
Bv 1 v 2 ; v 1 v 2 6 kv 1 k2EX kv 2 k2EX ; S2 Ni V i Nm V m Nm1 V m1 :
a xi 2T 2
and thus Assumption 1 is satised with L1 ab and L2 a. In higher b

m
 m 
dimensions, this assumption has to be checked. We further note that u1 is linear on sm and therefore, u 1  0.
sm
Assumption 2 is equivalent to L2 kyk2 6 yT Ab k y 6 U 2 kyk2 for all 
22 m1 
fk b k Similarly, u
1  0. Therefore sm1 is the only enriched ele-
y 2 R . Thus A 22 is uniformly positive denite in k and its sm2
eigenvalues are uniformly bounded. m1
ment, i.e., Kenr fsm1 g, and A22 A22 . Also, we can easily show
It is always possible to choose the diagonal matrix Dk such that  
 m   m1  
that u 1  u  . Let b xm bh with 0 < b < 1. Then
Assumption 3 is satised. For example, it is easy to check that 1
sm1 sm1
Assumption 3 is satised with L3 U 3 1 by choosing Dk from a direct computation, we have
n o   "
 #
k 1=2
k
diag dk1 ; d2
k
with d2 bjj . The Assumption 3 is trivially m1 hb1  b2 32 b  2b2 =3 hb2 1  b2 1 4b=6
A22 : 5:1
satised with L3 U 3 1 when Dk is 1  1. h hb2 1  b2 1 4b=6 hb2 1  b1 2b2 =3
m1
Clearly, A22 is associated with the vertices xm , xm1 . We choose the
Remark 4.15. In many practical situations, there are more than n o
m1 m1
n oni diagonal matrix Dm1 diag d1 ; d2 , where
one enrichments on a patch xi and the enrichments u  i
j ,
j1 m1 1=2 1=2 m1 1=2 1
ni > 1, could be almost linearly dependent. In such situations, d1 h b 1  b1 ; d2 h b 1  b1=2 : 5:2
we need to use a simple orthogonalization process to orthogonal- Then
n oni
ize the set u i
j on the patch xi . We can still prove Theorem b m1 Dm1 Am1 Dm1
j1 A 22 22
4.12, under the assumption that the orthogonal set of enrich- "
3  #
ments are linearly independent on each element si and si1 of 2
b  2b2 =3 b1=2 1  b1=2 1 4b=6
1=2 : 5:3
the patch xi . This assumption holds in many practical situations. b 1  b1=2 1 4b=6 1 2b2 =3
However, using more than one enrichment on a patch in higher b m1 are O1 for all 0 < b < 1. Also the
The diagonal elements of A 22
dimensions is more involved and it will be addressed in a forth- b m1 are k1 2  b=6 and k2 1 b=2. There-
eigenvalues of A 22
coming paper.
fore, recalling Remark 4.14, we have
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 101

1 1 3
kDm1 1 xk2 6 xT A22 x 6 kDm1 1 xk2 ; kDm 1 xk2 6 xT A22 x 6 kDm 1 xk2 :
m1 m
8x 2 R2 ; 5:4 5:5
6 32 32
and hence, Assumption 2 is satised with L2 16 and U 2 1.
Next, the matrix Ab m1 : Dm1 Am1 Dm1 is same as the matrix
We set D2 diagfd1 ; d2 g with di A22 1=2
ii . Clearly, the diago- 22 22

b 22 D2 A22 D2 are equal to 1. Recall that given in (5.3). The diagonal elements of A b m1 are O1 and its
nal elements of A 22

T 2 fxm ; xm1 g and Km Km1 fsm1 g. Therefore, eigenvalues are k1 2  b=6 and k2 1 b=2. Therefore
lm; m 1 1 and lm 1; m 1 2, where the index li; k for 1
kDm1 1 xk2 6 xT A22 x 6 kDm1 1 xk2 ;
m1
xi 2 T 2 and sk 2 Ki was dened just before (4.18). We also have 8x 2 R 2 :
6
Km Km1 fsm1 g. Therefore from (4.18), we have
h i2 h i2 Thus the above inequality together with (5.5) implies that Assump-
m1 m1
Dm d1 and Dm1 d2 . Also the vertices 1
tion 2 is satised with L2 32 and U 2 1 for all 0 < b < 1.
m m1
xm ; xm1 2 T 2 are associated with the diagonal elements The matrix A22 is assembled from the matrices A22 ; A22 and is
A22 jm jm ; A22 jm1 jm1 , respectively, of A22 , where jm 1; jm1 2. It given by
is easy to check that 2 h h
3
16 32
0
1< A22 1
11 Dm ; A22 1
22 Dm1 66 6h 2
3 

hb2 1b2 1
7
A22 6
4 16
h
32
hb1b
3
b  2b2
2 3 2
2b 7
5:
and the Assumption 3 is satised with L3 1 and U 3 6.

hb2 1b2 1
 hb2 1b
0 3 2
2b 3
1 2b2
We have shown in Remark 4.14 that the Assumption 1 is always
satised in 1-d; in this case L1 12 and U 1 2. Therefore, from The-
We choose D2 diagd1 ; d2 ; d3 with di A22 1=2 ii . Clearly the
2
orem 4.12, we have that KA Oh , and thus the GFEM with b 22 : D2 A22 D2 are equal to 1. Consider the
diagonal elements of A
S S 1 S 2 is indeed an SGFEM. We further note that Assumptions vertex xm 2 T 2 . Then Km fsm ; sm1 g and lm; m 2,
13 are satised for any 0 < b < 1, i.e., the constants lm; m 1 1. Also in this case, Ki Ki . Therefore, from (4.18),
L1 ; U 1 ; L2 ; U 2 ; L3 and U 3 are independent of b. Therefore h i2 h i2
m m1
2
KA Oh even when b  0 or b  1, i.e., when the point of dis- we have Dm d2 d1 . Similarly, we can show that
 h i2 h i2
continuity b of a1 x is close to the one of the vertices xi (see also Dm1 d1m and Dm1 d2
m1
. We also note that the vertices
Remark 5.1).
We next consider the (2.1) with ax a2 x. We again choose xm1 ; xm ; xm1 2 T 2 are associated with the diagonal elements
n Rx o A22 jm1 jm1 ; A22 jm jm ; A22 jm1 jm1 , respectively, of A22 , where
i
V i span 1; u1 xi1 1=a2 tdt . If the points of discontinuity
jm1 1; jm 2; jm1 3. An easy calculation yields
   o  o
b1 ; b2 of a2 x are separated, e.g., b1 2 sl and b2 2 sl with
 1 6 A22 1 1 1
11 Dm1 ; A22 22 Dm ; A22 33 Dm1 6 16:
jl  l j P 2, then we can again show that the GFEM with
S S 1 S 2 is an SGFEM, based on the arguments given above. Thus Assumption 3 is satised with L3 1 and U 3 16 for all
 o  o
Suppose there is an m such that b1 2 sm and b2 2 sm1 . More- 0 < b < 1. We have shown before that Assumption 1 is always sat-
 
over, suppose b1 xm1 h=2 and b2 xm bh with 0 < b < 1. ised in 1-d. Therefore from Theorem 4.12, we infer that
  2
Note that b1 is away from the vertices, whereas, b2 could be close KA Oh ; the result is true even when b  0 or b  1. Thus
n o
 i
to either xm (b  0) or xm1 (b  1). As before, let V i span u the GFEM with S S 1 S 2 is indeed an SGFEM.
1 ;
We remark that for ax a1 x or ax a2 x, we can show
clearly, V i f0g for i m  1; m; m 1. Therefore
that there exists  ni 2 V i such that ku  I xi u  
ni kExi Oh for
T 2 fxm1 ; xm ; xm1 g and
each xi 2 T 2 . Thus using the standard interpolation estimates and
X using Theorem 4.7, we have ku  uh kEX Oh, where uh is the
S2 Ni V i :
SGFEM solution.
im1;m;m1
We mention that the enrichments u  i
j used in SGFEM for inter-
 m1
We further note that u  m1
jsm1 u jsm2 0. Also it can be face problems are the same enrichments suggested in [35], and
1 1
thus the SGFEM for interface problems is same as the method pro-
 m1
shown that u 1
 m
jsm u  m  m1 js . Therefore
1 jsm and u1 jsm1 u1 m1 posed in [35]. We note that convergence in the presence of blend-
Kenr fsm ; sm1 g (i.e., sm ; sm1 are the only enriched elements), ing elements was addressed in [35], however, the conditioning of
m
and hence A22 is assembled from local stiffness matrices A22 and the method was not discussed in that paper.
m1
A22 .
From direct computation, we get m1
Remark 5.1. Note that A22 and thus A22 , A degenerate as b ! 0
" # or b ! 1. Let 0 be small, say, 0 1014 . We adjust the imple-
h h
m 16 32
A22 h h
; mentation when b 6 0 or 1  b 6 0 by setting b 0 or
32 32 1  b 0 , respectively. We emphasize that KA is bounded
m1 independently of b. h
and it is associated with the vertices xm1 and xm . The matrix A22
is same as in (5.1) and is associated with xm and xm1 . We choose
 
