Michele Gori - Lower Semicontinuity and Relaxation (2004)

Università degli Studi di Pisa
Facolt` a di Scienze Matematiche, Fisiche e Naturali

Scuola di Dottorato in Matematica
XVI Ciclo
Ph.D. Thesis in Mathematics
Michele Gori
Lower semicontinuity and relaxation
for integral and supremal functionals
Advisor: Prof. Giuseppe Buttazzo
Ph.D. School Director: Prof. Fabrizio Broglia
April 2004
Ringrazio Paolo Marcellini, che mi ha introdotto alla ricerca matematica, e Giuseppe Buttazzo,
punto di riferimento costante in questi anni di dottorato.
Ringrazio per la sua amicizia Francesco, con cui ho condiviso, n dagli anni del corso di laurea,
la passione per la matematica.
Ringrazio Anna Paola e Gigi per i loro consigli e Marco, Luca, Francesca e tutti gli altri colleghi
con cui ho condiviso questa esperienza.
Devo molto, inne, alla presenza costante e paziente di mia moglie Letizia, alla quale dedico
questo lavoro.
Contents
1 Introduction 1
2 Preliminary results 7
2.1 Notes on measure theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Notes on BV , Sobolev and Lipschitz functions . . . . . . . . . . . . . . . . . . . . 14
2.3 Some tools from convex analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Principal denitions and properties . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Recession functions and related topics . . . . . . . . . . . . . . . . . . . . . 19
3 Approximation of convex functions 29
3.1 De Giorgis and Serrins approximation methods . . . . . . . . . . . . . . . . . . . 29
3.2 Approximation of convex and demi-coercive functions . . . . . . . . . . . . . . . . 30
3.2.1 Approximation by means of convex cones . . . . . . . . . . . . . . . . . . . 30
3.2.2 Approximation by means of strictly convex functions . . . . . . . . . . . . . 34
4 Lower semicontinuity for integral functionals 39
4.1 A brief historical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Statement of the main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Proof of the main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.1 The condition on the geometric variable . . . . . . . . . . . . . . . . . . . . 44
4.3.2 The condition on the gradient variable . . . . . . . . . . . . . . . . . . . . . 52
4.4 A counterexample to the lower semicontinuity . . . . . . . . . . . . . . . . . . . . . 55
4.5 Appendix: the vectorial case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5.1 Some counterexamples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5.2 Lower semicontinuity theorems . . . . . . . . . . . . . . . . . . . . . . . . . 60
5 Lower semicontinuity for supremal functionals 65
5.1 A counterexample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2 Sucient conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 Necessary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6 Functionals dened on measures 73
6.1 Convex and level convex functionals . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.1.1 Simple results via convolutions . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.2 The integral functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.1.3 The supremal functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Non level convex functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2.1 Necessary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2.2 Sucient conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
i
ii CONTENTS
7 Applications to BV 93
7.1 Convex and level convex functionals . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.1.1 Lower semicontinuity and relaxation . . . . . . . . . . . . . . . . . . . . . . 93
7.1.2 Dirichlet problem for supremal functionals . . . . . . . . . . . . . . . . . . . 95
7.2 Non level convex functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2.1 A minimum problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Bibliography 100
Chapter 1
Introduction
In this Ph.D. Thesis we shall be concerned with some aspects of the central problem of the Calculus
of Variations which is to nd the minimum points of functionals. In our study we will consider, in
particular, the integral functional
I(u, ) =
_
f(x, u(x), u(x))dx, (1.1)

and the supremal functional
S(u, ) = ess sup
x
g(x, u(x), u(x)), (1.2)
where R
n
is an open set, f, g are two Borel functions from R R
n
with values in [0, ]
and u belongs to a suitable space of functions. In the following we will always denote by x, s and
the three variables of f and g, called geometric, function and gradient variable respectively.
Some of the most powerful tools to prove the existence of minimum points for such functionals
are the so called Direct Methods. These methods move by the fact that, given a topological space
(X, ), we have that a function F : X [0, ] admits a minimum point every time F is lower
semicontinuous (briey l.s.c.)
1
, that is,
x
h
, x X, x
h
x in = F(x) liminf
h
F(x
h
),
and there exists a compact
2
minimizing sequence, that is a sequence y
h
h=1
X which admits a
converging subsequence and such that
inf
xX
F(x) = lim
h
F(y
h
).
Indeed, in these hypotheses, we have that every limit point y
0
of y
h
h=1
satises the inequality
F(y
0
) liminf
h
F(y
h
) = inf
xX
F(x),
so that y
0
is just a minimum point of F.
Clearly the two conditions considered above work in opposite directions: indeed, roughly speak-
ing, the lower semicontinuity of F is simpler to prove when the topology has many open sets,
1
Actually this denition is the one of sequential lower semicontinuity: however, since we will always deal with
this notion, which may be dierent from the topological notion of lower semicontinuity given by means of open
sets, we simplify our notations with this little abuse. Note also that when the space (X, ) is a metric space, as for
instance R
n
endowed with the Euclidean distance, the two notions agree.
2
Even in this case, since we will always deal with the notion of sequential compactness, we can simplify our
notations.
1
2 CHAPTER 1. INTRODUCTION
while, on the contrary, we have more opportunities to prove the compactness of some minimizing
sequence when has few open sets.
Now, if we consider the functionals I and S given by (1.1) and (1.2), in order to apply the
Direct Methods, we have to nd a suitable topological space on which they can be dened and
such that it allows to prove the lower semicontinuity and the compactness. It is worth noting
that the notion of lower semicontinuity was introduced for the rst time in the framework of the
Calculus of Variations by Tonelli in 1913 just to treat functionals like (1.1) (see the monographs
[70] and [71]; for further details see also [62]).
Since our aim is to nd regular minimum points of I and S, certainly the rst attempt to
try is to dene them on C
1
(), that is, the set of the derivable functions on with continuous
rst derivatives, endowed with its natural topology. However, it is easy to see that this space is
not suitable to apply the Direct Methods since its topology, even if it makes easy the proof of
the lower semicontinuity of the considered functionals, requires too much conditions to prove the
compactness of the minimizing sequences. Thus, since we want I and S to be dened at least on
C
1
(), we need to nd more proper larger functional spaces on which they can be dened and in
which, this time, some minimum points can be found by more powerful compactness theorems.
The Sobolev spaces, whose theory has been developed to solve this particular kind of problems,
provide a suitable domain of denition for I and S, because, by the properties of their weak
(weak
) topologies, the compactness of the minimizing sequences can be obtained by requiring

simple conditions on the functionals, even if the lower semicontinuity could be more dicult to
prove.
The lower semicontinuity of I and S dened on Sobolev spaces has been deeply studied and
necessary and sucient conditions on f and g to obtain it have been found (see for instance De
Giorgi [35], Ioe [56] and Olech [65] for the integral setting and Barron and Jensen [14], Barron,
Jensen and Wang [15] and Barron and Liu [16] for the supremal setting). In particular, the
fundamental role of the convexity of f and of the level convexity of g in the gradient variable has
been understood
3
.
Assuming now that f and g guarantee the lower semicontinuity of I : W
1,p
loc
() [0, ] (with
1 < p < ) and S : W
1,
loc
() [0, ], we can simply prove the existence of a minimum by
letting, for instance,
f(x, s, ) c([s[
p
+[[
p
), (1.3)
where c > 0, and
g(x, s, )
([s[ +[[), (1.4)

where
: [0, ) [0, ) and lim

t
(t) = . Indeed, (1.3) and (1.4) guarantee that every

minimizing sequence of I and S is locally bounded in W
1,p
loc
() and W
1,
loc
() respectively and then
compact with respect to their weak (weak
) topologies.
Obviously, the minimum points found in this way can be considered only as generalized min-
imum points of the starting problem; the question of knowing if they really belong to C
1
() is
proper to the so called regularity theory
4
.
Nevertheless a large class of remarkable functionals, as the class of integral functionals in which
the integrand grows only linearly, does not satisfy the coercivity conditions (1.3) and (1.4) so that
3
As far as possible, from now on, f will denote a convex (in the gradient variable) function and g will denote
a level convex (in the gradient variable) one. However, we will always state all the properties of every function
considered in the statements and proofs.
4
We point out that, at least in the integral setting, it may happen that not only such a generalized minimum
point does not belong to C
1
() but also
inf
_
I(u, ) : u W
1,p
loc
()
_
< inf
I(u, ) : u C
1
()
.
When this particular situation occurs we say we have a Lavrentiev (or Gap) phenomenon (see [57], [60] and [73]).
Nevertheless the Sobolev spaces have such an important role that often a minimum point in these spaces is already
considered satisfactory.
3
the compactness theorems on the Sobolev spaces cannot be used anymore. Therefore, we would
like to extend again the domain of denition of our functionals on larger topological spaces as, for
instance, L
1
loc
(), C() or BV
loc
(), endowed with their natural topologies, in order to look for
the minimum points in these new spaces: clearly this strategy requires to prove more dicult lower
semicontinuity theorems, but it allows to have better compactness theorems (we will consider in
particular the case of BV
loc
() endowed with the w
-BV
loc
() topology).
However, functions in the spaces quoted above dont have a gradient representable as a function,
contrarily to the ones in Sobolev spaces. Thus, it is not clear what could be the meaning of I and
S when computed, for instance, on a function belonging to BV
loc
(). As a consequence, it must
be understood how we can build up extensions of I and S on these spaces in such a way that they
are l.s.c. with respect to their own topologies.
Before presenting one of the possible answers to this question, let us give a short background
on the area of the nonparametric surfaces. We follow Serrin [68].
Let R
2
be a bounded open set with Lipschitz boundary. If u A(), the set of the
piecewise ane functions on (that is, the functions such that their graph is a polyhedral surface;
see Chapter 2) the notion of area is elementary:
A(u, ) = Area(graph(u)) =
area of the plane faces =

_
_
1 +[u(x)[
2
dx.
However, it might be of interest a denition of the area that can be meaningful even for a continuous
curved surface. To this purpose, we could use the integral expression dened above but it is not so
clear which class of functions it can be really applied to. Then, as for the denition of the arc length,
we could dene the area of a continuous surface as the supremum of the areas of the approximating
polyhedra, but, as Schwartzs counterexample shows, this method leads to contradictions (see for
instance [27] page 540).
Lebesgue had the brilliant idea to dene the area of a continuous surface obtained as the graph
of u C(), as
/(u, ) = inf
_
liminf
h
A(u
h
, ) : u
h
A(), u
h
u in L
()
_
.
Subsequently Serrin proved that not only, as obvious, / is l.s.c. on C() with respect to the
uniform convergence, but also that, for every u A(), A(u, ) = /(u, ), that is / is really an
extension of A (see [68] Theorem 1): these ideas are the basis of the modern concept of relaxation.
For our purposes, since we focus our attention on the problem of the extension of functionals,
we introduce the notion of relaxation as follows
5
.
Let us consider a topological space (X, ), Y X be a dense subset with respect to and
F : Y [0, ]. We can always dene the following trivial extension of F on X, given by
F(x) =
_
F(x) if x Y,
if x , X Y,
but in general this functional is not l.s.c. on X with respect to . Nevertheless, by means of

F,
we can dene the functional R[](F) : X [0, ], called the relaxed functional of F on X with
respect to , as
R[](F)(x) = sup
_
G(x) : G

F, G is l.s.c. with respect to
_
;
note that the supremum of any family of l.s.c. functionals is l.s.c. too. Clearly, R[](F) is the
greatest functional less or equal to

F which is l.s.c. on X with respect to and it can be easily
5
As for the notion of lower semicontinuity, even in this case the denition that follows is the one of sequential
relaxation.
proved that it can be characterized, for every x X, as
R[](F)(x) = inf
_
liminf
h
F(x
h
) : x
h
Y, x
h
x in
_
. (1.5)
It also satises the equality
inf
xX
R[](F)(x) = inf
yY
F(y), (1.6)
that provides important information on the original minimum problem on Y .
However, even if the lower semicontinuity is now satised, it may happen that the functional
R[](F) just built up is not an extension of F on X since it could exist x Y such that R[](F)(x) <
F(x). In order to really have an extension we must have R[](F) = F on Y and this happens if
and only if F is l.s.c. on Y with respect to , that is
x
h
, x Y, x
h
x in = F(x) liminf
h
F(x
h
),
For these reasons, in order to obtain further extensions via relaxation of I and S on certain
topological spaces larger than the Sobolev ones, rst of all we have to understand the lower semicon-
tinuity properties of these functionals dened on Sobolev spaces with respect to the new considered
topologies. This argument is developed in Chapters 4 and 5.
In Chapter 4 we look for conditions on f in order to obtain the lower semicontinuity of I on
W
1,1
loc
() with respect to the L
1
loc
() convergence. This is a classical problem in the Calculus of
Variations, rst studied by Serrin in [69]. Here we propose slightly dierent versions of the results
contained in the papers of Gori and Marcellini [55], Gori, Maggi and Marcellini [54] and Gori and
Maggi [53], in which new and very mild conditions on f that guarantee the lower semicontinuity
of I are given.
The proofs of the main results of Chapter 4 are based on certain approximation theorems for
convex functions (depending continuously on a parameter), collected in Chapter 3. In particular,
we propose the proofs of two very recent theorems contained in [53] which provide new methods
of approximations by means of convex cones and of strictly convex functions.
In Chapter 5 we consider the functional S dened on W
1,
loc
() and we study its lower semicon-
tinuity with respect to the L
loc
() topology. Some conditions on g in this contest have been found
rst by Gori and Maggi [52]: in this chapter we improve these results, considering the problem of
the necessary conditions too.
As already stated, the theorems of Chapters 4 and 5 can be read as results that provide
hypotheses on f and g, that guarantee the equalities R
_
L
1
loc
(I) = I and R[L
loc
] (S) = S on
W
1,1
loc
() and W
1,
loc
() respectively. Nevertheless these results give no information about the
possibility to represent (and to easily compute) R
_
L
1
loc
(I) on L
1
loc
() W
1,1
loc
() as an integral
either R[L
loc
] (S) on C() W
1,
loc
() as a supremal
6
.
Moving from these remarks, in Chapter 7 we consider I and S dened on W
1,1
loc
() and we
approach the particular problem to understand if it is possible to nd one of their extensions on
BV
loc
() that is l.s.c. with respect to the w
-BV
loc
() topology and such that could be explicitly
represented on this space as an integral and as a supremal respectively.
We stress that the relaxation is not the unique strategy to extend the functionals I and S
on BV
loc
() in a lower semicontinuous way: indeed, another appropriate way to extend I and S,
dierent from the relaxation, is suggested by Serrin in [68] and [69] and described below.
First of all let us consider the following notion of convergence: given an open set and a
function u BV
loc
(), we say that a sequence u
h
h=1
converges to u in w
-BV
loc
( ) if there
exists a sequence
h
h=1
of open sets such that u
h
BV
loc
(
h
) and, for every open set
,
6
For what concerns the integral setting, the problem of the representation of R

L
1
loc
(I) has been deeply studied,

and several general integral formulas to represent R

L
1
loc
(I) at least on BV
loc
() have been found (see for instance
Dal Maso [30]).
5
we have

h
if h is large enough and u
h
u in w
-BV (
) as h . Then, for every

u BV
loc
(), we dene
I
(u, ) = inf
_
liminf
h
I(u
h
,
h
) : u
h
W
1,1
loc
(
h
), u
h
u in w
-BV
loc
( )
_
, (1.7)
and
S
(u, ) = inf
_
liminf
h
S(u
h
,
h
) : u
h
W
1,1
loc
(
h
), u
h
u in w
-BV
loc
( )
_
. (1.8)
The analogy between I
and S
and the functionals R[w
-BV
loc
] (I) and R[w
-BV
loc
] (S) is clear.
Moreover, for every u BV
loc
(), it holds
I
(u, ) R[w
-BV
loc
] (I)(u, ) and S
(u, ) R[w
-BV
loc
] (S)(u, ),
and, as we will see, in some cases even the equality holds.
In Chapter 7 we present the main results of the theory involving I
, developed by Serrin in
[69] and by Goman and Serrin in [49], and of the one about S
, developed by Gori in [51]. In

particular, when f and g depend only on the gradient variable (and of course they are convex and
level convex respectively), we prove not only that I
and S
extend I and S on BV
loc
() and
that they are l.s.c. with respect to the w
-BV
loc
() convergence, but also, that they can be also
represented on BV
loc
() as an integral and a supremal given by
7
_
f(u(x))dx +
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x),
and
_
ess sup
x
g(u(x))
_
_
[D
s
u[-ess sup
x
g
_
dD
s
u
d[D
s
u[
(x)
__
.
In the same chapter, by means of the representation formula found for S
and following the

analogous theory developed for the integral functionals by Anzellotti, Buttazzo and Dal Maso [8],
we dene also a generalized Dirichlet problem for supremal functionals dened on BV () and we
prove the existence of a minimum on BV () even if g is not coercive (that is, (1.4) does not hold):
in some sense, this existence result justies the extension of S on BV ().
It is worth noting that the lower semicontinuity theorems and the representations formulas on
BV
loc
() about the functionals (1.7) and (1.8), presented in Chapter 7, are obtained as corollaries
of more general results proved for analogous functionals dened on Radon measures: in fact in
Chapter 6 the well known theory for the integral functional (see [49], [69]) is presented together
with several new results about the supremal setting, which generalize the paper of Gori [51].
We stress that, even if these results about the integral functionals are classical, we present their
proofs together with the ones for the supremal functionals just to show the complete analogy that
there exists, even under a technical point of view, between integral and supremal settings.
Moreover in the same chapter, starting by the lower semicontinuity results obtained by Bou-
chitte and Buttazzo for non convex integral functionals dened on Radon measures (see [18]), we
approach the analogous problem for the non level convex supremal ones, nding necessary and
sucient conditions. When these results are applied to the setting of the functions of bounded
variation, they lead to the existence of a minimum point for a particular non level convex and one
dimensional problem, stated and solved in Chapter 7, which presents some similar aspects to the
celebrated Mumford-Shah image segmentation problem (see also Alicandro, Braides and Cicalese
[3]).
We conclude saying that Chapter 2 contains some brief notes on measure theory, functions of
bounded variation and convex analysis, in which the main denitions and theorems on these topics
7
See Chapter 2 for notations.
are recalled. However, in this chapter, we prove also a certain number of new propositions, most
of them about the ne properties of level convex functions, that are fundamental for the proofs of
the remaining chapters.
Chapter 2
Preliminary results
In this chapter we introduce most of the denitions and notations we will need, and we collect
several propositions and theorems that will be useful tools to prove the main results presented in
this thesis. In the following we will implicitly refer to this chapter for every (non trivial) symbol
used.
2.1 Notes on measure theory
Let us consider the measurable space (, B()) where R
n
is an open set and B() is the
-algebra of the Borel subsets of .
We say that K is well contained in (briey K ) if cl(K) , that is, the closure of K
is a subset of . Moreover, for every x R
n
and > 0, we set
B(x, ) = y R
n
: [x y[ < and B(x, ) = cl (B(x, )) .
We set also S
n1
= x R
n
: [x[ = 1.
If : B() [0, ] is a positive measure then it is called a positive Borel measure on .
Moreover is called a positive Radon measure (resp. nite positive Radon measure) on , if, for
every K B(), K , it is also (K) < (resp. () < ): the set of the positive Radon
measures on is denoted by /
+
(). Obviously the Lebesgue measure on R
n
, denoted with L
n
,
belongs to /
+
(R
n
) so as the standard Dirac measure centered on x R
n
denoted with
x
.
Let us x m N: if : B() R
m
is a vector measure then it is called a nite Radon measure
on while if : K B() : K R
m
and, for every K B(), K , is a vector
measure on B(K) then it is called a Radon measure on . We denote the set of Radon measures
(resp. nite Radon measures) with /
loc
(, R
m
) (resp. /(, R
m
))
1
.
Let us consider now /
loc
(, R
m
) /
+
(). For every B B() the total variation
2
of
on B is dened by
3
[[(B) = sup
_
_
_
r
j=1
[(B
j
)[ : B
j
B(), B
j
,
r
_
j=1
B
j
B, B
i
B
j
= , i ,= j
_
_
_
. (2.1)
It can be proved that [[ /
+
(), that /(, R
m
) implies [[() < and that if /
+
()
then [[ = .
1
Clearly /
+
() /
loc
(, R),
/
+
() : () <
/(, R) and /(, R

m
) /
loc
(, R
m
).
2
We present here the denition of total variation directly for Radon measures: it is simple to verify that, when
is a nite Radon measure, it is equivalent to the usual one (see for instance [7] page 3).
3
When x R
n
, [x[ means |x|
R
n the standard Euclidean norm on R
n
.
7
8 CHAPTER 2. PRELIMINARY RESULTS
Given now /
loc
(, R
m
) /
+
() we say that is concentrated on a set B B() if
[[( B) = 0. If
1
,
2
/
loc
(, R
m
) /
+
() we say that
1
is singular with respect to
2
(briey
1

2
) if there exist B
1
, B
2
B() such that
1
is concentrated in B
1
,
2
is concentrated
in B
2
and B
1
B
2
= .
If now we consider /
loc
(, R
m
) /
+
() and /
+
() we say that is absolutely
continuous with respect to (briey << ) if (B) = 0 implies [[(B) = 0. Moreover, xed
E B(), we denote with E the restriction of to E, that is, the element of /
loc
(, R
m
)
/
+
() dened, for every B B() (when , /(, R
m
) /
+
(), we need B too), as
(E) (B) = (E B). For every E B(R
n
) we will always denote with L
n
the measure L
n
E.
We write for short (x) instead of (x) and we denote
A
= x : (x) ,= 0 ,
the set of the atoms of . Finally, if /
+
(), we dene the support of as the set
spt() = cl
__
x : > 0, (B(x, )) > 0
__
,
while, if /
loc
(, R
m
), we set spt() = spt([[): note that is concentrated on spt().
Let us consider now a Borel function u : R, that is a function which is measurable with
respect to the -algebra B() (the same denition holds even if u : R
m
), and /
+
(). It
is surely well known the notion of integral on B B() of u with respect to . We prefer instead
to remember that the essential supremum on B B() of u with respect to is dened as
-ess sup
xB
u(x) =
_
_
inf
_
sup
xB\A
u(x) : A B(), A B, (A) = 0
_
if (B) ,= 0,
if (B) = 0,
pointing out that, if << L
n
, then
4
-ess sup
xB
u(x) ess sup
xB
u(x). (2.2)
We set also, for every a, b R, the notations a b = supa, b and a b = infa, b.
Let /
+
(), m 1 and u : R
m
be a Borel function: we say u L
1
(, R
m
) (resp.
L
(, R
m
)) if it is
_
[u(x)[d(x) < ,
_
resp. -ess sup
x
[u(x)[ <
_
,
and, in this case, we set
_
u(x)d(x) =
__
u
1
(x)d(x), . . . ,
_
u
m
(x)d(x)
_
R
m
,
where u = (u
1
, . . . , u
m
). When, for every K B(), K and (K) > 0
5
, u L
1
(K, R
m
)
(resp. L
(K, R
m
)), we say u L
1
loc,
(, R
m
) (resp. L
loc,
(, R
m
))
6
.
Fixed /
loc
(, R
m
), we say that a scalar valued (resp. R
m
-valued) Borel function u belongs
to L
1
() (resp. L
1
(, R
m
)) if
_
[u(x)[d[[(x) < ,
4
When the Lebesgues measure is considered, we write for short ess sup instead of /
n
-ess sup.
5
This last condition on K is clearly relevant only to dene L
loc,
(, R
m
).
6
When m = 1 we write for short L
1
(), L
(), L
1
loc,
() and L
loc,
() and when = /
n
, is omitted.
2.1. NOTES ON MEASURE THEORY 9
and, in this case, we set
_
u(x)d(x) =
__
u(x)d
i
(x), . . . ,
_
u(x)d
m
(x)
_
R
m
,
_
resp.
_
u(x)d(x) =
m
i=1
_
u
i
(x)d
i
(x) R
_
,
where = (
1
, . . . ,
m
) and, for every i 1, . . . , m,
_
u(x)d
i
(x) =
_
u(x)d
+
i
(x)
_
u(x)d
i
(x),
_
resp.
_
u
i
(x)d
i
(x) =
_
u
i
(x)d
+
i
(x)
_
u
i
(x)d
i
(x)
_
,
where
+
i
=
|i|+i
2
and
i
=
|i|i
2
belong to /
+
(). When, for every K B(), K ,
u L
1
(K) (resp. L
1
(K, R
m
)), we say u L
1
loc,
() (resp. L
1
loc,
(, R
m
)).
If we consider now /
+
() and a function u L
1
(, R
m
) (resp. L
1
loc,
(, R
m
)) we denote
with u the element of /(, R
m
) (resp. /
loc
(, R
m
)) dened, for every B B() (resp.
B B(), B ), as
(u )(B) =
_
B
u(x)d(x);
it is well known that [u [ = [u[ and u << . A quite standard result of measure theory is
that, given /
+
(), we can decompose in a unique way
7
a measure /
loc
(, R
m
) as
=
a
+
s
, (2.3)
where
a
L
1
loc,
(, R
m
) and
s
/
loc
(, R
m
) with
s
: we call
a
the absolutely
continuous part of while
s
its singular part. With the notation quoted above, we can also write
=
a
+
c
+
#
, (2.4)
where
c
=
s
( A
) is said the Cantor part of while

#
=
s
A
is said its atomic part.

These decompositions of obviously depend on even if in the notations
a
,
s
,
c
and
#
the
measure is not expressly named: however, every time one of these decompositions is used, the
measure will be clear by the context. Finally note that, if /(, R
m
), then
a
L
1
(, R
m
)
and
s
,
c
,
#
/(, R
m
).
The linear space /(, R
m
), endowed with the norm ||
M
= [[(), is a Banach space iden-
tiable with the dual of the linear space C
0
(, R
m
), (that is, the set of the continuous functions
dened on with values on R
m
with the property that, for every > 0, there exists K

such that, for every x K
, [(x)[ < ), by the duality

, ) =
_
(x)d(x) =
m
i=1
_
i
(x)d
i
(x),
while the linear space /
loc
(, R
m
) can be identied with the dual of the locally convex linear space
C
c
(, R
m
), (that is, the set of the continuous functions on with values on R
m
such that every
component has compact support), with the same duality. We remember that, given a function
7
This is the Radon-Nikod ym Theorem (see for instance [7] Theorem 1.28). The uniqueness of the decomposition
has to be interpreted in this way: if =
a
1
+
s
1
=
a
2
+
s
2
, then, for -a.e. x ,
a
1
(x) =
a
2
(x) and, for
every B B(), B ,
s
1
(B) =
s
2
(B).
C(, R
m
) (that is, the set of the continuous functions on with values on R
m
)
8
, its support
is dened as
spt() = cl (x : (x) ,= 0) .
For this, given
h
, /
loc
(, R
m
) (resp. /(, R
m
)), we say
h
in w
-/
loc
(, R
m
) (resp.
w
-/(, R
m
)) if, for every C
c
() (resp. C
0
()), we have
lim
h
_
(x)d
h
(x) =
_
(x)d(x).
Let us note that, for technical reasons, in the previous denition we have considered scalar valued
test functions : however, the denitions given agree with the ones obtained considering vector
value test functions and the duality before described.
It is suitable at this point to introduce also the following notion of convergence. Let
h
h=1
be a family of open sets and let
h
/
loc
(
h
, R
m
), /
loc
(, R
m
): we say
h
in
w
-/
loc
( , R
m
) if, for every K , we have K
h
if h is large enough and, for every
C
c
(),
lim
h
_
h
(x)d
h
(x) =
_
(x)d(x).
The following compactness theorem for Radon measures holds (see for instance [7] Theorem
1.59 and Corollary 1.60).
Theorem 2.1. Let
h
h=1
/
loc
(, R
m
) (resp. /(, R
m
)) be a sequence such that, for every
K ,
sup
_
[
h
[(K) : h N
_
<
_
resp. sup
_
[
h
[() : h N
_
<
_
.
Then there exists a subsequence
h
k
k=1
and /
loc
(, R
m
) (resp. /(, R
m
)) such that
h
k
in w
-/
loc
(, R
m
) (resp. w
-/(, R
m
)).
Given a measure /
loc
(, R
m
) and > 0, we dene the convolution of with step as
(x) =
_
B(x,)
n
k
_
x y
_
d(y) :
R
m
. (2.5)
In the previous denition k : R
n
[0, 1] is a convolution kernel, that is k C
c
(R
n
), for every
x R
n
, k(x) = k(x), spt(k) B(0, 1) and
_
R
n
k(x)dx = 1 (we require also that k(0) ,= 0), and
= x : d(x, ) > ,
where d(x, ) = inf[x y[ : y and is the topological boundary of .
It is well known that
, R
m
) and it can be simply proved using Fubinis Theorem
that
L
n
in w
-/
loc
( , R
m
) as 0 (see [7] Theorem 2.2). In the following, referring
to a convolution kernel k, we will set k
(x) =
n
k
_
1
x
_
.
Now we propose a theorem on measures that describes the point-wise behavior of the convolu-
tions of a measure and that will be fundamental in proving some results of the following chapters.
The principal tools used in its proof are the Besicovichs Derivation Theorem for Radon measures
and the Lebesgues points Theorem (see for instance [7] Theorem 2.22 and Corollary 2.23). This
theorem generalizes Theorem 3 in [51].
Theorem 2.2. Let /
loc
(, R
m
). Then, referring to the decomposition (2.3) with respect to
L
n
, we have
(i) for L
n
-a.e. x , lim
0
[
(x)
a
(x)[ = 0;
8
When m = 1 we write for short C(), C
0
() and Cc().
(ii) for [
s
[-a.e. x spt(
s
), there exists a sequence of positive numbers
h
h=1
, depending on
x and decreasing to zero, such that
9
lim
h
d
s
d[
s
[
(x)

h
(x)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= 0. (2.6)
For simplicity we prove Theorem 2.2 by means of two lemmas that, in our opinion, are inter-
esting on their own. The rst lemma is a simple fact from the measure theory and, in this form,
can be found in [6], Theorem 2.3.
Lemma 2.3. Let /
+
() such that L
n
. Then, for -a.e. x spt() and, for every
(0, 1), we have
n
limsup
0
(B(x, ))
(B(x, ))
1.
Proof. Let us set
() =
_
x spt() : lim
0
(B(x, ))
n
=
_
: (2.7)
since ( ()) = 0 and L
n
(()) = 0 (see [7] Theorem 2.22), we achieve the proof showing the
wanted relation for every x (). Let us suppose, by contradiction, there exist x
0
() and
0
(0, 1) such that
n
0
> limsup
0
(B(x
0
,
0
))
(B(x
0
, ))
.
Then there exists
0
such that, for every 0 <
0
, (B(x
0
,
0
))
n
0
(B(x
0
, )). If we call
() = (B(x
0
, )) we have that, for every 0 <
0
, (
0
)
n
0
(). Then, for every h N,
nh
0
(
h
0
0
) (
0
), thus
(B(x
0
,
h
0
0
))
(
h
0
0
)
n

n
0
(
0
) = (B(x
0
,
0
)) < .
If now h , then
h
0
0
0 and the left hand side of the previous inequality tends to innity:
thus the contradiction is found.
In particular, xed (0, 1), for -a.e. x spt(), there exists a sequence
h
h=1
, depending
on x and decreasing to zero, such that, for every h N,
(B(x,
h
))
n+1
(B(x,
h
)). (2.8)
Lemma 2.4. Let
s
/
loc
(, R
m
), /
+
() such that
s
L
n
and << L
n
. Then, for
[
s
[-a.e. x spt(
s
), there exists a sequence of positive numbers
h
h=1
, depending on x and
decreasing to zero, such that
lim
h
_
B(x,
h
)
k
h
(x y)d(y)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= 0 (2.9)
and
lim
h
_
B(x,
h
)
k
h
(x y)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= 0. (2.10)
9
We point out that, for [
s
[-a.e. x spt(
s
),
d
s
d|
s
|
(x) S
m1
and
_
B(x,
h
)
k
h
(x y)d[
s
[(y) ,= 0.
Proof. Obviously
s
. Thus, let us consider ([
s
[) dened as in (2.7), and the set
0
= ([
s
[)
_
x spt(
s
) : lim
0
(B(x, ))
[
s
[(B(x, ))
= 0, xis a Lebesgues point for
d
s
d[
s
[
_
.
Clearly [
s
[(
0
) = 0: then the proof is achieved if, for every x
0
, (2.9) and (2.10) hold.
By the properties of the convolution kernel k, we can nd (0, 1) and c, M > 0 such that, for
every x R
n
, k(x) M and, for every x B(0, ), k(x) c (remember that k(0) ,= 0). Thus, let
us x x
0
and, since
0
([
s
[), let us consider, with respect to [
s
[, the sequence
h
h=1
(
h
< ) given by (2.8). Then we have
0 lim
h
_
B(x,
h
)
k
h
(x y)d(y)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= lim
h
_
B(x,
h
)
k
_
xy
h
_
d(y)
_
B(x,
h
)
k
_
xy
h
_
d[
s
[(y)
lim
h
M
c
(B(x,
h
))
[
s
[(B(x,
h
))
= lim
h
M
c
(B(x,
h
))
[
s
[(B(x,
h
))

[
s
[(B(x,
h
))
[
s
[(B(x,
h
))
lim
h
M
c
n+1
(B(x,
h
))
[
s
[(B(x,
h
))
= 0,
and
0 lim
h
_
B(x,
h
)
k
h
(x y)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= lim
h
_
B(x,
h
)
k
_
xy
h
_
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
_
B(x,
h
)
k
_
xy
h
_
d[
s
[(y)
lim
h
M
c
_
B(x,
h
)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
[
s
[(B(x,
h
))
= lim
h
M
c
_
B(x,
h
)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
[
s
[(B(x,
h
))

