Lecture Notes on Elliptic Partial Di↵erential Equations

⇤
Luigi Ambrosio

Contents
1 Some basic facts concerning Sobolev spaces 3

2 Variational formulation of some PDEs 11

2.1 Elliptic operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Inhomogeneous boundary conditions . . . . . . . . . . . . . . . . . . . . . 14
2.3 Elliptic systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Necessary minimality conditions . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Lower semicontinuity of integral functionals 24

4 Regularity Theory 33
4.1 Nirenberg method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5 Decay estimates for systems with constant coefficients 43

6 Regularity up to the boundary 45

7 Interior regularity for nonlinear problems 49

8 Hölder, Morrey and Campanato spaces 51

9 XIX Hilbert problem and its solution in the two-dimensional case 57

10 Schauder theory 61

11 Regularity in Lp spaces 65
⇤
PhD course given in 2009-2010 and then in 2012-2013, 2014-2015, lectures typed by A.Carlotto and
A.Massaccesi

1
12 Some classical interpolation theorems 68

13 Lebesgue di↵erentiation theorem 70

14 Calderón-Zygmund decomposition 72

15 The BMO space 73

16 De Giorgi’s solution of Hilbert’s XIX problem 85

16.1 The basic estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
16.2 Some useful tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
16.3 Proof of Hölder continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

17 Regularity for systems 99

17.1 De Giorgi’s counterexample to regularity for systems . . . . . . . . . . . . 99

18 Partial regularity for systems 103

n
18.1 Partial regularity for systems: L (⌃(u)) = 0 . . . . . . . . . . . . . . . . 111
18.2 Hausdor↵ measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
18.3 Partial regularity for systems: H n 2+" (⌃(u)) = 0 . . . . . . . . . . . . . . 120

19 Some tools from convex and nonsmooth analysis 124

19.1 Subdi↵erential of a convex function . . . . . . . . . . . . . . . . . . . . . . 124
19.2 Convex functions and Measure Theory . . . . . . . . . . . . . . . . . . . . 130

20 Viscosity solutions 133

20.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
20.2 Viscosity versus classical solutions . . . . . . . . . . . . . . . . . . . . . . . 136
20.3 The distance function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
20.4 Maximum principle for semiconvex functions . . . . . . . . . . . . . . . . . 141
20.5 Existence and uniqueness results . . . . . . . . . . . . . . . . . . . . . . . . 146
20.6 Hölder regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

21 Regularity theory for viscosity solutions 154

21.1 The Alexandrov-Bakelman-Pucci estimate . . . . . . . . . . . . . . . . . . 154
21.2 The Harnack inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

Preface
Prerequisites: basic knowledge of Functional Analysis and Measure Theory, preferably
also a basic knowledge of Sobolev spaces of functions of one independent variable.

2
Br (x) Ball with center x and radius r (also Br = Br (0), B = B1 )
A⇢B Inclusion in the weak sense
AbB A ⇢ B (typically used for pairs of open sets)
Ln Lebesgue measure in Rn
C k (⌦) Functions continuously k-di↵erentiable in ⌦
Lp (⌦) Lebesgue Lp space
@u
@i u, @xi u, ri u, @xi
i-th partial derivative (weak or classical)
ru
R Gradient of u R
⌦
f dµ Mean integral value, namely ⌦ f dµ/µ(⌦)

1 Some basic facts concerning Sobolev spaces

In this book, we will make constant use of Sobolev spaces. Here, we will just summarize
the basic facts needed in the sequel, referring for instance to [4] or [1] for a more detailed
treatment of this topic.
Actually, it is possible to define Sobolev spaces in (at least) two di↵erent ways, whose
(partial) equivalence is discussed below.
Definition 1.1. Let ⌦ ⇢ Rn be an open and bounded domain and fix an exponent p with
1  p < 1. We can consider the class of regular functions C 1 ⌦ (i.e. the subset of
C 1 (⌦) consisting of functions u such that both u and ru admit a continuous extension
on @⌦) endowed with the norm
q
kukW 1,p =p kukpLp + krukpLp . (1.1)

We define the space H 1,p (⌦) to be the completion with respect to the W 1,p norm of C 1 (⌦).
For unbounded domains, including the whole space Rn , the definition is similar and
based on the completion of

u 2 C 1 (⌦) : u 2 Lp (⌦), |ru| 2 Lp (⌦) .

Notice that H 1,p (⌦) ⇢ Lp (⌦).

On the other hand, we can adopt a di↵erent viewpoint, inspired by the theory of
distributions.
Definition 1.2. Let ⌦ ⇢ Rn be an open domain and consider the space Cc1 (⌦) whose
elements will be called test functions. For i = 1, . . . , n, we say that u 2 L1loc (⌦) has i-th
derivative in weak sense equal gi 2 L1loc (⌦) if
Z Z
u@i ' dx = 'gi dx 8' 2 Cc1 (⌦). (1.2)
⌦ ⌦

3
Whenever such g1 , . . . , gn exist, we say that u is di↵erentiable in weak sense and we write
gi = @i u and
ru = @1 u, . . . , @n u).
For 1  p  1 we define the space W 1,p (⌦) as the subset of Lp (⌦) whose elements u are
weakly di↵erentiable with corresponding derivatives @i u also belonging to Lp (⌦).
It is clear that if gi exists, it must be uniquely determined up to Lebesgue negligible
sets, since h 2 L1loc (⌦) and
Z
h' dx = 0 8' 2 Cc1 (⌦)
⌦

implies h = 0. This implication can be easily proved by approximation, showing that the
property above is stable under convolution, namely h" = h ⇤ ⇢" satisfies
Z Z
h" ' dx = h' ⇤ ⇢" = 0 8' 2 Cc1 (⌦" ),
⌦" ⌦

where ⌦" is the (slightly) smaller domain

⌦" := {x 2 ⌦ : dist(x, @⌦) > "} , (1.3)

⇢" (x) = " n ⇢(x/") with ⇢ smooth, even and compactly supported in the unit ball and we
used the simmetry property (a consequence of Fubini’s theorem).
Z Z
(a ⇤ ⇢" )b dx = a(b ⇤ ⇢" ) dx. (1.4)

Obviously, classical derivatives are weak derivatives and the notation @i u (or, equivalently,
@u
@xi u, ri u or even @x i
) is justified.
Another classical way to relate weak and strong derivatives is via convolution: namely
if u has weak i-th derivative equal to g, then

@i (u ⇤ ⇢" ) = g ⇤ ⇢" in ⌦" , in the classical sense. (1.5)

Knowing the identity (1.5) for smooth functions, its validity can be easily extended con-
sidering both sides as weak derivatives and using (1.4):
Z Z Z Z Z
(u⇤⇢" )@i ' dx = u(@i ')⇤⇢" dx = u@i ('⇤⇢" ) dx = g'⇤⇢" dx = g ⇤⇢" ' dx
⌦ ⌦ ⌦ ⌦ ⌦

for all ' 2 Cc1 (⌦" ). Now, the smoothness of u ⇤ ⇢" tells us that the derivative in the left
hand side of (1.5) is (equivalent to) a classical one.
Another consequence of (1.5) is:

4
Theorem 1.3 (Constancy theorem). If u 2 L1loc (⌦) satisfies ru = 0 in the weak sense,
then for any ball B ⇢ ⌦ there exists a constant c 2 R such that u = c L n -a.e. in B. In
particular, if ⌦ is connected, u = c L n -a.e. in ⌦ for some c 2 R.

Proof. Again we argue by approximation, using the fact that (1.5) ensures that the func-
tion u ⇤ ⇢" are locally constant in ⌦" and taking the L1loc limit as " ! 0.
Notice also that Definition 1.2 covers the case p = 1, while it is not immediately clear
how to adapt Definition 1.1 to cover this case: usually H Sobolev spaces are defined for
p < 1 only.
In the next proposition we consider the relation of W 1,1 with Lipschitz functions. We
omit, for brevity, the simple proof, based once more on convolutions.

Proposition 1.4 (Lipschitz versus W 1,1 functions). If ⌦ ⇢ Rn is open, then Lip(⌦) ⇢

W 1,1 (⌦) and
kDukL1 (⌦)  Lip(u, ⌦). (1.6)
In addition, if ⌦ is convex then Lip(⌦) = W 1,1 (⌦) and equality holds in (1.6).

Since H 1,p (⌦) is defined by means of approximation by regular functions, for which
(1.2) is just the elementary “integration by parts formula”, it is clear that H 1,p (⌦) ⇢
W 1,p (⌦); in addition, the same argument shows that the weak derivative of u 2 H 1,p (⌦),
in the sense of W Sobolev spaces, is precisely the strong Lp (⌦, Rn ) limit of ruh , where
uh 2 C 1 (⌦) are strongly convergent to u. This allows to show by approximation some
basic calculus rules in H Sobolev spaces for weak derivatives, as the chain rule
0
r( u) = (u)ru 2 C 1 (R) Lipschitz with (0) = 0, u 2 H 1,p (⌦) , (1.7)

and, with a little more e↵ort (because one has first to show using the chain rule that
bounded H 1,p functions can be strongly approximated in H 1,p by equibounded C 1 (⌦)
functions) the Leibniz rule

r(uv) = urv + vru u, v 2 H 1,p (⌦) \ L1 (⌦) . (1.8)

On the other hand, we don’t have to prove the same formulas for the W Sobolev spaces:
indeed, using convolutions and a suitable extension operator described below (in the case
⌦ = Rn the proof is a direct application of (1.5), since in this case ⌦" = Rn ), one can
prove the following result:

Theorem 1.5 (H = W ). If either ⌦ = Rn or ⌦ is a bounded regular domain, then

H 1,p (⌦) = W 1,p (⌦) 1  p < 1. (1.9)

5
 With the word regular we mean that ⌦ is the epigrah of a Lipschitz function of (n 1)-
variables, written in a suitable system of coordinates, near to any boundary point.
However the equality H = W is not true in general, as the following example shows.
Example 1.6. In the Euclidean plane R2 , consider the open unit ball x2 +y 2 < 1 deprived
of one of its radii, say for instance the segment ⌃ given by ( 1, 0] ⇥ {0} . We can define on
this domain ⌦ a function v having values in ( ⇡, ⇡) and representing the angle in polar
coordinates. Fix an exponent 1  p < 2. It is immediate to see that v 2 C 1 (⌦) and that
its gradient is p-integrable, hence v 2 W 1,p (⌦). On the other hand, v 2 / H 1,p (⌦) because
the definition we have given would require the existence of regular approximations for
v up to the boundary: more precisely, one can easily show, using Fubini’s theorem and
polar coordinates, that any u 2 H 1,p (⌦) satisfies
1,p
! 7! u(rei! ) 2 Wloc (R) for L 1 -a.e. r 2 (0, 1). (1.10)
0 1 1,p
Indeed, if un 2 C
P(⌦) \ C (⌦) converge to u strongly in H (⌦) and (possibly extracting
a subsequence) n krun+1 run kp < 1, for all 2 (0, 1) the inequality |@✓ v|  |rv|/r
gives
Z 1 X ✓Z ⇡ ◆1/p X
@un+1 @un p
| | d✓ dr  1 1/p krun+1 run kp < 1.
n ⇡ @✓ @✓ n

Since > 0 is arbitrary, it follows that for L 1 -a.e. r 2 (0, 1) the 2⇡-periodic continuous
functions ✓ 7! un (rei✓ ) have derivatives strongly convergent in Lploc (R), and therefore
(by the fundamental theorem of calculus) are equicontinuous. Any limit point of these
functions in Lploc (R) must then be 2⇡-periodic, continuous and W 1,p . If, by contradiction,
we take u = v, a similar Fubini argument shows that un (rei✓ ) converge in Lp ( ⇡, ⇡) to
the function v for L 1 -a.e. r 2 (0, 1). But, the function v(r, ✓) = ✓ 2 ( ⇡, ⇡) has no
continuous 2⇡-periodic extension. Therefore we get a contradiction and v can’t be in
H 1,p (⌦).
Remark 1.7. Taking into account the example above, we mention the Meyers-Serrin
theorem [24], ensuring that, for any open set ⌦ ⇢ Rn and 1  p < 1, the identity
W 1,p
C 1 (⌦) \ W 1,p (⌦) = W 1,p (⌦) (1.11)

holds. The proof can be achieved by (1.5) and a partition of unity.

The previous example underlines the crucial role played by the boundary behaviour, when
we try to approximate a function in W 1,p by C 1 (⌦) (or even C 0 (⌦) \ C 1 (⌦)) functions.
In the Meyers-Serrin theorem, instead, no smoothness up to the boundary is required for
the approximating sequence. So, if we had defined the H spaces using C 1 (⌦) \ Lp (⌦)
functions with gradient in Lp (⌦) instead of C 1 (⌦) functions, the identity H = W would

6
be true unconditionally. In the case p = 1, the construction in the Meyers-Serrin theorem
provides for all u 2 W 1,1 (⌦) a sequence (un ) ⇢ C 1 (⌦) converging to u uniformly in ⌦,
with sup⌦ |run | convergent to kruk1 . Again, this might lead to a definition of H 1,1 for
which H 1,1 = W 1,1 unconditionally.
As it will be clear soon, we also need to define an appropriate subspace of H 1,p (⌦) in
order to work with functions vanishing at the boundary.
Definition 1.8. Given ⌦ ⇢ Rn open, we define H01,p (⌦) to be the completion of Cc1 (⌦)
with respect to the W 1,p norm.
It is clear that H01,p (⌦), being complete, is a closed subspace of H 1,p (⌦). Notice also
that H 1,p (Rn ) coincides with H01,p (Rn ). To see this, suffices to show that any function
u 2 C 1 (Rn ) with both |u| and |ru| in Lp (Rn ) belongs to H01,p (Rn ). We can indeed
approximate any such function u, strongly in H 1,p norm, by the compactly supported
functions R u, where R : Rn ! [0, 1] are smooth, 2-Lipschitz, identically equal to 1 on
B R and identically equal to 0 on Rn \ B R+1 .
We now turn to some classical inequalities.
Theorem 1.9 (Poincaré inequality, first version). Let ⌦ ⇢ Rn be an open bounded set
and p 2 [1, 1). Then there exists a constant C(⌦, p), depending only on ⌦ and p, such
that
kukLp  C(⌦, p) krukLp 8u 2 H01,p (⌦). (1.12)
In addition C(⌦)  C(n, p)diam(⌦).
The proof of this result can be simplified by means of these properties:
• H01,p (⌦) ⇢ H01,p (⌦0 ) if ⌦ ⇢ ⌦0 (monotonicity);

• If C(⌦, p) denotes the best constant, then C( ⌦, p) = C(⌦, p) (scaling invariance)

and C(⌦ + h, p) = C(⌦, p) (translation invariance).
The first fact is a consequence of the definition of the spaces H01,p in terms of regular
functions, while the second one (translation invariance is obvious) follows by:

u (x) = u ( x) 2 H01,p (⌦) 8u 2 H01,p ( ⌦). (1.13)

Proof. By the monotonicity and scaling properties, it is enough to prove the inequality
for ⌦ = Q ⇢ Rn where Q is the cube centered at the origin, with sides parallel to the
coordinate axis and length 2. We write x = (x1 , x0 ) with x0 = (x2 , . . . , xn ). By density,
we may also assume u 2 Cc1 (⌦) and hence use the following representation formula:
Z x1
0 @u
u(x1 , x ) = (t, x0 ) dt. (1.14)
1 @x1

7
Hölder’s inequality gives
Z 1 p
p 0 @u
|u| (x1 , x )  2 p 1
(t, x0 )dt (1.15)
1 @x1
and hence we just need to integrate w.r.t. x1 to get
Z 1 Z 1 p
p 0 @u
|u| (x1 , x ) dx1  2 p
(t, x0 )dt. (1.16)
1 1 @x1

Now, integrating w.r.t. x0 , repeating the previous argument for all the variables xj , j =
1, . . . , n and summing all such inequalities we obtain the thesis with C(Q, p)  2/n1/p .
⇤

Theorem 1.10 (Rellich). Let ⌦ be an open bounded subset with regular boundary and let
p 2 [1, 1). Then the immersion W 1,p (⌦) ,! Lp (⌦) is compact.
We do not give a complete proof of this result. Instead, we observe that it can be
obtained using an appropriate linear and continuous extension operator
T : W 1,p (⌦) ! W 1,p (Rn ) (1.17)
such that 8
< Tu = u in ⌦;
:
supp(T u) ⇢ ⌦0 ,
being ⌦0 a fixed bounded domain in Rn containing ⌦. When ⌦ is an halfspace the operator
can be achieved simply by a reflection argument; in the general case the construction relies
on the fact that the boundary of @⌦ is regular and so can be locally straightened by means
of Lipschitz maps (we will use these ideas later on, treating the boundary regularity of
solutions to elliptic PDE’s). The global construction is then obtained thanks to a partition
of unity.
The operator T allows basically a reduction to the case ⌦ = Rn , considered in the
next theorem.
Theorem 1.11. The immersion W 1,p (Rn ) ,! Lploc (Rn ) is compact, namely if (uk ) ⇢
W 1,p (Rn ) is bounded, then (un ) has limit points in the Lploc (Rn ) topology, and any limit
point belongs to W 1,p (Rn ).
Remark 1.12. It should be noted that the immersion W 1,p (Rn ) ,! Lp (Rn ) is obviously
continuous, but certainly not compact: just take a fixed element in W 1,p (Rn ) and sup-
ported in the unit square and consider the sequence of its translates along vectors ⌧h
with |⌧h | ! 1. Of course this is a bounded sequence in W 1,p (Rn ) but no subsequence
converges in Lp (Rn ) (indeed, all functions have the same Lp norm, while it is easily seen
that their Lploc limit is 0).

8
 Let us now briefly sketch the main points of the proof of this theorem, since some of
the ideas we use here will be often considered in the sequel.
Proof. Basically, it is enough to prove that a bounded family F ⇢ W 1,p (Rn ) is totally
bounded in Lploc (Rn ). To obtain this, observe firstly that given any Borel domain A ⇢ Rn
and any ' 2 W 1,p (A|h| ) we have

k⌧h ' 'kLp (A)  |h| kr'kLp (A|h| ) (1.18)

where A|h| is the |h| neighbourhood of the set A and ⌧h '(x) = '(x + h). By approxima-
tion, we can assume with no loss of generality that ' 2 C 1 (A|h| ). The inequality (1.18)
follows by the elementary representation
Z 1
(⌧h ' ')(x) = hr'(x + sh), hi ds (1.19)
0

since
Z Z 1
k⌧h ' 'kpLp (A)  |hr'(x + sh), hi|p ds dx (1.20)
A 0
Z 1Z
p
 |h| |r'(y)|p dy ds = |h|p kr'kpLp (A|h| ) (1.21)
0 A|h|

where we used the Cauchy-Schwarz inequality and Fubini’s theorem. Hence, denoting by
(⇢" )">0 any rescaled family of smooth mollifiers such that supp(⇢" ) ⇢ B(0, "), we have
that for any R > 0
sup k' ⇤ ⇢" 'kLp (BR ) ! 0 (1.22)
'2F

for " ! 0. In fact, since ' ⇤ ⇢" is a mean, weighted by ⇢" , of translates of '
Z
' ⇤ ⇢" = ⌧ y '⇢" (y) dy ,

by the previous result we deduce

Z !1/p
sup k' ⇤ ⇢" 'kLp (BR )  " sup |r'|p dx . (1.23)
'2F '2F BR+"

To conclude we just need to observe that the regularized family {' ⇤ ⇢" , ' 2 F} is rela-
tively compact in Lploc (Rn ) for any fixed " > 0. But this is easy since the Young inequality
implies
sup |' ⇤ ⇢" |  k'kL1 (BR+" ) k⇢" k1 (1.24)
BR

9
and similarly
sup |r(' ⇤ ⇢" )|  k'kL1 (BR+" ) kr⇢" k1 (1.25)
BR

so the claim follows by means of the Ascoli-Arzelá theorem. Notice that we used the
gradient bounds on elements of F only in (1.23). ⇤
We also need to mention another inequality due to Poincaré.

Theorem 1.13 (Poncaré inequality, second version). Let us consider a bounded, regular
and connected domain ⌦ ⇢ Rn and an exponent 1  p < 1, so that by Rellich’s theorem
we have the compact immersion W 1,p (⌦) ,! Lp (⌦). Then, there exists a constant C(⌦, p)
such that Z Z
p
|u u⌦ | dx  C |ru|p dx 8u 2 W 1,p (⌦) (1.26)
⌦ ⌦
R
where u⌦ = ⌦
u dx.
Proof. By contradiction, if the desired inequality were not true, exploiting its homogene-
ity and translation invariance we could find a sequence (un ) ⇢ W 1,p (⌦) such that

• (un )⌦ = 0 for all n 2 N;

R
• ⌦ |un |p dx = 1 for all n 2 N;
R
• ⌦ |run |p dx ! 0 for n ! 1.

By Rellich’s theorem there exists (up to a subsequence) a limit point u 2 Lp , that is

un ! u in Lp (⌦). It is now a general fact that if (run ) has some weak limit point g then
necessarily g = ru. Therefore, in this case we have by comparison ru = 0 in Lp (⌦) and
hence, by connectivity of the domain and the constancy theorem, we deduce that u must
be equivalent to a constant. By taking limits we see that u satisfies at the same time
Z Z
u dx = 0 and |u|p dx = 1, (1.27)
⌦ ⌦

which is clearly impossible. ⇤

Note that the previous proof is not constructive and crucially relies on the general
compactness result by Rellich.

Remark 1.14. It should be observed that the previous proof, even though very simple, is
far from giving the sharp constant for the Poincaré inequality (1.26). The determination
of the sharp constant is a difficult problem, solved only in very special cases (for instance
on intervals of the real line and p = 2, by Fourier analysis). Many more results are instead
available for the sharp constant in the Poincaré inequality (1.12).

10
2 Variational formulation of some PDEs
After the introductory section, whose main purpose was to fix the notation and recall
some basic tools of the theory of Sobolev spaces, we are now ready to discuss some basic
elliptic PDEs.
Let us consider the generalised Poisson equation
8 P
< u=f ↵ @↵ f↵ in ⌦;
(2.1)
:
u 2 H01,2 (⌦)

with data f, f↵ 2 L2 (⌦) for some fixed bounded and regular domain ⌦. This equation
has to be intended in a weak sense, that is, we look for u 2 H01,2 (⌦) satisfying
Z Z X
hru, r'i dx = (f ' + f↵ @↵ ') dx 8' 2 Cc1 (⌦). (2.2)
⌦ ⌦ ↵

Equivalently, by continuity of the bilinear form and density of Cc1 (⌦), the previous con-
dition could be requested for any ' 2 H01,2 (⌦).
In order to obtain existence
R weP just need to apply Riesz’s theorem to the associated
linear functional F (v) = ⌦ (f v + ↵ f↵ @↵ v) dx on the Hilbert space H01,2 (⌦) endowed
with the scalar product Z
(u, v) = hru, rvi dx (2.3)
⌦
which is equivalent to the usual one thanks to the Poincaré inequality (first version) proved
in Theorem 1.9.
We can consider many variants of the previous problem, basically by introduction of
one or more of the following elements:
• more general di↵erential operators instead of ;
• inhomogeneous or mixed boundary conditions;
• systems instead of single equations.
Our purpose now is to briefly discuss each of these situations.

2.1 Elliptic operators

The first variant is to consider scalar problems having the form
8 P ↵
P
< ↵, @↵ (A @ u) = f ↵ @↵ f↵ in ⌦;

:
u 2 H01,2 (⌦)

11
where, as before f, f↵ 2 L2 (⌦), and A 2 Rn⇥n is a constant matrix satisfying the following
requirements:
(i) A↵ 2 Rn⇥n is symmetric, that is A↵ = A ↵ ;

(ii) A has only positive eigenvalues, equivalently, A cI for some c > 0, in the sense of
quadratic forms.
Here and in the sequel we use the capital letter I to denote the identity n ⇥ n matrix. It
is not difficult to show that a change of independent variables, precisely u(x) = v(A 1 x),
transforms this problem into one of the form (2.1). For this reason it is convenient to deal
immediately with the case of a non-constant matrix A(x) 2 Rn⇥n satisfying:
(i) A is a Borel and L1 function defined on ⌦;

(ii) A(x) is symmetric for a.e. x 2 ⌦;

(iii) there exists a positive constant c such that

A(x) cI for a.e. x 2 ⌦ . (2.4)

As indicated above, the previous problem has to be intended in weak sense and precisely
Z Z X
hAru, r'i dx = (f ' + f↵ @↵ ') dx 8' 2 Cc1 (⌦). (2.5)
⌦ ⌦ ↵

By continuity and density, also in this case it is equivalent to require the validity of the
identity above for all ' 2 H01,2 (⌦). In order to obtain existence we could easily modify
the previous argument when |A| 2 L1 (⌦), using the equivalent scalar product
Z X
hu, vi := A↵ @↵ u@ v dx .
⌦ ↵,

However, in order to include also unbounded A’s, thus dropping assumption (i), we prefer
here to proceed di↵erently and introduce some ideas that belong to the so-called direct
method of the Calculus of Variations. Let us consider the functional F : H01,2 (⌦) ! R
Z Z XZ
1
F (v) = hArv, rvi dx f v dx f↵ @↵ v dx. (2.6)
⌦ 2 ⌦ ↵ ⌦

First we note that, thanks to the assumption (2.4) on A, for all " > 0 it holds
Z Z X Z
c 2 1 2 2 ✏
F (v) |rv| dx (|f | + |f↵ | ) dx v 2 + |rv|2 dx.
2 ⌦ 2" ⌦ ↵
2 ⌦

12
Choosing " < c/2 we get
Z Z X Z
c 1 ✏
F (v) |rv|2 dx 2
(|f | + 2
|f↵ | ) dx v 2 dx
4 ⌦ 2" ⌦ ↵
2 ⌦

and now, thanks to the Poincaré inequality, we can choose possibly " even smaller to get
Z Z X
c 2 1
F (v) |rv| dx (|f |2 + |f↵ |2 ) dx.
8 ⌦ 2" ⌦ ↵

In particular F is coercive, that is

lim F (v) = +1 (2.7)

kvk 1,2 !+1
H0 (⌦)

and consequently, in order to look for its minimum points it is enough to consider a closed
ball of H01,2 (⌦). Now, take any minimizing sequence (un ) of F : since H01,2 (⌦) is (being
Hilbert) a reflexive space we can assume, possibly extracting a subsequence, that un * u
for some u 2 H01,2 (⌦). Using Fatou’s lemma and the fact that uh ! u in H 1,2 implies
ruh(k) ! ru a.e. in ⌦ for a suitable subsequence h(k), it is not difficult to prove that
F is lower semicontinuous (we shall also prove this in Theorem 3.2, in a more general
framework). In addition, F is convex, being the sum of a linear and a convex functional.
It follows that F is also weakly lower semicontinuous, hence

F (u)  lim inf F (un ) = 1,2

inf F (2.8)
n!1 H0 (⌦)

and we conclude that u is a (global) minimizer of F . Actually, the functional F is strictly

convex and so u is its unique minimizer.
If A is bounded, since F is a C 1 functional on H01,2 (⌦) we get dF (u) = 0, where dF is
the di↵erential in the Gateaux sense of F :
F (u + "') F (u)
dF (u) ['] := lim 8' 2 H01,2 (⌦) .
"!0 "
Here a simple computation gives
Z Z XZ
dF (u) ['] = hAru, r'i dx f ' dx f↵ @↵ ' dx (2.9)
⌦ ⌦ ↵ ⌦

and the desired result follows. Even in the case when |A| 2 L1loc we can still di↵erentiate
the functional, but a priori only along directions in ' 2 Cc1 (⌦), and recover the weak
formulation of our PDE.

13
2.2 Inhomogeneous boundary conditions
We now turn to study the boundary value problem for u 2 H 1,2 (⌦)
8 P
< u=f ↵ @↵ f↵ in ⌦;

:
u=g on @⌦

with f, f↵ 2 L2 (⌦) and a suitable class of functions g 2 L2 (@⌦). Since the immersion
H 1,2 (⌦) ,! C(⌦) does not hold if n 2, the boundary condition has to be considered in
the weak sense described below.
Here and in the sequel, unless otherwise stated, we indicate with ⌦ an open, bounded
and regular subset of Rn .
Theorem 2.1. For any p 2 [1, 1) the restriction operator

T : C 1 (⌦) ! C 0 (@⌦) (2.10)

satisfies kT ukLp (@⌦)  C(p, ⌦)kukW 1,p (⌦) . Therefore it can be uniquely extended to a linear
and continuous operator from W 1,p (⌦) to Lp (@⌦).
Proof. We prove the result only in the case when ⌦ is the subgraph of a Lipschitz function
f inside the rectangle ⌦0 ⇥ (a, b), with ⌦ ⇢ Rn 1 open, with a0 = inf f > a, proving the
estimate on the portion
:= {(x0 , f (x0 )) : x0 2 ⌦0 }
of its boundary (here we use the notation x = (x0 , t) with x0 2 ⌦0 and t 2 (a, b)). The
general case can be easily achieved by a partition of unity argument.
By the fundamental theorem of calculus, for all t 2 (0, a0 a) we have
Z f (x0 ) p Z f (x0 )
0 0 0 0 0
p
|u(x , f (x ) t) u(x , f (x ))|  @xn u(x , r) dr  (b a) p 1
|@xn u(x0 , r)|p dr .
f (x0 ) t a

An integration w.r.t. x0 now gives

Z Z
|u(x0 , f (x0 ) t) u(x0 , f (x0 ))|p dx0  (b a) p 1
|@xn u|p dx ,
⌦0 ⌦
p
so that inserting the area element 1 + |rf (x0 )|2 and using the inequality |r + s|p 
2p 1 (|r|p + |s|p ) gives
Z Z Z
1 p p 1 0 0 p 0 p 1 p 1
p |u| d  2 |u(x , f (x ) t)| dx + 2 (b a) |@xn u|p dx ,
1+L 2
⌦0 ⌦

where L is the Lipschitz constant of f .

14
 Now we average this estimate with respect to t 2 (0, a0 a), together with the fact
that the determinant of the gradient of the map (x0 , t) 7! (x0 , f (x0 ) t) is identically equal
to 1, to get
Z Z Z
1 p 2p 1 p p 1 p 1
p |u| d  0 |u| dx + 2 (b a) |@xn u|p dx .
1 + L2 a a ⌦ ⌦

Because of the previous result, for u 2 W 1,p (⌦) we will interpret the boundary condi-
tion u|@⌦ = g as
T u = g. (2.11)
It can also be easily proved that T u is characterized by the identity
Z Z Z
@' @u
u dx = ' dx + ' T u ⌫i d 8' 2 C 1 (⌦) (2.12)
⌦ @xi ⌦ @xi @⌦

where ⌫ = (⌫1 , . . . , ⌫n ) is the unit normal vector, pointing out of ⌦. Indeed, using the
equality H 1,p (⌦) = W 1,p (⌦) of Theorem 1.5 one can start from the classical divergence
theorem with u 2 C 1 (⌦) and then argue by approximation.
Remark 2.2. It is possible to show that the previously defined restriction operator T is
not surjective if p > 1 and that its image can be described in terms of fractional Sobolev
spaces W s,p , characterized by the finiteness of the integral
Z Z
|u(x) u(y)|p
dxdy ,
|x y|n+sp

see [1], with s = 1 1/p. The borderline case p = 1 is special, and in this case Gagliardo
proved in [13] the surjectivity of T .
We can now mimic the argument described in the previous section in order to achieve
an existence result, provided the function g belongs to the image of T, that is there exists
e 2 W 1,2 (⌦) such that T u
a function u e = g. Indeed, if this is the case, our problem is
reduced to show existence of a solution for the equation
8 P
< v = fe e
↵ @↵ f↵ in ⌦;

:
v 2 H01,2 (⌦) .

where fe = f and fe↵ = f↵ @↵ u e. This is precisely the first problem we have discussed
above and so, denoted by v its unique solution, the function u = v + ue will satisfy both
our equation and the required boundary condition.

15
 Finally, let us discuss the so-called Neumann boundary conditions, involving the be-
haviour of the normal derivative of u on the boundary. We consider a problem of the
form 8 P P
↵
< ↵, @↵ (A @ u) + u = f ↵ @↵ f↵ in ⌦;

:
A↵ @ u⌫↵ = g on @⌦
with A↵ a real matrix and > 0 a fixed constant. For the sake of brevity, we just discuss
the case A↵ = ↵ so that the problem above becomes
8
>
< u + u = f in ⌦;

: @u = g
>
on @⌦.
@⌫
In order to give it a clear meaning, note that if u, v 2 C 1 (⌦) then
Z Z Z
@u
hru, rvi dx = v u dx + v d (2.13)
⌦ ⌦ @⌦ @⌫

and so in this case it is natural to ask that for any v 2 C 1 (⌦) the desired solution u
satisfies Z Z Z
[hru, rvi + uv] dx = vf dx + vg d . (2.14)
⌦ ⌦ @⌦

In order to obtain existence (and uniqueness) for this problem when g 2 L2 (@⌦), it is
enough to apply Riesz’s theorem to the bilinear form on H 1,2 (⌦)
Z
a(u, v) = [hru, rvi + uv] dx (2.15)
⌦

which is clearly equivalent to the standard HilbertR product on

R the same space (since
> 0) and the continuous linear functional F (v) = ⌦ vf dx + @⌦ vg d .

