*************************************

Introduction to Variational Methods in Partial

Differential Equations and Applications

A Summer Course at Michigan State University

(Math 890, Summer 2008)

by

Baisheng Yan

Department of Mathematics
Michigan State University
[email protected]
Contents

Course Description i
A. Motivating Examples i
B. Application to Physical Problems iv
C. Tentative Plans v
Chapter 1. Preliminaries 1
§1.1. Banach Spaces 1
§1.2. Bounded Linear Operators 5
§1.3. Weak Convergence and Compact Operators 7
§1.4. Spectral Theory for Compact Linear Operators 9
§1.5. Nonlinear Functional Analysis 11
Chapter 2. Sobolev Spaces 23
§2.1. Weak Derivatives and Sobolev Spaces 23
§2.2. Approximations and Extensions 27
§2.3. Sobolev Imbedding Theorems 39
§2.4. Equivalent Norms 47
Chapter 3. Existence Theory for Linear Problems 51
§3.1. Differential Equations in Divergence Form 51
§3.2. The Lax-Milgram Theorem 54
§3.3. Gårding’s Estimates and Existence Theory 55
§3.4. Symmetric Elliptic Operators 60
Chapter 4. Variational Methods for PDEs 65
§4.1. Variational Problems 65
§4.2. Multiple Integrals in the Calculus of Variations 66
§4.3. Direct Method for Minimization 70
§4.4. Minimization with Constraints 72
§4.5. Mountain Pass Theorem 86

iii
iv Contents

§4.6. Nonexistence and Radial Symmetry 90

Chapter 5. Weak Lower Semicontinuity on Sobolev Spaces 95
§5.1. The Convex Case 95
§5.2. Morrey’s Quasiconvexity 97
§5.3. Properties of Quasiconvex Functions 102
§5.4. Polyconvex Functions and Null-Lagrangians 108
§5.5. Existence in Nonlinear Elasticity 116
§5.6. Relaxation Principles 118
Bibliography 123
Index 125
Course Description

This course is intended to give an introduction to some important variational methods for
certain problems in partial differential equations (PDE) and applications. The course is
suitable for graduate students with some knowledge of partial differential equations. Since
this is an introduction course meant to convey important ideas of the methods, the rigorous
proof will be kept minimum; some details may be left to students as homework or class
presentation. Attendance, class participation and some extra reading of the references may
be necessary to truly learn the material.

A. Motivating Examples
Variational methods provide a solid basis for the existence theory of PDE and other applied
problems. We use some examples to introduce the main content of the course.

Example 1 – Dirichlet’s Principle. The starting example of variational method for

PDE is the Dirichlet principle for Laplace’s equation:

∆u = 0, u|∂Ω = f,

where Ω is a given bounded domain in Rn and f is a given function on the boundary of Ω.

This principle states that any classical solution u of this problem minimizes the Dirichlet
integral:
Z
I(u) = |∇u(x)|2 dx
Ω
among all smooth functions taking f on the boundary ∂Ω. Therefore, in order to solve the
problem, one tries to find a minimizer of the functional I among the mentioned class of
smooth functions. One of the most important methods for such a minimization problem
is the direct method of the calculus of variations, which originates from the Weierstrass
theorem. By, such a method, we take a minimizing sequence {uj } in the given class; i.e.,

lim I(uj ) = inf I(u);

j→∞

the infimum here is taken over all u in the given class. If this infimum is finite, then we
know that each component of {∇uj } is a bounded sequence in L2 (Ω). The weak convergence

i
ii Course Description

theorem implies that there exist subsequence {∇ujk } and G = (g1 , · · · , gn ) ∈ L2 (Ω) such
that
∇ujk * G weakly in L2 (Ω) and hence Ω |G|2 dx ≤ inf I(u).
R

Now the question is whether G renders a function in the given class; that is, does there
exist a function u in the given class such that G = ∇u? Since, in principle, G is only in
L2 (Ω), it is not clear whether such a u should exist or not. For this problem, the smoothness
of uj or even the fact that ∇uj converges strongly in L2 (Ω) to G (why?) would not help
much. The class in which we seek the minimizers plays an important role in guaranteeing
the existence of a minimizer. We need to have a larger class where I(u) is defined and a
minimizer can be found by this direct method. A minimizer in this larger class is only a
weak solution to the original Laplace equation. Is it smooth and a classical solution of the
Laplace equation? This is the regularity problem; we will not focus on such a problem in
this course. The extension of the admissible class leads use to the class of functions with
gradient in L2 (Ω) and such a class should be also complete (i.e., containing the limits of all
its Cauchy sequences). This motivates the study of Sobolev space H 1 (Ω) = W 1,2 (Ω) or the
general W m,p (Ω) spaces.

Example 2 – Lax-Milgram Method. The second example is on the Hilbert space

method (energy method) for second-order linear elliptic equations in divergence form:
Lu = f in Ω, u = 0 on ∂Ω,
where f is a given function in L2 (Ω) and Lu is a second-order linear elliptic operator:
n
X n
X
ij
Lu = − (a (x)uxi )xj + bi (x)uxi + c(x)u,
i,j=1 i=1

with ellipticity condition: for a constant θ > 0

n
X
ij ji
a =a ; aij (x)ξi ξj ≥ θ|ξ|2 , ∀ a.e. x ∈ Ω, ξ ∈ Rn .
i,j=1

A weak solution u is defined to be a function u ∈ H = H01 (Ω) for which

B[u, v] = (f, v)L2 (Ω) ∀ v ∈ H,
where B is the bilinear form associated with L:
 
Z Xn n
X
B[u, v] =  aij uxi vxj + bi uxi v + cuv  dx, u, v ∈ H.
Ω i,j=1 i=1

Note that f can also be assumed in the dual space H ∗ = H −1 (Ω); in this case, (f, v)L2 (Ω)
above is replaced by the pairing hf, vi between H ∗ and H. Note also that B[u, v] satisfies
the Gårding’s estimates (energy estimates):
|B[u, v]| ≤ αkukkvk, B[u, u] ≥ βkuk2 − γkuk2L2 (Ω) ,

where α > 0, β > 0 and γ ≥ 0 are constants (assuming aij , bi , c are L∞ (Ω) functions). The
Lax-Milgram theorem says that if γ = 0 then for any f ∈ H −1 (Ω) there exists a unique
u ∈ H such that B[u, v] = hf, vi for all v ∈ H. If B is also symmetric then the Lax-Milgram
theorem is simply the Riesz representation theorem. We will not consider the case
when γ > 0; the problem in this case involves the Fredholm alternative theorem.
A. Motivating Examples iii

Example 3 – Mountain Pass Method. Our second example is the use of critical point
theory to find a nontrivial solution to the semilinear elliptic equations of the following type:
∆u + |u|p−1 u = 0 in Ω, u = 0 on ∂Ω,
where 1 < p < n+2
n−2 (we will see why this special exponent appears here later). If we define
a functional on H = H01 (Ω) by
Z  
1 2 1 p+1
I(u) = |∇u| − |u| dx,
Ω 2 p+1
then any critical point of I on H will be a weak solution of the above problem. It can be seen
that I does not have finite infimum or superimum on H. Although there is a method solving
this problem based on minimizing functional I on a manifold of H (i.e., via minimization
with constraints, to be discussed later in the class), to study the critical points of I, we use
the mountain pass method, which is a major contribution in the nonlinear functional
analysis.
The functional I can be proved a C 1 function on H with derivative I 0 : H → H being a
locally Lipschitz function. A critical property of I is that it satisfies the so-called Palais-
Smale condition: every sequence {uj } in H with {I(uj )} bounded and I 0 (uj ) → 0 in H
is precompact in H. Furthermore, there exist positive constants r, a such that
I(u) ≥ a if kuk = r,
and there exists an element v ∈ H with kvk > r and I(v) ≤ 0. Note that I(0) = 0.
Therefore, both inside and outside the mountain range kuk = r there are points where I
takes a smaller value than it takes on the mountain range. Let Γ be the set of all continuous
passes connecting the two lower points 0 and v. Let
c = inf max I(u).
g∈Γ u∈g

Then the mountain pass theorem says that c is a critical point of I; that is, there exists
u ∈ H such that I(u) = c, I 0 (u) = 0. Note that c ≥ a and hence u 6= 0.

Example 4 – Weak Lower Semicontinuity. The direct method also works for the
minimization problem for many other integral functionals of the type
Z
I(u) = F (x, u(x), Du(x)) dx,
Ω
where u may be even a vector valued function. An important question in this regard is when
the functional I is lower semi-continuous with respect to weak convergence of W 1,p (Ω); that
is, when we have
I(u) ≤ lim inf I(uj ) whenever uj * u in W 1,p (Ω).
j→∞

This will motivate the study of various notion of convexity conditions. In the scalar case,
the weak lower semicontinuity of I is equivalent to the convexity of function F (x, u, ξ)
on the variable ξ ∈ Rn . In the vectorial case (where, say, u : Ω → RN for some N ≥ 2), the
weak lower semicontinuity problem is a difficult problem and involves Morrey’s notion of
quasiconvexity, which will also be discussed in the course. But, a sufficient condition for the
semicontinuity will be given in terms of null-Lagrangians; this will be the polyconvexity.
The vectorial case is closely related to the problems in nonlinear elasticity, harmonic
maps, liquid crystals, and other physical problems.
iv Course Description

B. Application to Physical Problems

The problems in nonlinear elasticity, liquid crystals and ferromagnetics will be disucced as
some physical problems which can be formulated and solved by variational methods.

Problem 1 – Nonlinear Elasticity. In continuum mechanics, a material occupying a

domain Ω ⊂ R3 is deformed by a map u to another domain in the same space. The material
at position x ∈ Ω is deformed to a point u(x) in the deformed domain u(Ω). The nonlinear
elasticity theory postulates that the total stored energy associated with the deformation u
is given by Z
I(u) = F (x, u(x), Du(x)) dx,
Ω
where Du(x) = (∂ui /∂xj ) is the 3 × 3-matrix of deformation gradient and F is the stored
energy density function. here, adj A and det A denote the cofactor matrix and the deter-
minant of matrix A. For elasticity problems, the constraint det Du(x) > 0 for a.e. x ∈ Ω
is always assumed; this renders an additional difficulty for the variational problems. Also,
material property and frame-indifference often prevent the function F (x, u, A) from being
convex in A. Nevertheless, the density function sometime can be written as (or is often
assumed to be)
F (x, u, A) = W (x, u, A, adj A, det A),
where adj A and det A denote the cofactor matrix and the determinant of matrix A, and
W (x, u, A, B, t) is a convex function of (A, B, t). Exactly, this means F is polyconvex. For
incompressible materials, the constraint det Du(x) = 1 is assumed. The relationship be-
tween weak convergence and determinant involves the compensated compactness, and
we will discuss this using the null-Lagrangians under a higher regularity assumption.

Problem 2 – Liquid Crystals. A liquid crystal is described by the orientation of the

line-like molecules. Such an orientation can be thought as a unit vector n(x) ∈ S 2 at each
material point x ∈ Ω, the domain occupied by the liquid crystal. The total energy, based
on the Oseen-Frank model, is given by
Z Z
I(n) = WOF (n, Dn) dx = (κ1 (div n)2 + κ2 (n · curl n)2 + κ3 (n × curl n)2 ) dx
Ω
ZΩ
+ κ4 (tr((Dn)2 ) − (div n)2 ) dx,
Ω

where κi (i = 1, 2, 3) are Frank’s curvature constants; the κ4 -term is a null-Lagrangian,

depending only on the boundary data of n. If first three κi > 0, then it can be shown that
the first part of WOF (n, A) is convex in A. If all ki are equal to a positive constant κ, then
I(n) reduces to the Dirichlet integral for harmonic maps: I(n) = Ω κ|Dn|2 dx. Note that
R

the constraint |n(x)| = 1 is lower-order in H 1 (Ω; S 2 ) and thus presents no problem when
using the direct method, but this nonconvex condition is the main obstacle for uniqueness
and regularity.

Problem 3 – Micromagnetics. In the theory of micromagnetics, one seeks the magne-

tization m : Ω ⊂ RN → RN of a body occupying the region Ω by minimizing the energy
functional
Z Z Z Z
α 1
I(m) = |∇m(x)|2 dx + ϕ(m(x)) dx − H · m(x) dx + |F (z)|2 dz
2 Ω Ω Ω 2 RN
C. Tentative Plans v

among all admissible magnetizations m satisfying

m ∈ L∞ (Ω), |m(x)| = 1 a.e. x ∈ Ω,
where F ∈ L2 (RN ; RN ) is the unique field determined by the simplified Maxwell’s equations:
curl F = 0, div(−F + mχΩ ) = 0 in RN .
Here α > 0 is a material constant, ϕ is the density of anisotropy that is minimized along
preferred crystallographic directions, H is a given external applied field (typically constant),
and F is the induced magnetic field on the whole RN related to m via Maxwell’s equations
above. The first term in the energy I(m) is called the exchange energy, the second term
called the anisotropy energy, the third the external interaction energy, and the last term
called the magnetostatic energy.
For large domains, the α-term of the energy can be dropped, which in any case makes the
problem similar to the liquid crystal problem. So we assume α = 0. Also, we will restrict
attention to the two-dimensional case N = 2, although N = 3 is the most physical case.
We will show that the minimization in this simplified case can be reformulated as a mini-
mization problem for nonconvex variational integral of the form
Z Z
1
J(u) = φ(∇u(x)) dx + |∇u(x)|2 dx,
Ω 2 R2 \Ω
where φ is a nonconvex function. Such a functional differs from what we considered above
since it is defined also on the complement of Ω. If time permits, I will discuss some new
results including the relaxation principle for J(u).

C. Tentative Plans
1. Preliminaries. I plan to start with some preliminary results in functional analysis
using some materials of a lecture note by Professor Dunninger; most of the materials can
be found in any textbook on real and functional analysis (also in Appendix D of Evans’
book, Partial Differential Equations, AMS Press).

2. Sobolev Spaces. I will do the Sobolev spaces using Evans’ book, Chapter 5, and some-
time refer to the new book Sobolev Spaces, Academic Press, by Adams and Fournier.

3. Variational Methods for PDEs. I will follow mainly Chapters 6 and 8 of Evans’
book on the variational methods discussed above, but will add some materials deemed to
be useful in understanding and developing the methods. If time permits, the maximum
principles and Sections 9.4.2 and 9.5.2 of Evans’ book may also be included.

4. Applications. I will briefly discuss the physical applied problems mentioned above
when a method can be applied to the problem. Some of the materials will come from sev-
eral research articles, including my own.
If time permits, I will discuss the nonconvex variational problems and relaxation prin-
ciple and the Young measure solutions to such problems; explicit example will be the
2-D micromagnetics problem.
 Chapter 1

Preliminaries

1.1. Banach Spaces

1.1.1. Vector Spaces. A (real) vector space is a set X, whose elements are called
vectors, and in which two operations, addition and scalar multiplication, are defined
as follows:
(a) To every pair of vectors x and y corresponds a vector x + y in such a way that
x+y =y+x and x + (y + z) = (x + y) + z.
X contains a unique vector 0 (the zero vector or origin of X) such that x+0 = x
for every x ∈ X, and to each x ∈ X corresponds a unique vector −x such that
x + (−x) = 0.
(b) To every pair (α, x), with α ∈ R and x ∈ X, corresponds a vector αx in such a
way that
1x = x, α(βx) = (αβ)x
and such that the two distributive laws
α(x + y) = αx + αy, (α + β)x = αx + βx
hold.

A nonempty subset M of a vector space X is called a subspace of X if αx + βy ∈ M

for all x, y ∈ M and all α, β ∈ R. A subset M of a vector space X is said to be convex if
tx + (1 − t)y ∈ M whenever t ∈ (0, 1), x, y ∈ M . (Clearly, every subspace of X is convex.)
Let x1 , . . . , xn be elements of a vector space X. The set of all α1 x1 + · · · + αn xn , with
αi ∈ R, is called the span of x1 , . . . , xn and is denoted by span{x1 , . . . , xn }. The elements
x1 , . . . , xn are said to be linearly independent if α1 x1 +· · ·+αn xn = 0 implies that αi = 0
for each i. If, on the other hand, α1 x1 + · · · + αn xn = 0 does not imply αi = 0 for each
i, the elements x1 , . . . , xn are said to be linearly dependent . An arbitrary collection of
vectors is said to be linearly independent if every finite subset of distinct elements is linearly
independent.
The dimension of a vector space X, denoted by dim X, is either 0, a positive integer
or ∞. If X = {0} then dim X = 0; if there exist linearly independent {u1 , . . . , un } such

1
2 1. Preliminaries

that each x ∈ X has a (unique) representation of the form

x = α1 u1 + · · · + αn un with αi ∈ R
then dim X = n and {u1 , . . . , un } is a basis for X; in all other cases dim X = ∞.

1.1.2. Normed Spaces. A (real) vector space X is said to be a normed space if to

every x ∈ X there is associated a nonnegative real number kxk, called the norm of x, in
such a way that
(a) kx + yk ≤ kxk + kyk for all x and y in X (Triangle inequality)
(b) kαxk = |α|kxk for all x ∈ X and all α ∈ R
(c) kxk > 0 if x 6= 0.
Note that (b) and (c) imply that kxk = 0 iff x = 0. Moreover, it easily follows from (a) that
|kxk − kyk| ≤ kx − yk for all x, y ∈ X.

1.1.3. Completeness and Banach Spaces. A sequence {xn } in a normed space X is

called a Cauchy sequence if, for each  > 0, there exists an integer N such that kxm −
xn k <  for all m, n ≥ N . We say xn → x in X if limn→∞ kxn − xk = 0 and, in this case, x
is called the limit of {xn }. X is called complete if every Cauchy sequence in X converges
to a limit in X.
A complete (real) normed space is called a (real) Banach space. A Banach space
is separable if it contains a countable dense set. It can be shown that a subspace of a
separable Banach space is itself separable.
Example 1.1. Let Ω be an open subset of Rn , n ≥ 1. The set C(Ω) of (real-valued)
continuous functions defined on Ω is an infinite dimensional vector space with the usual
definitions of addition and scalar multiplication:
(f + g)(x) = f (x) + g(x) for f, g ∈ C(Ω), x ∈ Ω
(αf )(x) = αf (x) for α ∈ R, f ∈ C(Ω), x ∈ Ω.
C(Ω̄) consists of those functions which are uniformly continuous on Ω. Each such function
has a continuous extension to Ω̄. C0 (Ω) consists of those functions which are continuous
in Ω and have compact support in Ω. (The support of a function f defined on Ω is the
closure of the set {x ∈ Ω : f (x) 6= 0} and is denoted by supp(f ).) The latter two spaces are
clearly subspaces of C(Ω).
For each n-tuple α = (α1 , . . . , αn ) of nonnegative integers, we denote by Dα the partial
derivative
D1α1 · · · Dnαn , Di = ∂/∂xi
of order |α| = α1 + · · · + αn . If |α| = 0, then D0 = I(identity).
For integers m ≥ 0, let C m (Ω) be the collection of all f ∈ C(Ω) such that Dα f ∈ C(Ω)
for all α with |α| ≤ m. We write f ∈ C ∞ (Ω) iff f ∈ C m (Ω) for all m ≥ 0. For
m ≥ 0, define C0m (Ω) = C0 (Ω) ∩ C m (Ω) and let C0∞ (Ω) = C0 (Ω) ∩ C ∞ (Ω). The spaces
C m (Ω), C ∞ (Ω), C0m (Ω), C0∞ (Ω) are all subspaces of the vector space C(Ω). Similar defini-
tions can be given for C m (Ω̄) etc.
For m ≥ 0, define X to be the set of all f ∈ C m (Ω) for which
X
kf km,∞ ≡ sup |Dα f (x)| < ∞.
Ω
|α|≤m
1.1. Banach Spaces 3

Then X is a Banach space with norm k · km,∞ . To prove, for example, the completeness
when m = 0, we let {fn } be a Cauchy sequence in X, i.e., assume for any ε > 0 there is a
number N (ε) such that for all x ∈ Ω

sup |fn (x) − fm (x)| < ε if m, n > N (ε).

x∈Ω

But this means that {fn (x)} is a uniformly Cauchy sequence of bounded continuous func-
tions, and thus converges uniformly to a bounded continuous function f (x). Letting m → ∞
in the above inequality shows that kfn − f km,∞ → 0.
Note that the same proof is valid for the set of bounded continuous scalar-valued func-
tions defined on a nonempty subset of a normed space X.

Example 1.2. Let Ω be a nonempty Lebesgue measurable set in Rn . For p ∈ [1, ∞), we
denote by Lp (Ω) the set of equivalence classes of Lebesgue measurable functions on Ω for
which
Z 1
p
p
kf kp ≡ |f (x)| dx < ∞.
Ω

(Two functions belong to the same equivalence class, i.e., are equivalent, if they differ
only on a set of measure 0.) Let L∞ (Ω) denote the set of equivalence classes of Lebesgue
measurable functions on Ω for which

kf k∞ ≡ ess-supx∈Ω |f (x)| < ∞.

Then Lp (Ω), 1 ≤ p ≤ ∞, are Banach spaces with norms k · kp . For p ∈ [1, ∞] we write
f ∈ Lploc (Ω) iff f ∈ Lp (K) for each compact set K ⊂ Ω.
For the sake of convenience, we will also consider Lp (Ω) as a set of functions. With this
convention in mind, we can assert that C0 (Ω) ⊂ Lp (Ω). In fact, if p ∈ [1, ∞), then as we shall
show later, C0 (Ω) is dense in Lp (Ω). The space Lp (Ω) is also separable if p ∈ [1, ∞). This
follows easily, when Ω is compact, from the last remark and the Weierstrass approximation
theorem.
Finally we recall that if p, q, r ∈ [1, ∞] with p−1 + q −1 = r−1 , then Hölder’s inequality
implies that if f ∈ Lp (Ω) and g ∈ Lq (Ω), then f g ∈ Lr (Ω) and

kf gkr ≤ kf kp kgkq .

Example 1.3. The Cartesian product X × Y , of two vector spaces X and Y , is itself a
vector space under the following operations of addition and scalar multiplication:

[x1 , y1 ] + [x2 , y2 ] = [x1 + x2 , y1 + y2 ]

α[x, y] = [αx, αy].

If in addition, X and Y are normed spaces with norms k · kX , k · kY respectively, then X × Y
becomes a normed space under the norm

k[x, y]k = kxkX + kykY .

Moreover, under this norm, X × Y becomes a Banach space provided X and Y are Banach
spaces.
4 1. Preliminaries

1.1.4. Hilbert Spaces. Let H be a real vector space. H is said to be an inner product
space if to every pair of vectors x and y in H there corresponds a real-valued function
(x, y), called the inner product of x and y, such that
(a) (x, y) = (y, x) for all x, y ∈ H
(b) (x + y, z) = (x, z) + (y, z) for all x, y, z ∈ H
(c) (λx, y) = λ(x, y) for all x, y ∈ H, λ ∈ R
(d) (x, x) ≥ 0 for all x ∈ H, and (x, x) = 0 if and only if x = 0.
For x ∈ H we set
(1.1) kxk = (x, x)1/2 .
Theorem 1.4. If H is an inner product space, then for all x and y in H, it follows that
(a) |(x, y)| ≤ kxk kyk (Cauchy-Schwarz inequality);
(b) kx + yk ≤ kxk + kyk (Triangle inequality);
(c) kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ) (Parallelogram law).

Proof. (a) is obvious if x = 0, and otherwise it follows by taking δ = −(x, y)/kxk2 in

0 ≤ kδx + yk2 = |δ|2 kxk2 + 2δ(x, y) + kyk2 .
This identity, with δ = 1, and (a) imply (b). (c) follows easily by using (1.1). 

Furthermore, by (d), equation (1.1) defines a norm on an inner product space H. If H

is complete under this norm, then H is said to be a Hilbert space.
Example 1.5. The space L2 (Ω) is a Hilbert space with inner product
Z
(f, g) = f (x)g(x) dx for all f, g ∈ L2 (Ω).
Ω

Theorem 1.6. Every nonempty closed convex subset S of a Hilbert space H contains a
unique element of minimal norm.

Proof. Choose xn ∈ S so that kxn k → d ≡ inf{kxk : x ∈ S}. Since (1/2)(xn + xm ) ∈ S,

we have kxn + xm k2 ≥ 4d2 . Using the parallelogram law, we see that
(1.2) kxn − xm k2 ≤ 2(kxn k2 − d2 ) + 2(kxm k2 − d2 )
and therefore {xn } is a Cauchy sequence in H. Since S is closed, {xn } converges to some
x ∈ S and kxk = d. If y ∈ S and kyk = d, then the parallelogram law implies, as in (1.2),
that x = y. 

If (x, y) = 0, then x and y are said to be orthogonal, written sometimes as x ⊥ y. For

M ⊂ H, the orthogonal complement of M , denoted by M ⊥ , is defined to be the set of
all x ∈ H such that (x, y) = 0 for all y ∈ M . It is easily seen that M ⊥ is a closed subspace
of H. Moreover, if M is a dense subset of H and if x ∈ M ⊥ , then in fact, x ∈ H ⊥ which
implies x = 0.
Theorem 1.7. (Projection) Suppose M is a closed subspace of a Hilbert space H. Then
for each x ∈ H there exist unique y ∈ M , z ∈ M ⊥ such that x = y + z. The element y is
called the projection of x onto M .
1.2. Bounded Linear Operators 5

Proof. Let S = {x − y : y ∈ M }. It is easy to see that S is convex and closed. Theorem 1.6
implies that there exists a y ∈ M such that kx − yk ≤ kx − wk for all w ∈ M . Let z = x − y.
For an arbitrary w ∈ M , w 6= 0, let α = (z, w)/kwk2 and note that
kzk2 ≤ kz − αwk2 = kzk2 − |(z, w)/kwk|2
which implies (z, w) = 0. Therefore z ∈ M ⊥ . If x = y 0 + z 0 for some y 0 ∈ M , z 0 ∈ M ⊥ , then
y 0 − y = z − z 0 ∈ M ∩ M ⊥ = {0}, which implies uniqueness. 

Remark. In particular, if M is a proper closed subspace of H, then there is a nonzero

element in M ⊥ . Indeed, for x ∈ H\M , let y be the projection of x on M . Then z = x − y
is a nonzero element of M ⊥ .

1.2. Bounded Linear Operators

1.2.1. Operators on a Banach Space. Let X, Y be real vector spaces. A map T : X →
Y is said to be a linear operator from X to Y if
T (αx + βy) = αT x + βT y
for all x, y ∈ D(T ) and all α, β ∈ R.
Let X, Y be normed spaces. A linear operator T from X to Y is said to be bounded if
there exists a constant m > 0 such that
(1.3) kT xk ≤ mkxk for all x ∈ X.
We define the operator norm kT k of T by
(1.4) kT k = sup kT xk = sup kT xk.
x∈X, kxk=1 x∈X, kxk≤1

The collection of all bounded linear operators T : X → Y will be denoted by B(X, Y ). We

shall also set B(X) = B(X, X) when X = Y . Observe that
kT Sk ≤ kT kkSk if S ∈ B(X, Y ), T ∈ B(Y, Z).
Theorem 1.8. If X and Y are normed spaces, then B(X, Y ) is a normed space with norm
defined by equation (1.4). If Y is a Banach space, then B(X, Y ) is also a Banach space.

Proof. It is easy to see that B(X, Y ) is a normed space. To prove completeness, assume
that {Tn } is a Cauchy sequence in B(X, Y ). Since
(1.5) kTn x − Tm xk ≤ kTn − Tm kkxk
we see that, for fixed x ∈ X, {Tn x} is a Cauchy sequence in Y and therefore we can define
a linear operator T by
T x = lim Tn x for all x ∈ X.
n→∞
If ε > 0, then the right side of (1.5) is smaller than εkxk provided that m and n are large
enough. Thus, (letting n → ∞)
kT x − Tm xk ≤ εkxk for all large enough m.
Hence, kT xk ≤ (kTm k + ε)kxk, which shows that T ∈ B(X, Y ). Moreover, kT − Tm k < ε
for all large enough m. Hence, limn→∞ Tn = T . 

The following theorems are important in linear functional analysis; see, e.g., [3].
Theorem 1.9. (Banach-Steinhaus) Let X be a Banach space and Y a normed space. If
A ⊂ B(X, Y ) is such that supT ∈A kT xk < ∞ for each fixed x ∈ X, then supT ∈A kT k < ∞.
6 1. Preliminaries

Theorem 1.10. (Bounded Inverse) If X and Y are Banach spaces and if T ∈ B(X, Y )
is one-to-one and onto, then T −1 ∈ B(Y, X).

1.2.2. Dual Spaces and Reflexivity. When X is a (real) normed space, the Banach
space B(X, R) will be called the (normed) dual space of X and will be denoted by X*. El-
ements of X* are called bounded linear functionals or continuous linear functionals
on X. Frequently, we shall use the notation hf, xi to denote the value of f ∈ X* at x ∈ X.
Using this notation we note that |hf, xi| ≤ kf k kxk for all f ∈ X*, x ∈ X.
Example 1.11. Suppose 1 < p, q < ∞ satisfy 1/p + 1/q = 1 and let Ω be a nonempty
Lebesgue measurable set in Rn . Then Lp (Ω)∗ = Lq (Ω). The case of p = ∞ is different. The
dual of L∞ is much larger then L1 .

The following results can be found in [3].

Theorem 1.12. (Hahn-Banach) Let X be a normed space and Y a subspace of X. As-
sume f ∈ Y ∗ . Then there exists a bounded linear functional f˜ ∈ X ∗ such that
hf˜, yi = hf, yi ∀ y ∈ Y, kf˜kX ∗ = kf kY ∗ .
Corollary 1.13. Let X be a normed space and x0 6= 0 in X. Then there exists f ∈ X ∗
such that
kf k = 1, hf, x0 i = kx0 k.

The dual space X ∗∗ of X ∗ is called the second dual space of X and is again a Banach
space. Note that to each x ∈ X we can associate a unique Fx ∈ X ∗∗ by Fx (f ) = hf, xi for
all f ∈ X ∗ . From Corollary 1.13, one can also show that kFx k = kxk. Thus, the (canonical)
mapping J : X → X**, given by Jx = Fx , is a linear isometry of X onto the subspace
J(X) of X ∗∗ . Since J is one-to-one, we can identify X with J(X).
A Banach space X is called reflexive if its canonical map J is onto X ∗∗ . For example,
all Lp spaces with 1 < p < ∞ are reflexive.

We shall need the following properties of reflexive spaces.

Theorem 1.14. Let X and Y be Banach spaces.
(a) X is reflexive iff X* is reflexive.
(b) If X is reflexive, then a closed subspace of X is reflexive.
(c) Let T : X → Y be a linear bijective isometry. If Y is reflexive, then X is
reflexive.

1.2.3. Bounded Linear Functionals on a Hilbert Space.

Theorem 1.15. (Riesz Representation) If H is a Hilbert space and f ∈ H*, then there
exists a unique y ∈ H such that
f (x) = hf, xi = (x, y) for all x ∈ H.
Moreover, kf k = kyk.

Proof. If f (x) = 0 for all x, take y = 0. Otherwise, there is an element z ∈ N (f )⊥ such

that kzk = 1. (Note that the linearity and continuity of f implies that N (f ) is a closed
subspace of H.) Put u = f (x)z −f (z)x. Since f (u) = 0, we have u ∈ N (f ). Thus (u, z) = 0,
which implies
f (x) = f (x)(z, z) = f (z)(x, z) = (x, f (z)z) = (x, y),
1.3. Weak Convergence and Compact Operators 7

where y = f (z)z. To prove uniqueness, suppose (x, y) = (x, y 0 ) for all x ∈ H. Then in
particular, (y − y 0 , y − y 0 ) = 0, which implies y = y 0 . From the Cauchy-Schwarz inequality
we get |f (x)| ≤ kxkkyk, which yields kf k ≤ kyk. The reverse inequality follows by choosing
x = y in the representation. 
Corollary 1.16. H is reflexive.

Let T : H → H be an operator on the Hilbert space H. We define the Hilbert space

adjoint T * : H → H as follows:
(T x, y) = (x, T *y) for all x, y ∈ H.
The adjoint operator is easily seen to be linear.
Theorem 1.17. Let H be a Hilbert space. If T ∈ B(H), then T * ∈ B(H) and kT k = kT *k.

Proof. For any y ∈ H and all x ∈ H, set f (x) = (T x, y). Then it is easily seen that
f ∈ H ∗ . Hence by the Riesz representation theorem, there exists a unique z ∈ H such that
(T x, y) = (x, z) for all x ∈ H, i.e., D(T ∗ ) = H. Moreover, kT ∗ yk = kzk = kf k ≤ kT kkyk,
i.e., T ∗ ∈ B(H) and kT ∗ k ≤ kT k. The reverse inequality follows easily from kT xk2 =
(T x, T x) = (x, T ∗ T x) ≤ kT xkkT ∗ kkxk. 

1.3. Weak Convergence and Compact Operators

1.3.1. Weak Convergence. Let X be a normed space. A sequence xn ∈ X is said to
be weakly convergent to an element x ∈ X, written xn * x, if hf, xn i → hf, xi for all
f ∈ X ∗.
Theorem 1.18. Let {xn } be a sequence in X.
(a) Weak limits are unique.
(b) If xn → x, then xn * x.
(c) If xn * x, then {xn } is bounded and kxk ≤ lim inf kxn k.

Proof. To prove (a), suppose that x and y are both weak limits of the sequence {xn } and
set z = x − y. Then hf, zi = 0 for every f ∈ X ∗ and by Corollary 1.13, z = 0. To prove (b),
let f ∈ X ∗ and note that xn → x implies hf, xn i → hf, xi since f is continuous. To prove (c),
assume xn * x and consider the sequence {Jxn } of elements of X ∗∗ , where J : X → X ∗∗
is the bounded operator defined above. For each f ∈ X ∗ , sup |Jxn (f )| = sup |hf, xn i| < ∞
(since hf, xn i converges). By the Banach-Steinhaus Theorem, there exists a constant c such
that kxn k = kJxn k ≤ c which implies {xn } is bounded. Finally, for f ∈ X ∗
|hf, xi| = lim |hf, xn i| ≤ lim inf kf kkxn k = kf k lim inf kxn k
which implies the desired inequality since kxk = supkf k=1 |hf, xi|. 

We note that in a Hilbert space H, the Riesz representation theorem implies that xn * x
means (xn , y) → (x, y) for all y ∈ H. Moreover, we have
(xn , yn ) → (x, y) if xn * x, yn → y.
This follows from the estimate
|(x, y) − (xn , yn )| = |(x − xn , y) − (xn , yn − y)| ≤ |(x − xn , y)| + kxn kky − yn k
and the fact that kxn k is bounded.
8 1. Preliminaries

The main result of this section is given by:

Theorem 1.19. If X is a reflexive Banach space, then the closed unit ball is weakly
compact, i.e., the sequence {xn }, with kxn k ≤ 1 has a subsequence which converges weakly
to an x with kxk ≤ 1.

1.3.2. Compact Operators. Let X and Y be normed spaces. An operator T : X → Y is

said to be compact if it maps bounded sets in X into relatively compact sets in Y , i.e., if
for every bounded sequence {xn } in X, {T xn } has a subsequence which converges to some
element of Y .
Since relatively compact sets are bounded, it follows that a compact operator is bounded.
On the other hand, since bounded sets in finite-dimensional spaces are relatively compact, it
follows that a bounded operator with finite dimensional range is compact. It can be shown
that the identity map I : X → X (kIxk = kxk) is compact iff X is finite-dimensional.
Finally we note that the operator ST is compact if (a) T : X → Y is compact and S : Y → Z
is continuous or (b) T is bounded and S is compact.
One of the main methods of proving the compactness of certain operators is based upon
the Ascoli theorem.
Let Ω be a subset of the normed space X. A set S ⊂ C(Ω) is said to be equicontinuous
if for each ε > 0 there exists a δ > 0 such that |f (x) − f (y)| < ε for all x, y ∈ Ω with
kx − yk < δ and for all f ∈ S.
Theorem 1.20. (Ascoli) Let Ω be a relatively compact subset of a normed space X and
let S ⊂ C(Ω). Then S is relatively compact if it is bounded and equicontinuous.

Remark. In other words, every bounded equicontinuous sequence of functions has a uni-
formly convergent subsequence.
Theorem 1.21. Let X and Y be Banach spaces. If Tn : X → Y are linear and compact
for n ≥ 1 and if limn→∞ kTn − T k = 0, then T is compact. Thus, compact operators form
a closed, but not a dense, subspace of B(X, Y ).

Proof. Let {xn } be a sequence in X with M = supn kxn k < ∞. Let A1 denote an infinite
set of integers such the sequence {T1 xn }n∈A1 converges. For k ≥ 2 let Ak ⊂ Ak−1 denote
an infinite set of integers such that the sequence {Tk xn }n∈Ak converges. Choose n1 ∈ A1
and nk ∈ Ak , nk > nk−1 for k ≥ 2. Choose ε > 0. Let k be such that kT − Tk kM < ε/4
and note that
kT xni − T xnj k ≤ k(T − Tk )(xni − xnj )k + kTk xni − Tk xnj k < ε/2 + kTk xni − Tk xnj k.
Since {Tk xni }∞ ∞
i=1 converges, {T xni }i=1 is a Cauchy sequence. 
Theorem 1.22. Let X and Y be normed spaces.
(a) If T ∈ B(X, Y ), then T is weakly continuous, i.e.,
xn * x implies T xn * T x.
(b) If T : X → Y is weakly continuous and X is a reflexive Banach space, then T
is bounded.
(c) If T ∈ B(X, Y ) is compact, then T is strongly continuous, i.e.,
xn * x implies T xn → T x.
1.4. Spectral Theory for Compact Linear Operators 9

(d) If T : X → Y is strongly continuous and X is a reflexive Banach space, then T

is compact.

Proof. (a) Let xn * x. Then for every g ∈ Y ∗

hg, T xn i = hT ∗ g, xn i → hT ∗ g, xi = hg, T xi.
(b) If not, there is a bounded sequence {xn } such that kT xn k → ∞. Since X is reflexive,
{xn } has a weakly convergent subsequence, {xn0 }, and so {T xn0 } also converges weakly. But
then {T xn0 } is bounded, which is a contradiction.
(c) Let xn * x. Since T is compact and {xn } is bounded, there is a subsequence {xn0 }
such that T xn0 → z, and thus T xn0 * z. By (a), T xn * T x, and so T xn0 → T x. Now
it is easily seen that every subsequence of {xn } has a subsequence, say {xn0 }, such that
T xn0 → T x. But this implies the whole sequence T xn → T x (See the appendix).
(d) Let {xn } be a bounded sequence. Since X is reflexive, there is a subsequence {xn0 }
such that xn0 * x. Hence T xn0 → T x, which implies T is compact. 
Theorem 1.23. Let H be a Hilbert space. If T : H → H is linear and compact, then T * is
compact.

Proof. Let {xn } be a sequence in H satisfying kxn k ≤ m. The sequence {T *xn } is therefore
bounded, since T * is bounded. Since T is compact, by passing to a subsequence if necessary,
we may assume that the sequence {T T *xn } converges. But then
kT *(xn − xm )k2 = (xn − xm , T T *(xn − xm ))
≤ 2mkT T *(xn − xm )k → 0 as m, n → ∞.
Since H is complete, the sequence {T *xn } is convergent and hence T * is compact. 

1.4. Spectral Theory for Compact Linear Operators

1.4.1. Fredholm Alternative.
Theorem 1.24. (Fredholm Alternative) Let T : H → H be a compact linear operator
on the Hilbert space H. Then equations (I − T )x = 0, (I − T ∗ )x∗ = 0 have the same finite
number of linearly independent solutions. Moreover,
(a) For y ∈ H, the equation (I − T )x = y has a solution iff (y, x∗ ) = 0 for every
solution x∗ of (I − T ∗ )x∗ = 0.
(b) For z ∈ H, the equation (I − T ∗ )x∗ = z has a solution iff (z, x) = 0 for every
solution x of (I − T )x = 0.
(c) The inverse operator (I − T )−1 ∈ B(H) whenever it exists.

1.4.2. Spectrum of Compact Operators. A subset S of a Hilbert space H is said to be

an orthonormal set if each element of S has norm 1 and if every pair of distinct elements
in S is orthogonal. It easily follows that an orthonormal
P set is linearly independent. An
orthonormal set S is said to be complete if x = φ∈S (x, φ)φ for all x ∈ H. It can be shown
that (x, φ) 6= 0 for at most countably many φ ∈ S. This series is called the Fourier series
for x with respect to the orthonormal set {φ}. Let {φi }∞
i=1 be a countable orthonormal set
PN 2
in H. Upon expanding kx − n=1 (x, φn )φn k , we arrive at Bessel’s inequality:
∞
X
|(x, φn )|2 ≤ kxk2 .
n=1
10 1. Preliminaries

Let T : D(T ) ⊂ H → H be a linear operator on the real Hilbert space H. The set ρ(T )
of all scalars λ ∈ R for which (T − λI)−1 ∈ B(H) is called the resolvent set of T . The
operator R(λ) = (T − λI)−1 is known as the resolvent of T . σ(T ) = R \ ρ(T ) is called the
spectrum of T . It can be shown that ρ(T ) is an open set and σ(T ) is a closed set. The
set of λ ∈ R for which there exists a nonzero x ∈ N (T − λI) is called the point spectrum
of T and is denoted by σp (T ). The elements of σp (T ) are called the eigenvalues of T and
the nonzero members of N (T − λI) are called the eigenvectors (or eigenfunctions if X
is a function space) of T .
If T is compact and λ 6= 0, then by the Fredholm alternative, either λ ∈ σp (T ) or
λ ∈ ρ(T ). Moreover, if H is infinite-dimensional, then 0 6∈ ρ(T ); otherwise, T −1 ∈ B(H)
and T −1 T = I is compact. As a consequence, σ(T ) consists of the nonzero eigenvalues of T
together with the point 0. The next result shows that σp (T ) is either finite or a countably
infinite sequence tending to zero.
Theorem 1.25. Let T : X → X be a compact linear operator on the normed space X.
Then for each r > 0 there exist at most finitely many λ ∈ σp (T ) for which |λ| > r.

1.4.3. Symmetric Compact Operators. The next result implies that a symmetric com-
pact operator on a Hilbert space has at least one eigenvalue. On the other hand, an arbi-
trary bounded, linear, symmetric operator need not have any eigenvalues. As an example,
let T : L2 (0, 1) → L2 (0, 1) be defined by T u(x) = xu(x).
Theorem 1.26. Suppose T ∈ B(H) is symmetric, i.e., (T x, y) = (x, T y) for all x, y ∈ H.
Then
kT k = sup |(T x, x)|.
kxk=1

Moreover, if H 6= {0}, then there exists a real number λ ∈ σ(T ) such that |λ| = kT k. If
λ ∈ σp (T ), then in absolute value λ is the largest eigenvalue of T .

Proof. Clearly m ≡ supkxk=1 |(T x, x)| ≤ kT k. To show kT k ≤ m, observe that for all
x, y ∈ H
2(T x, y) + 2(T y, x) = (T (x + y), x + y) − (T (x − y), x − y)
≤ m(kx + yk2 + kx − yk2 )
= 2m(kxk2 + kyk2 )
where the last step follows from the paralleogram law. Hence, if T x 6= 0 and y =
(kxk/kT xk)T x, then
2kxkkT xk = (T x, y) + (y, T x) ≤ m(kxk2 + kyk2 ) = 2mkxk2
which implies kT xk ≤ mkxk. Since this is also valid when T x = 0, we have kT k ≤ m. To
prove the ‘moreover’ part, choose xn ∈ H such that kxn k = 1 and kT k = limn→∞ |(T xn , xn )|.
By renaming a subsequence of {xn }, we may assume that (T xn , xn ) converge to some real
number λ with |λ| = kT k. Observe that
k(T − λ)xn k2 = kT xn k2 − 2λ(T xn , xn ) + λ2 kxn k2
≤ 2λ2 − 2λ(T xn , xn ) → 0.
We now claim that λ ∈ σ(T ). Otherwise, we arrive at the contradiction
1 = kxn k = k(T − λ)−1 (T − λ)xn k ≤ k(T − λ)−1 k k(T − λ)xn k → 0.
1.5. Nonlinear Functional Analysis 11

Finally, we note that if T φ = µφ, with kφk = 1, then |µ| = |(T φ, φ)| ≤ kT k which implies
the last assertion of the theorem. 

Finally we have the following result.

Theorem 1.27. Let H be a separable Hilbert space and suppose T : H → H is linear,

symmetric and compact. Then there exists a countable complete orthonormal set in H
consisting of eigenvectors of T .

1.5. Nonlinear Functional Analysis

In this final preliminary section, we list some useful results in nonlinear functional anal-
ysis. Lots of the proof and other results can be found in the volumes of Zeidler’s book
[41].

1.5.1. Contraction Mapping Theorem. Let X be a normed space. A map T : X → X

is called a contraction if there exists a number k < 1 such that
(1.6) kT x − T yk ≤ kkx − yk for all x, y ∈ X.

