ECOLE POLYTECHNIQUE

CENTRE DE MATH

EMATIQUES APPLIQU

EES
UMR CNRS 7641
91128 PALAISEAU CEDEX (FRANCE). Tel: 01 69 33 41 50. Fax: 01 69 33 30 11
http://www.cmap.polytechnique.fr/
A multiscale nite element
method
for numerical homogenization
Gregoire Allaire, Robert Brizzi
R.I. N
0
545 July 2004
A multiscale nite element method
for numerical homogenization
Gregoire ALLAIRE
CMAP, UMR-CNRS 7641, Ecole Polytechnique 91128 Palaiseau Cedex (France)
Robert BRIZZI
CMAP, UMR-CNRS 7641, Ecole Polytechnique 91128 Palaiseau Cedex (France)
Abstract
This paper is concerned with a multiscale nite element method for numerically solving
second order scalar elliptic boundary value problems with highly oscillating coecients. In
the spirit of previous other works, our method is based on the coupling of a coarse global
mesh and of a ne local mesh, the latter one being used for computing independently an
adapted nite element basis for the coarse mesh. The main new idea is the introduction
of a composition rule, or change of variables, for the construction of this nite element
basis. In particular, this allows for a simple treatment of high order nite element methods.
We provide optimal error estimates in the case of periodically oscillating coecients. We
illustrate our method on various examples.
1 Introduction
The goal of this paper is to build a multiscale nite element for performing numerical homog-
enization. The word multiscale is understood here in the practical sense that two dierent
meshes will be used: a ne mesh for computing locally and independently (i.e. allowing for an
easy parallelization) a nite element basis, and a coarse mesh for computing globally and at
low cost the solution of an elliptic partial dierential equation. By numerical homogenization
we mean that we compute, not only the mean eld solution of a highly heterogeneous problem,
but also the local uctuations which may be important in many applications. Recently there
has been many contributions on multiscale numerical methods, including [3], [6], [8], [9], [10],
[11], [12], [13], [14]. Our work is in the spirit of that of Hou and Wu [11].
Our model problem is a scalar elliptic partial dierential equation which arises in many
applications such as diusion in porous media, or composite materials. Let be a bounded
1
open set of R
n
and f L
2
() (or, more generally, f H
1
()). For simplicity, we consider
Dirichlet boundary conditions. Our model problem is to nd u

H
1
0
() solution of
_

_
divA

grad u

= f in
u

= 0 on
(1)
where A

= (a

ij
)
n
i,j=1
is a non-necessarily symmetric matrix of coecients which all belong to
L

(). We assume that A

is uniformly bounded and coercive. The notation > 0 stands for

some small scale in the problem (but there could well be several scales, so is just a convenient
notation for all the small scales of variation of A

). Solving numerically (1) is a dicult task

if the scale is small. Indeed, a good approximation is obtained with classical nite element
methods (or any other methods) only if the mesh size h is smaller than the nest scale, i.e.
h . As is well known, this is not satisfactory since CPU time as well as memory storage
grow polynomially with h
1
and soon become prohibitively too large.
There is one way out of this diculty in the special case of periodic heterogeneities or
oscillations of the tensor A

. In the periodic case, is the period, and A

is dened by
A

(x) = A
_
x

_
,
where y A(y) is a Y -periodic function where Y = (0, 1)
n
is the unit cube. It is a classical
result of homogenization theory (see e.g. [5]) that, for small , u

is approximated by
u

(x) u

(x) +
n

i=1

i
_
x

_
u

x
i
(x) (2)
and
u

(x) u

(x) +
n

i=1
(
y

i
)
_
x

_
u

x
i
(x), (3)
where
i
is the solution of the so-called cell problem
_

_
div
y
A(y)
_
e
i
+ grad
y

i
_
= 0 in Y,
y
i
(y) 0 Y periodic.
(4)
The numerical resolution of (1) is often replaced by the simpler one of the homogenized problem
_

_
divA

grad u

= f in
u

= 0 on .
(5)
where A

is a constant homogenized tensor given by the explicit formula A

e
i
=
_
Y
A(y)
_
e
i
+ grad
y

i
_
dy. The approximation of u

by u

can be improved by adding the

rst order corrector term u
1
. This additional contribution may not be so important in (2) but
is crucial in (3) since it is of the same order of magnitude than u

. Note in passing that (2)

and (3) can also be improved by adding a so-called boundary layer term since u

+ u
1
does
not satisfy the Dirichlet boundary condition on (see e.g. [4]).
In the non-periodic case, although there still exist an homogenized problem and approx-
imation formula similar to (2), (3), the homogenized matrix A

is unfortunately unknown a
priori. Therefore, one can not replace the numerical resolution of the original problem (1) by
that of the homogenized problem (5). Instead, many multiscale numerical methods have been
2
recently devised in order to solve directly (1) but at a price (in terms of CPU time and memory
storage) comparable to that of solving (5). Typically, a multiscale nite element method uses
a coarse mesh of size h > and an adapted nite element basis which incorporates the small
scale features of the oscillating tensor A

. The nite element basis is pre-computed locally on

each cell of the coarse mesh (and in parallel for computational eciency) by solving a local
version of (1). The number of degrees of freedom is thus not larger than for a classical nite
element method on the coarse mesh.
Let us describe the main idea behind our new multiscale nite element method. Although
it is designed for non-periodic homogenization problem, for the sake of simplicity we rst present
it in the context of periodic homogenization. A key idea is to remark that (2) looks like a rst
order Taylor expansion, namely it implies
u

(x) u

_
x +
_
x

__
(6)
where = (
1
, . . . ,
n
). Building on (6), we introduce a coarse mesh of size h > and a
classical conforming nite element basis
_

h
l
_
l
, and we dene an oscillating nite element basis
through the same composition rule

,h
l
(x) =
h
l
_
x +
_
x

__
.
Our method amounts to apply a standard Galerkin procedure to the variational formulation
of (1) with this oscillating nite element basis. The advantages of our method are at least
twofold. First, it is very easy to implement high order methods since the computation of
the oscillating functions
_
x

_
is independent of the order of the coarse mesh nite element
basis
_

h
l
_
l
. Second, the convergence analysis is somehow simpler since, roughly speaking, it
amounts to apply the change of variables x x +
_
x