[
s
[(B(x,
h
))
[
s
[(B(x,
h
))
lim
h
M
c
n+1
_
B(x,
h
)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
[
s
[(B(x,
h
))
= 0,
that prove (2.9) and (2.10).
Proof of Theorem 2.2. (i). Let us x /
loc
(, R
m
) and M > 0 such that, for every x R
n
,
k(x) M, then
[
(x)
a
(x)[ =
_
B(x,)
k
(x y)d(y)
a
(x)
_
B(x,)
k
(x y)[
a
(y)
a
(x)[dy +
_
B(x,)
k
(x y)d[
s
[(y)
M
n
_
B(x,)
[
a
(y)
a
(x)[dy +M
n
_
B(x,)
d[
s
[(y)
= M
n
_
B(x,)
[
a
(y)
a
(x)[dy +M
[
s
[(B(x, ))
n
,
that, for L
n
-a.e. x , goes to zero when 0.
(ii). Let us x /
loc
(, R
m
) and consider
s
and = [
a
(x)[ L
n
: since
s
L
n
and
<< L
n
we can apply Lemma 2.4 to
s
and . Thus, we can nd M with [
s
[(M) = 0
such that, for every x M, there exists a sequence
h
h=1
satisfying the conditions (2.9) and
(2.10) of Lemma 2.4. Then we simply end since, for every x M,
h
(x)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)

d
s
d[
s
[
(x)
_
B(x,
h
)
k
h
(x y)[
a
(y)[dy
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
+
_
B(x,
h
)
k
h
(x y)
d
s
d|
s
|
(y)
d
s
d|
s
|
(x)
d[
s
[(y)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
,
that goes to zero as h by the conditions (2.9) and (2.10).
We end this section with two useful propositions.
Proposition 2.5. Let
1
,
2
/
loc
(, R
m
) such that
1

2
. Then, for [
1
[-a.e. x ,
d1
d|1|
(x) =
d(
1
+
2
)
d|1+2|
(x).
Proof. Let A B() such that [
1
[(A) = [
2
[(A) = 0: clearly we can prove the wanted equality
only for [
1
[-a.e. x A. Since [
1
+
2
[ = [
1
[ + [
2
[ (see for instance Lemma 6.10),
1
<< [
1
[
and
1
,
2
<< [
1
+
2
[, we have that, for every B B(A),
1
(B) =
_
B
d
1
d[
1
[
(x)d[
1
[(x),
and
1
(B) =
_
B
d
1
d[
1
+
2
[
(x)d[
1
+
2
[(x) =
_
B
d
1
d[
1
+
2
[
(x)d[
1
[(x).
Thus, for [
1
[-a.e. x A,
d
1
d|
1
|
(x) =
d
1
d|
1
+
2
|
(x)
10
. We end noting that, with a similar argument,
for [
1
[-a.e. x A,
d
2
d|1+2|
(x) = 0.
Proposition 2.6. Let
1
,
2
/
loc
(, R
m
) such that
1

2
and let : [0, ] be a Borel
function. Then
_
(x)d[
1
+
2
[(x) =
_
(x)d[
1
[(x) +
_
(x)d[
2
[(x),
and
[
1
+
2
[-ess sup
x
(x) =
_
[
1
[-ess sup
x
(x)
_
_
[
2
[-ess sup
x
(x)
_
.
Proof. Let A B() such that [
1
[( A) = [
2
[(A) = 0. Since [
1
+
2
[ = [
1
[ +[
2
[, we have
_
(x)d[
1
+
2
[(x) =
_
A
(x)d[
1
+
2
[(x) +
_
\A
(x)d[
1
+
2
[(x)
=
_
A
(x)d[
1
[(x) +
_
\A
(x)d[
2
[(x) =
_
(x)d[
1
[(x) +
_
(x)d[
2
[(x),
and
[
1
+
2
[-ess sup
x
(x) =
_
[
1
+
2
[-ess sup
xA
(x)
_
_
[
1
+
2
[-ess sup
x\A
(x)
_
=
_
[
1
[-ess sup
xA
(x)
_
_
[
2
[-ess sup
x\A
(x)
_
=
_
[
1
[-ess sup
x
(x)
_
_
[
2
[-ess sup
x
(x)
_
.
10
Note that if /
+
(), : R
m
is a Borel function such that, for every B B(),
_
B
(x)d = 0 then,
for -a.e. x , (x) = 0.
2.2 Notes on BV , Sobolev and Lipschitz functions
Let us introduce now the notion of function of (locally) bounded variation (for more details see
[7] Chapter 3 or [40] Chapter 5). A Borel function u : R is said to have bounded variation
on , or briey u BV (), if u L
1
() and there exists Du /(, R
n
) such that, for every
C
c
(),
_
(x)dDu(x) =
_
u(x)(x)dx,
or, in other words, if u has a distributional gradient belonging to /(, R
n
).
Since Du /(, R
n
), following (2.3), it can be decomposed with respect to L
n
as
Du = u L
n
+D
s
u,
where u L
1
(, R
n
), D
s
u /(, R
n
) and D
s
u L
n
. When D
s
u = 0 we say u W
1,1
() and
if u L
() and u L
(, R
n
) too, we say u W
1,
().
When, for every open set
, u BV (
) (resp. u W
1,1
(
), u W
1,
(
)), we say
u BV
loc
() (resp. u W
1,1
loc
(), u W
1,
loc
()): in this case u L
1
loc
() (resp. u L
1
loc
(),
u L
loc
()) and Du /
loc
(, R
n
) (resp. Du = u L
n
with u L
1
loc
(, R
n
), Du = u L
n
with u L
loc
(, R
n
)). Finally we will consider sometimes functions belonging to W
1,1
(, R
m
)
or W
1,
(, R
m
): the denitions of these spaces are clear and then omitted.
Given m N and u
h
, u L
1
loc
(, R
m
) (resp. L
loc
(, R
m
)), we say u
h
u in L
1
loc
(, R
m
) (resp.
L
loc
(, R
m
)) if, for every K B(), K and L
n
(K) > 0,
lim
h
_
K
[u
h
(x) u(x)[dx = 0,
_
lim
h
ess sup
xK
[u
h
(x) u(x)[ = 0
_
,
while considering u
h
, u L
loc
(, R
m
), we say u
h
u in w
-L
loc
(, R
m
) if, for every K B(),
K and v L
1
(K),
lim
h
_
K
v(x)u
h
(x)dx =
_
K
v(x)u(x)dx.
The denitions of convergence in L
1
(, R
m
), L
(, R
m
) and w
-L
(, R
m
) are obvious.
It is well known the validity of the proposition below.
Proposition 2.7. Let u
h
h=1
L
loc
(, R
m
) (resp. L
(, R
m
)) be a sequence such that, for
every K B(), K ,
sup
_
ess sup
xK
[u
h
(x)[ : h N
_
<
_
resp. sup
_
ess sup
x
[u
h
(x)[ : h N
_
<
_
.
Then there exists a subsequence u
h
k
k=1
and u L
loc
(, R
m
) (resp. u L
(, R
m
)) such that
u
h
k
u in w
-L
loc
(, R
m
) (resp. w
-L
(, R
m
)).
If now u
h
, u BV
loc
() (resp. W
1,1
loc
()), we say u
h
u in w
-BV
loc
() (resp. W
1,1
loc
()) if
u
h
u in L
1
loc
() and Du
h
Du in w
-/
loc
(, R
n
) (resp. u
h
u in L
1
loc
(, R
n
)). Also in
this case it is clear what we mean for convergence in w
-BV () and W
1,1
().
The following compactness theorem for BV
loc
() functions holds (see [7] Theorem 3.23). For
the notion of set with Lipschitz boundary see [40] page 127.
Theorem 2.8. Let u
h
h=1
BV
loc
() be a sequence such that, for every K ,
sup
__
K
[u(x)[dx +[Du
h
[(K) : h N
_
< .
2.2. NOTES ON BV , SOBOLEV AND LIPSCHITZ FUNCTIONS 15
h
k
k=1
and u BV
loc
() such that u
h
k
u in w
-BV
loc
().
Moreover if is a bounded open set with Lipschitz boundary and
sup
__
[u(x)[dx +[Du
h
[() : h N
_
< ,
then u
h
h=1
BV (), u BV () and u
h
k
u in w
-BV ().
When = (a, b) R and u BV (a, b) it is meaningful, following (2.4), to decompose Du
/(a, b) with respect to L
1
as Du = u
L
1
+D
c
u+D
#
u where the notations are clear. If D
c
u = 0
we say u SBV (a, b). For this space the following simple compactness theorem holds (see [7]
Theorem 4.8: here we present a very special case).
Theorem 2.9. Let (a, b) R be an open bounded interval and let u
h
h=1
SBV (a, b)L
(a, b)
be a sequence such that
11
sup
hN
_
ess sup
x(a,b)
[u
h
(x)[ + ess sup
x(a,b)
[u
h
(x)[ + #(A
Du
)
_
< .
h
k
k=1
and u SBV (a, b) L
(a, b) such that u

h
k
u in
w
-BV (a, b) and u

h
k
u in w
-L
(a, b).
As for Radon measures, let
h
h=1
be a family of open sets and let u
h
L
1
loc
(
h
, R
m
), u
L
1
loc
(, R
m
): we say u
h
u in L
1
loc
( , R
m
) if, for every K , we have K
h
if h is large
enough and
lim
h
_
K
[u
h
(x) u(x)[dx = 0.
If instead u
h
BV
loc
(
h
) (resp. W
1,1
loc
(
h
)), u BV
loc
() (resp. W
1,1
loc
()), we say u
h
u in
w
-BV
loc
( ) (resp. W
1,1
loc
( )) if u
h
u in L
1
loc
( ) and Du
h
Du in w
-/
loc
( , R
n
)
(resp. u
h
u in L
1
loc
( , R
n
)).
Also in this case we suppose known the fundamental properties of the convolutions of a function
u belonging to L
1
loc
(), BV
loc
() or W
1,1
loc
(), and in particular the fact that, using the same
notations than for the measures,
u
(x) =
_
B(x,)
n
k
_
x y
_
u(y)dy :
R,
belongs to C
) and u
u in L
1
loc
( ).
Moreover, when u BV
loc
(), u
= (Du)
and (Du)
L
n
Du in w
-/
loc
( , R
n
), that
is, by denition, u
u in w
-BV
loc
( ), while if u W
1,1
loc
(), u
= (u)
and (u)
u
in L
1
loc
( , R
n
), that is, by denition, u
u in W
1,1
loc
( ).
Let now be a bounded open set in R
n
. We say that a function v W
1,
() is piecewise
ane in if there exists a family
j
N
j=1
of open disjoint subsets of such that
L
n
_
_

N
_
j=1
j
_
_
= 0,
and such that, for every j 1, ..., N, x
j
,
v(x) =
j
, x) +q
j
,
where
j
R
n
and q
j
R. We shall denote the space of such functions with A().
The next theorem is a particular case of Theorem 1.8, Chapter 2 in [29].
11
With #(A) we mean the number of element of A. Remember that A
Du
is the set of the atoms of Du.
Theorem 2.10. Let be a bounded open set with Lipschitz boundary and let u W
1,1
(). Then,
for every > 0, there exists v
A() such that

_
[u(x) v
(x)[dx < .
It is known that Lip
loc
() that is the space of the locally Lipschitz functions dened on , can
be identied with W
1,
loc
() (see [40] Theorem 5, page 131), thus, in our arguments, we will always
identify these two spaces. The following celebrated theorem due to Rademacher holds (see [40]
Theorem 2, page 81).
Theorem 2.11. Let u W
1,
loc
() and let
d
be the set of the points in which u is dieren-
tiable. Then L
n
(
d
) = 0. In particular
d
is dense in .
Let u W
1,
loc
() and let
d
be the set of the points in which u is dierentiable: for every
x , we set
12
,
u(x) = : x
k
x, x
k

d
, u(x
k
) , (2.11)
and
c
u(x) = co
_
u(x)
_
, (2.12)
and we call
c
u(x) the Clarkes gradient of u in x.
Some results about Clarkes gradient are presented here: the proof of Proposition 2.12 can
be found in Clarke [26] (see also Lebourg [58]) while Theorem 2.13 follows from Corollary 8.47,
Proposition 7.15 and Theorem 9.61 in [67].
Proposition 2.12. Let u W
1,
loc
(). Then the following properties hold for
c
u:
(i)
c
u does not change if, in its denition, we consider any
1

d
such that L
n
(
d
1
) = 0;
(ii) for every x ,
c
u(x) is nonempty, convex and compact;
(iii) the multi-valued map
c
u from to the nonempty, convex and compact subsets of R
n
is outer
semicontinuous, that is, if x
k
x,
k

c
u(x
k
) and
k
then
c
u(x);
(iv) if u C
1
() then, for every x ,
c
u(x) = u(x).
Theorem 2.13. Let u, u
h
W
1,
loc
() such that u
h
u with respect to the L
loc
() convergence
and let x
0
be such that u(x
0
) exists. Then there exists a subsequence u
h
k
k=1
and two
sequences x
k
k=1
,
k
k=1
such that, for every k N,
k

c
u
h
k
(x
k
) and it holds x
k
x
0
,
u
h
k
(x
k
) u(x
0
) and
k
u(x
0
).
2.3 Some tools from convex analysis
2.3.1 Principal denitions and properties
A set C R
m
is said convex if, for every , C, t (0, 1), we have t + (1 t) C while C is
said a cone if 0 C and, for every C, t > 0, it is t C.
If we consider any family of convex sets (resp. cones) C
i
iI
then

iI
C
i
is convex (resp. a
cone) too. Thus, given a set A R
m
, it is well dened the convex envelope of A as the set
co(A) =
_
C R
m
: A C, C is convex
_
,
12
See the next section for the denition of convex set and for the notation co(A), that means the convex envelope
of the set A.
2.3. SOME TOOLS FROM CONVEX ANALYSIS 17
so as the positive conic envelope of A given by the set
pos(A) = 0
_
C R
m
: A C, C is a cone
_
.
The next important theorem about the convex envelopes is due to Caratheodory (see [67]
Theorem 2.27).
Theorem 2.14. Let A R
m
, A ,= . Then
co(A) =
_
_
_
m+1
j=1
j
:
j
A,
j
0,
m+1
j=1
j
= 1
_
_
_
.
Given now A R
m
and
0
R
m
it is possible to dene also the positive conic envelope of A
with respect to
0
as
pos
0
(A) =
0
+ pos(A
0
),
and this set can be easily characterized as
pos
0
(A) =
0
(
0
) +
0
: > 0, A .
Let us note that if A is convex, then pos
0
(A) is convex too and that in general pos
0
(A) may fail
to be closed even if A it is (see for example [67] Chapter 3, Section G).
Let us give now some fundamental denitions: for our purposes we can conne ourselves to
consider only non negative functions dened on the whole space R
m
. We remember that a b =
supa, b and a b = infa, b.
Let : R
m
[0, ]:
is convex if, for every , R
m
, t (0, 1), (t + (1 t)) t() + (1 t)(). If,
whenever ,= , the previous inequality is strict we say that is strictly convex;
is level convex if, for every , R
m
, t (0, 1), (t + (1 t)) () ();
is sub-linear if, for every , R
m
, ( +) () +()
13
;
is sub-maximal if, for every , R
m
, ( +) () ();
is positively homogeneous of degree r [0, ) if, for every R
m
, t > 0, it is (t) = t
r
();
is demi-coercive if there exist a > 0, b 0 and R
m
such that, for every R
m
,
a[[ () +, ) +b;
is non constant on straight lines, briey n.c.s.l., if its restriction to any straight line is a
non constant function.
We remember that a convex and nite function is locally Lipschitzian (see De Giorgi [35]
Theorem 3, Section 2 or [40] Theorem 1, page 236), while this is trivially false for a level convex
function. However, it can be proved that both convex and level convex functions are dierentiable
L
n
almost everywhere (see Theorem 2.11 above for the convex case and Crouzeix [28] Theorem 1
for the level convex case).
Given a function : R
m
[0, ] we dene the domain of the set
dom() = R
m
: () < ,
13
Note that some authors (see for instance [49] or [67]) call sub-linear a function that satises the inequality just
introduced together with the positively homogeneity of degree 1 (see the denition below).
and when dom() ,= we say that is proper: we will always consider such functions in order to
avoid triviality. The epigraph of is the set
epi() = (, r) R
m
[0, ) : r () R
m+1
,
while the sub-level set of to level r [0, ) is the set
E
(r) = R
m
: () r R
m
.
The following three propositions will be used in several situations even if sometimes they will
not be expressly quoted. The proofs of Propositions 2.15 and 2.16 are trivial and then omitted
(for the rst part of Proposition 2.15 see Proposition 2.4 in [67]).
Proposition 2.15. A function f : R
m
[0, ] is convex if and only if epi(f) R
m+1
is convex.
A function g : R
m
[0, ] is level convex if and only if, for every r [0, ), E
g
(r) R
m
is
convex.
Proposition 2.16. Let f
h
: R
m
[0, ] be a sequence functions and f = supf
h
: h N. Then
(i) f is l.s.c. if, for every h N, f
h
is l.s.c.;
(ii) f is convex if, for every h N, f
h
is convex;
(iii) f is level convex if, for every h N, f
h
is level convex;
(iv) f is sub-linear if, for every h N, f
h
is sub-linear;
(v) f is sub-maximal if, for every h N, f
h
is sub-maximal.
Proposition 2.17. Let f : R
m
[0, +] be a proper and positively homogeneous of degree 1
function: then f is convex if and only if f is sub-linear. Let g : R
m
[0, +] be a proper and
positively homogeneous of degree 0 function: then g is level convex if and only if g is sub-maximal.
Proof. In order to prove the if part, let , R
m
and t (0, 1). We have
f(t + (1 t)) f(t) +f((1 t)) = tf() + (1 t)f(),
and
g(t + (1 t)) g(t) g((1 t)) = g() g().
In order to prove the only if part, let , R
m
. Then
f( +) = 2f
_
+
2
_
f() +f() and g( +) = g
_
+
2
_
g() g(),
and we end the proof.
We state now the celebrated Jensens inequality for convex functions and another Jensens type
inequality involving level convex functions whose simple proof can be found for instance in [15]
Theorem 1.2.
Theorem 2.18. Let f : R
m
[0, +] be proper, l.s.c and convex function, /
+
() with
() = 1 and L
1
(, R
m
). Then
f
__
(x)d(x)
_
f((x))d(x).
Theorem 2.19. Let g : R
m
[0, +] be a proper, l.s.c. and level convex function, /
+
()
with () = 1 and L
1
(, R
m
). Then
g
__
(x)d(x)
_
-ess sup
x
g((x)).
Finally let f : R
m
[0, ) be a convex function and
0
R
m
: the set
f(
0
) =
_
v R
m
: v,
0
) +f(
0
) f(), R
m
_
,
is called the sub-dierential of f at
0
and it is always a non empty, closed and convex set (see
[66] Section 23). The sub-dierential of f can be seen also as a multi-valued map from R
m
to
the class of non empty, closed and convex subset of R
m
and, from this point of view, f is outer
semicontinuous, that is, if
h
and v
h
f(
h
) is such that v
h
v then v f() (see [66]
Theorem 24.5).
2.3.2 Recession functions and related topics
Given a function : R
m
[0, ], the recession function of is the function dened, for every
R
m
, as
14
() = inf
_
liminf
h
(t
h
h
)
t
h
:
h
, t
h

_
,
and similarly, we dene also, for every R
m
,
() = inf
_
liminf
h
(t
h
h
) :
h
, t
h

_
.
Strictly related to
and
respectively are the functions dened, for every R

m
, as
15
0
() = sup
_
limsup
h
(t
h
)
t
h
: t
h
0
_
,
and
() = sup
_
limsup
h
(t
h
) : t
h
0
_
.
We underline that the denition of
has been already proposed in the work of Gori [51], while

the one of
appears here for the rst time.

Note also that, for every R
m
, we can nd suitable sequences
h
and t
h
(resp.
t
h
0) such that t
1
h
(t
h
h
)
() and (t
h
h
)
() (resp. t
1
h
(t
h
)
0
() and
(t
h
)
()).
We list now several propositions, involving the functions just introduced, which will be very
useful to prove, in particular, the theorems of Chapters 6 and 7. We remark that most of them
are new.
m
[0, ] be a proper function. Then f
is l.s.c., positively
homogeneous of degree 1 and f
(0) = 0. Moreover, if f is convex then f
is convex too.
Let g : R
n
[0, ] be a proper function. Then g
is l.s.c., positively homogeneous of degree 0

but not necessarily proper. Moreover, if g is level convex then g
is level convex too.

14
Note that the denition of recession function here presented agrees with the standard one given, for instance,
in [67] (see [67] Denition 3.17 and Theorem 3.21).
15
The denition of
0
follows [18] equation (2.7).
Proof. The proof of the part involving f
can be found in [67] Theorem 3.21.

Thus let us study g
. Let us x
h
,
0
R
m
such that
h

0
. Then, for every h N, there exist
two sequences t
j
h
j=1
R, and
j
h
j=1
R
m
such that t
j
h
,
j
h

h
and g(t
j
h
j
h
) g
(
h
)
as j . Thus, for every h N, there exists j
h
such that, t
j
h
h
h,
j
h
h

h

1
h
and
g
_
t
j
h
h

j
h
h
_
g
(
h
) +
1
h
and, unless to extract a subsequence, we can suppose that t
j
h
h
and
j
h
h

0
as h . Then
g
(
0
) liminf
h
g
_
t
j
h
h

j
h
h
_
liminf
h
_
g
(
h
) +
1
h
_
= liminf
h
g
(
h
),
and the lower semicontinuity is proved.
The proof of the positive homogeneity of degree 0 is very simple and it can be omitted.
Let us prove now the level convexity, that is, that for every r [0, ) the set : g
() r
is convex. If r < infg() : R
m
there is nothing to prove. Thus, xed r infg() : R
m
and E
g
(r) = : g() r we have
: g
() r =
_
: t
h
,
h
, such that lim
h
g(t
h
h
) r
_
=
>0
_
: t
h
,
h
such that h N, g(t
h
h
) r +
_
=
>0
_
:
h
0,
h
E
g
(r +) such that
h
h

_
=
>0
E
g
(r +).
For every > 0, E
g
(r + ) ,= and by denition we have that E
g
(r + ) is the so called horizon
cone of E
g
(r +) (see [67] Denition 3.3): since the horizon cone of a convex set is convex (see [67]
Theorem 3.6) we end the proof.
m
[0, ] be a proper, l.s.c. and positively homogeneous of degree
1 function. Then f = f
. Let g : R
m
[0, ] be a proper, l.s.c. and positively homogeneous of
degree 0 function. Then g = g
.
Proof. We make the proof only for g and g
since the other case is completely analogous. Let

R
m
: considering
h
= and t
h
we obtain g
() g(). In order to prove the converse

inequality we use the lower semicontinuity of g: indeed, if we consider
h
and t
h
such
that g(t
h
h
) g
(), then
g() liminf
h
g(
h
) = liminf
h
g(t
h
h
) = g
(),
and we achieve the proof.
m
[0, ] be a proper, l.s.c. and sub-linear function. Then f
0
is
l.s.c., positively homogeneous of degree 1, convex and, for every R
m
,
f
0
() = sup
t>0
f(t)
t
.
Let g : R
m
[0, ] be a proper, l.s.c. and sub-maximal function. Then g
is l.s.c., positively
homogeneous of degree 0, level convex and, for every R
m
,
g
() = sup
t>0
g(t).
Proof. Following Proposition 2.2 in [18], which proves the part of the proposition involving f, we
can easily prove the part about g.
Indeed, g
is clearly positively homogeneous of degree 0. Fixed now R

m
, for every r > 0,
we set
(r) = supg(t) : 0 < t r.
We have that is increasing and g
() = lim
r0
(r). However, by the sub-maximality of g we
have, for every r > 0,
(2r) = sup
0<t2r
g(t) = sup
0<tr
g(2t) sup
0<tr
g(t) = (r),
that implies that is constant on (0, ). Therefore
g
() = sup
t>0
g(t),
thus, in particular, being g
the supremum of a family of l.s.c. and sub-maximal functions, it is also

l.s.c. and sub-maximal. Finally, using Proposition 2.17, the level convexity of g
follows too.
Proposition 2.23. Let g : R
m
[0, ] be a proper, l.s.c. and level convex function such that,
for every R
m
0, g
() = . Then there exists a function
: [0, ) [0, ) such that,

for every R
m
,
g()
([[) and lim

t
(t) = . (2.13)
Proof. It is sucient to prove that, for every h N, there exists r
h
> 0 such that, for every
R
m
0, [[ > r
h
, we have g() > h. Let us suppose by contradiction that there exists h
0
N
such that, for every k N, we can nd
k
R
m
, [
k
[ > k, with f(
k
) h
0
. Then [
k
[ and,
unless to extract a (not relabelled) subsequence,

k
|
k
|
S
m1
. Moreover, for every M N,
there exists k
M
such that, for every k k
M
, we have [
k
[ > M. Then, considered
0
R
m
such
that g(
0
) < ,
g
_
0

M
[
k
[
0
+M

k
[
k
[
_
= g
__
1
M
[
k
[
_
0
+
M
[
k
[
k
_
g(
0
) g(
k
) g(
0
) h
0
< ,
and, by the lower semicontinuity of g, we nd, for every M N, the relation
g(
0
+M) liminf
k
g
_
0

M
[
k
[
0
+M

k
[
k
[
_
g(
0
) h
0
< .
However
> g(
0
) h
0
liminf
M
g
_
M
_
0
M
+
__
g
() = ,
and the contradiction is found.
Proposition 2.24. Let : R
m
[0, ] be a proper, l.s.c. and sub-maximal function
16
such that,
for every R
m
0,
() = . Then there exists a function

0
: (0, ) [0, ) such that,
for every R
m
0,
()
0
([[) and lim
t0
0
(t) = . (2.14)
Proof. It suces to prove that, for every h N, there exists
h
> 0 such that, for every R
m
0,
[[ <
h
, we have () > h. Let us suppose by contradiction that there exists h
0
N such that,
for every k N, we can nd
k
R
m
0, [
k
[
1
k
, with (
k
) h
0
. Then [
k
[ 0 and unless
16
In the following we will call every function that is sub-maximal but not level convex.
to extract a (not relabelled) subsequence,

k
|
k
|
S
m1
. Let t
h
h=1
(0, ) such that t
h
0
and
lim
h
(t
h
) = .
For every h, k N, there exists j
h,k
N such that
t
h
|
k
|
j
h,k
<
t
h
|
k
|
+ 1 and since t
h
k
|
k
|
t
h
as
k , then also j
h,k
k
t
h
as k . By the lower semicontinuity and the sub-maximality
of , we have
(t
h
) liminf
k
(j
h,k
k
) = liminf
k
_
_
j
h,k
i=1
k
_
_
liminf
k
(
k
).
Then, for every h N,
(t
h
) liminf
k
(
k
) h
0
< .
Taking the limit as h , we nd a contradiction and the proof is achieved.
m
[0, ] be a proper, l.s.c. and convex function. Let us dene the
function

f : R
m
R [0, ] in this way:
f(, ) =
_
_
_
f(
) if > 0,
f
() if = 0,
if < 0.
Then

f is proper, l.s.c., positively homogeneous of degree 1 and convex (in particular sub-linear)
on R
m+1
.
Proof. See [30] Theorem 3.1.
m
[0, ] be a proper, l.s.c. and level convex function. Let us dene
the function g : R
m
R [0, ] in this way:
g(, ) =
_
_
_
g(
) if > 0,
g
() if = 0,
if < 0.
(2.15)
Then g is proper, l.s.c., positively homogeneous of degree 0 and level convex (in particular sub-
maximal) on R
m+1
.
Proof. The functional g is clearly proper and positive homogeneous of degree 0. In order to prove
the lower semicontinuity we work in the following way. Let us x (
h
,
h
) (
0
,
0
): since g
is lower semicontinuous both on R
m
(0, ) (because of the lower semicontinuity of g) and on
R
m
(, 0), the lower semicontinuity inequality has to be proved only in the case in which
0
= 0. If this is the case, we can nd I
1
, I
2
, I
3
disjoint subsets of N such that I
1
I
2
I
3
= N and
such that, if h I
1
then
h
> 0, if h I
2
then
h
= 0 and if h I
3
then
h
< 0. Since
liminf
h
g(
h
,
h
) = inf
_
liminf
h,hI
i
g(
h
,
h
) : i 1, 2, 3, #(I
1
) =
_
,
it suces to prove the lower semicontinuity inequality of g only in the three cases in which I
i
= N,
i 1, 2, 3. However if, for every h N,
h
> 0, by the denition of g
,
g(
0
, 0) = g
(
0
) liminf
h
g
_
h
_
= liminf
h
g (
h
,
h
) ,
if, for every h N,
h
= 0, by the lower semicontinuity of g
(see Proposition 2.20),

g(
0
, 0) = g
(
0
) liminf
h
g
(
h
) = liminf
h
g (
h
,
h
) ,
and nally if, for every h N,
h
< 0 the inequality is trivially satised: thus the lower semiconti-
nuity of g is achieved.
Let us show now that g is level convex, that is, for every r [0, ), the set
_
(, ) : g(, ) r
_
=
_
(, ) : > 0, g(
1
) r
_
_
(, 0) : g
() r
_
= A B
is convex. We know that AB is closed and, since g
is l.s.c. and level convex, that B is a convex

and closed set. We note also that A is convex since A = or, if A ,= , we have that the convex
set E
g
(r) = : g() r ,= and A = (t, t) : E
g
(r), t > 0 that is convex
17
.
If A = then A B = B that is convex. If A ,= we are going to prove that A B = cl(A)
that is convex since A is convex. Clearly cl(A) A B: to prove the converse we only need to
prove that B cl(A).
In order to prove this let us x (
0
, 0) B and show that there exist two sequences
h

0
and
h
0 such that, for every h N, (
h
,
h
) A, that is, g(
h
1
h
) r. Since A ,= there exists
1
R
m
such that g(
1
) r: we claim that
1
+t
0
: t 0 E
g
(r).
If
0
= 0 there is nothing to prove; instead, supposing
0
,= 0, let x t 0 and consider
h

0
(
h
,= 0), t
h
([t
h
h
[ > t[
0
[) such that g(t
h
h
) g
(
0
): then, for every h N, by the level
convexity of g,
g
_
1
+t[
0
[
t
h
h
[t
h
h
[
_
g(t
h
h
) g(
1
),
since the point
1
+t[
0
[
t
h
h
|t
h
h
|
belongs to the segment joining
1
and t
h
h
. However
lim
h
_
1
+t[
0
[
t
h
h
[t
h
h
[
_
=
1
+t
0
,
and then, by the lower semicontinuity of g,
g(
1
+t
0
) liminf
h
(g(t
h
h
) g(
1
)) r,
that proves the claim.
At last setting, for every h N,
h
=
1+h0
h
and
h
=
1
h
, we have g(
h
,
h
) = g(
h
1
h
) r,
that is (
h
,
h
) A, and moreover
lim
h
(
h
,
h
) = (
0
, 0) cl(A),
that ends the proof.
Proposition 2.27. Let : R [0, ] be a continuous and sub-maximal function and let m =
sup() : R. Then, for every [0, ), () = m or, for every (, 0], () = m.
Proof. First of all we prove that (0) = m. Indeed, if this is not true then (0) is nite, there
exists > 0 such that (0) +2 m and
0
R such that (
0
) (0) +. Then, for every h N,
by the sub-maximality of , also
_
0
h
_
(0) +. But
0
h
0 as h thus, by continuity of
, (0) = lim
h
0
h
_
(0) + that is a contradiction.
Let us suppose now, again by contradiction, that there exist
1
,
2
> 0 such that (
1
)(
2
) <
m: by continuity of , we can suppose also that
1
,
2
Q. Writing
1
2
=
k
h
, where h, k N, we
have h
1
k
2
= 0 and
m = (0) = (h
1
k
2
) (
1
) (
2
) < m :
having found a contradiction, we achieve the proof.
17
Indeed, since A = pos (Eg(r) 1) \ (0, 0) R
m
R and (, ) A implies > 0, the convexity simply
follows.
Let us point out that if is only l.s.c. and sub-maximal on R, then the thesis of Proposition
2.27 is false: to verify this fact one can consider for instance the function dened, for every
R Z, as () = M > 0 and, for every Z, as () = 0.
The proofs of the following two propositions are simple and for this reason they are omitted
(Proposition 2.29 can be nd also in [28]).
m
[0, ] be a proper, l.s.c. and demi-coercive function, that is,
there exist a > 0, b 0 and R
m
m
, f() a[[ , ) b. Then,
for every R
m
, the following properties hold:
(i) f
() f
();
(ii) if f
() < then f
() = 0;
(iii) f
() a[[ , ).
Proposition 2.29. Let g : R [0, ] be a proper and l.s.c. function. Then g is level convex if
and only if g belongs to one of the three following classes:
(i) g is not decreasing on R;
(ii) g is not increasing on R;
(iii) there exists x
0
R such that g is not increasing on (, x
0
] and not decreasing on [x
0
, ).
In particular if g : R
m
[0, ] is a proper, l.s.c. and level convex function, then g(0) = infg() :
R
m
if and only if, for every S
m1
, the function t g(t) is not decreasing on (0, ).
The two propositions below describe some properties of the composition of a level convex
function with a strictly increasing one.
m
[0, ] be a proper, l.s.c and level convex function, s = supg() :
R
m
[0, ] and : [0, s] [0, s] be a continuous, strictly increasing function with s =
sup(t) : t [0, s] [0, ]. Then the composition g : R
m
[0, s] is proper, l.s.c. and level
convex. Moreover ( g)
= g
.
Proof. Clearly g is proper and l.s.c.. Let , R
m
and t (0, 1), then
( g)(t + (1 t)) = (g(t + (1 t))) (g() g()) = (g()) (g()),
that is g is level convex too. In order to prove that ( g)
= g
, we point out that is

bijective and
1
: [0, s] [0, s] is still continuous and strictly increasing. Let us x
0
R
m
and
let t
h
,
h

