Preface

The theory of homogenization of partial differential equations is a concept

that deals with the study of features that are different at different length
scales. For instance, in material science, homogenization deals with the study
of the macroscopic behaviour of a composite medium through its microscopic
properties. Fiberglass, bone etc. are examples of composite material. The
known and unknown quantities in the study of physical processes in a medium
with micro-structure depend on a small parameter ε = Ll , where L is the
macroscopic scale length of the dimension of a specimen of the medium and
l is the characteristic length of the medium configuration. The physical
parameters such as conductivity, elasticity etc. are discontinuous and switch
rapidly between different values across a small length scale ε. The study of
the limit, as ε → 0, is the aim of the mathematical theory of homogenization.
Though the case ε → 0 has no real physical meaning, it is important as a
tool for numerical computations. The origin of the word “homogenization”
is related to the question of replacing a heterogeneous medium by a fictitious
homogeneous one (the ‘homogenized’ material) for computational purposes.
Let Ω be an open bounded subset of Rn . The optimal design problem is
to find a subset Ω1 ⊂ Ω such that the “cost” functional
Z
J(a) = F (x, u(x)) dx
Ω

attains its minimum, where u(x) is the solution of the Dirichlet problem

−div(A(x)∇u(x)) = f (x) in Ω
(0.0.1)
u =0 on ∂Ω,
A(x) = a(x)I and (
α on Ω1
a(x) =
β on Ω \ Ω1 .

i
CHAPTER 0. PREFACE ii

The functional J is minimized over all subsets of Ω. The constants α, β

are the isotropic conductivity of the material Ω1 and its complement in Ω,
respectively. The optimal design problem concerns with finding the optimal
mix of the material constituents in Ω such that it optimizes some “property”
(say, energy) of the system given by J.
The existence of an optimal mixture of the material constituents was
proved in [Che75] under some regularity hypotheses on Ω. However, when no
regularity assumptions are made, then depending on the J, there may arise
a situation that we have no optimal mixture of the materials (cf. [Mur71,
Mur72]). The situation of no optimal solution corresponds to the case where
the corresponding conductivity coefficient a is no longer isotropic. This sit-
uation corresponds to the case where the material is mixed finely to form a
heterogeneous material. Physically, the material could be an alloy formed
by two material with conductivity α and β. Though, from microscopic point
of view this is still a mixture of two materials, from the macroscopic view
it behaves completely different, with new properties, from the original con-
stituents.
Thus, the situation of ‘no optimal solution’ motivates us to study the
Dirichlet problem (0.0.1) for a heterogeneous material Ω. This is a classical
second order elliptic boundary value problem and admits a unique solution.
However, note that for a heterogeneous material the coefficient a(x) oscillate
rapidly. Thus, when we try to compute the solution for (0.0.1), we need to use
grid or mesh at a scale much smaller than the scale of the mixture, which may
be practically impossible, leading to large errors in our computation. The
mathematical theory of homogenization ‘averages out’ the heterogeneities
and studies an ‘equivalent’ homogeneous fictitious material whose behaviour
reflects that of the original material, when the number of fibres is very large.
Homogenization, as a mathematical discipline, took shape only in the
last three decades but the physical ideas of homogenization date back at
least to [Poi22, Mos50, Max73, Cla79, Ray92]. A very good historical record
of works related to homogenization until 1975 can be found in [Bab76] and
the references therein.
An abstract theory of homogenization was introduced by S. Spagnolo
in a paper of 1967 (cf. [Spa67]) under the name of G-convergence1 (also
cf. [Spa68, GS73, Spa76]) and further generalised as H-convergence by L. Tar-
tar in [Tar77] and developed by F. Murat and L. Tartar (cf. [Mur78b, MT97]).
1
The terminology denoting the convergence of Green’s operators for boundary problems
CHAPTER 0. PREFACE iii

There is also a variational theory of homogenization, known as Γ-convergence,

proposed by Ennio De Giorgi in a sequence of papers (cf. [GS73, Gio75,
GF75]). For a thorough introduction to this theory we refer to [Gio84, Att84,
DM93, BD98]. The wide spread application and theory of homogenization
can also be found in [BLP78, JKO94, Hor97, CD99, CP99].
CHAPTER 0. PREFACE iv
Contents

Preface i

Notations vii

1 Asymptotic Expansion 1
1.1 Periodically Oscillating Functions . . . . . . . . . . . . . . . . 1
1.2 Second Order Elliptic Equation . . . . . . . . . . . . . . . . . 3
1.3 Periodic Boundary Conditions . . . . . . . . . . . . . . . . . . 4
1.4 Periodic Composite Material . . . . . . . . . . . . . . . . . . . 5
1.5 Asymptotic Expansion in Two Scales . . . . . . . . . . . . . . 7

2 Two-Scale Convergence 19
2.1 Vector-valued Function Spaces . . . . . . . . . . . . . . . . . . 19
2.2 Two-scale Convergence . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Classical Definition of Two-Scale Convergence . . . . . . . . . 32
2.4 Homogenization of Second Order Linear Elliptic Problems . . 33
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3 H-Convergence 39
3.1 Coercive Operators . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 H-Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Correctors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Generalised Energy Convergence . . . . . . . . . . . . . . . . . 58
3.5 Optimal Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4 Γ-Convergence 63
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Direct Method of Calculus of Variation . . . . . . . . . . . . . 63

v
CONTENTS vi

4.3 Sequential Γ-Convergence . . . . . . . . . . . . . . . . . . . . 68

4.4 Integral Representation: One Dimension . . . . . . . . . . . . 72

5 Bloch-Floquet Homogenization 77
5.1 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Schrödinger Operator with Periodic Potential . . . . . . . . . 78
5.2.1 Direct Integral Decomposition . . . . . . . . . . . . . . 79
5.3 Bloch Periodic Functions . . . . . . . . . . . . . . . . . . . . . 83
5.4 Bloch Transform . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4.1 Spectrum of Elliptic Operator . . . . . . . . . . . . . . 88
5.4.2 Regularity of λm (η) and φ1 (·, η) . . . . . . . . . . . . . 90
5.4.3 Taylor Expansion of Ground State . . . . . . . . . . . 92
5.5 Homogenization of Second order Elliptic Operator . . . . . . . 94

Appendices 101

Bibliography 103

Index 109
Notations

Symbols

Rn denotes the n-dimensional Euclidean space over R. {e1 , . . . , en } is the

standard basis of Rn

Ω denotes an open bounded subset of Rn

∂Ω denotes the boundary of Ω

|E| is the Lebesgue measure of a subset E ⊂ Rn . However, for a vector

n
n
pP
2
ξ ∈ R , |ξ| = i=1 ξi

M (α, β, Ω) denotes, for 0 < α < β, the class of all n × n matrices, A = A(x),
with L∞ (Ω) entries such that,

α|ξ|2 ≤ A(x)ξ · ξ and |A(x)ξ| ≤ β|ξ| for a.e. x ∀ξ ∈ Rn

t
A denotes the transpose of a matrix A

Function Spaces

D(Ω) or Cc∞ (Ω) is the class of all infinitely differentiable functions on Ω with
compact support

D0 (Ω) is the topological dual of D(Ω), the space of all distributions

Cper (Y ) denotes the class of Y -periodic functions in C(Rn )

H01 (Ω) is the closure of D(Ω) in W 1,2 (Ω) = (H 1 (Ω)) and its norm is denoted
by k.kH01 (Ω)

H −1 (Ω) is the dual space of H01 (Ω) and its norm is denoted by k · kH −1 (Ω)

vii
NOTATIONS viii

L∞ (Ω) is the space of all essentially bounded measurable functions and its
norm is denoted by k.k∞,Ω

Lp (Ω) is the space of all p-summable measurable functions and its norm is
denoted by k.kp,Ω (1 ≤ p < ∞)

Lp (Ω; X)
R denotes the class of all measurable functions f : Ω → X such that
p
Ω
kf (x)kX < ∞, where X is a Banach space

Lpper (Y ) denotes the class of Y -periodic functions in Lploc (Rn ) and its norm
denoted by k.kp,Y (1 ≤ p ≤ ∞)

W m,p (Ω) is the collection of all Lp (Ω) functions such that all distributional
derivatives upto order m are also in Lp (Ω) and its norm is denoted by
k.km,p,Ω
m,p
Wper (Y ) denotes the class of Y -periodic functions in W m,p (Rn ) and its norm
denoted by k · km,p,Y

General Conventions

h·, ·iX ? ,X denotes the duality pairing between X ? and X

→ will denote the convergence in the strong topology of the space

* will denote the convergence in the weak topology of the space

C0 is a generic positive constant independent of the parameters w.r.t

which a limit is taken; will be different in different inequalities

X? denotes the topological dual (space of continuous linear functionals)

of the space X
Chapter 1

Asymptotic Expansion

We begin the study of homogenization by considering a linear second order

elliptic problem in a domain with periodic structures. More precisely, the co-
efficients of the PDE have rapid periodic oscillation. The aim of this chapter
is to develop the two-scale asymptotic expansion. The mathematical justifi-
cation of the asymptotic expansion will be discussed in subsequent chapters.
The periodic framework models the case where the heterogeneities are very
small with respect to the size of the domain and are evenly distributed. This
is a realistic assumption for large class of applications. Some good references
on periodic homogenization are [BLP78, CD99, JKO94].

1.1 Periodically Oscillating Functions

Let us build tools required to model rapid oscillations of periodic functions.
Let us assume that Y = Πni=1 [0, li ) is a reference cell (or period) in Rn .

Definition 1.1.1. A function f : Rn → Rm is said to be Y -periodic if, for

all i = 1, 2, . . . , n, f (x + kyi ei ) = f (x) for a.e. x ∈ Rn and for all k ∈ Z.

For simplicity, we shall, henceforth, take the reference cell to be the unit
cube of Rn , i.e., Y = [0, 1]n . This is only for simplicity and to avoid carrying
the measure of Y , |Y |, in our calculations. Note that Rn = ∪k∈Zn (k + Y ),
is a disjoint union. Any function f : Y → R, defined a.e. on Y , may be
extended a.e. to Rn as a Y -periodic function. Let Lpper (Y ) denote the set
of all Lploc (Rn ) which are Y -periodic equipped with the norm of Lp (Y ). For
any f ∈ Lpper (Y ) and ε > 0, we may define a new function fε : Rn → R as

1
CHAPTER 1. ASYMPTOTIC EXPANSION 2

fε (x) = f (x/ε). Observe that fε is εY -periodic in Rn because, for all k ∈ Z

and i = 1, 2, . . . , n,
x  x
fε (x + kεei ) = f + kei = f = fε (x).
ε ε
The second equality is due to the Y -periodicity of f . Observe that fε on Rn
has increased number of oscillations, if any1 , compared to f .
Example 1.1. Consider the function f on [0, 1] defined as
(
1 [0, 1/2)
f (y) =
−1 [1/2, 1]

extended to all of R. Define, for any 0 < ε < 1, the new function fε (x) =
f (x/ε) on R. Note that the number of points of jump discontinuity for f in
[0, 1] is only one at y = 1/2. However, fε has more than one point of jump
discontinuity in [0, 1]. For instance, for ε = 1/2, the function fε has three
points of jump discontinuity, at x = 1/4, 1/2, 3/4, in [0, 1].
Example 1.2. Consider f (y) = sin(2πy) on [0, 1] extended to all of R. Define
fε (x) = sin(2πx/ε) on R. For any 0 < ε < 1, we see that the number of
oscillations on [0, 1] is increased for fε . For instance, for ε = 1/2, fε has twice
the number of oscillations, as that of f , on [0, 1].
Theorem 1.1.2. Let 1 ≤ p ≤ +∞ and f ∈ Lpper (Y ). Then fε (x) = f (x/ε),
for 0 < ε < 1, is bounded in any open cell R that contains any translation of
Y , i.e.,
|R|
kfε kpp,R ≤ C0 kf kpp,Y
|Y |
where C0 depends only on n, the dimension of Euclidean space.

Theorem 1.1.3. Let 1 ≤ p ≤ +∞, f ∈ Lpper (Y ) and set fε (x) = f (x/ε), for
0 < ε < 1, on Rn . For 1 ≤ p < ∞,
Z
1
fε * f (y) dy weakly in Lp (Ω)
|Y | Y
for any bounded open subset Ω ⊂ Rn . If p = ∞, then
Z
1
fε * f (y) dy weak-* in L∞ (Ω).
|Y | Y
1
constant functions have no oscillations
CHAPTER 1. ASYMPTOTIC EXPANSION 3

1.2 Second Order Elliptic Equation

Let Ω be an open bounded subset of Rn and let ∂Ω denote the boundary of
Ω. For any given 0 < α < β, let M (α, β, Ω) denote the class of all n × n
matrices, A = A(x), with L∞ (Ω) entries such that,
α|ξ|2 ≤ A(x)ξ.ξ ≤ β|ξ|2 a.e. x ∀ξ ∈ Rn .

Recall the following result on variational inequality on a Hilbert space.

Refer [KS00] for a complete theory on variational inequality.
Theorem 1.2.1. Let a(x, y) be a coercive bilinear form on H, K ⊂ H be
a closed and convex subset of H and f ∈ H ? . Then there exists a unique
solution x ∈ K to
a(x, y − x) ≥ hf, y − xi, ∀y ∈ K. (1.2.1)
The case K = H in, the above result, is popularly known as Lax-Milgram
result. In the case K = H and z ∈ H, by choosing y = x + z and y = x − z,
by turn, in (1.2.1), we have the equality a(x, z) = hf, zi for all z ∈ H and
for every given f ∈ H ? .
The Lax-Milgram result implies the existence and uniqueness of a weak
solution, u ∈ H01 (Ω), to the second order elliptic equation with Dirichlet
boundary condition,

−div(A(x)∇u(x)) = f (x) in Ω
(1.2.2)
u(x) = 0 on ∂Ω,
where A ∈ M (α, β, Ω) and f ∈ H −1 (Ω). In fact, one also has the estimate
1
kukH01 (Ω) ≤ kf kH −1 (Ω) . (1.2.3)
α
The bounded elliptic operator A : H01 (Ω) → H −1 (Ω), defined as A =
−div(A(x)∇), is an isomorphism and the norm of A−1 is not larger than
α−1 (cf. (1.2.3)). Moreover, the weak solution u of (1.2.2) can also be char-
acterized as the minimizer in H01 (Ω) of the functional J : H01 (Ω) → R defined
as Z
1
J(v) = A∇v.∇v dx − hf, viH −1 (Ω),H 1 (Ω) ,
2 Ω 0

i.e.,
J(u) = min 1
J(v).
v∈H0 (Ω)
CHAPTER 1. ASYMPTOTIC EXPANSION 4

1.3 Periodic Boundary Conditions

Let Y = [0, 1)n be the unit cell of Rn and let, for each i, j = 1, 2, . . . , n,
aij : Y → R and A(y) = (aij ). For any given f : Y → R, extended Y -
periodically to Rn , we want to solve the problem

−div(A(y)∇u(y)) = f (y) in Y
(1.3.1)
u is Y − periodic.
The condition u is Y -periodic is equivalent to saying that u takes equal values
on opposite faces of Y .
∞
Let us now identify the solution space for (1.3.1). Let Cper (Y ) be the
∞ n 1
set of all Y -periodic functions in C (R ). Let Hper (Y ) denote the closure
∞
of Cper (Y ) in the H 1 -norm. Being a second order equation, in the weak
1
formulation, we expect the weak solution u to be in Hper (Y ). Note that if
u solves (1.3.1) then u + c, for any constant c, also solves (1.3.1). Thus, the
1
solution will be unique up to a constant in the space Hper (Y ). Therefore, we
1
define the quotient space Wper (Y ) = Hper (Y )/R as solution space where the
solution is unique.
Solving (1.3.1) is to find u ∈ Wper (Y ), for any given f ∈ (Wper (Y ))? in
the dual of Wper (Y ), such that
Z
A∇u · ∇v dx = hf, vi(Wper (Y ))? ,Wper (Y ) ∀v ∈ Wper (Y ).
Y

The requirement that f ∈ (Wper (Y ))? is equivalent to saying that

Z
f (y) dy = 0
Y

because f defines a linear functional on Wper (Y ) and f (0) = 0, where 0 ∈

1
Hper (Y )/R. In particular, the equivalence class of 0 is same as the equivalence
class 1 and hence Z
f (y) dy = hf, 1i = hf, 0i = 0.
Y

Theorem 1.3.1. Let Y be unit open cell and let aij ∈ L∞ (Ω) such that the
matrix A(y) = (aij (y)) is elliptic with ellipticity constant α > 0. For any
f ∈ (Wper (Y ))? , there is a unique weak solution u ∈ Wper (Y ) satisfying
Z
A∇u · ∇v dx = hf, vi(Wper (Y ))? ,Wper (Y ) ∀v ∈ Wper (Y ).
Y
CHAPTER 1. ASYMPTOTIC EXPANSION 5

Note that the solution u we find from above theorem is an equivalence

class of functions which are all possible solutions. Any representative element
from the equivalence class is a solution. All the elements in the equivalence
differ by a constant. Let u be an element from the equivalence class and let
c be the constant Z
1
c= u(y) dy.
|Y | Y
R
Thus, we have u−c is a solution with zero mean value in Y , i.e., Y u(y) dy =
0. Therefore, rephrasing (1.3.1) as

 −div(A(y)∇u(y)) = f (y) in Y
u is Y − periodic
1
R

|Y | Y
u(y) dy = 0

we have unique solution u in the solution space

 Z 
1 1
Vper (Y ) = u ∈ Hper (Y ) | u(y) dy = 0 .
|Y | Y

Remark 1.3.2. By identifying the cell Y with an equivalent Torus, (1.3.1)

may be viewed as posed on the Torus. This formulation has the advantage
that the equation has no boundary condition because Torus has no boundary.

1.4 Periodic Composite Material

In this section, we mathematically model a periodic composite material. Let
Ω ⊂ Rn denote a periodic composite material. For simplicity, let us consider
the case of a composite material which is a mixture of two materials. Let
Y = [0, 1)n be the reference cell which is a mixture of materials Y1 and Y2
such that Y 1 ∪ Y 2 = Y and Y1 ∩ Y2 = ∅. For each 0 < ε < 1, the dilated cell
εY can be used to tile Rn , so that, Ω is also tiled using εY modelling the
periodic distribution of its constituents, for some ε very small (cf. Fig. 1.1).
Let us consider the second order elliptic problem with Dirichlet boundary
condition on Ω, for a given f ∈ H −1 (Ω),

−div(A(x)∇u) = f (x) in Ω
u =0 on ∂Ω,
CHAPTER 1. ASYMPTOTIC EXPANSION 6

εY
y

x
Figure 1.1: partition of Ω into ε-cells

where A(x) = (aij (x)) is a n × n matrix of measurable L∞ (Ω) functions such

that
α|ξ|2 ≤ A(x)ξ.ξ ∀ξ ∈ Rn .
We shall now observe that when Ω is a periodic composite material, the
functions aij , for all i, j = 1, 2, . . . , n, are rapidly oscillating periodic func-
tions. For each i, j ∈ {1, 2, . . . , n}, we are given measurable L∞ (Y ) functions
aij : Y → R with different values on Y1 and Y2 and
n
X
2
α|ξ| ≤ aij (y)ξi ξj a.e. y ∈ Y ∀ξ ∈ Rn .
i,j=1

The condition given above is called the ellipticity condition. We extend aij
to all of Rn , and for each 0 < ε < 1, we define the function aεij : Rn → R as
x
aεij (x) = aij a.e. x ∈ Rn
ε
and the n × n matrix Aε (x) = (aεij (x)) is in M (α, β, Ω). Thus, the Dirichlet
problem for the composite material Ω is given as, for a given f ∈ H −1 (Ω),

−div(Aε (x)∇uε (x)) = f (x) in Ω
(1.4.1)
uε = 0 on ∂Ω.
CHAPTER 1. ASYMPTOTIC EXPANSION 7

By Lax-Milgram result, there exists a unique solution uε ∈ H01 (Ω) such that

Z
Aε (x)∇uε (x) · ∇v(x) dx = hf, viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω)
0
Ω

and kuε kH01 (Ω) ≤ 1/αkf kH −1 (Ω) . Computing the solution, numerically, is
stable if the size of the grid is chosen smaller than ε. But for composite
materials, ε is very small and choosing grid smaller than ε leads to impossible
computable situation. Therefore, we study the limiting case, as ε → 0, of
the Dirichlet problem (1.4.1).

1.5 Asymptotic Expansion in Two Scales

Note that in the periodic set-up any x ∈ Ω has two reprsentations. One is
the macroscale representation x and the other is that x is in some translation
of the εY cell having the form x = εy for some y ∈ Y . Thus, any x ∈ Ω may
take two representations each in Ω and Y as x and y = xε , respectively. Thus,
the behaviour of a periodic composite material Ω, as given in (1.4.1), involves
two scales, viz., the “macroscopic or slow” scale x ∈ Ω and the “microscopic
or fast” scale y = x/ε ∈ Y . We intend to find uε (x) that solves (1.4.1).
Thus, our model suggests that uε depends on both the slow variable x and
fast variable y = x/ε, viewed as independent variables. This suggests us to
seek uε (x), with x ∈ Ω, in the form

 x  x  x
uε (x) = u0 x, + εu1 x, + ε2 u2 x, + ..., (1.5.1)
ε ε ε

where ui (x, y) are functions which are Y -periodic in the y-variable. The
partial derivative of a function φε (x) := φ(x, x/ε) is given as

∂φε (x) ∂φ(x, y) ∂φ(x, y) ∂yi ∂φ(x, y) ∂φ(x, y) 1 ∂φ(x, y)

= = + = + .
∂xi ∂xi ∂xi ∂xi ∂yi ∂xi ε ∂yi
CHAPTER 1. ASYMPTOTIC EXPANSION 8

The second order operator of the equation (1.4.1) can be rewritten as

n  
X ∂ ε ∂
−Aε = − aij (x)
i,j=1
∂xi ∂xj
n    
X ∂ 1 ∂ ∂ 1 ∂
= − + aij (y) +
i,j=1
∂xi ε ∂yi ∂xj ε ∂yj
n   n  
X ∂ ∂ −2
X ∂ ∂
= − aij (y) −ε aij (y)
i,j=1
∂xi ∂xj i,j=1
∂yi ∂yj
" n n #
X ∂  ∂
 X
∂

∂
− ε−1 aij (y) + aij (y)
i,j=1
∂x i ∂y j i,j=1
∂y i ∂xj
= A0 + ε−2 A2 + ε−1 A1 .

Substituting in (1.4.1), we get

(ε−2 A2 + ε−1 A1 + A0 ) u0 (x, y) + εu1 (x, y) + ε2 u2 (x, y) + . . . = f (x).

 

Now, equating like powers of ε both side, we get

A2 u0 = 0 (1.5.2)
A2 u1 + A1 u0 = 0 (1.5.3)
A2 u2 + A1 u1 + A0 u0 = f (x) (1.5.4)
A2 um+2 + A1 um+1 + A0 um = 0 ∀m ≥ 1. (1.5.5)

We first solve for u0 (·, y) in Y using the first equation


A2 u0 (·, y) = 0 in Y
u0 (·, y) is Y -periodic in y.

Since 0 ∈ (Wper (Y ))? , by Theorem 1.3.1, we have u0 (·, y) ∈ Hper1

(Y ) which
are unique up to constant. Thus, 0 ∈ Wper (Y ) is a solution. Therefore,
u0 (·, y) is independent of y and hence u0 (x, y) = u(x), a function of x.
We now proceed to find u1 using the problem

 A2 u1 (x, y) = −A1 u(x) in Y
u1 (·, y) is Y -periodic in y
 R
u (x, y) dy = 0.
Y 1
CHAPTER 1. ASYMPTOTIC EXPANSION 9

We first need to check that −A1 u ∈ (Wper (Y ))? or, equivalently,

Z
A1 u dy = 0.
Y

We simplify the RHS using the fact that u is a function of x,

n   n
X ∂ ∂u(x) X ∂aij (y) ∂u(x)
A1 u(x) = − aij (y) =−
i,j=1
∂yi ∂xj i,j=1
∂yi ∂xj
n n
!
X X ∂aij (y) ∂u(x)
= − .
j=1 i=1
∂y i ∂x j

Consider
Z
A1 u(x) dy = hA1 u(x), 1i
Y
Z X n  
∂ ∂u(x)
= − aij (y) · 1 dy
Y i,j=1 ∂yi ∂xj
n Z
X ∂u(x) ∂1
= aij (y) = 0.
i,j=1
∂xj Y ∂yi

Since −A1 u ∈ (Wper (Y ))? , by Theorem 1.3.1, we have u1 (·, y) ∈ Hper 1

(Y )
which are unique up to constant. Since the operator A2 depends on y
variable and uxj is independent of y, we are motivated to define, for each
j = 1, 2, . . . , n, the auxiliary periodic function χj as a solution to the prob-
lem
P ∂aij (y)
A2 χj (y) = − ni=1 ∂y

 i
in Y
χj (y) is Y -periodic in y (1.5.6)
 R
χ (y) dy = 0
Y j

or equivalently,

R j − yj ) = 0 in Y
 div(A(y)∇(χ
1
χ (y) dy = 0
|Y | Y j
χj − yj is Y -periodic.