_
to standard convergence result on
the coarse mesh
The content of the paper is as follows. In section 2, we recall some basic facts of ho-
mogenization theory and we give a precise statement about (6) in the general (non-periodic)
case. Section 3 is devoted to a precise denition of our multiscale nite element method. Its
convergence is then studied in section 4. Finally, in section 5 some numerical results are given.
2 Some results in homogenization theory
2.1 H-convergence and oscillating test functions
Let us recall some results of the H-convergence theory (for details see e.g. [15], [2]). Let /
n
be the linear space of square real matrices of order n and dene, for given positive constants
> 0 and > 0, the subspace of /
n
made of matrices which are coercive as well as their
inverses
/
,
=
_
M /
n
; M. [[
2
, M
1
. [[
2
, R
n
_
.
A sequence of matrices A

(; /
,
) is said to H-converge, when goes to zero, to a
homogenized matrix A

(; /
,
) if, for any right hand side f H
1
(), the sequence
of solutions u

of (1) satises
u

weakly in H
1
0
() ( 0),
A

grad u

grad u

weakly in L
2
()
n
( 0),
where u

denotes the solution of the homogenized equation (5). This denition makes sense
because of the following sequential compactness property [15].
3
THEOREM 2.1 Let (A

)
>0
be a sequence of matrices in L

(; /
,
). There exists a sub-
sequence, still denoted by , and a homogenized matrix A

(; /
,
) such that A

H-
converges to A

.
Except in periodic case, this abstract result does not give an explicit formula for the limit A

.
Actually, the homogenized tensor A

is dened as a limit in the distributional sense, namely

A

(x) grad w

i
A

e
i
in D

(; R
n
)
where (e
i
)
i=1,n
denotes the canonical basis of R
n
, and ( w

i
)
i=1,n
are the so-called oscillating
test functions which satisfy
w

i
x
i
weakly in H
1
() ( 0), (7)
and
g

i
= divA

grad w

i
g
i
= divA

e
i
strongly in H
1
() ( 0). (8)
The existence of such oscillating test functions is the key point in the proof of Theorem 2.1 but
they are neither explicit (they depend on A

) nor unique (they are unique up to the addition

of a sequence converging strongly to zero in H
1
()). For our purpose, we dene them as the
solutions of the following boundary value problems (j = 1, ..., n)
_

_
divA

grad w

j
= divA

e
j
in
w

j
= x
j
on .
(9)
These oscillating test functions are also useful for obtaining a corrector result.
THEOREM 2.2 Let (A

)
>0
be a sequence H-converging to A

in L

(; /
,
). Then,
grad u

=
n

i=1
w

i
u

x
i
+ r

, (10)
where the remainder r

converges strongly to zero in L

1
(; R
n
). Furthermore, if u

W
1,
(),
then r

converges strongly in L
2
(; R
n
).
REMARK 2.3 In the context of Theorem 2.2 it is clear that, if the homogenized solution is
smoother, say u W
2,
(), then
u

= u

+
n

i=1
( w

i
(x) x
i
)
u

x
i
+ r

(11)
where the remainder r

converges strongly to zero in H

1
(). In the sequel, we denote by
[

] =
__
w

j
x
i
__
i,j=1,n
(12)
the so-called corrector matrix.
Proof: The proof of (10) is classical (see e.g. [2]). The last statement of Theorem 2.2 is
simpler to prove, so we briey explain how to proceed. Using the coercivity of A

we are done
if we can show that
4
lim
0
_

_
grad u

]grad u

_
grad u

]grad u

_
dx = 0. (13)
Developing the scalar product in the integral (13), we obtain the following terms
_

]
t
A

] grad u

grad u

dx +
_

grad u

grad u

dx

grad u

] grad u

dx
_

] grad u

grad u

dx.
To pass to the limit in the rst term we use the facts (see Lemma 1.3.38 [2]) that [

]
t
A

]
converges to A

in D

(; M
n
), and that, thanks to Meyers theorem which implies a uniform
L
p
() bound (with p > 2) for [

], this convergence holds true weakly in L

1
(; M
n
). Thus,
since u

W
1,
(), we have
_

]
t
A

] grad u

grad u

dx
_

grad u

grad u

dx.
The second term is easy
_

grad u

grad u

dx =
_

f u

dx
_

f u

dx =
_

grad u

grad u

dx.
Applying the div-curl compensated compactness result [15] to the third term and using the
regularity assumptions as well as Meyers theorem, we get the desired result
_

grad u

] grad u

dx
_

grad u

grad u

dx.
The fourth term is treated in the same way and their combination gives zero.
2.2 A remark on the corrector result
The right hand side of formula (11) looks like the rst order Taylor expansion of u

at the point
w

(x) = ( w

1
(x), ..., w

n
(x)). It indicates that u

(x) may well be approximated by u

(x).
This remark is at the basis of the new form of the corrector result (Theorem 2.2) that we now
propose.
THEOREM 2.4 Let (A

)
>0
be a sequence H-converging to A

in L

(; /
,
). For f
H
1
(), let u

be the solution of (1) and u

be that of (5). Let w

j
be the family of oscillating
test functions dened in (9). Assume that u

W
2,
() and w

is uniformly bounded in
L
q
()
n
for any 2 q < +. Then
u

= u

+ r

, (14)
where the remainder term r

converges strongly to zero in H

1
0
().
REMARK 2.5 In space dimension n = 2, since H
1
() L
q
() for any 2 q < +, there
is no additional assumption on w

in Theorem 2.4. If w

is not uniformly bounded in any

L
q
()
n
, it is at least bounded in L
2

()
n
with 2

= 2n/(n 2) and we can still obtain the

strong convergence to zero of r

in W
1,n/(n1)
().
REMARK 2.6 The approximation of the principal part of the solution u

, i.e. u

, may
serve as a substitute for the approximation of the solution of problem (1). This idea is at the
root of the new multiscale nite element method described in this work.
5
Proof: It is not clear that w