m
Dm diag d1 ; d2
m m
with d1 d2 h
m 1=2
and Dm1 diag 5.2. Problems with singular solutions
 
m1 m2 m1 m2
d1 ; d2 with d1 ; d2 as given in (5.2). Then Let ax 1 in (2.1) and suppose f x be such that the solution u
" # of (2.1), (2.2) is of the form u xa u0 , where 12 < a < 32, a 1, and
1 1
b m : Dm Am Dm
A 16 32
: u0 is smooth with u0 0 0. Clearly u R H2 X. Let 0 < D < 1 and
22 22 1 1
32 16 set Xl : 0; D, Xr : D; 1. Then u 2 H2 Xr and jujH2 Xr
Clearly the diagonal elements of b m
A
are O1. The eigenvalues of 6 Cjxa j H2 Xr ju0 jH2 Xr . Clearly, jujH2 Xr depends on D and is extre-
22
b m are k1 1=32 and k2 3=32 and therefore (recall Remark 4.14),
A mely large for D  0.
22
102 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111


We consider T 1  T as before. Let T 2 : fxi 2 T : xi \ Xl ;g, Let sk 2 Kenr and set xk1=2 : k  12 h; xk1=2 is the mid-point of
where the patches xi have been dened before. Clearly, sk . We dene

x0 ; x1 2 T 2 . Let k 2 I be the largest index such that xi 2 T 2 for   a2 
 1 
  a2 
0 6 i 6 k  1. For xi 2 T 2 , let Gk jxa 00 xk1=2 j aa  1 k  h :
n o  2 
i i
V i span 1; u1 x  xi ; u2 xa jxi ; 
n o Note that for 1 6 j 6 k  1, sj1 2 Kenr implies sj 2 Kenr , and we
i have
and for xi 2 T n T 2 , let V i span 1; u1 x  xi . Clearly, V i f0g
for xi 2 T n T 2 . For xi 2 T 2 , we have !2a
Gj j 12
Vi
i
spanf  2u ri g 0; 16 6 32a : 5:6
Gj1 j  12
where ri : xa  I xi xa jxi ; recall I xi xa is the piecewise linear poly-
We now obtain a few results, which will be used to establish that
nomial that interpolates xa at the vertices fxi1 ; xi ; xi1 g of xi for the GFEM with S S 1 S 2 is an SGFEM.
i 0, and I x0 xa interpolates xa at fx0 ; x1 g. For an element sk  x i
(with xi 2 T 2 ), we dene rk : xa  Ik xa jsk , where Ik xa 2 P 1 sk Lemma 5.3. For xk 2 Kenr , there exist positive constants C 1 ; C 2 ,
interpolates xa
at xk1 , xk . Clearly, I xi xa Ik xa on sk  x i . It is also independent of k and h but may depend on a, such that
k 0
clear that r  X 0 on sk  x  i.
P
We dene S 2 k 1
i0 N i V i and we consider the GFEM based 3=2
krk 0 kL2 sk 3=2
C 1 h 6 6 C 2 h :
S S 1 S2 . We rst address the convergence of the GFEM solution Gk
uh . It is easy to show that for xi 2 T 2 , there exists 
ni 2 V i such that
ku  I xi u  ni kExi 6 Chju0 jH2 xi : Proof.

Also for xi 2 T n T 2 , from standard interpolation result we have 

(a) First let 2 6 k 6 k and let gx rk 0 x for x 2 sk . Then
a
ku  I xi ukExi 6 ChjujH2 xi 6 Chjx jH2 xi ju0 jH2 xi : max jg 0 xj max jrk 00 xj jaa  1jxk1
a2
x2sk
Therefore, from Theorem 4.7, there exists v 2 S1 S 2 such that  a2 !a2
" #1=2 1 a2 k  1
X X jaa  1j k  h
2 k  12
ku  v kEX 6 Ch ju0 j2H2 xi fjxa j2H2 xi ju0 j2H2 xi g
 2a  2a
xi 2T 2 xi 2T nT 2 k  12 3
h i1=2 Gk 6 Gk : M k : 5:7
6 Ch ju0 j2H2 X jxa j2H2 Xr : k1 2
Similarly,
Thus we have ku  uh kEX 6 Ch, where uh is the GFEM solution; note
 2a  2a
that C depends on jxa jH2 Xr and thus on D. k  12 3
c
We note that Xl is independent of h. However, if D h , c < 1 min jg 0 xj Gk P Gk : mk : 5:8
c 1c
k 4
(i.e., jXl j h ), then one can show that ku  uh kEX Oh . Thus
if we enrich only a xed number of patches in the neighborhood of We next note that g 0 does not change sign in sk and thus g is mono-
R xk
the singularity, we loose the optimal order of convergence. 
tonic in sk . Also, since sk gdx rk  0, there exists a unique
xk1
We now address the scaled condition number of the stiffness

matrix A of the GFEM. We note that the matrix A22 is assembled xk xk1 fh 2 sk with 0 < f < 1 such that g xk 0 and xk is char-


k
from element stiffness matrices A22 for the element sk , where sk acterized by
is enriched. We note that the set of enriched elements is given Z xk Z xk
by Kenr : fsk 2 fsl gl2Inf0g : xk 2 T 2 g. We further note that if jgjdx jgjdx: 5:9
sk 2 Kenr , then sj 2 Kenr for 1 6 j 6 k. Also from the denition of xk1 xk
 
k , it is clear that sk 2 Kenr and sj R Kenr for j P k 1. Now, for

 k
sk 2 Knenr , ok2 k , the matrices A22 are of the form We now obtain bounds on f, independent of k. Since g xk 0, it is
k
A22 blm
k k
; the entries blm are as given by clear from the mean value theorem, (5.7) and (5.8) that
l;m1
Z mk jx  xk j 6 min jg 0 kx  xk j 6 jgxj
k
blm Nk2l r0 Nk2m r0 dx: 6 max jg 0 kx  xk j 6 M k jx  xk j; x 2 sk : 5:10
sk
h i Consequently,
k k
Also,

since xk R T 2 , we have A22 b11 (an 1  1 matrix), where
k
b11 is given by the above expression. 2 Z xk 2
f2 h f2 h
mk 6 jgjdx 6 M k 5:11
k 2 xk1 2
Lemma 5.2. The entries of the matrix A22 are as follows:
Z Z
k 2 k 2 and
b11 N2k1 r0 dx; b22 N2k r0 dx;
sk sk 2 Z 2
Z 1  f2 h xk
1  f2 h
k k 2 mk 6 jgjdx 6 Mk : 5:12
b12 b21 N k1 N k r0 dx: 2 xk 2
sk
The proof is easy and we do not present it here. h Now from (5.9), (5.11), and (5.12), we have