Theorem 1.28. (Contraction Mapping) Let T : S ⊂ X → S be a contraction on the

closed nonempty subset S of the Banach space X. Then T has a unique fixed point, i.e.,
there exists a unique solution x ∈ S of the equation T x = x. Moreover, x = limn→∞ T n x0
for any choice of x0 ∈ S.

Proof. To prove uniqueness, suppose T x = x, T y = y. Since k < 1, we get x = y from

kx − yk = kT x − T yk ≤ kkx − yk.
To show that T has a fixed point we set up an iteration procedure. For any x0 ∈ S set
xn+1 = T xn , n = 0, 1, ...
Note that xn+1 ∈ S and xn+1 = T n+1 x0 . We now claim that {xn } is a Cauchy sequence.
Indeed, for any integers n, p
n+p−1
X
n+p n
kxn+p − xn k = kT x0 − T x0 k ≤ kT j+1 x0 − T j x0 k
j=n
n+p−1
X kn
≤ k j kT x0 − x0 k ≤ kT x0 − x0 k.
1−k
j=n

Hence as n → ∞, kxn+p − xn k → 0 independently of p, so that {xn } is a Cauchy sequence

with limit x ∈ S. Since T is continuous, we have
T x = lim T xn = lim xn+1 = x
n→∞ n→∞

and thus x is the unique fixed point. Note that the fixed point x is independent of x0 since
x is a fixed point and fixed points are unique. 

Our main existence result will be based upon the following so-called method of con-
tinuity or continuation method.
12 1. Preliminaries

Theorem 1.29. Let T0 , T1 ∈ B(X, Y ), where X is a Banach space and Y is a normed

space. For each t ∈ [0, 1] set
Tt = (1 − t)T0 + tT1
and suppose there exists a constant c > 0 such that for all t ∈ [0, 1] and x ∈ X
(1.7) kxkX ≤ ckTt xkY .
Then R(T1 ) = Y if R(T0 ) = Y .

Proof. Set S = {t ∈ [0, 1] : R(Tt ) = Y }. By hypothesis, 0 ∈ S. We need to show that

1 ∈ S. In this direction we will show that if τ > 0 and τ c(kT1 k + kT0 k) < 1, then
(1.8) [0, s] ⊂ S implies [0, s + τ ] ⊂ S.
(Note that any smaller τ works.) Since τ can be chosen independently of s, (1.8) applied
finitely many times gets us from 0 ∈ S to 1 ∈ S.
Let s ∈ S. For t = s + τ, Tt x = f is equivalent to the equation
(1.9) Ts x = f + τ T0 x − τ T1 x.
By (1.7), Ts−1 : Y → X exists and kTs−1 k ≤ c. Hence (1.9) is equivalent to

(1.10) x = Ts−1 (f + τ T0 x − τ T1 x) ≡ Ax
and for A : X → X we have for all x, y ∈ X
kAx − Ayk ≤ τ c(kT1 k + kT0 k)kx − yk.
By the contraction mapping theorem, (1.10) has a solution and this completes the proof. 

1.5.2. Nemytskii Operators. Let Ω be a nonempty measurable set in Rn and let f :

Ω × R → R be a given function. Assume
(i) for every ξ ∈ R, f (x, ξ) (as a function of x) is measurable on Ω
(ii) for almost all x ∈ Ω, f (x, ξ) (as a function of ξ) is continuous on R
(iii) for all (x, ξ) ∈ Ω × R
|f (x, ξ)| ≤ a(x) + b|ξ|p/q
where b is a fixed nonnegative number, a ∈ Lq (Ω) is nonnegative and 1 < p, q < ∞, 1/p +
1/q = 1. Note that p/q = p − 1. Then the Nemytskii operator N is defined by
N u(x) = f (x, u(x)), x ∈ Ω.
We have the following result needed later.
Lemma 1.30. N : Lp (Ω) → Lq (Ω) is continuous and bounded with
(1.11) kN ukq ≤ const (kakq + kukp/q
p ) for all u ∈ Lp (Ω)
and
Z
(1.12) hN u, vi = f (x, u(x))v(x)dx for all u, v ∈ Lp (Ω).
Ω

Here h·, ·i denotes the duality pairing between Lp (Ω) and Lq (Ω).
1.5. Nonlinear Functional Analysis 13

Proof. Since u ∈ Lp (Ω), the function u(x) is measurable on Ω and thus, Pn by (i) and P
(ii), the
function f (x, u(x)) is also measurable on Ω. From the inequality ( i=1 ξi ) ≤ c ni=1 ξir
r

and (iii) we get

|f (x, u(x))|q ≤ const(|a(x)|q + |u(x)|p ).
Integrating over Ω and applying the above inequality once more yields (1.11), which shows
that N is bounded.
To show that N is continuous, let un → u in Lp (Ω). Then there is a subsequence {un0 }
and a function v ∈ Lp (Ω) such that un0 (x) → u(x) a.e. and |un0 (x)| ≤ v(x) a.e. for all n.
Hence
Z
q
kN un0 − N ukq = |f (x, un0 (x)) − f (x, u(x))|q dx
Ω
Z
≤ const (|f (x, un0 (x))|q + |f (x, u(x))|q )dx
ZΩ
≤ const (|a(x)|q + |v(x)|p + |u(x)|p )dx.
Ω
By (ii), f (x, u (x)) − f (x, u(x)) → 0 as n → ∞ for almost all x ∈ Ω. The dominating
n0
convergence theorem implies that kN un0 − N ukq → 0. By repeating this procedure for any
subsequence of un0 , it follows that kN un − N ukq → 0 which implies that N is continuous.
Since N u ∈ (Lp (Ω))*, the integral representation (1.12) is clear. 

Remarks. (a) The following remarkable statement can be proved: If f satisfies (i) and (ii)
above and if the corresponding Nemytskii operator is such that N : Lp (Ω) → Lq (Ω), then
N is continuous, bounded and (iii) holds.
(b) If (iii) is replaced by
(iii)0 for all (x, ξ) ∈ Ω × R
|f (x, ξ)| ≤ a(x) + b|ξ|p/r
where b is a fixed nonnegative number, a ∈ Lr (Ω) is nonnegative and 1 < p, r < ∞, then
the above results are valid with q replaced by r. (i.e., N : Lp (Ω) → Lr (Ω).)
(c) We say that f satisfies the Caratheodory property, written f ∈ Car, if (i) and
(ii) above are met. If in addition (iii) is met, then we write f ∈ Car(p).

1.5.3. Differentiability. Let S be an open subset of the Banach space X. The functional
f : S ⊂ X → R is said to be Gateaux differentiable (G-diff) at a point u ∈ S if there
exists a functional g ∈ X* (often denoted by f 0 (u)) such that

d f (u + tv) − f (u)
= [f 0 (u)]v for all v ∈ X.

f (u + tv) = lim
dt t=0 t→0 t
The functional f 0 (u)
is called the Gateaux derivative of f at the point u ∈ S. If f is
G-diff at each point of S, the map f 0 : S ⊂ X → X* is called the Gateaux derivative of
f on S. In addition, if f 0 is continuous at u (in the operator norm), then we say that f
is C 1 at u. Note that in the case of a real-valued function of several real variables, the
Gateaux derivative is nothing more than the directional derivative of the function at u in
the direction v.

Let X, Y be Banach spaces and let A : S ⊂ X → Y be an arbitrary operator. A is

said to be Frechet differentiable (F-diff) at the point u ∈ S if there exists an operator
14 1. Preliminaries

B ∈ B(X, Y ) such that

lim kA(u + v) − Au − Bvk/kvk = 0.
kvk→0

The operator B, often denoted by A0 (u), is called the Frechet derivative of A at u. Note
that if A is Frechet differentiable on S, then A0 : S → B(X, Y ). In addition, if A0 is
continuous at u (in the operator norm), we say that A is C 1 at u.
Remark. If the functional f is F-diff at u ∈ S, then it is also G-diff at u, and the two
derivatives are equal. This follows easily from the definition of the Frechet derivative. The
converse is not always true as may be easily verified by simple examples from several variable
calculus. However, if the Gateaux derivative exists in a neighborhood of u and if f ∈ C 1 at
u, then the Frechet derivative exists at u, and the two derivatives are equal.

Example 1.31. (a) Let f (ξ) ∈ C(R). Then for k ≥ 0, the corresponding Nemytskii operator
N : C k (Ω̄) → C(Ω̄) is bounded and continuous. If in addition f (ξ) ∈ C 1 (R), then N ∈ C 1
and the Frechet derivative N 0 (u) is given by
[N 0 (u)v](x) = f 0 (u(x))v(x).

Note that for u, v ∈ C k (Ω̄), |N 0 (u)v|0 ≤ |f 0 (u)|0 |v|k and so N 0 (u) ∈ B(C k (Ω̄), C ( Ω̄)) with
kN 0 (u)k ≤ |f 0 (u)|0 . Clearly N 0 (u) is continuous at each point u ∈ C k (Ω̄). Moreover,
Z 1
0 d
|N (u + v) − N u − N (u)v|0 = sup | [ f (u(x) + tv(x)) − f 0 (u(x))v(x)]dt|
x 0 dt
Z 1
≤ |v|0 sup |f 0 (u(x) + tv(x)) − f 0 (u(x))|dt.
x 0

The last integral tends to zero since f0 is uniformly continuous on compact subsets of R.
More generally, let f (ξ) ∈ C k (R). Then the corresponding Nemytskii operator N :
C k (Ω̄)
→ C k (Ω̄) is bounded and continuous. If in addition f (ξ) ∈ C k+1 (R), then N ∈ C 1
with Frechet derivative given by [N 0 (u)v](x) = f 0 (u(x))v(x). Note that |uv|k ≤ |u|k |v|k for
u, v ∈ C k (Ω̄), and since C k (Ω̄) ⊂ C(Ω̄), the Frechet derivative must be of the stated form.

(b) Let f (ξ) ∈ C k+1 (R), where k > n/2. Then we claim that the corresponding Ne-
mytskii operator N : H k (Ω) → H k (Ω) is of class C 1 with Frechet derivative given by
[N 0 (u)v](x) = f 0 (u(x))v(x).
First, suppose u ∈ C k (Ω̄). Then N (u) ∈ C k (Ω̄) by the usual chain rule. If u ∈ H k (Ω),
let um ∈ C k (Ω̄) with kum − ukk,2 → 0. Since the imbedding H k (Ω) ⊂ C(Ω̄) is continuous,
um → u uniformly, and thus f (um ) → f (u) and f 0 (um ) → f 0 (u) uniformly and hence in L2 .
Furthermore, Di f (um ) = f 0 (um )Di um → f 0 (u)Di u in L1 . Consequently, by Theorem 2.10,
we have
Di f (u) = f 0 (u)Di u.
In a similar fashion we find
Dij f (u) = f 00 (u)Di uDj u + f 0 (u)Dij u
with corresponding formulas for higher derivatives.

1.5.4. Implicit Function Theorem. The following lemmas are needed in the proof of
the implicit function theorem.
1.5. Nonlinear Functional Analysis 15

Lemma 1.32. Let S be a closed nonempty subset of the Banach space X and let M be a
metric space. Suppose A(x, λ) : S × M → S is continuous and there is a constant k < 1
such that, uniformly for all λ ∈ M
kA(x, λ) − A(y, λ)k ≤ kkx − yk for all x, y ∈ S.
Then for each λ ∈ M, A(x, λ) has a unique fixed point x(λ) ∈ S and moreover, x(λ)
depends continuously on λ.

Proof. The existence and uniqueness of the fixed point x(λ) is of course a consequence of
the contraction mapping theorem. To prove continuity, suppose λn → λ. Then
kx(λn ) − x(λ)k = kA(x(λn ), λn ) − A(x(λ), λ)k
≤ kA(x(λn ), λn ) − A(x(λ), λn )k + kA(x(λ), λn ) − A(x(λ), λ)k
≤ kkx(λn ) − x(λ)k + kA(x(λ), λn ) − A(x(λ), λ)k.
Therefore
1
kx(λn ) − x(λ)k ≤ kA(x(λ), λn ) − A(x(λ), λ)k.
1−k
By the assumed continuity of A, the right side tends to zero as n → ∞, and therefore
x(λn ) → x(λ). 
Lemma 1.33. Suppose X, Y are Banach spaces. Let S ⊂ X be convex and assume A :
S → Y is Frechet differentiable at every point of S. Then
kAu − Avk ≤ ku − vk sup kA0 (w)k.
w∈S

In other words, A satisfies a Lipschitz condition with constant q = supw∈S kA0 (w)k.

Proof. For fixed u, v ∈ S, set g(t) = A(u + t(v − u)), where t ∈ [0, 1]. Using the definition
of Frechet derivative, we have
 
0 A(u + (t + h)(v − u)) − A(u + t(v − u))
g (t) = lim
h→0 h
 0 
hA (u + t(v − u))(v − u) + kh(v − u)kE
= lim
h→0 h
0
= A (u + t(v − u))(v − u).
Hence
kg(0) − g(1)k = kAu − Avk ≤ sup kg 0 (t)k
t∈[0,1]
which implies the desired result. 
Lemma 1.34. Let X be a Banach space. Suppose A : B(u0 , r) ⊂ X → X is a contraction,
with Lipschitz constant q < 1, where
r ≥ (1 − q)−1 kAu0 − u0 k.
Then A has a unique fixed point u ∈ B(u0 , r).

Proof. For u ∈ B(u0 , r)

kAu − u0 k ≤ kAu − Au0 k + kAu0 − u0 k ≤ qku − u0 k + (1 − q)r.
Since ku − u0 k ≤ r, A maps the ball B(u0 , r) into itself, and the result follows from the
contraction mapping theorem. 
16 1. Preliminaries

We now consider operator equations of the form A(u, v) = 0, where A maps a subset
of X × Y into Z. For a given [u0 , v0 ] ∈ X × Y we denote the Frechet derivative of A (at
[u0 , v0 ]) with respect to the first (second) argument by Au (u0 , v0 ) (Av (u0 , v0 )).
Theorem 1.35. (Implicit Function) Let X, Y, Z be Banach spaces. For a given [u0 , v0 ] ∈
X × Y and a, b > 0, let S = {[u, v] : ku − u0 k ≤ a, kv − v0 k ≤ b}. Suppose A : S → Z
satisfies the following:
(i) A is continuous.
(ii) Av (·, ·) exists and is continuous in S (in the operator norm)
(iii) A(u0 , v0 ) = 0.
(iv) [Av (u0 , v0 )]−1 exists and belongs to B(Z, Y ).
Then there are neighborhoods U of u0 and V of v0 such that the equation A(u, v) = 0 has
exactly one solution v ∈ V for every u ∈ U . The solution v depends continuously on u.

Proof. If in S we define
B(u, v) = v − [Av (u0 , v0 )]−1 A(u, v)
it is clear that the solutions of A(u, v) = 0 and v = B(u, v) are identical. The theorem will
be proved by applying the contraction mapping theorem to B. Since
Bv (u, v) = I − [Av (u0 , v0 )]−1 Av (u, v)
Bv (·, ·) is continuous in the operator norm. Now Bv (u0 , v0 ) = 0, so for some δ > 0 there is
a q < 1 such that
kBv (u, v)k ≤ q
for ku − u0 k ≤ δ, kv − v0 k ≤ δ. By virtue of Lemma 1.33, B(u, ·) is a contraction. Since
A is continuous, B is also continuous. Therefore, since B(u0 , v0 ) = v0 , there is an ε with
0 < ε ≤ δ such that
kB(u, v0 ) − v0 k ≤ (1 − q)δ
for ku − u0 k ≤ ε. The existence of a unique fixed point in the closed ball B(v0 , δ) follows
from Lemma 1.34 and the continuity from Lemma 1.32. 
Example 1.36. Let f (ξ) ∈ C 1,α (R), f (0) = f 0 (0) = 0, g(x) ∈ C α (Ω̄) and consider the
boundary value problem
(1.13) ∆u + f (u) = g(x) in Ω, u|∂Ω = 0.
Set X = Z = C α (Ω̄), Y = {u ∈ C 2,α (Ω̄) : u|∂Ω = 0} and
A(g, u) = ∆u + N (u) − g
where N is the Nemytskii operator corresponding to f . The operator A maps X × Y into
the space Z. Clearly A(0, 0) = 0 (A is C 1 by earlier examples) and
Au (0, 0)v = ∆v, v ∈ Y.
It is easily checked that all the conditions of the implicit function theorem are met. In
particular, condition (iv) is a consequence of the bounded inverse theorem. Thus, for a
function g ∈ C α (Ω̄) of sufficiently small norm (in the space C α (Ω̄)) there exists a unique
solution of (1.13) which lies near the zero function. There may, of course, be other solutions
which are not close to the zero function. (Note that the condition f 0 (0) = 0 rules out linear
functions.)
1.5. Nonlinear Functional Analysis 17

Remark. Note that the choice of X = Z = C(Ω̄), Y = {u ∈ C 2 (Ω̄) : u|∂Ω = 0} would fail
above since the corresponding linear problem is not onto. An alternate approach would be
to use Sobolev spaces. In fact, if we take X = Z = W k−2 (Ω), Y = W k (Ω) ∩ H01 (Ω) with k
sufficiently large, and if f (ξ) ∈ C k+1 (R), then as above, we can conclude the existence of a
unique solution u ∈ W k (Ω) provided kgkk−2,2 is sufficiently small. Hence, we get existence
for more general functions g; however, the solution u ∈ W k (Ω) is not a classical (i.e., C 2 )
solution in general.

1.5.5. Generalized Weierstrass Theorem. In its simplest form, the classical Weier-
strass theorem can be stated as follows: Every continuous function defined on a closed ball
in Rn is bounded and attains both its maximum and minimum on this ball. The proof
makes essential use of the fact that the closed ball is compact.
The first difficulty in trying to extend this result to an arbitrary Banach space X is
that the closed ball in X is not compact if X is infinite dimensional. However, as we shall
show, a generalized Weierstrass theorem is possible if we require a stronger property for the
functional.
A set S ⊂ X is said to be weakly closed if {un } ⊂ S, un * u implies u ∈ S, i.e.,
S contains all its weak limits. A weakly closed set is clearly closed, but not conversely.
Indeed, the set {sin nx}∞ 2
1 in L (0, π) has no limit point (because it cannot be Cauchy) so
it is closed, but zero is a weak limit that does not belong to the set. It can be shown that
every convex, closed set in a Banach space is weakly closed.
A functional f : S ⊂ X → R is weakly continuous at u0 ∈ S if for every sequence
{un } ⊂ S with un * u0 it follows that f (un ) → f (u0 ). Clearly, every functional f ∈ X*
is weakly continuous. A functional f : S ⊂ X → R is weakly lower semicontinu-
ous(w.l.s.c.) at u0 ∈ S if for every sequence {un } ⊂ S for which un * u0 it follows that
f (u0 ) ≤ lim inf n→∞ f (un ). According to Theorem 1.18, the norm on a Banach space is
w.l.s.c.. A functional f : S ⊂ X → R is weakly coercive on S if f (u) → ∞ as kuk → ∞
on S.
Theorem 1.37. Let X be a reflexive Banach space and f : C ⊂ X → R be w.l.s.c. and
assume
(i) C is a nonempty bounded weakly closed set in X or
(ii) C is a nonempty weakly closed set in X and f is weakly coercive on C.
Then
(a) inf u∈C f (u) > −∞;
(b) there is at least one u0 ∈ C such that f (u0 ) = inf u∈C f (u).
Moreover, if u0 is an interior point of C and f is G-diff at u0 , then f 0 (u0 ) = 0.

Proof. Assume (i) and let {un } ⊂ C be a minimizing sequence, i.e., limn→∞ f (un ) =
inf u∈C f (u). The existence of such a sequence follows from the definition of inf. Since X is
reflexive and C is bounded and weakly closed, there is a subsequence {un0 } and a u0 ∈ C
such that un0 * u0 . But f is w.l.s.c. and so f (u0 ) ≤ lim inf n→∞ f (un0 ) = inf u∈C f (u),
which proves (a). Since by definition, f (u0 ) ≥ inf u∈C f (u), we get (b).
Assume (ii) and fix u0 ∈ C. Since f is weakly coercive, there is a closed ball B(0, R) ⊂ X
such that u0 ∈ B ∩ C and f (u) ≥ f (u0 ) outside B ∩ C. Since B ∩ C satisfies the conditions
of (i), there is a u1 ∈ B ∩ C such that f (u) ≥ f (u1 ) for all u ∈ B ∩ C and in particular for
u0 . Thus, f (u) ≥ f (u1 ) on all of C.
18 1. Preliminaries

To prove the last statement we set ϕv (t) = f (u0 + tv). For fixed v ∈ X, ϕv (t) has a
local minimum at t = 0, and therefore hf 0 (u0 ), vi = 0 for all v ∈ X. 

The point u0 ∈ X is called a critical point of the functional f defined on X if f 0 (u0 )v =

0 for every v ∈ X.

Remark. Even though weakly continuous functionals on closed balls attain both their inf
and sup (which follows from the above theorem), the usual functionals that we encounter
are not weakly continuous, but are w.l.s.c.. Hence this explains why we seek the inf and
not the sup in variational problems.
1.5.5.1. Convex sets. A set C in the real normed space X is called convex if (1−t)u+tv ∈ C
for all t ∈ [0, 1], u, v ∈ C. The following result is needed later (see, e.g., [32]).
Theorem 1.38. A closed convex set in a Banach space is weakly closed.

1.5.6. Monotone Operators and Convex Functionals. Let A : X → X* be an oper-

ator on the real Banach space X.

A is monotone if
hAu − Av, u − vi ≥ 0 for all u, v ∈ X.

A is strongly monotone if
hAu − Av, u − vi ≥ cku − vkpX for all u, v ∈ X
where c > 0 and p > 1.

A is coercive if
hAu, ui
lim = +∞.
kuk→∞ kuk

Remark. A strongly monotone operator is coercive. This follows immediately from hAu, ui =
hAu − A0, ui + hA0, ui ≥ ckukpX − kA0kkukX .

Let C be a convex set in the real normed space X. A functional f : C ⊂ X → R is said

to be convex if
f ((1 − t)u + tv) ≤ (1 − t)f (u) + tf (v) for all t ∈ [0, 1], u, v ∈ C.
In the following we set
ϕ(t) = f ((1 − t)u + tv) = f (u + t(v − u))
for fixed u and v.
Lemma 1.39. Let C ⊂ X be a convex set in a real normed space X. Then the following
statements are equivalent:
(a) The real function ϕ : [0, 1] → R is convex for all u, v ∈ C.
(b) The functional f : C ⊂ X → R is convex.
(c) f 0 : C ⊂ X → X* (assuming f is G-diff on C) is monotone.
1.5. Nonlinear Functional Analysis 19

Proof. Assume ϕ is convex. Then

ϕ(t) = ϕ((1 − t) · 0 + t · 1) ≤ (1 − t)ϕ(0) + tϕ(1)
for all t ∈ [0, 1], which implies (b).
Similarly, if f is convex, then for t = (1 − α)s1 + αs2 , with α, s1 , s2 ∈ [0, 1], we have
ϕ(t) = f (u + t(v − u)) ≤ (1 − α)f (u + s1 (v − u)) + αf (u + s2 (v − u))
for all u, v ∈ C, which implies (a).
Fix u, v ∈ C. Then ϕ0 (t) = hf 0 (u + t(v − u)), v − ui. If f is convex, then ϕ is convex
and therefore ϕ0 is monotone. From ϕ0 (1) ≥ ϕ0 (0) we obtain
hf 0 (v) − f 0 (u), v − ui ≥ 0 for all u, v ∈ C
which implies (c).
Finally, assume f 0 is monotone. Then for s < t we have
1
ϕ0 (t) − ϕ0 (s) = hf 0 (u + t(v − u)) − f 0 (u + s(v − u)), (t − s)(v − u)i ≥ 0.
t−s
Thus ϕ0 is monotone, which implies ϕ, and thus f is convex. 
Theorem 1.40. Consider the functional f : C ⊂ X → R, where X is a real Banach space.
Then f is w.l.s.c. if any one of the following conditions holds:
(a) C is closed and convex; f is convex and continuous.
(b) C is convex; f is G-diff on C and f 0 is monotone on C.

Proof. Set
Cr = {u ∈ C : f (u) ≤ r}.
It follows from (a) that Cr is closed and convex for all r, and thus is weakly closed (cf.
Theorem 1.38). If f is not w.l.s.c., then there is a sequence {un } ⊂ C with un * u and
f (u) > lim inf f (un ). Hence, there is an r and a subsequence {un0 } such that f (u) > r and
f (un0 ) ≤ r (i.e., un0 ∈ Cr ) for all n0 large enough. Since Cr is weakly closed, u ∈ Cr , which
is a contradiction.
Assume (b) holds and set ϕ(t) = f (u + t(v − u)). Then by Lemma 1.39, ϕ : [0, 1] → R
is convex and ϕ0 is monotone. By the classical mean value theorem,
ϕ(1) − ϕ(0) = ϕ0 (θ) ≥ ϕ0 (0), 0<θ<1
i.e.,
f (v) ≥ f (u) + hf 0 (u), v − ui for all u, v ∈ C.
If un * u, then hf 0 (u), un − ui → 0 as n → ∞. Hence, f is w.l.s.c. 

1.5.7. Lagrange Multipliers. Let f, g : X → R be two functionals defined on the Banach

space X and let
Mc = {u ∈ X : g(u) = c}
for a given constant c. A point u0 ∈ Mc is called an extreme of f with respect to Mc if
there exists a neighborhood of u0 , U (u0 ) ⊂ X, such that
f (u) ≤ f (u0 ) for all u ∈ U (u0 ) ∩ Mc
or
f (u) ≥ f (u0 ) for all u ∈ U (u0 ) ∩ Mc .
20 1. Preliminaries

In the first case we say that f is (local) maximal at u0 with respect to Mc , while in the
second case f is (local) minimal at u0 with respect to Mc . A point u0 ∈ Mc is called an
ordinary point of the manifold Mc if its F-derivative g 0 (u0 ) 6= 0.
Let u0 be an ordinary point of Mc . Then u0 is called a critical point of f with respect
to Mc if there exists a real number λ, called a Lagrange multiplier, such that
f 0 (u0 ) = λg 0 (u0 ).
As we shall see, if u0 is an extremum of f with respect to Mc , and if u0 is an ordinary point,
then u0 is a critical point of f with respect to Mc . Note that if u0 is an extremum of f with
respect to X, then we can choose λ = 0, which implies the usual result.
Lemma 1.41. Let X be a Banach space. Suppose the following hold:
(i) f, g : X → R are of class C 1
(ii) For u0 ∈ X, we can find v, w ∈ X such that

(1.14) f 0 (u0 )v · g 0 (u0 )w 6= f 0 (u0 )w · g 0 (u0 )v.

Then f cannot have a local extremum with respect to the level set Mc at u0 .

Proof. Fix v, w ∈ X, and for s, t ∈ R consider the real-valued functions

F (s, t) = f (u0 + sv + tw), G(s, t) = g(u0 + sv + tw) − c.
Then
∂F ∂F
(0, 0) = f 0 (u0 )v, (0, 0) = f 0 (u0 )w
∂s ∂t
∂G ∂G
(0, 0) = g 0 (u0 )v, (0, 0) = g 0 (u0 )w
∂s ∂t
so that condition (1.14) is simply that the Jacobian |∂(F, G)/∂(s, t)| is nonvanishing at
(s, t) = (0, 0). Since F, G ∈ C 1 on R2 , we may apply the implicit function theorem to
conclude that a local extremum cannot occur at u0 . More precisely, assume w.l.o.g. that
Gt (0, 0) 6= 0. Since G(0, 0) = 0, the implicit function theorem implies the existence of a C 1
function φ such that φ(0) = 0 and G(s, φ(s)) = 0 for sufficiently small s. Moreover,
Gs (0, 0)
φ0 (0) = − .
Gt (0, 0)
Set z(s) = F (s, φ(s)) = f (u0 + sv + φ(s)w) and note that g(u0 + sv + φ(s)w) = c. Hence,
if to the contrary f has an extremum at u0 , then z(s) has a local extremum at s = 0. But,
an easy computation shows that Gt (0, 0)z 0 (0) = f 0 (u0 )v · g 0 (u0 )w − f 0 (u0 )w · g 0 (u0 )v 6= 0,
which is a contradiction. 
Theorem 1.42. (Lagrange) Let X be a Banach space. Suppose the following hold:
(i) f, g : X → R are of class C 1
(ii) g(u0 ) = c.
(iii) u0 is a local extremum of f with respect to the constraint Mc
Then either
(a) g 0 (u0 )v = 0 for all v ∈ X, or
(b) There exists λ ∈ R such that f 0 (u0 )v = λg 0 (u0 )v for all v ∈ X.
1.5. Nonlinear Functional Analysis 21

Proof. If (a) does not hold, then fix w ∈ X with g 0 (u0 )w 6= 0. By hypothesis and the above
lemma, we must have
f 0 (u0 )v · g 0 (u0 )w = f 0 (u0 )w · g 0 (u0 )v for all v ∈ X.
If we define λ = (f 0 (u0 )w)/(g 0 (u0 )w), then we obtain (b). 

More generally, one can prove the following:

Theorem 1.43. (Ljusternik) Let X be a Banach space. Suppose the following hold:
(i) g0 : X → R is of class C 1
(ii) gi : X → R are of class C 1 , i = 1, . . . , n
(iii) u0 is an extremum of g0 with respect to the constraint C:
C = {u : gi (u) = ci (i = 1, . . . , n)}
where the ci are constants.
Then there are numbers λi (not all zero) such that
n
X
(1.15) λi gi0 (u0 ) = 0.
i=0

As an application of Ljusternik’s theorem we have

Theorem 1.44. Let f, g : X → R be C 1 functionals on the reflexive Banach space X.
Suppose
(i) f is w.l.s.c. and weakly coercive on X ∩ {g(u) ≤ c}
(ii) g is weakly continuous
(iii) g(0) = 0, g 0 (u) = 0 only at u = 0.
Then the equation f 0 (u) = λg 0 (u) has a one parameter family of nontrivial solutions (uR , λR )
for all R 6= 0 in the range of g(u) and g(uR ) = R. Moreover, uR can be characterized as
the function which minimizes f (u) over the set g(u) = R.

Proof. Since g(u) is weakly continuous, it follows that MR = {u : g(u) = R} is weakly

closed. If MR is not empty, i.e., if R belongs to the range of g, then by Theorem 1.37,
there is a uR ∈ MR such that f (uR ) = inf f (u) over u ∈ MR . If R 6= 0 then it cannot
be that g 0 (uR ) = 0. Otherwise by (iii), uR = 0 and hence R = g(uR ) = 0, which is a
contradiction. Thus, by Ljusternik’s theorem, there exist constants λ1 , λ2 , λ21 + λ22 6= 0 such
that λ1 f 0 (uR ) + λ2 g 0 (uR ) = 0. Since uR is an ordinary point, it follows that λ1 6= 0, and
therefore λR = −λ2 /λ1 . 

Remark. In applying this theorem one should be careful and not choose g(u) = kuk, since
this g is not weakly continuous.

The following interpolation inequality, which is frequently referred to as Ehrling’s

inequality, will be needed in the next result.
Theorem 1.45. Let X, Y, Z be three Banach spaces such that
X ⊂ Y ⊂ Z.
22 1. Preliminaries

Assume that the imbedding X ⊂ Y is compact and that the imbedding Y ⊂ Z is continuous.
Then for each ε > 0, there is a constant c(ε) such that
(1.16) kukY ≤ εkukX + c(ε)kukZ for all u ∈ X.

Proof. If for a fixed ε > 0 the inequality is false, then there exists a sequence {un } such
that
(1.17) kun kY > εkun kX + nkun kZ for all n.
Without loss of generality we can assume kun kX = 1. Since the imbedding X ⊂ Y is
compact, there is a subsequence, again denoted by {un }, with un → u in Y . This implies
un → u in Z. By (1.17), kun kY > ε and so u 6= 0. Again by (1.17), un → 0 in Z, i.e., u = 0,
which is a contradiction. 
 Chapter 2

Sobolev Spaces

This chapter is devoted to a discussion of the necessary Sobolev function spaces which
permit a modern approach to the study of differential equations.

2.1. Weak Derivatives and Sobolev Spaces

2.1.1. Weak Derivatives. Let Ω be a nonempty open set in Rn . Suppose u ∈ C m (Ω) and
ϕ ∈ C0m (Ω). Then by integration by parts
Z Z
α |α|
(2.1) uD ϕdx = (−1) vϕdx, |α| ≤ m
Ω Ω

where v = Dα u. Motivated by (2.1), we now enlarge the class of functions for which the
notion of derivative can introduced.
Let u ∈ L1loc (Ω). A function v ∈ L1loc (Ω) is called the αth weak derivative of u if it
satisfies
Z Z
α |α| |α|
(2.2) uD ϕdx = (−1) vϕdx for all ϕ ∈ C0 (Ω).
Ω Ω

It can be easily shown that weak derivatives are unique. Thus we write v = Dα u to indicate
that v is the αth weak derivative of u. If a function u has an ordinary αth derivative lying
in L1loc (Ω), then it is clearly the αth weak derivative.

In contrast to the corresponding classical derivative, the weak derivative Dα u is defined

globally on all of Ω by (2.2). However, in every subregion Ω0 ⊂ Ω the function Dα u will
also be the weak derivative of u. It suffices to note that (2.2) holds for every function
|α|
ϕ ∈ C0 (Ω0 ), and extended outside Ω0 by assigning to it the value zero. In particular, the
weak derivative (if it exists) of a function u having compact support in Ω has itself compact
support in Ω and thus belongs to L1 (Ω).
We also note that in contrast to the classical derivative, the weak derivative Dα u is
defined at once for order |α| without assuming the existence of corresponding derivatives
of lower orders. In fact, the derivatives of lower orders may not exist as we will see in a
forthcoming exercise. However, in his book, Sobolev shows that if all weak derivatives exist
of a certain order, then all lower order weak derivatives exist.

23
24 2. Sobolev Spaces

Example 2.1. (a) The function u(x) = |x1 | has in the ball Ω = B(0, 1) weak derivatives
ux1 = sgn x1 , uxi = 0, i = 2, . . . , n. In fact, we apply formula (2.2) as follows: For any
ϕ ∈ C01 (Ω)
Z Z Z
|x1 |ϕx1 dx = x1 ϕx1 dx − x1 ϕx1 dx
Ω Ω+ Ω−
where Ω+ = Ω ∩ (x1 > 0), Ω− = Ω ∩ (x1 < 0). Since x1 ϕ = 0 on ∂Ω and also for x1 = 0,
an application of the divergence theorem yields
Z Z Z Z
|x1 |ϕx1 dx = − ϕdx + ϕdx = − (sgn x1 )ϕdx.
Ω Ω+ Ω− Ω

Hence |x1 |x1 = sgn x1 . Similarly, since for i ≥ 2

Z Z Z
|x1 |ϕxi dx = (|x1 |ϕ)xi dx = − 0ϕdx
Ω Ω Ω

|x1 |xi = 0 for i = 2, . . . , n. Note that the function |x1 | has no classical derivative with
respect to x1 in Ω.
(b) By the above computation, the function u(x) = |x| has a weak derivative u0 (x) =
sgn x on the interval Ω = (−1, 1). On the other hand, sgn x does not have a weak derivative
on Ω due to the discontinuity at x = 0.
(c) Let Ω = B(0, 1/2) ⊂ R2 and define u(x) = ln(ln(2/r)), x ∈ Ω, where r = |x| =
+ x22 )1/2 . Then u 6∈ L∞ (Ω) because of the singularity at the origin. However, we will
(x21
show that u has weak first partial derivatives.
First of all u ∈ L2 (Ω), for
Z Z 2π Z 1/2
2
|u| dx = r[ln(ln(2/r))]2 dr dθ
Ω 0 0
and a simple application of L’hopitals rule shows that the integrand is bounded and thus
the integral is finite. Similarly, it is easy to check that the classical partial derivative
− cos θ
ux1 = , where x1 = r cos θ
r ln(2/r)
also belongs to L2 (Ω). Now we show that the defining equation for the weak derivative is
met.
Let Ωε = {x : ε < r < 1/2} and choose ϕ ∈ C01 (Ω). Then by the divergence theorem
and the absolute continuity of integrals
Z Z  Z Z 
uϕx1 dx = lim uϕx1 dx = lim − ux1 ϕdx + uϕn1 ds
Ω ε→0 Ωε ε→0 Ωε r=ε

where n = (n1 , n2 ) is the unit outward normal to Ωε on r = ε. But (ds = εdθ)

Z Z 2π
| uϕn1 ds| ≤ |u| |ϕ|εdθ ≤ 2πεc ln(ln(2/ε)) → 0
r=ε 0
as ε → 0. Thus Z Z
uϕx1 dx = − ux1 ϕdx.
Ω Ω

The same analysis applies to ux2 . Thus u has weak first partial derivatives given by the
classical derivatives which are defined on Ω\{0}.
2.1. Weak Derivatives and Sobolev Spaces 25

2.1.2. Sobolev Spaces. For p ≥ 1 and k a nonnegative integer, we let

W k,p (Ω) = {u : u ∈ Lp (Ω), Dα u ∈ Lp (Ω), 0 < |α| ≤ k}
where Dα u denotes the αth weak derivative. When k = 0, W k,p (Ω) will mean Lp (Ω). It is
clear that W k,p (Ω) is a vector space.
A norm is introduced by defining
Z X
(2.3) kukk,p = kukW k,p (Ω) =( |Dα u|p dx)1/p
Ω |α|≤k

if 1 ≤ p < ∞. The space W k,p (Ω) is known as a Sobolev space of order k.

We also introduce the space W0k,p (Ω) which is defined to be the closure of the space
C0k (Ω) with respect to the norm k · kk,p . W0k,p (Ω) is also called a Sobolev space of order
k. As we shall see shortly, W k,p (Ω) 6= W0k,p (Ω) for k ≥ 1. (Unless Ω = Rn .)
Remark. The case p = 2 is special, since the spaces W k,2 (Ω), W0k,2 (Ω) will be Hilbert spaces
under the inner product
Z X
(u, v)k,2 = (u, v)W k,2 (Ω) = Dα uDα vdx.
Ω |α|≤k

Since we shall be dealing mostly with these spaces in the sequel, we introduce the special
notation:
H k (Ω) = W k,2 (Ω), H0k (Ω) = W0k,2 (Ω).
Theorem 2.2. W k,p (Ω) is a Banach space under the norm (2.3). If 1 < p < ∞, it is
reflexive and if 1 ≤ p < ∞, it is separable.

Proof. We first prove that W k,p (Ω) is complete with respect to the norm (2.3).
Let {un } be a Cauchy sequence of elements in W k,p (Ω), i.e.,
p
X Z
kun − um kk,p = |Dα un − Dα um |p dx → 0 as m, n → ∞.
|α|≤k Ω

Then for any α, |α| ≤ k, when m, n → ∞

Z
|Dα un − Dα um |p dx → 0
Ω
and, in particular, when |α| = 0
Z
|un − um |p dx → 0.
Ω
Since Lp (Ω) is complete, it follows that there are functions uα ∈ Lp (Ω), |α| ≤ k such that
Dα un → uα (in Lp (Ω)). Since each un (x) has weak derivatives (up to order k) belonging
to Lp (Ω), a simple limit argument shows that uα is the αth weak derivative of u0 . In fact,
Z Z Z Z
α α |α| α |α|
uD ϕdx ← un D ϕdx = (−1) ϕD un dx → (−1) uα ϕdx.
Ω Ω Ω Ω

Hence u0 ∈ W k,p (Ω) and kun − u0 kk,p → 0 as n → ∞.

Consider the map T : W 1,p (Ω) → (Lp (Ω))n+1 defined by
T u = (u, D1 u, . . . , Dn u).
26 2. Sobolev Spaces

If we endow the latter space with the norm

n+1
X
kvk = ( kvi kpp )1/p
i=1

(Lp (Ω))n+1 ,
for v = (v1 , . . . , vn+1 ) ∈ then T is a (linear) isometry. Now (Lp (Ω))n+1 is
reflexive for 1 < p < ∞ and separable for 1 ≤ p < ∞. Since W 1,p (Ω) is complete, its image
under the isometry T is a closed subspace of (Lp (Ω))n+1 which inherits the corresponding
properties as does W 1,p (Ω) (see Theorem 1.14). Similarly, we can handle the case k ≥ 2. 
Example 2.3. Let Ω be a bounded open connected set in Rn . Divide Ω into N open disjoint
subsets Ω1 , Ω2 , . . . , ΩN . Suppose the function u : Ω → R has the following properties:
(i) u is continuous on Ω̄.
(ii) For some i, Di u is continuous on Ω1 , Ω2 , . . . , ΩN , and can be extended contin-
uously to Ω̄1 , Ω̄2 , . . . , Ω̄N , respectively.
(iii) The surfaces of discontinuity are such that the divergence theorem applies.
Define wi (x) = Di u(x) if x ∈ ∪N
i=1 Ωi . Otherwise, wi can be arbitrary. We now claim
that wi ∈ Lp (Ω) is a weak partial derivative of u on Ω.
Indeed, for all ϕ ∈ C01 (Ω), the divergence theorem yields
Z XZ
uDi ϕdx = uDi ϕdx
Ω j Ωj
!
X Z Z
= uϕni dS − ϕDi udx
j ∂Ωj Ωj
Z
= − ϕDi udx.
Ω
Note that the boundary terms either vanish, since ϕ has compact support, or cancel out
along the common boundaries, since u is continuous and the outer normals have opposite
directions. Similarly, if u ∈ C k (Ω̄) and has piecewise continuous derivatives in Ω of order
k + 1, then u ∈ W k+1,p (Ω).

Remark. More generally, by using a partition of unity argument, we can show the following:
If O is a collection of nonempty open sets whose union is Ω and if u ∈ L1loc (Ω) is such that
for some multi-index α, the αth weak derivative of u exists on each member of O, then the
αth weak derivative of u exists on Ω.

Exercise 1. Consider the function u(x) = sgnx1 + sgnx2 in the ball B(0, 1) ⊂ R2 .
Show that the weak derivative ux1 does not exist, yet the weak derivative ux1 x2 does exist.
Exercise 2. (a) Let Ω be the hemisphere of radius R < 1 in Rn :
n
X
r2 ≡ x2i ≤ R2 , xn ≥ 0 (n ≥ 3).
i=1
Show that
u = (r(n/2−1) `nr)−1 ∈ H 1 (Ω).
(b) Let B = B(0, 1) be the open unit ball in Rn , and let
u(x) = |x|−α , x ∈ B.
2.2. Approximations and Extensions 27

For what values of α, n, p does u belong to W 1,p (B)?

2.2. Approximations and Extensions

2.2.1. Mollifiers. Let x ∈ Rn and let B(x, h) denote the open ball with center at x and
radius h. For each h > 0, let ωh (x) ∈ C ∞ (Rn ) satisfy
ωh (x) ≥ 0; ωh (x) = 0 for |x| ≥ h
Z Z
ωh (x)dx = ωh (x)dx = 1.
Rn B(0,h)
Such functions are called mollifiers. For example, let
(
k exp [(|x|2 − 1)−1 ], |x| < 1,
ω(x) =
0, |x| ≥ 1,
R
where k > 0 is chosen so that Rn ω(x) dx = 1. Then, a family of mollifiers can be taken as
ωh (x) = h−n ω(x/h) for h > 0.
Let Ω be a nonempty open set in Rn and let u ∈ L1 (Ω). We set u = 0 outside Ω. Define
for each h > 0 the mollified function
Z
uh (x) = ωh (x − y)u(y)dy
Ω
where ωh is a mollifier.
There are two other forms in which uh can be represented, namely
Z Z
(2.4) uh (x) = ωh (x − y)u(y)dy = ωh (x − y)u(y)dy
Rn B(x,h)

the latter equality being valid since ωh vanishes outside the (open) ball B(x, h). Thus
the values of uh (x) depend only on the values of u on the ball B(x, h). In particular, if
dist(x, supp(u)) ≥ h, then uh (x) = 0.
Theorem 2.4. Let Ω be a nonempty open set in Rn . Then
(a) uh ∈ C ∞ (Rn ).
(b) If supp(u) is a compact subset of Ω, then uh ∈ C0∞ (Ω) for all h sufficiently
small.

Proof. Since u is integrable and ωh ∈ C ∞ , the Lebesgue theorem on differentiating integrals

implies that for |α| < ∞
Z
α
D uh (x) = u(y)Dα ωh (x − y)dy
Ω
i.e., uh ∈ C ∞ (Rn ). Statement (b) follows from the remark preceding the theorem. 