() (see remark 2.10), so formula (14) makes sense if u

is rst extended by zero outside . Since u

is a Lipschitz function, u

belongs to H
1
().
Furthermore, since w

(x) = x a.e. on , the function u

belongs to H
1
0
(). We have
|u

|
H
1
0
()
= |grad u

grad ( u

)|
L
2
()
n
|grad u

] grad u

|
L
2
()
n + |[

] ( grad u

(grad u

) w

) |
L
2
()
n. (15)
The rst term in the right hand side of (15) goes to zero because of Theorem 2.2. The second
term is bounded by
|[

] ( grad u (grad u) w

) |
L
2
()
n |[

]|
L
p
(;Mn)
|grad u

(grad u

) w

|
L
p

()
n
with 1/p + 1/p

= 1/2. A Taylor expansion with integral rest yields

(grad u

) w

= grad u

+
_
1
0
u

(x +t( w

(x) x)).( w

(x) x) dt
and thus, we obtain
|(grad u

) w

grad u

|
L
p

()
n
|u

|
W
2,
()
| w

x|
L
p

(;R
n
)
n
.
By Meyers theorem there exists p > 2 such that |[

]|
L
p
(;Mn)
is uniformly bounded. By
assumption ( w

x) is bounded in any L
q
(; R
n
)
n
, 2 q < +, and since it converges strongly
to zero in L
2
(; R
n
)
n
, it also converges strongly in L
p

(; R
n
)
n
. All together this implies that
the second term in the right hand side of (15) goes to zero.
There is a converse statement of Theorem 2.4.
THEOREM 2.7 Let (A

)
>0
be a sequence H-converging to A

in L

(; /
,
). For f
H
1
(), let u

be the solution of (1) and w

be the family of oscillating test functions dened

in (9). If there exists a function u H
1
0
() W
1,
() such that
|u

u w

|
H
1
0
()
0 when 0,
then u = u

, the solution of the homogenized problem (5).

Proof: By assumption, u

admits the following representation formula, similar to (14),

u

= u w

+ r

, (16)
where r

converges strongly to zero in H

1
0
(). Let us consider the variational formulation of
(1)
a

(u

, v) =
_

grad u

grad v dx =
_

f v dx , v H
1
0
(). (17)
Substituting (16) into (17) gives
_

] (grad u) w

grad v dx +
_

grad r

grad v dx =
_

f v dx , v H
1
0
().
By assumption the second integral goes to zero, while the rst one converges to
_

u
v dx. Indeed, by H-convergence, A

] converges weakly to A

in L
2
(; M
n
) and, by the
Lebesgue dominated convergence theorem, (grad u) w

converges strongly to gradu in L

2
()
n
.
Thus, passing to the limit we obtain that u is a solution of the variational formulation of the
homogenized problem (5), i.e. u = u

.
6
2.3 An approximate variational formulation
The representation formula (14) for u

suggests an approximation of the variational formulation

(17). Indeed, it is equivalent to
a

( u

, v ) =
_

f v dx a

( r

, v ) , v H
1
0
(), (18)
where the last term goes to zero. Dropping it and choosing an adequate subspace of H
1
0
()
should yield a good approximation of (17). A rst possible choice of subspace is
_
v

H
1
0
() ; v H
1
0
() W
1,
() , v

= v w

_
,
but it is unfortunately not closed in H
1
0
(), so it can not be a Hilbert space. Another possibility,
which requires the additional regularity w

W
1,
(; R
n
), is
V

=
_
v

H
1
0
() ; v H
1
0
() , v

= v w

_
, (19)
which is a closed subspace of H
1
0
() since w

W
1,
(; R
n
) implies that v w

belongs
to H
1
0
() as soon as v does. We dened the approximate variational formulation as: nd
u H
1
0
() such that
a

(u w

, v w

) =
_

fv w

dx , v H
1
0
(). (20)
By the Lax-Milgram theorem (20) admits a unique solution u w

in V

. In the following, we
will call u the substituting homogenized solution. Remark that, u actually depends on but it
oscillates less compared to u

.
LEMMA 2.8 Assume w

W
1,
(; R
n
). Let u be the unique solution of (20) and u

be the
solution of the homogenized problem (5). Then
|(u u

) w

|
H
1
0
()
0 when 0.
REMARK 2.9 Of course, combining Lemma 2.8 and our corrector result Theorem 2.4, we
deduce that u w

is a good approximation of the solution u

of the original problem (1)

|u

u w

|
H
1
0
()
0 when 0.
Proof: Subtracting (20) from (18) with the same test function (v w

) V

H
1
0
() we
obtain
a

( (u u

) w

, v w

) = a

( r

, v w

).
Taking v = u u

, using coercivity and continuity of the bilinear form show that

|(u u

) w

|
H
1
0
()

1
| r

|
H
1
0
()
,
which goes to zero as does.
2.4 Change of variables
A reasonable question to ask is whether the mapping x w

(x) from into R

n
is a change
of variables, i.e. is one-to-one into . We do not know if it is true in general. Nevertheless we
have the following partial results.
LEMMA 2.10 Assume w

W
1,
()
n
. Then, the mapping w

is onto in the sense that

w

() .
7
Proof: Since w

= xon and W
1,
() C
0
() then, from topological degree theory [16],
deg( w

, , y) = deg(id, , y) y ,
where id denotes the canonical injection from into R
n
. On the other hand,
deg(id, , y) =
_

_
1 if y
0 if y , .
thus, for y , deg( w

, , y) ,= 0 and from topological degree property, we deduce that there

exists an x such that y = w

(x).
The question of injectivity is much more delicate. Let us simply recall the following result
of [1].
THEOREM 2.11 Assume R
2
is a bounded simply connected open set, whose boundary
is a convex closed curve. Assume further that A

(x) is symmetric and that A

does not
depend on x. Then, the mapping w

is a homeomorphism from into itself.

REMARK 2.12 The homogenized matrix A

is constant, for example, in the case of periodic

homogenization, i.e. A

(x) = A(
x

).
REMARK 2.13 In the case of small amplitude homogenization, one can prove that the mapping
w

is a homeomorphism. Indeed, by a perturbation argument, if A

(x) is close to a constant

matrix A
0
, then w

(x) is close to x and thus is one-to-one.