It is clear from above that for sk 2 Kenr and k k , the diagonal


2 Z xk Z 2
k k k f2 h xk
1  f2 h
elements b11 ; b22 > 0 and therefore A22 is associated with xk1 , xk . mk 6 jgjdx jgjdx 6 Mk ;
k k 2 xk1 xk 2
Also b11 > 0 and thus A22 is associated with xk 1 . A simple obser-
 
vation yields that the size of A22 is k  k . and thus
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 103

p
Mk 3=22a=2 Proof.
f 6 p p 2a=2
;
Mk mk 3=2 3=42a=2 
(a) Let 2 6 k 6 k and dene gx rk 0 x for x
2 sk . On sk , we
where we used the denition of mk and M k given in (5.8) and (5.7) have seen in the proof of Lemma 5.3 that g xk 0 where
respectively. Using a similar argument we obtain a lower bound of xk xk1 fh and 0 < fl 6 f 6 fr < 1. We have also seen that
f; we summarize the bounds of f as
mk jx  xk j 6 jgxj 6 M k jx  xk j; 8x 2 s k ; 5:14
fl 6 f 6 fr ; where
where
3=42a=2
fl mk Gk 3=42a ; M k Gk 3=22a :
3=22a=2 3=42a=2
3=22a=2 Let
and fr 2a=2 2a=2
: 5:13
3=2 3=4 1  fr
xk
xk1 fr h h:
2
Finally, from (5.10) and using the denition of M k (given in (5.7)),
we get Then
Z Z xk Z xk 1  fr
jgj2 dx jgj2 dx jgj2 dx jxk  xk j jxk1 fr h h  xk1  fhj
2
sk xk1 x
Z
k
Z 1  fr
xk
2 xk
2 P h: 5:15
6 M 2k xk  x dx M2k x  xk dx 2
xk1 xk
Also from the denition of xk , it is clear that gx 0 in xk ; xk and
3
M 2k h 3 3
thus from (5.14) and (5.15), we have
f 1  f 
3 1  fr
jgxj P mk jxk  xk j P mk h: 5:16
G2 3=222a h h 3 i
3
2
6 k fr 1  fl 3 ;
3 Therefore,
and similarly, we have Z xk Z xk

Z 3
jgj2 jlk j2 dx P jgj2 jlk j2 dx
m2 h xk1 xk
jgj dx P k f3 1  f3 
2
Z
sk 3 2 1  fr 2 xk
P m2k h jlk j2 dx: 5:17
G2k 3=422a h h 3 i 4
3
xk
P fl 1  fr 3 :
3 We make the change of variable y xhxk
 2
. Then
1fr
Thus R xi 2 R1
 k 0  x jlk j dx 1  fr h 0 j~lyj2 dy
r   2 Fy1 ; y2 : k2
 ;
3=2 L s k 3=2 y1 y22 2 y21 y22
C 1 h 6 6 C 2 h ; for 2 6 i 6 N;
Gk
r h i where
2a

where C 1 3=4 f3l 1  fr 3 =3 and C 2 3=22a
r
h i ~ly l x 1  fr hy
k k
f3r 1  fl 3 =3. 2

1  fr 1 fr 1  fr
y1 1  y y2 y :
(b) We now consider k 1. We note that on s1 0; h, 2 2 2

r1 0 x axa1  h
a1
: Thus Fy1 ; y2 is independent of k. We next note that Fy1 ; y2 is a
continuous function and Fby1 ; by2 Fy1 ; y2 . It is well known
By a direct computation, we get that the minimum of Fy1 ; y2 is attained on the compact set
Z 2a1 y21 y22 1. Hence there is a constant C min , independent of k but
h
a  12 h
jr1 0 j2 dx : may depend on fr , such that
0 2a  1
R xk 2
Therefore, 1  fr h x jlj dx
0< C 2min 6 Fy1 ; y2 k2 :
Rh 2 y1 y22
2a1
0
jr1 0 j2 dx a  12 h
2a4 42a Thus from (5.17), we have
G21 2a  1a2 a  12 h 2
Z
h
3
xk
1  fr 3
2
3 

: C : jgj2 jlk j2 dx P C 2min m2k h y1 y22
8
2a  1a2 242a xk1
  2 3

B1 G2k h y21 y22 ;
Hence we get the desired result with C 1 min C 1 ; C  and
C 2 max C 2 ; C  . h where B1 3=422a C 2min 1  fr 3 =8.
2

(b) We now consider k 1. On s1 0; h, we have

a1

gx axa1  h . It is easy to see that g x1 0, where
Lemma 5.4. Suppose sk 2 Kenr and let lk x be a linear function, x1 fh with f fa a1=1a . Since fa is increasing for
dened on sk , such that lk xk1 y1 and lk xk y2 . Then there exists 1
2
< a < 32 (with f redened for a 1), we have
a positive constant C 3 , independent of k and h but may depend on a, f 6 f
f3=2 2=32 .
such that


 k 0  3=2
2 1=2 x1 f h 1  f h=2. Since jgxj is increasing in x1 ; h , we
Set 
r  lk  2 P C 3 Gk h y1 y22 :
L s k have jgxj P gmin
jg
x1 j on 
x1 ; h. Therefore,
104 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

Z Z Z  
h
2 2
h
2 2
h
2 g2min G21 Next, recalling that A22 is k  k , we choose the diagonal matrix
jgj jl1 j dx > jgj jl1 j dx P g2min jl1 j dx 2 2a2
0 x1 x1 a2 ja  1j h =2 2a2
D2 diagd1 ; d2 ; . . . ; dk with dj A22 1=2
jj . Clearly, the diagonal
Z h Z h
2 elements of A b 22 D2 A22 D2 are equal to 1. Note that
jl1 j2 dx C 2 G21 h jl1 j2 dx; 
x1 x1 A22 jj ; 1 6 j 6 k , is associated with the vertex xj1 2 T 2 . Also,
0 j1 j 
where A22 11 b11 and A22 jj b22 b11 for 2 6 j 6 k .
 
h 
i2 Now, for xi 2 T 2 and i 0; k  1 (recall xi R T 2 for k 6 i 6 N),
242a g2min 242a a 1f
2

1 we have Ki fsi ; si1 g; li; i 2; li; i 1 1 and Ki Ki . Also
2
C : K0 fs1 g; l0; 1 1; K0 K0
 
and Kk 1 fsk g; lk  1; k
a2 a  12 h2a1 a2 a  12 
1; Kk 1 Kk 1 . Therefore, from (4.18), we have Di


As before, we can show that h i2 h i2 h i2

i i1  1
d2 d1 for xi 2 T 2 and i 1; k  1; also D0 d1
Z h
1  f h
2 1=2 h  i2
jl1 j2 dx P C min y1 y22 ; and Dk 1 d1
k
.
x1 2
We now show that Assumption 3 is satised.
and therefore,
Z h Proposition 5.6. Let A22 jj , 1 6 j 6 k , be the diagonal elements of

2 3
1=2
jgj2 jl1 j2 dx P B2 G21 h y21 y22 ; A22 and consider Di for xi 2 T 2 , dened above. Then Assumption 3 is
0

 satised.
2
where B2 C 2 C min 1  f h=2. Finally, dening C 3 min B1 ; B2
k 0
and recalling that g r  , we get the desired result. h 
 Proof. Let xi 2 T 2 and i 0; k  1; xi is associated with A22 ji ji ,
Now, for k k , consider the diagonal matrix Dk i1
  where ji i 1. Therefore using the denition of A22 i1;i1 ; d1 ,
k k k
diag d1 ; d2 with d1 d2 G1
k 3=2 b k Dk Ak Dk .
k h and set A i
22 22 and d2 , we have
b k (see Lemma 5.2) are
The diagonal elements of A 3 3
22 G2i h G2i1 h
Z Z A22 1 1
ji ji Di A22 i1;i1 Di : 5:18
k 1 2 k 1 2
i
b22
i1
b11
i
b22
i1
b11
b 11 3
N2k1 r0 dx; b22 3
N2k r0 dx:
G2k h sk G2k h sk
Now using Lemmas 5.4, 5.3, and 5.6, it is immediate that
Using Lemmas 5.4 and 5.3, it is clear that 2
Z 2 b22
i
2 C 3 b11
i1
G2i1 i1
b11 2
1 C 3 6 3
6 C 2 ; 22a
6 3
6 C 2 ;
C 3 6 k
b11 6 3
02
r dx 6 C 2 ; G2i h 3 G2i1 h G2i G2i1 h
3