With respect to a bounded set Ω we construct another set Ω(h) as follows: with each
point x ∈ Ω as center, draw a ball of radius h; the union of these balls is then Ω(h) . Clearly
Ω(h) ⊃ Ω. Moreover, uh can be different from zero only in Ω(h) .
Corollary 2.5. Let Ω be a nonempty bounded open set in Rn and let h > 0 be any number.
Then there exists a function η ∈ C ∞ (Rn ) such that
0 ≤ η(x) ≤ 1; η(x) = 1, x ∈ Ω(h) ; η(x) = 0, x ∈ (Ω(3h) )c .
Such a function is called a cut-off function for Ω.
28 2. Sobolev Spaces

Proof. Let χ(x) be the characteristic function of the set Ω(2h) : χ(x) = 1 for x ∈ Ω(2h) , χ(x) =
0 for x 6∈ Ω(2h) and set
Z
η(x) ≡ χh (x) = ωh (x − y)χ(y)dy.
Rn
Then Z
η(x) = ωh (x − y)dy ∈ C ∞ (Rn ),
Ω(2h)
Z
0 ≤ η(x) ≤ ωh (x − y)dy = 1,
Rn
and
(R
B(x,h) ωh (x − y)dy = 1, x ∈ Ω(h) ,
Z
η(x) = ωh (x − y)χ(y)dy =
B(x,h) 0, x ∈ (Ω(3h) )c .
In particular, we note that if Ω0 ⊂⊂ Ω, there is a function η ∈ C0∞ (Ω) such that η(x) = 1
for x ∈ Ω0 , and 0 ≤ η(x) ≤ 1 in Ω. 

Remark. Henceforth, the notation Ω0 ⊂⊂ Ω means that Ω0 , Ω are open sets and that Ω0 ⊂ Ω.

We shall have need of the following well-known result.

Theorem 2.6. (Partition of Unity) Assume Ω ⊂ Rn is bounded and Ω ⊂⊂ ∪N i=1 Ωi ,
where each Ωi is open. Then there exist C ∞ functions ψi (x)(i = 1, . . . , N ) such that
(a) 0 ≤ ψi (x) ≤ 1
(b) ψi has its support in Ωi
PN
(c) i=1 ψi (x) = 1 for every x ∈ Ω.

2.2.2. Approximation Theorems.

Lemma 2.7. Let Ω be a nonempty bounded open set in Rn . Then every u ∈ Lp (Ω) is
p-mean continuous, i.e.,
Z
|u(x + z) − u(x)|p dx → 0 as z → 0.
Ω

Proof. Choose a > 0 large enough so that Ω is strictly contained in the ball B(0, a). Then
the function 
u(x) if x ∈ Ω,
U (x) =
0 if x ∈ B(0, 2a) \ Ω
belongs to Lp (B(0, 2a)). For ε > 0, there is a function Ū ∈ C(B̄(0, 2a)) which satisfies the
inequality kU − Ū kLp (B(0,2a)) < ε/3. By multiplying Ū by an appropriate cut-off function,
it can be assumed that Ū (x) = 0 for x ∈ B(0, 2a)/B(0, a).
Therefore for |z| ≤ a, kU (x + z) − Ū (x + z)kLp (B(0,2a)) = kU (x) − Ū (x)kLp (B(0,a)) ≤ ε/3.
Since the function Ū is uniformly continuous in B(0, 2a), there is a δ > 0(δ < a) such that
kŪ (x + z) − Ū (x)kLp (B(0,2a)) ≤ ε/3 whenever |z| < δ. Hence for |z| < δ we easily see that
ku(x + z) − u(x)kLp (Ω) = kU (x + z) − U (x)kLp (B(0,2a)) ≤ ε. 
Theorem 2.8. Let Ω be a nonempty open set in Rn . If u ∈ Lp (Ω) (1 ≤ p < ∞), then
(a) kuh kp ≤ kukp
(b) kuh − ukp → 0 as h → 0.
2.2. Approximations and Extensions 29

If u ∈ C k (Ω̄), where Ω̄ is compact, then

(c) kuh − ukC k (Ω̄0 ) → 0 as h → 0,
where Ω0 ⊂⊂ Ω.
1/p 1/q
Proof. If 1 < p < ∞, let q = p/(p − 1). Then ωh = ωh ωh and Hölder’s inequality
implies
Z Z p/q
p p
|uh (x)| ≤ ωh (x − y)|u(y)| dy ωh (x − y)dy
Ω Ω
Z
≤ ωh (x − y)|u(y)|p dy
Ω
which obviously holds also for p = 1. An application of Fubini’s Theorem gives
Z Z Z  Z
p p
|uh (x)| dx ≤ ωh (x − y)dx |u(y)| dy ≤ |u(y)|p dy
Ω Ω Ω Ω
which implies (a).
To prove (b), let ω(x) = hn ωh (hx). Then ω(x) ∈ C ∞ (Rn ) and satisfies
ω(x) ≥ 0; ω(x) = 0 for |x| ≥ 1
Z Z
ω(x)dx = ω(x)dx = 1.
Rn B(0,1)

Using the change of variable z = (x − y)/h we have

Z
uh (x) − u(x) = [u(y) − u(x)]ωh (x − y)dy
B(x,h)
Z
= [u(x − hz) − u(x)]ω(z)dz.
B(0,1)

Hence by Hölder’s inequality

Z
p
|uh (x) − u(x)| ≤ d |u(x − hz) − u(x)|p dz
B(0,1)

and so by Fubini’s Theorem

Z Z Z
p
|uh (x) − u(x)| dx ≤ d ( |u(x − hz) − u(x)|p dx)dz.
Ω B(0,1) Ω

The right-hand side goes to zero as h → 0 since every u ∈ Lp (Ω) is p-mean continuous.
We now prove (c) for k = 0. Let Ω0 , Ω00 be such that Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. Let h0 be the
shortest distance between ∂Ω0 and ∂Ω00 . Take h < h0 . Then
Z
uh (x) − u(x) = [u(y) − u(x)]ωh (x − y)dy.
B(x,h)

If x ∈ Ω̄0 , then in the above integral y ∈ Ω̄00 . Now u is uniformly continuous in Ω̄00 and
ωh ≥ 0, and therefore for an arbitrary ε > 0 we have
Z
|uh (x) − u(x)| ≤ ε ωh (x − y)dy = ε
B(x,h)

provided h is sufficiently small. The case k ≥ 1 is handled similarly and is left as an

exercise. 
30 2. Sobolev Spaces

Remark. The following example shows that in (c) we cannot replace Ω0 by Ω. Let u ≡ 1 for
R1 Rh
x ∈ [0, 1] and consider uh (x) = 0 ωh (x − y)dy, where ωh (y) = ωh (−y). Now −h ωh (y)dy =
1 and so uh (0) = 1/2 for all h < 1. Thus uh (0) → 1/2 6= 1 = u(0). Moreover, for x ∈ (0, 1)
R x+h
and h sufficiently small, (x − h, x + h) ⊂ (0, 1) and so uh (x) = x−h ωh (x − y)dy = 1 which
implies uh (x) → 1 for all x ∈ (0, 1).
Corollary 2.9. Let Ω be a nonempty open set in Rn . Then C0∞ (Ω) is dense in Lp (Ω) for
all 1 ≤ p < ∞.

Proof. Suppose first that Ω is bounded and let Ω0 ⊂⊂ Ω. For a given u ∈ Lp (Ω) set
u(x), x ∈ Ω0

v(x) =
0, x ∈ Ω\Ω0 .
Then Z Z
p
|u − v| dx = |u|p dx.
Ω Ω\Ω0
By the absolute continuity of integrals, we can choose Ω0 so that the integral on the right
is arbitrarily small, i.e., ku − vkp < ε/2.
Since supp(v) is a compact subset of Ω, Theorems 2.4(b) and 2.8(b) imply that for h
sufficiently small, vh (x) ∈ C0∞ (Ω) with kv − vh kp < ε/2, and therefore ku − vh kp < ε.
If Ω is unbounded, choose a ball B large enough so that
Z
|u|p dx < ε/2
Ω\Ω0
where Ω0 = Ω ∩ B, and repeat the proof just given. 

We now consider the following local approximation theorem.

Theorem 2.10. Let Ω be a nonempty open set in Rn and suppose u, v ∈ L1loc (Ω). Then v =
Dα u iff there exists a sequence of C ∞ (Ω) functions {uh } with kuh − ukL1 (S) → 0, kDα uh −
vkL1 (S) → 0 as h → 0, for all compact sets S ⊂ Ω.

Proof. (Necessity) Suppose v = Dα u. Let S ⊂ Ω, and choose d > 0 small enough so that
the sets Ω0 ≡ S (d/2) , Ω00 ≡ S (d) satisfy Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. For x ∈ Rn define
Z Z
uh (x) = ωh (x − y)u(y)dy, vh (x) = ωh (x − y)v(y)dy.
Ω00 Ω00
Clearly, uh , vh ∈ ∞ n
C (R ) for h > 0. Moreover, from Theorem 2.8 we have kuh − ukL1 (S) ≤
kuh − ukL1 (Ω00 ) → 0.
Now we note that if x ∈ Ω0 and 0 < h < d/2, then ωh (x − y) ∈ C0∞ (Ω00 ). Thus by
Theorem 2.4 and the definition of weak derivative,
Z Z
α α |α|
D uh (x) = u(y)Dx ωh (x − y)dy = (−1) u(y)Dyα ωh (x − y)dy
00 Ω00
ZΩ
= ωh (x − y) · v(y)dy = vh (x).
Ω00
Thus, kDα uh − vkL1 (S) → 0.
|α|
(Sufficiency) Choose ϕ ∈ C0 (Ω) and consider a compact set S ⊃ supp(ϕ). Then as
h→∞
Z Z Z Z
uDα ϕdx ← uh Dα ϕdx = (−1)|α| ϕDα uh dx → (−1)|α| vϕdx
S S S S
2.2. Approximations and Extensions 31

which is the claim. 

If u is equal to a constant(a.e.) in Ω, then u has the weak derivative Dα u = 0, |α| > 0.

An application of Theorem 2.10 yields the converse:
Theorem 2.11. Let Ω be a bounded open connected set in Rn . If u ∈ L1loc (Ω) has a weak
derivative Dα u = 0 whenever |α| = 1, then u =const. a.e. in Ω.

Proof. Let Ω0 ⊂⊂ Ω. Then for x ∈ Ω0 and with uh as in Theorem 2.10, Dα uh (x) =

(Dα u)h (x) = 0 for all h sufficiently small. Thus uh = const = c(h) in Ω0 for such h. Since
kuh − ukL1 (Ω0 ) = kc(h) − ukL1 (Ω0 ) → 0 as h → 0, it follows that
kc(h1 ) − c(h2 )kL1 (Ω0 ) = |c(h1 ) − c(h2 )|mes(Ω0 ) → 0 as h1 , h2 → 0.
Consequently, c(h) = uh converges uniformly and thus in L1 (Ω0 ) to some constant. Hence
u = const(a.e.) in Ω0 and therefore also in Ω, by virtue of it being connected. 

We now note some properties of W k,p (Ω) which follow easily from the results of this
and the previous section.
(a) If Ω0 ⊂ Ω and if u ∈ W k,p (Ω), then u ∈ W k,p (Ω0 ).
(b) If u ∈ W k,p (Ω) and |a(x)|k,∞ < ∞, then au ∈ W k,p (Ω). In this case any weak
derivative Dα (au) is computed according to the usual rule of differentiating the
product of functions.
(c) If u ∈ W k,p (Ω) and uh is its mollified function, then for any compact set S ⊂
Ω, kuh − ukW k,p (S) → 0 as h → 0. If in addition, u has compact support in Ω, then
kuh − ukk,p → 0 as h → 0.

More generally, we have the following global approximation theorems. (See Meyers and
Serrin H = W . The proofs make use of a partition of unity argument.)
Theorem 2.12. Assume Ω is bounded and let u ∈ W k,p (Ω), 1 ≤ p < ∞. Then there exist
functions um ∈ C ∞ (Ω) ∩ W k,p (Ω) such that
um → u in W k,p (Ω).
In other words, C ∞ (Ω) ∩ W k,p (Ω) is dense in W k,p (Ω).
Theorem 2.13. Assume Ω is bounded and ∂Ω ∈ C 1 . Let u ∈ W k,p (Ω), 1 ≤ p < ∞. Then
there exist functions um ∈ C ∞ (Ω̄) such that
um → u in W k,p (Ω).
In other words, C ∞ (Ω̄) is dense in W k,p (Ω).

Exercise 1. Prove the product rule for weak derivatives:

Di (uv) = (Di u)v + u(Di v)
where u, Di u are locally Lp (Ω), v, Di v are locally Lq (Ω) (p > 1, 1/p + 1/q = 1).
Exercise 2. (a) If u ∈ W0k,p (Ω) and v ∈ C k (Ω̄), prove that uv ∈ W0k,p (Ω).
(b) If u ∈ W k,p (Ω) and v ∈ C0k (Ω), prove that uv ∈ W0k,p (Ω).
32 2. Sobolev Spaces

2.2.3. Chain Rules.

Theorem 2.14. Let Ω be a bounded open set in Rn . Let f ∈ C 1 (R), |f 0 (s)| ≤ M for all
s ∈ R and suppose u has a weak derivative Dα u for |α| = 1. Then the composite function
f ◦u has a weak derivative Dα (f ◦u) = f 0 (u)Dα u. Moreover, if f (0) = 0 and if u ∈ W 1,p (Ω),
then f ◦ u ∈ W 1,p (Ω).

Proof. According to Theorem 2.10, there exists a sequence {uh } ⊂ C 1 (Ω) such that kuh −
ukL1 (Ω0 ) → 0, kDα uh − Dα ukL1 (Ω0 ) → 0 as h → 0, where Ω0 ⊂⊂ Ω. Thus
Z Z
0
|f (uh ) − f (u)|dx ≤ sup |f | |uh − u|dx → 0 as h → 0
Ω0 Ω0
Z Z
0 α 0 α 0
|f (uh )D uh − f (u)D u|dx ≤ sup |f | |Dα uh − Dα u|dx
Ω0 Ω0
Z
+ |f 0 (uh ) − f 0 (u)||Dα u|dx.
Ω0
Since kuh − ukL1 (Ω0 ) → 0, there exists a subsequence of {uh }, which we call {uh } again,
which converges a.e. in Ω0 to u. Moreover, since f 0 is continuous, {f 0 (uh )} converges to f 0 (u)
a.e. in Ω0 . Hence the last integral tends to zero by the dominated convergence theorem.
Consequently, the sequences {f (uh )}, {f 0 (uh )Dα uh } tend to f (u), f 0 (u)Dα u respectively,
and the first conclusion follows by an application of Theorem 2.10 again.
Since f (0) = 0, the mean value theorem implies |f (s)| ≤ M |s| for all s ∈ R. Thus,
|f (u(x))| ≤ M |u(x)| for all x ∈ Ω and so f ◦ u ∈ Lp (Ω) if u ∈ Lp (Ω). Similarly,
f 0 (u(x))Dα u ∈ Lp (Ω) if u ∈ W 1,p (Ω), which shows that f ◦ u ∈ W 1,p (Ω). 
Corollary 2.15. Let Ω be a bounded open set in Rn . If u has an αth weak derivative
Dα u, |α| = 1, then so does |u| and
 Dα u

if u > 0
α
D |u| = 0 if u = 0
−Dα u if u < 0


i.e., Dα |u| = (sgn u)Dα u for u 6= 0. In particular, if u ∈ W 1,p (Ω), then |u| ∈ W 1,p (Ω).

Proof. The positive and negative parts of u are defined by

u+ = max{u, 0}, u− = min{u, 0}.

If we can show that Dα u+ exists and that

 α
α + D u if u > 0
D u =
0 if u ≤ 0
then the result for |u| follows easily from the relations |u| = u+ − u− and u− = −(−u)+ .
Thus, for h > 0 define
 1
(u2 + h2 ) 2 − h if u > 0
fh (u) =
0 if u ≤ 0.
Clearly fh ∈ C 1 (R) and fh0 is bounded on R. By Theorem 2.14, fh (u) has a weak derivative,
and for any ϕ ∈ C01 (Ω)
uDα u
Z Z Z
α α
fh (u)D ϕdx = − D (fh (u))ϕdx = − ϕ 1 dx.
Ω Ω u>0 (u2 + h2 ) 2
2.2. Approximations and Extensions 33

Upon letting h → 0, it follows that fh (u) → u+ , and so by the dominating convergence

theorem Z Z Z
+ α α
u D ϕdx = − ϕD udx = − vϕdx
Ω u>0 Ω
where
Dα u if u > 0

v=
0 if u ≤ 0
+
which establishes the desired result for u . 

Remark. Since u = u+ + u− , we have ∂u/∂xi = ∂u+ /∂xi + ∂u− /∂xi . Consequently,

∂u/∂xi = 0 a.e. on {u = c} = {x ∈ Ω : u(x) = c}.

The following result is of independent importance.

Theorem 2.16. un * u in W k,p (Ω), if and only if Dα un * Dα u in Lp (Ω) for all |α| ≤ k.
Theorem 2.17. Let f : R → R be Lipschitz continuous with f (0) = 0. Then if Ω is a
bounded open set in Rn , 1 < p < ∞ and u ∈ W01,p (Ω), we have f ◦ u ∈ W01,p (Ω).

Proof. Given u ∈ W01,p (Ω), let un ∈ C01 (Ω) with kun − uk1,p → 0 and define vn = f ◦ un .
Since un has compact support and f (0) = 0, vn has compact support. Also vn is Lipschitz
continuous, for
|vn (x) − vn (y)| = |f (un (x)) − f (un (y))|
≤ c|un (x) − un (y)|
≤ cn |x − y|.
Hence vn ∈ Lp (Ω).
Since vn is absolutely continuous on any line segment in Ω, its partial derivatives (which
exist almost everywhere) coincide almost everywhere with the weak derivatives. Moreover,
we see from above that |∂vn /∂xi | ≤ cn for 1 ≤ i ≤ n, and as Ω is bounded, ∂vn /∂xi ∈ Lp (Ω).
Thus vn ∈ W 1,p (Ω) and has compact support, which implies vn ∈ W01,p (Ω). From the
relation
|vn (x) − f (u(x))| ≤ c|un (x) − u(x)|
it follows that kvn − f ◦ ukp → 0. Furthermore, if ei is the standard ith basis vector in Rn ,
we have
|vn (x + hei ) − vn (x)| |un (x + hei ) − un (x)|
≤c
|h| |h|
and so
∂vn ∂un
lim sup k kp ≤ c lim sup k kp .
n→∞ ∂xi n→∞ ∂xi
But, {∂un /∂xi } is a convergent sequence in Lp (Ω) and therefore {∂vn /∂xi } is bounded in
Lp (Ω) for each 1 ≤ i ≤ n. Since kvn k1,p is bounded and W01,p (Ω) is reflexive, a subsequence
of {vn } converges weakly in W 1,p (Ω), and thus weakly in Lp (Ω) to some element of W01,p (Ω).
Thus, f ◦ u ∈ W01,p (Ω). 

Remark. In terms of the trace operator (defined later) we have γ0 (f ◦ u) = f ◦ γ0 (u).

Corollary 2.18. Let u ∈ W01,p (Ω). Then |u|, u+ , u− ∈ W01,p (Ω).

Proof. We apply the preceding theorem with f (t) = |t|. Thus |u| ∈ W01,p (Ω). Now
u+ = (|u| + u)/2 and u− = (u − |u|)/2. Thus u+ , u− ∈ W01,p (Ω). 
34 2. Sobolev Spaces

2.2.4. Extensions. If Ω ⊂ Ω0 , then any function u(x) ∈ C0k (Ω) has an obvious extension
U (x) ∈ C0k (Ω0 ). From the definition of W0k,p (Ω) it follows that the function u(x) ∈ W0k,p (Ω)
and extended as being equal to zero in Ω0 \Ω belongs to W0k,p (Ω0 ). In general, a function
u ∈ W k,p (Ω) and extended by zero to Ω0 will not belong to W k,p (Ω0 ). (Consider the function
u(x) ≡ 1 in Ω.) However, if u ∈ W k,p (Ω) has compact support in Ω, then u ∈ W0k,p (Ω) and
thus the obvious extension belongs to W0k,p (Ω0 ).

We now consider a more general extension result.

Theorem 2.19. Let Ω be a bounded open set in Rn with Ω ⊂⊂ Ω0 and assume k ≥ 1.
(a) If ∂Ω ∈ C k , then any function u(x) ∈ W k,p (Ω) has an extension U (x) ∈
W k,p (Ω0 ) into Ω0 with compact support. Moreover,
kU kW k,p (Ω0 ) ≤ ckukW k,p (Ω)
where the constant c > 0 does not depend on u.
(b) If ∂Ω ∈ C k , then any function u(x) ∈ C k (Ω̄) has an extension U (x) ∈ C0k (Ω0 )
into Ω0 with compact support. Moreover,
kU kC k (Ω̄0 ) ≤ ckukC k (Ω̄) , kU kW k,p (Ω0 ) ≤ ckukW k,p (Ω)
where the constant c > 0 does not depend on u.
(c) If ∂Ω ∈ C k , then any function u(x) ∈ C k (∂Ω) has an extension U (x) into Ω
which belongs to C k (Ω̄). Moreover
kU kC k (Ω̄) ≤ ckukC k (∂Ω)
where the constant c > 0 does not depend on u.

Proof. Suppose first that u ∈ C k (Ω̄). Let y = ψ(x) define a C k diffeomorphism that
straightens the boundary near x0 = (x01 , . . . , x0n ) ∈ ∂Ω. In particular, we assume there is a
ball B = B(x0 ) such that ψ(B ∩ Ω) ⊂ Rn+ (i.e., yn > 0), ψ(B ∩ ∂Ω) ⊂ ∂Rn+ . (e.g., we could
choose yi = xi − x0i for i = 1, . . . , n − 1 and yn = xn − ϕ(x1 , . . . , xn−1 ), where ϕ is of class
C k . Moreover, without loss of generality, we can assume yn > 0 if x ∈ B ∩ Ω.)
Let G and G+ = G ∩ Rn+ be respectively, a ball and half-ball in the image of ψ such
0
that ψ(x0 ) ∈ G. Setting ū(y) = u ◦ ψ −1 (y) and y = (y1 , . . . , yn−1 , yn ) = (y , yn ), we define
an extension Ū (y) of ū(y) into yn < 0 by
k+1
0 0
X
Ū (y , yn ) = ci ū(y , −yn /i), yn < 0
i=1
where the ci are constants determined by the system of equations
k+1
X
(2.5) ci (−1/i)m = 1, m = 0, 1, . . . , k.
i=1

Note that the determinant of the system (2.5) is nonzero since it is the Vandemonde deter-
minant. One verifies readily that the extended function Ū is continuous with all derivatives
up to order k in G. For example,
k+1
X
lim Ū (y) = ci ū(y 0 , 0) = ū(y 0 , 0)
y→(y 0 ,0)
i=1
2.2. Approximations and Extensions 35

by virtue of (2.5) with m = 0. A similar computation shows that

lim Ūyi (y) = ūyi (y 0 , 0), i = 1, . . . , n − 1.
y→(y 0 ,0)

Finally
k+1
X
lim Ūyn (y) = ci (−1/i)ūyn (y 0 , 0) = ūyn (y 0 , 0)
y→(y 0 ,0)
i=1
by virtue of (2.5) with m = 1. Similarly we can handle the higher derivatives.
Thus w = Ū ◦ ψ ∈ C k (B 0 ) for some ball B 0 = B 0 (x0 ) and w = u in B 0 ∩ Ω, (If x ∈ B 0 ∩ Ω,
then ψ(x) ∈ G+ and w(x) = Ū (ψ(x)) = ū(ψ(x)) = u(ψ −1 ψ(x)) = u(x)) so that w provides
a C k extension of u into Ω ∪ B 0 . Moreover,
sup |ū(y)| = sup |u(ψ −1 (y))| ≤ sup |u(x)|
G+ G+ Ω

and since x ∈ B 0 implies ψ(x) ∈ G

sup |Ū (ψ(x))| ≤ c sup |ū(y)| ≤ c sup |u(x)|.
B0 G+ Ω

Since a similar computation for the derivatives holds, it follows that there is a constant
c > 0, independent of u, such that
kwkC k (Ω̄∪B 0 ) ≤ ckukC k (Ω̄) .

Now consider a finite covering of ∂Ω by balls Bi , i = 1, . . . , N , such as B in the

preceding, and let {wi } be the corresponding C k extensions. We may assume the balls Bi
0
are so small that their union with Ω is contained in Ω . Let Ω0 ⊂⊂ Ω be such that Ω0 and
the balls Bi provide a finite open covering of Ω. Let {ηi }, i = 1, . . . , N , be a partition of
unity subordinate to this covering and set
X
w = uη0 + wi η i
0
with the understanding that wi ηi = 0 if ηi = 0. Then w is an extension of u into Ω and
has the required properties. Thus (b) is established.
(a) If u ∈ W k,p (Ω), then by Theorem 2.13, there exist functions um ∈ C ∞ (Ω̄) such that
um → u in W k,p (Ω). Let Ω ⊂ Ω00 ⊂ Ω0 , and let Um be the extension of um to Ω00 as given
in (b). Then
kUm − Ul kW k,p (Ω00 ) ≤ ckum − ul kW k,p (Ω)
which implies that {Um } is a Cauchy sequence and so converges to a U ∈ W0k,p (Ω00 ), since
Um ∈ C0k (Ω00 ). Now extend Um , U by 0 to Ω0 . It is easy to see that U is the desired
extension.
(c) At any point x0 ∈ ∂Ω let the mapping ψ and the ball G be defined as in (b). By
0 0
definition, u ∈ C k (∂Ω) implies that ū = u◦ψ −1 ∈ C k (G∩∂Rn+ ). We define Φ̄(y , yn ) = ū(y )
in G and set Φ(x) = Φ̄ ◦ ψ(x) for x ∈ ψ −1 (G). Clearly, Φ ∈ C k (B̄) for some ball B = B(x0 )
and Φ = u on B ∩ ∂Ω. Now let {Bi } be a finite covering of ∂Ω by balls such as B and let Φi
be the corresponding C k functions defined on Bi . For each i, we define the function Ui (x)
as follows: in the ball Bi take it equal to Φi , outside Bi take it equal to zero if x 6∈ ∂Ω and
equal to u(x) if x ∈ ∂Ω. The proof can now be completed as in (b) by use of an appropriate
partition of unity. 
36 2. Sobolev Spaces

2.2.5. Trace Theorem. Unless otherwise stated, Ω will denote a bounded open connected
set in Rn , i.e., a bounded domain.

Let Γ be a surface which lies in Ω̄ and has the representation

xn = ϕ(x0 ), x0 = (x1 , . . . , xn−1 )
where ϕ(x0 ) is Lipschitz continuous in Ū . Here U is the projection of Γ onto the coordinate
plane xn = 0. Let p ≥ 1. A function u defined on Γ is said to belong to Lp (Γ) if
Z
1
kukLp (Γ) ≡ ( |u(x)|p dS) p < ∞
Γ
where
Z Z n−1
X ∂ϕ 0 2 1 0
|u(x)|p dS = |u(x0 , ϕ(x0 ))|p [1 + ( (x )) ] 2 dx .
Γ U ∂xi
i=1
Thus Lp (Γ) reduces to a space of the type Lp (U ) where U is a domain in Rn−1 .
For every function u ∈ C(Ω̄), its values γ0 u ≡ u|Γ on Γ are uniquely given. The function
γ0 u will be called the trace of the function u on Γ. Note that u ∈ Lp (Γ) since γ0 u ∈ C(Γ).
On the other hand, if we consider a function u defined a.e. in Ω (i.e., functions are
considered equal if they coincide a.e.), then the values of u on Γ are not uniquely determined
since meas(Γ) = 0. In particular, since ∂Ω has measure 0, there exist infinitely many
extensions of u to Ω̄ that are equal a.e. We shall therefore introduce the concept of trace for
functions in W 1,p (Ω) so that if in addition, u ∈ C(Ω̄), the new definition of trace reduces
to the definition given above.
Lemma 2.20. Let ∂Ω ∈ C 0,1 . Then for u ∈ C 1 (Ω̄),
(2.6) kγ0 ukLp (∂Ω) ≤ ckuk1,p
where the constant c > 0 does not depend on u.

Proof. For simplicity, let n = 2. The more general case is handled similarly. In a neigh-
borhood of a boundary point x ∈ ∂Ω, we choose a local (ξ, η)-coordinate system, where the
boundary has the local representation
η = ϕ(ξ), −α ≤ ξ ≤ α
with the C 0,1 function ϕ. Then there exists a β > 0 such that all the points (ξ, η) with
−α ≤ ξ ≤ α, ϕ(ξ) − β ≤ η ≤ ϕ(ξ)
belong to Ω̄. Let u ∈ C 1 (Ω̄). Then
Z ϕ(ξ)
u(ξ, ϕ(ξ)) = uη (ξ, η)dη + u(ξ, t)
t
where ϕ(ξ) − β ≤ t ≤ ϕ(ξ). Applying the inequality (a + b)p ≤ 2p−1 (ap + bp ) together with
Hölder’s inequality we have
Z ϕ(ξ)
p p−1 p−1
|u(ξ, ϕ(ξ))| ≤ 2 β |uη (ξ, η)|p dη + 2p−1 |u(ξ, t)|p .
ϕ(ξ)−β

An integration with respect to t yields

Z ϕ(ξ)
p p−1
β|u(ξ, ϕ(ξ))| ≤ 2 [β p |uη (ξ, η)|p + |u(ξ, η)|p ]dη.
ϕ(ξ)−β
2.2. Approximations and Extensions 37

Finally, integration over the interval [−α, α] yields

Z α Z
p p−1
(2.7) β|u(ξ, ϕ(ξ))| dξ ≤ 2 (β p |uη |p + |u|p )dξdη
−α S

where S denotes a local boundary strip.

Suppose ϕ(·) is C 1 . Then the differential of arc length is given by ds = (1 + ϕ02 )1/2 dξ.
Addition of the local inequalities (2.7) yields the assertion (2.6). Now if ϕ(·) is merely
Lipschitz continuous, then the derivative ϕ0 exists a.e. and is bounded. Thus we also
obtain (2.6). 

Since C 1 (Ω̄) = W 1,p (Ω), the bounded linear operator γ0 : C 1 (Ω̄) ⊂ W 1,p (Ω) → Lp (∂Ω)
can be uniquely extended to a bounded linear operator γ0 : W 1,p (Ω) → Lp (∂Ω) such that
(2.6) remains true for all u ∈ W 1,p (Ω). More precisely, we obtain γ0 u in the following
way: Let u ∈ W 1,p (Ω). We choose a sequence {un } ⊂ C 1 (Ω̄) with kun − uk1,p → 0. Then
kγ0 un − γ0 ukLp (∂Ω) → 0.
The function γ0 u (as an element of Lp (∂Ω)) will be called the trace of the function
u ∈ W 1,p (Ω) on the boundary ∂Ω. (kγ0 ukLp (∂Ω) will be denoted by kukLp (∂Ω) .) Thus the
trace of a function is defined for any element u ∈ W 1,p (Ω).

The above discussion partly proves the following:

Theorem 2.21. (Trace) Suppose ∂Ω ∈ C 1 . Then there is a unique bounded linear operator
γ0 : W 1,p (Ω) → Lp (∂Ω) such that γ0 u = u|∂Ω for u ∈ C(Ω̄)∩W 1,p (Ω), and γ0 (au) = γ0 a·γ0 u
for a(x) ∈ C 1 (Ω̄), u ∈ W 1,p (Ω). Moreover, N (γ0 ) = W01,p (Ω) and R(γ0 ) = Lp (∂Ω).

Proof. Suppose u ∈ C(Ω̄) ∩ W 1,p (Ω). Then by Theorem 2.19, u can be extended into
Ω0 (Ω ⊂⊂ Ω0 ) such that its extension U ∈ C(Ω̄0 ) ∩ W 1,p (Ω0 ). Let Uh (x) be the mollified
function for U . Since Uh → U as h → 0 in both the norms k · kC(Ω̄) , k · kW 1,p (Ω) , we find that
as h → 0, Uh |∂Ω → u|∂Ω uniformly and Uh |∂Ω → γ0 u in Lp (∂Ω). Consequently, γ0 u = u|∂Ω .
Now au ∈ W 1,p (Ω) if a ∈ C 1 (Ω̄), u ∈ W 1,p (Ω) and consequently, γ0 (au) is defined. Let
{un } ⊂ C 1 (Ω̄) with kun − uk1,p → 0. Then

γ0 (aun ) = γ0 a · γ0 un

and the desired product formula follows by virtue of the continuity of γ0 .

If u ∈ W01,p (Ω), then there is a sequence {un } ⊂ C01 (Ω) with kun − uk1,p → 0. But
un |∂Ω = 0 and as n → ∞, un |∂Ω → γ0 u in Lp (∂Ω) which implies γ0 u = 0. Hence W01,p (Ω) ⊂
N (γ0 ). Now suppose u ∈ N (γ0 ). If u ∈ W 1,p (Ω) has compact support in Ω, then by an
earlier remark, u ∈ W01,p (Ω). If u does not have compact support in Ω, then it can be shown
that there exists a sequence of cut-off functions ηk such that ηk u ∈ W 1,p (Ω) has compact
support in Ω, and moreover, kηk u − uk1,p → 0. By using the corresponding mollified
functions, it follows that u ∈ W01,p (Ω) and N (γ0 ) ⊂ W01,p (Ω). Details can be found in
Mikhailov [26].
To see that R(γ0 ) = Lp (∂Ω), let f ∈ Lp (∂Ω) and let ε > 0 be given. Then there is a
u ∈ C 1 (∂Ω) such that ku − f kLp (∂Ω) < ε. If we let U ∈ C 1 (Ω̄) be the extension of u into Ω̄,
then clearly kγ0 U − f kLp (∂Ω) < ε, which is the desired result since U ∈ W 1,p (Ω). 
38 2. Sobolev Spaces

Remark. We note that the function u ≡ 1 belongs to W 1,p (Ω) ∩ C(Ω̄) and its trace on ∂Ω is
1. Hence this function does not belong to W01,p (Ω), which establishes the earlier assertion
that W01,p (Ω) 6= W 1,p (Ω).

Let u ∈ W k,p (Ω), k > 1. Since any weak derivative Dα u of order |α| < k belongs to
W 1,p (Ω),this derivative has a trace γ0 Dα u belonging to Lp (∂Ω). Moreover
kDα ukLp (∂Ω) ≤ ckDα uk1,p ≤ ckukk,p
for constant c > 0 independent of u.
Assuming the boundary ∂Ω ∈ C 1 , the unit outward normal vector n to ∂Ω exists and
is bounded. Thus, the concept of traces makes it possible to introduce, for k ≥ 2, ∂u/∂n
for u ∈ W k,p (Ω). More precisely, for k ≥ 2, there exist traces of the functions u, Di u so
that, if ni are the direction cosines of the normal, we may define
n
X
γ1 u = (γ0 (Di u))ni , u ∈ W k,p (Ω), k ≥ 2.
i=1

The trace operator γ1 : W k,p (Ω) → Lp (∂Ω) is continuous and γ1 u = (∂u/∂n)|∂Ω for u ∈
C 1 (Ω̄) ∩ W k,p (Ω).
For a function u ∈ C k (Ω̄) we define the various traces of normal derivatives given by
∂j u
γj u = |∂Ω , 0 ≤ j ≤ k − 1.
∂nj
Each γj can be extended by continuity to all of W k,p (Ω) and we obtain the following:
Theorem 2.22. (Trace) Suppose ∂Ω ∈ C k . Then there is a unique continuous linear
operator γ = (γ0 , γ1 , . . . , γk−1 ) : W k,p (Ω) → k−1 k−1−j,p (∂Ω) such that for u ∈ C k (Ω̄)
Q
j=0 W

∂j u
|∂Ω , j = 1, . . . , k − 1.
γ0 u = u|∂Ω , γj u =
∂nj
Moreover, N (γ) = W0k,p (Ω) and R(γ) = k−1 k−1−j,p (∂Ω).
Q
j=0 W

The Sobolev spaces W k−1−j,p (∂Ω), which are defined over ∂Ω, can be defined locally.

2.2.6. Green’s Identities. In this section we assume that p = 2.

Theorem 2.23. (Integration by Parts) Let u, v ∈ H 1 (Ω) and let ∂Ω ∈ C 1 . Then for
any i = 1, . . . , n
Z Z Z
(2.8) vDi udx = (γ0 u · γ0 v)ni dS − uDi vdx.
Ω ∂Ω Ω

(Di u, Di v are weak derivatives.)

Proof. Let {un } and {vn } be sequences of functions in C 1 (Ω̄) with kun − ukH 1 (Ω) →
0, kvn − vkH 1 (Ω) → 0 as n → ∞. Formula (2.8) holds for un , vn
Z Z Z
vn Di un dx = un vn ni dS − un Di vn dx
Ω ∂Ω Ω
and upon letting n → ∞ relation (2.8) follows. 
Corollary 2.24. Let ∂Ω ∈ C 1 .
2.3. Sobolev Imbedding Theorems 39

(a) If v ∈ H 1 (Ω) and u ∈ H 2 (Ω) then

Z Z Z
v∆udx = γ0 v · γ1 udS − (∇u · ∇v)dx (Green’s 1st identity).
Ω ∂Ω Ω

(b) If u, v ∈ H 2 (Ω) then

Z Z
(v∆u − u∆v)dx = (γ0 v · γ1 u − γ0 u · γ1 v)dS (Green’s 2nd identity).
Ω ∂Ω
Pn
In these formulas ∇u ≡ (D1 u, . . . , Dn u) is the gradient vector and ∆u ≡ i=1 Dii u is the
Laplace operator.

Proof. If in (2.8) we replace u by Di u and sum from 1 to n, then Green’s 1st identity is
obtained. Interchanging the roles of u, v in Green’s 1st identity and subtracting the two
identities yields Green’s 2nd identity. 

Exercise 1. Establish the following one-dimensional version of the trace theorem: If

u ∈ W 1,p (Ω), where Ω = (a, b), then

kukLp (∂Ω) ≡ (|u(a)|p + |u(b)|p )1/p ≤ const kukW 1,p (Ω)

where the constant is independent of u.

2.3. Sobolev Imbedding Theorems

We consider the following question: If a function u belongs to W k,p (Ω), does u automatically
belong to certain other spaces? The answer will be yes, but which other spaces depend upon
whether 1 ≤ kp < n, kp = n, n < kp < ∞.

The result we want to prove is the following:

Theorem 2.25. (Sobolev-Rellich-Kondrachov) Let Ω ⊂ Rn be bounded and open with

∂Ω ∈ C 1 . Assume 1 ≤ p < ∞ and k is a positive integer.
(a) If kp < n and 1 ≤ q ≤ np/(n − kp), then

W k,p (Ω) ⊂ Lq (Ω)

is a continuous imbedding; the imbedding is compact if 1 ≤ q < np/(n − kp).
Moreover,
(2.9) kukLq (Ω) ≤ CkukW k,p (Ω)
where the constant C depends only on k,p,n and Ω.
(b) If kp = n and 1 ≤ r < ∞, then

W k,p (Ω) ⊂ Lr (Ω)

is a compact imbedding and
(2.10) kukLr (Ω) ≤ CkukW k,p (Ω)
where the constant depends only on k,p,n and Ω.
40 2. Sobolev Spaces

(c) If kp > n and 0 ≤ α ≤ k − m − n/p, then

W k,p (Ω) ⊂ C m,α (Ω̄)
is a continuous imbedding; the imbedding is compact if 0 ≤ α < k − m − n/p.
Moreover,
(2.11) kukC m,α (Ω̄) ≤ CkukW k,p (Ω)
where the constant C depends only on k,p,n,α and Ω.
(d) Let 0 ≤ j < k, 1 ≤ p, q < ∞. Set d = 1/p − (k − j)/n. Then
W k,p ⊂ W j,q
is a continuous imbedding for d ≤ 1/q; the imbedding is compact for d < 1/q.
The above results are valid for W0k,p (Ω) spaces on arbitrary bounded domains Ω.

Remark. It is easy to check that the imbedding W 1,p (Ω) ⊂ Lp (Ω) is compact for all p ≥ 1
and all n.
A series of special results will be needed to prove the above theorem. Only selected
proofs will be given to illustrate some of the important techniques.

2.3.1. Gagliardo-Nirenberg-Sobolev Inequality. Suppose 1 ≤ p < n. Do there exist

constants C > 0 and 1 ≤ q < ∞ such that
(2.12) kukLq (Rn ) ≤ Ck∇ukLp (Rn )
for all u ∈ C0∞ (Rn )? The point is that the constants C and q should not depend on u.
We shall show that if such an inequality holds, then q must have a specific form. For
this choose any u ∈ C0∞ (Rn ), u 6≡ 0, and define for λ > 0
uλ (x) ≡ u(λx) (x ∈ Rn ).
Now Z Z Z
q q1
|uλ | dx = |u(λx)| dx = n |u(y)|q dy
IR n IR n λ IR n

and
λp
Z Z Z
p p p
|∇uλ | dx = λ |∇u(λx)| dx = n |∇u(y)|p dy.
IRn IRn λ IRn
Inserting these inequalities into (2.12) we find
1 λ
kukLq (Rn ) ≤ C k∇ukLp (Rn )
λn/q λn/p
and so
(2.13) kukLq (Rn ) ≤ Cλ1−n/p+n/q k∇ukLp (Rn ) .
But then if 1 − n/p + n/q > 0(< 0), we can upon sending λ to 0 (∞) in (2.13) obtain a
contradiction (u = 0). Thus we must have q = p* where
np
(2.14) p* =
n−p
is called the Sobolev conjugate of p. Note that then
1 1 1
(2.15) ∗
= − , p∗ > p.
p p n
Next we prove that the inequality (2.12) is in fact correct.
2.3. Sobolev Imbedding Theorems 41

Lemma 2.26. (Gagliardo-Nirenberg-Sobolev Inequality) Assume 1 ≤ p < n. Then

there is a constant C, depending only on p and n, such that

(2.16) kukLp∗ (Rn ) ≤ Ck∇ukLp (Rn )

for all u ∈ C01 (Rn ).

Proof. First assume p = 1. Since u has compact support, for each i = 1, . . . , n we have
Z xi
u(x) = uxi (x1 , . . . , xi−1 , yi , xi+1 , . . . , xn )dyi
−∞

and so
Z ∞
|u(x)| ≤ |∇u(x1 , . . . , yi , . . . , xn )|dyi (i = 1, . . . , n).
−∞

Consequently
n Z 1
n Y ∞ 
n−1
(2.17) |u(x)| n−1 ≤ |∇u(x1 , . . . , yi , . . . , xn )|dyi .
i=1 −∞

Integrate this inequality with respect to x1 :

n Z 1
Z ∞ n
Z ∞ Y ∞ 
n−1
|u(x)| n−1 dx1 ≤ |∇u|dyi dx1
−∞ −∞ i=1 −∞
1 n Z 1
Z ∞ 
n−1
Z ∞ Y ∞ 
n−1
= |∇u|dy1 |∇u|dyi dx1
−∞ −∞ i=2 −∞
1
1 n Z
!
Z ∞ 
n−1 Y ∞ Z ∞ n−1

≤ |∇u|dy1 |∇u|dx1 dyi

−∞ i=2 −∞ −∞

the last inequality resulting from the extended Hölder inequality in the appendix.
We continue by integrating with respect to x2 , . . . , xn and applying the extended Hölder
inequality to eventually find (pull out an integral at each step)
n Z 1
Z
n Y ∞ Z ∞ 
n−1
|u(x)| n−1 dx ≤ ··· |∇u|dx1 . . . dyi . . . dxn
IRn i=1 −∞ −∞
Z  n
n−1
= |∇u|dx
IRn

which is estimate (2.16) for p = 1.

Consider now the case that 1 < p < n. We shall apply the last estimate to v = |u|γ ,
where γ > 1 is to be selected. First note that

(γuγ−1 Di u)2

γ 2 if u ≥ 0
(Di |u| ) = γ−1 2 = (γ|u|γ−1 Di u)2 .
(−γ(−u) Di u) if u ≤ 0

Thus v ∈ C01 (Rn ), and

42 2. Sobolev Spaces

Z  n−1 Z
γn n
|u(x)| n−1 dx ≤ |∇|u|γ |dx
IRn IRn
Z
= γ |u|γ−1 |∇u|dx
IRn
Z  p−1 Z 1
p(γ−1) p p
p
≤ γ |u| p−1 dx |∇u| dx .
IRn IRn

We set
p(n − 1)
γ= >1
n−p
in which case
γn p(γ − 1) np
= = = p∗ .
n−1 p−1 n−p
Thus, the above estimate becomes
Z  1∗ Z 1
p p
p∗ p
|u| dx ≤C |∇u| dx .
IRn IRn

Theorem 2.27. Let Ω ⊂ Rn be bounded and open, with ∂Ω ∈ C 1 . Assume 1 ≤ p < n, and
∗
u ∈ W 1,p (Ω). Then u ∈ Lp (Ω) and
(2.18) kukLp∗ (Ω) ≤ CkukW 1,p (Ω)
where the constant C depends only on p,n and Ω.