0
such that ( g)(t
h
h
) ( g)
(
0
). Then, using the continuity of
1
,
( g)
(
0
) = lim
h
(g(t
h
h
)) =
1
_
lim
h
(g(t
h
h
))
_
=
_
lim
h
g(t
h
h
)
_
g
(
0
).
Conversely let

t
h
,

h

0
such that g
_
t
h
h
_
g
(
0
). Then, using the continuity of ,
( g
)(
0
) =
_
g
(
0
)
_
=
_
lim
h
g
_
t
h
h
_
_
= lim
h
_
g
_
t
h
h
__
( g)
(
0
)
and the equality is nally achieved.
m
[0, ] be a proper, l.s.c and level convex function, s = supg() :
R
m
[0, ] and K
g
= R
m
: g
() < s. Let us suppose that cl(K

g
) doesnt contains any
straight line. Then there exists a continuous and strictly increasing function : [0, s] [0, ]
such that the composition g : R
m
[0, ] is demi-coercive.
Proof. First of all let us note that, by the properties of g
, cl(K
g
) is a closed cone of R
m
. Moreover,
it is quite simple to prove that there exists a closed and convex cone C with non empty interior
such that it does not contain any straight line and cl(K
f
) C.
We claim that there exists
0
R
m
, such that
cl(K
f
) 0 C 0 R
m
:
0
, ) > 0. (2.16)
Following Proposition 1 in [53], let us dene C B
m
(0, 1) = C
m
and
0
=
_
C
m
d.
Moreover, for every S
m1
, let us dene the transformation r
as the reection of R
m
with
respect to the hyperplane
18
, that is in formulas, for every R

m
,
r
() = 2, ).
Now let us suppose by contradiction that there exists S
m1
C
m

0
. This implies that
19
0 =
0
, ) =
_
C
m
, )d =
_
R
m
1
C
m(), )d.
Setting H
+
= : , ) > 0 and H
= : , ) < 0 it follows,
0 =
_
H
+
1
C
m(), )d +
_
H
1
C
m(), )d.
We apply the change of variables = r
() to the second integral (note that r
(H
+
) = H
,
that r
preserves m-dimensional volumes, so that its Jacobian is equal to 1, and that r
is norm
preserving). We have,
_
H
1
C
m(), )d =
_
H
+
1
C
m(r
())r
(), )d.
Since, for every R
m
, r
(), ) = , ) we conclude
0 =
_
H
+
_
1
C
m() 1
C
m(r
())
_
, )d. (2.17)
Let us remark that, for every H
+
,
1
C
m() 1
C
m(r
()) 0. (2.18)
Indeed, this is trivial if C
m
or [[ > 1 (since r
is norm preserving). If (H
+
B
m
(0, 1))C
m
then r
() R
m
C
m
, since, if otherwise r
() C
m
, from C
m
and = 2, ) + r
(),
it would be C (note that the convex cones are closed under summation with non negative
coecients) and then in particular, by [[ 1, C
m
.
From (2.17), (2.18) and the denition of H
+
, for L
m
-a.e. H
+
, it is
1
C
m() 1
C
m(r
()) = 0.
Now H
+
C
m
and then, for the properties of the convex set with non empty interior, there
exists a sequence
k
k=1
C
m
, converging to and such that r
(
k
) C
m
. By the continuity of
18
Given a vector R
m
, then
= R
m
: , ) = 0.
19
In the following 1
K
denotes the characteristic function of the set K that is the function dened as 1
K
(x) = 1
if x K, 1
K
(x) = 0 if x , K.
r
and since C
m
is closed, it is r
() = C
m
. Then , C
m
that is t : t R C: thus
a contradiction is found and (2.16) is proved. Clearly (2.16) implies that there exists > 0 such
that
cl(K
f
) 0
_
R
m
0 :
_

[[
,
0
_
>
_
,
where, without loss of generality, we can also assume [
0
[ = 1.
Let us suppose at rst s < . For every h N, we set
C
h
=
_
: g() s
_
1
1
h
__
and D
h
= C
h

_
R
m
0 :
_

[[
,
0
_

_
.
Obviously C
h
h=1
and D
h
h=1
are two increasing sequences of sets. Moreover we have that
every D
h
is bounded. Indeed, if by contradiction there exists h
0
N such that D
h
0
is unbounded,
we can nd a sequence
j
j=1
D
h
0
such that [
j
[ and
j
|
j
|

0
S
m1
. Then
s
_
1
1
h
0
_
liminf
j
g
_
[
j
[

j
[
j
[
_
g
(
0
),
but since
0

_
R
m
0 :
_

||
,
0
_

_
it should be g
(
0
) = s and the contradiction is
found.
Let us dene now
(0) = 2 sup[[ : D
1
, ((t) = 0 if D
1
= ),
and, for every h N, t
_
s
_
1
1
h
_
, s
_
1
1
h+1
__
,
(t) = 2 sup[[ : D
h+1
, ((t) = 0 if D
h+1
= ).
Then : [0, s) [0, ) is clearly increasing and sup(t) : t [0, s) = : thus we can dene
also (s) = .
Let us x now
0

_
R
m
0 :
_

||
,
0
_

_
: if g(
0
) = 0 then
0
D
1
and
( g)(
0
) = (0) = 2 sup[[ : D
1
2[
0
[;
if 0 < g(
0
) < s then there exists h
0
N such that g(
0
)
_
s
_
1
1
h
0
_
, s
_
1
1
h
0
+1
__
and then
0
D
h0+1
and
( g)(
0
) = 2 sup[[ : D
h0+1
2[
0
[;
if at last g(
0
) = s then
( g)(
0
) = 2[
0
[.
Now we can easily prove that ( g) is demi-coercive. Indeed, xed R
m
0, if
_

||
,
0
_

then
( g)() +,
0
) 2[[ [[ = [[,
while, if
_

||
,
0
_
> , then
( g)() +,
0
)
_

[[
,
0
_
[[ > [[.
Thus, since trivially ( g)(0) 0, it holds that, for every R
m
, ( g)() +,
0
) [[.
We achieve the proof of the case s < choosing any function : [0, s] [0, ] which is
continuous, strictly increasing and such that, for every t [0, s], (t) (t): in this way g
is proper, l.s.c. and demi-coercive. The construction of the function is simple and it can be
omitted.
In order to treat the case s = , we can use the same argument once s
_
1
1
h
_
is changed with
h and the denition given for (0) is used to dene on [0, 1].
The two propositions below are based on the following simple equalities: if t
h
, l R, then
liminf
h
(t
h
l) =
_
liminf
h
t
h
_
l and limsup
h
(t
h
l) =
_
limsup
h
t
h
_
l.
m
[0, ] be a proper and Borel function and let l
j
, l R such that
l
j
l. If, for every j N, l
j
is l.s.c. then l is l.s.c. too.
Proof. Let
h
,
0
R
m
such that
h

0
. Then, for every j N,
( l)(
0
) ( l
j
)(
0
) liminf
h
( l
j
)(
h
) =
_
liminf
h
(
h
)
_
l
j
,
and, letting j , we obtain
( l)(
0
)
_
liminf
h
(
h
)
_
l = liminf
h
((
h
) l),
that completes the proof.
m
[0, ] be a proper and Borel function and let l R. Then
( l)
l.
Proof. Let us x
0
R
m
. If we consider t
h
0 such that (t
h
0
)
(
0
), then
( l)
(
0
) limsup
h
( l)(t
h
0
) =
_
limsup
h
(t
h
0
)
_
l =
(
0
) l.
Conversely let

t
h
0 such that ( l) (
t
h
0
) ( l)
(
0
). Then
(
0
) l
_
limsup
h
(
t
h
0
)
_
l = limsup
h
( l) (
t
h
0
) = ( l)
(
0
),
and the proof is achieved.
Finally we state the following theorem involving demi-coercive and convex functions (see Anzel-
lotti, Buttazzo and Dal Maso [8] Theorem 2.4) in which, in particular, the equivalence between
demi-coercivity and the property to be n.c.s.l. is proved.
m
[0, ] be a proper, l.s.c. and convex function and let
0
R
m
.
Then the following conditions are equivalent:
(a) f is demi-coercive;
(b) f is n.c.s.l.;
(c) there are no straight line containing
0
along which f is constant;
(d) the set R
m
: f
() = 0 contains no straight line;

(e) the set R
m
: 2f(
0
) = f(
0
+) +f(
0
) is bounded;
(f ) for every R
m
, ,= 0, there exists t > 0 such that 2f(
0
) < f(
0
+t) +f(
0
t);
(g) for every R
m
, ,= 0, we have f
() +f
() > 0;
(h) there exist a, b R, a > 0, such that, for every R
m
, f(
0
+) +f(
0
) a[[ b;
(i) there exists R
m
m
, ,= 0, f
() , ) > 0.
Chapter 3
Approximation of convex functions
In order to study the lower semicontinuity of the integral functional I given by (1.1) it is funda-
mental to be able to approximate from below the integrand f with more regular functions. For
this purpose many eorts has been spent to nd dierent strategies to build such approximation.
In this chapter we present several theorems on this topic: while in the rst part we state only the
classical results by De Giorgi and Serrin, the second part contains some recent theorems obtained
by Gori and Maggi in [53].
3.1 De Giorgis and Serrins approximation methods
Let R
d
be an open set and let f : R
n
[0, ]: we say that f has compact support on
if there exists
such that, for every (t, ) (
) R
n
, we have f(t, ) = 0.
The following approximation result was proved by De Giorgi (see [35] Theorem 3, Section 3).
Theorem 3.1. Let R
d
be an open set and f : R
n
[0, ) be a continuous function
with compact support in such that, for every t , f(t, ) is convex on R
n
. Then there exists a
sequence
q
q=1
C
c
(R
n
),
q
0 such that, setting
a
q,0
(t) =
_
R
n
f(t, )
_
(n + 1)
q
() +
n
h=1
h
()
_
d, (3.1)
and, for every h 1, . . . , n,
a
q,h
(t) =
_
R
n
f(t, )
h
()d, (3.2)
the sequence of functions given, for every j N, by
f
j
(t, ) = max
1jq
_
0, a
q,0
(t) +
n
h=1
a
q,h
(t)
h
_
,
satises the following conditions:
(i) for every j N, f
j
: R
n
[0, ) is a continuous function with compact support in
such that, for every t , f
j
(t, ) is convex on R
n
. Moreover, for every (t, ) R
n
,
f
j
(t, ) f
j+1
(t, ) and
f(t, ) = sup
jN
f
j
(t, );
29
30 CHAPTER 3. APPROXIMATION OF CONVEX FUNCTIONS
(ii) for every j N, there exists a constant M
j
> 0 such that, for every (t, ) R
n
,
[f
j
(t, )[ M
j
(1 +[[) , (3.3)
and, for every t ,
1
,
2
R
n
,
[f
j
(t,
1
) f
j
(t,
2
)[ M
j
[
1

2
[ . (3.4)
The next theorem is instead due to Serrin (see [69] Lemmas 5 and 8).
Theorem 3.2. Let R
d
n
[0, ) be a continuous function such
that, for every t , f(t, ) strictly convex on R
n
. Then, for every > 0 and
, there
exists a continuous function

f : R
n
[0, ) with compact support on
such that
1
:
(i) for every t ,

f(t, ) is convex;
(ii) for every (t, ) R
n
,

f(t, ) f(t, ) +;
(iii) for every t
, B
_
0,
1
_
, [
f(t, ) f(t, )[ ;
(iv)

f : R
n
R
n
exists, is continuous and there exists a constant M > 0 such that, for
every (t, ) R
n
,

f(t, )
M,
and, for every t
1
, t
2
, R
n
,

f(t
1
, )

f(t
2
, )
M[t
1
t
2
[(1 +[[).
3.2 Approximation of convex and demi-coercive functions
Let us introduce now two approximation results that will be instrumental in the future and that
can be found in [53] Theorem 4 and Theorem 5 respectively.
3.2.1 Approximation by means of convex cones
The following theorem says that a convex, demi-coercive function can be approximated by demi-
coercive cones. The advantage of this approximation is given by the fact that these cones, under
a certain point of view, behave better than the supporting hyper-planes.
Theorem 3.3. Let R
d
n
[0, ) be a l.s.c. (resp. continuous)
function such that, for every t , f(t, ) is convex and demi-coercive on R
n
. For every
0
R
n
,
let us dene P
0
f : R
n
[1, ) as
P
0
f(t, ) = inf
_
: (, ) pos
(0,1)
(epi f(t, ))
_
.
Then we have that:
(i) P
0
f is l.s.c. (resp. continuous) and, for every t , P
0
f(t, ) is convex, demi-coercive
and can be characterized as the greatest function less than or equal to f such that, for every
t , the map
(1 + P
0
f(t, +
0
))
is positively homogeneous of degree 1;
1
With
f(x, s, ) we mean the gradient of f(x, s, ).

3.2. APPROXIMATION OF CONVEX AND DEMI-COERCIVE FUNCTIONS 31
(ii) there exists a sequence
k
k=1
R
n
such that, for every (t, ) R
n
,
f(t, ) = sup
kN
P
k
f(t, ).
Proof. Let us prove (i). By the denition of P
0
f we have, for every t ,
epi
_
P
0
f(t, )
_
= cl
_
pos
(
0
,1)
(epi f(t, ))
_
.
Since, for every t , epi (f(t, )) epi (P
0
f(t, )) we obtain P
0
f f on R
n
and, since the
positive conic envelope of a convex set is convex too
2
, we have also that, for every t , P
0
f(t, )
is convex. The map (1 + P
0
f(t, +
0
)) clearly is positively homogeneous of degree 1.
In order to show that P
0
f is the greatest function less than or equal to f such that it satises
the property about the positive homogeneity of degree 1 asked in (i), we note that, if g is such that
g f and, for every t , (1 +g(t, +
0
)) is positively homogeneous of degree 1, then, for
every t ,
epi(g(t, )) cl
_
pos
(
0
,1)
(epi f(t, ))
_
= epi
_
P
0
f(t, )
_
,
so that g P
0
f.
It worth noting also that, for every (t, ) R
n
, 1 +P
0
f(t, +
0
) f
(t, )
3
. Indeed, by
means of Denitions 3.3 and 3.17 and Theorem 3.21 in [67], for every t , it holds
epi
_
f
(t, )
_
=
_
(, ) : (
h
,
h
) epi(f(t, )), t
h
such that
(
h
,
h
)
t
h
(, )
_
.
Thus let (, ), (
h
,
h
) and t
h
as above: since (
h
,
h
) epi(f(t, )) we have
(
h

0
,
h
+ 1)
t
h
+ (
0
, 1) pos
(0,1)
(epi f(t, ))
and then, passing to the limit as h , we obtain
( +
0
, 1) cl
_
pos
(0,1)
(epi f(t, ))
_
= epi
_
P
0
f(t, )
_
if and only if
(, ) epi
_
1 + P
0
f(t, +
0
)
_
,
that proves the wanted inequality.
Now let us point out that, by denition, we have
P
0
f(t, ) = inf
_
: (, ) pos
(
0
,1)
(epi f(t, ))
_
= inf
_
1 +( + 1) : 0, f(t, ), =
0
+(
0
)
_
= 1 + inf
_
+f
_
t,
0
+

0
_
: > 0
_
. (3.5)
Using this formula, by a simple computation, we can prove that, for every t , P
0
f(t, ) is
demi-coercive.
2
See [67] Chapter 3, Section G.
3
With f
(t, ) we mean (f(t, ))
().
Moreover (3.5) allows to achieve also the lower semicontinuity (resp. continuity) of P
0
f.
Indeed, let us consider the function

f : R
n
R [0, ) given, as in Proposition 2.25, by
f(t, , ) =
_
_
f
_
t,

_
if > 0,
f
(t, ) if = 0,
+ if < 0.
(3.6)
By Theorem 3.1 of [30],

f is l.s.c. on R
n
R and then the same property is had also by
f(t, , ) = +

f(t,
0
, ). By (3.5), the denition of f and the remark made above about f
we have
P
0
f(t, ) = 1 + inf
_
f(t, , ) : 0
_
,
but, since f(t, , ) and f(t, , 1) = 1 +f(t, ), it is in fact
P
0
f(t, ) = 1 + inf
_
f(t, , ) : 0 1 +f(t, )
_
.
In particular, by the lower semicontinuity of f, for every (t, ) there exists 0
(t,)
1 + f(t, )
such that
P
0
f(t, ) = 1 +f
_
t, ,
(t,)
_
.
Let us verify now the lower semicontinuity of P
0
f. Let (t
h
,
h
), (t
1
,
1
) R
n
such that
(t
h
,
h
) (t
1
,
1
) and suppose, without loss of generality, that
liminf
h
P
0
f(t
h
,
h
) = lim
h
P
0
f(t
h
,
h
).
Let us dene also
h
=
(t
h
,
h
)
,
1
=
(t
1
,
1
)
. If it is
h
, since
h
f(t
h
,
h
,
h
) = 1 + P
0
f(t
h
,
h
),
the lower semicontinuity inequality is trivially veried. If otherwise
h
is bounded we can suppose
h

0
and then
1 + P
0
f(t
1
,
1
) = f(t
1
,
1
,
1
) f(t
1
,
1
,
0
) liminf
h
f(t
h
,
h
,
h
) = liminf
h
(1 + P
0
f(t
h
,
h
)) ,
that completes the proof of the lower semicontinuity of P
0
f.
Finally, again by (3.5), we have also that the continuity of f implies the upper semicontinuity
(and then continuity) of P
0
f.
In order to prove (ii) we show at rst that
f(t, ) = sup
0
R
n
P
0
f(t, ). (3.7)
Without loss of generality, we can drop the dependence on t. Moreover we can reduce us to consider
the one dimensional case, as we can see by the following argument. Fixed S
n1
, we can dene,
for every R,
f
() = f(),
then, by the maximality property of P
0
f, we have that, for every
0
, R,
P
f() = P
0
f
().
If the approximation holds in dimension one we have, for every S
n1
and R,
f
() = sup
0R
P
0
f
()
so that,
f() = f
() = sup
0R
P
0
f
() = sup
0R
P
f() sup
0
R
n
P
0
f() f(),
that implies (3.7).
Therefore we prove (3.7) for n = 1. For this we consider a function f : R [0, ) which is
convex (and then continuous) and non constant (that is demi-coercive) and
0
R, and dene the
following set
f
(
0
) =
_
v R : v(
0
) 1 f(), R
_
.
Let us note that, for every
0
R, 0
f
(
0
). Moreover, for every ,
0
, v R, v ,= 0,
v(
0
) +f(
0
) = v
_

0

1 +f(
0
)
v
_
1, (3.8)
and then
v f(
0
) if and only if v
f
_
0

1 +f(
0
)
v
_
.
Let us dene
(
0
, v) =
0

1+f(0)
v
. For every
0
R, we set
f
(
0
) = inf
f
(
0
) 0 and
f
(
0
) = sup
f
(
0
) 0. By the maximality property of P
0
f, we have
P
0
f() = max
_
f
(
0
)(
0
) 1,
f
(
0
)(
0
) 1
_
. (3.9)
Since f is not constant and convex it is either lim
f() = or lim
f() = : we
consider the case in which only lim
f() = proving in this hypothesis that, for every R,

f() = sup
0
R
_
f
(
0
)(
0
) 1
_
: (3.10)
the other two cases (lim
f() = and lim

||
f() = ) follow immediately.
Let us consider the following (possibly empty) set,
A = R : f() = 0 ,
and claim that if A is not empty then it is connected and f is constant on A. If A =
0
the
claim is obvious. If instead there are
1
,
2
A,
1
<
2
, let us consider
1
< <
2
: by denition
of sub-dierential, f(
1
) = f(
2
) = m and f() m. However, by the convexity of f, f() m
too. Then, for every
1
< <
2
, we have f() = m and this implies f() = 0, that is A:
thus the claim is proved.
By (3.8) we immediately see that (3.10) holds on every RA with the supremum attained
on
0
=
(, v), for every choice of v ,= 0, v f(): if A = the proof is achieved.

Thus, let us suppose that A is non empty. Since f() as , we have that A is bounded
from above, so that we can consider
A
= supA < . By continuity of f, it is f(
A
) = m. By
choosing
h
A with
h

A
, we have also
0 = limsup
h
f(
h
) f(
A
).
Therefore 0 f(
A
) and, for every A, we have
A
and f() = f(
A
) = m. Let us x
A, and show the validity of (3.10). If
A
RA, there exists v
A
f(
A
) with v
A
> 0, such
that [0, v
A
] f(
A
) (remember that f(
A
) is convex). Then, considering v
h
= h
1
v
A
f(
A
),
we have
v
h
(
A
) +f(
A
) = v
h
(
(
A
, v
h
)) 1
f
(
(
A
, v
h
))(
(
A
, v
h
)) 1 f() = m,
and since the left hand side tends to f(
A
) = m as h , we conclude. If otherwise
A
A, let
us consider
h

A
, with
h
>
A
. We know that
limsup
h
f(
h
) f(
A
) = 0,
then we can choose v
h
0 with v
h
f(
h
). In particular it is
v
h
(
h
) +f(
h
)
f
(
(
h
, v
h
))(
(
h
, v
h
)) 1 f() = m,
and since the left hand side tends to f(
A
) = m as h , we complete the proof of (3.7).
It remains to show the validity of (ii) once it is known (3.7). This can be simply proved by
means of the following lemma.
Lemma 3.4. Let A R
n
and let ( be a set of l.s.c. functions from A to R. Let us dene, for
every x A, f(x) = sup
gG
g(x). Then there exists a sequence g
h
h=1
(, such that, for every
x A, f(x) = sup
hN
g
h
(x).
Since its proof consists only in a slight modication of the argument used in the proof of Lemma
9.2 in [43], we omit the details.
3.2.2 Approximation by means of strictly convex functions
As a consequence of Theorem 3.3 we show that the class of functions that can be represented as a
countable supremum of strictly convex ones is characterized by the demi-coercivity.
Theorem 3.5. Let R
d
be an open set and let f : R
n
[0, ) be a continuous function.
Then the two following conditions are equivalent:
(i) for every t , f(t, ) is convex and demi-coercive.
(ii) there exists a sequence f
j
j=1
such that, for every (t, ) R
n
,
f(t, ) = sup
jN
f
j
(t, ), (3.11)
where, for every j N, f
j
: R
n
[2, ) is a continuous function such that, for
every t , f
j
(t, ) is strictly convex in R
n
and, for every
, there exists a constant

C
j,
> 0 such that, for every (t, )
R
n
,
f
j
(t, ) C
j,
(1 +[[). (3.12)
Proof. The proof of (ii) (i) is obvious and can be omitted.
In order to prove the converse it is sucient to show that every conic type function given
in Theorem 3.3 can be approximated from below by a sequence of continuous functions, strictly
convex in the variable . Thus, we can suppose also that f(t, ) is positively homogeneous of degree
1.
We start proving that there exists a continuous function N : R
n+1
, such that, for every
t ,
epi(f(t, ))0 M R
n+1
: M, N(t)) > 0. (3.13)
Following exactly the rst part of Proposition 2.31, we can prove that the vector valued function
N(t) =
_
K(t)
NdN,
where K(t) = epi (f(t, )) B
n+1
(0, 1), satises just the above condition.
Then we only have to show that N(t) is continuous on . To this end let us x t
h
, t
0
,
t
h
t
0
, and consider
C = sup
_
f(t, ) : t t
0
t
h
h=1
, [[ 1
_
1 < ,
since it is the supremum of a continuous function on a compact set. For
S(t) =
_
(, ) R
n+1
: [[ 1, [0, C]
_
epi (f(t, )) ,
we have
[N(t
h
) N(t
0
)[ L
n+1
(K(t
h
)K(t
0
))
L
n+1
(S(t
h
)S(t
0
)) =
_
B
n
(0,1)
[f(t
h
, ) f(t
0
, )[d 0,
as h , where we used the notation AB = (A B) (B A). Thus also the continuity of
N(t) is proved.
Since it must be N(t) = ((t), a(t)) with (t) R
n
, a(t) > 0, then the inclusion (2.16) implies
that, for every R
n
0,
g(t, ) = f(t, ) +
(t), )
a(t)
> 0. (3.14)
The function c(t) = ming(t, v) : v S
n1
> 0 is continuous on and c(t)[[ g(t, ).
For every (0, 1), let us consider
h
(t, ) =
_
2
+c(t)
2
[[
2
, (t, ) R
n
.
Obviously, for every (0, 1), 0 h
(t, ) c(t)[[ and h
(t, ) c(t)[[ as 0. Moreover for

every (t, ) R
n
, (0, 1),
dh
d
(t, ) = 1 +

_
2
+c
2
(t)[[
2
0,
so that it follows h
(t, ) c(t)[[ as 0. Let us note also that, for every t , (0, 1) xed,
h
(t, ) is strictly convex. In order to prove this, we consider , R

n
, ,= , (0, 1); if [[ ,= [[
then
_
2
+c(t)
2
[ + (1 )[
2
2
+c(t)
2
([[ + (1 )[[)
2
<
_
2
+c(t)
2
[[
2
+ (1 )
_
2
+c(t)
2
[[
2
;
if [[ = [[ then [ + (1 )[ < [[ + (1 )[[, thus
_
2
+c(t)
2
[ + (1 )[
2
<
_
2
+c(t)
2
([[ + (1 )[[)
2

_
2
+c(t)
2
[[
2
+ (1 )
_
2
+c(t)
2
[[
2
.
Now we dene, for every (0, 1), f
: R
n
R as
f
(t, ) = (1 )g(t, ) +h
(t, )
(t), )
a(t)
, (t, ) R
n
.
This is a continuous function, strictly convex in the variable (as it is a sum of a convex function
and a strictly convex one). Clearly f
(t, ) f(t, ) as 0. Since h
(t, ) c(t)[[ g(t, ) we

have also
f
(t, ) g(t, )
(t), )
a(t)
= f(t, ).
For every (t, ) R
n
, (0, 1),
df
d
(t, ) = g(t, ) +h
(t, ) +
dh
d
(t, ) 0,
then f
(t, ) f(t, ) as 0. However, note that f
is not necessarily bounded from below, as

required in the statement: indeed, dening the open set
+
= t : (t) ,= 0, we have that,
for every t
+
, surely f
(t, ) 0, while, when t

+
, f
(t, ) could tend to as [[ .

In order to solve this problem is sucient to prove the existence of a continuous function
g
: R
n
[1, ) such that, for every t
+
, g
(t, ) is strictly convex, for every t

+
,
g
(t, ) = 0, and g
f. Indeed, in this case, we can dene, for every (t, ) R

n
, the function
f
(t, ) = f
(t, ) g
(t, ),
that is continuous, as the maximum between two continuous functions, f f
1, and for
every t , f
(t, ) is strictly convex, since, for every t

+
, it is the maximum of two strictly
convex functions while, for every t
+
, f
(t, ) = f
(t, ) that is strictly convex. Moreover

f
(t, ) f(t, ) as 0 since f
(t, ) f(t, ) as 0 and g
f.
For what concerns the existence of g
we can argue as follows. Let us dene, for every s R,

(s) = (s 0) 1 + exp
_
[s[
2
_
,
which is strictly convex, belongs to C
2
(R) and 1 (s) (s 0). Moreover, for every t
+
,
let us dene also b(t) =
(t)
a(t)
> 0 and (t) =

(t)
a(t)b(t)
. Clearly
f(t, ) c(t)[[ +b(t)(t), )
and, setting d(t) = (b(t) 1)(c(t) 1) > 0, we have, for every v S
n1
,
f(t, ) d(t)[[ +b(t)(t), ) d(t)v, ) +b(t)(t), ).
Then it follows that
4
, for every t
+
and v S
n1
,
f(t, )
1
n1
_
S
n1
_
e(t, v), ) 0
_
dH
n1
(v)
n1
_
S
n1
_
e(t, v), )
_
dH
n1
(v) =: g
(t, ),
where e(t, v) = d(t)v +b(t)(t). Thus, let us dene g
: R
n
R as
g
(t, ) =
_
_
_
1
n1
_
S
n1

_
e(t, v), )
_
dH
n1
(v) if t
+
,
0 if t
+
.
(3.15)
Clearly 1 g
f on R
n
and, by Lebesgue dominated convergence Theorem, it can be
veried that g
is continuous on
+
R
n
. If now (t
0
,
0
) (
+
) R
n
then lim
tt
0
b(t) =
lim
tt0
d(t) = 0 (we remark that a(t) is locally uniformly positive on ), thus lim
tt0
e(t, v) = 0
uniformly as v S
n1
. As a consequence lim
(t,)(t
0
,
0
)
g
(t, ) = 0 that implies that g
is continuous
on R
n
.
4
Here and in the following 1
n
denotes the usual n-dimensional Haussdorf measure and n = 1
n
(S
n
).
It remains to check that, for every t
+
, g
(t, ) is strictly convex. By the fact that C

2
(R)
and by Lebesgue dominated convergence Theorem, we have, for every t
+
, g
(t, ) C
2
(R
n
)
and
5
(t, ) =
1
n1
_
S
n1
_
e(t, v), )
_
e(t, v) e(t, v)dH
n1
(v).
Let us x (t, )
+
R
n
and w S
n1
: to achieve the proof it suces to show that
(t, )w, w
_
> 0.
Since
(t, )w, w
_
=
1
n1
_
S
n1
_
e(t, v), )
_
e(t, v), w)
2
dH
n1
(v),
we only have to prove that, for H
n1
-a.e. v S
n1
, it cannot be
e(t, v), w) = d(t)v +b(t)(t), w) = 0.
But if this holds, by continuity of e(t, ), for every v S
n1
, we also have
d(t)v +b(t)(t), w) = 0.
For v w
we deduce that w (t)
and then, for v = w, we get w = 0 , S

n1
.
The proof of the last property stated in (ii) is simple and can be omitted.
5
With
2
g(t, ) we mean the hessian matrix of g(t, ). Moreover denotes the standard tensorial product
between two vectors of R
n
.
Chapter 4
Lower semicontinuity for integral
functionals
4.1 A brief historical background
As announced in the introduction, this chapter is devoted to the problem of determining some new
conditions sucient to guarantee the lower semicontinuity, with respect to the L
1
loc
() convergence,
of functionals of the type
I(u, ) =
_
f(x, u(x), u(x))dx,

dened on the Sobolev space W
1,1
loc
(). This means to nd out conditions on f such that, for every
u
h
, u W
1,1
loc
(), u
h
u in L
1
loc
(), it follows
I(u, ) liminf
h
I(u
h
, ).
In the following the function f will satisfy the usual conditions
_
f : R R
n
[0, ],
for every (x, s) R, f(x, s, ) is convex in R
n
.
(4.1)
One of the rst results on this argument is an example due to Aronszajn (see [64] page 54)
which shows that conditions (4.1), together with the continuity of f, are not sucient for the
lower semicontinuity of I. Several years later, Serrin was able to prove the following fundamental
theorem in which some sucient conditions for the lower semicontinuity are presented (see [69]
Theorem 12).
Theorem 4.1. Let f be a continuous function satisfying (4.1) and one of the following conditions:
(a) for every (x, s) R, lim
||
f(x, s, ) = ;
(b) for every (x, s) R, f(x, s, ) is strictly convex in R
n
;
(c) for every i, j 1, . . . , n, the derivatives
f
xi
,
f
j
and