G2k h sk
and therefore,
where C 2 ; C 3 are independent of k and h. Similarly,
322a
3
Z 1 G2i h
k 1 02 6 6 : 5:19
C 3 6 b22 6 3
r dx 6 C 2 : 2C 2
2 i
b22
i1
b11
2
C 3 1 322a
G2k h sk
h i Similarly, we get

k k 3=2
We let Dk d1 with d1 G1 k h . Using similar arguments
h i
  k 

b k Dk Ak Dk b
k . 3
we show the C 3 6 b11 6 C 2 , where A 22 22 11 1 G2i1 h 1
b k
Thus the diagonal elements of A 22 are O1 for all sk 2 Kenr . 2 22a
6 i i1
6 2
;
k C 2 1 3 b22 b11 2C 2
We next show that the element matrices A22 satisfy the
Assumption 2. 3and combining (5.18), we infer that there exist constants L3 , U 3 ,
such that
Proposition 5.5. For sk 2 Kenr , the matrices Ak
22 satises Assump-
tion 2. L3 6 A22 1
j i j i Di 6 U 3 ;

where

 1 1
Proof. Suppose k k and let x x1 ; x2 T 2 R2 . Then, using L3 2
2
;
Lemma 5.2, we have 2C 2 C 2 1 322a
k k k k 1 322a
xT A22 x b11 x21 2b12 x1 x2 b22 x22 U3 2
2
:
Z h i 2C 3 C 3 1 322a
2
x21 N 2k1 2x1 x2 Nk1 Nk x22 N2k rk 0 dx 
sk Thus Assumption 3 hold for xi 2 T 2 ; i 0; k  1. The proofs for
Z Z 
xi 2 T 2 ; i 0; k  1 are simpler and we do not include them here. h
2
 2
x1 Nk1 x2 Nk 2 rk 0 dx 6 2 x21 x22 rk 0 dx;
sk sk
Based on Propositions 5.5 and 5.6, it is clear that Assumptions 2
and using Lemma 5.3, we have and 3 are satised. Assumption 1 always hold in 1-d. Thus from
2
3 Theorem 4.12, KA Oh and the GFEM with S S 1 S 2 is
xT A22 x 6 C 2 h G2k kxk2 C 2 kDk 1 xk2 :
k
an SFEM.
Next from Lemma 5.4, it is immediate that
Z 5.3. Problems with discontinuous solutions
2 3

x1 Nk1 x2 Nk 2 rk 0 dx P C 1 G2k h y21 y22 C 1 kDk 1 xk2 :
k
xT A22 x
sk
We now address a problem, which is different from (2.1). Let
k
Similar bounds fox A22 x2 for all x 2 R also hold. Thus Assumption 2
X 0; 1 and set Xl 0; c and Xr c; 1, where 0 < c < 1 is
xed. Consider
is satised with L2 C 1 and U 2 C 2 . h
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 105

( Z Z ) Z Z Z Z Z
HX : v 2 L2 X : v 0 v 1 0; v 0 2 dx < 1; and v 0 2 dx < 1 : ~ 00 2 dx
u ~ 00 2 dx
u ~00 2 dx
u u00 2 dx u00 2 dx 6 1:
Xl Xl X Xl Xr Xl Xr


Thus u ~ 2 H2 X and considering s uc uc
2

Hc  1 in (5.22), we
Then HX; k  kH is a Hilbert space, where
get the desired result. h
kv k2H : jv j2H1 Xl jv j2H1 Xr : Suppose c R T , i.e., c is not a vertex of the mesh. Therefore,
there exists an m such that c 2 sm1 and hence, c 2 xm \ xm1 .
We note that H10 X  HX and functions in HX may have jump We assume that m 1; N; this is always achieved for h small. Since
discontinuity at x c. u0 u1 0, we consider T 1 T n fx0 ; xN g (see Remark 4.6).
For f 2 L2 X, we consider the problem For 1 6 i 6 N  1, we consider V i
i
n spanf1; u1 ox  xi ;
u 2 HX; Bu; v Fv ; 8 v 2 HX; 5:20 ui2 Hc xgn and we oset V 0 span u0 1 x  x0 and
N
V N span u1 x  xN . Note that V i 2 Hxi for i 2 I (i.e., for
where xi 2 T ). Clearly, V i f0g for i 2 I n fm; m 1g. We set T 2 fxm ;
Z Z Z xm1 g  T and dene
Bu; v : u0 v 0 dx u0 v 0 dx and Fv : f v dx:
Xl Xr X S 2 Nm V 1m Nm1 V 1m1 :
The bilinear form B;  is coercive and bounded in HX. Also F is We consider the GFEM with S S 1 S 2 .
a bounded linear functional on HX. Thus the problem (5.20) has a m
Since u2 Hc is constant in sm , we have u  m
2 jsm 0. Similarly,
unique solution. m1
u
2 jsm2 0. Therefore Kenr fsm1 g. Moreover, the functions
If f and the solution u 2 HX of (5.20) are smooth in Xl and Xr ,
then u is the solution of the boundary value problem  m
u  m1 are discontinuous at x c, their values are zero at
2 ; u2  
 2m 
x xm ; xm1 , and u  m1
u 2

 .
 u00 f on Xl ; u00 f on Xr ; sm1 sm1
0  0 We assume that f is such that solution u 2 HX of (5.20) satis-
u0 u1 0 and u c u c 0:
es the assumptions of Lemma 5.7 and u s u ~ , where s is a step-
~ 2 H2 X. We now ad-
function with a discontinuity at x c and u
This problem mimics the problem with a crack in higher dimen-
sions, where the solution is discontinuous along the crack line away dress the convergence of the GFEM solution. We rst note that
from the crack-tip. Theorem 4.7 hold for u 2 HX with EX; Exi replaced by
We now give a characterization of the solution of (5.20). We HX; Hxi and with V i 2 Hxi . Now for xi 2 T n T 2 , we have
will use the Heaviside function u 2 H2 xi and from the standard interpolation result

1; 0 6 x < c; ~ j H 2 x :
ku  I xi ukHxi ku  I xi ukExi 6 ChjujH2 xi Chju 5:23
Hc x 5:21 i
1; c 6 x 6 1:
For xm 2 T 2 , it is easy to show that there exists  nm 2 V m such that
u  I xm u  n u  I xm u on xm . Therefore, ku  I xm u 
 m ~ ~ nm kL2 xm
6 Cjxm jku  I xm u   nm kExm , and from the standard interpolation
Lemma 5.7. Suppose u 2 HX such that u0 c u0 c 0 and result, we have

Z Z ku  I xm u  nm kHxm ku
~  I xm u
m
~ j H 2 x :
~kH1 x 6 Chju
m
5:24
00 2 00 2
u dx < 1; u dx < 1:
Xl Xl Similarly, there exists 
nm1 2 V m1 such that
Then
ku  I xm1 u  nm1 kHxm1 6 Chju
~ j H 2 x :
m1
~x;
ux sx u
Therefore combining (5.23), (5.24) and (5.3) and using the Theorem
~ 2 H2 X
where s is a step function with discontinuity at x c and u 4.7 with modications as mentioned above, we infer that there ex-
ists v 2 S S 1 S 2 such that
Proof. We rst note that uc and uc are well dened. We
dene ku  v kHX 6 Chju
~jH2 X :

uc  uc Therefore, ku  uh kHX Oh, where uh is the GFEM solution.

~ u
u Hc  1; 5:22
2 We next address the scaled condition number of the stiffness
where Hc x is given in (5.21). It is easy to check that matrix of the GFEM. Since sm1 is the only element in Kenr , we have
m1
A22 A22 .
u ~ c uc ;
~ jX ujX ; u Let c xm bh with 0 < b < 1. A direct computation yields that
l l

~ 0 jX u0 jX ;
u ~ 0 c u0 c :
and u " #
l l
2
m1 4 4  9b 6b =3  13 2b  2b2
Similarly, A22 :
h  1 2b  2b2 1  3b 6b2 =3
3
~ jXr ujXr uc  uc ; u
u ~ c uc ;
m1
Thus A22 is associated with n
vertices xm and o xm1 . We choose the
~ 0 jXr u0 jXr ;
u ~ 0 c u0 c :
and u m1 m1 m1 m1
diagonal matrix Dm1 diag d1 ; d2 with d1 d2
1=2
Note that u ~ c u~ c uc . We dene u~ c uc so that u ~ is h =2. Then
continuous at x c and thus is continuous on X. Also since " #
u0 c u0 c 0, it is clear from above that u ~ 0 c u
~ 0 c 0. b m1 m1 m1 m1 4  9b 6b2 =3  13 2b  2b2
~ 0 c 0 so that u ~ 0 is continuous at x c, and conse-
A 22 D A22 D :
We dene u  13 2b  2b2 1  3b 6b2 =3
quently u ~ 0 is continuous in X. Moreover,
106 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111

9 9
10 10

10 10
10 10

11 11
10 10


12 12
10 10

13 13
10 10

14
slope=2 14
slope=2
10 10

N N
15 15
10 10
2 3 4 5 2 3 4 5
10 10 10 10 10 10 10 10

(a) (b)
k
Fig. 1. Loglog plots of gk kxk  x xk is the computed solution of Ak xk b ; k 1; 2, associated with FEM with vertices xi ih and
^k k2 =kxk k2 where ^
xi N  ih; i 0; 1; . . . ; N, respectively. g1 , g2 have been computed and presented in (a) and (b), respectively, for N 100; 200; . . . ; 50000.