3 Denition of the multiscale nite element method
3.1 Approximation of the oscillating test functions
The idea of our multiscale nite element method is to solve the approximate variational formu-
lation (20) instead of the true one (17). This requires the computation of the oscillating test
functions which are not explicit since, in view of their denition (9), they depend on A

which
is unknown. Therefore, we need to introduce an adequate approximation procedure.
We introduce a coarse mesh of which, for simplicity, is assumed to be polyhedral. This coarse
mesh is a conformal triangulation T
h
such that
=
_
KT
h
K (21)
where the elements K satisfy diam(K) h. In practice the mesh size h is larger than the
space scale of oscillations , i.e. h > . For each K T
h
, let us dene w
,K
i
(i = 1, ..., n) as the
solution of
_

_
divA

grad w
,K
i
= divA

K
grad x
i
in K,
w
,K
i
= x
i
on K,
(22)
where A

K
is a local approximation of A

in K. The simplest approximation is to take A

K
constant in K : in such a case its precise value is irrelevant since the right hand side of (22)
cancels out. This will be our choice in the numerical examples of this paper. Nevertheless, it
is possible to take A

K
(x) as some varying local average of A

.
Collecting together these local approximations we dene w
,h
i
H
1
() by w
,h
i
= w
,K
i
for each K T
h
, and we set w
,h
=
_
w
,h
1
, . . . , w
,h
n
_
H
1
(; R
n
).
8
A numerical approximation of the local oscillating test functions dened in (22) is com-
puted by using a classical conforming nite element in each K T
h
. For each coarse mesh cell
K we introduce a local ne mesh T
K
h
where h

denotes the size of this ne mesh. Of course, we

have h

< h, but, since we want to resolve the oscillations of the tensor A

, we also have h

< .
Typically we use P
k
Lagrange nite elements for solving the local boundary value problems
(22). Since, from one cell K to the other, these problems are independent this may be done in
parallel. This procedure is very similar to that introduced in [11].
The hats used in our notation refer to exact solutions of boundary value problems : thus,
w

refers to the true oscillating test function and w

,h
refers to the collection over T
h
of all
solutions of (22). We shall drop the hat notation when we refer to the corresponding numerical
approximations: thus w
,K
i
is the P
k
nite element approximation of w
,K
i
, and w
,h
i
, w
,h
are
dened as above with the hat notation.
REMARK 3.1 The domain of denition of the local oscillating test functions w
,K
i
is exactly
the cell K. However, one can possibly dene it on a larger domain (even if only its restriction
to K is used). For instance, in order to prevent boundary layer eects we can devise a so-called
oversampling method in the spirit of [11] by solving (22) in a domain Q which is slightly larger
than K. In the context of [11] this yields to a non-conforming nite element method which is
analyzed in [9]. In our framework, the oversampling method is still a conforming nite element
method (see Remark 3.4).
REMARK 3.2 In many applications, like composite materials, the coecient matrix A

is dis-
continuous and thus the oscillating test function w
,h
does not have a second derivative. In such
a case, as is well known, using higher order nite elements does not improve the convergence
rate, so we content ourselves in using P
1
nite elements for computing w
,h
.
3.2 A multiscale nite element method
Let V
h
H
1
0
() be a nite dimensional subspace (dimV
h
= N
h
) corresponding to a conforming
nite element method dened on the coarse mesh (22). Typically we use P
k
Lagrange nite
elements. Let
_

h
l
_
l=1,...,N
h
denote a nite element basis of V
h
. In order to compute a numerical
approximation u
h
of the substituting homogenized solution u, we introduce an oscillating (or
multiscale) nite element basis dened by

,h
l
(x) =
h
l
w
,h
(x) , (l = 1, ..., N
h
). (23)
We therefore obtain a conformal nite element method associated to the coarse mesh T
h
and we
denote by V

h
H
1
0
() the space spanned by the functions
_

,h
l
_
l=1,...,N
h
. Roughly speaking,
V

h
is the space V
h
w
,h
.
From the approximate variational formulation (20), we deduce a numerical approximation:
nd u
h
w
,h
V

h
such that
a

(u
h
w
,h
, v
h
w
,h
) =
_

f v
h
w
,h
dx , v
h
w
,h
V

h
. (24)
We use Lagrange nite elements and consequently the degrees of freedom are the values at the
nodes n
K
j
(j = 1, ..., N
K
) of the elements K T
h
. For such an element K, let
,K
be the
associated local basis which is made of N
K
polynomials p
K
i
P
k
satisfying p
K
i
(n
K
j
) =
ij
.
The local oscillating nite element basis is

,K
i
(x) = p
K
i
w
,K
(x) (25)
9
and, since w
,K
(x) = x on K, it still satises

,K
i
(n
K
j
) =
ij
.
On the other hand, for each K T
h
,
(u
h
w
,K
)[
K
(x) = [
,K
(x)] u
K
h

where [
,K
(x)] = [
,K
1
, ...,
,K
N
K
] and u
K
h
is the column vector composed of values of u
h
at
the nodes of K.
REMARK 3.3 In the case of piecewise linear nite elements P
1
we recover the multiscale nite
element method previously introduced by T. Hou and X.-H. Wu [11]. Indeed, when the basis
functions p
K
i
belong to P
1
, by linearity the oscillating basis functions can be written

,K
i
(x) = p
K
i
(x) +
n

j=1
_
w
,K
j
(x) x
j
_
p
K
i
x
j
(x). (26)
A simple calculus shows that
divA

grad
,K
i
=
n

j=1
divA

grad w
,K
j

p
K
i
x
j
in K
and, if we choose A

K
constant in the denition (22), we obtain
_

_
divA

grad
,K
i
= 0 in K

,K
i
= p
i
on K
(27)
which is precisely the denition of the nite element basis in the multiscale method of T. Hou
and X.-H. Wu [11].
REMARK 3.4 As explained in Remark 3.1 we can devise an oversampling method in the spirit
of [11], [9]. If the local oscillating test functions w
,K
i
are computed from (22) in a domain
Q which is larger than K, the composition with
h
l
is still going to dene a conforming nite
element basis. However, the support of
,h
l
may now be dierent from that of
h
l
and its nodal
values are also dierent.
4 Convergence proof in the periodic case
Although for its practical implementation our multiscale numerical method does not make
any assumption on the possible type of heterogeneities or oscillations of the tensor A

, it is
convenient to analyze its convergence in the context of periodically oscillating coecients. In
this section, we assume that A

(x) = A(
x

) where A(y) is a periodic, uniformly coercive and

bounded, matrix in the unit cube Y = (0, 1)
n
. As is usual practice in the numerical analysis
of nite element methods, we assume that the tensor A(y) is smooth in order to use classical
result from this eld. However, from the point of view of homogenization theory, it is enough
to assume that A(y) is piecewise smooth, possibly discontinuous through smooth interfaces,
since this implies that the solutions of the cell problem (4) and those of (22) (which denes the
oscillating test functions) belong to W
1,
which is just enough for our analysis.
Recall that we use a P
k
Lagrange nite element method (k 1) on the coarse mesh of
size h, and, locally on each cell K of the coarse mesh, a P
k
Lagrange nite element method
on a ne mesh of size h

to compute the oscillating test functions. We always assume that

0 < h

< < h < 1.