Proof. Since ∂Ω ∈ C 1 , there exists an extension U ∈ W 1,p (Rn ) such that U = u in Ω, U

has compact support and
(2.19) kU kW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
Moreover, since U has compact support, there exist mollified functions um ∈ C0∞ (Rn ) such
that um → U in W 1,p (Rn ). Now according to Lemma 2.26,
kum − ul kLp∗ (Rn ) ≤ Ck∇um − ∇ul kLp (Rn )
∗
for all l, m ≥ 1; whence um → U in Lp (Rn ) as well. Since Lemma 2.26 also implies
kum kLp∗ (Rn ) ≤ Ck∇um kLp (Rn )
we get in the limit that
kU kLp∗ (Rn ) ≤ Ck∇U kLp (Rn ) .
This inequality and (2.19) complete the proof. 

Theorem 2.28. Let Ω ⊂ Rn be bounded and open. Assume 1 ≤ p < n, and u ∈ W01,p (Ω).
Then u ∈ Lq (Ω) and
kukLq (Ω) ≤ Ck∇ukLp (Ω)
for each q ∈ [1, p∗ ], the constant C depending only on p,q,n and Ω.
2.3. Sobolev Imbedding Theorems 43

Proof. Since u ∈ W01,p (Ω), there are functions um ∈ C0∞ (Ω) such that um → u in W 1,p (Ω).
We extend each function um to be 0 in Rn \Ω̄ and apply Lemma 2.26 to discover (as above)
kukLp∗ (Ω) ≤ Ck∇ukLp (Ω) .

Since |Ω| < ∞, we furthermore have

kukLq (Ω) ≤ CkukLp∗ (Ω)

for every q ∈ [1, p∗ ]. 

2.3.2. Morrey’s Inequality. We now turn to the case n < p < ∞. The next result shows
that if u ∈ W 1,p (Ω), then u is in fact Hölder continuous, after possibly being redefined on
a set of measure zero.

Theorem 2.29. (Morrey’s Inequality) Assume n < p < ∞. Then there exists a constant
C, depending only on p and n, such that
(2.20) kuk 0,1− n ≤ CkukW 1,p (Rn ) , ∀ u ∈ C 1 (Rn ).
C p (Rn )

Proof. We first prove the following inequality: for all x ∈ Rn , r > 0 and all u ∈ C 1 (Rn ),
rn |Du(y)|
Z Z
(2.21) |u(y) − u(x)| dy ≤ dy.
B(x,r) n B(x,r) |x − y|n−1
To prove this, note that, for any w with |w| = 1 and 0 < s < r,
Z s 
 d 
|u(x + sw) − u(x)| =   u(x + tw)dt
dt
Z0 s 
 
=  ∇u(x + tw) · wdt
Z 0s
≤ |∇u(x + sw)| dt.
0

Now we integrate w over ∂B(0, 1) to obtain

Z Z sZ
|u(x + sw) − u(x)| dS ≤ |∇u(x + sw)| dSdt
∂B(0,1) 0 ∂B(0,1)
|∇u(y)|
Z
= dy
B(x,s) |x − y|n−1
|∇u(y)|
Z
≤ dy.
B(x,r) |x − y|n−1

Multiply both sides by sn−1 and integrate over s ∈ (0, r) and we obtain (2.21).
To establish the bound on kukC 0 (Rn ) , we observe that, by (2.21), for x ∈ Rn ,
Z Z
1 1
|u(x)| ≤ |u(y) − u(x)| dy + |u(y)| dy
|B(x, 1)| B(x,1) |B(x, 1)| B(x,1)
! p−1
Z  Z 1/p (1−n)p
p

≤ C |∇u(y)|p dy |y − x| p−1 dy + CkukLp (Rn )

Rn B(x,1)

≤ CkukW 1,p (Rn ) .

44 2. Sobolev Spaces

To establish the bound on the semi-norm [u]γ , γ = 1 − np , take any two points x, y ∈ Rn .
Let r = |x − y| and W = B(x, r) ∩ B(y, r). Then
Z Z
1 1
(2.22) |u(x) − u(y)| ≤ |u(x) − u(z)| dz + |u(y) − u(z)| dz.
|W | W |W | W
Note that |W | = βrn , r = |x − y| and W ≤ min{ B(x,r) , B(y,r) }. Hence, using (2.21), by
R R R

Hölder’s inequality, we obtain

rn
Z Z Z
|u(x) − u(z)| dz ≤ |u(x) − u(z)| dz ≤ |Du(z)||x − z|1−n dz
W B(x,r) n B(x,r)
!1/p Z ! p−1
p
rn
Z (1−n)p
≤ |∇u(z)|p dz |z − x| p−1 dz
n B(x,r) B(x,r)
Z r  p−1
(1−n)p p
n n−1
≤ C r k∇ukLp (Rn ) s p−1 s ds
0
n+γ
≤ Cr k∇ukLp (Rn ) ,
where γ = 1 − np ; similarly,
Z
|u(y) − u(z)| dz ≤ C rn+γ k∇ukLp (Rn ) .
W

Hence, by (2.22),
|u(x) − u(y)| ≤ C |x − y|γ k∇ukLp (Rn ) .
This inequality and the bound on kukC 0 above complete the proof. 
Theorem 2.30. (Estimates for W 1,p , n < p ≤ ∞) Let Ω ⊂ Rn be bounded and open,
0,1− n
with ∂Ω ∈ C 1 . Assume n < p < ∞, and u ∈ W 1,p (Ω). Then u ∈ C p (Ω̄) a.e. and

kuk 0,1− n
p
≤ CkukW 1,p (Ω)
C (Ω̄)

where the constant C depends only on p, n and Ω.

Proof. Since ∂Ω ∈ C 1 , there exists an extension U ∈ W 1,p (Rn ) such that U = u in Ω, U

has compact support and
(2.23) kU kW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
Moreover, since U has compact support, there exist mollified functions um ∈ C0∞ (Rn ) such
that um → U in W 1,p (Rn ) (and hence on compact subsets). Now according to Morrey’s
inequality,
kum − ul kC 0,1−n/p (Rn ) ≤ Ckum − ul kW 1,p (Rn )
for all l, m ≥ 1; whence there is a function u∗ ∈ C 0,1−n/p (Rn ) such that um → u∗ in
C 0,1−n/p (Rn ). Thus u∗ = u a.e. in Ω. Since we also have
kum kC 0,1−n/p (Rn ) ≤ Ckum kW 1,p (Rn )
we get in the limit that
ku∗ kC 0,1−n/p (Rn ) ≤ CkU kW 1,p (Rn ) .
This inequality and (2.23) complete the proof. 
2.3. Sobolev Imbedding Theorems 45

2.3.3. General Cases. We can now concatenate the above estimates to obtain more com-
plicated inequalities.
Assume kp < n and u ∈ W k,p (Ω). Since Dα u ∈ Lp (Ω) for all |α| ≤ k, the Sobolev-
Nirenberg-Gagliardo inequality implies
kDβ ukLp∗ (Ω) ≤ CkukW k,p (Ω)
∗
if |β| ≤ k − 1, and so u ∈ W k−1,p (Ω). Moreover, kukk−1,p∗ ≤ ckukk,p . Similarly, we find
∗∗
u ∈ W k−2,p (Ω), where
1 1 1 1 2
∗∗
= ∗− = − .
p p n p n
Moreover, kukk−2,p∗∗ ≤ ckukk−1,p∗ . Continuing, we find after k steps that u ∈ W 0,q (Ω) =
Lq (Ω) for
1 1 k
= − .
q p n
The stated estimate (2.9) follows from combining the relevant estimates at each stage of
the above argument. In a similar manner the other estimates can be established.

2.3.4. Compactness. We now consider the compactness of the imbeddings. Before we

present the next result we recall some facts that will be needed. A subset S of a normed
space is said to be totally bounded if for each ε > 0 there is a finite set of open balls of
radius ε which cover S. Clearly, a totally bounded set is bounded, i.e., it is contained in a
sufficiently large ball. It is not difficult to see that a relatively compact subset of a normed
space is totally bounded, with the converse being true if the normed space is complete.
Moreover, a totally bounded subset of a normed space is separable.
Theorem 2.31. (Rellich-Kondrachov) Let Ω ⊂ Rn be bounded and open. Then for
1 ≤ p < n:
(a) The imbedding W01,p (Ω) ⊂ Lq (Ω) is compact for each 1 ≤ q < np/(n − p).
(b) Assuming ∂Ω ∈ C 1 , the imbedding W 1,p (Ω) ⊂ Lq (Ω) is compact for each 1 ≤
q < np/(n − p).
(c) Assuming ∂Ω ∈ C 1 , γ0 : W 1,p (Ω) → Lp (∂Ω) is compact.
If p > n, then
(d) Assuming ∂Ω ∈ C 1 , the imbedding W 1,p (Ω) ⊂ C α (Ω̄) is compact for each 0 ≤
α < 1 − (n/p).

Proof. We shall just give the proof for p = q = 2. The other cases are proved similarly.
(a) Since C01 (Ω) is dense in H01 (Ω), it suffices to show that the imbedding C01 (Ω) ⊂ L2 (Ω) is
compact. Thus, let S = {u ∈ C01 (Ω) : kuk1,2 ≤ 1}. We now show that S is totally bounded
in L2 (Ω).
For h > 0, let Sh = {uh : u ∈ S}, where uh is the mollified function for u. We claim
that Sh is totally bounded in L2 (Ω). Indeed, for u ∈ S, we have
Z
|uh (x)| ≤ ωh (z)|u(x − z)|dz ≤ (sup ωh )kuk1 ≤ c1 (sup ωh )kuk1,2
B(0,h)
and
|Di uh (x)| ≤ c2 sup |Di ωh |kuk1,2 , i = 1, . . . , n
so that Sh is a bounded and equicontinuous subset of C(Ω̄). Thus by the Ascoli Theorem,
Sh is relatively compact (and thus totally bounded) in C(Ω̄) and consequently also in L2 (Ω).
46 2. Sobolev Spaces

Now, by earlier estimates, we easily obtain

Z Z 
2 2
kuh − uk2 ≤ ωh (z) |u(x − z) − u(x)| dx dz
B(0,h) Ω

and
Z Z 1 2
du(x − tz) 
Z
2

|u(x − z) − u(x)| dx =  dt dx
Ω Ω

0 dt
Z Z 1 2
 
=  (−∇u(x − tz) · z)dt dx
 
Ω 0
Z  Z 1 
2
≤ |z| |∇u(x − tz)| dt dx ≤ |z|2 kuk21,2 .
2
Ω 0

Consequently, kuh − uk2 ≤ h. Since we have shown above that Sh is totally bounded in
L2 (Ω) for all h > 0, it follows that S is also totally bounded in L2 (Ω) and hence relatively
compact.
(b) Suppose now that S is a bounded set in H 1 (Ω). Each u ∈ S has an extension
U ∈ H01 (Ω0 ) where Ω ⊂⊂ Ω0 . Denote by S 0 the set of all such extensions of the functions
u ∈ S. Since kU kH 1 (Ω0 ) ≤ ckuk1,2 , the set S 0 is bounded in H01 (Ω0 ). By (a) S 0 is relatively
compact in L2 (Ω0 ) and therefore S is relatively compact in L2 (Ω).
(c) Let S be a bounded set in H 1 (Ω). For any u(x) ∈ C 1 (Ω̄), the inequality (2.7) with
p = 2 yields
c1
(2.24) kuk2L2 (∂Ω) ≤ kuk22 + c2 βkuk21,2
β
where the constants c1 , c2 do not depend on u or β. By completion, this inequality is
valid for any u ∈ H 1 (Ω). By (b), any infinite sequence of elements of the set S has a
subsequence {un } which is Cauchy in L2 (Ω): given ε > 0, an N can be found such that for
all m, n ≥ N, kum − un k2 < ε. Now we choose β = ε. Applying the inequality (2.24) to
um − un , it follows that the sequence of traces {γ0 un } converges in L2 (∂Ω).
(d) By Morrey’s inequality, the imbedding is continuous if α = 1 − (n/p). Now use the
fact that C β is compact in C α if α < β. 

Remarks. (a) When p = n, we can easily show that the imbedding in (a) is compact for all
1 ≤ q < ∞. Hence, it follows that the imbedding W01,p (Ω) ⊂ Lp (Ω) is compact for all p ≥ 1.
(b) The boundedness of Ω is essential in the above theorem. For example, let I = (0, 1)
and Ij = (j, j + 1). Let f ∈ C01 (I) and define fj to be the same function defined on Ij by
translation. We can normalize f so that kf kW 1,p (I) = 1. The same is then true for each fj
and thus {fj } is a bounded sequence in W 1,p (R). Clearly f ∈ Lq (R) for every 1 ≤ q ≤ ∞.
Further, if
kf kLq (R) = kf kLq (I) = a > 0
then for any j 6= k we have
Z j+1 Z k+1
kfj − fk kqLq (R) = |fj |q + |fk |q = 2aq
j k

and so fi cannot have a convergent subsequence in Lq (R). Thus none of the imbeddings
W 1,p (R) ⊂ Lq (R) can be compact. This example generalizes to n dimensional space and to
open sets like a half-space.
2.4. Equivalent Norms 47

2.4. Equivalent Norms

Two norms k · k and | · | on a vector space X are equivalent if there exist constants
c1 , c2 ∈ (0, ∞) such that
kxk ≤ c1 |x| ≤ c2 kxk for all x ∈ X.
Note that the property of a set to be open, closed, compact, or complete in a normed space
is not affected if the norm is replaced by an equivalent norm. A seminorm q on a vector
space has all the properties of a norm except that q(u) = 0 need not imply u = 0.

2.4.1. Equivalent Norms of W 1,p (Ω). The next result plays a crucial part in the next
chapter.
Theorem 2.32. Let ∂Ω ∈ C 1 and let 1 ≤ p < ∞. Set
Z X n
!1/p
kuk = |Di u|p dx + (q(u))p
Ω i=1

where q : W 1,p (Ω) → R is a seminorm with the following two properties:

(i) There is a positive constant d such that for all u ∈ W 1,p (Ω)
q(u) ≤ dkuk1,p .
(ii) If u = constant, then q(u) = 0 implies u = 0.
Then k · k is an equivalent norm on W 1,p (Ω).

Proof. First of all, it is easy to check that k · k defines a norm. Now by (i), it suffices to
prove that there is a positive constant c such that
(2.25) kuk1,p ≤ ckuk for all u ∈ W 1,p (Ω).
Suppose (2.25) is false. Then there exist un (x) ∈ W 1,p (Ω) such that (kun k1,p > nkun k. Set
vn = un /kun k1,p .)
(2.26) kun k1,p = 1 and 1 > nkun k.
According to Theorem 2.31, there is a subsequence, call it again {un }, which converges to u
in Lp (Ω). From (2.26) we have kun k → 0 and therefore un → u in W 1,p (Ω) and q(un ) → 0.
Hence, kuk1,p = 1 and by continuity, q(u) = 0. However, since Di u = 0 a.e. in Ω, u =
const., which must be zero by (ii) and this contradicts kuk1,p = 1. 
Example 2.33. Let ∂Ω ∈ C 1 . Assume a(x) ∈ C(Ω), σ(x) ∈ C(∂Ω) with a ≥ 0 (6≡ 0), σ ≥
0 (6≡ 0). Then the following norms are equivalent to k · k1,p on W 1,p (Ω):
Z X n Z p !1/p
|Di u|p dx +  udx
  R 
(2.27) kuk = with q(u) =  Ω udx .
Ω i=1 Ω

n
Z X Z p !1/p
p
  R 
(2.28) kuk = |Di u| dx +  γ0 udS  with q(u) =  ∂Ω γ0 udS  .
Ω i=1 ∂Ω

n
Z X Z !1/p
p p

1/p
σ|γ0 u|p dS
R
(2.29) kuk = |Di u| dx + σ|γ0 u| dS with q(u) = .
Ω i=1 ∂Ω
48 2. Sobolev Spaces

n
Z X Z !1/p
p p
R p

1/p
(2.30) kuk = |Di u| dx + a|u| dx with q(u) = Ω a|u| dx .
Ω i=1 Ω

In order to verify condition (i) of Theorem 2.32, one uses the trace theorem in (2.28) and
(2.29).
R
2.4.2. Poincaré’s Inequalities. If u has zero mean value, i.e., Ω udx = 0, then from
(2.27) it follows that
Z n
Z X
(2.31) |u(x)|p dx ≤ c |Di u|p dx, u ∈ W 1,p (Ω)
Ω Ω i=1

where the constant c > 0 is independent of u. This inequality is often referred to as

Poincaré’s inequality. We also note that (2.28) implies that (2.31) holds for u ∈ W01,p (Ω),
and therefore !1/p
Z X n
p
kuk1,p,0 = |Di u| dx
Ω i=1

defines an equivalent norm on W01,p (Ω).

2.4.3. Fourier Transform Methods. For a function u ∈ L1 (Rn ), we define the Fourier
transform of u by
Z
1
û(y) = e−ix·y u(x) dx, ∀ y ∈ Rn ,
(2π)n/2 Rn
and the inverse Fourier tranform by
Z
1
ǔ(y) = eix·y u(x) dx, ∀ y ∈ Rn .
(2π)n/2 Rn
Theorem 2.34. (Plancherel’s Theorem) Assume u ∈ L1 (Rn ) ∩ L2 (Rn ). Then û, ǔ ∈
L2 (Rn ) and
kûkL2 (Rn ) = kǔkL2 (Rn ) = kukL2 (Rn ) .

Since L1 (Rn ) ∩ L2 (Rn ) is dense in L2 (Rn ), we can use this result to extend the Fourier
transforms on to L2 (Rn ). We still use the same notations for them. Then we have
Theorem 2.35. (Property of Fourier Tranforms) Assume u, v ∈ L2 (Rn ). Then
(i) Rn uv̄ dx = Rn ûv̂¯ dy,
R R

(ii) Ddα u(y) = (iy)α û(y) for each multiindex α such that D α u ∈ L2 (Rn ),

ˇ
(iii) u = û.

Next we use the Fourier transform to characterize the spaces H k (Rn ).

Theorem 2.36. Let k be a nonnegative integer. Then, a function u ∈ L2 (Rn ) belongs to
H k (Rn ) if and only if
(1 + |y|k )û(y) ∈ L2 (Rn ).
In addition, there exists a constant C such that
C −1 kukH k (Rn ) ≤ k(1 + |y|k ) ûkL2 (Rn ) ≤ C kukH k (Rn )
for all u ∈ H k (Rn ).
2.4. Equivalent Norms 49

Using the Fourier transform, we can also define fractional Sobolev spaces H s (Rn ) for
any 0 < s < ∞ as follows
H s (Rn ) = {u ∈ L2 (Rn ) | (1 + |y|s ) û ∈ L2 (Rn )},
and define the norm by
kukH s (Rn ) = k(1 + |y|s ) ûkL2 (Rn ) .
From this we easily get the estimate
kukL∞ (Rn ) ≤ kûkL1 (Rn )
= k(1 + |y|s )û (1 + |y|s )−1 kL1 (Rn )
≤ k(1 + |y|s )ûkL2 (Rn ) k(1 + |y|s )−2 k2L1 (Rn )
≤ C kukH s (Rn ) ,
where C = k(1 + |y|s )−2 k2L1 (Rn ) < ∞ if and only if s > n2 . Therefore we have an easy
embedding, which is known valid for integers s by the previous Sobolev embedding theorem,
H s (Rn ) ⊂ L∞ (Rn ) if s > n2 .
 Chapter 3

Existence Theory for

Linear Problems

3.1. Differential Equations in Divergence Form

Henceforth, Ω ⊂ Rn denotes a bounded domain with boundary ∂Ω ∈ C 1 .

3.1.1. Linear Elliptic Equations. Consider the (Dirichlet) boundary value problem
(BVP)
(3.1) Lu = f in Ω, u = 0 on ∂Ω.
Here f is a given function in L2 (Ω) (or more generally, an element in the dual space of
H01 (Ω)) and L is a formal second-order differential operator in divergence form
given by

n
X n
X
Lu ≡ − Di (aij (x)Dj u) + bi (x)Di u + c(x)u
i,j=1 i=1

with real coefficients aij (x), bi (x) and c(x) only in L∞ (Ω) and aij (x) = aji (x) (i, j =
1, . . . , n). Moreover, L is assumed to be uniformly elliptic in Ω, i.e., there exists a
number θ > 0 such that for every x ∈ Ω and every real vector ξ = (ξ1 , . . . , ξn ) ∈ Rn
n
X n
X
(3.2) aij (x)ξi ξj ≥ θ |ξi |2 .
i,j=1 i=1

A function u ∈ H01 (Ω) is called a weak solution of (3.1) if the following holds
 
Z X n n
X Z
(3.3) B1 (u, v) ≡  aij Dj uDi v + ( bi Di u + cu)v  dx = f vdx, ∀ v ∈ H01 (Ω).
Ω i,j=1 i=1 Ω

Much of the material in this chapter is devoted to proving the existence of weak solutions;
the regularity problem will not be studied. Other problems can also be formulated in the
weak sense as above.

51
52 3. Existence Theory for Linear Problems

Exercise 1. Consider the following weak formulation: Given f ∈ L2 (Ω). Find u ∈

H 1 (Ω) satisfying Z Z
∇u · ∇vdx = f vdx for all v ∈ H 1 (Ω).
Ω Ω
Find the boundary value problem solved by u. What is the necessary condition for the
existence of such a u?

3.1.2. General Systems in Divergence Form. For N unknown functions, u1 , · · · , uN ,

we can write u = (u1 , · · · , uN ) and say u ∈ X(Ω; RN ) if each uk ∈ X(Ω), where X is a
symbol of any function spaces we learned. For example, if u ∈ W 1,p (Ω; RN ) then we use
Du to denote the N × n Jacobi matrix (∂uk /∂xi )1≤k≤N,1≤i≤n .
The (Dirichlet) BVP for a most general system of partial differential equations in di-
vergence form can be written as follows:
(3.4) − div A(x, u, Du) + b(x, u, Du) = F in Ω, u = 0 on ∂Ω,
where A(x, s, ξ) = (Aki (x, u, ξ)), 1 ≤ i ≤ n, 1 ≤ k ≤ N, and b(x, s, ξ) = (bk (x, u, ξ)),
1 ≤ k ≤ N, are given functions of (x, u, ξ) ∈ Ω × RN × MN ×n satisfying some structural
conditions, and F = (f k ), 1 ≤ k ≤ N , with each f k being a given functional in the dual
space of W01,p (Ω). The structural conditions will generally assure that both |A(x, u, Du)|
0 p
and |b(x, u, Du)| belong to Lp (Ω) for all u ∈ W 1,p (Ω; RN ), where p0 = p−1 .
A function u ∈ W01,p (Ω; RN ) is then called a weak solution of (3.4) if the following
holds
Z "X n
#
(3.5) Aki (x, u, Du)Di ϕ + bk (x, u, Du)ϕ dx = hf k , ϕi
Ω i=1

for all ϕ ∈ W01,p (Ω) and each k = 1, 2, · · · , N.

The system (3.4) is said to be linear if both A and b are linear in the variables (u, ξ);
that is,
X N
X
(3.6) Aki (x, u, Du) = akl
ij (x) Dj u
l
+ dkl l
i (x) u ,
1≤l≤N, 1≤j≤n l=1

X N
X
k
b (x, u, Du) = bkl
j (x) Dj u
l
+ ckl (x) ul .
1≤j≤n, 1≤l≤N l=1

In this case, as above for single equations, we work in the Hilbert space H01 (Ω; RN ), which
has the inner product defined by
X Z
(u, v) ≡ Di uk Di v k dx.
1≤i≤n, 1≤k≤N Ω

The pairing between H01 (Ω; RN ) and its dual is given by

N
X
hF, ui = hf k , uk i if F = (f k ) and u = (uk ).
k=1

Then a weak solution of linear system (3.4) is a function u ∈ H01 (Ω; RN ) such that
Z  
(3.7) B2 (u, v) ≡ akl D
ij j ul
Di v k
+ dkl l
i u D i v k
+ bkl
j D j ul k
v + ckl l k
u v dx = hF, vi
Ω
3.1. Differential Equations in Divergence Form 53

holds for all v ∈ H01 (Ω; RN ). Here the conventional summation notation is used.

There are some ellipticity conditions for the system (3.5) in terms of the leading coeffi-
cients A(x, u, ξ). Assume A is smooth on ξ and define

∂Aki (x, u, ξ)
Akl
ij (x, u, ξ) = , ξ = (ξjl ).
∂ξjl

The system (3.5) is said to satisfy the (uniform, strict) Legendre ellipticity condition
if there exists a ν > 0 such that, for all (x, s, ξ), it holds
n X
X N
(3.8) Akl k l
ij (x, s, ξ) ηi ηj ≥ ν |η|
2
for all N × n matrix η = (ηik ).
i,j=1 k,l=1

A weaker condition, obtained by setting η = q ⊗ p = (q k pi ) with p ∈ Rn , q ∈ RN , is the

following (strong) Legendre-Hadamard condition:
n X
X N
(3.9) Akl k l 2
ij (x, s, ξ) q q pi pj ≥ ν |p| |q|
2
∀ p ∈ Rn , q ∈ RN .
i,j=1 k,l=1

Note that for systems with linear leading terms A given by (3.6), the Legendre condition
and Legendre-Hadamard condition become, respectively,
n X
X N
(3.10) akl k l
ij (x) ηi ηj ≥ ν |η|
2
∀ η;
i,j=1 k,l=1

n X
X N
(3.11) akl k l 2
ij (x) q q pi pj ≥ ν |p| |q|
2
∀ p, q.
i,j=1 k,l=1

Remark. The Legendre-Hadamard condition does not imply the Legendre ellipticity condi-
tion. For example, let n = N = 2 and define constants akl
ij by

2
X
akl k l 2
ij ξi ξj ≡ det ξ +  |ξ| .
i,j,k,l=1

Since
2
X
akl k l 2 2 2
ij pi pj q q = det(q ⊗ p) +  |q ⊗ p| =  |p| |q| ,
i,j,k,=1

the Legendre-Hadamard condition holds for all  > 0. But, if 0 <  < 1/2, then the Legendre
ellipticity condition fails.

Exercise 1. Prove that the Legendre condition holds for this system if and only if
 > 1/2.
54 3. Existence Theory for Linear Problems

Remark. Let u = (v, w) and (x1 , x2 ) = (x, y). Then the system of differential equations
defined by akl
ij given above is

∆v + wxy = 0, ∆w − vxy = 0.

This system reduces to two fourth-order equations for v, w (where ∆f = fxx + fyy ):
2 ∆2 v − vxxyy = 0, 2 ∆2 w + wxxyy = 0.
We can easily see that both equations are elliptic if and only if  > 1/2.

Exercise 2. Formulate the biharmonic equation ∆2 u = f as a linear system and find

the appropriate B2 (u, v) in the definition of weak solutions.

3.2. The Lax-Milgram Theorem

Let H denote a real Hilbert space with inner product (·, ·) and norm k · k. A map B :
H × H → R is called a bilinear form if
B(αu + βv, w) = αB(u, w) + βB(v, w)
B(w, αu + βv) = αB(w, u) + βB(w, v)
for all u, v, w ∈ H and all α, β ∈ R.
Our first existence result is frequently referred to as the Lax-Milgram Theorem.
Theorem 3.1. (Lax-Milgram Theorem) Let B : H → H be a bilinear form. Assume
(i) B is bounded; i.e., |B(u, v)| ≤ αkukkvk, and
(ii) B is strongly positive; i.e., B(u, u) ≥ βkuk2 ,
where α, β are positive constants. Let f ∈ H ∗ . Then there exists a unique element u ∈ H
such that
(3.12) B(u, v) = hf, vi, ∀ v ∈ H.
1
Moreover, the solution u satisfies kuk ≤ β kf k.

Proof. For each fixed u ∈ H, the functional v 7→ B(u, v) is in H ∗ , and hence by the Riesz
Representation Theorem, there exists a unique element w = Au ∈ H such that
B(u, v) = (w, v) ∀ v ∈ H.
It can be easily shown that A : H → H is linear. From (i), kAuk2 = B(u, Au) ≤ αkukkAuk,
and hence kAuk ≤ αkuk for all u ∈ H; that is, A is bounded. Furthermore, by (ii),
βkuk2 ≤ B(u, u) = (Au, u) ≤ kAukkuk and hence kAuk ≥ βkuk for all u ∈ H. By the
Riesz Representation Theorem again, we have a unique w0 ∈ H such that hf, vi = (w0 , v)
for all v ∈ H and kf k = kw0 k. We will show that the equation Au = w0 has a (unique)
solution. There are many different proofs for this, and here we use the Contraction Mapping
Theorem. Note that the solution u to equation Au = w0 is equivalent to the fixed-point
of the map T : H → H defined by T (v) = v − tAv + tw0 (v ∈ H) for any fixed t > 0. We
will show for t > 0 small enough T is a contraction. Note that for all v, w ∈ H we have
kT (v) − T (w)k = k(I − tA)(v − w)k. We compute that for all u ∈ H
k(I − tA)uk2 = kuk2 + t2 kAuk2 − 2t(Au, u)
≤ kuk2 (1 + t2 α2 − 2βt).
3.3. Gårding’s Estimates and Existence Theory 55

We now choose t such that 0 < t < α2β2 . Then the expression in parentheses is positive and
less than 1. Thus the map T : H → H is a contraction on H and therefore has a fixed point.
This fixed point u solves Au = w0 and thus is the unique solution of (3.12); moreover, we
have kf k = kw0 k = kAuk ≥ βkuk and hence kuk ≤ β1 kf k. The proof is complete. 

For the bilinear forms B1 , B2 defined above, one can easily show the boundedness:
|Bj (u, v)| ≤ αkukkvk
for all u, v in the respective Hilbert spaces H = H01 (Ω) or H = H01 (Ω; RN ) for j = 1, 2.
The strong positivity (also called coercivity) for both B1 and B2 is not always guaranteed
and involves estimating on the quadratic form Bj (u, u); this is usually called Gårding’s
estimates. We will derive these estimates for both of them and state the corresponding
existence theorems below.

3.3. Gårding’s Estimates and Existence Theory

In the following, we assume all coefficients involved in the problems are in L∞ (Ω).

3.3.1. Theory for B1 .

Theorem 3.2. Assume the ellipticity condition (3.2) holds. Then, there are constants
β > 0 and γ ≥ 0 such that
(3.13) B1 (u, u) ≥ βkuk2 − γkuk2L2 (Ω)
for all u ∈ H = H01 (Ω).

Proof. Note that, by the ellipticity,

Z X n n
Z X
B1 (u, u) − ( bi Di u + cu)u dx ≥ θ |Di u|2 dx.
Ω i=1 Ω i=1

Let m = max{kbi kL∞ (Ω) | 1 ≤ i ≤ n}. Then

|(bi Di u, u)2 | ≤ mkDi uk2 kuk2
≤ (m/2)(εkDi uk22 + (1/ε)kuk22 )
where in the last step we used the arithmetic-geometric inequality |αβ| ≤ (ε/2)α2 +(1/2ε)β 2 .
Combining the estimates we find
B1 (u, u) ≥ (θ − mε/2)kDuk2L2 (Ω) − (k0 + mn/2ε)kuk2L2 (Ω) .
By choosing ε > 0 so that c − mε > 0 we arrive at the desired inequality, using the Poincare
inequality. 
Theorem 3.3. There is a number γ ≥ 0 such that for each λ ≥ γ and for each function
f ∈ L2 (Ω), the boundary value problem
Lu + λu = f (x) in Ω, u = 0 on ∂Ω
has a unique weak solution u ∈ H = H01 (Ω) which satisfies
kukH ≤ ckf k2
where the positive constant c is independent of f .
56 3. Existence Theory for Linear Problems

Proof. Take γ from (3.13), let λ ≥ γ and define the bilinear form
B λ (u, v) ≡ B1 (u, v) + λ(u, v)2 for all u, v ∈ H
which corresponds to the operator Lu + λu. Then B λ (u, v) satisfies the hypotheses of the
Lax-Milgram Theorem. 
Example 3.4. Consider the Neumann boundary value problem
∂u
(3.14) −∆u(x) = f (x) in Ω, = 0 on ∂Ω.
∂ν
A function u ∈ H 1 (Ω) is said to be a weak solution to (3.14 if
Z Z
(3.15) ∇u · ∇v dx = f v dx, ∀ v ∈ H 1 (Ω).
Ω Ω
Obviously,
R taking v ≡ 1 ∈ H 1 (Ω), a necessary condition to have a weak solution is
Ω f (x) dx = 0. We show this is also a sufficient condition for existence of the weak so-
lutions. Note that, if u is a weak solution, then u + c, for all constants c, is also a weak
solution. Therefore, to fix the constants, we consider the vector space
 Z 
H = u ∈ H 1 (Ω) 

u(x) dx = 0
Ω
equipped with inner product
Z
(u, v)H = ∇u · ∇v dx.
Ω
By the theorem on equivalent norms, it follows that H with this inner product, is indeed a
Hilbert space, and (f, u)L2 (Ω) is a bounded linear functional on H:
|(f, u)L2 (Ω) | ≤ kf kL2 (Ω) kvkL2 (Ω) ≤ kf kL2 (Ω) kvkH .
Hence the Riesz Representation Theorem implies that there exists a unique u ∈ H such
that
(3.16) (u, w)H = (f, w)L2 (Ω) , ∀ w ∈ H.
It follows that u is a weak solutionR to the Neumann problem since for any Rv ∈ H 1 (Ω) we
1
take w = v − c ∈ H, where c = |Ω| Ω vdx, in (3.16) and obtain (3.15) using Ω f dx = 0.

Example 3.5. Let us consider the nonhomogeneous Dirichlet boundary value problem
(3.17) −∆u = f in Ω, u|∂Ω = ϕ
where f ∈ L2 (Ω) and ϕ is the trace of a function w ∈ H 1 (Ω). Note that it is not sufficient
to just require that ϕ ∈ L2 (∂Ω) since the trace operator is not onto. If, for example,
ϕ ∈ C 1 (∂Ω), then ϕ has a C 1 extension to Ω̄, which is the desired w.
The function u ∈ H 1 (Ω) is called a weak solution of (3.17) if u − w ∈ H01 (Ω) and if
Z Z
∇u · ∇vdx = f vdx for all v ∈ H01 (Ω).
Ω Ω
Let u be a weak solution of (3.17) and set u = z + w. Then z ∈ H01 (Ω) satisfies
Z Z
(3.18) ∇z · ∇vdx = (f v − ∇v · ∇w)dx for all v ∈ H01 (Ω).
Ω Ω
Since the right hand side belongs to the dual space H01 (Ω)∗ , the Lax-Milgram theorem yields
the existence of a unique z ∈ H01 (Ω) which satisfies (3.18). Hence (3.17) has a unique weak
solution u.
3.3. Gårding’s Estimates and Existence Theory 57

Example 3.6. Now let us consider the boundary value (also called Dirichlet) problem for
the fourth order biharmonic operator:
∂u
∆2 u = f in Ω, u|∂Ω = |∂Ω = 0.
∂n
We take H = H02 (Ω). By the general trace theorem, H = H02 (Ω) = {v ∈ H 2 (Ω) : γ0 v =
γ1 v = 0}. Therefore, this space H is the right space for the boundary conditions.
Accordingly, for f ∈ L2 (Ω), a function u ∈ H = H02 (Ω) is a weak solution of the
Dirichlet problem for the biharmonic operator provided
Z Z
∆u∆vdx = f vdx ∀ v ∈ H.
Ω Ω

Consider the bilinear form Z

B(u, v) = ∆u∆vdx.
Ω
Its boundedness follows from the Cauchy-Schwarz inequality
|B(u, v)| ≤ k∆uk2 k∆vk2 ≤ dkuk2,2 kvk2,2 .
Furthermore, it can be shown that k∆uk2 defines a norm on H02 (Ω) which is equivalent to
the usual norm on H 2 (Ω). (Exercise!) Hence
B(u, u) = k∆uk22 ≥ ckuk22,2
and so, by the Lax-Milgram theorem (in fact, just the Riesz Representation Theorem), there
exists a unique weak solution u ∈ H.

Exercise 1. Denote by Hc1 the space

Hc1 = {u ∈ H 1 (Ω) : γ0 u = const}.
Note that the constant may be different for different u’s.
(a) Prove that Hc1 is complete.
(b) Prove the existence and uniqueness of a function u ∈ Hc1 satisfying
Z Z
(∇u · ∇v + uv)dx = f vdx for all v ∈ Hc1
Ω Ω

where f ∈ C(Ω̄).
(c) If u ∈ C 2 (Ω̄) satisfies the equation in (b), find the underlying BVP.

Exercise 2. Let Ω ⊂ Rn . Show that if u, v ∈ H02 (Ω), then

Z Xn Z
∆u∆vdx = Dij uDij vdx.
Ω i,j=1 Ω

Hence, deduce that k∆uk2 defines a norm on H02 (Ω) which is equivalent to the usual norm
on H02 (Ω).

Exercise 3. Let Ω = (1, +∞). Show that the BVP −u00 = f ∈ L2 (Ω), u ∈ H01 (Ω) does
not have a weak solution.
58 3. Existence Theory for Linear Problems

3.3.2. Theory for B2 . We will derive the Gårding estimates for B2 (u, u). For simplicity,
let H = H01 (Ω; RN ) and let (u, v)H and kukH be the inner product and norm defined above
on H. Define the bilinear form of the leading terms by
X n X N Z
A(u, v) = akl l k
ij (x) Dj u Di v dx.
i,j=1 k,l=1 Ω

Theorem 3.7. Assume that either coefficients akl kl

ij satisfy the Legendre condition or aij
are all constants and satisfy the Legendre-Hadamard condition. Then
A(u, u) ≥ ν kuk2H , ∀ u ∈ H.

Proof. In the first case, the conclusion follows easily from the Legendre condition. We
prove the second case when Aklij are constants satisfying the Legendre-Hadamard condition
n X
X N
akl k l 2 2
ij q q pi pj ≥ ν |p| |q| , ∀p ∈ Rn , q ∈ RN .
i,j=1 k,l=1

We prove
n X
X N Z Z
A(u, u) = akl l k
ij Dj u Di u dx ≥ ν |Du|2 dx
i,j=1 k,l=1 Ω Ω

for all u ∈ C0∞ (Ω; RN ). For these test functions u, we extend them onto Rn by zero outside
Ω and thus consider them as functions in C0∞ (Rn ; RN ). Define the Fourier transforms
for such functions u by
Z
−n/2
û(y) = (2π) e−i y·x u(x) dx; y ∈ Rn .
Rn
Then, for any u, v ∈ C0∞ (Rn ; RN ),
Z Z
u(x) · v(x) dx = û(y) · v̂(y) dy,
Rn Rn

D[ k ck
j u (y) = i yj u (y);

the last identity can also be written as Du(y) c = i û(y) ⊗ y. Now, using these identities, we
have Z Z
akl
ij D i uk
(x) D j ul
(x) dx = akl [k [l
ij Di u (y) Dj u (y) dy
n n
Z R RZ 
kl k l kl k l
= aij yi yj u (y) u (y) dy = Re
c b aij yi yj u (y) u (y) dy .
c b
Rn Rn
Write û(y) = η + iξ with η, ξ ∈ RN . Then
 
Re u (y) u (y) = η k η l + ξ k ξ l .
ck bl

Therefore, by the Legendre-Hadamard condition,

Xn X N  
Re aij yi yj u (y) u (y) ≥ ν |y|2 (|η|2 + |ξ|2 ) = ν |y|2 |û(y)|2 .
kl ck bl

i,j=1 k,l=1

Hence,
n X
X N Z
A(u, u) = akl k l
ij Di u (x) Dj u (x) dx
n
i,j=1 k,l=1 R
3.3. Gårding’s Estimates and Existence Theory 59

n X
X N Z 
= Re akl
ij yi yj
ck (y) ubl (y) dy
u
i,j=1 k,l=1 Rn
Z Z
2 2
≥ν |y| |û(y)| dy = ν |iû(y) ⊗ y|2 dy
Rn n
Z ZR
2
=ν |Du(y)|
c dy = ν |Du(x)|2 dx.
Rn Rn
The proof is complete. 
Theorem 3.8. Let B2 (u, v) be defined by (3.7). Assume
1) akl
ij ∈ C(Ω̄),
2) the Legendre-Hadamard condition holds for all x ∈ Ω; that is,
akl k l 2 2
ij (x) q q pi pj ≥ ν |p| |q| , ∀p ∈ Rn , q ∈ RN .

3) bkl kl kl ∞
i , c , di ∈ L (Ω).
Then, there exist constants λ0 > 0 and λ1 ≥ 0 such that
B2 (u, u) ≥ λ0 kuk2H − λ1 kuk2L2 , ∀ u ∈ H01 (Ω; RN ).

Proof. By uniform continuity, we can choose a small  > 0 such that

ν
|akl kl
ij (x) − aij (y)| ≤ , ∀ x, y ∈ Ω̄, |x − y| ≤ .
2
We claim
Z Z
ν
(3.19) akl
ij (x) D i uk
D j ul
dx ≥ |Du(x)|2 dx
Ω 2 Ω
for all test functions u ∈ C0∞ (Ω; RN ) with the diameter of the support diam(supp u) ≤ .
To see this, we choose any point x0 ∈ supp u. Then
Z Z
kl k l
aij (x) Di u Dj u dx = akl k l
ij (x0 ) Di u Dj u dx
Ω Ω
Z
akl kl k l


+ ij (x) − aij (x0 ) Di u Dj u dx
supp u
Z Z
2 ν
≥ν |Du(x)| dx − |Du(x)|2 dx,
Ω 2 Ω
which proves (3.19). We now cover Ω̄ with finitely many open balls {B/4 (xm )} with xm ∈ Ω
and m = 1, 2, ..., M. For each m, let ζm ∈ C0∞ (B/2 (xm )) with ζm (x) = 1 for x ∈ B/4 (xm ).
Since for any x ∈ Ω̄ we have at least one m such that x ∈ B/4 (xm ) and thus ζm (x) = 1, we
may therefore define
ζm (x)
ϕm (x) = PM 2
1/2 , m = 1, 2, ..., M.
j=1 ζj (x)
PM 2
Then m=1 ϕm (x) = 1 for all x ∈ Ω. (This is a special case of partition of unity.) We have
thus
M 
X 
(3.20) akl k l
ij (x) Di u Dj u = akl 2 k
ij (x) ϕm Di u Dj u
l

m=1
and each term (no summation on m)
akl 2 k l kl k l
ij (x) ϕm Di u Dj u = aij (x) Di (ϕm u ) Dj (ϕm u )
60 3. Existence Theory for Linear Problems

 
−akl
ij (x) ϕ m Di ϕm ul
D i uk
+ ϕ m D i ϕm uk
Dj ul
+ D i ϕ m D j ϕ m uk l
u .
Since ϕm u ∈ C0∞ (Ω ∩ B/2 (xm ); RN ) and diam(Ω ∩ B/2 (xm )) ≤ , we have by (3.19)
Z Z
ν
akl
ij (x) D (ϕ
i m uk
) D j (ϕ m ul
) dx ≥ |D(ϕm u)|2 dx.
Ω 2 Ω
Note also that
|D(ϕm u)|2 = ϕ2m |Du|2 + |Dϕm |2 |u|2 + 2 ϕm Di ϕm uk Di uk .
Therefore, we have by (3.20) and the fact that M 2
P
m=1 ϕm = 1 on Ω,
Z
akl k l
ij (x) Di u Dj u dx
Ω
Z
ν
≥ |Du|2 dx − C1 kukL2 kDukL2 − C2 kuk2L2 .
2 Ω
The terms in B2 (u, u) involving b, c and d can be estimated by kukL2 kDukL2 and kuk2L2 .
Finally, by all of these estimates and the inequality
1 2
ab ≤ a2 + b
4
we have B2 (u, u) ≥ λ0 kuk2H −λ1 kuk2L2 for all u ∈ H01 (Ω; RN ). This completes the proof. 

Note that the bilinear form B λ (u, v) = B2 (u, v) + λ (u, v)L2 satisfies the condition of
the Lax-Milgram theorem on H = H01 (Ω; RN ) for all λ ≥ λ1 ; thus, by the Lax-Milgram
theorem, we easily obtain the following existence result.
Theorem 3.9. Under the hypotheses of the previous theorem, for λ ≥ λ1 , the Dirichlet
problem for the linear system
(3.21) − div(A(x, u, Du)) + b(x, u, Du) + λu = F, u|∂Ω = 0
has a unique weak solution u in H01 (Ω; RN )
for any bounded linear functional F on H.
Moreover, the solution u satisfies kukH ≤ C kF k with a constant C depending on λ and the
L∞ -norms of the coefficients of linear terms A(x, s, ξ) and b(x, s, ξ) given above.
Corollary 3.10. Given λ ≥ λ1 as in the theorem, the operator K : L2 (Ω; RN ) → L2 (Ω; RN ),
where, for each F ∈ L2 (Ω; RN ), u = KF is the unique weak solution to the BVP (3.21)
above, is a compact linear operator.

Proof. By the theorem, kukH01 (Ω;RN ) ≤ CkF kL2 (Ω;RN ) . Hence K is a bounded linear opera-
tor from L2 (Ω; RN ) to H01 (Ω; RN ), which, by the compact embedding theorem, is compactly
embedded in L2 (Ω; RN ). Hence, as a linear operator from L2 (Ω; RN ) to L2 (Ω; RN ), K is
compact. 