2
f
xij
exist and are continuous.
Then the functional I is l.s.c. on W
1,1
loc
1
loc
() convergence.
The conditions (a), (b) and (c) quoted above are clearly independent, in the sense that we can
nd a continuous function f satisfying just one of them, but none of the other ones. However, the
39
40 CHAPTER 4. LOWER SEMICONTINUITY FOR INTEGRAL FUNCTIONALS
proof of Theorem 4.1 is essentially the same for every condition considered; indeed, the proof is
based on an approximation theorem for convex functions depending continuously on parameters
(see [69] Lemma 5) that can be applied, in particular, when f satises (a), (b) or (c). This fact
suggests the possibility to nd a suitable condition, implied independently by (a), (b) and (c),
which is still sucient for the lower semicontinuity of I. This problem, that is still open and seems
to be very dicult, is the argument of this chapter.
The program to unify in a unique condition (a), (b) and (c) has to start, rst of all, from the
analysis of the lower semicontinuity theorems and counterexamples we can nd in literature: in
this way it could be understood what kind of diculties we should expect.
The next theorem, again due to Serrin (see [69] Theorem 11), is relevant also because many
authors tried to improve it, dealing, in particular, with condition (i).
(i) there exists a modulus of continuity
1
such that, for every (x
1
, s
1
), (x
2
, s
2
) R, R
n
,
[f(x
1
, s
1
, ) f(x
2
, s
2
, )[ ([x
1
x
2
[ +[s
1
s
2
[) (1 +f(x
1
, s
1
, )) ; (4.2)
(ii) there exist two moduli of continuity and and a constant C > 0 such that, for every t > t
0
big enough, we have (t) Ct and, for every (x
1
, s
1
), (x
2
, s
2
) R, R
n
,
[f(x
1
, s
1
, ) f(x
2
, s
2
, )[ ([x
1
x
2
[) (1 +f(x
1
, s
1
, )) +([s
1
s
2
[). (4.3)
Then, for every u
h
, u W
1,1
loc
() such that u
h
u in L
1
loc
(), we have
I (u, ) liminf
h
I (u
h
, ) ,
assuming in addition, when the case (i) is considered, that u C().
The following lower semicontinuity result was proved by Dal Maso (see [30] Theorem 3.2)
and just moves in the direction described above: note that the assumption of continuity on the
limit function u of Theorem 4.2 is removed
2
, while some coercivity and growth conditions are now
introduced.
Theorem 4.3. Let be a bounded open set and f be a Borel function satisfying (4.1) such that,
for H
n
-a.e. (x
0
, s
0
) R
n
, f(x
0
, s
0
, ) is continuous on R
n
and, for every > 0, there exists
> 0 such that, for every (x, s) B(x
0
, ) B(s
0
, ), R
n
,
[f(x, s, ) f(x
0
, s
0
, )[ < (1 +[[). (4.4)
Let us suppose that, for every r > 0, there exist M > 0, m, A C
0
(), with, for every x ,
m(x) > 0, A(x) 0, and a L
1
(), such that, for every (x, s, ) [r, r] R
n
,
m(x) [[ a(x) f(x, s, ) M [[ +A(x).
Then, for every u
h
, u W
1,1
loc
() such that u
h
u in L
1
loc
(),
I (u, ) liminf
h
I (u
h
, ) ,
assuming in addition that u L
().
1
A function is said a modulus of continuity if : (0, ) (0, ) and lim
t0
(t) = 0
2
In fact Dal Maso was able to extend his analysis to u BV
loc
(), also considering sequences of integral
functionals which -converge
4.1. A BRIEF HISTORICAL BACKGROUND 41
In the same paper, Dal Maso proposed also a revisited form of the example by Aronszajn quoted
above.
A recent extension of Theorems 4.2(i) and 4.3 is due to Fonseca and Leoni (see [43] Theorem
1.1)
3
:
Theorem 4.4. Let f be a Borel function satisfying (4.1) and such that, for every (x
0
, s
0
) R
and for every > 0, there exists > 0 such that, for every (x, s) B(x
0
, ) B(s
0
, ) and R
n
,
f(x
0
, s
0
, ) f(x, s, ) (1 +f(x, s, )) . (4.5)
1,1
loc
1
loc
() convergence.
Note that the assumption (4.5) is a kind of lower semicontinuity of f with respect to (x, s)
R, uniform with respect to R
n
. This condition seems more natural than the continuity
conditions (4.2) and (4.4) required in Theorems 4.2(i) and 4.3, because of the following theorem
due to Fusco (see [45] Proposition 3.1) in which, dealing with the case f(x, s, ) = a(x) [[, the
author points out that a necessary condition for the lower semicontinuity of the functional I is the
lower semicontinuity of the function a.
Theorem 4.5. Let = (0, 1). Let f : R [0, ) dened, for every (x, ) R, as
f(x, ) = a(x) [[, where a : [0, ) is a bounded Borel function. Then, for every u W
1,1
loc
(),
R
_
L
1
loc
(I) (u, ) =
_
a(x) [u
(x)[ dx, (4.6)

where a(x) = sup
_
b(x) : b(x) a(x) for L
1
-a.e. x , b is l.s.c. on
_
. Thus in particular
I (u, ) =
_
a(x) [u
(x)[ dx
is l.s.c on W
1,1
loc
1
loc
() convergence if and only if a is l.s.c. on (that is,
there exists a l.s.c. function a : [0, ) such that, for L
1
-a.e. x , a(x) = a(x)).
Some researches had the aim to weaken the assumptions on f related to the dependence on s.
The next result, due to De Giorgi, Buttazzo and Dal Maso (see [37] Theorem 1), was the starting
point of this kind of results: here, roughly speaking, it is proved that, when the dependence on x
is dropped, the lower semicontinuity of I is always veried every time f is convex on the gradient
variable and it is regular enough to allow the composition f(u(x), u(x)) to be measurable.
Theorem 4.6. Let f : R R
n
[0, ] be a Borel function such that, for every s R, f(s, ) is
convex on R
n
, for every R
n
, f(, ) is measurable on R and, in particular, f(, 0) is l.s.c. on
R. If
limsup
||0
(f(s, 0) f(s, ))
+
[[
L
1
loc
(R), (4.7)
then the functional I is l.s.c. on W
1,1
loc
1
loc
() convergence.
Theorem 4.6 was generalized by Ambrosio [4], and subsequently by De Cicco [31], [32] to the
BV setting. In 1999 Fonseca and Leoni (see [43] Theorem 1.5) obtained the same conclusion of
Theorem 4.6 for integrands f, depending explicitly on the x variable too, under an assumption of
continuity of f on x uniform with respect to (s, ) R R
n
(similar to (4.5)), together with
a condition like (4.7) in which the variable x is also considered.
3
Also in this case the authors are able to consider u BV
loc
(): we quoted again the particular case when
u
h
, u W
1,1
loc
(), for a better comparison with the other results presented.
At this point it should be clear that the dependence of f on x must be treated carefully
in studying the lower semicontinuity of I with respect to the L
1
loc
() convergence. Of course
the dependence on x gives diculties not only technical because of the existence of Aronszajns
counterexample (in which f just depends on x). In particular, measurability of f with respect to
x is not appropriate thanks to Theorem 4.5.
In the paper of Gori and Marcellini [55] (see also Gori [50]) it was specically considered the
dependence of f with respect to x . In the results quoted above some qualied assumptions of
continuity, or lower semicontinuity, of f with respect to x uniformly on R
n
were supposed: in
[55] (see Theorem 1.6) it was proposed a new local condition that, in addition to the continuity of
f and (4.1), it is sucient for the lower semicontinuity.
Theorem 4.7. Let f be a continuous function satisfying (4.1) and such that, for every open set
HK RR
n
there exists a constant L = L
HK
such that, for every x
1
, x
2

,
s H and K,
[f(x
1
, s, ) f(x
2
, s, )[ L[x
1
x
2
[ . (4.8)
1,1
loc
1
loc
() convergence.
Condition (4.8) means that f is locally Lipschitz continuous with respect to x, locally with
respect to (s, ) and not necessarily globally, that is, the Lipschitz constant is not uniform for
(s, ) R R
n
. This allows us to obtain, as a corollary, an improvement of Serrins Theorem
4.1(c) since, when only the gradient
4
x
f exists and is continuous, this implies the Lipschitz
continuity of f with respect to x on the compact subsets of R R
n
(even if not necessarily
Lipschitz continuity of f on the full set R R
n
): Theorem 4.7 can be considered the rst strict
generalization of Theorem 4.1.
Moreover Theorem 4.7 underlines that hypothesis (c) in Theorem 4.1 seems to have a dierent
nature than hypotheses (a) and (b): indeed, roughly speaking, what assumption (c) really contains
is a regularity requirement in the geometric variable x, while (a) and (b) carry on some geometric
constraints on the convexity of f in the gradient variable .
Thus, in order to unify all the conditions (a), (b) and (c), we could try to understand if condition
(4.8) of Theorem 4.7 can be further improved and if hypotheses (a) and (b) come from the same
source: we give an answer to these questions in the following section.
4.2 Statement of the main theorem
Before stating the main result of this chapter, it is suitable to make some remarks. Let us note
that the condition (4.8) of Theorem 4.7 can be formulated in the following equivalent way: for
every (s, ) R R
n
,
f(, s, ) W
1,
loc
(),
and, for every
H K R R
n
HK
such that, for
every (s, ) H K,
ess sup
x
[
x
f(x, s, )[ L.
This way to read the condition (4.8) provides the right point of view to nd a generalization of
Theorem 4.7. Indeed, a new sucient condition can be found looking for a suitable summability
condition on the weak derivatives
x
f, rather than a qualied continuity assumption as, for
instance, Holder continuity in the x variable, that, as we will see in the following, it is not a
sucient condition for the lower semicontinuity (see Theorem 4.13).
Always referring to Theorem 4.1, it is worth noting that one of the main advantages of (c) (and
its generalizations) with respect to (a) and (b) is that this condition allows also to treat integrands
4
With xf(x, s, ) we mean the (in case weak) gradient of f(, s, ). As already said, an analogous denition
holds for
f(x, s, ) too.
4.3. PROOF OF THE MAIN THEOREM 43
f that are constant in the gradient variable for some pair (x, s). The common root of (a) and
(b) is just this one: they are hypotheses assumed to exclude the cases in which f(x, s, ) can be
constant on straight lines, that is (a) and (b) assure that f(x, s, ) satises what we called the
n.c.s.l. property in the gradient variable, condition that we know to be equivalent to the notion of
demi-coercivity (see Theorem 2.34).
Theorem 4.8, that of course generalizes Theorems 4.1 and 4.7, is the main result of this chapter
and it follows by the researches carried on by Gori, Maggi and Marcellini (see [54] Theorem 1.2)
and Gori and Maggi (see [53] Theorem 6): here the remarks just made take the form of a precise
statement.
(1) for every (s, ), f(, s, ) W
1,1
loc
(), and, for every open set
H K R R
n
,
HK
such that, for every (s, ) H K,
_
[
x
f(x, s, )[dx L;
(2) for every (x
0
, s
0
) R, it is that either, for every R
n
,
f(x
0
, s
0
, ) inf f(x, s, ) : (x, s, ) R R
n
,
or there exists > 0 such that, for every (x, s) B(x
0
, ) B(s
0
, ),
f(x, s, ) is demi-coercive on R
n
.
1,1
loc
1
loc
() convergence.
We point out that Theorem 4.8(1) is a less general version of Theorem 1.2 proved in [54]: we
prefer to present here this simplied result since, assuming f continuous instead of Caratheodory
and locally bounded function, many technical diculties can be avoided and the fundamental ideas
of the proof can be easier understood. Clearly we refer to the original paper [54] for further details.
Finally, the following simple and useful corollary holds.
Corollary 4.9. Let f be a continuous function satisfying (4.1) and one of the following conditions:
(1)
x
f exists and is continuous,
(2) for every (x, s) R, f(x, s, ) is demi-coercive on R
n
.
1,1
loc
1
loc
() convergence.
4.3 Proof of the main theorem
We start presenting a simple proposition that is useful to treat lower semicontinuity problems for
integral functionals. We state it in a very particular form that can be easily generalized to dierent
contests.
Proposition 4.10. Let R
n
be an open set and f : R R
n
[0, ] be a Borel function.
Let us consider, for every B B(), u W
1,1
loc
(), the functional
I(u, B) =
_
B
f(x, u(x), u(x))dx,
and suppose that, for every B open ball, I(, B) is l.s.c. on W
1,1
(B) with respect to the
L
1
(B) convergence. Then I(, ) is l.s.c. on W
1,1
loc
1
loc
() convergence.
Proof. By a simple modication of Corollary 2 page 28 in [40] we can nd a sequence B
i
i=1
of
disjoint open balls well contained in , such that
L
n
_

_
i=1
B
i
_
= 0.
Let us consider now u
h
, u W
1,1
loc
(), u
h
u in L
1
loc
(). Clearly, for every i N, we have also
u
h
, u W
1,1
(B
i
), u
h
u in L
1
(B
i
), and then, xed k N,
I
_
u,
k
_
i=1
B
i
_
=
k
i=1
I(u, B
i
)
k
i=1
liminf
h
I(u
h
, B
i
)
liminf
h
k
i=1
I(u
h
, B
i
) liminf
h
i=1
I(u
h
, B
i
) = liminf
h
I(u
h
, ),
where we have used the fact that f is non negative. Letting now k the lower semicontinuity
inequality is achieved.
In particular this proposition shows that, without loss of generality, in the following it will be
always allowed to consider bounded with Lipschitz boundary and prove the lower semicontinuity
of I on W
1,1
1
() convergence.
4.3.1 The condition on the geometric variable
The proof of Theorem 4.8(1) is structured as follows. We rst prove a lower semicontinuity result,
namely Lemma 4.11, under certain technical hypotheses: in particular we show the validity of a
chain rule involving Sobolev functions (see equation (4.18)) that represents the core of the lemma.
Subsequently, combining Lemma 4.11 and Theorem 3.1, we achieve the proof of Theorem 4.8(1).
As already stated, the proofs provided here are a slightly dierent and simplied version of the
original ones presented in [54].
Before proving the theorem, in order to complete the description of the lower semicontinuity
results involving conditions on the geometric variable, we have to remember that recently Fusco,
Giannetti and Verde [46], De Cicco, Fusco and Verde [33] and De Cicco and Leoni [34] proposed
several generalizations of Theorem 4.8(1) (or, better, of the more general Theorem 1.2 in [54]).
In [46] and [33] the authors proved in particular that the same set of hypotheses on f given in
Theorem 4.8(1) is still sucient to prove also the lower semicontinuity, with respect to the L
1
loc
()
convergence, of the standard integral functional, proposed by Dal Maso [30], that extends I on
BV
loc
(). In [34] instead, by means of a very sophisticated chain rule that is the main result of
the paper, more general conditions are found in order to obtain the lower semicontinuity of I on
W
1,1
loc
().
However, it is worth noting that all these results, as Theorem 4.7 and of course Theorem 4.8(1)
too, are proved by using, as a crucial step, the possibility to approximate f with the ane functions
built up by means of Theorem 3.1.
This fact underlines the power of Theorem 3.1 in studying the lower semicontinuity with re-
spect to the L
1
loc
() convergence of functionals in which the integrand f satises some regularity
conditions on the geometric variable.
Let us prove now the rst step of the proof of Theorem 4.8(1) (compare it with Lemma 4.1 in
[54]).
Lemma 4.11. Let R
n
be a bounded open set with Lipschitz boundary and f : R R
n
[0, ) be a continuous function satisfying (4.1) and such that:

(i) there exists an open set
H R such that, for every (x, s, ) (
)(RH)R
n
,
f(x, s, ) = 0,
(ii)
f : R R
n
R
n
exists and is continuous. Moreover, for every (s, ) R R
n
,
f(, s, ) W
1,1
(, R
n
) and, for every K R
n
, there exists L = L
K
such that, for
every (s, ) R K,
n
i=1
_
f
x
i
(x, s, )
dx L, (4.9)
(iii) there exists a constant M such that, for every (x, s, ) R R
n
,
[
f(x, s, )[ M, (4.10)
and, for every (x, s) R,
1
,
2
R
n
,
[
f(x, s,
1
)
f(x, s,
2
)[ M [
1

2
[ . (4.11)
1,1
1
() convergence.
Proof. Let u
h
, u W
1,1
() such that u
h
u in L
1
(). We will prove that
liminf
h
I(u
h
, ) I(u, ). (4.12)
Without loss of generality, we can assume that
liminf
h
I(u
h
, ) = lim
h
I(u
h
, ) < ,
and that u
h
converges almost everywhere to u in . Moreover, by (4.10) and (4.11), there exists
a constant M
such that, for every (x, s, ) R R

n
,
[f(x, s, )[ M
(1 +[[),
thus, in particular, we have I(u, ) < .
Since u W
1,1
() and since is supposed Lipschitz, xed > 0, by Theorem 2.10, there
exists v
A() such that

_
[u(x) v
(x)[dx , (4.13)
and by Fatous Lemma and the niteness of I(u, ), we can also choose v
such that
_
f(x, u(x), v
(x))dx
_
f(x, u(x), u(x))dx . (4.14)

Since v
A(), we can nd
j
N
j=1
disjoint open subsets of such that L
n
_

N
j=1

j
_
= 0
and
j
N
j=1
R
n
such that v
(x) =
j
when x
j
.
Now let us take a sequence of non negative functions
k=1
C
c
() such that, for every
k N, j 1, . . . , N,
k
C
c
(
j
) and, for L
n
-a.e. x ,
k
(x) 1. By Beppo Levis Theorem,
there exists an index k
such that, for every k k
, we have
_
k
(x)f(x, u(x), v
(x))dx
_
k
(x)f(x, u(x), u(x))dx 2. (4.15)
We write the dierence of the integrands in (4.12) in this way
f(x, u
h
(x), u
h
(x)) f(x, u(x), u(x)) = f(x, u
h
(x), u
h
(x)) f(x, u
h
(x), v
(x)) (4.16)
+f(x, u
h
(x), v
(x)) f(x, u(x), v
(x)) +f(x, u(x), v
(x)) f(x, u(x), u(x)).

By the convexity of f(x, s, ) with respect to we have
f(x, u
h
(x), u
h
(x)) f(x, u
h
(x), v
(x))
f(x, u
h
(x), v
(x)), u
h
(x) v
(x))
=
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
+
f(x, u(x), v
(x)), u(x)v
(x))+
f(x, u(x), v
(x))
f(x, u
h
(x), v
(x)), v
(x)).
Using the inequality just found together with (4.16), multiplying for
k
and integrating over , we
obtain
_
k
(x)
_
f(x, u
h
(x), u
h
(x)) f(x, u(x), u(x))
_
dx
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dx
+
_
k
(x)
f(x, u(x), v
(x)), u(x) v
(x))dx
+
_
k
(x)
f(x, u(x), v
(x))
f(x, u
h
(x), v
(x)), v
(x))dx
+
_
k
(x)
_
f(x, u
h
(x), v
(x)) f(x, u(x), v
(x))
_
dx
+
_
k
(x)
_
f(x, u(x), v
(x)) f(x, u(x), u(x))

_
dx.
We remember that, by (4.10), for every (x, s) R and for every v
,
[
f(x, s, v
(x))[ M;
then, by (4.13), we have
_
k
(x)
f(x, u(x), v
(x)), u(x) v
(x))dx M
_
[u(x) v
(x)[ dx M.
Moreover, being (x, s)
k
(x)f(x, s, v
(x)) and (x, s)
k
(x)
f(x, s, v
(x)) continuous and

bounded functions, by the Lebesgues dominated convergence theorem we have,
lim
h
_
k
(x)
_
f(x, u
h
, v
) f(x, u, v
)
_
dx = 0,
lim
h
_
k
(x)
f(x, u, v
f(x, u
h
, v
), v
)dx = 0,
and in conclusion, by means of (4.15), we obtain that, for every > 0 and for every v
, k k
,
liminf
h
_
k
(x)
_
f(x, u
h
(x), u
h
(x)) f(x, u(x), u(x))
_
dx
liminf
h
_
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dxM2.
Then we claim that the proof is complete if we show that, for every xed v
and k k
, we have
lim
h
_
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dx = 0. (4.17)
Indeed, since 0
k
(x) 1, we shall have that, for every k k
,
liminf
h
_
f(x, u
h
(x), u
h
(x))dx
_
k
(x)f(x, u(x), u(x))dx M 2,
and letting k , by Beppo Levis Theorem, we obtain
liminf
h
_
f(x, u
h
(x), u
h
(x))dx
_
f(x, u(x), u(x))dx M 2.

Now the dependence from v
is vanished so that letting 0 we gain (4.12).

Thus it remains to prove (4.17): we stress that it suces to prove this relation when v
and k
are xed. In order to achieve this we prove that
5
_
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dx
=
_
i=1
_
_
u
h
(x)
u(x)
x
i
_
k
(x)
f(x, s, v
(x))
_
(i)
ds
_
dx. (4.18)
In order to prove this, xed j 1, . . . , N, let us consider the continuous function
g
j
(x, s) =
k
(x)
f(x, s,
j
) :
j
R R
n
,
with spt(g
j
)
j
H, that we can considered dened on the whole space R
n
R (extending g
j
to zero out of
j
H). For > 0 we dene
6
g
j,
(x, s) =
_
B(x,)
k
(x y)g
j
(y, s)dy : R
n
R R
n
.
Clearly, for every > 0, g
j,
C
c
(R
n
R, R
n
) and, if <
j
small enough, then spt(g
j,
)
j
H
too. Moreover, by the properties of convolutions we have also that, for every s R, g
j,
(, s)
C
c
(R
n
, R
n
).
We claim now that there exists C
j,
> 0 such that, for every x, y R
n
, s R,
[g
j,
(x, s) g
j,
(y, s)[ C
j,
[x y[. (4.19)
Indeed, for every i 1, . . . , n, (x, s) R
n
R, we have
g
j,
x
i
(x, s)
_
B(x,)
k
(x y)
g
j
x
i
(y, s)dy

n
_
R
n
g
j
x
i
(x, s)
dx

n
_
sup
x
[
k
(x)[
__
R
n
[
f(x, s,
j
)[ dx +
n
_
R
n
f
x
i
(x, s,
j
)
dx

n
M
_
sup
x
[
k
(x)[
_
L
n
(
j
) +
n
L
Kj
=
n
L
j
, (4.20)
where we have used the properties (i), (4.9) with constant L
K
j
where K
j
=
j
and (4.10): this
obviously implies (4.19).
It is worth noting that by the previous inequality we can deduce that, for every s R,
_
R
n
g
j
x
i
(x, s)
dx L
j
, (4.21)
5
Till the end of the proof with the notation
(i)
we will denote the i-th component of the vector R
n
.
6
We refer to the notations for the convolutions introduced in Chapter 2.
and then, by means of the properties of the convolutions with respect to the L
1
norm, for every
> 0, s R, we have also
_
R
n
g
j,
x
i
(x, s)
dx L
j
. (4.22)
The validity of (4.19) implies, by standard arguments
7
, for every <
j
, u, v W
1,1
(
j
), the
equality
_
j
g
j,
(x, v(x)), v(x)) g
j,
(x, u(x)), u(x))dx =
_
j
_
n
i=1
_
v(x)
u(x)
g
(i)
j,
x
i
(x, s)ds
_
dx.
Now we prove that, passing to the limit as 0 in the previous formula we can obtain the same
relation for g
j
too. In order to prove the equality
lim
0
_
j
g
j,
(x, v(x)), v(x))dx =
_
j
g
j
(x, v(x)), v(x))dx,
since the boundedness of g and the summability of v, we only need to prove that
lim
0
g
j,
(x, v(x)) = g
j
(x, v(x)), (4.23)
for L
n
-a.e. x
j
: but this easily follows by the property of the convolutions together with the
continuity of g
j
. Thus it remains to prove that, for every i 1, . . . , n,
lim
0
_
j
_
_
v(x)
u(x)
g
(i)
j,
x
i
(x, s)ds
_
dx =
_
j
_
_
v(x)
u(x)
g
(i)
j
x
i
(x, s)ds
_
dx.
Indeed
_
_
v(x)
u(x)
g
(i)
j,
x
i
(x, s)ds
_
v(x)
u(x)
g
(i)
j
x
i
(x, s)ds
dx
_
j
_
_
H
g
(i)
j,
x
i
(x, s)
g
(i)
j
x
i
(x, s)
ds
_
dx
=
_
H
_
_
j
g
(i)
j,
x
i
(x, s)
g
(i)
j
x
i
(x, s)
dx
_
ds,
that goes to zero as 0 since, by (ii) and the standard properties of the convolutions, we have,
for every xed s H,
lim
0
_
j
g
(i)
j,
x
i
(x, s)
g
(i)
j
x
i
(x, s)
dx = 0,
and moreover, using (4.21) and (4.22), it is, for every > 0, s H,
_
j
g
(i)
j,
x
i
(x, s)
g
(i)
j
x
i
(x, s)
dx
_
j
g
(i)
j,
x
i
(x, s)
dx +
_
j
g
(i)
j
x
i
(x, s)
dx 2L
j
,
that is, it is bounded on H uniformly on : then we are in the position to apply the Lebesgues
dominated convergence Theorem. We conclude that, for every j 1, . . . , N, u, v W
1,1
(
j
),
also
_
j
g
j
(x, v(x)), v(x)) g
j
(x, u(x)), u(x))dx =
_
j
_
n
i=1
_
v(x)
u(x)
g
(i)
j
x
i
(x, s)ds
_
dx.
7
The following relation can be proved considering at rst u, v C
(
j
) W
1,1
(
j
) and then arguing by
approximation.
Hence (4.18) immediately follows since
_
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dx
=
N
j=1
_
j
g
j
(x, v(x)), v(x)) g
j
(x, u(x)), u(x))dx =
N
j=1
_
j
_
n
i=1
_
v(x)
u(x)
g
(i)
j
x
i
(x, s)ds
_
dx
=
_
i=1
_
_
u
h
(x)
u(x)
x
i
_
k
(x)
f(x, s, v
(x))
_
(i)
ds
_
dx.
Now by (4.18) we have can prove (4.17): indeed
k
(x)
_
f(x, u
h
(x), v
(x)), u
h
(x))
f(x, u(x), v
(x)), u(x))
_
dx
i=1
N
j=1
_
_
u
h
(x)
u(x)
k
x
i
(x)
f(x, s,
j
)[ ds
dx +
n
i=1
N
j=1
_
_
u
h
(x)
u(x)
f
x
i
(x, s,
j
)
ds
dx.
Using (4.10), we have
n
i=1
N
j=1
_
_
u
h
(x)
u(x)
k
x
i
(x)
f(x, s,
j
)[ ds
dx nM
_
sup
x
[
k
(x)[
__
[u
h
(x) u(x)[ dx,
which goes to zero as h . On the other hand,
n
i=1
N
j=1
_
j
_
u
h
(x)
u(x)
f
x
i
(x, s,
j
)
ds
dx
n
i=1
N
j=1
_
D
j,h
f
x
i
(x, s,
j
)
dxds,
where
D
j,h
=
_
(x, s)
j
H : minu
h
(x), u(x) s maxu
h
(x), u(x)
_
.
But now
lim
h
[D
j,h
[ lim
h
_
j
[u
h
(x) u(x)[ dx = 0,
and, if we dene K =
j
N
j=1
and consider L
K
as in hypothesis (4.9) we obtain, for every j
1, . . . , N, i 1, . . . , n,
_
j
R
f
x
i
(x, s,
j
)
dxds =
_
H
_
_
f
x
i
(x, s,
j
)
dx
_
ds [H[ L
K
.
Hence, for every j 1, . . . , N, i 1, . . . , n,
lim
h
_
D
j,h
f
x
i
(x, s,
j
)
dxds = 0,
from which we conclude that (4.17) holds and this ends the proof.
Now we are ready for the proof of Theorem 4.8(1). We start modifying the integrand f in order
to satisfy the hypotheses of Theorem 3.1, then we manipulate De Giorgis approximating functions
in order to apply on them Lemma 4.11.
Proof of Theorem 4.8(1). By Proposition 4.10 we can suppose bounded and study the lower
semicontinuity on W
1,1
1
() convergence. Let
i
i=1
be an increasing
sequence of smooth functions with compact support in R, converging point-wise to 1 in R.
Since
i
(x, s)f(x, s, ) is an increasing sequence of functions which point-wise converges to f(x, s, ),
by a standard argument, it is sucient to prove the stated lower semicontinuity assuming directly
that there exists
H R such that, for every (x, s, ) (
)(RH)R
n
, f(x, s, ) = 0.
In this way, by the hypotheses of Theorem 4.8(1), we have that, for every (s, ) R R
n
,
f(, s, ) W
1,1
(), (4.24)
and, for every K R
n
K
such that, for every (s, ) R K,
_
[
x
f(x, s, )[ dx L,
or, equivalently,
n
i=1
_
f
x
i
(x, s, )
dx L. (4.25)
Since f has compact support in (x, s), it is possible to approximate it with an increasing sequence
f
j
(x, s, )
j=1
as in Theorem 3.1. We would like to apply Lemma 4.11 to such functions, but this
is not possible. Thus we denote by k a convolution kernel in R
n
and we consider the functions
g
j
(x, s, ) =
_
R
n
k
j
()f
j
(x, s, )d,
where, setting M
j
the constant related to f
j
in the statement of Theorem 3.1, we choose
j
=
(jM
j
)
1
. By the Lipschitz continuity (3.4) of f
j
with respect to R
n
, we have that, for every
(x, s, ) R R
n
,
[g
j
(x, s, ) f
j
(x, s, )[
_
R
n
[f
j
(x, s, ) f
j
(x, s, )[ k
j
()d
M
j
_
B(0,
j
)
[[ k
j
()d
1
j
,
so that
f
j
(x, s, )
2
j
g
j
(x, s, )
1
j
f
j
(x, s, ) f(x, s, ). (4.26)
By the Beppo Levis Theorem, for every u W
1,1
() we have
lim
j
_
f
j
(x, u(x), u(x)) dx =
_
f (x, u(x), u(x)) dx,

and thus, if we consider the sequence of integrals
I
j
(u, ) =
_
_
g
j
(x, u(x), u(x))
1
j
_
dx, (4.27)
by (4.26) we obtain that I
j
(u, ) converges, as j , to the main integral I(u, ), and that, at
the same time, I
j
(u, ) I(u, ) for every j N. Therefore
I(u, ) = sup
jN
I
j
(u, ),
and then, since the supremum of a family of l.s.c. functionals is l.s.c. too, it suces to show that
every I
j
(u, ) is lower semicontinuous on W
1,1
1
() convergence: since
is bounded, this is the same to prove this property for the functional
_
g
j
(x, u(x), u(x)) dx.
In order to achieve this we shall invoke Lemma 4.11: let us verify that, for every j N, g
j
satises
its hypotheses.
Clearly every g
j
is non negative, continuous, convex in the gradient variable and has compact
support in (x, s), being a convolution in of f
j
f. Furthermore
g
j
exists and is continuous.
Next we verify (4.10) and (4.11). By the Lipschitz continuity (3.4) of f
j
, we have, for every
(x, s) R, R
n
[g
j
(x, s,
1
) g
j
(x, s,
2
)[
_
R
n
[f
j
(x, s,
1
) f
j
(x, s,
2
)[ k
j
()d M
j
[
1

2
[ ,
that implies [
g
j
(x, s, )[ M
j
. By the denition of convolution
[
g
j
(x, s,
1
)
g
j
(x, s,
2
)[
_
R
n
[f
j
(x, s,
1
) f
j
(x, s,
2
)[
k
j
()
d M
j
T
j
[
1