2.3 Elliptic systems

In order to deal with systems, we first need to introduce an appropriate notation. We
will consider functions u : ⌦ ⇢ Rn ! Rm and, consequently, we will use Greek letters
(say ↵, , . . .) in order to indicate the starting domain of such maps (so that ↵, 2
{1, 2, . . . , n}), while we will use Latin letters (say i, j, k, . . .) for the target domain (and
hence i, j 2 {1, 2, . . . , m}). In many cases, we will need to work with four indices matrices
(i.e. rank four tensors) like A↵ij , whose meaning should be clear from the context. Our
first purpose now is to see whether it is possible to adapt some ellipticity condition (having

16
the form A cI for some c > 0) to the vector-valued case. The first idea is to define the
Legendre condition X ↵
Aij ⇠↵i ⇠ j c |⇠|2 8⇠ 2 Rm⇥n (2.16)
↵, ,i,j

where R m⇥n
indicates, as above, the space of m ⇥ n real matrices. Let us apply it in order
to obtain existence and uniqueness for the system
8 P ↵ j
P ↵
< ↵, ,j @↵ (Ai,j @ u ) = fi ↵ @ ↵ fi i = 1, . . . , m
:
u 2 H01 (⌦; Rm )

with data fi , fi↵ 2 L2 (⌦).1 The weak formulation of the problem is obviously
Z X Z "X X
#
A↵ij @ uj @↵ 'i dx = f i 'i + fi↵ @↵ 'i dx (2.17)
⌦ i,j,↵, ⌦ i i,↵

m
for every ' 2 [Cc1 (⌦)] and i = 1, . . . , m. Now, if the matrix A↵ij is symmetric with respect
to the transformation (↵, i) ! ( , j) (which is implied, for instance, by the symmetries
in (↵, ) and (i, j)), then it defines a scalar product on H01 (⌦; Rm ) by the formula
Z X
(', ) = A↵ij @↵ 'i @ j dx. (2.18)
⌦ i,j,↵,

If, moreover, A satisfies the Legendre condition (2.16) for some c > 0, it is immediate to
see that this scalar product is equivalent to the standard one (with A↵i,j = ↵ ij ) and so
we are led to apply again Riesz’s theorem to conclude the proof.
From now on, we will often adopt Einstein’s summation convention on repeated in-
dices, using it without explicit mention.
It should be noted that in the proof of some existence result (and, in particular, in
the scalar case) the symmetry hypothesis w.r.t. the transformation (↵, i) ! ( , j) is not
necessary, since we can exploit the following:
Theorem 2.3 (Lax-Milgram). Let H be a (real) Hilbert space and let a : H ⇥ H ! R a
bilinear, continuous and coercive form so that

a(u, u) |u|2 8u 2 H ,
0
for some > 0. Then for any F 2 H there exists uF 2 H such that a(uF , v) = F (v) for
all v 2 H.
1
Note that we sometimes omit the Sobolev exponent when this is equal to two: for instance H01 (⌦)
stands for H01,2 (⌦).

17
Proof. By means of the standard Riesz’s theorem it is possible to define a linear operator
T : H ! H such that
a(u, v) = hT u, vi 8u, v 2 H
and such T is continuous since

kT uk2 = hT u, T ui = a(u, T u)  C kuk kT uk ,

where C is a constant of continuity for a(·, ·) and hence kT k  C. Now we introduce the
auxiliary bilinear form
a(u, v) = hT T ⇤ u, vi = hT ⇤ u, T ⇤ vi ,
e
which is obviously symmetric and continuous. Moreover, thanks to the coercivity of a we
have that e
a is coercive too, because
p
kuk2  a(u, u) = hT u, ui = hu, T ⇤ ui  kuk kT ⇤ uk = kuk ea(u, u)

and so e
a(u, u) 2
kuk2 . Since e
a determines an equivalent scalar product on H we can
apply again Riesz Theorem to obtain a vector u0F 2 H such that

a(u0F , v) = F (v) 8 v 2 H .
e

a the thesis is achieved setting uF = T ⇤ u0F :

By the definitions of T and e

F (v) = ã(u0F , v) = hT ⇤ u0F , T ⇤ vi = hT uF , vi = a(uF , v) 8v 2 H .

⇤
As indicated above, we now want to formulate a di↵erent notion of ellipticity for the
vector case. To this aim, it is useful to analyse more in detail the scalar case. We have
the two following conditions:

(E) A I that is hAv, vi |v|2 for all v 2 Rm⇥n (ellipticity);

R R
(C) aA (u, u) = ⌦ hAru, rui dx ⌦
|ru|2 dx for all u 2 H01 (⌦; Rm ) (coercivity).

It is obvious by integration that (E) ) (C) and we may wonder about the converse
implication. As we will see below, this holds in the scalar case (m = 1) and fails in the
vectorial case (m > 1).

Proposition 2.4. Let (C) and (E) as above. Then, (C) is equivalent to (E).

18
Proof. Let is prove that (C) implies (E). The computations become more transparent if
we work with functions having complex values, and so let us define for any u, v 2 H01 (⌦; C)
Z Z n
X
⌦ ↵
aA (u, v) = Aru, rv dx = A↵ @↵ u@ u dx .
⌦ ⌦ ↵, =1

A simple computation shows that (here ru 2 Cn stands for r<u + ir=u, where <u and
=u are respectively the real and imaginary part of u)

<aA (u, u) = aA (<u, <u) + aA (=u, =u) .

Hence, (C) implies Z

<aA (u, u) |ru|2 dx . (2.19)
⌦

Now consider a function ' 2 Cc1 (⌦) and define u⌧ (x) = '(x)ei⌧ x·⇠ . We have that
Z Z
1 2 ↵ ↵
<aA (u⌧ , u⌧ ) = ' A ⇠↵ ⇠ dx + o⌧ = A ⇠↵ ⇠ '2 dx + o⌧
⌧2 ⌦ ⌦

with o⌧ ! 0 as ⌧ ! +1, and

Z Z
1 2
2
|ru⌧ | dx = '2 |⇠|2 dx + o⌧ (1) .
⌧ ⌦ ⌦

Hence, exploiting our coercivity assumption and letting ⌧ ! +1 in (2.19) we get

Z
↵ 2
A ⇠↵ ⇠ |⇠| '2 dx 0 (2.20)
⌦

which immediately implies the thesis (it is enough to choose ' not identically zero).
Actually, every single part of our discussion is still true in the case when A↵ = A↵ (x)
is Borel and L1 function in ⌦ and we can conclude that (E) holds for a.e. x 2 ⌦: we just
need to choose, in the very last step, an appropriate sequence of rescaled and normalized
mollifiers concentrating around x0 , for any Lebesgue point x0 of A. The conclusion comes,
in this situation, by Lebesgue di↵erentiation theorem.
For the reader’s convenience we recall here some basic facts concerning Lebesgue points
(see also Section 13). Given f 2 L1loc (Rn ) and x0 2 Rn , we say that x0 is a Lebesgue point
for f if there exists 2 R such that
Z
lim |f (y) | dy = 0. (2.21)
r#0 Br (x0 )

19
In this case is unique and it is sometimes written

= fe(x0 ) = g
lim f (x). (2.22)
x!x0

Notice that both the notion of Lebesgue point and fe are invariant in the Lebesgue
equivalence class of f . The Lebesgue di↵erentiation theorem says that for L n -a.e. x0 2 Rn
the following two properties hold: x0 is a Lebesgue point of f and fe(x0 ) = f (x0 ). Notice
however that the validity of the second property at a given x0 does depend on the choice
of a representative of f in the Lebesgue equivalence class.
Going back to the previous discussion, it is very interesting to note that the argument
above does not give a complete equivalence when m > 1: in fact, the coercivity condition
Z
aA (u, u) |ru|2 dx u 2 H 1 (⌦; Rm ) (2.23)
⌦

can be applied to test functions having the form u⌧ (x) = '(x)bei⌧ x·a with a 2 Rn and
b 2 Rm and implies the Legendre-Hadamard condition

A↵ij ⇠↵i ⇠ j |⇠|2 for all ⇠ = a ⌦ b , (2.24)

that is the Legendre condition restricted to rank one matrices ⇠↵i = a↵ bi . Explicit exam-
ples show that the Legendre-Hadamard condition is in general strictly weaker than the
Legendre condition.

Example 2.5. When m = n = 2, consider the tensor A↵ij implicitly defined by

A↵ij ⇠↵i ⇠ j = det(⇠) + " |⇠|2 (2.25)

with " 0. Since t 7! det(M + tN ) is linear for any rank one matrix N , the Legendre-
Hadamard condition with = " is fulfilled. On the other hand our quadratic form,
restricted to diagonal matrices with eigenvalues t and t, equals

t2 + 2t2 " .

It follows that the Legendre condition with = 0 fails when 2" < 1.

Nevertheless, the Legendre-Hadamard condition is sufficient to imply coercivity:

Theorem 2.6 (Gårding). Assume that A↵ij satisfies the Legendre-Hadamard condition
for some positive constant and the symmetry condition A↵ij = A ↵ ji. Then aA (u, u)
R
|ru|2 dx for all u 2 H 1 (Rn ; Rm ).

20
 In the proof of Gårding’s theorem, we denote by S(Rn ) the Schwartz space of smooth
C-valued functions that decay at 1 faster than any polynomial, and by ' b and ' e the
Fourier transform of ' and its inverse, respectively
Z
n/2
b = (2⇡)
'(⇠) '(x)e ix·⇠ dx (2.26)

and Z
n/2
e
'(x) = (2⇡) '(⇠)eix·⇠ d⇠ . (2.27)

We will also make use of the Plancherel identity:

Z Z
'b b d⇠ = ' dx 8', 2 S(Rn ) . (2.28)

Proof. By density it is enough to prove the result when u 2 [Cc1 (Rn )]m . In this case we
use the representation Z
n/2
u(⇠) = (2⇡) '(x)e ix·⇠ dx ,
Rn
m
b
that is u(⇠) = '(⇠) for some ' 2 [S(Rn )] . Consequently,

@↵ uj (⇠) = ix[ j
↵' ,

hence
Z Z Z
@uj @ul
aA (u, u) = A↵jl d⇠ = i 2
A↵jl x[ j dl
↵ ' x ' d⇠ = A↵jl (x↵ 'j )(x 'l ) dx ,
Rn @⇠↵ @⇠ Rn Rn

the last passage being due to the Plancherel identity (2.28). Now we can apply our
hypothesis to get
A↵jl a↵ bj a bl |a|2 |b|2
with a = x 2 Rn and b = ' 2 Cn , so that
Z
aA (u, u) |x|2 |'(x)|2 dx . (2.29)
Rn

If we perform the same steps with ↵ jl in place of A↵jl we see at once that
Z Z
2
|ru| (⇠) d⇠ = |x|2 |'(x)|2 dx . (2.30)
Rn Rn

Comparing (2.29) and (2.30) we conclude the proof. ⇤

21
Remark 2.7. Gårding’s theorem marks in some sense the di↵erence between pointwise
and integral inequalities. It is worth mentioning some related inequalities that are typi-
cally nonlocal, namely, they do not arise from the integration of a pointwise inequality.
An important example is Korn’s inequality:
Z Z p
p ru + (ru)t
|ru| dx  c(n, p) dx for all u 2 Cc1 (Rn , Rn ) , (2.31)
R n R n 2

for p 2 (1, 1). A variant of this example is the Korn-Poincaré inequality: if ⌦ is an open,
bounded and regular set in Rn and p 2 (1, 1), then
Z Z p
p ru + (ru)t
inf |u(x) Bx c| dx  C(⌦, p) dx . (2.32)
c2Rm , B t = B ⌦ ⌦ 2

2.4 Necessary minimality conditions

The importance of the Legendre-Hadamard condition is also clear from a variational
perspective. Indeed, let u : ⌦ ⇢ Rn ! Rm be a locally Lipschitz function, that is
1,1
u 2 Wloc (⌦; Rm ), fix a Lagrangian L and define a functional
Z
F (u, ⌦) = L(x, u, ru) dx .
⌦

We say that u is a local minimizer for F if

F (u, A)  F (v, A) 1,1

for all v 2 Wloc (⌦; Rm ) such that {v 6= u} b A b ⌦. (2.33)

We will make the following standard assumptions on the Lagrangian: we assume that
L : ⌦ ⇥ Rm ⇥ Rm⇥n ! R is Borel and, denoting the variables as (x, s, p), we assume that
L is of class C 1 in (s, p) with

sup (|L| + |Ls | + |Lp |) < 1 (2.34)

K

for any domain K = ⌦0 ⇥ {(s, p)| |s| + |p|  R} with R > 0 and ⌦0 b ⌦. In this case it is
possible to show that the map
Z
t 7! L(x, u + t'ru + tr') dx
⌦0

1,1
is of class C 1 for all u, ' 2 Wloc (⌦; Rm ) and ⌦0 b ⌦, and its derivative equals
Z
Ls (x, u + t', ru + tr') · ' + Lp (x, u + t', ru + tr') · r' dx
⌦0

22
(the assumption (2.34) is needed to di↵erentiate under the integral sign). As a conse-
quence, if a locally Lipschitz function u is a local minimizer and ' 2 Cc1 (⌦0 ; Rm ), since
F (u, ⌦0 )  F (u + t', ⌦0 ) we can di↵erentiate at t = 0 to obtain
Z "X X i
#
@'
Lsi (x, u, ru)'i + Lp↵i (x, u, ru) dx = 0 . (2.35)
⌦ 0
i ↵,i
@x ↵

Hence, exploiting the arbitrariness of ', we obtain the Euler-Lagrange equations in the
weak sense:
@
Lp↵ (x, u, ru) = Lsi (x, u, ru) i = 1, 2, . . . , m .
@x↵ i
Exploiting this idea, we can associate to many classes of PDEs appropriate energy func-
tionals, so that the considered problem is nothing but the Euler-Lagrange equation for
the corresponding functional. For instance, neglecting the boundary conditions (that can
actually be taken into account by an appropriate choice of the ambient functional space),
equations having the form
u = g(x, u) (2.36)
derive from the functional
Z s
1
L(x, s, p) = |p|2 g(x, r) dr . (2.37)
2 0

Adding stronger hypotheses on the Lagrangian L, in analogy with what has been done
above, i.e. requiring that

sup (|Lss | + |Lsp | + |Lpp |) < 1

K

for any domain K = ⌦0 ⇥{(s, p)| |s| + |p|  R} with ⌦0 b ⌦, we can find another necessary
minimality condition corresponding to
d2
F (u + t') 0,
dt2 t=0

namely
Z
0  (', ') = [Ar'r' + Br' · ' + C' · '] dx 8' 2 Cc1 (⌦; Rm ) , (2.38)
⌦

where the dependence on x and all indices are omitted for brevity and
8
>
<A(x) = Lpp (x, u(x), ru(x)) ;
B(x) = Lps (x, u(x), ru(x)) ; (2.39)
>
:
C(x) = Lss (x, u(x), ru(x)) .

23
We can finally obtain pointwise conditions on the local minimizer u by means of the
following theorem, whose proof can be obtained arguing as in the proof that coercivity
implies ellipticity (Proposition 2.4).

Theorem 2.8. Consider the bilinear form on H01 (⌦; Rm ) defined by

Z
⇥(u, v) = (Arurv + Bru · v + Cu · v) dx , (2.40)
⌦

where A = A↵ij (x), B = Bij↵ (x) and C = Cij (x) are Borel and L1 functions. If ⇥(u, u)
0 for all u 2 H01 (⌦; Rm ) then A(x) satisfies the Legendre-Hadamard condition with = 0
for a.e. x 2 ⌦.

Hence, in our case, we find that Lpp (x, u(x), ru(x)) satisfies the Legendre-Hadamard
condition with = 0 for a.e. x 2 ⌦.

3 Lower semicontinuity of integral functionals

Tonelli’s theorem is a first powerful tool leading to an existence result for minimizers of
integral functionals of the form
Z
F (u) := L(x, u(x), ru(x)) dx (3.1)
⌦

in suitable function spaces (including for instance the boundary conditions).

Before stating Tonelli’s theorem, we recall some useful facts about uniformly integrable
maps. A comprehensive treatment of this subject can be found for instance in [28], see
also [3, Theorem 1.38].

Theorem 3.1 (Dunford-Pettis). Let (X, A, µ) be a finite measure space and let F ⇢
L1 (X, A, µ). Then the following facts are equivalent:

(i) the family F is sequentially relatively compact w.r.t. the weak-L1 topology;

(ii) there exists : [0, 1) ! [0, 1], with (t)/t ! +1 as t ! 1, such that
Z
(|f |) dµ  1 8f 2 F ;
X

(iii) F is uniformly integrable, i.e.

Z
8" > 0 9 >0 s.t. µ(A) < =) |f | dµ < " 8f 2 F .
A

24
Theorem 3.2 (Tonelli). Let L(x, s, p) : ⌦ ⇥ Rm ⇥ Rm⇥n be a Lagrangian satisfying the
following properties:
(1) L is non-negative;
(2) L is lower semicontinuous w.r.t. s and the partial derivatives Lp↵i exist and are
continuous w.r.t. s;
(3) p 7! L(x, s, ·) is convex2 .
Then any sequence (uh ) ⇢ W 1,1 (⌦; Rm ) converging to u in L1 (⌦; Rm ) with |ruh | uniformly
integrable satisfies the lower semicontinuity inequality
lim inf F (uh ) F (u) . (3.2)
h!1
Proof. We start by noticing that there is a subsequence uh(k) such that
lim inf F (uh ) = lim F (uh(k) )
h!1 k!1

and, possibly extracting one more subsequence,

uh(k) ! u a.e. in ⌦ .
Thanks to the Dunford-Pettis Theorem we can also assume the weak-L1 convergence
ruh(k) * g in L1 (⌦; Rm⇥n ) .
Passing to the limit in the integration by parts formula, this immediately implies that u
belongs to W 1,1 (⌦; Rm ) and that ru = g.
Thanks to Egorov’s Theorem, for all " > 0 there exists a compact subset K" ⇢ ⌦ such
that
• |⌦ \ K" | < ";
• Lp (x, uh(k) (x), ru(x)) ! Lp (x, u(x), ru(x)) uniformly on K" ;
• Lp (x, u(x), ru(x)) is bounded on K" .
Because of the convexity hypothesis (3) and the non-negativity of L, we can estimate
Z
lim inf F (uh ) = lim L(x, uh(k) (x), ruh(k) (x)) dx
h!1 k!1 ⌦
Z
lim inf L(x, uh(k) (x), ruh(k) (x)) dx
k!1
Z K"
⇥ ⌦ ↵⇤
lim inf L(x, uh(k) (x), ru(x)) + Lp (x, uh(k) (x), ru(x)), ruh(k) (x) ru(x) dx
k!1 K"
Z Z
⇥ ⌦ ↵⇤
L(x, u(x), ru(x)) dx + lim inf Lp (x, u(x), ru(x)), ruh(k) (x) ru(x) .
K" k!1 K"
2
We will see that this assumption can be considerably weakened.

25
Hence, the weak convergence ruh(k) * ru ensures that
Z
lim inf F (uh ) L(x, u(x), ru(x)) dx
h!1 K"

and as " ! 0 we achieve the desired inequality (3.2). ⇤

Before stating the following corollary we recall Rellich’s theorem (see Theorem 1.10)
which provides the compactness of the inclusion W 1,1 (⌦) ⇢ L1 (⌦) whenever ⌦ ⇢ Rn is
an open, bounded and regular set.
Corollary 3.3. Let ⌦ ⇢ Rn be an open, bounded and regular set and let L be a Borel
Lagrangian satisfying hypotheses (2), (3) from Theorem 3.2 and
(1’) L(x, s, p) (|p|) + c|s| for some : [0, 1) ! [0, 1] with lim (t)/t = 1, c > 0.
t!1

Then the problem

min F (u)| u 2 W 1,1 (⌦; Rm )
admits a solution.
Proof. It is a classical application of the direct method of Calculus of Variations, where
hypothesis (10 ) provides the sequential relative compactness of sublevels {F  t} with
respect to the so-called sequential weak-W 1,1 topology (i.e. strong convergence in L1 of
the functions and weak convergence in L1 of their gradients) and semicontinuity is given
by Theorem 3.2. ⇤
At this point one could ask whether the convexity assumption in Theorem 3.2 is na-
tural. The answer is negative: as the Legendre-Hadamard condition is weaker than the
Legendre condition, here we are in an analogous situation and Example 2.5 fits again. Let
us define a weaker, although less transparent, convexity condition, introduced by Morrey.
Definition 3.4 (Quasiconvexity). A continuous function F : Rm⇥n ! R is said to be
quasiconvex at A 2 Rm⇥n if for all ⌦ ⇢ Rn open and bounded it holds
Z
F (A + r') dx F (A) 8 ' 2 Cc1 (⌦; Rm ) . (3.3)
⌦

We say that F is quasiconvex if it is quasiconvex at every point A 2 Rm⇥n .

Remark
R 3.5. Obviously we can replace the left-hand side in (3.3) with the quantity
{r'6=0}
F (A + r') dx: this follows from the equality
Z ✓ ◆ Z
|{r' 6= 0}| |{r' 6= 0}|
F (A + r') dx = 1 F (A) + F (A + r') dx .
⌦ |⌦| |⌦| {r'6=0}

This proves that the dependence from ⌦ of this notion is only seeming. Another way to
see this relies on the observation that whenever (3.3) is valid for ⌦, then:

26
 • it is valid for every ⌦0 ⇢ ⌦, thanks to the previous observation;
• it is valid for x0 + ⌦, for x0 2 Rn and > 0, considering the transformation
'(x) 7! '(x0 + x)/ .
Finally, a simple approximation argument gives
Z
F (A + r') dx F (A) 8 ' 2 Cc1 (⌦; Rm ) . (3.4)
⌦

The definition of quasiconvexity is related to Jensen’s inequality, which we briefly

recall here.
Theorem 3.6R (Jensen). Let us consider a probability measure µ on a convex domain
X ⇢ Rp , with X |y| dµ(y) < 1, and a convex, lower semicontinuous function F : X !
R [ {+1}. Then Z ✓Z ◆
F (y) dµ(y) F y dµ(y) . (3.5)
X X

Notice that the inequality above makes sense: either F ⌘ +1 or it is finite at least
one point. In the second case the negative part of F has at most linear growth and the
integral in the left hand side makes sense, finite or infinite.
Now, let f 2 L1 (⌦, Rm⇥n ) and consider the law µ of the map f with respect to the
rescaled Lebesgue measure L n /L n (⌦). If F : Rm⇥n ! R[{+1} is lower semicontinuous
and convex, thanks to Jensen’s inequality one has
Z Z ✓Z ◆ ✓Z ◆
F (f (x)) dx = F (y) dµ(y) F y dµ(y) = F f dx . (3.6)
⌦ Rm⇥n Rm⇥n ⌦

Quasiconvexity should be considered as a weak version of convexity: indeed, if F is convex

then the inequality (3.6) holds for all maps f , thanks to Jensen’s inequality; on the other
hand the condition (3.3) concerns only gradient maps (more precisely gradients of maps
coinciding with an affine function on the boundary of the domain). If we go back to the
formulation (3.5), we should say that quasiconvexity should be understood as (3.5) for
measures µ in Rmn generated by gradient maps.
Proposition 3.7. Any convex lower semicontinuous function F : Rm⇥n ! R [ {+1} is
quasiconvex.
Proof. Fix ' 2 Cc1 (⌦; Rm ) and consider the law µ of the map x 7! f (x) = A + r'(x)
with respect to the rescaled Lebesgue measure L n /L n (⌦). Since r' is bounded the
measure µ has compact support and one has
Z Z
y dµ(y) = A + r'(x) dx = A .
Rn ⌦

From (3.6) we conclude. ⇤

27
Remark 3.8. The following chain of implications holds:

convexity =) quasiconvexity =) Fpp (A) satisfies Legendre-Hadamard with =0.

All these notions are equivalent when either n = 1 or m = 1; more generally:

• An integration by parts easily yields

Z
(A + detr') dx = A|⌦| 8' 2 Cc1 (⌦; Rn ) .
⌦

Hence, Example 2.5 provides a quasiconvex function that is not convex when n =
m = 2, and considering the determinant of a 2 ⇥ 2 minor the example fits also the
case min{m, n} 2;

• when max{n, m} 3 and min{n, m} 2, there exist highly nontrivial examples

showing that the Legendre-Hadamard condition does not imply quasiconvexity;

• the equivalence between Legendre-Hadamard condition and quasiconvexity is still

open for n = m = 2.

Let us recall that we introduced quasiconvexity as a “natural” hypothesis to improve

Morrey-Tonelli’s theorem. The following Theorem 3.12 confirms this fact.

Definition 3.9 (w⇤ -convergence in W 1,1 ). Let us consider an open set ⌦ ⇢ Rn and
fk 2 W 1,1 (⌦). We write fk ! f in w⇤ -W 1,1 (⌦) if

• fk ! f uniformly in ⌦;

• krfk kL1 is uniformly bounded.

⇤
Proposition 3.10. If fk ! f in w⇤ -W 1,1 (⌦), then f 2 W 1,1 (⌦) and rfk * rf .

This is a direct consequence of the fact that (rfk ) is sequentially compact in the w⇤ -
topology of L1 , and any weak⇤ limit provides a weak derivative of f (hence f 2 W 1,1 ,
the limit is unique and the whole sequence of derivatives w⇤ -converges). Obviously an
analogous statement holds for Rm -valued maps.
Before stating Morrrey’s lower semicontinuity theorem we give a quick proof of Rademacher’s
di↵erentiability theorem.

Theorem 3.11 (Rademacher). Any locally Lipschitz function f : ⌦ ⇢ Rn ! Rm is

di↵erentiable L n -a.e. and its di↵erential coincides L n -a.e. with the weak gradient.

28
 R
Proof. Fix a Lebesgue point x0 of the weak gradient ru, namely Br (x0 ) |ru L| dx ! 0
as r # 0 for some linear map L : Rn ! Rm . We shall prove that f is di↵erentiable at x0
and that the (classical) gradient rf at x0 is equal to L.
First of all, it is easy to see that this property can be equivalently stated as follows:

fr (y) ! L(y) uniformly on B 1 as r # 0 ,

where fr (y) = (f (x0 + ry) f (x0 ))/r are the rescaled maps. Notice, that fr are equi-
Lipschitz in B 1 and equi-bounded (because fr (0) = 0), hence fr is relatively compact in
C 0 (B 1 ) as r # 0. Hence, suffices to show that any limit point f0 (y) = limi fri (y) coincides
with L(y). A simple change of variables shows that (understanding here gradients as weak
gradients!) Z Z
|rfr L| dy = |rf L| dx .
B1 Br (x)

It follows that rfr ! L in L1 (B1 ; Rm⇥n ), hence rf0 = L L n -a.e. in B1 . By the

constancy theorem we get f0 (y) = L(y) + c for some constant c, which obviously should
be 0 because f0 (0) = limi fri (0) = 0.

Theorem 3.12 (Morrey). Assume that L : ⌦ ⇥ Rm ⇥ Rmn ! [0, 1) is continuous,

and that the functional F in (3.1) is lower semicontinuous w.r.t. the w⇤ -W 1,1 (⌦; Rm )
convergence at some function u. Then L(x, u(x), ·) is quasiconvex at ru(x) for almost
every x 2 ⌦.
Conversely, under the same assumptions on L, if L(x, s, ·) is quasiconvex for all (x, s) 2
⌦ ⇥ Rm , then F is lower semicontinuous w.r.t. w⇤ -W 1,1 (Q; Rm ) convergence.
Proof. (Necessity of quasiconvexity) It is sufficient to prove the result for any Lebesgue
point x0 2 ⌦ of ru. The main tool is a blow-up argument: if Q = ( 1/2, 1/2)n is the
unit cube centered at 0, Qr (x0 ) = x0 + rQ ⇢ ⌦ and v 2 W01,1 (Q, Rm ), we set
Z
Fr (v) := L(x0 + ry, u(x0 + ry) + rv(y), ru(x0 + ry) + rv(y)) dy .
Q

The formal limit as r # 0 of Fr , namely

Z
F0 (v) := L(x0 , u(x0 ), ru(x0 ) + rv(y)) dy
Q

is lower semicontinuous at v = 0 with respect to the w⇤ -W 1,1 (Q; Rm ) convergence because

of the following two facts:

29
 • each Fr is lower semicontinuous at 0 with respect to the w⇤ -W 1,1 (Q; Rm ) conver-
gence, indeed
Z
1
Fr (v) = n L (x, u(x) + rv ((x x0 )/r) , ru(x) + rv ((x x0 )/r)) dx
r Qr (x0 )
✓ Z ◆
1
= n F (u + rv ((x x0 )/r)) L(x, u(x), ru(x)) dx ;
r ⌦\Qr (x0 )

• being x0 a Lebesgue point for ru, for any bounded sequence (vh ) ⇢ W01,1 (Q; Rm )
it is easily checked that the continuity of L gives

lim sup |Fr (vh ) F0 (vh )| = 0 .

r!0+ h

Let us introduce the auxiliary function

H(p) := L(x0 , u(x0 ), ru(x0 ) + p) .

Given a test function ' 2 Cc1 (Q, Rm ), we work with the 1-periodic function such that
|Q = ' and the sequence of highly oscillating (because h 1 -periodic) functions

1
vh (x) := (hx) ,
h
⇤
which obviously converge uniformly to 0. Since rvh (x) = r (hx) we have also vh * 0
in W 1,1 (Q; Rm ), so that thanks to the lower semicontinuity of F0 at 0 one has
Z Z
n
H(0) = F0 (0)  lim inf H(rvh (x)) dx = lim inf h H(r (y)) dy
h!1 Q h!1 Qh
Z Z
= H(r (y)) dy = H(r'(y)) dy ,
Q Q

which is exactly the quasiconvexity property for L(x, u(x), ·) at ru(x0 ).

(Sufficiency of quasiconvexity) We split the proof in several steps, reducing ourselves
to progressively simpler cases. First, since any open set ⌦ can be monotonically ap-
proximated by bounded open sets with closure contained in ⌦, we can assume that ⌦ is
bounded and that L 2 C 0 (⌦ ⇥ Rm ⇥ Rmn ). Since ⌦ can be written as the disjoint union
of half-open disjoint cubes, by the superadditivity of the lim inf we can also assume that
⌦ = Q is a n-cube with side length `. We also set

M := sup |(x, ruh (x))| : x 2 ⌦, h 2 N .

30
Now, considering the decomposition
⇥ ⇤
L(x, uh (x), ruh (x)) = L(x, uh (x), ruh (x)) L(x, u(x), ruh (x)) + L(x, u(x), ruh (x))

we see immediately that we need only to consider Lagrangians L1 (x, p) independent of s

(just take L1 (x, p) = L(x, u(x), p)).
The next step is to reduce ourselves to Lagrangians independent of x. To this aim, con-
sider a modulus of continuity for L1 in the ball B M and a decomposition of Q in 2kn cubes
Qi with side length `/2k and centers xi . Then, adding and subtracting L(xi , ruh (xi )) and
using once more the superadditivity of lim inf, yields
Z X Z p
n` X n
lim inf L1 (x, ruh (x)) dx lim inf L1 (xi , ruh (x)) dx !( k ) L (Qi ).
h!1 Q i
h!1 Q i
2 j
P
Since i L n (Qi ) = `n , if we are able to show that for any i it holds
Z Z
lim inf L1 (xi , ruh (x)) dx L1 (xi , ru(x)) dx
h!1 Qi Qi

we obtain
Z XZ p
n`
lim inf L1 (x, ruh (x)) dx L1 (xi , ru(x)) dx !( )
h!1 Q i Qi 2k
Z p
n`
L1 (x, ru(x)) dx 2!( ).
Q 2k

As k ! 1 we recover the liminf inequality.