3.4. Symmetric Elliptic Operators

3.4.1. Symmetric Elliptic Operators. In what follows, we assume Ω is a bounded
domain. We consider the operator
Lu = − div A(x, Du) + c(x)u, u ∈ H01 (Ω; RN ),
where A(x, Du) is a linear system defined with A(x, ξ), ξ ∈ MN ×n , given by
X
Aki (x, ξ) = akl l
ij (x)ξj .
1≤l≤N, 1≤j≤n
3.4. Symmetric Elliptic Operators 61

Here akl ∞
ij (x) and c(x) are given functions in L (Ω).
The bilinear form associated to L is
 
Z XN X n
(3.22) B(u, v) =  akl l k
ij (x)Dj u Di v + c(x)u · v
 dx, u, v ∈ H01 (Ω; RN ).
Ω k,l=1 i,j=1

In order for B to be symmetric on H01 (Ω; RN ), that is, B(u, v) = B(v, u) for all u, v ∈
H01 (Ω; RN ), we need the following symmetry condition:
(3.23) akl lk
ij (x) = aji (x), ∀ i, j = 1, 2, · · · , n; k, l = 1, 2, · · · , N.
We also assume the Gårding inequality holds (see Theorem 3.8 for sufficient conditions):
(3.24) B(u, u) ≥ σkuk2H 1 − µkuk2L2 , ∀ u ∈ H01 (Ω; RN ),
0

where σ > 0 and µ ∈ R are constants.

3.4.2. The Compact Inverse. For each F ∈ L2 (Ω; RN ), define u = KF to be the unique
weak solution in H01 (Ω; RN ) of the BVP
Lu + µu = F in Ω, u|∂Ω = 0.
By Theorem 3.9 and Corollary 3.10, this K is well defined and is a compact linear operator
on L2 (Ω; RN ). Sometime, we write K = (L+µI)−1 . Here I denotes the identity on L2 (Ω; RN )
and also the identity embedding of H01 (Ω; RN ) into L2 (Ω; RN ).
Theorem 3.11. K : L2 (Ω; RN ) → L2 (Ω; RN ) is symmetric and positive; that is,
(KF, G)L2 = (KG, F )L2 , (KF, F )L2 ≥ 0, ∀ F, G ∈ L2 (Ω; RN ).
Furthermore, given λ ∈ R and F ∈ L2 (Ω; RN ), u ∈ H01 (Ω; RN ) is a weak solution of
Lu − λu = F if and only if [I − (λ + µ)K]u = KF.

Proof. Let u = KF and v = KG. Then

(u, G)L2 = B(v, u) + µ(v, u)L2 = B(u, v) + µ(v, u)L2 = (v, F )L2 ,
proving the symmetry. Also, by (3.24),
(3.25) (KF, F )L2 = (u, F )L2 = B(u, u) + µkuk2L2 ≥ σkuk2H 1 = σkKF k2H 1 ≥ 0.
0 0

Finally, u ∈ H01 (Ω; RN )

is a weak solution of Lu − λu = F if and only if Lu + µu =
F + (λ + µ)u, which is equivalent to the equation u = K[F + (λ + µ)u] = KF + (λ + µ)Ku;
that is, [I − (λ + µ)K]u = KF. 

3.4.3. Orthogonality Conditions. From Theorem 3.11 above, the BVP

(3.26) Lu = F, u|∂Ω = 0
has a solution if and only if KF ∈ R(I − µK) = [N (I − µK)]⊥ .
If µ = 0, the Lax-Milgram theorem already implies that the problem (3.26) has a unique
weak solution u = KF. If µ 6= 0, then it is easy to see that
N = N (I − µK) = {v ∈ H01 (Ω; RN ) | Lv = 0}.
Note that, by the Fredholm Alternative Theorem, this null space N is of finite dimension
k, and let {v1 , v2 , · · · , vk } be a basis of this null space. Then (3.26) is solvable if and only
if the following orthogonality condition holds:
Z
F · Gi dx = 0, i = 1, 2, · · · , k,
Ω
62 3. Existence Theory for Linear Problems

where Gi = Kvi ; that is, Gi is the unique weak solution to the BVP: LGi + µGi = vi , in
H01 (Ω; RN ).

3.4.4. Eigenvalue Problems. A number λ ∈ R is called a (Dirichlet) eigenvalue of

operator L if the BVP problem

Lu − λu = 0, u|∂Ω = 0

has nontrivial weak solutions in H01 (Ω; RN ); these nontrivial solutions are called the eigen-
functions corresponding to eigenvalue λ.
From Theorem 3.11, we see that λ is an eigenvalue of L if and only if equation (I − (λ +
1
µ)K)u = 0 has nontrivial solutions u ∈ L2 (Ω; RN ); this exactly says that λ 6= −µ and λ+µ
is an eigenvalue of operator K. Since, by (3.25), K is strictly positive, all eigenvalues of K
consist of a countable set of positive numbers tending to zero and hence the eigenvalues of
L consist of a set of numbers {λj }∞j=1 with −µ < λ1 ≤ λ2 ≤ · · · ≤ λj → ∞.

Theorem 3.12. (Eigenvalue Theorem) Assume (3.23) and (3.24) with µ = 0. Then the
eigenvalues of L consist of a countable set Σ = {λk }∞
k=1 , where

0 < λ1 ≤ λ2 ≤ λ3 ≤ · · ·

and
lim λk = ∞.
k→∞

Note that λ1 is called the (Dirichlet) principal eigenvalue of L. Let wk be an eigenfunction

corresponding to λk satisfying kwk kL2 (Ω;RN ) = 1. Then {wk }∞k=1 forms an orthonormal
2 N
basis of L (Ω; R ).

Note that

(3.27) λ1 = min B(u, u).

u∈H01 (Ω;RN ), kukL2 =1

Theorem 3.13. Let N = 1 and let w1 be an eigenfunction corresponding to the principal

eigenvalue λ1 . Then, either w1 (x) > 0 for all x ∈ Ω or w1 (x) < 0 for all x ∈ Ω. Moreover,
the eigenspace corresponding to λ1 is one-dimensional.

Proof. The first part of the theorem relies on the characterization (3.27) and a maximum
principle which we do not study here. We prove only the second part. Let w be another
eigenfunction. Then, either w(x) > 0 for all x ∈ Ω or w(x) < 0 for all x ∈ Ω. Let t ∈ R be
such that
Z Z
w(x) dx = t w1 (x)dx.
Ω Ω

Note that u = w − tw1 is also a solution to Lu = λ1 u. We claim u ≡ 0 and hence w = tw1 ,

proving the eigenspace is one-dimensional. Suppose u 6= 0. Then u is another eigenfunction
corresponding to λ1 . Then, by theR theorem, we would have either u(x) > 0 for all x ∈ Ω or
u(x) < 0 for all x ∈ Ω and hence Ω u(x)dx 6= 0, which is a contradiction. 
3.4. Symmetric Elliptic Operators 63

Remark. Why is the eigenvalue problem important? In one specific case, one can use
eigenvalues and eigenfunctions to study some evolution (i.e. time-dependent) problems. For
example, to solve the initial boundary value problem (IBVP) for the time-dependent
parabolic equation:
(3.28) ut + Lu = 0, u|∂Ω = 0, u(x, 0) = ϕ(x),
one can try to find the special solutions of the form: u(x, t) = e−λt w(x); this reduces to the
eigen-problem: Lw = λw. Therefore, for each pair (λi , wi ) of eigenvalue and eigenfunction,
one obtains a special solution to the evolution equation given by ui (x, t) = e−λi t wi (x). Then
one proceeds to solve the IBVP (3.28) by finding the solution of the form
X
u(x, t) = ai e−λi t wi (x),
i
where ai is determined by the eigen-expansion of the initial data ϕ:
X
ϕ(x) = ai wi (x).
i
However, this course does not study the parabolic or other time-dependent problems.
 Chapter 4

Variational Methods
for PDEs

4.1. Variational Problems

This chapter and the next will discuss some new methods for solving the boundary value
problem for some partial differential equations. All these problems can be written in the
abstract form:

(4.1) A[u] = 0 in Ω, B[u] = 0 on ∂Ω,

where A[u] denotes a given PDE for unknown u and B[u] is a given boundary value condition.
There is, of course, no general theory for solving such problems.
The Calculus of Variations identifies an important class of problems which can be solved
using relatively simple techniques from nonlinear functional analysis. This is the class of
variational problems, where the operator A[u] can be formulated as the first variation
(“derivative”) of an appropriate “energy” functional I(u) on a Banach space X; that is,
A[u] = I 0 (u). In this way, A : X → X ∗ and the equation A[u] = 0 can be formulated weakly
as

hI 0 (u), vi = 0, ∀ v ∈ X.

The advantage of this new formulation is that solving problem (4.1) (at least weakly) is
equivalent to finding the critical points of I on X. The minimization method for a
variational problem is to solve the problem by finding the minimizers of the related energy
functional. In this chapter and the next, we shall only study the variational problems on
the Sobolev space X = W 1,p (Ω; RN ).
We should also mention that many of the physical laws in applications arise directly as
variational principles. However, although powerful, not all PDE problems can be formulated
as variational problems; there are lots of other (nonvariational) important methods for
studying PDEs.

65
66 4. Variational Methods for PDEs

4.2. Multiple Integrals in the Calculus of Variations

Consider the multiple integral functional
Z
(4.2) I(u) = F (x, u(x), Du(x)) dx,
Ω
where F (x, s, ξ) is a given function on Ω × RN × MN ×n .

4.2.1. First Variation and Euler-Lagrange Equations. Suppose F (x, s, ξ) is contin-

uous and is also smooth in s and ξ. Assume u is a nice (say, u ∈ C 1 (Ω̄; RN )) minimizer of
I(u) with its own boundary data; that is, u is a map such that
I(u) ≤ I(u + tϕ)
for all t ∈ R1and ϕ ∈ C0∞ (Ω; RN ).
Then by taking derivative of I(u + t ϕ) at t = 0 we see
that u satisfies
Z
Fξk (x, u, Du) Di ϕk (x) + Fsk (x, u, Du) ϕk (x) dx = 0


(4.3)
i
Ω
for all ϕ ∈ C0∞ (Ω; RN ). (Summation notation is used here.) The left-hand side is called
the first variation of I at u. Since this holds for all test functions, we conclude after
integration by parts:
(4.4) − div A(x, u, Du) + b(x, u, Du) = 0,
where A, b are defined by
(4.5) Aki (x, s, ξ) = Fξk (x, s, ξ), bk (x, s, ξ) = Fsk (x, s, ξ).
i

The coupled differential system in divergence form for u is called the system of Euler-
Lagrange equations for the functional I(u).

4.2.2. Second Variation and Legendre-Hadamard Conditions. If F, u are suffi-

ciently smooth (e.g. of class C 2 ) then, at the minimizer u, we have
d2 
I(u + tϕ) ≥ 0.

dt2

t=0
This implies
(4.6)
Z h i
Fξk ξl (x, u, Du) Di ϕk Dj ϕl + 2Fξk sl (x, u, Du) ϕl Di ϕk + Fsk sl (x, u, Du) ϕk ϕl dx ≥ 0
i j i
Ω
for all ϕ ∈ C0∞ (Ω; RN ). The left-hand side of this inequality is called the second variation
of I at u.
We can extract useful information from (4.6). Note that routine approximation argu-
ment shows that (4.6) is also valid for all functions ϕ ∈ W01,∞ (Ω; RN ) (that is, all Lipschitz
functions vanishing on ∂Ω). Let ρ : R → R be the periodic zig-zag function of period 1
defined by
(4.7) ρ(t + 1) = ρ(t) (t ∈ R),

(4.8) ρ(t) = t if 0 ≤ t ≤ 12 ; ρ(t) = 1 − t 1

2 ≤ t ≤ 1.
if
Given any vectors p ∈ Rn , q ∈ RN and  > 0, define ϕ ∈ W01,∞ (Ω; RN ) by
x·p
ϕ(x) =  ρ( )ζ(x)q, ∀ x ∈ Ω,

4.2. Multiple Integrals in the Calculus of Variations 67

where ζ ∈ C0∞ (Ω) is a given scalar test function. Note that Di ϕk (x) = ρ0 ( x·p k
 )pi q ζ + O()
as  → 0+ . Substitute this ϕ into (4.6) and let  → 0+ and we obtain
 
Z Xn X N
 Fξk ξl (x, u, Du) pi pj q k q l  ζ 2 dx ≥ 0.
i j
Ω i,j=1 k,l=1

Since this holds for all ζ ∈ C0∞ (Ω), we deduce

n X
X N
(4.9) Fξk ξl (x, u, Du) pi pj q k q l ≥ 0, ∀ x ∈ Ω, p ∈ Rn , q ∈ RN .
i j
i,j=1 k,l=1

This is the weak Legendre-Hadamard condition for F at the minimum point u.

4.2.3. Ellipticity Conditions and Convexities. We now consider the ellipticity of the
Euler-Lagrange equation (4.4), where A(x, s, ξ), b(x, s, ξ) are given by (4.5) and F (x, s, ξ) is
C 2 in ξ. In this case, the Legendre ellipticity condition (3.8) and the Legendre-Hadamard
condition (3.9) defined in the previous chapter reduce to, respectively:

(4.10) Fξk ξl (x, s, ξ) ηik ηjl ≥ ν |η|2 ∀η ∈ MN ×n ;

i j

(4.11) Fξk ξl (x, s, ξ) q k q l pi pj ≥ ν|p|2 |q|2 ∀q ∈ RN , p ∈ Rn .

i j

Obviously, (4.10) implies (4.11). We also have the following equivalent conditions.
Lemma 4.1. Let F be C 2 in ξ. Then, the conditions (4.10) and (4.11) are equivalent to
the following conditions, respectively:
ν
(4.12) F (x, s, η) ≥ F (x, s, ξ) + Fξk (x, s, ξ) (ηik − ξik ) + |η − ξ|2 ;
i 2
ν 2 2
(4.13) F (x, s, ξ + q ⊗ p) ≥ F (x, s, ξ) + Fξk (x, s, ξ) pi q k + |p| |q|
i 2
for all x ∈ Ω, s, q ∈ RN , ξ, η ∈ MN ×n and p ∈ Rn .

Proof. Let ζ = η − ξ and f (t) = F (x, s, ξ + tζ). Then, by Taylor’s formula,

Z 1
0
f (1) = f (0) + f (0) + (1 − t) f 00 (t) dt.
0
Note that
f 0 (t) = Fξk (x, s, ξ + tζ) ζik , f 00 (t) = Fξk ξl (x, s, ξ + tζ) ζik ζjl .
i i j

From this and the Taylor formula, inequalities (4.12) and (4.13) are equivalent to (4.10)
and (4.11), respectively. 

A function F (x, s, ξ) is said to be convex in ξ ∈ MN ×n if

F (x, s, tξ + (1 − t)η) ≤ tF (x, s, ξ) + (1 − t)F (x, s, η)
for all x, s, ξ, η and 0 ≤ t ≤ 1. While F (x, s, ξ) is said to be rank-one convex in ξ if the
function f (t) = F (x, s, ξ + t q ⊗ p) is convex in t ∈ R1 for all x, s, ξ and q ∈ RN , p ∈ Rn .
Obviously, a convex function is always rank-one convex.
We easily have the following result.
68 4. Variational Methods for PDEs

Lemma 4.2. Let F (x, s, ξ) be C 2 in ξ. Then the convexity of F (x, s, ξ) in ξ is equivalent

to (4.10) with ν = 0, while the rank-one convexity of F (x, s, ξ) in ξ is equivalent to (4.11)
with ν = 0.

Remark. Rank-one convexity does not imply convexity. For example, take n = N ≥ 2, and
F (ξ) = det ξ. Then F (ξ) is rank-one convex but not convex in ξ. (Exercise!) Later on, we
will study other convexity conditions related to the energy functionals given by (4.2).

4.2.4. Structural Conditions. Distributional solutions to the Euler-Lagrange equations

(4.4) can be defined as long as A(x, u, Du) and b(x, u, Du) are in L1loc (Ω; RN ), but we need
some structural conditions on F (x, s, ξ) so that the weak solutions to the BVP
(4.14) − div A(x, u, Du) + b(x, u, Du) = 0 in Ω, u = ϕ on ∂Ω,
can be defined and studied in W 1,p (Ω; RN ). These conditions are also sufficient for the
functional I to be Gateaux-differentiable on W 1,p (Ω; RN ).
Standard Growth Conditions. We assume F (x, s, ξ) is C 1 in (s, ξ) and
(4.15) |F (x, s, ξ)| ≤ c1 (|ξ|p + |s|p ) + c2 (x), c2 ∈ L1 (Ω);
p
(4.16) |Ds F (x, s, ξ)| ≤ c3 (|ξ|p−1 + |s|p−1 ) + c4 (x), c4 ∈ L p−1 (Ω);
p
(4.17) |Dξ F (x, s, ξ)| ≤ c5 (|ξ|p−1 + |s|p−1 ) + c6 (x), c6 ∈ L p−1 (Ω),
where c1 , c3 , c5 are constants.
Theorem 4.3. Under the standard conditions above, the functional I : W 1,p (Ω; RN ) → R
is Gateaux-differentiable and, for u, v ∈ W 1,p (Ω; RN ), the directional derivative hI 0 (u), vi
is exactly given by
Z
0
Fξk (x, u, Du) Di v k (x) + Fsk (x, u, Du) v k (x) dx


(4.18) hI (u), vi =
i
Ω
(as usual, summation notation is used here).

Proof. Given u, v ∈ X = W 1,p (Ω; RN ), let h(t) = I(u + tv). Then, by (4.15), h is finite
valued, and for t 6= 0 we have
h(t) − h(0) F (x, u + tv, Du + tDv) − F (x, u, Du)
Z Z
= dx ≡ F t (x) dx,
t Ω t Ω
where for almost every x ∈ Ω,
Z t
t 1 1 d
F (x) = [F (x, u + tv, Du + tDv) − F (x, u, Du)] = F (x, u + sv, Du + sDv) ds.
t t 0 ds
Clearly,
lim F t (x) = Fξk (x, u, Du) Di v k (x) + Fsk (x, u, Du) v k (x) a.e.
t→0 i

We also write
1 th
Z i
F t (x) = Fξk (x, u + sv, Du + sDv) Di v k (x) + Fsk (x, u + sv, Du + sDv) v k (x) ds.
t 0 i

ap bq 1 1
Using conditions (4.16), (4.17), and Young’s inequality: ab ≤ p + q , where p + q = 1, we
obtain that, for all 0 < |t| ≤ 1,
|F t (x)| ≤ C1 (|Du|p + |Dv|p + |u|p + |v|p ) + C2 (x), C2 ∈ L1 (Ω).
4.2. Multiple Integrals in the Calculus of Variations 69

Hence, by the Lebesgue dominated convergence theorem,

Z Z h i
h0 (0) = lim F t (x) dx = Fξk (x, u, Du) Di v k (x) + Fsk (x, u, Du) v k (x) dx.
t→0 Ω Ω
i

This proves the theorem. 

Dirichlet Classes. Given ϕ ∈ W 1,p (Ω; RN ), we define the Dirichlet class of ϕ to be the set
Dϕ = {u ∈ W 1,p (Ω; RN ) | u − ϕ ∈ W01,p (Ω; RN )}.

4.2.5. Weak Solutions of Euler-Lagrange Equations. Under the standard growth

conditions, we say u ∈ W 1,p (Ω; RN ) is a weak solution to the BVP of Euler-Lagrange
equations (4.14) if u ∈ Dϕ and
Z
Fξk (x, u, Du) Di v k (x) + Fsk (x, u, Du) v k (x) dx = 0


(4.19)
i
Ω

for all v ∈ W01,p (Ω; RN ).

Theorem 4.4. Under the standard growth conditions above, any minimizer u ∈ Dϕ of
I(u) = min I(v)
v∈Dϕ

is a weak solution of the BVP for the Euler-Lagrange equation (4.14).

Proof. This follows from Theorem 4.3. 

4.2.6. Minimality and Uniqueness of Weak Solutions. We study the weak solutions
of Euler-Lagrange equations under the hypotheses of convexity and certain growth condi-
tions. For this purpose, we consider a simple case where F (x, s, ξ) = F (x, ξ) satisfies, for
some 1 < p < ∞,

(4.20) |Fξk (x, ξ)| ≤ µ (χ(x) + |ξ|p−1 ) ∀ x ∈ Ω, ξ ∈ MN ×n ,

i
p
where µ > 0 is a constant and χ ∈ L p−1 (Ω) is some function. Let
Z
I(u) = F (x, Du(x)) dx.
Ω

Theorem 4.5. Let F (x, ξ) be C2and convex in ξ. Let u ∈ W 1,p (Ω; RN ) be a weak solution
of the Euler-Lagrange equation of I and I(u) < ∞. Then u must be a minimizer of I in
the Dirichlet class Du of W 1,p (Ω; RN ). Furthermore, if F satisfies the Legendre condition
(4.10) (with ν > 0), then u is the unique minimizer of I in Du .

Proof. Since u is a weak solution of the Euler-Lagrange equation of I, it follows that

Z
(4.21) Fξk (x, Du(x)) Di v k dx = 0
i
Ω
p
for all v ∈ W01,p (Ω; RN ). The growth condition (4.20) implies Fξk (x, Du) ∈ L p−1 (Ω). Now
i
let v ∈ W 1,p (Ω; RN ) with v − u ∈ W01,p (Ω; RN ). By the convexity condition, we have
ν
(4.22) F (x, η) ≥ F (x, ξ) + Fξk (x, ξ) (ηik − ξik ) + |η − ξ|2 , ∀ ξ, η,
i 2
70 4. Variational Methods for PDEs

where ν = 0 if F is only convex in ξ and ν > 0 if F satisfies the Legendre condition. This
implies
Z Z Z
F (x, Dv) dx ≥ F (x, Du) dx + Fξk (x, Du) Di (v k − uk ) dx
i
Ω Ω Ω
Z
ν
+ |Du − Dv|2 dx.
2 Ω
Since u is a weak solution, we have
Z
Fξk (x, Du) Di (v k − uk ) dx = 0.
i
Ω
Therefore, it follows that
Z
ν
(4.23) I(v) ≥ I(u) + |Du − Dv|2 dx ≥ I(u)
2 Ω

for all v ∈W 1,p (Ω; RN )with v − u ∈ W01,p (Ω; RN ).

This shows that u is a minimizer of I in
the Dirichlet class Du .
Now assume ν > 0 and v ∈ Du is another such minimizer of I. Then from (4.23) we easily
obtain Du = Dv in Ω and hence v ≡ u since u − v ∈ W01,p (Ω; RN ). The proof is now
completed. 

4.3. Direct Method for Minimization

4.3.1. Weak Lower Semicontinuity. Let X = W 1,p (Ω; RN ). A set C ⊂ X is called
weakly closed if {uν } ⊂ C, uν * u implies u ∈ C. For instance, by Theorem 1.38, all
Dirichlet classes Dϕ are weakly closed.
A functional I : X → R̄ is called weakly lower semicontinuous(w.l.s.c.) on X if for
every u0 ∈ X and every sequence {uν } weakly convergent to u0 in X it follows that
I(u0 ) ≤ lim inf I(uν ).
ν→∞

I is called weakly coercive on a (unbounded) set C in X if I(u) → ∞ as kuk → ∞ on C.

4.3.2. Direct Method in the Calculus of Variations. The following theorem is a

special case of the generalized Weierstrass theorem.
Theorem 4.6. (Existence of Minimizers) Let I : C ⊆ X = W 1,p (Ω; RN ) → R be w.l.s.c.
and weakly coercive on a nonempty weakly closed set C in X. Assume 1 < p < ∞. Then
there is at least one u0 ∈ C such that I(u0 ) = inf u∈C I(u). In this case we say u0 ∈ C is a
minimizer of I on C.

Proof. This theorem is a special case of Theorem 1.37 under assumption (ii) there; the proof
involves the basic ideas of what is known as the direct method of calculus of variations. We
explain this method by giving a proof of this theorem. First of all, by the weak coercivity,
one can easily show that inf u∈C I(u) is finite; so take a sequence {uν }, called a minimizing
sequence, such that
lim I(uν ) = inf I(u).
ν→∞ u∈C
Then the weak coercivity condition implies that {uν } must be bounded in X. Since 1 <
p < ∞ and thus X = W 1,p (Ω; RN ) is reflexive, there exists a subsequence of {uν }, denoted
4.3. Direct Method for Minimization 71

by {uνj }, and u0 ∈ X such that uνj * u0 weakly in X. The weak closedness of C implies
u0 ∈ C. Now the w.l.s.c. of I implies
I(u0 ) ≤ lim inf I(uνj ) = inf I(u).
j→∞ u∈C

This implies u0 is a minimizer of I over C. 

4.3.3. An Example: p-Laplace Equations. As an example, we consider the BVP for

nonlinear p-Laplace equations (p ≥ 2):
n
X
(4.24) − Di (|∇u|p−2 Di u) + f (x, u) = 0 in Ω, u = 0 on ∂Ω,
i=1
where f : Ω × R → R be a given function satisfying the Carathéodory property, i.e.,
for every s ∈ R, f (x, s) (as a function of x) is measurable on Ω, and for almost all x ∈ Ω,
f (x, s) (as a function of s) is continuous on R. Note that when p = 2, (4.24) becomes the
semilinear elliptic problem.
Define Z s
F (x, s) = f (x, t) dt
0
and the functional I on X = W01,p (Ω) by
Z  
1 p
I(u) = |∇u| + F (x, u) dx.
Ω p
Assume the following growth conditions are satisfied:
(4.25) F (x, s) ≥ −c1 |s| − c2 (x),
(4.26) |F (x, s)| ≤ c3 (x) + c4 |s|p ,
(4.27) |f (x, s)| ≤ c5 (x) + c6 |s|p−1
for all x ∈ Ω and s ∈ R, where c1 , c4 , c6 are nonnegative constants, and c2 , c3 ∈ L1 (Ω), and
p
c5 ∈ L p−1 (Ω) are given functions. Note that (4.27) implies (4.26).
Theorem 4.7. Under these conditions, the functional I has a minimizer on X = W01,p (Ω)
and hence (4.24) has a weak solution.

Proof. We only need to check I is w.l.s.c and weakly coercive on X. We write

Z Z
1
I(u) = I1 (u) + I2 (u) = |∇u|p dx + F (x, u) dx.
p Ω Ω
Note that pI1 (u) = kukp1,p,0 on X which is w.l.s.c. And since the embedding X ⊂ Lp (Ω)
is always compact, by (4.26), we can show that I2 is in fact continuous under the weak
convergence. Hence I is w.l.s.c. on X. By (4.25), we have
1 1
I(u) ≥ kukp1,p,0 − c1 kukL1 (Ω) − C ≥ kukp1,p,0 − ckuk1,p,0 − C,
p p
and hence I(u) → ∞ if kukX = kuk1,p,0 → ∞. This proves the weak coercivity of I. The
result follows then from Theorem 4.6. 

Exercise 1. Let n ≥ 3 and 2 ≤ p < n. Show the theorem is valid if (4.27) above is
replaced by
|f (x, s)| ≤ a(x) + b |s|q ,
72 4. Variational Methods for PDEs

q+1
where b ≥ 0 is a constant, a ∈ L q (Ω) and 1 ≤ q < p∗ − 1.

4.4. Minimization with Constraints

In some cases, we need to minimize functional I under certain constraints. If I is a multiple
integral functional on X = W 1,p (Ω; RN ) defined before, there may be constraints given in
terms of one of the following:
R
(4.28) J(u) = 0, where J(u) = Ω G(x, u) dx.

(4.29) h(u(x)) = 0, ∀ a.e. x ∈ Ω.

(4.30) M (Du(x)) = 0, ∀ a.e. x ∈ Ω.

All these lead to some PDEs involving the Lagrange multipliers. For different types of
constraints, the Lagrange multiplier comes in significantly different ways.

4.4.1. Nonlinear Eigenvalue Problems. The following theorem was proved in the Pre-
liminaries.
Theorem 4.8. Let X be a Banach space. Let f, g : X → R be of class C 1 and g(u0 ) = c.
Assume u0 is a local extremum of f with respect to the constraint g(u) = c. Then either
g 0 (u0 )v = 0 for all v ∈ X, or there exists λ ∈ R such that f 0 (u0 )v = λg 0 (u0 )v for all v ∈ X;
that is, u0 is a critical point of f − λg.

If u0 6= 0 then the corresponding λ is called an eigenvalue for the nonlinear eigenvalue

problem: f 0 (u) = λg 0 (u) and u0 is the corresponding eigenfunction.

We have following applications.

Theorem 4.9. Let k(x), l(x) ∈ C(Ω̄) with l(x) > 0 on Ω̄. Then, for each R ∈ (0, ∞), the
problem
n+2
(4.31) ∆u + k(x)u + λl(x)|u|τ −1 u = 0 in Ω, u|∂Ω = 0 (1 < τ < )
n−2
1 τ +1 dx
R
has a (weak) solution pair (uR , λR ) with τ +1 Ω l(x)|uR | = R.

Proof. Define the functionals

Z Z
1 1
f (u) = (|∇u|2 − k(x)u2 ) dx, g(u) = l(x)|u|τ +1 dx.
2 Ω τ +1 Ω
We showed earlier that f is w.l.s.c., and g is weakly continuous on H01 (Ω) (it is here that
n+2
the assumption τ < n−2 is used to guarantee that the embedding H01 (Ω) ⊂ Lτ +1 (Ω) be
compact). Now we show that f is weakly coercive relative to g on H01 (Ω); that is, weakly
coercive on the sets {u ∈ H01 (Ω) | g(u) ≤ c} for all constants c > 0. First of all, by Hölder’s
inequality we have
Z τ +1 Z
τ
|u| dx ≤ |Ω| |u|τ +1 dx
Ω Ω
|Ω|τ
Z
≤ l(x)|u|τ +1 dx ≤ Cg(u)
α Ω
4.4. Minimization with Constraints 73

where α = inf l(x). Hence kuk1 is uniformly bounded if g(u) ≤ const. Furthermore, by
Ehrling’s inequality, for each ε > 0, there is an absolute constant c(ε) > 0 such that

kuk22 ≤ εkuk21,2,0 + c(ε)kuk21 for all u ∈ H01 (Ω).

Thus
1 1
f (u) ≥ (1 − ε max |k|)kuk21,2,0 − c(ε) max |k|kuk21 .
2 2
If we choose ε small enough so that 1 − ε max |k| > 0, then f (u) → ∞ as kuk1,2,0 → ∞ with
g(u) ≤ const.
It easily follows that g 0 (u) = 0 iff u = 0. Moreover, since the range of the functional
g(u) is [0, ∞), we can apply Theorem 4.8 to conclude that for each R ∈ (0, ∞) there exists
a nontrivial weak solution (uR , λR ) of (4.31) with g(uR ) = R. Moreover, if βR = f (uR ) =
min f (u) subject to g(u) = R, then it is easily seen that λR = (τ2β R
+1)R . 

Corollary 4.10. For 1 < τ < (n + 2)/(n − 2) there exists a nontrivial weak solution of

(4.32) ∆u + |u|τ −1 u = 0 in Ω, u|∂Ω = 0.

Proof. By Theorem 4.9, there exist uR 6= 0, λR > 0 such that

∆uR + λR |uR |τ −1 uR = 0 in Ω, uR |∂Ω = 0.

Set uR = kv (k > 0 is to be determined). Then k∆v + λR k τ |v|τ −1 v = 0, and so if we

choose k to satisfy λR k τ −1 = 1, it follows that

∆v + |v|τ −1 v = 0 in Ω, v|∂Ω = 0

has a nontrivial weak solution in H01 (Ω). 

Remarks. (i) Another method to show the existence of nontrivial weak solution to (4.32) is
given later by the Mountain Pass Theorem.
n+2
(ii) The nonexistence of nontrivial classical solutions to problem (4.32) when τ ≥ n−2
will be studied later for certain domains with simple topological property.
(iii) For domains like annulus, problem (4.32) always has nontrivial solutions for all
τ > 1. See the following exercise.

Exercise 1. Let Ω be the annulus 0 < a < r < b, r = |x| and suppose τ > 1. Prove
that the BVP
∆u + |u|τ −1 u = 0 in Ω, u|∂Ω = 0
has a nontrivial solution.
Hint: Minimize the functional
Z b
f (u) = (u0 )2 rn−1 dr, r = |x|
a

for all u in the set

  Z b 
C= u∈ H01 (a, b) |u| τ +1 n−1
r dr = 1 .

a
74 4. Variational Methods for PDEs

4.4.2. Harmonic Maps and Liquid Crystals. We now consider the Dirichlet energy
Z
1
I(u) = |Du|2 dx
2 Ω
for u ∈ H 1 (Ω; RN ) with point-wise constraint |u(x)| = 1 for almost every x ∈ Ω. Let
C = {u ∈ Dϕ | |u(x)| = 1 a.e.},
where Dϕ is a Dirichlet class. We assume C is non-empty. Then C is weakly closed in
H 1 (Ω; RN ). We have the following result.
Theorem 4.11. There exists u ∈ C satisfying
I(u) = min I(v).
v∈C
Moreover, u satisfies
Z Z
(4.33) Du · Dv dx = |Du|2 u · v dx
Ω Ω
for each v ∈ H01 (Ω; RN ) ∩ ∞
L (Ω; RN ).

Remark. In this case, we see u is a weak solution to the harmonic map equation:
−∆u = |Du|2 u in Ω.
The Lagrange multiplier in this case is the function λ = |Du|2 , instead of a constant.

Proof. The existence of minimizers follows by the direct method as above. Given any
v ∈ H01 (Ω; RN ) ∩ L∞ (Ω; RN ), let  be such that ||kvkL∞ (Ω) ≤ 12 . Define
u(x) + v(x)
w (x) = , h() = I(w ).
|u(x) + v(x)|
Note that w ∈ C and h(0) = I(u) = minC I ≤ h() for sufficiently small ; hence, h0 (0) = 0.
Computing this h0 (0) and using the fact (Du)T u = 0 (which follows from |u| = 1) yield the
weak form of harmonic equation (4.33). 

Liquid Crystals. Let n = N = 3. Then the harmonic map Dirichlet energy can be considered
as a special case of the Oseen-Frank energy for liquid crystals defined earlier:
Z Z
I(u) = F (u, Du) dx = (κ1 (div u)2 + κ2 (u · curl u)2 + κ3 (u × curl u)2 ) dx
Ω
ZΩ
+ κ4 (tr((Du)2 ) − (div u)2 ) dx,
Ω
where Ω is a bounded domain in R3 ,
u : Ω → R3 , div u = trDu is the divergence of u and
curl u = ∇ × u denotes the curl vector of u in R3 . In the first part of the total energy
I(u), the κ1 -term represents the splay energy, κ2 -term represents the twist energy and
κ3 -term represents the bending energy, corresponding to the various deformations of the
nematic director u with |u(x)| = 1; the κ4 -term is a null-Lagrangian, depending only
on the boundary data of u (see Chapter 5).
From the algebraic relation |Du − (Du)T |2 = 2| curl u|2 one easily sees that
|Du|2 = | curl u|2 + tr((Du)2 ).
Furthermore, since |u(x)| = 1, it easily follows that
| curl u|2 = (u · curl u)2 + |u × curl u|2 .
4.4. Minimization with Constraints 75

If κ = min{κ1 , κ2 , κ3 } then it easily follows that

Z Z
(4.34) I(u) ≥ κ |Du|2 dx + (κ4 − κ) [tr((Du)2 ) − (div u)2 ] dx.
Ω Ω

If κ > 0, then it can be shown that the first part of F (u, ξ) is convex and quadratic in ξ, and
hence the first part of the energy I is w.l.s.c. on H 1 (Ω; R3 ) by Tonelli’s theorem (Theorem
5.1 of Chapter 5). By (4.34), one can easily obtain that
Z
|Du(x)|2 dx ≤ c0 I(u) + Cϕ , ∀ u ∈ Dϕ ⊂ H 1 (Ω; R3 ).
Ω

Therefore, by the direct method, we have established the following existence result for
minimizers.

Theorem 4.12. Let κi > 0 for i = 1, 2, 3. If C = {u ∈ Dϕ | |u(x)| = 1 a.e.} 6= ∅, then there

exists at least one u ∈ C such that

I(u) = min I(v).

v∈C

1
Remark. If all four κi are
R equal2 to 2 , then I(u) reduces to the Dirichlet integral for har-
1
monic maps: I(u) = 2 Ω |Du| dx.

4.4.3. Stokes’ Problem. Let Ω ⊂ R3 be open, bounded and simply connected. Given
f ∈ L2 (Ω; R3 ), the Stokes’ problem

−∆u = f − ∇p, div u = 0 in Ω, u=0 on ∂Ω,

can be solved by minimizing the functional

Z
1
I(u) = ( |Du|2 − f · u) dx
Ω 2
over the set
C = {u ∈ H01 (Ω; R3 ) | div u = 0 in Ω}.
The function p in the equation is the corresponding Lagrange multiplier, known as the pres-
sure. Pressure does not appear in the variational problem, but arises due to the constraint
div u = 0.

4.4.4. Incompressible Elasticity. Suppose u : Ω → R3 represents the displacement of

an elastic body occupying the domain Ω ⊂ R3 before deformation. Assume the body is
incompressible, which means

det Du(x) = 1 ∀ x ∈ Ω.

The stored energy is given by an integral functional

Z
I(u) = F (x, u, Du) dx.
Ω

The suitable space to work in this case is W 1,p (Ω; R3 ) with p ≥ 3. For further details, see
next chapter.
76 4. Variational Methods for PDEs

4.4.5. A Nonlocal Problem: Ferromagnetism. (See James & Kinderlehrer [23], and
Pedregal & Yan [34] for more results.)
We study a minimization problem in ferromagnetism which involves a constraint and
a nonlocal term. After some simplifications, the total energy I(m) in ferromagnetism is
given by
Z Z
1
(4.35) I(m) = ϕ(m(x)) dx + |∇u(x)|2 dx,
Ω 2 Rn

where Ω is a bounded domain with C 1 boundary ∂Ω, m ∈ L∞ (Ω; Rn ) with |m(x)| = 1

a.e. x ∈ Ω (that is, m ∈ L∞ (Ω; S n−1 )), representing the magnetization of a ferromagnetic
material occupying the domain Ω, ϕ is a given function representing the anisotropy of
the material, and u is the nonlocal stray energy potential determined by m over the whole
space Rn by the simplified Maxwell equation:
(4.36) div(−∇u + mχΩ ) = 0 in Rn ,
1 (Rn ) with ∇u ∈ L2 (Rn ; Rn ). The
where χΩ is the characteristic function of Ω, and u ∈ Hloc
equation (4.36) is understood in the weak sense:
Z Z
∇u · ∇ζ dx = m · ∇ζ dx ∀ ζ ∈ H01 (Rn ).
Rn Ω
Note that any two weak solutions of (4.36) can only differ by a constant. Let B be a fixed
open ball containing Ω̄ in Rn . We have the following result.
Theorem 4.13. For each m R∈ L2 (Ω; Rn ), there exists a unique weak solution u = T m ∈
1 (Rn ) to (4.36) such that 2 n 1
Hloc B u(x)dx = 0. Moreover, T : L (Ω; R ) → H (B) is linear
and bounded with
kT mkH 1 (B) ≤ C kmkL2 (Ω;Rn ) , ∀ m ∈ L2 (Ω; Rn ).

Proof. Let
Z
1
X = {u ∈ Hloc (Rn ) | 2 n
∇u ∈ L (R ; R ),n
u(x) dx = 0}.
B
Given m ∈ L2 (Ω; Rn ), define a functional J : X → R by
Z Z
1
J(u) = |∇u|2 dx − m · ∇u dx, u ∈ X.
2 Rn Ω
We solve the minimization problem: inf u∈X J(u). Note that X is simply a nonempty set
of functions and has no topology defined. But we can always take a minimizing sequence
uk ∈ X such that
lim J(uk ) = inf J(u) < ∞.
k→∞ u∈X
By Cauchy’s inequality with , it follows that
Z Z Z
1 2 2
J(u) ≥ |∇u| dx −  |∇u| dx − C |m|2 dx.
2 Rn Rn Ω
1
Taking  = 4 yields
Z
1
J(u) ≥ |∇u|2 dx − C, ∀ u ∈ X.
4 Rn
Therefore {∇uk } is bounded in L2 (Rn ; Rn ) and hence, via a subsequence, converges weakly
to some F ∈ L2 (Rn ; Rn ); this F can be written as F = ∇ū for a unique u ∈ X. (The
last fact needs a little more analysis; see the Hodge decompositions below!) Using this
4.4. Minimization with Constraints 77

weak convergence, we see that ū ∈ X is a minimizer of J over X. It also satisfies the

Euler-Lagrange equation
Z Z
(4.37) ∇ū · ∇ζ = m · ∇ζ ∀ζ ∈ X.
Rn Ω
This equation also holds for all ζ ∈ 1 n
H0 (R ).
From this equation, the uniqueness of mini-
mizers and the linear dependence of ū on m follow. Define ū = T m. Furthermore, using
ζ = ū ∈ X in (4.37), we also have
Z Z
2
|∇ū| dx = m · ∇ū dx ≤ kmkL2 (Ω;Rn ) k∇ūkL2 (B;Rn ) .
Rn Ω
Hence
k∇ūkL2 (B;Rn ) ≤ kmkL2 (Ω;Rn ) .
R
Since B ū dx = 0, by Poincaré’s inequality,
kūkH 1 (B) ≤ C kmkL2 (Ω;Rn ) .
The theorem is proved. 

From (4.37), we also see that

Z Z
1
(4.38) I(m) = ϕ(m) dx + m · ∇ū dx,
Ω 2 Ω
where ū = T m is defined in the theorem above. The following result is a special case of
some general theorem. We leave the easy proof to the student.
Lemma 4.14. Let T : L2 (Ω; Rn ) → H 1 (B) be defined above. Then T mk * T m weakly in
H 1 (B) whenever mk * m weakly in L2 (Ω; Rn ).
Theorem 4.15. Assume ϕ(m) ≥ 0 on S n−1 and the minimal set ϕ−1 (0) contains at least
{±m0 }. Then
inf I(m) = 0.
m∈L∞ (Ω;S n−1 )

Proof. It suffices to find a sequence {mk } in L∞ (Ω; S n−1 ) such that

Z Z
1
I(mk ) = ϕ(mk (x))dx + mk (x) · ∇uk (x) dx → 0
Ω 2 Ω
as k → ∞, where uk = T mk is the solution in X of the simplified Maxwell equation (4.36).
To this end, let η ∈ S n−1 be such that η ⊥ m0 and define
mk (x) = ρ(kx · η)m0 + (1 − ρ(kx · η))(−m0 ), x ∈ Ω,
where ρ(t) is a periodic function of period 1 and
1 1
0 ≤ t ≤ ; ρ(t) = 0
ρ(t) = 1 < t < 1.
2 2
It is easy to check wk * 0 weakly in L2 (Ω; Rn ). Let uk = T wk . Then, by Lemma 4.14
above, uk * 0 in H 1 (B). Note that, by compact embedding theorems, H 1 (B) ⊂⊂ L2 (Ω)
and H 1 (B) ⊂⊂ L2 (∂Ω) are compact; hence uk → 0 strongly in both L2 (Ω) and L2 (∂Ω).
We now compute by the divergence theorem that
Z Z Z
mk · ∇uk dx = uk mk · ν dS − uk div mk dx,
Ωj ∂Ωj Ωj
78 4. Variational Methods for PDEs

where ν is the unit outward normal on the boundary and the formula is valid on each piece
Ωj of Ω̄ where mk is constant m0 or −m0 ; hence div mk = 0 on each Ωj . Moreover, mk ·ν = 0
on ∂Ωj \ ∂Ω. Hence we have
Z  Z
|uk (x)| dS ≤ |∂Ω|1/2 kuk kL2 (∂Ω) → 0 as k → ∞.
 
 mk · ∇uk dx ≤
 
Ω ∂Ω

Finally, noting that ϕ(mk (x)) = 0, we arrive at

Z Z
1
I(mk ) = ϕ(mk (x))dx + mk (x) · ∇uk (x) dx → 0
Ω 2 Ω
as k → ∞. 
Theorem 4.16. Assume ϕ−1 (0) = {±m0 }. Then there exists no minimizer of I(m) on
L∞ (Ω; S n−1 ).