CHAPTER 1. ASYMPTOTIC EXPANSION 10

Substituting this in the equation of u1 , we get

n
X ∂u(x)
A2 u1 (x, y) + A1 u(x) = A2 u1 (x, y) + A2 χj (y)
j=1
∂xj
n
!
X ∂u(x)
= A2 u1 (x, y) + A2 χj (y)
j=1
∂xj
n
!
X ∂u(x)
= A2 u1 (x, y) + χj (y) .
j=1
∂xj

Therefore, we have

n
X ∂u(x)
u1 (x, y) = − χj (y) + ũ(x)
j=1
∂xj

for some function ũ(x). Finally, we solve for u2 in the problem


 A2 u2 (x, y) = f (x) − A1 u1 (x, y) − A0 u(x) in Y
u2 (·, y) is Y -periodic in y
 R
u (x, y) dy = 0.
Y 2

For the above equation to be solvable, by Theorem 1.3.1, it is necessary that

f (x) − A1 u1 (x, y) − A0 u(x) ∈ (Wper (Y ))? or, equivalently,

Z Z
f (x) = (A1 u1 (x, y) + A0 u(x)) dy.
Y Y
CHAPTER 1. ASYMPTOTIC EXPANSION 11

To check this, let us first consider

Z
I := A1 u1 (x, y) dy
Y
n Z     
X ∂ ∂u1 ∂ ∂u1
= − aij (y) + aij (y) dy
i,j=1 Y ∂x i ∂y j ∂y i ∂x j
n Z   Z 
X ∂ ∂u1 ∂u1 ∂1
= − aij (y) dy + aij (y) dy
i,j=1 Y ∂x i ∂y j Y ∂x j ∂y i

n
(Z " n
!# )
X ∂ ∂ X ∂u(x)
= − aij (y) ũ(x) − χk (y) dy
i,j=1
∂xi Y ∂yj k=1
∂xk
n Z   
X ∂ ∂χk (y) ∂u(x)
= aij (y) dy
i,j,k=1
∂xi Y ∂yj ∂xk
n Z    2
X ∂χk (y) ∂ u(x)
= aij (y) dy .
i,j,k=1 Y ∂yj ∂xi ∂xk

Next, we consider
Z Z X n  
∂ ∂u(x)
A0 u(x) dy = − aij (y) dy
Y Y i,j=1 ∂xi ∂xj
n Z  2
X ∂ u(x)
= − aij (y) dy .
i,j=1 Y ∂x i ∂x j

For convenience, we rewrite the above relation by replacing the index j with
k to get
Z n Z  2
X ∂ u(x)
A0 u(x) dy = − aik (y) dy .
Y i,k=1 Y ∂xi ∂xk

Therefore, we need to check that

n
"Z n
! #
∂ 2 u(x)
Z X X ∂χk (y)
f (x) dy = aij (y) − aik (y) dy
Y i,k=1 Y j=1
∂yj ∂xi ∂xk
n
X ∂ 2 u(x)
f (x) = − a0ik ,
i,k=1
∂xi ∂xk
CHAPTER 1. ASYMPTOTIC EXPANSION 12

where !
Z n
1 X ∂χk (y)
a0ik = aik (y) − aij (y) dy. (1.5.7)
|Y | Y j=1
∂yj

Proposition 1.5.1. The n × n matrix A0 with the entries (1.5.7) satisfies

the ellipticity condition.

Proof. Let aY : Wper (Y ) × Wper (Y ) → R be a bilinear form defined as

Z
aY (φ, ψ) = A(y)∇y φ(y) · ∇y ψ(y) dy.
Y

Now, by the weak formulation of (1.5.6), χj ∈ Wper (Y ) satisfies aY (χj , ψ) =

aY (yj , ψ) forall ψ ∈ Wper (Y ). Choosing the test function ψ to be χi , we get
aY (χj − yj , χi ) = 0. On the other hand,
Z
0
|Y |aij = A(y)∇y (yj − χj (y)) · ∇y yi dy = aY (yj − χj , yi ).
Y

Hence, we see that |Y |a0ij = aY (yj − χj , yi − χi ). Therefore, for any non-zero

vector ξ ∈ Rn ,
n
X
A0 ξ · ξ = a0ij ξj ξi
i,j=1
n Z
X 1
= A(y)∇y (ξj [yj − χj (y)]) · ∇y (ξi [yi − χi (y)]) dy
i,j=1
|Y | Y
Z n
! n
!
1 X X
= A(y)∇y ξj [yj − χj ] · ∇y ξi [yi − χi ] dy
|Y | Y j=1 i=1

α|ξ|2
Z
≥ |∇y η|2 dy = α0 |ξ|2
|Y | Y

where Z
α
α0 = |∇y η|2 dy
|Y | Y
Pn ξi
and η(y) := i=1 |ξ| (yi − χi (y)). To show the ellipticity of (a0ij ), it is enough
to show that α0 > 0. On the contrary, suppose α0 = 0 then |∇y η(y)| = 0
for a.e. y ∈ Y . Since Y is connected, η is constant on Y , say C0 . Thus,
CHAPTER 1. ASYMPTOTIC EXPANSION 13

Pn Pn
i=1 ξi yi = i=1 ξi χi (y) + |ξ|C0 . Since χi is such that its average is zero, we
obtain
n Z
X 1
yi dy = |ξ|C0 .
i=1
|Y | Y

This is not possible for non-zero ξ due to the periodicity of χi . WHY???????

From the result proved abvoe and the fact that f ∈ H −1 (Ω), the equation

n
X ∂ 2 u(x)
− a0ik = f (x) (1.5.8)
i,k=1
∂xi ∂xk

is admits a unique solution, by Lax-Milgram result, u(x) ∈ H01 (Ω). Observe

that the form of (1.5.8) is similar to (1.4.1), but without ε dependence. Thus,
(1.5.8) is called the homogenized form (1.4.1). The homogenized operator
(or effective coefficient) A0 = (a0ik ) is computed by first computing χk in the
cell Y using (1.5.6) and using (1.5.7) to compute A0 . Note that a0ik are all
constant and hence the homogenized equation has constant coefficients. But
we caution here that this is very specific to the periodic case.
Now that we have checked that f (x) − A1 u1 (x, y) − A0 u(x) ∈ (Wper (Y ))? ,
1
by Theorem 1.3.1, we have u2 (·, y) ∈ Hper (Y ) which are unique up to con-
stant. We wish to solve for u2 using the equation

A2 u2 (x, y) = f (x) − A1 u1 (x, y) − A0 u(x) in Y.

Simplifying, as before, we get

n
" n #
X X ∂χk (y) ∂ 2 u(x)
A1 u1 (x, y) + A0 u(x) = aij (y) − aik (y)
i,k=1 j=1
∂yj ∂xi ∂xk
n  
X ∂ ∂u1
− aij (y)
i,j=1
∂yi ∂xj
CHAPTER 1. ASYMPTOTIC EXPANSION 14

We first compute the term corresponding to u1 ,

n n 
∂ 2 u1
  
X ∂ ∂u1 X ∂aij (y) ∂u1
aij (y) = + aij (y)
i,j=1
∂y i ∂x j i,j=1
∂y i ∂x j ∂yi ∂xj
n n
X ∂aij (y) ∂ ũ(x) X ∂aij (y) ∂ 2 u(x)
= − χk (y)
i,j=1
∂yi ∂xj i,j,k=1
∂yi ∂xj ∂xk
n
X ∂χk (y) ∂ 2 u(x)
− aij (y)
i,j,k=1
∂yi ∂xj ∂xk
n n
X ∂aij (y) ∂ ũ(x) X ∂[aij (y)χk (y)] ∂ 2 u(x)
= −
i,j=1
∂yi ∂xj i,j,k=1
∂yi ∂xj ∂xk

Therefore the equation for u2 becomes, A2 u2 (x, y) =

n
X ∂aij (y) ∂ ũ(x)
f (x) +
i,j=1
∂yi ∂xj
n
" n  #
∂ 2 u(x)

X X ∂χk (y) ∂[aji (y)χk (y)]
− aij (y) + − aik (y) .
i,k=1 j=1
∂yj ∂yj ∂xi ∂xk

Now, using the homogenized equation (1.5.8) for f in the above relation, we
get A2 u2 (x, y) =
n
" n  #
∂ 2 u(x)

X X ∂χk (y) ∂[aji (y)χk (y)]
− aij (y) + − aik (y) + a0ik
i,k=1 j=1
∂yj ∂yj ∂xi ∂xk
n
X ∂aij (y) ∂ ũ(x)
+
i,j=1
∂yi ∂xj

As before, we are motivated to define, for each i, k = 1, 2, . . . , n, the

auxiliary periodic function θik as a solution to the problem
 P  
∂[aji (y)χk (y)]

 A2 θik (y) = aik (y) − a0ik − nj=1 aij (y) ∂χ∂yk (y)
j
+ ∂yj
in Y
 R θik (y) is Y -periodic in y
θ (y) dy = 0.

Y ik
(1.5.9)
CHAPTER 1. ASYMPTOTIC EXPANSION 15

Substituting the auxiliary problems (1.5.9) and (1.5.6) in the equation of

u2 , we get
n n
!
2
X ∂ u(x) X ∂ ũ(x)
A2 u2 (x, y) = A2 θik − χj (y)
i,k=1
∂xi ∂xk j=1 ∂xj

Therefore, we have
n n
X ∂ 2 u(x) X ∂ ũ(x)
u2 (x, y) = θik − χj (y) + ũ2 (x)
i,k=1
∂xi ∂xk j=1 ∂xj

for some function ũ2 (x). This way one can proceed for all values of uk , using
the recurrence relation (1.5.5). Finally, substituting the values of uk (x, y) in
(1.5.1), we have
n
!
X ∂u(x)
uε (x) = u(x) + ε − χj (y) + ũ(x)
j=1
∂x j

n n
!
2
X ∂ u(x) X ∂ ũ(x)
+ε2 θik − χj (y) + ũ2 (x) + . . .
i,k=1
∂xi ∂xk j=1 ∂xj

The following theorem summarises the result we have obtained above.

Theorem 1.5.2. Let f ∈ H −1 (Ω) and uε be the solution of (1.4.1). Then

there exists u ∈ H01 (Ω) such that

uε (x) → u(x) as ε → 0

where u ∈ H01 (Ω) is the unique solution of (1.5.8) and A0 = (a0ij ) is a

matrix with constant entries and is elliptic. Further, the effective coefficients
a0ij depend only on the matrix A (we started with), and not on any other
data, viz., f and Ω etc.

Mathematically the result obtained is not precise, for we ignored various

crucial issues during our computation. For instance, we have conveniently
assumed the differentiability of the L∞ (Y ) functions aij on Y . Also, we
have at few places conveniently swapped derivative and integral. In spite of
these pitfalls the asymptotic expansion approach gives a fair idea on what is
expected to be the homogenized form of (1.4.1).
CHAPTER 1. ASYMPTOTIC EXPANSION 16

Example 1.3. Let us understand the above method in the one dimensional
case. Let a : [0, 1] → R be a function such that 0 < α ≤ a(x) < β a.e.
in [0, 1]. Thus, a satisfies the ellipticity condition and is in L∞ (Y ), and is
extended periodically to all of R and aε (x) = a(x/ε), for x ∈ (a, b) ⊂ R. Let
Y = (0, 1) and the equation (1.4.1) takes the form

(aε (x) dudx

ε (x)
 d
− dx ) = f (x) in (a, b)
uε (a) = uε (b) = 0.

We know that the asymptotic expansion of uε involves the function u which

is a solution to the homogenized equation (1.5.8). Thus, to compute u we
need to find the effective coefficient a0 such that
 
d du(x)
− a0 (x) = f (x) in (a, b)
dx dx

We already know that a0 can be computed by finding the function χ that

solves (1.5.6), in one dimension which takes the form
  
dχ(y)
 −
 dy d
a(y) dy
= − da(y)
dy
in (0, 1)
 R χ(y) is Y -periodic in y
χ(y) dy = 0.

Y

Simplifying the differential equation, we get

  
d dχ(y)
− a(y) −1 =0
dy dy

and, therefore, a(y)χ0 (y) = a(y) + c, for some constant c. This first order
differential equation will admit a periodic solution iff
Z   Z
c 1
1+ dy = 1 + c dy = 0.
Y a(y) Y a(y)

Note that due to the ellipticity condition, a(y) > 0 for all y, and hence there
is no division by zero. The effective coefficient is given by formula (1.5.7)
Z   Z
dχ(y)
a0 = a(y) − a(y) dy = (a(y) − a(y) − c) dy = −c.
Y dy Y
CHAPTER 1. ASYMPTOTIC EXPANSION 17

Hence, Z −1
1
a0 = dy .
Y a(y)
The interesting fact to be observed here is that the effective coefficient a0 is
the inverse of the weak limit in Lp (Ω) of 1/aε rather that the weak limit of
aε , as one would expect. So, the homogenized coefficient is not always same
as taking the averages.
We caution that for the homogenized equation (1.5.8) to be solvable we
need to check that the effective coefficients a0ik satisfy the ellipticity condition
and are bounded. We shall address these questions after we give a precise
mathematical formulation of deriving the homogenized equation.
CHAPTER 1. ASYMPTOTIC EXPANSION 18
Chapter 2

Two-Scale Convergence

The aim of this chapter is to give a rigorous treatment of the formal asymp-
totic expansion of homogenization problems with periodic coefficients, dis-
cussed in previous chapters. The two-scale method justifies the formal expan-
sion. The notion of two-scale convergence was introduced by G. Nguetseng
(cf. [Ngu89]) in 1989 and further developed by G. Allaire (cf. [All92, All94,
LNW02]).
In Chapter 1, we studied (1.4.1) by introducing a formal two-scale asymp-
totic expansion for the solution uε , i.e.,
∞
X  x
i
uε (x) = ε ui x, .
i=0
ε

The above power series expansion was proposed, expecting that the solution
uε will exhibit a two-scale oscillation, viz., in x and y = xε variables. This is
due to the presence of two scales in (1.4.1). The formal two-scale asymptotic
expansion method is not rigorous. Two-scale convergence method incorpo-
rates the lessons learnt from the two-scale asymptotic expansion and gives
rise to a rigorous justification of the homogenization process. In this chap-
ter, we shall consider more general coefficients A(x, y) = (aij (x, y)) defined
on Ω × Y instead of A(y).

2.1 Vector-valued Function Spaces

Let X be a Banach space. Let D0 (Ω; X) denote the class of all linear contin-
uous X-valued functions on D(Ω). For 1 ≤ p ≤ ∞, let Lp (Ω; X) denote the

19
CHAPTER 2. TWO-SCALE CONVERGENCE 20

kf (x)kpX < ∞.
R
class of all measurable functions f : Ω → X such that Ω

Theorem 2.1.1. For 1 ≤ p ≤ ∞, the space Lp (Ω; X) is a Banach space

w.r.t the norm Z 1/p
p
kf kp,Ω,X = kf (x)kX dx .
Ω

Further, for 1 < p < ∞, if X is reflexive then Lp (Ω; X) is reflexive. Also,

for 1 ≤ p < ∞, if X is separable then Lp (Ω; X) is separable.
Theorem 2.1.2 (Pettis’ theorem (cf. [Yos95])). Let X be a separable Banach
space X. A function f : Ω → X is measurable if and only if the real-valued
functions x 7→ hG, f (x)i is measurable for every G ∈ X ? , the dual of X.
If X is chosen to be the Banach space Cper (Y ), the set of all Y -periodic
functions in C(Rn ), then f ∈ Lp (Ω; Cper (Y )) implies that, for each x ∈ Ω,
f (x) : Rn → R is a Y -periodic function. Thus, f can be seen as a real-valued
two variable function on Ω × Rn .
Theorem 2.1.3 (cf. See [All92, LNW02])). A function f ∈ L1 (Ω; Cper (Y ))
if and only if there exists a zero measure subset E ⊂ Ω such that:
(a) for any x ∈ Ω\E, the function y 7→ f (x, y) is continuous and Y -periodic;

(b) for any y ∈ Y , the function x 7→ f (x, y) is measurable;

(c) the map x 7→ supy∈Y |f (x, y)| is in L1 (Ω), i.e.,

Z
sup |f (x, y)|dx < ∞.
Ω y∈Y

Proof. Suppose f ∈ L1 (Ω; Cper (Y )), then (a) and (c) are obvious from def-
initions and it only remains to prove (b). Since the map f : Ω → Cper (Y )
is measurable, by Pettis’ theorem, x 7→ hG, f (x)i is measurable for every
G ∈ [Cper (Y )]? . In particular, for
R any fixed y ∈ Y , choose G to be the Dirac
measure δy at y, i.e., hδy , gi = Y g(t) dδy = g(y), for all g ∈ Cper (Y ). Thus,

x 7→ hδy , f (x)i = f (x, y)

is measurable.
Conversely, if (a), (b) and (c) are satisfied then it only remains to prove
that the map f : Ω → Cper (Y ) is measurable. By Pettis’ theorem, it is
CHAPTER 2. TWO-SCALE CONVERGENCE 21

enough to prove that x 7→ hG, f (x)i is measurable, for every G ∈ [Cper (Y )]? .
Without loss of generality, assume G to be a positive functional because any
G can be split into positive and negative functionals. By RieszR representation,
there is a unique positive measure µG such that hG, gi = Y g dµG , for all
g ∈ Cper (Y ). We now approximate G by a sequence of functionals Gk which
are finite linear combination of Dirac measures. Let {Yi }k1 be a partition of
Y into k-disjoint cubes of side length k1 and λi := µG (Yi ). Let yi ∈ Yi , χYi
be the characteristic function of Yi in Y , extended periodically to Rn and
δi := δyi be the Dirac measure at yi . Define Gk as,

k Z
X k
X
hGk , gi := g(y) d(λi δi ) = λi g(yi ).
i=1 Y i=1

Note that x 7→ hGk , f (x)i = ki=1 λi f (x, yi ) is measurable because it is sum

P
of measurable functions. We now claim that

lim hGk , f (x)i = hG, f (x)i

k→∞

which will imply that x 7→ hG, f (x)i is measurable. Note that

k
X k
X
hGk , f (x)i = λi f (x, yi ) = f (x, yi )µG (Yi )
i=1 i=1
Z k
!
X
= f (x, yi )χYi (y) dµG (y).
Y i=1

For eachP x ∈ Ω, let Ak (x) be the class of all simple functions of the form
sx (y) = ki=1 αi (x)χYi (y) and sx (y) ≤ f (x, y), for all y ∈ Y . Also, let S(x)
be the class of all simple functions satisfying sx (y) ≤ f (x, y), for all y ∈ Y .
CHAPTER 2. TWO-SCALE CONVERGENCE 22

Then
Z k
!
X
lim hGk , f (x)i = lim f (x, yi )χYi (y) dµG (y)
k→∞ k→∞ Y i=1
" Z #
≤ lim sup sx (y) dµG (y)
k→∞ sx ∈Ak (x) Y
Z
≤ sup sx (y) dµG (y)
sx ∈S(x) Y
Z
≤ f (x, y) dµG (y)
Y
= hG, f (x)i.
By taking −f instead of f in the above argument and by linearity of duality,
we obtain the reverse inequality and, hence, the equality.
For 1 ≤ p < ∞, let Xp (Ω; Y ) generically denote one of the follow-
ing spaces: Lp (Ω; Cper (Y )), Lpper (Y ; C(Ω)), C(Ω̄; Cper (Y )). The demand of
smoothness in one of the variables is mandatory. The space Xp (Ω; Y ) is a
separable Banach space and Xp (Ω; Y ) is dense in Lp (Ω × Y ).
Theorem 2.1.4. Let 1 ≤ p < ∞. For any φ ∈ Lp (Ω; Cper (Y )), the functions
φε (x) := φ x, xε are measurable in x and satisfies:

(a)
kφε kp,Ω ≤ kφkLp (Ω;Cper (Y )) (2.1.1)
and, hence, φε ∈ Lp (Ω);
(b) Z
ε 1
φ * φ(·, y) dy weakly in Lp (Ω). (2.1.2)
|Y | Y
1
In particular, for p = 2, kφε k22,Ω → |Y |
kφk22,Ω×Y .
Proof. We shall prove the result only for Xp (Ω; Y ) = L1 (Ω; Cper (Y )) because
the proof is similar in all other cases. By Theorem 2.1.3, the functions φε
are Caratheodory (cf. [All92, ET74]) and hence they are measurable. The
inequality (2.1.1) is easy to conclude because
Z  x Z
ε
kφ k1,Ω = |φ x, | dx ≤ sup |φ(x, y)| dx = kφkL1 (Ω;Cper (Y )) .
Ω ε Ω y∈Y
CHAPTER 2. TWO-SCALE CONVERGENCE 23

We shall now prove (2.1.2). Consider the partition {Yi }k1 of Y and yi ∈ Yi as
in Theorem 2.1.3. Let χYi be the characteristic function of Yi in Y , extended
periodically to Rn . Define the step functions
k
X
φk (x, y) = φ(x, yi )χYi (y).
i=1

Note that the map x 7→ φk (x, yi ) is in L1 (Ω). Define

k
 x X x
φεk (x) := φk x, = φ(x, yi )χYi .
ε i=1
ε

But x Z
1
χYi * χYi (y) dy weak-* in L∞ (Ω).
ε |Y | Y

Therefore, for each fixed k ∈ N and ψ ∈ L∞ (Ω),

Z k
X Z x 
lim φεk (x)ψ(x) dx = lim φ(x, yi )χYi ψ(x) dx
ε→0 Ω i=1
ε→0 Ω ε
k Z Z
1 X
= φ(x, yi )ψ(x) χYi (y) dy dx
|Y | i=1 Ω Y
Z Z
1
= φk (x, y)ψ(x) dy dx.
|Y | Ω Y

Thus, (2.1.2) is true for step functions. But, for each fixed k ∈ N and
ψ ∈ L∞ (Ω),
Z  Z  Z
ε 1
φ (x) − φ(x, y) dy ψ(x) dx ≤ |φε (x) − φεk (x)| |ψ(x)| dx
Ω |Y | Y
ZΩ Z
ε 1
+ φk − φk dy |ψ| dx
Ω |Y | Y
Z
1
+ |φk − φ| |ψ(x)| dy dx.
|Y | Ω×Y

As ε → 0, the second term, consisting of step functions, converges to zero.

In the first and third term are smaller than its supremum w.r.t y-variable.
CHAPTER 2. TWO-SCALE CONVERGENCE 24

Thus,
Z  Z  Z
ε 1
lim φ − φ dy ψ dx ≤ 2 sup |φ(x, y) − φk (x, y)| |ψ(x)| dx
ε→0 Ω |Y | Y Ω y∈Y
= 2 k(φ − φk )ψkL1 (Ω;Cper (Y )) .

Define
gk (x) := sup |φ(x, y) − φk (x, y)| |ψ(x)|.
y∈Y

By the continuity of φ in y-variable, gk (x) → 0, as k → ∞, pointwise

for a.e x ∈ Ω. Further, gk (x) ≤ 2 supy∈Y |φ(x, y)| |ψ(x)| in L1 (Ω). Hence,
by Lebesgue’s dominated convergence result, k(φ − φk )ψkL1 (Ω;Cper (Y )) → 0.
Thus, (2.1.2) is proved.

Theorem 2.1.5 (cf. [BM]). Let 1 ≤ p < ∞. Suppose φ(x, y) = φ1 (x)φ2 (y)
such that φ1 ∈ Ls (Ω) and φ2 ∈ Ltper (Y ) with 1 ≤ s, t < ∞ and 1s + 1t = p1 .
Then φε (x) = φ x, xε ∈ Lp (Ω) and

Z
ε φ1 (·)
φ * φ2 (y) dy weakly in Lp (Ω).
|Y | Y

2.2 Two-scale Convergence

The notion of two-scale convergence was introduced by G. Nguetseng (cf.
[Ngu89]) for L2 -spaces and, then, generalized to Lp spaces in [All92, LNW02].
x

2
Recall that a sequence uε (x) := u x, ε in LR (Ω), with u being Y -periodic in
second variable, will weakly converge to |Y1 | Y u(x, y) dy (cf. Theorem 1.1.3).
Thus, weak limit of an oscillating sequence does not capture the oscillations.
Two-scale convergence is a generalization of weak convergence such that the
limit captures the oscillations.

Definition 2.2.1. Let 1 < p, q < ∞ such that p1 + 1q = 1. A sequence

{uε } ⊂ Lp (Ω) is said to two-scale converge to u ∈ Lp (Ω × Y ) if
Z  x Z Z
1
uε (x)φ x, dx → u(x, y)φ(x, y) dy dx, (2.2.1)
Ω ε |Y | Ω Y
2s
for all φ ∈ Lq (Ω; Cper (Y )). The convergence is denoted as uε * u.
CHAPTER 2. TWO-SCALE CONVERGENCE 25

Theorem 2.2.2 (Uniqueness). The two-scale limit is unique.

Proof. If u, v ∈ Lp (Ω × Y ) are two distinct two-scale limits of a sequence
{uε } ⊂ Lp (Ω), then
Z Z
[u(x, y) − v(x, y)]φ(x, y) dy dx = 0
Ω Y
q
for all φ ∈ L (Ω; Cper (Y )). Thus, u = v a.e. in Ω × Y .
Example 2.1. Let uε (x) := u x, xε , where u ∈ Lp (Ω × Y ) is smooth and

Y -periodic in y-variable. For any φ ∈ Lq (Ω; Cper (Y )), the product uφ ∈

L1 (Ω; Cper (Y )). By the Y -periodicity of uφ and Theorem 2.1.4, it follows
2s
that uε * u.
Example 2.2. Let uε := u x, εx2 , where u ∈ Lp (Ω × Y ) is a smooth and Y -

periodic in the second variable. For any φ ∈ Lq (Ω; Cper (Y )), set ψ(x, y, z) :=
u(x, z)φ(x, y). Recall that if ψ : Ω × Rn × Rn → R is a function which is
Y -periodic in both second variable and third variable, then
 x x Z Z
1
ψ x, , 2 * 2
ψ(x, y, z) dz dy weakly in Lp (Ω).
ε ε |Y | Y Y
Therefore,
Z  Z Z Z
x   x 1
u x, 2 φ x, dx → u(x, z)φ(x, y) dz dy dx.
Ω ε ε |Y |2 Ω Y Y
Equivalently, Z
 x  2s 1
u x, 2 * u(x, z) dz.
ε |Y | Y
Thus, the two-scale limit is same as the Lp weak limit. More generally, if
Yi = Y , for all i = 1, 2, . . . , k,
 x Z Z
x  2s 1
u x, 2 , . . . , k * ... u(x, y2 , . . . , yk ) dy2 . . . dyk
ε ε |Y |k−1 Y2 Yk

which is same as the Lp weak limit. If ψ : Ω × Rn · · · × Rn → R is a

k + 1 variable function which is Y -periodic in each of the i + 1 variable, for
i = 1, 2, . . . , k, then
 x Z Z
x
ψ x, , . . . , k * ... ψ(x, y1 , . . . , yk ) dy1 . . . dyk weakly in Lp (Ω).
ε ε Y1 Yk
CHAPTER 2. TWO-SCALE CONVERGENCE 26

Remark 2.2.3. Recall that uε * u weakly in Lp (Ω) if, for all φ ∈ Lq (Ω),
Z Z
uε (x)φ(x) dx → u(x)φ(x) dx.
Ω Ω

Thus, the usual weak convergence in Lp (Ω) hides (averages out) the effect
of oscillations in uε . In order to capture the oscillations
of the form xε ,
one has to treat uε with test functions of the form φ x, xε . This was the
motivation behind the definition of two-scale convergence. Also, note that,
as seen in the above example, the test function φ x, xε is not good enough

to capture higher order oscillations of the form εxk for k ≥ 2. To capture

these oscillations, one may need to use test functions of the form φ x, εxk .