Hence, we are led to show the lower semicontinuity property for Lagrangians L2 (p) =
L1 (xi , p) independent of x. In this proof we shall use the fact that continuous quasiconvex
functions are locally Lipschitz. This property can be obtained noticing that bounded
convex functions w are Lipschitz, with the quantitative estimate
supB2r (x) w inf B2r (x) w
Lip(w, Br (x))  ,
r
and quasiconvex functions g satisfy the Legendre-Hadamard condition, hence g(p) as
function of p↵i is convex.
Now, let us consider a quasiconvex Lagrangian L2 (p). We consider two cases: first,
the case when the limit function u is affine and then, by a blow-up argument again, the
general case. Assume now that u is affine, let A = ru and consider a smooth function

31
 : ⌦ ! [0, 1] with compact support. We can apply the quasiconvexity inequality to
' = (uh u) and the local Lipschitz property with R(p) = L2 (p) L2 (0) to get
Z
R(A)  R((1 )A + ruh + (uh u) ⌦ r ) dx
⌦ Z Z Z
 C(|A| + kruh k1 ) (1 ) dx + Ckr k1 |uh u| dx + R(ruh ) ,
⌦ ⌦ ⌦
so that passing to the limit first as h ! 1 and then as " 1 gives the result.
Finally, we consider the generalR case, using Rademacher’s theorem and a blow-up
argument. Assume that the lim inf ⌦ L2 (ruh ) dx is a limit, that we call L, and consider
the family of measures µh := L2 (ruh )L n . Being this family bounded, we can assume
with no loss of generality that µh weakly converge, in the duality with Cc (⌦), to some
measure µ. Recall that the evaluations on compact sets K and open sets A are respectively
upper and lower semicontinuous w.r.t. weak convergence, i.e.
µ(K) lim sup µh (K), µ(A)  lim inf µh (A) . (3.7)
h!1 h!1

In particular µ(⌦)  L, so that if we show that µ L2 (ru)L n we are done. By

Lebesgue’s di↵erentiation theorem for measures, suffices to show that
µ(B r (x0 ))
lim inf L2 (ru(x0 )) for a.e. x0 2 ⌦ . (3.8)
r#0 !n rn
We shall prove this property at any di↵erentiability point x0 of u. To this aim, let ri ! 0
be a sequence on which the liminf is achieved, and " > 0. For any i we can choose hi i
so large that
Z Z
rin ri
L2 (ruhi ) dx  µ(B ri (x0 )) + , |uhi u| dx  . (3.9)
Bri (x0 ) i Bri (x0 ) i
Now, rescale as follows
uhi (x0 + ri y) u(x0 ) u(x0 + ri y) u(x0 )
vi (y) := , wi (y) :=
ri ri
to obtain functions vi satisfying
Z Z
µ(B ri (x0 )) 1
L2 (rvi ) dy  + , |vi wi | dy ! 0 .
B1 rin i B1

Since wi (y) ! ru(x0 )(y) uniformly in B 1 , thanks to the di↵erentiability assumption, we

obtain that vi converge to the linear function y 7! ru(x0 )y in L1 (B1 ; Rm ). Therefore
Z
µ(B ri (x0 )) 1
lim inf n
lim inf L2 (rvi ) dy !n L2 (ru(x0 )) .
i!1 ri i!1 B1 i
⇤

32
 The previous result shows that quasiconvexity of the Lagrangian is equivalent to se-
quential lower semicontinuity of the integral functional in the weak⇤ -W 1,1 convergence.
However, in many problems of Calculus of Variations only L↵ bounds, with ↵ < 1, are
available on the gradient. A remarkable improvement of Morrey’s result is the following:
Theorem 3.13 (Acerbi-Fusco). Suppose that a Borel Lagrangian L(x, s, p) is continuous
in (s, p) and satisfies

0  L(x, s, p)  C(1 + |s|↵ + |p|↵ ) 8(x, s, p) 2 ⌦ ⇥ Rm ⇥ Rmn

for some ↵ > 1 and some constant C. Suppose also that the map p 7! L(x, s, p) is
quasiconvex for all (x, s). Then F is sequentially lower semicontinuous w.r.t. the weak
W 1,↵ (⌦; Rm )-topology.

4 Regularity Theory
We begin studying the local behaviour of (weak) solutions of the system of equations
⇢
@↵ A↵ij @ uj = fi @↵ Fi↵ i = 1, . . . , m
(4.1)
u 2 Hloc (⌦; R )
1 m

with A↵ij 2 L1 (⌦), fi 2 L2loc (⌦) and Fi↵ 2 L2loc (⌦). From now on we shall use | · | for the
Hilbert-Schmidt norm of matrices and tensors, even though some estimates would still be
valid with the (smaller) operator norm.
Theorem 4.1 (Caccioppoli-Leray inequality). If the Borel coefficients A↵ij satisfy the
Legendre condition (L) with > 0 and

sup |A↵ij (x)|  ⇤ < 1 ,

x2⌦

then there exists a positive constant c = c( , ⇤) such that for any ball BR (x0 ) b ⌦ and
any k 2 Rm it holds
Z Z Z Z
2 2 2 2 2
c |ru| dx  R |u(x) k| dx+R |f (x)| dx+ |F (x)|2 dx .
BR/2 (x0 ) BR (x0 ) BR (x0 ) BR (x0 )
(4.2)
Before proceeding to the proof, some remarks are in order.
Remark 4.2. (1) The validity of (4.2) for all k 2 Rm depends on the translation in-
variance of the PDE. Moreover, the inequality (and the PDE as well) has a natural
scaling invariance: if we think of u as an adimensional quantity, then all sides have
dimension lengthn 2 , because f ⇠ length 2 and F ⇠ length 1 .

33
 (2) The Caccioppoli-Leray inequality is meaningful because for a general function u the
gradient ru can not be controlled by the variance of u! Precisely because of this
fact we can expect that several useful (regularity) informations can be drawn from
it. We will see indeed that CL inequalities are very “natural” and useful in the
context of regularity theory.
Remark 4.3 (Absorbtion scheme). In the regularity theory it often happens that one
can estimate, for some ↵ < 1,
A  BA↵ + C .
The absorption scheme allows to bound A in terms of B, C and ↵ only and works as
follows: by the Young inequality
b "p ap bq 1 1
ab = "a  + q (with + = 1)
" p "q p q
for p = 1/↵ one obtains
"p A B q
A  BA↵ + C  + q +C.
p "q
"p 1
Now, if we choose " = "(p) sufficiently small, so that  , we get
p 2
Bq
A2 + 2C .
"q q
Let us prove Theorem 4.1.
Proof. Without loss of generality, we can consider x0 = 0 and k = 0. As typical in
regularity theory, we choose test functions depending on the solution u itself, namely
:= u⌘ 2
where ⌘ 2 Cc1 (BR ), ⌘ ⌘ 1 in BR/2 , 0  ⌘  1 and |r⌘|  4/R.
Since u solves (4.1), we have that
Z Z Z
Arur f F ·r =0 (4.3)

where integrations are understood to be on BR . Moreover

r = ⌘ 2 ru + 2⌘u ⌦ r⌘ , (4.4)
so completing (4.3) with (4.4) we obtain
Z Z Z Z Z
2 2
⌘ Aru ru + 2 ⌘Aru (u ⌦ r⌘) f ⌘ F ru 2 ⌘F (u ⌦ r⌘) = 0 . (4.5)

Let us deal with each addendum separately.

34
 • By the Legendre condition
Z Z
⌘ 2
A↵ij i
@↵ u @ u j
⌘ 2 |ru|2 .
BR BR

• We have
Z Z Z
8⇤
2 ⌘Aru (u ⌦ r⌘)  2 ⌘|A||ru||u||r⌘|  (⌘|ru|) |u|
R
Z Z
4⇤" 2 2 4⇤
 ⌘ |ru| + |u|2 ,
R R"
where the first estimate is due to Schwarz inequality, the second one relies on the
boundedness of coefficients A↵ij and the estimate on |r⌘|, and the third one is based
on the Young inequality.

• By the Young inequality

Z Z Z Z
2 i 1 2 R2
⌘ |fi u |  |f ||u|  |u| + |f |2 .
BR BR 2R2 BR 2 BR

• Similarly, since ⌘ 4  ⌘ 2 , one has

Z Z Z
1
⌘ 2 |Fi↵ @↵ ui |  2
⌘ |ru| + 2
|F |2 .
4

• Again by the same arguments (Schwarz inequality, estimate on |r⌘| and Young
inequality)
Z Z Z Z
↵ i 8 2 4
2 ⌘|Fi u |@↵ ⌘|  |F ||u|  4 |F | + 2 |u|2 .
BR R BR BR R BR

From (4.5) it follows that

Z Z
2 2
⌘ |ru|  ⌘ 2 Aru ru
BR BR
Z Z Z Z
2
= 2 ⌘Aru (u ⌦ r⌘) + f + ⌘ F ru + 2 ⌘F (u ⌦ r⌘)
BR BR BR BR
Z Z
4⇤" 2 2
 ⌘ |ru| + ⌘ 2 |ru|2 (4.6)
R BR 4 BR
✓ ◆Z Z ✓ ◆Z
4⇤ 1 4 2 R2 2 1
+ + + |u| + |f | + +4 |F |2 .
R" 2R2 R2 BR 2 BR BR

35
By choosing " sufficiently small, in such a way that 4⇤"/R = /4, one can absorb line
(4.6), and the thesis follows noticing that
Z Z
2 2
⌘ |ru| |ru|2 .
BR BR/2

Remark 4.4 (Widman’s hole-filling technique). There exists a sharper version of the
Caccioppoli-Leray inequality, let us illustrate it in the simpler case f = 0, F = 0. Indeed,
since
4
|r⌘|  B \B ,
R R R/2
following the proof of Theorem 4.1 one obtains
Z Z
c
|ru(x)|2 dx  2 |u(x) k|2 dx . (4.7)
BR/2 R BR \BR/2
R
Setting k := BR/2 u, the Poincaré inequality in the domain B1 \ B 1/2 and a scaling
argument give Z Z
|ru(x)|2 dx  c |ru(x)|2 dx . (4.8)
BR/2 BR \BR/2
R
Adding to (4.8) the term c BR/2
|ru(x)|2 dx, we get
Z Z
2
(c + 1) |ru(x)| dx  c |ru(x)|2 dx .
BR/2 BR

Setting ✓ := c/(c + 1) < 1, we obtained a decay inequality

Z Z
2
|ru(x)| dx  ✓ |ru(x)|2 dx .
BR/2 BR

k 1
Iterating (4.7) and interpolating (i.e. considering the integer k such that 2 R<r
2 k R), it is not difficult to get
Z ⇣ r ⌘↵ Z
2 ↵
|ru(x)| dx  2 |ru(x)|2 dx 0<rR (4.9)
Br R BR

with (1/2)↵ = ✓, i.e. ↵ = ln2 (1/✓). When n = 2, this implies that u 2 C 0,↵/2 , as we will
see.
The following is another example of “unnatural” inequality, which provides additional
informations on functions that satisfy it.

36
Definition 4.5 (Reverse Hölder’s inequality). Let ↵ 2 (1, 1). A non-negative function
f 2 L↵loc (⌦) satisfies a reverse Hölder’s inequality with exponent ↵ if there exists a constant
c > 0 such that Z ✓Z ◆↵
↵
f c f 8BR (x) b ⌦ .
BR/2 (x) BR (x)

For the sake of completeness, we now recall the Sobolev inequalities. Detailed proofs
will be provided later on: concerning the cases p = n and p > n, we will see them in the
more general context of Morrey’s theory. We will treat the case p < n while dealing with
De Giorgi’s solution of Hilbert’s XIX problem, since slightly more general versions of the
Sobolev inequality are needed there.
Theorem 4.6 (Sobolev inequalities). Let ⌦ be either the whole space Rn or a bounded
regular domain.
• If p < n, denoting with p⇤ := nnpp > p the Sobolev conjugate exponent (characterized
also by p1⇤ = p1 n1 ), we have the continuous immersion
⇤
W 1,p (⌦) ,! Lp (⌦) .

• If p = n, the inclusion of W 1,n (⌦) in the space BM O(⌦) of functions of bounded

mean oscillation provides exponential integrability in bounded subsets of ⌦.3

• If p > n,
W 1,p (⌦) ,! C 0,1 n/p
(⌦) .

Remark 4.7. Combining the Poincaré inequality with the inequality

✓Z ◆1/p⇤ ✓Z ◆1/p ✓Z ◆1/p
p⇤ p p
|v v|  cI |v v| dx + B1 |rv| ,
B1 B1

⇤
coming from the continuity of the embedding W 1,p ,! Lp , we get
✓Z ◆1/p⇤ ✓Z ◆1/p
p⇤ p
|v v| c B1 |rv| ,
B1

for some constant c depending on cI and cP . By a rescaling argument this gives

✓Z ◆1/p⇤ ✓Z ◆1/p
p⇤ p
|u u|  cR |ru| . (4.10)
BR BR

3
The result is basically sharp, as the example of ( ln |x|)↵ 2 W 1,n (B1 ) for n > 1 and ↵ 2 (0, 1 1/n)
shows.

37
If u solves (4.1) with f = F = 0, combining (4.10) with the CL inequality when p⇤ = 2
(that is, p = 2n/(n + 2) < 2), we write
Z !1/2 ✓ Z ◆1/2 ✓Z ◆1/p
2 2 p
cCL R |ru|  |u u|  cR |ru| .
BR/2 BR BR

This way we proved that |ru|p satisfies a reverse Hölder’s inequality with exponent ↵ =
2/p > 1 and C = c/cCL , that is
Z !1/2 ✓Z ◆1/p
2 p
|ru| C |ru| .
BR/2 BR

Remark 4.8 (Embedding for higher order Sobolev spaces). Recall first that higher order
Sobolev spaces W k,p (⌦) are recursively defined (k 1 integer, 1  p  1)
W k,p (⌦) := u 2 W 1,p (⌦) : ru 2 W k 1,p
(⌦; Rn ) .
Together with the Sobolev embedding in Theorem 4.6, with p > n, another way to gain
continuity is using the Sobolev spaces W k,p , with k sufficiently large. In fact, we can
arbitrarily expand the chain
⇤ ⇤ )⇤
W 2,p ,! W 1,p ,! L(p .
Iterating the ⇤ operation k-times we get
1 1 k
= ,
p⇤···⇤ p n
therefore if k > [ np ] (where [·] denotes the integer part) we obtain W k,p ⇢ C 0,↵ with any
positive ↵ with ↵ < 1 n/p + [n/p].

4.1 Nirenberg method

For the moment let us consider a (local) solution u to the Poisson equation
u=f f 2 L2loc (⌦) .
2
Our aim is to prove that u belongs to Hloc (⌦).
When we talk about an a priori estimate, we mean this argument: suppose that we
@u 1
already know that @x i
2 Hloc (⌦), then it is not difficult to check (using the fact that
higher order weak derivative, as well as classical ones, commute) that this function solves
✓ ◆
@u @f
=
@xi @xi

38
in a weak sense. For any ball BR (x0 ) b ⌦, by the Caccioppoli-Leray inequality we get,
Z ✓ ◆ 2 Z 2 Z
@u c @u
r  2 + |f |2 . (4.11)
BR/2 (x0 ) @xi R BR (x0 ) @xi BR (x0 )

We have chosen the Poisson equation because constant coefficients di↵erential operators
commute with convolution, so in this case the a priori regularity assumption can be a
posteriori removed. Indeed, estimate (4.11) applies to u ⇤ ⇢" with f ⇤ ⇢" in place of f ,
since u ⇤ ⇢" satisfies
(u ⇤ ⇢" ) = f ⇤ ⇢" .
2
Passing to the limit as " ! 0 we obtain that u 2 Hloc (⌦) and that the same inequality
1
holds for u, starting from the assumption u 2 Hloc (⌦).
The situation is much more complex when the coefficients A↵ij are not constant and
therefore di↵erentiation provides a worse right hand side in the PDE. Nirenberg’s idea is
to introduce partial discrete derivatives

u(x + hei ) u(x) ⌧h,i u u

h,i u(x) := = (x) .
h h
Remark 4.9. Some basic properties of di↵erentiation are still true and easy to prove:

• (sort of) Leibniz property

h,i (ab) = (⌧h,i a) h,i b +( h,i a)b ;

• integration by parts (relying ultimately on the translation invariance of Lebesgue

measure)
Z Z
'(x) h,i u(x) dx = u(x) h,i '(x) dx 8 ' 2 Cc1 (⌦), |h| < dist(supp ', @⌦) .
⌦ ⌦

In the next lemma we show that membership to W 1,p with p > 1 can be characterized in
terms of uniform Lp bounds on h,i u; notice that one implication was already established
in (1.18).

Lemma 4.10. Consider u 2 Lploc (⌦), with 1 < p  1 and fix i 2 {1, . . . , n}. The partial
@u
derivative @x i
belongs to Lploc (⌦) if and only if
Z
8⌦ b ⌦
0
9 c(⌦ )0
s.t. ( h,i u) '  c(⌦0 )k'kLp0 (⌦0 ) 8 ' 2 Cc1 (⌦0 ) ,
⌦0

with 1/p + 1/p0 = 1.

39
Proof. The first implication has been proved in (1.18), because we know that h,i u is
bounded in Lploc (⌦) when h ! 0, so we can conclude with Hölder’s inequality.
Now fix ⌦0 b ⌦,
Z Z Z
@'
u dx = lim u h,i ' dx = lim ( h,i u) ' dx  c(⌦0 )k'kLp0 (⌦0 ) ;
⌦0 @xi h!0 ⌦0 h!0 ⌦0

0
because of the duality relation between Lp (⌦0 ) and Lp (⌦0 ), the weak derivative @xi u exists
and belongs to Lploc (⌦). ⇤
Let us see how Lemma 4.10 contributes to regularity theory, still in the simplified case
1
of the Poisson equation. Suppose f 2 Hloc (⌦) in the Poisson equation, then translation
invariance and linearity allow us to write

⌧h,i u = ⌧h,i f =) ( h,i u) = h,i f .

Thanks to Lemma 4.10, h,i f is bounded in L2loc (⌦), then by the Caccioppoli-Leray
inequality |r h,i u| is bounded in L2loc (⌦).
As h,i (ru) = r h,i u is bounded in L2loc (⌦; Rn ), thanks to Lemma 4.10 again (applied
componentwise) we get
@
(ru) 2 L2loc (⌦; Rn ) .
@xi
After these preliminaries about Nirenberg’s method, we are now ready to prove the
main result concerning H 2 regularity.

Theorem 4.11. Let ⌦ be an open regular domain in Rn . Consider a function A 2

0,1
(⌦; Rm ⇥n ) such that A(x) := A↵ij (x) satisfies the Legendre-Hadamard condition for
2 2
Cloc
a given constant > 0. Then, for every choice of subsets ⌦0 b ⌦00 b ⌦ there exists a
constant c := c(⌦0 , ⌦00 , A) such that
Z ⇢Z Z
2 2 2
⇥ 2 ⇤
|r u| dx  c |u| dx + |f | + |rF |2 dx
⌦0 ⌦00 ⌦00

1
for all u 2 Hloc (⌦; Rm ) weak solution of the equation

div(Aru) = f div(F )

with data f 2 L2loc (⌦; Rm ) and F 2 Hloc

1
(⌦; Rm⇥n ).

In order to simplify the notation, in the following proof let s denote the unit vector
corresponding to a given fixed direction and consequently ⌧h := ⌧h,s and h := h,s .

40
Remark 4.12. Although the thesis concerns a generic domain ⌦0 b ⌦, it is enough to
prove it for balls inside ⌦. More precisely, if 2R < dist(⌦0 , @⌦), we just need to prove the
inequality
Z ⇢Z Z
2 2 2
⇥ 2 ⇤
|r u| dx  c |u| dx + |f | + |rF |2 dx
BR/2 (x0 ) B2R (x0 ) B2R (x0 )

for any x0 2 ⌦0 . The general result can be easily obtained by a compactness and covering
argument.
Notice also that the statement as given is redundant, since the term div(F ) can always
be absorbed into f . We will see however that the optimal estimate is obtained precisely
doing the opposite, i.e. considering heuristically f as a divergence.
Proof. We assume x0 = 0 and, by the previous remark, F = 0 (possibly changing f ). In
addition, we prove the result under the stronger assumption that the Legendre condition
with constant holds uniformly in ⌦.
First note that the given equation is equivalent, by definition, to the identity
Z Z
Arur' dx = f ' dx (4.12)
⌦ ⌦

for all ' 2 Cc1 (⌦; Rm ). If we apply it to the test function ⌧ h ' with |h| ⌧ 1 and we do
a change of variable, we find
Z Z
⌧h (Aru)r' dx = ⌧h f ' dx . (4.13)
⌦ ⌦

Subtracting (4.12) to equation (4.13) and dividing by h, we get (thanks to the discrete
Leibniz property)
Z Z Z
(⌧h A)r( h u)r' dx = ( h f )' dx ( h A)rur' dx ,
⌦ ⌦ ⌦

which is nothing but the weak form of the equation

div((⌧h A)rv) = f 0 div(G) (4.14)

for v = h u and with data f 0 := h f and G := ( h A)ru.

Now, the basic idea of the proof will be to use the Caccioppoli-Leray inequality. How-
ever, a direct application of the CL inequality would lead to an estimate having the L2
norm of f 0 on the right hand side, and we know from Lemma 4.10 that this norm can be
1
uniformly bounded in h only if f 2 Hloc . Hence, rather than applying CL directly, we will
revisit its proof, trying to get estimates depending only on the L2 norm of f (heuristically,
we see f 0 as a divergence).

41
To this aim, take a cut-o↵ function ⌘ compactly supported in BR , with 0  ⌘  1,
identically equal to 1 on BR/2 and such that |r⌘|  4/R, and insert in (4.14) the test
function := ⌘ 2 h u = ⌘ 2 v with |h| < R/2.
Using Young inequality as in Theorem 4.1 (see (4.6)), we get
Z Z
3 2 2 4⇤"
⌘ |rv|  ⌘ 2 |rv|2
4 BR R BR
✓ ◆Z Z ✓ ◆Z
4⇤ 4 2 2 1
+ + |v| + ⌘ v hf + +4 |G|2 ,
R" R2 BR BR BR

with ⇤ depending only on A. As in the proof of Theorem 4.1, we absorb the term with
k⌘rvk2L2 (BR ) in the left side of the inequality, so that, up to some constant c > 0 depending
on ( , ⇤, R), we get
Z Z Z Z
2 2 2
c ⌘ |rv| dx  |v| dx + |G| dx + ⌘ 2 v h f dx .
2
(4.15)
B2R BR BR

We now study each term of (4.15) separately. Firstly

Z Z
2
|v| dx  |ru|2 dx
BR BR+h

by means of (1.18). The right hand side can in turn be estimated using the classical
Caccioppoli-Leray inequality for u between the balls B3R/2 and B2R : it gives an upper
bound of the desired form.
R
Concerning the term ⌘ 2 v h f dx, by means of discrete integration by parts and Young
inequality, we can write
Z Z Z
2 2 2 1
⌘ v h f dx  "˜ | h (⌘ v)| dx + |f |2 dx . (4.16)
BR BR "˜ BR

The first term in the right hand side of (4.16) can be estimated with (since |r⌘|2  64/R2 )
Z Z Z
2 2 4 2 128
|r(⌘ v)| dx  2 ⌘ |rv| dx + 2 |v|2 dx ,
BR+h BR+h R BR+h
4 2
R ⌘  ⌘2 we can absorb the
so that choosing " sufficiently small and using the inequality
first term and use once more the CL inequality to estimate BR+h |v| dx.
The term involving the integral kGk2L2 (BR ) can be estimated in the very same way, using
this time also the local Lipschitz assumption on A to bound h A, so that finally we put
together all the corresponding estimates to obtain the thesis (the conclusion comes from
Lemma 4.10 and then letting h ! 0 in the estimate involving v = h u). ⇤

42
Remark 4.13. It should be clear from the proof that the previous result only concerns
interior regularity and cannot be used in order to get information about the behaviour
of the function u near the boundary @⌦. In other terms, we can not guarantee that the
constant c remains bounded as ⌦0 invades ⌦ (so that R ! 0), even if global regularity
assumptions on A, u, f and F are made. The issue of boundary regularity requires
di↵erent techniques that will be described later on.

5 Decay estimates for systems with constant coeffi-

cients
Our next target towards the development of a regularity theory is now to derive some
decay estimates for constant coefficients di↵erential operators. Let A = A↵ij be a matrix
satisfying the Legendre-Hadamard condition for some > 0, let ⇤ = |A| and consider the
problem 8
< div(Aru) = 0
: 1
u 2 Hloc (⌦; Rm ) .
Then, these two inequality hold for any Br (x0 ) ⇢ BR (x0 ) b ⌦ :
Z ⇣ r ⌘n Z
2
|u| dx  c(n, , ⇤) |u|2 dx (5.1)
Br (x0 ) R BR (x0 )
Z ⇣ r ⌘n+2 Z
2
|u ur,x0 | dx  c(n, , ⇤) |u uR,x0 |2 dx (5.2)
Br (x0 ) R BR (x0 )

with c(n, , ⇤) depending only on n, and ⇤.

Here us,x0 denotes as usual the mean value of u on Bs (x0 ).
Proof of (5.1). By a standard rescaling argument, it is enough ⇥ ⇤ to study the case
R = 1. For the sequel, let ⇥k ⇤be the smallest integer such that k > n2 (and consequently
n
H k ,! C 0,↵ with ↵ = k 2
). First of all, by the Caccioppoli-Leray inequality, we have
that Z Z
2
|ru| dx  c1 |u|2 dx .
B1/2 (x0 ) B1 (x0 )
1,2
Now, for any ↵ 2 {1, 2, . . . , n}, we know that @↵ u 2 Hloc (⌦) by Theorem 4.11, and since
the matrix A has constant coefficients it will solve the same equation. Hence, we can
iterate the argument in order to get an estimate having the form
Z X Z
2
|r u|  ck |u|2 dx
B2 k (x0 ) | |k B1 (x0 )

43
for some constant ck > 0. Consequently, thanks to our choice of the integer k, we can
find a constant  such that
Z
sup |u|2 dx   |u|2 dx .
B2 k (x0 ) B1 (x0 )

In order to conclude the proof, it is better to consider two cases. If r  2 k , then

Z Z
2 n 2 n
|u| dx  !n r sup |u|  !n r |u|2 dx ,
Br (x0 ) B2 k (x0 ) B1 (x0 )

where !n denotes the Lebesgue measure of the unit ball in Rn . Hence, for this
R case we have
the thesis, provided c( , ⇤) !n . If r 2 (2 k , 1), then it is clear that Br (x0 ) |u|2 dx 
R
B1 (x0 )
|u|2 dx and so, since we have a lower bound for r, we just need to choose c( , ⇤)
such that c( , ⇤) 2kn .
We can now prove the second inequality, that concerns the notion of variance of the
function u on a ball.
Proof of (5.2). Again, it is useful to study two cases separately. If r  R/2, then by
the Poincaré inequality there exists a constant c(n) such that
Z Z
2
|u ux0 ,r | dx  c(n)r 2
|ru|2 dx
Br (x0 ) Br (x0 )

and so
Z ◆n Z ✓
2 r 2
|u ux0 ,r | dx  c(n)r |ru|2 dx
Br (x0 ) R/2 BR/2 (x0 )
✓ ◆n+2 Z
r
 c(n, , ⇤) |u uR,x0 |2 dx
R/2 BR (x0 )

respectively by the previous result applied to the gradient ru and finally by the Cacciop-
poli-Leray inequality. For the case R/2 < r  R we need to use the following fact, that
will be discussed below: the mean value ux0 ,r is a minimizer for the function
Z
m7 ! |u m|2 dx . (5.3)
Br (x0 )

If we give this for granted, the conclusion is easy because

Z Z ⇣ r ⌘n+2 Z
|u ur,x0 |2 dx  |u uR,x0 |2 dx  2n+2 |u uR,x0 |2 dx.
Br (x0 ) Br (x0 ) R BR (x0 )

44
 Let us go back to the study of
Z
inf |u m|p dx
m2R ⌦

for 1  p < 1 and u 2 Lp (⌦) where ⌦ is any open, bounded domain in Rn . As we

pointed out above, this problem is easily solved, when p = 2, by the mean value u⌦ : it
suffices to notice that
Z Z Z
2
|u m| dx = |u| dx 2m u dx + m2 L n (⌦) .
2
⌦ ⌦ ⌦

Nevertheless, this is not true in general, for p 6= 2. Of course

Z Z
p
inf |u m| dx  |u u⌦ |p dx
m ⌦ ⌦

but we also claim that, for any m 2 R, we have

Z Z
p
|u u⌦ | dx  2 p
|u m|p dx . (5.4)
⌦ ⌦

Since the problem is clearly translation invariant, it is sufficient to prove inequality (5.4)
for m = 0. But in this case
Z Z Z Z
p p 1 p p 1 p p
|u u⌦ | dx  2 |u| dx + 2 |u⌦ | dx  2 |u|p dx ,
⌦ ⌦ ⌦ ⌦

thanks to the elementary inequality |a + b|p  2p 1 |a|p + |b|p and to the fact that
Z Z
p
|u⌦ | dx  |u|p dx
⌦ ⌦

which is a standard consequence of the Hölder (or Jensen) inequality.

6 Regularity up to the boundary

Let us first consider a simple special case. Suppose we have to deal with the problem
8
< u=f
(6.1)
:
u 2 H01 (R) ,

where R := ( a, a)n 1 ⇥ (0, a) is a rectangle in Rn with sides parallel to the coordinate

axes. Let us use coordinates x = (x0 , xn ) with x0 2 Rn 1 and assume f 2 L2 (R). The

45
rectangle R0 = ( a/2, a/2)n 1 ⇥ (0, a/2) is not relatively compact in R, nevertheless via
Nirenberg’s method we may find estimates of the form
Z Z Z
2 c 2
|@xs ru| dx  2 |ru| dx + c |f |2
R 0 a R R

for s = 1, 2, . . . , n 1, provided u = 0 on R \ {xn = 0}. Indeed, we are allowed to use

test functions ' = ⌘ 2 h,s u, where the support of ⌘ can touch the hyperplane {xn = 0}
(because of the homogeneous Dirichlet boundary condition on u). The equation (6.1) may
be rewritten as
@ 2u
= x0 u + f
@x2n
and here the right hand side x0 u + f is in L2 (R0 ). We conclude that also the missing
second derivative in the xn direction is in L2 , hence u 2 H 2 (R0 ). Notice that this argument
requires only the validity of the homogeneous Dirichlet condition on the portion {xn = 0}
of the boundary of R. In addition, this homogeneous Dirichlet condition also ensures that
all functions
@u
i = 1, . . . , n 1
@xi
have 0 trace on {xn = 0}, and this is crucial for the iteration of this argument with higher
order derivatives (see also Theorem 6.2 below).
Now we want to use this idea in order to study the regularity up to the boundary for
problems like 8
< div(Aru) = f + divF
:
u 2 H01 (⌦; Rm )
under the following hypotheses:

• f 2 L2 (⌦; Rm );

• F 2 H 1 (⌦; Rm⇥n );
2 ⇥n2
• A 2 C 0,1 (⌦; Rm );

• A(x) satisfies the Legendre-Hadamard condition uniformly in ⌦;

• ⌦ has a C 2 boundary, in the sense that it is, up to a rigid motion, locally the
subgraph of a C 2 function.

Theorem 6.1. Under the previous assumptions, the function u belongs to H 2 (⌦; Rm ) and
⇥ ⇤
kukH 2  c(⌦, A, n) kf k2 + kF kH 1

46
Proof. Since we already have the interior regularity result at our disposal, it suffices to
show that for any x0 2 @⌦ there exists a neighbourhood U of x0 in ⌦ such that u 2 H 2 (U ).
Without loss of generality we assume x0 = 0. We consider first the case of a flat boundary.
Step 1. (Flat boundary) By applying Nirenberg’s method as described above for the
case of the constant coefficient operator we get @x↵ ui 2 H 1 (R0 ) for ↵ = 1, 2, . . . , n 1
and i = 1, 2, . . . , m, and
✓ ✓ ◆◆ ✓ ◆ ✓ ◆
@u @f @F @A
div Ar = + div + div ru . (6.2)
@x↵ @x↵ @x↵ @x↵
Anyway, we cannot include in the previous conclusion the second derivatives @x2n xn ui and
here we really need to refine a bit the strategy seen above for the Poisson equation. Ac-
tually, this is not complicated because the equation readily implies that @xn (Ann j
ij @xn u ) 2
L2 (R0 ) for any i 2 {1, 2, . . . , m}. Formally this implies, by the Leibniz rule, that Ann 2
ij @xn xn u
j
2 0
belong to L (R ); this is formal because one of the factors is only a distribution (not yet
a function). To make this rigorous, we use the di↵erence quotients in the xn direction
and the discrete Lebniz rule: since by Lemma 4.10 the di↵erence quotients h (Ann j
ij @xn u )
have uniformly bounded L2 norm in Rh0 = {x 2 R : dist(x, @R0 ) > h}, we obtain that the
same is true for Ann ij
j nn
h @xn u . Since the matrix Aij is invertible with detAij
nn m
(as a
consequence of the Legendre-Hadamard condition) we get
Z
lim sup | h @xn uj |2 dx < 1
h!0+ 0
Rh

which gives @x2n xn uj 2 L2 (R0 ).