4 4
10 10

6 6
10 10

8 8
10 10

10 10
10 10

12 12
10 10

slope=2 slope=2
14 14
10 10

16
N 16
N
10 10
1 2 3 4 5 6 7 1 2 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10 10 10 10 10
(a) (b)
140

120

100

80

R
60
e

40

20

20

N
40 1 2 3 4 5 6
10 10 10 10 10 10 10
7

(c)
Fig. 2. (a) Loglog plot of ga kx  ^ xa is the computed solution of Problem 2a (using superLU). (b) Loglog plot of gb kx  ^
xa k2 =kxk2 with respect to N, where ^ xb k2 =kxk2
with respect to N, where ^ xb is the computed solution of Problem 2b (using MUMPS). (c) Semi-log plot of 100  gc  gb =gb with respect to N, where gc kx  ^ xc k2 =kxk2
and ^xc is the computed solution of Problem 1c (using MUMPS with 128 processors). Proportionally distributed 1931 values of N in the interval 10; 107  are used in all the
gures.
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111 107

2 3
10 10

0 4
10 10

2 5
10 10

4 6

10 10

6 7
10 10

8
slope=4 8
10 10

10
10
10
9
slope=4

12
N 10
N
10 10
1 2 3 4 2 3
10 10 10 10 10 10

(a) (b)
Fig. 3. Plots of g kx  ^
xk2 =kxk2 where ^ x is the computed solution of the linear system Ax b of Problem 3, associated with the GFEM with vertices xi ih, i 0; 1; . . . ; N. In
(a), we used N 50; 100; 150; . . . ; 10000, and in (b), we used every value of N in the interval [100, 1000] to show the detail.

9 11
10 10

10
10
12
10
11
10

12 13
10 10

13
10
14
10
14
10
slope=2 slope = 2
15
N 15
N
10 10
1 2 3 4 2 3
10 10 10 10 10 10

(a) (b)
Fig. 4. Plots of g kx  ^
xk2 =kxk2 where ^ x is the computed solution of the linear system Ax b of Problem 4, associated with the SGFEM with vertices xi ih, i 0; 1; . . . ; N. In
(a), we used N 50; 100; 150; . . . ; 10000, and in (b), we used every value of N in the interval [100, 1000] to show the detail.

The diagonal elements of A b m1

are O1 for any 0 < b < 1. The Also xm ; xm1 2 T 2 are associated with A22 ym ym ; A22 ym1 ym1 ,
22

eigenvalues of Ab m1
are k 5
 2b 2b2  T and k2 56  2b respectively, where ym 1; and ym1 2. It is easy to check that
22 1 6
q
2b2 T, where T 16 13  84b 228b2  288b3 144b4 (obtained 3 24
6 A22 1 1
11 Dm ; A22 22 Dm1 6 :
4 5
from MAPLE). It can be shown that
Thus Assumption 3 is satised with L3 3=4 and U 3 24=5. There-
p p 2
1 5 13 1 5 13 fore from Theorem 4.12, we have KA Oh , where A is the
6 k1 6  ; 6 k2 6 : stiffness matrix, for all 0 < b < 1.
6 6 6 2 6 6

Thus, as before, we have

p! 6. Conclusion
1 2 5 13
kD m1 1 T m1
 xk 6 x A22 x 6 kDm1 1 xk2 ; 8 x 2 R2 ; The GFEM uses special enrichment functions, based on the
6 6 6
p available (or extracted) information on the unknown solution of
and the Assumption 2 is satised with L2 16 and U 2 56 13
6
. the underlying variational problem. The use of special enrichment
1=2
We choose D2 diagfd1 ; d2 g with di A22 ii . Clearly, the functions gives rise to the excellent convergence properties of the
b
diagonal elements of A 22 D2 A22 D2 are equal to 1. As in the rst GFEM. In fact, for a given problem, it is possible to choose several
example (i.e., when ax a1 x) in Section 5.1, we have classes of enrichment functions such that the GFEM, employing
T 2 fxm ; xm1 g and Km Km1 fsm1 g. Therefore, lm; m 1 each of these enrichment classes, will yield excellent convergence
1 and lm 1; m 1 2. Also Km Km1 fsm1 g and there- properties. However, GFEM employing some of these classes of
h i2 enrichments could be ill-conditioned, i.e., there could be severe
m1 2 m1
fore from (4.18), we have Dm d1  and Dm1 d2 . loss of accuracy in the computed solution of the linear system asso-
108 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111
ciated with the GFEM. The loss of accuracy could be much more
than that experienced in a standard FEM. In this paper, we have
presented and analyzed a modication of the GFEM the stable
GFEM (SGFEM), which does not have the problem with severe loss
of accuracy. SGFEM has all the advantages of the GFEM, namely, it
has excellent approximation property and does not require
enrichment on all the patches. Moreover, the method does not
use a ramp-function in the presence of blending elements for cer-
tain problems, e.g., interface problems, and is also very robust with
respect to the parameters of the enrichments (e.g., the parameter b
in Sections 5.1 and 5.3). The loss of accuracy is characterized by the
scaled condition number and is expressed through Hypothesis H,
which was validated based on various examples.
The abstract framework developed in this paper has been ap-
plied to a one-dimensional problem for the clarity of exposition.
This framework could also be applied to higher dimensional prob-
lems, which will be reported in a forthcoming paper.
Acknowledgement
We thank Professor C. Armando Duarte of the Department of Ci-
vil and Environmental Engineering, University of Illinois at Urbana-
Champaign, and Professor John E. Osborn of the Department of
Mathematics, University of Maryland at College Park, for fruitful
discussions on certain aspects of this paper. We also thank Mr. Karl
Schulz of PECOS, ICES, University of Texas at Austin, for his help on
the use of superLU and MUMPS on the Lonestar system, Texas Ad-
vanced Computing Center, and also for various illuminating discus-
sions and interpretation of the results, presented in this paper.
Appendix A
Validation and Verication (V & V) is a fairly new eld and is
still in its developing stage [1,43,7,38]. Suppose a mathematical
model of some Reality (e.g., a physical, chemical or biological
system or process), formulated for a particular goal or purpose,
is given. The objective of V & V is to assess whether the predic-
tions based on the computed solution of a mathematical model
are reliable enough so that they could be the basis for certain
decisions related to the goal.
Validation is the process of building condence on the mathe-
matical model [1,43,38]. The process is of course constrained by
the cost, available time, and skills, as explicitly underlined in
[46]. It is based on a set of properly selected problems and their
mathematical models for which experimental data is available.
These problems are called validation problems and they are chosen
with varying level of complexity; more complex problems are clo-
ser to the Reality. Of course, obtaining the experimental data for
the validation problems with increasing complexity is increasingly
costly. The prediction based on the computed solution of these
problems is then compared with the experimental data. The
assessment of the difference is based on a specied tolerance and
a suitably selected metric (could be more than one) relative to
the specic goal. If the measure of the difference is larger than
the tolerance for any validation problem, the mathematical model
is rejected. If none of the validation models are rejected, then one