The basic idea is that one has to treat with test functions with same order
of oscillations. This is called the multi-scale or reiterated homogenization.
Remark 2.2.4. Suppose uε admits an asymptotic expansion
∞
X  x
uε (x) = εi ui x,
i=0
ε

where ui ’s are Y -periodic and smooth in the second variable. Then, by

2s
Example 2.1, uε * u0 , the first term in the expansion.
2s
Theorem 2.2.5. If uε converges to u strongly in Lp (Ω) then uε * u. In
particular, the two-scale limit is independent of y-variable.
Proof. Let uε → u strongly in RLp (Ω). For any φ ∈ Lq (Ω; Cper (Y )), let
φε (x) := φ x, xε and φ̄(x) := |Y1 | Y φ(x, y) dy. Then

Z
uε (x)φε (x) − u(x)φ̄(x) dx
 
≤ kuε − ukp,Ω kφε kq,Ω
Ω
Z
u(x) φε (x) − φ̄(x) dx .
 
+
Ω

By (2.1.1), kφε kq,Ω is uniformly bounded. Moreover, the strong convergence

of uε implies that the first term goes to zero. By (2.1.2) and u ∈ Lp (Ω), the
second term goes to 0.
Example 2.3. The converse of above result is not true, i.e., two-scale con-
vergence need not imply strong convergence. Consider the function u ∈
CHAPTER 2. TWO-SCALE CONVERGENCE 27

L2 ([0, 1] × [0, 1]),
defined as u(x, y) = sin(2πy), and define the sequence
uε (x) := sin 2πxε
in L2 [0, 1]. Because {uε } converge weakly to 0 in L2 [0, 1]
(periodic oscillating function weakly converges to average), if it strong con-
verges then the limit must be 0. But k sin(2πx/ε)k2,[0,1] = 1/2 and, hence, do

not strongly converge. The two-scale limit of the sequence uε (x) := sin 2πx ε
is u(x, y) = sin(2πy) on [0, 1] × [0, 1].
p 2s p
Theorem 2.2.6.
1
R For any sequence uε ⊂pL (Ω), if uε * u with u ∈ L (Ω×Y )
then uε * |Y | Y u(x, y)dy weakly in L (Ω). In particular, if the two-scale
limit u is independent of y then the two-scale limit and weak limit coincide.
2s
Proof. Let uε * u. Then, in particular, for any φ ∈ Lq (Ω) ⊂ Lq (Ω; Cper (Y ))
(φ independent of y),
Z Z Z
1
uε (x)φ(x) dx → u(x, y)φ(x) dy dx.
Ω |Y | Ω Y
Thus, Z
1
uε * u(·, y)dy weakly in Lp (Ω).
|Y | Y
If u(x, y) = u(x) then |Y1 | Y u(x) dy = u(x). Thus, weak limit and two-scale
R

limit coincide for y independent functions.

We have noted that the two-scale convergence is intermediary between
strong and weak convergences in Lp (Ω). If the weak and two-scale limits are
different then it means that there is more information in the two-scale limit,
than the weak limit, about the oscillations in the sequence.
Example 2.4. The converse of the above result is not true, i.e., weak conver-
gence need not imply two-scale convergence. Consider the sequence {un } ⊂
L2 [0, 1] defined as
(
sin(2πnx) if n is odd
un (x) =
cos(2πnx) if n is even.

This sequence converges weakly to zero in L2 ([0, 1] but does not two-scale
converge.

Corollary 2.2.7. If a sequence {uε } ⊂ Lp (Ω) two-scale converges then it is

bounded in Lp (Ω).
CHAPTER 2. TWO-SCALE CONVERGENCE 28

Proof. If uε two-scale converges then it also weakly converges in Lp (Ω). Any

weakly convergent sequence in Lp (Ω) is norm bounded.
Theorem 2.2.8 (Compactness Theorem). Let {uε } be a bounded sequence
in Lp (Ω). Then there exists a subsequence of {uε } (still denoted by ε) and a
2s
u ∈ Lp (Ω × Y ) such that uε * u.
Proof. Step 1 For each uε , define Lε : Lq (Ω; Cper (Y )) → R, as
Z  x
Lε (ψ) = uε (x)ψ x, dx.
Ω 
Note that Lε is a continuous linear

functional on Lq (Ω; Cper (Y )). For
q x q
any ψ ∈ L (Ω; Cper (Y )), ψ x, ε ∈ L (Ω). Now, by Hölder’s inequal-
ity,
Z  x
|Lε (ψ)| ≤ uε (x)ψ x, dx
Ω 
x
≤ kuε kp,Ω kψ(x, )kq,Ω
ε
≤ kuε kp,Ω kψ(x, y)kLq (Ω;Cper (Y )) .

The last inequality follows from Theorem 2.1.4. Since {uε } is bounded
in Lp (Ω), there is a constant C0 > 0 (independent of ε) such that
kuε kp,Ω ≤ C0 . Thus, the sequence {Lε } is bounded in the dual of
Lq (Ω; Cper (Y )).

Step 2 Recall that Lq (Ω; Cper (Y )) is separable. Thus, by Banach-Alaoglu

theorem, there is a L ∈ [Lq (Ω; Cper (Y ))]? and a subsequence of {Lε }
such that
Lε (ψ) → L(ψ) ∀ψ ∈ Lq (Ω; Cper (Y )).

Step 3 Passing to the limit, as ε → 0, in the inequality of Step 1, we obtain

|L(ψ)| ≤ C0 kψ(x, y)||Lq (Ω;Cper (Y )) ∀ψ ∈ Lq (Ω; Cper (Y )).

Step 4 By the density of Lq (Ω; Cper (Y )) in Lq (Ω × Y ), we extend L as a

bounded linear functional to all of Lq (Ω×Y ) and denote the extension
by L̃. Then

|L̃(ψ)| ≤ C0 kψ(x, y)kq,Ω×Y ∀ψ ∈ Lq (Ω × Y ).

CHAPTER 2. TWO-SCALE CONVERGENCE 29

By Riesz representation theorem, L̃ ∈ [Lq (Ω×Y )]? , may be identified

with an element v ∈ Lp (Ω × Y ). Then, for every ψ ∈ Lq (Ω; Cper (Y )),
Z  x
lim uε (x)ψ x, dx = lim Lε (ψ) = L(ψ)
ε→0 Ω  ε→0
Z
= v(x, y)ψ(x, y) dy dx.
Ω×Y

Thus, uε two-scale converges to u(x, y) := |Y |v(x, y).

Corollary 2.2.9. Every weakly convergent sequence in Lp (Ω) has a two-scale

converging subsequence.

Theorem 2.2.10. Let {uε } be a bounded sequence in Lp (Ω). Then, along a

subsequence,
Z  x Z Z
uε (x)φ x, dx → u(x, y)φ(x, y) dy dx
Ω ε Ω Y

for all φ such that φ(x, y) = φ1 (x)φ2 (y) where φ1 ∈ Ls (Ω) and φ2 ∈ Lt (Ω)
with 1 ≤ s, t < ∞ and 1s + 1t = 1q .

Proof. If φ is in the variable separable form the smoothness hypothesis on

one of the variable of φ may be relaxed and the measurability of φ x, xε
may be derived. The proof of the result is again by approximation (cf.
[LNW02]).

Recall that the Lp -norm is lower-semicontinuous w.r.t the weak topology,

i.e., if uε converges weakly to u in Lp (Ω) then

kukp,Ω ≤ lim inf kuε kp,Ω .

ε→0

A similar result holds for two-scale convergence.

Theorem 2.2.11. Let {uε } ⊂ Lp (Ω) two-scale converge to u ∈ Lp (Ω × Y ).

Then
1
lim inf kuε kp,Ω ≥ kukp,Ω×Y ≥ kūkp,Ω , (2.2.2)
ε→0 |Y |
where ū = |Y1 | Y u(x, y) dy.
R
CHAPTER 2. TWO-SCALE CONVERGENCE 30

Proof. If u ∈ Lp (Ω) then |u|p−2 u ∈ Lq (Ω × Y ). Now, choose a sequence

ψk ∈ Lq (Ω; Cper (Y )) that converges to |u|p−2 u strongly in Lq (Ω × Y ). By
Young’s inequality, we obtain
Z  x Z Z  x q
p
p uε (x)ψk x, dx ≤ |uε (x)| dx + (p − 1) ψk x, dx.
Ω ε Ω Ω ε
Fix k and pass to limit, as ε → 0, to obtain
Z Z Z
p
u(x, y)ψk (x, y) dx dy ≤ lim inf |uε (x)|p dx
|Y | Ω Y ε→0
ZΩ Z
p−1
+ |ψk (x, y)|q dy dx.
|Y | Ω Y
Now, passing to limit, as k → ∞, we obtain
Z Z Z
p p
|u(x, y)| dy dx ≤ lim inf |uε (x)|p dx
|Y | Ω Y ε→0
ZΩ Z
p−1
+ |u(x, y)|p dy dx
|Y | Ω Y
Z Z Z
1 p
|u(x, y)| dy dx ≤ lim inf |uε (x)|p dx.
|Y | Ω Y ε→0 Ω

This implies the first inequality of (2.2.2). The second inequality in (2.2.2)
follows from Jensen’s inequality,
Z Z p Z Z
p
p
|Y | kūkp,Ω = u(x, y) dy dx ≤ |u(x, y)|p dy dx = kukpp,Ω×Y .
Ω Y Ω Y

m,p
Let Wper (Y ) denote the class of functions in W m,p (Rn ) which are Y -
periodic.
Theorem 2.2.12 (Compactness in W 1,p ). For any given 1 ≤ p < ∞, let uε
be a bounded sequence in W 1,p (Ω). Then there is exists a u ∈ W 1,p (Ω) and
u1 ∈ Lp (Ω; Wper
1,p
(Y )) such that, for a subsequence (still denoted by ε),

uε * u weakly in W 1,p (Ω)

2s
uε * u in Lp (Ω)
2s
∇uε * ∇u + ∇y u1 in [Lp (Ω)]n .
CHAPTER 2. TWO-SCALE CONVERGENCE 31

Proof. Let {uε } be bounded in W 1,p (Ω). By weak compactness, uε * u

weakly in W 1,p (Ω) and, thus, strongly in Lp (Ω). Hence, u is also the two-
scale limit of {uε } in Lp (Ω). Also, {∇uε } is bounded in [Lp (Ω)]n . Hence, by
two-scale compactness theorem, there exists v ∈ [Lp (Ω × Y )]n such that
Z Z Z
∂uε  x  1
φ x, dx → vi (x, y)φ(x, y) dy dx, (2.2.3)
Ω ∂xi ε |Y | Ω Y
for all φ ∈ Lq (Ω; Cper (Y )). In particular, (2.2.3) is valid for all φ in the dense
∞ ∞
subset D(Ω; Cper (Y )). Consider Φ ∈ [D(Ω; Cper (Y ))]n such that divy Φ = 0.
Then
Z  x Z h  x  x i
−1
∇uε · Φ x, dx = − uε divx Φ x, + ε divy Φ x, dx
Ω ε ε ε
ZΩ  x
= − uε divx Φ x, dx
ΩZ Z ε
as ε→0 −1
−→ u(x)divx Φ(x, y) dy dx
|Y | Ω Y
Z Z
1
= ∇u(x) · Φ(x, y) dy dx.
|Y | Ω Y
Thus, using (2.2.3), we observe that
Z Z
[v(x, y) − ∇u(x)] · Φ(x, y) dy dx = 0
Ω Y

∞
for all Φ ∈ [D(Ω; Cper (Y ))]n such that divy Φ = 0. By a classical result (cf.
[Tem79, GR81]), v(x, y) − ∇u(x) is a gradient with respect to y, i.e., there
is a u1 ∈ Lp (Ω; Wper
1,p
(Y )) such that

v(x, y) − ∇u(x) = ∇y u1 (x, y).

Recall that the strong convergence of uε to u in Lp (Ω) is characterized

by both the weak convergence, i.e., uε * u in Lp (Ω) and norm convergence,
i.e., kuε kp,Ω → kukp,Ω . This motivates the definition of strong two-scale
convergence.
Definition 2.2.13. We say {uε } ⊂ Lp (Ω) strongly two-scale converges to
2s 2s
u ∈ Lp (Ω × Y ), denoted as uε → u if uε * u and kuε kp,Ω → |Y1 | kukp,Ω×Y .
CHAPTER 2. TWO-SCALE CONVERGENCE 32

The strong two-scale convergence is weaker than strong convergence in

p
L (Ω) and stronger than two-scale convergence.
Recall that the product of two sequences, one converging strongly and the
other weakly, is the product of its strong and weak limits. A similar result
is valid in the context of two-scale convergence too.
Theorem 2.2.14. Let uε and vε be two sequences in Lp (Ω) and Lq (Ω),
2s 2s
respectively, such that uε → u, for u ∈ Lp (Ω × Y ), and vε * v, for
v ∈ Lq (Ω × Y ). Then the product uε vε converges in the distribution sense,
i.e., Z
1
uε vε * u(·, y)v(·, y) dy weak-* in D0 (Ω).
|Y | Y
Further, if u ∈ Lp (Ω; Cper (Y )) then
 ·
lim uε − u ·, = 0.
ε→0 ε p,Ω

2.3 Classical Definition of Two-Scale Conver-

gence
In the early stages of the development of two-scale convergence, the test func-
tion space for φ, in the definition of two-scale convergence, was restricted to
∞
the sub-class D(Ω; Cper (Y )). This was later replaced (cf. [LNW02]) with
q
L (Ω; Cper (Y )) to have the implication of weak convergence. Because if
∞
D(Ω; Cper (Y )) is used as the test function class, then the two-scale conver-
gence of {uε } will not imply its weak convergence (and, hence, boundedness
in Lp (Ω)).
Example 2.5. Consider un : (0, 1) → R defined as
(
n if 0 < x < n1
un (x) :=
0 if n1 < x < 1.
∞
Then, for all φ ∈ D((0, 1); Cper (0, 1)),
Z 1 Z 1/n Z 1 z 
un (x)φ(x, nx) dx = n φ(x, nx) dx = φ , z dz → 0
0 0 0 n
due to the compact support of φ in (0, 1). But {un } is not bounded
R1 in
2 2
L (0, 1). Also, un do not weakly converge to 0 in L (0, 1) because 0 un g = 1
where g ≡ 1 is in L2 (0, 1).
CHAPTER 2. TWO-SCALE CONVERGENCE 33

∞
Example 2.6. The choice of C(Ω; Cper (Y )) as test function space will, also,
not yield any better result. Let ũε be the periodic extension of uε , defined
above, to all of R. Define vε : (0, 1) → R as
(
ũε ( x ) if 41 < x < 34
vε (x) =
0 otherwise.
∞
Then, for all φ ∈ C((0, 1); Cper (0, 1)),
Z 1  x Z 1Z 1
vε (x)φ x, dx → u(x, y)φ(x, y) dy dx
0 ε 0 0
where (
1 if 41 < x < 3
4
u(x, y) =
0 otherwise.
Note that {vε } is not bounded in L2 (0, 1).
Theorem 2.3.1. Let {uε } be a bounded sequence in Lp (Ω) and
Z  x Z Z
1 ∞
uε (x)φ x, dx → u(x, y)φ(x, y) dy dx ∀φ ∈ D(Ω; Cper (Y )).
Ω ε |Y | Ω Y
2s
Then uε * u.
Proof. The idea is to approximate φ ∈ Lq (Ω; Cper (Y )) by a sequence φk ∈
∞
D(Ω; Cper (Y )) and use the uniform bound of {uε }.
Theorem 2.3.2 (Compactness). Let {uε } be a bounded sequence in Lp (Ω).
Then, along a subsequence,
Z  x Z Z
uε (x)φ x, dx → u(x, y)φ(x, y) dy dx ∀φ ∈ Lqper (Y ; C(Ω̄)).
Ω ε Ω Y
∞
Proof. Use the density of D(Ω; Cper (Y )) in Lqper (Y ; C(Ω̄)).

2.4 Homogenization of Second Order Linear

Elliptic Problems
Recall from § 1.4 that a Dirichlet problem for a periodic composite material
Ω is given as, for a given f ∈ H −1 (Ω),

−div(Aε (x)∇uε (x)) = f (x) in Ω
(2.4.1)
uε = 0 on ∂Ω
CHAPTER 2. TWO-SCALE CONVERGENCE 34

where Aε (x) = aij x, xε is in M (α, β, Ω × Y ) and aij : Ω × Y → R. We

know there exists a unique solution uε ∈ H01 (Ω), by Lax-Milgram result, such
that
Z
Aε (x)∇uε (x) · ∇v(x) dx = hf, viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω)
0
Ω

and kuε kH01 (Ω) ≤ 1/αkf kH −1 (Ω) . Since uε is uniformly bounded in H01 (Ω),
by Theorem 2.2.12, there exists a u ∈ H01 (Ω) and u1 ∈ L2 (Ω; Hper1
(Y )) such
that, for a subsequence

 uε * u
 weakly in H01 (Ω) and strongly in L2 (Ω);
2s
uε * u in L2 (Ω); (2.4.2)

 ∇u * 2s 2 n
ε ∇u + ∇y u1 in [L (Ω)] .
∞
Consider the test functions φ ∈ D(Ω) and φ1 ∈ D(Ω; Cper (Y )) and choose
x


v(x) = φ(x) + εφ1 x, ε in the weak formulation above to obtain
Z  x D  x E
A x, ∇uε · (∇φ + ε∇φ1 + ∇y φ1 ) dx = f, φ + εφ1 x, .
Ω ε ε H −1 (Ω),H01 (Ω)
Note that φ + εφε1 * φ weakly in H01 (Ω). Thus, the term in RHS converges,
i.e.,
hf, φ + εφε1 iH −1 (Ω),H 1 (Ω) → hf, φiH −1 (Ω),H01 (Ω) .
0

Note that {Aε (x)∇uε (x)} is bounded in [L2 (Ω)]n and, by (2.1.1), {∇φε1 } is
bounded in [L2 (Ω)]n . Using Hölder’s inequality, we obtain
Z  x
A x, ∇uε · ε∇φ1 dx ≤ Cε.
Ω ε
Therefore, Z  x
lim A x, ∇uε · ε∇φ1 dx = 0.
ε→0 Ω ε
Let us denote ψ(x, y) := t A(x, y)(∇φ(x) + ∇y φ1 (x, y)). Then
Z  x   x  Z  x
A x, ∇uε · ∇φ(x) + ∇y φ1 x, dx = ∇uε · ψ x, dx.
Ω ε ε Ω ε
Since ψ is a two-scale test function, one may pass to limit in the RHS above.
Thus,
Z Z
ε ε→0 1
Aε ∇uε · (∇φ + ∇y φ1 ) dx → [∇u + ∇y u1 (x, y)] · ψ(x, y) dy dx
Ω |Y | Ω×Y
CHAPTER 2. TWO-SCALE CONVERGENCE 35

and, therefore,
Z
1
A(x, y)(∇u + ∇y u1 ) · (∇φ + ∇y φ1 (x, y)) dy dx = hf, φiH −1 (Ω),H01 (Ω)
|Y | Ω×Y
(2.4.3)
∞
for all φ ∈ D(Ω) and φ1 ∈ D(Ω; Cper (Y )). Thus, by density, for all φ ∈ H01 (Ω)
and φ1 ∈ L2 (Ω; Hper
1
(Y )). In particular, by choosing φ1 ≡ 0, we get
( h i
1
R
−divx |Y | Y
A(x, y)(∇u(x) + ∇y u1 (x, y)) dy = f in Ω
(2.4.4)
u =0 on Ω

and by choosing φ ≡ 0 we get


−divy [A(x, y)(∇u(x) + ∇y u1 (x, y))] = 0 in Ω × Y
(2.4.5)
u1 (x, y) is Y -periodic in y.

Both (2.4.4) and (2.4.5) are called the coupled two-scale homogenized system
of equations. The system (2.4.4) and (2.4.5) have a unique pair of solution
(u, u1 ) in H := H01 (Ω) × L2 (Ω; Hper
1
(Y )/R), due to Lax-Milgram result. The
space H is a Hilbert space with the norm

k(φ, φ1 )k2H = k∇φk2L2 (Ω) + k∇y φ1 k2L2 (Ω×Y ) .

Theorem 2.4.1. Let uε be the unique solution of (2.4.1). Then there exists
(u, u1 ) ∈ H satisfying (2.4.2) and (u, u1 ) is the unique solution of the two-
scale system (2.4.4) and (2.4.5).

Proof. Observe that (2.4.3) is the weak formulation of (2.4.4) and (2.4.5).
We introduce the bilinear form B : H × H → R defined as
Z
B[(φ, φ1 ), (ψ, ψ1 )] = A(x, y)(∇φ(x)+∇y φ1 (x, y))·(∇ψ(x)+∇y ψ1 (x, y))
Ω×Y

and the linear form L : H → R defined as

L(φ, φ1 ) = hf, φiH −1 (Ω),H01 (Ω) .

Note that B is H-elliptic and continuous. Also, L is continuous on H. Thus,

by Lax-Milgram result, (2.4.3) has a unique solution (u, u1 ) ∈ H. Further,
by uniqueness of solution, the convergence holds for entire sequence.
CHAPTER 2. TWO-SCALE CONVERGENCE 36

The coupled two-scale system (2.4.4) and (2.4.5) can be decoupled. Using
(2.4.5), one can represent u1 in terms of u. This is then used in (2.4.4) to
get the homogenized equation of u. Recall that {ej }n1 denotes the standard
basis vectors Rn . Freezing x as a parameter in (2.4.5) and substituting
Pn of ∂u
∇u(x) = j=1 ∂xj (x)ej , we get

−divy [A(x, y)∇y u1 (x, y)] = nj=1 ∂x∂u

 P
j
(x)divy [A(x, y)ej ] in Y
u1 (x, y) is Y -periodic in y.

The RHS motivates us to introduce, for each index j = 1, 2, . . . , n, the cell

problem. Let χj (x, y) be the solution of

−divy [A(x, y)∇y χj (x, y)] = −divy [A(x, y)ej ] in Y
(2.4.6)
χj (x, y) is Y -periodic in y.

Observe that above equation is same as (1.5.6) for each fixed x ∈ Ω. Thus,
there is a function ũ(x) such that
n
X ∂u
u1 (x, y) + (x)χj (x, y) = ũ(x).
j=1
∂x j

Substituting u1 in (2.4.4), we get a equation for u


−div(A0 (x)∇u(x)) = f in Ω
(2.4.7)
u = 0 on ∂Ω,

where A0 (x) = (a0ij (x)) is given as

Z n
!
1 X ∂χk (x, y)
a0ik (x) = aik (x, y) − aij (x, y) dy.
|Y | Y j=1
∂yj

Note that this is the precise form we obtained in (1.5.7) except that the
coefficient depends on x, as well. Further,
Z n
!
1 X ∂χ k (x, y)
a0ik (x) = aik (x, y) − aij (x, y) dy
|Y | Y j=1
∂y j
Z
1
= A(x, y)(ek − ∇y χk ) · ei dy.
|Y | Y
CHAPTER 2. TWO-SCALE CONVERGENCE 37

But, by taking χi as a test function in (2.4.6), we get

Z
A(x, y)[ek − ∇y χk ] · ∇y χi dy = 0.
Y

Thus, we can write

Z
1
a0ik = A(x, y)[ek − ∇y χk ] · [ei − ∇y χi ] dy.
|Y | Y

Thus, we have shown the following result.