Step 2. (Straightnening of the boundary) There exist h 2 C 2 (Rn 1 ) and V = ( b, b)n
such that (up to a rigid motion, choosing the hyperplane {xn = 0} as the tangent one to
@⌦ at 0)
⌦ \ V = {x 2 V : xn > h(x0 )} .
Consequently, we can define the change of variables x0n = xn h(x0 ) and the function
H(x0 , xn ) = (x0 , xn h(x0 ))) that maps ⌦ \ V onto H(⌦ \ V ), which contains a rectangle
R = ( a, a)n 1 ⇥ (0, a). We set ⌦0 := H 1 (R) ⇢ V \ ⌦ and U := H 1 (R0 ), with
R0 = ( a/2, a/2)n 1 ⇥ (0, a/2).
It is clear that H is invertible and, called G its inverse, both H and G are C 2 functions.
Moreover rH is a triangular matrix with det(rH) = 1. Besides, the maps G and H induce
isomorphisms between both H 1 and H 2 spaces (via change of variables in the definition
of weak derivative, as we will see in a moment). To conclude, it suffices to show that
v = u G belongs to H 2 (R0 ; Rm ). To this aim, we check that v solves in R the PDE
8
e
< div(Arv) = fe + divFe
(6.3)
: 0
v = 0 on {xn = 0} \ R

47
where of course the boundary condition has to be interpreted in the weak sense and
⇥ ⇤
fe = f G, Fe = (F · DH) G, A e = DH · A · (DH)t G

(here contractions are understood with respect to the greek indices, the only ones involved
in the change of variables, see (6.4) below). These formulas can be easily derived by an
elementary computation, starting from the weak formulation of the problem and apply-
ing a change of variables in order to express the di↵erent integrals in terms of the new
coordinates. For instance
Z Z
i
fi (x)' (x) dx = fi G(y)'i G(y) det(rG(y)) dy
⌦0 R

just letting x = G(y), but then det(rG) = 1 and we can set ' = H so that equivalently
= ' G and Z Z
i
fi (x)' (x) dx = fei (y) i (y) dy .
⌦0 R

The computation for Fe or A e is less trivial, but there is no conceptual difficulty. We just
see the first one:
Z Z
↵ @'i @'i
Fi (x) (x) dx = Fi↵ (G(y)) (G(y)) det(rG(y)) dy
⌦0 @x↵ R @x↵
Z
@ i @H
= Fi↵ (G(y)) (y) (G(y)) dy
R @y @x↵

which leads to the conclusion. Note that here and above the arbitrary test function '
has been replaced by the arbitrary test function . However, we should ask whether the
conditions on A (for instance, the Legendre-Hadamard condition) still hold true for A.e
e above. In
This is the case and we can verify it directly by means of the expression of A
fact, ✓ 0 0◆

e↵0 0 @H ↵ ↵ @H
Aij = A G (6.4)
@x↵ ij @x
a 2 Rn and b 2 Rm
and so, for any e
✓ 0 ◆✓ 0 ◆
e↵0 0 (y)e @H ↵ @H
Aij a 0 bi bj = A↵ij (G(y))
a↵ 0 e (G(y))e
a↵ 0 (G(y))e
a 0 bi bj
@x↵ @x
2
a|2 |b|2
|rH(G(y))e (rH(G(y))) 1
a|2 |b|2
|e

since clearly
1 2
a|2  (rH(G(y)))
|e a| 2 .
|rH(G(y))e

48
Hence, A e satisfies the Legendre-Hadamard condition for an appropriate constant 0 > 0
depending on and H, and of course A e 2 C 0,1 (R).
Through this transformation of the domain, we can finally apply Step 1 and find that
v 2 H 2 (R0 ). Coming back to the original variables we obtain the H 2 regularity of u.
⇤
If both the boundary and the data are sufficiently regular, this method can be iterated
to get the following theorem.
Theorem 6.2. Assume, in addition to the hypotheses above, that f 2 H k (⌦; Rm ) and
2 2
also F 2 H k+1 (⌦; Rm⇥n ), A 2 C k,1 (⌦, Rm ⇥n ) with ⌦ such that @⌦ 2 C k+2 . Then u 2
H k+2 (⌦; Rm ).
We are not going to present the detailed proof of the previous result, but the basic
idea consists in di↵erentiating the starting equation with respect to each fixed direction
to get an equation having the form of (6.3), as in (6.2), provided we set Fe = @x
@F
↵
@A
+ @x ↵
ru.

7 Interior regularity for nonlinear problems

So far, we have just dealt with linear problems and the richness of di↵erent situations
was only based on the possibility of varying the elliptic operator, the boundary conditions
and the number of dimensions involved in the equations. We see now that Nirenberg’s
technique is particularly appropriate to deal also with nonlinear PDE’s, as those arising
from Euler-Lagrange equations of non-quadratic functionals.
Consider a function F 2 C 2 (Rm⇥n ) and assume the following:
(i) there exists a constant C > 0 such that |D2 F (⇠)|  C for any ⇠ 2 Rm⇥n ;
(ii) F satisfies a uniform Legendre condition, i.e. @p↵i @p F (p)⇠i↵ ⇠j |⇠|2 for all ⇠ 2
j
Rm⇥n , for some > 0 independent of p 2 Rm⇥n .
@2F
Let Bi↵ := @F
@p↵
and A↵ij := @p↵
and notice that A↵ij is symmetric with respect to the
i i @pj
transformation (↵, i) ! ( , j).
Let ⌦ ⇢ Rn be an open domain and let u 2 Hloc 1
(⌦; Rm ) be a local minimizer of the
functional Z
w 7 ! I(w) := F (rw) dx .
⌦
The implication
F 2 C1 ) u 2 C1
is strongly related to Hilbert’s XIX problem (initially posed in 2-dimensions space and in
the category of analytic functions). In the sequel we will first treat the case n = 2 and
much later the case n 3, which is significantly harder.

49
 Recall that u is a local minimizer for I if, for any
Z Z
u 2 Hloc (⌦; R ), spt(u u ) ⇢ ⌦ b ⌦ =)
1 m 0 0
F (ru0 ) dx F (ru) dx .
⌦0 ⌦0

If this is the case, we have already seen how the Euler-Lagrange equation can be ob-
tained: considering perturbations of the form u0 = u + tr' with ' 2 Cc1 (⌦, Rm ) one can
prove (using the fact that the regularity assumptions on F allow di↵erentiation under the
integral sign) that
Z Z
d @'i
0= F (ru + tr') dx = Bi↵ (ru) dx .
dt ⌦ t=0 ⌦ @x↵

Now, suppose s is a fixed coordinate direction (and let es be the corresponding unit vector)
and h > 0 a small positive increment: if we apply the previous argument to a test function
having the form ⌧ h ', we get
Z
@'i
⌧h (Bi↵ (ru)) dx = 0
⌦ @x↵
and consequently, subtracting this to the previous one
Z
↵ @'i
h,s (B i (ru)) dx = 0 .
⌦ @x↵
However, as a consequence of the regularity of F , we can write
Z 1
↵ ↵ d ↵
Bi (ru(x + hes )) Bi (ru(x)) = Bi (tru(x + hes ) + (1 t)ru(x)) dt
0 dt
Z 1  j
↵ @u @uj
= Aij (tru(x + hes ) + (1 t)ru(x)) dt (x + hes ) (x)
0 @x @x

and setting Z 1
e↵ (x)
A := A↵ij (tru(x + hes ) + (1 t)ru(x)) dt
ij,h
0
we rewrite the previous condition as
Z j
e↵ (x) @ h,s u @'i
A ij,h (x) (x) dx = 0 .
⌦ @x @x↵

Hence, w = h,s u solves the equation

eh rw) = 0 .
div(A (7.1)

50
 It is obvious by the definition that A e↵ (x) satisfies both the Legendre condition for
ij,h
the given constant > 0 and a uniform upper bound on the L1 norm. Therefore we can
apply the Caccioppoli-Leray inequality to the problem (7.1) to obtain constants C1 and
C2 , not depending on h, such that
Z Z
2 C1
|r( h,s u)| dx  2 | h,s u|2 dx  C2
BR (x0 ) R B2R (x0 )

for any BR (x0 ) ⇢ B2R (x0 ) b ⌦. Consequently, by Lemma 4.10 we deduce that
2
u 2 Hloc (⌦; Rm ). (7.2)

Moreover, we have that

• ! @u/@xs in L2loc (this is clearly true if u is regular and then we can exploit
h,s u
the fact that the operators h,s are equibounded, still by Lemma 4.10);

• @u/@xs satisfies, in a weak sense, the equation

✓ ◆
@u
div A(ru)r =0. (7.3)
@xs
In fact
h!0
A↵ij (tru(x + hes ) + (1 t)ru(x)) ! A↵ij (ru(x))
in Lp for any 1  p < 1, as an easy consequence of the continuity of translations
in Lp and the continuity of A.
In order to solve Hilbert’s XIX problem, we would like to apply a classical result
by Schauder saying that if w is a weak solution of the problem div(Brw) = 0, then
B 2 C 0,↵ ) w 2 C 1,↵ , and so u 2 C 2,↵ . But we first need to improve the regularity of
1
B(x) = A(ru(x)). As a matter of fact, at this point we just know that A(ru) 2 Hloc ,
0,↵
while we need A(ru) 2 C . When n = 2 we can apply Widman’s technique (see (4.9))
to the PDE (7.3) to obtain Hölder regularity of ru, both in the scalar and in the vectorial
case. The situation is much harder in the case n > 2, since this requires deep new ideas:
the celebrated theory by De Giorgi-Nash-Moser which solves the problem in the scalar
case. We will see that in the vectorial case new difficulties arise.

8 Hölder, Morrey and Campanato spaces

In this section we introduce the Hölder spaces C 0,↵ , the Morrey spaces Lp, and the
Campanato spaces Lp, . All these spaces are relevant, besides the standard Lebesgue
spaces, in the regularity theory, as we will see.

51
Definition 8.1 (Hölder spaces). Given A ⇢ Rn , u : A ! Rm and ↵ 2 (0, 1] we define
the ↵-Hölder semi-norm on A as
|u(x) u(y)|
kuk↵,A := sup .
x6=y2A |x y|↵

We say that u is ↵-Hölder in A, and write u 2 C 0,↵ (A; Rm ), if kuk↵,A < 1.

If ⌦ ⇢ Rn is open, we say that u : ⌦ ! Rm is locally ↵-Hölder if for any x 2 ⌦ there
exists a neighbourhood Ux b ⌦ such that kuk↵,Ux < 1. The corresponding vector space is
0,↵
denoted by Cloc (⌦; Rm ).
If k 2 N, the space of functions of class C k (⌦; Rm ) with all i th derivatives ri u with
|i|  k in C 0,↵ (⌦; Rm ) will be denoted by C k,↵ (⌦; Rm ).

Remark 8.2. The spaces C k,↵ (⌦; Rm ) are Banach when endowed with the norm
X
kukC k,↵ = ri u C 0,↵ .
|i|k

Definition 8.3 (Morrey spaces). Assume ⌦ ⇢ Rn open, 0 and 1  p < 1. We say

that f 2 Lp (⌦) belongs to Lp, (⌦) if
Z
sup r |f |p dx < +1
0<r<d⌦ , x0 2⌦ ⌦(x0 ,r)

where ⌦(x0 , r) := ⌦ \ Br (x0 ) and d⌦ is the diameter of ⌦. It is easy to verify that

✓ Z ◆1/p
p
kf kLp, := sup r |f | dx
0<r<d⌦ , x0 2⌦ ⌦(x0 ,r)

is a norm on Lp, (⌦).

Remark 8.4. We mention here some of the basic properties of the Morrey spaces Lp, :

(i) Lp, (⌦; R) are Banach spaces, for any 1  p < 1 and 0;

(ii) Lp,0 (⌦; R) = Lp (⌦; R);

(iii) Lp, (⌦; R) = {0} if > n;

(iv) Lp,n (⌦; R) ⇠ L1 (⌦; R);

(v) Lq,µ (⌦; R) ⇢ Lp, (⌦; R) if ⌦ is bounded, q p and (n )/p (n µ)/q.

52
Note that the condition (n )/p (n µ)/q can also be expressed by asking  c
with the critical value c defined by the equation (n c )/p = (n µ)/q. The proof of the
first result is standard, the second statement is trivial, while the third and fourth ones are
immediate applications of Lebesgue Di↵erentiation Theorem. Finally the last one relies
on Hölder inequality:
✓Z ◆ ✓Z ◆p/q
p q
|f | dx  |f | dx (!n rn )(1 p/q)
⌦(x,r) ⌦(x,r)

= C(n, p, q) kf kpLq,µ rµp/q+n(1 p/q)

= C(n, p, q) kf kpLq,µ r c .

Definition 8.5 (Campanato spaces). Assume ⌦ ⇢ Rn open, > 0, 1  p < 1. A

function f 2 Lp (⌦) belongs to the Campanato space Lp, if
Z
p
kf kLp, := sup r |f (x) fx0 ,r |p dx < 1 , (8.1)
x0 2⌦, 0<r<d⌦ ⌦(x0 ,r)

where, as before, d⌦ is the diameter of ⌦ and

Z
fx0 ,r := f (x) dx . (8.2)
⌦(x0 ,r)

The mean fx0 ,r defined in (8.2) might not be optimal in the calculation of the sort of
p-variance in (8.1), anyway it gives equivalent results, thanks to (5.4).

Remark 8.6. As in Remark 8.4, we briefly highlight the main properties of Campanato
spaces.

(i) As defined in (8.1), k · kLp, is merely a seminorm because constants have null Lp,
norm. If ⌦ is connected, then Lp, modulo constants is a Banach space.

(ii) Lq,µ ⇢ Lp, when ⌦ is bounded, p  q and (n )/p (n µ)/q.

(iii) C 0,↵ ⇢ Lp,n+↵p , because

Z
|f (x) fx0 ,r |p dx  kf kpC 0,↵ r↵p L n (B(x0 , r)) = kf kpC 0,↵ !n rn+↵p .
⌦(x0 ,r)

We will see that a converse statement holds (namely functions in these Campanato
spaces have Hölder continuous representatives in their Lebesgue equivalence class),
and this is very useful: we can replace the pointwise definition of Hölder spaces with
an integral one.

53
 Actually, Campanato spaces are interesting only when n, exactly because of
their relationship with Hölder spaces. On the contrary, if < n, Morrey spaces and
Campanato spaces are basically equivalent. In the proof of this and other results we need
a mild regularity assumption on ⌦, namely the existence of c⇤ > 0 satisfying

L n (⌦ \ Br (x0 )) c⇤ r n 8x0 2 ⌦, 8r 2 (0, d⌦ ) . (8.3)

For instance this assumption includes domains which are locally subgraphs of Lipschitz
functions, while it rules out domains with outer cusps.
Theorem 8.7. Let ⌦ ⇢ Rn be an open bounded set satisfying (8.3) and let 0  < n.
Then the spaces Lp, and Lp, are equivalent, i.e.

k · kLp, ' k · kLp, + k · kLp .

Proof. All through the proof we denote by c a generic constant depending from the
constant c⇤ in (8.3) and from n, p, . We allow it to vary, even within the same line.
Without using the hypothesis on , we easily prove that Lp, ⇢ Lp, : trivially Jensen’s
inequality ensures Z Z
p
|fx0 ,r | dx  |f (x)|p dx ,
⌦(x0 ,r) ⌦(x0 ,r)

thus we can estimate

Z ✓Z Z ◆
|f (x) fx0 ,r |p dx  2p 1 p
|f (x)| dx + p
|fx0 ,r | dx
⌦(x0 ,r) ⌦(x0 ,r) ⌦(x0 ,r)
Z
p
 2 |f (x)|p dx .
⌦(x0 ,r)
R
Conversely, we would like to estimate r ⌦(x0 ,r)
|f (x)|p dx with kf kLp, + kf kp for
every 0 < r < d⌦ and every x0 2 ⌦. As a first step, by triangular inequality we separate
Z Z
p
|f (x)| dx  2 p 1
|f (x) fx0 ,r |p dx+crn |fx0 ,r |p  c r kf kpLp, + rn |fx0 ,r |p ,
⌦(x0 ,r) ⌦(x0 ,r)

so we took out the problematic summand |fx0 ,r |p .

In order to estimate |fx0 ,r |p , let us bring in an inequality involving means on concentric
balls: when x0 2 ⌦ is fixed and 0 < r < ⇢ < d⌦ , it holds
Z
n p
c⇤ !n r |fx0 ,r fx0 ,⇢ |  |fx0 ,r fx0 ,⇢ |p dx
⌦(x0 ,r)
✓Z Z ◆
p 1 p p
 2 |fx0 ,r f (x)| dx + |f (x) fx0 ,⇢ | dx
⌦(x0 ,r) ⌦(x0 ,r)

 2 p 1
kf kpLp, r +⇢ 2 p
kf kpLp, ⇢ ,

54
thus we obtained that
n
⇣ ⇢ ⌘ np n
|fx0 ,r fx0 ,⇢ |  ckf kLp, r p ⇢ p = ckf kLp, ⇢ p . (8.4)
r
(k+1)
Now fix a radius R > 0: if r = 2 R and ⇢ = 2 k R, inequality (8.4) means that
✓ ◆ n
R p
|fx0 ,R/2k+1 fx0 ,R/2k |  ckf kLp, , (8.5)
2k
and, adding up when k = 0, . . . , N 1, it means that
n ✓ ◆ n
n 2N p 1 R p
|fx0 ,R/2N fx0 ,R |  ckf kLp, R p
n  ckf kLp, . (8.6)
2 p 1 2N
Let us go back to our purpose of estimating |fx0 ,r |p : we choose R 2 (d⌦ /2, d⌦ ) and
N 2 N such that r = R/2N . By triangular inequality
|fx0 ,r |p  2p 1
(|fx0 ,r fx0 ,R |p + |fx0 ,R |p ) ;
since
|fx0 ,R |  c(d⌦ )kf kLp ,
the only thing left to conclude is to apply inequality (8.6) in this case:
|fx0 ,r fx0 ,R |p  ckf kpLp, r n
,
that is all we needed to conclude that
Z
r |f |p  c kf kpLp, + dn⌦ kf kpLp .
⌦(x0 ,r)

Remark 8.8. When the dimension of the domain space is n, the Campanato space L1,n is
very important in harmonic analysis and elliptic regularity theory: after John-Nirenberg
seminal paper, this space is called BM O (bounded mean oscillation). It consists of the
space of all functions f : ⌦ ! R such that there exists a constant C satisfying the
inequality Z
|f (x) fx0 ,r | dx  Crn 8 r 2 (0, d⌦ ), 8 x0 2 ⌦ .
⌦(x0 ,r)

Notice that L (⌦) ( BM O(⌦): for example, consider ⌦ = (0, 1) and f (x) = ln x. For
1

any a, r > 0 it is easy to check that

Z a+r Z a+r ✓ ◆
a
| ln t ln(a + r)| dt = (ln(a + r) ln t) dt = r + a ln r,
a a a+r

55
 R a+r
hence ln x 2 BM O(⌦). For simplicity, we replaced the mean a ln s ds with ln(a + r),
but, up to a multiplicative factor 2, this does not make a di↵erence. On the contrary
/ L1 (⌦).
ln x 2
Theorem 8.9 (Campanato). With the previous notation, when n <  n+p Campanato
spaces Lp, are equivalent to Hölder spaces C 0,↵ with ↵ = ( n)/p. Moreover, if ⌦ is
connected and > n + p, then Lp, is equivalent to the set of constants.
Proof. As in the proof of Theorem 8.7, the letter c denotes a generic constant depending
on the exponents, the space dimension n and the constant c⇤ in (8.3).
Let = n + ↵p. We already observed in Remark 8.6 that C 0,↵ ⇢ Lp, , so we need to
prove the converse inclusion: given a function f 2 Lp, , we are looking for a representative
in the Lebesgue equivalence class of f which belongs to C 0,↵ .
Recalling inequality (8.5) with fixed radius R > 0 and x 2 ⌦, we obtain that the
sequence (fx,R/2k ) has the Cauchy property. Hence we define
Z
˜
f (x) := lim f (y) dy .
k!1 ⌦(x,R/2k )

Clearly
Z Z
|f (y) p
fx,R/2k | dy ! 0 =) |f (y) f˜(x)|p dy ! 0 , (8.7)
⌦(x,R/2k ) ⌦(x,R/2k )

but since c⇤ rn  L n (⌦(x, r))  !n rn , for r 2 R/2k+1 , R/2k we have

Z n Z
2 ! n
|f (y) f˜(x)| dy 
p
|f (y) f˜(x)|p dy ,
⌦(x,r) c ⇤ k
⌦(x,R/2 )

so that (8.7) implies that

Z
|f (y) f˜(x)|p dy ! 0 as r # 0 .
⌦(x,r)

In particular, f˜ does not depend on the chosen initial radius R. Let us prove that

f˜ 2 C 0,↵ (⌦) .

We employ again an inequality from the proof of Theorem 8.7: letting N ! 1 in (8.6),
we get that
|f˜(x) fx,R |  ckf kLp, R↵
with ↵ = ( n)/p; consequently, given x, y 2 ⌦ and choosing R = 2|x y|,

|f˜(x) f˜(y)|  |f˜(x) fx,R | + |fx,R fy,R | + |fy,R f˜(y)|  c|x y|↵ + |fx,R fy,R | .

56
The theorem will be proved if we can estimate |fx,R fy,R |. To this aim, we use the
inclusion ⌦(y, R/2) ⇢ ⌦(x, R) to get
Z
n n p
c⇤ 2 R |fx,R fy,R |  |fx,R fy,R |p ds
⌦(y,R/2)
✓Z Z ◆
p 1 p p
 2 |f (s) fx,R | ds + |f (s) fy,R |
⌦(x,R) ⌦(y,R)
p p
 2 kf kLp, R ,

and finally
n
|fx,R fy,R |  ckf kLp, R p  c|x y|↵ .
⇤
The following inclusions follow by the Hölder and the Poincaré inequalities, respec-
tively.
Proposition 8.10 (Inclusions between Lebesgue and Morrey spaces, Morrey and Cam-
1,n/p0
panato spaces). For all p 2 (1, 1), Lploc (⌦) ⇢ Lloc (⌦). In addition

|ru| 2 Lp,
loc (⌦) =) u 2 Lp,
loc
+p
(⌦) . (8.8)
0,↵
Corollary 8.11 (Sobolev embedding for p > n). If p > n, then W 1,p (⌦) ⇢ Cloc (⌦), with
1,p 0,↵
↵ = 1 n/p. If ⌦ is bounded and regular, then W (⌦) ⇢ C (⌦).
Proof. By the previous proposition we get
1,p 1,n/p0
u 2 Wloc =) |ru| 2 Lloc (⌦) = L1,n n/p
(⌦) = L1,n 1+↵
(⌦) . (8.9)

Applying (8.8) and (8.9), we get u 2 L1,n+↵

loc
0,↵
(⌦), so that u 2 Cloc (⌦). If ⌦ is bounded
and regular we apply this inclusion to a W 1,p extension of u to obtain the global C 0,↵
regularity. ⇤

9 XIX Hilbert problem and its solution in the two-

dimensional case
Let ⌦ ⇢ Rn open, let F 2 C 3 (Rm⇥n ) and let us consider a local minimizer u of the
functional Z
v 7! F (rv) dx (9.1)
⌦
2
as in Section 2.4. We assume that r F (p) satisfies the Legendre condition (2.16) with
> 0 independent of p and is uniformly bounded.

57
 We have seen that u satisfies the Euler-Lagrange equations, for (9.1) they are
@
Fp↵i (ru) = 0 i = 1, . . . , m . (9.2)
@x↵
We have also seen in Section 7 how, di↵erentiating (9.2) along the direction xs , one can
obtain ✓ ◆
@ @ 2 uj
Fp↵ p (ru) =0 i = 1, . . . , m . (9.3)
@x↵ i j @x @xs
In the spirit of Hilbert’s XIX problem, we are interested in the regularity properties
of u. Fix s 2 {1, . . . , n}, let us call
@u
w(x) := (x) 2 L2 (⌦, Rm ) ,
@xs
A(x) := r2 F (ru(x)) ,
thus (9.3) can be written as
✓ ◆
@ @ 2 uj
div (Arw) = Fp↵ p (ru) =0. (9.4)
@x↵ i j @x @xs
1
Since w 2 Hloc (⌦; Rm ) by (7.2), we can use the Caccioppoli-Leray inequality for w, in
the sharp version of Remark 4.4. Combining it with the Poincaré inequality (choosing k
equal to the mean value of w on the ball BR (x0 ) \ BR/2 (x0 )), we obtain
Z Z Z
2 2 2
|rw| dx  cR |w k| dx  c |rw|2 dx ,
BR/2 (x0 ) BR (x0 )\BR/2 (x0 ) BR (x0 )\BR/2 (x0 )
R
thus, adding c |rw|2 dx to both sides, we get
BR/2 (x0 )
Z Z
2 c
|rw| dx  |rw|2 dx .
BR/2 (x0 ) c + 1 BR (x0 )
Now, if ✓ := c/c + 1 < 1 and ↵ = log2 ✓, we can write the previous inequality as
Z ✓ ◆↵ Z
2 1
|rw| dx  |rw|2 dx . (9.5)
BR/2 (x0 ) 2 BR (x0 )

In order to get a power decay inequality from (9.5), we state this basic iteration lemma.
Lemma 9.1. Consider a non-decreasing function f : (0, R0 ] ! R satisfying
⇣ ⇢ ⌘ ✓ 1 ◆↵
f  f (⇢) 8 ⇢  R0 .
2 2
Then ⇣ r ⌘↵
↵
f (r)  2 f (R) 8 0 < r  R  R0 .
R
58
Proof. Fix r < R  R0 and choose a number N 2 N such that
R R
<r .
2N +1 2N
It is clear from the iteration of the hypothesis that
✓ ◆ ✓ ◆↵N
R 1
f  f (R) ,
2N 2

thus, by monotonicity,
N ↵N
f (r)  f 2 R 2 f (R) = 2↵ 2 ↵(N +1)
f (R) < 2↵ (r/R)↵ f (R) .

⇤
Thanks to Lemma 9.1, we are ready to transform (9.5) in
Z ⇣ ⇢ ⌘↵ Z
2
|rw| dx  c |rw|2 dx 80 < ⇢  R ,
B⇢ (x0 ) R BR (x0 )

therefore |rw| 2 L2,↵

loc (⌦). So, as we remarked in the proof of Corollary 8.11, this gives
w 2 L2,↵+2
loc (⌦). All these facts are true in any number n of space dimensions, but when
n = 2 we can apply Campanato Theorem to get
0,↵/2
w 2 Cloc (⌦) .
1,↵/2 0,↵/2 2 2
Since s is arbitrary, it follows that u 2 Cloc (⌦) and A = r2 F (ru) 2 Cloc (⌦; Rm ⇥n ).
The Schauder theory that we will consider in the next section (just apply Theorem 10.4
to @xs u, solving the PDE (9.4)) will allow us to conclude that
2,↵/2
u 2 Cloc (⌦) .

As long as F is sufficiently regular, the iteration of this argument solves XIX Hilbert’s
regularity problem in the C 1 category.
We close this section with a more technical but useful iteration lemma in the same
spirit of Lemma 9.1.

Lemma 9.2 (Iteration Lemma). Consider a non-decreasing real function f : (0, R0 ] ! R

which satisfies for some coefficients A > 0, B 0 and exponents ↵ > the following
inequality h⇣ ⇢ ⌘↵ i
f (⇢)  A + " f (R) + BR 8 0 < ⇢  R  R0 . (9.6)
R

59
If
✓ ◆ ↵↵
1
" (9.7)
2A
for some 2 ( , ↵), then
h⇣ ⇢ ⌘ i
f (⇢)  c(↵, , , A) 8 0 < ⇢  R  R0 .
f (R) + B⇢ (9.8)
R
Proof. Without loss of generality, we assume A > 1/2. We choose ⌧ 2 (0, 1) such that
2A⌧ ↵ = ⌧ , (9.9)
thus (9.7) gives the inequality
"  ⌧↵ . (9.10)
The following basic estimate uses the hypothesis (9.6) jointly with (9.9) and (9.10):
f (⌧ R)  A(⌧ ↵ + ")f (R) + BR
 2A⌧ ↵ f (R) + BR = ⌧ f (R) + BR . (9.11)
The iteration of (9.11) easily gives
f (⌧ 2 R)  ⌧ f (⌧ R) + B⌧ R  ⌧ 2 f (R) + ⌧ BR + B⌧ R
= ⌧ 2 f (R) + BR ⌧ (1 + ⌧ ).
It now can be easily proven by induction that
N
X1
N N (N 1) 1 ⌧ N( )
f (⌧ R)  ⌧ f (R) + BR ⌧ ⌧ k( )
= ⌧ N f (R) + BR ⌧ (N 1)
.
k=0
1 ⌧( )

So, given 0 < ⇢  R  R0 , if N verifies

⌧ N +1 R < ⇢  ⌧ N R ,
we conclude choosing the constant c(↵, , , A) in such a way that the last line in the
following chain of inequalities holds:
BR ⌧ (N 1)
f (⇢)  f (⌧ N R)  ⌧ N f (R) +
1 ⌧( )
⌧ 2
 ⌧ ⌧ (N +1) f (R) + BR ⌧ (N +1)
1 ⌧( )
⇣⇣ ⇢ ⌘ ⌘ ⌧ 2
< ⌧ f (R) + B⇢
R ⇣⇣ ⌘ 1 ⌧ ( ) ⌘
⇢
 c(↵, , , A) f (R) + B⇢ .
R
⇤

60
Remark 9.3. The fundamental gain in Lemma 9.2 is the passage from R to ⇢ and the
removal of ", provided that " is small enough. These improvements can be obtained at
the price of passing from the power ↵ to the worse power < ↵.

10 Schauder theory
We are treating Schauder theory in a local form in ⌦ ⇢ Rn , just because it would be
too long and technical to deal also with boundary regularity (some ideas are analogous
to those used in Section 6). We shall describe first a model result for constant coefficient
operators, and then we will consider the case of Hölder continuous coefficients.
We recall the usual PDE we are studying, in a divergence form:
8
< div (Aru) = divF in ⌦ ;
(10.1)
: 1
u 2 Hloc (⌦; Rm ) .

Theorem 10.1. If A↵ij are constant and satisfy the Legendre-Hadamard condition for
some > 0, then for all µ < n + 2 it holds

F 2 L2,µ
loc (⌦) =) ru 2 L2,µ
loc (⌦) .

Proof. In this proof, c = c(n, , |A|) and its value can change from line to line. Since
the estimates we make are local, we assume with no loss of generality that F 2 L2,µ (⌦).
Let us fix a ball BR b ⌦ with center x0 2 ⌦ and compare with u the solution v of the
homogeneous problem 8
< div(Arv) = 0 in BR ;
(10.2)
:
v=u in @BR .
Since rv belongs to H 1 for previous results concerning H 2 regularity and its components
@v
@x↵
solve the same problem (because we supposed to have constant coefficients), we can
use the decay estimates (5.1) and (5.2).
So, if 0 < ⇢ < R, (5.2) provides us with the following inequality:
Z ⇣ ⇢ ⌘n+2 Z
2
|rv(x) (rv)⇢ | dx  c |rv(x) (rv)R |2 dx . (10.3)
B⇢ R BR

Now we try to employ (10.3) to get some estimate for u, the original “non-homogeneous”,
solution of (10.1). Obviously, we can write

u=w+v ,

61
where w 2 H01 (BR ; Rm ). Thus (first using ru = rv + rw, then the minimality of the
mean and (10.3), eventually rv = ru rw and (rw)R = 0)
Z
|ru(x) (ru)⇢ |2 dx
B⇢
Z Z !
 2 |rw(x) (rw)⇢ |2 dx + |rv(x) (rv)⇢ |2 dx
B⇢ B⇢
Z ⇣ ⇢ ⌘n+2Z
 2 |rw(x) (rw)R |2 dx + c |rv(x) (rv)R |2 dx
B R BR
Z ⇢ ⇣ ⇢ ⌘n+2Z
 c |rw(x)|2 dx + c |ru(x) (ru)R |2 dx .
BR R BR

The auxiliary function Z

f (⇢) := |ru(x) (ru)⇢ |2 dx
B⇢

is non decreasingR because of the minimality property of the mean (ru)⇢ , when one min-
imizes m 7! B⇢ |ru(x) m|2 dx. In order to get that f satisfies the hypothesis of
R
Lemma 9.2, we have to estimate BR |rw|2 dx. We can consider w as a function in
H 1 (Rn ) (null out of ⌦) so, by Gårding inequality (choosing the test function ' = w),
Z Z
2
|rw(x)| dx  c Arw(x)rw(x) dx
BR BR
Z Z
= c F (x)rw(x) dx = c (F (x) FR )rw(x) dx (10.4)
BR

because div(Arw)
R = divF by linearity. Applying Young inequality to (10.4) and then
absorbing BR |rw|2 dx in the left side of (10.4), we get
Z Z
2
|rw(x)| dx  c |F (x) FR |2 dx  ckF k2L2,µ Rµ ,
BR BR

because F 2 L2,µ .
Therefore we obtained the decay inequality of Lemma 9.2 for f with ↵ = n + 2, =µ
and " = 0, then ⇣ ⇢ ⌘µ
f (⇢)  c f (R) + c⇢µ ,
R
that is ru 2 L2,µ . ⇤

Corollary 10.2. With the previous notation, when µ = n + 2↵, Theorem 10.1 and Cam-
panato Theorem 8.9 yield that
F 2 C 0,↵ =) ru 2 C 0,↵ .

