Lecture Notes CoV 2021

Lecture notes on
Advanced Calculus of Variations

Marco Bonacini∗
Summer Semester 2021

University of Trento
These notes cover the main topics discussed in the course “Advanced Calculus of Varia-
tions”, given by the author at the University of Trento in the Summer Semester 2021, and
are only for the use of the students in the class.
These notes are mainly based on the books [2, 3, 4, 6] and further sources which are not
always mentioned specifically.
∗
E-mail: marco.bonacini@unitn.it
1
Contents
Introduction 4
1 Minimization of integral functionals among Lipschitz functions 7

1.1 Reminder: Lipschitz functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Existence of minimizers: preliminary results . . . . . . . . . . . . . . . . . . . 9
1.3 Maximum principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 The bounded slope condition and existence of minimizers . . . . . . . . . . . 14
2 Direct methods 18
2.1 An overview of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.1 Weak derivatives and Sobolev spaces . . . . . . . . . . . . . . . . . . . 18
2.1.2 Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.3 Boundary conditions and traces . . . . . . . . . . . . . . . . . . . . . . 23
2.1.4 Embedding theorems and Poincaré inequality . . . . . . . . . . . . . . 24
2.1.5 Duals and weak convergence . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 The direct method of the Calculus of Variations . . . . . . . . . . . . . . . . . 30
2.2.1 Lower semicontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.2 Compactness and existence of minimizers . . . . . . . . . . . . . . . . 32
2.2.3 The direct method in weak topologies . . . . . . . . . . . . . . . . . . 33
2.3 Minimum problems for integral functionals . . . . . . . . . . . . . . . . . . . 35
2.3.1 Functionals defined on Lp . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.2 Functionals defined on W1,p . . . . . . . . . . . . . . . . . . . . . . . 38
2.4 Necessity of convexity for weak lower semicontinuity . . . . . . . . . . . . . . 46
2.5 Integral functionals: the general case . . . . . . . . . . . . . . . . . . . . . . . 52
3 Relaxation 56
3.1 Relaxation in an abstract setting . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Convex envelope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3 Relaxation of integral functionals . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Characterization by sequences of the relaxation . . . . . . . . . . . . . . . . . 70
4 Euler-Lagrange equations and regularity 74

4.1 First variation and Euler-Lagrange equations . . . . . . . . . . . . . . . . . . 74
4.2 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3 Regularity of minimizers: Hilbert’s XIX problem and De Giorgi - Nash Theorem 82
4.3.1 An elliptic, quasilinear equation for the partial derivatives . . . . . . . 83
4.3.2 De Giorgi - Nash Theorem . . . . . . . . . . . . . . . . . . . . . . . . 86
2
5 Γ-convergence (a short intro) 88
5.1 Towards the definition of Γ-convergence . . . . . . . . . . . . . . . . . . . . . 88
5.2 One-dimensional examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3 Main properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4 An application to integral functionals . . . . . . . . . . . . . . . . . . . . . . 96
Bibliography 99
3
Introduction
The classical problem of the Calculus of Variations, which is the main topic that will be
discussed in this course, is the minimization of an integral functional, that is a map which is
typically of the form Z
F : u 7−→

f x, u(x), ∇u(x) dx (0.1)
Ω
defined on a suitable class of functions u : Ω ⊂ Rn → Rm . The goal is to solve the minimiza-
tion problem Z

min f x, u(x), ∇u(x) dx (0.2)
u∈U Ω
that is to find a minimizer ū ∈ U such that F (ū) ≤ F (v) for every v ∈ U. The space U
might include, for instance, suitable boundary conditions, for instance Dirichlet boundary
conditions: u = u0 on ∂Ω, where u0 is prescribed.
One of the motivations for the study of this problem comes form its physical meaning:
in Mechanics, a functional of the form (0.1) might represent, for instance, the energy of an
elastic body, and minimizers correspond to equilibrium configurations of the system.
As a toy example, along this introduction we can keep in mind the Dirichlet integral
Z
F (u) := |∇u|2 dx (0.3)
Ω
and its minimization over the set U := {u ∈ C 1 (Ω) : u = u0 on ∂Ω}.
The investigation of (0.2) has a long history which goes back to the classical theory
developed in ’700 initially for one-dimensional problems (Euler, Lagrange. . . ), that is for
functionals defined on functions u : R → Rm of one independent variable (possibly with m ≥
1). Several classical examples fall in this setting, for instance the brachistochrone problem,
the problem of geodesics, the problem of minimal surfaces of revolution. . .
The classical approach to solve the minimum problem (0.2) is based on the notion of
critical point and goes back to the idea which all students learn in the Analysis 1 course: to
find minimizers of a differentiable function g : [a, b] → R, one can look for critical points, that
is points where the first derivative of g vanishes. Indeed, in an interior minimum or maximum
point x0 we necessarily have g 0 (x0 ) = 0. (Notice that to follow this strategy we need an extra
regularity assumption of g, which is strictly speaking not necessary to formulate the minimum
problem). The problem is therefore reduced to the solution of an equation (g 0 (x0 ) = 0):
minimizers will be found among its possible solutions.
This approach can be extended to the case of an integral functional such has (0.1) and
leads to the definition of critical points of F as the solutions to a suitable (system of) ordinary
4
differential equation in the one-dimensional case n = 1, or partial differential equation for
n > 1, the so-called Euler-Lagrange equation associated to F .
Let us see how this strategy works explicitly in the example (0.3). In order to derive
the Euler-Lagrange equation satisfied by a minimizer, we mimic what we do in Analysis 1
for functions of one variable: we introduce a small increment of the independent variable (a
“variation”), and we take the limit of incremental quotients. Suppose that ū is a minimizer
of the Dirichlet integral (0.3), with respect to the boundary condition ū = u0 on ∂Ω. Fix
another function v ∈ C 1 (Ω) vanishing on the boundary, v = 0 on ∂Ω, and for t ∈ R consider
the one-parameter family of variations ut := ū + tv, and the corresponding values of the
functional Z
g(t) := F (ut ) = |∇(ū + tv)|2 dx, t ∈ R.
Ω
Notice that ut is an admissible competitor for all t, that is ut ∈ U satisfies the same boundary
conditions as ū. Therefore the function of one variable g(t) has a minimum point at t = 0:
g(0) ≤ g(t) for all t, and hence
g(t) − g(0)
0 = g 0 (0) = lim
Z t
t→0
Z
1 2 2
= lim |∇ū + t∇v| dx − |∇ū| dx
t→0 t Ω Ω
Z Z Z
= lim 2 ∇ū · ∇v dx + t |∇v|2 dx = 2 ∇ū · ∇v dx.
t→0 Ω Ω Ω
Therefore we obtain that a minimizer ū satisfies

Z
∇ū · ∇v dx = 0 for every v ∈ C 1 (Ω) with v = 0 on ∂Ω.
Ω
Integrating by parts (assuming that ū is sufficiently regular) we then obtain

Z
(−∆ū)v dx = 0 for every v ∈ C 1 (Ω) with v = 0 on ∂Ω,
Ω
which eventually yields to the partial differential equation

(
−∆ū = 0 in Ω,
(0.4)
ū = u0 on ∂Ω.
In other words, we found that a (sufficiently regular) minimizer of the Dirichlet energy (0.3)
solves the Laplace equation (0.4).
This approach can be adapted to more general integral functionals in the form (0.1), as
we will see, and leads to explicit solutions obtained by solving the associated Euler-Lagrange
equation. In the classical Calculus of Variations a large number of one-dimensional problems
were solved in this way, by means of ad hoc techniques to solve the corresponding ordinary
differential equation. For the higher dimensional case, however, this method shows insuperable
limitations, due to the complexity of the associated partial differential equation which in most
cases cannot be solved explicitly. Notice also that this strategy has some caveat:
• tipically, to take the derivative you will need some regularity assumption on the inte-
grand f (in the example this was obvious, as f (ξ) = |ξ|2 );
5
• more significantly, we are assuming a priori that a minimizer ū exists (and also has
some regularity), which is in general not granted!
Of course, the argument works also in the opposite way: you might be interested in finding a
solution to some partial differential equation; if the equation is the Euler-Lagrange equation
associated to some functional, then you might try to find (by other methods) a minimizer of
the functional, which will be the solution to the equation that you were looking for (notice,
however, that in principle you might not obtain all the solutions).
The idea at the core of the modern approach of the Calculus of Variations starts with the
observation that any function f : [a, b] → R defined on a bounded and closed interval in R
has a maximum as a minimum, as a consequence of Weierstrass Theorem. The proof goes as
follows: you take a minimizing sequence (xn )n∈N ⊂ [a, b], that is a sequence of points with
the property that
f (xn ) → inf f (x) as n → ∞
x∈[a,b]
(which always exists by definition of infimum). Then, the compactness of the interval [a, b]
allows you to extract a convergent subsequence xnk → x0 . By continuity of f you can then
deduce that that f (xnk ) → f (x0 ), which allows to conclude that x0 is a minimum point of f ,
and the minimum of f is attained. Actually, by inspecting this argument it is easy to realize
that you do not really need the continuity, but only the lower semicontinuity of f , which is
a much weaker property:
f (x0 ) ≤ lim inf f (xn ) for any sequence xn → x0 .
n→∞
The compactness of the space and the lower semicontinuity of the function are still sufficient
to guarantee the existence of a minimizer.
Therefore, to generalize this method to a minimization problem such as (0.2) we need the
class U to be compact, and the functional F to be lower semicontinuous with respect to the
same topology. This is the core of the so-called direct method of the Calculus of Variations,
which amounts to find a topological space which is compact, and such that the topology is
sufficiently rich to guarantee the lower semicontinuity of the functional. However, this is not
always a trivial task, as the last two properties are in competition. It is usually required to
enlarge the space on which F is originally defined.
In the case of integral functionals as in (0.1), we surely need a space with a notion of
differentiability, at least in a weak sense. Thanks to the contribution of several mathematicians
(Beppo Levi, Tonelli, Sobolev, Morrey, Caccioppoli, Stampacchia. . . ), we arrive at a “good”
formulation of the minimum problem for a large class of functionals in the Sobolev spaces of
functions which possess a “weak derivative” with some integrability properties, endowed with
the weak topology.
If we do the things in the right way, with this method we obtain the existence of a
minimizer, which however might have very weak regularity properties. The issue of regularity
of minimizers is therefore a central problem in this approach: once you have established the
existence of a minimizer in the enlarged space by means of the abstract direct method, it
is desirable to prove that the minimizer actually belongs to the original space and is more
regular, as a consequence of its minimality property. Think about the case of the Dirichlet
integral (0.3): if a minimizer exists, then we know that it satisfies the Euler-Lagrange equation
(0.4), i.e. the minimizer is a harmonic function, much more regular than what you originally
required!
6
Chapter 1
Minimization of integral functionals

among Lipschitz functions
The first problem that we consider in these notes is the following minimization problem for
an integral functional in the class of Lipschitz functions. As main reference for this chapter,
see [6, Chapter 1].
Let Ω ⊂ Rn be open and bounded, and let f : Rn → R. We consider the functional
Z
F (u) := f (∇u(x)) dx for u ∈ Lip(Ω), (1.1)
Ω
and the corresponding minimum problem

n o
min F (u) : u ∈ Lip(Ω), u = ϕ on ∂Ω (1.2)
where ϕ is a fixed boundary datum. More precisely, our goal will be to determine sufficient
conditions on Ω and on f which guarantee the existence of a minimizer, that is a function
ū ∈ Lip(Ω) such that ū = ϕ on ∂Ω and
F (ū) ≤ F (v) for every v ∈ Lip(Ω) with v = ϕ on ∂Ω.
The method that we will see in this chapter is independent on the growth of the function
f , and in particular it can be applied to the area functional
Z p
F (u) := 1 + |∇u(x)|2 dx,
Ω
which we should consider as a prototype example. In this case the integrand f has only linear
growth at infinity, in the sense that
c1 |ξ| ≤ f (ξ) ≤ c2 (1 + |ξ|).
The minimization of functionals with linear growth is often a subtle problem, as the more
refined methods that we will see in the next chapter usually fail do to the lack of coercivity. On
the other hand, with the strategy developed in this chapter we need to consider an integrand
f depending only on the gradient of u, and the function u needs to be scalar.
In this notes, for a Lebesgue measurable set Ω ⊂ Rn we will denote its n-dimensional
Lebesgue measure by the symbols Ln (Ω) or |Ω|.
7
1.1 Reminder: Lipschitz functions
We recall that a function u : E → R, with E ⊂ Rn , is Lipschitz continuous if there exists a
constant L > 0 such that
|u(x) − u(y)| ≤ L|x − y| for all x, y ∈ E.
We denote by Lip(E) the class of Lipschitz continuous functions on E. Obviously, Lipschitz
continuous functions are continuous, with a uniform modulus of continuity. For a Lipschitz
function u ∈ Lip(E), we denote its Lipschitz constant by
|u(x) − u(y)|
Lip(u) := sup ,
x,y∈E |x − y|
x6=y
and, for k > 0, we denote the set of Lipschitz continuous functions with Lipschitz constant
bounded by k by
Lipk (E) := u ∈ Lip(E) : Lip(u) ≤ k .
Finally, if ϕ ∈ Lip(∂E), we define

Lip(Ē; ϕ) := u ∈ Lip(Ē) : u = ϕ on ∂E .
Recall that, by Rademacher’s Theorem, a Lipschitz continuous function u is differentiable
almost everywhere, so that ∇u(x) is well-defined for a.e. x.
Exercise 1. Let Ω ⊂ Rn be open and let u ∈ Lipk (Ω). Prove that |∇u(x)| ≤ k for almost
every x ∈ Ω.[1]
We also recall the following extension result (which holds more in general if E is an
arbitrary metric space, with the same proof).
Theorem 1.1 (McShane). Let E ⊂ Rn and let f ∈ Lip(E). Then there exists a function
f˜ ∈ Lip(Rn ) such that f˜(x) = f (x) for all x ∈ E, and Lip(f˜) = Lip(f ).
Proof. Denoting by L := Lip(f ), we define for x ∈ Rn
f˜(x) := inf f (y) + L|x − y| : y ∈ E .

First notice that f˜(x) > −∞ for all x ∈ Rn . Indeed, fix a point x0 ∈ E; then for all x ∈ Rn
and y ∈ E we have
f (y) + L|x − y| = f (y) + L|(x − x0 ) − (y − x0 )|
≥ f (y) + L|y − x0 | − L|x − x0 |
≥ f (x0 ) − L|x − x0 |,
where we used the Lipschitz continuity of f in the last inequality.
Next, if x ∈ E, we have f (x) ≤ f (y) + L|x − y| for all y ∈ E, which yields f (x) ≤ f˜(x);
since the opposite inequality is trivial, we conclude that f˜(x) = f (x) for all x ∈ E.
Finally, if ξ, η ∈ Rn then
f˜(ξ) = inf f (y) + L|ξ − y| : y ∈ E

≤ inf f (y) + L|ξ − η| + L|η − y| : y ∈ E
= f˜(η) + L|ξ − η|,
from which we easily obtain that Lip(f˜) = L.
8
We finally recall the following compactness theorem.
Theorem 1.2 (Ascoli-Arzelà). Let K ⊂ Rn be a compact set, and let (fj )j∈N be a sequence
of continuous functions fj : K → R such that:
(i) (fj )j is equibounded, that is there exists M ∈ R such that
sup |fj (x)| ≤ M for all j ∈ N;

x∈K
(ii) (fj )j is uniformly equicontinuous, that is for all ε > 0 there exists δ > 0 such that
|fj (x) − fj (y)| < ε for all x, y ∈ K with |x − y| < δ, for all j ∈ N.
Then there exist f ∈ C 0 (K) and a subsequence (fjk )k such that fjk → f uniformly on K.
Notice that any sequence of equi-Lipschitz functions (uj )j (that is, with uniformly bounded
Lipschitz constant: supj Lip(uj ) < +∞) satisfies the equicontinuity assumption.
1.2 Existence of minimizers: preliminary results

We now discuss the existence of minimizers for problem (1.2). We first consider a modified
problem, in which we introduce and artificial constraint on the Lipschitz constant of admissible
functions.
Proposition 1.3. Let Ω ⊂ Rn be open and bounded, let ϕ ∈ Lip(Ω) be a fixed boundary
datum, and let f : Rn → R be of class C 1 (Rn ) and convex. Then for every k ≥ Lip(ϕ) the
minimum problem
Z
min f (∇u(x)) dx : u ∈ Lipk (Ω), u = ϕ on ∂Ω (1.3)
Ω
has a solution.
Proof. Let Lipk (Ω; ϕ) := {u ∈ Lipk (Ω) : u = ϕ on ∂Ω} and let
m := inf{F (u) : u ∈ Lipk (Ω; ϕ)}.
We want to prove that there exists ū ∈ Lipk (Ω; ϕ) such that F (ū) = m. As a preliminary
remark, we observe that, since by Exercise 1 we have |∇u(x)| ≤ k for a.e. x ∈ Ω, for all
u ∈ Lipk (Ω; ϕ), then Z
|F (u)| ≤ |f (∇u)| dx ≤ |Ω| | max |f (ξ)|,
Ω ξ∈Bk
which in particular implies that m ∈ R.

Step 1: compactness. Let (uj )j be a minimizing sequence, that is uj ∈ Lipk (Ω; ϕ) and
F (uj ) → m as j → ∞. We wish to apply Ascoli-Arzelà Theorem 1.2 to the sequence (uj )j .
Let us check the assumptions:
9
• (uj )j is equibounded: for all x ∈ Ω and y ∈ ∂Ω we have
|uj (x)| − |ϕ(y)| = |uj (x)| − |uj (y)| ≤ |uj (x) − uj (y)| ≤ k|x − y|,
which implies
|uj (x)| ≤ |ϕ(y)| + k|x − y| for all x ∈ Ω, y ∈ ∂Ω
and in turn, as Ω and ϕ are bounded by assumption,
sup |uj | ≤ sup |ϕ| + k diam Ω

Ω ∂Ω
(where diam Ω := sup{|x − y| : x, y ∈ Ω}). This shows equiboundedness.

• (uj )j is uniformly equicontinuous: this follows easily from the fact that the uj are all
equi-Lipschitz continuous, that is
|uj (x) − uj (y)| ≤ k|x − y| for all x, y ∈ Ω.
We can the apply Ascoli-Arzelà Theorem 1.2 and deduce that there exists ū ∈ C 0 (Ω) and a
subsequence (uj` )` such that uj` → ū uniformly in Ω as ` → ∞. From the uniform convergence
it is easy to deduce that Lip(ū) ≤ k and ū = ϕ on ∂Ω, so that ū ∈ Lipk (Ω; ϕ).
Step 2: minimality. To conclude the proof, we need to show that F (ū) = m. Since f is
convex and of class C 1 , we have
f (η) ≥ f (ξ) + h∇f (ξ), η − ξi for all ξ, η ∈ Rn ,
so that by applying the previous inequality with ξ = ∇ū(x) and η = ∇uj` (x), and integrating
over Ω, we obtain
Z Z Z
f (∇uj` ) dx ≥ f (∇ū) dx + h∇f (∇ū), ∇uj` − ∇ūi dx. (1.4)
Ω Ω Ω
Now, since ∇f (∇ū) ∈ L∞ (Ω; Rn ), by density of smooth functions in L1 we have that for
every ε > 0 there exists gε ∈ C 1 (Ω; Rn ) such that
Z

∇f (∇ū) − gε dx < ε. (1.5)
Ω
From (1.4) we obtain

Z Z Z
f (∇uj` ) dx − f (∇ū) dx ≥ h∇f (∇ū), ∇uj` − ∇ūi dx
Ω ZΩ Ω Z
= hgε , ∇uj` − ∇ūi dx + h∇f (∇ū) − gε , ∇uj` − ∇ūi dx
ZΩ ZΩ

≥ hgε , ∇uj` − ∇ūi dx − ∇f (∇ū) − gε ∇uj − ∇ū dx
`
Ω Ω
(1.5)
Z
≥ hgε , ∇uj` − ∇ūi dx − 2kε
ZΩ
= − (uj` − ū)divgε dx − 2kε,
Ω
10
where we used integration by parts in the last passage. Hence
Z
F (uj` ) ≥ F (ū) − (uj` − ū)divgε dx − 2kε.
Ω
By sending ` → ∞, recalling that (uj )j was a minimizing sequence and that uj` → ū uniformly
we obtain
m = lim F (uj` ) ≥ F (ū) − 2kε,
`→∞
and as ε is arbitrary we conclude that F (ū) = m, that is what we had to show.
Remark 1.4. The C 1 assumption on f in Proposition 1.3 (which is used only in Step 3 of the
above proof ) could actually be removed, by using the lower semicontinuity results for integral
functionals which we will prove in Chapter 2.
Now that we have established the existence of a minimum point for the problem (1.3),
we would like to obtain a minimizer also for the initial problem (1.2), that is to remove the
restriction on the Lipschitz constant. The following proposition gives a sufficient condition
which guarantees that a minimizer for (1.3) is also a minimizer for (1.2).
Proposition 1.5. Under the assumptions of Proposition 1.3, let ū ∈ Lipk (Ω; ϕ) be a mini-
mizer for problem (1.3). If Lip(ū) < k, then ū is also a minimizer for (1.2).
Proof. Let v ∈ Lip(Ω; ϕ) and set, for ε > 0, vε := ū + ε(v − ū). For ε sufficiently small we
have
Lip(vε ) ≤ Lip(ū) + ε Lip(v − ū) < k,
therefore vε ∈ Lipk (Ω; ϕ) is admissible for problem (1.3), and by minimality of ū
Z
F (ū) ≤ F (vε ) =

f ε∇v + (1 − ε)∇ū dx
Z Ω Z
≤ ε f (∇v) dx + (1 − ε) f (∇ū) dx = εF (v) + (1 − ε)F (ū),
Ω Ω
where we used in particular the convexity of f in the second line. This shows that F (ū) ≤
F (v) for all v ∈ Lip(Ω; ϕ), that is ū solves (1.2).
1.3 Maximum principle

Definition 1.6. We say that u ∈ Lipk (Ω) minimizes F in Lipk (Ω) with respect to its own
boundary conditions (or equivalently that u is a minimizer of F in Lipk (Ω; u)) if
F (u) = min F (v) : v ∈ Lipk (Ω), v = u on ∂Ω .

Proposition 1.7 (Maximum principle). In addition to the assumptions of Proposition 1.3,

assume that f is strictly convex, that is
f (λξ + (1 − λ)η) < λf (ξ) + (1 − λ)f (η) for all ξ, η ∈ Rn , ξ 6= η, and λ ∈ (0, 1).
Let u, v ∈ Lipk (Ω) be minimizers of F in Lipk (Ω) with respect to their own boundary condi-
tions. Then
sup |u − v| = max |u − v|. (1.6)
Ω ∂Ω
11
Proof. We prove the proposition in two steps. In the first step we prove a comparison principle,
which is the key property leading to the maximum principle in the statement.
Step 1: comparison principle. We claim that, if u ≤ v on ∂Ω, then u ≤ v in Ω.
To prove the claim, consider u ∧ v := min{u, v}, u ∨ v := max{u, v}. These functions
belong to Lipk (Ω), u ∧ v = u on ∂Ω, u ∨ v = v on ∂Ω, and
( (
∇u a.e. on {u < v}, ∇u a.e. on {u ≥ v},
∇(u ∧ v) = ∇(u ∨ v) = (1.7)
∇v a.e. on {u ≥ v}, ∇v a.e. on {u < v},
see Exercise 2 below. Then

Z Z Z
f (∇(u ∧ v)) dx = f (∇u) dx + f (∇v) dx,
Ω {u<v} {u≥v}
Z Z Z
f (∇(u ∨ v)) dx = f (∇v) dx + f (∇u) dx.
Ω {u<v} {u≥v}
By adding the previous two equalities and using the minimality of u and v we obtain
Z Z Z Z
f (∇u) dx + f (∇v) dx ≤ f (∇(u ∧ v)) dx + f (∇(u ∨ v)) dx
Ω Ω ZΩ Z Ω
= f (∇u) dx + f (∇v) dx,

Ω Ω
which implies that F (u) = F (u ∧ v), F (v) = F (u ∨ v). Since by strict convexity the
minimizer is unique (see Exercise 3 below), we conclude that u = u ∧ v, that is u ≤ v in Ω.
Step 2: maximum principle. Denote by ε := max∂Ω |u − v|. Since v + ε, v − ε ∈ Lipk (Ω) and
F (v + ε) = F (v − ε) = F (v), we have that v + ε and v − ε are minimizers of F in Lipk (Ω)
with respect to their own boundary conditions. Moreover v − ε ≤ u ≤ v + ε on ∂Ω. We can
then apply the comparison principle and obtain that
v − ε ≤ u ≤ v + ε in Ω,
that is supΩ |u − v| ≤ ε = max∂Ω |u − v|. The other inequality is immediate by continuity.
Exercise 2. Let u, v ∈ Lipk (Ω). Prove that u ∨ v, u ∧ v ∈ Lipk (Ω) and that (1.7) holds.[2]
Exercise 3. Let f be a strictly convex function. Prove that the minimum problem (1.2) (or
(1.3)) has at most one solution.[3]
Remark 1.8. The previous result, as well as the uniqueness of minimizers, does not hold
if we assume that f is just convex and not strictly convex. As a counterexample, consider
f (ξ) := max{0, |ξ| − M } for some fixed M > 0, and the associated functional F . Then f is
convex but not strictly convex. Clearly every function u ∈ LipM (Ω) minimizes F with respect
to its own boundary conditions.
Recall that we already established the existence of a minimizer ū for the problem (1.3),
where we imposed the constraint Lip(u) ≤ k. Also recall that, in order to show that ū is
also a solution to the original problem (1.2), by Proposition 1.5 it is sufficient to show the
strict inequality Lip(ū) < k. Thanks to the maximum principle, in the following proposition
we deduce a useful expression which will be used to estimate the Lipschitz constant of a
minimizer.
12
Proposition 1.9. Let Ω and f satisfy the assumptions of Proposition 1.3, and assume also
that f is strictly convex. Let ū be the minimizer of (1.3), given by Proposition 1.3. Then
|ū(x) − ū(y)| |ū(x) − ū(y)|

Lip(ū) := sup = sup . (1.8)
x,y∈Ω |x − y| x∈Ω |x − y|
x6=y y∈∂Ω
x6=y
Proof. Denote by L the right-hand side of (1.8). We have to show that Lip(ū) ≤ L, as the
other inequality is trivial. Therefore, we need to prove that given any x, y ∈ Ω there exist
x0 ∈ Ω and y0 ∈ ∂Ω such that
|ū(x) − ū(y)| |ū(x0 ) − ū(y0 )|
≤ . (1.9)
|x − y| |x0 − y0 |
Let τ := y − x and uτ (z) := ū(z + τ ). We remark that:
(a) uτ is defined for z ∈ Ωτ := {z ∈ Rn : z + τ ∈ Ω};
(b) since x + τ = y ∈ Ω, we have x ∈ Ω ∩ Ωτ ;
(c) uτ (x) = ū(y);
(d) uτ minimizes F in Lipk (Ωτ ; uτ );
(e) if v minimizes F in Lipk (Ω; v) and A ⊂ Ω is open, then v|A minimizes F in Lipk (A; v).
Indeed, assume by contradiction that there exists w ∈ Lipk (A; v) such that
Z Z
f (∇w) dx < f (∇v) dx.
A A
Then the function (

w(z) if z ∈ A,
w̃(z) :=
v(z) if z ∈ Ω\A
belongs to Lipk (Ω; v) (exercise) and

Z Z Z Z
f (∇w̃) dx = f (∇w) dx + f (∇v) dx < f (∇v) dx,
Ω A Ω\A Ω
which contradicts the minimality of v.
We are now ready to prove (1.9). We apply the maximum principle (Proposition 1.7) in
Ω ∩ Ωτ : by using properties (d) and (e) we have that both ū and uτ minimize F in Ω ∩ Ωτ ,
with respect to their own boundary conditions; hence
max |ū(z) − uτ (z)| = max |ū(z) − uτ (z)| = |ū(z0 ) − uτ (z0 )|

z∈Ω∩Ωτ z∈∂(Ω∩Ωτ )
for some z0 ∈ ∂(Ω ∩ Ωτ ). In particular, by (b) we deduce that
|ū(x) − uτ (x)| ≤ |ū(z0 ) − uτ (z0 )|.
Since ∂(Ω ∩ Ωτ ) = (∂Ω ∩ Ωτ ) ∪ (∂Ωτ ∩ Ω), we eventually have two possibilities:
13
• if z0 ∈ ∂Ω ∩ Ωτ , then z0 + τ ∈ Ω. In this case define y0 := z0 ∈ ∂Ω, x0 := z0 + τ ∈ Ω,
and we have
|ū(x) − ū(y)| (c) |ū(x) − uτ (x)| |ū(z0 ) − uτ (z0 )| |ū(y0 ) − ū(x0 )|
= ≤ = ,
|x − y| |x − y| |τ | |x0 − y0 |
which is the goal (1.9);
• if z0 ∈ ∂Ωτ ∩ Ω, then z0 + τ ∈ ∂Ω. In this case define y0 := z0 + τ ∈ ∂Ω, x0 := z0 ∈ Ω,

and we have
|ū(x) − ū(y)| (c) |ū(x) − uτ (x)| |ū(z0 ) − uτ (z0 )| |ū(x0 ) − ū(y0 )|
= ≤ = ,
|x − y| |x − y| |τ | |x0 − y0 |
which proves the goal (1.9) also in this case.
The proof is concluded.
1.4 The bounded slope condition and existence of minimizers