10
THEOREM 4.1 Let u

be the exact solution of (1) and u

h
u
h
w
,h
be the numerical solution
of (24). Assume that u

W
k+1,
() and
i
W
1,
(Y ). There exists a constant C
independent of and h such that
|u

h
|
H
1
0
()
C
_
h
k
+
_

h
+
_
h

_
k

_
. (28)
REMARK 4.2 Using an oversampling method as explained in Remarks 3.1 and 3.4 would im-
prove slightly estimate (28) by replacing the
_
/h term in its right hand side (which is due to
boundary layer eects) with /h. However, the resonance eect (i.e. the fact that method does
not converge if h ) does not disappear.
Proof: From Ceas lemma [7] applied to (24), there exists a constant C independent of
and h such that
|u

h
|
H
1
0
()
C inf
v

h
V

h
|u

h
|
H
1
0
()
. (29)
Dene
h
as the V
h
-interpolation operator:
h
v(x) =
N
h

l=1
v(n
l
)
h
l
(x) where n
l
denotes the
nodes associated to the P
k
nite element method. In the same way

h
denotes the V

h
-
interpolation operator:

h
v(x) =
N
h

l=1
v(n
l
)
,h
l
(x). It satises

h
v = (
h
v) w
,h
.
In (29) we choose v

h
=

h
u

where u

is the exact solution of the homogenized problem (5).

This yields
|u

h
|
H
1
0
()
C |u

h
u

|
H
1
0
()
. (30)
Introducing the rescaled solution of the cell problem (4)
w

(x) = x +
_
x

_
, (31)
we bound the right hand side of (30)
|u

h
u

)|
L
2
()
n C
_
|u

(u

) |
L
2
()
n
+ |( u

h
u

) w

|
L
2
()
n
+ |
h
u

( w

w
,h
)|
L
2
()
n
+ |
h
u

( w
,h
w
,h
)|
L
2
()
n
_
.
(32)
The upper bound (32), which gives the order of convergence of our multiscale nite element
method, is made of four terms. The rst one is related to a corrector result in periodic ho-
mogenization. The second one is linked to an interpolation result for the coarse mesh P
k
nite
element method. The third one is related to an homogenization result for the local oscillating
test functions. Finally the fourth term is concerned with an error estimate for the P
k
nite
element method used to compute the local oscillating test functions.
The rst term in the right hand side of (32) is bounded thanks to Lemma 4.4. The second
term is
|[ w

]( u

h
u

) w

|
L
2
()
n |Id+[
y
]|
L

(Y )
|( u

h
u

) w

|
L
2
()
n. (33)
To estimate the right hand side of (33) we perform a Taylor expansion with integral rest
( u

h
u

) w

= ( u

h
u

)(x) +
_
1
0
( u

h
u

)
_
x +t(
x

)
_

_
x

_
dt
11
and thus (33) is bounded by
|Id + [
y
]|
L

(Y )
_
|u

h
u

|
H
1
()
+ |u

h
u

|
W
2,
()
||
L
2
(Y )
_
|Id + [
y
]|
L

(Y )
|u

|
W
k+1,
()
_
h
k
+ h
k1
||
L
2
(Y )
_
by standard interpolation results for Lagrange P
k
nite elements [7]. Remark that the above
Taylor expansion is valid only if the interpolate
h
u

admits a second derivative, which is true

if the nite element order is k 2. For k = 1 the argument must be slightly changed, but,
since in such a case our method coincides with that in [11], [12], we do not give unnecessary
details.
The third term in the right hand side of (32) is
|[( w

w
,h
)](
h
u

) ( w

w
,h
)|
L
2
()
n |u

|
W
1,
()
|( w

w
,h
)|
L
2
()
n. (34)
To estimate the right hand side of (34) we write
|( w

w
,h
)|
2
L
2
()
n =

KT
h
|( w

w
,h
)|
2
L
2
(K)
n
and we use Lemma 4.3 for each cell K: actually, w
,h
converges to x in H
1
(K) weakly as
goes to 0 (for xed h), so w

is exactly the sum of the homogenized limit and of the corrector

term. Since K is of size h, its surface [K[ is of order h
n1
and the total number of cells K is
of order h
n
. Thus, we obtain
|( w

w
,h
)|
2
L
2
()
n C

h
.
Finally the fourth term in the right hand side of (32) is
|[( w
,h
w
,h
)](
h
u

) ( w
,h
w
,h
)|
L
2
()
n |u

|
W
1,
()
|( w
,h
w
,h
)|
L
2
()
n. (35)
In the right hand side of (35) we have the dierence between an exact cell solution w
,h
and
its numerical approximation w
,h
. By standard interpolation results for Lagrange P
k
nite
elements [7] (valid if the tensor A(y) is smooth enough), we thus obtain
|( w
,h
w
,h
)|
2
L
2
()
n =

KT
h
|( w
,h
w
,h
)|
2
L
2
(K)
n C(h

)
2k

KT
h
[ w
,h
[
2
H
k

+1
(K)
.
Then, assuming again that A(y) is smooth enough, the oscillating test function is also smooth
and satises
[ w
,h
[
H
k