62
 In the next theorem we consider the case of variable, but continuous, coefficients,
proving in this case a Lp,µ regularity of |ru| with µ < n; as we have seen, the Poincaré
inequality then provides Hölder regularity at least of u if µ + p > n.
Theorem 10.3. Considering again (10.1), suppose that A↵ij 2 C(⌦) and A satisfies a
(locally) uniform Legendre-Hadamard condition for some > 0. If F 2 L2,µ
loc with µ < n,
then |ru| 2 L2,µ
loc .
Naturally, since µ < n, Campanato spaces and Morrey spaces coincide, so that we
used Morrey spaces for simplicity.
Proof. Here is an example of Korn’s technique of freezing of coefficients. We use the
same convention on c of the previous proof, namely c = c(n, , sup |A|).
Fix a point x0 2 ⌦ and define
F̃ (x) := F (x) + (A(x0 ) A(x)) ru(x) ,
so that the solution u of (10.1) solves
div(A(x0 )ru(x)) = divF̃ (x) with F̃ (x) := F (x) + (A(x0 ) A(x))ru(x) .
Write u = v + w, where v solves the homogeneous PDE (10.2) with frozen coefficients
A(x0 ). Using (5.1) for v we obtain
Z ⇣ ⇢ ⌘n Z Z
2 2
|ru(x)| dx  c |rv(x)| dx + c |rw(x)|2 dx
B⇢ R BR B
⇣ ⇢ ⌘n Z Z R
 c |rv(x)|2 dx + c |F̃ (x) FR |2 dx .
R BR BR

Thanks to the continuity property of A, there exists a (local) modulus of continuity ! of

A which allows us to estimate
Z Z Z
2 2 2
|F̃ (x)| dx  2 |F (x) FR | dx + 2! (R) |ru(x)|2 dx . (10.5)
BR BR BR

Consequently, as F 2 L2,µ
loc ,
Z Z
2 µ 2
|F̃ (x) FR | dx  c̃R + 2! (R) |ru(x)|2 dx
BR BR

with c̃ depending only on kF kL2,µ . We are ready to use Lemma 9.2 with f (⇢) :=
R 2
loc

B⇢
|ru(x)| dx, ↵ = n, = µ < n and " = ! 2 (R): it tells us that if R is under a
threshold depending only on c, ↵, , ! and kF kL2,µ we have
loc
⇣ ⇢ ⌘µ
f (⇢)  c f (R) + c⇢µ ,
R
so that |ru| 2 L2,µ
loc . ⇤

63
 We can now prove Schauder theorem for elliptic PDE’s in divergence form. In the
non-divergence form the result is (in the scalar case)
X @ 2u
A↵ 2 C 0,↵ =) u 2 C 2,↵ , (10.6)
↵,
@x ↵ @x

if A is of class C 0,↵ . The proof follows similar lines, i.e. starting for second derivative decay
estimates for constant coefficient operators, and then freezing the coefficients. Notice also
that both (10.6) and Theorem 10.4 below are easily seen to be optimal, considering 1-
dimensional ODE’s au00 = f or (au0 )0 = f 0 .
Theorem 10.4 (Schauder). Suppose that the coefficients A↵ij (x) of the PDE (10.1) belong
to C 0,↵ (⌦) and A satisfies a (locally) uniform Legendre-Hadamard in ⌦ for some > 0.
Then the following implication holds
0,↵ 0,↵
F 2 Cloc =) ru 2 Cloc ,
that is to say
F 2 L2,n+2↵
loc =) ru 2 L2,n+2↵
loc .
Proof. With the same idea of freezing coefficients (and the same notation, too), we
estimate by (5.1)
Z ⇣ ⇢ ⌘n+2 Z Z
2 2
|ru(x) (ru)⇢ | dx  c |ru(x) (ru)R | dx + c |F̃ (x) FR |2 dx .
B⇢ R BR BR
(10.7)
Additionally, the Hölder property of A makes us rewrite (10.5) as
Z Z Z
2 2 2↵
|F̃ (x) FR | dx  2 |F (x) FR | dx + cR |ru(x)|2 dx . (10.8)
BR BR BR
0,↵
Since F 2 Cloc , we obtain
Z Z
2 n+2↵ 2↵
|F̃ (x) FR | dx  cR + cR |ru(x)|2 dx .
BR BR

Theorem 10.3 with µ = n ↵ < n tells us that |ru| 2 L2,µ , thus

Z
|F̃ (x) FR |2 dx  cRn+2↵ + cRn+↵ . (10.9)
BR

Adding (10.9) to (10.7) and applying Lemma 9.2 with exponents n + 2 and n + ↵, we
get ru 2 L2,n+↵ , so that ru 2 C 0,↵/2 , in particular |ru| is locally bounded. Using this
information we can improve (10.9) as follows:
Z
|F̃ (x) FR |2 dx  cRn+2↵ .
BR
Now we reach the conclusion, again by Lemma 9.2 with exponents n + 2 and n + 2↵.
⇤

64
11 Regularity in Lp spaces
In this section we deal with elliptic regularity in the category of Lp spaces, obviously a
natural class of spaces besides Morrey, Hölder and Campanato spaces.

Lemma 11.1. In a measure space (⌦, F, µ), consider a F-measurable function f : ⌦ !

[0, 1] and set
F (t) := µ ({x 2 ⌦ : f (x) > t}) .
The following equalities hold for 1  p < 1:
Z Z 1
p
f (x) dµ(x) = p tp 1 F (t) dt (11.1)
Z ⌦ Z0 1
f p (x) dµ(x) = p tp 1 F (t) dt + sp F (s) 0s<1. (11.2)
{f >s} s

Proof. It is a simple consequence of Fubini’s Theorem that

Z Z Z f (x) ! Z 1 ✓Z ◆
p p 1
f (x) dµ(x) = p t dt dµ(x) = p tp 1
{t<f (x)} dµ(x) dt
⌦ ⌦ 0 0
Z 1
= p tp 1 F (t) dt .
0

Equation (11.2) follows from (11.1) applied to the function f {f >s} . ⇤

Theorem 11.2 (Markov inequality). In a measure space (⌦, F, µ), a function f 2

Lp (⌦, F, µ) satisfies (with the convention 0 ⇥ 1 = 0)
Z
p
t µ ({|f | t})  |f |p dµ 8t 0 . (11.3)
⌦
Proof. We begin with the trivial pointwise inequality

s {g s} (x)  g(x) 8x 2 ⌦ (11.4)

for g nonnegative. Thus, integrating (11.4) in ⌦ we obtain

Z
sµ ({g s})  g dµ .
⌦

The thesis follows choosing s = tp and g = |f |p . ⇤

65
 The Markov inequality inspires the definition of a space which is weaker than Lp , but
still keeps (11.3).

Definition 11.3 (Marcinkiewicz space). Given a measure space (⌦, F, µ) and an exponent
1  p < 1, the Marcinkiewicz space Lpw (⌦, µ) is defined by

Lpw (⌦, µ) := f : ⌦ ! R F-measurable sup tp µ ({|f | > t}) < 1 .

t>0

We denote4 with kf kpLpw the smallest constant c satisfying

tp µ ({|f | > t})  c 8t > 0 .

Remark 11.4. If µ is a finite measure, then

q<p =) Lp ⇢ Lpw ⇢ Lq .

The first inclusion is due to Markov inequality (11.2), on the other hand, if f 2 Lpw , then
Z Z 1 ✓Z 1 Z 1 ◆
q q 1 q 1 q 1
|f | dµ(x) = q t F (t) dt  q t F (t) dt + t F (t) dt
⌦ 0 0 1
Z 1
q
 qµ(⌦) + q tq 1 kf kpLpw t p dt = qµ(⌦) + kf kpLpw .
1 p q

Definition 11.5 (Maximal operator). When f 2 L1loc (Rn ) we define the maximal function
Mf by Z
Mf (x) := sup |f (y)| dy , (11.5)
Qr (x) Qr (x)

where Qr (x) is the n-dimensional cube with center x and side length r.

It is easy to check that at Mf (x) f˜(x) at Lebesgue points, so that Mf f L n -a.e.

in Rn . On the other hand, it is important to remark that the maximal operator M does
not map L1 into L1 .

Example 11.6. In dimension n = 1, consider f = [0,1] 2 L1 . Then

1
Mf (x) = when |x| 1,
2|x|

/ L1 . In fact, it is easy to prove that Mf 2 L1 implies |f | = 0 L n -a.e. in Rn .

so Mf 2
4
Pay attention to the lack of subadditivity of k · kLpw : the notation is misleading, this is not a norm!
For instance both 1/x and 1/(1 x) have weak L1 norm equal to 1 on ⌦ = (0, 1), but their sum has weak
L1 norm strictly greater. On the other hand, it is easily seen that kf + gkLpw  2kf kLpw + 2kgkLpw

66
 However, if f 2 L1 , the maximal operator Mf belongs to the weaker Marcinkiewicz
space L1w , as we are going to see in Theorem 11.8. We first recall the Vitali covering
theorem, in a version valid in any metric space.

Lemma 11.7 (Vitali). Let F be a finite family of balls in a metric space (X, d). Then,
there exists G ⇢ F, made of disjoint balls, satisfying
[ [
B⇢ B̂ .
B2F B2G

Here, for B ball, B̂ denotes the ball with the same center and triple radius.
Proof. The initial remark is that if B1 and B2 are intersecting balls then B1 ⇢ B c2 ,
provided the radius of B2 is larger than the radius of B1 . Assume that the family of balls
is ordered in such a way that their radii are non-increasing. Pick the first ball B1 , then
pick the first ball among those that do not intersect B1 and continue in this way, until
either there is no ball left or all the balls left intersect one of the chosen balls. The family
G of chosen balls is, by construction, disjoint. If B 2 F \ G, then B has not been chosen
because it intersects one of the balls in G; the first of these balls Bf has radius larger
than the radius of B (otherwise B would have been chosen before Bf ), hence B ⇢ B cf .
⇤

Theorem 11.8 (Hardy-Littlewood maximal theorem). The maximal operator Mf de-

fined in (11.5) satifies

kMf kL1w  3n kf kL1 8f 2 L1 (Rn ) .

Proof. Fix t > 0 and a compact set K ⇢ {Mf > t}: by inner regularity of the Lebesgue
measure we will reach the conclusion showing that
3n
L n (K)  kf kL1 .
t
Since K ⇢ {Mf > t}, for any x 2 K there exists a radius r(x) such that
Z
|f (y)| dy t(r(x))n .
Qr(x) (x)

Compactness allows us to cover K with a finite number of cubes

[
K⇢ Qr(xi ) (xi ) ,
i2I

67
then Vitali’s lemma stated for the distance induced by the sup norm in Rn allows us to
find J ⇢ I such that the cubes Qr(xj ) (xj ), j 2 J, are pairwise disjoint and
[ [
Q3r(xj ) (xj ) Qri (xi ) K .
j2J i2I

We conclude that
X Z
n n n 3n X 3n
L (K)  3 (r(xj ))  f (y) dy  kf kL1 .
j2J
t i2I Qr(x ) (xi ) t
i

12 Some classical interpolation theorems

In the sequel, we will make extensive use of some classical interpolation theorems, that
are basic tools in Functional and Harmonic Analysis.
Assume (X, F, µ) is a measure space. For the sake of brevity, we will say that a linear
operator T mapping a vector space D ⇢ Lp (X, µ) into Lq (X, µ) is of type (p, q) if it is
continuous with respect to the Lp Lq topologies. If this happens, obviously T can be
extended (by Hahn-Banach) to a linear continuous operator from Lp (X, µ) to Lq (X, µ)
and the extension is unique if D is dense.
The inclusion Lp \ Lq ⇢ Lr for p  q and r 2 [p, q] can be better specified with the
following result.
Theorem 12.1 (Riesz-Thorin interpolation). Let p, q 2 [1, 1] with p  q and T :
Lp (X, µ) \ Lq (X, µ) ! Lp (X, µ) \ Lq (X, µ) a linear operator which is both of type (p, p)
and (q, q). Then T is of type (r, r) for all r 2 [p, q].
We do not give the proof of this theorem, whose proof follows the lines of the more
general Marcinkiewicz theorem below (a standard reference is [26]). In the sequel we shall
consider operators T that are not necessarily linear, but Q-subadditive for some Q 0,
namely
|T (f + g)|  Q(|T (f )| + |T (g)|) 8f, g 2 D .
For instance, the maximal operator is 1-subadditive. We also say that a space D of real-
valued functions is stable under truncations if f 2 D implies f {|f |<k} 2 D for all k > 0
(all Lp spaces are stable under truncations).
Definition 12.2 (Strong and weak (p, p) operators). Let s 2 [1, 1], D ⇢ Ls (X, µ) a
linear subspace and let T : D ⇢ Ls (X, µ) ! Ls (X, µ), not necessarily linear. We say
that T is of strong type (s, s) if kT (u)ks  Ckuks for all u 2 D, for some constant C

68
independent of u.
If s < 1, we say that T is of weak type (s, s) if
kukss
µ ({x : |T u(x)| > ↵})  C s 8↵ > 0, u 2 D
↵
for some constant C independent of u and ↵. Finally, by convention, T is called of weak
type (1, 1) if it is of strong type (1, 1).
We can derive an appropriate interpolation theorem even in the case of weak continuity.
Theorem 12.3 (Marcinkiewicz Interpolation Theorem). Assume that p, q 2 [1, 1] with
p < q, D ⇢ Lp (X, µ) \ Lq (X, µ) is a linear space stable under truncations and T : D !
Lp (X, µ) \ Lq (X, µ) is Q-subadditive, of weak type (p, p) and of weak type (q, q).
Then T is of strong type (r, r) for all r 2 (p, q).
Remark 12.4. The most important application of the previous result is perhaps the
study of the boundedness of maximal operators (see the next Remark). In that case, one
typically works with p = 1 and q = 1 and we limit ourselves to prove the theorem under
this additional hypothesis.
Proof. We can truncate f 2 D as follows:
f = g + h, g(x) = f (x) {|f | s} (x), h(x) = f (x) {|f |> s} (x) ,

where is an auxiliary parameter to be fixed later. By assumption g 2 D \ L1 (X, µ)

while h 2 D \ L1 (X, µ) by linearity of D. Hence
|T (f )|  Q|T (g)| + Q|T (h)|  QA1 s + Q|T (h)|
with A1 as the operator norm of T acting from D \ L1 (X, µ) into L1 (X, µ). Choose
so that QA1 = 1/2, therefore
s
{|T (f )| > s} ⇢ {|T (h)| > }
2Q
and so
✓ ◆Z ✓ ◆Z
s 2A1 Q 2A1 Q
µ {|T (f )| > s}  µ {|T (h)| > }  |h| dµ  |f | dµ ,
2Q s X s {|f |> s}

where A1 is the constant appearing in the weak (1, 1) estimate. By integration of the
previous inequality, we get
Z 1 Z 1Z
p 1
p s µ ({|T (f )| > s}) ds  2A1 Qp sp 2 |f | dµ ds
0 0 {|f | s}

and by means of the Fubini-Tonelli Theorem we finally get

Z Z |f (x)|/ !
2A1 Qp
kT (f )kpp  2A1 Qp sp 2 ds |f (x)| dµ(x) = kf kpp
X 0 (p 1) p 1

and the conclusion follows. ⇤

69
Remark 12.5 (The limit case p = 1). In the limit case p = 1 we can argue similarly to
find
Z 1
µ ({|T (f )| > s}) ds
1
Z Z |f (x)|/ ! Z
1
 2A1 Q s ds |f (x)| dµ(x) = 2A1 Q |f | log |f | dµ.
{|f | } 1 {f }

Therefore, a slightly better integrability of |f | provides at least integrability of |T (f )| on

bounded sets.
Remark 12.6. As a byproduct of the previous result, we have that the maximal operator
M defined in the previous section is of strong type (p, p) for any p 2 (1, 1] (and only
of weak type (1, 1)). These facts, which have been derived for simplicity in the standard
Euclidean setting, can be easily generalized, for instance to pseudo-metric spaces (i.e.
when the distance fulfils only the triangle and symmetry assumptions) endowed with a
doubling measure, that is a measure µ such that µ(B2r (x))  µ(Br (x)) for some constant
not depending on the radius and the center of the ball. Notice that in this case the
constant in the weak (1, 1) bound of the maximal operator does not exceed 2 , since
µ(B3r (x))  2 µ(Br (x)).

13 Lebesgue di↵erentiation theorem

In this section, we want to give a direct proof, based on the (1, 1)-weak continuity of the
maximal operator M, of the classical Lebesgue di↵erentiation theorem.
Theorem 13.1. Let (X, d, µ) be a metric space with a finite doubling measure on its Borel
-algebra and p 2 [1, 1). If f 2 Lp (µ) then for µ-a.e. x 2 X we have that
Z
lim |f (y) f (x)|p dµ(y) = 0.
r#0 Br (x)

Proof. Let ⇢ Z
⇤t := x 2 X| lim sup |f (y) f (x)|p dµ(y) > t .
r#0 Br (x)

The thesis can be achieved showing that for any t > 0 we have µ(⇤t ) = 0, since the stated
property holds out of [n ⇤1/n . Now, we can exploit the metric structure of X in order to
approximate f in L1 (µ) norm by means of continuous and bounded functions: for any
" > 0 we can write f = g + h with g 2 Cb (X) and khkpLp  t". Hence, it is enough to
prove that for any t > 0 we have µ(At ) = 0 where
⇢ Z
At := x 2 X| lim sup |h(y) h(x)|p dµ(y) > t .
r#0 Br (x)

70
This is easy, because by definition
⇢ ⇢
t t
At ⇢ |h|p > [ M(|h|p ) >
2p+1 2p+1
and, if we consider the corresponding measures, we have (taking Remark 12.6 into account)
2p+1 2p+1
µ(At )  khkpLp + M khkpLp  2p+1 (1 + M )"
t t
where M is the constant in the weak (1, 1) bound. Since " > 0 is arbitrary we get the
thesis. ⇤

Remark 13.2. All the previous results have been derived for the maximal operator
defined in terms of centered balls, that is
Z
M f (x) = sup f (y) dy
r>0 Br (x)

and the Lebesgue di↵erentiation theorem has been stated according to this setting. How-
ever, it is clear that we can generalize everything to any metric space (X, d, µ) with a
finite doubling measure and a suitable family of sets F := [x2X Fx with
Z
MF f (x) = sup f (y) dy
A2Fx A

provided there exists a universal constant C > 0 such that

for all A 2 Fx there exists r > 0 such that A ⇢ Br (x) and µ(A) Cµ(Br (x)) . (13.1)
Indeed, even though one might define the maximal operator with this larger family of
mean values, suffices just to notice that
Z Z
1
|f (y) f (x)| dµ(y)  |f (y) f (x)| dµ(y) ,
A C Br (x)
provided Br (x) is chosen according to (13.1).
In Euclidean spaces, an important example to which the previous remark applies, in
connection with Calderón-Zygmund theory, is given by
Fx := {Q cube, x 2 Q} ,
consequently Lebesgue theorem gives
Z
lim |f (y) f (x)|p dy = 0
x2Q, |Q|!0 Q

for a.e. x 2 Rn , Notice that requiring |Q| ! 0 (i.e. diam(Q) ! 0) is essential to “factor”
continuous functions as in the proof of Theorem 13.1.

71
14 Calderón-Zygmund decomposition
We need to introduce another powerful tool, that will be applied to the study of the BM O
spaces. Here and below Q will indicate an open cube in Rn and similarly Q0 or Q00 .

Theorem
R 14.1. Let f 2 L1 (Q), f 0 and consider a positive real number ↵ such that
Q
f dx  ↵. Then, there exists a finite or countable family of open cubes {Qi }i2I with
Qi ⇢ Q and sides parallel to the ones of Q, such that

(i) Qi \ Qj = ; if i 6= j;
R
(ii) ↵ < Qi f dx  2n ↵ 8i;

(iii) f  ↵ a.e. on Q \ [i Qi .

Remark 14.2. The remarkable (and useful) aspect of this decomposition is that the
“bad” set {f > ↵} is almost all packed inside a family of cubes, carefully chosen in such
a way that still the mean values inside the cubes is of order ↵. As a consequence of the
existence of this decomposition, we have
X XZ
n
↵ L (Qi ) < f dx  kf k1 .
i i Qi

The proof is based on a so-called stopping-time argument.

Proof. Divide the cube Q in 2n subcubes by means of n bisections of Q with hyperplanes
parallel to the sides of the cube itself. We will call this process dyadic decomposition.
Then
R
• if Qi f > ↵ we do not divide Qi anymore;

• else we iterate the process on Qi .

At each step we collect the cubes that verify the first condition and put together all
such cubes, thus forming a countable family. The first two properties are obvious
R by
construction: indeed, if Qi is a chosen cube then its parent cube Q̃i satisfies Q̃i f  ↵,
R
which gives easily Qi f  2n ↵. For the third one, note that if x 2 Q \ [i Qi , then there
R
exists a sequence of subcubes (Q ej ) with x 2 \j Q
ej and L n Q ej ! 0, e f dx  ↵.
Qj
Thanks to the Lebesgue di↵erentiation theorem we get f (x)  ↵ for a.e. x 2 Q \ [i Qi .
⇤

72
Remark 14.3 (Again in the limit case p = 1). Using the Calderon-Zygmund decompo-
sition, for ↵ > kf k1 we can reverse somehow the weak (1, 1) estimate:
Z XZ
|f | dx  |f | dx  2n ↵L n (Qi )  2n ↵L n {M |f | > ↵/2n } ,
{|f |>↵} i Qi

because the cubes Qi are contained in {M |f | > ↵/2n }. Using this inequality we can also
reverse theRimplication of Remark 12.5, namely assuming with no loss of generality that
f 0 and f dx = 1:
Z Z 1Z Z 1 Z
1
f log f dx = f dxdt = f dx
{f >1} 0 {log f >t} 1 s {f >s}
Z 1 Z
n s 1 +
 2 L {M f > } ds = 2 (M f ) dx .
1 2 2

15 The BMO space

Given a cube Q ⇢ Rn , we define
⇢ Z
1
BM O(Q) := u 2 L (Q)| sup |u uQ0 | dx < 1 ,
Q0 ⇢Q Q0

where uQ0 denotes the mean value of u on Q0 . We also define the seminorm kukBM O
as the supremum in the right hand side. An elementary argument replacing balls with
concentric cubes shows that BM O(Q) ⇠ L1,n , that is the two spaces consist of the same
elements and the corresponding semi-norms are equivalent. Here we recall the inclusion
already discussed in Remark 8.8.

Theorem 15.1. For any cube Q ⇢ Rn the following inclusion holds:

W 1,n (Q) ,! BM O(Q).

Proof. First, notice that W 1,n (Q) ,! {u| |ru| 2 L1,n 1 (Q)}, as an immediate conse-
quence of the Hölder inequality. Then, by Poincaré inequality, there exists a dimensional
constant C > 0 such that for any Q0 ⇢ Q with sides of length h
Z Z
|u uQ0 | dx  Ch |ru| dx  C |ru|L1,n 1 hn .
Q0 Q0

73
 However, it should be clear that the previous inclusion is far from being an equality
as elementary examples show, see Remark 8.8. We shall extend now to n-dimensional
spaces the example in Remark 8.8, stating first a simple sufficient (and necessary, as we
will see) condition for BMO.

Proposition 15.2. Let u : Q ! R be a measurable function such that, for some b > 0,
B 0, the following property holds:

8 C ⇢ Q cube, 9 aC 2 R s.t. L n C \ {|u aC | > }  Be b

|C| 8 0. (15.1)

Then u 2 BM O(Q).

The proof of the proposition above is simple, since

Z Z Z 1
1 B
|u uC | dx  |u aC | dx = L n C \ {|u aC | } d  |C| .
2 C C 0 b

Example 15.3. Thanks to Proposition 15.2 we can check that ln |x| 2 BM O (0, 1)n .
Indeed, ln |x| satisfies (15.1) (the parameters b and B will be made precise later). To see
this, fix a cube C, with h the length of the side of C. We define, respectively,

⇠ := max |x| , ⌘ := min |x| , aC := ln ⇠ ,

x2C x2C

so that ✓ ◆
⇠
aC u = ln 0.
|x|
We estimate the Lebesgue measure of C \ {⇠ |x|e }: naturally we can assume that
⇠ ⌘e , otherwise there is nothing to prove, so
p
⇠e ⌘ ⇠ diam(C) ⇠ nh ,

then p
nh
⇠ .
1 e
Finally p
1 n 1 ( n)n !n
L (C \ {|u aC | })  n L n B⇠e  e n
,
hn h (1 e )n
so that distinguishing
p n the cases  1 and > 1 we see that (15.1) holds with b = n and
n 1 n
B = max{e , ( n) !n (1 e ) }.

The following theorem by John and Nirenberg was first proved in [21].

74
Theorem 15.4 (John-Nirenberg, first version). There exist constants c1 , c2 depending
only on the dimension n such that

L n ({|u uQ | > t})  c1 e c2 t/kukBM O

L n (Q) 8u 2 BM O(Q) \ {0} . (15.2)

Remark 15.5. In the proof we present here, we will find explicitly c1 = e and c2 =
1/(2n e). However, these constants are not sharp.
Before presenting the proof, we discuss here two very important consequences of this
result.
Corollary 15.6 (Exponential integrability of BM O functions). For any c < c2 there
exists K(c, c1 , c2 ) such that
Z
ec|u uQ |/kukBM O dx  K(c, c1 , c2 ) 8u 2 BM O(Q) \ {0} .
Q

Proof. It is a simple computation:

Z Z 1 Z 1
c|u uQ | cc1
e dx = c ect L n ({|u uQ | > t}) dt  cc1 e(c c2 )t
dt = ,
Q 0 0 c2 c

where we assumed kukBM O(Q) = 1, L n (Q) = 1 and we used the John-Nirenberg inequal-
ity. ⇤

Remark 15.7 (Better integrability of W 1,n functions). The previous theorem tells that
the class BM O (and hence also W 1,n ) has exponential integrability properties. This result
can be in part refined by the celebrated Moser-Trudinger inequality, that we quote here
without proof.
1/(n 1)
For any n > 1 set ↵n := n!n 1 . and consider a bounded domain ⌦ in Rn , with
n > 1. Then
⇢Z Z
n/(n 1) 1,n
C(⌦) := sup exp ↵n |u| dx : u 2 W0 (⌦), |ru|n dx  1 < 1 .
⌦ ⌦

This inequality has been first proved in [27].

Theorem 15.8. If p 2 [1, 1) we have
✓Z ◆1/p
p
|u uQ | dx  c(n, p)kukBM O 8u 2 BM O(Q) .
Q

Consequently the following isomorphisms hold:

Lp,n (Q) ⇠ BM O(Q) ⇠ L1,n (Q) . (15.3)

75
 The proof of Theorem 15.8 relies on a simple and standard computation, similar
to the one presented before in order to get exponential integrability. Indeed, assuming
kukBM O = 1, (15.2) gives
Z Z 1 Z 1
p n p 1
|u uQ | dx = p L {|u uQ | > s} s ds  c1 p e c2 s sp 1 ds .
Q 0 0

We can now conclude this section, by proving the John-Nirenberg inequality (15.2).
Proof. By homogeneity, we can assume without loss of generality that kukBM O = 1. Let
↵ > 1 be a parameter, to be specified later. We claim that it is possible to define, for any
k 1 a countable family of subcubes Qki i2I contained in Q such that
k

(i) |u(x) uQ |  2n k↵ a.e. on Q \ [i2Ik Qki ;

P n k k n
(ii) i2Ik L (Qi )  ↵ L (Q).

The combination of linear growth in (i) and geometric decay in (ii) leads to the exponential
decay of the repartition function: indeed, choose k such that 2n ↵k  t < 2n ↵(k + 1), then

L n ({|u uQ | > t})  L n ({|u uQ | > 2n ↵k})  ↵ k L n (Q)

k c2 t
by the combined use of the previous properties. Now we want ↵  c1 e for all
t 2 [2n ↵k, 2n ↵(k + 1)), which is certainly verified if
k c2 2n ↵(k+1)
↵ = c1 e

and consequently we determine the constants c1 , c2 , requiring

n↵ c2 2n ↵
e c2 2 = ↵, c1 e =1.

By the first relation c2 = log ↵/(2n ↵) and we maximize with respect to ↵ > 1 to find
1
↵ = e, c1 = e, c2 = .
2n e
Now we just need to prove the claim. If k = 1 we simply apply the Calderón-Zygmund
decomposition to f = |u uQ | for the level ↵ and get a collection {Q1i }i2I1 . We have to
verify that the required conditions are verified. Condition (ii) follows by Remark 14.2,
while (i) is obvious since |u(x) uQ |  ↵ a.e. out of the union of Q1i by construction.
But, since kukBM O = 1, we also know that
Z
8i 2 I1 |u uQ1i | dx  1 < ↵ ,
Q1i

76
hence we can iterate the construction, by applying the Calderón-Zygmund decomposition
to each of the functions |u uQ1i | with respect to the corresponding cubes Q1i . In this way,
we find a family of cubes Q2i,l , each contained in one of the previous ones. Moreover
Remark 14.2 and the induction assumption give
X X1Z X1 1
n 2
L (Qi,l )  |u uQ1i | dx  L n (Q1i )  2 L n (Q) ,
i,l i
↵ Q1i i
↵ ↵

which is (ii). In order to get (i), notice that

! !
[ [ [ [
Q\ Q2i,l ⇢ Q\ Q1i [ (Q1i \ Q2i,l )
i i l

so for the first set in the inclusion the thesis is obvious by the case k = 1. For the second
one, we first observe that
Z
|uQ uQ1i |  |uQ u| dx  2n ↵
Q1i

and consequently, since |u uQ1i |  ↵ on Q1i \ [l Q2i,l we get

|u(x) uQ |  |u(x) uQ1i | + |uQ1i uQ |  ↵ + 2n ↵  2n · 2↵ .

With minor changes, we can deal with the general case k > 1 and this is what we need
to conclude the argument and the proof. ⇤
The John-Nirenberg theorem stated in Theorem 15.4 can be extended considering the
p
L norms, so that the case of BM O maps corresponds to the limit as p ! 1.
Theorem 15.9 (John-Nirenberg, second version). For any p 2 [1, 1) and u 2 Lp (Q)
define
( ✓Z ◆p )
X
p n
Kp (u) := sup L (Qi ) |u(x) uQi | dx | {Qi } partition of Q .
i Qi

Then there exists a constant c = c(p, n) such that

ku uQ kLpw  c(p, n)Kp (u) .

The proof of Theorem 15.9 is basically the same as Theorem 15.4, the goal being to
prove the polynomial decay
c(p, n)
|{|u uQ | > t}|  Kp (u) t>0
tp
77
instead of an exponential decay.
The following important result improves the classical interpolation theorems in Lp
spaces, replacing L1 with BM O. This is crucial for the application to elliptic PDE’s, as
we will see.

Theorem 15.10 (Stampacchia’s interpolation). Let D ⇢ L1 (Q; Rs ) be a linear space

and p 2 [1, 1). Consider a linear operator T : D ! BM O(Q0 ), continuous with respect
to the norms (L1 (Q; Rs ), BM O(Q0 )) and (Lp (Q; Rs ), Lp (Q0 )). Then for every r 2 [p, 1)
the operator T is continuous with respect to the (Lr (Q; Rs ), Lr (Q0 )) topologies.
Proof. For simplicity we assume s = 1 (the proof is the same in the general case). We
fix a partition {Qi } of Q and we regularize the operator T with respect to {Qi } (even if
we do not write the dependence of T̃ from {Qi } for brevity):
Z
T̃ (u)(x) := |T u(y) (T u)Qi | dy 8 x 2 Qi .
Qi

We claim that T̃ satisfies the assumptions of Marcinkiewicz theorem. Indeed

(1) T̃ is obviously 1-subadditive;

(2) L1 ! L1 continuity holds by the inequality

Z
kT̃ ukL1 = sup |T u(y) (T u)Qi | dy  kT ukBM O  ckukL1 ;
i Qi

(3) Lp ! Lp continuity holds too, in fact, by Jensen’s inequality,

X ✓Z ◆p
p n
kT̃ ukLp = L (Qi ) |T u(y) (T u)Qi | dy
i Qi
XZ
 |T u(y) (T u)Qi |p dy
i Qi
XZ
 2p 1
(|T u(y)|p + |(T u)Qi |p ) dy  2p kT ukpLp  c2p kukpLp .
i Qi

Thanks to Marcinkiewicz theorem the operator

T̃ : D ⇢ Lr (Q) ! Lr (Q0 ) (15.4)

is continuous for every r 2 [p, 1], and its continuity constant c can be bounded indepen-
dently of the chosen partition {Qi }.

78
 In order to get information from Theorem 15.9, for r 2 [p, 1), we estimate
X ✓Z ◆r
r n
Kr (T u) = sup L (Qi ) |T u(y) (T u)Qi | dy = sup kT̃{Qi } ukrLr  ckukLr ,
{Qi } i Qi {Qi }

where we used the continuity property of T̃ : Lr (Q) ! Lr (Q0 ) stated in (15.4). Therefore,
by Theorem 15.9, we get
kT u (T u)Q kLrw  c(r, n, T )kukLr 8u 2 D .
Since u 7! (T u)Q obviously satisfies a similar Lrw estimate, we conclude that kT ukLrw 
c(r, n, T )kukLr for all u 2 D. Again, thanks to Marcinkiewicz theorem, with exponents
0 0
p and r, we have that the continuity Lr ! Lr holds for every r0 2 [p, r). Since r is
arbitrary, we got our conclusion. ⇤
We are now ready to employ these harmonic analysis tools to the study of regularity
in Lp spaces for elliptic PDEs, considering first the case of constant coefficients. Suppose
that ⌦ ⇢ Rn is an open, bounded set with Lipschitz boundary @⌦, suppose that the
coefficients A↵ij satisfy the Legendre-Hadamard condition with > 0 and consider the
divergence form of the PDE
⇢
div(Aru) = divF
(15.5)
u 2 H01 (⌦; Rm ) .
In the spirit of Theorem 15.10, we define
T F := ru.
Thanks to Campanato regularity theory, we already got the continuity of T : L2, ! L2,
when 0  < n + 2, thus choosing = n and using the isomorphism (15.3) we see that
T is continuous as an operator
T : L1 (⌦; Rm⇥n ) ! BM O(⌦; Rm⇥n ) . (15.6)
Remark 15.11. Let us remark the importance of weakening the norm in the target
space in (15.6): we passed from L1 (for which, as we will see, no estimate is possible) to
BM O. For BM O the regularity result for PDEs is true and Theorem 15.10 allows us to
interpolate between 2 and 1.
We are going to apply Theorem 15.10 with D = L1 (⌦; Rs ) and s = m ⇥ n. By the
global Caccioppoli-Leray inequality (see Theorem 6.1) we obtain the second hypothesis
of Theorem 15.10: T : L2 (⌦; Rm⇥n ) ! L2 (⌦; Rm⇥n ) is continuous. Therefore
T : D ! Lp (⌦; Rm⇥n ) (15.7)
is (Lp , Lp )-continuous if p 2 [2, 1). Since the (unique) extension of T to the whole of Lp
still maps F into ru, with u solution to (15.5), we have proved the following result:

79
Theorem 15.12. For all p 2 [2, 1) the operator F 7! ru in (15.5) maps Lp (⌦; Rm⇥n )
into Lp (⌦; Rm⇥n ) continuously.

Our intention is now to extend the previous result for p 2 (1, 2), by a duality argument.