Thanks to the preliminary results discussed in the previous sections, we are now in position to
prove the existence of solutions to the original problem (1.2). The existence and the Lipschitz
regularity of minimizers are guaranteed by an additional assumption on the boundary datum
ϕ: we require that for all x ∈ ∂Ω we can find two hyperplanes through the point (x, ϕ(x)) in
Rn+1 such that the graph of ϕ is included between them, and the slope of the hyperplanes is
uniformly bounded with respect to x ∈ ∂Ω.
Definition 1.10 (Bounded slope condition). We say that the domain Ω and the boundary
datum ϕ : ∂Ω → R satisfy the bounded slope condition if there exists M > 0 such that for all
x ∈ ∂Ω there are a, b ∈ Rn , with |a|, |b| ≤ M , so that
ϕ(x) + a · (y − x) ≤ ϕ(y) ≤ ϕ(x) + b · (y − x) for all y ∈ ∂Ω.
The bounded slope condition is clearly satisfied if ϕ is an affine function. We remark that
this is essentially an assumption on the domain Ω: indeed, if ϕ is not affine then a 6= b for all
x ∈ ∂Ω; therefore
for all x ∈ ∂Ω there exists αx ∈ Rn \{0} such that αx · (y − x) ≥ 0 for all y ∈ ∂Ω.
(simply by choosing α = b − a). This means that, for every x ∈ ∂Ω, there is a half-space
Vx := {y ∈ Rn : αx · (y − x) ≥ 0} such that Ω ⊂ Vx and x ∈ ∂Vx . Hence each point in ∂Ω has
a supporting hyperplane, and Ω is convex.
We can now prove the main result of this chapter.
Theorem 1.11. Let Ω ⊂ Rn be open and bounded, and let f : Rn → R be of class C 1 and
convex. Let ϕ : ∂Ω → R be such that (Ω, ϕ) satisfy the bounded slope condition. Then there
exists a solution to the minimum problem
n o
min F (u) : u ∈ Lip(Ω), u = ϕ on ∂Ω , (1.10)
where F is the functional defined in (1.1).
14
Proof. We first observe that the bounded slope condition implies that ϕ ∈ Lip(∂Ω) with
Lip(ϕ) ≤ M : indeed for every x, y ∈ ∂Ω we have
−M |x − y| ≤ a · (y − x) ≤ ϕ(y) − ϕ(x) ≤ b · (y − x) ≤ M |x − y|,
that is
|ϕ(x) − ϕ(y)| ≤ M |x − y| for all x, y ∈ ∂Ω.
Then by McShane Theorem 1.1 there exists at least one extension u ∈ LipM (Ω) such that
u = ϕ on ∂Ω, so that the minimum problem (1.10) makes sense as the class of competitors is
not empty. We divide the proof of the theorem into three steps.
Step 1. If u is an affine function, then u minimizes F in Lip(Ω; u); that is, every affine
function is a minimizer with respect to its own boundary conditions.
Indeed, let u(x) = a·x+b be an affine function, with a ∈ Rn , b ∈ R. For every v ∈ Lip(Ω; u)
we have by the divergence theorem and Jensen’s inequality
Z Z
1 1
f (a) = f ∇v dx ≤ f (∇v) dx,
|Ω| Ω |Ω| Ω
and therefore Z Z
F (u) = f (∇u) dx = f (a)|Ω| ≤ f (∇v) dx = F (v).
Ω Ω
Step 2. We prove the theorem under the additional assumption that f is strictly convex.
Let k > M , with M given by the bounded slope condition. We remarked above that
Lip(ϕ) ≤ M ; let then ū ∈ Lipk (Ω; ϕ) be a minimizer of F on Lipk (Ω; ϕ), whose existence is
given by Proposition 1.3. By Proposition 1.9 we have
|ū(x) − ū(y)| |ϕ(x) − ū(y)|
Lip(ū) = sup = sup . (1.11)
x∈∂Ω |x − y| x∈∂Ω |x − y|
y∈Ω y∈Ω
x6=y x6=y
Thanks to the bounded slope condition, for fixed x ∈ ∂Ω we have ϕ(y) ≤ ϕ(x) + b · (y − x)
for some b ∈ Rn with |b| ≤ M < k. The function v(y) := ϕ(x) + b · (y − x) is affine, with
slope smaller than k, and thanks to the first step it minimizes F with respect to its boundary
condition. Since ū = ϕ ≤ v on ∂Ω, we can then apply the comparison principle to get
ū(y) ≤ v(y) = ϕ(x) + b · (y − x) for all y ∈ Ω,
and by the same argument we also have
ū(y) ≥ ϕ(x) + a · (y − x) for all y ∈ Ω.
Therefore
−M |x − y| ≤ a · (x − y) ≤ ū(y) − ϕ(x) ≤ b · (y − x) ≤ M |x − y|,
that is
|ū(y) − ϕ(x)|
≤M for all x ∈ ∂Ω and y ∈ Ω.
|x − y|
By (1.11) we conclude that Lip(ū) ≤ M < k, and in turn by Proposition 1.5 that ū is a
solution to (1.10).
15
Step 3. We finally remove the strict convexity assumption on f . For j ∈ N, define fj (ξ) :=
f (ξ) + 1j |ξ|2 , ξ ∈ Rn . It is easily seen that fj is strictly convex, and by the previous step for
every j ∈ N there exists uj ∈ Lip(Ω; ϕ) which solves
n Z o
min Fj (u) := fj (∇u) dx : u ∈ Lip(Ω), u = ϕ on ∂Ω ,
Ω
and moreover Lip(uj ) ≤ M for all j. Since the functions uj are equi-Lipschitz continuous, we
can apply Ascoli-Arzelà Theorem 1.2 and obtain that, up to subsequences, uj → u uniformly
on Ω, for some u ∈ LipM (Ω) with u = ϕ on ∂Ω.
It only remains to show that u is a minimizer of (1.10). For every v ∈ Lip(Ω; ϕ) we have
Z Z Z Z Z
1
f (∇uj ) dx ≤ fj (∇uj ) dx ≤ fj (∇v) dx = f (∇v) dx + |∇v|2 dx.
Ω Ω Ω Ω j Ω
By arguing as in the second step of the proof of Proposition 1.3, using the uniform convergence
of uj to u, the C 1 regularity and the convexity of f , we can show that
Z Z
lim inf f (∇uj ) dx ≥ f (∇u) dx,
j→∞ Ω Ω
hence by passing to the limit as j → ∞ in the previous inequality we conclude that F (u) ≤
F (v) for all v ∈ Lip(Ω; ϕ), as desired.
Remark 1.12 (A sufficient condition for the bounded slope condition). We have seen that
if the bounded slope condition is satisfied and ϕ is not affine, then Ω must be convex. On
the other hand, convexity (or even strict convexity) alone is not sufficient to guarantee the
bounded slope condition. A sufficient condition is the following, which geometrically means
that the boundary of Ω cannot lie too close to its tangent hyperplane: assume that
a) Ω ⊂ Rn is open, bounded, convex and with smooth boundary;
b) there exists a constant c1 > 0 such that
c1 ν(x) · (y − x) ≥ |x − y|2 for all x, y ∈ ∂Ω,
where ν(x) denotes the interior unit normal to ∂Ω at x;
c) ϕ is of class C 2 (Rn ).
Then the bounded slope condition is satisfied. Indeed for all x, y ∈ ∂Ω
ϕ(y) = ϕ(x) + ∇ϕ(x) · (y − x) + R(x, y) with |R(x, y)| ≤ c2 |x − y|2
for some c2 > 0 independent of x and y. Then it is sufficient to take
b := ∇ϕ(x) + c1 c2 ν(x), a := ∇ϕ(x) − c1 c2 ν(x)
and the bounded slope condition is satisfied with this choice.
16
Notes
[1]
Solution to Exercise 1. Let x ∈ Ω be a point of differentiability for u (by Rademacher’s Theorem, almost
every point has this property); we show that |∇u(x)| ≤ k. Suppose that |∇u(x)| 6= 0, otherwise the conclusion
∇u(x)
is trivial. Let ξ := |∇u(x)| ; then ξ ∈ Rn , |ξ| = 1, and
u(x + hξ) − u(x) k|x + hξ − x|

|∇u(x)| = h∇u(x), ξi = ∂ξ u(x) = lim ≤ lim = k.
h→0 h h→0 |h|
[2]
Solution to Exercise 2. It is easy to verify that u ∨ v, u ∧ v ∈ Lipk (Ω). To show formula (1.7), we only need
to care about the set {u = v}: indeed {u < v} and {u > v} are open sets on which u ∨ v and u ∧ v coincide
with either u or v, and the proof is immediate. We are then left with the proof that
∇(u ∨ v) = ∇(u ∧ v) = ∇u = ∇v a.e. on {u = v}.
This follows from the following property: if w ∈ Lip(Ω), then
∇w = 0 a.e. on {w = 0}. (1.12)
Let us prove (1.12) in dimension n = 1, that is in the case Ω = I is an interval. Since w is Lipschitz and we
want to prove (1.12) almost everywhere in {w = 0}, we can assume without loss of generality (by Rademacher’s
theorem) that w is differentiable at every point of {w = 0}. We define E0 := {x ∈ I : w(x) = 0, |w0 (x)| > 0}
and we need to show that E0 has measure zero. For k ∈ N let
n |x − y| o
Ek := x ∈ I : w(x) = 0, |w(y)| ≥ for all y ∈ (x − k1 , x + k1 ) ∩ I .
k
Observe that E0 ⊂ k Ek . Now, it is clear that if x ∈ Ek , then (x − k1 , x + k1 ) ∩ Ek = {x}, since w(y) 6= 0 for
S
all y ∈ (x − k1 , x + k1 ); it follows immediately that the measure of Ek is zero, and therefore also the measure
of E is zero.
The proof of (1.12) in dimension n > 1 follows from the one-dimensional case using Fubini’s Theorem:
indeed setting Ei := {z ∈ Ω : w(z) = 0, |∂i w(z)| > 0}, for i ∈ {1, . . . , n}, we have
Z
L1 t ∈ R : (x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) ∈ Ei dx1 . . . dxi−1 dxi+1 . . . dxn = 0

|Ei | =
Rn−1
(indeed the function inside the integral is equal to zero, by using the fact that (1.12) holds in dimension one
and considering the one-dimensional restrictions of w to the lines in the direction xi ).
Solution to Exercise 3. Let u and v be two minimum points of F in Lip(Ω; ϕ), with u 6= v. Then ∇u 6= ∇v
[3]
on a set of positive measure. Since u+v

2
∈ Lip(Ω; ϕ) and by strict convexity we have
u Z
v 1 1
min F ≤ F + = f ∇u + ∇v dx
Lip(Ω;ϕ) 2 2 Ω 2 2
Z Z
1 1
< f (∇u) dx + f (∇v) dx
2 Ω 2 Ω
1 1 1 1
= F (u) + F (v) = min F + min F = min F ,
2 2 2 Lip(Ω;ϕ) 2 Lip(Ω;ϕ) Lip(Ω;ϕ)
which is a contradiction.
17
Chapter 2
Direct methods
2.1 An overview of Sobolev Spaces

Sobolev spaces turn out to be the proper setting in which to apply the direct method of the
Calculus of Variations to obtain existence of minimizers of integral functionals, which is the
main goal of this chapter. We start with a basic introduction to the main definitions and
properties of Sobolev spaces, assuming that the reader is already familiar with this theory
and referring to [4] for the proofs of the results stated below. For a deeper and more complete
discussion of Sobolev spaces we refer to the monograph [7].
2.1.1 Weak derivatives and Sobolev spaces

Let Ω ⊂ Rn be an open set. We denote by Cc∞ (Ω) the space of infinitely differentiable
functions ϕ : Ω → R with compact support in Ω, which will often be referred to as test
functions.
In order to introduce the Sobolev spaces, we weaken the notion of derivative by requiring
the validity of the integration by parts formula. This will enlarge the class of C 1 functions to
a space containing less smooth functions but with better compactness properties.
Definition 2.1 (Weak derivative). Let u ∈ L1loc (Ω). We say that a function v ∈ L1loc (Ω) is
the weak derivative of u with respect to xi , i ∈ {1, . . . , n}, if
Z Z
∂ϕ
u dx = − vϕ dx for all ϕ ∈ Cc∞ (Ω), (2.1)
Ω ∂xi Ω
∂u
and we will denote it by the symbols v = ∂xi = ∂i u.
The weak derivative, if it exists, is unique (up to a set of measure zero). Moreover it is
easily seen that for a function u ∈ C 1 (Ω) the notion of weak derivative coincides with the
classical notion of derivative. We will denote the vector of the weak first derivatives of u by
the symbol Du(x) = ∇u(x) := (∂1 u(x), . . . , ∂n u(x)).
Example 2.2. Let n = 1, Ω = (0, 2), and let u be the function

(
x if 0 < x ≤ 1,
u(x) =
1 if 1 < x < 2.
18
We claim that the weak derivative of u is the function
(
1 if 0 < x ≤ 1,
v(x) =
0 if 1 < x < 2.
Indeed for every test function ϕ ∈ Cc∞ (0, 2):

Z 2 Z 1 Z 2 Z 1 Z 2
0 0 0
uϕ dx = xϕ (x) dx + ϕ (x) dx = ϕ(1) − ϕ(x) dx + ϕ(2) − ϕ(1) = − vϕ dx,
0 0 1 0 0
hence v = u0 in the weak sense.
Example 2.3. Let n = 1, Ω = (0, 2), and let u be the function

(
x if 0 < x ≤ 1,
u(x) =
2 if 1 < x < 2.
Then u does not admit a weak derivative: indeed, if by contradiction there is a function
v ∈ L1loc (0, 2) satisfying (2.1), then for all test functions ϕ ∈ Cc∞ (0, 2)
Z 2 Z 2 Z 1 Z 2
− vϕ dx = uϕ0 dx = xϕ0 (x) dx + 2 ϕ0 (x) dx
0 0 0 1
Z 1
= ϕ(1) − ϕ(x) dx + 2ϕ(2) − 2ϕ(1)
0
Z 1
=− ϕ dx − ϕ(1).
0
Now we apply the previous identity to a sequence of test functions ϕk ∈ Cc∞ (0, 2) such that
0 ≤ ϕk ≤ 1, ϕk (1) = 1 for all k, and ϕk (x) → 0 as k → ∞ for all x 6= 1: then
Z 2 Z 1
− vϕk dx = − ϕk dx − ϕk (1),
0 0
and we obtain a contradiction by passing to the limit as k → ∞.
Definition 2.4 (Sobolev space W 1,p (Ω)). Let Ω ⊂ Rn be an open set, and let 1 ≤ p ≤ +∞.
We define the Sobolev space
n o
W 1,p (Ω) := u ∈ Lp (Ω) : ∂i u ∈ Lp (Ω) for all i = 1, . . . , n (2.2)
consisting of all functions u ∈ Lp (Ω) such that all the weak derivatives of order one exist and
belong to Lp (Ω). For u ∈ W 1,p (Ω) we define the norm
n Z
!1
Z p
∂u p
X
p
kukW 1,p (Ω) = |u| dx + dx (1 ≤ p < +∞),
∂xi

Ω i=1 Ω (2.3)
n
∂u
X
kukW 1,∞ (Ω) = kukL∞ (Ω) + .
∂xi L∞ (Ω)

i=1
19
The space W 1,p (Ω) with the norm k · kW 1,p (Ω) is a Banach space. In the special case p = 2
we usually write
H 1 (Ω) := W 1,2 (Ω), (2.4)
as this is an Hilbert space with the scalar product
Z n Z
X ∂u ∂v
(u, v)H 1 := uv dx + dx for u, v ∈ H 1 (Ω). (2.5)
Ω Ω ∂xi ∂xi
i=1
1,p
We say that u ∈ Wloc (Ω) if u ∈ W 1,p (Ω0 ) for all Ω0 ⊂⊂ Ω. Given a sequence uk ∈ W 1,p (Ω)
and u ∈ W (Ω), by (strong) convergence in W 1,p (Ω) we mean convergence in norm, i.e.
1,p
uk → u in W 1,p (Ω) ⇐⇒ kuk − ukW 1,p (Ω) → 0,

1,p
uk → u in Wloc (Ω) ⇐⇒ kuk − ukW 1,p (Ω0 ) → 0 for all Ω0 ⊂⊂ Ω.
More in general, we can consider weak derivatives of higher order and the corresponding
Sobolev spaces. It is convenient to use the multiindex notation: a n-dimensional multiindex
α is a vector α = (α1 , . . . , αn ) with αi ∈ N; the length of α is |α| := α1 + . . . + αn . For a
function ϕ ∈ Cc∞ (Rn ) and a multiindex α we let
∂ |α| ϕ
Dα ϕ(x) := (x).
∂xα1 1 · · · ∂xαnn
Definition 2.5. Let u ∈ L1loc (Ω) and let α be a multiindex. We say that a function v ∈ L1loc (Ω)
is the α-weak derivative of u if
Z Z
uDα ϕ dx = (−1)|α| vϕ dx for all ϕ ∈ Cc∞ (Ω), (2.6)
Ω Ω
and we will denote it by the symbol v = Dα u.
For 1 ≤ p ≤ +∞ and k ∈ N we define the
Sobolev space
n o
W k,p (Ω) := u ∈ Lp (Ω) : Dα u ∈ Lp (Ω) for all multiindex α with |α| ≤ k (2.7)
(and H k (Ω) := W k,2 (Ω)), with norm

!1
X Z p
kukW k,p (Ω) = |Dα u|p dx (1 ≤ p < +∞),

|α|≤k Ω
(2.8)
X
α
kukW k,∞ (Ω) = kD ukL∞ (Ω) .
|α|≤k
The following proposition deals with the structure of W k,p (Ω) as function space.
Proposition 2.6. Let Ω ⊂ Rn be an open set, k ∈ N. Then:
(i) if 1 < p < +∞, W k,p (Ω) is a reflexive Banach space;
(ii) if 1 ≤ p < +∞, W k,p (Ω) is separable.
Exercise 4. Let n = 2, Ω = (−1, 1) × (−1, 1), u(x, y) = |x|y. Show that u ∈ W 1,1 (Ω).[4]
Exercise 5. Let Ω = B1 (0) ⊂ Rn be the open unit ball in Rn , n ≥ 2. Prove that the function
u(x) = |x|−α , α > 0, belongs to W 1,p (Ω) if and only if α < n−p / W 1,p (Ω)
p . In particular, u ∈
for p ≥ n.[5]
20
2.1.2 Approximation
We next discuss the possibility of approximating Sobolev functions with smooth functions.
We start with the following interior approximation result, which can be proved by means of
a convolution argument.
Theorem 2.7 (Meyers-Serrin). Let Ω ⊂ Rn be open, 1 ≤ p < +∞, k ∈ N. Given u ∈

W k,p (Ω), there exists a sequence (uj )j ⊂ W k,p (Ω) ∩ C ∞ (Ω) such that uj → u in W k,p (Ω).
Remark 2.8. The theorem is false for p = +∞. Indeed, if (uk )k ⊂ C ∞ (Ω) ∩ W 1,∞ (Ω) is
such that kuk − ukW 1,∞ (Ω) → 0, then necessarily u ∈ C 1 (Ω).
Notice that in Theorem 2.7 we do not assert that uj ∈ C ∞ (Ω): we do not have a control
of the regularity of u up to the boundary of Ω! In order to obtain an approximation result
with functions in C ∞ (Ω), we need to require some regularity of ∂Ω.
Definition 2.9 (Lipschitz boundary). Let Ω ⊂ Rn be open and bounded. We say that Ω has
Lipschitz boundary if for every point x0 ∈ ∂Ω we can find an orthogonal coordinate system
(y1 , . . . , yn ), an open rectangle A × B ⊂ Rn (with A open rectangle in Rn−1 , B = (a, b) open
interval in R), and a Lipschitz function ψ : A → B such that x0 ∈ A × B and
Ω ∩ (A × B) = (y1 , . . . , yn ) ∈ Rn : (y1 , . . . , yn−1 ) ∈ A, a < yn < ψ(y1 , . . . , yn−1 ) .

Notice that not only ∂Ω has to be locally the graph of a Lipschitz function in some
coordinate system, but the domain Ω has to lie on one side of the graph. For instance, the
set Ω = R2 \{(x1 , 0) : x1 ∈ R} does not have Lipschitz boundary, even if ∂Ω is a straight line.
Theorem 2.10 (Global approximation). Let Ω ⊂ Rn be open and bounded with Lipschitz
boundary, 1 ≤ p < +∞, k ∈ N. Then the space C ∞ (Ω) is dense in W k,p (Ω): given u ∈
W k,p (Ω), there exists a sequence (uj )j ⊂ C ∞ (Ω) such that uj → u in W k,p (Ω) as j → ∞.
Weak derivatives enjoy several of the properties of classical derivatives: in particular

linearity and Leibniz’s formula for the derivative of the product.
Exercise 6. Let Ω ⊂ Rn be open, 1 ≤ p ≤ +∞. Prove that if u ∈ W 1,p (Ω) and ψ ∈ Cc∞ (Rn ),
then uψ ∈ W 1,p (Ω) with ∇(uψ) = ψ∇u + u∇ψ.
Also the chain rule remains valid in the framework of Sobolev spaces:
Proposition 2.11 (Chain rule). Let Ω ⊂ Rn be an open set, 1 ≤ p < +∞. Let f : R → R
be a C 1 function, with f 0 bounded; assume also f (0) = 0 if Ω has infinite measure. Then for
every u ∈ W 1,p (Ω) we have f ◦ u ∈ W 1,p (Ω) with D(f ◦ u) = f 0 (u)Du.
Proof. By boundedness of f 0 we have |f (t)| ≤ C(1 + |t|) for all t ∈ R, therefore if |Ω| is finite
Z Z
|f ◦ u|p dx ≤ C p (1 + |u|)p dx ≤ 2p C p |Ω| + kukpLp (Ω) < ∞.
Ω Ω
Hence f ◦ u ∈ Lp (Ω). If Ω has infinite measure, the assumptions f (0) = 0 yields |f (t)| ≤ C|t|,
so that the same conclusion follows as before.
21
We next look at the weak derivative of f ◦ u. Let uk ∈ C ∞ (Ω) be given by Theorem 2.7,
with uk → u in W 1,p (Ω). By the classical chain rule we have for all ϕ ∈ Cc∞ (Ω)
Z Z Z
(f ◦ uk )∂i ϕ dx = − ∂i (f ◦ uk )ϕ dx = − f 0 (uk )∂i uk ϕ dx. (2.9)
Ω Ω Ω
We now pass to the limit as k → ∞ in the previous identity: since

f ◦ uk (x) − f ◦ u(x) ≤ sup |f 0 (t)| |uk (x) − u(x)|,

t∈R
and uk → u in Lp (Ω), we obtain f ◦ uk → f ◦ u in Lp (Ω) and hence

Z Z
lim (f ◦ uk )∂i ϕ dx = (f ◦ u)∂i ϕ dx.
k→∞ Ω Ω
Moreover, by possibly passing to a subsequence we can assume that uk → u almost everywhere

in Ω, so that f 0 (uk ) → f 0 (u) almost everywhere. Therefore
Z Z Z
f 0 (uk )∂i uk ϕ dx − 0
0
f (uk )∂i uk − f 0 (uk )∂i u|ϕ| dx

f (u)∂ i u ϕ dx≤

Ω Ω Ω
Z
0
f (uk ) − f 0 (u)|∂i u||ϕ| dx

+
Ω Z
0
≤ sup |f (t)| |∂i uk − ∂i u||ϕ| dx
t∈R Ω
Z
0
f (uk ) − f 0 (u)|∂i u||ϕ| dx → 0

+
Ω
as k → ∞ by the Dominated Convergence Theorem. Hence we can pass to the limit in (2.9)
and we obtain
Z Z
(f ◦ u)∂i ϕ dx = − f 0 (u)∂i u ϕ dx for all ϕ ∈ Cc∞ (Ω).
Ω Ω
This identity shows that ∂i (f ◦ u) = f 0 (u)∂i u in the weak sense. Finally, ∂i (f ◦ u) ∈ Lp (Ω)
by boundedness of f 0 and since ∂i u ∈ Lp (Ω).
Lemma 2.12 (Stampacchia). Let Ω ⊂ Rn be an open set, 1 ≤ p < +∞. Let u ∈ W 1,p (Ω).
Then u+ , u− , |u| ∈ W 1,p (Ω), where
u+ := max{u, 0}, u− := min{u, 0}, (2.10)
with ( (
+ Du(x) if u(x) > 0, + 0 if u(x) ≥ 0,
Du (x) = Du − (x) =
0 if u(x) ≤ 0, Du(x) if u(x) < 0.
It follows that Du = 0 almost everywhere on the set {x ∈ Ω : u(x) = c}, where c ∈ R is any
given constant.
Proof (sketch). Consider the functions
(√
t2 + ε2 − ε if t > 0,
fε (t) :=
0 if t ≤ 0.
22
Notice that fε ∈ C 1 (R) with |fε0 | ≤ 1. By Proposition 2.11 we have fε ◦ u ∈ W 1,p (Ω) and
Z Z
u(x)∂i u(x)
(fε ◦ u)∂i ϕ dx = − p ϕ(x) dx for all ϕ ∈ Cc∞ (Ω).
2
(u(x)) + ε 2
Ω {x∈Ω : u(x)>0}
By letting ε → 0 we obtain by dominated convergence

Z Z
+
u ∂i ϕ dx = − ∂i u ϕ dx for all ϕ ∈ Cc∞ (Ω),
Ω {x∈Ω : u(x)>0}
which proves the formula for the first derivative of u+ . The claims for u− and |u| follows
easily by the relations u− = −(−u)+ and |u| = u+ − u− .
To prove the last claim, consider without loss of generality c = 0. Since u = u+ + u− , we
have Du = Du+ + Du− , and both Du+ and Du− vanish almost everywhere on {u = 0}.
Notice that, if Ω has finite measure, the conclusions of Proposition 2.11 and Lemma 2.12
hold also for p = +∞, as W 1,∞ (Ω) ⊂ W 1,p (Ω) for all p < +∞.
2.1.3 Boundary conditions and traces

In formulating classical problems in the framework of Sobolev spaces (for instance minimum
problems, partial differential equations. . . ), it would be desirable to impose some boundary
conditions on the solutions. However, Sobolev functions are in principle defined only up to
a Lebesgue-negligible set, and typically the boundary of the domain will have zero Lebesgue
measure. We can however introduce a suitable subspace of W 1,p (Ω) in order to deal with
boundary conditions “in a weak sense”.
Definition 2.13. Given Ω ⊂ Rn open, 1 ≤ p < +∞, and k ≥ 1, we denote by
k·kW k,p
W0k,p (Ω) := Cc∞ (Ω)
the closure of Cc∞ (Ω) in W k,p (Ω). We also write
H0k (Ω) := W0k,2 (Ω).
By definition a function u belongs to W0k,p (Ω) if and only if it can be approximated

by smooth functions compactly supported in Ω, i.e. if and only if there exists a sequence
ϕj ∈ Cc∞ (Ω) such that ϕj → u in W k,p (Ω).
Remark 2.14. We prefer not to define the space W0k,∞ (Ω). Indeed, in this case the con-
vergence of a sequence of smooth functions in the norm k · kW k,∞ (Ω) is just the uniform con-
vergence of the function and all its derivatives up to order k. Since the uniform convergence
preserves continuity, the limit of a sequence of functions in Cc∞ (Ω) in the norm k · kW k,∞ (Ω)
is necessarily a function of class C k (Ω).
The idea is to interpret W01,p (Ω) as the space of Sobolev functions u ∈ W 1,p (Ω) vanishing
on the boundary ∂Ω. However, we cannot properly write “u = 0 on ∂Ω”, as Sobolev functions
are always defined up to a set of Lebesgue measure zero. It turns out that we can define a
proper notion ot trace of a Sobolev function on ∂Ω, provided that the set Ω is sufficiently
regular: in the case of Lipschitz boundary, we can assign “boundary values” to a Sobolev
function by means of a trace operator.
23
Theorem 2.15 (Trace). Let Ω ⊂ Rn be open and bounded with Lipschitz boundary, and let
1 ≤ p < +∞. Then there exists a bounded linear operator
Tr : W 1,p (Ω) → Lp (∂Ω)
such that:
(i) Tr u = u|∂Ω if u ∈ W 1,p (Ω) ∩ C 0 (Ω),
(ii) k Tr ukLp (∂Ω) ≤ CkukW 1,p (Ω) for all u ∈ W 1,p (Ω), for a constant C depending only on p
and Ω.
We call Tr u the trace of u. We can then characterize the space W01,p (Ω) as the space of
functions whose trace on ∂Ω vanishes.
Theorem 2.16. Let Ω ⊂ Rn be open and bounded with Lipschitz boundary, 1 ≤ p < +∞.
Then
W01,p (Ω) = u ∈ W 1,p (Ω) : Tr u = 0 on ∂Ω .

(2.11)
Remark 2.17 (Non-homogeneous boundary conditions). Notice that, if we want to impose

a non-homogeneous boundary condition of the form “u = g on ∂Ω”, where g ∈ W 1,p (Ω) is
given, we could work in the space g + W01,p (Ω) := {u ∈ W 1,p (Ω) : u − g ∈ W01,p (Ω)}. This
formulation does not require any regularity of ∂Ω.
Theorem 2.18. For 1 ≤ p < +∞ we have
W 1,p (Rn ) = W01,p (Rn ).
Exercise 7. Let Ω ⊂ Rn be open and bounded with Lipschitz boundary. Prove that for every
u ∈ W 1,p (Ω) and ϕ ∈ C 1 (Ω) the following integration by parts formula holds:
Z Z Z
uDϕ dx = − ϕ Du dx + Tr u ϕν dHn−1
Ω Ω ∂Ω
where ν ∈ Sn−1 is the exterior unit normal to ∂Ω.[6]
Exercise 8. Let Ω ⊂ Rn be open and bounded with Lipschitz boundary. Let u ∈ W 1,p (Ω) and
v ∈ W 1,p (Rn \Ω), 1 ≤ p < +∞. Prove that the function
(
u(x) if x ∈ Ω,
w(x) =
v(x) if x ∈ Rn \Ω
belongs to W 1,p (Rn ) if and only if Tr u = Tr v on ∂Ω.[7]
2.1.4 Embedding theorems and Poincaré inequality

The main result of this section shows that the fact that u belongs to W 1,p (Ω) (that is, the
integrability of |u|p and of |Du|p ) allows to increase the integrability of u (that is, u ∈ Lq (Ω)
for some q > p); if the exponent p is sufficiently large, we also obtain that (a representative
of) the function u is even continuous. In other words, we have an embedding of W 1,p (Ω) into
some other space X.
24
The crucial tools in order to obtain such embeddings are “Sobolev-type inequalities” of
the form kukX ≤ CkukW 1,p (Ω) , which allow to control the norm of a function u in some space
X in terms of the Sobolev norm. In order to give an idea of the type of estimates that hold
true for Sobolev functions, we work for the moment in Ω = Rn and we assume 1 ≤ p < n,
and we try to see whether we can establish an estimate of the form
kukLq (Rn ) ≤ CkDukLp (Rn ) for all u ∈ Cc∞ (Rn ) (2.12)
for some constant C = C(n, p, q) and for some exponent 1 ≤ q < +∞.
Suppose that (2.12) holds; we can find a relation between q, p, n by a scaling argument:
let u ∈ Cc∞ (Rn ), u 6= 0, and let uλ (x) := u(λx) for λ > 0. The two terms on the left-hand side
and on the right-hand side of (2.12) scale differently with respect to λ: indeed by changing
variables
Z 1 Z 1
q q
q −n q −n
kuλ kLq (Rn ) = |u(λx)| dx =λ q |u(y)| dy = λ q kukLq (Rn ) ,
Rn Rn
Z 1 Z 1
p p
p 1− n p 1− n
kDuλ kLp (Rn ) = |λDu(λx)| dx =λ p |Du(y)| dy =λ p kukLq (Rn ) ,
Rn Rn
Hence by plugging uλ into (2.12) we get
−n 1− n
λq kukLq (Rn ) ≤ Cλ p kDukLp (Rn ) .
Since this inequality must be true for all λ > 0, we see that it must be − nq = 1 − np , that is
np
q = p∗ :=
n−p
(notice that p∗ > p). The previous scaling argument gives that (2.12) can possibly be true
only for this particular choice of q; the fact that (2.12) is actually true is the content of the
following theorem.
Theorem 2.19 (Gagliardo-Nirenberg-Sobolev inequality). Assume 1 ≤ p < n. Then there
exists a constant C = C(p, n) depending only on p and n such that
kukLp∗ (Rn ) ≤ CkDukLp (Rn ) for all u ∈ Cc1 (Rn ), (2.13)

np
where p∗ := n−p .
The condition of compact support is really needed here, as the counterexample u ≡ 1

shows. The inequality tells you that a control of the Lp -norm of the gradient of u yields
a gain of summability of u. In the same spirit we have the following Sobolev Embedding
Theorem. In order to state the theorem, we recall that given two Banach spaces X, Y , with
Y ⊂ X, we say that:
• the embedding Y ,→ X is continuous if there is a constant C such that kukX ≤ CkukY
for all u ∈ Y ;
• the embedding Y ,→ X is compact (in symbols Y ⊂⊂ X) if it is continuous and

every sequence (uk )k ⊂ Y bounded in Y (i.e. supk kuk kY < +∞) has a subsequence
converging in X.
25
Theorem 2.20 (Sobolev embedding theorem). Let Ω ⊂ Rn be open and bounded with Lips-
chitz boundary. Then the following hold.
(i) If 1 ≤ p < n, there exists a constant C = C(p, n, Ω) such that for every u ∈ W 1,p (Ω)
np
kukLp∗ (Ω) ≤ CkukW 1,p (Ω) , where p∗ := , (2.14)
n−p
∗
that is, the embedding W 1,p (Ω) ,→ Lp (Ω) is continuous. Moreover for all 1 ≤ r < p∗
the embedding W 1,p (Ω) ,→ Lr (Ω) is compact.
(ii) If p = n, for all 1 ≤ r < +∞ the embedding W 1,p (Ω) ,→ Lr (Ω) is compact.
(iii) If n < p ≤ +∞, there exists a constant C = C(p, n, Ω) such that for every u ∈ W 1,p (Ω)
n
kukC 0,α (Ω) ≤ CkukW 1,p (Ω) , where α = 1 − . (2.15)
p
that is, the embedding W 1,p (Ω) ,→ C 0,α (Ω) is continuous.
The case (iii) is usually referred to as Morrey Theorem and says that, if u ∈ W 1,p (Ω) with
p > n, then a representative of u is Hölder continuous in Ω. We recall here that a function u
is Hölder continuous in Ω with exponent α ∈ (0, 1) if
|u(x) − u(y)| ≤ C|x − y|α for all x, y ∈ Ω. (2.16)
The space of Hölder continuous functions on Ω is denoted by C 0,α (Ω), with norm
|u(x) − u(y)|
kukC 0,α (Ω) := kukL∞ (Ω) + sup (2.17)
x,y∈Ω |x − y|α
x6=y
Remark 2.21 (Rellich-Kondrachov). The compactness of the embedding W 1,p (Ω) ,→ Lr (Ω)
in Theorem 2.20(i) (1 ≤ p < n, 1 ≤ r < p∗ ) is usually referred to as Rellich-Kondrachov
Compactness Theorem. Notice that this holds in particular for r = p.
Since by Ascoli-Arzelà Theorem the embedding C 0,α (Ω) ,→ C 0 (Ω) is compact, we also have
W 1,p (Ω) ⊂⊂ C 0 (Ω) for p > n.
In conclusion, in all cases if Ω ⊂ Rn is open, bounded with Lipschitz boundary we have

the compact embedding
W 1,p (Ω) ⊂⊂ Lp (Ω) for all 1 ≤ p ≤ +∞, (2.18)
that is, if (uk )k ⊂ W 1,p (Ω) is a sequence with supk kuk kW 1,p (Ω) < +∞, then there exists a
subsequence (ukj )j and a function u ∈ Lp (Ω) such that ukj → u in Lp (Ω).
Example 2.22. The assumption of Lipschitz boundary in the previous theorems is essential.
As a counterexample, consider the domain Ω = {(x, y) ∈ R2 : 0 < x < 1, 0 < y < x2 } ⊂ R2 ,
and the function u(x, y) = xα . Then u ∈ Lq (Ω) if and only if
Z Z 1
q
|u(x, y)| dx dy = xqα+2 dx < +∞ if and only if qα + 2 > −1.
Ω 0
26
Similarly, one finds that u ∈ W 1,2 (Ω) if and only if 2(α − 1) + 2 > −1, that is α > − 12 .
Therefore if q > 6 we can find α > − 12 such that qα + 2 ≤S−1, that is u ∈ W 1,2 (Ω)\Lq (Ω).
As a counterexample in general dimension, let Ω = ∞ i=1 Bri (xi ), where the centers xi
are chosen such that xi → x̄, and thePradii ri > 0 are such that all the balls are disjoint.
Let u : Ω → R be defined by u(x) = i ai χBri (xi ) (x) for suitable constants ai ∈ R. Then
DuP≡ 0 in Ω, and u ∈ W 1,p (Ω) if and only if i |ai |p rin < +∞, while u ∈ Lr (Ω) if and only
P
if i |ai |r rin < +∞. If p < r, we can find a sequence of coefficients (ai )i such that the first
condition is satisfied but not the second, so that u ∈ W 1,p (Ω)\Lr (Ω).
We conclude this subsection by recalling the Poincaré inequality, which allows to control
the Lp -norm of a function u in terms of the Lp -norm of its gradient.
Theorem 2.23 (Poincaré-Wirtinger inequality). Let Ω ⊂ Rn be open, bounded and connected,
with Lipschitz boundary, and let 1 ≤ p ≤ +∞. Then for all u ∈ W 1,p (Ω) we have
Z Z Z
u(x) − uΩ p dx ≤ Cp,Ω 1
|Du(x)|p dx,

where uΩ := u(x) dx (2.19)
Ω Ω |Ω| Ω
and Cp,Ω is a constant depending only on p, n, and Ω.
Proof. By contradiction, assume that for every k ∈ N there exists a function uk ∈ W 1,p (Ω)
such that
kuk − (uk )Ω kLp (Ω) > kkDuk kLp (Ω) .
By translation and normalization we can assume without loss of generality that (uk )Ω = 0 and
kuk kLp (Ω) = 1 for all k. Hence the contradiction assumption implies that kDuk kLp (Ω) ≤ k1 .
In particular, the sequence (uk )k is bounded in W 1,p (Ω) and by Remark 2.21 there is a
subsequence (ukj )j and a function u ∈ Lp (Ω) such that ukj → u in Lp (Ω). In particular, u
satisfies
uΩ = 0, kukLp (Ω) = 1.
However, for all ϕ ∈ Cc∞ (Ω) and for all i = 1, . . . , n we have
Z Z Z
u ∂i ϕ dx = lim ukj ∂i ϕ dx = − lim ∂i ukj ϕ dx = 0
Ω j→∞ Ω j→∞ Ω
since kDuk kLp (Ω) ≤ k1 → 0. Hence Du = 0 almost everywhere in Ω and, as Ω is connected,

this yields that u is constant in Ω, and, in turn, that u = 0 almost everywhere in Ω (as
uΩ = 0). This contradicts the fact that kukLp (Ω) = 1.
Theorem 2.24 (Poincaré inequality in W01,p ). Let Ω ⊂ Rn be open and bounded, and let
1 ≤ p < n. Then there exists a constant C = C(p, q, n, Ω) such that for every u ∈ W01,p (Ω)
Z 1 Z 1
q p
u(x)q dx p
for all 1 ≤ q ≤ p∗ .