+1
(K)
C
k

_
[K[.
Collecting together all four terms, (32) yields
|u

h
u

)|
L
2
()
n C
_

+ (h
k
+h
k1
) +
_

h
+
_
h

_
k

_
,
which implies (28).
We now recall a classical corrector result in periodic homogenization [5] where the depen-
dence on the size of the domain is made explicit (for a proof, see [11]). Let be a smooth
bounded open set, f L
2
() and g H
1
(). Dene the original problem
_
divA

grad v

= f in ,
v

= g on ,
and its homogenized limit
_
divA

grad v

= f in ,
v

= g on .
12
LEMMA 4.3 There exists a constant C, which is independent of , and the data f, g, such
that
_
_
_
_
_
v

i=1

i
_
x

_
v

x
i
(x)
_
_
_
_
_
H
1
()
C

_
[[ |v

|
W
2,
()
.
The next lemma is a quantitative version of Theorem 2.4 in the periodic case. Remark
however that it involves the solution of the cell problem instead of the oscillating test function
w

dened by (9).
LEMMA 4.4 Let w

be dened by (31). Assume that u

W
2,
() and
i
W
1,
(Y ).
Then, there exists a constant C, independent of , such that
|u

|
H
1
0
()
C

.
Proof: We have
|u

( u

)|
L
2
()
n |u

[ w

] u

|
L
2
()
n
+ |[ w

] ( u

(u

) w

) |
L
2
()
n.
(36)
Since [ w

(x)] = Id + [
y
]
_
x

_
, the rst term in the right hand side of (36) is bounded by

by a classical corrector result (see Lemma 4.3). The second term is bounded by
|Id + [
y
]|
L

(Y )
|u

(u

) w

|
L
2
()
n.
A Taylor expansion with integral rest yields
(u

) w

= u

+
_
1
0
u

_
x +t(
x

)
_

_
x

_
dt
and thus, we obtain
|(u

) w

|
L
2
()
n |u

|
W
2,
()
| |
L
2
()
n
which gives the desired result.
5 Numerical results
In this section, we experimentally study the convergence and the accuracy of our multiscale
method through numerical computations. For the sake of comparison we rst implemented the
method of T. Hou and X.-H. Wu [11], based on the direct numerical computation of the base
functions dened in (27). We checked that our multiscale method in the P
1
case, which is the-
oretically equivalent to the method of T. Hou and X.-H. Wu, does indeed coincide numerically
although the implementations are quite dierent. The main novelty is the implementation of
our P
2
multiscale method (denoted by P
2
-MSFEM). As can be expected from the error estimate
(28), its numerical results give a better approximation than the P
1
method.
We rst conduct numerical experiments in the periodic setting with a smooth scalar
conductivity tensor (taken from [11])
A

(x) = a(x/) Id
where
a(x/) = 1/(2 +P sin(2x
1
/))(2 +P sin(2x
2
/)).
In this formula, P is used as a contrast parameter (P = 1.8 for the numerical results presented
here). The right hand side of the Dirichlet boundary value problem is f = 1. All computations
are performed on the unit square domain which is uniformly meshed by triangular nite
13
elements (the coarse mesh of size h). Each triangle in this coarse mesh is again meshed by
triangular nite elements (the local ne mesh of size h

<< h). To make comparisons, we

computed a reference solution (hopefully converged) on a very ne mesh of 10
6
degrees of
freedom, with a classical P
1
-nite element method (denoted by P
1
-FEM). We also compared
our solution to the approximate one given by the rst two terms of the two-scale asymptotic
expansion. The homogenized conductivity is given by A

= 1/2(4 P
2
)
1/2
.
Secondly, since the main interest of numerical homogenization is to compute approximate
solutions in a non-periodic setting, we consider an heterogeneous composite material made of
a pseudo random distribution of spherical inclusions in a background matrix. In such a case,
there is of course no reference solution.
Let us recall once more that we are interested in the case h > since the opposite case
h < is covered by the classical nite element method in a much better way.
5.1 Periodic setting
Comparison with the two-scale asymptotic expansion A rst obvious comparison is
made between the two-scale asymptotic expansion and the P
2
-multiscale nite element approx-
imation. This latter solution is computed on a coarse mesh with h = 2 10
1
and a ne mesh
h

= 4 10
3
. The asymptotic expansion (denoted by P
1
-FEM AE) is built from P
1
nite
element approximations of the homogenized solution and of the oscillating functions (see (2)).
These latter ones are computed on the unit cell Y = (0, 1)
2
with periodic boundary conditions
(see (3)). Figure 1 shows as expected a good approximation.
0 1 0.5
-0.1
0
-0.05
P1-FEMAE
P2-MsFEM
0.1 0.05 0.15
-0.06
-0.05
-0.04
-0.03
-0.055
-0.045
-0.035
-0.025 P1-FEMAE
P2-MsFEM
Figure 1: Cross-sections at y = 0.5 of the rebuilt solution from asymptotic expansion and
multiscale approximation (left) with a close-up (right) : = 10
2
, h = 1/5 and h

= h/500.
Resonance eects and optimal mesh scale As we saw in the previous section, the theo-
retical bound of the speed of convergence of the method takes the form, for k = 2, k

= 1 and
h

=
h
M
(with M = 500 in our experiments)
g

(h) = h
2
+
_

h
+
h
M
.
Although g

(h) is only an upper bound of the true error, it indicates that there exists an
optimal h

() mesh size (which depends on k too) such that the numerical error, or at least g

,
is minimum. If we assume that M is very large, the order of this optimal value is
h

()
1/5
. (37)
14
According to [11] the resonance eect is the occurrence of large errors when the grid size
h and the heterogeneities scale are close. This eect is also predicted by the above formula
and is inherent to the multiscale framework used to compute the oscillating functions : the
decomposition of the initial boundary value problems into independent smaller boundary value
problems set on the coarse mesh element is somehow arbitrary.
The theoretical bound of the order of convergence, g

(h), is not small when the discretiza-

tion step h is close to (which is supposed to be very small). Therefore it is not a good idea to
take h too small. Rather, in view of (37), the optimal discretization step h

() is larger than
. This is indeed conrmed by our numerical experiments shown on gure 2 where one can
see that the behavior of the true numerical error (right) is close to that of the upper bound
g

(h) (left). It clearly indicates that there exists an optimal mesh size h

(larger than ) which

minimizes the numerical error.
h
g
_
e
p
s
(
h
)
0 0.1 0.2 0.3 0.4 0.5
0.2
0.4
0.6
0.8
1
EPS=10-2
EPS=2.10-2
EPS=4.10-2
EPS=8.10-2
h
|
|
U
e
p
s
,
h