Lemma 15.13 (Helmholtz decomposition). If p 2 and B is a matrix satisfying the

Legendre-Hadamard inequality, a map G 2 L (⌦; R
p m⇥n
) can always be written as a sum

G = Br + G̃ , (15.8)

where (understanding the divergence w.r.t. the spatial components)

div(G̃) = 0 in ⌦

and, for some constant c⇤ > 0, the following inequality holds:

kr kLp  c⇤ kGkLp . (15.9)

Proof. It is sufficient to solve in H01 (⌦; Rm ) the PDE

div(Br ) = divG .

and set G̃ := G Br . The estimate (15.9) is just a consequence of Theorem 15.12.

⇤
Fix q 2 (1, 2), so that its conjugated exponent p is larger than 2, and set D :=
L (⌦; Rm⇥n ). Our aim is to prove that T : L2 ! Lq is (Lq , Lq )-continuous. We are going
2

to show that, for every F 2 D, T F belongs to (Lp )0 ⇠ Lq . In the chain of inequalities

that follows we are using A⇤ , that is the adjoint matrix of A, which certainly keeps the
Legendre-Hadamard property. Lemma 15.13 is used in order to decompose the generic
function G 2 Lp as in (15.8), so
Z
sup hT F, Gi = sup T F (x)G(x) dx
kGkLp 1 kGkLp 1
Z ⇣ ⌘
⇤
= sup ru(x) A r (x) + G̃(x) dx
kGkLp 1
Z
 sup (Aru(x)) r (x) dx
kr kLp c⇤
Z
= sup F (x)r (x) dx  c⇤ kF kLq .
kr kLp c⇤

If we approximate now F 2 Lq in the Lq topology by functions Fn 2 L2 we can use

the (Lq , Lq )-continuity to prove existence of weak solutions to the PDE in H01,q , when

80
the right hand side is Lq only. Notice that the solutions
R obtained in this way have no
variational character anymore, since their energy Aruru dx is infinite (for this reason
they are sometimes called very weak solutions). Since the variational characterization
is lacking, the uniqueness of these solutions needs a new argument, based on Helmholtz
decomposition.
Theorem 15.14. For all q 2 (1, 2) there exists a continuous operator T : Lq (⌦; Rm⇥n ) !
H01,q (⌦; Rm ) mapping F to the unique weak solution u to (15.5).
Proof. We already illustrated the construction of a solution u, by a density argument
and uniform Lq bounds. To show uniqueness, it suffices to show that u 2 H01,q and that
div(Aru) = 0 implies u = 0. To this aim, we define G = |ru|q 2 ru 2 Lp and we apply
Helmholtz decomposition G = A⇤ r + G̃ with 2 H01,p and G̃ 2 Lp divergence-free. By
aR density argumentRw.r.t. u and w.r.t. (notice that the exponents are dual) we have
G̃ru dx = 0 and Arur dx = 0, hence
Z Z Z Z
q ⇤
|ru| dx = Gru dx = A r ru dx = Arur dx = 0.
⌦ ⌦ ⌦ ⌦

Remark 15.15 (General Helmholtz decomposition). Thanks to Theorem 15.14, the

Helmholtz decomposition showed above is possible for every p 2 (1, 1).
Remark 15.16 (W 2,p estimates). By di↵erentiating the equation and multiplying by
cut-o↵ functions, we easily see that Theorem 15.12 and Theorem 15.14 yield
div(Aru) = f, |ru| 2 Lploc , f 2 Lploc =) 2,p
u 2 Wloc .
Remark 15.17 (No L1 bound is possible). As it was claimed above, let us show here that
T does not map L1 into L1 , with ⌦ = B1 ⇢ Rn . First we prove that this phenomenon
occurs if T is known to be discontinuous, then we prove that T is indeed discontinuous.
To check the first claim, let (⌦k ) be a countable family of pairwise disjoint closed balls
contained in ⌦: by a scaling argument we can find (since also the rescaled operators of
T on ⌦i are discontinuous) functions Fk 2 L1 (⌦i ; Rm⇥n ) with kFk k1 = 1 and solutions
uk 2 H01 (⌦k ; Rm ) to the equation (15.5) with kruk k1 k. Then it is easily shown (for
instance by approximation with finite families of balls) that the function
(
uk (x) if x 2 ⌦k
u(x) :=
0 if x 2 ⌦ \ [k ⌦k

belongs to H01 (⌦; Rm ), solves the equation with datum

(
Fk (x) if x 2 ⌦k
F (x) :=
0 if x 2 ⌦ \ [k ⌦k ,

81
but its gradient is patently not bounded.
So, it remains to prove that T is necessarily discontinuous, which we will do restricting
our discussion to the scalar case for the sake of simplicity. By the same duality argument
used before, if T were continuous we would get an estimate of the form

krukL1  ckF kL1

whenever u 2 H01 (⌦; Rm ) solves equation (15.5) for m = 1.

Hence, a standard approximation argument (based on convolution of the right hand
side, and Rellich compactness theorem) would imply the existence, for any vector-valued
measure µ in ⌦, of solutions of bounded variation, i.e., functions u 2 L1 (⌦; R), whose
weak gradient Du = (D1 u, . . . , Dn u) is a vector-valued measure satisfying
XZ XZ
↵
A @x↵ dD u = @x↵ dµ↵ 8 2 Cc1 (⌦; R). (15.10)
↵, ⌦ ↵ ⌦

and
|Du|(⌦)  c|µ|(⌦), (15.11)
where |µ| (resp. |Du|) denote the total variation of the measure µ (resp. Du). On the
other hand, we claim that the inequality (15.11) can’t be true. In fact, when n = 2 and
m = 1, consider the identity matrix A↵ := ↵ and the corresponding Laplace equation

v= 0 , (15.12)

where 0 is the Dirac measure supported in 0. The well-known fundamental solution of

(15.12) is
log |x|
v(x) = 2 C 1 (R2 \ {0}) ,
2⇡
1,p
so that v 2 Wloc (R2 ) for any p < 2, with rv(x) = (2⇡) 1 x/|x|2 , and (understanding
the second derivative in the pointwise sense) |r2 v| 2
/ L1 (⌦), since
✓ ◆
2 1 x⌦x
r v(x) = I 2 .
2⇡|x|2 |x|2

For any ⌘ 2 Cc1 (⌦) with ⌘ ⌘ 1 on B1/2 we have

(@x↵ (v⌘)) = @ x↵ ( 0 + v ⌘ + 2hrv, r⌘i)

so if we introduce the vector measure µ whose components are defined by

µ1 = 0 + v ⌘ + 2hrv, r⌘i, µ2 = 0,

82
we have that the function w = @x1 (⌘v) 2 L1 (R2 ) is a distributional solution in R2 to the
equation X
w= @ x↵ µ ↵ .
↵

It follows that ũ = w u, with u as in (15.10) is a distributional solution to Laplace

equation in B1 , and therefore standard properties of harmonic functions (for instance the
mean value property and a convolution argument applied to ũ) imply that ũ is equivalent
in B1 to a smooth function. By the properties of u and ũ, it follows that the distributional
derivative of w = ũ u is locally representable in ⌦ by a measure with finite total variation.
By our choice of ⌘, this implies the same for @x1 v in B1/2 , and a similar argument gives
the same property for @x2 v. Since |r2 v| is not summable in B1/2 , we have reached a
contradiction.

Now we move from constant to continuous coefficients, using Korn’s technique.

Theorem 15.18. In an open set ⌦ ⇢ Rn let u 2 Hloc

1
(⌦; Rm ) be a solution to the PDE

div(Aru) = f + divF
2 2
with coefficients A 2 C(⌦; Rn m ) which satisfy a uniform Legendre-Hadamard condition
for some > 0. Moreover, if p 2 (1, 1), let us suppose that F 2 Lploc and f 2 Lqloc , where
the Sobolev conjugate exponent q ⇤ = qn/(n q) coincides with p. Then |ru| 2 Lploc (⌦).
Proof. We give the proof for p 2 (the other cases come again by duality). Let us fix
s 2 and let us show that
⇤
|ru| 2 Ls^p
loc (⌦) =) |ru| 2 Lsloc^p (⌦) . (15.13)

Proving (15.13) ends the proof because |ru| 2 L2loc (⌦) (case s = 2) and in finitely many
steps s⇤ becomes larger than p.
Fix a point x0 2 ⌦ and a radius R > 0 such that BR (x0 ) b ⌦: we choose a cut-o↵
function ⌘ 2 Cc1 (BR (x0 )), with 0  ⌘  1 and ⌘ ⌘ 1 in BR/2 (x0 ).
⇤
We claim that ⌘u belongs
⇤
to H01,s ^p (BR (x0 )) if R ⌧ 1, as it is the unique fixed point of a
contraction in H01,s ^p (BR (x0 )), that we are going to define and study in some steps. This
⇤
implies, in particular, that |ru| 2 Ls ^p (BR/2 (x0 )).
(1) We start localizing the equation. Replacing ' with ⌘' in the PDE, by algebraic

83
computations we obtain
Z
A(x)r(⌘u)(x)r'(x) dx
BR (x0 )
Z
= A(x) (⌘(x)ru(x) + u(x) ⌦ r⌘(x)) r'(x) dx
BR (x0 )
Z
= A(x) (ru(x)r(⌘')(x) + u(x) ⌦ r⌘(x)r'(x) ru(x) (r⌘(x)'(x))) dx
BR (x0 )
Z
= f (x)⌘(x)'(x)+F (x)r(⌘')(x)+A(x) (u(x)⌦r⌘(x)r'(x) ru(x)r⌘(x)'(x)) dx
BR (x0 )
Z
= f˜(x)'(x) + F̃ (x)r'(x) dx ,
BR (x0 )

defining
f˜(x) := f (x)⌘(x) + F (x)r⌘(x) A(x)ru(x)r⌘(x)
and
F̃ (x) := F (x)⌘(x) + A(x)u(x) ⌦ r⌘(x) .
Thus ⌘u satisfies

div(A(x0 )r(⌘u)) = f˜ + div[F̃ + (A A(x0 ))r(⌘u)]. (15.14)

(2) In order to write f˜ in divergence form, let us consider the problem

⇢
w = f˜
.
w 2 H01 (⌦; Rm )

Thanks to the previous Lp regularity result for constant coefficients PDEs, since f˜ 2 Ls^q
loc
(because we assumed that |ru| 2 Ls^p 2 s^q
loc ), we have |r w| 2 Lloc (see also Remark 15.16).
(s^q)⇤
By Sobolev immersion we get |rw| 2 Lloc , hence
⇤ ⇤ ⇤
|rw| 2 Lsloc^q = Lsloc^p .

Now we define ⇤
F ⇤ (x) := F̃ (x) + rw(x) 2 Lsloc^p .
⇤
(3) Let E = H01,s ^p (BR (x0 ); Rm ) and let us define the operator ⇥ : E ! E which
associates to each V 2 E the function v 2 E that solves

div (A(x0 )rv) = divF ⇤ div ((A(x0 ) A)rV ) . (15.15)

84
 ⇤
The operator ⇥ is well-defined because |F ⇤ | 2 Ls ^p (BR (x0 )) (we saw this in step (2))
and we can take advantage of regularity theory for constant coefficients operators. The
operator ⇥ is a contraction, in fact
1
kr(v1 v2 )kE  ck (A(x0 ) A) r(V1 V2 )kE  kr(V1 V2 )kLs⇤ ^p (BR (x0 ))
2
if R is sufficiently small, according to the continuity of A. Here we use the fact that the
constant c in the first inequality is scale invariant, so it can be “beaten” by the oscillation
of A in BR (x0 ), if R is small enough.
Let us call v⇤ 2 E the unique fixed point of (15.15). According to (15.14), ⌘u already
1,s^p 1,s⇤ ^p
solves (15.15), but in the larger space H0 . Thus ⌘u 2 H0 if we are able to show
that v⇤ = ⌘u, and to see this it suffices to show that uniqueness holds in the larger space
as well.
Consider the di↵erence v 0 := v⇤ ⌘u 2 H01,s^p (BR (x0 ); Rm ) ⇢ H01 (BR (x0 ); Rm ): v 0 is a
weak solution of
div (A(x)rv 0 ) = 0 ,
hence v 0 ⌘ 0 (we can indeed use the variational characterization of the solution). This
concludes the proof. ⇤

16 De Giorgi’s solution of Hilbert’s XIX problem

16.1 The basic estimates
We briefly recall here the setting of Hilbert’s XIX problem, that has already been described
and solved in dimension 2.
We deal with local minimizers v of scalar functionals
Z
v7 ! F (rv) dx
⌦

where F 2 C 2, (Rn ) (at least, for some > 0) satisfies the following ellipticity property:
there exist two positive constants  ⇤ such that ⇤I r2 F (p) I for all p 2 Rn (this
implies in particular that |r2 F | is uniformly bounded). We have already seen that under
these assumptions it is possible to derive the Euler-Lagrange equations divFp (rv) = 0.
By di↵erentiation, for any direction s 2 {1, . . . , n}, the equation for u := @v/@xs is
✓ ◆
@ @
Fp↵ p (rv) u =0.
@x↵ @x

85
Recall also the fact that, in order to obtain this equation, we needed to work with the
approximation h,s v and with the interpolating operator
Z 1
Aeh (x) := Fpp (trv(x + hes ) + (1 t)rv(x)) dt
0
and to exploit the Caccioppoli-Leray inequality.
One of the striking ideas of De Giorgi was basically to split the problem, that is to
deal with u and v separately, as rv is only involved in the coefficients of the equation for
u. The key point of the regularization procedure is then to show that under no regularity
assumption on rv (i.e. not more than measurability), if u is a solution of this equation,
0,↵
then u 2 Cloc (⌦), with ↵ depending only on n and on the ellipticity constants , ⇤. If
this is true, we can proceed as follows:
u 2 C 0,↵ ) v 2 C 1,↵ ) Fpp (rv) 2 C 0,↵ ) u 2 C 1,↵ ,
where the implications rely upon the fact that Fpp is Hölder continuous and on the
Schauder estimates of Theorem 10.4. Since u is any partial derivative of v, we eventually
get v 2 C 2,↵ . If F is more regular, by continuing this iteration (now using Shauder
regularity for PDE’s whose coefficients are C 1, , C 2, and so on) we obtain
F 2 C1 ) v 2 C1
and also, by the tools developed in [20], that F 2 C ! ) v 2 C ! , which is the complete
solution of the problem raised by Hilbert.
Actually, we have solved this problem in the special case n = 2, since, by means of
Widman’s technique, we could prove that |ru| 2 L2,↵ and hence u 2 L2,↵+2 for some
↵ > 0. This is enough, if n = 2, to conclude that u 2 C 0,↵/2 .
First of all, let us fix our setting. Let ⌦ be an open domain in Rn , 0 <  ⇤ < 1
and let A↵ be a Borel symmetric matrix satisfying a.e. the condition I  A(x)  ⇤I.
1
We want to show that if u 2 Hloc solves the problem
div (A(x)ru(x)) = 0
0,↵
then u 2 Cloc . Some notation is needed: for B⇢ (x) ⇢ ⌦ we define
A(k, ⇢) := {u > k} \ B⇢ (x) ,
where the dependence on the center x can be omitted. This should not create confusion,
since we will often work with a fixed center. In this section, we will derive many func-
tional inequalities, but typically we are not interested in finding the sharpest constants,
but only on the functional dependence of these quantities. Therefore, in order to avoid
complications of the notation we will use the same symbol (generally c) to indicate di↵er-
ent constants, possibly varying from one passage to the next one. However we will try to
indicate the functional dependence explicitly whenever this is appropriate and so we will
use expressions like c(n) or c(n, , ⇤) many times.

86
Theorem 16.1 (Caccioppoli inequality on level sets). For any k 2 R and B⇢ (x) ⇢
BR (x) b ⌦ we have
Z Z
2 c
|ru| dy  (u k)2 dy (16.1)
A(k,⇢) (R ⇢)2 A(k,R)
with c = 16⇤2 / 2 .
Remark 16.2. It should be noted that the previous theorem generalizes the Caccioppoli-
Leray inequality, since we do not ask ⇢ = R/2 and we introduce the sublevels.
1,1
Theorem 16.3 (Chain rule). If u 2 Wloc (⌦), then for any k 2 R the function (u k)+
1,1
belongs to Wloc (⌦). Moreover we have that r(u k)+ = ru a.e. on {u > k} , while
r(u k)+ = 0 a.e. on {u  k} .
Proof. Since this theorem is rather classical, we just sketch the proof. By the arbitrari-
ness of u, the problem is clearly translation-invariant and we can assume
p without loss of
generality k = 0. Consider the family of functions defined by '" (t) := t + "2 " for t 0
2

and identically zero elsewhere, whose derivatives are uniformly bounded and converge to
1 1,1
{t>0} . Moreover, let (un ) be a sequence of Cloc functions approximating u in Wloc . We
have that for any n 2 N and " > 0 the classical chain rule gives r ['" (un )] = '0" (un )run .
Passing to the limit as n ! 1 gives r ['" (u)] = '0" (u)ru. Now, we can pass to the limit
as " # 0 and use the dominated convergence theorem to conclude that ru+ = {u>0} ru.
⇤
We can come to the proof of the Caccioppoli inequality on level sets.
Proof. Let ⌘ be a cut-o↵ function supported in BR (x), with ⌘ ⌘ 1 on B ⇢ (x) and |r⌘| 
2/(R ⇢). If we apply the weak form of our equation to the test function ' := ⌘ 2 (u k)+
we get
Z Z
2
⌘ Aruru dy = 2 ⌘Arur⌘(u k)+ dy
A(k,R) B (x)
Z R Z
⇤ 2 2 4"⇤
 ⌘ |Du| dy + (u k)2 dy
" A(k,R) (R ⇢)2 A(k,R)
for any " > 0, by our upper bound and by Young inequality. Here we set " = 2⇤/ so
that, thanks to the uniform ellipticity assumption, we obtain
Z Z Z
8⇤2
2
⌘ Aruru dy  2 2
⌘ |ru| dy + 2
(u k)2 dy .
A(k,R) 2 A(k,R) (R ⇢) A(k,R)
Since on the smaller ball ⌘ is identically equal to 1, we eventually get
Z Z
16⇤2
2
|ru| dy  2 2
(u k)2 dy ,
A(k,⇢) (R ⇢) A(k,R)

which is our thesis. ⇤

87
 The second great idea of De Giorgi was that (one-sided) regularity could be achieved
for all functions satisfying the previous functional inequality, regardless of the fact that
these were solutions to an elliptic equation. For this reason he introduced a special class
of objects.
Definition 16.4 (De Giorgi’s class). We define the De Giorgi class DG+ (⌦) as follows:
DG+ (⌦) := {u | 9 c 2 R s.t. 8 k 2 R, Br (x) b BR (x) b ⌦, u satisfies (16.1) } .
In this case, we also define c+
DG (u) to be the minimal constant larger than 1 for which
the condition (16.1) is verified.
Remark 16.5. From the previous proof, it should be clear that we do not really require
u to be a solution, but just a sub-solution of our problem. In fact, we have proved that
16⇤2
div (Aru)  0 in D0 (⌦) =) u 2 DG(⌦), c+
DG (u)  2
.

In a similar way, the class DG (⌦) (corresponding to supersolutions) and cDG (u) could
be defined by
Z Z
2 c
|ru| dy  2
(u k)2 dy
{u<k}\B⇢ (x) (R ⇢) {u<k}\BR (x)

and obviously u 7! u maps DG+ (⌦) in DG (⌦) bijectively, with c+

DG (u) = cDG ( u).

The main part of the program by De Giorgi can be divided into two steps:
(i) If u 2 DG+ (⌦), then it satisfies a strong maximum principle in a quantitative form
(more precisely the L2 to L1 estimate in Theorem 16.8);
0,↵
(ii) If both u and u belong to DG+ (⌦), then u 2 Cloc (⌦).
Let us start by discussing the first point. We define these two crucial quantities:
Z
U (h, ⇢) := (u h)2 dy, V (h, ⇢) := L n (A(h, ⇢)) .
A(h,⇢)

Theorem 16.6. The following properties hold:

(i) both U and V are non-decreasing functions of ⇢, and non-increasing functions of h;
(ii) for any h > k and 0 < ⇢ < R the following inequalities hold:
1
V (h, ⇢)  U (k, ⇢),
(h k)2
c(n) · c+
DG (u)
U (k, ⇢)  U (k, R)V 2/n (k, ⇢) .
(R ⇢)2

88
Proof. The first statement the first inequality in the second statement are trivial, since
Z Z
2 2
(h k) V (h, ⇢) = (h k) dy  (u k)2 dy
A(h,⇢) A(h,⇢)
Z
 (u k)2 dy = U (k, ⇢) .
A(k,⇢)

For the second inequality, let us introduce a Lipschitz cut-o↵ function ⌘ supported in
B(R+⇢)/2 (x) with ⌘ ⌘ 1 on B ⇢ (x) and |r⌘|  4/(R ⇢). We need to note that
Z Z
4c+
DG (u)
⌘ 2 |r(u k)+ |2 dy  (u k)2 dy
B(R+⇢)/2 (R ⇢)2 A(k,R)

and Z Z
+ 2 2 16
((u k) ) |r⌘| dy  (u k)2 dy .
B(R+⇢)/2 (R ⇢)2 A(k,R)

Combining these two inequalities, since c+

DG (u) 1, we get
Z Z
+ 2 40c+
DG (u)
|r(⌘(u k) )| dy  (u k)2 dy
B(R+⇢)/2 (R ⇢)2 A(k,R)

and by the Sobolev embedding inequality with the function ⌘(u k)+ this implies
✓Z ◆2/2⇤ Z
2⇤ c(n) · c+
DG (u)
(u k) dy  (u k)2 dy
A(k,⇢) (R ⇢)2 A(k,R)

for some constant c(n) depending on the dimension n. In order to conclude, we just need
to apply Hölder’s inequality, in fact
Z ✓Z ◆2/2⇤
2 2⇤
U (k, ⇢) = (u k) dy  (u k) dy V (k, ⇢)2/n
A(k,⇢) A(k,⇢)

with p = 2⇤ /2 = n/(n 2), p0 = n/2. ⇤

The previous inequalities can be slightly weakened, writing
1
V (h, ⇢)  U (k, R),
(h k)2
c(n) · c+
DG (u)
U (h, ⇢)  U (k, R)V 2/n (k, R)
(R ⇢)2
and we shall use these to obtain the quantitative maximum principle.

89
 We can view these inequalities as joint decay properties of U and V ; in order to get
the decay of a single quantity, it is convenient to define ' := U ⇠ V ⌘ for some choice of the
(positive) real parameters ⇠, ⌘ to be determined. We obtain:

C⇠ 1
U ⇠ (h, ⇢)V ⌘ (h, ⇢)  U ⇠+⌘ (k, R)V 2⇠/n (k, R).
(h k) (R ⇢)2⇠
2⌘

where C := c(n) · c+DG (u), a convention that will be systematically adopted in the sequel.
Since we are looking for some decay inequality for ', we look for solutions (✓, ⇠, ⌘) to the
system
2⇠
⇠ + ⌘ = ✓⇠, = ✓⌘ .
n
Setting ⌘ = 1 (by homogeneity this choice is not restrictive), we get ⇠ = n✓/2 and we can
use the first equation to get r
1 1 2
✓= + + . (16.2)
2 4 n
Note that ✓ > 1 : this fact will play a crucial role in the following proof. In any case, we
get the decay relation

C⇠ 1
'(h, ⇢)  '✓ (k, R) .
(h k) (R ⇢)2⇠
2⌘

Theorem 16.7. Let u 2 DG+ (⌦), BR0 (x) b ⌦. For any h0 2 R there exists d =
d(h0 , R0 , c+
DG (u)) such that '(h0 + d, R0 /2) = 0. Moreover, we can take

'(h0 , R0 )✓ 1
d2 = c0 (n)[c+
DG (u)]
n✓/2
,
R0n✓

with the constant c0 (n) depending only on the dimension n. In particular u  h0 + d

L n -a.e. on BR0 /2 (x).

Corollary 16.8 (L2 to L1 estimate). If u 2 DG+ (⌦), then for any BR0 (x) ⇢ ⌦ and for
any h0 2 R
✓ Z ◆1/2 ✓ ◆(✓ 1)/2
00 1 V (h0 , R0 )
ess sup u  h0 + c (n)[c+
DG (u)]
n✓/4
(u 2
h0 ) dy .
BR0 /2 (x) !n R0n A(h0 ,R0 ) R0n
Proof. This corollary comes immediately from Theorem 16.7, once we express ' in terms
of U and V and recall that ⇠ + 1 = ✓⇠ (that is ⇠(✓ 1) = 1), by means of simple algebraic
computations. ⇤

90
Remark 16.9. From Corollary 16.8 with h0 = 0, we can get the maximum principle for
u, as anticipated above. In fact
Z
+ 2 + n✓/2
ess sup (u )  q(n)[cDG (u)] u2 dy
BR0 /2 (x) BR0 (x)

with q(n) easily estimated in terms of c00 (n) and !n .

We are now ready to prove Theorem 16.7, the main result of this section.
Proof. Define kp := h0 + d d/2p and Rp := R0 /2 + R0 /2p+1 , so that kp " (h0 + d) while
Rp # R0 /2. Here d 2 R is a parameter to be fixed in the sequel. From the decay inequality
for ' we get
" ✓ p+2 ◆2⇠ ✓ p+1 ◆2 #
2 2
'(kp+1 , Rp+1 )  '(kp , Rp ) '(kp , Rp )✓ 1 C ⇠
R0 d

and letting p := 2µp '(kp , Rp ) this becomes

h i
p+1  µ ⇠ 4⇠+2 p(2⇠+2)
p 2 C 2 2 R0 2⇠ d 2 2 µp(✓ 1) ✓ 1
p .

This is true for any µ 2 R but we fix it so that (2⇠ +2) = µ(✓ 1), leading to a cancellation
of two factors in the previous inequality. Having chosen µ, if we choose d as follows
2⇠
2µ C ⇠ 24⇠+2 ✓ 1
0 R0 d
2
=1
then 1  0 . Hence, 2µ C ⇠ 24⇠+2 1✓ 1 R0 2⇠ d 2  1 and the decay inequality yields 2  1 .
By induction, it follows that p  0 , 8p 2 N. In that case, '(kp , Rp )  2 µp '(h0 , R0 ) !
0 and, since by monotonicity
'(h0 + d, R0 /2)  '(kp , R0 /2)  '(kp , Rp ) ,
we get the thesis. But the previous condition on d is satisfied if
d2 c0 (n)[c+
DG (u)]
n✓/2
R0 2⇠ ✓ 1
0

and the desired claim follows. ⇤

We are now in position to discuss the notion of oscillation, which will be crucial for
the conclusion of the argument by De Giorgi.
Definition 16.10. Let ⌦ ⇢ Rn be an open set, Br (x) ⇢ ⌦ and u : ⌦ ! R a measurable
function. We define its oscillation on Br (x) as
!(Br (x))(u) := ess sup u ess inf u.
Br (x) Br (x)

When no confusion arises, we will omit the explicit dependence on the center of the ball,
thus identifying !(r) = !(Br (x)).

91
 It is an immediate consequence of the previous results that if u 2 DG+ (⌦) \ DG (⌦),
then ✓Z ◆ ✓Z ◆
1/2 1/2
2 2
ess sup u  ⇣ u dy , ess inf u  ⇣ u dy
Br/2 (x) Br (x) Br/2 (x) Br (x)

for a constant ⇣, which is a function of the dimension n and of cDG (u). Here and in the
sequel we shall denote by cDG (u) the maximum of c+ DG (u) and cDG (u) and by DG(u) the
intersection of the spaces DG+ (⌦) and DG (⌦).
Consequently, under the same assumptions,
✓Z ◆1/2
2
!(Br/2 (x))(u)  2⇣ u dy .
Br (x)

Let us see the relation between the decay of the oscillation of u and the Hölder regu-
larity of u. We prove this result passing through the theory of Campanato spaces (a more
elementary proof is based on the observation that the Lebesgue representative defined at
approximate continuity points is Hölder continuous).
Theorem 16.11. Let ⌦ ⇢ Rn be open, c 0, ↵ 2 (0, 1] and let u : ⌦ ! R be a measurable
0,↵
function such that for any Br (x) ⇢ ⌦ we have !(Br (x))  cr↵ . Then u 2 Cloc (⌦), that
0,↵
is, there exists in the Lebesgue equivalence class of u a Cloc representative.
Proof. By definition of essential extrema, for L n -a.e. y 2 Br (x) we have that
ess inf Br (x) u  u(y)  ess supBr (x) u. These inequalities imply ess inf Br (x) u  uBr (x) 
ess supBr (x) and hence that L n -a.e. in Br (x) the inequality |u uBr (x) |  cr↵ holds. We
0,↵
have proved that u 2 L2,n+2↵ (⌦), but this gives u 2 Cloc (⌦) (regularity is local since no
assumption is made on ⌦), which is the thesis. ⇤
This theorem motivates our interest in the study of oscillation of u, that will be carried
on by means of some tools we have not introduced so far.

16.2 Some useful tools

De Giorgi’s proof of Hölder continuity is geometric in spirit and ultimately based on the
isoperimetric inequality. Notice that, as we will see, the isoperimetric inequality is also
underlying the Sobolev inequalities, which we used in the proof of the sup estimate for
functions in DG+ (⌦).
We will say that a set E ⇢ Rn is regular if it is locally the epigraph of a C 1 function.
In this case, it is well-known that by local parametrizations and a partition of unity, we
can define n 1 (@E), the (n 1)-dimensional surface measure of @E.
Of course, regular sets are a very unnatural (somehow too restrictive) setting for
isoperimetric inequalities, but it is sufficient for our purposes. We state without proof
two isoperimetric inequalities:

92
Theorem 16.12 (Isoperimetric inequality). Let E ⇢ Rn be a regular set such that
n 1 (@E) < 1. Then
1⇤
min {L n (E), L n (Rn \ E)}  c(n) [ n 1 (@E)]

with c(n) a dimensional constant.

It is also well-known that the best constant c(n) in the previous inequality is
1⇤ ⇤
L n (B1 )/[ n 1 (@B1 )] = !n /[n!n ]1 ,
that is, balls have the best isoperimetric ratio.
Theorem 16.13 (Relative isoperimetric inequality). Let ⌦ ⇢ Rn be an open and bounded
set, with @⌦ Lipschitz. Let E ⇢ ⌦ with ⌦ \ @E 2 C 1 . Then
⇤
min {L n (E), L n (⌦ \ E)}  c(⌦) [ n 1 (⌦ \ @E)]1 .
Let us introduce another classical tool in Geometric Measure Theory.
Theorem 16.14 (Coarea formula). Let ⌦ ⇢ Rn be open and u 2 C 1 (⌦) be non-negative,
then Z Z 1
|ru| dx = n 1 (⌦ \ {u = t}) dt .
⌦ 0
Remark 16.15. It should be observed that the right-hand side of the previous formula
is well-defined, since by the classical Sard’s theorem
u 2 C 1 (⌦) =) L 1 {u(x) : x 2 ⌦, ru(x) = 0} = 0 .
By the implicit function theorem this implies that almost every sublevel set {u < t} is
regular.
Proof. A complete proof will not be described here since it is far from the main purpose
of these lectures, however we sketch the main points. The interested reader may consult,
for instance, [12]. R R1
We first prove ⌦ |ru| dx  0 n 1 (⌦ \ {u = t}) dt. Consider the pointwise identity
Z 1
u(x) = {u>t} (x) dt
0
that implies
Z Z Z
|ru| dx = sup hru, 'i dx = sup u div' dx
⌦ '2Cc1 , |'|1 ⌦ '2Cc1 , |'|1 ⌦
Z 1 ✓Z ◆
= sup (div') {u>t} dx dt
'2Cc1 , |'|1 0 ⌦
Z Z !
1
 sup div' dx dt.
0 '2Cc1 , |'|1 {u>t}

93
Hence, by the Gauss-Green theorem (with ⌫t outer normal to {u > t}) we obtain
Z Z 1 Z ! Z 1
|ru| dx  sup h', ⌫t i d n 1 dt  n 1 (⌦ \ {u = t}) dt ,
⌦ 0 '2Cc1 , |'|1 ⌦\{u=t} 0

again exploiting the fact that for a.e. t the set {u = t} is the (regular) boundary of
{u > t}.
Let us consider the converse inequality, namely
Z Z 1
|ru| dx n 1 (⌦ \ {u = t}) dt .
⌦ 0

It is not restrictive to assume that ⌦ is a cube. This is trivial (with equality) if u

is continuous and piecewise linear, since on each part of a triangulation of ⌦ the coarea
formula is just Fubini’s Theorem. The general case is obtained by approximation, choosing
piecewise affine functions which converge to u in W 1,1 (⌦) and using Fatou’s lemma and
the lower semicontinuity of E 7! n 1 (⌦ \ @E) (this, in turn, follows by the sup formula
we already used in the proof of the first inequality). We omit the details. ⇤
In order to deduce the desired Sobolev embeddings, we need a technical lemma.