2
[ ,
where
T
j
=
_
R
n
j
()
d.
Therefore assumptions (4.10) and (4.11) are satised with M
j
= maxM
j
, M
j
T
j
.
It remains to prove the weak derivability of g
j
in the x variable and to verify (4.9). We shall
start examining such properties for the coecients a
q,h
in Theorem 3.1. From (3.2) we have that
a
q,h
is continuous. For every C
c
(), by (4.24) we have
_
a
q,h
(x, s)

x
i
(x)dx =
_
_
R
n
f(x, s, )
h
()d
_

x
i
(x)dx
=
_
R
n
__
f(x, s, )

x
i
(x)dx
_
h
()d =
_
R
n
__
f
x
i
(x, s, )(x)dx
_
h
()d
=
_
__
R
n
f
x
i
(x, s, )
h
()d
_
(x)dx;
furthermore, thanks to the main assumption (4.25), we obtain
_
a
q,h
x
i
(x, s)
dx =
_
_
R
n
f
x
i
(x, s, )
h
()d
dx
_
sup
R
n
[
q
()[
_
_
spt()
__
f
x
i
(x, s, )
dx
_
d
_
sup
R
n
[
q
()[
_
L
n
(spt(
q
)) L
spt(
q
)
.
Hence a
q,h
(, s) W
1,1
() and there exists a constant R
q,h
such that, for every s R,
_
a
q,h
x
i
(x, s)
dx R
q,h
.
The same analysis carried on for a
q,h
applies unchanged to a
q,0
, so that a
q,0
(x, s) is a continuous
function such that, for every s R, a
q,0
(, s) W
1,1
(), and there exists a constant R
q,0
such that
_
a
q,0
x
i
(x, s)
dx R
q,0
.
In particular a
q,0
(, s) +
n
h=1
a
q,h
(, s)
h
W
1,1
(), from which,
f
j
(, s, ) = max
q{1,...,j}
_
0, a
q,0
(, s) +
n
h=1
a
q,h
(, s)
h
_
W
1,1
(),
and in the end we have also that
g
j
(, s, ) W
1,1
(, R
n
), with
g
j
x
i
(x, s, ) =
_
R
n
f
j
x
i
(x, s, )k
j
()d.
Since,
_
f
j
x
i
(x, s, )
dx
j
q=1
_
_
x
i
_
a
q,0
(x, s) +
n
h=1
a
q,h
(x, s)
h
_
dx
_
q=1
_
a
q,0
x
i
(x, s)
dx +[[
j
q=1
n
h=1
_
a
q,h
x
i
(x, s)
dx S
j
(1 +[[) ,
for a suitable constant S
j
(depending on R
q,0
, R
q,h
, j), we deduce
_
g
j
x
i
(x, s, )
dx
_
__
R
n
j
()
f
j
x
i
(x, s, )
d
_
dx
_
B
j
(0)
k
j
()
__
f
j
x
i
(x, s, )
dx
_
d
_
B
j
(0)
S
j
(1 +[ [)
j
()
d S
j
T
j
(1 +[[ +
j
).
Hence, xed K R
n
, if L
j,K
= nS
j
T
j
(1 + T +
j
), (where T is equal to the radius of a ball
centered at the origin and containing K), we have, for every (s, ) R K,
n
i=1
_
g
j
x
i
(x, s, )
dx L
j,K
,
that is g
j
satises the condition (4.9) of Lemma 4.11 too. Thus Lemma 4.11 applies to every g
j
,
providing the lower semicontinuity of each I
j
. By the previous arguments this ends the proof of
Theorem 4.8(1).
4.3.2 The condition on the gradient variable
In this section we use Theorem 3.5, together with the blow up method by Fonseca and M uller [44]
and the original argument of Serrin [69], in order to prove Theorem 4.8(2): as already said the
following proof can be found in [53] Theorem 6.
It is worth noting that Theorem 4.8(2) has been recently generalized by Maggi [59] who proved
in particular that, under the same set of hypotheses of Theorem 4.8(2), the standard extension
of the integral functional I on BV
loc
() proposed by Dal Maso [30] is lower semicontinuous on
BV
loc
1
loc
() convergence.
Proof of Theorem 4.8(2). By Proposition 4.10 we can assume bounded and then also f 2. By
the usual blow up method (see for example the rst part of the proof of Theorem 1.1 of Fonseca
and Leoni [43]), the problem is reduced to the following one: given (x
0
, s
0
,
0
) RR
n
,
h
0
and w
h
W
1,1
(Q
n
) such that w
h
w
0
in L
1
(Q
n
), where Q
n
= (0, 1)
n
and w
0
(y) =
0
, y), we
have to prove
lim
h
_
Q
n
f(x
0
+
h
y, s
0
+
h
w
h
(y), w
h
(y))dy f(x
0
, s
0
,
0
). (4.28)
If it is f(x
0
, s
0
, ) inff(x, s, ) : (x, s, ) R R
n
, then (4.28) is trivially veried.
Otherwise there exists > 0 such that f satises the hypotheses of Theorem 3.5 with =
B(x
0
, ) B(s
0
, ): clearly we can suppose, for every h N,
h
< and
ess sup
yQ
n
[w
0
(y)[

h
.
In particular we have, for every (x, s, ) B(x
0
, ) B(s
0
, ) R
n
,
f(x, s, ) = sup
jN
f
j
(x, s, ), (4.29)
where every f
j
is continuous, with values in [0, ), and such that f
j
(x, s, ) is strictly convex
on R
n
. We know also that, taking small enough, we can nd C
j
(x, s, ) B(x
0
, ) B(s
0
, ) R
n
,
f
j
(x, s, ) C
j
(1 +[[). (4.30)
Let us x now > 0 and
(0, ). We can apply Theorem 3.2 to each f

j
, with
= B(x
0
,
)
B(s
0
,
), to nd

f
j
: B(x
0
, ) B(s
0
, ) R
n
[0, ) continuous, convex in the gradient variable
and satisfying (i), (ii), (iii), (iv) and (v) of Theorem 3.2 with constant M
j
.
We need to truncate the functions w
h
in order apply the described approximation. Let us
consider the sets
E
h
=
_
y Q
n
: [w
h
(y)[

h
_
, E
+
h
=
_
y Q
n
: w
h
(y) >

h
_
, E
h
=
_
y Q
n
: w
h
(y) <

h
_
,
and dene the sequence
v
h
(y) =
_
_
_
w
h
(y), y E
h
,
h
, y E
+
h
,
h
, y E
h
.
(4.31)
It is v
h
W
1,1
(Q
n
), v
h
w
0
in L
1
(Q
n
) and lim
h
L
n
(Q
n
E
h
) = 0.
By the denition of E
h
and v
h
, by (4.29), (4.30) and property (ii) of

f
j
, we have
_
Q
n
f(x
0
+
h
y, s
0
+
h
w
h
(y), w
h
(y))dy
_
E
h
f
j
(x
0
+
h
y, s
0
+
h
w
h
(y), w
h
(y))dy
=
_
E
h
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))dy
=
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))dy
_
Q
n
\E
h
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), 0)dy
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))dy C
j
L
n
(Q
n
E
h
)
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))dy C
j
L
n
(Q
n
E
h
) . (4.32)
Let us consider the function g
j,h
: Q
n
R R
n
dened, for (y, s) Q
n
R, as
g
j,h
(y, s) =

f
j
(x
0
+
h
y, s
0
+
h
s,
0
).
By Theorem 3.2 it is a continuous and bounded function with compact support, and it satises an
uniform Lipschitz condition, that is, for every y, z Q
n
, s R,
[g
j,h
(y, s) g
j,h
(z, s)[ M
j
h
(1 +[
0
[)[y z[.
Now, for every y Q
n
, the convexity of

f
j
in the variable implies
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))

f
j
(x
0
+
h
y, s
0
+
h
v
h
(y),
0
)
+g
j,h
(y, w
0
(y)) g
j,h
(y, v
h
(y)),
0
) +g
j,h
(y, v
h
(y)), v
h
(y)) g
j,h
(y, w
0
(y)),
0
),
and then
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y), v
h
(y))dy
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y),
0
)dy
+
_
Q
n
g
j,h
(y, w
0
(y)) g
j,h
(y, v
h
(y)),
0
)dy +
_
Q
n
g
j,h
(y, v
h
(y)), v
h
(y)) g
j,h
(y, w
0
(y)),
0
)dy.
Let us evaluate the limit as h of the right hand side of the previous inequality. Since, for
L
n
-a.e. x Q
n
, v
h
(x) w
0
(x) (and then
h
v
h
(x) 0), by Fatous Lemma we have
liminf
h
_
Q
n
f
j
(x
0
+
h
y, s
0
+
h
v
h
(y),
0
)dy

f
j
(x
0
, s
0
,
0
),
while, by the continuity and boundedness of g
j,h
, it follows
lim
h
_
Q
n
g
j,h
(y, w
0
(y)) g
j,h
(y, v
h
(y)),
0
)dy = 0.
Let us show now that also
lim
h
_
Q
n
g
j,h
(y, v
h
(y)), v
h
(y)) g
j,h
(y, w
0
(y)),
0
)dy = 0.
In order to see this let us note that, by a mollication argument similar to the one used to prove
(4.18) in the proof of Theorem 4.8(1), the following formula holds
8
_
Q
n
g
j,h
(y, v
h
(y)), v
h
(y)) g
j,h
(y, w
0
(y)),
0
)dy =
_
Q
n
_
n
i=1
_
v
h
(y)
w
0
(y)
g
(i)
j,h
y
i
(y, s)ds
_
dy,
from which follows

_
Q
n
g
j,h
(y, v
h
(y)), v
h
) g
j,h
(y, w
0
(y)),
0
)dy
i=1
_
Q
n
_
v
h
(y)
w
0
(y)
g
(i)
j,h
y
i
(y, s)
ds
dy nM
j
h
(1 +[
0
[)
_
Q
n
[w
0
v
h
[dy,
that goes to zero as h .
If now is so small that [
0
[ 1/, property (iii) of

f
j
implies

f
j
(x
0
, s
0
,
0
) f
j
(x
0
, s
0
,
0
) ,
and hence, by (4.32), it is that
liminf
h
_
Q
n
f(x
0
+
h
y, s
0
+
h
w
h
(y), w
h
(y))dy f
j
(x
0
, s
0
,
0
) 2.
Thus, it suces to let tend to zero and then take the supremum on j N to conclude the
proof.
Remark 4.12. Note that the same proof allows also to consider the limit function u BV
loc
(),
and then to give a direct lower bound of the relaxed functional R[L
1
loc
](I) on BV
loc
().
8
In the following (g
j,h
)
(i)
denotes the i-th component of g
j,h
.
4.4. A COUNTEREXAMPLE TO THE LOWER SEMICONTINUITY 55
4.4 A counterexample to the lower semicontinuity
As explained in the rst section of this chapter, Aronszajns example shows that conditions (4.1),
together with the continuity of the integrand f, are not sucient to guarantee the lower semicon-
tinuity on W
1,1
1
() convergence of the integral functional I. However,
Theorem 4.7 shows that if f satises (4.8) too, that is f is locally Lipschitz continuous in x
uniformly for (x, ) belonging to a compact set, then the lower semicontinuity is achieved.
Theorem 4.8(1), in which suitable summability conditions on
x
f are required, improves con-
dition (4.8): nevertheless it is interesting to understand what we can expect supposing, instead of
the Lipschitz continuity of f, its Holder continuity with respect to x, that is, there exists (0, 1)
such that, for every compact set K R R
n
, we can nd a constant L = L
K
such that, for
every (x
1
, s, ), (x
2
, s, ) K,
[f(x
1
, s, ) f(x
2
, s, )[ L[x
1
x
2
[
. (4.33)
By modifying a one dimensional (that is n=1) version of Aronszajns Example proposed by Dal
Maso
9
, Gori, Maggi and Marcellini solved this problem showing that, for every (0, 1), there
exists a one dimensional integral functional for which the lower semicontinuity fails and such that
its integrand satises (4.1) and (4.33) (see [54] Example 4.2, see also [55] Theorem 2.1 in which
this fact was proved using multiple integrals, that is n 2.).
Here we present a simplied version just of Example 4.2 in [54].
Theorem 4.13. Let > 0 and = (, 1 + ). For every (0, 1), there exist a sequence
u
h
h=1
W
1,
(), such that u
h
0 in L
(), and a function

10
b : R R (both of them
depending on ) such that:
(i) b is bounded and uniformly continuous in R;
(ii) there exists a constant L such that, for every x
1
, x
2
and s R,
[b(x
1
, s) b(x
2
, s)[ L[x
1
x
2
[
; (4.34)
(iii) setting, for every (x, s, ) R R,
f(x, s, ) = [b(x, s) 1[ ,
we have that f(x, s, ) is continuous, satises (4.1) and the Holder continuity property (4.33)
and
lim
h
I(u
h
, ) = lim
h
_
f(x, u
h
(x), u
h
(x))dx <
_
f(x, 0, 0)dx = I(0, ). (4.35)

In particular the functional I is not lower semicontinuous on W
1,
() with respect to the
L
() convergence.
Proof. Let us x (0, 1) and choose k N such that k(1 ) 2 (then k > 2).
9
Note that the original example by Aronszajn is related to a double integral, that is, n = 2, and Aronszajns
integrand f(x, ) does not explicitly depends on s. A one-dimensional version of Aronszajns example was known
to Dal Maso, who gave us some handwritten notes on the subject: however, in this case, the integrand f depends
on all the three variables x, s and .
10
In this section denotes cl().
First of all let us build up the sequence u
h
. Let us x h 2N =
_
2i : i N
_
with h h
0
,
where
11
h
0

1
2
1
4
, and dene u
h
on in this way: for every r
_
0, 1, ..., h
k
1
_
,
u
h
(x) =
_
_
4h
k
h(h+1)
x +
1
h+1

4r
h(h+1)
if x
_
r
h
k
,
r
h
k
+
1
4h
k
,
1
h
if x
_
r
h
k
+
1
4h
k
,
r
h
k
+
2
4h
k
4h
k
h(h+1)
x +
1
h
+
4r+2
h(h+1)
if x
_
r
h
k
+
2
4h
k
,
r
h
k
+
3
4h
k
,
1
h+1
if x
_
r
h
k
+
3
4h
k
,
r+1
h
k
,
and, for every x (, 0] [1, 1 +), u
h
(x) =
1
h+1
. Trivially we have u
h
W
1,
(), u
h
0 in
L
() and, since we are considering h 2N, the graphs of the functions u

h
are disjoint. Moreover
an easy computation gives
lim
h
_
[u
h
(x)[ dx = lim
h
2h
k
h(h + 1)
= .
In order to dene the function b let us consider, for every h 2N, h h
0
and r
_
0, 1, ..., h
k
1
_
,
the subsets of R given by
A
r
h
=
__
x,
4h
k
h(h + 1)
x +
1
h + 1

4r
h(h + 1)
_
: x
_
r
h
k
,
r
h
k
+
1
4h
k
__
,
B
r
h
=
__
x,
4h
k
h(h + 1)
x +
1
h
+
4r + 2
h(h + 1)
_
: x
_
r
h
k
+
2
4h
k
,
r
h
k
+
3
4h
k
__
,
C
h
=
_
1
2h
+
1
2(h 1)
_
,
D
1
= , 1 + R, and D
2
=
_
(, 0]
_
1
2h
0
+
1
2(h
0
1)
, +
__
,
whose geometric meaning, with respect to the sequence u
h
: h 2N, h h
0
, is clear. Let us
consider now the closed set
S =
_
_
_
h,r
A
r
h
_
_
_
_
_
h,r
B
r
h
_
_
_
_
h
C
h
_
D
1
D
2
,
and dene the function b on S as
b (x, s) =
_
_
h(h+1)
4h
k
if (x, s) A
r
h
,
h(h+1)
4h
k
if (x, s) B
r
h
,
0 if (x, s)
_
h
C
h
_
D
1
D
2
.
Since lim
h
h(h+1)
4h
k
= 0, we have that b is continuous on S; moreover note that, on A
r
h
and on B
r
h
, b
assumes the value of the inverse of u
h
on the points in which the graph of u
h
intersects these sets.
Now we extend b on R. Let us consider s
0

h
_
1
h+1
,
1
h
_
and observe that S
_
s
0
_
contains only a nite number of points among which there are (, s
0
) and (1 +, s
0
): thus, in
11
This value is chosen only for technical reasons, in order to compute the value (4.36) in a simpler way.
4.4. A COUNTEREXAMPLE TO THE LOWER SEMICONTINUITY 57
the segment s
0
, we dene b by a linear interpolation between every pair of adjacent points
in which the value of b is known.
For every h 2N, h h
0
, the described extension allows us to build up b continuous in the
strip R
h
=
_
1
h+1
,
1
h
_
. We have, in particular, that b is dened in every set of the form
_
1
j
_
,
where j N and j h
0
, if h
0
is even, or j h
0
+ 1 if h
0
is odd.
Now we complete the extension of b setting, for every h 2N, h h
0
b (x, s) =
_
_
b
_
x,
1
h
_ __
1
h
s
_
2h(h + 1) + 1
if (x, s) E
h
,
b
_
x,
1
h+1
___
1
h+1
s
_
2(h + 1)(h + 2) + 1
_
if (x, s) F
h
,
where
E
h
=
_
1
h
,
1
2h
+
1
2(h 1)
_
and F
h
=
_
1
2(h + 2)
+
1
2(h + 1)
,
1
h + 1
_
.
Since this extension makes b continuous on R and since, for every (x, s) D
1
D
2
, b(x, s) = 0,
we have that b is uniformly continuous on R and then condition (i) of the theorem is veried.
We want now to prove that b satises condition (ii). Fixed h 2N, h h
0
we have that, by
construction, the value
M
h
= sup
_
[b (x, s) b (y, s)[
[x y[
: (x, s) , (y, s) R
h
E
h
F
h
; x ,= y
_
(4.36)
is in fact a maximum which is obtained by the pairs of points of the type
_
r
h
k
+
1
4h
k
,
1
h
_
,
_
r
h
k
+
2
4h
k
,
1
h
_
, or
_
r
h
k
+
3
4h
k
,
1
h + 1
_
,
_
r + 1
h
k
,
1
h + 1
_
, (4.37)
for every r
_
0, 1, ..., h
k
1
_
12
. Computing (4.36), using any of the previous pair of points, we
have
M
h
=
2h(h + 1)
4h
k
4
h
k
=
4
2

h(h + 1)
h
kk

4
2

2h
2
h
kk
4
h
2
h
kk
,
and since, by denition k (1 ) 2, we have
M
h
4
h
2
h
kk
4
that doesnt depend on h. Thus, since outside

h
(R
h
E
h
F
h
), that is in D
1
D
2
, we have
b 0 we conclude that surely b satises (4.34) for L = 4
.
Finally let us prove that the function f(x, s, ) = [b(x, s) 1[ satises condition (iii). Clearly
f is continuous and satises (4.1) and (4.33). Moreover
I (0, ) =
_
[b (x, 0) 0 1[ dx = 1 + 2
while, for every h 2N, h h
0
,
I (u
h
, ) =
_
[b (x, u
h
(x)) u
h
(x) 1[ dx =
_
_
h
k
j=1
2
1
4h
k
_
_
+ 2 =
1
2
+ 2.
Thus
lim
h
I (u
h
, ) =
1
2
+ 2 < 1 + 2 = I (u, )
and (4.35) is also veried.
12
For the pair of points
,
1
h+1
0,
1
h+1
and
1
1
4h
k
,
1
h+1
1 +,
1
h+1
, because of the choice of h

0
, the
value of the ratio in (4.36) is surely less then the one of the pairs in (4.37).
The theorem just proved provides also a useful information about Theorem 4.8(2). Indeed, it
is simple to verify that the function f of Theorem 4.13 is such that, for every (x
0
, s
0
) R,
either, for every R,
f(x
0
, s
0
, ) 1,
or there exists > 0 such that, for every (x, s) B(x
0
, ) B(s
0
, ),
f(x, s, ) is demi-coercive on R.
There is no contradiction with Theorem 4.8(2) because it is inf f = 0: however, this shows how
the hypotheses of Theorem 4.8(2) are close to be sharp.
We note also that, when n = 1, the demi-coercivity reduces to ask that
f(x, s, ) is not constant, (4.38)
for every xed pair (x, s). However, if in the hypotheses of Theorem 4.8(2) we replace the demi-
coercivity condition with (4.38) and assume n 2, then the lower semicontinuity will fail again.
This fact can be deduced by the original Aronszajns counterexample [64] in which a continuous
function : (0, 1)
2
S
1
is built up in such a way that the integral functional generated by
f(x, ) = [(x), )[, where (x, ) (0, 1)
2
R
2
, it is not lower semicontinuous (see also [30]
Example 4.1). Clearly f(x, ) satises (4.38) but fails to be demi-coercive.
4.5 Appendix: the vectorial case
In this section we deal with a function f satisfying the conditions
_
f : R
m
R
mn
[0, ],
for every (x, s) R
m
, f(x, s, ) is convex in R
mn
,
(4.39)
where is an open set of R
n
, n, m 1. As in the scalar case we want to nd conditions on f such
that, for every u
h
, u W
1,1
loc
(, R
m
) such that u
h
u in L
1
loc
(, R
m
),
I(u, ) liminf
h
I(u
h
, ). (4.40)
The lower semicontinuity results by Serrin [69] (see Theorems 4.1 and 4.2) were proved for the
scalar case m = 1: some eorts have been spent to understand the question of their validity in the
vectorial setting
13
, that is, when m 2.
4.5.1 Some counterexamples
Eisen [38] showed with an example that Theorem 4.1(c) is false in the vectorial case: obviously
the same example shows that the same holds for Theorems 4.7 and 4.8(1) and Corollary 4.9(1)
too. We propose here a simplied version of the example quoted above (see Gori and Maggi [52]
Example 2.1) that, as we will explain in the next chapter, it is meaningful even in the context of
the supremal functionals.
Example 4.14. Let f : R
2
R
2
[0, ) be a function dened, for every (s
1
, s
2
,
1
,
2
) = (s, )
R
2
R
2
, as
f(s
1
, s
2
,
1
,
2
) = (
1
s
2

2
s
1
1)
2
.
13
However, in studying the lower semicontinuity problems in the vectorial setting, the notion of convexity in
the gradient variable of f is replaced by the more general and natural one of quasiconvexity (see Morrey [63] and
Dacorogna [29]).
4.5. APPENDIX: THE VECTORIAL CASE 59
Clearly f is non negative, smooth and convex in the gradient variable = (
1
,
2
). Let us consider
the sequence of functions in C
((0, 1), R
2
) given, for every h N, by
u
h
(x) =
_
1
2h
sin(2h
2
x),
1
2h
cos(2h
2
x)
_
.
An easy computation gives u
h
0 in L
((0, 1), R
2
) as h . However, for every h N,
I(u
h
, (0, 1)) = 0, while I(0, (0, 1)) = 1 so that the lower semicontinuity fails. Note in particular
that, for every x (0, 1), f(u
h
(x), u
h
(x)) = 0.
Later Cern y and Mal y [25] proved also that Theorem 4.1(b) (and of course Theorem 4.8(2) and
Corollary 4.9(2) too) is false in the vectorial setting.
2
R
2
1
, s
2
,
1
,
2
) = (s, )
R
2
R
2
, as
f(s
1
, s
2
,
1
,
2
) =

2
2
+ exp(
1
)
exp((s
2
1
+s
2
2
)
2
)
.
Clearly f is non negative, smooth and convex in the gradient variable = (
1
,
2
). Let us consider
the sequence of functions in W
1,
((0, 1), R
2
) given, for every h N, by u
h
(x) = (u
h,1
(x), u
h,2
(x)),
where, for every x [0,
1
h
],
u
h,1
(x) =
_
_
h
4
x if x
_
0,
1
h
3
,
hcos
_
h
3
4
_
x
1
h
3
_
_
if x
_
1
h
3
,
5
h
3
,
h h
3
_
h
1
h
_ _
x
5
h
3
_
if x
_
5
h
3
,
6
h
3
,
h
2
h
2
6
_
1
h
x
_
if x
_
6
h
3
,
1
h
,
and
u
h,2
(x) =
_
_
0 if x
_
0,
1
h
3
,
hsin
_
h
3
4
_
x
1
h
3
_
_
if x
_
1
h
3
,
5
h
3
,
0 if x
_
5
h
3
,
1
h
,
and extended by periodicity on (0, 1). A computation gives u
h
0 in L
((0, 1), R
2
) as h
and
lim
h
I(u
h
, (0, 1)) =
1
e
.
Since I(0, (0, 1)) = 1 the lower semicontinuity fails.
Following Theorem 4.1, it remains only to understand the situation about condition (a). The
following example given by Cern y and Mal y [24] shows that, when f is lower semicontinuous (but
not continuous), even if it is coercive, the lower semicontinuity may fail. However, they are not
able to consider f continuous as in Theorem 4.1(a): as we will see this is due to the fact that in
these hypotheses a lower semicontinuity theorem can be proved (see Theorem 4.19).
2
R
2
1
, s
2
,
1
,
2
) = (s, )
R
2
R
2
, as
f(s
1
, s
2
,
1
,
2
) =
_
_
_
[
1
1[ +[
2
[ +
1
|s
2
|
+[s
1
[ +[s
2
[ if s
2
,= 0,
2[
1
1[ +[
2
[ +[s
1
[ +[s
2
[ if s
2
= 0.
Clearly f is non negative, lower semicontinuous and convex in the gradient variable = (
1
,
2
).
Moreover, for every (s
1
, s
2
,
1
,
2
) R
2
R
2
,
f(s
1
, s
2
,
1
,
2
) [s
1
[ +[s
2
[ +[
1
[ +[
2
[ 1,
that is f is coercive. Let us consider the sequence of functions in W
1,1
((0, 1), R
2
) given, for every
h N, by u
h
(x) = (u
h,1
(x), u
h,2
(x)), where, for every x [0,
1
h
],
u
h,1
(x) =
_
_
_
x if x
_
0,
1
h

1
h
4
,
_
h
3
1
_ _
1
h
x
_
if x
_
1
h

1
h
4
,
1
h
,
and
u
h,2
(x) =
_
_
0 if x
_
0,
1
h

1
h
4
,
_
h
4
_
x
1
h
+
1
h
4
_ _
1
h
x
_
if x
_
1
h

1
h
4
,
1
h
,
and extended by periodicity on (0, 1). We can see that u
h
0 in L
((0, 1), R
2
) as h and
lim
h
I(u
h
, (0, 1)) = 1.
Since I(0, (0, 1)) = 2 the lower semicontinuity fails.
4.5.2 Lower semicontinuity theorems
Examples 4.14 and 4.15 seem to indicate that, when the vectorial case is considered, we need a
coercivity assumption on the integrand in order to obtain the lower semicontinuity. Moreover, as
Example 4.16 says, even if this condition is veried, f must be more regular than lower semicon-
tinuous. Nevertheless under the investigation of the scalar case made in the previous sections, we
became aware of the fact that, in particular, two results without coercivity assumptions hold in the
vector-valued case too, once provided the dependence on the s variable is dropped (note that the
dependence on s is fundamental in the counterexamples proposed). The rst one is a vector-valued
version of Theorem 4.7 (see Gori, Maggi and Marcellini [54] Theorem 5.1).
Theorem 4.17. Let f be a continuous function satisfying (4.39) such that f = f(x, ) and, for
every open set
K R
mn
K
such that, for every
x
1
, x
2

, K,
[f(x
1
, s, ) f(x
2
, s, )[ L[x
1
x
2
[ . (4.41)
1,1
loc
(, R
m
) with respect to the L
1
loc
(, R
m
) convergence.
Proof. By Proposition 4.10, that can be applied also to the vectorial setting, we can suppose
bounded and with Lipschitz boundary. Moreover let us assume at rst that f satises the following
conditions too:
(i) there exists an open set
, such that, for every x
and for every R

mn
,
f(x, ) = 0;
(ii)
f(x, ) exists, is continuous on R

mn
and such that, for every R
mn
f(, )
W
1,
(, R
mn
);
(iii) there exists a constant M such that, for every (x, ) R
mn
,
[
f(x, )[ M, (4.42)
and, for every x ,
1
,
2
R
mn
,
[
f(x,
1
)
f(x,
2
)[ M [
1

2
[ . (4.43)
Let u
h
, u W
1,1
loc
(, R
m
) such that u
h
u in L
1
loc
(, R
m
). Without loss of generality, we can
assume that
liminf
h
I(u
h
, ) = lim
h
I(u
h
, ) < ,
and that u
h
converges almost everywhere to u in . Moreover, from (4.42) and (4.43), we have
I(u, ) < .
As in the proof of Lemma 4.11, for every > 0, there exists v
A(, R
m
) such that,
_
[u(x) v
(x)[ dx , and
_
f(x, v
(x)) dx
_
f(x, u(x))dx . (4.44)

Since v
A(, R
m
), we consider
j
N
j=1
such that, for every j 1, . . . , N, x
j
,
v
(x) =
j
R
mn
in
j
. Now let us consider the sequence of functions
k=1
C
c
() as in
Lemma 4.11 and, using the same argument, we obtain the lower semicontinuity if we prove that,
for every v
and k xed,
lim
h
_
k
(x)
_
f(x, v
(x)) u
h
(x)
f(x, v
(x)) u(x)
_
dx = 0. (4.45)
Observing that, for every j 1, . . . , N,
k
(x)
f(x,
j
) W
1,
(
j
, R
mn
) C
c
(
j
, R
mn
), we
have

k
(x)
_
f(x, v
(x)) u
h
(x)
f(x, v
(x)) u(x)
_
dx
j=1
_
j
k
(x)
f(x,
j
) (u
h
(x) u(x))dx
j=1
n
i=1
m
q=1
_
j
k
(x)
f
q,i
(x,
j
)

x
i
_
u
(q)
h
(x) u
(q)
(x)
_
dx
=
N
j=1
n
i=1
m
q=1
x
i
_
k
(x)
f
q,i
(x,
j
)
_
_
u
(q)
h
(x) u
(q)
(x)
_
dx
j=1
n
i=1
m
q=1
C
i,j
_
u
(q)
h
(x) u
(q)
(x)
dx,
which goes to zero as h , since, for every i 1, . . . , n and j 1, . . . , N,
C
i,j
= sup
_
x
i
(
k
(x)
f(x,
j
))
: x
j
_
< .
In order to treat the general case, we only have to use the same (in fact simpler) argument used
in the proof of Theorem 4.8(1).
The next proposition states that Corollary 4.9(2) holds in the vectorial case if f = f(x, ) (see
Gori, Maggi [53] Proposition 3).
Proposition 4.18. Let f be a continuous function satisfying (4.39) such that f = f(x, ) and,
for every x , f(x, ) is convex and demi-coercive on R
mn
. Then the functional I is l.s.c. on
W
1,1
loc
(, R
m
1
loc
(, R
m
) convergence.
Proof. We consider at rst the case in which f(x, ) is strictly convex. The proof of Theorem
4.1(b), valid in the scalar case, is based on Lemmas 4, 5, 7, 8 of [69]. However, Lemmas 4, 5, 8 are
also true in the vectorial setting. Thus, Lemma 7 is the only one which makes Serrins Theorem
4.1(b) work only in the scalar setting: however, if f does not depend on s, we can easily extend
Lemma 7 to the vectorial setting and nd that Theorem 4.1(b) is true also in this more general
contest.
Once known the lower semicontinuity for the strictly convex case we end in a standard way
applying Theorem 3.5.
Clearly Examples 4.14 and 4.15 show that, if we allow the presence of the s variable, Proposition
4.18 and Theorem 4.17 are false.
The following result, proved by Fonseca and Leoni in [42] (see Theorem 1.1 therein), shows
that if f is coercive and satises some regularity conditions then the lower semicontinuity can be
proved. In particular this theorem implies the vectorial version of Theorem 4.1(a).
Theorem 4.19. Let f be a l.s.c. function satisfying (4.39). Let us suppose that, for every
(x
0
, s
0
) R
m
either, for every R
mn
, f(x
0
, s
0
, ) = 0, or there exists c
0
,
0
> 0 and a
continuous function g : B((x
0
, s
0
),
0
) R
mn
such that the composition
f(x, s, g(x, s)) L
(B((x
0
, s
0
),
0
)) , (4.46)
and, for every (x, s) B((x
0
, s
0
),
0
),
f(x, s, ) c
0
[[
1
c
0
.
1,1
loc
(, R
m
1
loc
(, R
m
) convergence.
A question posed in [42] is about the necessity of assumption (4.46). This hypothesis comes out
by the use of the approximation theorem by Ambrosio (see [4] Lemma 1.5, Statement A). Example
4.16 shows that such assumption in general cannot be dropped. However, it may be interesting
to note that, if we assume a local coerciveness with superlinear growth, then Theorem 4.19 holds
without (4.46), as stated in the next proposition (see Gori, Maggi and Marcellini [54] Proposition
5.4).
Proposition 4.20. Let f be a l.s.c. function satisfying (4.39). Let us suppose that, for every
(x
0
, s
0
) R
m
either, for every R
mn
, f(x
0
, s
0
, ) = 0, or there exists c
0
,
0
> 0, p
0
> 1
such that, for every (x, s) B((x
0
, s
0
),
0
),
f(x, s, ) c
0
[[
p0
1
c
0
. (4.47)
1,1
loc
(, R
m
1
loc
(, R
m
) convergence.
Proof. The usual blow up method reduce the problem of the lower semicontinuity to prove the
inequality
liminf
h
_
Q
n
f(x
0
+
k
y, s
0
+
h
w
h
(y), w
h
(y))dy f(x
0
, s
0
,
0
),
where (x
0
, s
0
,
0
) RR
n
,
h
0 and w
h
h=1
W
1,1
(Q
n
) is a sequence such that w
h
w
0
in L
1
(Q
n
), where Q
n
= (0, 1)
n
and w
0
(y) =
0
, y).
This trivially holds if, for the pair (x
0
, s
0
) it is f(x
0
, s
0
, ) = 0. Otherwise there exist
0
, c
0
> 0,
p
0
> 1 such that, for every (x, s, ) B((x
0
, s
0
),
0
) R
mn
,
f(x, s, ) c
0
[[
p
0
1
c
0
.
If we set M = B((x
0
, s
0
),
0
) we are in the hypotheses of Lemma 1.5, Statement B by Ambrosio
[4] and thus there exist continuous functions a
h
: M R and b
h
: M R
mn
such that
f(x, s, ) = sup
hN
a
h
(x, s) +b
h
(x, s) ,
for every (x, s, ) B
0
(x
0
, s
0
) R
mn
. Now we can conclude the proof working as in Theorem 1.1
of [42] quoted above.
Finally we note that, under the assumption (4.47), if the coerciveness constant doesnt depend
on s, we can consider the case in which f is only Borel measurable with respect to x. The following
result follows Proposition 5.6 in Gori and Marcellini [55].
m
R
nm
be a Borel function such that, for L
n
-a.e. x ,
f(x, , ) is l.s.c., for every s R, f(x, s, ) is convex and, for every (s, ) R
m
R
mn
,
f(x, s, ) a(x) [[
p
, (4.48)
where p > 1 and a : [0, ] is a Borel function such that
a
1
p1
L
1
loc
(). (4.49)
1,1
loc
(, R
m
1
loc
(, R
m
) convergence.
Proof. Let u
h
, u W
1,1
loc
(, R
m
) such that u
h
u in L
1
loc
(, R
m
). We have to prove that
liminf
h
I (u
h
, ) I (u, ) . (4.50)
To this aim we can assume that
liminf
h
I (u
h
, ) = lim
h
I (u
h
, ) = C < +.
By Holder inequality and by the coercivity assumption (4.48), for every B B(), h N, we have
_
B
[u
h
(x)[ dx =
_
B
a(x)
1
p
[u
h
(x)[ a(x)
1
p
dx
__
B
a(x) [u
h
(x)[
p
dx
_1
p
__
B
a(x)
1
p
p
p1
dx
_
(p1)
p
__
B
f (x, u
h
(x), u
h
(x)) dx
_1
p
__
B
a(x)
1
p1
dx
_
p1
p
C
1
p
__
B
a(x)
1
p1
_
p1
p
.
Then u
h
h=1
is a sequence locally equi-integrable on and in fact u
h
u in w-W
1,1
loc
(, R
m
):
therefore we can apply the lower semicontinuity theorem by Ioe (see [56]).
Chapter 5
Lower semicontinuity for supremal
functionals
As stated in the introduction, in this chapter we consider the problem, rst approached by Gori
and Maggi in [52], to understand in which cases the functional
S(u, ) = ess sup
x
g(x, u(x), u(x)),
is l.s.c. on W
1,
loc
loc
() convergence.
The more natural problem of the lower semicontinuity of S with respect to the w
-W
1,
loc
()
convergence has been deeply studied by several authors (see Barron and Jensen [14], Barron and
Liu [16] and Barron, Jensen and Wang (see [15]) and it has been immediately pointed out the
fundamental role of the level convexity of g in the gradient variable.
One of the main results on this topic is due to Barron, Jensen and Wang (see [15] Theorem
3.4), who provide for supremal functionals a result similar to the classical one of Ioe [56] in the
integral setting.
Theorem 5.1. Let g : RR
n
[0, ] be a proper and Borel function such that, for L
n
-a.e.
x , g(x, , ) is l.s.c. and, for every s R, g(x, s, ) is level convex. Then the functional S is
l.s.c. on W
1,
loc
() with respect to the w
-W
1,
loc
() convergence.
In order to prove this theorem the authors show rst the fact that the lower semicontinuity
of a supremal functional can be always reduced to the lower semicontinuity of the elements of a
suitable family of integrals whose integrands can also be innite. Then they conclude proving that
the hypotheses of Theorem 5.1 allow to apply to every functional of the considered family just
Ioes theorem quoted above.
Considering now the lower semicontinuity of S with respect to the L
loc
() convergence, it is
quite simple to understand that the integral reduction approach by Barron, Jensen and Wang is
not suitable anymore. Indeed, the theory of lower semicontinuity for integrals with respect to this
kind of convergence requires the integrand to be regular enough in the lower order variables and,
in particular, to be nite (as seen in Chapter 4).
Thus the problem must be approached in a dierent way. Following [52], by means of suitable
tools from the convex analysis, we are able to nd simple conditions on g sucient for the lower
semicontinuity, that generalize Theorem 1.3 in [52]; moreover we carry on also an analysis on the
necessary conditions. Example 5.2, Theorem 5.5 and Theorem 5.8 below are the main results of
this approach.
Finally we stress the fact that in the whole chapter we are dealing with scalar valued functions
u : R. However, the situation in the vectorial setting, at least about the sucient conditions,
65
66 CHAPTER 5. LOWER SEMICONTINUITY FOR SUPREMAL FUNCTIONALS
is claried by Example 4.14. Indeed, it shows that Theorems 5.4 and 5.5 dont hold in the vectorial
case, that is, when g : R
m
R
mn
[0, ] and u W
1,
loc
(, R
m
).
5.1 A counterexample
An important remark about Theorem 5.1 is that it holds even if we consider the lower semiconti-
nuity of S on W
1,1
loc
() with respect to the w-W
1,1
loc
() convergence.
In Example 5.2 below we consider an upper semicontinuous function g = g(x, ) : (0, 1) R
[0, ) such that, for every x (0, 1), g(x, ) is level convex and such that the supremal functional
S, related to g, is not l.s.c. on a suitable sequence of functions in C
([0, 1]) converging with respect

to the w
-BV (0, 1) convergence. This means that if in Theorem 5.1 we want to consider weaker
convergences than the w-W
1,1
loc
(), like the w
-BV
loc
() convergence, we have to strengthen the
measurability assumption of g on the geometric variable.
This example, that can be found in [52] Example 2.2, suggests that a suitable regularity con-
dition on g, in order to obtain the lower semicontinuity of S, is its global lower semicontinuity on
R R
n
(obviously together with the level convexity on the gradient variable).
Example 5.2. Let us show that there exist a closed set K [0, 1] of L
1
-positive measure, a
sequence u
h
h=1
C
([0, 1]) such that u

h
0 in L
(0, 1) and with

sup
hN
__
1
0
[u
h
(x)[dx
_
< , (5.1)
and 0 < c < d, 0 < <
dc
2
such that, for every x K, h N,
0 < c + < u
h
(x) < d . (5.2)
In this way, setting f() =
_