Proof. Suppose m̄ is a minimizer. Then I(m̄) = inf I = 0 and hence ϕ(m̄(x)) = 0 and
T m̄ = 0. Therefore,
m̄(x) = χE (x)m0 + (1 − χE (x))(−m0 ), div(m̄χΩ ) = 0,
where E = {x ∈ Ω | m̄(x) = m0 } is a measurable subset of Ω. Therefore
Z
(4.39) f (x)∇ζ(x) · m0 dx = 0 ∀ ζ ∈ C0∞ (Rn ),
Rn

where f (x) = χΩ (x)[2χE (x) − 1] = 2χE (x) − χΩ (x). Let x = x0 + tm0 where x0 ⊥ m0 and
write f (x) = g(x0 , t). In (4.39), by change of variables, we have
Z
g(x0 , t)ζt (x0 , t) dx0 dt = 0 ∀ ζ ∈ C0∞ (Rn ).
Rn

This implies the weak derivative gt (x0 , t) ≡ 0 on Rn , and hence g(x0 , t) = h(x0 ) is indepen-
dent of t. But h(x0 ) = f (x0 + tm0 ) vanishes for large t and hence h ≡ 0. So f ≡ 0 on Rn .
However f (x) ∈ {±1} for x ∈ Ω. This is a contradiction. 

4.4.6. The Hodge Decompositions on Rn . We now present a short theory for Hodge
decompositions on Rn with elementary proofs. See [38] for some related results.
Let X = L2 (Rn ; Rn ) denote the Hilbert space with the inner product and norm defined
by Z Z
1 1 n n
(u, v) = (u v + · · · + u v ) dx = u · v dx; kuk = (u, u)1/2 .
Rn Rn
For u ∈ L2 (Rn ; Rn ), we define div u, curl u = (curl u)ij as distributions as follows:
Z Z
1 n
hdiv u, ϕi = − (u ϕx1 + · · · + u ϕxn ) dx = − u · ∇ϕ dx;
Rn Rn
Z
h(curl u)ij , ϕi = − (ui ϕxj − uj ϕxi ) dx, ∀ ϕ ∈ C0∞ (Rn ).
Rn
1,1
Note that, if u ∈ Wloc (Rn ; Rn ), then div u = tr Du and curl u = Du − (Du)T . In the case
n = 2 or n = 3, the operator curl u can be identified as follows:
curl u ≈ ∇⊥ · u = ∇ · u⊥ = div(u⊥ ) = u1x2 − u2x1 (n = 2);
curl u ≈ ∇ × u = (u3x2 − u2x3 , u1x3 − u3x1 , u2x1 − u1x2 ) (n = 3).
4.4. Minimization with Constraints 79

Note that in the sense of distribution, the Laplacian of m ∈ X (for each component of m)
can be written as
(4.40) ∆m = ∇(div m) + div(curl m),
which means
n
X
i
∆m = ∂xi (div m) + ∂xj (curl m)ij ∀ i = 1, 2, · · · , n.
j=1

Define the subspaces of divergence-free and curl-free fields as follows:

Xdiv = {u ∈ L2 (Rn ; Rn ) | div u = 0 in the sense of distribution};
Xcurl = {u ∈ L2 (Rn ; Rn ) | curl u = 0 in the sense of distribution}.
It is easy to see they are closed linear subspaces of X = L2 (Rn ; Rn ). First we have the
following result.
Lemma 4.17. Xdiv ∩ Xcurl = {0}.

Proof. Let m ∈ Xdiv ∩ Xcurl . Then curl m = div m = 0 in the sense of distributions. Hence,
by (4.40) above, ∆m = 0 holds also in the sense of distributions. Hence m ∈ C ∞ (Rn ; Rn )
and each of its components mi is a harmonic function in Rn . Sine each mi also belongs to
L2 (Rn ), the mean value property and Hölder’s inequality imply that
Z
1
|mi (x)| ≤ |mi | dy ≤ c R−n/2 kmk
|BR (x)| BR (x)
for any x ∈ Rn and R > 0, where BR (x) = {y ∈ Rn | |y − x| < R} denotes the ball of radius
R and center x. Letting R → ∞ shows mi = 0 and hence m = 0. 

To characterize the spaces Xdiv and Xcurl and prove the Hodge decomposition theorem,
we introduce the linear function spaces:
Y = {f ∈ L2loc (Rn ) | ∇f ∈ X}
and, for n = 3,
M = {m ∈ L2loc (R3 ; R3 ) | ∇ × m ∈ L2 (R3 ; R3 )}.
It is easy to see that ∇f ∈ Xcurl for all f ∈ Y and, for n = 3, ∇ × m ∈ Xdiv for all m ∈ M.
The converse is also true; we have the following results.
Theorem 4.18. There exists a uniform constant Cn > 0 such that, for each v ∈ Xcurl ,
there exists a f ∈ Y satisfying that
Z Z
1 2
(4.41) (a) v = ∇f ; (b) sup n+2 |f | dx ≤ Cn |v|2 dx,
R≥1 R BR (0) Rn

where BR (y) = {x ∈ Rn | |x − y| < R} denotes the ball centered y of radius R.

Theorem 4.19. Let n = 3. There exists a uniform constant C > 0 such that, for every
w ∈ Xdiv , there exists a m ∈ M satisfying that
Z Z
1
(4.42) (a) w = ∇ × m; (b) sup 3 |m|2 dx ≤ C |w|2 dx,
R≥1 R QR (0) Rn

where QR (y) = {x ∈ R3 | |xi −yi | < R, i = 1, 2, 3} denotes the cubes centered y of side-length
2R.
80 4. Variational Methods for PDEs

The estimate (4.41b) in Theorem 4.18 above is not sharp as we will obtain some better
estimates later. However, the estimate (b) in both (4.41) and (4.42) does provide a way to
represent a curl-free v or divergence-free field w by some local function f or m, with f being
viewed as the potential function of v and m the velocity of w; initially, these local functions
have only been defined as the Schwartz distributions. The estimate (4.41b) also suggests
that we equip the space Y with the norm kf k∗ defined by
Z Z
2 2 1
(4.43) kf k∗ = |∇f (x)| dx + sup n+2 |f (x)|2 dx.
Rn R≥1 R BR
Let Y1 be the subspace of Y defined by
1
Y1 = {f ∈ Hloc (Rn ) | kf k∗ < ∞}.
It can be easily shown that Y1 is indeed a Banach space with the norm defined. We shall
try to find the minimal subspace Z of Y1 for which the gradient operator ∇ : Z → Xcurl is
bijective. In fact such a Z can be completely determined when n ≥ 3.
Theorem 4.20. Let Y2 be the closure of C0∞ (Rn ) in Y1 . Then, for n ≥ 3,
2n
Y2 = {f ∈ L n−2 (Rn ) | ∇f ∈ L2 (Rn ; Rn )}.
Furthermore, Y2 has the equivalent norms
kf k∗ ≈ k∇f kL2 (Rn ) ∀ f ∈ Y2 ,
and the gradient operator ∇ : Y2 → Xcurl is bijective.
The proof of this theorem relies on the Sobolev-Gagliardo-Nirenberg inequality for
H 1 (Rn ) functions when n ≥ 3 (see [2, 16, 21]); note that in this case the finite num-
ber 2∗ = n−22n
is the Sobolev conjugate of 2. In the case n = 2, there is no such a Sobolev-
Gagliardo-Nirenberg inequality; instead, there is a John-Nirenberg-Trudinger type of BMO-
estimates for functions with gradient in L2 (R2 ) (see [21, 35]. However, we shall try to avoid
the BMO-estimates. One of the minimal subspaces of Y1 on which the gradient operator is
bijective can be characterized as follows.
Theorem 4.21. Let n = 2 and Z be the closure in Y1 of the subspace
S = {ϕ − hϕiρ | ϕ ∈ H 1 (R2 )},
R
where hϕiρ = R2 ϕ(x)ρ(x) dx and ρ(x) is the weight function defined by
 
1 1
(4.44) ρ(x) = χ{|x|≤1} + 4 χ{|x|>1} .
2π |x|
Then Z has the equivalent norms kf k∗ ≈ k∇f kL2 (R2 ) ∀ f ∈ Z, and the gradient operator
∇ : Z → Xcurl is bijective.
Given any u ∈ X = L2 (Rn ; Rn ), for each R > 0, let BR = BR (0) and consider the
following minimization problem:
Z
(4.45) inf |∇ϕ − u|2 dx.
ϕ∈H01 (BR ) BR

By the direct method, this problem has a unique solution, which we denote by ϕR , also
extended by zero to all Rn . This sequence {ϕR } is of course uniquely determined by u ∈ X.
It also satisfies the following properties:
1
R
(4.46) Rn (∇ϕR − u) · ∇ζ dx = 0 ∀ ζ ∈ H0 (Ω), Ω ⊆ BR ,
(4.47) k∇ϕR kL2 (Rn ) ≤ kuk.
4.4. Minimization with Constraints 81

Lemma 4.22. Given u ∈ X, it follows that ∇ϕR → v in X as R → ∞ for some v ∈ X

and that v ∈ Xcurl is uniquely determined by u. Moreover, this v satisfies
kv − uk = 0 min kv 0 − uk;
v ∈Xcurl

therefore, v = u if u ∈ Xcurl .

Proof. First of all, we claim ∇ϕR * v weakly in X as R → ∞. Let v 0 , v 00 be the weak

limits of any two subsequences {∇ϕR0 } and {∇ϕR00 }, where R0 , R00 are two sequences going
to ∞. We would like to show v 0 = v 00 , which shows that ∇ϕR * v as R → ∞. Note that it
is easily seen that v 0 , v 00 ∈ Xcurl and, by (4.46) above,
Z Z
(4.48) (v 0 − u) · ∇ζ = (v 00 − u) · ∇ζ = 0 ∀ Ω ⊂⊂ Rn , ∀ ζ ∈ H01 (Ω).
Rn Rn
This implies div(v 0
− u) = div(v 00
− u) = 0 and hence div(v 0 − v 00 ) = 0; therefore, v 0 − v 00 ∈
Xdiv ∩ Xcurl = {0} by Lemma 4.17 above. We denote this weak limit by v ∈ Xcurl . Note
that, by (4.48), div(v − u) = 0. Hence if curl u = 0 then v − u ∈ Xdiv ∩ Xcurl = {0}; hence
v = u. We now prove ∇ϕR → v in X as R → ∞. Taking ζ = ϕR ∈ H01 (BR ) in (4.48) and
letting R → ∞ we have
Z
(4.49) (v − u) · v dx = 0.
Rn
Using ζ = ϕR in (4.46), taking R → ∞ and by weak limit, we have
Z Z Z
2
lim |∇ϕR | dx = v·u= |v|2 dx.
R→∞ Rn Rn Rn
This implies ∇ϕR → v strongly in L2 (Rn ; Rn ). We now show that
(4.50) kv − uk = 0 min kv 0 − uk.
v ∈Xcurl

Given any ∈ Xcurl , choose the sequence ϕ̃R corresponding to v 0 − v. Since div(v − u) = 0,
v0
it easily follows that
(v 0 − v, v − u) = lim (∇ϕ̃R , v − u) = 0.
R→∞
Hence kv 0 − uk2 = kv − v 0 k2 + 2(v 0 − v, v − u) + kv − uk2 ≥ kv − uk2 ; this proves (4.50). 

We can now prove the following Hodge decomposition theorem.

Theorem 4.23. (Hodge decomposition) For each u ∈ X, there exist unique elements
⊥ , X
v ∈ Xcurl , w ∈ Xdiv such that u = v + w. Moreover, Xcurl = Xdiv ⊥
div = Xcurl . Conse-
quently, X = L2 (Rn ; Rn ) = Xdiv ⊕ Xcurl .

Proof. Given u ∈ X, let v ∈ Xcurl be defined as above, and let w = u − v. Then u = v + w

and, by (4.48) above, w ∈ Xdiv . We now show that v, w are unique. Suppose u = v 0 + w0
for another pair v 0 ∈ Xcurl and w0 ∈ Xdiv . Then m = v − v 0 = w0 − w ∈ Xcurl ∩ Xdiv . Hence
v 0 = v and w0 = w.
⊥ , given any v ∈ X 1
To show Xcurl = Xdiv curl , w ∈ Xdiv , let ϕR ∈ H0 (BR ) be the sequence
determined by u = v as above. Since ∇ϕR → v and
Z
(w, ∇ϕR ) = w(x) · ∇ϕR (x) = 0,
Rn
⊥ . This shows X
it follows easily that (w, v) = 0; hence v ∈ Xdiv ⊥ ⊥
curl ⊆ Xdiv . Assume u ∈ Xdiv .
We will show u ∈ Xcurl . By the first part of this result, let u = v + w, where v ∈ Xcurl ,
82 4. Variational Methods for PDEs

w ∈ Xdiv . Then 0 = (u, w) = (v, w) + |w|2 = |w|2 ; hence w = 0 and u = v ∈ Xcurl . This
⊥ . Similarly, it follows X
proves Xcurl = Xdiv ⊥
div = Xcurl . 
Theorem 4.24. Let Y1 be the space with the norm k·k∗ defined above. Then Y1 is a Banach
space. Moreover, the gradient operator ∇ : Y1 → Xcurl is surjective; more precisely, for any
v ∈ Xcurl , there exists a f ∈ Y1 such that
v = ∇f, kf k∗ ≤ Cn kvk.

Proof. The proof that Y1 is a Banach space follows directly by the definition and will not
be given here. We prove the rest of the theorem.
Given v ∈ Xcurl , let v = v ∗ ρ be the smooth approximation of v. Then v ∈ Xcurl ∩
C ∞ (Rn ; Rn ). Define
Z 1
f (x) = v (tx) · x dt.
0
Then one can easily compute that
Z 1 Z 1
j
∂xj f (x) = v (tx)dt + (∂xj v )(tx) · tx dt
0 0
Z 1 Z 1
j
= v (tx)dt + ∇vj (tx) · tx dt
0 0
Z 1 Z 1
j d j
= v (tx)dt + [v (tx)]t dt
0 0 dt
Z 1  Z 1 
j j
1 j
= v (tx)dt + v (tx)t0 − v (tx)dt
0 0
= vj (x),
where we have used the condition ∂xj v = ∇vj since curl v = 0. This proves ∇f (x) = v (x)
for all x ∈ Rn . Therefore, for all x, y ∈ Rn ,
Z 1
f (x + y) − f (y) = v (y + tx) · x dt.
0
Hence
Z 1
2
2
 
|f (x + y) − f (y)| =  v (y + tx) · x dt
0
Z 1
2
≤ |x| |v (y + tx)|2 dt.
0

Integrating this inequality over x ∈ BR (0) = BR , we obtain

Z Z 1 Z 
2 2 2
|f (z) − f (y)| dz ≤ R |v (y + tx)| dx dt
BR (y) 0 BR
!
Z 1 Z
= R2 |v (z)|2 dz t−n dt
0 BtR (y)
!
Z 1 Z
1
= Rn+2 |v (z)|2 dz dt
0 |BtR (y)| BtR (y)

≤ Rn+2 M(|v |2 )(y),

4.4. Minimization with Constraints 83

where M(h) is the maximal function of h (see Stein [35]). Since |v |2 ∈ L1 (Rn ), it follows
that
5n 5n
Z Z
n 2 2
m{y ∈ R | M(|v | )(y) > α} ≤ |v | dx ≤ |v|2 dx.
α Rn α Rn
Let
E = {y ∈ B1 | M(|v |2 )(y) ≤ α0 },
where we choose
2 · 5n
Z
α0 = |v|2 dx.
|B1 | Rn
Then it follows that |E | ≥ 12 |B1 | for all . Therefore, it is a simple exercise to show that
there exists a sequence k → 0 and a point y0 ∈ B1 such that y0 ∈ ∩∞ k=1 Ek ; that is,
2 · 5n
M(|vk |2 )(y0 ) ≤ α0 = kvk2 , ∀ k = 1, 2, · · · .
|B1 |
Using this y0 we define a new sequence
gk (z) = fk (z) − fk (y0 ), z ∈ Rn .
Then, for all R ≥ 1, we have
2 · 5n
Z Z
2
|gk (z)| dz ≤ |gk (z)|2 dz ≤ (2R)n+2 kvk2 .
BR B2R (y0 ) |B1 |
By using diagonal subsequences, there exists a subsequence gkj and a function f ∈ L2loc (Rn )
such that gkj * f weakly as kj → ∞ on all balls BR (0), R > 0. This function f must
satisfy ∇f = v ∈ L2 (Rn ; Rn ) and
Z Z
1 2
sup n+2 |f (x)| dx ≤ Cn |v(x)|2 dx;
R≥1 R BR Rn

hence kf k∗ ≤ Ckvk. This completes the proof. 

Proof of Theorem 4.19. Given w ∈ Xdiv , let w = w ∗ ρ be the smooth approximation

of w. Then w ∈ Xdiv ∩ C ∞ (R3 ; R3 ). Define
m (x, c) = (p (x, c), q (x, c), 0), x ∈ R3 , c ∈ [−1, 1],
where
Z x3 Z x2
(4.51) p (x, c) = w2 (x1 , x2 , s) ds − w3 (x1 , t, c) dt;
c 0
Z x3
(4.52) q (x, c) = − w1 (x1 , x2 , s) ds.
c
Since div w = 0, one can easily verify that
w (x) = ∇ × m (x, c), ∀ c ∈ [−1, 1].
We now estimate
Z |x3 |+1
2 2
|q (x, c)| ≤ (|x3 | + 1) |w1 (x1 , x2 , s)|2 ds.
−|x3 |−1

Integrating this inequality over the cube QR (0) = {x ∈ R3 | |xi | < R, i = 1, 2, 3}, we obtain
Z Z
2 2
(4.53) |q (x, c)| dx ≤ 2R(R + 1) |w1 (x)|2 dx.
QR (0) R3
84 4. Variational Methods for PDEs

Next we write p (x, c) = g (x, c) − f (x0 , c) with x0 = (x1 , x2 ), where

Z x3 Z x2
2 0
(4.54) g (x, c) = w (x1 , x2 , s) ds; f (x , c) = w3 (x1 , t, c) dt.
c 0

For g (x, c), we have the same estimate as q (x, c):

Z Z
2 2
(4.55) |g (x, c)| dx ≤ 2R(R + 1) |w2 (x)|2 dx.
QR (0) R3

For f (x0 , c), we easily estimate that

Z Z
0 0
(4.56) 2
|f (x , c)| dx ≤ 2R 2
|w3 (x0 , c)|2 dx0 = 2R2 H (c).
|x0i |<R R2

|w3 (x0 , c)|2 dx0 ∈ L1 (R). It follows that

R
Note that H (c) = R2
Z Z
1 1
m{c ∈ R | H (c) > α} ≤ H (c) dc = |w3 (x)|2 dx.
α R α R3 
Let E = {c ∈ [−1, 1] | H (c) ≤ α0 }, where α0 = R3 |w3 (x)|2 dx. Then it follows that
R

|E | ≥ 1 for all . Therefore, as above, there exists a sequence k → 0 and a point c0 ∈ [−1, 1]
such that c0 ∈ ∩∞ k=1 Ek ; that is,
Z
Hk (c0 ) ≤ |w3 (x)|2 dx ∀ k = 1, 2, · · · .
R3

Hence by (4.56)
Z Z
2 3
(4.57) |f (x, c0 )| dx ≤ 4R |w3 (x)|2 dx.
QR (0) R3

Using this c0 we define a new sequence uk (x) = mk (x, c0 ). Then we have wk = curl uk
and, for all R ≥ 1, by (4.53)-(4.56),
Z
(4.58) |uk (x)|2 dx ≤ CR3 kwk2 .
QR (0)

By using diagonal subsequences, there exists a subsequence ukj and a function m ∈ L2loc (R3 )
such that ukj * m weakly as kj → ∞ on all cubes |xi | < R, R > 0. This field m must
satisfy w = curl m ∈ L2 (R3 ; R3 ); hence m ∈ M. Moreover, by (4.58),
Z Z
1
sup 3 |m(x)|2 dx ≤ C |w(x)|2 dx.
R≥1 R QR (0) R3

This completes the proof. 

Proof of Theorem 4.20. Let n ≥ 3 and Y1 , Y2 be the spaces defined above, and define
2n
W = {f ∈ L n−2 (Rn ) | ∇f ∈ L2 (Rn ; Rn )}.
We proceed with several lemmas to prove Theorem 4.20.

Lemma 4.25. Y2 ⊆ W. Moreover

(4.59) kf k∗ ≤ Ck∇f kL2 (Rn ) ∀ f ∈ Y2 .
4.4. Minimization with Constraints 85

Proof. Let f ∈ Y2 . Then there exists a sequence fj ∈ C0∞ (Rn ) such that kfj − f k∗ → 0 as
j → ∞. Therefore k∇fj kL2 → k∇f kL2 . By Sobolev-Galiardo-Nirenberg inequality,
kfj kL2∗ (Rn ) ≤ C k∇fj kL2 (Rn ) ∀ j.
∗
Hence fj * g ∈ L2 (Rn ). Since fj → f in L2 (BR ) for all R > 0. We have f = g. Hence
f ∈ W. Furthermore, by Hölder’s inequality,
kfj kL2 (BR ) ≤ cn R2 kfj kL2∗ (BR ) ≤ C R2 k∇fj kL2 (Rn ) .
Hence, by taking limits as j → ∞, it follows that
Z Z
1
sup n+2 |f |2 dx ≤ C |∇f (x)|2 dx,
R≥1 R BR Rn

which proves (4.59). 

Lemma 4.26. W ⊆ Y2 .

Proof. Let f ∈ W. Define fj = f ρj , where ρj ∈ W 1,∞ (Rn ) defined by ρj (x) = 1 on |x| ≤ j,

ρj (x) = 0 on |x| ≥ 2j and ρj (x) is linear in |x| for j ≤ |x| ≤ 2j. Then fj ∈ Y2 . It can be
easily shown that
lim kfj − f k∗ = 0,
j→∞
which proves f ∈ Y2 and hence W ⊆ Y2 . 
Lemma 4.27. ∇ : Y2 → Xcurl is surjective.

Proof. Given any v ∈ Xcurl , let ϕR ∈ H01 (BR ) be the function determined as in the
minimization problem (4.45) above with u = v. Then ∇ϕR → v in X = L2 (R2 ; R2 ) as
R → ∞. Lemma 4.25 implies that {ϕR } is a Cauchy sequence in Y2 and hence its limit f
belongs to Y2 and satisfies ∇f = v. This completes the proof. 

Proof of Theorem 4.21. We first prove the following result.

Lemma 4.28. Let ρ(x) be defined as above. Then for all ϕ ∈ H 1 (R2 ),
Z Z
2
(4.60) |ϕ(x) − hϕiρ | ρ(x) dx ≤ β |∇ϕ(x)|2 dx.
R2 R2

Proof. First of all, by Poincaré’s inequality,

Z Z
2
(4.61) |ψ(x) − (ψ)0 | dx ≤ C |∇ψ(x)|2 dx
B1 B1

for all ψ ∈ H 1 (B1 ), where (ψ)0 is the average value of ψ on B1 ; that is,
Z Z
1
(ψ)0 = ψ(x) dx = 2 ψ(x)ρ(x) dx.
π B1 {|x|<1}

Given ϕ ∈ H 1 (R2 ), let ψ(x) = ϕ( |x|x2 ). Then ψ ∈ H 1 (B1 ). Using the above Poincaré
inequality for this ψ, after change of variable, one obtains that
Z Z
2 −4
(4.62) |ϕ(y) − (ψ)0 | |y| dy ≤ C |∇ϕ(y)|2 dy,
{|y|>1} {|y|>1}
where
Z Z Z
1 1
(ψ)0 = ψ(x) dx = ϕ(y)|y|−4 dy = 2 ϕ(x)ρ(x) dx.
π B1 π {|y|>1} {|x|>1}
Combining (4.61) for ψ(x) = ϕ(x) with (4.62) we obtain (4.60). 
86 4. Variational Methods for PDEs

Note that, for all f ∈ Y1 ,

Z Z
1
(4.63) sup 4
|f (x)|2 dx ≤ 2π |f (x)|2 ρ(x) dx.
R≥1 R BR R2

Let
S = {ϕ − hϕiρ | ϕ ∈ H 1 (R2 )}.
Then the previous lemma and (4.63) imply
Z Z Z
1 2 2
sup 4 |f | dx ≤ 2π |f | ρ dx ≤ C |∇f |2 dx ∀ f ∈ S.
R≥1 R BR R2 R2

We have thus proved the following result.

Corollary 4.29. Let Z be the closure of S in Y1 . Then kf k∗ ≈ k∇f kL2 for all f ∈ Z.
Furthermore, for all f ∈ Z,
Z Z Z
2
f (x)ρ(x) dx = 0, |f (x)| ρ(x) dx ≤ C |∇f |2 dx,
R2 R2 R2

where C is a constant independent of f.

Finally, we prove the following result to complete the proof of Theorem 4.21.

Corollary 4.30. ∇ : Z → Xcurl is surjective.

Proof. The proof is similar to that of Lemma 4.27 above. Given any v ∈ Xcurl , let ϕR ∈
H01 (BR ) be the function determined as in the minimization problem (4.45) above with u = v.
Let fR = ϕR − hϕR iρ . Then fR ∈ S and ∇fR = ∇ϕR → v in X = L2 (R2 ; R2 ) as R → ∞.
Corollary 4.29 implies that {fR } is a Cauchy sequence in Y1 and hence its limit f belongs
to the closure Z of S and satisfies ∇f = v. This completes the proof. 

4.5. Mountain Pass Theorem

In this section we consider another approach to noncoercive problems of the type dealt with
in the previous section.

4.5.1. The Palais-Smale Condition. We first introduce the following Palais-Smale

compactness condition (PS), which plays an essential role in the results to follow.
The Palais-Smale Condition (PS). Let E : X → R be G-diff on the Banach space X. We
say E satisfies the Palais-Smale condition (PS) if the following holds: whenever {un } is a
sequence in X such that E(un ) is bounded and kE 0 (un )k → 0, then {un } has a convergent
subsequence.
Remark. The (PS) condition is not satisfied by very smooth functions very often. For
example, the function E : R → R with E(u) = cos u does not satisfy (PS), which can be
easily seen by considering the sequence un = nπ. Similarly, the function E(u) = c does not
satisfy (PS). It can be shown that if E is F-diff on the Banach space X (not necessarily
reflexive) and is bounded below and satisfies (PS), then E attains its minimum value. But
we study the case where functionals are neither bounded from above or below.
4.5. Mountain Pass Theorem 87

4.5.2. Deformation Lemma.

Lemma 4.31. (Deformation) Let E : X → R be a C 1 functional satisfying (PS). Let
c, s ∈ R and define
Kc = {u ∈ X : E(u) = c, E 0 (u) = 0}
As = {u ∈ X : E(u) ≤ s}.
Assume Kc = ∅. Then there exists an ε̄ > 0 and a continuous function η : [0, 1] × X → X
such that for all 0 < ε ≤ ε̄,
(i) η(0, u) = u for all u ∈ X,
(ii) η(t, u) = u for all t ∈ [0, 1], u 6∈ E −1 ([c − ε, c + ε]),
(iii) E(η(t, u)) ≤ E(u) for all t ∈ [0, 1], u ∈ X,
(iv) η(1, Ac+ε ) ⊂ Ac−ε .

This lemma shows that if c is not a critical level, then we can nicely deform the set Ac+ε
into Ac−ε for some ε > 0.

4.5.3. The Mountain Pass Theorem. We now prove the main theorem of this section.
Theorem 4.32. (Ambrosetti-Rabinowitz) Let E : X → R be a C 1 functional satisfying
(PS) on the real Banach space X. Let u0 , u1 ∈ X, c0 ∈ R and R > 0 be such that
(i) ku1 − u0 k > R
(ii) E(u0 ), E(u1 ) < c0 ≤ E(v) for all v such that kv − u0 k = R.
Then E has a critical point u with E(u) = c, c ≥ c0 ; the critical value c is defined by
(4.64) c = inf sup E(p(t))
p∈K t∈[0,1]

where K denotes the set of all continuous maps p : [0, 1] → X with p(0) = u0 and p(1) = u1 .

Remark. Think of the graph of E as a landscape with a low spot at u0 , surrounded by a

ring of mountains. Beyond these mountains lies another low point at u1 . Note that every
path p connecting u0 to u1 has to cross the sphere {v : kv − u0 k = R} since u1 lies outside
the sphere. Moreover, on this sphere the value of E is at least c0 . Hence the maximum
value of E(p(t)) for any such p is at least c0 . Hence c ≥ c0 . An important aspect of the
MPT is that the critical point u is distinct from u0 and u1 . Hence, if u0 already satisfies
E 0 (u0 ) = 0 by some other method, then u will be a second solution of E 0 (u) = 0.

Proof. Let c be as defined in (4.64). If it is not a critical value, then Kc = ∅. Let η and
ε̄ be as in the deformation lemma. Now by condition (ii) of the theorem, we can choose ε
small enough so that 0 < ε < ε̄ and E(u0 ), E(u1 ) 6∈ [c − ε, c + ε] (since c ≥ c0 ). Let p ∈ K
and define the path ξ : [0, 1] → X by
ξ(t) = η(1, p(t)).
Since p(0) = u0 and p(1) = u1 , it follows, by the choice of ε, that
ξ(0) = η(1, u0 ) = u0 , ξ(1) = η(1, u1 ) = u1
using condition (ii) of the lemma. Thus ξ ∈ K. Now, we can choose p ∈ K such that
max E(p(t)) < c + ε.
t∈[0,1]
88 4. Variational Methods for PDEs

Since p(t) ∈ Ac+ε , by (iv) of the lemma, ξ(t) ∈ Ac−ε . Thus

max E(ξ(t)) ≤ c − ε
t∈[0,1]

which contradicts the definition of c. Hence Kc 6= ∅. 

4.5.4. Saddle Point Solutions. In order to illustrate these ideas we apply the MPT to
the problem
(4.65) ∆u + f (x, u) = 0 in Ω, u|∂Ω = 0
where f ∈ Car satisfies the growth condition
|f (x, z)| ≤ c(1 + |z|τ ), |fz (x, z)| ≤ c(1 + |z|τ −1 )
with 1 < τ < (n + 2)/(n − 2). Moreover, we assume f (x, 0) = 0 and
0 ≤ F (x, z) ≤ γf (x, z)z
Rz
for some constant γ < 1/2, where F (x, z) = 0 f (x, s)ds. Finally we assume
a|z|τ +1 ≤ |F (x, z)| ≤ A|z|τ +1
for constants 0 < a ≤ A. Associated with (4.65) is the functional I defined by
Z
1
I(u) = |∇u|2 dx − J(u),
2 Ω
where Z
J(u) = F (x, u)dx.
Ω
The following lemma simplifies some of the technical details that follow.
Lemma 4.33. Let I and J be defined on H = H01 (Ω) as above. Then both I and J are of
C 1 on H and
(a) J 0 : H → H ∗ is compact.
(b) If {un } is a bounded sequence in H such that I 0 (un ) → 0 as n → ∞, then {un }
has a convergent subsequence.

Proof. The norm Ω |∇u|2 dx is obviously of C ∞ on H. J is a Nemytskii operator with F

R

being C 1 in u, and hence is of C 1 on H with J 0 : H → H ∗ given by

Z
0
hJ (u), vi = f (x, u(x)) v(x) dx, ∀ u v ∈ H.
Ω
We have the estimate
kJ 0 (u) − J 0 (w)kH ∗ = sup |hJ 0 (u) − J 0 (w), vi|
v∈H, kvk≤1
≤ kf (·, u) − f (·, w)k(τ +1)/τ kvkτ +1
≤ C kf (·, u) − f (·, w)k(τ +1)/τ , ∀ u w ∈ H.
Let {un } be a bounded sequence in H. Then by choosing a subsequence if necessary,
un * u in H and un → u in Lτ +1 (Ω). So J 0 (un ) converges to J 0 (u) and hence J 0 : H → H ∗
is compact.
Let A : H → H* denote the duality map defined by
Z
hAu, vi = ∇u · ∇v dx for all u, v ∈ H.
Ω
4.5. Mountain Pass Theorem 89

Since hI 0 (u), vi = hAu, vi − hJ 0 (u), vi, and A−1 exists and is bounded on H ∗ (by the Lax-
Milgram theorem), it follows that
A−1 I 0 (u) = u − A−1 J 0 (u).
Let {un } be a bounded sequence in H with I 0 (un ) → 0 as n → ∞. Since A−1 is continuous
and J 0 is compact, by passing to a subsequence if necessary, we have un = A−1 I 0 (un ) +
A−1 J 0 (un ) converges. 

Remark. From the above proof we see that the compactness of the embedding H = H01 (Ω) ⊂
Lτ +1 (Ω) is necessary for proving (PS).
Theorem 4.34. The boundary value problem (4.65) has at least one nontrivial solution
u ∈ H01 (Ω).

Proof. Consider the C 1 functional I(u) above, which we rewrite as

1
I(u) = kuk21,2,0 − J(u), where kuk21,2,0 = Ω |∇u|2 dx,
R
2
1
for u ∈ H0 (Ω). In order to apply MPT, we first choose u0 = 0. Clearly, I(u0 ) = 0. Now we
show that I|∂B(0,R) ≥ c0 for some R, c0 > 0. In view of the embedding H01 (Ω) ⊂ Lτ +1 (Ω),
we have Z
J(u) ≤ A |u|τ +1 dx ≤ ckukτ1,2,0
+1
.
Ω
Thus for all u satisfying kuk1,2,0 = R
R2
I(u) ≥ − cRτ +1 ≥ c0 > 0
2
provided we take R sufficiently small. Next, let v = tu, where u 6= 0 is a fixed function in
H01 (Ω), and t > 0 is to be selected. Then
t2
I(v) = kuk21,2,0 − J(tu)
2
t2
Z
2 τ +1
≤ kuk1,2,0 − at |u|τ +1 dx
2
< 0
for t > 0 and sufficiently large. Moreover, v ∈ H01 (Ω)verb|kB̄(0, R). Finally, we must verify
the Palais-Smale condition. Let {un } ⊂ H01 (Ω) be a sequence such that |I(un )| ≤ c for all
n and I 0 (un ) → 0. If we can show that {kun k1,2,0 } is bounded, then according to Lemma
4.33, I satisfies (PS). Thus, for the given sequence
Z
1 2
I(un ) = kun k1,2,0 − F (x, un )dx
2 Ω
and Z
0 2
hI (un ), un i = kun k1,2,0 − un f (x, un )dx
Ω
and so Z
1
kun k21,2,0 ≤ I(un ) + γ un f (x, un )dx
2 Ω
≤ c + γ(kun k21,2,0 + kI 0 (un )k kun k1,2,0 ).
Now suppose kun k1,2,0 → ∞. Then from the above inequality we arrive at the contradiction
1/2 ≤ γ. Thus, by MPT, there exists a critical point u 6= 0 of I which is a weak solution of
(4.65). 
90 4. Variational Methods for PDEs

4.6. Nonexistence and Radial Symmetry

We now establish a nonexistence result which shows that the boundary value problem (4.31)
(with k = 0, l = 1) does not have a positive solution for τ > (n + 2)/(n − 2).

4.6.1. Pohozaev’s Identity. We shall need the following fundamental identity:

Theorem 4.35. (Pohozaev) Let u ∈ C 2 (Ω̄), u|∂Ω = 0. Then
Z Z
2
(4.66) un (x · ν)dS = [2(∇u · x) + (n − 2)u]∆udx
∂Ω Ω
∂u
where ν = (ν1 , · · · , νn ) denotes the outward unit normal to ∂Ω and un = ∂ν = ∇u · ν
denotes the normal derivative of u on ∂Ω.

Proof. By an easy calculation we have the following identity

2∆u(∇u · x) = div[2(∇u · x)∇u − |∇u|2 x] + (n − 2)|∇u|2 .
An application of the divergence theorem gives
Z Z
[2(∇u · x)un − |∇u|2 x · ν]dS = [2∆u(∇u · x) − (n − 2)|∇u|2 ]dx.
∂Ω Ω
|∇u|2 dx
R R
Since u = 0 on ∂Ω, Ω u∆udx + Ω = 0 and uxi = νi un , i.e., ∇u = un ν, which
combined with the above integral identity yields (4.66). 
Corollary 4.36. Let u ∈ C 2 (Ω̄) be a solution of the boundary value problem
∆u + f (u) = 0 in Ω, u|∂Ω = 0.
Ru
Let F (u) = 0 f (t)dt. Then u satisfies the equation
2−n
Z Z Z
1
(4.67) u2n (x · ν)dS = n F (u)dx + uf (u)dx.
2 ∂Ω Ω 2 Ω

Proof. From (4.66) we have

2−n
Z Z Z
1 2
u (x · ν)dS = ∆u(∇u · x)dx + uf (u)dx
2 ∂Ω n Ω 2 Ω
so it suffices to show that
Z Z
∆u(∇u · x)dx = n F (u)dx.
Ω Ω
In this direction, we first note that ∇F = f ∇u; hence
∆u(∇u · x) = −f (u)(∇u · x) = −(∇F · x).
But ∇ · (xF ) = nF + (x · ∇F ) and therefore
Z Z Z
∆u(∇u · x)dx = nF dx − ∇ · (xF )dx.
Ω Ω Ω
However, the second integral on the right is zero, as may be seen by applying the divergence
theorem and noting that F = 0 on ∂Ω since u is. 

4.6.2. Star-shaped Domains. Let Ω be an open set containing 0. We say Ω is star-

shaped (with respect to 0) if, for each x ∈ Ω̄, the line segment {λx : 0 ≤ λ ≤ 1} lies in
Ω̄.
4.6. Nonexistence and Radial Symmetry 91

Remark. Clearly if Ω is convex and 0 ∈ Ω, then Ω is star-shaped. An annulus is not

star-shaped since x · ν < 0 on the boundary of the inner circle.
Lemma 4.37. Assume ∂Ω is C 1 and Ω is star-shaped with respect to 0 ∈ Ω. Then x·ν(x) ≥
0 for all x ∈ ∂Ω, where ν(x) is the outward unit normal at x ∈ ∂Ω.

Proof. Given x0 ∈ ∂Ω, since ∂Ω is C 1 , there exists a ball B = Bε (x0 ) and a C 1 function φ
on B such that
Ω ∩ B = {x ∈ B | φ(x) < 0}, ∂Ω ∩ B = {x ∈ B | φ(x) = 0}.
∇φ(x0 )
Note that the outward unit normal at x0 is given by ν(x0 ) = |∇φ(x 0 )|
. Let δ > 0 be
sufficiently small so that λx0 ∈ B and hence, from Ω being star-shaped, λx0 ∈ Ω̄ ∩ B for all
λ ∈ [1 − δ, 1]. Consider h(λ) = φ(λx0 ) defined on λ ∈ [1 − δ, 1]. Then h has the maximum 0
at right-end point 1 and hence
h0 (1− ) = ∇φ(x0 ) · x0 ≥ 0.
This proves x0 · ν(x0 ) ≥ 0 for all x0 ∈ ∂Ω. 

4.6.3. Nonexistence of Classical Solutions.

Theorem 4.38. Let Ω be star-shaped with respect to x = 0. Then the problem
n+2
(4.68) ∆u + |u|τ −1 u = 0 in Ω, u|∂Ω = 0 (τ > )
n−2
has no nontrivial C 2 (Ω̄) solution. Furthermore, the problem
n+2
(4.69) ∆u + uτ = 0 in Ω, u|∂Ω = 0 (τ ≥ )
n−2
has no positive C 2 (Ω̄) solution.

Proof. Suppose u is a nontrivial C 2 (Ω̄) solution of (4.68). Applying formula (4.67) with
f (u) = |u|τ −1 u (and thus F (u) = τ +1
1
|u|τ +1 ), and using Lemma 4.37, we obtain

|u|τ +1 2−n
Z Z Z
1 2
(4.70) u (x · ν)dS = n dx + |u|τ +1 dx
2 ∂Ω n τ + 1 2
 Ω Z Ω
n n−2 τ +1
(4.71) = − |u| dx ≥ 0,
τ +1 2 Ω
n n−2 n+2
which yields τ +1 ≥2 and hence a contradiction: τ ≤ n−2 .
If u is a positive C 2 (Ω̄) solution of (4.69) with τ = n+2
n−2 . Since
u = 0 on ∂Ω, by a sharp
maximum principle (see Serrin’s Maximum Principle below), we have un ≤ −σ on ∂Ω for a
positive number σ. By (4.71), we obtain
Z Z
σ2 x · ν dS ≤ u2n (x · ν) dS = 0,
∂Ω ∂Ω
from which we have a desired contradiction:
|x|2
Z Z
0= x · ν dS = ∆( ) dx = n|Ω|.
∂Ω Ω 2
This completes the proof. 
92 4. Variational Methods for PDEs

4.6.4. Radial Symmetry of Solutions on a Ball. Let Ω = B be the open unit ball in
Rn . We shall study the nonlinear problem
(4.72) ∆u + f (u) = 0 in Ω, u|∂Ω = 0.
We are interested in positive solutions: u > 0 in Ω. We assume that f : R → R is Lipschitz,
but is otherwise arbitrary. Our intention is to prove that u is necessarily radial, i.e., u(x)
depends only on r = |x|. This is a remarkable conclusion, since we are making essentially
no assumptions on the nonlinearity. The technique we use is due to Gidas, Ni & Nirenberg
[20], and is based upon an extension of the maximum principle and the method of moving
planes.

4.6.5. Serrin’s Maximum Principle.

Lemma 4.39. (Serrin’s Maximum Principle) Let Ω be a bounded domain in Rn . As-
sume u ∈ C 2 (Ω̄) satisfies
∆u + a(x)u ≤ 0, u ≥ 0 (6≡ 0) in Ω
where a(x) is bounded. If x0 ∈ ∂Ω, u(x0 ) = 0, and Ω satisfies the interior ball condition at
x0 , then the normal derivative un (x0 ) < 0. Moreover, u > 0 in Ω.

Proof. Set a = a+ + a− , where a+ = max(a, 0), a− = min(a, 0). Thus

∆u + a− (x)u ≤ −a+ u ≤ 0 in Ω
and the conclusions follows from the strong maximum principle and the boundary point
principle. 

4.6.6. Moving Plane Method. Let Ω+ denote the open upper half ball Ω ∩ {xn > 0},
and Ω− the open lower half ball Ω ∩ {xn < 0}.
Lemma 4.40. Let u ∈ C 2 (Ω̄) be a positive solution of (4.72). Then
uxn < 0 in Ω near ∂Ω+ .

Proof. Fix any point x0 ∈ ∂Ω+ and let ν = (ν1 , ..., νn ) denote the outer unit normal to
∂Ω+ at x0 . Note that νn > 0. We claim that uxn < 0 in Ω near x0 .
We shall give the proof under the assumption f (0) ≥ 0. Then
0 = −∆u − f (u) + f (0) − f (0)
Z 1
∂
≤ −∆u − f (su(x))ds
0 ∂s
≤ −∆u + cu
R1
for c(x) = − 0 f 0 (su(x))ds. According to Lemma 4.39, uν (x0 ) = ∇u(x0 ) · ν < 0. Since
∇u is parallel to ν on ∂Ω and νn > 0, we conclude that uxn (x0 ) < 0, and thus uxn <
0 in Ω near x0 . 

We now apply the method of moving planes.

(i) If 0 ≤ λ ≤ 1, set Pλ ≡ {x ∈ Rn : xn = λ}.
(ii) Let xλ ≡ (x1 , . . . , xn−1 , 2λ − xn ) denote the reflection of x in Pλ .
(iii) Let Eλ ≡ {x ∈ Ω : λ < xn < 1}.
4.6. Nonexistence and Radial Symmetry 93

Theorem 4.41. Let u ∈ C 2 (B̄) be a positive solution of (4.72). Then u is radial, i.e.,
u(x) = v(r) (r = |x|)
for some strictly decreasing function v : [0, 1] → [0, ∞).

Proof. Consider for each 0 ≤ λ < 1 the statement

(4.73) u(x) < u(xλ ) for all x ∈ Eλ .
According to Lemma 4.40, we see that this statement is valid for each λ < 1, λ sufficiently
close to 1. Set
λ0 ≡ inf{0 ≤ λ ≤ 1 : (4.73) holds for each λ ≤ µ < 1}.
We will show that λ0 = 0. If to the contrary λ0 > 0, set
w(x) = u(xλ0 ) − u(x) (x ∈ Eλ0 ).
Then
−∆w = f (u(xλ0 )) − f (u(x)) = −cw in Eλ0
for Z 1
c(x) = − f 0 (su(xλ0 ) + (1 − s)u(x))dx.
0
As w ≥ 0 in Eλ0 , we deduce from Lemma 4.39 (applied to Eλ0 ) that
w > 0 in Eλ0 , wxn > 0 on Pλ0 ∩ Ω.
Thus
(4.74) u(x) < u(xλ0 ) in Eλ0
and
(4.75) uxn < 0 on Pλ0 ∩ Ω.
Using (4.75) and Lemma 4.40 we conclude
(4.76) u(xn ) < 0 on Pλ0 −ε ∩ Ω for all 0 ≤ ε ≤ ε0
if ε0 is sufficiently small. Then (4.74) and the continuity of u imply
(4.77) u(x) < u(xλ0 −ε ) in Eλ0 −ε for all 0 ≤ ε ≤ ε0
if ε0 is small enough. Assertion (4.77) contradicts our choice of λ0 .
Since λ0 = 0, we see that
u(x1 , . . . , xn−1 , −xn ) ≥ u(x1 , . . . , xn ) for all x ∈ Ω+ .
A similar argument in Ω− proves
u(x1 , . . . , xn−1 , −xn ) ≤ u(x1 , . . . , xn ) for all x ∈ Ω+ .
Thus u is symmetric in the plane P0 and uxn = 0 on P0 .
This argument applies as well after any rotation of coordinate axes, and so the Theorem
follows. 
 Chapter 5

Weak Lower
Semicontinuity on
Sobolev Spaces

As discussed before, in many variational problems, weak lower semicontinuity is an essential

condition for existence of minimizers using the minimization method. In this chapter, we
study the conditions for weak lower semicontinuity of a multiple integral functional I(u) in
the Sobolev space W 1,p (Ω; RN ). Assume
Z
I(u) = F (x, u, Du) dx.
Ω

Recall that I is called weakly lower semicontinuous on W 1,p (Ω; RN ) if

(5.1) I(ū) ≤ lim inf I(uν ) whenever uν * ū weakly in W 1,p (Ω; RN ).
ν→∞

5.1. The Convex Case

5.1.1. Tonelli’s Theorem. We first prove a semicontinuity result of Tonelli.

Theorem 5.1. Let F (x, s, ξ) ≥ 0 be smooth and convex in ξ. Assume F, Fξ are both
continuous in (x, s, ξ). Then the functional I(u) defined above is sequentially weakly (weakly
star if p = ∞) lower semicontinuous on W 1,p (Ω; RN ) for all 1 ≤ p ≤ ∞.