Lemma 16.16. Let G : [0, 1) ! [0, 1) a non-increasing measurable function. Then for
any ↵ 1 we have Z 1 ✓Z 1 ◆↵
↵ 1 1/↵
↵ t G(t) dt  G (t) dt .
0 0
Proof. It is sufficient to prove that for any T > 0 we have the finite time inequality
Z T ✓Z T ◆↵
↵ 1 1/↵
↵ t G(t) dt  G (t) dt . (16.3)
0 0

Since G is non-increasing, we can observe that

Z t
1/↵
G (t)  G1/↵ (s) ds ,
0

which is equivalent to
✓Z t ◆↵ 1
↵ 1 1/↵
t G(t)  G (s) ds G1/↵ (t) .
0

Then, multiplying both sides by ↵, (16.3) follows by integration. ⇤

94
 We are now ready to derive the Sobolev inequalities stated in Theorem 4.6.

Theorem 16.17 (Sobolev embedding, p = 1). For any u 2 W 1,1 (Rn ) we have that
✓Z ◆1/1⇤ Z
1⇤
|u| dx  c(n) |ru| dx .
Rn Rn

Consequently, we have the following continuous embeddings:

⇤
(1) W 1,1 (Rn ) ,! L1 (Rn );
⇤
(2) for any ⌦ ⇢ Rn open, regular and bounded W 1,1 (⌦) ,! L1 (⌦).
Proof. By Theorem 16.3 it is possible to reduce the thesis to the case u 0, and
1
smoothing reduces the proof to the case u 2 C . Under these assumptions we have
Z Z 1 ✓Z 1 ◆1⇤
1⇤ ⇤ 1/(n 1) n n 1/1⇤
u dx = 1 t L ({u > t}) dt  L ({u > t}) dt
Rn 0 0

thanks to Lemma 16.16. Consequently, the isoperimetric inequality and the coarea for-
mula give
Z ✓Z 1 ◆1 ⇤ ✓Z ◆1 ⇤
1⇤
u dx  c(n) n 1 {u = t} dt = c(n) |ru| dx .
Rn 0 Rn

The continuous embedding in (2) follows by the global one in (1) applied to an extension
of u (recall that regularity of ⌦ yields the existence of a continuous extension operator
from W 1,1 (⌦) to W 1,1 (Rn )). ⇤

Theorem 16.18 (Sobolev embeddings, 1 < p < n). For any u 2 W 1,p (Rn ) we have that
✓Z ◆1/p⇤ ✓Z ◆1/p
p⇤ p
|u| dx  c(n, p) |ru| dx .
Rn Rn

Consequently, the have the following continuous embeddings:

⇤
(1) W 1,p (Rn ) ,! Lp (Rn );
⇤
(2) for any ⌦ ⇢ Rn open, regular and bounded W 1,p (⌦) ,! Lp (⌦).
Proof. Again, it is enough to study the case u 0. We can exploit the case p = 1 to get
✓Z ◆1/1⇤ Z
↵1⇤
u dx  c(n) ↵u↵ 1 |ru| dx 8↵ > 1
Rn Rn

95
and, by Hölder’s inequality, the right hand side can be estimated from above with
Z 1/p0 Z 1/p
(↵ 1)p0
c(n)↵ u dx |ru|p dx .
Rn Rn

Now, choose ↵ such that ↵1⇤ = (↵ 1)p0 . Consequently

✓Z ◆1/1⇤ 1/p0 ✓Z ◆1/p
↵1⇤ p
u dx  c(n, p) |ru| dx ,
Rn Rn

but 1/1⇤ 1/p0 = 1/p⇤ , ↵1⇤ = p⇤ and the claim follows. The second part of the statement
can be obtained as in Theorem 16.17. ⇤
We will also make use of the following refinement of the Poincaré inequality in W01,1 :
even though no assumption is made on the behaviour of u at the boundary of the domain,
⇤
it is still possible to control the L1 norm with the gradient.
Theorem 16.19. Let u 2 W 1,1 (BR ) with u 0 and suppose that L n ({u = 0})
n
L (BR )/2. Then
✓Z ◆1/1⇤ Z
1⇤
u dx  c(n) |ru| dx .
BR BR
⇤
Proof. This result is the local version of the embedding W 1,1 ,! L1 . Hence, in order
to give the proof, it is just needed to mimic the previous argument substituting the
isoperimetric inequality with the relative isoperimetric inequality, that is, here
⇤
L n (BR \ {u > t})  c(n) n 1 [L n (BR \ {u = t})]1 .

We leave the details to the reader. ⇤

16.3 Proof of Hölder continuity

We divide the final part of the proof in two parts.
Lemma 16.20 (Decay of V ). Let ⌦ ⇢ Rn be open and let u 2 DG+ (⌦). Suppose that
B2r b ⌦ and k0 < ess supB2r (u)  M satisfies
1
V (k0 , r)  L n (Br ) , (16.4)
2
then the sequence of levels k⌫ = M (M k0 )/2⌫ for ⌫ 0 satisfies
✓ ◆2(n 1)/n
V (k⌫ , r) c(n) c+
DG (u)
 .
rn ⌫

96
Proof. Take two levels h, k such that M h k k0 and define u := u ^ h u ^ k =
(u ^ h k)+ . By construction u 0 and since u 2 W 1,1 (⌦) we also have u 2 W 1,1 (⌦). It
is also clear that ru 6= 0 only on A(k, r) \ A(h, r). Notice that
1 n
L n ({u = 0} \ Br ) L n ({u  k} \ Br ) L n ({u  k0 } \ Br ) L (Br )
2
and so we can apply the relative version of the critical Sobolev embedding and Hölder’s
inequality to get
Z ✓Z ◆1⇤
1⇤ n 1⇤
(h k) L (A(h, r)) = u dy  c(n) |ru| dy
A(h,r) Br
Z
= c(n) |ru| dy
A(k,r)\A(h,r)
✓Z ◆1⇤ /2
2 ⇤ /2
 c(n) |ru| dy L n (A(k, r) \ A(h, r))1 .
A(k,r)

We can now exploit the De Giorgi property of u that is

Z Z
2 c+
DG (u)
|ru| dy  2
(u k)2 dy  (M k)2 !n c+
DG (u)r
n 2
A(k,r) r B2r

in order to obtain
⇤
(h k)2 L n (A(h, r))2/1  c(n)c+
DG (u)(M k)2 rn 2 (V (k, r) V (h, r)) . (16.5)

Here we can conclude the proof by applying (16.5) for h = ki+1 and k = ki , so that
⌫
X
2/1⇤ ⇤
⌫V (k⌫ , r)  V (ki , r)2/1
i=1
⌫
X
 4c(n)c+
DG (u)r
n 2
[V (ki , r) V (ki+1 , r)]
i=1
 4c(n)c+
DG (u)!n r
2n 2
.

Theorem 16.21 (C 0,↵ regularity). Let ⌦ ⇢ Rn be open and let u 2 DG(⌦). Then
0,↵
u 2 Cloc (⌦), with 2↵ = log2 1 2 (⌫+2) ,

⌫ = 2c(n) [cDG (u)](n✓ 1)/(✓ 1)

(16.6)

and ✓ > 1 given by (16.2), solution to the equation n✓(✓ 1) = 2.

97
Proof. Pick an R > 0 such that B2R (x) b ⌦ and consider for any r  R the functions
m(r) := ess inf Br (x) (u) and M (r) := ess supBr (x) (u). Moreover, set !(r) = M (r) m(r)
and µ(r) := (m(r) + M (r)) /2. We apply the previous lemma to the sequence k⌫ :=
M (2r) !(2r)
2⌫+1
, but to do this we should check the hypothesis (16.4), which means

1
L n ({u > µ(2r)} \ Br (x))  L n (Br (x)).
2
Anyway, either L n ({u > µ(2r)} \ Br (x))  12 L n (Br (x)) or L n ({u < µ(2r)} \ Br (x)) 
1
2
L n (Br (x)). The second case is analogous, provided we work with u instead of u, and
it is precisely here that we need the assumption that both u and u belong to DG+ (⌦).
Using Lemma 16.20 it is easily seen that the choice of ⌫ as in (16.6), with c(n) large
enough, provides
✓ ◆
00
⇥ + ⇤n✓/4 V (k⌫ , r) (✓ 1)/2 1
c (n) cDG (u)  ,
rn 2
where c00 (n) is the dimensional constant in Theorem 16.8. Moreover, this choice of ⌫ has
been made independently of of r and R (this is crucial for the validity of the scheme
below).
Now apply the maximum principle in Theorem 16.8 to u with radii r/2 and r and
h0 = M (2r) !(2r)2⌫+1
= k⌫ (for the previous choice of ⌫) to obtain
⇣r⌘ ✓ ◆(✓ 1)/2
00
⇥ ⇤n✓/4 V (h0 , r)
M  h0 + c (n) c+
DG (u) (M (2r) h0 )
2 rn

and, by the appropriate choice of ⌫ that has been described, we deduce

⇣r⌘ M (2r) h0 M (2r) + h0 1
M  h0 + = = M (2r) !(2r).
2 2 2 2⌫+2
If we subtract the essential minimum m(2r) and use m(r/2) m(2r) we finally get
⇣r⌘ ✓ ◆
1
!  !(2r) 1
2 2⌫+2

which is the desired decay estimate. By the standard iteration argument5 , we find
✓ ◆↵
↵ r
!(r)  4 !(R) 0<rR
R

for 2↵ = log2 1 2 (⌫+2)

and the conclusion follows from Theorem 16.11. ⇤

5
We refer to Lemma 9.1, with the obvious changes.

98
17 Regularity for systems
17.1 De Giorgi’s counterexample to regularity for systems
In the previous section we saw De Giorgi’s regularity result for solutions u 2 H 1 (⌦) of
the elliptic equation
div (A(x)ru(x)) = 0
0,↵
with bounded Borel coefficients A satisfying I  A  ⇤I. It turned out that u 2 Cloc (⌦),
with ↵ = ↵(n, , ⇤).
It is natural to investigate about similar regularity properties for systems, still under
no regularity assumption on A (otherwise, Schauder theory is applicable). In 1968, in [8],
Ennio De Giorgi provided a counterexample showing that the scalar case is special. De
Giorgi’s example not only solves an elliptic PDE, but it is also the minimum of a convex
variational problem.
When m = n, consider
u(x) := x|x|↵ . (17.1)
We will show in (17.7), (17.8) and (17.9) that, choosing
!
n 1
↵= 1 p , (17.2)
2 (2n 2)2 + 1

the function u is the solution of the Euler-Lagrange equation

P associated with the uniformly
convex functional (here r · u stands for the divergence i @xi ui )
Z ✓ ◆2
x⌦x
L(u) := (n 2)r · u(x) + n ru(x) + |ru(x)|2 dx . (17.3)
B1 |x|2

If n / L1 (B1 ), because
3 then |u| 2
! ✓ ◆
n 1 3 1
↵= 1 p 1 p >1
2 (2n 2)2 + 1 2 17

and this provides a counterexample not only to Hölder regularity, but also to local boun-
dedness of solutions. In the case n = 2 we already know from Widman’s technique (see
Remark 4.4) that u is locally Hölder continuous, nevertheless De Giorgi’s example will
show that this regularity cannot be improved Rto local Lipschitz.
Calling A(x) the matrix such that L(u) = B1 hA(x)ru, rui dx, we remark that A has
a discontinuity at the origin (determined by the term x ⌦ x/|x|2 ).

99
 The Euler-Lagrange equation associated to (17.3) is the following (in the weak distri-
butional sense): for every h = 1, . . . , n it must be
n n
!
@ X @ut X xs xt @ut
0 = (n 2) (n 2) +n (17.4)
@xh t=1
@xt s,t=1
|x|2 @xs
n
" n n
!#
X @ xh xk X @ut X xs xt @ut
+ n (n 2) +n (17.5)
k=1
@xk |x|2 t=1
@xt s,t=1
|x|2 @xs
n
X @ 2 uh
+ . (17.6)
k=1
@x2k

We are going to prove in a few steps that u is the unique minimizer of L, with boundary
data given by u itself, and that u solves the Euler-Lagrange equations. The steps are:
(i) u, as defined in (17.1), belongs to C 1 (B1 \ {0}; Rn ) and solves in B1 \ {0} the
Euler-Lagrange equations;
(ii) u 2 H 1 (B1 ; Rn ) and is also a weak solution in B1 of the system.
Let us perform step (i). Fix h 2 {1, . . . , n}, and use extensively the identity
@
|x|↵ = ↵xh |x|↵ 2
.
@xh
Then |x|↵ = (n↵ + ↵2 2↵)|x|↵ 2
and
@
(xh |x|↵ ) = xh |x|↵ + |x|↵ = (↵n + ↵2 )xh |x|↵ 2
(17.7)
@xh
and this is what we need to put in (17.6) when u is given by (17.1). For both (17.4) and
(17.5) we have to calculate
Xn
@
(xt |x|↵ ) = (n + ↵)|x|↵ ,
t=1
@x t

and n n
X xs xt @ut X xs xt
2
= 2
↵xs xt |x|↵ 2
+ st |x|
↵
= (↵ + 1)|x|↵ .
s,t=1
|x| @x s s,t=1
|x|
Therefore (17.4) is given by
n n
!
@ X @ut X xs xt @ut
(n 2) (n 2) +n = ↵(n 2)[(n 2)(n+↵)+n(↵+1)]xh |x|↵ 2
.
@xh t=1
@xt s,t=1
|x|2 @xs
(17.8)

100
In order to compute the term (17.5) we first get

Xn
@
xh xk |x|↵ 2
= (n + ↵ 1)xh |x|↵ 2
,
k=1
@x k

and therefore we obtain

n
" n n
!#
X @ xh xk X @ut X xs xt @ut
n 2
(n 2) +n
k=1
@x k |x| t=1
@x t s,t=1
|x|2 @xs
= n(n + ↵ 1)[(n 2)(n + ↵) + n(↵ + 1)]xh |x|↵ 2
. (17.9)

Putting together (17.7), (17.8) and (17.9), u(x) = x|x|↵ solves the Euler-Lagrange equa-
tion if and only if ⇣ n ⌘2
(2n 2)2 ↵ + + ↵n + ↵2 = 0 ,
2
which leads to the choice (17.2) of ↵.
Let us now perform step (ii), checking first that u 2 H 1 . As |ru(x)| ⇠ |x|↵ and
2↵ > n, it is easy to show that |ru| 2 L2 (B1 ). Moreover, for every ' 2 Cc1 (B1 \ {0})
we have classical integration by parts formula
Z Z
ru(x)'(x) dx = u(x)r'(x) dx . (17.10)

Thanks to Lemma 17.1 below, we are allowed to approximate in H 1 (B1 ; Rn ) norm every
' 2 Cc1 (B1 ) with a sequence ('k ) ⇢ Cc1 (B1 \ {0}). Then we can pass to the limit in
(17.10) because |ru| 2 L2 (B1 ) to obtain u 2 H 1 (B1 ; Rm ). Now, using the fact that the
Euler-Lagrange PDE holds in the weak sense in B1 \ {0} (because it holds in the classical
sense), we have Z
A(x)ru(x)r'(x) dx = 0 (17.11)
B1

for every ' 2 Cc1 (B1 \ {0}; Rn ). Using Lemma 17.1 again, we can extend (17.11) to every
' 2 Cc1 (B1 ; Rn ), thus obtaining the validity of the Euler-Lagrange PDE in the weak sense
in the whole ball.
Finally, since the functional L in (17.3) is convex, the Euler-Lagrange equation is
satisfied by u if and only if u is a minimizer of L(u) with boundary condition

u(x) = x in @B1 .

This means that De Giorgi’s counterexample holds not only for solution of the Euler-
Lagrange equation, but also for minimizers.

101
Lemma 17.1. Assume that n > 2. For every ' 2 Cc1 (B1 ) there exists 'k 2 Cc1 (B1 \{0})
such that 'k tends to ' strongly in W 1,2 (B1 ).
Proof. Consider 2 Cc1 (Rn ) with ⌘ 1 on B 1 , then rescale setting k (x) := (kx).
2
Set 'k := '(1 k ); in L topology we have ' 'k = ' k ! 0 and (r') k ! 0. Since
r(' 'k ) = (r') k + 'r k ,
the thesis is equivalent to verify that
Z
'(x)2 |r k (x)|
2
dx ! 0 ,
B1

but
Z Z
2 2 2 2
'(x) |r k (x)| dx  (sup ' )k |r (kx)|2 dx
B1 B1
Z
2 2 n
 (sup ' )k |r (x)|2 dx ! 0 ,
Rn

where we used the fact that n > 2. ⇤

We conclude noticing that the restriction n 3 in the proof of Lemma 17.1 is not
really needed. Indeed, when n = 2 we have
⇢Z
inf |r (x)|2 dx | 2 Cc1 (B1 ), = 1 in a neighbourhood of 0 = 0 . (17.12)

Let us prove (17.12): we first prove that

⇢Z 1
inf r|a0 (r)|2 dr | a(0) = 1, a(1) = 0 =0,
0

considering radial functions (x) = a(|x|). We can take a (r) := 1 r , so

Z 1
!0
r|a0 (r)|2 dr = !0.
0 2
Then, considering suitable approximations of a , for instance min{(1 r ), 1 }/(1 )
and their mollifications (which are equal to 1 in a neighbourhood of 0) we prove (17.12).
Using (17.12) to remove the point singularity also in the case n = 2, it follows that
the functional L(u) and its minimizer are a counterexample to Lipschitz regularity.
In a more general perspective, we recall that the p-capacity of a compact set K ⇢ Rn
is defined by
⇢Z
inf |r |p dx | 2 Cc1 (Rn ), ⌘ 1 in a neighbourhood of K .
Rn

We proved that singletons have null 2-capacity in Rn for n 2.

102
18 Partial regularity for systems
As we have seen with De Giorgi’s counterexample, it is impossible to expect an “every-
where” regularity result for elliptic systems: the main idea is to pursue a di↵erent goal, a
“partial” regularity result, away from a singular set. This strategy goes back to De Giorgi
himself, and it was implemented for the first time in the study of minimal surfaces.
Definition 18.1 (Regular and singular sets). For a generic function u : ⌦ ! R we call
regular set of u the set

⌦reg (u) := x 2 ⌦ 9 r > 0 s.t. Br (x) ⇢ ⌦ and u 2 C 1 (Br (x)) .

Analogously, the singular set is

⌃(u) := ⌦ \ ⌦reg (u) .

The set ⌦reg (u) is obviously the largest open subset A of ⌦ such that u 2 C 1 (A).
Briefly, let us recall here the main results we have already obtained for elliptic systems.

(a) If we are looking at the problem

R from the variational point of view, studying local
1 2 m⇥n
minimizers u 2 Hloc of v 7! ⌦ F (Dv) dx, with F 2 C (R ), |D2 F (p)|  ⇤, we
already have the validity of the Euler-Lagrange equations. More precisely, if
Z Z
F (ru(x)) dx  F (rv(x)) dx 8 v s.t. {u 6= v} b ⌦0 b ⌦ ,
⌦0 ⌦0

then
@
Fp↵i (ru) = 0 8 i = 1, . . . , m .
@x↵
(b) If F satisfies a uniform Legendre condition for some > 0, by Nirenberg method we
have ru 2 Hloc 1
(⌦; Rm⇥n ) and (by di↵erentiation of the (EL) equations with respect
to x )
✓ ◆
@ @ 2 uj
Fp↵ p (ru) = 0 8 i = 1, . . . , m, = 1, . . . , n . (18.1)
@x↵ i j @x @x

Definition 18.2 (Uniform quasiconvexity). We say that F is -uniformly quasiconvex if

Z Z
F (A + r'(x)) F (A) dx |r'|2 dx 8 ' 2 Cc1 (⌦; Rm ) .
⌦ ⌦

In this section we shall provide a fairly complete proof of the following result, following
with minor variants the original proof in [10].

103
Theorem 18.3 (Evans). If F 2 C 2 (Rm⇥n ) is -uniformly quasiconvex with > 0 and
satisfies
|r2 F (p)|  ⇤ 8p 2 Rm⇥n , (18.2)
for some ⇤ > 0. Then any local minimizer u belongs to C 1, (⌦reg ) for some =
(n, m, , ⇤) and
L n (⌦ \ ⌦reg ) = 0 .
The following list summarizes some results in the spirit of Theorem 18.3. At this stage
we should point out that the growth condition (18.2) is a bit restrictive if we want to allow
the standard examples of quasiconvex functions, i.e. convex functions of determinants of
minors of ru; it includes for instance functions of the form
s X
2
F (ru) := |ru| + 1 + (M ru)2
M

where M ru is a 2 ⇥ 2 minor of ru.

A more general growth condition considered in [10] is
|r2 F (p)|  C0 1 + |p|q 2
with q 2, (18.3)
which leads to the estimates |rF (p)|  C1 (1 + |p|q 1 ) and |F (p)|  C2 (1 + |p|q ).
(i) If r2 F I for some > 0, then Giaquinta and Giusti (see [16] and [18]) proved
a much stronger estimate on the size of the singular set, namely (here H k denotes
the Hausdor↵ measure, that we will introduce later on)
Hn 2+"
(⌃(u)) = 0 8" > 0 .

(ii) If r2 F I for some > 0 and it is globally uniformly continuous, then we have
even H n 2
(⌃(u)) = 0.
(iii) If u is locally Lipschitz, then Kristensen and Mingione proved in [23] that there
exists > 0 such that
H n (⌃(u)) = 0 .

(iv) On the contrary, when n = 2 and m = 3, there exists a Lipschitz solution u for
the system @x@↵ Fp↵i (ru) (with F smooth and satisfying the Legendre-Hadamard
condition), provided in [25], such that
⌦reg (u) = ; .
This last result clarifies once for all that partial regularity can be expected for (local)
minimizers only. We will see how local minimality (and not only the validity of the
Euler-Lagrange equations) plays a role in the proof of Evans’ result.

104
 We will start with a decay lemma relative to constant coefficients operators.

Lemma 18.4. There exists a constant C⇤ = C⇤ (n, m, , ⇤) 2 (0, 1) such that, for every
constant matrix A satisfying the Legendre-Hadamard condition with and the inequality
|A|  ⇤, any solution u 2 H 1 (Br ; Rm ) of

div(Aru) = 0 in Br

satisfies
Z Z
2 2
|ru(x) (ru)B↵r | dx  C⇤ ↵ |ru(x) (ru)Br |2 dx 8↵ 2 (0, 1) .
B↵r Br

Proof. As a consequence of what we proved in the section about decay estimates for
systems with constant coefficients, considering (5.2) with ⇢ = ↵r and ↵ < 1, we have that
Z ⇣ ↵r ⌘n+2 Z
2
|ru(x) (ru)B↵r | dx  c(n, m, , ⇤) |ru(x) (ru)Br |2 dx . (18.4)
B↵r r Br

It is enough to consider the mean of (18.4), so that

Z Z
2 2
|ru(x) (ru)B↵r | dx  c(n, m, , ⇤)↵ |ru(x) (ru)Br |2 dx .
B↵r Br

1
Definition 18.5 (Excess). For any function u 2 Hloc (⌦; Rm ) and any ball B⇢ (x) b ⌦ the
excess of u in B⇢ (x) is defined by
Z !1/2
Exc (u, B⇢ (x)) := |ru(y) (ru)B⇢ (x) |2 dy .
B⇢ (x)

When we consider functions F satisfying the more general growth condition (18.3),
then we should modify the definition of excess as follows, see [10]:
Z
2
Exc (u, B⇢ (x)) = 1 + |ru(y) (ru)B⇢ (x) |q 2 |ru(y) (ru)B⇢ (x) |2 dy.
B⇢ (x)

However, in our presentation we will cover only the case q = 2.

Remark 18.6 (Properties of the excess). We list here the basic properties of the excess,
they are trivial to check.

105
 (i) Any additive perturbation by an affine function p(x) does not change the excess,
that is
Exc (u + p, B⇢ (x)) = Exc (u, B⇢ (x)) .

(ii) The excess is positively 1-homogeneous, that is for any number 0

Exc ( u, B⇢ (x)) = Exc (u, B⇢ (x)) .

(iii) We have the following scaling property:

✓ ◆
u(⇢ · +x0 )
Exc , B1 (0) = Exc (u, B⇢ (x0 )) .
⇢

Remark 18.7. The name “excess” is inspired by De Giorgi’s theory of minimal surfaces,
presented in [6] and [7], see also [15] for a modern presentation. The excess of a set E at
a point is defined (for regular sets) by
Z
Exc (E, B⇢ (x)) := |⌫E (y) ⌫E (x)|2 d n 1 (y) ,
B⇢ (x)\@E

where ⌫E is the inner normal of the set E. The correspondence between Exc (u, B⇢ (x))
and Exc (E, B⇢ (x)) can be made more evident seeing near x the set @E as the graph
associated to a function
p u, in a coordinate system where ru(x) = 0. Indeed, the identity
⌫E = ( ru, 1)/ 1 + |ru|2 and the area formula for graphs give
Z Z p Z
2
|⌫E (y) ⌫E (x)| d n 1 (y) = 2 2
1 + |ru(z)| 1 dz ⇠ |ru(z)|2 dz ,
B⇢ (x)\@E ⇡(B⇢ (x)\@E) B⇢ (z)

where ⇡(B⇢ (x) \ @E) denotes the projection of the B⇢ (x) \ @E on the hyperplane.

The main ingredient in the proof of Evans’ theorem will be the decay property of the
excess: there exists a critical threshold such that, if the decay in the ball is below the
threshold, then decay occurs in the smaller balls.

Theorem 18.8 (Excess decay). Let F be as in Theorem 18.3. For every M 0 and all
↵ 2 (0, 1/4) there exists "0 = "0 (n, m, , ⇤, M, ↵) > 0 satisfying the following implication:
if
R
(a) u 2 H 1 (Br (x); Rm ) is a local minimizer in Br (x) of v 7! F (rv) dx,

(b) |(ru)Br (x) |  M ,

(c) Exc (u, Br (x)) < "0 ,

106
then
Exc (u, B↵r (x))  Ce ↵ Exc (u, Br (x))
with Ce depending only on (n, m, , ⇤). When r2 F is uniformly continuous, condition
(b) is not needed for the validity of the implication and "0 is independent of M .
Proof. We choose Ce in such a way that 16C⇤2 CP C ⇤ < Ce2 , where C⇤ is the constant
of Lemma 18.4, CP is the constant in the Poincaré inequality and C ⇤ is the constant of
Proposition 18.9 below.
The proof is by contradiction, assuming that the statement fails for some ↵ and M
(for simplicity we keep F fixed in the contradiction argument, but a slightly more complex
proof would give the stronger result): in step (ii) we will normalize the excesses, obtaining
functions wk with Exc (wk , B↵ (0)) Ce ↵ while Exc (wk , B1 (0)) = 1. Each wk is a solution
of
@
Fp↵i (rwk ) = 0 .
@x↵
We will see in step (iii) that, passing through the limit as k ! 1, any limit point w1
w.r.t. the weak H 1 topology solves
⇣ ⌘
div Fp↵ p (p1 )rw1 = 0.
i j

Using Lemma 18.4 in combination with Proposition 18.9 we will reach the contradiction.
(i) By contradiction, we have M 0, ↵ 2 (0, 1/4) and local minimizers uk : ⌦ ! Rm in
Brk (xk ) with
"k := Exc (uk , Brk (xk )) ! 0
satisfying
(ruk )Brk (xk )  M (18.5)
but
Exc (uk , B↵rk (xk )) > Ce ↵ Exc (uk , Brk (xk )) 8k 2 N .
(ii) Suitably rescaling and translating the functions uk , we can assume that xk = 0,
rk = 1 and (uk )B1 = 0 for all k. Setting pk := (ruk )B1 , the hypothesis (18.5) gives, up to
subsequences,
pk ! p1 2 Rm⇥n . (18.6)
We start here a parallel and simpler path through this proof, in the case when r2 F is
uniformly continuous: in this case no uniform bound on pk is needed and we can replace
(18.6) with
2 2
r2 F (pk ) ! A1 2 Rm ⇥n . (18.7)
Notice that (18.7) holds under (18.6), simply with A1 = r2 F (p1 ). Notice also that,
in any case, A1 satisfies a (LH) condition with constant (this can be achieved using

107
oscillating test functions, as we did to show that quasi-convexity implies the Legendre-
Hadamard condition) and |A1 |  ⇤
We do a second translation in order to annihilate the mean of the gradients of mini-
mizers: let us define
vk (x) := uk (x) pk (x) ,
so that (vk )B1 = 0 and (rvk )B1 = 0. According to property (i) of Remark 18.6 the excess
does not change, so still
Exc (vk , B1 ) = "k ! 0
and
Exc (vk , B↵ ) > Ce ↵ "k .
During these operations, we need not lose sight of the variational problem we are solving,
for example every function vk minimizes the integral functional associated to

p 7! F (p + pk ) F (pk ) rF (pk )p .

In order to get some contradiction, our aim is to find a “limit problem” with some decaying
property. Let us define
vk
wk := k 2 N.
"k
It is trivial to check that (wk )B1 = (rwk )B1 = 0, moreover

Exc (wk , B1 ) = 1 andExc (wk , B↵ ) > Ce ↵ . (18.8)

R
The key point of the proof is that wk is a local minimizer of v 7! Fk (rv) dx, where

1
Fk (p) := [F ("k p + pk ) F (pk ) rF (pk )"k p] .
"2k

Here we used the fact local minimality w.r.t. to an integrand F is preserved if we multiply
F by a positive constant or add to F an affine function.
(iii) We now study both the limit of Fk and the limit of wk , as k ! 1. Since Fk 2
C 2 (Rm⇥n ), by Taylor expansion we are able to identify a limit Lagrangian, given by
1
F1 (p) = hA1 p, pi ,
2
to which Fk (p) converge uniformly on compact subsets of Rm⇥n . Indeed, this is clear
with A1 = r2 F (p1 ) in the case when pk ! p1 ; it is still true with A1 given by
(18.7) when r2 F is uniformly continuous, writing Fk (p) = 12 hr2 F (pk + ✓"k p)p, pi with
✓ = ✓(k, p) 2 (0, 1).

108
Once we have the limit problem defined by F1 , we drive our attention to wk : it is
a bounded sequence in H 1,2 (B1 ; Rm ) because the excesses are constant, so by Rellich
theorem we have that (possibly extracting one more subsequence)

wk ! w1 in L2 (B1 ; Rm )

and, as a consequence,
rwk * rw1 in L2 (B1 ; Rm ) . (18.9)
The analysis of the limit problem now requires the verification that w1 solves the Euler
equation associated to F1 . We need just to pass to the limit in the (EL) equation satisfied
by wk , namely
X Z 1 ✓ @F @F
◆ i
@
↵
(pk + "k rwk (x)) ↵
(pk ) (x) dx = 0 8 ' 2 Cc1 (B1 ; Rm ) .
" @pi
↵,i B1 k
@pi @x↵

Writing the di↵erence quotient of rF with the mean value theorem and using r2 F (pk ) !
A1 we obtain
Z
hA1 rw1 (x), r'(x)i dx = 0 8 ' 2 Cc1 (B1 ; Rm ) , (18.10)
B1

provided we show that (here ✓ = ✓(x, ↵, ) 2 (0, 1))

XXZ @ 2F
lim | ↵ (pk + ✓"k rwk ) (A1 )↵ij | dx = 0 .
k!1
↵, i, j B1 @pi pj

This can be obtained splitting the integral into the regions {|rwk |  L} and {|rwk | > L},
with L fixed. The first contribution goes to zero, thanks to the convergence of pk to p1
or, when pk is possibly unbounded, thanks to the uniform continuity of r2 F . The second
contribution tends to 0 as L " 1 uniformly in k, since |r2 F |  ⇤ and krwk k2  1.
(iv) Equality (18.10) means that

div (A1 rw1 ) = 0

in a weak sense: since the equation has constant coefficients we can apply Lemma 18.4 to
get
Z Z
2 2 2
|rw1 (x) (rw1 )B2↵ | dx  4C⇤ ↵ |rw1 (x)|2 dx  4C⇤2 ↵2 . (18.11)
B2↵ B1

On the other hand, using Proposition 18.9 below we get

Z
2 2 2 C⇤
Ce ↵ < Exc (wk , B↵ )  2 |wk (wk )2↵ (rwk )2↵ (x)|2 dx ,
↵ B2↵

109
hence passing to the limit as k ! 1 gives
Z
Ce2 4
↵  |w1 (w1 )2↵ (rw1 )2↵ (x)|2 dx .
C⇤ B2↵

On the other hand, the Poincaré inequality and (18.11) gives

Z Z
2 2
|w1 (w1 )2↵ (rw1 )2↵ (x)| dx  4CP ↵ |rw1 (rw1 )2↵ |2 dx  16CP C⇤2 ↵4 .
B2↵ B2↵
Taking into account our definition of Ce we have reached a contradiction.

⇤
The following proposition can be considered as a nonlinear Caccioppoli inequality. It
can be derived without using the Euler-Lagrange equation (which would not help) and
using the minimality instead.
Proposition 18.9 (Caccioppoli inequality for minimizers). There exists C ⇤ = C(n, m, , ⇤)
such that if F is -quasiconvex with |r2 F |  ⇤ and if u is a local minimizer in ⌦, then
Z Z
2 C⇤
|ru A| dx  2 |u a A(x x0 )|2 dx
Br/2 (x0 ) r Br (x0 )

for all balls Br (x0 ) b ⌦, all A 2 Rm⇥n and a 2 Rm .