-

U
r
e
f
|
|
H
1
0.1 0.2 0.3 0.4 0.5
0.1
0.05
0.15
EPS=8.10-2
EPS=4.10-2
EPS=2.10-2
EPS=10-2
Figure 2: Predicted (left) and computed (right) error estimate as a function of h for dierent
values of .
Convergence toward the homogenized solution We x the ratio M = h/h

(equal to
500) large enough so that M >
1
. Then, for dierent values of and for the optimal mesh
size h

() we compute the P
2
-MSFEM approximation of u

. One can check on gure 3 that

this numerical approximation of u

clearly converges toward the homogenized solution.

x
0 1 0.5
-0.1
0
-0.05
EPS=8.10-2
EPS=4.10-2
EPS=2.10-2
EPS=10-2
U*
x
0.1 0.2 0.3 0.4
-0.12
-0.11
-0.1
-0.09
-0.08
-0.07
-0.06
EPS=8.10-2
EPS=4.10-2
EPS=2.10-2
EPS=10-2
U*
Figure 3: Convergence toward the homogenized solution for experimental h

with = 8.10
2
,
4.10
2
, 2.10
2
, 10
2
: cross-section at y = 0.5 (left), and close-up (right).
15
Eects of the boundary conditions for the oscillating functions Imposing Dirichlet
boundary conditions for the oscillating functions on each element K is a convenient, albeit
arbitrary, choice. As a matter of fact it is one of the main problems of the method. Since the
oscillating function w
,h
(x) is equal to the identity x on all boundary nodes on K, see (22),
it can not oscillate on theses boundaries. Therefore our multiscale method cannot mimic the
oscillating properties of the true solution locally on the boundaries of the coarse mesh cells K.
This eect is easily seen when we consider cross-sections of the solution along the sides of the
elements of the coarse mesh (see gure 4).
x
U
_
e
p
s
(
x
,
y
=
0
.
4
)
0 1 0.5
-0.1
0
-0.05
P2-MsFEM
P1-FEM reference
x
U
_
e
p
s
(
x
,
y
=
0
.
4
)
0.09 0.1 0.095 0.105
-0.054
-0.052
-0.05
-0.048
-0.046
P2-MsFEM
P1-FEM
Cross sections along the side of the elements
Figure 4: Cross-section (left) and close-up (right) at y = 0.4 of the reference and multiscale
solutions: = 10
2
, h

= 1/5.
On the other hand, the oscillating behavior of the solution is well captured inside the coarse
mesh elements where the oscillations of the oscillating function w
,h
(x) are fully developed. The
cross-sections of gure 5 show a very good approximation of the solution.
x
U
e
p
s
(
x
,
y
=
0
.
5
)
0 1 0.5
-0.1
0
-0.05
P2-MsFEM
P1-FEMreference
x
|
|
U
e
p
s
,
h

-

U
r
e
f
|
|
H
1
0.18 0.19 0.2 0.21 0.22 0.23 0.24
-0.1
-0.09
-0.08
-0.095
-0.085
P2-MsFEM
P1-FEM
Figure 5: Cross section (left) and close-up (right) at y = 0.5 of the reference and multiscale
solutions: = 10
2
, h

= 1/5.
Very often in the literature, comparisons are made only on the value of the unknown u

and
not on the values of its partial derivatives. In the periodic setting, the asymptotic expansion
(2) shows that this comparison is too naive and simple because the corrector term is of order
which is precisely small. A good comparison just implies that the homogenized solution is well
captured. On the other hand, the gradient asymptotic expansion (3) shows that the corrector
term is of order 1 for the partial derivatives. Thus, a good comparison of the gradient implies
that, not only the homogenized solution is well captured, but also the local uctuations due
to the corrector term. It is therefore very important, in the periodic setting at least, to make
precise comparisons for the gradient eld gradu

or the ux A

gradu

. This is done on gures 6

and 7 where we see a remarkably good agreement, except at those nodes located at the coarse
16
cell boundaries. As already explained, this last eect is due to the applied ane boundary
conditions for the oscillating functions w
,h
. Indeed, the overshoots and undershoots on gures
6 and 7 take place precisely at the interface between two coarse mesh elements.
x
D
U
e
p
s
/
D
x
(
x
,
y
=
0
.
5
)
0 1 0.5
-1
0
1
-0.5
0.5
P2-MsFEM
P1-FEM
H
x
D
U
e
p
s
/
D
x
(
x
,
y
=
0
.
5
)
0.28 0.29 0.3 0.31 0.32 0.33 0.34
-0.4
-0.3
-0.2
-0.1
0
P2-MsFEM
P1-FEMreference
H*
Figure 6: Cross-section (left) and close-up (right) of the partial derivative U

/x at y = 0.5 of
the reference and multiscale solutions: = 10
2
, h

= 1/5. Here, H stands for the homogenized

solution.
x
D
U
e
p
s
/
D
y
(
x
,
y
=
0
.
5
)
0 1 0.5
-0.3
-0.2
-0.1
0
0.1
0.2
0.3 P2-MsFEM
P1-FEMreference
Figure 7: Cross-section of the partial derivative U

/y at y = 0.5 of the reference and

multiscale solutions: = 10
2
, h

= 1/5.
5.2 Two-phase composite material
Our multiscale nite element method is not designed for periodic homogenization but rather
for general numerical homogenization when no explicit asymptotic expansion of the solution
is known. As a model problem we consider the study of the conductivity properties (or anti-
plane elasticity properties) of a two-phase composite material made of spherical inclusions
in a background matrix. Both phases are isotropic with a high conductivity a

(x) = 100
in the inclusions and a lower one a

(x) = 1 in the matrix. The inclusions are randomly

distributed in the computational domain = (0, 1)
2
. On purpose we choose a large number
of 10
6
inclusions so that a direct computation is out of reach with a standard nite element
method. This numerical experiment corresponds to a minimum distance between two inclusions
of = 8 10
4
and a particle diameter of /2.
The right hand side of the elliptic boundary value problem is equal to zero and the solution
satises mixed boundary conditions :
Dirichlet ones on the sides denoted by
0
(u

= 0 on lower side and u

= 1 on the upper
one),
17

1

1

0
(a) (b)
Figure 8: (a) The coarse mesh of = (0, 1)
2
, (b) Close-up on the inclusions: h = 1/33 and
h

= h/500.
Homogeneous Neumann ones take place on the left and right sides denoted by
1
(a

grad u

n = 0 where n is the normal unit vector to ).