Theorem 2.4.2. Let uε be the unique solution of (2.4.1). Then uε converges
to u weakly in H01 (Ω), where u is the unique solution of the homogenized
equation (2.4.7).
Remark 2.4.3. One may replace the smoothness condition on A = (aij ) by
that of admissibility (or strong two scale limit) of A, i.e. the coefficients
satisfy Z Z Z
 x 2 1
lim aij x, dx = aij (x, y)2 dy dx
ε→0 Ω ε |Y | Ω Y
for all 1 ≤ i, j ≤ n. In such situations one
may appeal
to Theorem 2.2.14 for
x x
passing to limits. In particular, if aij x, ε = aij ε , then no assumption is
required as the admissibility condition is trivially satisfied.
Remark 2.4.4. The two scale homogenized system (2.4.4) and (2.4.5) is a
coupled system of equations with two unknowns, u and u1 , in x and y (macro-
scopic and microscopic, respectively) variables. Seemingly complicated, it is
a well-posed system and has a unique solution. Further, it was possible to
decouple the system to recover the homogenized equation. This was due to
the simple nature of the equation considered. For other types of equations
the decoupling may not be possible or may produce very complicated equa-
tions, viz., integro-differential equations. The homogenized equation may
pose existential issues while the two-scale form, though with twice the num-
ber of unknowns and variables, may give a solution. Thus, the presence of the
microscopic variables in the two-scale homogenized problem may double the
size of the equation but simplifies the structure. In some cases, decoupling
might introduce strange effects, viz., memory or non-local effects.
Theorem 2.4.5 (Corrector). If ∇y u1 (x, x/ε) ∈ [L2 (Ω)]n then kξε k2,Ω → 0,
where ξε := ∇uε (x) − ∇u(x) − ∇y u1 (x, x/ε). Thus, if both u1 and u are in
H 1 (Ω) then
kuε (x) − u(x) − εu1 (x, x/ε)kH 1 (Ω) → 0.
CHAPTER 2. TWO-SCALE CONVERGENCE 38

2
Proof. If A is smooth, say, A ∈ C(Ω; L∞ per (Y ))
n
then, by the regularity of
j 2
χ , the function u1 (x, x/ε) ∈ L (Ω). Thus, u1 may be used as a test
function
for the two-scale convergence. Let ξε := ∇uε − ∇u(x) − ∇y u1 x, xε . Then,
Z Z
Aε (x)ξε · ξε dx = f (x)uε (x) dx
Ω ZΩ h  x i h  x i
+ Aε ∇u + ∇y u1 x, · ∇u + ∇y u1 x, dx
Ω ε ε
Z h  x i
− (Aε + t Aε )(x)∇uε (x) · ∇u(x) + ∇y u1 x, dx.
Ω ε
Using the coercivity of A and passing to the two-scale limit, we obtain
Z
2
α lim sup kξε k2,Ω ≤ f (x)u(x) dx
ε→0 Ω Z Z
1
− A(x, y)[∇u + ∇y u1 ] · [∇u + ∇y u1 ] dy dx
|Y | Ω Y

The term on right-hand side is zero (why!), thus completing the proof.

2.5 Summary
Chapter 3

H-Convergence

In this chapter, we consider the non-periodic situation as opposed to those

considered in previous chapters. Let us set the environment for non-periodic
case.

3.1 Coercive Operators

Definition 3.1.1. We say a linear operator A : X → X ? is bounded or
continuous, if there is a constant 0 < β < +∞ such that

kAxkX ? ≤ βkxkX , ∀x ∈ X. (3.1.1)

Let B(X, X ? ) denote the set of all linear bounded homomorphisms from
X to X ? . The norm on B(X, X ? ) is given as,

kAxkX ?
kAkB(X,X ? ) = sup .
x∈X kxkX

Definition 3.1.2. An operator A : X → X ? is said to be coercive or X-

elliptic, if there is a constant 0 < α such that

hAx, xiX ? ,X ≥ αkxk2X , ∀x ∈ X. (3.1.2)

Theorem 3.1.3. Let X be a reflexive Banach space. Any coercive operator

A ∈ B(X, X ? ) is an isomorphism.

39
CHAPTER 3. H-CONVERGENCE 40

Proof. It is enough to show A is bijective. Observe that

αkxk2X ≤ hAx, xiX ? ,X ≤ kAxkX ? kxkX

and hence,
αkxkX ≤ kAxkX ? . (3.1.3)

Step 1 (Claim: A is injective). Let Ax1 = Ax2 . Then, from (3.1.3), αkx1 −
x2 kX ≤ 0. Therefore, x1 = x2 .

Step 2 (Claim: Im(A) is closed). Suppose {Axn } is a Cauchy sequence in

X ? . Then, by (3.1.3), {xn } is a Cauchy sequence in X. Since X is
complete, there is a x0 ∈ X such that xn → x0 (as n → ∞) in X.
Thus, from (3.1.1), we have

kA(xn − x0 )kX ? → 0.

Consequently, Ax0 is the limit of the Cauchy sequence {Axn }. Thus,

Im(A) is closed.

Step 3 (Claim: A is surjective). Let Im(A) 6= X ? . By Hahn-Banach theo-

rem, for the closed subspace Im(A) of X ? , there is a non-zero func-
tional vanishing on Im(A). Thus, there is a non-zero z ∈ X ?? ∼ =X
such that
hAx, ziX ? ,X = 0, ∀x ∈ X.
In particular, hAz, ziX ? ,X = 0 and, by (3.1.2), z = 0 which is a
contradiction. Thus A is surjective.

Corollary 3.1.4. Let A ∈ B(X, X ? ) be a coercive operator. Then, for any

f ∈ X ? , the equation Au = f has a unique solution.

Remark 3.1.5. Note that, by Theorem 3.1.3, A−1 is an isomorphism in

L(X ? , X) and
1
kA−1 f kX ≤ kf kX ? , ∀f ∈ X ? .
α
Thus, A−1 ∈ B(X ? , X).
CHAPTER 3. H-CONVERGENCE 41

Definition 3.1.6. Let X be a reflexive Banach space and X ? be its topo-

logical dual. A sequence of coercive operators {Aε } in B(X, X ? ) is said to
G-converge to A0 if
ε→0
hg, A−1 −1
ε f i −→ hg, A0 f i ∀f, g ∈ X ? .

The above definition defines a topology in B(X, X ? ). Note that G-

convergence of a sequence of operators is nothing but the weak operator
topology (WOT) convergence of the inverse operators.
Theorem 3.1.7 (G compactness). Let X be a separable reflexive Banach
space and X ? be its topological dual. Let {Aε } be a sequence of equi-coercive,
uniformly bounded operators in B(X, X ? ), then the sequence is G-compact,
i.e., there exists a subsequence {Aδ } of {Aε } and A0 such that G-converges
to a coercive A0 , as δ → 0.
Proof. By remark 3.1.5, we know that {A−1 ?
ε } ⊂ B(X , X) is uniformly
bounded. Let {fk }∞ ?
1 ⊂ X be the countable dense subset of X .
?
?
We shall now construct an operator L : X → X as follows: Note that,
for each fixed k, kA−1
ε fk kX is bounded uniformly w.r.t ε. Thus, by weak-
compactness of unit ball (Banach-Alaoglu result), there is a subsequence
{A−1 f } converging to, say, some u1 . Set Lf1 = u1 . Now, extract a weak
ε1 1
j

convergence subsequence {A−1 f } from {A−1

ε2j 2
f } that weakly converges to
ε1j 2
u2 . Set Lf2 = u2 . Proceeding this way, by extracting subsequence, at every
stage, we define Lfk = uk . For each k, by choosing the diagonal sequence
A−1
εj
, we have
j

lim hg, A−1 f i = hg, Lfk i ∀g ∈ X ? .

εj k
k→∞ j

We next show that L is bounded on {fk }. Due to the equi-coercivity of

Aε , we have for each k,

hfk , A−1 f i ≥ αkA−1

εj k
f k2 .
εj k X
j j

Thus, by weak lower semi-continuity of norm, we get

αkLfk k2X ≤ α lim inf kA−1 f k2

εj k X
j→∞ j

≤ lim hfk , A−1 f i

εjj k
j→∞
= hfk , Lfk i ≤ kfk kX ? kLfk kX .
CHAPTER 3. H-CONVERGENCE 42

Thus, kLfk kX ≤ 1/αkfk kX ? and L is bounded on the dense subset of X ? . Let

f ∈ X ? . Since {fk } is dense there is a sequence fm → f in X ? as m → ∞.
Since Lfm is bounded in X, for a subsequence, it converges weakly in X to,
say, u? . Note that

hg, A−1
εj
f i − hg, u? i = hg, A−1
εj
f − A−1 f + A−1
εj m
f − Lfm + Lfm − u? i.
εj m
j j j j

For large m, one can make the RHS as small as possible. Thus,

lim hg, A−1

εj
f i = hg, u? i.
j→∞ j

Since the choice of u? is independent of the choice of the subsequence {fm },

we set Lf = u? , for all f ∈ X ? . We leave it as an exercise to show that L is
linear.
We now claim that L is coercive. We know that L is bounded and kLk ≤
1/α. By the equi-coercivity of Aε , we have
1 α
hf, A−1
εj
f i ≥ αkA−1
εj
f k2X ≥ α 2
kAεj A−1 2
j f kX ≥ kf k2X ? .
j j β j εj β2
The constant β is the bound of Aε . Passing to limit as j → ∞, we get
hf, Lf i ≥ βα2 kf k2X ? . Thus, L is coercive and, by remark 3.1.5, we have
β2
kL−1 k ≤ α
. Moreover, L−1 is also coercive since

hf, A−1
εj
f i ≥ αkA−1
εj
f k2X
j j

implies that
αkLf k2X ≤ hf, Lf i
or equivalently, hL−1 u, ui ≥ αkuk2X . We set A0 = L−1 .

3.2 H-Convergence
Let Ω be an open bounded subset of Rn . Let 0 < α < β and M (α, β, Ω)
denote the set of all n × n matrices A(x) = (aij (x)) of functions such that

α|ξ|2 ≤ A(x)ξ · ξ and |A(x)ξ| ≤ β|ξ| for a.e. x ∈ Ω and for all ξ ∈ Rn .

Observe that the class M (α, β, Ω) is closed under transpose of matrices.

Given a sequence of matrices {Aε } ⊂ M (α, β, Ω), define the operator Aε :
CHAPTER 3. H-CONVERGENCE 43

H 1 (Ω) → H −1 (Ω) as Aε = −div(Aε ∇). For any f ∈ H −1 (Ω), the second

order elliptic problem

−div(Aε ∇uε ) = f in Ω
(3.2.1)
uε = 0 on ∂Ω

has a unique solution, by Lax-Milgram result, satisfying the estimate

1
kuε kH01 (Ω) ≤ kf kH −1 (Ω) . (3.2.2)
α
Hence there exists a subsequence such that

uε * u0 weakly in H01 (Ω).

The uniqueness of uε follows from (3.2.2). The bounded elliptic operator

Aε = −div(Aε ∇) from H01 (Ω) into H −1 (Ω) is an isomorphism and the norm
of (Aε )−1 is not larger than α−1 (cf. (3.2.2)). Moreover, we also know that
the solution uε of (3.2.3) can be characterized as the minimizer of
Z
1
Jε (v) = Aε ∇v.∇v dx − hf, viH −1 (Ω),H 1 (Ω)
2 Ω 0

in H01 (Ω). Since Aε ∈ M (α, β, Ω), Aε : H01 (Ω) → H −1 (Ω) is an uniformly

bounded, equi-coercive sequence of operators with constants β and α, re-
spectively. Thus, by Theorem 3.1.3 and Theorem 3.1.7, A−1ε exists and there
is a A0 to which Aε G-converges. Does there exist a matrix A0 such that
A0 = −div(A0 ∇)?

Definition 3.2.1. A sequence {Aε } ⊂ M (α, β, Ω) is said to H-converges to

H
a matrix A0 : Ω → [L∞ (Ω)]n×n , denoted as Aε * A0 , if for every sequence
fε → f strongly in H −1 (Ω), the solution uε of

−div(Aε ∇uε ) = fε in Ω
(3.2.3)
uε = 0 on ∂Ω

is such that
uε * u0 weakly in H01 (Ω) and (3.2.4a)

Aε ∇uε * A0 ∇u0 weakly in (L2 (Ω))n , (3.2.4b)

CHAPTER 3. H-CONVERGENCE 44

where u0 is the unique solution of


−div(A0 ∇u0 ) = f in Ω
(3.2.5)
u0 = 0 on ∂Ω.

The matrix A0 is called the H-limit of the sequence {Aε }.

Before we embark on the problem of finding H-limit of a given sequence,

we highlight some useful properties of H-convergence.

Theorem 3.2.2 (Uniqueness). The H-limit of a sequence {Aε } is unique.

Proof. Let A0 be a H-limit of {Aε }. Let ω ⊂⊂ ω0 ⊂ Ω and φ ∈ D(ω0 ) such

that φ ≡ 1 on ω. For each λ ∈ Rn , we define fλ = −div[A0 ∇{(λ · x)φ(x)}]
and let uλε be the unique solution

−div(Aε ∇uλε ) = fλ in ω0
uλε = 0 on ∂ω0 .

Then, by definition of H-convergence,

uλε * (λ · x)φ(x) weakly in H01 (ω0 )

Aε ∇uλε * A0 ∇{(λ · x)φ(x)} weakly in (L2 (ω0 ))n .

If A1 is another H-limit of {Aε } then, corresponding to the same fλ and by
H-convegence
uλε * v0λ weakly in H01 (ω0 )
Aε ∇uλε * A1 ∇v0λ weakly in (L2 (ω0 ))n .
By uniqueness of weak limits, v0λ (x) = (λ · x)φ(x) and A0 ∇{(λ · x)φ(x)} =
A1 ∇{(λ · x)φ(x)} in ω0 . Thus, in ω, ∇{(λ · x)φ(x)} = λ and, hence, A0 = A1
in ω.

Corollary 3.2.3 (Local Property). If {Aε } and {Bε } are two sequences in
M (α, β, Ω) which H-converges to A0 and B0 , respectively, such that Aε = Bε
in ω ⊂ Ω, for all ε, then A0 = B0 in ω.
H H
Theorem 3.2.4 (Transpose). If Aε * A0 then t Aε * t A0 .
CHAPTER 3. H-CONVERGENCE 45

Proof. Let ω ⊂⊂ Ω and g ∈ H −1 (ω). Let vε be the solution of


−div(t Aε ∇vε ) = g in ω
vε = 0 on ∂ω.

Then, upto a subsequence, there is a v ∈ H01 (ω) and η ∈ [L2 (ω)]n such that

vε * v weakly in H01 (ω) and

t
Aε ∇vε * η weakly in (L2 (ω))n .
Further, −div(η(x)) = g(x) a.e. in ω. For any f ∈ H −1 (ω), let uε be the
solution of (3.2.3) and, by H-convergence, there is a unique u0 ∈ H01 (ω)
satisfying (3.2.5). For any φ ∈ Cc∞ (ω), using vε φ as a test function in (3.2.3),
one obtains
Z
lim Aε ∇uε · ∇vε φ dx = − limhdiv(Aε ∇uε ), φvε iH −1 (ω),H01 (ω)
ε→0 ω ε→0
Z
− lim Aε ∇uε · ∇φvε dx
ε→0 ω
= −hdiv(A0 ∇u0 ), φviH −1 (ω),H01 (ω)
Z
− A0 ∇u0 · ∇φv dx
Z ω
= A0 ∇u0 · ∇vφ dx.
ω

Thus, Aε ∇uε · ∇vε * A0 ∇u0 · ∇v weak-* in D0 (Ω). Using uε φ as a test

function in the equation for vε , one obtains
Z
lim t Aε ∇vε · ∇uε φ dx = − limhdiv(t Aε ∇vε ), φuε iH −1 (ω),H01 (ω)
ε→0 ω ε→0
Z
− lim t Aε ∇vε · ∇φuε dx
ε→0 ω
Z
= −hdiv(η), φu0 iH −1 (ω),H01 (ω) − η · ∇φu0 dx
Z ω

= η · ∇u0 φ dx.
ω

Thus, t Aε ∇vε ·∇uε * η ·∇u0 weak-* in D0 (Ω). Since Aε (x)∇uε (x)·∇vε (x) =
t
Aε (x)∇vε (x) · ∇uε (x), we have A0 (x)∇u0 (x) · ∇v(x) = η(x) · ∇u0 (x) a.e. in
CHAPTER 3. H-CONVERGENCE 46

ω. Because the elliptic operator is an isomorphism, as f varies in H −1 (ω), u0

varies in H01 (ω). Therefore, for any λ ∈ Rn , we can choose a u0 ∈ H01 (ω) such
that ∇u0 (x) = λ on ω1 ⊂⊂ ω. Thus, for any λ ∈ Rn , t A0 (x)∇v(x)·λ = η(x)·λ
a.e. in ω1 and t A0 (x)∇v(x) = η(x) a.e. in ω. Hence, the limit v satisfies the
equation −div(t A0 (x)∇v(x)) = g and t A0 is the H-limit of t Aε . Moreover, by
the uniqueness of H-limit, t A0 is unique and, hence, v is the unique limit for
all subsequences. Thus, the convergences is true for the entire sequence.
Example 3.1 (One Dimension). Let us find the H-limit for a one dimension
problem. Let Ω = (a, b), f ∈ L2 (a, b) and let aε : (a, b) → R be a function
satisfying the ellipticity 0 < α ≤ aε (x) ≤ β a.e. and is in L∞ (a, b). Note
that the ellipticity condition implies that 1/aε is in L∞ (a, b). The equation
to be homogenized is
− dx (aε (x) dudx
ε (x)
 d
) = f (x) in (a, b)
uε (a) = uε (b) = 0.
There exists a unique solution uε ∈ H01 (a, b), by Lax-Milgram result, such
that
Z
duε (x) dv(x)
aε (x) · dx = hf, viH −1 (a,b),H 1 (a,b) , ∀v ∈ H01 (a, b)
Ω dx dx 0

and kuε kH01 (a,b) ≤ (C/α)kf kL2 (a,b) . By Eberlein-Šmuljan theorem, there exists
a subsequence of {uε }, also denoted by uε , such that
uε * u weakly in H01 (a, b),
for some u ∈ H01 (a, b). We need to find the homogenized equation which u
solves. Set ξε := aε u0ε . Note that ξε is bounded in L2 (a, b), because aε is
bounded in L∞ (a, b) and u0ε is bounded in L2 (a, b). From the equation, we
have that −ξε0 (x) = f (x) and hence ξε is bounded in H 1 (a, b). Therefore,
there exists a ξ ∈ H 1 (a, b) such that for a subsequence of ξε (denoted by
itself),
ξε * ξ weakly in H 1 (a, b).
Thus, by compact imbedding of H 1 (a, b) in L2 (a, b), ξε converges to ξ strongly
in L2 (a, b). Note that if either aε or u0ε converges strongly, then ξ is the
product of the limits of aε and u0ε . But, in general, this need not be the case.
Recall the u0ε converges to u0 weakly in L2 (a, b), therefore
1
ξε * u0 weakly in L2 (a, b).
aε
CHAPTER 3. H-CONVERGENCE 47

Since 1/aε is bounded in L∞ (a, b), for a subsequence,

1
* b(x) weak-* in L∞ (a, b).
aε (x)
Then, u0 = b(x)ξ. Also, the constant sequence ξε0 converges weakly to ξ 0 in
L2 (a, b) implies that −ξ 0 (x) = f (x). Hence,
 
d 1 du(x)
− = f (x).
dx b(x) dx

u is already in H01 (a, b) satisfying the boundary condition and the effective
coefficient is a0 = 1/b. The effective coefficient is bounded in L∞ (a, b) and
satisfies the ellipticity condition. Note that the choice of b depends on the
subsequence chosen and u is not unique. All the above arguments were for a
subsequences, extracted sufficiently.
However, if aε = a(x/ε) such that a is periodic. Then
Z 1
1
b= dy
0 a(y)

for every subsequence of aε , and a0 = b−1 . Thus, u is a unique solution

and all the convergences above are true for the entire sequence and not just
for subsequences (cf. Eberlein-Šmuljan result). Thus, in the one dimensional
case, the H-limit of a sequence {aε } ⊂ M (α, β, Ω) is the inverse of the weak-*
limit in L∞ of the inverses of aε .
Example 3.2 (Layering). Let Ω be a bounded open subset of Rn . Let uε be
the solution of (3.2.3) with fε = f in L2 (Ω), for all ε. The coefficient matrix
is such that Aε (x) = Aε (x1 ) and Aε = (aεij ). Therefore, upto a subsequence,
there is a u ∈ H01 (Ω) and ξ ∈ [L2 (Ω)]n such that

uε * u weakly in H01 (Ω) and

Aε ∇uε * ξ weakly in (L2 (Ω))n .

P 
ε ε
Also, −div(ξ) = f . Set ξε := Aε ∇uε = a
j=1 ij (x )u
1 xj (x) . Since
i
−div(ξε ) = f (x), it follows that
n
∂ξε1 X ∂ξ i
ε
− =f+ .
∂x1 i=2
∂x i
CHAPTER 3. H-CONVERGENCE 48

Let Ω1 be the projection of Ω in R and Ω2 be the projection of Ω in Rn−1

such that Ω = Ω1 × Ω2 . Define gε : Ω1 → L2 (Ω2 ) as gε (x1 )(x2 , . . . , xn ) :=
ξε1 (x1 , . . . , xn ). Thus, gε ∈ L2 (Ω1 ; L2 (Ω2 )). Note that

∂gε
∈ L2 (Ω1 ; H −1 (Ω2 ))
∂x1
because
∂gε ∂ξ 1
= ε.
∂x1 ∂x1
Thus, {gε } is bounded in H 1 (Ω1 ; H −1 (Ω2 )) and, by Aubin-Lions compactness
theorem, relatively compact in L2 (Ω1 ; H −1 (Ω2 )). Thus, gε → g strongly in
L2 (Ω1 ; H −1 (Ω2 )) and g = ξ 1 . Using the equation for ξε1 (x), we get

1 X ∂  aε1j (x1 ) 
ε 1 ε
ux1 (x) = ε ξ (x) − u (x) (3.2.6)
a11 (x1 ) ε j=2
∂xj aε11 (x1 )

and, for 2 ≤ i ≤ n,

aε (x1 1 aεi1 (x1 )aε1j (x1 )

X ∂   
ξεi (x) = εi1 ξε (x) + ε
aij − ε
u (x) . (3.2.7)
a11 (x1 ) j=2
∂xj aε11 (x1 )

Also, by the ellipticity of Aε , we have the following weak-* convergences in

L∞ (Ω)
1 1
ε
* b11 := ,
a11 a11
aεi1 ai1
* ci := for 2 ≤ i ≤ n,
aε11 a11
aε1j a1j
ε
* dj := for 2 ≤ j ≤ n,
a11 a11
aεi1 aε1j ai1 a1j
aεij − ε
* eij := aij − for 2 ≤ i, j ≤ n.
a11 a11
Multiplying φ ∈ Cc∞ (Ω) on both sides of (3.2.6), passing to limit, we get
Z n
φ 1
X a1j
ux1 (x)φ(x) dx = hξ , iL2 (Ω1 ;H −1 (Ω2 )),L2 (Ω1 ;H01 (Ω2 )) − hu(x), φxj i.
Ω a11 j=2
a11
CHAPTER 3. H-CONVERGENCE 49

This yields ξ 1 (x) = nj=1 a1j uxj . Similarly, multiplying φ ∈ Cc∞ (Ω) on both
P
sides of (3.2.7) and, passing to limit, one gets for 2 ≤ i ≤ n

ai1 (x1 ) 1 X ai1 (x1 )a1j (x1 )


i
ξ (x) = ξ (x) + aij − uxj (x)
a11 (x1 ) j=2
a11 (x1 )
n
X
= aij (x1 )uxj .
j=1

Thus, ξ = A0 ∇u where A0 = (aij ).

Theorem 3.2.5 (H-compactness). For any given sequence Aε ⊂ M (α, β, Ω)

there is a A0 ∈ M (α, β 2 /α, Ω) and a subsequence {Aε0 } such that Aε0 H-
H
converges to A0 , i.e., Aε0 * A0 .

Proof. Let O be an open bounded subset of Rn such that Ω ⊂ O. Let {Mε } ⊂

M (α, β, O) such that Mε = t Aε in Ω. For instance, one may choose Mε = αI
in O \ Ω. Define the operator Tε : H 1 (O) → H −1 (O) as Tε = −div(Mε ∇).
Since Mε ∈ M (α, β, O), Tε : H01 (O) → H −1 (O) is a uniformly bounded, equi-
coercive operator with constants β and α, respectively. By Theorem 3.1.7,
there is a T0 : H01 (O) → H −1 (O), with ellipticity constant α and bound β 2 /α,
such that, for a subsequence,
ε→0
hg, Tε−1 f i −→ hg, T0−1 f i ∀f, g ∈ H −1 (O).

Consider the projection function Pi : O → R defined as Pi (x) := xi , for

all i = 1, 2, . . . , n. Note that Pi ∈ H 1 (O) \ H01 (O). For any φ ∈ D(O),
φPi ∈ H01 (O). Define Fi := T0 (φPi ) ∈ H −1 (O) and wεi := Tε−1 Fi ∈ H01 (O).
This means that, for every g ∈ H −1 (O),
ε→0
hg, wεi i = hg, Tε−1 Fi i −→ hg, T0−1 Fi i = hg, φPi i,

i.e., wεi * φPi weakly in H01 (O). Choose φ ∈ D(O) such that φ ≡ 1 in Ω,
then

(i) φPi = Pi in Ω;

(ii) wεi * Pi weakly in H 1 (Ω);

(iii) −div(t Aε ∇wεi ) = Fi in Ω.

CHAPTER 3. H-CONVERGENCE 50

The last equation implies that there is a ηi ∈ [L2 (Ω)]n such that, for a
subsequence, t Aε ∇wεi * ηi weakly in [L2 (Ω)]n . Also, −div(ηi ) = Fi in Ω. The
fact that ∇wεi weakly converges to ei motivates to define t A0 ei = ηi . Thus,
set A0 (x) = aij (x) where aij (x) = (ηi )j . Let ω ⊂⊂ Ω be compactly contained
in Ω. Define the operator Aε : H01 (ω) → H −1 (ω) as Aε = −div(Aε ∇). By
Theorem 3.1.7, there is a A0 : H01 (ω) → H −1 (ω) such that, for a subsequence,
ε→0
hg, A−1 −1
ε f i −→ hg, A0 f i ∀f, g ∈ H
−1
(ω).
−1
For a fixed f ∈ H −1 (ω), set uε := A−1
ε f and u0 := A0 f . Then uε * u0
−1 −1
weakly in H0 (ω). Set ξε := Aε ∇(Aε ) : H (ω) → [L (ω)]n . Note that
1 2

β
kξε f k2,ω ≤ βkA−1
ε f k1,2,ω ≤ kf k−1,2,ω
α
is bounded in [L2 (ω)]n . Arguing as in the proof of Theorem 3.1.7, one can
conclude that there is a ξ0 : H −1 (ω) → [L2 (ω)]n such that

ξε f * ξ0 f weakly in [L2 (ω)]n .