Proof. We need only to prove I(u) is w.l.s.c. on W 1,1 (Ω; RN ). To this end, assume {uν }
is a sequence weakly convergent to u in W 1,1 (Ω; RN ). We need to show
I(u) ≤ lim inf I(uν ).
ν→∞

By the Sobolev embedding theorem it follows that (via a subsequence) uν → u in L1 (Ω; RN ).

We can also assume uν (x) → u(x) for almost every x ∈ Ω. Now, for any given δ > 0 we
choose a compact set K ⊂ Ω such that
(i) uν → u uniformly on K and |Ω \ K| < δ (by Egorov’s theorem);
(ii) u, Du are continuous on K (by Lusin’s theorem).

95
96 5. Weak Lower Semicontinuity on Sobolev Spaces

Since F (x, s, ξ) is smooth and convex in ξ, it follows that

F (x, s, η) ≥ F (x, s, ξ) + Fξk (x, s, ξ) (ηik − ξik ) ∀ξ, η ∈ MN ×n .

i

Therefore, since F ≥ 0,
Z
I(uν ) ≥ F (x, uν , Duν ) dx
K
Z h i
≥ F (x, uν , Du) + Fξk (x, uν , Du) (Di ukν − Di uk )
i
ZK Z
= F (x, uν , Du) + Fξk (x, u, Du) (Di ukν − Di uk )
i
K K
Z
+ [Fξk (x, uν , Du) − Fξk (x, ukν , Du)] (Di ukν − Di uk ).
i i
K

Since F (x, s, ξ) is uniformly continuous on bounded sets and uν (x) → u(x) uniformly on K
we have
Z Z
lim F (x, uν , Du) dx = F (x, u, Du) dx,
ν→∞ K K

lim kFξk (x, uν , Du) − Fξk (x, ukν , Du)kL∞ (K) = 0.

ν→∞ i i

Now since Fξk (x, u, Du) is bounded on K and Di ukν converges to Di uk weakly in L1 (Ω) as
i
ν → ∞, we thus have
Z
lim Fξk (x, u, Du) (Di ukν − Di uk ) dx = 0.
ν→∞ K i

From these estimates, noting that, for any two sequences {aν }, {bν },

(5.2) lim inf (aν + bν ) ≥ lim inf aν + lim inf bν ,

ν→∞ ν→∞ ν→∞

we have
Z
lim inf I(uν ) ≥ F (x, u, Du).
ν→∞ K

If F (x, u, Du) ∈ L1 (Ω), i.e., I(u) < ∞, then for any given  > 0, we use Lebesgue’s absolute
continuity theorem to determine δ > 0 so that
Z Z
F (x, u, Du) ≥ F (x, u, Du) − , ∀E ⊂ Ω, |Ω \ E| < δ.
E Ω

On the other hand, if I(u) = ∞ then for any given large number M > 0 we choose δ > 0
so that
Z
F (x, u, Du) dx > M, ∀E ⊂ Ω, |Ω \ E| < δ.
E

In any of these two cases, by letting either  → 0 or M → ∞, we obtain

lim inf I(uν ) ≥ I(u).

ν→∞

The theorem is proved. 

5.2. Morrey’s Quasiconvexity 97

5.1.2. Existence in the Convex Case. Using the theorem, we obtain the following
existence result for convex functionals.
Theorem 5.2. In addition to the hypotheses of the previous theorem, assume there exists
1 < p < ∞ such that
F (x, s, ξ) ≥ c |ξ|p − C(x),
where c > 0, C ∈ L1 (Ω) are given. If for some ϕ ∈ W 1,p (Ω; RN ), I(ϕ) < ∞, then mini-
mization problem inf u∈Dϕ I(u) has a minimizer in the Dirichlet class Dϕ .

Proof. This follows from the abstract existence Theorem 4.6 in the previous chapter. 

Remark. Both theorems in this section hold for more general functions F (x, s, ξ). For ex-
ample, we can replace the continuity condition by the Carathéodory condition.

5.2. Morrey’s Quasiconvexity

In this section, we will derive a condition which, under the mild general assumption, will
be the “right” (necessary and sufficient) condition for the weak lower semicontinuity for
integral functionals on Sobolev spaces. This will be Morrey’s quasiconvexity condition;
see Morrey [27, 28]. Please be warned that there is at least one other quasiconvexity in the
analysis that has a totally different meaning.

5.2.1. Lipschitz Convergence. Note that W 1,∞ (Ω; RN ) can be identified with the space
of all Lipschitz maps from Ω to RN . A sequence {uν } converges to u in the weak star
topology of W 1,∞ (Ω; RN ) if and only if {uν } converges to u in the sense of Lipschitz
convergence; that is,
1) uν → u uniformly in C(Ω̄; RN );
2) the Lipschitz norms of uν and u are bounded.

5.2.2. Quasiconvexity as Necessary Condition. The following result, mainly due to

Morrey [27], gives the necessary condition for the lower semicontinuity under the Lipschitz
convergence of the multiple integral
Z
I(u) = F (x, u(x), Du(x)) dx.
Ω

Theorem 5.3. Assume F (x, s, ξ) is continuous on Ω̄ × RN × MN ×n . Assume the functional

I(u) is s.l.s.c. with respect to Lipschitz convergence on W 1,∞ (Ω; RN ). Then the following
condition holds for all x0 ∈ Ω, s0 ∈ RN , ξ0 ∈ MN ×n and all φ ∈ W01,∞ (Ω; RN ) :
Z
(5.3) F (x0 , s0 , ξ0 ) ≤ − F (x0 , s0 , ξ0 + Dφ(x)) dx.
Ω

Proof. Let Q be a fixed open cube containing Ω̄ with center x̄ and side-length 2L. We
prove this theorem by several lemmas.
Lemma 5.4. Suppose
Z
(5.4) F (x0 , s0 , ξ0 ) ≤ − F (x0 , s0 , ξ0 + Dφ(x)) dx
Q

holds for all φ ∈ W01,∞ (Ω; RN ). Then (5.3) holds.

98 5. Weak Lower Semicontinuity on Sobolev Spaces

Proof. For any φ ∈ W01,∞ (Ω; RN ), we extend φ by zero onto Q; then φ ∈ W01,∞ (Q; RN ).
Inserting it into (5.4) yields (5.3). 

In the following, let x0 ∈ Ω, s0 ∈ RN , ξ0 ∈ MN ×n be given. Define ũ(x) = s0 + ξ0 ·

(x − x0 ). Let also φ ∈ W01,∞ (Q; RN ) be given. Assume Q0 ⊂⊂ Ω is an arbitrarily given
cube containing x0 with side-length 2l. For any positive integer ν we divide each side of
Q0 into 2ν intervals of equal length, each being equal to 2−ν+1 l. This divides Q0 into 2nν
small cubes {Qνj } with j = 1, 2, ..., 2nν . Denote the center of each cube Qνj by x̄νj and define
a function uν : Ω → RN as follows:
( nν
ũ(x) if x ∈ Ω \ ∪2j=1 Qνj ;
uν (x) = −ν ν
ũ(x) + 2 L l φ x̄ + 2 l L (x − x̄νj ) if x ∈ Qνj , 1 ≤ j ≤ 2nν .

We easily see that uν → ũ uniformly on Ω. Moreover,

nν
if x ∈ Ω \ ∪2j=1 Qνj ;

ξ0
Duν (x) = ν
ξ0 + Dφ x̄ + 2 l L (x − x̄νj ) if x ∈ Qνj , 1 ≤ j ≤ 2nν .

Therefore {Duν } is uniformly bounded and, by definition, {uν } converges to ũ as ν → ∞

in the sense of Lipschitz convergence. Note that
Z
I(ũ) = F (x, ũ(x), ξ0 ) dx
Ω
and that Z
I(uν ) = F (x, uν (x), Duν ) dx
Ω
Z Z
= F (x, ũ, ξ0 ) dx + F (x, uν , Duν ) dx.
Ω\Q0 Q0
Therefore, by the lower semicontinuity of I, we thus have
Z Z
(5.5) F (x, ũ, ξ0 ) dx ≤ lim inf F (x, uν , Duν ) dx.
Q0 ν→∞ Q0

From the uniform continuity of F (x, s, ξ) on bounded sets and the fact that uν → ũ uni-
formly on Ω we have
Z Z
(5.6) lim inf F (x, uν , Duν ) dx = lim inf F (x, ũ, Duν ) dx.
ν→∞ Q0 ν→∞ Q0

We now compute
2nν Z
2ν L
Z 
X

F (x, ũ, Duν ) dx = F x, ũ, ξ0 + Dφ x̄ + (x − x̄νj ) dx
Q0 Qνj l
j=1
nν
2
2ν L
Z 
X

= F x̄νj , ũ(x̄νj ), ξ0 + Dφ x̄ + (x − x̄νj ) dx + o(1)
Qνj l
j=1
2 nν n Z
X l
F x̄νj , ũ(x̄νj ), ξ0 + Dφ(y) dy + o(1)


= ν
2 L Q
j=1

2 nν
X
(5.7) = F̃ (x̄νj ) |Qνj | + o(1),
j=1
5.2. Morrey’s Quasiconvexity 99

where o(1) → 0 as ν → ∞, and

Z
F̃ (x) = − F (x, ũ(x), ξ0 + Dφ(y)) dy.
Q

This function is continuous on Q0

and the sum in (5.7) is simply the Riemann sum of the
integral of F̃ over Q0 . Therefore, we arrive at
Z Z
lim F (x, ũ, Duν ) dx = F̃ (x) dx,
ν→∞ Q0 Q0

which by (5.6) implies Z Z

F (x, ũ(x), ξ0 ) dx ≤ F̃ (x) dx.
Q0 Q0
This inequality holds for any cube Q0 ⊂⊂ Ω containing x0 ; therefore,
F (x0 , ũ(x0 ), ξ0 ) ≤ F̃ (x0 ).
This is nothing but
Z
F (x0 , s0 , ξ0 ) ≤ − F (x0 , s0 , ξ0 + Dφ(y)) dy.
Q
Finally, the proof of Theorem 5.3 is complete. 

Motivated by this theorem, we have the following definition of quasiconvex functions

in the sense of Morrey.
Morrey’s Quasiconvexity. A function F : MN ×n → R̄ is called quasiconvex (in the sense
of Morrey) if
Z
(5.8) F (ξ) ≤ − F (ξ + Dφ(x)) dx
Ω

holds for all φ ∈ W01,∞ (Ω; RN ).

5.2.3. Quasiconvexity as Sufficient Condition. The following result, due to Meyers,

will be useful to relax the zero boundary condition on φ.
Theorem 5.5. Let F : MN ×n → R be continuous and quasiconvex. For every bounded
set Q ⊂ Rn and every sequence {zν } in W 1,∞ (Q; RN ) converging to zero in the sense of
Lipschitz convergence, we have
Z
F (ξ) ≤ lim inf − F (ξ + Dzν (x)) dx
ν→∞ Q

for every ξ ∈ MN ×n .

Proof. Let Qk = {x ∈ Q | dist(x, ∂Q) > 1/k}. Then Qk ⊂⊂ Q and |Q\Qk | → 0 as k → ∞.

Choose a cut-off function ζk ∈ C0∞ (Q) such that

0 ≤ ζk ≤ 1, ζk  = 1, Mk = kDζk kL∞ < ∞.
Qk

Since zν → 0 uniformly on Q we can choose a subsequence {νk } such that

kzνk kL∞ ≤ (Mk + 1)−1 ∀ k = 1, 2, ...
and we may also assume
Z Z
lim F (ξ + Dzνk (x)) dx = lim inf F (ξ + Dzν (x)) dx.
k→∞ Q ν→∞ Q
100 5. Weak Lower Semicontinuity on Sobolev Spaces

Define φk = ζk zνk . Then φk ∈ W01,∞ (Q; RN ) and we can use them as test functions in the
definition of quasiconvexity to obtain
Z
|Q| F (ξ) ≤ F (ξ + Dφk (x)) dx
Q
Z Z
= F (ξ + Dzνk ) + F (ξ + ζk Dzνk + zνk ⊗ Dζk )
Qk Q\Qk
Z
= F (ξ + Dzνk (x)) dx + k ,
Q
where Z
k = [F (ξ + ζk Dzνk + zνk ⊗ Dζk u) − F (ξ + Dzνk (x))] dx.
Q\Qk
Since F (ξ) is bounded on bounded sets and |Q \ Qk | → 0 as k → ∞, we easily have k → 0
as k → ∞. Therefore,
Z
|Q| F (ξ) ≤ lim inf F (ξ + Dzν (x)) dx.
ν→∞ Q
This completes the proof. 

We now prove the sufficiency of quasiconvexity for the lower semicontinuity of the func-
tional Z
I(u) = F (x, u(x), Du(x)) dx
Ω
under the Lipschitz convergence on Ω.
Theorem 5.6. Assume F (x, s, ξ) is continuous on Ω̄ × RN × MN ×n and is quasiconvex in
ξ. Then the functional I defined above is s.l.s.c. with respect to Lipschitz convergence on
Ω.

Proof. Let {zk } be any sequence converging to 0 in the sense of Lipschitz convergence on
Ω, and let u ∈ W 1,∞ (Ω; RN ) be any given function. We need to show
Z Z
(5.9) F (x, u, Du) ≤ lim inf F (x, u + zk , Du + Dzk ).
Ω k→∞ Ω
For any given  > 0, we choose finitely many disjoint cubes Qj contained in Ω such that
Z
I(u) ≤ F (x, u, Du) dx + 
∪Qj

and Z
I(u + zk ) ≥ F (x, u + zk , Du + Dzk ) dx − ,
∪Qj
for all k = 1, 2, · · · . In what follows, we prove for each cube Q = Qj
Z
IQ (u) ≡ F (x, u, Du) dx ≤ lim inf IQ (u + zk ).
Q k→∞

This, by (5.2), will certainly imply the conclusion of the theorem. To this end, for each
positive integer ν, we divide Q into small cubes {Qνj } with center x̄νj as in the proof of
Theorem 5.3: nν
2[
Q= Qνj ∪ E, |E| = 0.
j=1
5.2. Morrey’s Quasiconvexity 101

Define Z Z
(u)νj = − u(x) dx, (Du)νj = − Du(x) dx,
Qνj Qνj
and nν nν
2
X 2
X
ν
U (x) = (u)νj · χQνj (x), ν
M = (Du)νj · χQνj (x).
j=1 j=1
Note that
kU ν kL∞ + kM ν kL∞ ≤ kukW 1,∞
and that the sequences {U ν } and {M ν } converge almost everywhere to u and Du on Q as
ν → ∞, respectively. We now estimate IQ (u + zk ).
Z
IQ (u + zk ) = F (x, u + zk , Du + Dzk ) = ak + bνk + cνk + dν + IQ (u),
Q
where
Z
ak = [F (x, u + zk , Du + Dzk ) − F (x, u, Du + Dzk )] dx,
Q
2nν
Z
X
bνk = [F (x, u, Du + Dzk ) − F (x̄νj , (u)νj , (Du)νj + Dzk )] dx,
j=1 Qνj

2nν
Z
X
cνk = [F (x̄νj , (u)νj , (Du)νj + Dzk ) − F (x̄νj , (u)νj , (Du)νj )] dx,
j=1 Qνj

2nν
Z
X
ν
d = [F (x̄νj , (u)νj , (Du)νj ) − F (x, u, Du)] dx.
j=1 Qνj

By the uniform continuity of F (x, s, ξ) on bounded sets and the pointwise convergence of
{U ν } and {M ν } we have
lim ak = 0, lim dν = 0
k→∞ ν→∞
and limν→∞ bνk = 0 uniformly with respect to k. We apply Theorem 5.5 to each Qνj to
obtain, by (5.2),
lim inf cνk ≥ 0
k→∞
for all ν = 1, 2, · · · . Therefore, again by (5.2),
lim inf IQ (u + zk ) ≥ IQ (u),
k→∞
as desired. The proof is complete. 

5.2.4. Weak Lower Semicontinuity on Sobolev Spaces. Quasiconvexity is also the

“right” condition for weak lower semicontinuity of integral functionals on W 1,p (Ω; RN ). The
most general theorem in this direction is the following theorem due to Acerbi & Fusco [1].
Theorem 5.7. (Acerbi-Fusco’s Theorem) Let F (x, s, ξ) be a Carathéodory function.
Assume for some 1 ≤ p < ∞
0 ≤ F (x, s, ξ) ≤ a(x) + C (|s|p + |ξ|p ),
where C > 0 isR a constant and a(x) ≥ 0 is a locally integrable function in Ω. Then func-
tional I(u) = Ω F (x, u, Du) dx is w.s.l.s.c. on W 1,p (Ω; RN ) if and only if F (x, s, ξ) is
quasiconvex in ξ.
102 5. Weak Lower Semicontinuity on Sobolev Spaces

5.2.5. Existence in the General Case.

Theorem 5.8 (Existence of minimizers). Let F (x, s, ξ) be Carathéodory and quasiconvex
in ξ and satisfy
max{0, c |ξ|p − C(x)} ≤ F (x, s, ξ) ≤ a(x) + C (|s|p + |ξ|p )
for some 1 < p < ∞, where c > 0 is a constant and C(x), a(x) are given integrable functions
in Ω. Then, for any ϕ ∈ W 1,p (Ω; RN ), the minimization problem
Z
min F (x, u(x), Du(x)) dx
u∈Dϕ Ω
has a minimizer in the Dirichlet class Dϕ .

5.3. Properties of Quasiconvex Functions

5.3.1. Domain Independence. We prove that quasiconvexity is independent of the do-
main Ω.
Theorem 5.9. Let F : MN ×n → R̄ be such that (5.8) holds for all φ ∈ W01,∞ (Ω; RN ). Then
for any bounded open set G ⊂ Rn with |∂G| = 0 one has
Z
(5.10) F (ξ) ≤ − F (ξ + Dψ(y)) dy, ∀ξ ∈ MN ×n
G

holds for all ψ ∈ W01,∞ (G; RN ).

Proof. Let G ⊂ Rn be any bounded open set with |∂G| = 0, and ψ ∈ W01,∞ (G; RN ).
Assume ȳ ∈ G. For any x ∈ Ω and  > 0 let
G(x, ) = {z ∈ Rn | z = x +  (y − ȳ) for some y ∈ G}.
Then there exists an x > 0 such that x ∈ G(x, ) ⊂ Ω for all x ∈ Ω and 0 <  < x . This
means the family 
{G(x, )  x ∈ Ω, 0 <  < x }
covers Ω in the sense of Vitali covering. Therefore, there exists a countable disjoint
subfamily {G(xj , j )} and a set E of measure zero such that
∞
[
(5.11) Ω= G(xj , j ) ∪ E.
j=1

We now define a function φ : Ω → RN as follows.

if x ∈ ∞
( S
0 j=1 ∂[G(xj , j )] ∪ E,
φ(x) = x−xj

j ψ ȳ + j if x ∈ G(xj , j ) for some j.

One can verify that φ ∈ W01,∞ (Ω; RN ) and

x − xj

Dφ(x) = Dψ ȳ + ∀x ∈ G(xj , j ).
j
Therefore, from (5.8), it follows that
Z
F (ξ) |Ω| ≤ F (ξ + Dφ(x)) dx
Ω
∞ Z  
X x − xj

= F ξ + Dψ ȳ + dx
G(xj ,j ) j
j=1
5.3. Properties of Quasiconvex Functions 103

∞
X Z
= nj F (ξ + Dψ(y)) dy
j=1 G

|Ω|
Z
= F (ξ + Dψ(y)) dy,
|G| G
where the last equality follows since, by (5.11), ∞ n
P
j=1 j = |Ω|/|G|. We have thus proved
(5.10). 

In the following, let Σ be the unit cube in Rn ; that is

Σ = {x ∈ Rn | 0 < xα < 1, α = 1, 2, · · · , n}.
Note that |Σ| = 1. Let f : Rn → RN be any given function. We say f is Σ-periodic
if f (· · · , xα , · · · ) is 1-periodic in xα for all α = 1, 2, · · · , n. Quasiconvexity can be also
characterized by the following condition.
Theorem 5.10. Let F : MN ×n → R be continuous. Then F is quasiconvex if and only if
Z
(5.12) F (ξ) ≤ F (ξ + Dφ(x)) dx ∀ξ ∈ MN ×n
Σ
for all Σ-periodic Lipschitz functions φ ∈ W 1,∞ (Rn ; RN ).

Proof. Since any function ψ ∈ W01,∞ (Σ; RN ) can be extended as a Σ-periodic function on
Rn , we easily see that (5.12) implies (5.10) for G = Σ thus the quasiconvexity of F . We
have only to prove (5.12) holds if F is quasiconvex. Let φ ∈ W 1,∞ (Rn ; RN ) be a Σ-periodic
function. Define
1
φj (x) = φ(jx)
j
for all j = 1, 2, · · · . It is easily seen that φj → 0 in the sense of Lipschitz convergence on
W 1,∞ (Σ; RN ). Therefore the theorem of Meyers, Theorem 5.5, and the quasiconvexity of F
implies Z
F (ξ) ≤ lim inf F (ξ + Dφj (x)) dx.
j→∞ Σ
Note that Z Z
F (ξ + Dφj (x)) dx = F (ξ + Dφ(jx)) dx
Σ Σ
Z
= j −n F (ξ + Dφ(y)) dy
jΣ
and that, besides a set of measure zero,
j n
[
jΣ = (x̄ν + Σ),
ν=1
where x̄ν are the left-lower corner points of the subcubes obtained by dividing the sides of
jΣ into j-equal subintervals. Since Dφ(x) is Σ-periodic, we thus have
Z Xjn Z Z
n
F (ξ + Dφ(y))dy = F (ξ + Dφ(y))dy = j F (ξ + Dφ(x))dx,
jΣ ν=1 x̄ν +Σ Σ

and therefore Z
F (ξ) ≤ F (ξ + Dφ(x)) dx
Σ
as needed; the proof is complete. 
104 5. Weak Lower Semicontinuity on Sobolev Spaces

5.3.2. Convexity vs Quasiconvexity.

Lemma 5.11 (Jensen’s inequality). Let (E, µ) be a measure space with total mass µ(E) = 1
and let h : E → RL be an integrable function on E. If G : RL → R is a convex function,
then Z  Z
G h(x) dµ ≤ G(h(x)) dµ.
E E

Proof. Let F = E h(x) dµ. Since G : RL → R is convex, there exists lF ∈ RL (note that
R

lF = DG(F ) for almost every F ) such that

G(A) ≥ G(F ) + lF · (A − F ) ∀A ∈ RL .
Then G(h(x)) ≥ G(F ) + lF · (h(x) − F ) for all x ∈ E, and integrating over x ∈ E yields
Z Z
G(h(x)) dµ ≥ G(F ) + lF · (h(x) − F ) dµ = G(F ),
E E
which proves Jensen’s inequality. 
Theorem 5.12. If F : MN ×n → R is convex, then F is quasiconvex.

Proof. This follows easily from Jensen’s inequality and the divergence theorem given above.


5.3.3. Quasiconvexity vs Rank-one Convexity. Recall that F : MN ×n → R̄ is rank-

one convex if for any ξ ∈ MN ×n , q ∈ RN , p ∈ Rn the function f (t) = F (ξ + tq ⊗ p) is a
convex function of t ∈ R.

We have the following result.

Theorem 5.13. Every finite-valued quasiconvex function is rank-one convex.

Proof. Let F be quasiconvex. We need to show that for any ξ ∈ MN ×n , q ∈ RN , p ∈ Rn

the function f (t) = F (ξ + tq ⊗ p) is a convex function of t ∈ R; that is, for all 0 < θ < 1
and t, s ∈ R,
(5.13) f (θt + (1 − θ)s) ≤ θf (t) + (1 − θ)f (s).
which is equivalent to
(5.14) ˜ ≤ θF (ξ˜ + aq ⊗ p) + (1 − θ)F (ξ˜ + bq ⊗ p),
F (ξ)
where
ξ˜ = ξ + [θt + (1 − θ)s]q ⊗ p, a = (1 − θ)(t − s), b = θ(s − t).
Assume t > s; so a > 0 and b < 0. Let ζ(τ ) be the periodic Lipschitz function of peroid
1 on R satisfying ζ(τ ) = aτ for 0 ≤ τ ≤ θ and ζ(τ ) = b(τ − 1) for θ ≤ τ ≤ 1. Let G be a
cube of unit volume which is bounded between two planes {x · p = 0} and {x · p = 1}. For
x ∈ Rn , define
|x|∞ = max{|xi | | 1 ≤ i ≤ n}, δ(x) = dist∞ (x, ∂G) = inf{|x − y|∞ | y ∈ ∂G}.
Then δ ∈ W01,∞ (G) and ∇δ(x) ∈ {±ei | i = 1, 2, · · · , n}. Define functions by
uk (x) = k −1 ζ(kx · p) q, φk (x) = min{k −1 ζ(kx · p), δ(x)} q, k = 1, 2, · · · .
Note that Duk (x) ∈ {aq ⊗ p, bq ⊗ p}, φk ∈ W01,∞ (G; RN ) and
lim |{x ∈ G | Duk (x) 6= Dφk (x)}| = 0.
k→∞
5.3. Properties of Quasiconvex Functions 105

Hence, by (5.8) and noting that Duk (x) and Dφk (x) take only on finitely many values,
Z
˜
F (ξ) ≤ lim F (ξ˜ + Dφk (x)) dx
k→∞ G
Z
= lim F (ξ˜ + Duk (x)) dx
k→∞ G

= θF (ξ˜ + aq ⊗ p) + (1 − θ)F (ξ˜ + bq ⊗ p).

This completes the proof. 

Remark. Another proof in the case of continuous F is as follows. Let F be quasiconvex. If

F is of class C 2 , then Z
f (t) = F (ξ + tDφ(x)) dx
Ω
takes its minimum at t = 0. Therefore f 00 (0) ≥ 0; that is,
Z
Fξk ξl (ξ) Di φk (x)Dj φl (x) dx ≥ 0
i j
Ω
for all φ ∈ C0∞ (Ω; RN ). This implies the weak Legendre-Hadamard condition:
Fξk ξl (ξ) q k q l pi pj ≥ 0 ∀ q ∈ RN , p ∈ Rn ,
i j

which is equivalent to that F is rank-one convex. For a continuous F , let F  = F ∗ ρ be

the regularization of F . Then F  is of class C ∞ and can be shown to be quasiconvex, and
hence F  satisfies (5.14) for all  > 0; letting  → 0 yields that F satisfies (5.14) and thus
is rank-one convex.
Lemma 5.14. Let n ≥ 2 and let A, B ∈ MN ×n be such that rank(A − B) = 1. Define
(
0 ξ ∈ {A, B},
F (ξ) =
∞ ξ∈ / {A, B}.
Then F is quasiconvex convex, but not rank-one convex.

Proof. The rank-one convexity of F would imply

0 ≤ F (λA + (1 − λ)B) ≤ λF (A) + (1 − λ)F (B) = 0
for all 0 < λ < 1. Hence F is not rank-one convex. To see F is quasiconvex, given ξ ∈ MN ×n
and φ ∈ W01,∞ (Σ; RN ), we need to show
Z
(5.15) F (ξ) ≤ F (ξ + Dφ(x)) dx.
Σ
Since the integral on the right-hand side takes only two values of {0, ∞}, we only need to
prove the inequality when Z
F (ξ + Dφ(x)) dx = 0.
Σ
In this case, we must have ξ + Dφ(x) ∈ {A, B} for almost every x ∈ Σ. We claim, in this
case, one must have ξ ∈ {A, B} and hence (5.15) is valid. To prove this, let ξ + Dφ(x) =
χE (x)A + (1 − χE (x))B, where E = {x ∈ Σ | ξ + Dφ(x) = A}. Integrating this over Σ yields
ξ = λA + (1 − λ)B, where λ = |E| ∈ [0, 1]. Hence
Dφ(x) ∈ {A − ξ, B − ξ} = {(1 − λ)(A − B), λ(A − B)} a.e. x ∈ Σ.
Let A − B = q ⊗ p with p ∈ S n−1 .
Let e ∈ S n−1
be a vector perpendicular to p. Then
Dφ(x) e = 0 for almost every x ∈ Σ. This implies φ(x + te) is independent of t as long as
106 5. Weak Lower Semicontinuity on Sobolev Spaces

x + te ∈ Σ̄; hence, choosing t such that x + te ∈ ∂Σ, we have φ(x) = 0 for all x ∈ Σ. Hence,
ξ ∈ {A, B}, as claimed. 
Theorem 5.15. Every finite-valued quasiconvex function is locally Lipschitz continuous.

Proof. Let F : MN ×n → R be quasiconvex. Then f (t) = F (ξ + tq ⊗ p) is a finite convex

function of t ∈ R; hence f is locally Lipschitz continuous. From this the theorem follows. 
Lemma 5.16. If F is a polynomial of degree two (quadratic polynomial) then F is quasi-
convex if and only if F is rank-one convex.

Proof. Assume F is a rank-one convex quadratic polynomial. We show F is quasiconvex.

Since subtraction of an affine function from a function does not change the quasiconvexity
or rank-one convexity, we thus assume F is a homogeneous quadratic polynomial given by
F (ξ) = Akl k l
ij ξi ξj (summation notation is used here and below)
with Akl
ij are constants. Note that the rank-one convexity is equivalent to the weak Legendre-
Hadamard condition:
X n X N
Akl k l
ij q q pi pj ≥ 0.
i,j=1 k,l=1
Also,
(5.16) F (ξ + η) = F (ξ) + Akl k l kl l k kl k l
ij ξi ηj + Aij ξj ηi + Aij ηi ηj .

Hence the linear system defined by constants Ãαβ αβ

ij = Aij + δij δ
αβ satisfies the Legendre-

Hadamard condition, where  > 0 and δij , δ αβ are the usual delta notation. Using the
Fourier transform as before, we can show that
Z
Akl k l
ij Di φ (x)Dj φ (x) dx ≥ 0
Rn

for all φ ∈ C0∞ (Rn ; RN ). Hence by (5.16)

Z Z
F (ξ + Dφ(x)) dx = F (ξ) + Akl k l
ij Di φ (x)Dj φ (x) dx ≥ F (ξ),
Σ Rn

holds for all φ ∈ C0∞ (Σ; RN ). This proves the quasiconvexity of F. 

Using the definition of null-Lagrangians given later, we have the following.

Lemma 5.17. A rank-one convex third degree polynomial must be a null-Lagrangian and
thus quasiconvex.

Proof. Let F be a rank-one convex third degree polynomial. Then the polynomial f (t) =
F (ξ + tq ⊗ p) is convex and of degree ≤ 3 in t, and hence the degree of f (t) cannot be 3.
Note that the coefficient of t2 term in f is half of
(5.17) Fξk ξl (ξ) pi pj q k q l ≥ 0,
i j

which holds for all ξ, p, q. Since Fξk ξl (ξ) is linear in ξ, condition (5.17) implies
i j

Fξk ξl (ξ) pi pj q k q l ≡ 0.
i j

Therefore f (t) = F (ξ +tq⊗p) is affine in t and hence F is rank-one affine. Consequently, the
result follows from the fact that a rank-one affine function must be a null-Lagrangian. 
5.3. Properties of Quasiconvex Functions 107

5.3.4. Šverák’s Example. The following example of Šverák settles a long-standing open
problem raised by Morrey.
Theorem 5.18 (Šverák [36]). If n ≥ 2, N ≥ 3 then there exists a rank-one convex function
F : MN ×n → R which is not quasiconvex.

Proof. We only prove the theorem for n = 2, N = 3. Consider the periodic function
u : R2 → R3 by
1
u(x) = (sin 2πx1 , sin 2πx2 , sin 2π(x1 + x2 )).
2π
Then  
cos 2πx1 0
Du(x) =  0 cos 2πx2 .
cos 2π(x1 + x2 ) cos 2π(x1 + x2 )
Let L be the linear subspace of M = M3×2 ; that is,
   
 r 0  
L = [r, s, t] ≡  0 s   r, s, t ∈ R .

t t
 

Note that a matrix ξ = [r, s, t] ∈ L is of rank ≤ 1 if and only if at most one of {r, s, t} is
nonzero. Define g : L → R by g([r, s, t]) = −rst. Using formula 2 cos α · cos β = cos(α + β) +
cos(α − β), we easily obtain
1 1

g(Du(x)) = − − sin 4π(x1 + x2 ) + cos 4πx1 + cos 4πx2
4 4
and hence by a direct computation that
Z Z 1Z 1
1
(5.18) g(Du(x)) dx = g(Du(x)) dx1 dx2 = − .
Σ 0 0 4
We now extend g to the whole M. Let P : M → L be the orthogonal projection onto L and
consider the fourth degree polynomials:
(5.19) F,k (ξ) = g(P ξ) +  (|ξ|2 + |ξ|4 ) + k |ξ − P ξ|2 .
Lemma 5.19. For each  > 0 there exists a k = k > 0 such that F,k is rank-one convex.

Proof. We use contradiction method. Suppose otherwise; then there exists an 0 > 0 such
that F0 ,k is not rank-one for each k > 0. Hence there exist sequences ξk ∈ M, pk ∈ R2 , qk ∈
R3 with |pk | = |qk | = 1 such that
∂ 2 F0 ,k (ξk )
(5.20) qkα qkβ pki pkj ≡ D2 F0 ,k (ξk ) [qk ⊗ pk , qk ⊗ pk ] ≤ 0, ∀ k = 1, 2, · · · .
∂ξiα ∂ξjβ
Computing f 00 (0) for f (t) = F,k (ξ + tη) yields
D2 F,k (ξ) [η, η] = f 00 (0)
= D2 g(P ξ) [P η, P η] + 2 |η|2 +  (4|ξ|2 |η|2 + 8(ξ · η)2 ) + 2k |η − P η|2 .
The term D2 g(P ξ) is linear in ξ; the third term is quadratic and positive definite in ξ if
η 6= 0 (this is the reason the |ξ|4 -term is needed for F,k ). Using this formula with ξ = ξk
and η = qk ⊗ pk and from (5.20), we deduce {ξk } is bounded as k → ∞. Assume, via
subsequence,
ξk → ξ,¯ qk → q̄, pk → p̄ as k → ∞.
108 5. Weak Lower Semicontinuity on Sobolev Spaces

Since D2 F,j (ξ) [η, η] ≤ D2 F,k (ξ) [η, η] for all k ≥ j, we deduce

(5.21) ¯ [P (q̄ ⊗ p̄), P (q̄ ⊗ p̄)] + 20 + 2j |P (q̄ ⊗ p̄) − q̄ ⊗ p̄ | ≤ 0

D2 g(P ξ)
for all j = 1, 2, · · · . Hence P (q̄ ⊗ p̄) = q̄ ⊗ p̄; that is, q̄ ⊗ p̄ ∈ L. This implies q̄ ⊗ p̄ = [a, b, c],
where at most one of a, b, c is nonzero. Therefore, function
t 7→ g(P (ξ¯ + tq̄ ⊗ p̄)) = g(P ξ¯ + tq̄ ⊗ p̄)
is affine in t, and hence the first term in (5.21) vanishes. This yields the desired contradiction
0 ≤ 0. The lemma is proved. 

We now complete the proof of Šverák’s theorem. Let u be the periodic function above.
We choose  > 0 small enough such that
Z
1
 (|Du(x)|2 + |Du(x)|4 ) dx < .
Σ 4
Let F (ξ) = F,k (ξ) be a rank-one function determined by the previous lemma. Since
Du(x) ∈ L, it follows by (5.18) that
Z Z Z
F (Du(x)) dx = g(Du) +  (|Du|2 + |Du|4 ) < 0 = F (0).
Σ Σ Σ

This shows that F is not quasiconvex by Theorem 5.10 above. The theorem is now proved.


5.4. Polyconvex Functions and Null-Lagrangians

Unlike the convexity and rank-one convexity, quasiconvexity is a global property since the
inequality (5.8) is required to hold for all test functions. It is thus generally impossible to
verify whether a given function F (ξ) is quasiconvex. We have already seen that every convex
function is quasiconvex. However, there is a class of functions which are quasiconvex but
not necessarily convex. This class, mainly due to Morrey, has been called the polyconvex
functions in J. M. Ball [4]. In order to introduce the polyconvex functions, we need some
notation.

5.4.1. Determinant and Adjugate Matrix. Let ξ be a n×n square matrix. We denote
by det ξ and adj ξ the determinant and adjugate matrix of ξ, respectively, which satisfy
the following relation:
ξ (adj ξ) = (adj ξ) ξ = (det ξ) I,
where I is the n × n identity matrix. From this relation, we have
∂ det ξ
(5.22) = (adj ξ)T ,
∂ξ
where η T is the transpose matrix of η.

Lemma 5.20. For all u ∈ C 2 (Ω̄; Rn ) it follows that div(adj Du(x))T = 0; that is,
n
X
Di [(adj Du(x))ik ] = 0 (k = 1, 2, · · · , n).
i=1
5.4. Polyconvex Functions and Null-Lagrangians 109

Proof. This result follows from a general result on null-Lagrangians later, but we present
a direct proof here. Note that by the identity above, we have (det ξ)I = ξ T (adj ξ)T . For
ξ = Du(x) this implies
n
Di (uk (x)) (adj Du(x))jk ,
X
(det Du(x)) δij = i, j = 1, 2, · · · , n.
k=1
Differentiating this identity with respect to xj and summing over j = 1, 2, · · · , n, we have
n n
(Dj Di uk ) (adj Du)jk + (Di uk ) Dj [(adj Du)jk ]
X X
δij (adj Du)m k
k Dj (Dm u ) =
j,k,m=1 k,j=1

for i = 1, 2, · · · , n. This identity simplifies to read

 
n n
Dj [(adj Du)jk ] = 0
X X
(Di uk )  (i = 1, 2, · · · , n).
k=1 j=1

In short, this can be written as

(5.23) Du(x)T [div(adj Du(x))T ] = 0, x ∈ Ω.
Now, if det Du(x0 ) 6= 0 then by (5.23), div(adj Du(x0 ))T = 0. If instead det Du(x0 ) = 0,
we choose a sequence ν → 0 such that det(Du(x0 ) + ν I) 6= 0 for all ν. Use (5.23) with
ũ = u + ν x we have det Dũ(x0 ) 6= 0 and hence
0 = div(adj Dũ(x0 ))T = div(adj(Du(x0 ) + ν I))T = 0
for all ν → 0. Hence div(adj(Du(x0 )))T = 0. 

5.4.2. Subdeterminants. Let σ = min{n, N }. Given an integer k ∈ [1, σ], for any two
ordered sequences of integers
1 ≤ i1 < i2 < · · · < ik ≤ N, 1 ≤ α1 < α2 < · · · < αk ≤ n,
i
letJαi11iα2 2···ibe the determinant of the k × k matrix whose (q, p) position element is ξαqp
k
···αk (ξ)
for all 1 ≤ p, q ≤ k. Note that, by the usual notation,
∂(ui1 , ui2 , · · · , uik )
 iq 
i1 i2 ···ik ∂u
Jα1 α2 ···αk (Du(x)) = = det .
∂(xα1 , xα2 , · · · , xαk ) ∂xαp
Let J (ξ) be the collection of all Jαi11iα2 2···i
···αk (ξ) for all k ∈ [1, σ] with the same ordered integral
k

sequences {iq }, {αp } for all ξ. In this way, J defines a function from MN ×n to RL , where
σ   
X N n
L = L(n, N ) = .
k k
k=1

Theorem 5.21. Let J (ξ) be defined as above, and let Σ be the unit cube in ⊂ Rn . Then it
follows that Z


J ξ + Dφ(x) dx = J (ξ)
Σ
for all φ ∈ C0∞ (Σ; RN ) and ξ ∈ MN ×n .

Proof. Since each J (ξ) is given by a k × k-determinant, without loss of generality, we

12···k (ξ), where 1 ≤ k ≤ σ = min{n, N }. For
only prove this identity for J (ξ) = Jk (ξ) = J12···k
simplicity, set u(x) = ξx + φ(x). Let
x0 = (x1 , · · · , xk ), x00 = (xk+1 , · · · , xn ) if k + 1 ≤ n.
110 5. Weak Lower Semicontinuity on Sobolev Spaces

Let Σ0 , Σ00 be the unit cubes in x0 , x00 variables, respectively. Fix x00 ∈ Σ00 , for t ≥ 0,
consider maps Vt , Ut : Σ0 → Rk such that

Vti (x0 ) = txi + (ξx)i , Uti (x0 ) = txi + ui (x0 , x00 ).

We can choose t > 0 sufficiently large so that Vt , Ut are both diffeomorphisms on Σ0 , and
therefore
Z Z Z Z
0 0 0 0
det(DUt (x )) dx = dy = dy = det(DVt (x0 )) dx0 .
Σ0 Ut (Σ0 ) Vt (Σ0 ) Σ0

Since both sides are polynomials of t of degree k, it follows that this equality holds for all
t. When t = 0 this implies
Z Z
0 00 0
(5.24) Jk (ξ + Dφ(x , x )) dx = Jk (ξ) dx0 .
Σ0 Σ0

Integrating (5.24) over x00 ∈ Σ00 we deduce

Z
(5.25) Jk (ξ + Dφ(x)) dx = Jk (ξ),
Σ

completing the proof. 

5.4.3. Polyconvex Functions. A function F (ξ) on MN ×n is called a polyconvex func-

tion if there exists a convex function G : RL → R̄ such that F (ξ) = G(J (ξ)) for all
ξ ∈ MN ×n ; that is, F = G ◦ J on MN ×n .
Remark. For a polyconvex function we may have different convex functions in its repre-
sentation. For example, let n = N = 2 and F (ξ) = |ξ|2 − det ξ. In this case, consider
J (ξ) = (ξ, det ξ) ∈ R5 . Then we have

F (ξ) = G1 (J (ξ)), F (ξ) = G2 (J (ξ)),

where
G1 (ξ, t) = |ξ|2 − t, G2 (ξ, t) = (ξ11 − ξ22 )2 + (ξ12 + ξ21 )2 + t
are both convex functions of (ξ, t).

Theorem 5.22. A polyconvex function is quasiconvex.

Proof. Let F : MN ×n → R be a polyconvex function. Then there exists a convex function

G : RL → R such that F (ξ) = G(J (ξ)) for all ξ. Given ξ ∈ MN ×n and φ ∈ C0∞ (Σ; RN ), let
h(x) = J (ξ + Dφ(x)). Then Jensen’s inequality implies
Z  Z
G h(x) dx ≤ G(h(x)) dx.
Σ Σ

By the theorem above, the left-hand side is G(J (ξ)) = F (ξ) and therefore
Z Z Z
F (ξ) ≤ G(h(x)) dx = G(J (ξ + Dφ(x))) dx = F (ξ + Dφ(x)) dx,
Σ Σ Σ

proving that F is quasiconvex. 

5.4. Polyconvex Functions and Null-Lagrangians 111

5.4.4. Null-Lagrangians. A smooth function F : Ω × RN × MN ×n → R is called a null-

Lagrangian if the system of Euler-Lagrange equations (4.4) for the energy
Z
I(u) = F (x, u, Du) dx
Ω
is satisfied by all smooth functions u : Ω → RN ; that is, for all u ∈ C 2 (Ω̄; RN ), the equation
n
X ∂
(5.26) F k (x, u(x), Du(x)) = Fsk (x, u(x), Du(x)), ∀ x ∈ Ω̄,
∂xi ξi
i=1
holds for all k = 1, 2, · · · , N.
Example 5.23. Let f : Rn → R be C 1 , and let F (s, ξ) = f (s) det ξ for s ∈ Rn , ξ ∈ Mn×n .
Then, by (5.22),
Fsk (s, ξ) = fsk (s) det ξ, Fξk (s, ξ) = f (s) (adj ξ)ik .
i

Hence for any u ∈ C 2 (Ω̄; Rn ), by Lemma 5.20 above, we have

n n
X ∂ X ∂
Fξk (u(x), Du(x)) = [f (u(x)) (adj Du(x))ik ]
∂xi i ∂xi
i=1 i=1
X n
= fsj (u(x)) ujxi (adj Du(x))ik
i,j=1
X n
= fsj (u(x)) (det Du(x))δjk
j=1
= fsk (u(x)) (det Du(x))
for all k = 1, 2, · · · , n. Hence F (s, ξ) = f (s) det ξ is a null-Lagrangian.