(dc)
2
_
2
and g(x, ) = 1
K
(x)f(), we obtain
S(0, (0, 1)) =
_
d c
2
_
2
>
_
d c
2

_
2
liminf
h
S(u
h
, (0, 1)).
In order to do this, let us x t > 3 and, for every 0 < a < b < 1 with b a > t
1
, we set
T
h
([a, b]) =
_
a,
b a
2

1
2t
h
_
_
b a
2
+
1
2t
h
, b
_
.
Then we dene, by induction on h N,
c
1
= T
1
([0, 1]) =
_
I
1
i
_
2
i=1
, c
h
=
2
h1
_
i=1
T
h
_
I
h1
i
_
=
_
I
h
i
_
2
h
i=1
,
where, for every h N, the intervals I
h
i
are ordered in such a way that supI
h
i
< inf I
h
i+1
. Finally
we set
K
h
=
2
h
_
i=1
I
h
i
, and K =
h=1
K
h
.
Clearly K is a closed set, and since
L
1
_
I
h
i
_
=
1
2
h
_
1
1
t
h1
k=0
_
2
t
_
k
_
,
5.2. SUFFICIENT CONDITIONS 67
we have L
1
(K) = (t 3)/(t 2) (K is simply one of the standard variants of the Cantors Set).
Fixed now h N, for every i 1, . . . , 2
h
, let us dene u
h
on I
h
i
as the ane function that
takes the value 0 on the left extreme of I
h
i
and the value 2
h1
on the right extreme. Then we
extend u
h
on [0, 1] in a smooth way, under the constraint that, for every i 1, . . . , 2
h
the total
variation of u
h
on the interval between I
h
i
and I
h
i+1
is controlled by 3 2
h1
.
Clearly, for every x (0, 1), it is [u
h
(x)[ 3 2
h1
, and then u
h
0 in L
(0, 1). Moreover

the total variation can be estimated, looking carefully at the construction, by
_
1
0
[u
h
(x)[ dx
_
2
h
+ (2
h
1)
_
3
2
h+1
,
that implies (5.1). Finally, for every h N and x K
h
, it is
u
h
(x) =
1
2
h+1

1
L
1
_
I
h
i
_ =
1
2
_
1
1
t
h1
k=0
_
2
t
_
k
_,
thus, in particular, since we have
lim
h
_
1
1
t
h1
k=0
_
2
t
_
k
_
=
t 3
t 2
(0, 1),
we can choose c, d and satisfying (5.2).
Remark 5.3. Example 5.2 provides a counterexample also in the integral setting: indeed, if we
consider the integral functional I in (1.1) built up starting from the function g of Example 5.2 it
says that Ioes lower semicontinuity Theorem (see [56]) doesnt hold with respect to w
-BV ()
convergence, at least with upper semicontinuity with respect to the geometric variable x. This fact
is classically shown with a counterexample by Carbone and Sbordone [23] (see also [22]).
5.2 Sucient conditions
As Example 5.2 suggests, we will consider the case in which g is l.s.c. on RR
n
and, of course,
level convex in the gradient variable.
By means of Theorem 2.13 we can prove our lower semicontinuity results. The rst one is quite
surprising: it shows that in the scalar case we still have lower semicontinuity with respect to the
L
loc
() convergence on C
1
sequences, even if level convexity is dropped (see [52] Theorem 1.1).
Theorem 5.4. Let g : RR
n
[0, ] be a proper and l.s.c. function and let u
h
h=1
C
1
(),
u W
1,
loc
() such that u
h
u in L
loc
(). Then
S(u, ) liminf
h
S(u
h
, ).
Proof. Let us take t > liminf
h
S(u
h
, ) (we can suppose t < ). By the continuity of u
h
and
u
h
, the function g(x, u
h
(x), u
h
(x)) of the x variable is l.s.c. on , thus the essential supremum
in the denition of S is a point-wise supremum, that is,
S(u
h
, ) = sup
x
g(x, u
h
(x), u
h
(x)).
Then there exists h
t
N such that, for every h h
t
, x , we have g(x, u
h
(x), u
h
(x)) t.
Let now
d
the set in which u is dierentiable and x x
0

d
: we know that L
n
(
d
) = 0
and that, applying Theorem 2.13 and Proposition 2.12(iv), there exists a subsequence u
h
k
k=1
and a sequence x
k
k=1
, such that x
k
x
0
, u
h
k
(x
k
) u(x
0
) and u
h
k
(x
k
) u(x
0
). Hence,
by the lower semicontinuity of g, we have, for every x
0

d
,
g(x
0
, u(x
0
), u(x
0
)) liminf
h
g(x
h
, u
h
(x
h
), u
h
(x
h
)) t,
that is S(u, ) t: then the thesis follows.
In order to prove a general lower semicontinuity theorem we need to assume the level convexity
of g in the gradient variable. We prove that if g is l.s.c. and level convex in then S is lower
semicontinuous on W
1,
loc
loc
() topology. Let us note that this theorem
generalizes Theorem 1.3 in [52].
Theorem 5.5. Let g : R R
n
[0, ] be a proper, l.s.c. function such that, for every
(x, s) R, g(x, s, ) is level convex. Then the functional S is l.s.c. in W
1,
loc
() with respect to
the L
loc
() convergence.
Proof. Let u, u
h
W
1,
loc
() such that u
h
u in L
loc
(). We can suppose that
liminf
h
S(u
h
, ) = lim
h
S(u
h
, ) = L < +.
Let now t > L: then there exists h
t
t
, S(u
h
, ) t. In particular
we can nd
1
, with L
n
(
1
) = 0, such that, for every h h
t
, x
1
, we have
g(x, u
h
(x), u
h
(x)) t and u(x), u
h
(x) exist.
We claim now that the l.s.c. and the level convexity of g with respect to the gradient variable
imply that, for every h h
t
, x ,
c
u
h
(x), we have g(x, u
h
(x), ) t.
Indeed, let us consider h h
t
, x and the set u
h
(x) as in (2.12) built up with respect to
1
. By the l.s.c. of g we trivially have that, for every x , u
h
(x), g(x, u
h
(x), ) t.
Now, by Proposition 2.12(i), we still have
c
u
h
(x) = co
_
u
h
(x)
_
and then, by Theorem 2.14,
for every
c
u
h
(x), there exist
j
n+1
j=1
u
h
(x) and
j
n+1
j=1
, with 0
j
1 and
n+1
j=1

j
=
1, such that =
n+1
j=1

j
j
. Thus, by the denition of level convexity,
g(x, u
h
(x), ) = g
_
_
x, u
h
(x),
n+1
j=1
j
_
_
n+1
j=1
g(x, u
h
(x),
j
) t,
and the claim is proved.
Fixed now x
0

1
we have that, by Theorem 2.13, there exists a subsequence u
h
k
k=1
and
two sequences x
k
k=1
,
k
k=1
such that, for every k N,
k

c
u
h
k
(x
k
) and it holds x
k
x
0
,
u
h
k
(x
k
) u(x
0
) and
k
u(x
0
). Then
(x
k
, u
h
k
(x
k
),
k
) (x
0
, u(x
0
), u(x
0
))
and by the lower semicontinuity of g it is g(x
0
, u(x
0
), u(x
0
)) t. Thus S(u,
1
) = S(u, ) t.
Since this holds for every t > L it follows S(u, ) L and we get the proof.
5.3 Necessary conditions
We end this chapter trying to prove that the level convexity of g in the gradient variable is a
necessary condition for the lower semicontinuity of S. What we would like it happens, in order to
nd a theorem stating necessary and sucient conditions, is that supposing g l.s.c. on RR
n
and that, for every open set
, S(u,
) is l.s.c. on W
1,
loc
(
loc
(
)
convergence, then g is level convex in the variable.
However, the situation is more complicated it seems. We propose here an example that shows
a particular situation we can nd.
5.3. NECESSARY CONDITIONS 69
Example 5.6. Let = (0, 1) and g : RR [2, ) be a continuous function such that, for
every (x, s) R, g(x, s, ) is level convex. By Theorem 5.5, for every
open set, S(u,
)
is l.s.c on W
1,
loc
(
loc
(
) convergence. Let us dene the following closed

set
A = (x, x, ) : x , (, 0] R R,
and let us consider a continuous function h : A [0, 1] such that, for every x , h(x, x, ) is not
level convex in the variable. We set now
g
(x, s, ) =
_
g(x, s, ) if (x, s, ) ( R R) A,
h(x, s, ) if (x, s, ) A;
by construction we have that g
is a l.s.c. function that doesnt satisfy the hypotheses of Theorem

5.5 about the level convexity. Note also that g
cannot be obtain by g by means of a trivial

modication of g on a set of the type M R R
n
, where M and L
1
(M) = 0.
Let us denote with S
the supremal functional associated to g
and prove that, for every

open set, S
(u,
) is l.s.c. on W
1,
loc
(
loc
(
) convergence.
In order to do this we claim at rst that, for every function u W
1,
loc
(), dened
d
= x
: u
(x) exists, the set

Q(u) = x : (x, u(x), u
(x)) A
d
has L
1
-null measure: let us note at rst that every point of Q(u) is isolated. Indeed, for every
x Q(u) we have that u
(x) exists, u(x) = x and u
(x) 0: then, in particular, we have

lim
yx
u(y) u(x)
y x
0.
It follows that we can nd a neighbor V (x) of x such that, for every y V (x) x, u(y) ,= y, else
there exists y
n
x such that u(y
n
) = y
n
that implies
lim
n
u(y
n
) u(x)
y
n
x
= lim
n
y
n
x
y
n
x
= 1.
Now let us dene, for every i N,
Q
i
(u) =
_
x Q(u) : B1
i
(x) Q(u) = x
_
.
Trivially there are only a nite number of elements in every Q
i
(u) and since Q(u) =

i=1
Q
i
(u)
we have that Q(u) is countable thus, in particular, of L
1
-null measure and the claim is proved.
From this fact it follows that, for every open set
and for every function u W

1,
loc
(
),
S(u,
) = S
(u,
). Indeed, we have
S
(u,
) = ess sup
x
(x, u(x), u
(x)) = ess sup

x(
d
\Q(u))
g(x, u(x), u
(x)) =
ess sup
x(
d
\Q(u))
g(x, u(x), u
(x)) = S(u,
).
Thus S
(u,
) is l.s.c. on W
1,
loc
(
loc
(
) convergence even if g
is not level
convex in the gradient variable.
Example 5.6 show that, starting from a l.s.c. function g such that its associated supremal
functional is l.s.c., without further information on the regularity of g, we have no hope to prove
nothing about its convexity property.
However, at the same time, the example shows that it is natural to identify two l.s.c. functions
g, g
with the property that, for every u W

1,
loc
() and for L
n
-a.e. x ,
g(x, u(x), u(x)) = g
(x, u(x), u(x)),

since their supremal functionals agree. This identication clearly denes an equivalence relationship
that we denote by .
From this new point of view, given for instance a l.s.c. function g, the level convexity as a
necessary condition to the lower semicontinuity have to be interpreted in the following way: there
exists a l.s.c. function g
such that g g
and, for every (x, s) R, g
(x, s, ) is level convex.

In the following we will study this relationship restricted to a particular class of functions
satisfying same regularities and we try to understand the form of the equivalence classes. Our
purpose, in order to prove necessary conditions, is to nd a particular class of functions in which
every equivalence class for the relationship is a singleton. Evidently, by Example 5.6, we cannot
consider only l.s.c. functions but we need much more regularity.
The following simple proposition identify suitable regularity conditions.
Proposition 5.7. Let g, g
: R R
n
[0, ] be proper and l.s.c. functions such that, for
every R
n
, g(, , ) and g
(, , ) are continuous on R and let us suppose that g g
. Then
g = g
.
Proof. Let us x (x
0
, s
0
,
0
) R R
n
and u(x) = s
0
+
0
, x x
0
). Then, since g g
, we
have that for L
n
-a.e. x , g(x, u(x),
0
) = g
(x, u(x),
0
). In particular this holds for a dense
subset of thus, by continuity, we conclude that g(x
0
, s
0
,
0
) = g
(x
0
, s
0
,
0
).
The following theorem shows that, considering the set of functions of the proposition above,
the level convexity is just a necessary condition for lower semicontinuity.
Theorem 5.8. Let g : R R
n
[0, ] be a proper and l.s.c. function such that, for every
R
n
, g(, , ) is continuous on R and such that, for every
open set, S(,
) is
l.s.c. on W
1,
loc
(
) with respect to the w
-W
1,
loc
() convergence. Then, for every (x, s) R,
g(x, s, ) is level convex.
Proof. We argue by contradiction. If there exists (x
0
, s
0
) R such that g(x
0
, s
0
, ) is not level
convex, then there exists r
0
R such that E
g(x
0
,s
0
,)
(r
0
) is not convex in R
n
. Thus there are
1
,
2
E
g(x
0
,s
0
,)
(r
0
) and t (0, 1) such that
0
= t
1
+ (1 t)
2
, E
g(x
0
,s
0
,)
(r
0
).
Let us consider now the function u(x) = s
0
+
0
, x x
0
) W
1,
loc
(R
n
). Its quite simple to nd
a sequence u
h
W
1,
loc
(R
n
) such that u
h
u in w
-W
1,
loc
(R
n
) and such that, for every h N,
when u
h
exists, its value is
1
or
2
. Such a sequence can be built up working in this way. Let
us consider the function v : R R
n
dened componentwise as
v
i
(y) =
_

1,i
y (x
0,i
, x
0,i
+t),
2,i
y (x
0,i
+t, x
0,i
+ 1),
where i 1, . . . , n, and extended on R by periodicity. Dened now v
h
(y) = v(hy) : R R
n
, we
can consider the functions
u
h
(x) = s
0
+
n
i=1
_
x
i
x0,i
v
h,i
(y)dy W
1,
loc
(R
n
).
Its simple to verify that u
h
u in w
-W
1,
loc
(R
n
). Then, using this sequence, for every > 0, we
have
r
0
< g(x
0
, s
0
,
0
) S (u, B(x
0
, )) liminf
h
S (u
h
, B(x
0
, )) .
5.3. NECESSARY CONDITIONS 71
Dening 2 = g(x
0
, s
0
,
0
) r
0
> 0, by the continuity of g(, , ) and the hypotheses on u
h
, if h is
big enough and is small enough, we have, for every x B(x
0
, ),
g(x, u
h
(x),
1
) r
0
+

2
, g(x, u
h
(x),
2
) r
0
+

2
.
Thus it follows
r
0
+ liminf
h
S (u
h
, B(x
0
, )) r
0
+

2
and we have found a contradiction.
Chapter 6
Functionals dened on measures
In the rst part of this chapter we study some particular integral and supremal functionals dened
on the space of Radon measures. Let f, g : R
m
[0, ] be proper and Borel functions and let us
consider the following functionals dened, for every /
loc
(, R
m
), as
I
m
(, ) =
_
f(
a
(x))dx +
_
_
d
s
d[
s
[
(x)
_
d[
s
[(x), (6.1)
and
S
m
(, ) =
_
ess sup
x
g(
a
(x))
_
_
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x)
__
, (6.2)
where the measure has been decomposed as in (2.3) with respect to the Lebesgues measure L
n
.
The main purpose of the rst part of this chapter is to prove that, if we suppose f l.s.c. and
convex and g l.s.c. and level convex, then I
m
and S
m
are l.s.c. on /
loc
(, R
m
) with respect to
the w
-/
loc
(, R
m
) convergence.
The functional I
m
has been deeply studied by several authors and in particular by Goman and
Serrin [49], Ambrosio and Buttazzo [5], Bouchitte [17], Bouchitte and Valadier [21] and De Giorgi,
Ambrosio and Buttazzo [36] (see also the book of Buttazzo [22]). In these papers necessary and
sucient conditions on f for the lower semicontinuity of I
m
have been found even when f depends
on the geometric variable x.
On the contrary, the functional S
m
dened above seems to be new.
The strategy we follow in order to prove the lower semicontinuity of I
m
and S
m
is essentially
the same of Serrin [69] and, in particular, of Goman and Serrin [49], where the study of the
functional I
m
is approached for the rst time. Indeed, as already shown in [51], the arguments of
these papers can be adapted to the supremal setting too
1
.
Before stating the main theorem concerning I
m
and S
m
, let us dene, for every /
loc
(, R
m
)
and with respect to the decomposition (2.3) of with respect to L
n
, the following functionals
I
m
(, ) =
_
f(
a
(x))dx, (6.3)
S
m
(, ) = ess sup
x
g(
a
(x)), (6.4)
I
m
(, ) = inf
_
liminf
h
I
m
(
h
,
h
) :
h
/
loc
(
h
, R
m
),
h
<< L
n
,
h
w
-/
loc
( , R
m
)
_
,
(6.5)
1
Note that in [51] it is considered the less general case in which Sm is dened on BV
loc
(), that is = Du (see
Chapter 7).
73
74 CHAPTER 6. FUNCTIONALS DEFINED ON MEASURES
and
S
m
(, ) = inf
_
liminf
h
S
m
(
h
,
h
) :
h
/
loc
(
h
, R
m
),
h
<< L
n
,
h
w
-/
loc
( , R
m
)
_
.
(6.6)
The following theorem shows the complete analogy between the functionals (6.1) and (6.2).
Theorem 6.1. Let f, g : R
m
[0, ] be proper and Borel functions and consider the two following
pairs of conditions:
(a) f is l.s.c. and convex on R
m
,
(b) for every open set R
n
, the functional I
m
is l.s.c. on /
loc
(, R
m
) with respect to the
w
-/
loc
(, R
m
) convergence,
and
(i) g is l.s.c. and level convex on R
m
,
(ii) for every open set R
n
, the functional S
m
is l.s.c. on /
loc
(, R
m
w
-/
loc
(, R
m
) convergence.
Then (a) is equivalent to (b) so as (i) to (ii). Moreover when the previous conditions are satised,
for every /
loc
(, R
m
), we have
I
m
(, ) = I
m
(, ) = lim
0
I
m
(
L
n
,
), (6.7)
S
m
(, ) = S
m
(, ) = lim
0
S
m
(
L
n
,
), (6.8)
where
denotes the convolution of , and I

m
and S
m
are convex and level convex respectively on
/
loc
(, R
m
).
The second part of the chapter is devoted to the study of the lower semicontinuity of other two
classes of functionals dened, for every /
loc
(, R
m
), as
I
m
(, ) =
_
f(
a
(x))dx +
_
_
d
c
d[
c
[
(x)
_
d[
c
[(x) +
xA
(
#
(x)), (6.9)
and
S
m
(, ) =
_
ess sup
x
g(
a
(x))
_
_
[
c
[-ess sup
x
g
_
d
c
d[
c
[
(x)
__
_

xA
(
#
(x))
_
, (6.10)
where f, g, , : R
m
[0, ] are Borel functions.
First of all we remark that I
m
and S
m
generalize I
m
and S
m
respectively: in fact, when is
positively homogeneous of degree 1 and is positively homogeneous of degree 0, we have
xA
(
#
(x)) =
_
A
_
d
#
d[
#
[
(x)
_
d[
#
[(x) and
xA
(
#
(x)) = [
#
[-ess sup
xA
_
d
#
d[
#
[
(x)
_
.
If we consider now = f
and = g
, applying Propositions 2.5 and 2.6, we obtain, for every

/
loc
(, R
m
), I
m
(, ) = I
m
(, ) and S
m
(, ) = S
m
(, ): then, when this particular case
is considered, we have that Theorem 6.1 already solves the problem of the lower semicontinuity of
I
m
and S
m
.
6.1. CONVEX AND LEVEL CONVEX FUNCTIONALS 75
It is also clear that, even if we suppose that f and g are convex and level convex, I
m
and S
m
could fail to be convex and level convex on /
loc
(, R
m
) because of the presence of the functions
and : for this reason, with a little abuse, we will call I
m
the non convex functional and S
m
the
non level convex one, contrarily to I
m
and S
m
.
Finally let us note also that the value of and for = 0 does not enter in the computation of
I
m
and S
m
, thus these two functions could be dened only on R
m
0: however, to simplify several
arguments, we prefer to dene and on the whole space R
m
adding the condition (0) = (0) = 0
that, as noted, doesnt reduce the generality of the statements.
The lower semicontinuity properties of I
m
have been studied for the rst time by Bouchitte
and Buttazzo in [18] and other aspects, as the problem of the relaxation or the one of the integral
representation, have been analyzed by the same authors in subsequent works (see [19], [20]). In the
supremal case instead, this kind of study has not been developed, even if a functional very similar
to S
m
but dened on BV
loc
(a, b) has been recently studied by Alicandro, Braides and Cicalese in
[3] (see also Chapter 7).
Among the results proved for the functional I
m
it is interesting to quote the following one in
which we can nd both necessary and sucient conditions (it follows from Theorem 3.3 [18] and
Theorem 2.3 [19]).
m
[0, ] and : R
m
[0, ] be proper and Borel functions with
(0) = 0. Then the following conditions are equivalent:
(a) f is l.s.c. and convex on R
m
, is l.s.c and sub-linear on R
m
with (0) = 0 and, for every
R
m
0, f
() =
0
(),
(b) for every open set R
n
, the functional I
m
is l.s.c. on /
loc
(, R
m
w
-/
loc
(, R
m
) convergence.
Here, keeping in mind the previous theorem proved for I
m
, we propose some analogous results on
S
m
. We are able to prove in particular the two following theorems where we give some necessary
and sucient conditions for the lower semicontinuity: even in this case it is clear the analogy
between the integral and the supremal setting.
m
[0, ] and : R
m
[0, ] be proper and Borel functions with (0) =
0. Let us suppose that, for every open set R
n
, the functional S
m
is l.s.c. on /
loc
(, R
m
) with
respect to the w
-/
loc
(, R
m
) convergence. Then g is l.s.c. and level convex on R
m
and, dened
l = infg() : R
m
, l is l.s.c. and sub-maximal on R
m
. Moreover, for every R
m
0,
g
() = ( l)
().
m
[0, ] be a l.s.c. and level convex function, l = infg() : R
m
and : R
m
[0, ] such that l is a l.s.c and sub-maximal function with (0) = 0. Let us
suppose that, for every R
m
0, g
() = ( l)
() = . Then the functional S

m
is l.s.c. on
/
loc
(, R
m
-/
loc
(, R
m
) convergence.
As we can see, we are not able to nd conditions on g and that are both necessary and
sucient for the lower semicontinuity of S
m
: our conjecture is that in Theorem 6.4 the condition
g
= ( l)
= could be replaced with the weaker condition g
= ( l)
that we already
know, thanks to Theorem 6.3, to be necessary and, as Alicandro, Braides and Cicalese show in [3]
Theorem 4.4, to be also sucient at least in dimension one.
6.1 Convex and level convex functionals
This section is devoted to the proof of Theorem 6.1. Let us start considering the implications
(b) (a) and (ii) (i).
Proof of Theorem 6.1 (b) (a). Let us consider in the following = Q
n
= (0, 1)
n
the unit cube
in R
n
: clearly L
n
(Q
n
) = 1. Let us x
h
,
0
R
m
,
h

0
and let
h
=
h
L
n
,
0
=
0
L
n
. We
have
h

0
in w
-/
loc
(Q
n
, R
m
) and then
f(
0
) = I
m
(
0
, Q
n
) liminf
h
I
m
(
h
, Q
n
) = liminf
h
f(
h
),
that is f is l.s.c..
Let us consider now , R
m
, t (0, 1) and a sequence B
h
h=1
B(Q
n
) such that, for every
h N, L
n
(B
h
) = t and 1
B
h
(x) t1
Q
n(x) in w
-L
(Q
n
) (see [2] Proposition 4.2 and Remark 4.3).
If
h
= 1
B
h
(x) L
n
+ 1
Q
n
\B
h
(x) L
n
then
h

0
= (t + (1 t)) L
n
in w
-/
loc
(Q
n
, R
m
)
thus
f(t + (1 t)) = I
m
(
0
, Q
n
) liminf
h
I
m
(
h
, Q
n
)
= liminf
h
_
L
n
(B
h
)f() +L
n
(Q
n
B
h
)f()
_
= tf() + (1 t)f(),
that is f is convex.
Proof of Theorem 6.1 (ii) (i). Let
h
,
0
R
m
,
h

0
and let
h
=
h
L
n
,
0
=
0
L
n
. We
have
h

0
in w
-/
loc
(R
n
, R
m
) and then
g(
0
) = S
m
(
0
, R
n
) liminf
h
S
m
(
h
, R
n
) = liminf
h
g(
h
),
that is g is l.s.c..
Let us consider now , R
m
, t (0, 1) and a sequence B
h
h=1
B(R
n
) such that 1
B
h
(x)
t1
R
n(x) in w
-L
(R
n
) (see [2] Proposition 4.2 and Remark 4.3). If
h
= 1
B
h
(x)L
n
+1
R
n
\B
h
(x)
L
n
then
h

0
= (t + (1 t)) L
n
in w
-/
loc
(R
n
, R
m
) thus
g(t + (1 t)) = S
m
(
0
, R
n
) liminf
h
S
m
(
h
, R
n
) = g() g(),
that is g is level convex.
The proofs of the implications (a) (b) and (i) (ii) of Theorem 6.1 are more dicult and
for this reason their are developed in three steps. In the rst one we prove that, for every
/
loc
(, R
m
), I
m
(, ) and S
m
(, ) can be computed considering only the convolutions of as in
(6.7) and (6.8) (see Theorem 6.6). This fact implies, in a very simple way, the second step that is the
proof that I
m
and S
m
are l.s.c. on /
loc
(, R
m
-/
loc
(, R
m
) convergence
(see Proposition 6.7) and the properties of convexity and level convexity on /
loc
(, R
m
) (see
Proposition 6.8). At last in the third step we prove the equalities I
m
= I
m
and S
m
= S
m
(see
Theorems 6.16 and 6.17) that complete the proof of Theorem 6.1.
6.1.1 Simple results via convolutions
Let us start studying I
m
and S
m
. Lemma 6.5 and Theorem 6.6 are completely inspired to Lemma
1 and Theorem 1 of [69] (see also [51]).
Lemma 6.5. Let f : R
m
[0, +] be a proper, l.s.c and convex function and g : R
m
[0, +]
be a proper, l.s.c and level convex function. Let /
loc
(, R
m
),
h
/
loc
(
h
, R
m
),
h
<< L
n
such that
h
in w
-/
loc
( , R
m
). Then, for every > 0, we have
I
m
(
L
n
,
) liminf
h
I
m
(
h
,
h
).
and
S
m
(
L
n
,
) liminf
h
S
m
(
h
,
h
).
Proof. Let us x > 0 and K
. We have that (
h
)
point-wise on K as h .
Indeed, if x K we have that, if h is large enough, K
h,
and
[(
h
)
(x)
(x)[ =
_
B(x,)
k
(x y)d
h
(y)
_
B(x,)
k
(x y)d(y)
that converges to zero as h by denition of w
-/
loc
( , R
m
) convergence.
Let us prove the relation for I
m
. Let us x x K and apply Theorem 2.18 with = k
(xy)L
n
:
we have
f(
(x)) liminf
h
f
_
(
h
)
(x)
_
= liminf
h
f
_
_
B(x,)
a
h
(y)d(y)
_
liminf
h
_
B(x,)
f(
a
h
(y))d(y).
Then integrating on K and using Fubinis and Fatous Theorems
_
xK
f(
(x))dx
_
xK
liminf
h
_
_
B(x,)
k
(x y)f(
a
h
(y))dy
_
dx
liminf
h
_
K
_
_
B(x,)
k
(x y)f(
a
h
(y))dy
_
dx liminf
h
_
K+B(0,)
_
_
B(y,)
k
(x y)dx
_
f(
a
h
(y))dy
liminf
h
_
h
f(
a
h
(x))dx = liminf
h
I
m
(
h
,
h
),
and we conclude by taking K
.
The argument is simpler for S
m
. Let us x x K and apply Theorem 2.19 and (2.2) with
= k
(x y) L
n
<< L
n
, in order to get
g(
(x)) liminf
h
g
_
(
h
)
(x)
_
liminf
h
ess sup
yB(x,)
g(
a
h
(y)) liminf
h
S
m
(
h
,
h
).
Then
sup
xK
g(
(x)) liminf
h
S
m
(
h
,
h
),
and we conclude by taking K
.
By means of Lemma 6.5 we nd a representation formula for I
m
and S
m
via convolutions,
proving in this way the rst step of the proof of Theorem 6.1.
m
[0, +] be a proper, l.s.c. and convex function and g : R
m
[0, +]
be a proper, l.s.c. and level convex function. Then, for every /
loc
(, R
m
), we have
I
m
(, ) = lim
0
I
m
(
L
n
,
) and S
m
(, ) = lim
0
S
m
(
L
n
,
). (6.11)
In particular the limits in the right hand sides exist.
Proof. Let us consider S
m
. For every /
loc
(, R
m
),
L
n
in w
-/
loc
( , R
m
) as 0
and
L
n
<< L
n
so that, by Lemma 6.5,
limsup
0
S
m
(
L
n
,
) S
m
(, ) liminf
0
S
m
(
L
n
,
).
The same argument prove the equality also for the integral case.
By means of Theorem 6.6 we are able to prove the two following propositions that complete
the second step of the proof of Theorem 6.1.
m
m
[0, +] be a proper, l.s.c. and level convex function. Then I
m
and S
m
are l.s.c. on /
loc
(, R
m
)
with respect to the w
-/
loc
(, R
m
) convergence.
Proof. Since the same proof works in both cases we make it for the functional S
m
. Let ,
h

/
loc
(, R
m
) be such that
h
in w
-/
loc
(, R
m
). Then, by Theorem 6.6, there exists a
sequence
h
0 such that, for every h N,
S
m
_
(
h
)
h
L
n
,
h
_
S
m
(
h
, ) +
1
h
. (6.12)
Now we claim that (
h
)
h
L
n
in w
-/
loc
( , R
m
). Indeed, xed C
c
(), if h is so
large that spt() + B(0,
h
)
h
and by using the usual properties of the convolutions (see [7]
equation (2.3) page 42), we have
(x)(
h
)
h
(x)dx
_
(x)d(x)
(x)(
h
)
h
(x)dx
_
(x)d
h
(x)
(x)d
h
(x)
_
(x)d(x)
h
(x) (x)[ d
h
(x) +
(x)d
h
(x)
_
(x)d(x)
.
Now the rst term goes to zero as h since
h
uniformly on and there exists a
constant M > 0 such that, for every h N, [
h
[(spt() +B(0,
h
)) M; the second term goes to
zero by denition of
h
h=1
.
Then using the denition of S
m
and (6.12) we obtain its lower semicontinuity since
S
m
(, ) liminf
h
S
m
_
(
h
)
h
L
n
,
h
_
liminf
h
S
m
(
h
, ),
which ends the proof.
m
m
[0, +] be a proper, l.s.c. and level convex function. Then, for every
1
,
2
/
loc
(, R
m
) and
for every t (0, 1),
I
m
_
t
1
+ (1 t)
2
,
_
tI
m
(
1
, ) + (1 t)I
m
(
2
, ),
that is, I
m
is convex on the linear space /
loc
(, R
m
), and
S
m
_
t
1
+ (1 t)
2
,
_
S
m
(
1
, ) S
m
(
2
, ),
that is, S
m
is level convex on the linear space /
loc
(, R
m
).
Proof. Fixed > 0, by the convexity of f, we have, for every x
,
f
_
(t
1
+ (1 t)
2
)
(x)
_
= f
_
t(
1
)
(x) + (1 t)(
2
)
(x)
_
tf
_
(
1
)
(x)
_
+ (1 t)f
_
(
2
)
(x)
_
,
and then
I
m
_
(t
1
+ (1 t)
2
)
L
n
,
_
tI
m
_
(
1
)
L
n
,
_
+ (1 t)I
m
_
(
2
)
L
n
,
_
.
Analogously, xed > 0, by the level convexity of g, we have, for every x
,
g
_
(t
1
+ (1 t)
2
)
(x)
_
= g
_
t(
1
)
(x) + (1 t)(
2
)
(x)
_
g
_
(
1
)
(x)
_
g
_
(
2
)
(x)
_
,
and then
S
m
_
(t
1
+ (1 t)
2
)
L
n
,
_
S
_
(
1
)
L
n
,
_
S
_
(
2
)
L
n
,
_
.
Passing to the limit as 0 we complete the proof applying Theorem 6.6.
6.1.2 The integral functional
In this section we prove the equality I
m
= I
m
. The proof is based on several facts about the
relations between sub-linear functions and measures: we propose here a slight revisited form of the
results proved by Goman and Serrin in [49].
Fixed m N, let us consider T : R
m
[0, ] and the following sets of hypotheses:
T is proper, l.s.c. and sub-linear; (6.13)
T is positively homogeneous of degree 1; (6.14)
there exists a constant C > 0 such that, for every R
m
, T() C[[. (6.15)
loc
(, R
m
), and set, for every B B(),
T
(B) = sup
_
_
_
r
j=1
T ((B
j
)) : j B
j
B(), B
j
,
r
_
j=1
B
j
B, B
i
B
j
= if i ,= j
_
_
_
.
(6.16)
Note that if T() = [[ then T
agrees with [[, the total variation of given by (2.1).