We claim that ξ0 = A0 ∇(A−10 ). Observe that

Z Z Z
i i
ξε f · ∇wε dx = Aε ∇uε · ∇wε dx = ∇uε · t Aε ∇wεi dx.
ω ω ω

But both the LHS and RHS converge weak-* in D0 (ω). For all φ ∈ Cc∞ (ω),
using integration by parts
Z Z
i
(ξε f · ∇wε )φ dx * (ξ0 f · ei )φ dx
ω ω

and Z Z
t
(∇uε · Aε ∇wεi )φ dx * (∇u0 · t A0 ei )φ dx.
ω ω
−1
Thus, for a.e. in ω and any f ∈ H (ω),

ξ0 f = A0 ∇u0 = A0 ∇(A−1
0 f ).

Hence, ξ0 = A0 ∇(A−1
0 ).

Corollary 3.2.6. Let {Aε } ⊂ M (α, β, Ω) and fε → f strongly in H −1 (Ω). If

uε is a solution of (3.2.3) then both (3.2.4a) and (3.2.4b) are satisfied where
u0 solves (3.2.5) and A0 is a H-limit of Aε .
CHAPTER 3. H-CONVERGENCE 51

Proof. We know that uε ∈ H01 (Ω), by Lax-Milgram result, is such that

Z
Aε (x)∇uε (x) · ∇v(x) dx = hfε , viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω)
0
Ω

and kuε kH01 (Ω) ≤ (1/α)kfε kH −1 (Ω) . By Eberlein-Šmuljan theorem, there exists
a subsequence of {uε }, also denoted by uε , such that (3.2.4a) is satisfied, for
some u0 ∈ H01 (Ω).
Set ξε (x) = Aε (x)∇uε (x). Note that ξε is bounded in (L2 (Ω))n , because
the entries of Aε are bounded in L∞ (Ω) and ∇uε is bounded in (L2 (Ω))n .
Therefore, there exists a ξ0 ∈ (L2 (Ω))n such that for a subsequence of ξε
(denoted by itself),
ξε * ξ weakly in (L2 (Ω))n .
Note that, in contrast, in the one-dimensional case we had the strong con-
vergence in L2 (Ω), because we had the boundedness of ξε in H 1 (Ω).
Passing to the limit, as ε → 0, in the weak formulation of (3.2.3)
Z
ξε ∇v dx = hfε , viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω),
0
Ω
R
we get Ω ξ0 ∇v dx = hf, viH −1 (Ω),H 1 (Ω) for all v ∈ H01 (Ω). Thus, −div(ξ0 ) = f
0
in Ω. Our proof is done if we show that ξ0 = A0 ∇u0 .
The compactness of H-convergence implies the existence of a matrix
2 H
A0 ∈ M (α, βα , Ω) such that, for a subsequence (still denoted by ε), Aε * A0 .
H
Further, t Aε * t A0 . Note that, as usual, ξε is a product of two weak con-
verging sequences and finding their limit is not trivial. This was cleverly
overcome in one dimensional case. For the general case, Tartar came up with
idea of using the adjoint of Aε to define some useful test functions, called the
method of oscillating test function.
For each 1 ≤ i ≤ n, let wεi ∈ H 1 (Ω) be a solution of

−div(t Aε ∇wεi ) = −div(t A0 ei ) in Ω
(3.2.8)
wεi = xi on ∂Ω.

The function wεi ∈ H 1 (Ω), for all 1 ≤ i ≤ n and satisfies the following
properties
 i
 wε * xi weakly in H 1 (Ω),
t
Aε ∇wεi * t A0 ei weakly in (L2 (Ω))n ,
div(t Aε ∇wεi ) converges strongly in H −1 (Ω).

CHAPTER 3. H-CONVERGENCE 52

Note that for any φ ∈ D(Ω), φwεi ∈ H01 (Ω). Thus, in particular choosing
v = φwεi in the weak formulation of (3.2.3), we get
Z
i
fε , φwε H −1 (Ω),H 1 (Ω) = ξε · ∇(φwεi ) dx
0
ZΩ Z
= ξε · (∇φ)wε dx + ξε · (∇wεi )φ dx.
i
Ω Ω

Note that the last term involves product of two weak converging sequences
in (L2 (Ω))n . To overcome this difficulty, we use φuε as a test function in
(3.2.8) to get
Z Z
t t
A0 ei ∇(φuε ) dx = Aε ∇wεi ∇(φuε ) dx
Z ZΩ ZΩ
t t t
A0 ei (∇φ)uε dx + A0 ei (∇uε )φ dx = Aε ∇wεi ∇φuε dx
Ω Ω ΩZ

+ t Aε ∇wεi ∇uε φ dx
Z Z Z Ω
t
A0 ei (∇φ)uε dx + t A0 ei (∇uε )φ dx = t
Aε ∇wεi ∇φuε dx
Ω Ω ΩZ

+ ξε · (∇wεi )φ dx
Ω

Therefore, we have
Z Z
fε , φwεi H −1 (Ω),H 1 (Ω) = ξε · (∇φ)wεi
dx + t A0 ei (∇φ)uε dx
0
ΩZ Ω Z

+ t A0 ei (∇uε )φ dx − t Aε ∇wεi ∇φuε dx.

Ω Ω

Now passing to limit, as ε → 0, both sides we get

Z Z
hf, φxi iH −1 (Ω),H 1 (Ω) = ξ0 · (∇φ)xi dx + t A0 ei (∇φ)u0 dx
0
ΩZ Ω Z

+ t A0 ei (∇u0 )φ dx − t A0 ei (∇φ)u0 dx
Z Z Ω Z Ω
t
ξ0 · ∇(φxi ) dx = ξ0 · (∇φ)xi dx + A0 ei (∇u0 )φ dx
Ω Z ZΩ Ω

ξ0 · ei φ dx = A0 ∇u0 · ei φ dx.
Ω Ω
CHAPTER 3. H-CONVERGENCE 53

Hence ξ0 = A0 ∇u0 and u0 solves (3.2.5).

Note that the choice of A0 depends on the subsequence chosen and u0
is not unique. All the above arguments were for a subsequences, extracted
suitably.
Example 3.3 (Periodic Case). If Aε (x) = (aεij (x)) where aεij (x) = aij xε a.e.

x ∈ Rn such that aij : Y = (0, 1)n → R, extended Y -periodically to all

Rn and restricted to Ω. In the periodic case, one can explicitly compute
the matrix A0 , as seen in the informal asymptotic expansion. Note that,
in the proof of the above theorem, the matrix A0 is obtained as a limit
of a converging subsequence of {Aε }, which existed due to compactness of
H-convergence. This A0 was then used in defining the functions wεi using
(3.2.8). In the periodic case, we expect the function wεi to be periodic and
hence solve (3.2.8) in Y , instead of Ω. For each i = 1, 2, . . ., we begin by
solving the cell equation
t i

 −divy ( A∇y (w (y) − yi )) = 0 in Y
wi (y) is Y -periodic in y (3.2.9)
1
R

|Y | Y
wi (y) dy = 0

Set wεi (x) = εwi xε . The reason behind having a factor of ε while defining

wεi is to avoid the factor of 1/ε while computing its first derivative. Thus,
h  x i x 1 x
∇x wεi (x) = ∇x εwi = ε∇y wi = ∇y w i
ε ε ε ε
and the vector x x
t
Aε (x)∇wεi (x) t
= A ∇y w i
ε ε
2 n
in (L (Ω)) is also Y -periodic. Therefore, by Theorem 1.1.3, we have that
Z
t i 1 t
Aε (x)∇wε (x) * A(y)∇y wi (y) dy weakly in (L2 (Ω))n .
|Y | Y

Recall that our aim is to identify A0 such that (3.2.8) is satisfied. Consider
the function φ ∈ D(Ω) and set φε (y) = φ(εy) for y ∈ (0, 1)n and extended
to all of Rn . Using φε ∈ D(Rn ) as a test function in the cell equation of wi
above, we get
Z Z
t i t
A(y)∇y w (y) · ∇y φε (y) dy = A(y)ei · ∇y φε (y) dy
Rn Rn
CHAPTER 3. H-CONVERGENCE 54

Using the change of variable x = εy and ∇y = (1/ε)∇x , we get

Z x Z
t 1 dx x 1 dx
A ∇x wεi (x) · ∇x φ(x) = t
A ei · ∇x φ(x)
Ω ε ε ε ε ε ε
Z ZΩ
t
Aε (x)∇wεi (x) · ∇φ(x) dx = t
Aε (x)ei · ∇φ(x) dx.
Ω Ω

By passing to the limit, as ε → 0,

Z  Z 
1 t i


A(y) ∇y w (y) − ei dy · ∇φ(x) dx = 0.
Ω |Y | Y

We need to define A0 such that (3.2.8) is satisfied. Thus, we set

Z Z  
t 1 t i t
A0 ei = A(y)∇y w (y) dy − A(y) dy ei ,
|Y | Y Y

and, for all φ ∈ D(Ω)

Z
t
A0 ei · ∇φ(x) dx = 0.
Ω

By density, the above equality is true for all φ ∈ H01 (Ω). Therefore,
Z Z  
1 i
A0 ei = A(y)∇y χ (y) dy − A(y) dy ei ,
|Y | Y Y

where χi solves the cell equation (3.2.9) where t A is replaced with A. Note
that formula is same as (1.5.7).
Note that the A0 obtained is independent of the choice of the subsequence
of Aε . Thus, u is a unique solution and all the convergences (3.2.4a) and
(3.2.4b) are true for the entire sequence and not just for subsequences (cf.
Eberlein-Šmuljan).
Recall that in the one dimensional case, we encountered the problem of
product of two weak converging sequences (recall the sequence ξε ). In the one
dimensional case, it was easy overcome this constraint by other means. How-
ever, the same idea would fail in higher dimension. The following theorem,
popular as compensated compactness, is a fix of the problem in higher dimen-
sions and is due to F. Murat and L. Tartar (cf. [Mur78a, Mur79, Tar79]).
CHAPTER 3. H-CONVERGENCE 55

Lemma 3.2.7 (div-curl lemma). Let uε and vε be two sequences in (L2 (Ω))n
such that
uε * u0 weakly in (L2 (Ω))n

vε * v0 weakly in (L2 (Ω))n .

If {div uε } is compact in H −1 (Ω) and {curl vε }1 is bounded in (L2 (Ω))n×n ,
then
uε vε → u0 v0 weak* in D0 (Ω).

H
Theorem 3.2.8 (Energy convergence). If Aε * A0 then
Z Z
Aε ∇uε .∇uε dx → A0 ∇u0 .∇u0 dx (3.2.10)
Ω Ω

where uε and u0 are, respectively, the unique solution of (3.2.3) and (3.2.5).

The energy convergence also amounts to saying that the quadratic forms
associated with the operators converge, i.e., hAε uε , uε i → hA0 u0 , u0 i. In
section §4.3 (cf. Lemma 4.1), we will observe that this is actually subject to
a special type of convergence called the Γ-convergence.
The energy functional (cf. (3.2.10)) involves a product of two weakly
converging sequences and we have shown that the limit of the product is
equal to the product of the limit. This property is not true, in general, and
was possible due to the div-curl lemma.

3.3 Correctors
We have from (3.2.4a) that

∇uε * ∇u0 weakly in (L2 (Ω))n .

In general, the above convergence is not strong. However, by adjusting the

term ∇u0 , we get a strong convergence (cf. Theorem 3.3.3). This adjustment
is done by introducing the corrector matrix.
 
1 ∂vi ∂vj
for any v ∈ (L2 (Ω))n , (curlv)ij = ∂xj − ∂xi
CHAPTER 3. H-CONVERGENCE 56

The corrector matrices are obtained by looking for functions χiε ∈ H 1 (Ω),
for 1 ≤ i ≤ n, with the following properties:
 i
 χε * xi weakly in H 1 (Ω),
Aε ∇χiε * A0 ei weakly in (L2 (Ω))n , (3.3.1)
div(Aε ∇χiε ) converges strongly in H −1 (Ω).


One procedure to build a function with above properties is by defining χiε ∈

H 1 (Ω), for 1 ≤ i ≤ n, as a solution of

−div(Aε ∇χiε ) = −div(A0 ei ) in Ω
(3.3.2)
χiε = xi on ∂Ω.

Then the corrector matrix Dε ∈ (L2 (Ω))n×n is defined as Dε ei = ∇χiε for

1 ≤ i ≤ n. For other choices of χiε , we may have D̃ε . But they are “unique”
in the sense that
Dε − D̃ε → 0 in [L2loc (Ω)]n×n .
Some interesting properties of the corrector functions are given by the fol-
lowing proposition, the proof of which can be found in [CD99, MT97].

Theorem 3.3.1. Let Aε ∈ M (α, β, Ω), χiε be a function with properties

(3.3.1) and Dε ei = ∇χiε . Also, let Aε H-converge to A0 , then the following
are true:

(a) Dε * I weakly in [L2 (Ω)]n×n , where I is the identity matrix.

(b) Aε Dε * A0 weakly in [L2 (Ω)]n×n .

(c) t Dε Aε Dε * A0 weak* in [D0 (Ω)]n×n .

Proof. Note that {Dε } is bounded in [L2 (Ω)]n×n . Let Φ ∈ [Cc∞ (Ω)]n . Then
n
X
Φ(x) = Φi (x)ei
i=1

where Φi ∈ Cc∞ (Ω). Note that

Z n Z
X n Z
X
Dε Φ dx = Dε ei Φi dx = ∇χiε Φi dx.
Ω i=1 Ω i=1 Ω
CHAPTER 3. H-CONVERGENCE 57

Therefore,
Z n Z
X Z
lim Dε Φ dx = ei Φi dx = Φ dx.
ε→0 Ω Ω Ω
i=1

Lemma 3.3.2. Let Aε * A0 and fε → f in H −1 (Ω). If uε := A−1 ε fε and

u0 := A−1 f then, for any Φ ∈ [Cc∞ (Ω)]n and φ ∈ Cc∞ (Ω),
Z Z
[Aε (∇uε − Dε Φ) · (∇uε − Dε Φ)]φ dx → [A0 (∇u0 − Φ) · (∇u0 − Φ)]φ dx.
Ω Ω

Proof. Let Φ ∈ [Cc∞ (Ω)]n . Then

n
X
Φ(x) = Φi (x)ei
i=1

where Φi ∈ Cc∞ (Ω). Consider

Z Z
[Aε (∇uε − Dε Φ) · (∇uε − Dε Φ)]φ dx = (Aε ∇uε · ∇uε )φ dx
Ω Ω
n Z
X
− (Aε ∇uε · Dε ei )Φi φ dx
i=1 Ω
Xn Z
− (Aε Dε ei · ∇uε )Φi φ dx
Ω
Zi=1
+ (Aε Dε ei · Dε ej )Φi Φj φ dx
Ω

Passing to the limit, as ε → 0, we have the result.

The interest of the corrector matrix Dε is the following theorem:
H
Theorem 3.3.3 (cf. [CD99]). Let Aε * A0 . Let uε solve (3.2.3) and fε
converge strongly to f . If uε weakly converge to u0 in H 1 (Ω), then

∇uε − Dε ∇u0 → 0 strongly in (L1loc (Ω))n .

Moreover, if Dε ∈ (Lr (Ω))n×n , kDε k(Lr (Ω))n ≤ C0 for 2 ≤ r ≤ +∞ and

∇u0 ∈ (Ls (Ω))n , 2 ≤ s < +∞, then

∇uε − Dε ∇u0 → 0 strongly in (Ltloc (Ω))n ,

CHAPTER 3. H-CONVERGENCE 58

 rs
where t = min 2, r+s . If u0 is a solution of (3.2.5) then

∇uε − Dε ∇u0 → 0 strongly in (Lt (Ω))n .

Proof. Observe that if u0 ∈ Cc∞ (Ω) then we choose Φ = ∇u0 in the previous
lemma and the result is true. For any δ > 0, choose Φ ∈ Cc∞ (Ω) such that

k∇u0 − φk(Ls (Ω))n ≤ δ ∀s < ∞.

1 1
Therefore, for q
= s
+ 1r ,

k∇u0 − φk[Lq (Ω)]n ≤ C0 δ.

For q ≥ 1, we note that both r, s ≥ 2. Let ω ⊂⊂ Ω and take φ ∈ Cc∞ (Ω)

such that φ = 1 on ω and 0 ≤ φ ≤ 1. Consider
Z
2
lim sup αk∇uε − Dε Φk2,ω ≤ lim [Aε (∇uε − Dε Φ) · (∇uε − Dε Φ)]φ dx
ε→0 ε→0 Ω
Z
= [A0 (∇u0 − Φ) · (∇u0 − Φ)]φ dx
Ω
β2 C 2β 2
≤ k∇u0 − Φk22,Ω ≤ 0 δ 2 .
α α
Because
∇uε − Dε ∇u0 = (∇uε − Dε Φ) − (Dε ∇u0 − Dε Φ)
and, choosing t = min{2, q},
 
β
lim sup k∇uε − Dε ∇u0 k[Lt (ω)]n ≤ + 1 C0 δ.
ε→0 α

3.4 Generalised Energy Convergence

A question of similar interest is to know the limit of k∇uε k22,Ω . One knows
that this quantity is uniformly bounded and hence, at least for a subsequence,
converges. We know that the limit is not k∇u0 k22,Ω , since we know from
the above theorem that uε does not converge to u0 strongly in H01 (Ω). We
would like to know the limit and whether it can be expressed in terms of
CHAPTER 3. H-CONVERGENCE 59

the functionR u0 . More generally, the problem can be framed as identifying

the limit of Ω Bε ∇uε .∇uε dx where Bε is a family of matrices in M(c, d, Ω).
More precisely, does there exist a matrix B 0 ∈ M(c0 , d0 , Ω) such that, at least
for a subsequence, we have
Z Z
Bε ∇uε .∇uε dx → B 0 ∇u0 .∇u0 dx?
Ω Ω

The convergence question posed above is answered when Bε = Aε (cf.

(3.2.10)), in which case, it has been observed that B ] = A0 , the H-limit of
Aε . The general problem was studied by Kesavan and Rajesh in [KR02].

Theorem 3.4.1. Let Aε ∈ M(a, b, Ω), Bε ∈ M(c, d, Ω) and χiε be a function

with properties (3.3.1) and Dε ei = ∇χiε . Also, let Aε H-converge to A0 , then
the following are true:

(a) There exists a B ] (depending only on {Aε } and {Bε }) such that
t
Dε Bε Dε * B ] weak* in (D0 (Ω))n×n . (3.4.1)

(b) If Bε = Aε for all ε, then B ] = A0 .

(c) If Bε ’s are symmetric, then B ] is symmetric.

(d) B ] ∈ M c, d( ab )2 , Ω .

The existence of the matix B ] , mentioned in the above proposition, was

shown by Kesavan and Vanninathan, for the periodic case (cf. [KV77]), and
by Kesavan and Saint Jean Paulin in the general case (cf. [KP97]), in the
process of homogenizing an optimal control problem.
It was observed that the required B 0 is actually the B ] obtained in Propo-
sition 3.4.1 and thus
Z Z
Bε ∇uε .∇uε dx → B ] ∇u0 .∇u0 dx. (3.4.2)
Ω Ω

Therefore, if C is the positive square root of the matrix B ] when Bε = I, for

all ε > 0, then
k∇uε k22,Ω → kC∇u0 k22,Ω .
CHAPTER 3. H-CONVERGENCE 60

3.5 Optimal Bounds

The computation of H-limit A0 is, in general, not easy to characterise for
a given sequence of matrices {Aε }. We ask if it is possible to conclude
something in the simple case when Aε = aε (x)I, i.e., are isotropic and the
domain is a (non-periodic) mixture of two materials. Let Ω = Ωε1 ∪ Ωε2 such
that Ωε1 ∩ Ωε2 = ∅. Let
(
a1 if x ∈ Ωε1
aε (x) =
a2 if x ∈ Ωε2

such that 0 < α ≤ aε (x) ≤ β a.e. on Ω. Without loss of generality,

assume a1 ≤ a2 . If 1Ωε1 denotes the characteristic function of Ωε1 then
aε (x) = a1 1Ωε1 (x) + a2 (1 − 1Ωε1 (x)). Since 1Ωε1 is bounded in L∞ (Ω), there
is a θ ∈ L∞ (Ω) such that

1Ωε1 * θ weak-* in L∞ (Ω).

Therefore, aε * a1 θ(x) + a2 (1 − θ(x)) weak-* in L∞ (Ω). Note that Aε (x) =

aε (x)I is symmetric, therefore, its H-limit A0 (for a subsequence) is also
symmetric and is in M (α, β, Ω). Thus, A0 admits strictly positive eigenvalues
λ1 (x), . . . , λn (x) defined in Ω.
Theorem 3.5.1. The eigenvalues λ1 (x), . . . , λn (x) of A0 satisfy the following
inequalities, a.e. in Ω:

A∗ (x) ≤ λi (x) ≤ A(x) ∀i = 1, 2, . . . , n;

n
X 1 1 n−1
≤ ∗ + ;
i=1
λi (x) − a1 A (x) − a1 A(x) − a1
and n
X 1 1 n−1
≤ ∗
+
i=1
a2 − λi (x) a2 − A (x) a2 − A(x)
where  −1
∗ θ(x) 1 − θ(x)
A (x) := +
a1 a2
and
A(x) := a1 θ(x) + (1 − θ(x))a2 .
CHAPTER 3. H-CONVERGENCE 61

µ(Ωε )
In the two dimensional case and when θ = µ(Ω)1 , the above result has a
nice geometric interpretation. For each θ ∈ (0, 1), consider the set of points
 
a1 a2
(X(θ), Y (θ)) := , a2 − (a2 − a1 )θ .
(a2 − a1 )θ + a1

Note that (X(0), Y (0)) = (a2 , a2 ) and (X(1), Y (1)) = (a1 , a1 ). The points
subscribe to a concave hyperbola H1 given by Y = a1 + a2 − a1Xa2 , for a1 ≤
X, Y ≤ a2 , above the line Y = X. The reflection of H1 along Y = X line
gives a convex hyperbola H2 with equation
a1 a2
Y = .
a1 + a2 − X

Let G(θ) := (A∗ , A∗ ) and M (θ) := (A, A) be the points on Y = X line. Then
the points G(θ), M (θ), (A∗ , A) ∈ H1 and (A, A∗ ) ∈ H2 form a square. The
condition
A∗ (x) ≤ λi (x) ≤ A(x) ∀i = 1, 2
implies that the eigenvalues of the H-limit A0 are contained in the square
describes above. The remaining two inequalities imply that the eigenvalues
are contained between two hyperbolas H3 and H4 given by the equation
1 1 1 1
H3 : + = ∗ +
X − a1 Y − a1 A − a1 A − a1

and
1 1 1 1
H4 : + = + .
a2 − X a2 − Y a2 − A∗ a2 − A
The well-known Hashin-Shtrikman (Clausius-Mossoti or Lorentz-Lorenz
or Maxwell-Garnett) inequalities is the two dimensional case with A0 is
isotropic. In this case, the inequalities reduce to
   
θa1 + (1 − θ)a2 + a2 θa1 + (1 − θ)a2 + a1
a1 ≤ λ ≤ a2 .
(1 − θ)a1 + θa2 + a1 (1 − θ)a1 + θa2 + a2
CHAPTER 3. H-CONVERGENCE 62
Chapter 4

Γ-Convergence

4.1 Motivation
Recall that the weak solution u of (1.2.2) can also be characterized as the
minimizer in H01 (Ω) of the functional
Z
1
J(v) = A∇v.∇v dx − hf, viH −1 (Ω),H 1 (Ω) ,
2 Ω 0

i.e.,
J(u) = min
1
J(v).
v∈H0 (Ω)

Thus, the problem of studying the asymptotic behaviour of the second order
elliptic problem 
−div(Aε ∇uε ) = f in Ω
uε = 0 on ∂Ω,
with {Aε } ⊂ M (α, β, Ω) is equivalent to finding a functional J on H01 (Ω)
whose minimum is the solution of the homogenized elliptic equation such
that both the minimizers and minima of Jε converge to the minimizers and
minima of J. Thus, we need to study the convergence of functionals such
that the minimizers and minima converge.

4.2 Direct Method of Calculus of Variation

To motivate the direct method of calculus of variation, we begin by recalling
a basic result due to Karl Weierstrass, known as the extreme value theorem.

63
CHAPTER 4. Γ-CONVERGENCE 64

Our aim, in this section, will be to generalise this basic result to an arbitrary
topological space. We recall the proof to motivate some definitions.
Theorem 4.2.1 (Extreme value theorem). Any real valued continuous func-
tion f on a closed bounded interval [a, b] attains its minimum.
Proof. We show f has a lower bound. Suppose f is not bounded below then
there exists a sequence {xn } ⊂ [a, b] such that f (xn ) > n. Since the sequence
xn is bounded, by Bolzano-Weierstrass theorem, there exists a convergent
subsequence {xnk } and let x be its limit. Thus, by continuity of f , f (xnk )
converges to f (x) which is a contradiction since f (xnk ) > nk ≥ k. Thus
there exists a infimum (greatest lower bound) C such that f (x) ≥ C for all
x ∈ [a, b]. It now remains to show that there exists a x ∈ [a, b] such that
f (x) = C.
Let {yn } ⊂ [a, b] be a sequence such that f (yn ) ≤ C + 1/n. Since C ≤
f (yn ) for all n, we have that f (yn ) converges to C (thus the sequence yn
is called minimizing sequence). By applying Bolzano-Weierstrass theorem
again, there exists a convergent subsequence {ynk } and let y be its limit.
Using continuity of f again, f (ynk ) converges to f (y). Thus f (y) = C.
Moreover, since [a, b] is closed y ∈ [a, b].
We shall, henceforth, concentrate on the minimum of the function f ,
since the corresponding result for maximum can be obtained by applying the
results to −f .
Definition 4.2.2. Let X be a topological space. A function F : X → R =
R ∪ {−∞, +∞} is said to be lower semicontinuous (lsc) at a point x ∈ X if

F (x) = sup inf F (y).