could have condence that the mathematical model realistically
describes the Reality, with respect to the goal, beyond the scope
of the chosen validation problems. The level of condence will be
based on the tolerance as well as the number and the selection
of the validation problems. We mention that the set of the valida-
tion problems is nite, their selection has a large subjective com-
ponent, and a philosophical question about the justication of
the condence in the mathematical model could certainly be raised
(see [28]).
Numerical algorithms and their properties obtained from the
mathematical analysis are always based on various assumptions
that are not satised when the algorithm is implemented on a
computer. For example, innite precision arithmetic is often as-
sumed while describing a numerical algorithm or stating an
inference about the algorithm obtained from the analysis.
However, this assumption is always violated by a computer
working with nite precision arithmetic. The output from the
computer implementation of the algorithm may also depend,
for example, on the package in which the algorithm have been
implemented, the compiler, the processor, the computing plat-
form with single or multiple processors, among other factors.
Consequently, the output may vary even when the same out-
come is predicted by the mathematical analysis for two different
algorithms. For example, suppose the solution of the linear sys-
tem Ax b is sought using algorithms of the form
PTi APi z PTi b; x Pi z, where Pi is a permutation matrix for
i 1; 2. Both the algorithms should yield that solution x, how-
ever, the computed solutions could be different (see Problems
1a,b, below). Thus the implementation of a numerical algorithm
in a computer is analogous to a Reality; the goal is to obtain a
particular quantity of interest for a particular purpose (related to
a decision). The mathematical model of this Reality is the
inference obtained from the mathematical analysis, or other
statements based on the inference, about obtaining the quantity
of interest from the algorithm. Therefore, the process of valida-
tion of the inference has to be performed to have condence
in the inference or a statement based on the inference.
We have briey formulated the following hypothesis in the
Introduction. Let Ax b; x; b 2 Rn be a linear system, where the
n  n matrix A belongs to a class of sparse matrices that include
the stiffness matrices associated with FEM, GFEM, or SGFEM. Let
^
x be the computed solution of the linear system, obtained from
an elimination method, e.g., some variant of Gaussian elimination.
Moreover, ^ x is computed in nite precision arithmetic with ma-
chine precision . Let H DAD where D is a diagonal matrix with
Dii A1=2
ii ; clearly, Hii 1. Recall the scaled condition number
KA of A is given by KA : j2 H, where j2 H kHk2 kH1 k2 is
the condition number of H based on the k  k2 vector norm. Also re-
call g : kx  ^ xk2 =kxk2 .
Hypothesis H. For n, not small,
g  Cnb KA; b  0; 7:1
where ^x has been computed in a computing environment satisfying
the IEEE standard for oating point arithmetic (with the guard di-
git), there is no overow or underow during the computation of
^x, and C, b do not depend on n as well as other factors mentioned
before.
 small, such
The  in (7.1) means there exist 0 < C 1 ; C 2 and 0 < b
that g Cnb KA with C 1 6 C 6 C 2 and jbj 6 b.  Also this hypothe-
sis addresses the range N for which not (almost) all digits of accu-
racy is lost (see Problem 3a).
Hypothesis H is based on certain mathematical inferences (re-
sults), which we will discuss later. The validation of (7.1) with re-
spect to the tolerance s fs1 ; s2 g means that C 2 =C 1 6 s1 and
 6 s2 . Note that s2 is primary and should be small for condence
b
in (7.1), however s1 could be allowed to be larger. The set of vali-
dation problems consists of stiffness matrices of FEM, GFEM,
SGFEM, and other similar matrices, e.g., arising in nite difference
method, applied to solve various linear elliptic variational prob-
lems of increasing complexity. For condence in the Hypothesis,
we require that (7.1) is not rejected for any of the validation prob-
lems relative to the given tolerance s. We note that it is possible to
select a tolerance such that the hypothesis is not rejected, however,
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111
the tolerance have to be admissible (e.g., reasonably small) for the
decision making process. In our case, the decision will be whether
to accept the SGFEM over the standard GFEM. We note that the
class of matrices for which the hypothesis will be validated is not
precisely dened, similar to a class of complex physical or engi-
neering problem.
We now give a theoretical rationale for (7.1). There is a lot of
literature available on the accuracy of the computed solutions of
the linear system Ax b. We particularly mention the classic [52]
and a modern book [26] with an excellent survey of the theoret-
ical results in the area. Typically, the loss of accuracy in the
numerical solution due to round-offs is analyzed by the backward
error analysis. This analysis shows that the computed solution is
the exact solution of a perturbed linear system, and it provides
estimates of the perturbations in terms of the data of the linear
system. A bound on the loss of accuracy in the computed solu-
tion, measured by g dened before, is then obtained using the
perturbation estimates.
It is well known from a standard perturbation argument that for
a full matrix A,
g 6 f nj2 A; 7:2
where  is the machine precision and f n depends on the algo-
rithm used to solve Ax b (see e.g., [24,27]). In Hypothesis H,
we hypothesize that j2 A is replaced by KA. We also hypoth-
 where C 1 , C 2 , and b
esize that C 1 nb 6 f n 6 C 3 nb and jbj 6 b,  are
as dened before. It is important to note that in the mathemati-
cal literature, only an upper bound of g is available; in contrast,
the Hypothesis H addresses both the upper and lower bounds of
g.
Consider the linear system DADz Db, where D is a diagonal
matrix with Dii 2gi in the rage of the oating point system.
Clearly x Dz. We now cite the following old result of Bauer [8]:
Theorem 7.1. Let ^x, ^z be the computed solutions of the linear systems
Ax b and DADz Db, respectively, obtained from an elimination
method with no pivoting. Furthermore, we assume that there is no
overow or underow in the computation of ^ x, ^z. Then all the digits of
^
x and ^
xD are same, where ^ xD D^z.
We note that the result of the above Theorem is not true if the
diagonal elements of D are not binary. However in that situation,
the quantities kx  ^ xD k, and k^
xk, kx  ^ x^xD k are of the same
order.
We next note that it is possible to nd a diagonal matrix such