≤C |Du(x)| dx , (2.20)
Ω Ω
Comparing this statement with Theorem 2.20(i), we notice two important facts: (1) when
dealing with the spaces W01,p (Ω), the regularity assumption on ∂Ω can be removed; (2) in the
estimate (2.20) only the gradient of u appears on the right-hand side of the inequality, while
in (2.14) there is the full W 1,p -norm of u: in view of this estimate, on W01,p (Ω) the norm
kDukLp (Ω) is equivalent to kukW 1,p (Ω) (if Ω is bounded).
27
Proof of Theorem 2.24. For u ∈ W01,p (Ω) there exists a sequence uk ∈ Cc∞ (Ω) such that
uk → u in W 1,p (Ω). Extend each uk as a C ∞ -function on Rn by setting uk = 0 on Rn \Ω.
Then by applying the Gagliardo-Nirenberg-Sobolev inequality (Theorem 2.19) we get
kuk kLp∗ (Ω) = kuk kLp∗ (Rn ) ≤ CkDuk kLp (Rn ) = CkDuk kLp (Ω) .
By passing to the limit as k → +∞ we obtain the inequality in the statement. for q = p∗ .1

For q < p∗ , simply use that fact that as Ω is bounded we have kukLq (Ω) ≤ CkukLp∗ (Ω) .
Exercise 9 (Poincaré inequality on balls). Let 1 ≤ p ≤ +∞. Prove that there is a constant
C = C(p, n) such that for every ball Br (x) ⊂ Rn we have
ku − (u)Br (x) kLp (Br (x)) ≤ CrkDukLp (Br (x))
for all u ∈ W 1,p (Br (x)).[8]
2.1.5 Duals and weak convergence

Along this subsection Ω ⊂ Rn will be an open set. In order to introduce the weak topology
on W 1,p (Ω), we first need to identify the dual of this space. Let us first recall that, for
1 ≤ p < +∞, by Riesz Representation Theorem the dual of Lp (Ω) can be identified with
0
Lp (Ω), where
1 1
+ =1
p p0
(p0 = ∞ if p = 1). This means that every bounded linear functional L : Lp (Ω) → R can be
0
uniquely represented by a function v ∈ Lp (Ω), in the sense that
Z
L(u) = uv dx for all u ∈ Lp (Ω)
Ω
(and every functional of this form is a bounded linear functional on Lp (Ω)). In particular,
Lp (Ω) is reflexive for 1 < p < +∞. With this characterization of the dual (Lp (Ω))0 we can
give the following definition.
Definition 2.25 (Weak convergence in Lp ). Let (uk )k ⊂ Lp (Ω), u ∈ Lp (Ω). We say that:
• if 1 ≤ p < +∞, uk converges to u weakly in Lp (Ω) (in symbols uk * u) if
Z Z
0
lim uk v dx = uv dx for all v ∈ Lp (Ω); (2.21)
k→∞ Ω Ω
∗
• if p = +∞, uk converges to u weakly* in L∞ (Ω) (in symbols uk * u) if
Z Z
lim uk v dx = uv dx for all v ∈ L1 (Ω). (2.22)
k→∞ Ω Ω
Strong convergence always implies weak convergence. We recall a few useful properties of
weak convergence (for the proof, see any book on Functional Analysis, for instance [1]).
1
Notice that, in order to show that kuk kLp∗ (Ω) → kukLp∗ (Ω) , one can use again the Gagliardo-Nirenberg-
Sobolev inequality kuk −uj kLp∗ (Ω) ≤ CkDuk −Duj kLp (Ω) to obtain that (uk )k is a Cauchy sequence in Lp∗ (Ω),
and therefore it converges to u in Lp∗ (Ω).
28
Proposition 2.26. Let 1 ≤ p ≤ +∞. Then the following properties hold (replacing “weak
convergence” by “weak* convergence” in the case p = +∞):
(i) any weakly convergent sequence is bounded, i.e.
uk * u weakly in Lp (Ω) =⇒ sup kuk kLp (Ω) < ∞; (2.23)

k
(ii) the norm is lower semicontinuous with respect to weak convergence, i.e.
uk * u weakly in Lp (Ω) =⇒ kukLp (Ω) ≤ lim inf kuk kLp (Ω) ; (2.24)
k→∞
0
(iii) if uk * u weakly in Lp (Ω) and vj → v strongly in Lp (Ω), then
Z Z
lim uk vk dx = uv dx ; (2.25)
k→∞ Ω Ω
(iv) for 1 < p < +∞, if uk * u in Lp (Ω) and kuk kLp (Ω) → kukLp (Ω) , then uk → u strongly
in Lp (Ω).
Exercise 10. Prove the implication (iv) in Proposition 2.26 in the case p = 2.[9]
We now turn to the duals of the Sobolev spaces W 1,p (Ω). The idea is to regard W 1,p (Ω)
as a closed subspace of the space Lp (Ω; Rn+1 ) of functions f = (f1 , . . . , fn+1 ) : Ω → Rn+1
such that fi ∈ Lp (Ω) for all i, via the map
T : u ∈ W 1,p (Ω) 7−→ (u, ∂1 u, . . . , ∂n u) ∈ Lp (Ω; Rn+1 ).
Assume that 1 ≤ p < +∞. With the previous identification, any bounded linear functional
L ∈ (W 1,p (Ω))0 can be extended by Hahn-Banach Theorem to a bounded linear functional
on Lp (Ω; Rn+1 ), for which we can use Riesz Representation Theorem as before: namely, if
0
L ∈ (W 1,p (Ω))0 then there exist f0 , f1 , . . . fn ∈ Lp (Ω), with p1 + p10 = 1, such that
Z n Z
X
L(u) = uf0 dx + ∂i u fi dx for all u ∈ W 1,p (Ω) (2.26)
Ω i=1 Ω
(however, notice that this representation is not unique, as the extension of L will not be unique
in general). Conversely, every functional of the form (2.26) is an element of (W 1,p (Ω))0 . In
view of this result, we can characterize the weak convergence in W 1,p (Ω).
Definition 2.27 (Weak convergence in W 1,p ). Let (uk )k ⊂ W 1,p (Ω), u ∈ W 1,p (Ω). Then:
• if 1 ≤ p < +∞, uk converges to u weakly in W 1,p (Ω) (in symbols uk * u) if and only
if uk * u weakly in Lp (Ω) and ∂i uk * ∂i u weakly in Lp (Ω) for all i = 1, . . . , n;
∗
• if p = +∞, uk converges to u weakly* in W 1,∞ (Ω) (in symbols uk * u) if and only if
∗ ∗
uk * u weakly* in L∞ (Ω) and ∂i uk * ∂i u weakly* in L∞ (Ω) for all i = 1, . . . , n.2
2
This definition might be a bit misleading, since we speak of “weak*-convergence” even if W 1,∞ is not the
dual of some space.
29
The properties in Proposition 2.26 remain valid also for the weak convergence in W 1,p .
We conclude this overview by discussing the compactness properties of W 1,p (Ω) with respect
to weak convergence, which follow from abstract results in Functional Analysis. Clearly the
following statement holds also by replacing W 1,p (Ω) by Lp (Ω).
Theorem 2.28 (Compactness). Assume that (uk )k ⊂ W 1,p (Ω) is a bounded sequence in
W 1,p (Ω), that is supk kuk kW 1,p (Ω) < ∞.
• if 1 < p < +∞, then there exist a subsequence (ukj )j and u ∈ W 1,p (Ω) such that ukj * u
weakly in W 1,p (Ω).
∗
• if p = +∞, then there exist a subsequence (ukj )j and u ∈ W 1,∞ (Ω) such that ukj * u
weakly* in W 1,∞ (Ω).
Proof. Case 1 < p < ∞. Recall that a bounded sequence in a reflexive Banach space is
sequentially weakly compact, that is there exists a weakly convergent subsequence. The
conclusion follows by applying this result to X = W 1,p (Ω), which is reflexive for 1 < p < ∞.
Case p = ∞. Recall that by Banach-Alaoglu Theorem, if X is a Banach space and X 0
its dual, the unit ball BX 0 = {x ∈ X 0 : kxkX 0 ≤ 1} in X 0 is compact in the weak* topology.
Furthermore, if X is separable then the weak* topology is metrizable on BX 0 .
From the previous two properties, it follows that if X is separable and (xk )k ⊂ X 0 is
bounded in X 0 , then there is a weakly* convergent subsequence. The conclusion follows by
applying this result to X 0 = L∞ (Ω), which is the dual of the separable space L1 (Ω).
Remark 2.29. The previous theorem fails for p = 1, as W 1,1 (Ω) is not reflexive. As a
counterexample, consider the domain Ω = (−1, 1) ⊂ R and the sequence

−1 if x ∈ (−1, −1/k),

uk (x) = kx if x ∈ (−1/k, 1/k),

1 if x ∈ (1/k, 1).

2.2 The direct method of the Calculus of Variations

Let (X , τ ) be a topological space. We consider the minimization of a functional
F : X → R = R ∪ {±∞},
that is the problem of finding a point x0 ∈ X such that F(x0 ) ≤ F(x) for all x ∈ X . In this
case we write F(x0 ) = inf x∈X F(x) = minx∈X F(x).
The reason to consider functionals possibly taking the value +∞ is that typically we are
interested in minimizing a functional F over a subspace K of X ; this is then equivalent to
minimize the extended real valued functional F = F + χK , where
(
0 if x ∈ K,
χK (x) =
+∞ if x ∈ X \K
is the indicator function of K. The main idea of direct methods is to choose a suitable
topology on X which guarantees compactness of the space and lower semicontinuity of the
functional F.
30
2.2.1 Lower semicontinuity
We start by recalling the notion of semicontinuity and some related properties.
Definition 2.30 (Lower semicontinuity). A functional F : X → R is lower semicontinuous

in X if
F(x) ≤ lim inf F(y) for all x ∈ X , (2.27)
y→x
where, denoting by N (x) the family of all open neighbourhoods of the point x,
lim inf F(y) := sup inf F(y). (2.28)

y→x U ∈N (x) y∈U \{x}
Equivalently, F is l.s.c. if for every t < F(x) there exists a neighbourhood U ∈ N (x) such
that F(y) > t for all y ∈ U .
Definition 2.31 (Sequential lower semicontinuity). The functional F is sequentially lower

semicontinuous on X if for every x ∈ X
F(x) ≤ lim inf F(xn ) for all xn → x. (2.29)

n→∞
Exercise 11. A set A ⊂ X is sequentially open if for every x ∈ A and for every sequence
xn → x, one has that xn ∈ A for all n sufficiently large. A set C ⊂ X is sequentially closed
if for every sequence xn ∈ C with xn → x, one has x ∈ C. Consider the functions
( (
1 if x ∈ A, 0 if x ∈ C,
1A (x) := χC (x) := (2.30)
0 if x ∈
/ A, +∞ if x ∈ / C.
Prove that:
A is (sequentially) open ⇐⇒ 1A is (sequentially) l.s.c.

C is (sequentially) closed ⇐⇒ χC is (sequentially) l.s.c.
The two notions of lower semicontinuity coincide if the topological space satisfies the first
axiom of countability (each point has a countable base of neighbourhoods); this is the case,
for instance, in metric spaces. However for more general topological spaces the sequential
lower semicontinuity is a weaker condition: if F is l.s.c., then it is also sequentially l.s.c., but
the converse implication is not true, as the following example shows.
Example 2.32. We show with a counterexample that the sequential lower semicontinuity
does not imply the lower semicontinuity in the topological sense. In view of the previous
exercise, it is sufficient to construct a set which is sequentially closed but not closed (this can
always be done in every infinite dimensional Hilbert space endowed with the weak topology).
Let X = H01 (Ω) endowed with the weak topology of H 1 (Ω), with Ω ⊂ Rd open and bounded,
and let
C = {u ∈ H01 (Ω) : kukL2 (Ω) ≥ 1}.
Given a sequence (un )n ⊂ C such that un * u in H01 (Ω), we have by Rellich Theorem that
un → u in L2 (Ω) and hence kukL2 (Ω) = limn→∞ kun kL2 (Ω) ≥ 1, so that u ∈ C. Therefore C
is sequentially weakly closed.
31
However, C is not closed: indeed we can show that the function u ≡ 0 belongs to the
closure of C in the weak topology. To prove this, we show that every weak neighbourhood of 0
contains points of C. A neighbourhood basis of 0 in the weak topology is given by the sets of
the form n o
U = v ∈ H01 (Ω) : |hfi , vi| ≤ ε for all i = 1, . . . , N
for fixed ε > 0, f1 , . . . , fN ∈ (H 1 (Ω))0 , N ∈ N. Now since the space has infinite dimension
we can find v0 6= 0 such that hfi , v0 i = 0 for all i = 1, . . . , N : then tv0 ∈ U for all t ∈ R, and
ktv0 kL2 (Ω) = tkv0 kL2 (Ω) ≥ 1 for t sufficiently large. Hence tvo ∈ C ∩ U 6= ∅.
Exercise 12 (Equivalent characterizations of lower semicontinuity). Let F : X → R. The

following properties are equivalent:
(i) F is (sequentially) l.s.c.;
(ii) the sublevel set {x ∈ X : F(x) ≤ t} is (sequentially) closed for all t ∈ R;
(iii) the epigraph epi(F) := {(x, t) ∈ X × R : F(x) ≤ t} is (sequentially) closed.[10]
Exercise 13. Let (Fi )i∈I be a family of (sequentially) l.s.c. functionals on X . Prove that
F(x) := supi Fi (x) is (seq.) l.s.c. on X .[11]
2.2.2 Compactness and existence of minimizers

In addition to lower semicontinuity, the second main ingredient to establish the existence of
minimizers for functionals F : X → R is a compactness condition on the space X . Also in
this case we give two versions, a topological condition and a sequential condition.
Definition 2.33 (Compactness). A subset K of a topological space X is compact if every
open covering of K has a finite subcovering.
Definition 2.34 (Sequential compactness). A subset K of a topological space X is sequen-
tially compact if every sequence (xn )n ⊂ K has a convergent subsequence xnk → x ∈ K.
Also in this case, the two notions agree in metric spaces, but are in general different, as
the following examples from [8] show.
Example 2.35. Let X be the set of ordinals less than the first uncountable ordinal, and let
B be the collection of sets of the form {x : x < a}, {x : a < x < b}, and {x : a < x}. Then
B is a base for a topology on X , and X is sequentially compact but not compact.
Example 2.36. Let X = {f : [0, 1] → [−1, 1]} with the topology of pointwise convergence.
The space X can be identified with [−1, 1][0,1] with the product topology, which is compact as
product of compact spaces. However, X is not sequentially compact: for instance, the sequence
fn (x) = sin(nx) does not have any convergent subsequence.
The core of the direct method of the Calculus of Variations is the following Weierstrass
Theorem. We present a topological version and a sequential version of the theorem.
Theorem 2.37 (Weierstrass, topological version). Let X be a compact topological space, and
let F : X → R be a lower semicontinuous functional. Then there exists x ∈ X such that
F(x) = min F.
X
32
Proof. The compactness of X implies the following property: if C is any collection of closed
sets with the finite intersection property (i.e. any finite subcollection of C has a nonempty
intersection), then C has nonempty intersection.
Let I := inf X F and assume without loss of generality that I < +∞. For I < t < +∞,
the sublevel sets
Ct := x ∈ X : F(x) ≤ t
are nonempty (as t > I), closed (as F is l.s.c.), and T satisfy the finite intersection property.
Then by compactness of X the intersection C∞ := t>I Ct is not empty. Let then x0 ∈ C∞ :
as x0 ∈ Ct for all t, we have F(x0 ) ≤ t for all t > I, and therefore F(x) ≤ inf X F.
Theorem 2.38 (Weierstrass, sequential version). Let X be a sequentially compact topological

space, and let F : X → R be a sequentially lower semicontinuous functional. Then there
exists x ∈ X such that
F(x) = min F.
X
Proof. Assume without loss of generality that inf X F < +∞. Let {xn }n ⊂ X be a minimizing
sequence, that is limn→∞ F(xn ) = inf X F (which always exists). By sequential compactness,
we can extract a converging subsequence {xnk }k , with x = limk→∞ xnk ∈ X . By sequential
lower semicontinuity,
F(x) ≤ lim inf F(xnk ) = inf F,
k→∞ X
which concludes the proof.
It is clear from the proofs that the previous theorems hold under the weaker assumption
that there exists t0 > inf X F such that the sublevel set {F ≤ t0 } is (sequentially) compact.
Definition 2.39 (Coercivity). A functional F : X → R is (sequentially) coercive if for all
t ∈ R the sublevel set {x ∈ X : F(x) ≤ t} is (sequentially) relatively compact in X (that is,
its closure is compact).
Theorem 2.40 (Tonelli). Let F : X → R be (sequentially) coercive and (sequentially) lower
semicontinuous. Then there exists x ∈ X such that
F(x) = min F.
X
Remark 2.41. Notice that, in addition to the existence of a minimizer, Theorem 2.38 shows
that any minimizing sequence for F has a convergent subsequence, and that any limit point
of a subsequence of a minimizing sequence is a minimizer of F.
2.2.3 The direct method in weak topologies

The two requirements of Weierstrass Theorem pull in different directions: in order to get
compactness one should in general weaken the topology, while it is easier to get lower semi-
continuity if the topology is stronger. A good compromise is often the weak topology of a
Banach space.
Let X be a Banach space. Assume that F is a weakly l.s.c. functional: then, from the
fact that any sequence converging with respect to the strong topology also converges in the
weak topology we deduce that
xn → x strongly in X =⇒ xn * x weakly in X =⇒ F(x) ≤ lim inf F(xn ),

n→∞
33
where the last implication follows by the weak l.s.c. of F. This shows that the weak lower
semicontinuity of a functional implies its strong lower semicontinuity:
F is weakly l.s.c. =⇒ F is strongly l.s.c.
(this property is true more in general for two comparable topologies: if a functional is l.s.c.
with respect to a given topology, then it is also l.s.c. with respect to any stronger topol-
ogy). The converse of this implication is in general not true; however, under the additional
assumption of convexity of the functional, the two properties are actually equivalent.
Definition 2.42. Let X be a Banach space. A map F : X → R is convex if
F(tx + (1 − t)y) ≤ tF(x) + (1 − t)F(y) (2.31)
for all x, y ∈ X such that F(x) < +∞, F(y) < +∞, for all t ∈ (0, 1). The function F is
strictly convex if (2.31) holds with strict inequality if x 6= y.
Exercise 14. Let F : X → R be a strictly convex function. Prove that F has at most one
minimum point in X .[12]
We remark that F is convex if and only if the epigraph epi(F) ⊂ X × R (see Exercise 12)
is convex. We then obtain the following proposition.
Proposition 2.43. Let X be a Banach space, and let F : X → R be convex. Then
F is weakly l.s.c. ⇐⇒ F is strongly l.s.c.
Proof. The proof follows from the fact that, as a consequence of Hahn-Banach Theorem, a
convex set is strongly closed if and only if it is weakly closed (see [1, Theorem 3.7]). Hence,
being epi(F) a convex set, we have
F strongly l.s.c. ⇐⇒ epi(F) strongly closed
⇐⇒ epi(F) weakly closed
⇐⇒ F is weakly l.s.c.
which proves the proposition.
As a consequence of the previous proposition we have the following existence theorem.

Theorem 2.44. Let X be a reflexive Banach space, and let F : X → R be convex, strongly
l.s.c., and such that
lim F(x) = +∞. (2.32)
kxk→∞
Then there exists x ∈ X such that

F(x) = min F.
X
Proof. By assumption, for every t ∈ R there exists R > 0 such that F(x) > t for every x with
kxk > R: in particular
{F ≤ t} ⊂ B R .
Since B R is compact in the weak topology (by reflexivity), {F ≤ t} is also weakly compact.
Hence F is weakly coercive.
By Proposition 2.43 F is also weakly lower semicontinuous. Hence F fulfills the assump-
tion of Theorem 2.40.
34
0
Example 2.45 (p-laplacian). Let Ω ⊂ Rn be open and bounded, 1 < p < +∞. Let g ∈ Lp (Ω).
We consider the minimum problem
Z Z
p 1,p
min F(u) := |Du| dx − p gu dx : u ∈ W0 (Ω) . (2.33)
Ω Ω
Notice that F(u) is well-defined for u ∈ W 1,p (Ω) by Hölder inequality. We aim at proving the
existence of a minimizer by means of Theorem 2.44.
Since 1 < p < +∞, the space X := W01,p (Ω) is reflexive. Let us check the coercivity
condition (2.32). By Hölder inequality, and recalling that kDukLp (Ω) is an equivalent norm
in W01,p (Ω) by Poincaré inequality, we have
Z
p
F(u) = kDukLp (Ω) − p gu dx
Ω
p
≥ kDukLp (Ω) − pkgkLp0 (Ω) kukLp (Ω)
≥ c1 kukpW 1,p (Ω) − pcg kukW 1,p (Ω) −→ +∞ as kukW 1,p (Ω) → +∞,
that is, F is weakly coercive in W01,p (Ω).

We need to check weak lower semicontinuity: if uk * u weakly in W01,p (Ω), then
kDukpLp (Ω) ≤ lim inf kDuk kpLp (Ω)

k→∞
(since the norm is weakly lower semicontinuous), and

Z Z
lim guk dx = gu dx
k→∞ Ω Ω
0
(since g ∈ Lp (Ω) and uk * u weakly in Lp (Ω)). Hence F is sequentially weakly lower
semicontinuous. By Theorem 2.44 we therefore conclude that there exists ū ∈ W01,p (Ω) such
that F(ū) = minu∈W 1,p (Ω) F(u).
0
We will see later that minimizers of (2.33) are weak solutions to the p-laplacian equation
(
−div |Du|p−2 Du = g in Ω,

(2.34)
u=0 on ∂Ω.
2.3 Minimum problems for integral functionals

We would like to apply the abstract results of the previous section to integral functionals of
the form Z
F(u) := f (x, u(x), Du(x)) dx
Ω
and prove the existence of minimizers for such functionals under suitable assumptions on f .
As a main reference for the contents of this and the following sections, see [2, Chapter 3].
35
2.3.1 Functionals defined on Lp
Let Ω ⊂ Rn be an open set, 1 ≤ p < +∞, m ≥ 1. We first consider functionals depending
only on the function u, but not on Du: we define the functional
Z
G(u) := g(x, u(x)) dx for u ∈ Lp (Ω; Rm ) (2.35)
Ω
where g : Ω × Rm → R satisfies the following assumptions:

(g1) g is measurable with respect to L(Ω) ⊗ B(Rm ), where L(Ω) denotes the Lebesgue σ-
algebra on Ω and B(Rm ) the Borel σ-algebra on Rm ;
(g2) the map s 7→ g(x, s) is lower semicontinuous on Rm , for almost every fixed x ∈ Ω;
(g3) g(x, s) ≥ −a(x) − b|s|p for some a ∈ L1 (Ω), b ∈ R.

Proposition 2.46. Under the assumptions (g1)–(g3), the functional G in (2.35) is well-
defined on Lp (Ω; Rm ) and is lower semicontinuous with respect to the strong topology of
Lp (Ω; Rm ). Assuming in addition that
(g4) the map s 7→ g(x, s) is convex on Rm , for almost every fixed x ∈ Ω,
then G is convex and lower semicontinuous with respect to the weak topology of Lp (Ω; Rm ).
Proof. The map x 7→ (x, u(x)) is measurable from (Ω, L(Ω)) to (Ω × Rm , L(Ω) ⊗ B(Rm ));
then by the measurability assumption on g the composition x 7→ g(x, u(x)) is measurable. By
decomposing g = g + − g − in positive and negative part, we observe that 0 ≤ g − (x, u(x)) ≤
a(x) + b|u(x)|p ∈ L1 (Ω). Then the integral in (2.35) is well-defined, possibly taking the value
+∞, as Z Z Z
g(x, u(x)) dx := g + (x, u(x)) dx − g − (x, u(x)) dx
Ω Ω Ω
(as the second integral on the right-hand side is finite).
We next check the lower semicontinuity of G in the strong topology (we can use the
sequential definition, as the strong topology is metrizable). Let uk → u strongly in Lp (Ω; Rm );
we shall prove that
G(u) ≤ lim inf G(uk ).
k→∞
Extract a subsequence (ukj )j such that
lim inf G(uk ) = lim G(ukj ).

k→∞ j→∞
In this way, we are allowed to pass to any further subsequence, without altering the limit
value of G(ukj ) (without this step, passing to a subsequence would have possibly increased
the lim inf of G(uk )). We can also assume that ukj → u almost everywhere in Ω (by taking a
further subsequence). Then by (g2)
g(x, u(x)) ≤ lim inf g(x, ukj (x)) for almost every x ∈ Ω
j→∞
and also
g(x, u(x)) + a(x) + b|u(x)|p ≤ lim inf g(x, ukj (x)) + a(x) + b|ukj (x)|p

j→∞
36
By applying Fatou’s Lemma (we have nonnegative functions)
Z Z Z Z
p p
G(u) + a(x) dx + b |u(x)| dx ≤ lim inf G(ukj ) + a(x) dx + b |ukj | dx ,
Ω Ω j→∞ Ω Ω
and since ukj → u strongly in Lp (Ω) we conclude that
G(u) ≤ lim inf G(ukj ) = lim inf G(uk ),

j→∞ k→∞
which proves the strong lower semicontinuity of G.