U ∈N (x) y∈U

F is lower semicontinuous on X if F is lower semicontinuous at each point

x ∈ X.
Remark 4.2.3. Let X be a topological space satisfying first axiom of count-
ability. Then a function F : X → R is lower semicontinuous at x ∈ X
iff
F (x) ≤ lim inf F (xn )
n→∞

for every sequence {xn } converging to x ∈ X. This is the sequential charac-

terisation of the lower semi-continuity.
CHAPTER 4. Γ-CONVERGENCE 65

Note that one can, in fact, weaken the hypothesis of the extreme value
theorem.
Exercise 4.1. Prove the Extreme value theorem when f is lower semicontiu-
ous. Replace lower semicontinuity hypothesis with upper semicontinuity to
obtain the maximizer.
Exercise 4.2. Show that if F is lower semicontinuous then the sublevel set
{F ≤ α} := {x ∈ X : F (x) ≤ α} is closed for all α ∈ R.
We have already seen that the continuity property in Extreme Value
theorem can be relaxed to lower semicontinuity. Another crucial element of
the proof is the Bolzano-Weierstrass theorem which is about the compactness
of the interval [a, b].

Definition 4.2.4. A function F : X → R is coercive on X if the closure of

the sublevel set {F ≤ α} := {x ∈ X : F (x) ≤ α} is compact in X for every
α ∈ R.

Exercise 4.3. Show that if F is a coercive functional on X and G ≥ F , then

G is coercive.
Exercise 4.4. If F is coercive then there is a non-empty compact set K such
that
inf F (x) = inf F (x).
x∈X x∈K

Definition 4.2.5. A minimizing sequence for F : X → R in X is a sequence

{xn } in X such that
inf F (y) = lim F (xn )
y∈X n→∞

Theorem 4.2.6. Let X be a topological space. Assume that the function

F : X → R is coercive and lower semicontinuous. Then F has a minimizer
in X.

Proof. If F is identically +∞ or −∞, then every point of X is a minimum

point for F . If F takes the value −∞, then all those points are minimizers
of F . Suppose now that F is not identically +∞ and F > −∞. Let {xn } be
a sequence in X such that

lim F (xn ) = inf F (y) := d.

n→∞ y∈X
CHAPTER 4. Γ-CONVERGENCE 66

The existence of such a sequence is clear. Without loss of generality, we can

assume that F (xn ) < +∞ for all n. Let α := supn F (xn ) < +∞. Moreover,
since F is coercive, the sublevel set {F ≤ α} is compact and hence there is a
subsequence {xk } of {xn } which converges to a point x ∈ X. Since F is lsc
we obtain
d = inf F (y) ≤ F (x) ≤ lim inf F (xk ) = d.
y∈X k→∞

Thus, F (x) = d and hence is the minimizer of F in X. which proves our

theorem.

Definition 4.2.7. A family of functionals {Fn } on X is said to be equi-

coercive, if for every α ∈ R, there is a compact set Kα of X such that the
sublevel sets {Fn ≤ α} ⊆ Kα for all n.

Exercise 4.5. If {Fn } is a family of equi-coercive, then there is a non-empty

compact K (independent of n) such that

inf Fn (x) = inf Fn (x).

x∈X x∈K

Proposition 4.2.8. A family of functions Fn on X is equi-coercive if and

only if there exists a lower semicontinuous coercive function Ψ : X → R such
that Fn ≥ Ψ on X, for every n.

Proof. Let Ψ : X → R be a lower semicontinuous coercive function such that

Fn ≥ Ψ on X, for every n. Set Kα := {Ψ ≤ α}. Kα is closed and compact
because of the lsc and coercivity of Ψ, respectively. Moreover, {Fn ≤ α} ⊆
Kα , for all n. Thus, Fn are equi-coercive.
Conversely, let Fn be equi-coercive. Then, for each α ∈ R, there is a
compact set Kα such that {Fn ≤ α} ⊆ Kα , for all n. We shall now define
Ψ : X → R as
(
+∞, if x 6∈ Kα , ∀α ∈ R
Ψ(x) =
inf{α | x ∈ Kβ for all β > α}.

We now show that Ψ ≤ Fn for all n. Let x ∈ X. If Fn (x) = +∞, for

all n, then by definition, Ψ(x) = Fn (x) = +∞. Otherwise, let Fk be a
subfamily such that Fk (x) = βk < ∞. Thus, x ∈ Kβk for all k and hence
Ψ(x) = inf k {βk } ≤ Fn (x). Thus, Ψ(x) ≤ F (x), for every x ∈ X. It now
CHAPTER 4. Γ-CONVERGENCE 67

remains to show that Ψ is lsc and coercive. Note that any x ∈ {Ψ ≤ α}

implies x ∈ Kβ for all β > α. Therefore, the sublevel

{Ψ ≤ α} = ∩β>α Kβ

is an arbitrary intersection compact sets and hence is closed and compact.

Definition 4.2.9. Let X be a vector space. We say a function F : X → R

is convex if
F (tx + (1 − t)y) ≤ tF (x) + (1 − t)F (y)
for every t ∈ (0, 1) and for every x, y ∈ X such that F (x) < +∞ and
F (y) < +∞. We say a function F : X → R is strictly convex if F is not
identically +∞ and

F (tx + (1 − t)y) < tF (x) + (1 − t)F (y)

for every t ∈ (0, 1) and for every x, y ∈ X such that x 6= y, F (x) < +∞ and
F (y) < +∞.

If F is constant, then one can see that X is the set of all minimizers of F .
We now show that with strict convexity the minimizer, if exists, is unique.

Proposition 4.2.10. Let X be a vector space. Let F : X → R be a strictly

convex function. Then F has at most one minimizer in X.

Proof. If x and y are two minimizers of F in X, then

F (x) = F (y) = d := min F (z) < +∞.

z∈X

If x 6= y, by strict convexity we have

F (tx + (1 − t)y) < tF (x) + (1 − t)F (y) = d, ∀t ∈ (0, 1).

This contradicts the fact that d is a minimum of F . Therefore x = y.

Thus, combining Theorem 4.2.6 and Proposition 4.2.10, we have the fol-
lowing theorem.

Theorem 4.2.11. Let X be a topological vector space and let F : X → R be

a lower semicontinuous, coercive and strictly convex functional, then F has
a unique minimizer.
CHAPTER 4. Γ-CONVERGENCE 68

4.3 Sequential Γ-Convergence

The notion of Γ-convergence was introduced by Ennio De Giorgi in a sequence
of papers (cf. [GS73, Gio75, GF75]). An excellent account of this concept is
the book of Dal Maso [DM93] and A. Braides [Bra02].

Definition 4.3.1. A function F is said to be the Γ-limit of Fn (denoted as

Γ
Fn → F ) w.r.t the topology of X, if F = F + = F − , where

(i)
F − (x) = sup lim inf inf Fn (y).
U ∈N (x) n→∞ y∈U

(ii)
F + (x) = sup lim sup inf Fn (y).
U ∈N (x) n→∞ y∈U

We say F − is the Γ-lower limit and F + is the Γ-upper limit.

Remark 4.3.2. If X is a topological space satisfying first axiom of countabil-

ity, the Γ-limit can be characterised as satisfying the following two conditions:

(i) For every x ∈ X and for every sequence {xn } converging to x in X, we

have
lim inf Fn (xn ) ≥ F (x).
n→∞

(ii) For every x ∈ X, there exists a sequence {xn } converging to x in X

(called the Γ-realising sequence) such that

lim Fn (xn ) = F (x).

n→∞

Γ Γ
Exercise 4.6. Show that if Fn → F , Gn → G and Fn ≤ Gn , for each n, then
F ≤ G.
Exercise 4.7. Show that if Fn Γ-converges to F , then F is lower semicontin-
uous.
Exercise 4.8. Let X be a topological vector space. Show that if Fn : X → R
is convex for each n, then Γ-lim supn Fn is convex. Also show that the Γ-
lim inf n Fn is, in general, not convex.
CHAPTER 4. Γ-CONVERGENCE 69

Exercise 4.9. Compute the Γ-limit of a constant sequence Fn = F on X.

Theorem 4.3.3. Let X be a topological space and Fn be a family functions
on X.
1. If U is an open subset of X, then

inf F + (x) ≥ lim sup inf Fn (x).

x∈U n x∈U

2. If K is a compact subset of X, then

inf F − (x) ≤ lim inf inf Fn (x).

x∈K n x∈K

Proof. 1. Let x ∈ U . Then, from the definition of Γ-upper limit which

says F (x) is sup over all neighbourhoods of x, we have

F + (x) ≥ lim sup inf Fn (y).

n→∞ y∈U

Therefore,
inf F + (x) ≥ lim sup inf Fn (y).
x∈U n→∞ y∈U

2. Since F − is lsc and by the compactness of K, F − attains its minimum

on K (cf. Theorem 4.2.6). Set d := lim inf n inf x∈K Fn (x) and let xn
be a sequence (extracting subsequence, if necessary) in K such that
limn Fn (xn ) = d. Thus, there is a subsequence xk which converges to
some x ∈ K. Therefore, for every neighbourhood U of x, inf y∈U Fk (y) ≤
Fk (xk ) for infinitely many k. Now, taking lim inf both sides,

lim inf inf Fk (y) ≤ lim inf Fk (xk ) = d

k y∈U k

and taking supremum over all neighbourhoods U of x, we still have

F − (x) = sup lim inf inf Fk (y) ≤ d.

U k y∈U

Now, since x ∈ K, inf x∈K F − (x) ≤ d.

Theorem 4.3.4 (Fundamental Theorem of Γ-convergence). Let X be a topo-

logical space. Let {Fn } be a equi-coercive family of functions and let Fn Γ-
converges to F in X, then
CHAPTER 4. Γ-CONVERGENCE 70

(i) F is coercive.

(ii) limn→∞ dn = d, where dn = inf x∈X Fn (x) and d = inf x∈X F (x). That
is, the minima converges.

(iii) The minimizers of Fn converge to a minimizer of F .

Proof. Since {Fn } are equi-coercive, by Proposition 4.2.8, there is a lsc, co-
ercive function Ψ on X such that Fn ≥ Ψ. Now, by Exercise 4.6, F ≥ Ψ and
by Exercise 4.3 F is coercive.
Now, by putting U = X in Theorem 4.3.3, we get d ≥ lim supn dn . We
now need to show that d ≤ lim inf n dn . If Fn are all not identically +∞, then
lim inf n dn < +∞. Set lim inf n dn = α. By the equi-coercivity of Fn , there is
a compact set Kα such that {Fn ≤ α} ⊆ Kα , for all n. Consider,

d ≤ inf F (y) ≤ lim inf inf Fn (y)

y∈Kα n y∈Kα
= lim inf inf Fn (y)
n y∈X
= lim inf dn .
n

Thus, lim supn dn ≤ d ≤ lim inf n dn and hence, limn dn = d.

Since F is coercive and lsc (Γ-limit is always lsc), then by Theorem 4.2.6,
F attains its minimum. Let x∗n be a minimizer of Fn , then since Fn are
equi-coercive x∗n belong to a compact set K of X and hence converges up to
a subsequence. Let x∗n → x∗ in X. We need to show that F (x∗ ) = d. By
Γ-lower limit,
F (x∗ ) ≤ lim inf Fn (x∗n ) = lim inf dn = d.
n n

But, d ≤ F (x∗ ). Hence d = F (x∗ ).

Theorem 4.3.5 (Compactness). If X is a topological space satisfying second

axiom of countability then any sequence of functionals Fn : X → R has a Γ-
convergent subsequence.

Proof. Let {Uk }k∈N be a countable base for the topology of X. For each k,
let dnk = inf y∈Uk Fn (y). Thus, {dnk }n is a sequence in R which is compact,
hence has a subsequence {dm k }m whose limit as m → ∞ exists in R. Thus,
for each k, we have subsequence {dm k }m whose limit as m → ∞ exists in R.
CHAPTER 4. Γ-CONVERGENCE 71

Choose the diagonal sequence dkk whose limit exists n R as k → ∞. In other

words, we have chosen a subsequence Fk of Fn such that

lim dkk = lim inf Fk (y).

k→∞ k→∞ y∈Uk

Now, define F (x) = supU ∈N (x) limk→∞ inf y∈Uk Fk (y) and we have by definition
Fk Γ-converges to F .
H Γ
Example 4.1. Let Aε * A0 then we wish to show that Jε * J in the weak
topology of H01 (Ω) where
Z
Jε (u) = Aε ∇u.∇u dx
Ω

and Z
J(u) = A0 ∇u.∇u dx.
Ω

Let u ∈ H01 (Ω). We need to find a sequence {uε } in H01 (Ω) such that uε
converges to u weakly in H01 (Ω) and limε→0 Jε (uε ) = J(u). Let uε ∈ H01 (Ω)
be the solution of
−div(Aε ∇uε ) = −div(A0 ∇u). (4.3.1)
1
Then,
R R H-convergence that uε * u weakly in H0 (Ω) and
it follows from
A ∇uε .∇uε dx → Ω A0 ∇u.∇u dx. Thus, we have shown the existence of
Ω ε
a sequence {uε } converging weakly to u in H01 (Ω) such that

lim Jε (uε ) = J(u).

ε→0

Now, let wε ∈ H01 (Ω) be a sequence such that wε * u weakly in H01 (Ω).
Then, the solution uε obtained in (4.3.1) minimizes the functional
Z
1
Jε (v) − A0 ∇u.∇v dx.
2 Ω

Hence, in particular, we have

Z Z Z
1 1
Aε ∇wε .∇wε dx − A0 ∇u.∇wε dx ≥ Aε ∇uε .∇uε dx
2 Ω Ω 2 Ω
Z
− A0 ∇u.∇uε dx
Ω
CHAPTER 4. Γ-CONVERGENCE 72

and taking liminf on both sides of above inequality we have

lim inf Jε (wε ) ≥ J(u).
ε→0

Γ
Hence Jε * J in the weak topology of H01 (Ω).
In the above example, we assume the H-convergence of the matrix coeffi-
cients to describe the Γ-limit. A general question of interest is the following:
If for any sequence of functionals, by compactness, there is a Γ-limit, then
under what conditions one can get an integral representation of Γ-limit. In
the next section, we describe the situation in one-dimension.

4.4 Integral Representation: One Dimension

For any given 1 < p < ∞ and c1 , c2 , c3 > 0, let F = F(p, c1 , c2 , c3 ) be the
class of all functionals F : W 1,p (Ω) → [0, +∞) such that
Z
F (u) = f (x, ∇u(x)) dx
Ω

where f : Ω × Rn → [0, +∞)

H 1. is a Borel function such that ξ 7→ f (x, ξ) is convex for all x ∈ Ω,
H 2. and satisfies the growth conditions of order p
c1 |ξ|p − c2 ≤ f (x, ξ) ≤ c3 (1 + |ξ|p ), ∀x ∈ Ω, ξ ∈ Rn .
Exercise 4.10. If f satisfies H1 and H2, then f satisfies the local Lipschitz
condition
|f (x, ξ) − f (x, ζ)| ≤ k(1 + |ξ|p−1 + |ζ|p−1 )|ξ − ζ| ∀ξ, ζ ∈ Rn .
The constant k depends only on c3 and p.
We take n = 1 in the dimension of Euclidean space and set Ω = (a, b).
Observe that any functional in F is invariant by addition of a constant c,
i.e., F (u + c) = F (u). Thus, it is sufficient to characterize in the space
X = {u ∈ W 1,p (Ω) | u(b) = 0}
equipped with Lp norm instead of W 1,p (Ω). Since X is embedded in L∞ (a, b),
L1 (a, b) ⊂ X ? .
CHAPTER 4. Γ-CONVERGENCE 73

Proposition 4.4.1. Let X = {u ∈ W 1,p (Ω) | u(b) = 0} equipped with Lp

norm. Let F ∈ F and consider its integrand f as a function on X, then
F ? : X ? → R is given as
Z b  Z x 
? ?
F (φ) = f x, − φ(t) dt dx, ∀φ ∈ L1 (a, b).
a a

Proof. Let us assume f (x, ·) ∈ C 1 (R) for all x ∈ (a, b). Due to the growth
conditions and continuity of f ,

f ? (x, ξ ? ) = sup{ξ ? · ξ − f (x, ξ)} = max{ξ ? · ξ − f (x, ξ)}.

ξ∈R ξ∈R

Thus, if ζ is the point at which maximum is attained, then

∂f
f ? (x, ζ ? ) = ζ ? · ζ − f (x, ζ) if and only if ζ ? − (x, ζ) = 0. (4.4.1)
∂ζ
Let φ ∈ L1 (a, b), define Φ ∈ W 1,1 (a, b) as,
Z x
Φ(x) = − φ(t) dt.
a

Note that Φ0 = −φ and Φ(a) = 0. Thus, the convex conjugate of F is given

as
Z b 
? 0
F (φ) = sup [φ(x)v(x) − f (x, v (x)] dx
v∈X a
Z b 
0 0
= sup [Φ(x)v (x) − f (x, v (x)] dx (integration by parts)
v∈X a
Z b 
0 0
= max [Φ(x)v (x) − f (x, v (x)] dx
v∈X a
Z b
= [Φ(x)u0 (x) − f (x, u0 (x)] dx.
a

By computing Euler equations, we have Φ− ∂f ∂u

(x, u0 ) = c, for some constant c.
But Φ(a) = 0 and ∂u (a, u0 (a)) = 0, implies that c = 0 and thus, Φ = ∂f
∂f
∂u
(x, u0 )
? 0
a.e. on (a, b). By choosing ζ = Φ(x) and ζ = u (x) in (4.4.1), we have
∂f
Φ(x) = (x, u0 (x)) if and only if f ? (x, Φ(x)) = Φ(x)u0 (x) − f (x, u0 (x)).
∂u
CHAPTER 4. Γ-CONVERGENCE 74

Hence,
Z b
?
Φ(x)u0 (x) − f (x, u0 (x) dx


F (φ) =
a
Z b
= f ? (x, Φ(x) dx
a
Z b  Z x 
?
= f x, − φ(t) dt dx
a a

Now, for any f satisfying H1 and H2, we define

Z b
fε (x, ξ) = ρε (x − y)f (y, ξ) dy,
a

where ρε are the sequence of mollifiers. Observe that fε are convex in the
second variable and, by Jensen’s inequality, fε ≥ f . Also, observe that
limε fε? (x, ξ ? ) = f ? (x, ξ ? ) for all x ∈ (a, b) and ξ ? ∈ R. We have, for each ε,
Z b  Z x 
? ?
Fε (φ) = fε x, − φ(t) dt dx ∀φ ∈ L1 (a, b).
a a

Now, by dominated convergence theorem and F ? ≥ Fε? , we get

Z b  Z x 
? ? ?
F (φ) ≥ lim Fε (φ) = f x, − φ(t) dt dx.
k a a

Also, by the convex conjugate definition, f ? (x, ξ ? ) ≥ ξ ? ξ − f (x, ξ) for all

x, ξ, ξ ? . Now, choose ξ ? = Φ(x), ξ = v 0 , where v ∈ X and integrate both
sides of above inequality,
Z b Z b
?
f (x, Φ(x)) dx ≥ (Φ(x)v 0 (x) − f (x, v 0 (x))) dx
a a
Z b
= (φ(x)v(x) − f (x, v 0 (x))) dx.
a
Rb
Taking supremum over v ∈ V , we have F ? (φ) ≤ a
f ? (x, Φ(x)) dx.
Proposition 4.4.2. Let gn : Ω × Rn → [0, +∞) satisfy hypotheses H1 and
H2, for all n. If gn (·, ξ) weak* converges to g(·, ξ) for all ξ ∈ R, then
gn (·, v(·)) weak* converges to g(·, v(·)), for all v ∈ C([a, b]).
CHAPTER 4. Γ-CONVERGENCE 75

Proof. Let v ∈ C([a, b]) and φ ∈ L1 (a, b). Also, let (xi−1 , xi ) be k number of
partitions of (a, b) for i = 1, 2, . . . , k such that x0 = a and xk = b. Consider,
Z b k
X Z xi
(gn (x, v) − g(x, v)) φ dx ≤ [gn (x, v) − gn (x, v(xi ))] φ dx
a i=1 xi−1
k
X Z xi
+ [gn (x, v(xi )) − g(x, v(xi ))] φ dx
i=1 xi−1

Xk Z xi
+ [g(x, v(xi )) − g(x, v(x))] φ dx
i=1 xi−1

The second term converges to zero, by hypothesis, and by uniform local

Lipschitz continuity (cf. Exercise 4.10 of gn and g, we have the result.

Lemma 4.4.3. Let gn : Ω×Rn → [0, +∞) satisfy hypotheses H1 and H2, for
all n. Then, there exists a subsequence of {gn } and a g : (a, b)×R → [0, +∞)
such that gn (·, ξ) weak* converges to g(·, ξ) for all ξ ∈ R.

Theorem 4.4.4. Let {Fn } be a sequence in F with integrand fn and F ∈ F

with integrand f . Then the following statements are equivalent:

1. Fn (·, I) Γ-converges to F (·, I) in W 1,p (I), for all open intervals I of

(a, b).

2. fn? (·, ξ ? ) weak* converges to f ? (·, ξ ? ), for all ξ ? ∈ R.

The proof of above lemma and theorem are being skipped and can be
found in [Bra02].
Example 4.2. Let 0 < α ≤ aε (x) ≤ β < +∞ and g ∈ L2 (a, b). Let Fε :
H01 (a, b) → R be defined as
Z b 
1 0 2
Fε (u) = aε (x)|u | − gu dx.
a 2

The Euler-Lagrange equations yields that the minimizers uε ,

(
d
aε (x) du


− dx dx
ε
= g in (a, b)
uε (a) = uε (b) = 0.
CHAPTER 4. Γ-CONVERGENCE 76

ξ2
Now, set fε (x, ξ) := aε (x)|ξ|2 . Then, fε? (x, ξ ? ) = 4aε (x)
. But, for each ξ ? ∈
ξ2
Rn , fε? (·, ξ ? ) converges weak* in L∞ (a, b) to f ? (·, ξ ), where f ? (x, ξ ? ) =
?
4b(x)
and
1 1
* .
aε (x) b(x)
Chapter 5

Bloch-Floquet Homogenization

5.1 Fourier Transform

In this chapter we assume the functions to be complex valued. Recall that
−∆ : H 2 (Rn ) ⊂ L2 (Rn ) → L2 (Rn ) is an unbounded, self-adjoint operator
whose spectral decomposition is well-known. The “generalised” eigenfunc-
tions1 are the plane waves or Fourier waves eıξ·x , for each ξ ∈ Rn , and |ξ|2
is an eigenvalue for each ξ ∈ Rn giving the spectrum to be [0, ∞). In other
words, −∆(eıx·ξ ) = |ξ|2 eıx·ξ .
Theorem 5.1.1. Given any f ∈ L2 (Rn ) there is a unique fˆ ∈ L2 (Rn ) such
that Z
1
f (x) = fˆ(ξ)eıξ·x dξ.
(2π)n/2 Rn
Also, for any f, g ∈ L2 (Rn ),
Z Z
f (x)g(x) dx = fˆ(ξ)ĝ(ξ) dξ.
Rn Rn

In particular, the Fourier transform f 7→ fˆ is an isometry from L2 (Rn ) to

L2 (Rn ).
For any f ∈ L1 (Rn ), its Fourier transform fˆ : Rn → C is given as
Z
ˆ 1
f (ξ) = f (x)e−ıξ·x dx.
(2π)n/2 Rn
1
For each ξ ∈ Rn , eıξ·x are not elements of L2 (Rn ) but they span L2 (Rn )

77
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 78

The Fourier transform map F : L1 (Rn ) → L∞ (Rn ) is defined as F(f ) = fˆ.

Note that F is a bounded linear with kFk ≤ 1, since kfˆk∞ ≤ kf k1 . The
definition is generalised to Schwartz class. In this sense, the Fourier transform
can be generalised to L2 (Rn ). The Fourier transform will change a differential
equation in to an algebraic equation. For instance, −∆u = f will tranform,
on applying Fourier transform, to
n Z
∂ 2 u(x) −ıx·ξ
Z
1 1 X
fˆ(ξ) = −ıx·ξ
f (x)e dx = − e dx
(2π)n/2 Rn (2π)n/2 j=1 Rn ∂x2j
n Z
1 X ∂u(x) −ıx·ξ
= (−ıξ j ) e dx (Integration by parts)
(2π)n/2 j=1 Rn ∂xj
n Z
X 1
= − (−ıξj )2
n/2
u(x)e−ıx·ξ dx (Integration by parts)
j=1
(2π) Rn

= |ξ|2 û(ξ).

More generally, any m-th order linear differential equation with constant
α
P
coefficients P (D)u = f where P (D) = |α|≤m aα D will transform in to
an algebraic eqaution P (ıξ)û(ξ) = fˆ(ξ). The Laplacian is a particular case
of the elliptic operator −∆ + c(x) with c ≡ 0. For c(x) 6= 0 (without loss
of generality assume c(x) ≥ 0), the Bloch theorem gives the generalised
eigenfunction for −∆ + c(x) when c is Y -periodic, for any given reference cell
Y ⊂ Rn .