Proposition 6.9. Let T : R
m
[0, ] be a function satisfying (6.13). Then, for every
/
loc
(, R
m
), T
is a positive Borel measure. Moreover if T satises (6.15) too, T
/
+
().
Proof. Let us note at rst that if B
1
, B
2
B() and B
1
B
2
then 0 T
(B
1
) T
(B
2
) and that
T
() = 0. Let us consider now a sequence E

i
i=1
B() such that, for every i ,= j, E
i
E
j
= .
Then dened B =
i=1
E
i
let us show that T
(B) =
i=1
T
(E
i
). Indeed, since T 0, for every
q N,
T
(B)
q
i=1
T
(E
i
) and then T
(B)
i=1
T
(E
i
).
In order to prove the converse let B
j
r
j=1
B() as in (6.16). Then, for every j = 1, . . . , r,
(B
j
) =
i=1
(E
i
B
j
) and, by the sub-linearity and the lower semicontinuity of T,
r
j=1
T((B
j
)) =
r
j=1
T
_
lim
q
q
i=1
(E
i
B
j
)
_
j=1
liminf
q
T
_
q
i=1
(E
i
B
j
)
_
j=1
liminf
q
q
i=1
T ((E
i
B
j
)) =
r
j=1
i=1
T((E
i
B
j
)) =
i=1
r
j=1
T((E
i
B
j
))
i=1
T
(E
i
),
and, taking now the supremum of the left hand side with respect to the families B
j
r
j=1
B()
as in (6.16), we nd the opposite inequality.
Lemma 6.10. Let T : R
m
[0, ] be a function satisfying (6.13). If , /
loc
(, R
m
) and
= + then T
+T
, equality holding when .

Proof. Let us x B B(). Let us suppose at rst that T
(B) = . Then, for every M N, we

can nd B
j
r
j=1
as in (6.16) such that
M
r
j=1
T((B
j
))
r
j=1
T((B
j
)) +T((B
j
)) T
(B) +T
(B),
then T
(B) + T
(B) = . If instead T
(B) < , for every > 0, there exists B

j
r
j=1
as in
(6.16) such that
T
(B)
r
j=1
T((B
j
)) +.
Then
T
(B)
r
j=1
T((B
j
)) +
r
j=1
T
_
(B
j
) +(B
j
)
_
+
j=1
T((B
j
)) +
r
j=1
T((B
j
)) + T
(B) +T
(B) +,
and, letting 0, we nd the wanted inequality. If now we know that we can nd two
disjoint sets B
, B
B(), such that is concentrated on B
and is concentrated on B
: thus
T
(B) = T
(B B
) and T
(B) = T
(B B
),
and then, since T
is a measure,
T
(B) +T
(B) = T
(B B
) +T
(B B
) = T
(B).
This completes the proof.
Theorem 6.11. Let T : R
m
[0, ] be a function satisfying (6.13) and (6.14), /
loc
(, R
m
)
and u L
1
loc,
(), u 0. Then we have, for every B B(), B ,
T
__
B
u(x)d(x)
_
_
B
u(x)dT
(x).
Proof. Let B B(), B (so that
_
B
u(x)d[[(x) < ) and let us suppose at rst that u is
a simple function on B, that is
u(x) =
r
j=1
c
j
1
Bj
(x),
where c
j
0 and B
j
r
j=1
is a partition of B (that obviously is admissible in (6.16)). Then by the
sub-linearity and the positively homogeneity of degree 1 of T,
T
__
B
u(x)d(x)
_
= T
_
_
r
j=1
c
j
(B
j
)
_
_
j=1
c
j
T((B
j
))
r
j=1
c
j
T
(B
j
) =
_
B
u(x)dT
(x).
If now u is not simple then there exists an increasing sequence u
h
h=1
of simple positive functions
on B such that, in particular,
_
B
u
h
(x)d(x)
_
B
u(x)d(x). Then, for every h N,
T
__
B
u
h
(x)d(x)
_
_
B
u
h
(x)dT
(x)
and, passing to the limit for h , we end by the lower semicontinuity of T (used to treat the
left hand side) and by Beppo Levis Theorem (used to treat the right hand side).
m
[0, ] be a function satisfying (6.13) and (6.14) and let
h

/
loc
(
h
, R
m
), /
loc
(, R
m
) such that
h
in w
-/
loc
( , R
m
). Then
T
() liminf
h
T
h
(
h
).
Proof. Let us consider B B(), B and > 0 and nd B
j
r
j=1
T
(B)
r
j=1
T((B
j
)) +.
Fixed k N we can nd, for every j 1, . . . , r, a compact set F
k
j
and an open set G
k
j
such that
F
k
j
B
j
, [[(B
j
F
k
j
) <
1
k
, and F
k
j
G
k
j
, [[(G
k
j
F
k
j
) <
1
k
.
We can also suppose that G
k
j
r
j=1
are disjoint subsets well contained in . Let us consider now,
for every j 1, . . . , r, a function
k
j
C
c
(G
k
j
) such that 0
k
j
1 and
k
j
1 on F
k
j
. Then
(B
j
)
_
k
j
(x)d(x)
(B
j
) (F
k
j
)
(F
k
j
)
_
k
j
(x)d(x)
1
k
+
_
G
k
j
\F
k
j
k
j
(x)d(x)
2
k
.
Then, by considering the lower semicontinuity of T there exists k
N such that
T((B
j
)) T
__
j
(x)d(x)
_
+

r
,
and then
T
(B)
r
j=1
T
__
j
(x)d(x)
_
+ 2.
Now there exists h
, by the convergence of the measures

h
, the
lower semicontinuity of T and Theorem 6.11, we have
T
(B)
r
j=1
T
__
j
(x)d
h
(x)
_
+ 3
r
j=1
_
j
(x)dT
h
(x) + 3
=
_
h
r
j=1
j
(x)dT
h
(x) + 3 T
h
(
h
) + 3.
Then
T
(B) liminf
h
T
h
(
h
) + 3,
and we end the proof taking the supremum on B .
m
[0, ] be a function satisfying (6.13) and (6.15). Then T is
C-Lipschitz on R
m
and T(0) = 0.
Proof. Clearly (6.15) implies that T(0) = 0. In order to prove the Lipschitz inequality let us
consider , R
m
: we have
T() = T( + ) T() +T( ) T() +C[ [,
that is T() T() C[ [. Since the inequality holds for every , R
m
, it implies the
wanted Lipschitz inequality.
m
[0, ] be a function satisfying (6.13) and (6.15) and let
1
,
2

/
loc
(, R
m
). Then, for every B B(), B ,
[T
1
(B) T
2
(B)[ C[
1

2
[(B).
Proof. Let us consider B B(), B (so that T
1
(B), T
1
(B) < ) and > 0. We can nd
(considering suitable renements) B
j
r
j=1
T
1
(B)
r
j=1
T(
1
(B
j
)) + and T
2
(B)
r
j=1
T(
2
(B
j
)) +.
Then, by Lemma 6.13,
[T
2
(B) T
1
(B)[
j=1
T(
1
(B
j
))
r
j=1
T(
2
(B
j
)) + 2
C
r
j=1
[
1
(B
j
)
2
(B
j
)[ + 2 C[
1

2
[(B) + 2,
and we end by letting 0.
m
[0, ] be a function satisfying (6.13), (6.14) and (6.15),
/
+
() and /
loc
(, R
m
). Then, for every B B(),
T
(B) =
_
B
T(
a
(x))d(x) +T
s (B),
where has been decomposed as in (2.3) with respect to .
Proof. By Lemma 6.10, T
= T
+ T
s. Thus we have to prove that, for every B B(),

T
(B) =
_
B
T(
a
(x))d(x). Let us suppose at rst that B B(), B (so that T
(B) <
) and that
a
is a simple function on B, that is
a
(x) =
k
i=1
c
i
1
E
i
(x),
where E
i
k
i=1
B() is a partition of B and c
i
k
i=1
R
m
. Let us x > 0 and nd B
j
r
j=1
as
in (6.16) such that
T
(B)
r
j=1
T
_
(
a
)(B
j
)
_
+.
Then
r
j=1
k
i=1
T
_
(
a
)(E
i
B
j
)
_
T
(B)
r
j=1
k
i=1
T
_
(
a
)(E
i
B
j
)
_
+. (6.17)
But, by (6.14), we have also
r
j=1
k
i=1
T
_
(
a
)(E
i
B
j
)
_
=
r
j=1
k
i=1
T
_
c
i
(E
i
B
j
)
_
=
r
j=1
k
i=1
T(c
i
)(E
i
B
j
)
=
k
i=1
T(c
i
)(E
i
) =
_
B
T(
a
(x))d(x),
and since in (6.17) is arbitrary we end.
If now
a
is not simple let us consider a sequence
a
h
h=1
of simple functions on B such that
a
h

a
in L
1
(B, R
m
). Then, if we dene
h
=
a
h
/(B, R
m
), we have, by Lemma 6.13,
_
B
T(
a
(x)) T(
a
h
(x))d(x)
C
_
B
[
a
(x)
a
h
(x)[d(x),
and, by Lemma 6.14,
[T
(B) T
h
(B)[ C
_
E
[
a
(x)
a
h
(x)[d(x).
Since the right hand side of both the inequalities go to zero as h and since, for every h N,
T
h
(B) =
_
B
T(
a
h
(x))d(x), it follows that also T
(B) =
_
B
T(
a
(x))d(x).
If now B is not well contained in we consider an increasing sequence B
i
i=1
B() with
B
i
and B
i
and we end by approximation using the property of the integrals and of the
elements of /
+
().
Note that applying two times Theorem 6.15 we nd in particular
T
(B) =
_
B
T(
a
(x))d(x) +
_
B
T
_
d
s
d[
s
[
(x)
_
d[
s
[(x).
After these preliminary results on the sub-linear functionals we have the right tools to prove
the equality between I
m
and I
m
. Theorem 6.16, together with Theorem 6.6 and Propositions 6.7
and 6.8, proves the part of Theorem 6.1 concerning the functionals I
m
, I
m
and I
m
.
m
[0, ] be a proper, l.s.c. and convex function. Then, for every
/
loc
(, R
m
), I
m
(, ) = I
m
(, ).
Proof. Let us x /
+
(). Given /
loc
(, R
m
) let us consider (, ) /
loc
(, R
m+1
)
and dene the measure

f
(,)
where

f : R
m+1
[0, ] is the function given by Proposition 2.25
that satises (6.13) and (6.14). We claim at rst that, for every /
loc
(, R
m
), B B(),
f
(,)
(B) =
_
B
f(
a
(x))d +f
s (B), (6.18)
where f
s is well dened since f
is l.s.c. and sub-linear (see Proposition 2.20).

Clearly we can write (, ) = (
a
, 1) + (
s
, 0) and, by Lemma 6.10,
f
(,)
=

f
(
a
,1)
+

f
(
s
,0)
=

f
(
a
,1)
+f
s .
Thus it remains to show that,
f
(
a
,1)
(B) =
_
B
f(
a
(x))d.
It is well known that there exists an increasing sequence f
h
h=1
of convex functions such that
f
h
f, [f
h
()[ C
h
_
1 +[[
2
and, for every [[ h, f
h
() = f(). Then, for every 0,
f
h
(, ) C
h
[[ C
h
[(, )[, that is,

f
h
satises also (6.15) on R
m
[0, ]: since the (m + 1)-th
component of (
a
, 1) is a positive measure it is simply to verify that also for these functions
Theorem 6.15 holds
2
, that is,
f
h,(
a
,1)
(B) =
_
B
f
h
(
a
(x), 1)d =
_
B
f
h
(
a
(x))d.
2
Indeed, one could develop all the results till now proved considering T : R
m
[0, ] [0, ] and
/
loc
(, R
m+1
) with
m+1
/
+
().
By Beppo Levis Theorem we obtain
lim
h
_
B
f
h
(
a
(x))d =
_
B
f(
a
(x))d
We end if we prove that also
lim
h
f
h,(
a
,1)
(B) =

f
(
a
,1)
(B).
Suppose at rst that

f
(
a
,1)
(B) < . Fixed > 0, there exists B
j
r
j=1
B(), B
j
B, such
that
f
(
a
,1)
(B)
r
j=1
f
_
(
a
)(B
j
), (B
j
)
_
+.
Since
a
<< we have that (B
j
) = 0 implies (
a
)(B
j
) = 0. Thus we claim that there exists
h
such that, for every h h
, j 1, . . . , r,
f
_
(
a
)(B
j
), (B
j
)
_
=

f
h
_
(
a
)(B
j
), (B
j
)
_
.
Indeed, for every [[ h,

f
h
(, ) =

f(, ) and then, if h is big enough we have, for every
j = 1, . . . , r, [
a
(B
j
)[ h[(B
j
)[ and the claim follows. Therefore
f
(
a
,1)
()
r
j=1
f
_
(
a
)(B
j
), (B
j
)
_
+
j=1
f
h
_
(
a
)(B
j
), (B
j
)
_
+

f
h,(
a
,1)
() +,
that implies, as h and later 0,
f
(
a
,1)
() liminf
h
f
h,(
a
,1)
().
The other inequality is obvious since, for every (, ) R
m+1
,

f
h
(, )

f(, ). The case
f
(
a
,1)
(E) = can be proved using a similar argument so that (6.18) is achieved.
loc
(, R
m
), = L
n
and =
a
L
n
+
s
: applying the claim just
proved both to f and to f
(with = [
s
[) we obtain in particular
f
(,L
n
)
() =
_
f(
a
(x))dx +f
s () =
_
f(
a
(x))dx +
_
_

s
[
s
[
(x)
_
d[
s
[(x) = I
m
(, ).
We end, by means of Theorem 6.6, proving that also
f
(,L
n
)
() = lim
0
_
f(
(x))dx.
Indeed, by Theorem 6.12 (that can be applied to

f too) and the claim proved we have
f
(,L
n
)
() liminf
0
f
(,1)L
n(
) = liminf
0
f
L
n(
) = liminf
0
_
f(
(x))dx.
On the other hand, using Theorem 6.11,
_
f(
(x))dx =
_
f(
(x), 1)dx =
_
f
_
_
B(x,)
k
(x y)d(, L
n
)(y)
_
dx
_
_
B(x,)
k
(x y)d
f
(,L
n
)
(y)
_
dx
_
_
_
B(y,)
(x y)dx
_
d
f
(,L
n
)
(y)

f
(,L
n
)
(),
then
limsup
0
_
f(
(x))dx

f
(,L
n
)
(),
and the proof is nally achieved.
6.1.3 The supremal functional
We prove now the equality S
m
= S
m
. As for the integral case Theorem 6.17, together with Theorem
6.6 and Propositions 6.7 and 6.8, proves the part of Theorem 6.1 involving S
m
, S
m
and S
m
.
m
[0, ] be a proper, l.s.c. and level convex function. Then, for every
/
loc
(, R
m
), S
m
(, ) = S
m
(, ).
Proof. Let us suppose at rst that g is positively homogeneous of degree 0.
Let us x /
loc
(, R
m
) and show at rst that S
m
(, ) S
m
(, ). By Theorem 2.2(i)
there exists N such that L
n
(N) = 0 and for every x N,
(x)
a
(x) as 0. Then,
using Theorem 6.6, for every x N,
g(
a
(x)) liminf
0
g(
(x)) liminf
0
sup
x
g(
(x)) = S
m
(, ),
and then
S
m
(, ) = ess sup
x
g(
a
(x)) sup
x\N
g(
a
(x)) S
m
(, ).
Thus it remains to show that
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x)
_
S
m
(, ).
By Theorem 2.2(ii) there exists M with [
s
[(M) = 0 and such that, for every x M,
there exists a sequence
h
h=1
, depending on x and decreasing to zero, such that,
lim
h
d
s
d[
s
[
(x)

h
(x)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
= 0. (6.19)
Then, using Proposition 2.21 and Theorem 6.6, for every x M,
g
_
d
s
d[
s
[
(x)
_
= g
_
d
s
d[
s
[
(x)
_
liminf
h
g
_

h
(x)
_
B(x,
h
)
k
h
(x y)d[
s
[(y)
_
= liminf
h
g (
h
(x)) lim
h
S
m
_
h
L
n
,
h
_
= S
m
(, ).
In conclusion,
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x)
_
sup
x\M
g
_
d
s
d[
s
[
(x)
_
S
m
(, ),
and we achieve S
m
(, ) S
m
(, ).
In order to prove the converse inequality let us x > 0. By Theorem 2.19 and Propositions
2.17 and 2.21 we have, for every x
,
g(
(x)) = g
_
_
B(x,)
k
(x y)d(y)
_
= g
_
_
B(x,)
k
(x y)
a
(y)dy +
_
B(x,)
k
(x y)
d
s
d[
s
[
(y)d[
s
[(y)
_
g
_
_
B(x,)
k
(x y)
a
(y)dy
_
g
_
_
B(x,)
k
(x y)
d
s
d[
s
[
(y)d[
s
[(y)
_
_
ess sup
yB(x,)
g(
a
(y))
_
_
[
s
[- ess sup
yB(x,)
g
_
d
s
d[
s
[
(y)
_
B(x,)
k
(x y)d[
s
[(y)
__
_
ess sup
x
g(
a
(x))
_
_
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x)
__
= S
m
(, ),
thus,
sup
x
g(
(x)) S
m
(, ).
The wanted inequality S
m
(, ) S
m
(, ) is achieved as 0, invoking again Theorem 6.6.
In the general case, let us consider the function g dened in (2.15): by Proposition 2.26 we know
that g is l.s.c., positively homogeneous of degree 0 and level convex. Let now x /
loc
(, R
m
)
and consider (, L
n
) /(, R
m+1
): its decomposition with respect to L
n
is clearly given by
(, L
n
) = (
a
(x), 1) L
n
+
_
d
s
d[
s
[
(x), 0
_
[
s
[.
We have also that (, L
n
)
R
m+1
is given by (, L
n
)
(x) = (
(x), 1). Then applying the

rst step to g and (, L
n
) we have
m
_
(, L
n
),
_
=

S
m
_
(, L
n
),
_
,
where

S
m
and

S
m
denotes the functionals (6.6) and (6.2) built up considering g and dened on
the space /
loc
(, R
m+1
). But now we have:
m
_
(, L
n
),
_
= lim
0
sup
x
g
_
(, L
n
)
(x)
_
= lim
0
sup
x
g(
(x), 1) = lim
0
sup
x
g(
(x)) = S
m
(, ),
and, again by Proposition 2.21,
S
m
_
(, L
n
),
_
=
_
ess sup
x
g(
a
(x), 1)
_
_
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x), 0
__
=
_
ess sup
x
g(
a
(x))
_
_
[
s
[-ess sup
x
g
_
d
s
d[
s
[
(x)
__
= S
m
(, ).
This achieves the proof.
6.2 Non level convex functionals
In this section we propose the proof of the Theorems 6.3 and 6.4 related to the functional S
m
in
(6.10).
6.2. NON LEVEL CONVEX FUNCTIONALS 87
6.2.1 Necessary conditions
Proof of Theorem 6.3. The fact that g is l.s.c. and level convex can be proved using the same
argument of the implication (ii) (i) of Theorem 6.1. In order to prove the other parts of the
theorem let us consider
j
j=1
R
m
such that g(
j
) l = infg() : R
m
.
Let us study the function . Let x
0
R
n
and
h
,
0
R
m
0 such that
h

0
and, xed
j N, let
h
=
h

x
0
+
j
L
n
, and
0
=
0

x
0
+
j
L
n
.
Then
h

0
in w
-/
loc
(R
n
, R
m
) and
(
0
) g(
j
) = S
m
(
0
, R
n
) liminf
h
S
m
(
h
, R
n
) = liminf
h
((
h
) g(
j
)).
This implies that, for every j N, g(
j
) is l.s.c. on R
m
since also the lower semicontinuity in
= 0 trivially follows from (0) = 0: by using Proposition 2.32, we obtain that the same holds for
l.
Let , R
m
0 such that ,= and x
h
, x
0
R
n
, x
h
,= x
0
, such that x
h
x
0
. Moreover,
xed j N, set
h
=
x
h
+
x
0
+
j
L
n
, and
0
= ( +)
x
0
+
j
L
n
.
Since
h

0
in w
-/
loc
(R
n
, R
m
) we have
( +) g(
j
) = S
m
(
0
, R
n
) liminf
h
S
m
(
h
, R
n
)
= liminf
h
() () g(
j
) = () () g(
j
).
Then, since (0) = 0, we have that, for every j N, g(
j
) is sub-maximal on R
m
that implies,
once j , that l satises the same property too.
Let us prove now the inequality g
( l)
. Let
0
R
m
0 and let
h

0
, t
h
such
that g(t
h
h
) g
(
0
). Thus we have that there exists a sequence of positive numbers r
h
h=1
such that, for every h N, t
h
L
n
(B(0, r
h
)) = 1. Fixed j N and setting
h
= t
h
_
h

j
t
h
_
1
B(0,r
h
)
(x) L
n
+
j
L
n
, and
0
=
0

0
+
j
L
n
,
we have
h

0
in w
-/
loc
(R
n
, R
m
) and then
(
0
) g(
j
) = S
m
(
0
, R
n
) liminf
h
S
m
(
h
, R
n
) = lim
h
g(t
h
h
) g(
j
) = g
(
0
) g(
j
).
This fact proves, as j , that, for every
0
R
m
0, ( l)(
0
) g
(
0
). If now we consider
a sequence s
h
0 such that ( l)
(
0
) = lim
h
( l)(s
h
0
), since g
is positively homogeneous
of degree 0, we have
( l)
(
0
) = lim
h
( l)(s
h
0
) liminf
h
g
(s
h
0
) = g
(
0
),
and we nd the desired inequality.
We end proving g
( l)
. Let us consider
0
R
m
0 and let t
h
0, t
h
< 1, such that
(t
h
0
)
(
0
). Let us x = Q
n
= (0, 1)
n
and dene, for every k N,
G
k
=
_
x Q
n
: x
i
=
q
k + 1
, q 1, . . . , k, i 1, . . . , n
_
:
note that
3
#(G
k
) = k
n
. Fixed now M N, we claim that there exists an sequence k
h
h=1
N,
depending on M and such that k
h
and, for every h N, M t
h
k
n
h
< 2
n
M. Obviously this
happen if, for every h N, there exists k
h
N such that
_
Mt
1
h
_ 1
n
k
h
< 2
_
Mt
1
h
_ 1
n
,
and since
_
Mt
1
h
_ 1
n
> 1 and
_
Mt
1
h
_ 1
n
as h , k
h
can be found. Then, unless to extract
a (not relabelled) subsequence, we have t
h
k
n
h
s
M
[M, 2
n
M].
Fixed j N, set the following elements of /
loc
(Q
n
, R
m
)
h
=
xG
k
h
t
h
k
n
h
k
n
h
0

x
+
j
L
n
and
0
= s
M
0
L
n
+
j
L
n
.
It is a quite standard result that
h

0
in w
-/
loc
(Q
n
, R
n
), thus,
g(s
M
0
+
j
) = S
m
(
0
, Q
n
) liminf
h
S
m
(
h
, Q
n
) = liminf
h
[ (t
h
0
) g(
j
)] =
(
0
) g(
j
).
Then, letting j and using Proposition 2.33, we have
g
_
s
M
_
0
+

j
s
M
__
= g(s
M
0
+
j
)
(
0
) l = ( l)
(
0
),
and since this relation holds for every M N, taking the limit for M also s
M
and then
we obtain g
(
0
) ( l)
(
0
).
6.2.2 Sucient conditions
Here we prove a theorem in which sucient conditions for the lower semicontinuity of S
m
are
found: we treat directly the case in which g and depends on x too and we will deduce Theorem
6.4 as a corollary of our main result.
The following lemma is based on Lemma 3.6 in [18].
Lemma 6.18. Let k N and
/
k
= /(, R
m
) : #(spt()) k.
Then /
k
is sequentially closed with respect to the w
-/(, R
m
) topology. Moreover if l [0, )
and : R
m
[0, ] is a proper and Borel function such that l is l.s.c. on R
m
and,
for every x , (x, ) l is sub-maximal on R
m
and (x, 0) = 0, then the functional
(, ) =
xA
_
(x, (x)) l
_
,
is l.s.c on /
k
with respect to the w
-/(, R
m
) topology.
Proof. Let
h
h=1
/
k
,
h
in w
-/(, R
m
). Then, for every h N, there exist two
sequences x
h
i
k
i=1
(with, for every h N, x
h
i
,= x
h
j
if i ,= j) and
h
i

k
i=1
R
m
such that
h
=
k
i=1
h
i

x
h
i
.
Clearly there exists M > 0 such that, for every i 1, . . . , k, h N, [
h
i
[ M. Following
Lemma 3.6 in [18], let I 1, . . . , k the (possibly empty) set of the index i such that x
h
i
h=1
3
If A is a set we denote with #(A) the cardinality of A.
admits a convergent subsequence. Then we can nd x
i
iI
,
i
iI
R
m
(note that it could
be
i
= 0) and a (not relabelled) subsequence of
h
h=1
such that, for every i I, x
h
i
x
i
and
h
i

i
for h . On the other hand, if i , I it is easy to verify that
h
i

x
h
i
0 in
w
-/(, R
m
). Therefore we have
=
iI
i

x
i
/
k
,
with = 0 if I = . Note that, in particular, it may happen that there are i, j I such that
x
i
= x
j
. The lower semicontinuity of is trivial if = 0 (since by denition (0, ) = l), while if
,= 0 it simply follows by using the lower semicontinuity and the sub-maximality of l, indeed
(, ) =
xA
_
(x, (x)) l
_
iI
_
(x
i
,
i
) l
_
liminf
h
iI
_
(x
h
i
,
h
i
) l
_
liminf
h
k
i=1
_
(x
h
i
,
h
i
) l
_
= liminf
h
(
h
, ),
and the proof is achieved.
Later we need a weak version of Theorem 4.1 proved by Acerbi, Buttazzo and Prinari [2].
m
[0, ] be a Borel function such that, for L
n
-a.e. x , g(x, )
is l.s.c. and level convex on R
m
. Then the functional
S(u, ) = ess sup
x
g(x, u(x))
is l.s.c. on L
(, R
m
-L
(, R
m
) convergence.
We can now prove the main result of the section. Note that, when we are dealing with g :
R
m
[0, ], g
(x, ) means (g(x, ))
().
m
[0, ] be a proper and Borel function such that, for L
n
-a.e.
x , g(x, ) is l.s.c. and level convex on R
m
and there exists a function
: [0, ) [0, )
such that, for every x and R
m
,
g(x, )
([[) and lim

t
(t) = .
Moreover, setting l = infg(x, ) : (x, ) R
m
, let : R
m
[0, ] be a proper and
Borel function such that l is l.s.c. on R
m
, for every x , (x, ) l is sub-maximal on
R
m
and (x, 0) = 0, and there exists a function
0
: (0, ) [0, ) such that, for every x ,
R
m
0,
(x, ) l
0
([[) and lim
t0
0
(t) = .
Then the functional
S
m
(, ) =
_
ess sup
x
g(x,
a
(x))
_
_
[
c
[-ess sup
x
g
_
x,
d
c
d[
c
[
(x)
__
_

xA
(x,
#
(x))
_
is l.s.c. on /
loc
(, R
m
-/
loc
(, R
m
) convergence.
Proof. Let us suppose at rst that R
n
is bounded and let us prove the lower semicontinuity
of S
m
on /(, R
m
-/(, R
m
) convergence. Consider ,
h
/(, R
m
)
such that
h
in w
-/(, R
m
): then there exists L > 0 such that [
h
[(), [[() L. Without
loss of generality we can suppose
liminf
h
S
m
(
h
, ) = lim
h
S
m
(
h
, ),
and that there exists a constant M > 0 such that, for every h N, S
m
(
h
, ) M. This fact
implies that there exists M
> 0 such that, for every h N,

(i)
c
h
= 0,
(ii) for L
n
-a.e. x , [
a
h
(x)[ M
,
(iii) for every x A
h
, [
#
h
(x)[
1
M
.
By (ii) we obtain
a
h
(x)
h=1
L
(, R
m
) and
sup
_
ess sup
x
[
a
h
(x)[ : h N
_
< .
Then by Proposition 2.7 there exists a subsequence (not relabelled) and a function u L
(, R
m
),
such that
a
h
(x) u(x) in w
-L
(, R
m
). (6.20)
Moreover, since, for every h N, [
h
[() L, we have also [
#
h
[() L thus, writing
#
h
=
xA
#
h
(x)
x
, it follows, for every h N,
L
#
h
() =
xA
#
h
(x)

1
M
#(A
h
) .
Then there exists k N such that, for every h N, #(A
h
) k. Then, for every h N,
#
h
/
k
and, as said, [
#
h
[() L. Thus, by Theorem 2.1, there exists a subsequence (not relabelled) and
a measure /(, R
m
) such that
#
h
in w
-/(, R
m
). (6.21)
By Lemma 6.18 we know /
k
.
We claim now that the condition
h
in w
-/(, R
m
) together with (6.20) and (6.21) gives
that = u L
n
+, that is
a
= u and
#
= . Indeed, for every C
0
(),
_
(x)d
h
(x) =
_
(x)
a
h
(x)dx +
_
(x)d
#
h
(x),
and, by (6.20) (noting that C
0
() L
1
() since L
n
() < ) and (6.21) we have that
h

u L
n
+ in w
-/(, R
m
). The uniqueness of the limit allows to achieve the claim.
Now, by Theorem 6.19, we have
ess sup
x
g(x,
a
(x)) liminf
h
ess sup
x
g(x,
a
h
(x)),
while, by Lemma 6.18, we have
xA
_
(x,
#
(x)) l
_
liminf
h
xA
h
_
(x,
#
h
(x)) l
_
.
Then
liminf
h
S
m
(
h
, ) = liminf
h
_
ess sup
x
g(x,
a
h
(x))
_
_
_

xA
h
_
(x,
#
h
(x)) l
_
_
_
_
ess sup
x
g(x,
a
(x))
_
_

xA
_
(x,
#
(x)) l
_
_
= S
m
(, ),
and we get the lower semicontinuity.
In the general case let us consider ,
h
/
loc
(, R
m
) such that
h
in w
-/
loc
(, R
m
).
Then, for every open set
,
h
in w
-/(
, R
m
) and then
S
m
(,
) liminf
h
S
m
(
h
,
) liminf
h
S
m
(
h
, ).
At least, since
supS
m
(,
) :
= S
m
(, ),
the proof is nally achieved.
The proof of Theorem 6.4 stated at the beginning of this chapter easily follows from Theorem
6.20 and Propositions 2.23 and 2.24.
Chapter 7
Applications to BV
In this chapter we present some consequences of the results proved in Chapter 6 when they are
applied to the setting of the functions of bounded variation.
7.1 Convex and level convex functionals
7.1.1 Lower semicontinuity and relaxation
The following theorem easily follows by Theorem 6.1 and it provides the precise statement of what
we announced in the introduction about the extensions of integral and supremal functionals on
BV
loc
(). The integral part of this theorem was proved in the paper of Goman and Serrin (see
[49] Theorem 5) while the supremal part was proved by Gori in (see [51] Theorem 5).
n
[0, +] be a proper, l.s.c and convex function and g : R
n
[0, +]
be a proper, l.s.c. and level convex function. Let us consider the following functionals dened, for
every u BV
loc
(), as
I(u, ) =
_
f(u(x))dx +
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x), (7.1)
and
S(u, ) =
_
ess sup
x
g(u(x))
_
_
[D
s
u[-ess sup
x
g
_
dD
s
u
d[D
s
u[
(x)
__
. (7.2)
Then I and S are lower semicontinuous on BV
loc
() with respect to the w
-BV
loc
() convergence
and, for every u BV
loc
(),
I(u, ) = I
(u, ) = lim
0
I(u
) and S(u, ) = S
(u, ) = lim
0
S(u
), (7.3)
where the functionals I
and S
are given by (1.7) and (1.8), while u
denotes the convolution of

u.
Proof. Since, when u
h
, u BV
loc
() and u
h
u in w
-BV
loc
(), we have Du
h
Du in w
-
/
loc
(, R
n
) and since, for every u BV
loc
(), u
u in w
-BV
loc
( ) (and then u
L
n
=
(Du)
Du in w
-/
loc
( , R
n
)) we end applying directly Theorem 6.1.
A simple consequence of the equalities (7.3) is that, in particular, for every u BV ()
1
,
I(u, ) R[w
-BV ] (I) (u, ) and S(u, ) R[w
-BV ] (S) (u, ).