Finally, if g(x, ·) is convex it is easy to check that also G is convex. Then the weak lower
semicontinuity of G follows by Proposition 2.43.
If it is possible to apply the previous result to both g and −g, then the functional G will
be continuous.
Definition 2.47. Let g : Ω × Rm → R. We say that g satisfies the Carathéodory condition

if
(g1)0 the map x 7→ g(x, s) is measurable with respect to L(Ω), for every fixed s ∈ Rm ;
(g2)0 the map s 7→ g(x, s) is continuous on Rm , for almost every fixed x ∈ Ω.
Proposition 2.48. Assume that g satisfies the Carathéodory condition and
|g(x, s)| ≤ a(x) + b|s|p for some a ∈ L1 (Ω), b ∈ R.
Then the functional G defined in (2.35) is continuous with respect to the strong topology.
Proof. We need to prove that g satisfies the measurability condition (g1): once this is done,
it would be possible to apply Proposition 2.46 to both g and −g and to obtain the continuity
of G.
In order to prove measurability, fix k ∈ N and subdivide Rm into countably many cubes
(k) (k) (k) (k) S (k)
(Qi )i∈N , where each Qi is a cube of side k1 , Qi ∩ Qj 6= ∅ for i 6= j, and i Qi = Rm .
(k) (k)
For all k and i fix also a point si ∈ Qi . Then consider the sequence of functions
∞
(k)
X
gk (x, s) := g(x, si )1Q(k) (s).
i
i=1
Each function gk is L(Ω) ⊗ B(Rm )-measurable, as it is a countable sum of L(Ω) ⊗ B(Rm )-

measurable functions. As gk (x, s) → g(x, s) as k → ∞ for every x ∈ Ω and s ∈ Rm by
continuity of g, the measurability of g follows. Notice that this argument does not work if
we assume only lower semicontinuity of s 7→ g(x, s): this is why in that case we need the
assumption (g1), that is measurability of g in both variables.
Given a function g satisfying the Carathéodory condition, under which assumptions is

the composition v(x) = g(x, u(x)) a function of Lr (Ω) for every u ∈ Lp (Ω)? Under which
assumptions is the map u ∈ Lp (Ω) 7→ v ∈ Lr (Ω) continuous?
37
Theorem 2.49 (Nemitski operators). Let p, r ≥ 1 and let g : Ω × Rm → R satisfy the
Carathéodory condition and the bound
p
|g(x, s)| ≤ a(x) + b|s| r
for some a ∈ Lr (Ω), b ≥ 0. Then the Nemitski operator T : u 7→ g(·, u(·)) is continuous from
Lp (Ω) to Lr (Ω).
Proof. The function v(x) = T u(x) is measurable and

Z Z Z Z
|v(x)|r dx = |g(x, u(x))|r dx ≤ Cr |a(x)|r dx + Cr br |u(x)|p dx < ∞,
Ω Ω Ω Ω
hence T : Lp (Ω) → Lr (Ω) is well-defined. To check the continuity of T , fix u0 ∈ Lp (Ω) and
observe that Z
kT u − T u0 krLr (Ω) = |g(x, u(x)) − g(x, u0 (x))|r dx. (2.36)
Ω
Let h(x, s) = |g(x, s) − g(x, u0(x))|r
then h is a Carathéodory function satisfying the bound
(g30 ), hence by Proposition 2.48 the right-hand side of (2.36) is continuous with respect
to strong Lp -convergence. Therefore we obtain that kT u − T u0 kLr (Ω) → 0 as u → u0 in
Lp (Ω).
2.3.2 Functionals defined on W1,p

Let Ω ⊂ Rn be an open set, 1 ≤ p < +∞. We now consider functionals defined on the Sobolev
space W 1,p (Ω), depending on Du: we define the functional
Z
F(u) := f (x, Du(x)) dx for u ∈ W 1,p (Ω) (2.37)
Ω
where f : Ω × Rn → R satisfies the following assumptions:
(f1) f is measurable with respect to L(Ω) ⊗ B(Rn );
(f2) the map ξ 7→ f (x, ξ) is lower semicontinuous on Rn , for almost every x ∈ Ω;
(f3) f (x, ξ) ≥ −a(x) − b|ξ|p for some a ∈ L1 (Ω), b ∈ R.
Proposition 2.50. Under the assumptions (f1)–(f3), the functional F in (2.37) is well-
defined on W 1,p (Ω) and is lower semicontinuous with respect to the strong topology of W 1,p (Ω).
Assuming in addition that
(f4) the map ξ 7→ f (x, ξ) is convex on Rn , for almost every x ∈ Ω,
then F is convex and lower semicontinuous with respect to the weak topology of W 1,p (Ω).
Proof. The functional G : Lp (Ω; Rn ) → R defined by G(w) := Ω f (x, w(x)) dx satisfies the
R
assumptions of Proposition 2.46. Then the conclusion follows since F(u) = G(Du) and if
uk → u in W 1,p (Ω) then Duk → Du in Lp (Ω).
38
We now have some sufficient conditions which guarantee that integral functionals of the
form (2.35) of (2.37) are (strongly or weakly) lower semicontinuous. We next discuss how to
apply the direct method to prove the existence of minimum points for problems of the form
Z Z
min f (x, Du(x)) dx + g(x, u(x)) dx . (2.38)
u∈W 1,p (Ω) Ω Ω
We assume that f : Ω × Rn → R satisfies the following conditions:
(f1) f is measurable with respect to L(Ω) ⊗ B(Rn );
(f2) the map ξ 7→ f (x, ξ) is lower semicontinuous on Rn , for almost every x ∈ Ω;
(f3)0 f (x, ξ) ≥ a0 (x) + b0 |ξ|p for some a0 ∈ L1 (Ω), b0 > 0;
(f4) the map ξ 7→ f (x, ξ) is convex on Rn , for almost every x ∈ Ω.
We assume that g : Ω × R → R satisfies the following conditions:
(g1) g is measurable with respect to L(Ω) ⊗ B(R);
(g2) the map s 7→ g(x, s) is lower semicontinuous on R, for almost every fixed x ∈ Ω;
(g3)0 g(x, s) ≥ a1 (x) + b1 |s|p for some a1 ∈ L1 (Ω), b1 > 0.
Notice that we replaced the assumptions (f3) and (g3) with stronger assumption (the coeffi-
cients b0 and b1 are positive!), which is a p-growth condition from below. The importance of
this assumption is to guarantee coerciveness of the functional.
Theorem 2.51. Let Ω ⊂ Rn be open and bounded with Lipschitz boundary, 1 < p < +∞.
If f and g satisfy the assumptions (f1), (f2), (f3)0 , (f4) and (g1), (g2), (g3)0 respectively,
then the minimum problem (2.38) has a solution.
Proof. We establish the existence of a minimizers by the direct method of the calculus of
variations (in particular, we apply Theorem 2.44). We work in the space W 1,p (Ω) endowed
with the weak topology: notice that the space is reflexive, as 1 < p < +∞. We let
Z Z
H(u) := f (x, Du(x)) dx + g(x, u(x)) dx .
Ω
| {z } |Ω {z }
F (u) G(u)
We need to prove weak lowersemicontinuity and coerciveness of H.

Lower semicontinuity. The lower semicontinuity of F with respect to the weak topology of
W 1,p follows by Proposition 2.50.
The functional G is l.s.c. with respect to the strong topology of Lp (Ω) by Proposition 2.46.
We show that this implies that G is sequentially weakly l.s.c. in W 1,p (Ω). Indeed, let uk * u
weakly in W 1,p (Ω). The sequence (uk )k is bounded in W 1,p (Ω) and, by Rellich Theorem
(2.18) (recall that ∂Ω is Lipschitz), uk → u strongly in Lp (Ω).3 Hence, since G is strongly
l.s.c. in Lp (Ω),
G(u) ≤ lim inf G(uk ).
k→∞
3
Exercise: why we do not need to extract a subsequence here?
39
Then H = F + G is sequentially weakly lower semicontinuous in W 1,p (Ω).
Coercivity. Since W 1,p (Ω) is reflexive, coercivity of H with respect to the weak topology is
equivalent to prove that
lim H(u) = +∞.
kukW 1,p (Ω) →+∞
We have thanks to (f3)0 and (g3)0

Z Z
H(u) = f (x, Du(x)) dx + g(x, u(x)) dx
Z Ω Z Ω Z Z
≥ a0 (x) dx + b0 |Du(x)|p dx + a1 (x) dx + b1 |u(x)|p dx
Ω Ω Ω Ω
≥ −ka0 kL1 (Ω) − ka1 kL1 (Ω) + min{b0 , b1 }kukpW 1,p (Ω) ,
from which the claim follows by positivity of the constants b0 and b1 .
Remark 2.52. In the previous proof, the weak lower semicontinuity of G could be proved
also without the assumption of Lipschitz boundary, by using the assumption (g3)0 . Indeed,
let uk * u weakly in W 1,p (Ω). For every Ω0 ⊂⊂ Ω with ∂Ω0 Lipschitz, we have by Rellich
Theorem that uk → u strongly in Lp (Ω0 ) and therefore
Z Z
g(x, u(x)) dx ≤ lim inf g(x, uk (x)) dx
Ω0 k→∞ 0
ZΩ Z
≤ lim inf g(x, uk (x)) dx − a1 (x) dx.
k→∞ Ω Ω\Ω0
By letting Ω0 → Ω we obtain the conclusion.
Exercise 15. Let Ω ⊂ Rn be open and bounded, 1 < p < +∞. Assume that f satisfies the
assumptions (f1), (f2), (f3)0 , (f4). Show that the minimum problem
Z
min f (x, Du(x)) dx
u∈W01,p (Ω) Ω
has a solution.[13]
We next discuss the case with boundary conditions:

Z Z
min f (x, Du(x)) dx + g(x, u(x)) dx . (2.39)
u∈W01,p (Ω) Ω Ω
In this case we can weaken the coercivity assumption (g3)0 and allow for a negative constant
b1 , provided that it is not too large: that is, the coercivity in W01,p (Ω) follows just from the
growth condition on f .
Theorem 2.53. Let Ω ⊂ Rn be open and bounded, 1 < p < +∞. Assume that f satisfies the
assumptions (f1), (f2), (f3)0 , (f4), and that g satisfies (g1), (g2) and
(g3)00 g(x, s) ≥ a1 (x) + b1 |s|p for some a1 ∈ L1 (Ω) and b1 > − Cb0Ω ,
40
where CΩ is the constant in the Poincaré inequality in W01,p (Ω). Then the minimum problem
(2.39) has a solution.
Proof. The weak lowersemicontinuity of the functional is still satisfied, as in the previous
theorem. We assume without los of generality that b1 ≤ 0, as the case b1 > 0 follows by the
previous theorem (since W01,p (Ω) is a closed subspace of W 1,p (Ω)). We prove coerciveness:
we have for all u ∈ W01,p (Ω)
Z Z
f (x, Du(x)) dx + g(x, u(x)) dx
Ω Z Ω Z Z Z
p
≥ a0 (x) dx + b0 |Du(x)| dx + a1 (x) dx + b1 |u(x)|p dx
Ω Ω Z Ω ZΩ
≥ −ka0 kL1 (Ω) − ka1 kL1 (Ω) + b0 |Du(x)|p dx + b1 CΩ |Du(x)|p dx
Ω Ω
≥ −ka0 kL1 (Ω) − ka1 kL1 (Ω) + b0 + b1 CΩ kDukpLp (Ω) ,

and the previous quantity tends to +∞ as kukW 1,p (Ω) → +∞ since b0 + b1 CΩ > 0.
0
Remark 2.54. We can always consider linear lower order terms in Theorem 2.53, that is
0 1 1
g(x, s) = ψ(x)s where ψ ∈ Lp (Ω), + 0 = 1.
p p
Indeed by Young inequality, for every ε > 0 we have
0
εp |s|p |ψ(x)|p
g(x, s) ≥ − − ,
p p0 εp0
εp b0
therefore the assumption (g3)00 is satisfied by choosing ε such that p < CΩ .
We next discuss the case of non homogeneous boundary conditions. We recall that in the
framework of Sobolev spaces a boundary condition of the form “u = u0 on ∂Ω” can be written
in weak form as “u − u0 ∈ W01,p (Ω)”. This formulation does not require any regularity of ∂Ω.
We have the following result.
Theorem 2.55. Let Ω ⊂ Rn be open and bounded, 1 < p < +∞, and let u0 ∈ W 1,p (Ω).
Assume that f satisfies the assumptions (f1), (f2), (f3)0 , (f4), and that g satisfies (g1),
(g2), (g3)00 . Then the minimum problem
Z Z
min f (x, Du(x)) dx + g(x, u(x)) dx (2.40)
u∈u0 +W01,p (Ω) Ω Ω
has a solution.
Proof. The idea is to reformulate the problem in the space W01,p (Ω) and to apply Theo-
rem 2.53. Indeed, letting v = u − u0 , the minimum problem (2.40) is equivalent to
Z Z
min f (x, Dv(x) + Du0 (x)) dx + g(x, v(x) + u0 (x)) dx , (2.41)
v∈W01,p (Ω) Ω Ω
41
which is of the same form as (2.39) for the new functional
Z Z
F (v) := ˜
f (x, Dv(x)) dx + g̃(x, v(x)) dx,
Ω Ω
with
f˜(x, ξ) := f (x, ξ + Du0 (x)),
g̃(x, s) := g(x, s + u0 (x)).
We only need to check that f˜ and g̃ satisfy all the assumptions (f1), (f2), (f3)0 , (f4), (g1),
(g2), and (g3)00 . It is easily seen that the only assumptions that require a proof are the lower
bounds (f3)0 and (g3)00 , as all the other conditions are immediate. For every ξ ∈ Rn and for
ε ∈ (0, 1) we have by convexity of the map ξ 7→ |ξ|p
|ξ|p = |ξ + Du0 (x) − Du0 (x)|p
−Du0 (x) p

ξ + Du0 (x)
= (1 − ε) +ε
1−ε ε
p
Du0 (x) p

ξ + Du0 (x)
≤ (1 − ε) + ε

1−ε ε
1 1
= p−1
|ξ + Du0 (x)|p + p−1 |Du0 (x)|p .
(1 − ε) ε
Therefore
f˜(x, ξ) = f (x, ξ + Du0 (x)) ≥ a0 (x) + b0 |ξ + Du0 (x)|p
1 − ε p−1
≥ a0 (x) − b0 |Du0 (x)|p + b0 (1 − ε)p−1 |ξ|p
ε
≥ ã0 (x) + b̃0 |ξ|p ,
with 1 − ε p−1
ã0 := ã0 − b0 |Du0 |p ∈ L1 (Ω), b̃0 := b0 (1 − ε)p−1 > 0.
ε
Hence f˜ satisfies (f3)0 (with a constant b̃0 which can be taken as close as we want to b0 ). A
similar estimate can be obtained for g̃: assuming that b1 ≤ 0 (the case b1 > 0 does not give
any problem) we have
p p |s|p

≤ |u0 (x)| +
p
u0 (x) s
|s + u0 (x)| = ε + (1 − ε) p−1
,
ε 1−ε ε (1 − ε)p−1
hence
g̃(x, s) = g(x, s + u0 (x)) ≥ a1 (x) + b1 |s + u0 (x)|p
b1 b1
≥ a1 (x) + p−1 |u0 (x)|p + |s|p
ε (1 − ε)p−1
≥ ã1 (x) + b̃1 |s|p ,
with
b1 b1
ã1 := a1 + |u0 |p ∈ L1 (Ω), b̃1 := .
εp−1 (1 − ε)p−1
To conclude that also (g3)00 holds, we only need to check the condition b̃1 > − Cb̃0Ω . This is
true since by assumption b1 > − Cb0Ω and b̃0 → b0 , b̃1 → b1 as ε → 0.
42
We conclude this section by showing with some examples that all the assumptions are
actually necessary in order to get existence of minimizers.
Example 2.56. For ε > 0 the problem
Z 1 Z 1
0 2 2 2
min |u | dx − π (1 + ε) |u| dx .
u∈H01 (0,1) 0 0
| {z }
F (u)
does not have a solution. Indeed, let uk (x) = k sin(πx) ∈ H01 (0, 1): then
Z 1 Z 1 Z 1
F(uk ) = π 2 k 2 | cos(πx)|2 dx−π 2 (1+ε)k 2 | sin(πx)|2 dx = −επ 2 k 2 | sin(πx)|2 dx → −∞
0 0 0
as k → +∞, that is inf H01 (Ω) F = −∞. The problem here is that the assumption (g3)00 is not
satisfied, as π −2 is exactly the Poincaré constant in (0, 1).
Example 2.57 (Lack of coercivity). Let us see that Theorem 2.51 does not hold if we remove
the assumption b1 > 0 in (g3)0 . A similar counterexample can be obtained if we remove the
assumption b0 > 0 in (f3)0 . Consider
Z Z
1
min |Du|2 dx + 2
dx .
u∈W 1,2 (Ω) Ω Ω 1+u
| {z }
F (u)
We have inf F = 0, which can be proved by taking the Rminimizing sequence uk (x) = k, k ∈ N.
1
However the infimum is not attained, since F(u) ≥ Ω 1+u 2 dx > 0 for all u. The problem
here is that the functional is not coercive, and indeed the minimizing sequence (uk )k is not
relatively compact.
Example 2.58 (Lack of reflexivity). Let us show that in the case p = 1 we cannot apply
the existence result: indeed, in this case the space W 1,1 (Ω) is not reflexive and the condition
(2.32) does not guarantee the coercivity of the functional with respect to the weak topology:
there are bounded sequences in W 1,1 which do not have weakly convergent subsequences.
As a counterexample, consider the problem
Z 1 Z 1
0

min (1 + |x|)|u (x)| dx + λ u − sgn x dx .
1,1
u∈W (−1,1) −1 −1
| {z }
F (u)
We can show that, if λ > 2, then inf F = 2 but there is no minimizer.

Notice first that, for all u ∈ W 1,1 (−1, 1), the truncation4 τ (u) := (u ∧ 1) ∨ (−1) is in
1,1
W (−1, 1) and F(u) ≥ F(τ (u)). Hence in the minimum problem we can assume without
loss of generality that −1 ≤ u ≤ 1.
Let us show that F(u) ≥ 2 for all u ∈ W 1,1 (−1, 1). We distinguish three cases.
• If u ≥ 0 on (−1, 0) then
Z 0
F(u) ≥ λ u − sgn x dx ≥ λ > 2.
−1
4
We use the notation a ∨ b := max{a, b}, a ∧ b := min{a, b}.
43
• Similarly if u ≤ 0 on (0, 1) then
Z 1
F(u) ≥ λ u − sgn x dx ≥ λ > 2.
0
• It remains to discuss the case where there exist x1 ∈ (−1, 0) and x2 ∈ (0, 1) such that
u(x1 ) < 0 and u(x2 ) > 0. We can assume without loss of generality that u(x1 ) ≤ u(x)
for all x ∈ (−1, 0), and u(x2 ) ≥ u(x) for all x ∈ (0, 1). Then
Z 1 Z x2 Z x2
0 0 0

(1 + |x|)|u (x)| dx ≥ |u (x)| dx ≥
u (x) dx = u(x2 ) − u(x1 ),
−1 x1 x1
and
Z 0
Z 1
λ u − sgn x dx ≥ λ(u(x1 ) + 1), λ u − sgn x dx ≥ λ(1 − u(x2 )). (2.42)
−1 0
Hence
F(u) ≥ u(x2 ) − u(x1 ) + λ(u(x1 ) + 1) + λ(1 − u(x2 ))

= (λ − 1)u(x1 ) − (λ − 1)u(x2 ) + 2λ
≥ −(λ − 1) − (λ − 1) + 2λ = 2.
Therefore inf F ≥ 2. To show that inf F = 2, it is sufficient to consider the minimizing

sequence 
1
−1 if − 1 < x < − k ,

uk (x) := kx if − k1 < x < k1 ,

1 if k1 < x < 1,

for which
Z 1/k Z 1/k
uk − sgn x dx = 2 + 1 + λ o(1) → 2.

F(uk ) = k (1 + |x|) dx + λ
−1/k −1/k k
However, the minimum is not attained, since the only possibility to have F(u) = 2 is to have
equalities in (2.42), which forces u(x) = u(x1 ) for all x ∈ (−1, 0) and u(x) = u(x2 ) for all
x ∈ (0, 1); however, such a function does not belong to W 1,1 (−1, 1).
In this case the sublevel sets of F are bounded, but not weakly compact as the space is not
reflexive: the minimizing sequence uk converges to u(x) = 1(0,1) (x) − 1(−1,0) (x), which is not
in W 1,1 (−1, 1).
Exercise 16. Prove that

Z 1
inf (1 + |x|)|u0 (x)| dx : u ∈ W 1,1 (−1, 1), u(−1) = −1, u(1) = 1 = 2
−1
and that the infimum is not attained.[14]
44
Example 2.59 (Lack of convexity). Let n = 1, Ω = (0, 1), W (ξ) := (ξ 2 − 1)2 and consider
the minimum problem Z 1 Z 1
0
min W (u ) dx + u4 dx .
u∈W01,4 (0,1) 0 0
| {z }
F (u)
Notice that f (x, ξ) = W (ξ) is not convex. We have inf F = 0: indeed, F(u) ≥ 0, and
a minimizing sequence is obtained by taking tooth saw functions uk ∈ W01,4 (0, 1) such that
|u0k | = 1 almost everywhere and kuk kL∞ (0,1) → 0, see Figure 2.1.
u1
u2
uk
x
0 1
Figure 2.1: The recovery sequence in Example 2.59.
However, the minimum is not attained: if F(u) = 0, then both terms in the functional
would be zero, which would imply u = 0 almost everywhere and u0 ∈ {±1} almost everywhere,
which is not possible.
Notice that F is not lower semicontinuous: for the minimizing sequence considered before,
we have uk * 0 but F(0) = 1 > 0 = limk→∞ F(uk ).
Example 2.60 (Lack of convexity). Let Ω := (0, 1) × (0, 1) ⊂ R2 and consider
Z h i
2
min (∂x u)2 − 1 + (∂y u)4 dx dy .
u∈W01,4 (Ω) Ω
| {z }
F (u)
We can show that inf F = 0. Then a minimizer does not exist, since if F(u) = 0 then ∂y u ≡ 0,
that is u would be constant on vertical lines, and hence u ≡ 0 by the boundary conditions;
however F(0) = 1.
To prove that inf F = 0, first observe that F(u) ≥ 0 for all u. Fix a cut-off function
ϕ ∈ Cc∞ (0, 1), 0 ≤ ϕ ≤ 1, ϕ ≡ 1 on (η, 1 − η), and |ϕ0 | ≤ η2 , for some η ∈ (0, 12 ). Then let
vk (x, y) := uk (x)ϕ(y),
where uk are the one-dimensional tooth saw functions in Figure 2.1. Then vk ∈ W01,4 (Ω) and
setting Ωη := (0, 1) × (η, 1 − η) we have
Z h i
2
(∂x vk )2 − 1 + (∂y vk )4 dx dy = 0,
Ωη
Z h i Z Z
2 2
(∂x vk )2 − 1 + (∂y vk )4 dx dy = (u0k (x)ϕ(y))2 − 1 + (uk (x)ϕ0 (y))4
Ω\Ωη Ω\Ωη Ω\Ωη
2 4 1 4 2
≤ |Ω\Ωη | + |Ω\Ωη | ≤ 2η + 3 4 .
η 2k η k
45
It follows that lim supk→∞ F(vk ) ≤ 2η for every η > 0, hence inf F = 0.
2.4 Necessity of convexity for weak lower semicontinuity

In the previous section we have seen, through some examples, that if we drop the assumption
of convexity of the integrand ξ 7→ f (x, ξ) the weak lowersemicontinuity of the associated
functional is no longer guaranteed. This is not a case: indeed, in the next theorem we show
that convexity is a necessary condition for lower semicontinuity.
Theorem 2.61 (Necessity of convexity). Let Ω ⊂ Rn be open and bounded, f : Ω × Rn → R

be a Carathéodory function, that is
• the map x 7→ f (x, ξ) is measurable with respect to L(Ω), for every fixed ξ ∈ Rn ;
• the map ξ 7→ f (x, ξ) is continuous on Rn , for almost every fixed x ∈ Ω.
Assume also that
• for every R > 0 there exists gR ∈ L1 (Ω) such that
|f (x, ξ)| ≤ gR (x) for a.e. x ∈ Ω and for all ξ ∈ Rn with |ξ| ≤ R. (2.43)
Consider the functional F : W 1,∞ (Ω) → R defined by

Z
F(u) := f (x, Du(x)) dx for u ∈ W 1,∞ (Ω). (2.44)
Ω
If F is sequentially weakly* lower semicontinuous on W 1,∞ (Ω), then for almost every x ∈ Ω
the function
ξ 7→ f (x, ξ) is convex on Rn .
Remark 2.62. Notice that if u ∈ W 1,∞ (Ω), then there exists R > 0 such that kDuk∞ ≤ R,
that is |Du(x)| ≤ R for almost every x ∈ Ω. Hence in view of (2.43) we have that
|f (x, Du(x))| ≤ gR (x) ∈ L1 (Ω)
and since f (·, Du(·)) is measurable (as f is Carathéodory), we obtain that f (·, Du(·)) is inte-
grable in Ω. Therefore the functional F is well-defined on W 1,∞ (Ω).
Remark 2.63. Recall the Definition 2.27 of weak* convergence in W 1,∞ (Ω). The assumption
of sequential weak* lower semicontinuity of F means that
∗
for every uk * u weakly* in W 1,∞ (Ω) F(u) ≤ lim inf F(uk ).
k→∞
This is the weakest possible assumption: indeed, since Ω is bounded, if F is sequentially weakly
l.s.c. in W 1,p (Ω) for some 1 ≤ p < +∞, then it is also seq. weakly* l.s.c. in W 1,∞ (Ω) and
therefore the theorem can be applied.
Remark 2.64. The theorem was firstly proved by Tonelli in dimension n = 1; in this case,
u can be vector-valued, that is u : R → Rm with m ≥ 1. For n > 1, u has to be scalar-valued:
for n > 1 and m > 1 the theorem does not hold.
46
For the proof of the theorem we will need the following lemma concerning the asymptotic
behaviour of sequences of highly oscillating functions.
Lemma 2.65 (Riemann-Lebesgue). Let g ∈ L∞ (R) be 1-periodic, that is g(x + 1) = g(x) for
all x ∈ R. Define gk (x) := g(kx), for k ∈ N. Then
Z 1
∗
gk * m weakly* in L∞ (R), where m := g(x) dx.
0
Proof. Assume without loss of generality that m = 0 (otherwise consider g̃(x) = g(x) − m).
We want to prove that
Z
gk (x)ϕ(x) dx → 0 for all ϕ ∈ L1 (R), (2.45)
R
and by density is sufficient to prove (2.45) for ϕ ∈ Cc∞ (R). Let

Z x
G(x) := g(t) dt.
0
Since g ∈ L∞ (R), we have that G is differentiable almost 0

R 1everywhere on R and G (x) = g(x)
for a.e. x ∈ R. Moreover, since g is 1-periodic with 0 g(x) dx = 0, it follows that G is
one-periodic, and therefore G is bounded.
Let Hk (x) := G(kx); by the chain rule Hk0 (x) = kG0 (kx) = kg(kx) for a.e. x, hence
Z Z

gk (x)ϕ(x) dx = g(kx)ϕ(x) dx

R
RZ
1 0

= Hk (x)ϕ(x) dx
k R
Z
1 0
c
= − Hk (x)ϕ (x) dx ≤ → 0
k R k
where we used the fact that Hk is bounded in the last passage. This proves (2.45).
Remark 2.66. One can prove a n-dimensional version of the Riemann-Lebesgue Lemma:
if g ∈ L1loc (Rn ) is 1-periodic in the coordinate directions, that is g(x + ei ) = g(x) for all
i = 1, . . . , n and for a.e. x ∈ Rn , then setting gε (x) := g( xε ) the following implications hold:
• if g ∈ Lploc (Rn ) with 1 ≤ p < +∞, then gε * m weakly in Lp (Ω) as ε → 0, for all
Ω ⊂ Rn open and bounded,
∗
• if g ∈ L∞ (Rn ) then gε * m weakly* in L∞ (Rn ) as ε → 0,
R
where m := (0,1)n g(x) dx.
Proof of Theorem 2.61. For a fixed ξ ∈ Rn we define the linear function
uξ (x) := ξ · x, Duξ (x) = ξ. (2.46)
Fix ξ1 , ξ2 ∈ Rn and λ ∈ (0, 1), and let ξ := λξ1 + (1 − λ)ξ2 . The goal is to show that
f (x, ξ) ≤ λf (x, ξ1 ) + (1 − λ)f (x, ξ2 ) (2.47)
47
which is the convexity of the map ξ 7→ f (x, ξ). In order to prove the claim, the first step will
be to construct a sequence uk ∈ W 1,∞ (Ω) such that as k → ∞
∗
uk * uξ weakly* in W 1,∞ (Ω), (2.48)
F(uk ) → λF(uξ1 ) + (1 − λ)F(uξ2 ). (2.49)

Indeed, if we construct such a sequence then by sequential lower semicontinuity of F with
respect to weak* convergence we obtain
(2.49)
F(uξ ) ≤ lim inf F(uk ) = λF(uξ1 ) + (1 − λ)F(uξ2 ),
k→∞
or explicitly Z Z Z
f (x, ξ) dx ≤ λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx. (2.50)
Ω Ω Ω
This is the integral version of the goal (2.47); then in a final step we will need to localize the
previous inequality.
Step 1: construction of uk (case n = 1). We first show how to construct a sequence uk
satisfying the conditions (2.48)–(2.49) in the one-dimensional case. The idea is to approximate
the linear function uξ , with slope ξ, with piecewise affine oscillating functions with slope taking
alternatively the values ξ1 and ξ2 . More precisely, for k ∈ N let
h λ h1 1 λ h2 2 λ
Ak := 0, ∪ , + ∪ , + ∪ ...
k k k k k k k
hλ 1 h1 λ 2
Bk := R\Ak = , ∪ + , ∪ ...
k k k k k
and let uk ∈ W 1,∞ (Ω) be defined by the conditions
(
ξ1 if x ∈ Ak
uk (0) = 0, u0k (x) = ξ1 1Ak (x) + ξ2 1Bk (x) =
ξ2 if x ∈ Bk
(see Figure 2.2). Notice that by construction

c
|uk (x) − uξ (x)| ≤ →0 as k → ∞,
k
that is uk → u strongly in L∞ (and hence also weakly* in L∞ ). In order to prove (2.48), we
∗
need to show that u0k * u0ξ = ξ weakly* in L∞ . To this aim, notice that 1A1 coincides with
the 1-periodic extension on R of the function 1[0,λ) ; since 1Ak (x) = 1A1 (kx), we obtain by
Riemann-Lebesgue Lemma 2.65
Z 1
∗
1Ak * 1A1 (x) dx = λ weakly* in L∞ (R),
0
Z 1 (2.51)
∗ ∞
1Bk * 1B1 (x) dx = 1 − λ weakly* in L (R),
0
and therefore
∗
u0k = ξ1 1Ak + ξ2 1Bk * λξ1 + (1 − λ)ξ2 = ξ weakly* in L∞ (R)
48
y
uk
uk+1
uξ
x
λ 1 1 λ 2
k k k + k k
Figure 2.2: The sequence (uk )k constructed in Step 1.
∗
that is u0k * u0ξ weakly* in L∞ (R). This shows the condition (2.48).
We next show that for the sequence uk also the condition (2.49) holds. Indeed we have
Z Z Z
F(uk ) = f (x, Duk (x)) dx = f (x, ξ1 ) dx + f (x, ξ2 ) dx
Ω Ω∩Ak Ω∩Bk
Z Z
= f (x, ξ1 )1Ak (x) dx + f (x, ξ2 )1Bk (x) dx
ZΩ Z Ω
→ f (x, ξ1 )λ dx + f (x, ξ2 )(1 − λ) dx = λF(uξ1 ) + (1 − λ)F(uξ2 ),

Ω Ω
where we can pass to the limit thanks to the convergences (2.51), since f (·, ξi ) ∈ L1 (Ω) by
assumption (2.43).
Step 2: construction of uk (case n > 1). We next discuss how to extend the previous
construction in the higher-dimensional case n > 1. Without loss of generality, we fix the
coordinate system so that the vector ξ1 − ξ2 is in the direction e1 . We denote the generic
point x ∈ Rn as x = (x1 , x0 ) ∈ R × Rn−1 . With this notation
ξ1 = (ξ11 , ξ10 ), ξ2 = (ξ21 , ξ20 ) with ξ10 = ξ20
and
ξ = λξ1 + (1 − λ)ξ2 = (ξ 1 , ξ 0 ), ξ 1 = λξ11 + (1 − λ)ξ21 , ξ 0 = ξ10 = ξ20 .
Then uξ (x) = ξ · x = ξ 1 x1 + ξ 0 · x0 = uξ1 (x1 ) + uξ0 (x0 ); the idea is to use the one-dimensional
construction in the previous step to approximate uξ1 , while keeping uξ0 fixed. More precisely,
we let
ũk (x1 , x0 ) := uk (x1 ) + ξ 0 · x0
where uk are the one-dimensional functions constructed in the previous step; in particular,
we have uk → uξ1 in L∞ , so that ũk → uξ in L∞ . Furthermore
(
ξ11 + ξ 0 = ξ1 if (x1 , x0 ) ∈ Ãk := Ak × Rn−1 ,
Dũk (x) =
ξ21 + ξ 0 = ξ2 if (x1 , x0 ) ∈ B̃k := Bk × Rn−1 .
49
∗ ∗
We have 1Ãk * λ and 1B̃k * 1 − λ weakly* in L∞ by Lemma 2.65 (using Fubini’s Theorem),
∗
and therefore Dũk * λξ1 + (1 − λ)ξ2 = ξ. This shows that the functions ũk satisfy the
condition (2.48); the other condition (2.49) can be proved exactly as in the one-dimensional
case.
Step 3: localization. Summing up, the construction performed in the first two steps yields
the inequality (2.50), which is valid for every ξ1 , ξ2 ∈ R and for every λ ∈ (0, 1), with
ξ = λξ1 + (1 − λ)ξ2 . We next want to localize the previous inequality: we show that for every
open set U ⊂ Ω we have
Z Z Z
f (x, ξ) dx ≤ λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx. (2.52)
U U U
Notice that this is not a trivial consequence of the previous step: indeed, F is weakly* l.s.c.
in W 1,∞ (Ω), but it is not guaranteed that F is weakly* l.s.c. also in W 1,∞ (U ).
Let V ⊂⊂ U be open, and let ϕ ∈ Cc∞ (U ) be a cut-off function, such that 0 ≤ ϕ ≤ 1,
ϕ ≡ 1 in V . Let wk := ϕuk + (1 − ϕ)uξ , where uk is the sequence satisfying (2.48)–(2.49)
∗
constructed in the previous steps. Then we easily have wk * uξ weakly* in W 1,∞ (Ω), and
by lower semicontinuity
Z Z Z
f (x, ξ) dx + f (x, ξ) dx = F(uξ ) ≤ lim inf F(wk ) = lim inf f (x, Dwk (x)) dx
U Ω\U k→∞ k→∞ Ω
Z Z Z
= lim inf f (x, Duk (x)) dx + f (x, Dwk (x)) dx + f (x, ξ) dx
k→∞ V U \V Ω\U
Z Z Z
≤ lim sup f (x, Duk (x)) dx + lim sup f (x, Dwk (x)) dx + f (x, ξ) dx
k→∞ V k→∞ U \V Ω\U
Z Z
=λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx
V Z V Z
+ lim sup f (x, Dwk (x)) dx + f (x, ξ) dx
k→∞ U \V Ω\U
from which we get