5.2 Schrödinger Operator with Periodic Po-

tential
Definition 5.2.1. Let {ei } be the canonical basis for Rn . Let Y = Πni=1 [0, `i )
be a reference cell (or period) in Rn . A function f : Rn → R is said to be
Y -periodic if f (x + ei pi `i ) = f (x) for a.e. x ∈ Rn and all p ∈ Zn , for all
i = 1, 2, . . . , n.

Consider the Schrödinger operator −∆ + c(x) where c is a periodic func-

tion, i.e., for some ` = (`i ) ∈ Rn and p ∈ Zn , c(x + ei `i pi ) = c(x). Let
L : S(R) → S(R) be the operator L := −∆ + c(x). The operator L has large
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 79

symmetry group. Define, for each p ∈ Zn ,

n
X
[U (p)v](x) := v(x + pi `i ).
i=1

Then U (p)L = LU (p). In fact, U (p)e−ıLs = e−ıLs U (p). Let us first consider
the one dimension situation with c ∈ Cc∞ (R) with bounded derivatives and
L : S(R) → S(R) defined as
d2
L := − + c(x).
dx2
If c is 2π-periodic and, hence, c admits a uniformly convergent Fourier series
X
c(x) = cm eımx
m∈Z

where Z π
1
cm = c(x)e−ımx dx.
2π −π

If u ∈ S(R) then
Z
1
\
Lu(x)(ξ) = ξ 2 û(ξ) + √ c(x)u(x)e−ıξx dx
2π R
Z !
1 X
= ξ 2 û(ξ) + √ cm eımx u(x)e−ıξx dx
2π R m∈Z
Z
X 1
2
= ξ û(ξ) + cm √ u(x)e−ı(ξ−m)x dx
m∈Z
2π R
X
= ξ 2 û(ξ) + cm û(ξ − m).
m∈Z

Thus, Lu(ξ)
c depends only on the values û(ξ − m) for all m ∈ Z. But recall
that û(ξ − m) = eıxm
\ u(x)(ξ). This suggests that the operator L depends on
the “modulation” by all m ∈ Z.

5.2.1 Direct Integral Decomposition

Let H be a separable Hilbert space and (X, µ) be a σ-finite measure space.
Let L2 (X, µ; H) is the Hilbert space of square integrable H-valued functions.
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 80

If µ is a sum of point measures at finite set of points x1 , . . . , xk then, any

f ∈ L2 (X, µ; H), is determined by the k-tuple (f (x1 ), . . . , f (xk )). Thus,
L2 (X, µ; H) is isomorphic to the direct sum ⊕ki=1 H. For more general µ, one
may define a kind of “continuous direct sum” called the constant fiber direct
integral and write Z ⊕
L2 (X, µ; H) = H dµ.
X

Definition 5.2.2. A function T (·) : X → L(H) is measurable iff, for each

φ, ψ ∈ H, hφ, T (·)ψi is measurable. L∞ (X, µ; L(H)) denotes the equivalence
class (with a.e.) of measurable functions from X to L(H) with
kT k∞ = ess supx∈X kT (x)kL(H) < ∞.
R⊕
Definition 5.2.3. A bounded operator T on H = X H dµ is said to be de-
composed by the direct integral decomposition iff there is T (·) ∈ L∞ (X, µ; L(H))
such that, for all ψ ∈ H,
(T ψ)(x) = T (x)ψ(x).
We then say T is decomposable and
Z ⊕
T = T (x) dµ(x).
X

The T (x) are called the fibers of T .

Theorem 5.2.4. Let H = l2 and
Z ⊕
H= H dx.
(− 21 , 12 ]

For η ∈ − 12 , 21 , let Lη : l2 → l2 be defined as


X
(Lη (z))k = (η + k)2 zk + cm zk−m .
m∈Z

Define T : L2 (R) → H by
[(T f )(η)]k = fˆ(η + k).
d 2
2
For L = − dx 2 + c(x) on L (R),

Z ⊕
−1
T LT = Lη dη.
(− 12 , 21 ]
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 81

When c ≡ 0, the eigenvalues and eigenfunctions of Lη are (η +k)2 and the

Fourier transform of eı(η+k)x , respectively. This suggests that Lη is related to
d2
− dx 2 on [0, 2π) with the boundary condition u(2π) = e
ı2πη
u(0) and u0 (2π) =
eı2πη u0 (0).
Lemma 5.2.5. Let H = L2 [0, 2π) and
Z ⊕
H= H dη.
(− 21 , 12 ]

Then T : S(R) → H given by

X 1 1
(T f )η (x) = eı2πmη f (x + 2πm) η ∈ (− , ] x ∈ [0, 2π)
m∈Z
2 2

which extends uniquely to an unitary operator on L2 (R). Moreover,

Z ⊕ 
d2 d2
  
−1
T − 2 T = − 2 dη (5.2.1)
dx (− 12 , 21 ] dx η
 2
d d2 2
where − dx 2 is the operator − dx 2 on L [0, 2π) with boundary condition
η

u(2π) = eı2πη u(0) u0 (2π) = eı2πη u0 (0).

Proof. Let us note that T is well defined. For any f ∈ S(R), the series in
RHS is convergent. For any f ∈ S(R), T f ∈ S(R) because
 
Z 1 Z 2π X ∞ 2
2
 e−ı2πmη f (x + 2πm) dx dη
− 12 0 m=−∞
" !Z 1
#
Z 2π X 2
= f (x + 2πm)f (x + 2πp) e−ı2π(p−m)η dη dx
0 m,p∈Z − 12

( by Fubini’s Theorem)
Z 2π X ! Z
2
= |f (x + 2πm)| dx = |f (x)|2 dx.
0 m∈Z R

Thus, T is well defined and admits a unique isometry extension. To see that
T is onto H, we compute T ? . For any g ∈ H, x ∈ [0, 2π] and m ∈ Z
Z 1
2
?
(T g)(x + 2πm) = eı2πmη gη (x) dη.
− 12
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 82

Further,
Z
?
kT gk22 = |(T ? g)(y)|2 dy
R
!
Z2π X
= |(T ? g)(2πm + x)|2 dx
0 m∈Z
2
!
Z 2π X Z 2π
= eı2πmη gη (x) dθ dx
0 m∈Z 0
Z 2π Z 2π 
2
= |gη (x)| dθ dx (Plancherel’s identity)
0 0
= kgk2 .

Finally, to prove (5.2.1), let G be the operator on the right-hand side of

(5.2.1). We shall show that if f ∈ S(R), then T f ∈ D(G) and T (−f 00 ) =
G(T f ). Since −d2 /dx2 is essentially self-adjoint on S(R) and G is self-adjoint,
(5.2.1) will follow. So, suppose f ∈ S(Rn ), then T f is given by the convergent
sum as in the statement. Thus, T f ∈ C ∞ (0, 2π) with (T f )0η (x) = (T fη0 (x)
and similarly for higher derivatives. Further, it is clear that
X
(T f )θ (2π) = e−ı2πmη f (2π(m + 1))
m∈Z
X
= e−ı2π(m−1)η f (2πm) = eı2πη (T f )η (0).
m∈Z

2
Similarly, (T f )0η (2π) = eı2πη (T fη )0 (0). Thus, for each η, (T f )η ∈ D((− dx
d
2 )η )

and
d2
 
− 2 (T f ) = U (−f 00 )η .
dx η

We conclude that T f ∈ D(G) and G(T f ) = U (−f 00 ). This proves (5.2.1).

Theorem 5.2.6 (Direct Integral Decomposition of Periodic Schrödinger op-

erator). Let
 c be a bounded measurable function on R with period 2π. For
1 1
η ∈ − 2 , 2 , let
d2
 
Lη = − 2 + c(x)
dx η
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 83

be an operator on L2 [0, 2π]. Let T be given by

 
X
ı2πmη 1 1
(T f )η (x) = e f (x + 2πm) η ∈ − , and x ∈ [0, 2π).
m∈Z
2 2

Then Z ⊕
d2
 
−1
T − 2 +c T = Lη dη.
dx (− 12 , 21 ]

Proof. Let c be the η-independent operator acting on the fiber H = L2 [0, 2π)
by (cη f )(x) = c(x)f (x) for 0 ≤ x ≤ 2π. It is sufficient to prove that
Z ⊕
−1
T cT = cη dη.
(− 12 , 12 ]

For f ∈ S(R),
X
(T cf )η (x) = e−ı2πmη c(x + 2πm)f (x + 2πm)
m∈Z
X
= c(x) e−ı2πmη f (x + 2πm)
m∈Z
= cη (T f )η (x).

The second last equality is due to the periodicity of c.

5.3 Bloch Periodic Functions

The Bloch transform is a generalization of Fourier transform that leaves the
periodic functions invariant, in some sense. Let us begin by considering a
generalization of periodic functions.

Definition 5.3.1. Let Y = Πni=1 [0, `i ) be a reference cell (or period) in Rn .

For each η ∈ Rn , a function f : Rn → R is said to be (η, Y )-Bloch periodic
if f (x + ` · p) = eı2πp·η f (x) for a.e. x ∈ Rn and for all p ∈ Zn .

Note that the case η = 0 corresponds to the usual notion of Y -periodic

functions. Note that the boundary condition remains unchanged if η is re-
placed with η + k, for any k ∈ Zn . Hence, it is sufficient to consider η ∈ Y ?
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 84

where Y ? = (− 21 , 12 ]n . The cell Y ? is called the reciprocal cell and, in Physics

literature, Y ? is known as the first Brillouin zone.
We shall assume that Y = [0, 2π)n and, for j, k = 1, 2, . . . , n, ajk : Y → R
is such that ajk ∈ L∞ per (Y ). Let A(y) = (ajk (y)) ∈ M (α, β, Y ) and is a
symmetric matrix, i.e., ajk (y) = akj (y). One can extend ajk to entire Rn as a
Y -periodic function. Also, c is a Y -periodic function such that c(y) ≥ c3 > 0.
We are interested in the spectral resolution of closure of the operator A =
−div(A(y)∇) + c(y) in L2 (Rn ).
By Bloch Theorem, it is enough to study the (η, Y )-Bloch periodic eigen-
value problem, for each η ∈ Rn , i.e.,
Definition 5.3.2. For any fixed (momentum) vector η ∈ Y ? , consider the
eigenvalue problem: given a symmetric A ∈ M (α, β, Y ), find λ(η) ∈ C and
non-zero ψ(·; η) : Rn → R such that

Aψ(y; η)) = λ(η)ψ(y; η) in Rn
(5.3.1)
ψ(y + 2π`) = e2πı`·η ψ(y) ` ∈ Zn , y ∈ Rn .
The eigenvalues ψ are known as Bloch waves associated with A and the
eigenvalues λ are called Bloch eigenvalues.
Suppose η ∈ Y ? have rational components and η = (η1 , . . . , ηn ). Recall
that there is a homeomorphism from Y ? to S 1 . Thus, eı2πηj ∈ S 1 . In this
sense, the Bloch periodicity condition has the form e2πıp·η = ω p where ω ∈
[S 1 ]n and ω p = ω1p1 ω2p2 . . . ωnpn . For any m ∈ Zn , let Dm ⊂ [S 1 ]n be the
collection of all ω ∈ [S 1 ]n such that its j-th component is the mj -th root of
unity. Thus, ω m = 1 for all ω ∈ Dm . The spectral problem (5.3.1) may be
seen as a sequence of spectral problems, i.e., for each m ∈ Zn , we define ψm
as 
Aψm (y) = λm ψm (y) in Rn
ψm (y + 2πm) = ψ(y) y ∈ Rn .
Note
Qn that in the above boundary condition ψ is Ym -periodic where Ym =
2
i=1 [0, 2πmi ). The space of spectral decomposition is L (Y
per m ) which admits
2 2
the orthogonal decomposition Lper (Ym ) = ⊕ω∈Dm Lper (ω, Y ) where

L2per (ω, Y ) = {ψ ∈ L2loc (Rn ) | ψ(y + 2π`) = ω ` ψ(y) ∀` ∈ Zn , y ∈ Rn }.

Thus, we observe that the above space consists of (η, Y )-Bloch Periodic func-
tions. For any irrational η can be approximated by rationals by varying m
and noting that the sets of roots of unity is dense in S 1 .
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 85

5.4 Bloch Transform

n
Theorem 5.4.1 (Bloch Decomposition). Let Y = [0, 2π)n and Y ? = − 21 , 12 .
Given a f ∈ L2 (Rn ) there is a unique function, called Bloch Transform,
fb ∈ L2 (Y × Y ? ) such that
Z
f (y) = fb (y, η)eıη·y dη.
Y?

Also, for any f, g ∈ L2 (Rn ), the Parseval formula holds, i.e.,

Z Z Z
f (y)g(y) dy = fb (y, η)gb (y, η) dy dη.
Rn Y Y?

In particular, the Bloch transform f 7→ fb is an isometry from L2 (Rn ) to

L2 (Y × Y ? ).
Proof. For any f ∈ D(Rn ) and for each η ∈ Y ? , define
X
fb (y; η) := f (y + 2πp)e−ı(y+2πp)·η .
p∈Zn

The sum is well defined because it has finite number of terms because f has
compact support. Note that fb (y; η) is Y -periodic in y variable because
X
fb (y + 2π; η) := f (y + 2πp)e−ı(y+2πp)·η = fb (y; η).
p+1∈Zn

Similarly, eıy·η fb (y; η) is Y ? -periodic in η variable because, for k ∈ Zn ,

X
eıy·(η+k) fb (y; η + k) = eıy·(η+k) f (y + 2πp)e−ı(y+2πp)·(η+k)
p∈Zn
X
= eıy·(η+k) f (y + 2πp)e−ı(y+2πp)·η e−ı(y+2πp)·k
p∈Zn
X
= eıy·(η+k) e−ıy·k f (y + 2πp)e−ı(y+2πp)·η e−ı2πp·k
p∈Zn
= eıy·η fb (y; η).

In the above relation we have used the fact that eı2πp·k = 1. Observe that
X
eıy·η fb (y; η) = f (y + 2πp)e−2ıπp·η .
p∈Zn
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 86

Thus,
Z X Z
ıy·η
e fb (y; η) dη = f (y) + f (y + 2πp) e−2ıπp·η dy
Y ? Y ?
p∈Zn
p6=0

e−ıπp − eıπp
X  
= f (y) − f (y + 2πp) dy
p∈Zn
2ıπp1 . . . pn
p6=0
= f (y).

Therefore, we have proved the results for all functions in D(Rn ). Similarly,
one can prove the Parseval’s formula for functions in D(Rn ). The Bloch
transform is a linear map on D(Rn ) bounded on L2 (Rn ). Define B : D(Rn ) →
L2 (Y × Y ? ) as Bf = fb . B is a bounded operator w.r.t L2 -norm. Consider
Z Z X
2
kBf k2 = kfb k2 ≤ 2
|f (y + 2πp)e−ı(y+2πp)·η |2 dη dy
Y Y ? p∈Zn
Z X Z 
2 −ı(y+2πp)·η 2
= |f (y + 2πp)| |e | dη dy
Y p∈Zn Y?
XZ
?
= |Y | |f (y + 2πp)|2 dy
p∈Zn Y
Z
= |f (y)|2 dy = kf k22 .
Rn

We can unitarily extend B to all of L2 (Rn ). Thus, by density of D(Rn ) in

L2 (Rn ), the Bloch transform extends to L2 (Rn ) and Parseval’s formula holds
true.

Remark 5.4.2. Note that, for each fixed η ∈ Y ? , y 7→ fb (y, η) is extended

Y -periodic to Rn and, for each fixed y ∈ Y , η 7→ eıη·y fb (y, η) is extended
Y ? -periodic to Rn . Thus, the Bloch transform may be seen as an isometry
from L2 (Rn ) to L2 (Rn × Rn ).

Remark 5.4.3. The Bloch transform is a “modulation” of Zak transform.

The Zak transform for any f ∈ D(Rn ) is defined as
X
fz (y; η) := f (y + 2πp)e−ı2πp·η
p∈Zn
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 87

and extended unitarily to to L2 (Rn ). Further, fb (y; η) = e−ıy·η fz (y; η) for all
f ∈ D(Rn ).
The following theorem explains the sense in which the Bloch transform
leaves the periodic functions invariant.
Theorem 5.4.4 (Invariance of Periodic Functions). Let Y = [0, 2π)n and
c : Y → C be such that c ∈ L∞ (Y ) extended Y -periodically to Rn . For any
f ∈ L2 (Rn ), (cf )b (y; η) = c(y)fb (y; η).
Proof. It is enough to prove the result for f ∈ D(Rn ). Consider
X
(cf )b (y; η) = c(y + 2πp)f (y + 2πp)e−ı(y+2πp)·η
p∈Zn
X
= c(y) f (y + 2πp)e−ı(y+2πp)·η
p∈Zn
= c(y)fb (y; η).

By density the result is true for any f ∈ L2 (Rn ).

Theorem 5.4.5. For any f ∈ H 1 (Rn ), (∇y f )b (y; η) = (∇y + ıη)fb (y; η).
Proof. It is enough to prove the result for f ∈ D(Rn ). Consider
X
(∇y f )b (y; η) = [∇y f (y + 2πp)] e−ı(y+2πp)·η
p∈Zn
X
∇y f (y + 2πp)e−ı(y+2πp)·η
 
=
p∈Zn
X
+ıη f (y + 2πp)e−ı(y+2πp)·η
p∈Zn
= [∇y + ıη] fb (y; η).

For any f ∈ L2 (Rn ), consider the equation Au = f in Rn . Applying

Bloch transform to this equation, using Theorems 5.4.4 and 5.4.5, we obtain
a family of equations, indexed by η ∈ Y ? , with periodic boundary conditions:

A(η)ub (y; η) = fb (y; η) in Rn
(5.4.1)
ub (y + 2π`; η) = ub (y; η) ` ∈ Zn y ∈ Rn ,
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 88

where A(η) is the shifted operator, denoted as

n    
X ∂ ∂
A(η) := − + ıηj ajk (y) + ıηk + c(y).
j,k=1
∂y j ∂y k

The shifted operator equation admits a solution (being a periodic problem)

1
in Hper (Y ) and a corresponding Poincaré inequality holds true, i.e., for all
u ∈ Hper (Y ) and η ∈ Y ? ,
1

c (k∇uk2,Y + |η|kuk2,Y ) ≤ k∇u + ıuηk2,Y ≤ k∇uk2,Y + |η|kuk2,Y .

5.4.1 Spectrum of Elliptic Operator

The spectral decomposition of A, in one dimension periodic media, was first
studied by Floquet (cf. [Flo83]) and much later, in crystal lattice, by Bloch.
We shall compute the spectral decomposition of A in L2 (Rn ) via the spectral
decomposition of the shifted operator A(η). Consider the eigenvalue problem

A(η)φ(y; η) = λ(η)φ(y; η) in Rn
(5.4.2)
φ(y + 2π`) = φ(y) ` ∈ Zn , y ∈ Rn

Theorem 5.4.6 (Periodic Eigen Value problem). There exists a sequence of

pairs (λm , φm ) satisfying

Aφ(y) = λφ(y) in Rn
(5.4.3)
φ(y + 2π`) = φ(y) ` ∈ Zn , y ∈ Rn

where {λm } are positive real eigenvalues and {φm (y)} are the corresponding
eigenvectors, for each m ∈ N, such that {φm } form an orthonormal basis
of L2per (Y ) and 0 ≤ λ1 ≤ λ2 ≤ . . . diverges and each eigenvalue has finite
multiplicity.

Remark 5.4.7. By Theorem 5.4.6, for each fixed η ∈ Y ? , there exists

a sequence of pairs (λm , φm ) satisfying (5.4.2) where {λm (η)} are positive
real eigenvalues and {φm (y; η)} are the corresponding eigenvectors, for each
m ∈ N, such that {φm (·; η)} form an orthonormal basis of L2per (Y ) and
0 ≤ λ1 (η) ≤ λ2 (η) ≤ . . . diverges and each eigenvalue has finite multiplicity.
By varying η ∈ Y ? , we obtain the spectral resolution of A in L2 (Rn ). The
set {eıy·η φm (y, η); m ∈ N, η ∈ Y ? }forms a ‘generalised’ basis of L2 (Rn ). As a
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 89

consequence, L2 (Rn ) can be identified with L2 (Y ? ; `2 (N)). A acts as a mul-

tiplication operator: A[eıy·η φm (y, η)] = λm (η)eıy·η φm (y, η). The spectrum of
A, denoted as σ(A), coincides with the Bloch spectrum and denoted as σb .
The Bloch spectrum is defined as the union of the images of all the mappings
λm (η), i.e.,  
σb := ∪∞
m=1 inf λm (η), sup λm (η) .
η∈Y ? η∈Y ?

The spectrum has a band structure. In contrast to the homogeneous case,

σ(A) need not fill up the entire [0, ∞) and there may be gaps.
Theorem 5.4.8. For any f ∈ L2 (Rn ), its Bloch transform is given as
∞
X
fb (y; η) = fbm (η)φm (y; η)
m=1

where, {φm } are the eigenfunctions corresponding to the shifted operator A(η)
and fbm (η), for each η ∈ Y ? , is the m-th Bloch coefficient of f defined as
Z
m
fb (η) := f (y)e−ıy·η φm (y; η) dy.
Rn

Proof. It is enough to prove the result for f ∈ D(Rn ). Recall that, for each
η ∈ Y ? , fb (·; η) ∈ L2per (Y ). Hence, by spectral decomposition of A(η),
∞
X
fb (y; η) = fbm (η)φm (y; η),
m=1

where Z
fbm (η) = fb (y; η)φm (y; η) dy.
Y
But,
Z X
fbm (η) = f (y + 2πp)e−ı(y+2πp)·η φm (y; η) dy
Y p∈Zn
Z X
= f (y + 2πp)e−ı(y+2πp)·η φm (y + 2πp); η) dy
Y p∈Zn
Z
= f (y)e−ıy·η φm (y; η) dy.
Rn
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 90

Remark 5.4.9. The Bloch inversion formula can rewritten as:

Z Z ∞
X
ıy·η
f (y) = e fb (y; η) dη = eıy·η fbm (η)φm (y; η) dη.
Y? Y? m=1

Further, the Plancherel formula holds:

Z Z ∞
X
2
|f (y)| dy = |fbm (η)|2 dη. (5.4.4)
Rn Y ?
m=1

Remark 5.4.10 (Algebraic Formula for Solution). For each m ∈ N and

η ∈ Y ? , multiply φm (y; η) on both sides of (5.4.1) to obtain
Z "∞ # Z X ∞
X
k
A(η) ub (η)φk (y; η) φm (y; η) dy = fbk (η)φk (y; η)φm (y; η) dy
Y k=1 Y k=1
∞
Z X
ukb (η)φk (y; η)λm (η)φm (y; η) dy = fbm (η)
Y k=1

um m
b (η)λm (η) = fb (η)
fbm (η)
um
b (η) = .
λm (η)

Set ψm (y; η) := {eıy·η φm (y; η)}. Then, for each η ∈ Y ? , ψm (·; η) forms a
basis of L2 (Rn ). Thus, L2 (Rn ) can be identified with L2 (Y ? ; `2 (N)). Let us
compute ψ(y + 2π`):

ψm (y + 2π`) = eıy·η e2πı`·η φm (y + 2π`)

= eıy·η e2πı`·η φm (y)
= e2πı`·η ψm (y).

5.4.2 Regularity of λm (η) and φ1 (·, η)

Theorem 5.4.11. For all m ≥ 1, η 7→ λm (η) is a Lipschitz function.

Proof. Consider the quadratic form associated with A(η):

Z   
∂v ∂v
a(v, v; η) = ajk (y) + ıηk v + ıηj v dy.
Y ∂yk ∂yj
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 91

The quadratic form admits a decomposition as follows:

a(v, v; η) = a(v, v; η 0 ) + R(v, v; η, η 0 )

where
Z Z
∂v ∂v
R= ajk (y) (ıηj − ıηj0 )v dy + ajk (y)(ıηk − ıηk0 )v dy
Y ∂yk ∂yj
ZY
+ ajk (y)(ηk ηj − ηk0 ηj0 )|v|2 dy.
Y

By Cauchy-Schwarz’s inequality,
Z
0
|R| ≤ C0 |η − η | (|∇v|2 + |v|2 ) dy.
Y

By min-max principle,

a(v, v; η)
λm (η) = min max
1
W ⊂Hper (Y ) v∈W kvk22,Y

1
where W is a m-dimensional subspace of Hper (Y ). Using the estimate on R,
we deduce that
λm (η) ≤ λm (η 0 ) + C0 |η − η 0 |
for a suitable constant C0 . Interchanging η and η 0 , we obtain

|λm (η) − λm (η 0 )| ≤ C0 |η − η 0 |.

Theorem 5.4.12 (Analyticity). There is a δ > 0 such that λ1 (η) is analytic

in the open ball Bδ (0) centred at origin and radius δ. Further, one can choose
a corresponding unit eigenvector φ1 (y; η) satisfying

(i) η 7→ φ1 (·; η) from Y ? to Hper

1
(Y ) is analytic on Bδ (0).

(ii) φ1 (y; 0) := |Y |−1/2 := (2π)−n/2 .