Figure 8 shows the approximated P
2
-MSFEM solution u

into a coarse element. It varies

almost linearly as the homogenized solution. As noticed before, a much more interesting
numerical result is the prole of the ux eld a

grad u

.
Figure 9: (a) The coarse mesh, (b) Flux density i.e. |a

grad u

|
R
2 in an element of the coarse
mesh and (c) a close-up.
As already explained, the ane boundary conditions for the oscillating functions breaks
their necessary oscillating character on the coarse cell boundaries. This generates localized
large errors for the gradient at those boundaries (see gures 6 and 7). In the present case, since
the conductivity jumps between the two phases, the errors are unacceptably large. Therefore,
in order to circumvent this diculty, we implemented another kind of boundary conditions
which gives a better approximation. Following an idea of [11] we solve 1-d elliptic problems on
each line line segment of K and these 1-d solutions are used as boundary conditions for the
oscillating functions w
,h
. The boundary conditions for those 1-d problems are of Dirichlet type:
the nodal values at the corners of K must be equal to x. In numerical practice, we approximate
the value of the oscillating functions on each side of K by piecewise linear functions. The
inclusions which intersect the sides of this element generate a partition composed of segments.
18
If s denotes the curvilinear coordinate on the coarse element boundary, on each segment we take
w
,K
i
= ps + q onK where p and q are constants. At the ends of two contiguous segments, we
write the conditions of continuity of the functions and of the tangential derivatives (i.e. along
the boundary of the element). This last condition is true only if the normal unit vector to K
coincide with the unit vector supported by the side of the element.
y
D
U
_
e
p
s
/
D
x
(
x
=
0
.
5
,
y
)
0.49 0.5 0.51 0.495 0.505 0.515
-1
0
-0.5
0.5
P2-MsFEM
y
D
U
_
e
p
s
/
D
y
(
x
=
0
.
5
,
y
)
0.49 0.5 0.51 0.495 0.505 0.515
0
1
2
0.5
1.5
P2-MsFEM
Figure 10: Close-up of the cross-sections of the partial derivatives U

/x(x = 0.5, y) and

U

/y(x = 0.5, y) of the multiscale solution : = 8 10

4
, h

= 1/33.
This modication of the boundary conditions for the oscillating functions w
,h
greatly
improves the precision of the multiscale nite element method at the interface betweens two
coarse mesh cells. For example, gure 9 displays the norm of the ux vector. At a very small
scale, one can clearly see the diusion channels between close inclusions. This proves that our
method is able to reproduce the ne details of the local uctuations.
Acknowledgments. The work of both authors has been supported by the GdR MoMaS
CNRS-2439 sponsored by ANDRA, BRGM, CEA and EDF whose support is gratefully ac-
knowledged.
References
[1] G. ALESSANDRINI, V. NESI, Univalent -harmonic mappings, Arch. Ration. Mech.
Anal. 158, 155171 (2001).
[2] G. ALLAIRE, Shape Optimization by the homogenization method, Applied Mathematical
Sciences, 146, Springer, (2002).
[3] T. ARBOGAST, Numerical subgrid upscaling of two-phase ow in porous media, in Nu-
merical treatment of multiphase ows in porous media, Lecture Notes in Physics, vol. 552,
Chen, Ewing and Shi eds., pp.35-49 (2000).
[4] I. BABUSKA, Solution of interface problems by homogenization I, II, III, SIAM J. Math.
Anal. 7, pp. 603634, pp. 635-645 (1976) and 8, pp. 923-937 (1977).
[5] A. BENSOUSSAN, J.L. LIONS and G. PAPANICOLAOU, Asymptotic analysis for peri-
odic structures, Studies in Mathematics and its applications, Vol. 5, North-Holland Pub-
lishing company, (1978).
[6] Y. CAPDEBOSCQ, M. VOGELIUS, Wavelet Based Homogenization of a 2 Dimensional
Elliptic Problem, to appear.
19
[7] P. CIARLET, The nite element methods for elliptic problems, North-Holland, Amsterdam
(1978).
[8] W. E, B. ENGQUIST, The heterogeneous multiscale methods, Commun. Math. Sci. 1,
87132 (2003).
[9] Y. EFENDIEV, T. HOU, X.-H. WU, Convergence of a nonconforming multiscale nite
element method, SIAM J. Numer. Anal. 37, 888910 (2000).
[10] Y. EFENDIEV, X.-H. WU, Multiscale nite element methods for problems with highly
oscillatory coecients, Numerische Math., vol.90(3), 459-486 (2002).
[11] T. Y. HOU, X.-H. WU, A multiscale nite element method for elliptic problems in com-
posite materials and porous media, Journal of computational physics 134, 169-189, (1997).
[12] T. Y. HOU, X.-H. WU, Z. CAI, Convergence of a multiscale nite element method for
elliptic problems with rapidly oscillating coecients, Math. of Comp. 68, 913-943, (1999).
[13] A.-M. MATACHE, I. BABUSKA, C. SCHWAB, Generalized p-FEM in homogenization,
Numer. Math. 86, 319375 (2000).
[14] A.-M. MATACHE, C. SCHWAB, Two-scale FEM for homogenization problems, M2AN
Math. Model. Numer. Anal. 36, 537572 (2002).
[15] F. MURAT, L. TARTAR, H-convergence, in Topics in the mathematical modeling of com-
posite materials, A. Cherkaev and R.V. Kohn eds., series : Progress in Nonlinear Dif-
ferential Equations and their Applications, Birkha user, Boston 1997. French version :
mimeographed notes, seminaire dAnalyse Fonctionnelle et Numerique de lUniversite
dAlger (1978).
[16] P. H. RABINOWITZ, Theorie du degre topologique et applications ` a des probl`emes aux
limites non lineaires, redige par H. BERESTYCKI, Publications du Laboratoire dAnalyse
Numerique, Universite de Paris VI, 4, place Jussieu, Paris, N
o
75010.
20