Z Z Z Z
f (x, ξ) dx ≤ λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx + lim sup f (x, Dwk (x)) dx.
U V V k→∞ U \V
To compute the last term on the right-hand side, we choose R > 0 such that |ξ1 | < R − 1 and
|ξ2 | < R − 1. With this choice we have

|Dwk | = ϕDuk + (1 − ϕ)ξ + Dϕ(uk − uξ )

≤ ϕDuk + (1 − ϕ)ξ + Dϕ(uk − uξ )
ckDϕk∞
≤R−1+ <R
k
for k large enough, therefore by assumption (2.44)
Z Z
lim sup f (x, Dwk (x)) dx ≤ gR (x) dx,
k→∞ U \V U \V
50
and in turn
Z Z Z Z
f (x, ξ) dx ≤ λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx + gR (x) dx
U V V U \V
Z Z Z
≤λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx + 2 gR (x) dx.
U U U \V
By letting V ↑ U , the last integral goes to zero (as gR ∈ L1 (Ω) is independent of U and V ),
so that (2.52) holds for every U ⊂ Ω open.
Step 4: conclusion. For every ξ1 , ξ2 ∈ Rn and λ ∈ (0, 1), setting ξ = λξ1 + (1 − λ)ξ2 , we have
Z Z Z
f (x, ξ) dx ≤ λ f (x, ξ1 ) dx + (1 − λ) f (x, ξ2 ) dx for every U ⊂ Ω open. (2.53)
U U U
There exists a Lebesgue-negligible set Nξ1 ,ξ2 ,λ (depending on ξ1 , ξ2 , λ) such that |Nξ1 ,ξ2 ,λ | = 0
and every point x ∈ Ω\Nξ1 ,ξ2 ,λ is a Lebesgue point for the functions f (·, ξ1 ), f (·, ξ2 ), f (·, ξ):
hence for such points x ∈ Ω\Nξ1 ,ξ2 ,λ , by choosing U = Br (x) in (2.53) and dividing by |Br (x)|
Z Z Z
f (y, ξ) dy ≤ λ f (y, ξ1 ) dy + (1 − λ) f (y, ξ2 ) dy
Br (x) Br (x) Br (x)
and by passing to the limit as r → 0
f (x, ξ) ≤ λf (x, ξ1 ) + (1 − λ)f (x, ξ2 ) for every x ∈ Ω\Nξ1 ,ξ2 ,λ .
Notice that this is not yet the conclusion: we want that (2.46) holds for every x ∈ Ω\N ,
for some set N with |N | = 0, but here we just proved the property for a set N which depends
on ξ1 , ξ2 and λ! To conclude, let
[
N := Nξ1 ,ξ2 ,λ .
ξ1 ,ξ2 ∈Qn
λ∈Q∩(0,1)
Then |N | = 0 and for every ξ1 , ξ2 ∈ Qn and λ ∈ Q ∩ (0, 1),
f (x, ξ) ≤ λf (x, ξ1 ) + (1 − λ)f (x, ξ2 ) for every x ∈ Ω\N.
By density and by continuity of ξ 7→ f (x, ξ), the previous inequality then holds for every
ξ1 , ξ2 ∈ Rn , for every λ ∈ (0, 1), and for every x ∈ Ω\N .
Remark 2.67. Under the assumptions of Theorem 2.61, if F is sequentially weakly* contin-
uous in W 1,∞ (Ω), then F is linear:
n Z
X
F(u) = fi (x)∂i u(x) dx
i=1 Ω
0
with fi ∈ Lp (Ω).
51
2.5 Integral functionals: the general case
Until now we have considered functionals depending separately on the function u or on its
gradient Du, in the form (2.35) or (2.37) (or the sum of two functionals of this form). This
is not the most general case, which would be to consider functionals of the form
Z
H(u) := f (x, u(x), Du(x)). (2.54)
Ω
The following theorem, of which we omit the complex proof (see for instance [2, Theo-
rem 3.23]), provides a sufficient condition for the weak lower semicontinuity of such a func-
tional.
Theorem 2.68 (Ioffe-Olech). Let Ω ⊂ Rn be an open set, and let
f : Ω × Rm × Rd → [0, +∞]
satisfy the following assumptions:
(i) f is measurable with respect to L(Ω) ⊗ B(Rm ) ⊗ B(Rd );
(ii) for a.e. x ∈ Ω the map (s, ξ) 7→ f (x, s, ξ) is lower semicontinuous;
(iii) for a.e. x ∈ Ω, for all s ∈ Rm the map ξ 7→ f (x, s, ξ) is convex.
Then the functional
Z
F(u, v) := f (x, u(x), v(x)) dx, u ∈ L1 (Ω; Rm ), v ∈ L1 (Ω; Rd )
Ω
is sequentially lower semicontinuous with respect to the topology s−L1 (Ω; Rm )⊗w−L1 (Ω; Rd ):
that is, if uk → u strongly in L1 (Ω; Rm ) and vk * v weakly in L1 (Ω; Rd ), then
F(u, v) ≤ lim inf F(uk , vk ).
k→∞
Remark 2.69. Notice that we considered the weakest possible assumption: indeed, if Ω is
bounded the lower semicontinuity in the statement implies the lower semicontinuity with re-
spect to the Lp -topology, for all p > 1.
Corollary 2.70. Let Ω ⊂ Rn be an open set, and let f : Ω × Rm × Rm×n satisfy the assump-
tions of Theorem 2.68. Then the functional H : W 1,1 (Ω; Rm ) → [0, ∞] defined by (2.54) is
sequentially lower semicontinuous with respect to the weak convergence in W 1,1 (Ω; Rm ).
Proof. Assuming that ∂Ω is Lipschitz, if uk * u weakly in W 1,1 (Ω; Rm ) then by Rellich
Theorem uk → u strongly in L1 (Ω; Rm ), and Duk * Du weakly in L1 (Ω; Rn×m ): then by
Theorem 2.68
H(u) = F(u, Du) ≤ lim inf F(uk , Duk ) = lim inf H(uk ).
k→∞ k→∞
If ∂Ω is not Lipschitz, then approximate Ω from the interior by an increasing sequence of
Lipschitz domains Ωk ⊂⊂ Ω. By the previous argument the functionals
Z
Hk (u) := f (x, u(x), Du(x)) dx, u ∈ W 1,1 (Ω),
Ωk
are sequentially weakly lower semicontinuous in W 1,1 (Ω). Then also H = supk Hk has the
same property.
52
Remark 2.71. For functionals H as in (2.54), defined on vector-valued functions u : Ω →
Rm the convexity of f with respect to ξ is no longer a necessary condition for lower semicon-
tinuity. In this case, the right notion is quasi-convexity, see [2] for more on this subject.
Remark 2.72. The previous theorems can be extended to the case of integrands f not neces-
sarily positive, but satisfying a lower bound f (x, s, ξ) ≥ a(x) with a ∈ L1 (Ω).
Notes
[4]
Solution to Exercise 4. Clearly u ∈ L1 (Ω). We compute the weak partial derivatives of u according to the
definition. Let ϕ ∈ Cc∞ (Ω). Then
Z 1 Z 1 Z 1 Z 0 Z 1
− u(x, y)∂x ϕ(x, y) dx dy = − y − x∂x ϕ(x, y) dx + x∂x ϕ(x, y) dx dy
−1 −1 −1 −1 0
Z 1 Z 0 Z 1
=− y ϕ(x, y) dx − ϕ(x, y) dx dy
−1 −1 0
Z 1 Z 1
= y sgn(x)ϕ(x, y) dx dy.
−1 −1
This shows that ∂x u(x, y) = sgn(x)y in the weak sense. For the partial derivative in the variable y we have
Z 1 Z 1 Z 1 Z 1
− u(x, y)∂y ϕ(x, y) dx dy = − |x| y∂y ϕ(x, y) dy dx
−1 −1 −1 −1
Z 1
y=1
Z 1
=− |x| yϕ(x, y) y=−1 − ϕ(x, y) dy dx
−1 −1
Z 1 Z 1
= |x|ϕ(x, y) dx dy,
−1 −1
This shows that ∂y u(x, y) = |x| in the weak sense. Finally, it is easily seen that ∂x u, ∂y u ∈ L1 (Ω), hence
u ∈ W 1,1 (Ω).
[5]
Solution to Exercise 5. Notice first that u is smooth away from x = 0, with (classical) partial derivatives
given by
αxi α
∂i u(x) = − α+2 , |Du(x)| = (x 6= 0).
|x| |x|α+1
Hence |Du| ∈ L1 (B1 ) if α + 1 < n. Fix ϕ ∈ Cc∞ (B1 ) and ε > 0. We have, denoting by ν = (ν1 , . . . , νn ) the
exterior unit normal on ∂Bε ,
Z Z Z
u ∂i ϕ dx = − ∂i u ϕ dx − uϕνi dHn−1
B1 \Bε B1 \Bε ∂Bε
and (since α < n − 1)

Z Z
n−1 −α
|ϕ| dx ≤ εn−1−α kϕkL∞ (B1 ) Hn−1 (∂B1 ) → 0

uϕν i dH ≤ ε as ε → 0,
∂Bε ∂Bε
therefore Z Z
u ∂i ϕ dx = − ∂i u ϕ dx.
B1 B1
By definition of weak derivative, this shows that the weak derivative of u in B1 coincides with the classical
derivative ∂i u(x) = − |x|αx i
α+2 . We then need to check the integrability of u and ∂i u:
Z Z Z 1 Z Z 1
|u|p dx = |x|−αp dx = ρ−αp dHn−1 dρ = Hn−1 (∂B1 ) ρn−1−αp dρ,
B1 B1 0 ∂Bρ 0
hence u ∈ Lp (B1 ) if and only if n − 1 − αp > −1, that is α < np . Furthermore

Z Z Z 1
|Du|p dx = α|x|−(α+1)p dx = αHn−1 (∂B1 ) ρn−1−(α+1)p dρ,
B1 B1 0
53
hence Du ∈ Lp (B1 ) if and only if n − 1 − (α + 1)p > −1, that is (α + 1)p < n.
[6]
Solution to Exercise 7. Let uk ∈ C ∞ (Ω) be an approximating sequence given by Theorem 2.10, with
uk → u in W 1,p (Ω). By the standard integration by parts formula we have for all k ∈ N and for all ϕ ∈ C 1 (Ω)
Z Z Z
uk Dϕ dx = − ϕ Duk dx + Tr uk ϕν dHn−1 .
Ω Ω ∂Ω
We can pass to the limit as k → ∞ in all the terms in the previous equality: in particular, for the first term we
use the convergence uk → u in Lp (Ω); for the second term we use the convergence Duk → Du in Lp (Ω; Rn );
for the last term, by continuity of the trace operator (Theorem 2.15) and |ν| = 1 we have
Z Z Z
n−1 n−1 n−1

Tr uk ϕν dH − Tr u ϕν dH ≤
Tr(uk − u)ϕν dH
∂Ω ∂Ω ∂Ω
≤ k Tr(uk − u)kLp (Ω) kϕkLp0 (Ω)
≤ Ckuk − ukW 1,p (Ω) → 0.
[7]
Solution to Exercise 8. By using Exercise 7 we have for all ϕ ∈ Cc∞ (Rn )
Z Z Z
wDϕ dx = uDϕ dx + vDϕ dx
Rn Ω Rn \Ω
Z Z Z Z
=− ϕ Du dx + Tr u ϕν dHn−1 − ϕ Dv dx − Tr v ϕν dHn−1
Ω ∂Ω Rn \Ω ∂Ω
Z Z Z
Tr u − Tr v ϕν dHn−1

=− ϕ Du dx − ϕ Dv dx +
Ω Rn \Ω ∂Ω
where ν denotes the exterior unit normal to ∂Ω. If Tr u = Tr v, then from the previous identity we deduce
that Z Z Z
wDϕ dx = − ϕ Du dx − ϕ Dv dx,
Rn Ω Rn \Ω
n
which implies that the weak derivative of w in R is
(
Du in Ω,
Dw =
Dv in Rn \Ω,
and in particular it follows that Dw ∈ Lp (Rn ; Rn ), that is w ∈ W 1,p (Rn ). Conversely, if w ∈ W 1,p (Rn ),
assume by contradiction that Tr u 6= Tr v. We can find a sequence ϕk ∈ Cc∞ (Rn ) with kϕk k∞ ≤ 1, ϕk → 0
almost everywhere, and such that
Z
Tr u − Tr v ϕk νi dHn−1 ≥ c > 0

for all k and i = 1, . . . , n
∂Ω
therefore
Z Z
∂i w ϕk dx = − w ∂i ϕk dx
Rn n
Z R Z Z
Tr u − Tr v ϕk ν dHn−1

= ϕk Du dx + ϕk Dv dx −
Ω Rn \Ω ∂Ω
which gives a contradiction by passing to the limit as k → ∞.

[8]
Solution to Exercise 9. See [4, p. 291].
[9]
Solution to Exercise 10. Since L2 (Ω) is an Hilbert space with scalar product hu, vi =
R
Ω
uv dx, we have
kuk − uk2L2 (Ω) = kuk k2L2 − 2huk , ui + kuk2L2 −→ kuk2L2 − 2hu, ui + kuk2L2 = 0,
where we used property (iii) in Proposition 2.26 to deduce that huk , ui → hu, ui.
[10]
Solution to Exercise 12. We prove only the topological version.
54
“(i)⇒(ii)”: we show that {F > t} is open. Let x0 be such that F(x0 ) > t; then by lower semicontinuity
t < F(x0 ) ≤ lim inf F(y) = sup inf F(y),

y→x0 U ∈N (x0 ) y∈U
and therefore there exists a neighbourhood U ∈ N (x0 ) such that F(y) > t for all y ∈ U , that is U ⊂ {F > t}.
Hence {F > t} is open.
“(ii)→(iii)”: we show that (epi(F))c := (X ×R)\epi(F) is open. Let (x0 , t0 ) ∈ (epi(F))c , that is F(x0 ) > t0 .
For ε > 0 sufficiently small we have F(x0 ) > t0 + ε, and by (ii) there exists a neighbourhood U ∈ N (x0 ) such
that F (x) > t0 + ε for all x ∈ U . Then U × (t0 − ε, t0 + ε) is a neighbourhood of (x0 , t0 ) contained in (epi(F))c .
“(iii)→(i)”: let x0 ∈ X , and let us prove that F is l.s.c. at x0 , that is for every ε > 0 there exists U ∈ N (x0 )
such that F(y) ≥ F(x0 ) − ε for all y ∈ U . We have (x0 , F(x0 ) − ε) ∈ (epi(F))c which is open, therefore there
exists a neighbourhood V of (x0 , F(x0 ) − ε) in the product topology such that F(x) > t for all (x, t) ∈ V .
Then there exists U ∈ N (x) such that U × {F (x0 ) − ε} ⊂ V , and F(x) > F(x0 ) − ε for all x ∈ U , as desired.
[11]
Solution to Exercise 13. Fix t < F(x). Then there exists i ∈ I such that t < Fi (x); by lower semicontinuity
of Fi , we can find U ∈ N (x) such that Fi (y) > t for all y ∈ U . Therefore for every t < F(x) there exists
U ∈ N (x) such that F(y) ≥ Fi (y) > t for all y ∈ U , which proves that F is lower semicontinuous.
To prove the sequential version of the statement, let xn → x; then for every i ∈ I we have by sequential
lower semicontinuity of Fi that
Fi (x) ≤ lim inf Fi (xn ) ≤ lim inf F(xn ),

n→∞ n→∞
from which it follows that F(x) ≤ lim inf n→∞ F(xn ) by taking the supremum with respect to i ∈ I.
[12]
Solution to Exercise 14. If x and y are two minimum points of F, with x 6= y, then by strict convexity
x y 1 1 1 1
min F ≤ F + < F(x) + F(y) = min F + min F = min F,
X 2 2 2 2 2 X 2 X X
which is a contradiction.
Solution to Exercise 15. The space W01,p (Ω) is reflexive and the functional is weakly lower semicontinuous
[13]
by Proposition 2.50. We only need to prove coerciveness: since by Poincaré inequality the norm kDukLp (Ω) is
an equivalent norm on W01,p (Ω), for all u ∈ W01,p (Ω) we have
Z
f (x, Du(x)) dx ≥ −ka0 kL1 (Ω) + b0 kDukpLp (Ω) → +∞ as kukW 1,p (Ω) → +∞.
0
Ω
[14]
Solution to Exercise 16. We have for all u ∈ X := {u ∈ W 1,1 (−1, 1) : u(−1) = −1, u(1) = 1}
Z 1 Z 1
|u0 (x)| dx ≥ u0 (x) dx = 2,

F(u) ≥
−1 −1
hence inf F ≥ 2. A minimizing sequence is given by the functions


1
−1 if − 1 < x < − k ,

uk (x) := kx if − k1 < x < k1 ,

1 if k1 < x < 1.

R 1/k 1
Indeed F(uk ) = k −1/k
(1 + |x|) = 2 + k
→ 2. However, a minimizer cannot exists since if u ∈ X satisfies
F(u) = 2, then
Z 1
|x||u0 (x)| dx = 0 ⇒ |u0 | = 0 a.e.
−1
but in this case the boundary conditions cannot be satisfied.
55
Chapter 3
Relaxation
In view of Theorem 2.61, in the case of a non-convex integrand the associated integral func-
tional is not lower semicontinuous. In particular we cannot guarantee that cluster points of
minimizing sequences are minimizers. The notion of relaxation allows to treat the case of
functionals which are not l.s.c. by considering their lower semicontinuous envelope: in this
way, we can describe the behaviour of minimizing sequences in term of minimum points of
some other functional.
After discussing the relaxation procedure in a general topological space in Section 3.1, we
will compute the lower semicontinuous envelope of integral functionals of the form
Z
F(u) := f (x, Du(x)) dx, u ∈ W 1,p (Ω),
Ω
with respect to the weak topology of W 1,p (Ω), 1 ≤ p < +∞. As already observed, it is
particularly interesting to consider this question if ξ 7→ f (x, ξ) is not convex. We will see in the
main result of this section that the relaxation of such a functional requires the convexification
of the integrand.
3.1 Relaxation in an abstract setting

In this section, X will denote a topological space and F : X → R a generic functional.
Definition 3.1 (Lower semicontinuous envelope). The lower semicontinuous envelope of F

(or relaxation of F) is the functional F : X → R defined by
n o
F(x) := sup G(x) : G is l.s.c., G ≤ F , (3.1)
that is, F is the greatest lower semicontinuous functional below F.
Notice that the definition makes sense, as the set on the right-hand side of (3.1) is not
empty (G ≡ −∞ is admissible). We remark that F is characterized by the following properties:
(i) F is lower semicontinuous, as the supremum of an arbitrary family of l.s.c. functions is

l.s.c. (see Exercise 13);
(ii) F ≤ F;
56
(iii) if G is l.s.c. and G ≤ F, then G ≤ F.
Example 3.2. Consider the function f : R → R defined by

(
5 if x ≥ 2,
f (x) :=
3 if x < 2.
Then f is not lower semicontinuous: indeed, lim inf x→2 f (x) = 3 < f (2). The lower semi-
continuous envelope of f is the function
(
5 if x > 2,
g(x) :=
3 if x ≤ 2.
Indeed, g(x) ≤ f (x) for all x, and g is l.s.c.; therefore g ≤ f¯. Let us prove the opposite
inequality. For x 6= 2 the inequality is trivial, since f¯(x) ≤ f (x) = g(x). For x = 2, we have
f¯(2) ≤ lim inf y→2 f (y) = 3 = g(2).
This procedure can be adapted to any family of functions which is “stable” with respect to
the supremum operation, for instance for convex functions: by the same idea you can define
the convex envelope of a function. A remarkable difference is that, while the convexification is
a global operation, the relaxation is a local operation, that is the value of F(x) depends only
on the behaviour of F in a neighbourhood of x. This is shown by the following proposition.
Proposition 3.3. For every x ∈ X we have
F(x) = sup inf F(y). (3.2)

U ∈N (x) y∈U
Proof. Denote by H the right-hand side of (3.2); let us prove that F = H.

Step 1: H is lower semicontinuous. By Definition 2.30, we need to show that
H(x) ≤ lim inf H(y) = sup inf H(y).

y→x U ∈N (x) y∈U \{x}
Fix t < H(x). By definition of H, there exists a neighbourhood U ∈ N (x) such that
inf y∈U F(y) > t. Take z ∈ U ; then U ∈ N (z) and H(z) ≥ inf y∈U F(y) > t.
Therefore we have shown that, for every t < H(x), there exists U ∈ N (x) such that for
every z ∈ U we have H(z) > t, that is, H is lower semicontinuous.
Step 2: H ≤ F. This is immediate, since for every U ∈ N (x) we have inf y∈U F(y) ≤ F(x).
Step 3: F ≤ H. By the first two steps, it immediately follows that H ≤ F. Let us prove the
opposite inequality. Since F is lower semicontinuous and F ≤ F, we have
F(x) = sup inf F(y) ≤ sup inf F(y) = H(x),

U ∈N (x) y∈U U ∈N (x) y∈U
that is the conclusion.
As usual, if the space satisfies the first axiom of countability we can characterize the lower
semicontinuous envelope in terms of sequences.
57
Proposition 3.4. Assume that X satisfies the first axiom of countability, let F : X → R,
and let F be its relaxation. Then:
(i) for every x ∈ X and for every sequence xk → x, we have F(x) ≤ lim inf k→∞ F(xk );
(ii) for every x ∈ X there exists a sequence xk → x such that F(x) = limk→∞ F(xk ).
Corollary 3.5. If X satisfies the first axiom of countability and F : X → R, then

n o
F(x) = min lim inf F(xk ) : xk → x . (3.3)
k→∞
Proof of Proposition 3.4. Notice that (i) is valid in a general topological space: we will not
use the countability assumption in the proof.
Proof of (i). Since F is lower semicontinuous, it is in particular sequentially lower semicon-
tinuous, therefore
F(x) ≤ lim inf F(xk ) ≤ lim inf F(xk ),
k→∞ k→∞
the last inequality following from F ≤ F.

Proof of (ii). Assume F(x) < +∞. Let (Uk )k be a countable neighbourhood base of the
point x such that Uk+1 ⊂ Uk for every k. Let (tk )k ⊂ R be a decreasing sequence such that
tk → F(x), with tk > F(x) for all k. Then, since
tk > F(x) ≥ inf F(y),

y∈Uk
we can choose a point xk ∈ Uk such that F(xk ) < tk . Then it is easy to see that xk → x and
(i)
F(x) = lim tk ≥ lim sup F(xk ) ≥ lim inf F(xk ) ≥ F(x),
k→∞ k→∞ k→∞
from which the conclusion follows.
Theorem 3.6 (Relaxation). Let X be a topological space, F : X → R, and let F be the lower
semicontinuous envelope of F. Assume that F is coercive. Then:
(i) the minimum minX F is attained;
(ii) minX F = inf X F;
(iii) assume also that X satisfies the first axiom of countability and F 6≡ +∞; then, if x is a
minimum point of F, there exists a minimizing sequence (xk )k for F such that xk → x;
conversely, if (xk )k is a minimizing sequence for F, then there exists a convergent
subsequence xkj → x, and x is a minimum point of F.
Proof. We first show that (independently of the coerciveness of F)
inf F = inf F. (3.4)

X X
Let G be the constant function which is identically equal to inf X F. Then G is obviously
lower semicontinuous, and G ≤ F; then G ≤ F, which in turn implies that inf X F ≤ inf X F.
The opposite inequality is trivial, since F ≤ F.
58
We next show that, if F is coercive, then F is coercive. By Definition 2.39, for every t ∈ R
the set {F ≤ t} is compact. We need to show that {F ≤ t} is relatively compact. We claim
that
{F ≤ t} ⊂ {F ≤ t0 } for all t0 > t. (3.5)
Indeed, let t0 > t, and let x ∈ {F ≤ t}, that is F(x) ≤ t. By (3.2), for every U ∈ N (x) we
have inf y∈U F(y) ≤ t < t0 ; that is, for every U ∈ N (x) there exists y ∈ U such that F(y) ≤ t0 ,
or equivalently y ∈ {F ≤ t0 }. This shows that x ∈ {F ≤ t0 }, that is (3.5) holds.
Finally, we have that {F ≤ t} ⊂ {F ≤ t0 }, and, since the second set is compact by coer-
civity of F, it follows that also the first set is compact. Therefore F is coercive.
We can now prove the properties in the statements. The existence of a minimizer for F
follows by Tonelli’s Theorem 2.40, since F is lower semicontinuous and coercive. Moreover
the second claim follows from (3.4).
Let us show the first part of (iii). Let x be a minimum point of F. By Proposition 3.4,
there exists a sequence xk → x such that
lim F(xk ) = F(x) = min F = inf F,

k→∞ X X
that is, (xk )k is a minimizing sequence for F.

Let us eventually prove the second part of (iii). Let (xk )k be a minimizing sequence, that
is F(xk ) → inf X F. For t > inf X F, we have xk ∈ {F ≤ t} for k sufficiently large, and by
coerciveness there exists a convergent subsequence xkj → x. Then by Proposition 3.3
F(x) ≤ lim inf F(xkj ) = lim F(xk ) = inf F = min F,

j→∞ k→∞ X X
that is x is a minimizer for F.
From the theorem it follows that minimum points of F are all and only the cluster points
of minimizing sequence for F.
Exercise 17. Let X be a topological space, let F : X → R, and let G : X → R be continuous.

6 F + G).[15]
Prove that F + G = F + G. (Notice that, for general functionals, F + G =
Exercise 18. Let X be a topological space, F : X → R. Prove that:
(i) for every s ∈ R \

{F ≤ s} = {F ≤ t}
t>s
(where the bar denotes the closure in X);
(ii) the epigraph of F coincides with the closure in X × R of the epigraph of F.
Prove the same properties assuming that X satisfies the first axiom of countability and using
the characterization in Proposition 3.4 of F.[16]
59
3.2 Convex envelope
We define the convexification procedure, which plays an essential role in the relaxation of
integral functionals.
Definition 3.7 (Convex envelope). Let X be a vector space and let f : X → R ∪ {+∞}.
Assume that there exists a ∈ X ∗ and b ∈ R such that f (x) ≥ ha, xi + b. Then we define the
convex envelope (or convexification) of f as the function
n o
co(f )(x) := sup g(x) : g : X → R ∪ {+∞}, g convex, g ≤ f . (3.6)
Notice that the definition makes sense, as the set on the right-hand side of (3.6) is not
empty by the assumption. We remark that co(f ) is characterized by the following properties:
(i) co(f ) is convex, as the supremum of an arbitrary family of convex functions is convex;
(ii) co(f ) ≤ f ;
(iii) if g is convex and g ≤ f , then g ≤ co(f ).
In other words, co(f ) is the greatest convex function which is below f . We remark that,
differently from the procedure of relaxation, convexification is not a local operation: if f1 = f2
in a neighbourhood U of a point x, it is not guaranteed that co(f1 )(x) = co(f2 )(x).
Proposition 3.8. Let f be as in Definition 3.7. Then

N
X
co(f )(x) = inf λi f (xi ) : N ∈ N, x1 , . . . , xN ∈ X ,
i=1
(3.7)
N
X N
X
λ1 , . . . , λN > 0, λi = 1, x = λi xi .
i=1 i=1
Proof. Let x ∈ X , N ∈ N, λi > 0, xi ∈ X be such that N

P PN
i=1 λi = 1 and i=1 λi xi = x.
Denote by g the right-hand side of (3.7). Since co(f ) is convex and co(f ) ≤ f we have
N
X N
X
co(f )(x) ≤ λi co(f )(xi ) ≤ λi f (xi ),
i=1 i=1
therefore co(f ) ≤ g. Conversely, we have that g ≤ f (trivially) and if we prove that g is

convex, we can conclude that g ≤ co(f ) by definition of co(f ). Therefore we only need to
show that g is convex.
We fix x, y ∈ X and t ∈ (0, 1), and we need to show that
g(tx + (1 − t)y) ≤ tg(x) + (1 − t)g(y). (3.8)
Notice first that g > −∞, thanks to the lower bound on f in the assumption. We can assume
without loss of generality that g(x) < +∞ and g(y) < +∞ (otherwise the inequality (3.8) is
trivial). Fix ε > 0; by definition of g, there exist:
60
PN PN
• N ∈ N, λi > 0, xi ∈ X with i=1 λi = 1, i=1 λi xi = x such that
N
X
λi f (xi ) < g(x) + ε,
i=1
PM PM
• M ∈ N, µi > 0, yi ∈ X with i=1 µi = 1, i=1 µi yi = y such that
M
X
µi f (yi ) < g(y) + ε.
i=1
Then we consider the coefficients tλ1 , . . . , tλN , (1−t)µ1 , . . . , (1−t)µM , and the points x1 , . . . , xN ,
y1 , . . . , yM : we have
N
X M
X N
X M
X
tλi + (1 − t)µi = 1, tλi xi + (1 − t)µi yi = tx + (1 − t)y,
i=1 i=1 i=1 i=1
and therefore
N
X M
X
g(tx + (1 − t)y) ≤ tλi f (xi ) + (1 − t)µi f (yi )
i=1 i=1

≤ t g(x) + ε + (1 − t) g(y) + ε = tg(x) + (1 − t)g(y) + ε.
We obtain (3.8) by letting ε → 0.
Remark 3.9. If x = N
P PN
i=1 λi xi , by setting yi = xi − x we have i=1 λi yi = 0, and from (3.7)
we obtain the equivalent representation
N
X
co(f )(x) = inf λi f (x + yi ) : N ∈ N, y1 , . . . , yN ∈ X ,
i=1
(3.9)
N
X N
X
λ1 , . . . , λN > 0, λi = 1, λi yi = 0 .
i=1 i=1
In the case X = Rn , if the function f : Rn → R is continuous, then by density we can replace

the infimum in (3.9) by the infimum on a countable set:
N
X
co(f )(x) = inf λi f (x + yi ) : N ∈ N, y1 , . . . , yN ∈ Qn ,
i=1
(3.10)
N
X N
X
λi ∈ Q ∩ (0, 1), λi = 1, λi yi = 0 .
i=1 i=1
3.3 Relaxation of integral functionals

Thanks to the tool of convexification, we are now ready to state and prove the main result of
this chapter.
61
Theorem 3.10. Let Ω ⊂ Rn be open and bounded with Lipschitz boundary, 1 ≤ p < +∞,
and let f : Ω × Rn → [0, +∞) satisfy the Carathéodory condition, that is
• the map x 7→ f (x, ξ) is measurable with respect to L(Ω), for every fixed ξ ∈ Rn ;
• the map ξ 7→ f (x, ξ) is continuous on Rn , for almost every fixed x ∈ Ω.

Assume also that
• 0 ≤ f (x, ξ) ≤ a(x) + b|ξ|p for some a ∈ L1 (Ω), a ≥ 0, b ≥ 0.
Let F : W 1,p (Ω) → R be the functional defined by
Z
F(u) := f (x, Du(x)) dx, u ∈ W 1,p (Ω). (3.11)
Ω
Then the lower semicontinuous envelope of F, with respect to the weak topology of W 1,p (Ω),
is the functional Z
F(u) = coξ (f )(x, Du(x)) dx, (3.12)
Ω
where coξ (f )(x, ·) is the convex envelope of the map ξ 7→ f (x, ξ) for a.e. fixed x ∈ Ω.
We also have an equivalent version of the theorem, in which we include boundary condi-
tions. Notice that this second theorem does not follow from the previous one, but requires a
complete proof.
Theorem 3.11. Let Ω ⊂ Rn be open and bounded, 1 ≤ p < +∞, and let f : Ω×Rn → [0, +∞)
satisfy the assumptions of Theorem 3.10. Fix w ∈ W 1,p (Ω) and let Fw : W 1,p (Ω) → R be the
functional defined by
Z
 f (x, Du(x)) dx if u − w ∈ W 1,p (Ω),
0
Fw (u) := Ω (3.13)
+∞ otherwise.