1 1
that j2 DAD 6 j2 A. For example, for l > 0, let A .
0 l
Then j2 A 1=l is large for l small. Let l v2d , where
 
1=2 < v < 2. Consider D
1 0 1 2d=2 ,
d=2 . Then DAD
0 2 0 v
and 1 6 j2 DAD < 2 and consequently, j2 DAD < j2 A for l
small. Let D be the diagonal matrix such that
 
j2 D AD minD j2 DAD (minimum over all diagonal matrices
D with binary diagonal elements), then from the above theorem
and (7.2), we have
g 6 f nj2 D AD  6 f nj2 A:
Thus j2 D AD provides more accurate information about g than
j2 A. But in general, it is not easy to nd either D or j2 D AD .
In Hypothesis H, we used KA j2 DAD, where D is a diagonal
1=2
matrix with Dii Aii and Dii may not be binary. We note, how-
ever, that not using a binary only inuences C 1 , C 2 by factors of 1/
2 and 2 respectively. We also mention that in the literature
[14,26], an upper bound of the form (7.2) for g is available with
f n Cn2 and j2 A replaced by KA for symmetric positive de-
nite linear systems solved by Cholesky decomposition. In Hypothe-
109
sis H, we used f n Cnb ; b  0, based on our computational
experience.
We now consider a set of validation problems, whose exact
solution (experimental data) is known. The solution to these prob-
lems will be computed on various computers using double preci-
sion, i.e., with 16 digits of accuracy.
Problem 1: We consider approximating the solution ux x of
the problem u00 x 0; x 2 0; 1; u0
0; u1 1; by the FEM using piecewise linear nite
elements.
Problem 1a: We use the FE mesh vertices xi ih for i 0; 1; . . . ; N
and h 1=N. The FE solution is same as the exact
solution u of the problem. Let the associated linear
1
systems be A1 x1 b . The exact solution vector
1 1
x is known, namely, xi ih; i 1; 2; . . . ; N. We
will solve the linear system by the standard LU
decomposition algorithm for sparse matrices with-
out partial pivoting.
Problem 1b: We use the mesh vertices xi N  i 1h,
i 0; 1; . . . ; N. The FE solution is same as the exact
solution u and let the associated linear system be
2 2
A2 x2 b ; it is known that xi N  i 1
h; i 1; . . . ; N. Note that the elements of x2 are the
permuted elements of x1 and thus kx1 k2 kx2 k2 .
We will solve the linear system by the same algo-
rithm as Problem 1a.
The computations are performed on a Dell Latitude PC with
INTEL CORE (TM)2 CPU, 1.20GHZ.
Problem 2: We approximate the solution ux 1 of the prob-
lem u00 x 0; x 2 0; 1; u0 1; u1 1; by the
piecewise linear FEM based on the mesh vertices
as in Problem 1a. Let the associated linear system
be Ax b. It is clear that the exact solution is given
by xi 1; i 1; 2; . . . ; N.
Problem 2a: The linear system is solved by a sparse matrix direct
solver superLU [29] on a single processor.
Problem 2b: The linear system is solved by a sparse matrix direct
solver MUMPS [2] on a single processor.
Problem 2c: The linear system is solved by MUMPS, using parallel
computation, on 128 processors.The computations
were performed on the Lonestar system at Texas
Advanced Computing Center. Lonestar is a Linux
based cluster comprised of 1888 compute nodes
connected via high speed quad-data rate inniband,
with each compute node containing two hex-core
socket (INTEL Xeon 5680 processors) for an aggre-
gate system size of 22656 cores. Each core runs at
a peak of 3.33GHZ.
Problem 3: We consider approximating the solution ux x2 of
the problem u00 x 2; x 2 0; 1, u0 0; u0 1
2, by the GFEM based on S S 1 S 2 (see (3.4)). We
i
use ni 1 and u1 x x2 ; i 0; 1; . . . ; N. We order
0 1
the shape functions as N 0 u1 ; N 1 ; N 1 u1 ; N 2 ;
N
. . . ; N N ; N N u1 and suppose the associated stiffness
matrix is Ax b, where A is of the order 2N 1.
The GFEM solution is same as the exact solution u.
It is easy to see that x2i1 1; i 0; 2; ; N and
x2i 0; i 1; 2; . . . ; N. The linear system is solved
by the same algorithm and on the same platform
as in Problem 1a.
Problem 4: We consider approximating the solution of the same
problem in Problem 3 by the SGFEM based on
S S 1 S 2 (see (4.10)) with ni 1, T 2 T , and
u i 2 2
1 x  I xi x . We order the shape functions as
110 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111
0 1 N
N0 u 1 ; N1 ; N1 u
 1 ; N2 ; . . . ; NN ; NN u
 1 and suppose the
associated stiffness matrix is Ax b, where A is of
the order 2N 1. The GFEM solution is same as the
exact solution u and it is easy to see that
x2i1 1; i 0; 2; ; N and x2i ih2 ; i 1; 2; . . . ; N.
The linear system is solved by the same method
and on the same platform as in Problem 1a.
We will now validate Hypothesis H based on the validation
problems described above. We will consider the tolerance
s s1 ; s2 , with s1 400 and s2 0.
x1 and ^
Let ^ x2 be the computed solutions of the linear systems
1 2
A1 x1 b and A2 x2 b of Problem 1a and Problem 1b,
respectively. It can be shown that for large N, KA1 KA2
 0:4N 2 . We have computed and presented the loglog plots of
the relative errors gk kxk  ^ xk k2 =kxk k2 , k 1; 2, with respect
to N 100; 200; . . . ; 50000 in Fig. 1. We have observed that
k k
C 1 0:4N 2  6 gk 6 C 2 0:4N 2  for k 1; 2 with C 2 =C 1 6 120 < s1
(note s2 0). Thus we do not reject Hypothesis H. Note that we
did not reject the hypothesis based only on the subset of meshes
with the values of N, mentioned above. Moreover, it is interesting
to note that the plots of g1 and g2 are quite different. Thus the
computed solution is affected by changing the order of the FE mesh
vertices, in spite of the fact that kx1 k2 kx2 k2 .
In Problem 2, we solve the linear system Ax b using two dif-
ferent software superLU and MUMPS; we also implement MUMPS
on multiple processors. Let ^ xa ; ^
xb ; ^
xc be the computed solutions
of Problems 2a, 2b, and 2c, respectively. These solutions were com-
puted for 10 6 N 6 107 , with 90 values of N in the range 10; 102 ,
with 400 values of N in the range 102 ; 103 , and 360 values of N in
the range 10i ; 10i1 ; i 3; 4; 5; 6, and with N 107 . We presented
the loglog plots of gk kx  ^ xk k2 =kxk2 ; k a; b, for the values of
N given before, in Fig. 2a and b respectively. We observed that for
N P 100, C 2 =C 1 6 200 < s1 for both the problems. Thus we do not
reject the Hypothesis H for N P 100. Note that for N 6 100, Fig. 2a
and b suggest that b  1. It is also clear from Fig. 2b that the
implementation of the algorithm in MUMPS changes drastically
for N > 5  103 ; this is not the case with superLU, as seen in
Fig. 2a. Thus the computed solution depends on the software pack-
age, as mentioned before. For Problem 2c, we did not display the
loglog plot of gc kx  ^ xc k2 =kxk2 as it would be very similar
to the plot of gb in Fig. 2b. However, we computed Re
100gc 1018 . This is likely the effect of various cancelations. h
gb =gb the signed relative difference percent and pre-
sented the semi-log plot of Re in Fig. 2c for the same values of N,
given before. For N 6 5  103 , we see that Re  0 and values of Re
starts to oscillate for N > 5  103 . This indicates that the imple-
mentation in MUMPS changes drastically. Fig. 2c also suggests that
gc is larger than gb for most values on N, and gc gets closer to
gb as N increases.
Let ^
x be the solution of the linear system Ax b of Problem 3.
We have KA ON 4 (see Section 3.1). The loglog plot of
g kx  ^xk2 =kxk2 with respect to N 50; 100; 150; . . . ; 10000 have
been presented in Fig. 3a. In Fig. 3b, we show the details in the
range 100 6 N 6 1000, where we have presented the loglog plot
of g for every value of N in this range. Based on both these data
(i.e., the values of g for every value of N in the range
100 6 N 6 1000 and for N 1050; 1100; 1150; . . . ; 10000),we have
observed that C 1 N 4 6 g 6 C 2 N 4 with C 2 =C 1 6 340 < s1 (note
s2 0). Thus we do not reject the Hypothesis H, again based on
the subset of meshes with the values of N mentioned above. It is
important to note that in Problem 3a, all the digits of accuracy
were lost for N P 9000, and thus the Hypothesis H does not ad-
dress the value of N P 9000. We also computed g for every value
of N in the range 9000 6 N 6 11000; g was of the order 1 and oscil-
lated around 1.
Let ^
x be the computed solution of the linear system Ax b of
2
Problem 4. We have shown in this paper that KA Oh . We
have presented the loglog plot of g kx  ^ xk2 =kxk2 , with respect
to N 50; 100; 150; . . . ; 10000 in Fig. 4a, and for every value of N
in the range 100 6 N 6 1000 in Fig. 4b. Based on both these data
(i.e., the values of g for every value of N in the range 100 6 N
6 1000 and N 1050; 1100; 1150; . . . ; 10000), we observed that
C 1 N 2 6 g 6 C 2 N 2 with C 2 =C 1 6 240 < s1 (note s2 0). Thus we
do not reject the Hypothesis H (based on meshes with these values
of N).
Thus we did not reject the Hypothesis H for any validation prob-
lems with respect to the tolerance s1 400 and s2 0. But we
would reject the Hypothesis H if we choose s1 300, since
C 2 =C 1 6 340is1 in Problem 3. However, if the values of g for every
value of N in the range [100,1000] were not available (see Fig. 3b),
then we will have C 2 =C 1 6 250 < s1 , and we thus we would not re-
ject Hypothesis H. Hence validation depends on the values of N, i.e.,
on the number of validation problems considered, since each value
of N (in each of Problems 1, 2, 3, and 4) constitutes a separate val-