We now prove the following boundary dependence property for null-Lagrangians.

R
Theorem 5.24. Let I(u) = Ω F (x, u, Du) dx and F be a null-Lagrangian. Assume u, v ∈
C 2 (Ω̄; RN ) satisfy u(x) = v(x) for all x ∈ ∂Ω. Then I(u) = I(v). Moreover, if F satisfies
|F (x, s, ξ)| ≤ c(|s|p + |ξ|p ) + C(x) with c > 0 and C ∈ L1 (Ω), then I(u) = I(v) for all
u, v ∈ W 1,p (Ω; RN ) with u − v ∈ W01,p (Ω; RN ).

Proof. Define h(t) = I(U (t, ·)) for t ∈ [0, 1], where U (t, x) = tu(x) + (1 − t)v(x). Then
Z  
h0 (t) = Fξk (x, U (t, x), DU (t, x))Di (uk − v k ) + Fsk (x, U (t, x), DU (t, x))(uk − v k ) dx.
i
Ω
Since u − v, U (t, ·) ∈ C 2 (Ω̄; RN ) and u − v = 0 on ∂Ω, using the divergence theorem and
(5.26) with u = U (t, ·) we have h0 (t) = 0 for all t ∈ (0, 1) and hence h(0) = h(1) and the first
part of the result follows. The second part follows by the continuity of I on W 1,p (Ω; RN )
under the given assumption. 

We now study the null-Lagrangians depending only on variable in MN ×n .

Theorem 5.25. Let F : MN ×n → R be continuous and let
Z
IΩ (u) = F (Du(x)) dx,
Ω
where Ω is any smooth bounded domain in Rn . Then the following conditions are equivalent:
(1) F is of C 2 and is a null-Lagrangian;
112 5. Weak Lower Semicontinuity on Sobolev Spaces

+ Dϕ(x)) dx for all ξ ∈ MN ×n and ϕ ∈ C0∞ (Σ; RN );

R
(2) F (ξ) = Σ F (ξ
(3) IΩ (u) = IΩ (u + φ) for all u ∈ C 1 (Ω̄; RN ) and φ ∈ C0∞ (Ω; RN );
(4) IΩ is continuous with respect to the Lipschitz convergence on W 1,∞ (Ω; RN );
(5) there exists a linear function L : RL → R such that F (ξ) = L(J (ξ)) for all ξ ∈
MN ×n .

Proof. We will not show the equivalence of (5) to other conditions, but it is important to
know that null-Lagrangians can only be the linear combination of subdeterminants. We
prove other equivalent conditions.
It is easy to see that (3) implies (2). Note that (2) is equivalent to that both F and −F
are quasiconvex; the latter is equivalent to that both IΩ and −IΩ are lower semicontinuous
with respect to the Lipschitz convergence on W 1,∞ (Ω; RN ). Therefore, (4) is equivalent to
(2). It remains to show (2) implies (1) implies (3). Let us first show that (1) implies (3).
To this end, given u ∈ C 2 (Σ̄; RN ) and φ ∈ C0∞ (Σ; RN ), let
Z
f (t) = F (Du(x) + tDφ(x)) dx.
Σ
Then, by (1), Z
f 0 (t) = Fξαi (Du(x) + tDφ(x)) Dα φi (x) dx = 0.
Σ
Hence f is a constant function; hence f (0) = f (1), which proves (3). The proof of that (2)
implies (1) will follow from several lemmas proved below. 
Lemma 5.26. If F is of C 2 , then (2) implies (1).

Proof. Note that (2) implies

Z
F (ξ + Dφ(x)) dx = F (ξ) |Ω| ∀ φ ∈ C0∞ (Ω; RN ).
Ω

Since F is of C 2 , this implies that, for any φ, ψ ∈ C0∞ (Ω; RN ), the function
Z
f (t) = F (t Dφ(x) + Dψ(x)) dx
Ω

is constant and of C 2. Therefore, f 0 (0)

= 0, which gives
Z
Fξk (Dψ(x)) Di φk (x) dx = 0.
i
Ω

Now given u ∈ C 2 (Ω̄; RN ), we can select a sequence {ψν } in C0∞ (Ω; RN ) such that ψν → u
in C 2 (supp φ; RN ). Using the identity above with ψ = ψν and letting ν → ∞ yield
Z
Fξk (Du(x)) Di φk (x) dx = 0,
i
Ω
which, by the divergence theorem, shows the Euler-Lagrange equation for I is satisfied by
u ∈ C 2 (Ω̄; RN ), and hence F is a null-lagrangian. 

We say F is rank-one affine if F (ξ + tq ⊗ p) is affine in t for all ξ ∈ MN ×n , q ∈ RN , p ∈

Rn .
Lemma 5.27. If F satisfies (2) then F is rank-one affine.
5.4. Polyconvex Functions and Null-Lagrangians 113

Proof. Since (2) is equivalent to that F and −F are rank-one convex, which is equivalent
to that F is rank-one affine. 

We need some notation. Let µiα = ei ⊗ eα , where {ei } and {eα } are the standard bases
of RN and Rn , respectively. For each 1 ≤ k ≤ σ = min{n, N } and 1 ≤ i1 , · · · , ik ≤ N, 1 ≤
α1 , · · · , αk ≤ n, we define, inductively,
Fαi11 (ξ) = F (ξ + µiα11 ) − F (ξ),
i ···i i ···i
Fαi11···i 1 k−1 ik 1 k−1
···αk (ξ) = Fα1 ···αk−1 (ξ + µαk ) − Fα1 ···αk−1 (ξ).
k

Note that if F is a polynomial it follows

Fαi11···i k i1 ik
···αk (ξ) = ∂ F (ξ)/∂ξα1 · · · ∂ξαk .
k

Indeed, we have the same permutation invariance property.

Lemma 5.28. Let {10 , · · · , k 0 } be any permutation of {1, · · · , k}. Then
i 0 ···i
Fαi11···i 1 k 0
···αk (ξ) = Fα10 ···αk0 (ξ).
k

Proof. We use induction on k. Assume this is true for all k ≤ s − 1. Let {10 , 20 , · · · , s0 } be
a permutation of {1, 2, · · · , s}. We need to show
i 0 ···i
Fαi11···i s 1 s 0
(5.27) ···αs (ξ) = Fα10 ···αs0 (ξ).

By definition
i ···i i ···i
Fαi11···i 1 s−1 is 1 s−1
···αs (ξ) = Fα1 ···αs−1 (ξ + µαs ) − Fα1 ···αs−1 (ξ).
s

By induction assumption, (5.27) holds if s0 = s. We thus assume s0 < s. In this case, by

induction assumption,
i ···î 0 ···is0 i ···î 0 ···is0
Fαi11···i s 1 s
···αs (ξ) = Fα1 ···α̂ (ξ + µiαss ) − Fα11 ···α̂s (ξ)
s0 ···αs0 s0 ···αs0

(where the m̂ means omitting m)

i ···î 0 ··· i i ···î 0 ···
+ µiαss + µαss0 ) − Fα11 ···α̂s + µiαss )
0
= Fα11 ···α̂s (ξ (ξ
s0 ··· s0 ···

i ···î 0 ··· i 0 i ···î 0 ···

−Fα11 ···α̂s (ξ + µαss0 ) + Fα11 ···α̂s (ξ)
s0 ··· s0 ···

i ···î 0 ···is i 0 i ···î 0 ···is

= Fα11 ···α̂s (ξ + µαss0 ) − Fα11 ···α̂s (ξ),
s0 ···αs s0 ···αs

which, by induction assumption, equals

i 0 ···i 0 i 0 i 0 ···i 0 i 0 ···i 0
Fα110 ···α(s−1)
(s−1)0
(ξ + µαss0 ) − Fα110 ···α(s−1)
(s−1)0
(ξ) = Fα110 ···αss0 (ξ).
This proves the induction procedure and hence the lemma. 

Lemma 5.29. If F is rank-one affine, then all Fαi11···i k

···αk are also rank-one affine. Moreover,
j i1 ···ik j
(5.28) Fαi11···i i1 ···ik
···αk (ξ + t µβ ) = Fα1 ···αk (ξ) + t Fα1 ···αk β (ξ).
k

Proof. Use induction again on k. 

Lemma 5.30. If F satisfies (2) then F (ξ) is a polynomial in ξ.
114 5. Weak Lower Semicontinuity on Sobolev Spaces

Proof. If F satisfies (2) then F is rank-one affine. Write

N X
X n
ξ= ξαi µiα .
i=1 α=1
Then a successive use of the previous lemma shows that F (ξ) is a polynomial of degree at
most nN in ξ with coefficients determined by Fαi11···i k
···αk (0). 

Remark. The function F (ξ) = (tr ξ)2 − tr ξ 2 is a null-Lagrangian on Mn×n . To see this, note
n n n n
that F (ξ) = ( ξii )( ξkk ) − ξik ξki = Pik (ξ), where Pik (ξ) = ξii ξkk − ξki ξik . For each
P P P P
i=1 k=1 i,k=1 i,k=1
pair (i, k), i 6= k, Pik (ξ) is a 2 × 2 subdeterminant of ξ. Hence F (ξ) is a null-Lagrangian.

5.4.5. Compensated Compactness of Null-Lagrangians. We prove a compensated

compactness property of the null-Lagrangians. For nonlinear operators with similar prop-
erties we refer to Coifman et al [12].
Theorem 5.31. Let Jk (Du) be any k × k subdeterminant. Let {uj } be any sequence weakly
convergent to ū in W 1,k (Ω; RN ) as j → ∞. Then Jk (Duj ) → Jk (Dū) in the sense of
distribution in Ω; that is,
Z Z
lim Jk (Duj (x)) φ(x) dx = Jk (Dū(x)) φ(x) dx
j→∞ Ω Ω
for all φ ∈ C0∞ (Ω).

Proof. We prove this by induction on k. Obviously, the theorem is true when k = 1.

Assume it holds for Js with s ≤ k − 1. We need to show it also holds for s = k. Without
loss of generality, we may assume
∂(u1 , u2 , · · · , uk )
Jk (Du(x)) = .
∂(x1 , x2 , · · · , xk )
For any smooth function u, we observe that Jk (Du) is actually a divergence:
k  ν+1 ∂(u2 , · · · , uk )

X ∂ 1 (−1)
(5.29) Jk (Du(x)) = u ,
∂xν ∂(x1 , · · · , xˆν , · · · , xk )
ν=1
where xˆν , again, means deleting xν . Let
(ν) (−1)ν+1 ∂(u2 , · · · , uk )
Jk−1 (Du(x)) = .
∂(x1 , · · · , xˆν , · · · , xk )
Then (5.29) implies
Z k Z
(ν)
X
(5.30) Jk (Du(x)) φ(x) dx = u1 (x) Jk−1 (Du(x)) Dν φ(x) dx.
Ω ν=1 Ω

By density argument, this identity still holds if u ∈ W 1,k (Ω; RN ). Suppose uj * ū in

W 1,k (Ω; RN ). By the Sobolev embedding theorem, uj → ū in Lk (Ω; RN ). Moreover, by the
(ν) (ν)
induction assumption, Jk−1 (Duj ) → Jk−1 (Dū) in the sense of distribution. Note that since
(ν) k k
sequence {Jk−1 (Duj )} is also bounded in L k−1 (Ω) it also weakly converges in L k−1 (Ω); by
(ν)
density the weak limit must be Jk−1 (Dū). We can then apply (5.30) to conclude
Z Z
lim Jk (Duj (x)) φ(x) dx = Jk (Dū(x)) φ(x) dx,
j→∞ Ω Ω
5.4. Polyconvex Functions and Null-Lagrangians 115

as desired. The proof is complete. 

Corollary 5.32. Let Jk (Du) be any k × k subdeterminant and p > k be a number. Let {uj }
be any sequence weakly convergent to ū in W 1,p (Ω; RN ) as j → ∞. Then Jk (Duj ) * Jk (Dū)
p
weakly in L k (Ω).
p
Proof. Note that fj = Jk (Duj ) is bounded in L k (Ω) and hence, via a subsequence, fj * f¯
p
for some f¯ weakly in L k (Ω), which also implies fj → f¯ in distribution. Hence f¯ ≡ Jk (Du)
p
and the whole sequence fj * Jk (Du) in L k (Ω). 

Remark. The following example shows that the weak convergence uj * ū in W 1,k (Ω; RN )
does not imply the weak convergence Jk (Duj ) * Jk (Dū) in L1 (Ω).

Example 5.33. (Ball and Murat [7]). Let B be the unit open ball in Rn . Consider, for
j = 1, 2, · · · , the radial mappings

 jr if 0 ≤ r ≤ 1/j,
Uj (|x|)
uj (x) = x, Uj (r) = 2 − jr if 1/j ≤ r ≤ 2/j,
|x|
0 if 2/j ≤ r < 1.


Computation shows that uj * 0 in W 1,n (B; Rn ) as j → ∞. But

det Duj (x) = (Uj (r)/r)n−1 Uj0 (r)
for a.e. x ∈ B, where r = |x|, and hence
Z
| det Duj (x)| dx = C
|x|<2/j

is a constant independent of j. This shows that {det Duj } is not equi-integrable in B, and
therefore it does not converges weakly in L1 (B), or in L1loc (B). Therefore, although uj * ū
in W 1,k (Ω; RN ), Jk (Duj ) → Jk (Dū) in the sense of distribution, and {Jk (Duj )} is bounded
in L1 (Ω), but it is not true that Jk (Duj ) * Jk (Dū) weakly in L1 (Ω).
However, if we have det Duj (x) ≥ 0 a.e. in Ω and uj → ū weakly in W 1,n (Ω; Rn ) then
det Duj * det Dū weakly in L1loc (Ω). This is the well-known theorem of S. Müller [29]. We
will not prove this result here, but will present an easier result which can be proved without
using Müller’s theorem and Theorem 5.31.
Theorem 5.34. Let uj ∈ Dϕ , the Dirichlet class of ϕ ∈ W 1,n (Ω; Rn ). Assume uj * ū in
W 1,n (Ω; Rn ) and det Duj (x) ≥ 0 a.e. in Ω. Then det Duj * det Dū weakly in L1 (Ω).

Proof. Note that, since det Duj is bounded in L1 (Ω), the weak convergence is equivalent
to
Z Z
(5.31) lim det Duj (x) dx = det Dū(x) dx
j→∞ E E
for all measurable sets E ⊆ Ω. By Theorem 5.24, we have
Z Z Z
(5.32) det Duj (x) dx = det Dū(x) dx = det Dϕ(x) dx.
Ω Ω Ω
Hence (5.31) holds when E = Ω. Let E ⊂ Ω. Let
F (x, ξ) = χE (x) max{0, det ξ},
116 5. Weak Lower Semicontinuity on Sobolev Spaces

where χE is the characteristic function of E. Then F (x, ξ) is polyconvex and hence

quasiconvex in ξ. Note that
0 ≤ F (x, ξ) ≤ C |ξ|n .
Hence by Theorem 5.7 the functional
Z
I(u) = F (x, Du)dx
Ω
is w.l.s.c. on W 1,n (Ω; Rn ). Using the fact
F (x, Duj (x)) = χE (x) det Duj (x), F (x, Dū(x)) ≥ χE (x) det Dū(x) ∀ a.e. x ∈ Ω,
this w.l.s.c. implies
Z Z
(5.33) det Dū(x) dx ≤ lim inf det Duj (x) dx.
E j→∞ E
This is also valid for E c = Ω \ E; that is,
Z Z
(5.34) det Dū(x) dx ≤ lim inf det Duj (x) dx.
Ec j→∞ Ec
Note that Z Z Z
det Duj dx = det Duj dx − det Duj dx.
Ec Ω E
Hence (5.32) and (5.34)imply
Z Z
det Dū dx ≥ lim sup det Duj dx,
E j→∞ E

which, combined with (5.33), implies (5.31). 

5.5. Existence in Nonlinear Elasticity

5.5.1. Hyperelastic Materials. In nonlinear elasticity, for a hyperelastic material,
it is assumed that a stored energy density function F (x, u, Du) exists such that the
continuum mechanical quantities can be given by this function. In particular, the total
stored energy is defined by
Z
I(u) = F (x, u(x), Du(x)) dx,
Ω
where Ω ⊂ R3is a bounded smooth domain occupied by the hyperelastic material in the
reference configuration, and u : Ω → R3 is the deformation (u(x) representing displacement
of material point x ∈ Ω in the deformed configuration).
The stored energy density function F (x, s, ξ) should satisfy several assumptions due to
continuum mechanics principles and material properties; some of these assumptions exclude
the possibility of F being convex in ξ and almost all practical problems require that F (x, s, ξ)
satisfy the following:

(5.35) F (x, s, ξ) = ∞ if det ξ ≤ 0, lim F (x, s, ξ) = ∞.

det ξ→0+

The first is due to the orientation-preserving assumption and the second is due to the
fact that infinite energy is needed to press a solid material into zero-volume.
Therefore, for nonlinear elasticity problems, we often need to consider extended-
valued energy functionals.
5.5. Existence in Nonlinear Elasticity 117

5.5.2. Polyconvex Energy Density Functions. We assume the energy density function
F (x, s, ξ) satisfies the following assumptions:
(1) F (x, s, ξ) = +∞ if and only if det ξ ≤ 0;
(2) F (x, s, ξ) is continuous in (x, s) ∈ Ω̄×R3 and polyconvex convex in ξ ∈ M3×3 , det ξ >
0; that is, there exists a continuous function W (x, s, ξ, η, τ ) which is convex in
(ξ, η, τ ) ∈ M3×3 × M3×3 × R+ such that, for all x, s,
F (x, s, ξ) = W (x, s, ξ, adj ξ, det ξ) for all ξ ∈ M3×3 with det ξ > 0;
(3) F (x, s, ξ) ≥ c0 |ξ|3 − C(x) for all ξ ∈ M3×3 , where c0 > 0 and C ∈ L1 (Ω) are given.
Note that these assumptions are compatible with the assumption (5.35) above.

5.5.3. Existence Theorem. Let I(u) be the energy functional defined by

Z
I(u) = F (x, u(x), Du(x)) dx,
Ω

where F satisfies the previous assumptions. Let ϕ ∈ W 1,3 (Ω; R3 ) be a given function such
that I(ϕ) < ∞. Then we have the following existence result.

Theorem 5.35. (Existence in Nonlinear Elasticity) There exists at least one mini-
mizer ū ∈ Dϕ with det Dū(x) > 0 for almost every x ∈ Ω such that
I(ū) = min I(u).
u∈Dϕ

Proof. Let {uν } be a minimizing sequence in Dϕ . Since lim I(uν ) = inf Dϕ I ≤ I(ϕ) < ∞,
from Assumption (1), it follows that
det Duν (x) > 0 a.e.
Now Assumption (3) implies that {|Duν |} is bounded in L3 (Ω) and hence from the fixed
Dirichlet boundary condition, {uν } is bounded in W 1,3 (Ω; R3 ). Therefore, via a subsequence,
we assume uν * ū in W 1,3 (Ω; R3 ) as ν → ∞. Hence, by Corollary 5.32 and Theorem 5.34,
(5.36) adj Duν * adj Dū in L3/2 (Ω), det Duν * det Dū in L1 (Ω).
It is easy to see that det Dū(x) ≥ 0 for almost every x ∈ Ω. We now show det Dū(x) > 0
a.e. in Ω. (We follow Pedregal.) Let
h(τ ) = inf{F (x, s, ξ) | (x, s, ξ) ∈ Ω̄ × R3 × M3×3 , det ξ = τ }.
Then h(τ ) is continuous and h(τ ) = ∞ for τ ≤ 0. Note that
Z Z
sup h(det Duν (x)) dx ≤ F (x, uν (x), Duν (x)) dx < ∞.
ν Ω Ω

Let E = {x ∈ Ω | det Dū(x) = 0}. Define

w(x) = lim inf det Duν (x), x ∈ E.
ν→∞

By Fatou’s lemma,
Z Z Z
0≤ w(x)dx ≤ lim inf det Duν (x)dx = det Dū(x)dx = 0.
E\{w=0} ν→∞ E\{w=0} E\{w=0}
118 5. Weak Lower Semicontinuity on Sobolev Spaces

Hence w(x) = 0 a.e. in E. Using Fatou’s lemma again,

Z Z
h(0) = lim inf h(det Duν (x)) dx
E E ν→∞
Z
≤ lim inf h(det Duν (x))dx < ∞.
ν→∞ E
Since h(0) = ∞, we must have |E| = 0. This proves that det Dū(x) > 0 a.e. in Ω. Note
that at every point x where det Duν (x) > and det Dū(x) > 0, by the polyconvex convexity
assumption (2), there exists a function W (x, s, J) which is, for all given (x, s), convex on
J = (ξ, η, τ ), τ > 0 such that F (x, s, ξ) = W (x, s, J(ξ)) for all x, s, ξ with det ξ > 0, where
J(ξ) = (ξ, adj ξ, det ξ). Hence
F (x, uν , Duν ) − F (x, ū, Dū)
= [W (x, uν , J(Dū)) − W (x, ū, J(Dū))] + [W (x, uν , J(Duν )) − W (x, uν , J(Dū))]
≥ [W (x, uν , J(Dū)) − W (x, ū, J(Dū))] + WJ (x, uν , J(Dū)) · (J(Duν ) − J(Dū)).
We can then follow the similar steps as in the proof of the Tonelli theorem (Theorem 5.1)
to show that
I(ū) ≤ lim inf I(uν ).
ν→∞
This proves that ū is a minimizer. 

Remark. For more general cases, see Müller, Tang & Yan [31].

5.6. Relaxation Principles

There are many application problems which do not satisfy the convexity conditions. In
such cases, minimizers may not be found as we did before using the direct method, but the
variational methods may still help to study the problem.

5.6.1. Non-quasiconvex Problems. In general, we consider the multiple integral func-

tional Z
I(u) = F (x, u(x), Du(x)) dx,
Ω
where Ω is a bounded domain in Rn , u : Ω → RN , and F (x, s, ξ) is a given function not
quasiconvex in ξ. We give two examples of one-dimensional problem.
Example 5.36. Consider the classical example due to Bolza: to minimize the functional
Z 1
I(u) = [((u0 )2 − 1)2 + u2 ]dx
0
among all Lipschitz functions u satisfying boundary conditions u(0) = u(1) = 0. Let
 
1  [kx] 1 
uk (x) = − x − −  ,
2k k 2k
where [a] stands for the integer part of a ∈ R. Then uk → 0 in the Lipschitz convergence,
and I(uk ) → 0 as k → ∞. Hence the infimum of I over all such u is 0. Note that I(0) = 1.
Hence
I(0) > lim inf I(uk ),
k→∞
and so I is not lower semicontinuous in the Lipschitz convergence. Also I does not have
minimizers since I(u) = 0 would imply |u0 | = 1 a.e and u ≡ 0.
5.6. Relaxation Principles 119

Example 5.37. Let f (ξ) = (ξ 2 − 1)2 for ξ ∈ R and consider the functional
Z 1 Z 1
I(u) = f (u0 (x)) dx = ((u0 )2 − 1)2 dx
0 0

and minimize it over all Lipschitz functions u satisfying boundary conditions u(0) = 0, u(1) =
a. We show that this nonconvex problem has minimizers for all values of a; moreover, the
minimizer is unique if |a| ≥ 1 and there exist infinitely many minimizers if |a| < 1.

Proof. Let first of all |a| < 1. Then it can be easily seen that there are infinitely many
Lipschitz functions u such that u0 (x) ∈ {−1, 1} with u(0) = 0 and u(1) = a; any of such
functions is a minimizer of I with minimum value zero. Now let |a| ≥ 1. Note that the
strict inequality
f (ξ) − [f (a) + f 0 (a)(ξ − a)] > 0
holds for all ξ 6= a if |a| > 1, or holds for all ξ ∈ / {−1, 1} if |a| = 1. Therefore, for all
Lipschitz functions u with u(0) = 0, u(1) = a it follows that
Z 1 Z 1
I(u) = f (u0 (x)) dx ≥ f (a) + f 0 (a)(u0 (x) − a) dx = f (a);
0 0

hence the minimum value of I for such u’s is f (a) and, certainly, ū(x) = ax is a minimizer.
We show this is the unique minimizer. Suppose v is any minimizer of I over all Lipschitz
functions u with u(0) = 0, u(1) = a; that is,
Z 1
I(v) = f (v 0 (x)) dx = f (a)
0

with v(0) = 0 and v(1) = a. If a = 1 then since f (1) = 0 it must follow that v 0 (x) ∈ {−1, 1}
R1
a.e. x ∈ (0, 1) and 0 (1−v 0 (x)) dx = 0 and hence v 0 (x) = 1 on (0, 1). This implies v(x) = x.
Similarly if a = −1 then v(x) = −x. Now assume |a| > 1. Since f (ξ)−f (a)−f 0 (a)(ξ −a) > 0
for all ξ 6= a, from
Z 1
[f (v 0 (x)) − f (a) − f 0 (a)(v 0 (x) − a)] dx = 0,
0

it follows that v 0 (x) = a for a.e. x ∈ (0, 1). Hence v(x) = ax. 

For this example, we see that the function

Z 1 (
0 c f (ξ) if |ξ| ≥ 1
g(ξ) = inf f (ξ + w (x)) dx = f (ξ) =
w∈W01,∞ (0,1) 0 0 if |ξ| < 1
is the convexification of function f ; that is, the largest convex function less than or equal
to f .

5.6.2. Quasiconvexification. Given a function F : MN ×n → R, we call the largest quasi-

convex function less than or equal to F the quasiconvexication or quasiconvex envelope
of F. We denote the quasiconvexification of F by F qc .
Theorem 5.38. Let F ≥ 0 be continuous. Then
Z
(5.37) qc
F (ξ) = inf F (ξ + Dφ(x)) dx, ξ ∈ MN ×n .
φ∈W01,∞ (Σ;RN ) Σ
120 5. Weak Lower Semicontinuity on Sobolev Spaces

Proof. For any quasiconvex function G ≤ F,

Z Z
G(ξ) ≤ G(ξ + Dφ(x)) dx ≤ F (ξ + Dφ(x)) dx.
Σ Σ
Hence G(ξ) ≤ F qc (ξ) by the definition of F qc . It remains to prove that F qc is itself quasi-
convex. We first observe that
Z
qc
(5.38) F (ξ) = inf − F (ξ + Dφ(x)) dx
φ∈W01,∞ (Ω;RN ) Ω

for any open set Ω ⊂ Rn with |∂Ω| = 0. This can be proved by using the Vitali covering ar-
gument as before. We now claim that, for all piecewise affine Lipschitz continuous functions
φ ∈ W01,∞ (Σ; RN ), it follows that
Z
qc
(5.39) F (ξ) ≤ F qc (ξ + Dφ(x)) dx.
Σ
To see this, let Σ = ∪i Ωi ∪ E be such that |E| = 0 and each Ωi is an open set with |∂Ωi | = 0
and such that φ ∈ W01,∞ (Σ; RN ) takes constant gradients Dφ = Mi on each Ωi . Let  > 0
be given. By the above remark on the definition of F qc , there exists ψi ∈ W01,∞ (Ωi ; RN )
such that Z
F qc (ξ + Mi ) ≥ − F (ξ + Mi + Dψi (x)) dx − .
Ωi
With each ψi being extended to Σ, we set ψ = φ + i ψi . Then ψ ∈ W01,∞ (Σ; RN ) and we
P
have Z X
F qc (ξ + Dφ(x)) dx = |Ωi |F qc (ξ + Mi )
Σ i
Z
≥ F (ξ + Dψ(x)) dx −  ≥ F qc (ξ) − ,
Σ
and the (5.39) follows. From this we can show F qc is rank-one convex (using “sawtooth”
like piecewise affine functions) and therefore is locally Lipschitz continuous (since it is
convex in each coordinate direction). Since piece-wise affine Lipschitz functions are dense
in W01,p (Σ; RN ), (5.39) holds for all φ ∈ W01,p (Σ; RN ), and hence F qc is quasiconvex. The
proof of the theorem is complete. 

5.6.3. Weak Lower Semicontinuous Envelopes. Suppose F (x, s, ξ) is a Caratheodory

function and satisfies
0 ≤ F (x, s, ξ) ≤ C (|ξ|p + |s|p + 1) ∀x ∈ Ω, s ∈ RN , ξ ∈ MN ×n
˜ on W 1,p (Ω; RN )
for some 1 ≤ p < ∞. We are interested in the largest w.l.s.c. functional I(u)
which is less than or equal to the functional
Z
I(u) = F (x, u(x), Du(x)) dx.
Ω
˜
This functional I(u) is called the (weak lower semicontinuous) envelope or relaxation
of I in the weak topology of W 1,p (Ω; RN ). It turns out under the condition given above,
˜
I(u) is given by the integral functional of the quasiconvexification of F with respect to ξ.
˜
Theorem 5.39. Under the given condition, the envelope I(u) of I in the weak topology of
1,p N
W (Ω; R ) is given by
Z
˜
I(u) = F qc (x, u(x), Du(x)) dx,
Ω
5.6. Relaxation Principles 121

where F qc (x, s, ·) is the quasiconvexification of F (x, s, ·) for given (x, s).

ˆ
Proof. Let I(u) be the integral of F qc (x, u(x), Du(x)) over Ω. Since F qc is quasiconvex and
ˆ
satisfies the same growth condition as F , the functional I(u) is thus (sequential) w.l.s.c. on
W (Ω; R ) by the theorem of Acerbi and Fusco. Therefore, Iˆ ≤ I.
1,p N ˜ To prove the other
direction, we first assume there exists a Caratheodory function g(x, s, ξ) such that
Z
(5.40) ˜
I(u) = g(x, u(x), Du(x)) dx ∀ u ∈ W 1,p (Ω; RN ).
Ω

Then g(x, s, ·) must be quasiconvex and g ≤ F , and thus g ≤ F qc ; this proves I˜ ≤ I. ˆ

However, the proof of integral representation (5.40) is beyond the scope of this lecture and
is omitted; see e.g., Acerbi & Fusco [1], Buttazzo [9], or Dacorogna [13]. 

5.6.4. Relaxation Principle. We have the following theorem; the proof of the theorem
is also omitted; see, e.g., Acerbi & Fusco [1], Buttazzo [9], or Dacorogna [13].
Theorem 5.40 (Relaxation Principle). Assume
c |ξ|p ≤ F (x, s, ξ) ≤ C (|ξ|p + |s|p + 1)
holds for some constants c > 0, C > 0, p > 1. Then
Z Z
inf F (x, u, Du) dx = min F qc (x, u, Du) dx
u∈Dϕ Ω u∈Dϕ Ω

for any ϕ ∈ W 1,p (Ω; RN ), where Dϕ is the Dirichlet class of ϕ.

Remark. The passage from F to F qc is called relaxation. The relaxation principle above
seems too nice for the nonconvex variational problems since it replaces a nonconvex prob-
lem which may not have any solution by a quasiconvex problem that has solutions. But,
there are costs for this: we may lose some information about the minimizing sequences; a
minimizer so obtained for the relaxed problem may not characterize what seems more in-
teresting in applications the finer and finer patterns of minimizing sequences. In the phase
transition problems for certain materials, it accounts for the loss of information about the
microstructures by a macroscopic effective processing (relaxation). For more information,
we refer to Ball & James [6], Müller [30] and the references therein.

5.6.5. Existence of Minimizers for a Nonconvex Problem. Consider

Z
I(u) = f (∇u(x)) dx
Ω
where Ω is a bounded domain in Rn , u : Ω → R is a scalar function, and f : Rn → R is a
given continuous function.
We define the subgradient of f at a point ξ by
∂f (ξ) = {l ∈ Rn | f (η) ≥ f (ξ) + l · (η − ξ), ∀ η ∈ Rn }.
Each l ∈ ∂f (ξ) is called a subdifferential of f at ξ. Note that ∂f (ξ) may be an empty set
for some ξ.
For a set of points ξ1 , ξ2 , · · · , ξq in Rn , we denote by co{ξ1 , · · · , ξq } the convex hull
defined by ( q q
)
X  X
conv{ξ1 , · · · , ξq } = λi ξi  ∀ λi ≥ 0, λi = 1 .

i=1 i=1
122 5. Weak Lower Semicontinuity on Sobolev Spaces

The following theorem is due to A. Cellina [10, 11]. We do not give a proof of this
theorem, but simply remark that even without convexity assumption the minimization may
still have solutions. However, the existence can not follow from the direct method, but has
to rely on different methods.
Theorem 5.41. Given ξ ∈ Rn , the minimization problem
Z
inf f (∇u)dx
u∈W 1,1 (Ω), u|∂Ω=ξx Ω
has a minimizer if and only if either ∂f (ξ) 6= ∅, or there exist ξ1 , ξ2 , · · · , ξq such that
ξ ∈ int conv{ξ1 , · · · , ξq } and ∩qi=1 ∂f (ξi ) 6= ∅.

Remark. By the relaxation principle, any minimizer u of the above problem must be a
minimizer for the relaxed problem and satisfy
(5.41) ∇u(x) ∈ K = {ξ ∈ Rn | f (ξ) = f # (ξ)}, a.e. x ∈ Ω,
where f#
is the convexification of f . The proof of Cellina’s theorem relies on constructing
minimizers u of the relaxed problem satisfying the first-order partial differential relation
(5.41).
Bibliography

[1] E. Acerbi and N. Fusco, Semicontinuity problems in the calculus of variations, Arch. Rational Mech.
Anal., 86 (1984), 125–145.
[2] R. A. Adams and J. Fournier, “Sobolev Spaces,” Second Edition. Academic Press, New York, 2003.
[3] A. Bruckner, J. Bruckner and B. Thomson, “Real Analysis,” Prentice-Hall, 1997.
[4] J. M. Ball, Convexity conditions and existence theorems in nonlinear elasticity, Arch. Rational Mech.
Anal., 63 (1977), 337–403.
[5] J. M. Ball, J. C. Currie and P. J. Olver, Null Lagrangians, weak continuity, and variational problems
of arbitrary order, J. Funct. Anal., 41 (1981), 135–174.
[6] J. M. Ball and R. D. James, Proposed experimental tests of a theory of fine microstructures and the
two well problem, Phil. Trans. Roy. Soc. London, A 338 (1992), 389–450.
[7] J. M. Ball and F. Murat, W 1,p -Quasiconvexity and variational problems for multiple integrals, J.
Funct. Anal., 58, (1984), 225–253.
[8] V. Barbu and Th. Precupanu, “Convexity and Optimization in Banach Spaces,” Reidel, Boston,
1986.
[9] G. Buttazzo, “Semicontinuity, Relaxation and Integral Representation in the Calculus of Variations,”
Pitman Res. Notes Math. Ser. 207, Harlow, 1989.
[10] A. Cellina, On minima of a functional of the gradient: necessary conditions, Nonlinear Anal., 20
(1993), 337–341.
[11] A. Cellina, On minima of a functional of the gradient: sufficient conditions, Nonlinear Anal., 20
(1993), 343–347.
[12] R. Coifman, P.-L. Lions, Y. Meyer and S. Semmes, Compensated compactness and Hardy spaces, J.
Math. Pures Appl., 72(9) (1993), 247–286.
[13] B. Dacorogna, “Direct Methods in the Calculus of Variations,” Springer, New York, 1989.
[14] D. Dunninger, Lecture Notes on Applied Analysis, (unpublished) Michigan State University
[15] L. C. Evans, Quasiconvexity and partial regularity in the calculus of variations, Arch. Rational Mech.
Anal., 95 (1986), 227–252.
[16] L. C. Evans, “Partial Differential Equations,” AMS-GSM, Vol. 19, 1998.
[17] M. Giaquinta, “Multiple Integrals in the Calculus of Variations and Nonlinear Elliptic Systems,”
Princeton University Press, 1983.
[18] M. Giaquinta, “Introduction to Regularity Theory for Nonlinear Elliptic Systems,” Birkhäuser, 1993.
[19] M. Giaquinta and E. Giusti, On the regularity of the minima of certain variational integrals, Acta
Math., 148 (1982), 31–46.
[20] B. Gidas, W.-M. Ni and L. Nirenberg, Symmetry and related properties via the maximum principle,
Comm. Math. Physics 68 (1979), 209–243.

123
124 Bibliography

[21] D. Gilbarg and N. S. Trudinger, “Elliptic Partial Differential Equations of Second Order,” 2nd ed.,
Springer-Verlag, Berlin, 1984.
[22] R. Hart, D. Kinderlehrer and F.H. Lin, Existence and partial regularity of static liquid crystal con-
figurations, Commun. Math. Phys., 105 (1986), 547–570.
[23] R. D. James and D. Kinderlehrer, Frustration in ferromagnetic materials, Cont. Mech. Thermodyn.,
2 (1990), 215–239.
[24] F.-H. Lin, On nematic liquid crystals with variable degree of orientation, Comm. Pure Appl. Math.,
44(3) (1991), 453–486.
[25] N. Meyers, Quasi-convexity and lower semicontinuity of multiple variational integrals of any order,
Trans. Amer. Math. Soc., 119 (1965), 125–149.
[26] V. P. Mikhailov, “Partial Differential Equations,” Mir, 1978.
[27] C. B. Morrey, Jr., Quasiconvexity and the lower semicontinuity of multiple integrals, Pacific J. Math.,
2 (1952), 25–53.
[28] C. B. Morrey, Jr., “Multiple Integrals in the Calculus of Variations,” Springer, New York, 1966.
[29] S. Müller, Higher integrability of determinants and weak convergence in L1 , J. Reine Angew. Math.,
412 (1990), 20–34.
[30] S. Müller, Variational models for microstructure and phase transitions, in “Calculus of Variations
and Geometric Evolution Problems,” (Cetraro, 1996), pp. 85–210, Lecture Notes in Math. 1713,
Springer, 1999.
[31] S. Müller, Q. Tang and B. Yan, On a new class of elastic deformations not allowing for cavitation,
Ann. Inst. Henri Poincaré, Analyse non linéaire, 11(2) (1994), 217–243.
[32] L. Nirenberg, “Functional Analysis.” Courant Institute Lecture Notes, New York University, 1960.
[33] P. Pedregal, “Parametrized Measure and Variational Principles.” Birkhauser, Basel, 1997.
[34] P. Pedregal and B. Yan, On two-dimensional ferromagnetism, Preprint.
[35] E. Stein, “Harmonic Analysis.” Princeton University Press, Princeton, 1993.
[36] V. Šverák, Rank-one convexity does not imply quasiconvexity, Proc. Roy. Soc. Edinburgh Sect. A,
120 (1992), 185–189.
[37] L. Tartar, The compensated compactness method applied to systems of conservation laws, in “Systems
of Nonlinear Partial Differential Equations”, NATO ASI Series, Vol. 103, D. Reidel, Dordrecht, 1983.
[38] R. Temam, “Navier-Stokes Equations: Theory and Numerical Analysis.” Revised Edition. Elsevier
Science Publishers, Amsterdam, 1984.
[39] K. Yosida, “Functional Analysis,” 6th edition, Springer, New York, 1980.
[40] L. C. Young, “Lectures on the Calculus of Variations and Optimal Control Theory,” Saunders, 1969
(reprint by Chelsea, 1980).
[41] E. Zeidler, “Nonlinear Functional Analysis and its Applications,” Vol.1-4, Springer.
Index

p-Laplace equations, 71 eigenvalue, 62

Eigenvalue Theorem Theorem, 62
Acerbi-Fusco’s Theorem Theorem, 101 eigenvalues, 10
adjugate matrix, 108 eigenvectors, 10
Ambrosetti-Rabinowitz Theorem, 87 equicontinuous, 8
Ascoli Theorem, 8 Estimates for W 1,p , n < p ≤ ∞ Theorem, 44
Euler-Lagrange equations, 66
Banach space Existence in Nonlinear Elasticity Theorem, 117
reflexive, 6 Existence of Minimizers Theorem, 70
separable, 2 extended-valued, 116
Banach-Steinhaus Theorem, 5 extreme, 19
Bessel’s inequality, 9 maximal, 20
biharmonic, 57 minimal, 20
bilinear form, 54
bounded, 54 ferromagnetism, 76
strongly positive, 54 first variation, 65, 66
Bounded Inverse Theorem, 6 fixed point, 11
Fourier series, 9
Caratheodory property, 13 Fourier transform, 48
Cartesian product, 3 Fourier transforms, 58
Cauchy sequence, 2 Frechet differentiable, 13
Cauchy-Schwarz inequality, 4 Fredholm Alternative Theorem, 9
characteristic function, 116 function
coercivity, 55 cut-off function, 27
continuation method, 11 mollified function, 27
Contraction Mapping Theorem, 11 p-mean continuous, 28
convex, 18, 67 support, 2
convex hull, 121 trace, 36, 37
convexification, 119, 122 functional
critical point, 18, 20 convex, 18
critical points, 65 weakly coercive, 17, 70
weakly continuous, 17
weakly lower semicontinuous, 17
Deformation Lemma, 87
determinant, 108
differential operator Gårding’s estimates, 55
uniformly elliptic, 51 Gagliardo-Nirenberg-Sobolev Inequality Lemma, 41
direct method, 70 Gateaux differentiable, 13
Dirichlet class, 69 Gateaux-differentiable, 68
dual space, 6 Green’s 1st identity, 39
second dual space, 6 Green’s 2nd identity, 39

Ehrling’s inequality, 21 Hahn-Banach Theorem, 6

eigenfunctions, 10, 62 harmonic map equation, 74
eigenspace, 62 harmonic maps, 75

125
126 Index

Hilbert space, 4 Plancherel’s Theorem Theorem, 48

Hodge decomposition Theorem, 81 Pohozaev Theorem, 90
Hodge decomposition theorem, 81 Poincaré’s inequality, 48
Hodge decompositions, 76 point spectrum, 10
hyperelastic material, 116 polyconvex, 108, 110
principal eigenvalue, 62
Implicit Function Theorem, 16 projection, 4
inner product, 4 Projection Theorem, 4
inverse Fourier tranform, 48 Property of Fourier Tranforms Theorem, 48

Lagrange multiplier, 20 quasiconvex, 99

Lagrange multipliers, 72 quasiconvex envelope, 119
Lagrange Theorem, 20 quasiconvex functions, 99
Lax-Milgram Theorem, 54 quasiconvexication, 119
Lax-Milgram Theorem Theorem, 54 quasiconvexity, 97
Legendre ellipticity condition, 53
Legendre-Hadamard, 106 rank-one convex, 67, 104
Legendre-Hadamard condition, 53, 67 regularity, 51
linear functionals relaxation, 121
bounded linear functionals, 6 Rellich-Kondrachov Theorem, 45
continuous linear functionals, 6 resolvent set, 10
Lipschitz convergence, 97 Riesz Representation Theorem, 6
liquid crystals, 74
Ljusternik Theorem, 21 second variation, 66
second-order differential operator in divergence form,
method of continuity, 11 51
minimization method, 65 semilinear, 71
minimizer, 70 seminorm, 47
minimizing sequence, 70 sequence
Morrey’s Inequality Theorem, 43 weakly convergent, 7
Serrin’s Maximum Principle Lemma, 92
Neumann boundary, 56 set
nonexistence, 73 convex, 1
nonlinear eigenvalue problem, 72 totally bounded, 45
nonlinear functional analysis, 11 weakly closed, 17, 70
norm, 2 simplified Maxwell equation, 76
equivalent, 47 Sobolev conjugate, 40
null-Lagrangian, 74, 111 Sobolev space of order k, 25
null-Lagrangians, 106 Sobolev-Rellich-Kondrachov Theorem, 39
spectrum, 10
operator star-shaped, 90
bounded, 5 stored energy density function, 116
compact, 8 structural conditions, 68
contraction, 11 subdifferential, 121
Hilbert space adjoint, 7 subgradient, 121
Laplace operator, 39
linear operator, 5 total stored energy, 116
Nemytskii operator, 12 Trace Theorem, 37, 38
operator norm, 5 Triangle inequality, 2, 4
resolvent, 10
strongly continuous, 8 variational problems, 65
weakly continuous, 8 vector space, 1
ordinary point, 20 Banach space, 2
orientation-preserving, 116 basis, 2
orthogonal, 4 complete, 2
orthogonal complement, 4 dimension, 1
orthogonality condition, 61 inner product space, 4
orthonormal, 9 normed space, 2
complete, 9 subspace, 1
orthonormal basis, 62 Vitali covering, 102
Oseen-Frank energy, 74
weak derivative, 23
Palais-Smale compactness condition, 86 weak solution, 51, 69
Parallelogram law, 4 weakly compact, 8
Partition of Unity Theorem, 28 weakly lower semicontinuous, 70, 95