Then the lower semicontinuous envelope of Fw , with respect to the weak topology of W 1,p (Ω),
is the functional
Z
 co (f )(x, Du(x)) dx if u − w ∈ W 1,p (Ω),
ξ 0
F w (u) = Ω (3.14)
+∞ otherwise.

Remark 3.12. The second theorem does not follow directly from the first one. Indeed, notice
that we can write Fw = F + χw+W 1,p (Ω) . In general, if X is a metric space, F : X → R,
0
and K ⊂ X , the equality F + χK = F + χK is not trivial, since χK is not continuous (recall
Exercise 17). If you try to check the two conditions in Proposition 3.4:
(i) let xk → x, we would like to have F(x) + χK (x) ≤ lim inf k→∞ (F(xk ) + χK (xk )): this
is true if K is closed;
(ii) the problem is to construct a sequence xk → x such that F(xk )+χK (xk ) → F(x)+χK (x):
you have for free a sequence such that F(xk ) → F(x), but you have to guarantee that
xk ∈ K for every k, which might not be true!
62
We premise a couple of lemmas to the proof of the theorems. In the first, we show that a
convex and locally bounded function is locally Lipschitz continuous.
Lemma 3.13. Let X be a Banach space, U ⊂ X be open and convex, and let f : U → R
be a convex function. Let x0 ∈ U and R > 0 such that B2R (x0 ) ⊂ U . Assume that there
are constants m, M ∈ R such that m ≤ f (x) ≤ M for all x ∈ B2R (x0 ). Then f is Lipschitz
−m
continuous in BR (x0 ), with Lip(f ) ≤ MR .
Proof. We first prove the lemma in the one-dimensional case. In this case, the proof reduces
to the following statement: let 0 < a < b, and let f : [−b, b] → [m, M ] be a convex function;
then
M −m
|f (x) − f (y)| ≤ |x − y| for all x, y ∈ [−a, a].
b−a
Assume without loss of generality that −a ≤ x < y ≤ a. Assume also that f (x) < f (y) (the
other case is analogous). By convexity, the point (b, f (b)) is above the line passing through
the points (x, f (x)), (y, f (y)). Therefore (see Figure 3.1) the slope of the line passing through
the points (x, f (x)), (y, f (y)), which is the incremental quotient f (y)−f
y−x
(x)
, must be bounded
−m
from above by the slope of the line passing through the points (a, m), (b, M ), which is Mb−a .
The required estimate follows: this proves the lemma in the one-dimensional case.
M
B2R (x0 )
BR (x0 )
x0 y
−b −a x y a b
Figure 3.1: Construction for the proof of Lemma 3.13. Left: the slope of the blue line must
be smaller than the slope of the red line.
In the general case, the idea is to reduce to the one-dimensional case by considering the
restriction of f to the line passing through x and y. Let g(t) := f (x + t(y − x)), for t ∈ R.
We have g(0) = f (x), g(1) = f (y). Let α0 < α < 0 < β < β 0 the values of the parameter t
corresponding to the intersections of the line {x + t(y − x)} with the boundaries of the balls
BR (x0 ), B2R (x0 ), as in Figure 3.1. Then, since m ≤ g(t) ≤ M for all t ∈ (α0 , β 0 ), by the
one-dimensional case we have that
M −m
|f (y) − f (x)| = |g(1) − g(0)| ≤ .
β0 − β
To conclude the proof, it is enough to show that β 0 1−β ≤ |x−y|

R . By setting x1
:= x + β(y − x) ∈
0 0
∂BR (x0 ), x2 := x + β (y − x) ∈ ∂B2R (x0 ), we have (β − β)|x − y| = |x1 − x2 | ≥ R, as
claimed.
63
Lemma 3.14. Under the assumptions of Theorem 3.10, the function coξ (f ) satisfies the
Carathéodory condition.
Proof. Notice that, as ξ 7→ f (x, ξ) is continuous for a.e. x ∈ Ω, in view of Remark 3.9 we can
write the representation formula
N
X
coξ (f )(x, ξ) = inf λi f (x, ξ + ηi ) : N ∈ N, η1 , . . . , ηN ∈ Qn ,
i=1
N
X N
X
λi ∈ Q ∩ (0, 1), λi = 1, λi ηi = 0
i=1 i=1
which holds forPa.e. x ∈ Ω and for every ξ ∈ Rn . For every N , {λi } and {ηi } fixed the
N n
map (x, ξ) 7→ i=1 λi f (x, ξ + ηi ) is measurable with respect to L(Ω) ⊗ B(R ), since f is
Carathéodory (we proved this before: see the proof of Proposition 2.48). Then coξ (f ) is the
infimum of a countable family of measurable functions, and therefore it is measurable with
respect to L(Ω) ⊗ B(Rn ).
To prove the lemma, it remains to show that the map ξ 7→ coξ (f )(x, ξ) is continuous for
a.e. x. By assumption we have
0 ≤ f (x, ξ) ≤ a(x) + b|ξ|p
which implies, since the map ξ 7→ a(x) + b|ξ|p is convex and therefore coincides with its
convexification,
0 ≤ coξ (f )(x, ξ) ≤ a(x) + b|ξ|p .
By applying Lemma 3.13 to coξ (f )(x, ·) in B2R (0), with m = 0, M = a(x)+b(2R)p , we obtain
a(x) + b(2R)p
|coξ (f )(x, ξ1 ) − coξ (f )(x, ξ2 )| ≤ |ξ1 − ξ2 | for all ξ1 , ξ2 ∈ BR (0),
R
which implies the continuity of ξ 7→ coξ (f )(x, ξ).
Proof of Theorem 3.10 and Theorem 3.11. Define for u ∈ W 1,p (Ω) the functionals
Z
G(u) := coξ (f )(x, Du(x)) dx, (3.15)
Ω
Z
 co (f )(x, Du(x)) dx
ξ if u − w ∈ W01,p (Ω),
Gw (u) = Ω (3.16)
+∞ otherwise.

The goal is to show that G = F and Gw = F w .

Since coξ (f ) is a Carathéodory function and is convex in the second variable, the func-
tionals G, Gw are well-defined, and by the results in Section 2.3 (see Proposition 2.50) they are
lower semicontinuous with respect to the weak topology of W 1,p (Ω). Moreover G(u) ≤ F(u)
and Gw (u) ≤ Fw (u) for all u ∈ W 1,p (Ω), since coξ (f ) ≤ f . Therefore
G ≤ F, Gw ≤ F w .
We need to prove the opposite inequalities.
64
Step 1. We define a “fake” convexification
n
co1ξ (f )(x, ξ) := inf λf (x, ξ + ξ 1 ) + (1 − λ)f (x, ξ + ξ 2 ) : ξ 1 , ξ 2 ∈ Rn , λ ∈ (0, 1),
o
λξ 1 + (1 − λ)ξ 2 = 0
n
= inf λf (x, ξ + ξ 1 ) + (1 − λ)f (x, ξ + ξ 2 ) : ξ 1 , ξ 2 ∈ Qn , λ ∈ (0, 1) ∩ Q,
o
λξ 1 + (1 − λ)ξ 2 = 0
where the second equality follows from the fact that f is by assumption continuous with
respect to the variable ξ. Since we can write co1ξ (f ) as the infimum on a countable family, it
follows that
co1ξ (f ) is measurable with respect to L(Ω) ⊗ B(Rn ). (3.17)
By iteration, we define for k ≥ 0
cok+1
ξ (f )(x, ξ) := co1ξ (cokξ (f ))(x, ξ).
Let us omit in this first step the dependence on x, which plays the role of a fixed parameter.
The sequence (cokξ (f ))k is decreasing, since co1ξ (f ) ≤ f . We claim that
cokξ (f ) → coξ (f ) as k → ∞. (3.18)
Denote by g the limit of cokξ (f ), which exists by monotonicity. Since coξ (f ) ≤ cokξ (f ) for
every k, we have coξ (f ) ≤ g. Let us prove the opposite inequality. In order to do so, we show
that g is convex: for fixed ε > 0, ξ 1 , ξ 2 ∈ Rn , λ ∈ (0, 1), we can find k ∈ N such that
g(ξ 1 ) + ε > cokξ (f )(ξ 1 ), g(ξ 2 ) + ε > cokξ (f )(ξ 2 ),
and therefore
g λξ 1 + (1 − λ)ξ 2 ≤ cok+1 (f ) λξ 1 + (1 − λ)ξ 2

ξ
≤ λ cokξ (f )(ξ 1 ) + (1 − λ) cokξ (f )(ξ 2 )
≤ λ g(ξ 1 ) + ε + (1 − λ) g(ξ 2 ) + ε

≤ λg(ξ 1 ) + (1 − λ)g(ξ 2 ) + ε,
which shows that g is convex since ε is arbitrary.

Therefore g is a convex function and g ≤ f , hence g ≤ coξ (f ) (as the convex envelope
is the largest convex function below f ). Since we already proved the opposite inequality, we
conclude that the claim (3.18) holds.
Step 2. We remark that we cannot say that the “fake” convexification co1ξ (f ) is convex in
the second variable, neither continuous (indeed, in the proof of the continuity of coξ (f ) in
Lemma 3.14 we heavily used the convexity). However, we can say that
co1ξ (f ) is upper semicontinuous, (3.19)
since the infimum of a family of continuous functions is upper semicontinuous.
65
Concerning cokξ (f ), by using the upper semicontinuity we can compute the infimum in its
definition on a countable dense set, so that also for this function we obtain
cokξ (f ) is L(Ω) ⊗ B(Rn )-measurable and upper semicontinuous. (3.20)
We now set
Z
 cok (f )(x, Du(x)) dx
ξ if u − w ∈ W01,p (Ω),
Fwk (u) := Ω
+∞ otherwise.

This definition makes sense by (3.20). We claim that
F w ≤ Fw1 , Fw1 ≤ Fw2 , Fw2 ≤ Fw3 , . . . (3.21)
Let us postpone the proof of the claim (3.21), and let us show that it implies the conclusion
of the theorem. Indeed, if (3.21) is true then we
F w ≤ Fw1 ≤ Fw ,
so that taking the lower semicontinuous envelope of each functional in the previous chain of
inequalities we obtain F w ≤ Fw1 ≤ F w , that is Fw1 = F w . Now we iterate:
(3.21)
F w = Fw1 ≤ Fw2 ≤ Fw1 ,
(3.21)
F w = Fw2 ≤ Fw3 ≤ Fw2 , . . .
By induction we obtain that for every k ∈ N and u ∈ W 1,p (Ω) such that u − w ∈ W01,p (Ω)
Z Z
k k k→∞
F w (u) ≤ Fw (u) = coξ (f )(x, Du(x)) dx −→ coξ (f )(x, Du(x)) dx = Gw (u)
Ω Ω
where the convergence follows by the monotone convergence theorem, using (3.18) and the
bounds on f . This proves that F w (u) ≤ Gw (u) for u ∈ w +W01,p (Ω) (the case u ∈
/ w +W01,p (Ω)
is trivial). The proof of the inequality F ≤ G is analogous.
Step 3. To conclude the proof, it only remains to show the claim (3.21). Notice that it is
enough to show that
F w ≤ Fw1 for all u ∈ w + W01,p (Ω). (3.22)
We assume for simplicity that w = 0. It is enough to prove (3.22) for every u in a set
D ⊂ W01,p (Ω) which is dense with respect to the strong topology. Indeed, let u ∈ W01,p (Ω)
and take an approximating sequence (uk )k ⊂ D, uk → u strongly in W 1,p (Ω). Then, if (3.22)
holds in D, we have
(3.22)
F 0 (u) ≤ lim inf F 0 (uk ) ≤ lim inf F01 (uk ) ≤ F01 (u). (3.23)
k→∞ k→∞
To prove the last inequality (which means that F01 is upper semicontinuous with respect to
the strong convergence in W 1,p (Ω)) we can argue as follows: up to subsequences, we can
66
assume that Duk (x) → Du(x) pointwise for a.e. x ∈ Ω, therefore since co1ξ (f )(x, ·) is upper
semicontinuous by (3.19) we get
co1ξ (f )(x, Du(x)) ≥ lim sup co1ξ (f )(x, Duk (x)) for a.e. x ∈ Ω.
k→∞
In turn, for a.e. x ∈ Ω

h i
−co1ξ (f )(x, Du(x)) + a(x) + b|Du(x)|p ≤ lim inf −co1ξ (f )(x, Duk (x)) + a(x) + b|Duk (x)|p .
k→∞
Notice that the quantity in brackets is nonnegative, in view of the assumption on f (which
holds also for co1ξ (f )); then by integrating the previous inequality on Ω and applying Fatou’s
Lemma we get
Z Z Z Z
1 p 1 p
−F0 (u) + a(x) dx + b |Du| dx ≤ lim inf −F0 (uk ) + a(x) dx + b |Duk | dx ,
Ω Ω k→∞ Ω Ω
from which the last inequality in (3.23) follows. Therefore (3.23) shows that it is sufficient to
prove (3.22) for u in a strongly dense set D, which will be our next task.
Step 4. We choose as dense set D the set of piecewise affine functions:
n
D := u ∈ W01,p (Ω) : there exist Ω1 , . . . , Ωk ⊂ Ω open, pairwise disjoint,
[ k o (3.24)
Ln Ω\ Ωi = 0, u|Ωi affine for every i = 1, . . . , k .
i=1
It can be proved that this set is dense in W01,p (Ω) with respect to the strong convergence in
W 1,p (Ω), see [2, Theorem 12.15].1 We need to prove (3.22) for every u ∈ D.
Let U ⊂ Ω be open, and let v = uξ + c be an affine function in U , where uξ (x) := x · ξ.
Let ε > 0. We claim that
there exist uk ∈ W 1,p (U ) such that uk * v weakly in W 1,p (U ), uk = v near ∂U,

(3.25)
Z Z
lim sup f (x, Duk (x)) dx ≤ co1ξ (f )(x, ξ) dx + ε.
k→∞ U U
If this is true, then we can prove (3.22) for u ∈ D: indeed, if u ∈ D let Ω1 , . . . Ωm be a

partition of Ω such that u|Ωi = uξi + ci for some ξi ∈ Rn , ci ∈ R. Then, using (3.25) with
U = Ωi and v = u|Ωi we obtain, for every i = 1, . . . , m, a sequence (uik )k satisfying (3.25).
By defining vk := uik in Ωi , we obtain a function vk ∈ W01,p (Ω) (notice that you do not create
1
Here we are discussing the problem with boundary conditions, Theorem 3.11, and therefore it is enough to
consider a dense set in W01,p (Ω). For the case without boundary conditions, Theorem 3.10, we need to consider
a dense set in W 1,p (Ω): also in this case, it can be proved that piecewise affine functions are dense in W 1,p (Ω),
provided that ∂Ω is Lipschitz, see for instance [7, Theorem 11.40]. This is the reason why Theorem 3.10 has
an extra assumption on Ω.
67
jumps on the boundary of Ωi , since uik = u near ∂Ωi ), with vk * u weakly in W 1,p (Ω), and
Z
F 0 (u) ≤ lim inf F0 (vk ) = lim inf f (x, Dvk (x)) dx
k→∞ k→∞ Ω
m Z
X
= lim inf f (x, Duik (x)) dx
k→∞ Ωi
i=1
m
X Z
≤ lim sup f (x, Duik (x)) dx
i=1 k→∞ Ωi
m Z
(3.25) X
≤ co1ξ (f )(x, ξi ) dx + mε
i=1 Ωi
Z
= co1ξ (f )(x, Du(x)) dx + mε = F01 (u) + mε.
Ω
Since ε is arbitrary, we obtain F 0 (u) ≤ F01 (u) for all u ∈ D.

Step 5. The final missing step is to show that the construction (3.25) is possible. Assume
without loss of generality that v = uξ for ξ ∈ Rn . We need to approximate (in the weak
topology) the affine function uξ in an “optimal way”, in the sense that we want that the
energies F(uk ) “approximate” the energy F 1 (uξ ). The idea is to approximate the slope ξ of
uξ with a fine microstructure, as we did in the proof of Theorem 2.61.
Fix ε > 0. Recalling the definition of co1ξ (f ) as an infimum over a countable set of
parameters (ξk1 , ξk2 , λk )k∈N , we define inductively the Lebesgue measurable sets
n o
E1 := x ∈ U : co1ξ (f )(x, ξ) + ε > λ1 f (x, ξ + ξ11 ) + (1 − λ1 )f (x, ξ + ξ12 ) ,
n k−1
[ o
Ek := x ∈ U \ Ei : co1ξ (f )(x, ξ) + ε > λk f (x, ξ + ξk1 ) + (1 − λk )f (x, ξ + ξk2 ) .
i=1
By definition of co1ξ (f ), every point x ∈ U belongs to one of the sets Ei , that is U = ∞

S
i=1 Ei
is a measurable partition of U . We want to replace the sets Ek by open sets. To do so, we use
absolute continuity of theR integral: since f (·, ξ) ∈PL1 (U ), for every ε > 0 there exists δ > 0
such that if |A| < δ then A f (x, ξ) dx < ε. Since ∞ i=1 |Ei | = |U | < +∞, we can find k0 ∈ N
such that
∞
X
|Ek | < δ.
k=k0 +1
For i = 1, . . . , k0 we can find a compact set Ki ⊂ Ei such that |Ei \Ki | < kδ0 , and pairwise
disjoint open sets Ui such that Ki ⊂ Ui and |Ui \Ki | < ηi , where ηi > 0 is a small parameter
to be chosen. Fix a cut-off function ϕi ∈ Cc∞ (Ui ), ϕi = 1 in Ki , 0 ≤ ϕi ≤ 1.
By using the “zig-zag” construction introduced in the proof of Theorem 2.61, we can split
each set Ui as
∗ ∗
Ui = Ui1,k ∪ Ui2,k ∪ Nik with |Nik | = 0, 1U 1,k * λi , 1U 2,k * 1 − λi weakly* in L∞ as k → ∞,
i i
and construct functions

(
∗ ξ + ξi1 in Ui1,k ,
uki * uξ weakly* in W 1,∞
(Ui ) with Duki =
ξ + ξi2 in Ui2,k .
68
We finally glue together the functions uki by means of the cut-off functions ϕi : we define
k0
X k0
X
uk := ϕi uki + 1 − ϕi uξ .
i=1 i=1
By construction we have uk ∈ W 1,p (U ), uk * uξ weakly in W 1,p (U ), and uk = uξ near ∂U .

Observe also that, since the sets Ui are pairwise disjoint, then on Ui all the functions ϕj
vanish except for ϕi , hence
uk = ϕi uki + (1 − ϕi )uξ on Ui .
Now
Z Z Z
f (x, Duk ) dx = f (x, Duki ) dx + f (x, ϕi Duki + (1 − ϕi )ξ + (uki − uξ )Dϕi ) dx
Ui Ki Ui \Ki
Z Z
1
≤ f (x, ξ + ξi )1U 1,k (x) dx + f (x, ξ + ξi2 )1U 2,k (x) dx
Ki i Ki i
Z h i
a(x) + bcp |ξ|p + |ξi1 |p + |ξi2 |p + bcp ci |uki − uξ |p dx.

+
Ui \Ki
Recall that |Ui \Ki | < ηi ; by choosing ηi such that

Z
ε ε
bcp |ξ|p + |ξi1 |p + |ξi2 |p |Ui \Ki | <

a(x) dx < , ,
Ui \Ki 2k 0 2k 0
we obtain for every i = 1, . . . , k0

Z
lim sup f (x, Duk ) dx
k→∞ Ui
Z Z
1
≤ lim sup f (x, ξ + ξi )1U 1,k (x) dx + lim sup f (x, ξ + ξi2 )1U 2,k (x) dx
k→∞ Ki i k→∞ Ki i
Z
ε
+ + lim sup bcp ci |uki − uξ |p dx
k0 k→∞ Ui \Ki
Z h i ε
= λi f (x, ξ + ξi1 ) + (1 − λi )f (x, ξ − ξi2 ) dx +
k0
ZKi
ε
≤ co1ξ (f )(x, ξ) dx + ε|Ki | + .
Ki k 0
By summing over i = 1, . . . , k0 we obtain

Z k0
X Z
lim sup Sk0 f (x, Duk ) dx ≤ lim sup f (x, Duk ) dx
k→∞ i=1 Ui i=1 k→∞ Ui
k0 Z
X ε
≤ co1ξ (f )(x, ξ) dx + ε|Ki | +
Ki k0
Zi=1
≤ co1ξ (f )(x, ξ) dx + ε|U | + ε.
U
69
On the other hand, the part outside the sets Ki , i = 1, . . . , k0 , is negligible:
Z Z Z
Sk f (x, Duk ) dx = Sk f (x, ξ) dx ≤ Sk f (x, ξ) dx
U\ 0 Ui U\ 0 Ui U\ 0 Ki
i=1 i=1 i=1
Z Z
= S∞ f (x, ξ) dx + Sk0 f (x, ξ) dx ≤ ε + ε
i=k0 +1 Ei i=1 Ei \Ki
S∞ Sk 0
since | i=k0 +1 Ei | < δ, | i=1 Ei \Ki |
< δ. We finally have proved that
Z Z
lim sup f (x, Duk ) dx ≤ co1ξ (f )(x, ξ) dx + ε|U | + 3ε
k→∞ U U
which completes the proof.
3.4 Characterization by sequences of the relaxation

In Theorem 3.10 and Theorem 3.11 we have computed the lower semicontinuous envelope
of an integral functional with respect to the weak topology of W 1,p (Ω). However, the weak
topology is not metrizable and we are not allowed to use the sequential characterization of the
relaxation in Proposition 3.4. Notice that this would be important in order to characterize
the behaviour of minimizing sequences as in part (iii) of Theorem 3.6.
However, recall that the metrizability of the weak topology is closely related to the sepa-
rability of the space, thanks to the following abstract abstract result in Functional Analysis
(for a proof, see [1, Theorem 3.29]).
Theorem 3.15. Let X be a Banach space such that X 0 is separable. Then the restriction
to any ball BR := {x ∈ X : kxk < R} of the weak topology is metrizable. A distance which
generates the weak topology is given by
∞
X |hfi , x − yi|
d(x, y) := ,
2n
n=1
where (fn )n is a dense sequence in the unit ball of the dual {f ∈ X 0 : kf k ≤ 1}.
We recall that the theorem can be applied if, for instance, X is reflexive and separable,
as this implies that X 0 is reflexive and separable ([1, Corollary 3.27]).
In view of the previous result, we are allowed to use the sequential characterization of
the relaxation if we work on balls: this is not restrictive if we assume that the functional is
coercive. In this case we can still prove the existence of a recovery sequence, that is property
(ii) in Proposition 3.4 (while property (i) is always true in a general topological space).
Proposition 3.16. Let X be a separable reflexive Banach space, and let F : X → R be weakly
coercive. Then the sequential characterization of the l.s.c. envelope F given in Proposition 3.4
holds, namely:
for every x ∈ X there exists a sequence xk * x such that F(x) = lim F(xk ).
k→∞
70
Proof. The result is trivial if F(x) = +∞. Assume then that F(x) < +∞ and let t ∈ R be
such that F(x) < t. Let Xt denote the weak closure of the sublevel set {F ≤ t}: since F
is weakly coercive, the set Xt is weakly compact and therefore bounded (Exercise 19): this
implies that the restriction to Xt of the weak topology is metrizable, by Theorem 3.15. In
the following, we endow Xt with the topology obtained as the restriction to Xt of the weak
topology of X .
Now we have
t > F(x) = sup inf F(y)
U ∈Nw (x) y∈U
(where Nw (x) denotes the family of weak neighbourhoods of x in X ), and therefore
inf F(y) < t for every U ∈ Nw (x).

y∈U
This implies that:
• for every U ∈ Nw (x) there exists y ∈ U such that F(y) < t, that is Xt ∩ U 6= ∅. This
means that every weak neighbourhood of x has nonempty intersection with Xt , that is
x belongs to the weak closure of Xt . Since Xt is by definition weakly closed, we obtain
x ∈ Xt ;
• since t > inf y∈U F(y) = inf y∈U ∩Xt F(y) for every U ∈ Nw (x), we have
F(x) = sup inf F(y).

U ∈Nw (x) y∈U ∩Xt
Let now Ft denote the restriction of F to Xt . Recall that the neighbourhoods of x in Xt are
of the form U ∩ Xt , for U ∈ Nw (x). Then the relaxation F t of Ft with respect to the topology
of Xt is
F t (x) = sup inf F(y) = F(x).
U ∈Nw (x) y∈U ∩Xt
Now for F t (x) we can use the sequential characterization, since it is the relaxation of Ft
with respect to a metrizable topology: therefore there exists a sequence (xk )k ⊂ Xt such that
xk * x (in Xt , and therefore weakly in X ) and
lim F(xk ) = F t (x) = F(x),

k→∞
which is what we had to prove.
Exercise 19. Let X be a Banach space and let K ⊂ X be weakly compact. Prove that K is
bounded in the norm of X .[17]
Going back to the setting of integral functionals, assume that:
• 1 < p < ∞ (notice that we now exclude p = 1!),

0
• w ∈ W 1,p (Ω), g ∈ Lp (Ω) are given,
71
• f : Ω × Rn → R ∪ {+∞} is a Carathéodory function satisfying the bounds
a0 (x) + b0 |ξ|p ≤ f (x, ξ) ≤ a(x) + b|ξ|p
for some a0 ∈ L1 (Ω), b0 > 0, a ∈ L1 (Ω), a ≥ 0, b ≥ 0.

Consider the functionals defined for u ∈ W 1,p (Ω) by
Z
 f (x, Du(x)) dx if u − w ∈ W 1,p (Ω), Z
0
Fw (u) := Ω G(u) := g(x)u(x) dx,

+∞ otherwise, Ω
and the minimum problem

n o
(I) := inf Fw (u) + G(u) : u ∈ W 1,p (Ω) . (3.26)
Then we have:
(a) the relaxation of Fw + G with respect to the weak topology of W 1,p (Ω) is F w + G (using
Exercise 17, since G is continuous with respect to the weak topology);
(b) by Theorem 3.11, the l.s.c. envelope F w is given by

Z
 co (f )(x, Du(x)) dx if u − w ∈ W 1,p (Ω),
ξ 0
F w (u) := Ω
+∞ otherwise;

(c) the space W 1,p (Ω) is separable and reflexive (since p > 1), and Fw is weakly coercive
on W 1,p (Ω) thanks to the growth assumption from below.
Therefore we can apply the abstract Theorem 3.6 (also part (iii), since by (c) we can apply
Proposition 3.16) and obtain that the minimum
n o
(M ) := min F w (u) + G(u) : u ∈ W01,p (Ω) (3.27)
is attained, with
(M ) = (I).
Moreover every minimizing sequence for (I) has a subsequence which weakly converges to a
minimum point of (M ), and conversely if u is a minimum point of (M) then there exists a
minimizing sequence (uk ) for (I) such that uk * u weakly in W 1,p (Ω).
Notes
[15]
Solution to Exercise 17. We first remark that F + G is well-defined, as G takes values in R. Then by
Definition 3.1
n o
F + G = sup H : H is l.s.c., H ≤ F + G
n o
= sup H − G : H is l.s.c., H ≤ F + G + G
n o
= sup H − G : H − G is l.s.c., H − G ≤ F + G = F + G.
72
[16]
Solution to Exercise 18. By (3.5) we have immediately that
\
{F ≤ s} ⊂ {F ≤ t}.
t>s
T
To show the opposite inclusion, let x ∈ t>s {F ≤ t}, and assume by contradiction that x ∈/ {F ≤ s}, that is
F(x) > s. Then F(x) > s + ε for some ε > 0. By (3.2) there exists U ∈ N (x) such that inf y∈U F(y) > s + ε.
Then U ∩ {F ≤ s + ε} = ∅. This contradicts the fact that x ∈ {F ≤ s + ε}.
We next prove the second property. Since F ≤ F we immediately have epi(F) ⊂ epi(F). Moreover, since
F is lower semicontinuous, its epigraph is closed (Exercise 12): therefore epi(F) ⊂ epi(F). We now prove
the opposite inclusion. Let (x, t) ∈ epi(F), that is F(x) ≤ t. We show that any neighbourhood of (x, t) in
X × R has nonempty intersection with epi(F), which proves that (x, t) ∈ epi(F). We can consider a generic
neighbourhood of (x, t) of the form U × (t − ε, t + ε), where U ∈ N (x) and ε > 0. By Proposition 3.3,
since F(x) ≤ t we have inf y∈U F(y) ≤ t, hence there exists y ∈ U such that F(y) ≤ t + 2ε . Therefore
(y, t + ε/2) ∈ (U × (t − ε, t + ε)) ∩ epi(F), as desided.
[17]
Solution to Exercise 19. The result follow as an application of Banach-Steinhaus Theorem (see [1, Theo-
rem 2.2]). Indeed, for x ∈ X consider the linear map ϕx : X 0 → R defined by
hϕx , x0 i := hx0 , xi for every x0 ∈ X 0 .
It is well-known that kϕx k = supkx0 kX 0 ≤1 |hϕx , x0 i| = kxkX .