idation problem. But as mentioned before, the choice of the toler-
ance depends on the type of decision related to the goal. For
example in this paper, we have to decide whether to accept SGFEM
over the standard GFEM. In this case, we may allow s1 to be bigger;
in fact if s1 500, we still accept SGFEM over GFEM since the value
of g for GFEM will be much larger than the g of SGFEM for large N.
We summarize by stating that
(a) we have condence in Hypothesis H, based on the chosen
validation problems (Problems 14). We underline that we
have also considered other 2- and 3-dimensional validation
problems for the Hypothesis H, which we do not present
in this paper. We will present a more substantial validation
of Hypothesis H in a future publication.
(b) Because of our condence in Hypothesis H, we prefer the use
of SGFEM over GFEM, since linear system of SGFEM is less
prone to the loss of accuracy than the linear system of the
GFEM, when solved using an elimination method.
Remark 7.2. As mentioned before, all the computations presented
here were performed with 1016 accuracy. However, all the gures,
presented above, indicate that that the apparent accuracy is about
References
[1] American Society of Mechanical Engineers, New York. ASME guide for
Verication and Validation in Computational Solid Mechanics, 2006. V&V 10.
[2] P.R. Amestoy, I.S. Duff, J. Koster, J.Y. LExcellent, A fully asynchronous
multifrontal solver using distributed dynamic scheduling, SIMAX 23 (2001)
1541.
[3] I. Babuska, R. Lipton. Optimal local approximation spaces for generalized nite
element methods with application to multiscale problems. Technical Report
10-12, ICES, University of Texas at Austin, 2010.
[4] I. Babuka, U. Banerjee, J. Osborn, Generalized nite element methods: Main
ideas, results, and perspective, Int. J. Comput. Methods 1 (1) (2004) 137.
[5] I. Babuka, G. Caloz, J. Osborn, Special nite element methods for a class of
second order elliptic problems with rough coefcients, SIAM J. Numer. Anal. 31
(1994) 945981.
[6] I. Babuka, J.M. Melenk, The partition of unity nite element method, Int. J.
Numer. Meth. Engrg. 40 (1997) 727758.
[7] I. Babuka, J.T. Oden, Verication and Validation in computational engineering
and science: basic concepts, Comput. Methods Appl. Mech. Engrg. 193 (2004)
40574066.
[8] F.L. Bauer, Optimal scaling of matrices and the importance of the minimal
condition, in: C.M. Popplewell (Ed.), Information Processing 62 IFIP Congress
1962, North-Holland, Amsterdam, 1963, pp. 198201.
[9] T. Belytschko, T. Black, Elastic crack growth in nite elements with minimal
remeshing, Int. J. Numer. Meth. Engrg. 45 (1999) 601620.
[10] S.E. Benzley, Representation of singularities with isoparametric nite
elements, Int. J. Numer. Meth. Engrg. 8 (1974) 537545.
 I. Babuka, U. Banerjee / Comput. Methods Appl. Mech. Engrg. 201204 (2012) 91111
[11] H. Blum, M. Dobrowoski, On nite element methods for elliptic equations on
domains with corners, Computing 28 (1982) 5363.
[12] E. Byskov, The calculation of stress intensity factors using nite element with
cracked element, Int. J. Fract. Mech. 6 (1970) 159167.
[13] C. Daux, N. Moes, J. Dolbow, N. Sukumar, T. Belytschko, Arbitrary branched and
intersecting cracks with extended nite element method, Int. J. Numer. Meth.
Engrg. 48 (2000) 17411760.
[14] J. Demmel. On oating point error in cholesky. Technical Report CS-89-87,
Dept. of Computer Science, Univ. of Tennessee, 1989.
[15] J. Dolbow, N. Mos, T. Belytschko, Modeling fracture in Mindlin-Reissner plates
with the extended nite element method, J. Solids Struct. 37 (2000) 7161
7183.
[16] J.E. Dolbow, An Extended Finite Element Method with Discontinuous
Enrichment for Applied Mechanics. Ph.D. Thesis, Northwestern University,
1999.
[17] C.A. Duarte, J.T. Oden, An h-p adaptive method using clouds, Comput. Methods
Appl. Mech. Engrg. 139 (1996) 237262.
[18] C.A. Duarte, J.T. Oden, H-p Clouds An h-p Meshless Method, Numer. Methods
Partial Differ. Equat. 12 (1996) 673705.
[19] P. Esser, J. Grande, A. Reusken, An extended nite element method applied to
levitated droplet problems, Int. J. Numer. Meth. Engrg 84 (2010) 757773.
[20] M. Farsad, F.J. Vernerey, H.S. Park, An extended nite element/level set method
to study surface effects on the mechanical behavior and properties of
nanomaterials, Int. J. Numer. Meth. Engrg 84 (2010) 14661489.
[21] G. Fix, S. Gulati, G.I. Wakoff, On the use of singular functions with nite
element approximations, J. Comp. Phys. 13 (1973) 209228.
[22] T.-P. Fries, A corrected XFEM approximation without problems in blending
elements, Int. J. Numer. Meth. Engrg. 75 (2008) 503532.
[23] T-P Fries, T. Belytschko, The extended/generalized nite element method: An
overview of the method and its applications, Int. J. Numer. Meth. Engrg. 84
(2010) 253304.
[24] G.H. Golub, C.F. Van Loan, Matrix Computations, The Johns Hopkins University
Press, Baltimore, USA, 1996.
[25] M. Griebel, M.A. Schweitzer, A Particle-Partition of Unity method Part VI:
Adaptivity, in: M. Griebel, M.A. Schweitzer (Eds.), Meshfree Methods for
Partial Differential Equations III, Lecture Notes on Computer Science and
Engineering, Vol. 26, Springer, 2006, pp. 121148.
[26] N.J. Higham, Accuracy and Stability of Numerical Algorithms, SIAM,
Philadelphia, 2002.
[27] D. Kincaid, W. Cheney, Numerical Analysis; Mathematics of Scientic
Computing, American Mathematical Society, 2002.
[28] G.B. Kleindorfer, L. ONeill, R. Ganeshan, Validation in similation: various
positions in the philosophy of science, Management Science 44 (1998) 1087
1099.
[29] X.S. Li, J.W. Demmel, SuperLU-DIST: A scalable distributed-memory sparse
direct solver for unsymmetric linear systems, ACM Trans, Mathematical
Software 29 (2003) 110140.
[30] C. Lu, B. Shanker, Generalized nite element method for vector electromafnetic
problems, IEEE Transactions on Antennas and Propagation 55 (2007) 1369
1381.
[31] A.M. Matache, I. Babuska, C. Schwab, Generalized p-FEM in homogenization,
Numer. Math. 86 (2000).
111
[32] J.M. Melenk, On Generalized Finite Element Methods, Ph.D. Thesis, University
of Maryland, 1995.
[33] J.M. Melenk, I. Babuka, The partition of unity nite element method: Basic
theory and application, Comput. Methods Appl. Mech. Engrg. 139 (1996) 289
314.
[34] A. Menk, S.P.A. Bordas, A robust preconditioning technique for the extended
nite element method, Int. J. Meth. Engrg. 85 (2011) 16091632.
[35] N. Moes, M. Cloirec, P. Cartraud, J.F. Remacle, A computational approach to
handle complex microstructure geometries, Comput. Methods Appl. Mech.
Engrg. 192 (2003) 31633177.
[36] N. Moes, J. Dolbow, T. Belytschko, A nite element method for crack without
remeshing, Int. J. Numer. Meth. Engrg. 46 (1999) 131150.
[37] A. Nouy, A. Clment, eXtended stochastic nite element method for the
numerical simulation of heterogegeous materials with random material
interfaces, Int. J. Numer. Meth. Engng 83 (2010) 13121344.
[38] W.L. Oberkampf, Ch. J. Roy, Verication and Validation in Scientic Computing,
Cambridge University Press, New York, 2010.
[39] J.T. Oden, C.A.M. Duarte. Clouds, Cracks and FEMs, in: B. Daya Reddy (Ed.),
Recent Developments in Computational and Applied Mechanics, 1997.
[40] J.T. Oden, C.A.M. Duarte, O.C. Zienkiewicz, A new cloud-based hp nite element
method, Comput. Methods Appl. Mech. Engrg. 153 (1998) 117126.
[41] P. OHara, C.A. Duarte, T. Eason, Generalized nite element analysis for three-
dimensional problems exhibiting sharp thermal gradients, Comput. Methods
Appl. Mech. Engrg. 198 (2009) 18571871.
[42] A.K. Rao, I.S. Raju, A.V.K. Murthy, A powerful hybrid method in nite element
analysis, Int. J. Numer. Meth. Engrg. 3 (1971) 389403.
[43] P.J. Roache, Fundamentals of Verication and Validation, Hermosa Publisher,
Albuquerque, NM, 2009.
[44] M.A. Schweitzer, A parallel multilevel partition of unity method for elliptic
partial differential equations, Lecture Notes in Computational Science, vol. 29,
Springer, 2003.
[45] A. Simone, C.A. Duarte, E. Van der Giessen, A generalized nite element
method for polycrystals with discontinuous grain boundaries, Int. J. Numer.
Meth. Engrg. 67 (2006) 11221145.
[46] Simulation Interoperability Standards Organization, Orlando, FL, Guide for
generic methodology for Vericatin and Validation (V&V) and acceptance of
models, simulations, and data, 2007.
[47] G. Strang, G. Fix, An Analysis of the Finite Element Method, 2nd ed., Wellesley-
Cambridge, 2008.
[48] T. Strouboulis, I. Babuka, K. Copps, The design and analysis of the generalized
nite element method, Comput. Methods Appl. Mech. Engrg. 181 (2000) 43
69.
[49] T. Strouboulis, K. Copps, I. Babuka, The generalized nite element method: an
example of its implementation and illustration of its performance, Int. J.
Numer. Methods Engrg. 47 (2000) 14011417.
[50] T. Strouboulis, K. Copps, I. Babuka, The generalized nite element method,
Comput. Methods Appl. Mech. Engrg. 190 (2001) 40814193.
[51] N. Sukumar, N. Moes, B. Moran, T. Belytschko, Extended nite element method
for three dimensional crack modelling, Int. J. Numer. Methods Engrg. 48 (11)
(2000) 15491570.
[52] J.H. Wilkinson, The Algebraic Eigenvalue Problem, The Oxford University Press,
1988.