R
(iii) kφ1 (·; η)k2,Y = 1 and Y φ1 (y; η) dy = 0 for each η ∈ Bδ .
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 92

5.4.3 Taylor Expansion of Ground State

Observe that (5.4.1) is a polynomial of degree two w.r.t η variable. Let
Tm (η) : L2 (Y ) → L2 (Y ) be defined as
Tm (η)(φ) = A(η)φ − λm φ.
For a fixed m ∈ N, let us compute the j-th first partial derivative of (5.4.2)
w.r.t η to get
∂φm ∂A(η) ∂φm ∂λm
A(η) + φm = λm + φm .
∂ηj ∂ηj ∂ηj ∂ηj
Thus,
∂φm ∂A(η) ∂λm
Tm (η) = − φm + φm
∂ηj ∂ηj ∂ηj
∂λm
= ıej A(∇y + ıη)φm + (∇y + ıη) · (ıAej φm ) + φm .
∂ηj
There exists a solution to the above equation which is unique upto an additive
multiple of φm . Hence, the RHS satisfies the compatibility condition or
Fredhölm alternative. Therefore,
Z
∂φm
Tm (η) φ dy = 0
Y ∂ηj m
yields a formula for ∇η λm (η m ) in terms of φm . Thus,
 
∂λm ∂A(η)
(η) = φm (·; η), φm (·; η) .
∂ηj ∂ηj
Similarly, by computing the j-th second partial derivative of (5.4.2) w.r.t η,
we get
∂ 2 φm
 
∂φm ∂φm
Tm (η) = ıej A(∇y + ıη) + (∇y + ıη) · ıAej
∂ηj ∂ηk ∂ηk ∂ηk
 
∂φm ∂φm
+ıek A(∇y + ıη) + (∇y + ıη) · ıAek
∂ηj ∂ηj
∂λm ∂λm ∂λm ∂λm
+ + − ej Aek φm − ek Aej φm
∂ηj ∂ηk ∂ηk ∂ηj
∂ 2 λm
+ φm .
∂ηk ∂ηj
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 93

There exists a solution to the above equation which is unique upto an additive
multiple of φm . Hence, the RHS satisfies the compatibility condition or
Fredhölm alternative. Therefore,
∂ 2 φm
Z
Tm (η) φ dy = 0
Y ∂ηj ∂ηk m
yields a formula for the Hessian matrix Dη2 λm (η m ) in terms of φm . Thus,

1 ∂ 2 λm
  
1 ∂A(η) ∂λm ∂φm
(η) = hajk φm , φm i + − , φm
2 ∂ηj ∂ηk 2 ∂ηj ∂ηj ∂ηk
  
1 ∂A(η) ∂λm ∂φm
+ − , φm .
2 ∂ηk ∂ηk ∂ηj
Let us summarise the properties of the eigenvalues λm (η) and eigenvectors
φm (y; η).

(a) All odd order derivatives of λ1 (η) at η = 0 vanish.

(b) All odd order derivatives of φ1 (·, η) at η = 0 are purely imaginary. For
instance, the first order derivatives at η = 0 are given by
∂φ1
(y; 0) = ı|Y |−1/2 wj (y),
∂ηj
1
where wj ∈ Hper (Y ) is the unique solution of the cell problem
(
∂a
Awj = nk=1 ∂yjkk in Rn ,
P
1
R
|Y | Y
wj (y) dy = 0.

(c) All even order derivatives of φ1 (·; η) at η = 0 are real.

(d) Second order derivatives of λ1 (η) at η = 0 are given by
1 ∂ 2 λ1
(0) = a0jk , ∀j, k = 1, ..., n,
2 ∂ηj ∂ηk
where a0jk are the homogenized coefficients defined by
Z " n
#
1 X ∂wm
ajk + ajm .
|Y | Y m=1
∂y m
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 94

Theorem 5.4.13. The origin is a critical point of the first Bloch eigenvalue,
i.e., ∂λ
∂ηj
1
(0) = 0 for all j = 1, ..., n.Further, the Hessian of λ1 at η = 0 is
given by
1 ∂ 2 λ1
(0) = a0jk ∀j, k = 1, ..., n.
2 ∂ηj ∂ηk
The derivatives of the first Bloch mode can also be calculated and they are as
follows:
∂φ1 1
(y; 0) = ı|Y |− 2 wj (y) ∀j = 1, ..., n.
∂ηj
1
Proof. Use the information λ1 (0) = 0 and φ1 (y; 0) = |Y |− 2 in the Taylor
expansion with η = 0.

5.5 Homogenization of Second order Elliptic

Operator
Let Aε = −divx (A(x/ε)∇x ) be the elliptic opertor with periodically oscillat-
ing coefficients. If ξ corresponds to the Fourier variable corresponding to x
then εξ corresponds to the Fourier variable corresponding to x/ε. Recall that,
for each m ∈ N, {λm (η)} and {eıy·η φm (y; η)} are the eigenvalues and eigenvec-
tors, respectively, of A = −divy (A(y)∇y ). We employ the change of variables,
y = x/ε and η = εξ, in the equation A[eıy·η φm (y; η)] = λm (η)eıy·η φm (y; η) to
obtain h  x i x 
ε2 Aε eıx·ξ φm ; εξ = λm (εξ)eıx·ξ φm ; εξ .
ε ε
Thus, the eigenvalues and eigenvectors of Aε are ε−2 λm (εξ) and eıx·ξ φm (x/ε; εξ).
Set λεm (ξ) := ε−2 λm (εξ) and φεm (x; ξ) := φm (x/ε; εξ). Hence, the Bloch trans-
form of f ∈ L2 (Rn ), for each x ∈ Rn and ε > 0, is
∞
X
fbε (x; ξ) = fbm,ε (ξ)φεm (x; ξ)
m=1

where, for each m ∈ N, ε > 0 and ξ ∈ ε−1 Y ? , the m-th Bloch coefficient of f
is Z
m,ε −n/2
fb (ξ) = ε f (x)e−ıx·ξ φεm (x; ξ) dx.
Rn
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 95

Thus, the inverse formula is

Z ∞
X
f (x) = ε n/2
fbm,ε (ξ)eıx·ξ φεm (x; ξ) dξ.
ε−1 Y ? m=1

The εn/2 is a normalising factor appearing because the Lebesgue measure of

ε−1 Y ? is ε−n . The Parseval identity holds: for any f, g ∈ L2 (Rn )
Z Z X∞
ε −n
f (x)g(x) dx = fbm,ε (ξ)gbm,ε (ξ) dξ.
Rn ε−1 Y ? m=1

Applying the Bloch transform, the equation Aε uε = f transforms in to a

set of algebraic equations, indexed by m ≥ 1, λεm (ξ)um,ε m,ε
b (ξ) = fb (ξ) for all
ξ ∈ ε−1 Y ? (cf. Remark 5.4.10). Our aim is to pass to the limit in the system
of algebraic equations. We first claim that one can neglect all the equations
corresponding to m ≥ 2.
Proposition 5.5.1. Let
Z ∞
X
vε (x) = ε n/2
um,ε
b (ξ)e
ıx·ξ ε
φm (x; ξ) dξ.
ε−1 Y ? m=2

Then kvε k2,Rn ≤ C0 ε.

Proof. Since Z Z
Aε uε uε dx = f (x)uε (x) dx.
Rn Rn
The LHS is bounded and, applying Parseval Identity, we get
Z Z ∞
X
β 2
|∇uε | dx ≥ ε n
fbm,ε (ξ)um,ε
b (ξ) dξ
Rn ε−1 Y ? m=1
Z X∞
= εn λεm (ξ)|um,ε 2
b (ξ)| dξ
ε−1 Y ? m=1
Z ∞
X
= ε n−2
λm (η)|um,ε 2
b (ξ)| dξ
ε−1 Y ? m=1
Z X∞
≥ εn−2 λm (η)|um,ε 2
b (ξ)| dξ
ε−1 Y ? m=2
Z ∞
X
(N )
≥ εn−2 λ2 |um,ε 2
b (ξ)| dξ.
ε−1 Y ? m=2
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 96

The last inequality is a consequence of the min-max principle yielding, for

m ≥ 2,
(N )
λm (η) ≥ λ2 (η) ≥ λ2 > 0 ∀η ∈ Y ? ,
(N )
where λ2 is the second eigenvalue of the eigenvalue problem for A in the
cell Y with Neumann boundary condition on ∂Y . Then
Z ∞
X
ε n
|um,ε 2 2
b (ξ)| dξ ≤ C0 ε .
ε−1 Y ? m=2

By Plancherel’s identity, the left side is equal to kvε k2,Rn .

Remark 5.5.2. Consider the algebraic equation corresponding to m = 1,

i.e.,
λε1 (ξ)u1,ε 1,ε −1 ?
b (ξ) = fb (ξ) ∀ξ ∈ ε Y .

Multiplying both sides by εn/2 , we get

ε−2 λ1 (εξ)εn/2 u1,ε

b (ξ) = ε
n/2 1,ε
fb (ξ) ∀ξ ∈ ε−1 Y ? .

Expanding λ1 (εξ) by Taylor’s formula around ξ = 0, we get

" n
#
1 X ∂ 2 λ1
(0)ξj ξk + O(εξ 3 ) εn/2 u1,ε
b (ξ) = ε
n/2 1,ε
fb (ξ)
2 j,k=1 ∂ηj ηk

Passing to the limit as ε → 0 to get

n
1 X ∂ 2 λ1
(0)ξj ξk û0 (ξ) = fˆ(ξ).
2 j,k=1 ∂ηj ηk

Setting
1 ∂ 2 λ1
a0jk = (0)
2 ∂ηj ηk
2u
Then nj,k=1 a0jk ξk ξj uˆ0 (ξ) = fˆ(ξ) and A0 u0 := − nj,k=1 a0jk ∂x∂ j ∂x
P P 0
k
= f (x).
The only flaw in the above argument is that in passing to limit we have
not checked uniform compact support of the sequence. To overcome this
difficulty we use cut-off function technique to localize the equation.
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 97

Proposition 5.5.3 (First Bloch Transform tends to Fourier Transform). Let

{gε } ⊂ L2 (Rn ) be a sequence such that there is a fixed compact set K ⊂ Rn
n
such that supp(gε ) ⊆ K for all ε. If gε * g weakly in L2 (Rn ) then ε 2 gb1,ε * ĝ
weakly in L2loc (Rn ).

Proof. The first Bloch transform gb1,ε (ξ), a priori defined for

ε−1 ε−1 n
ξ ∈ ε−1 Y ? = (− , )
2 2

can be extended by zero outside ε−1 Y ? . We write

Z
n 1,ε x
ε 2 gb (ξ) = gε (x)e−ıx·ξ φ1 ( ; 0) dx
Rn ε
Z  
−ıx·ξ x x
+ gε (x)e φ1 ( ; εξ) − φ1 ( ; 0) dx.
K ε ε
−1
Since φ1 (y; 0) = |Y | 2 = (2π)−n/2 , the first term is nothing but the Fourier
transform of gε and so it converges weakly to ĝ(ξ) in L2 (Rn ). By Cauchy-
Schwarz inequality and the regularity of the first Bloch eigenfunction η 7→
φ1 (·, η) ∈ L2per (Y ) at η = 0, the second term is bounded by

Z  21
x x
kgε k2,Rn |φ1 ( ; εξ) − φ1 ( ; 0)|2 dx ≤ C0 kφ1 (y; εξ) − φ1 (y; 0)k2,Y .
K ε ε

By Lipschitz continuity of η 7→ φ1 (·, η), the second term in the right side is
bounded above by C0 εξ. Thus, if |ξ| ≤ M then it is bounded above by cM ε
and so, in particular, it converges to zero in L∞ n
loc (R ).

Theorem 5.5.4. Let Ω ⊂ Rn be an arbitrary, not necessarily bounded, do-

main. Consider a sequence uε * u0 weakly in H 1 (Ω) and Aε u0 = f in Ω
with f ∈ L2 (Ω). Then u0 satisfies A0 u0 = f in Ω. In fact, Aε ∇uε * A0 ∇u0
weakly in L2 (Ω).

Proof. Let φ ∈ D(Ω) be arbitrary. If uε satisfies Aε uε = f in Ω then consider

its localization φuε satisfies

Aε (φuε ) = φf + gε + hε in Rn ,
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 98

where
n n
X ∂φ X ∂ 2φ
gε = −2 σjε − aεjk uε ,
j=1
∂xj j,k=1 ∂xj ∂xk
n
X ∂uε
σjε (x) = aεjk ,
k=1
∂xk
n
X ∂aεjk ∂φ
hε = − uε .
j,k=1
∂x j ∂xk

Using the arguments given in the remark above, we can pass to the limit
above, since φuε is bounded in H 1 (Rn ). Neglecting all the harmonics cor-
responding to m ≥ 2 and considering only the m = 1 yields at the limit

n
1 X ∂ 2 λ1 [ n 1,ε n 1,ε
(0)ξj ξk (φu0 )(ξ) = (φf )(ξ) + lim ε 2 gb (ξ) + lim ε 2 ĥb (ξ).
d
2 j,k=1 ∂ηj ∂ηk ε→0 ε→0

(5.5.1)
The sequence σjε is bounded in L2 (Ω). Therefore, we can extract a subse-
quence (still denoted by ε) which is weakly convergent in L2 (Ω). Let σj0
denote its limit and its extension by zero outside Ω. Using this convergence
and the definition of gε , we see that
n n
X ∂φ X ∂ 2φ
gε * g0 := −2 σj0 − M(ajk ) u0 weakly in L2 (Rn ),
j=1
∂xj j,k=1 ∂xj ∂xk

where M(ajk ) is the average of ajk on Y . Therefore,

n
ε 2 gb1,ε (ξ) * ĝ0 (ξ) weakly in L2loc (Rn ).

A similar argument fails for {h1,ε 2 n

b } because hε is not bounded in L (R ). We
decompose
Z x 
n 1,ε
ε 2 hb (ξ) = hε (x)e−ıx·ξ φ1 , 0 dx
Rn ε
Z    x 
−ıx·ξ x 
+ hε (x)e φ1 ; εξ − φ1 ;0 dx.
Rn ε ε
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 99

Using the Taylor expansion of φ1 (y; η) at η = 0, the second term is equal to

n Z
" n #
X ∂ajk x ∂φ
  X ∂φ1 x 
−ε−1 (x)uε (x)e−ıx·ξ ε ; 0 ξ` + O(ε2 ξ 2 ) dx,
n ∂yj ε ∂xk ∂η` ε
j,k=1 R `=1

which evidently converges to

n   Z
X ∂ajk ∂φ1 ∂φ
− M (y; 0) ξ` u0 e−ıx·ξ dx.
j,k,`=1
∂yj ∂η` Rn ∂xk

strongly in L∞ n
loc (R ). On the other hand, after integraing by parts, the first
term in the RHS of the decomposition of εn/2 h1,ε b becomes
n Z  2
∂φ ∂uε

X ∂ φ ∂φ x 
ε
ajk uε + − ıξj uε e−ıx·ξ φ1 ; 0 dx.
j,k=1 Rn ∂x j ∂xk ∂x k ∂xj ∂x k ε
1
Choosing φ1 (y; 0) = |Y |− 2 , it is easily seen that the above integral converges
weakly in L2 (Rn ) to
n Z
∂ 2φ
 
− 12
X ∂φ
|Y | M(ajk ) u0 − ıξj M(ajk ) u0 e−ıx·ξ dx
j,k=1 R
n ∂x j ∂x k ∂x k
n Z
− 12
X ∂φ −ıx·ξ
+ |Y | σk0 e dx.
k=1 R
n ∂xk

Using this information in (5.5.1) and using Theorem 5.4.13, we conclude that
n n Z
X
0 − 12
X ∂φ −ıx·ξ
[
ajk ξj ξk (φu0 )(ξ) = (φf )(ξ) − |Y |
d σk0 e dx
j,k=1 k=1 R
n ∂xk
n Z
X
− 12 0 ∂φ
−ı ξj |Y | ajk u0 e−ıx·ξ dx.
j,k=1 Rn ∂x k

This can be rewritten as

n Z
−1
X ∂φ −ıx·ξ
[A\
0 (φu0 )](ξ) = (φf )(ξ) − |Y | 2
d σk0 e dx
k=1 Rn ∂xk
n Z
X 1 ∂φ
−ı ξj |Y |− 2 a0jk u0 e−ıx·ξ dx.
j,k=1 Rn ∂xk
CHAPTER 5. BLOCH-FLOQUET HOMOGENIZATION 100

This is the localized homogenized equation in the Fourier space. Taking in-
verse Fourier transform of the above equation, we obtain
n n  
X ∂φ X
0 ∂ ∂φ
A0 (φu0 ) = φf − σk0 − a u0 in Rn .
k=1
∂xk j,k=1 jk ∂xj ∂xk

On the other hand, we can calculate A0 (φu0 ) directly:

n 
∂ 2φ

0 ∂φ ∂u0
X
A0 (φu0 ) = − a0jk u0 + 2ajk + φA0 u0 in Rn .
j,k=1
∂xj ∂xk ∂xj ∂xk

A comparison of the above two equation yields

n n
!
X X ∂u0 ∂φ
φ(A0 u0 − f ) = a0jk − σj0 in Rn .
j=1 k=1
∂xk ∂xj

Since the above relation is true for all φ in D(Ω), the desired conclusions
follow. In fact, let us choose φ(x) = φ0 (x)eımx·ν , where ν is a unit vector
in Rn and φ0 (x) ∈ D(Ω) is fixed. Letting m → ∞ in the resuling relation
and varying the unit vector ν, we can easily deduce, successively, that σj0 =
Pn 0 ∂u0
k=1 ajk ∂xk in Ω and A0 u0 = f in Ω.
Appendices

101
Bibliography

[All92] G. Allaire, Homogenization and two-scale convergence, SIAM J.

Math. Anal. 23 (1992), no. 6, 1482–1518.

[All94] , Two-scale Convergence: A New Method in Periodic Ho-

mogenization, Nonlinear Partial Differential Equations and Their
Applications, Pitman Research Notes Math Series 302, 1994,
pp. 1–14.

[Att84] H. Attouch, Variational convergence for functions and operators,

Pitman, London, 1984.

[Bab76] Ivo Babuška, Homogenization and its application. mathematical

and computational problems, Numerical solution of Partial Dif-
ferential Equations-III (Proc. of third SYNSPADE, 1975) (Bert
Hubbard, ed.), Academic Press, 1976, pp. 89–116.

[BD98] Andrea Braides and Anneliese Defranceschi, Homogenization of

Multiple Integrals, Oxford lecture series in mathematics and its
applications, no. 12, Oxford University Press, 1998.

[BLP78] A. Bensoussan, J. L. Lions, and G. Papanicolaou, Asymptotic anal-

ysis for periodic structures, North Holland, Amsterdam, 1978.

[BM] J. Ball and F. Murat, W 1,p quasi convexity and variational prob-
lems for multiple integrals, Journal of Functional Analysis 58.

[Bra02] Andrea Braides, Γ-Convergence for Beginners, Oxford Lecture se-

ries in Mathematics and its Applications, vol. 22, Oxford Univer-
sity Press, 2002.

103
BIBLIOGRAPHY 104

[CD99] Doina Cioranescu and Patrizia Donato, An Introduction to Ho-

mogenization, Oxford lecture series in mathematics and its appli-
cations, no. 17, Oxford University Press, 1999.
[Che75] D. Chenais, On the existence of a solution in a domain identifica-
tion problem, J. Math. Anal. Appl. 52 (1975), 189–219.
[Cla79] R. Clausius, Die mech. wa̋rme theorie, Vieweg Bd 2 (1879), 62.
[CP99] Doina Cioranescu and Jeannine Saint Jean Paulin, Homogeniza-
tion of Reticulated Structures, Applied Mathematical Sciences, no.
136, Springer-Verlag, 1999.
[DM93] Gianni Dal Maso, An introduction to Γ-Convergence, Progress in
Nonlinear Differential Equations and Their Applications, vol. 8,
Birkhäuser Boston, 1993.
[ET74] I. Ekeland and R. Temam, Analyse convexe et problèmes variation-
nels, Dunod Gauthier-Villars (English Edition: North-Holland,
1976), 1974.
[Flo83] G. Floquet, Sur les équations différentielles linéaires à coefficients
périodiques, Ann. École Norm. Sér 2 12 (1883), 47–89.
[GF75] Ennio De Giorgi and T. Franzoni, Su un tipo di convergenza vari-
azionale, Atti Acad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur.
58 (1975), 842–850.
[Gio75] Ennio De Giorgi, Sulla convergenza di alcune successioni di inte-
grali del tipo dell’area, Rendi Condi di Mat. 8 (1975), 277–294.
[Gio84] , G-operators and Γ-Convergence, Proceedings of the inter-
national congress of mathematicians (Warsazwa, August 1983),
PWN Polish Scientific Publishers and North Holland, 1984,
pp. 1175–1191.
[GR81] V. Girault and P. A. Raviart, Finite element methods for navier-
stokes equations, Lecture Notes in Mathematics 749 (1981).
[GS73] Ennio De Giorgi and Sergio Spagnolo, Sulla convergenza degli in-
tegrali dell’energia per operatori ellittici del secondo ordine, Boll.
Un. Mat. It. 8 (1973), 391–411.
BIBLIOGRAPHY 105

[Hor97] Ulrich Hornung (ed.), Homogenization and Porous Media, Inter-

disciplinary applied mathematics, no. 6, Springer-Verlag, 1997.

[JKO94] V. V. Jikov, S. M. Kozlov, and O. A. Oleinik, Homogenization of

Differential Operators and Integral Functionals, Springer-Verlag,
1994, Translated by G. A. Yosifian.

[KP97] S. Kesavan and Jeannine Saint Jean Paulin, Homogenization of an

optimal control problem, SIAM J. Control Optim. 35 (1997), no. 5,
1557–1573.

[KR02] S. Kesavan and M. Rajesh, On the limit matrix obtained in the

homogenization of an optimal control problem, Proc. Indian Acad.
Sci. (Math. Sci.) 112 (2002), no. 2, 337–346.

[KS00] David Kinderlehrer and Guido Stampacchia, An introduction to

variational inequalities and their applications, Classics in Applied
Mathematics, vol. 31, SIAM, 2000.

[KV77] S. Kesavan and M. Vanninathan, L’homogénéisation d’un problème

de contrôle optimal, C. R. Acad. Sci. Paris, Série A 285 (1977),
441–444.

[LNW02] D. Lukkassen, G. Nguetseng, and P. Wall, Two-scale convergence,

Int. J. of Pure and Appl. Math. 2 (2002), no. 1, 35–86.

[Max73] C. Maxwell, Treatise on electricity and magnetismus, Oxford Uni-

versity Press 1 (1873), 365.

[Mos50] O. F. Mosotti, Discussione analitica sul influenze che l’azione di

in mezzo dielettrico ha sulla distribuzione dell’ electtricita alla su-
perficie di pin corpi electtici disseminati in esso, Mem. Di Math
et Di Fisica in Modena 24 (1850), no. 2, 49.

[MT97] François Murat and Luc Tartar, H-convergence, Topics in the

Mathematical modelling of Composite materials (Andrej Cherkaev
and Robert Kohn, eds.), Progress in Nonlinear Differential Equa-
tions and Their Applications, vol. 31, Birkhäuser, Boston, 1997,
English translation of [Tar77, Mur78b], pp. 21–43.
BIBLIOGRAPHY 106

[Mur71] François Murat, Un contre-exemple pour le problème du contrôle

dans les coefficients, C. R. Acad. Sci. Paris, Série A 273 (1971),
708–711.

[Mur72] , Théorèmes de non-existence pour des problèmes de

contrôle dans les coefficients, C. R. Acad. Sci. Paris, Série A 274
(1972), 395–398.

[Mur78a] , Compacité par compensation, Ann. Sc. Norm. Sup. Pisa

5 (1978), 489–507.

[Mur78b] , H-convergence, Séminaire d’Analyse Fonctionnelle et

Numérique de l’Université d’Alger, March 1978, mimeographed
notes. English translation in [MT97].

[Mur79] , Compacité par compensation ii, Proceedings of the inter-

national meeting on recent methods in non linear analysis (Rome,
May 1978) (Ennio De Giorgi, E. Magenes, and U. Mosco, eds.),
Pitagora Editrice, Bologna, 1979, pp. 245–256.

[Ngu89] G. Nguetseng, A general convergence result for a functional related

to the theory of homogenization, SIAM J. Math. Anal. 20 (1989),
no. 3, 608–623.

[Poi22] S. D. Poisson, Second mêm. sur la théorie de magnetisme, Mem.

de L’Acad. de France 5 (1822).

[Ray92] J. W. Rayleigh, On the influence of obstacles in rectangular order

upon the properties of a medium, Phil. Mag. 34 (1892), no. 5, 481.

[Spa67] Sergio Spagnolo, Sul limite delle soluzioni di problemi di Cauchy

relativi all’equazione del calore, Ann. Sc. Norm. Sup. Pisa 21
(1967), 657–699.

[Spa68] , Sulla convergenza di soluzioni di equazioni paraboliche ed

ellittiche, Ann. Sc. Norm. Sup. Pisa 22 (1968), 571–597.

[Spa76] , Convergence in energy for elliptic operators, Numeri-

cal solution of Partial Differential Equations-III (Proc. of third
SYNSPADE, 1975) (Bert Hubbard, ed.), Academic Press, 1976,
pp. 469–498.
BIBLIOGRAPHY 107

[Tar77] Luc Tartar, Cours Peccot au Collège de France, partially written

in [Mur78b]. English translation in [MT97], March 1977.

[Tar79] , Compensated compactness and applications to partial dif-

ferential equations, Non linear analysis and mechanics, Heriot-
Watt Symposium, Volume IV (R. J. Knops, ed.), Research Notes
in Mathematics, vol. 39, Pitman, Boston, 1979, pp. 136–212.

[Tem79] R. Temam, Navier-stokes equations, North-Holland, 1979.

[Yos95] K. Yosida, Functional analysis, Springer-Verlag, 1995.

BIBLIOGRAPHY 108
Index

convergence sequence
H-, 43 minimizing, 65
energy, 55 Spagnolo, ii
correctors, 55
Tartar, ii, 54
De Giorgi, iii two-scale
strong convergence, 31
energy two-scale converge, 24
convergence, 55 two-scale convergence, 19, 24
functional, 55
Vanninathan, 59
function
convex, 67
strictly convex, 67

homogenization, i
periodic, 1
reiterated, 26

Kesavan, 59

Lax-Milgram theorem, 3
lemma
div-curl, 55
limit
H-, 44

multi-scale, 26
Murat, ii, 54

Rajesh, 59

Saint Jean Paulin, 59

109