We want to apply Banach-Steinhaus Theorem to the family of continuous, linear operators (ϕx )x∈K from
X 0 to R. We need to check the hypothesis
sup |hϕx , x0 i| = sup |hx0 , xi| < ∞ for every x0 ∈ X 0 .

x∈K x∈K
This condition is satisfied since the image x0 (K) of the weakly compact set K through the weakly continuous
function x0 is compact in R, and hence bounded. Therefore by Banach-Steinhaus Theorem we conclude that
sup kϕx k = sup kxkX < ∞,

x∈K x∈K
that is, K is bounded in the norm of X .
73
Chapter 4
Euler-Lagrange equations and

regularity
In this chapter we discuss the connection between the minimization of an integral functional,
as considered in the previous chapters, and a corresponding partial differential equation, the
Euler-Lagrange equation. We will give some ideas of how this point of view can yield the
regularity of the solution, which is in principle very weak since we had to enlarge the space
in order to find a minimizer.
4.1 First variation and Euler-Lagrange equations

Let Ω ⊂ Rn be an open set. We fix some notation to be used in the following. Let
f : Ω × R × Rn → R
be a given function of class C 1 . We will denote the variables of f by (x, s, ξ) ∈ Ω × R × Rn ,

so that f (x, s, ξ) = f (x1 , . . . , xn , s, ξ1 , . . . , ξn ). We also write
∂f
Ds f := , Dξ f := (Dξ1 f, . . . , Dξn f )
∂s
for the partial derivative of f with respect to s and for the gradient of f with respect to
ξ = (ξ1 , . . . , ξn ) respectively. We consider the integral functional
Z
F(u) := f (x, u(x), Du(x)) dx.
Ω
In the next theorem we show that, under suitable growth assumptions on the integrand f
and its derivatives, any minimizer of F is a weak solution to a suitable partial differential
equation.
Theorem 4.1 (Euler-Lagrange equations). Let Ω ⊂ Rn be open and bounded, 1 < p < +∞,
u0 ∈ W 1,p (Ω). Let f : Ω × R × Rn → R satisfy the following assumptions:
• f is of class C 1 (globally);
• the map ξ 7→ f (x, s, ξ) is convex, for every x ∈ Ω and s ∈ R;
74
• a0 (x) + C1 |ξ|p ≤ f (x, s, ξ) ≤ C2 1 + |s|p + |ξ|p for some a0 ∈ L1 (Ω), 0 < C1 < C2 ;

• |Ds f (x, s, ξ)| ≤ C2 (1 + |s|p−1 + |ξ|p−1 ), |Dξ f (x, s, ξ)| ≤ C2 (1 + |s|p−1 + |ξ|p−1 ).
Then there exists a solution ū ∈ W 1,p (Ω) of the minimum problem

Z
1,p
min F(u) := f (x, u(x), Du(x)) dx : u ∈ u0 + W0 (Ω) . (4.1)
Ω
Moreover, every minimizer ū of (4.1) satisfies the following identity:

Z h i
Dξ f (x, ū(x), Dū(x)) · Dv(x) + Ds f (x, ū(x), Dū(x)) v(x) dx = 0 (4.2)
Ω
for all v ∈ W01,p (Ω), which is the weak form of the Euler-Lagrange equations.
Proof. The existence of minimizers follows by the direct method: indeed, the functional is
sequentially lower semicontinuous with respect to the weak convergence in W 1,p (Ω), thanks
to Corollary 2.70. Moreover, the functional is sequentially weakly coercive: indeed, suppose
that uk ∈ u0 + W 1,p (Ω) is a sequence with bounded energy, i.e. F(uk ) ≤ C for every k. Then,
by the growth condition from below on f , we find
Z Z Z
p
a0 (x) dx + C1 |Duk | dx ≤ f (x, uk (x), Duk (x)) dx ≤ C,
Ω Ω Ω
from which we obtain that

sup kDuk kLp (Ω) < ∞. (4.3)
k
In turn, by applying Poincaré inequality (Theorem 2.24) to the sequence uk − u0 ∈ W01,p (Ω)
we find
kuk kLp (Ω) ≤ kuk − u0 kLp (Ω) + ku0 kLp (Ω) ≤ Cp,Ω kDuk − Du0 kLp (Ω) + ku0 kLp (Ω) ≤ C. (4.4)
From (4.3) and (4.4) it follows that the sequence (uk )k is bounded in W 1,p (Ω), and since
this space is reflexive (p > 1) there is a weakly convergent subsequence ukj * u, with
u − u0 ∈ W01,p (Ω). This proves that the functional is coercive.
We next turn to the second part of the statement: we show that any minimizer ū satisfies
(4.2). Let v ∈ W01,p (Ω) and consider, for t ∈ R, the function ut := ū + tv. Since ut − u0 ∈
W01,p (Ω), ut is admissible in the minimum problem (4.1) and by minimality of ū we have
F(ū) ≤ F(ut ) for every t ∈ R. Therefore the function g : R → R (of one real variable)
Z
g(t) := F(ut ) = f (x, ū + tv(x), Dū(x) + tDv(x)) dx
Ω
has a minimum point at t = 0. If we show that g is differentiable at t = 0, it follows that

g 0 (0) = 0. Let us consider the difference quotients of g: we would like to compute
g(t) − g(0)
Z
1 h i
lim = lim f (x, ū + tv, Dū + tDv) − f (x, ū, Dū) dx. (4.5)
t→0 t t→0 t Ω
75
To compute the previous limit, we apply Lebesgue’s Dominated Convergence Theorem. In-
deed, we have as t → 0
1h i
f (x, ū + tv, Dū + tDv) − f (x, ū, Dū) → Ds f (x, ū, Dū)v + Dξ f (x, ū, Dū) · Dv
t
pointwise in Ω, by the differentiability assumption on f . Furthermore, in view of the growth
conditions we have for t ∈ (0, 1)
f (x, ū + tv, Dū + tDv) − f (x, ū, Dū) 1 t d
Z

= f (x, ū + σv, Dū + σDv) dσ
t t 0 dσ
Z t
1
≤ Ds f (x, ū + σv, Dū + σDv)v + Dξ f (x, ū + σv, Dū + σDv) · Dv dσ
t 0
1 t
Z
C2 1 + |ū + σv|p−1 + |Dū + σDv|p−1 |v| + |Dv| dσ

≤
t 0
≤ C 1 + |ū|p−1 + |v|p−1 + |Dū|p−1 + |Dv|p−1 |v| + |Dv| ∈ L1 (Ω).

| {z }| {z }
∈Lp0 (Ω) ∈Lp (Ω)
Therefore it is possible to apply the Dominated Convergence Theorem in (4.5) and obtain
that
g(t) − g(0)
Z
0
0 = g (0) = lim = Ds f (x, ū, Dū) v + Dξ f (x, ū, Dū) · Dv dx
t→0 t Ω
for all v ∈ W01,p (Ω), which yields (4.2).
We now clarify in which sense the integral identity (4.2) is the weak formulation of a
partial differential equation. Suppose, in addition to the assumptions of Theorem 4.1, that f
is of class C 2 and that
ū is a minimizer of class C 2 .
Notice that this is a very strong assumption, since by the direct method we can only prove
existence of minimizers in the Sobolev space W 1,p (Ω)! We can choose any test function
v ∈ Cc∞ (Ω) in (4.2):
Z X n
Dξi f (x, ū(x), Dū(x))Dxi v(x) + Ds f (x, ū(x), Dū(x)) v(x) dx = 0.
Ω i=1
Under the previous assumptions we can integrate by parts (there are no boundary terms since
v has compact support in Ω) and we obtain
Z X n
Dxi Dξi f (x, ū(x), Dū(x)) +Ds f (x, ū(x), Dū(x)) v(x) dx = 0 for all v ∈ Cc∞ (Ω).

−
Ω i=1
Recall now the following result.

Lemma 4.2 (Fundamental Lemma of the Calculus of Variations). Let Ω ⊂ Rn be open, let
f ∈ L1loc (Ω) satisfy Z
f (x)v(x) dx = 0 for every v ∈ Cc∞ (Ω). (4.6)
Ω
Then f = 0 almost everywhere in Ω.
76
By applying the lemma to our case we conclude that
n
X
− Dxi Dξi f (x, ū(x), Dū(x)) + Ds f (x, ū(x), Dū(x)) = 0 in Ω, (4.7)
i=1
which is the Euler-Lagrange equation associated to the functional F. Therefore every regular
minimizer ū of F solves the boundary value problem
(
−div Dξ f (x, ū, Dū) + Ds f (x, ū, Dū) = 0 in Ω,
(4.8)
ū = u0 on ∂Ω.
Remark 4.3. In general, the Euler-Lagrange equation (4.8) might have other solutions which
are not minimizers of F (in this case, they will be called critical points of F). There is an
exception: suppose that the joint map (s, ξ) 7→ f (x, s, ξ) is jointly convex for each x. In this
case every weak solution is in fact a minimizer.
To see this, suppose that ū is a weak solution, that is ū satisfies (4.2). By convexity, we
have for all ξ, η ∈ Rn and for all s, r ∈ R
f (x, r, η) ≥ f (x, s, ξ) + Ds f (x, s, ξ)(r − s) + Dξ f (x, s, ξ) · (η − ξ).
Then, for every w ∈ u0 + W01,p (Ω), by choosing ξ = Dū(x), η = Dw(x), s = ū(x), r = w(x)
in the previous inequality we find
f (x, w, Dw) ≥ f (x, ū, Dū) + Ds f (x, ū, Dū)(w − ū) + Dξ f (x, ū, Dū) · (Dw − Dū),
and integrating over Ω we obtain

Z
F(w) ≥ F(ū) + Ds f (x, ū, Dū)(w − ū) + Dξ f (x, ū, Dū) · (Dw − Dū) dx,
Ω
where the last integral is equal to zero by (4.2), since w − ū ∈ W01,p (Ω) is an admissible test
function. Therefore F(ū) ≤ F(w) for every w ∈ u0 + W01,p (Ω), that is ū is a minimizer.
Example 4.4 (Dirichlet energy and p-Laplacian). Take f (x, s, ξ) = f (ξ) = 21 |ξ|2 . Then the
functional is the Dirichlet energy
Z
1
F(u) = |Du(x)|2 dx, u ∈ W 1,2 (Ω),
2 Ω
and the Euler-Lagrange equation is simply
−∆u = 0 in Ω
(or in weak form Ω Du · Dv dx = 0 for all v ∈ Cc∞ (Ω)).

R
p
More in general, for p ∈ (1, ∞) we can consider f (ξ) = |ξ|p and the corresponding Euler-
Lagrange equation
−div |Du|p−2 Du = 0 in Ω

(the operator on the left-hand side is called p-Laplacian, see Example 2.45).
77
Rs
Example 4.5 (Nonlinear Poisson equation). Let f (x, s, ξ) = f (s, ξ) = 12 |ξ|2 − 0 g(y) dy,
where g : R → R is a given smooth function. In this case the corresponding Euler-Lagrange
equation becomes
−∆u = g(u) in Ω.
p
Example 4.6 (Minimal surfaces equation). Let f (x, s, ξ) = f (ξ) = 1 + |ξ|2 , so that
Z p
F(u) = 1 + |∇u(x)|2 dx.
Ω
Notice that the value of the functional F(u) represents the area of the graph of a function
u : Ω → R. The associated Euler-Lagrange equation is

∇u
−div p = 0 in Ω.
1 + |∇u|2
The operator on the right-hand side represents (up to a constant) the mean curvature of
the graph of u. We then obtain the principle that area-minimizing surfaces have zero mean
curvature.
Example 4.7 (General elliptic equations in divergence form). Let Ω ⊂ Rn be an open and
bounded set, and consider a density f : Ω × R × Rn of the following form.
n
1 X
f (x, s, ξ) := aij (x)ξi ξj − g(x)s,
2
i,j=1
where:
• aij ∈ L∞ (Ω),
• aij = aji for all i, j ∈ {1, . . . , n},
• (ellipticity condition) there exists λ > 0 such that

n
X
aij (x)ξi ξj ≥ λ|ξ|2 for almost every x ∈ Ω, for every ξ ∈ Rn ,
i,j=1
• g ∈ L2 (Ω) is a given function.

The associated functional is
n Z Z
1 X
F(u) = aij (x)Dxi u(x)Dxj u(x) dx − u(x)g(x) dx.
2 Ω Ω
i,j=1
By Theorem 2.53 and Remark 2.54, one obtains that the minimum problem
min F(u)
u∈H01 (Ω)
has a solution ū. Furthermore, ū is a weak solution of the associated Euler-Lagrange equation
( P
− ni,j=1 Dxi aij Dxj ū = g in Ω,

ū = 0 on ∂Ω.
78
4.2 Constraints
In this section we consider an example of constrained minimum problem, in which we seek to
minimize an integral functional subject to an integral constraint.
Let Ω ⊂ Rn be open and bounded. We consider for simplicity the Dirichlet energy
Z
1
F(u) := |∇u(x)|2 dx
2 Ω
(but we could work with more general integral functionals, as those considered in the previous
section). Let also Z
G(u) := g(u(x)) dx,
Ω
where g ∈ C ∞ (R) satisfies the following bound:
|g 0 (s)| ≤ C1 1 + |s|

for all s ∈ R, (4.9)
where C1 is a positive constant. Notice that from (4.9) it also follows
|g(s)| ≤ C2 1 + |s|2

for all s ∈ R,
which in particular implies that the functional G is well-defined and finite for u ∈ L2 (Ω). We
are then interested in the constrained minimum problem
min{F(u) : u ∈ H01 (Ω), G(u) = 0}. (4.10)
We also introduce the admissible class
A := {u ∈ H01 (Ω) : G(u) = 0}. (4.11)
Theorem 4.8. Assume that A =

6 ∅. Then the minimum problem (4.10) has a solution.
Proof. Let (uk )k be a minimizing sequence, so that uk ∈ A and F(uk ) → inf u∈A F(u). Since
k∇ukL2 (Ω) is an equivalent norm on H01 (Ω) by the Poincaré inequality, and supk F(uk ) =
supk 21 k∇uk k2L2 (Ω) < ∞, the sequence (uk )k is bounded in H01 (Ω) and therefore there exists a
weakly convergent subsequence:
ukj * ū weakly in H01 (Ω),
for some ū ∈ H01 (Ω). By lower semicontinuity
F(ū) ≤ lim inf F(ukj ) = inf F.

j→∞ A
To conclude, we only need to show that ū ∈ A, that is G(u) = 0. Notice first that for s < t
we have thanks to the assumption (4.9)
Z t Z t
0

|g(s) − g(t)| ≤ |g (z)| dz ≤ C1 1 + |z| dz ≤ C1 1 + |t| + |s| |t − s|.
s s
79
By Rellich Theorem we have that ukj → u strongly in L2 (Ω), and therefore
Z
|G(ū)| = |G(ū) − G(ukj )| ≤ |g(ū(x)) − g(ukj (x))| dx
Z Ω

≤ C1 1 + |ū(x)| + |ukj (x)| |ū(x) − ukj (x)| dx
Ω
Z 1 Z 1
2 2 2
2
≤ C1 |ū − ukj | dx 1 + |ū| + |ukj | dx →0 as j → ∞,
Ω Ω
and therefore G(ū) = 0.
Theorem 4.9 (Lagrange multiplier). Let Ω ⊂ Rn be open, bounded, connected, with Lipschitz
boundary. Let ū ∈ A be a minimizer for (4.10). Then there exists λ ∈ R such that
Z Z
∇ū · ∇v dx = λ g 0 (ū)v dx for all v ∈ H01 (Ω), (4.12)
Ω Ω
that is, ū is a weak solution to the nonlinear boundary value problem

(
−∆ū = λg 0 (ū) in Ω,
ū = 0 on ∂Ω.
Proof. We first assume that g 0 (ū) is not equal to zero almost everywhere in Ω. We can then
find a function w ∈ H01 (Ω) such that
Z
g 0 (ū)w dx 6= 0. (4.13)
Ω
Fix any function v ∈ H01 (Ω). Notice that we cannot argue as in the proof of Theorem 4.1 and
consider the variation ū + tv, since this function in principle does not belong to the admissible
class A! The idea si then to consider a two-parameter family of variations
uτ,σ (x) := ū + τ v + σw, τ, σ ∈ R,
and to use the Implicit Function Theorem to choose σ = σ(τ ) so that uτ,σ(τ ) satisfies the
constraint G(uτ,σ(τ ) ) = 0.
Step 1. With the previous idea in mind, consider the function
Z
j(τ, σ) := G(uτ,σ ) = g(ū + τ v + σw) dx.
Ω
We have j(0, 0) = G(ū) = 0. Moreover, j is of class C 1 and

Z Z
∂j 0 ∂j
(τ, σ) = g (ū + τ v + σw)v dx, (τ, σ) = g 0 (ū + τ v + σw)w dx.
∂τ Ω ∂σ Ω
Therefore by (4.13) Z
∂j
(0, 0) = g 0 (ū)w dx 6= 0.
∂σ Ω
80
We can then apply the Implicit Function Theorem to obtain that there exists a function
φ : R → R such that φ(0) = 0 and j(τ, φ(τ )) = 0 for all τ ∈ (−τ0 , τ0 ), for some small τ0 > 0.
By differentiating we find
∂j ∂j
(τ, φ(τ )) + (τ, φ(τ ))φ0 (τ ) = 0 for all τ ∈ (−τ0 , τ0 ),
∂τ ∂σ
hence
∂j R 0
0 (0, 0) g (ū)v dx
φ (0) = − ∂τ
∂j
= −RΩ 0 . (4.14)
∂σ (0, 0) Ω g (ū)w dx
Step 2. We now have the one-parameter family of functions uτ,φ(τ ) = ū + τ v + φ(τ )w, defined
for τ ∈ (−τ0 , τ0 ), which satisfy uτ,φ(τ ) ∈ H01 (Ω) and G(uτ,φ(τ ) ) = j(τ, φ(τ )) = 0. Hence
uτ,φ(τ ) ∈ A for all τ ∈ (−τ0 , τ0 ).
The functions uτ,φ(τ ) are therefore admissible in the minimum problem (4.10) and by mini-
mality of ū we have
F(ū) ≤ F(uτ,φ(τ ) ) for all τ ∈ (−τ0 , τ0 ).
Therefore the function i : (−τ0 , τ0 ) → R, i(τ ) := F(uτ,φ(τ ) ) has a minimum at τ = 0, and
therefore
Z

0 = i0 (0) = ∇ū + τ ∇v + φ(τ )∇w · ∇v + φ0 (τ )∇w dx

τ =0
ZΩ
∇ū · ∇v + φ0 (0)∇w dx

=
Ω R 0
g (ū)v dx
Z Z
(4.14) Ω
= ∇ū · ∇v dx − R 0 ∇ū · ∇w dx.
Ω Ω g (ū)w dx Ω
R
∇ū·∇w dx
Therefore, defining λ := RΩ
0 (which is independent of v), we find
Ω g (ū)w dx
Z Z
∇ū · ∇v dx = λ g 0 (ū)v dx,
Ω Ω
as desired.
Step 3. To conclude the proof, it remains to consider the case where g 0 (ū) = 0 almost
everywhere in Ω. We have that D(g(ū)) = g 0 (ū)Du = 0 almost everywhere on Ω. Then,
by Poincaré inequality (Theorem 2.23) it follows that g(ū) is equal to a constant almost
everywhere in Ω, and this constant must be zero since
Z
0 = G(ū) = g(ū) dx.
Ω
Therefore g(ū) = 0 almost everywhere in Ω. By the boundary condition (ū ∈ H01 (Ω)) it
follows that g(0) = 0 and therefore G(0) = 0, which implies that 0 ∈ A. Hence ū = 0 almost
everywhere, or otherwise F(0) = 0 < F(ū) which violates the minimality of ū in A. Then
(4.12) is trivially satisfied.
81
4.3 Regularity of minimizers: Hilbert’s XIX problem and De
Giorgi - Nash Theorem
We have seen in the previous section that, in order to deduce that a minimizer ū of some
integral functional F is a strong solution of the corresponding Euler-Lagrange equation, we
assumed a priori that ū is smooth (at least C 2 ), a fact that in general is not guaranteed.
Indeed, in order to establish the existence of a minimizer via the direct method we had to
work in the larger Sobolev space W 1,p (Ω), so that in general we can only say that a minimizer
ū will be a weak solution of the Euler-Lagrange equation, in the sense of the identity (4.2).
This is a general issue which is connected to the approach to existence of minimizers via
the direct method: one can usually prove the existence of a solution having, in general, less
regularity than expected, since in order to gain compactness of minimizing sequences it is often
necessary to enlarge the space of admissible configurations. A branch of the research in the
Calculus of Variations has been devoted to the regularity theory for minimizers of variational
problems, that is to establish a posteriori the regularity of a minimizer, once its existence has
been proved.
The goal of this section is to give some ideas of this topic, without entering too much into
the technical details, by discussing one of the most deep and important results in regularity
theory, in the framework of intergral functionals.
Hilbert’s XIX problem. Is it true that any minimizer of a “regular integral” of the Calculus
of Variations Z
F(u) := f (∇u(x)) dx (4.15)
Ω
is an analytic function, given that f is analytic?
In other words, Hilbert’s problem asks whether minimizers of an integral functional of the
form (4.15) inherit the regularity properties of the integrand f . In order to better understand
the problem, we need to explain the meaning of “regular integral”. The precise setting is the
following.
Let Ω ⊂ Rn be open and bounded, and let f ∈ C ∞ (Rn ) satisfy the following assumptions:
(i) f is uniformly convex, that is there exists λ > 0 such that

n
X ∂2f
(ξ)ηi ηj ≥ λ|η|2 for all ξ, η ∈ Rn , (4.16)
∂ξi ∂ξj
i,j=1
(ii) the second derivatives of f are uniformly bounded, that is there exists Λ > 0 such that
2
∂ f
for all ξ ∈ Rn .

∂ξi ∂ξj ≤ Λ
(ξ) (4.17)

We consider the associated functional F, defined in (4.15), and we assume that ū ∈ H 1 (Ω) is
a minimizer of F with respect to its own boundary conditions, that is
F(ū) ≤ F(v) for all v ∈ H 1 (Ω) such that v − ū ∈ H01 (Ω). (4.18)
82
Remark 4.10. Notice that, in view of the assumption (4.16) and by a Taylor expansion, we
have the bound from below
λ 2
f (ξ) ≥ f (0) + ∇f (0) · ξ + |ξ| for all ξ ∈ Rn .
2
Then, using Young’s inequality
ε|ξ|2 |∇f (0)|2

∇f (0) · ξ ≥ − −
2 2ε
and choosing ε := λ2 , we find the coercivity condition
λ 2
f (ξ) ≥ |ξ| + β for all ξ ∈ Rn , for some β ∈ R. (4.19)
4
On the other hand, using the assumption (4.17), we have the growth conditions from above
(compare with the assumptions of Theorem 4.1)
|f (ξ)| ≤ C 1 + |ξ|2 , |∇f (ξ)| ≤ C 1 + |ξ| for all ξ ∈ Rn .

(4.20)
Exercise 20. Use the direct method to prove that for a given boundary datum u0 ∈ H 1 (Ω)
the functional F has a unique minimizer ū in the class X := {u ∈ H 1 (Ω) : u − u0 ∈ H01 (Ω)},
and that the minimizer ū is a weak solution of the Euler-Lagrange equation
− div ∇f (∇ū) = 0 in Ω.[18]

(4.21)
Hilbert’s problem asks what is the regularity of a minimizer ū, satisfying (4.18): the direct
method only gives a solution in H 1 (Ω), but can we say that ū is actually smooth?
4.3.1 An elliptic, quasilinear equation for the partial derivatives

The first step towards the regularity of a minimizer ū is to observe that the partial derivatives
of ū are weak solutions of a quasilinear, elliptic partial differential equation. We define the
coefficients
∂2f
aij (x) := (∇ū(x)), A(x) := D2 f (∇ū(x)) = (aij (x))i,j=1,...,n (4.22)
∂ξi ∂ξj
for x ∈ Ω and i, j ∈ {1, . . . , n}. Notice that the coefficients aij depend on the gradient of the
solution ū. They are symmetric (aij = aji ), bounded, measurable functions by (4.17),
aij ∈ L∞ (Ω) with kaij kL∞ (Ω) ≤ Λ, (4.23)
and they satisfy an uniform ellipticity condition by (4.16),

n
X
A(x)[ξ, ξ] = aij (x)ξi ξj ≥ λ|ξ|2 for every ξ ∈ Rn . (4.24)
i,j=1
We then have the following result.
83
2 (Ω) and its partial derivative v := ∂ ū
Theorem 4.11. Let ū satisfy (4.18). Then ū ∈ Hloc k ∂xk
(1 ≤ k ≤ n) satisfies
Z
A(x)[∇vk (x), ∇ϕ(x)] dx = 0 for every ϕ ∈ Cc∞ (Ω), (4.25)
Ω
that is, vk is a weak solution to the elliptic equation

n
X
− Di aij (x)Dj vk (x) = 0. (4.26)
i,j=1
Proof. Recall that ū ∈ H 1 (Ω) solves the Euler-Lagrange equation (4.21), whose weak formu-
lation is Z
∇f (∇ū) · ∇ϕ dx = 0 for every ϕ ∈ H01 (Ω). (4.27)
Ω
The proof uses the method of difference quotients, see [4, Section 5.8.2].
Step 1. Fix a direction k ∈ {1, . . . , n}. For h ∈ R we define the difference quotient in the
direction k
ū(x + hek ) − ū(x)
τh,k ū(x) := .
h
Take now any test function ϕ ∈ H01 (Ω) with supp ϕ ⊂⊂ Ω. If |h| is sufficiently small, the
function τ−h,k ϕ is compactly supported in Ω and can be used as a test function in (4.27):
Z
1
0= ∇f (∇ū(x)) · ∇ ϕ(x − hek ) − ϕ(x) dx
h Ω
n Z
X 1
= Di f (∇ū(x)) · Di ϕ(x − hek ) − ϕ(x) dx (4.28)
h Ω
i=1
n Z
X 1
= Di f (∇ū(x + hek )) − Di f (∇ū(x)) · Di ϕ(x) dx
Ω h
i=1
where the second equality follows by a change of variable. We can write by using the Funda-
mental Theorem of Calculus
1 1Z 1 d
Di f (∇ū(x + hek )) − Di f (∇ū(x)) = Di f s∇ū(x + hek ) + (1 − s)∇ū(x) ds
h h 0 ds
n Z 1
X 1 2

= Dij f s∇ū(x + hek ) + (1 − s)∇ū(x) · Dj ū(x + hek ) − Dj ū(x) ds
h 0
j=1
n n
X (h) Dj ū(x + hek ) − Dj ū(x) X (h)
= bij = bij τh,k (Dj ū(x))
h
j=1 j=1
n
(h)
X
= bij Dj (τh,k ū(x)),
j=1
where we defined the coefficients

Z 1
(h) 2

bij := Dij f s∇ū(x + hek ) + (1 − s)∇ū(x) ds.
0
84
We now insert the previous identity in (4.28) and, setting Ωh := {x ∈ Ω : dist (x, ∂Ω) > h},
we finally obtain that
n Z
(h)
X
for every ϕ ∈ H01 (Ω) with supp ϕ ⊂ Ωh .

bij Dj τh,k ū Di ϕ dx = 0 (4.29)
i,j=1 Ω
In other words, the incremental quotient wh = τh,k ū is a weak solution of the elliptic equation
n
(h)
X
− Di bij Dj wh = 0 locally in Ω. (4.30)
i,j=1
(h)
Notice that the coefficients bij are bounded and uniformly elliptic, as follows immediately
from their definition and from the assumptions (4.16)–(4.17):
n
(h) (h)
X
|bij | ≤ Λ , bij ξi ξj ≥ λ|ξ|2 . (4.31)
i,j=1
Step 2. Since wh is a weak solution to the elliptic equation (4.30), with the coefficients
satisfying the estimates (4.31) uniformly with respect to h, we can show that for every open
sets Ω0 ⊂⊂ Ω00 ⊂⊂ Ω there is a constant C = C(Ω0 , Ω00 , λ, Λ) such that for every h with |h|
sufficiently small it holds Z Z
|∇wh |2 dx ≤ C |wh |2 dx. (4.32)
Ω0 Ω00
To obtain this estimate, take a cut-off function ψ ∈ Cc∞ (Ω00 ), ψ ≥ 0, ψ ≡ 1 in Ω0 , and test
the weak formulation (4.29) of the equation with ϕ = ψ 2 wh : we find
Z n Z n Z
(h) (h)
X X
2 2 2
λ |∇wh | ψ dx ≤ bij Dj wh Di wh ψ dx = −2 bij wh ψDj wh Di ψ dx
Ω i,j=1 Ω i,j=1 Ω
Z 1 Z 1
2 2
≤ 2Λ |∇wh |2 ψ 2 dx |∇ψ|2 wh2 dx ,
Ω Ω
which gives
4Λ2
Z Z Z Z
|∇wh |2 dx ≤ |∇wh |2 ψ 2 dx ≤ 2 |∇ψ|2 wh2 dx ≤ C(Ω0 , Ω00 , Λ, λ) wh2 dx.
Ω0 Ω λ Ω Ω00
This proves the estimate (4.32).

Step 3. We now want to pass to the limit as h → 0 in (4.29).
Notice first that, since wh = τh,k ū and ū ∈ H 1 (Ω), by standard properties of the difference
quotients (see [4, Section 5.8.2]) the right-hand side in (4.32) is bounded by C Ω |∇ū|2 ,
R
uniformly with respect to h. Therefore since the gradient commutes with the operator τh,k
we obtain the uniform bound Z
|τh,k (∇ū)|2 dx ≤ C.
Ω0
By using again standard properties of difference quotients ([4, Section 5.8.2]) this implies that
∇ū ∈ H 1 (Ω0 ) for every Ω0 ⊂⊂ Ω, that is u ∈ Hloc
2 (Ω), and moreover as h → 0
τh,k (∇ū) → Dk (∇ū) = ∇(Dk ū) = ∇vk in L2loc (Ω)
85
(recall that we set vk = Dk ū).
We can finally pass to the limit in (4.29): indeed by Lebesgue’s Dominated Convergence
(h) 2 f (∇ū) = a as h → 0 almost everywhere in Ω, and therefore
Theorem we have that bij → Dij ij
we find
Xn Z
aij Dj vk Di ϕ dx = 0 for every ϕ ∈ H01 (Ω) with supp ϕ ⊂⊂ Ω,
i,j=1 Ω
which implies (4.25).
4.3.2 De Giorgi - Nash Theorem

2 (Ω) and that its partial derivative v := ∂ ū
In view of Theorem 4.11, we know that ū ∈ Hloc ∂xk ,
1 (Ω) of the quasilinear, elliptic equation
for k = 1, . . . , n, is a weak solutions in Hloc
n
X
2

− Di aij Dj v = 0, with aij (x) := Dij f (∇ū(x)). (4.33)
i,j=1
The coefficients aij are in L∞ (Ω) and uniformly elliptic by (4.23)–(4.24). The idea is now
to exploit the regularity theory for solutions to an equation of this form to obtain further
∂ ū
regularity of v = ∂x k
and, in turn, of ū. In particular, we would like to apply the following
regularity result (see for instance [5, Theorem 5.19]).
0,α
Theorem 4.12 (Shauder regularity). Assume that aij ∈ Cloc (Ω) satisfy (4.23)–(4.24). Let
1 1,α
v ∈ Hloc (Ω) be a weak solution to the equation (4.33). Then v ∈ Cloc (Ω).
∂ ū 2,α
This theorem, applied to v = ∂x k
, would yield ū ∈ Cloc (Ω). The main problem here is
that the coefficients (aij )ij in the equation (4.33) are only measurable and bounded, while in
order to apply the previous regularity theorem we would need Hölder continuity of aij !
1,α
Suppose, for a moment, to know for some reason that the solution ū is of class Cloc (Ω),
for some α ∈ (0, 1). Then we could deduce that the coefficients aij = Dij 2 f (∇ū) are more
regular - in particular not just in L∞ , but actually Hölder continuous! It would then be
∂ ū
possible to apply Theorem 4.12 to v = ∂x k
, which is a weak solution to (4.33), and obtain
∂ ū 1,α 2,α
that ∂x k
∈ Cloc (Ω), and hence ū ∈ Cloc (Ω). In turn, by a standard bootstrap argument it
would be possible to improve the regularity of ū up to C ∞ , if f ∈ C ∞ .
This was the state of the art about Hilbert’s XIX problem in the ’50, when Ennio De
Giorgi and John Nash became interested in this question. In particular, it is clear that the
previous strategy has a gap:
• by Theorem 4.11, we know that a minimizer ū is in Hloc 2 (Ω) and its partial derivatives
satisfy the elliptic equation (4.33), with coefficients aij ∈ L∞ (Ω);

1,α 0,α
• if we knew that ū ∈ Cloc (Ω), then aij ∈ Cloc (Ω), we could apply Shauder Theorem 4.12
2,α
to v = Dk ū and obtain that ū ∈ Cloc (Ω). In particular, ū would be a strong solution of
the Euler-Lagrange equation (4.21).
De Giorgi and Nash were able to prove, independently and at the same time, a regularity
result for solution to an elliptic equation of the form (4.33) with bounded and measurable
coefficients, thus filling the missing step and providing a positive answer to Hilbert’s problem.
86
Theorem 4.13 (De Giorgi - Nash). Assume that aij ∈ L∞ (Ω) satisfy (4.23)–(4.24). Let
1 (Ω) be a weak solution to the equation (4.33). Then there exists α ∈ (0, 1) (depending
v ∈ Hloc
0,α
on λ, Λ, and n) such that v ∈ Cloc (Ω).
A proof of the theorem, which follows essentially the original strategy of De Giorgi, can
be found for instance in [5, Theorem 8.13].
The contribution of De Giorgi and Nash to the solution of Hilbert’s problem was a sort
of bridge between the first step of the strategy outlined before (regularity in Sobolev spaces)
and the second step (Shauder regularity theory in Hölder spaces). Notice also that, although
De Giorgi - Nash Theorem is stated as a result for linear partial differential equations, it
actually provides regularity for the solutions to nonlinear equations such as (4.21).
Notes
[18]
Solution to Exercise 20. Thanks to the estimates (4.19)–(4.20), we can directly apply Theorem 4.1 with
p = 2 (notice that the convexity of f follows by (4.16)). In particular, the theorem yields the existence of a
minimizer ū which satisfies the Euler-Lagrange equation
Z
Dξ f (Dū(x)) · Dv(x) dx = 0 for all v ∈ H01 (Ω),
Ω
which is the weak formulation of (4.21). Uniqueness follows by strict convexity of f , see Exercise 14.
87
Chapter 5
Γ-convergence (a short intro)
The notion of Γ-convergence, introduced by E. De Giorgi, is a notion of convergence of

functionals suitable to treat minimization problems. Consider, for instance, a situation in
which one has a family of minimization problems depending on a (small or big) parameter ε
min{Fε (u) : u ∈ Xε }, Fε : Xε → R, (5.1)
which present a “limit” behaviour in some asymptotic regime ε → ε0 . The parameter might
have different meanings (discretization parameter, scaling factor. . . ). The idea is to replace
this family of problems by a single, “effective” problem (independent of ε)
min{F(u) : u ∈ X }, F : X → R, (5.2)
which captures the relevant behaviour of minimizers of the functionals Fε . In particular, we
would like to guarantee that, if uε is a minimizer of (5.1), then uε → u0 as ε → ε0 , with u0
minimizer of (5.2), and Fε (uε ) → F(u0 ). In this way the study of the limit problem (5.2)
gives an approximate description of the solutions to (5.1) in the asymptotic regime.
From the converse point of view, another situation is the following: suppose that you
are interested in a minimum problem (5.2), and you want to approximate it by a family of
problems (5.1) so that the previous convergence properties are satisfied. The advantage here
is that the approximating problems might be easier to study - for instance, numerically.
5.1 Towards the definition of Γ-convergence