1
Here we refer to the notations given in the Introduction (see in particular (1.5)) and we consider I and S
dened on W
1,1
() that, as known, is a subset of BV () dense with respect to the w
-BV convergence (so as

C
() W
1,1
(); see [7] Theorem 3.9).
93
94 CHAPTER 7. APPLICATIONS TO BV
The propositions below show two particular cases in which also the opposite inequalities hold, and
thus I and S are just the relaxed functionals of I and S. Compare Proposition 7.2 with Theorem
2 by Serrin [68].
In the following, by I
|C
()
and S
|C
()
we mean the functionals I and S restricted to the
space C
(), that is the set of the innitely derivable functions with every derivative uniformly
continuous on : this space, in the cases considered here, is dense on BV () with respect to the
w
-BV () convergence.
Proposition 7.2. Let R
n
be a bounded open set with Lipschitz boundary and f : R
n
[0, )
be a convex function such that, for every R
n
,
f() c[[ +d,
where c, d > 0. Then, for every u BV (),
I(u, ) = R[w
-BV ]
_
I
|C
()
_
(u, ),
thus, in particular, I(u, ) = R[w
-BV ] (I) (u, ).

Proof. Let us x u BV () and
R
n
: then we can consider u BV (
) an extension
of u on
(see [40] Theorem 1, page 183). Clearly, if 0 < <

0
small enough, u
(),
u
u in w
-BV () (see [7] Lemma 3.24) and u
= u
on
. We claim now that

lim
0
[I( u
, ) I(u
)[ = 0.
Indeed
[I( u
, ) I(u
)[
_
\
f( u
(x))dx c
_
\
[ u
(x)[dx +L
n
(
).
By Theorem 2.2(b) in [7], we know that
_
\
[ u
(x)[dx [D u[
_
(
) +B(0, )
_
, (7.4)
where (
)+B(0, ) = x
: d(x,
) < . Since [D u[(
) < and (
)+B(0, )
as 0 then the left hand side of (7.4) tends to zero and the claim is achieved. Thus
I(u, ) = lim
0
I(u
) = lim
0
I( u
, ) R[w
-BV ]
_
I
|C
()
_
(u, )
and we end the proof being the opposite inequality satised.
Proposition 7.3. Let = B(0, 1) R
n
and g : R
n
[0, ] be a proper, l.s.c. and level convex
function such that g(0) = 0. Then, for every u BV (),
S(u, ) = R[w
-BV ]
_
S
|C
()
_
(u, ),
thus, in particular, S(u, ) = R[w
-BV ] (S) (u, ).

Proof. Let us x u BV (), consider the convolution u
: B(0, 1) R (with <

1
4
) and set,
for every x B(0, 1), u
(x) = u
((1 2)x): clearly u
(B(0, 1)).
We claim at rst that u
u in w
-BV (B(0, 1)). Indeed, we simply have that, by means of the

properties of the convolutions and the change of variables formula, there exists a constant C > 0
such that, for every 0 < <
1
4
,
_
B(0,1)
[ u
(x)[dx +
_
B(0,1)
[ u
(x)[dx
1
(1 2)
n
_
B(0,1)
[u(x)[dx +
1
(1 2)
n1
[Du[ (B(0, 1)) dx C.
Thus, by Theorem 2.8 there exists v BV (B(0, 1)) such that u
v in w
-BV (B(0, 1)). We prove

that, for L
n
-a.e. x B(0, 1), u(x) = v(x) showing that, for L
n
-a.e. x B(0, 1), u
(x) u(x).
Indeed, if M = supk(x) : x R
n
we have
[ u
(x) u(x)[ =
_
B((12)x,)
k
((1 2)x y)u(y)dy u(x)
_
B((12)x,)
k
((1 2)x y)[u(y) u(x)[dy

M
n
_
B(x,3)B(0,1)
[u(y) u(x)[dy.
Now if is small enough (that is 3 1 [x[), B(x, 3) B(0, 1) = B(x, 3) and then
[ u
(x) u(x)[
3
n
M
(3)
n
_
B(x,3)
[u(y) u(x)[dy,
that, for L
n
-a.e. x B(0, 1), goes to zero as 0 by the Lebesgues points Theorem.
Once the claim is proved, applying Proposition 2.29 to g, we have, for every x B(0, 1),
g ( u
(x)) = g
_
(1 2)u
((1 2)x)
_
g
_
u
((1 2)x)
_
sup
xB(0,1)
g (u
(x)) ,
and then
sup
xB(0,1)
g ( u
(x)) sup
xB(0,1)
g (u
(x)) .
This allows to achieve the thesis since
S(u, ) = lim
0
S(u
) liminf
0
S( u
, ) R[w
-BV ]
_
S
|C
()
_
(u, )
and being the opposite inequality satised.
Remark 7.4. It is quite simple to understand that the proposition above holds also if we set
to be a bounded convex set. We have written down the proof for the unit ball in order to simplify
many technical details.
7.1.2 Dirichlet problem for supremal functionals
Following the work of Anzellotti, Buttazzo and Dal Maso [8] in which the same problem is consid-
ered for the integral functionals, we shall introduce a generalized Dirichlet problem for a particular
class of supremal functionals dened on BV (). The formulation we proposed is similar to the
one used in the nonparametric Plateaus problem (see [47]).
We have to underline that the functional (7.5) of Theorem 7.5 does not arise from the relaxation
on BV (), with respect to the w
-BV () convergence, of the functional (7.2) dened on C

1
()
(that is the space of the function belonging to C
1
() with xed boundary data ) but it is
simply suggested by the analogy with the integral setting. However, when the hypotheses of
Theorem 7.5 are considered, we suppose it is a good candidate to represent just the functional
R[w
-BV ()]
_
S
|C
1
()
_
.
In the following, given a function u BV (), we will always consider the restriction of u to
in the trace sense (see for instance [40] Theorem 1 page 177).
Theorem 7.5. Let be an open and bounded subset of R
n
with Lipschitz boundary, g : R
n
[0, ]
be a proper, l.s.c. and level convex function and L
1
H
n1
(). Let us consider, for every
u BV (), the functional
S
(u, ) = S(u, )
_
H
n1
-ess sup
x
g
_
((x) u(x))(x)
_
_
, (7.5)
where S is given by (7.2) and : S
n1
is the vector eld of the outer normal vectors to
2
.
Assume that there exist an open and bounded subset of R
n
with Lipschitz boundary

and w W
1,1
(
) such that = w
|
and
ess sup
x
\
g(w(x)) = infg() : R
n
. (7.6)
Then S
is l.s.c. on BV () with respect to the w
-BV () convergence.
Moreover if s = supg() : R
n
[0, ] and K
g
= R
n
: g
() < s is such that cl(K

g
)
does not contain any straight line, then the problem
minS
(u, ) : u BV () , (7.7)
admits at least a solution.
Proof. Let u, u
h
BV () such that u
h
u in w
-BV () and dene

u(x) =
_
u(x) if x ,
w(x) if x
,
and u
h
(x) =
_
u
h
(x) if x ,
w(x) if x
.
It is easy to see that u, u
h
BV (
) and u
h
u in w
-BV (
). Moreover, since
D u = ( u) H
n1
and D u
h
= ( u
h
) H
n1
(see [47] and [8] Theorem 3.1), we have also

S
(u, ) = S( u,
) and S
(u
h
, ) = S( u
h
,
).
Then, by means of Theorem 7.1, it is
S
(u, ) = S( u,
) liminf
h
S( u
h
,
) = liminf
h
S
(u
h
, )
and the rst part of the theorem is proved.
Before proving the second part of the theorem we need a remark. Let : [0, s] [0, ] be
a continuous and strictly increasing function: it is clear that u BV () is a minimum point of
S
(u, ) if and only if it is a minimum point of (S
(u, )). However, a simple computation gives

that, for every u BV (),
(S
(u, )) =
_
ess sup
x
( g)(u(x))
_
_
[D
s
u[-ess sup
x
( g
)
_
dD
s
u
d[D
s
u[
(x)
__
_
H
n1
-ess sup
x
( g
)
_
((x) u(x))(x)
_
_
.
Thus, by Proposition 2.30, we have that (S
(u, )) is in fact the functional (7.5) related to

g : R
n
[0, ]. Then, invoking now Proposition 2.31, we can suppose, without loss of
2
This vector eld can be dened since is Lipschitz so that Rademachers theorem can be applied.
generality, that g is demi-coercive, that is, there exist a > 0, b 0 and R
n
such that, for every
R
n
,
g() a[[ , ) b.
If, for every u BV (), S
(u, ) = there is nothing to prove. Thus let us consider any

u BV () such that S
(u, ) < : we claim that, setting c =

1
3L
n
()

1
3H
n1
()
,
S
(u, ) (7.8)
c
__
g(u(x))dx +
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x) +
_
_
((x) u(x))(x)
_
H
n1
(x)
_
.
Indeed, clearly it holds
ess sup
x
g(u(x))
1
L
n
()
_
g(u(x))dx. (7.9)
Moreover since
[D
s
u[-ess sup
x
g
_
dD
s
u
d[D
s
u[
(x)
_
< ,
it holds, for [D
s
u[-a.e. x , g
_
dD
s
u
d|D
s
u|
(x)
_
< and then, by Proposition 2.28(ii), we have also,
for [D
s
u[-a.e. x , g
_
dD
s
u
d|D
s
u|
(x)
_
= 0: then, in particular,
1
L
n
()
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x) = 0. (7.10)
At last using Proposition 2.28(i) we have
H
n1
-ess sup
x
g
_
((x) u(x))(x)
_
1
H
n1
()
_
_
((x) u(x))(x)
_
dH
n1
(x). (7.11)
Putting together (7.9), (7.10) and (7.11) we obtain, for every u BV () such that S
(u, ) < ,
S
(u, )
1
L
n
()
_
g(u(x))dx
1
L
n
()
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x)
1
H
n1
()
_
_
((x) u(x))(x)
_
H
n1
(x)
c
__
g(u(x))dx +
_
_
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x) +
_
_
((x) u(x))(x)
_
dH
n1
(x)
_
,
and (7.8) is proved.
By means of (7.8) and Proposition 2.28(iii)
3
we nd
S
(u, ) ac
_
[u(x)[dx c
_
, u(x))dx bcL
n
() +ac[D
s
u[()+
c
_
_
,
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x) +ac
_
[(x) u(x)[dH
n1
(x)+
c
_
, ((x) u(x))(x))dH
n1
(x),
3
Compare with Theorems 2.7 and 3.2 of [8].
and then, using the fact that
4
, for every u BV (),
_
, u(x))dx +
_
_
,
dD
s
u
d[D
s
u[
(x)
_
d[D
s
u[(x) =
_
, u(x)(x))dH
n1
(x)
and that there exists
5
a constant A = A() > 0 such that, for every u BV (),
_
[u(x)[dx A
_
[Du[() +
_
[u(x)[dH
n1
(x)
_
,
it follows
S
(u, )
ac
2
_
1
1
A
__
[Du[() +
_
[u(x)[dx +
_
[u(x)[dH
n1
(x)
_
dc, (7.12)
where
d =
_
[(x)[dH
n1
(x) +
_
, (x)(x))dH
n1
(x).
If now we consider a minimizing sequence u
h
h=1
such that, for every h N, S
(u
h
, ) <
we can prove its compactness only noting that (7.12) allows just to apply Theorem 2.8: this fact,
together with the lower semicontinuity of the functional S
, guarantees the existence of a minimum

for (7.7).
Remark 7.6. In the hypotheses of Theorem 7.5, if there exists u BV () such that S
(u, ) < ,
then the same holds for every solution u of (7.7) too. Then, for H
n1
-a.e. x ,
((x) u(x))(x) K
g
= R
n
: g
() < .
Since g
is level convex and positively homogeneous of degree 0 (see Proposition 2.20), K

g
is clearly
a convex cone and then it follows that, for H
n1
-a.e. x x : (x) , K
g
, (x) u(x) 0
and, for H
n1
-a.e. x x : (x) , K
g
, (x) u(x) 0.
Therefore we conclude that, for H
n1
-a.e. x x : (x) , K
g
(K
g
), (x) = u(x).
It is also clear that, arguing as in the proof of Theorem 7.5, we cannot avoid the presence of
the condition (7.6). Nevertheless we can easily nd suitable classes of examples in which (7.6) is
satised.
Consider for instance the case in which = B(0, 1) R
2
and g : R
2
[0, ) is dened as
g(
1
,
2
) = (
1
0) +(
2
0). If is the restriction to S
1
of a function w C
1
(B(0, 1 +)) (where
> 0) such that, for every x B(0, 1 +) B(0, 1), w(x) (
1
,
2
) :
1
0,
2
0, then (7.6)
holds.
7.2 Non level convex functionals
7.2.1 A minimum problem
In this section we nd a simple application of Theorem 6.20 to the BV setting where obviously
only the one dimensional case is considered: the result here presented should be compared with
the work of Alicandro, Braides and Cicalese [3] in which an analogous problem is deeply analyzed.
For this let (a, b) R, g, w, : (a, b) R [0, ] be proper Borel functions. We want to
minimize the following functional
L(u, K) =
_
ess sup
x(a,b)\K
g(x, u
(x))
_
_
ess sup
x(a,b)\K
w(x, u(x))
_
_

xK
(x, u(x+) u(x))
_
, (7.13)
4
See Theorem 1 pg 177 [40] and [47]
5
See [47] and equation (3.4) of [8].
on the class
/ =
_
(u, K) : K = a = t
0
< . . . < t
m
= b (a, b), m N, i 1, . . . , m u W
1,
(t
i1
, t
i
)
_
.
Here u(x+) and u(x) are the right and the left limit of u in x that always exist for every x K
when (u, K) /. Let us note the analogy between this problem and the classical Mumford-Shah
image segmentation one (see [7] Chapter 6).
In order to solve this problem let us note at rst that if (u, K) / then u SBV (a, b).
For this we give a relaxed formulation of the functional L extending its denition on the space
SBV (a, b) in the obvious way: for every u SBV (a, b), decomposing its distributional derivative
as Du = u
L
1
+D
#
u /(a, b), where u
L
1
(a, b) and D
#
u is purely atomic, we set
L(u, (a, b)) =
_
ess sup
x(a,b)
g(x, u
(x))
_
_
ess sup
x(a,b)
w(x, u(x))
_
_

xA
Du
(x, D
#
u(x))
_
. (7.14)
Clearly, when (u, K) /, L(u, K) = L(u, (a, b)), thus
inf
uSBV (a,b)
L(u, (a, b)) inf
(u,K)A
L(u, K).
The following theorem shows that, with suitable hypotheses on g, w, , the functional L admits
a minimum on /. Note that, in order to obtain the compactness we need a particular coercivity
hypothesis on , that, as Proposition 2.27 shows, is more regular than it seems.
Theorem 7.7. Let g, w, : (a, b) R [0, ] be proper and Borel functions. Let us suppose that,
for L
1
-a.e. x (a, b), g(x, ), w(x, ) are l.s.c. and level convex on R and there exists a function
: [0, ) [0, ) such that, for every x (a, b) and R,

g(x, ) w(x, )
([[) and lim

t
(t) = .
Moreover, setting l = infg(x, ) : (x, ) (a, b)R, let us suppose that l is l.s.c. on (a, b)R,
for every x (a, b), (x, ) l is sub-maximal on R and (x, 0) = 0, and there exists a function
0
: (0, ) [0, ) such that,
_
(x, )
0
() if x (a, b), > 0,
(x, ) = if x (a, b), < 0,
and lim
t0
0
(t) = .
Then the problem
minL(u, K) : (u, K) /
admits at least a solution.
Proof. We start considering the functional L dened by (7.14) and proving, by using the Direct
Methods, that it has a minimum on SBV (a, b). For this let us consider a minimizing sequence
u
h
h=1
SBV (a, b): we can suppose that there exists a constant M > 0 such that, for every
h N, L(u
h
, (a, b)) M (we can suppose L , ). We want to prove at rst the compactness of
this sequence.
Using the coercivity property of g, w and , we can nd a constant M

u SBV (a, b) such that L(u, (a, b)) M (in particular, for every u
h
), we have
(i) for L
1
-a.e. x (a, b), [u
(x)[ M
,
(ii) for L
1
-a.e. x (a, b), [u(x)[ M
,
(iii) for every x A
Du
(the set of the atoms of Du), D
#
u(x)
1
M
: this condition implies, in

particular, that on every jump point the function u increases.
Then, considering the sequence u
h
h=1
, we are in the position to prove the compactness of
this sequence using Theorem 2.9 if we are able to prove that there exists C > 0 such that, for
every h N, #(A
Du
h
) C. In order to prove this we consider the formula, referred to a good
representative of u
h
, given by
u
h
(b) u
h
(a+) =
_
b
a
u
(x)dx +
xA
Du
h
D
#
u
h
(x)
that implies, by means of (iii),
#(A
Du
h
)
1
M
xA
Du
h
D
#
u
h
(x) [u
h
(b) u
h
(a+)[ +
_
b
a
[u
h
(x)[dx 2M
+ (b a)M
= C.
Thus there exists a subsequence (not relabelled) and u SBV (a, b) L
(a, b) such that

u
h
u in w
-BV (a, b), and u

h
u in w
-L
(a, b).
Let us show now the lower semicontinuity of L on this sequence. Obviously Du
h
D u in w
-
/(a, b) and, by Theorem 6.20, for the functional
_
ess sup
x(a,b)
g(x, u
(x))
_
_

xA
Du
(x, D
#
u(x))
_
,
the lower semicontinuity holds. By Theorem 6.19 also the functional
ess sup
x(a,b)
w(x, u(x)),
is lower semicontinuous on the considered sequence. Then
L( u, (a, b)) liminf
h
L(u
h
, (a, b)),
that is, L admits a minimum on SBV (a, b) given by u. Let us note now that since (i), (ii) and (iii)
hold for u too, clearly #(A
D u
) C and, for L
1
-a.e. x (a, b), [u
(x)[ M
: then ( u, A
D u
) /
and the thesis is achieved.
Bibliography
[1] E. Acerbi - G. Buttazzo - N. Fusco, Semicontinuity and relaxation for integrals depending
on vector valued functions, Journal Math. Pures et Appl., 62 (1983), 371387.
[2] E. Acerbi - G. Buttazzo - N. Prinari, The class of functionals which can be represented
by a supremum, J. Convex Analysis, 9 (2002), 225-236.
[3] R. Alicandro - A. Braides - M. Cicalese, L
energies on discontinuous functions,

Preprint.
[4] L. Ambrosio, New lower semicontinuity results for integral functionals, Rend. Accad. Naz.
Sci. XL 11 (1987), 1-42.
[5] L. Ambrosio - G. Buttazzo, Weak lower semicontinuous envelope of functions dened on
a space of measures, Ann. Mat. pura appl., 150 (1988), 311-340.
[6] L. Ambrosio - G. Dal Maso, On the relaxation in BV (, R
m
) of quasi-convex integrals, J.
Func. Anal., 109 (1992), 76-97.
[7] L. Ambrosio - N. Fusco - D. Pallara, Functions of bounded variation and free disconti-
nuity problems, Oxford University Press, Inc. New York, 2000.
[8] G. Anzellotti - G. Buttazzo - G. Dal Maso, Dirichlet problem for demi-coercive func-
tionals, Nonlinear Anal., 10 (1986), 603-613.
[9] G. Aronsson, Minimization problems for the functional sup
x
F(x, f(x), f
(x)), Ark. Mat., 6

(1965), 33-53.
[10] G. Aronsson, Extension of functions satisfying Lipschitz conditions, Ark. Mat., 6 (1967),
551-561.
x
F(x, f(x), f
(x)) II, Ark. Mat.,

6 (1969), 409-431.
x
F(x, f(x), f
(x)) III, Ark. Mat.,

7 (1969), 509-512.
[13] J.P. Aubin - H. Frankowska, Set Valued Analysis, Birkhauser, Basel, 1990.
[14] E.N. Barron - R.R. Jensen, Relaxed minimal control, SIAM J. Control and Optimization
33 n.4 (1995), 1028-1039.
[15] E.N. Barron - R.R. Jensen - C.Y. Wang, Lower semicontinuity of L
functionals, Ann.
I. H. Poincare - AN 18 (2001), 495-517.
[16] E.N. Barron - W. Liu, Calculus of variations in L
, Appl. Math. Optim., 35 (1997),

237-263.
101
102 BIBLIOGRAPHY
[17] G. Bouchitt e, Representation integrale de fonctionelles convexes sur un espace de measures,
Ann. Univ. Ferrara, 18 (1987), 113-156.
[18] G. Bouchitt e - G. Buttazzo, New lower semicontinuity results for nonconvex functionals
dened on measures, Nonlinear Anal., 15 (1990), 679-692.
[19] G. Bouchitte - G. Buttazzo, Integral representation of nonconvex functionals dened on
measures, Ann. Inst. Henri Poincare, 9 (1992), 101-117.
[20] G. Bouchitte - G. Buttazzo, Relaxation for a class of nonconvex functionals dened on
measures, Ann. Inst. Henri Poincare, 10 (1993), 345-361.
[21] G. Bouchitt e - M. Valadier, Integral representation of convex functionals on a space of
measures, J. funct. Analysis, 80 (1988), 398-420.
[22] G. Buttazzo, Semicontinuity, relaxation and integral representation in the Calculus of Vari-
ations, Pitman Res. Notes Math. Ser., 207, Longman, Harlow (1989).
[23] L. Carbone - C. Sbordone, Some properties of -limits of integral functionals, Ann. Mat.
Pura Appl., 122 (1979), 1-60.
[24] R. Cern y - J. Mal y, Counterexample to lower semicontinuity in the Calculus of Variations,
Math. Z., 248 (2001), 689-694.
[25] R. Cern y - J. Mal y, Yet more counterexample to lower semicontinuity in the Calculus of
Variations, J. Convex Analysis, 9 (2002), 295-299.
[26] F.H. Clarke, Generalized gradients and applications, Trans. Amer. Math. Soc., 205 (1975),
247-262.
[27] R. Courant - F. John, Introduction to Calculus and Analysis, Volume II, Wiley-Interscience
Publication, 1974.
[28] J.P. Crouzeix, A review of continuity and dierentiability properties of quasiconvex functions
on R
n
, in Research Notes in Mathematics, J.P. Aubin - R.B. Vinter, Convex analisys and
Optimization, Pitman Advanced Publishing Program, 1982.
[29] B. Dacorogna, Direct methods in the Calculus of Variations, Springer-Verlag, Berlin, 1989.
[30] G. Dal Maso, Integral representation on BV () of -limits of variational integrals,
Manuscripta Math., 30 (1980), 387-416.
[31] V. De Cicco, A lower semicontinuity result for functionals dened on BV (), Ricerche Mat.,
39 (1990), 293-325.
[32] V. De Cicco, Lower semicontinuity for certain integral functionals on BV (), Boll. Un.
Mat. Ital., 5-B (1991), 291-313.
[33] V. De Cicco - N. Fusco - A. Verde, On L
1
-lower semicontinuity in BV (), Preprint,
Dipartimento di Matematica e Applicazioni R. Caccioppoli, Napoli.
[34] V. De Cicco - G. Leoni, A chain rule in L
1
(div; ) and its applications to lower semicon-
tinuity, to appear in Calculus of Variations (2003).
[35] E. De Giorgi, Teoremi di semicontinuità nel calcolo delle variazioni, Istituto Nazionale di
Alta Matematica, Roma (1968).
BIBLIOGRAPHY 103
[36] E. De Giorgi - L.Ambrosio - G. Buttazzo, Integral representation and relaxation for
functionals dened on measures Atti Accad. Naz. Lincei, Cl. Sci. Fis. Mat. Natur. Rend., 81
(1987), 7-13.
[37] E. De Giorgi - G. Buttazzo - G. Dal Maso, On the lower semicontinuity of certain
integral functionals, Atti Accad. Naz. Lincei, Cl. Sci. Fis. Mat. Natur. Rend. 74 (1983), 274-
282.
[38] G. Eisen, A counterexample for some lower semicontinuity results, Math. Z., 162 (1978),
241-243.
[39] I. Ekeland - R. T emam, Analyse convexe et problèmes variationnels, Dunod, Paris, 1974.
[40] L.C. Evans - R.F. Gariepy, Measure theory and ne properties of functions, Studies in
advanced mathematics, CRC Press, Boca Raton, Florida, 1992.
[41] H. Federer, Geometric measure theory, Springer-Verlag, Berlin, 1969.
[42] I. Fonseca - G. Leoni, Some remarks on lower semicontinuity, Indiana Univ. Math. J., 49
(2000), 617-635.
[43] I. Fonseca - G. Leoni, On lower semicontinuity and relaxation, Proc. Roy. Soc. Edinburgh,
131A (2001), 519-565.
[44] I. Fonseca - S. M uller, Quasi-convex integrands and lower semicontinuity in L
1
, SIAM J.
Math. Anal., 23 (1992), 1081-1098.
[45] N. Fusco, Dualità e semicontinuità per integrali del tipo dellarea, Rend. Accad. Sci. Fis.
Mat. Napoli, 46 (1979), 81-90.
[46] N. Fusco - F. Giannetti - A. Verde, Remark on the L
1
-lower semicontinuity for integral
functionals in BV , Preprint n. 15 (2002), Dipartimento di Matematica e Applicazioni R.
Caccioppoli, Napoli.
[47] E. Giusti, Minimal surfaces and functions of bounded variation, Birkhauser, Basel, 1984.
[48] E. Giusti, Metodi diretti nel calcolo delle variazioni, Unione Matematica Italiana, Bologna,
1994.
[49] C. Goffman - J. Serrin, Sublinear functions of measure and variational integrals, Duke
Math J., 31 (1964), 159-178.
[50] M. Gori, I teoremi di semicontinuità di Serrin e loro estensioni, Degree Thesis in Mathe-
matics, University of Firenze, July 2000.
[51] M. Gori, On the denition of a supremal functional on BV , Preprint n.2.449.1417 (October
2002) University of Pisa.
[52] M. Gori - F. Maggi, On the lower semicontinuity of supremal functionals, ESAIM: COCV,
9 (2003), 135-143.
[53] M. Gori - F. Maggi, The common root of the geometric conditions in Serrins lower semi-
continuity theorem, to appear in Ann. di Mat. Pura e Appl..
[54] M. Gori - F. Maggi - P. Marcellini, On some sharp conditions for lower semicontinuity
in L
1
, Dierential and Integral Equations, 16, n.1 (2003), 51-76.
[55] M. Gori - P. Marcellini, An extension of the Serrins lower semicontinuity theorem, J.
Convex Analysis 9 (2002), 475-502.
104 BIBLIOGRAPHY
[56] A.D. Ioffe, On lower semicontinuity of integral functionals, SIAM J. Control Optim., 15
(1977), 521-538.
[57] M. Lavrentiev, Sur quelques problèmes du calcul des variations, Ann. Mat. Pura Appl., 4
(1926), 107-124.
[58] G. Lebourg, Generic dierentiability of lipschitzian functions, Trans. Amer. Math. Soc., 256
(1979), 125-144.
[59] F. Maggi, On the relaxation on BV of certain non coercive integral functionals, to appear
in J. Convex Analysis.
[60] B. Mani` a, Sopra un esempio di Lavrentie, Boll. Un. Mat. Ital., 13 (1934), 146-153.
[61] P. Marcellini, Some problems of semicontinuity and of convergence for integrals of the
calculus of variations, Recent Methods in Nonlinear Analisys, De Giorgi et al. editors, Rome,
Pitagora Ed., 1978, 205-221.
[62] P. Marcellini, Alcuni recenti sviluppi nei problemi 19-esimo e 20-esimo di Hilbert, Bollettino
U.M.I., (7) 11-A (1997), 323-352.
[63] C. B. Morrey, Multiple Integrals in the Calculus of Variations, Die Grundlehren der math-
ematischen Wissenschaften 130, Springer-Verlag New York, New York, 1966.
[64] C. Y. Pauc, La methode metrique en calcul des variations, Hermann, Paris, 1941.
[65] C. Olech, A characterization of L
1
-weak lower semicontinuity of integral functionals, Bull.
De lAcademie Polonaise Des Sciences, 25 n.2 (1977), 135-142.
[66] R. T. Rockafellar, Convex Analysis, Princeton Mathematical Series 28, Princeton Uni-
versity Press, Princeton, New Jersey, 1970.
[67] R.T. Rockafellar - R.J-B. Wets, Variational Analysis, Die Grundlehren der mathema-
tischen Wissenschaften 317, Springer-Verlag, Berlin, 1998.
[68] J. Serrin, On the area of curved surfaces, Amer. Math. Monthly, 68 (1961), 435-440.
[69] J. Serrin, On the denition and properties of certain variational integrals, Trans. Amer.
Math. Soc., 101 (1961), 139-167.
[70] L. Tonelli, Fondamenti di Calcolo delle Variazioni, Volume primo, Zanichelli, Bologna,
1921.
[71] L. Tonelli, Fondamenti di Calcolo delle Variazioni, Volume secondo, Zanichelli, Bologna,
1923.
[72] L. Tonelli, Sur la semi-continuite des integrales doubles du calcul des variations, Acta Math.,
53 (1929), 325-346.
[73] V.V. Zhikov, Averaging of functionals of the Calculus of Variations and elaticity theory,
Math. USSR Izv., 29 (1987), 33-66.

Michele Gori - Lower Semicontinuity and Relaxation (2004)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Michele Gori - Lower Semicontinuity and Relaxation (2004)

Uploaded by

Copyright:

Available Formats

Universit`a degli Studi di Pisa

Facolt` a di Scienze Matematiche, Fisiche e Naturali

f(x, u(x), u(x))dx, (1.1)

) topologies, the compactness of the minimizing sequences can be obtained by requiring

([s[ +[[), (1.4)

: [0, ) [0, ) and lim

(t) = . Indeed, (1.3) and (1.4) guarantee that every

area of the plane faces =

(I) = I and R[L

(I) has been deeply studied,

) as h . Then, for every

and the functionals R[w

, developed by Gori in [51]. In

and following the

/(, R) and /(, R

) is said the Cantor part of while

is said its atomic part.

, [(x)[ < ), by the duality

(a, b) such that u

-BV (a, b) and u

A() such that

respectively are the functions dened, for every R

has been already proposed in the work of Gori [51], while

appears here for the rst time.

(0) = 0. Moreover, if f is convex then f

is l.s.c., positively homogeneous of degree 0

is level convex too.

can be found in [67] Theorem 3.21.

since the other case is completely analogous. Let

() g(). In order to prove the converse

is clearly positively homogeneous of degree 0. Fixed now R

the supremum of a family of l.s.c. and sub-maximal functions, it is also

() = . Then there exists a function

: [0, ) [0, ) such that,

([[) and lim

() = . Then there exists a function

(see Proposition 2.20),

is l.s.c. and level convex, that B is a convex

, we point out that is

() < s. Let us suppose that cl(K

, that is in formulas, for every R

() to the second integral (note that r

preserves m-dimensional volumes, so that its Jacobian is equal to 1, and that r

() = 0 contains no straight line;

such that, for every (t, ) (

f(x, s, ) we mean the gradient of f(x, s, ).

(t, ) we mean (f(t, ))

f() = proving in this hypothesis that, for every R,

f() = and lim

(, v), for every choice of v ,= 0, v f(): if A = the proof is achieved.

, there exists a constant

(t, ) c(t)[[ and h

(t, ) c(t)[[ as 0. Moreover for

(t, ) is strictly convex. In order to prove this, we consider , R

(t, ) f(t, ) as 0. Since h

(t, ) c(t)[[ g(t, ) we

(t, ) f(t, ) as 0. However, note that f

is not necessarily bounded from below, as

(t, ) 0, while, when t

(t, ) could tend to as [[ .

(t, ) is strictly convex, for every t

f. Indeed, in this case, we can dene, for every (t, ) R

(t, ) is strictly convex, since, for every t

(t, ) that is strictly convex. Moreover