We now try to derive the definition of Γ-convergence in a natural way, by looking for conditions
which guarantee the convergence of minima and minimizers. In this derivation, we are inspired
by the direct method of the Calculus of Variations.
Consider a sequence of minimum problems, for j ∈ N,
min{Fj (u) : u ∈ Xj }, Fj : Xj → [0, +∞]. (5.3)
Without loss of generality, we can consider functionals defined over the same space, that is
Xj = X for all j. Indeed, we can always take a larger space X such that Xj ⊂ X for all j,
and extend the functionals Fj to X by setting
(
Fj (u) if u ∈ Xj ,
Fej (u) :=
+∞ otherwise in X \Xj .
88
In this way we have inf Xj Fj = inf X Fej and we can work in a fixed space X , the same for all
functionals of the sequence.
Let then (X , d) be a metric space and Fj : X → [0, +∞]. Let (uj )j∈N be a “generalized
minimizing sequence”, that is such that
Fj (uj ) − inf Fj → 0 (5.4)
X
1
(in particular, we can assume 0 ≤ Fj (uj ) − inf X Fj ≤ j ). We now would like to have the
following properties:
• compactness of the sequence (uj )j : for this, we assume that the functionals (Fj )j are
equicoercive, that is
for every t ∈ R there exists K ⊂ X compact such that {Fj ≤ t} ⊂ K for all j ∈ N.
(5.5)
If this property is satisfied, then we can deduce that uj ∈ K for all j ∈ N, for some
compact set K ⊂ X . Therefore, up to subsequences we have that uj → ū as j → ∞.
We would like to deduce that ū is a minimizer of some limit functional F.
• Lower bound for the sequence of infima: we would like to have

F(ū) ≤ lim inf Fj (uj ) = lim inf inf Fj . (5.6)
j→∞ j→∞ X
This is guaranteed if we assume the liminf inequality

F(v) ≤ lim inf Fj (vj ) for all v ∈ X and for all vj → v. (5.7)
j→∞
• Upper bound for the sequence of infima: we would like to have

lim sup inf Fj ≤ inf F ≤ F(ū). (5.8)
j→∞ X X
This is surely guaranteed if we assume that lim supj→∞ (inf X Fj ) ≤ F(u) for all u ∈ X .
By localizing thus condition around u, we can assume the stronger condition that
for all u ∈ X , for all δ > 0 lim sup inf{Fj (v) : d(v, u) < δ} ≤ F(u).
j→∞
This is in turn equivalent to the limsup inequality

for every u ∈ X there exists vj ∈ X , vj → u such that lim sup Fj (vj ) ≤ F(u). (5.9)
j→∞
If we now have all the previous properties, that is the equicoercivity (5.5), the liminf inequality
(5.7), and the limsup inequality (5.9), we obtain that for a generalized minimizing sequence
(uj )j it holds uj → ū (by equicoercivity) and
(5.7)
inf F ≤ F(ū) ≤ lim inf Fj (uj )
X j→∞

= lim inf inf Fj
j→∞ X
(5.8)
≤ lim sup inf Fj ≤ inf F.
j→∞ X X
89
Therefore all the inequalities in the previous chain are equalities. We conclude that
F(ū) = inf F
X
(that is, there exists a minimizer of the limit functional F), that uj → ū (that is, any gener-
alized minimizing sequence converges to a minimizer of the limit functional), and inf X Fj →
inf X F. In particular, we have the convergence of minima and minimizers.
The previous discussion motivates the following definition.
Definition 5.1 (Γ-convergence). Let (X , d) be a metric space, Fj , F : X → R. We say that

Γ
the sequence (Fj )j∈N Γ-converges to F as j → ∞ in X , and we write Fj → F, if the following
two conditions are satisfied:
(i) (liminf inequality) for every x ∈ X and for every xj ∈ X such that xj → x it holds
F(x) ≤ lim inf Fj (xj );

j→∞
(ii) (limsup inequality) for every x ∈ X there exists a sequence x̄j ∈ X such that x̄j → x
and
F(x) ≥ lim sup Fj (x̄j ).
j→∞
The sequence (x̄j )j in the property (ii) of the above definition is usually called recovery
sequence. The definition can also be given pointwise, by saying that F(x) = Γ−limj→∞ Fj (x)
if the conditions (i)–(ii) hold at x.
The liminf inequality expresses the fact that, whatever sequence (xj )j we take to approx-
imate a point x, the value of Fj (xj ) is in the limit larger than F(x). The limsup inequality
implies that this lower bound is sharp, that is, it is attained along a particular sequence (x̄j )j .
Notice that the condition (ii) can be equivalently written in the following form:
for all x ∈ X there exist x̄j ∈ X such that x̄j → x and F(x) = lim Fj (xj ).
j→∞
Example 5.2 (Γ-limit of a constant sequence). In the particular case of a constant sequence
Γ
Fj = F for every j, it is in general not true that Fj → F as j → ∞. Indeed, it it was true
Γ
that Fj → F, then by the liminf inequality we would have
F(x) ≤ lim inf Fj (xj ) = lim inf F(xj )

j→∞ j→∞
for every sequence xj → x, that is, F would be lower semicontinuous. Therefore, if F is not
lower semicontinuous, then the constant sequence Fj = F does not Γ-converge to F.
Exercise 21. Show that, if F is lower semicontinuous, then the constant sequence Fj = F
Γ-converges to F.[19]
90
5.2 One-dimensional examples
The Example 5.2 shows that the Γ-limit might be different from the pointwise limit, even if
the pointwise limit exists. We discuss a few more one-dimensional examples that show, in
particular, that there is no relation between pointwise convergence and Γ-convergence. In all
the following examples we take X = R.
Γ
Example 5.3. Let fj (x) := cos(jx) for j ∈ N. Then fj → −1 as j → ∞ (but the pointwise
limit of fj does not exists).
Indeed, the liminf inequality holds trivially. For the limsup inequality, given x ∈ R and
2k −1 2k +1
j ∈ R consider kx,j ∈ Z such that x ∈ [ x,jj π, x,jj π). Then a recovery sequence is given
2kx,j −1
by x̄j := j π: indeed x̄j → x and fj (x̄j ) = −1 for all j.
Figure 5.1: The sequence in Example 5.3.
Example 5.4. Let


−jx
 if 0 < x < 1j , (
1 0 if x 6= 0,
fj (x) := jx − 2 if j < x < 2j , g(x) :=
 −1 if x = 0.
0 otherwise,

Γ Γ
Then fj → 0 pointwise, and fj → g. Notice that, if we take gj = −fj , then gj → 0. Therefore
in general
Γ − lim(−Fj ) 6= −Γ − lim Fj !
j j
−1
91
Example 5.5. Let fj (x) := jxejx . Then (fj )j Γ-converges to the function

0
 if x < 0,
f (x) := −e −1 if x = 0,

+∞ if x > 0.

Indeed:
• Let x < 0. If xj → x, then xj < x
2 for j sufficiently large and
jx jx
lim inf fj (xj ) ≥ lim inf fj (x/2) = lim inf e 2 = 0 = f (x),
j→∞ j→∞ j→∞ 2
which shows the liminf inequality. As a recovery sequence, take xj = x for every j, and
fj (xj ) → 0 = f (x).
• Let x > 0. If xj → x, then xj > x
2 for j sufficiently large and
jx jx
lim inf fj (xj ) ≥ lim inf fj (x/2) = lim inf e 2 = +∞ = f (x),
j→∞ j→∞ j→∞ 2
which shows the liminf inequality. As a recovery sequence, take xj = x for every j, and
fj (xj ) → +∞ = f (x).
• Let x = 0. Since min fj = −e−1 for every j, and f (0) = −e−1 , the liminf inequality is
trivially satisfied. To check the limsup inequality, take the recovery sequence xj = − 1j :
then xj → 0 and fj (xj ) = −e−1 → f (0).
−e−1
2 x2
Example 5.6. Let fj (x) := jxe−2j . Then (fj )j Γ-converges to the function
1
(
− 21 e− 2 if x = 0,
f (x) :=
0 if x 6= 0.
Exercise 22. Let fj (x) := arctan(jx). Compute the Γ-limit of the sequence (fj )j .[20]
92
y
1
− 12 e− 2
5.3 Main properties

We next discuss a few basic properties of Γ-convergence which follow immediately from its
definition.
Proposition 5.7. If Fj → F uniformly on U ⊂ X , with U open, and F is lower semicon-

Γ
tinuous, then Fj → F in U .
Proof. Given x ∈ U and xj → x, we have that xj ∈ U for j large enough and
|Fj (xj ) − F(xj )| ≤ sup |Fj (z) − F(z)| → 0 as j → ∞

z∈U
by uniform convergence. Then since F is lower semicontinuous
lim inf Fj (xj ) = lim inf F(xj ) ≥ F(x).

j→∞ j→∞
This proves the liminf inequality.

To show the limsup inequality, given x ∈ U , it is sufficient to take the constant sequence
x̄j = x as a recovery sequence: indeed by uniform convergence Fj (x̄j ) = Fj (x) → F(x).
Γ
Proposition 5.8 (Stability under continuous perturbations). If Fj → F and G : X → R is
Γ
continuous, then Fj + G → F + G.
Proof. Given x ∈ X and xj → x, we have

lim inf Fj + G (xj ) = lim inf Fj (xj ) + lim G(xj ) ≥ F(x) + G(x),
j→∞ j→∞ j→∞
Γ
which shows the liminf inequality. Given x ∈ X , since Fj → F there exists a sequence x̄j → x
such that Fj (x̄j ) → F(x). Then (Fj + G)(x̄j ) → (F + G)(x) by continuity of G, and hence
(x̄j )j is a recovery sequence for Fj + G.
Γ Γ Γ
Example 5.9. In general, it is not true that if Fj → F and Gj → G, then Fj + Gj → F + G.
As a counterexample, consider X = R, Fj (x) := sin2 (jx), Gj (x) := cos2 (jx). Show as an
Γ Γ
exercise that Fj + Gj → 1 (trivial), but Fj , Gj → 0.
The next theorem formalizes the initial discussion which motivates the notion of Γ-
convergence, and shows - under a equicoercivity condition - that it implies convergence of
minima and minimizers.
93
Theorem 5.10 (Fundamental theorem of Γ-convergence). Let (X , d) be a metric space, and
Γ
let Fj , F : X → R. Assume that Fj → F and that the functionals Fj are equi-mildly coercive,
that is
there exists K ⊂ X , K 6= ∅ compact, such that inf Fj = inf Fj for all j ∈ N. (5.10)
X K
Then:
• the minimum minX F is attained,
• lim inf Fj = min F,

j→∞ X X
• if (xj )j is such that limj→∞ Fj (xj ) = limj→∞ inf X Fj , then every limit point of (xj )j is
a minimizer of F.
Proof. By equi-mild coercivity, there exists a sequence (x̃j )j with x̃j ∈ K and
lim inf Fj (x̃j ) = lim inf inf Fj .

j→∞ j→∞ X
As K is compact and X is a metric space, there exist a subsequence (x̃jk )k such that x̃jk →
x̄ ∈ K, and we can also assume that
lim inf Fj (x̃j ) = lim Fjk (x̃jk ).

j→∞ k→∞
We then define the sequence

(
xjk if j = jk for some k,
xj :=
x̄ otherwise.
In this way xj → x̄ and by Γ-convergence
inf F ≤ F(x̄) ≤ lim inf Fj (xj ) ≤ lim inf Fjk (x̃jk ) = lim inf inf Fj .
X j→∞ k→∞ j→∞ X
On the other hand, given any x ∈ X there exists a recovery sequence yj → x such that
F(x) ≥ lim sup Fj (yj ) ≥ lim sup inf Fj ,

j→∞ j→∞ X
hence by taking the infimum over x
inf F ≥ lim sup inf Fj ≥ lim inf inf Fj ≥ F(x̄) ≥ inf F.

X j→∞ X j→∞ X X
This shows that x̄ is a minimizer of the limit functional and that the first two properties in
the statement hold. Finally, to prove the last statement it is sufficient to repeat the previous
argument for the sequence (xj )j .
Remark 5.11. As a particular case of the previous theorem, we have that if xj is a minimum
point of Fj for every j, then up to subsequences xj converges to a minimizer of F. Conversely,
in general it is not true that any minimizer of F is the limit of a sequence of minimizers:
Γ
consider for instance X = Rn , Fj (x) = 1j |x|, Fj → F ≡ 0.
94
Example 5.12. Notice that Γ-convergence does not imply convergence of local minimizers.
For instance, consider X = R and fj (x) := ex + sin(jx), which Γ-converge to the function
f (x) := ex − 1.
Definition 5.13. Let (X , d) be a metric space, Fj : X → R, x ∈ R. We define

n o
Γ − lim inf Fj (x) := inf lim inf Fj (xj ) : xj → x ,
j→∞ j→∞
n o
Γ − lim sup Fj (x) := inf lim sup Fj (xj ) : xj → x .
j→∞ j→∞
Notice that, by a diagonal argument, the two infima in Definition 5.13 are attained. It is
easy to check that
Γ − lim inf Fj (x) ≤ Γ − lim sup Fj (x),
j→∞ j→∞
and that, if they coincide, then the sequence Fj Γ-converges to their common value:
Γ
Γ − lim inf Fj (x) = Γ − lim sup Fj (x) = λ =⇒ Fj (x) → λ.
j→∞ j→∞
Proposition 5.14. The functions x 7→ Γ − lim inf j→∞ Fj (x), x 7→ Γ − lim supj→∞ Fj (x) are
lower semicontinuous. In particular the Γ-limit, if it exists, is always lower semicontinuous.
Proof. Denote by F 0 (x) := Γ − lim inf j Fj (x). We prove sequential lower semicontinuity: let
xk → x, and let us show that F 0 (x) ≤ lim inf k F 0 (xk ).
By Definition 5.13, for every k ∈ N we can find a sequence (xjk )j such that xjk → xk as
j → ∞ and
F 0 (xk ) = lim inf Fj (xjk ).
j→∞
By induction, for ever k ∈ N we can find an index jk (with jk > jk−1 for all k) such that
1 1
d(xjkk , xk ) < , Fjk (xjkk ) < F 0 (xk ) + .
k k
Consider then the sequence, for j ∈ N,
(
xjkk if j = jk for some k ∈ N,
yj :=
x otherwise.
Since d(xjkk , x) ≤ 1
k + d(xk , x) → 0 as k → ∞, we have that yj → x as j → ∞, and
F 0 (x) ≤ lim inf Fj (yj ) ≤ lim inf Fjk (yjk ) = lim inf Fjk (xjkk )
j→∞ k→∞ k→∞
1
≤ lim inf F 0 (xk ) + = lim inf F 0 (xk ),
k→∞ k k→∞
which proves the lower semicontinuity of F 0 . The proof of the lower semicontinuity of the
Γ − lim sup is analogous.
The theory of relaxation can be seen as a particular case of Γ-convergence, in view of the
following result (see also Example 5.2 and Exercise 21).
95
Γ
Proposition 5.15 (Relaxation). If Fj = F for every j ∈ N, where F : X → R, then Fj → F,
where F is the relaxation of F.
Proof. Denote by F 0 := Γ − lim inf j Fj , F 00 := Γ − lim supj Fj .

We first notice that F 00 ≤ F (by definition), and F 00 is lower semicontinuous (by Propo-
sition 5.14), hence F 00 ≤ F (since the relaxation is the largest functional with these two
properties).
To conclude, it is sufficient to show that F 0 ≥ F. Consider any functional G which is lower
semicontinuous and G ≤ F. Then, setting Gj = G for all j, we have Gj ≤ Fj and therefore
Γ − lim inf Gj ≤ F 0
j→∞
and since G is lower semicontinuous, Γ − lim inf j Gj = G (by Exercise 21). Hence G ≤ F 0 for
every lower semicontinuous functional G such that G ≤ F; taking the supremum over such G,
we obtain F ≤ F 0 , as desired.
Notice that, as a consequence of Proposition 5.15 and Theorem 5.10 we obtain the main
theorem of the relaxation: if F is mildly coercive, then
min F = inf F
X X
where the first minimum is attained. Moreover the minimizers of F are exactly all the limits
of converging minimizing sequences of F.
We finally mention without proof the following compactness result (see [3, Theorem 8.5]).
Proposition 5.16 (Compactness). Let (X , d) be a separable metric space and let Fj : X → R.

Γ
Then there exists a subsequence (Fjk )k and a functional F such that Fjk → F as k → ∞.
Exercise 23. Let (X , d) be a metric space. We say that a sequence of nonempty sets (Ej )j∈N ,
Ej ⊂ X , converges to E ⊂ X in the sense of Kuratowski if the following two conditions are
satisfied:
(i) if (Ejk )k is a subsequence and xjk ∈ Ejk , xjk → x, then x ∈ E;
(ii) for every x ∈ E there exists a sequence xj → x with xj ∈ Ej for every j.

Γ
Prove that Ej → E in the sense of Kuratowski if and only if χEj → χE .[21]
Γ
Exercise 24. Let Fj , F : X → R. Prove that Fj → F if and only if epi(Fj ) converges to
epi(F) in the sense of Kuratowski in X × R.[22]
5.4 An application to integral functionals

We conclude these introductory notes by discussing an important application of Γ-convergence
in the framework of integral functionals.
We consider an open and bounded set Ω ⊂ Rn and a function f : Rn × Rn → [0, +∞)
satisfying the following properties:
(a) f is Borel-measurable;
96
(b) f is periodic in the first variable, i.e. f (x + ei , ξ) = f (x, ξ) for every i ∈ {1, . . . , n};
(c) c1 |ξ|p ≤ f (x, ξ) ≤ c2 (1 + |ξ|p ) for some c1 , c2 > 0, p ∈ (1, ∞).

We consider the family of functionals, for ε > 0,
Z
 f x , ∇u(x) dx

if u ∈ W 1,p (Ω),
Fε (u) := Ω ε (5.11)
+∞ if u ∈ Lp (Ω)\W 1,p (Ω).

The underlying physical setting is the following: Ω represents the region occupied by a com-
posite material with a periodic microstructure at scale ε > 0, very small compared with the
macroscopic size of the object Ω. The function u describes the state of the material (for
instance, it could be the elastic deformation of the body, although to model such situation it
would be more appropriate to consider vector-valued functions u), and the preferred states
are selected as minimizers of the associated stored energy Fε (u).
The homogenization problem consists in the study of the limit behaviour of the energies
as the scale parameter ε tends to zero. Due to the oscillatory behaviour of the function
f ( xε , ξ), this is a typical case in which one has only weak convergence and the pointwise
limit differs from the Γ-limit. “The mathematical theory of homogenization deals with the
overall response of composite materials, like stratified or fibred materials, matrix-inclusion
composites, porous media, materials with many small holes or fissures. All these structures
are strongly heterogeneous, if observed at a microscopic level, but exhibit the typical behaviour
of an ideal homogeneous material, when they are observed on a macroscopic scale” [3].
Theorem 5.17 (Homogenization). Under the assumptions listed above, there exists a convex
function fhom : Rn → [0, +∞) such that the sequence (Fε )ε Γ-converges, as ε → 0, to the
functional Z
1,p (Ω),
hom (∇u(x)) dx if u ∈ W
 f
F(u) := Ω (5.12)

+∞ p
if u ∈ L (Ω)\W (Ω) 1,p
with respect to the strong topology of Lp (Ω).

The fact that the density fhom of the Γ-limit does not depend explicitly on the space
variable x can be interpreted as the fact that the behaviour of the ε-periodic material, at a
macroscopic scale, is similar to that of a homogeneous material.
Notes
[19]
Solution to Exercise 21. We check the two conditions in the definition of Γ-convergence for Fj = F. If
xj → x, then trivially F(x) ≤ lim inf j F(xj ) by lower semicontinuity. Given x ∈ X , it is sufficient to choose as
recovery sequence xj = x for all j.
[20] Γ
Solution to Exercise 22. We show that fj → g, where g is the function
(
− π if x ≤ 0,
g(x) := π 2
2
if x > 0.
Indeed:
• Let x < 0. Since fj ≥ − π2 , the liminf inequality is trivial. As a recovery sequence, take xj = x for every
j, so that fj (xj ) = arctan(jx) → − π2 = g(x) as j → ∞.
97
• Let x > 0. If xj → x, then xj > x
2
for j sufficiently large and
π
lim inf fj (xj ) ≥ lim inf fj (x/2) = lim inf arctan(jx/2) = = g(x),
j→∞ j→∞ j→∞ 2
which shows the liminf inequality. As a recovery sequence, take xj = x for every j, so that fj (xj ) =
arctan(jx) → π2 = g(x).
• Let x = 0. Since fj ≥ − π2 and g(0) = − π2 , the liminf inequality is trivially satisfied. To√check the limsup
inequality, take the recovery sequence xj = − √1j : then xj → 0 and fj (xj ) = arctan(− j) → − π2 = g(0)
as j → ∞.
[21]
Solution to Exercise 23. Recall that χE (x) = 0 if x ∈ E and χE (x) = +∞ if x ∈
/ E. Assume that Ej → E
in the sense of Kuratowki. Then we have:
• for every xj → x, χE (x) ≤ lim inf j χEj (xj ). Indeed, if x ∈ E the previous inequality is trivial, since
χE (x) = 0; if x ∈
/ E, then by the first condition in the definition of Kuratowski convergence we have
that xj ∈
/ Ej for all j large enough, therefore lim inf j χEj (xj ) = +∞ = χE (x);
• given x ∈ X , there exists x̄j → x such that χEj (xj ) → χE (x). Indeed, if x ∈ E then by the second
condition in the definition of Kuratowksi convergence there exists a sequence x̄j → x with x̄j ∈ Ej , and
therefore limj χEj (xj ) = 0 = χE (x); if instead x ∈/ E, then by the first condition in the definition of
Kuratowski convergence we have x ∈ / Ej for all j large enough, hence we can take as recovery sequence
the constant sequence x̄j = x for all j and limj χEj (x) = +∞ = χE (x).
Γ Γ
This proves that χEj → χE . Conversely, suppose that χEj → χE . Then we have:
• let (Ejk )k be a subsequence such that xjk ∈ Ejk , xjk → x. Then define the sequence (yj )j by setting
yj = xjk if j = jk for some k, yj = x otherwise. Then yj → x and therefore χE (x) ≤ lim inf j χEj (yj ) =
lim inf k χEjk (xjk ) = 0, therefore x ∈ E;
• let x ∈ E, then there exists a recovery sequence x̄j → x such that χEj (x̄j ) → χE (x) = 0. Therefore
χEj (xj ) = 0 for all j ≥ j0 for some j0 ∈ N sufficiently large, hence x̄j ∈ Ej for j ≥ j0 . Then the sequence
xj defined by xj = x̄j for j ≥ j0 , and xj arbitrary point of Ej for j < j0 , satisfies the requirement.
This proves that Ej → E in the sense of Kuratowski.
[22] Γ
Solution to Exercise 24. Assume that Fj → F. We check the two conditions in the definition of Kuratowski
convergence.
• Let (xjk , tjk ) ∈ epi(Fjk ) for some subsequence jk , with (xjk , tjk ) → (x, t). We need to show that
(x, t) ∈ epi(F), that is F(x) ≤ t. Construct a sequence (yj , tj )j by setting
(
(xjk , tjk ) if j = jk for some k,
(yj , tj ) :=
(x, t) otherwise.
Then (yj , tj ) → (x, t) and by Γ-convergence
F(x) ≤ lim inf Fj (xj ) ≤ lim inf Fjk (xjk ) ≤ lim inf tjk = t.
j→∞ k→∞ k→∞
• Let (x, t) ∈ epi(F), that is F(x) ≤ t. By Γ-convergence there exists a recovery sequence xj → x such
that Fj (xj ) → F(x). Let εj := |Fj (xj ) − F(x)| → 0. Then Fj (xj ) ≤ F(x) + εj ≤ t + εj , that is
(xj , t + εj ) ∈ epi(Fj ) for every j and (xj , t + εj ) → (x, t).
Assume now that epi(Fj ) → epi(F) in the Kuratowski sense. We check the two condition in the definition of
Γ-convergence.
• Let xj → x. Choose a subsequence (xjk )k such that lim inf j→∞ Fj (xj ) = limk→∞ Fjk (xjk ) = `. Then
for every ε > 0 we have F(xjk ) ≤ ` + ε for k sufficiently large, therefore (xjk , ` + ε) ∈ epi(Fjk ) for k
large and by Kuratowski convergence (x, ` + ε) ∈ epi(F), hence
F(x) ≤ ` + ε = lim inf Fj (xj ) + ε.
j→∞
As ε is arbitrary we obtain the liminf inequality.

• Let x ∈ X . Since (x, F(x)) ∈ epi(F), by Kuratowski convergence we can find a sequence (xj , tj ) →
(x, F(x)) such that (xj , tj ) ∈ epi(Fj ) for all j. Then lim supj Fj (xj ) ≤ lim supj tj = F(x), hence (xj )j
is a recovery sequence.
98
Bibliography
[1] H. Brezis, Functional analysis, Sobolev spaces and partial differential equations.
Springer, New York, 2011.
[2] B. Dacorogna, Direct methods in the calculus of variations. Second edition. Applied
Mathematical Sciences, 78. Springer, New York, 2008.
[3] G. Dal Maso, An introduction to Γ-convergence. Birkhäuser, Boston, 1993.
[4] L.C. Evans Partial differential equations. Second edition. Graduate Studies in Mathe-
matics, 19. American Mathematical Society, Providence, RI, 2010.
[5] M. Giaquinta, L. Martinazzi, An introduction to the regularity theory for elliptic sys-
tems, harmonic maps and minimal graphs. Second edition. Appunti della Scuola Normale
Superiore di Pisa, Edizioni della Normale, Pisa, 2012.
[6] E. Giusti, Direct methods in the calculus of variations. World Scientific Publishing Co.,
Inc., River Edge, NJ, 2003.
[7] G. Leoni, A first course in Sobolev spaces. Second edition. Graduate Studies in Mathe-
matics, 181. American Mathematical Society, Providence, RI, 2017.
[8] H.L. Royden, Real analysis. Collier-Macmillan, New York, 1968.
99

Lecture Notes CoV 2021

Uploaded by

Lecture Notes CoV 2021

Uploaded by

Lecture notes on

Advanced Calculus of Variations

Summer Semester 2021

1 Minimization of integral functionals among Lipschitz functions 7

4 Euler-Lagrange equations and regularity 74

and its minimization over the set U := {u ∈ C 1 (Ω) : u = u0 on ∂Ω}.

Therefore we obtain that a minimizer ū satisfies

Integrating by parts (assuming that ū is sufficiently regular) we then obtain

which eventually yields to the partial differential equation

Minimization of integral functionals

and the corresponding minimum problem

F (ū) ≤ F (v) for every v ∈ Lip(Ω) with v = ϕ on ∂Ω.

c1 |ξ| ≤ f (ξ) ≤ c2 (1 + |ξ|).

(i) (fj )j is equibounded, that is there exists M ∈ R such that

sup |fj (x)| ≤ M for all j ∈ N;

1.2 Existence of minimizers: preliminary results

Proof. Let Lipk (Ω; ϕ) := {u ∈ Lipk (Ω) : u = ϕ on ∂Ω} and let

m := inf{F (u) : u ∈ Lipk (Ω; ϕ)}.

which in particular implies that m ∈ R.

|uj (x)| ≤ |ϕ(y)| + k|x − y| for all x ∈ Ω, y ∈ ∂Ω

and in turn, as Ω and ϕ are bounded by assumption,

sup |uj | ≤ sup |ϕ| + k diam Ω

(where diam Ω := sup{|x − y| : x, y ∈ Ω}). This shows equiboundedness.

|uj (x) − uj (y)| ≤ k|x − y| for all x, y ∈ Ω.

f (η) ≥ f (ξ) + h∇f (ξ), η − ξi for all ξ, η ∈ Rn ,

From (1.4) we obtain

1.3 Maximum principle

F (u) = min F (v) : v ∈ Lipk (Ω), v = u on ∂Ω .

Proposition 1.7 (Maximum principle). In addition to the assumptions of Proposition 1.3,

see Exercise 2 below. Then

= f (∇u) dx + f (∇v) dx,

that is supΩ |u − v| ≤ ε = max∂Ω |u − v|. The other inequality is immediate by continuity.

|ū(x) − ū(y)| |ū(x) − ū(y)|

Let τ := y − x and uτ (z) := ū(z + τ ). We remark that:

(a) uτ is defined for z ∈ Ωτ := {z ∈ Rn : z + τ ∈ Ω};

(b) since x + τ = y ∈ Ω, we have x ∈ Ω ∩ Ωτ ;

(c) uτ (x) = ū(y);

(d) uτ minimizes F in Lipk (Ωτ ; uτ );

Then the function (

belongs to Lipk (Ω; v) (exercise) and

which contradicts the minimality of v.

max |ū(z) − uτ (z)| = max |ū(z) − uτ (z)| = |ū(z0 ) − uτ (z0 )|

for some z0 ∈ ∂(Ω ∩ Ωτ ). In particular, by (b) we deduce that

|ū(x) − uτ (x)| ≤ |ū(z0 ) − uτ (z0 )|.

Since ∂(Ω ∩ Ωτ ) = (∂Ω ∩ Ωτ ) ∪ (∂Ωτ ∩ Ω), we eventually have two possibilities:

• if z0 ∈ ∂Ωτ ∩ Ω, then z0 + τ ∈ ∂Ω. In this case define y0 := z0 + τ ∈ ∂Ω, x0 := z0 ∈ Ω,

1.4 The bounded slope condition and existence of minimizers

ϕ(x) + a · (y − x) ≤ ϕ(y) ≤ ϕ(x) + b · (y − x) for all y ∈ ∂Ω.

where F is the functional defined in (1.1).

−M |x − y| ≤ a · (y − x) ≤ ϕ(y) − ϕ(x) ≤ b · (y − x) ≤ M |x − y|,

ū(y) ≤ v(y) = ϕ(x) + b · (y − x) for all y ∈ Ω,

and by the same argument we also have

ū(y) ≥ ϕ(x) + a · (y − x) for all y ∈ Ω.

a) Ω ⊂ Rn is open, bounded, convex and with smooth boundary;

b) there exists a constant c1 > 0 such that

c1 ν(x) · (y − x) ≥ |x − y|2 for all x, y ∈ ∂Ω,

where ν(x) denotes the interior unit normal to ∂Ω at x;

Then the bounded slope condition is satisfied. Indeed for all x, y ∈ ∂Ω

ϕ(y) = ϕ(x) + ∇ϕ(x) · (y − x) + R(x, y) with |R(x, y)| ≤ c2 |x − y|2

for some c2 > 0 independent of x and y. Then it is sufficient to take

b := ∇ϕ(x) + c1 c2 ν(x), a := ∇ϕ(x) − c1 c2 ν(x)

and the bounded slope condition is satisfied with this choice.

u(x + hξ) − u(x) k|x + hξ − x|

∇(u ∨ v) = ∇(u ∧ v) = ∇u = ∇v a.e. on {u = v}.

This follows from the following property: if w ∈ Lip(Ω), then

∇w = 0 a.e. on {w = 0}. (1.12)

on a set of positive measure. Since u+v

2.1 An overview of Sobolev Spaces

2.1.1 Weak derivatives and Sobolev spaces

Example 2.2. Let n = 1, Ω = (0, 2), and let u be the function

Indeed for every test function ϕ ∈ Cc∞ (0, 2):

hence v = u0 in the weak sense.

Example 2.3. Let n = 1, Ω = (0, 2), and let u be the function

and we obtain a contradiction by passing to the limit as k → ∞.

uk → u in W 1,p (Ω) ⇐⇒ kuk − ukW 1,p (Ω) → 0,

(and H k (Ω) := W k,2 (Ω)), with norm