Proof. By translation invariance we can assume a = 0, x0 = 0. Let r/2  t < s  r and
let ⇣ 2 Cc1 (Bs ) with ⇣ ⌘ 1 on Bt , 0  ⇣  1 and |r⇣|  2(s t). Define = ⇣(u Ax),
= (1 ⇣)(u Ax), so that + = u Ax gives
r + r = ru A.
From the -uniform quasiconvexity we get
Z Z
2
F (A) + |r | dx  F (A + r ) dx
Bs Bs
Z
= F (ru r ) dx (18.12)
Z Bs
 F (ru) rF (ru)r + C|r |2 dx ,
Bs

with C = C(⇤). On the other hand, since u is a local minimum, we have

Z Z
F (ru) dx  F (ru r ) dx
Bs Bs
Z
= F (A + r ) dx (18.13)
Bs
Z
 F (A) + rF (A)r + C|r |2 dx .
Bs

110
Combining (18.12) with (18.13) we get
Z Z
2
|r | dx  |rF (A) rF (ru)||r | + C|r |2 dx ,
Bs Bs

so that (using also that ⌘ 0 on Bt )

Z Z
2
|ru A| dx  C |ru A||r | + |r |2 dx ,
Bt Bs \Bt

with C = C( , ⇤).
Now, since |r |  |ru A| + 2|u Ax|/(s t), we get
Z Z Z
2 2 D
|ru A| dx  D |ru A| dx + |u Ax|2 dx
Bt Bs \Bt (s t)2 Br

for some new constant D = D( , ⇤). Now we apply the hole-filling technique to get
Z Z Z
2 2 D
|ru A| dx  ✓ |ru A| dx + |u Ax|2 dx .
Bt Bs (s t)2 Br
with ✓ = D/(D +1) < 1. At this point, since the inequality is true for all r/2  t  s  r,
a standard iteration scheme gives the result. Indeed, let ⌧ 2 (0, 1) with ✓ < ⌧ 2 and define
ti = (1 ⌧ i /2)r, so that t0 = r/2, ti " r and ti+1 ti = r(1 ⌧ )⌧ i /2. By iteration of the
inequality Z Z
2 4D
|ru A| dx  ✓ |ru A|2 dx + 2 ⌧ 2i
Bt i Bti+1 r (1 ⌧ )2
we get
Z Z N
X1
2 N 2 4D
|ru A| dx  ✓ |ru A| dx + 2 (✓/⌧ 2 )i
Bt 0 BtN r (1 ⌧ )2 i=0
Z
N 4D⌧ 2
 ✓ |ru A|2 dx +
Br r2 (1 ⌧ )2 (⌧ 2 ✓)
for any integer N 1. As N ! 1 we get the result.

18.1 Partial regularity for systems: L n (⌃(u)) = 0

After proving Theorem 18.8 about the decay of the excess, we will see how it can be used
to prove partial regularity for systems.
We briefly recall that ⌦reg (u) denotes the largest open set contained in ⌦ where u :
⌦ ! Rm admits a C 1 representative, while ⌃(u) := ⌦ \ ⌦reg (u). Our aim is to show that
for a solution of an elliptic system the following facts:

111
 • L n ⌃(u) = 0;
• H n 2+" ⌃(u) = 0 for all " > 0 in the uniformly convex case and H n 2
(⌃(u)) = 0
if r2 F is also uniformly continuous.
In order to exploit Theorem 18.8 and prove that L n ⌃(u) = 0, we fix once for all
the constant ↵ 2 (0, 1/4) in such a way that Ce ↵ < 1/2 (recall that Ce depends only on
the dimensions and on the ellipticity constants). Then, we fix M 0, so that there is an
associated "0 = "0 (n, m, , ⇤, M ) for which the decay property of the excess applies with
halving of the excess from the scale r to the scale ↵r.
Definition 18.10. We will call

⌦M (u) := x 2 ⌦ 9Br (x) b ⌦ with (ru)Br (x) < M1 and Exc (u, Br (x)) < "1

where
M1 := M/2 (18.14)
and "1 verifies
2n/2 "1  "0 (18.15)
and for ↵ 2 (0, 1/4) fixed, chosen in such a way that Ce ↵ < 1/2,

(2n+1 + ↵ n 1+n/2
2 )"1  M . (18.16)

Remark 18.11. The set ⌦M (u) ⇢ ⌦ of Definition 18.10 is open, since the inequalities
are strict. Moreover, by Lebesgue approximate
R continuity theorem (that is, if f 2 Lp (⌦),
then for L -almost every x one has Br (x) |f (y) f (x)|p dy ! 0 as r # 0), it is easy to
n

see that
L n ({|ru| < M1 } \ ⌦M (u)) = 0 . (18.17)
Finally, using (18.17), we realize that
! !
[ [
Ln ⌦\ ⌦M (u) = L n ⌦\ {|ru| < M1 } =0. (18.18)
M 2N M 2N

By the previous remark, if we are able to prove that

⌦M (u) ⇢ ⌦reg 8M > 0, (18.19)

we obtain L n (⌃(u)) = 0. So, the rest of this section will be devoted to the proof of the
inclusion above, with M fixed.
Fix x 2 ⌦M (u), according to Definition 18.10 there exists r > 0 such that Br (x) b ⌦,
|(ru)Br (x) | < M1 and Exc (u, Br (x)) < "1 . We will prove that

Br/2 (x) ⇢ ⌦reg (u) ,

112
so let us fix y 2 Br/2 (x).
(1) Thanks to our choice of "1 (see property (18.15) of Definition 18.10) we have
Z !1/2
Exc u, Br/2 (y) = |ru(z) (ru)Br/2 (y) |2 dz
Br/2 (y)
Z !1/2
 |ru(z) (ru)Br (x) |2 dz
Br/2 (y)
✓Z ◆1/2
n/2 2
 2 |ru(z) (ru)Br (x) | dz = 2n/2 Exc (u, Br (x)) < "0
Br (x)

so, momentarily ignoring the hypothesis that |(ru)Br/2 (y) | should be bounded by M (we
are postponing this to point (2) of this proof), Theorem 18.8 gives tout court
1 1
Exc u, B↵r/2 (y)  Exc u, Br/2 (y) < "0 ,
2 2
thus, just iterating Theorem 18.8, we get

Exc u, B↵k r/2 (y)  2 k Exc u, Br/2 (y) . (18.20)

As we have often seen through these notes, we can apply an interpolation argument to a
sequence of radii with ratio ↵ to obtain
✓ ◆µ ✓ ◆µ
µ ⇢ µ ⇢
Exc (u, B⇢ (y))  ↵ Exc u, Br/2 (y)  ↵ "0 8⇢ 2 (0, r/2], y 2 Br/2 (x)
r/2 r/2

with µ = (log2 (1/↵)) 1 . We conclude that the components of ru belong to the Cam-
panato space L2,n+2µ (Br/2 (x)) and then u belongs to C 1,µ (Br/2 (x)).
(2) Now that we have explained how the proof runs through the iterative application of
Theorem 18.8, we deal with the initially neglected hypothesis, that is |(ru)Br/2 (y) | < M
and, at each subsequent step, |(ru)B↵k r/2 (y) | < M . Remember that in part (1) of this
proof we never used (18.14) and (18.16).
Since x 2 ⌦M (u) and r fulfills Definition 18.10, for the first step it is sufficient to use the

113
triangular inequality in (18.21) and Hölder’s inequality in (18.22): in fact we can estimate
Z
(ru)Br/2 (y) = ru(z) (ru)Br (x) dz + (ru)Br (x)
Br/2 (y)
Z
 ru(z) (ru)Br (x) dz + (ru)Br (x) (18.21)
Br/2 (y)
✓ Z ◆
2n
 ru(z) (ru)Br (x) dz + (ru)Br (x)
!n rn Br (x)
✓Z ◆1/2
n 2
 2 ru(z) (ru)Br (x) dz + (ru)Br (x) (18.22)
Br (x)

 2n Exc (u, Br (x)) + (ru)Br (x) < 2n "1 + M1 < M . (18.23)

We now show inductively that for every integer k 1

k 1
X
n n n/2 j
(ru)B↵k r/2 (y)  M1 + 2 "1 + ↵ "1 2 2 . (18.24)
j=0

If we recall (18.14) and (18.16), it is clear that (18.24) implies

(ru)B↵k r/2 (y)  M

for every k 1.
The first step (k = 1) follows from (18.23), because, estimating as in (18.21) and (18.22),
we immediately get
Z
(ru)B↵r/2 (y)  ru(z) (ru)Br/2 (y) dz + (ru)Br/2 (y)
B↵r/2 (y)
n
 ↵ Exc u, Br/2 (y) + (ru)Br/2 (y)
n n/2
 ↵ 2 "1 + 2 n "1 + M 1 .

Being the first step already proved, we fix our attention on the (k + 1)th step. With the
same procedure, we estimate again
Z
(ru)B↵k+1 r/2 (y)  ru(z) (ru)B↵k r/2 (y) dz + (ru)B↵k r/2 (y)
B↵k+1 r/2 (y)
n
 ↵ Exc u, B↵k r/2 (y) + (ru)B↵k r/2 (y)
k 1
X
n n/2 k n n n/2 j
 ↵ 2 "1 + M 1 + 2 " 1 + ↵ "1 2 2 (18.25)
j=0

114
where (18.25) is obtained joining the estimate on the excess (18.20) with the inductive
hypothesis (18.24).
In order to carry out our second goal, namely to prove that

Hn 2+"
⌃(u) = 0 8" > 0 ,

we need some basic results concerning Hausdor↵ measures.

18.2 Hausdor↵ measures

Definition 18.12. Consider a subset B ⇢ Rn , k 0 and fix 2 (0, 1]. The so-called
pre-Hausdor↵ measures H k are defined by
(1 1
)
X [
H k (B) := ck inf [diam(Bi )]k B ⇢ Bi , diam(Bi ) < ,
i=1 i=1

while H k is defined by
H k (B) := lim H k (B) , (18.26)
!0

the limit in (18.26) being well defined because 7! H k (B) is non-increasing. The constant
ck 2 (0, 1) will be conveniently fixed in Remark 18.14.
It is easy to check that H k is the counting measure when k = 0 (provided c0 = 1)
and H k is identically 0 when k > n.
The spherical Hausdor↵ measure S k has a definition analogous to Definition 18.12,
but only covers made with balls are allowed, so that

H k  S k  2k H k , H k  S k  2k H k . (18.27)

Remark 18.13. Simple but useful properties of Hausdor↵ measures are:

(i) The Hausdor↵ measures are translation invariant, that is

H k (B + h) = H k (B) 8 B ⇢ Rn , 8 h 2 Rn ,

and (positively) k-homogeneous, that is

H k ( B) = k
H k (B) 8 B ⇢ Rn , 8 > 0 .

(ii) The Hausdor↵ measures are countably subadditive, which means that whenever we
have a countable cover of a subset B, namely B ⇢ [i2I Bi , then
X
H k (B)  H k (Bi ) .
i2I

115
(iii) For every set A ⇢ Rn the map B 7! H k (A \ B) is -additive on Borel sets, which
means that whenever we have a countable pairwise disjoint cover of a Borel set B
by Borel sets Bi , we have
X
H k (A \ B) = H k (A \ Bi ) .
i2I

(iv) Having fixed the subset B ⇢ Rn and > 0, we have that

k k0 0
k > k0 =) H k (B)  H k (B) . (18.28)
In particular, looking at (18.28) when ! 0, we deduce that
0
H k (B) < +1 =) H k (B) = 0
or, equivalently,
0
H k (B) > 0 =) H k (B) = +1 .
Remark 18.14. When k is an integer, the choice of ck is meant to be consistent with
the usual notion of k-dimensional area: if B is a Borel subset of a k-dimensional plane
⇡ ⇢ Rn , 1  k  n, then we would like that
L⇡k (B) = H k (B) , (18.29)
where L⇡k is the k-dimensional Lebesgue measure on ⇡ ⇠ Rk . It is useful to remember
the isodiametric inequality among all sets with prescribed diameter, balls have the largest
volume: more precisely, if !k := L k (B1 (0)), for every Borel subset B ⇢ Rk there holds
✓ ◆k
k diam(B)
L (B)  !k . (18.30)
2
Thanks to (18.30), it can be easily proved that equality (18.29) holds if we choose
!k
ck = k .
2
Recall also that !k can be computed by the formula !k = ⇡ k/2 / (1 + k/2), where is
Euler’s function: Z 1
(t) := st 1 e s ds .
0
More generally, with this choice of the normalization constant, if B is contained in an
embedded C 1 -manifold M of dimension k in Rn , then
H k (B) = k (B)

where k is the classical k-dimensional surface measure defined on Borel subsets of M by

local parametrizations and partitions of unity.

116
Proposition 18.15. Consider a locally finite measure µ 0 on the family of Borel sets
B(Rn ) and, fixing t > 0, set
⇢
µ(B r (x))
B := x | lim sup >t , (18.31)
r!0 !k rk

then B is a Borel set and

µ(B) tS k (B) .
Moreover, if µ vanishes on H k -finite sets, then H k (B) = 0.

A traditional proof of Proposition 18.15 is based on Besicovitch covering theorem,

whose statement is included below for the sake of completeness. We present instead a
proof based on a more general and robust covering theorem, valid in general metric spaces.

Theorem 18.16 (Besicovitch). There exists an integer ⇠ = ⇠(n) with the following prop-
erty: if A ⇢ Rn is bounded and ⇢ : A ! (0, 1), there exist sets A1 , . . . , A⇠(n) ⇢ A such
that

(a) for all j = 1, . . . , ⇠, the balls in {B⇢(x) (x)}x2Aj are pairwise disjoint;

(b) the ⇠ families still cover the set A, that is

0 1
⇠
[ [
A⇢ @ B⇢(x) (x)A .
j=1 x2Aj

Let us introduce now the general covering theorem.

Definition 18.17 (Fine cover). A family F of closed balls in a metric space (X, d) is a
fine cover of a set A ⇢ X if

inf r > 0|B r (x) 2 F = 0 for all x 2 A .

Theorem 18.18. Fix k 0, consider a fine cover F of A ⇢ X, with (X, d) metric space.
Then there exists a countable and pairwise disjoint subfamily F 0 = {B i }i 1 ⇢ F such that
at least one of the following conditions holds:
P
1
(i) [r(Bi )]k = 1,
i=1
✓ ◆
k
S
1
(ii) H A\ Bi = 0.
i=1

117
Proof. The subfamily F 0 is chosen inductively, beginning with F0 := F. Surely, there
exists a closed ball, let us call it B 1 , such that
1
r B1 > sup r(B)| B 2 F0 .
2
Now put
F1 := {B 2 F0 | B \ B 1 = ;} ,
and choose among them a ball B 2 2 F1 such that
1
r(B 2 ) > sup r(B)| B 2 F1 .
2
If we try to go on analogously, the only chance by which the construction has to stop
is that for some l 2 N the family Fl = ;, so we are getting (because the cover is fine)
that the union of the chosen balls covers the whole of A and therefore option (ii) in the
statement.
Otherwise, assuming that the construction does not stop, we get a family F 0 = {B i }i 1 =
{B ri (yi )}i 1 . We prove that if (i) does not hold, and in particular diam(B i ) ! 0, then
we have to find (ii) again. S
Fix an index i0 2 N: for every x 2 A \ i10 B i there exists a ball B r(x) (x) 2 F such
that
i0
[
B r(x) (x) \ Bi = ; ,
i=1

because F is a fine cover of A and the complement of [i10 B i is open in X. On the other
hand, we claim that there exists an integer i(x) > i0 such that

B r(x) (x) \ B i(x) 6= ; . (18.32)

In fact if
B r(x) (x) \ B i = ; 8 i > i0 , (18.33)
then
r(x)
ri 88 i > i0 (18.34)
2
but ri ! 0, so (18.34) leads to a contradiction. Without loss of generality, we can
think that i(x) is the first index larger than i0 for which (18.32) holds, too. Since, by
construction, ri(x) > 12 sup{r(B)|B 2 Fi(x) 1 } (and B r(x) (x) 2 Fi(x) 1 by the minimality
of i(x)), then r(x)  2ri(x) .
Since the balls intersect, the inequality d(x, yi(x) )  r(x) + ri(x)  3ri(x) gives

B r(x) (x) ⇢ B 5ri(x) (yi(x) )

118
and therefore
i0
[ 1
[
A\ Bi ⇢ B 5ri (yi ) . (18.35)
i=1 i=i0 +1

Choosing i0 such that 10ri < for every i > i0 , (18.35) says that
1
! i0
! 1
[ [ X
Hk A\ Bi  H k
A\ Bi  !k (10ri )k .
i=1 i=1 i=i0 +1

We conclude remarking that when ! 0, i0 ! +1 and

1
! 1
[ X
Hk A\ B i  lim !k (10ri )k = 0 .
i0 !1
i=1 i=i0 +1

⇤
Now we are able to prove Proposition 18.15.
Proof. Intersecting B with balls, one easily reduces to the case of a bounded set B.
Hence, we can assume B bounded and µ finite measure. Fix > 0, an open set A B
and consider the family

F := B r (x) r < /2, B r (x) ⇢ A, µ (Br (x)) > t!k rk , (18.36)

that is a fine cover of B. Applying Theorem 18.18, we get a subfamily F 0 ⇢ F whose

elements we will denote by
B i = B ri (xi ) .
First we exclude possibility (i) of Theorem 18.18: as a matter of fact
1
X 1
1 X µ(A)
rik < µ(B i )  <1.
i=1
t!k i=1 t!k

Since (ii) holds and we can compare H k with S k via (18.27), to get
1
! 1 1
[ X 1X µ(A)
S k (B)  S k Bi  !k rik < µ(B i )  , (18.37)
i=1 i=1
t i=1
t

As # 0 we get tS k (B)  µ(A) and the outer regularity of µ gives tS k (B)  µ(B).
Finally, the last statement of the proposition can be achieved noticing that the in-
equality (18.37) gives that S k (B) is finite; if we assume that µ vanishes on sets with
finite k-dimensional measure we obtain that µ(B) = 0; applying once more the inequality
we get S k (B) = 0. ⇤

119
18.3 Partial regularity for systems: H n 2+"
(⌃(u)) = 0
Aware of the usefulness of Proposition 18.15 for our purposes, we are now ready to obtain
that if F 2 C 2 (Rm⇥n ) satisfies the Legendre condition for some > 0 and satisfies also

|r2 F (p)|  ⇤ < 1 8 p 2 Rm⇥n

then we have a stronger upper bound on the size of the singular set, namely

Hn 2+"
(⌃(u)) = 0 8" > 0 , (18.38)

where, as usual, ⌃(u) := ⌦ \ ⌦reg (u).

Let us remark that, with respect to the first partial regularity result and with respect
to Evans Theorem 18.3, we slightly but significantly changed the properties of the system,
replacing the weaker hypothesis of uniform quasiconvexity with the Legendre condition
for some positive (i.e. uniform convexity). In fact, thanks to the Legendre condition
the sequence h,s (ru) satisfies an equielliptic family of systems, then, via Caccioppoli
inequality the sequence h,s (ru) is uniformly bounded in L2loc . The existence of second
derivatives in L2loc is useful to estimate the size of the singular set.
We will also obtain a stronger version of (18.38) for systems in which r2 F is uniformly
continuous, we will see it in Corollary 18.21.
As for the strategy: in Proposition 18.19 we are going to split the singular set ⌃(u)
in two other sets, ⌃1 (u) and ⌃2 (u), and then we are going to estimate separately the
Hausdor↵ measure of each of them with the aid of Proposition 18.20 and Theorem 18.23,
respectively.

Proposition 18.19. Consider, as previously, a variational problem defined by F 2

C 2 (Rm⇥n ) with |r2 F |  ⇤, satisfying the Legendre condition for some > 0. If u is
a local minimizer of such a problem, define the sets
⇢ Z
2 n
⌃1 (u) := x 2 ⌦ lim sup r |r2 u(y)|2 dy > 0
r!0 Br (x)

and ⇢
⌃2 (u) := x 2 ⌦ lim sup (ru)Br (x) = +1 .
r!0

Then ⌃(u) ⇢ ⌃1 (u) [ ⌃2 (u). If in addition r2 F is uniformly continuous, we have ⌃(u) ⇢

⌃1 (u).
Proof. Fix x 2 ⌦ such that x 2
/ ⌃1 (u) [ ⌃2 (u), then

• there exists M1 < 1 such that (ru)Br (x) < M1 for arbitrarily small radii r > 0;

120
 • thanks to Poincaré inequality
Z
2 2 n
Exc (u, Br (x))  C(n)r |r2 u(y)|2 dy ! 0 ;
Br (x)

thus for some M = M (M1 , n, m, , ⇤) > 0 we have that x 2 ⌦M (u), where ⌦M (u) has
been specified in Definition 18.10, and ⌦M (u) ⇢ ⌦reg due to (18.19).
The second part of the statement can be achieved noticing that, in the case when r2 F
is uniformly continuous, no bound on |(ru)Br (x) | is needed in the decay theorem and in
the characterization of the regular set. ⇤
2,2
Proposition 18.20. If u 2 Wloc (⌦), we have that

Hn 2
(⌃1 (u)) = 0 .
Proof. Let us employ Proposition 18.15 with the absolutely continuous measure µ :=
|r2 u|2 L n . Obviously we choose k = (n 2) and we have that µ vanishes on sets with
finite H n 2 -measure. The thesis follows when we observe that
[1 ⇢
µ(B r (x)) 1
⌃1 (u) = x 2 ⌦ lim sup n 2
> .
⌫=1
r!0 !n 2 r ⌫

⇤
By the second part of the statement of Proposition 18.19 we get:
Corollary 18.21. If we add the uniform continuity of D2 F to the hypotheses of Propo-
sition 18.20, we can conclude that

Hn 2
(⌃(u)) = 0 . (18.39)

The estimate on the Hausdor↵ measure of ⌃2 (u) is a bit more complex and passes
through the estimate of the Hausdor↵ measure of the so-called approximate discontinuity
set Sv of a function v.
Definition 18.22. Given a function v 2 L1loc (⌦), we put
⇢ Z
⌦ \ Sv := x 2 ⌦ 9 z 2 R s.t. lim |v(y) z| dy = 0 .
r#0 Br (x)

When such a z exists, it is unique and we will call it approximate limit of v at the point
x.
Theorem 18.23. If v 2 W 1,p (⌦), 1  p  n, then

Hn p+"
(Sv ) = 0 8" > 0 .

121
 Notice that the statement is trivial in the case p > n, by the Sobolev Embedding
Theorem (i.e. Sv = ;): as p increases the Hausdor↵ dimension of the approximate
discontinuity set moves from n 1 to 0.
Applying this theorem to v = ru 2 H 1,2 (⌦; Rm⇥n ), p = 2, we get that H n 2+" (⌃2 (u)) =
0.
Proof. (1) Fix 0 < ⌘ < ⇢, we claim that
Z ⇢ Z Z
n!n (v)B⌘ (x) (v)B⇢ (x)  (n 1) t n |rv(y)| dy dt + ⇢ (n 1)
|rv(y)| dy ;
0 Bt (x) B⇢ (x)
(18.40)
we will show this in the part (3) of this Rproof.
Suppose that x is a point for which Bt (x) |rv(y)| dy = o(tn 1+" ) for some " > 0, then
R
it is also true that ⇢ (n 1) B⇢ (x) |rv(y)| dy ! 0 and the sequence (v)Br (x) admits a limit
z as r ! 0 because it is a Cauchy sequence. Thanks to the Poincaré inequality
Z Z
(n 1) r!0
|v(y) (v)Br (x) | dy  C(n)r |rv(y)| dy ! 0 ,
Br (x) Br (x)

therefore Z
r!0
|v(y) z| dy ! 0 ,
Br (x)

that is to say, x 2
/ Sv . This chain of implications means that, for all " > 0,
⇢ Z
⌦ \ Sv x2⌦ |rv(y)| dy = o(tn 1+" ) . (18.41)
Bt (x)

(2) In order to refine (18.41) suppose that

Z
|rv(y)|p dy = o(tn p+"
)
Bt (x)

for some " > 0, then, by Hölder’s inequality,

Z
0
|rv(y)| dy  o(tn/p 1+"/p
)tn/p = o(tn 1+"/p
).
Bt (x)

For this reason we can deduce from (18.41) the inclusion

⇢ Z
⌦ \ Sv x2⌦ |rv(y)|p dy = o(tn p+" ) 8" > 0 . (18.42)
Bt (x)
R
In view of Proposition 18.15, the complement of the set {x 2 ⌦ | Bt (x) |rv(y)|p dy =
o(tn p+" )} is H n p+" -negligible, hence the jump set Sv is H n p+" -negligible, too.

122
 (3) This third part is devoted to the proof of (18.40); for the sake of simplicity we put
x = 0. Let us consider the characteristic function B1 ; since we would like to di↵erentiate
the map Z ✓ ◆
n y
⇢ 7! ⇢ v(y) dy ,
⇢
a possible proof of (18.40) is based on a regularization of , di↵erentiation and passage
to the limit.
We produce instead a direct proof based on a ad hoc calibration: we need a vector
field with supp ⇢ B ⇢ whose divergence almost coincides with the operator acting on
v in left member of (18.40), that is
n n
div = n ⌘ B⌘ ⇢ B⇢ . (18.43)

Therefore,
n n n n
(x) := x ⌘ ^ |x| ⇢ ^ |x|
verifies (18.43) and, with the notation µ = |rv| B⇢ L n , there holds
Z Z Z
n n
v(y) dy v(y) dy = v(y)div (y) dy (18.44)
⌘ n B⌘ ⇢n B⇢
Z Z Z
= (y) · rv(y) dy  | (y)||rv(y)| dy  |y| (n 1) dµ(y) (18.45)
B⇢ R n
Z 1 Z 1
(n 1)
= µ |y| > t dt = (n 1) s n µ(Bs ) ds (18.46)
0 0
Z ⇢ Z Z 1 Z
n n
= (n 1) s |rv(y)| dy ds + (n 1) s |rv(y)| dy ds
0 Bs ⇢ B⇢
Z ⇢ Z Z
n (n 1)
= (n 1) s |rv(y)| dy ds + ⇢ |rv(y)| dy ,
0 Bs B⇢

where we pass from (18.44) to (18.45) by the divergence theorem, from (18.45) to (18.46)
by Cavalieri’s principle and then it is all change of variables and Fubini’s theorem. ⇤

Remark 18.24. In the case p = 1 it is even possible to prove that Sv is -finite with
respect to H n 1 , so the measurement of the discontinuity set with the scale of Hausdor↵
measures is sharp. On the contrary, in the case p > 1 the right scale for the measurement
of the approximate discontinuity set are the so-called capacities.

123
19 Some tools from convex and nonsmooth analysis
19.1 Subdi↵erential of a convex function
In this section we briefly recall some classical notions and results from convex and nons-
mooth analysis, which will be useful in dealing with uniqueness and regularity results for
viscosity solutions to partial di↵erential equations.
In the sequel we consider a convex open subset ⌦ of Rn and a convex function u : ⌦ !
R. Recall that u is convex if

u (1 t)x + ty  (1 t)u(x) + tu(y) 8x, y 2 ⌦, t 2 [0, 1] .

If u 2 C 2 (⌦) this is equivalent to say that r2 u(x) 0, in the sense of symmetric operators,
for all x 2 ⌦.

Definition 19.1 (Subdi↵erential). For each x 2 ⌦, the subdi↵erential @u(x) is the set

@u(x) := {p 2 Rn |u(y) u(x) + hp, y xi 8 y 2 ⌦} .

Obviously @u(x) = {ru(x)} at any di↵erentiability point.

Remark 19.2. According to Definition 19.1, it is easy to show that

u(x + tv) u(x)

@u(x) = {p 2 Rn | lim inf hp, vi 8v 2 Rn } . (19.1)
+ t!0 t
Indeed, when p 2 @u(x) the relation

u(x + tv) u(x)

hp, vi
t
passes through the limit. Conversely, let us recall the monotonicity property of di↵erence
quotients of a convex function, i.e.

u(x + t0 v) u(x) (1 t0 /t) u(x) + (t0 /t)u(x + tv) u(x) u(x + tv) u(x)
 = ,
t0 t0 t
(19.2)
for any 0 < t0 < t. Hence, for every y 2 ⌦, we have (choosing t = 1, v = y x)

u(x + t0 v) u(x) o(t0 )

u(y) u(x) hp, y xi + .
t0 t0
The same monotonicity property (19.2) yields that the lim inf in (19.1) is a limit.

Remark 19.3. The following properties are easy to check:

124
 (i) The graph of the subdi↵erential, i.e. {(x, p)|p 2 @u(x)} ⇢ ⌦ ⇥ Rn , is closed, in
fact convex functions are continuous (suffices, by (ii) below, to show that they are
locally bounded to obtain even local Lipschitz continuity).

(ii) Convex functions are locally Lipschitz in ⌦; to see this, fix a point x0 2 ⌦ and
x, y 2 Br (x0 ) b BR (x0 ) b ⌦. Thanks to the monotonicity of di↵erence quotients
seen in (19.2), we can estimate

u(y) u(x) u(yR ) u(x) osc(u, B R (x0 ))

  ,
|y x| |yR x| R r

where yR 2 @BR (x0 ) is on the halfline starting from x and containing y. Reversing
the roles of x and y we get

osc(u, B R (x0 ))
Lip(u, Br (x0 ))  .
R r
This proves the local Lipschitz continuity and we can use this information to replace
B r (x0 ) by BR (x0 ), or even Br (x0 ) by B r (x0 ) in the inequality above. Equivalently

osc(u, BR (x0 ))
ess sup |ru|  ,
Br (x0 ) R r

because of (1.6).

(iii) As a consequence of (ii) and Rademacher’s Theorem, @u(x) 6= ; for all x 2 ⌦. In

addition, a convex function u belongs to C 1 if and only if @u(x) is a singleton for
every x 2 ⌦. Indeed, if {xh } are di↵erentiability points of u such that xh ! x and
ru(xh ) has at least two distinct limit points, then @u(x) is not a singleton. Hence
ru has a continuous extension to the whole of ⌦ and u 2 C 1 .

(iv) Given convex functions fk : ⌦ ! R, locally uniformly converging in ⌦ to f , and

xk ! x 2 ⌦, any sequence (pk ) with pk 2 @fk (xk ) is bounded (by the local Lipschitz
condition) and any limit point p of (pk ) satisfies

p 2 @f (x) .

In fact, it suffices to pass to the limit as k ! 1 in the inequalities

fk (y) fk (xk ) + hpk , y xk i 8y 2 ⌦ .

As a first result of nonsmooth analysis, we state the following theorem.

125
Theorem 19.4 (Nonsmooth mean value theorem). Consider a convex function f : ⌦ ! R
and a couple of points x, y 2 ⌦. There exist z in the closed segment between x and y and
p 2 @f (z) such that
f (x) f (y) = hp, x yi .
Proof. Choose a positive convolution kernel ⇢ with support contained in B 1 and define
the sequence of functions f" := f ⇤ ⇢" , which are easily seen to be convex in the set ⌦" in
(1.3), because
Z
f" ((1 t)x + ty) = f ((1 t)x + ty "⇠)⇢(⇠) d⇠
Z ⌦

 ((1 t)f (x "⇠) + tf (y "⇠)) ⇢(⇠) d⇠

⌦
= (1 t)f" (x) + tf" (y) ;

moreover f" ! f locally uniformly. Thanks to the classical mean value theorem for
regular functions, for every " > 0 there exists z" = (1 ✓" )x + ✓" y, with ✓" 2 (0, 1), such
that
f" (x) f" (y) = hp" , x yi .
with p" = rf" (z" ) 2 @f" (z" ). Since (z" , p" ) are uniformly bounded as " ! 0, we can find
"k ! 0 with ✓"k ! ✓ 2 [0, 1] and p"k ! p. Remark 19.3(iv) allows us to conclude that
p 2 @f ((1 ✓)x + ✓y) and
f (x) f (y) = hp, x yi .
⇤
As an application of the nonsmooth mean value theorem, we can derive a pointwise
version of Remark 19.3(iii). Notice that we will follow a similar idea to achieve second
order di↵erentiability.

Proposition 19.5. If f : ⌦ ! R is convex, then f is di↵erentiable at x 2 ⌦ if and only

if @f (x) is a singleton. If this is the case, @f (x) = {rf (x)}.
Proof. One implication is trivial. For the other one, assume that @f (x) = {p} and notice
that closure of the graph of @f and the local Lipschitz property of f give that xh ! x
and ph 2 @f (xh ) imply ph ! p. Then, the nonsmooth mean value theorem gives

f (y) f (x) = hpxy , y xi = hp, y xi + hpxy , x yi = hp, y xi + o(|y x|) .

126
Remark 19.6. Recall that a continuous function f : ⌦ ! R is convex if and only if its
Hessian r2 f is non-negative, i.e. for every non-negative ' 2 Cc1 (⌦) and every ⇠ 2 Rn
there holds Z
@ 2'
f (x) 2 (x) dx 0 .
⌦ @⇠
This result is easily obtained by approximation by convolution, because, still in the weak
sense,
r2 (f ⇤ ⇢" ) = r2 f ⇤ ⇢" .
Although we shall not need this fact in the sequel, except in Remark 19.17, let us
mention, for completeness, that the positivity condition on the weak derivative r2 f im-
plies that this derivative is representable by a symmetric matrix-valued measure. To see
this, it suffices to apply the following result to the second derivatives r2⇠⇠ f :
Lemma 19.7. Consider a positive distribution T 2 D 0 (⌦), i.e.

8 ' 2 Cc1 (⌦), ' 0 =) hT, 'i 0.

Then there exists a locally finite non-negative measure µ in ⌦ such that

Z
hT, i = dµ 8 2 Cc1 (⌦) .
⌦

Proof. Fix an open set ⌦0 b ⌦, define K := ⌦0 and choose a non-negative cut-o↵ function
' 2 Cc1 (⌦) with '|K ⌘ 1. For every test function 2 Cc1 (⌦0 ), since (k kL1 ' ) 0
and T is a positive distribution, we have

hT, i  hT, k kL1 'i = C(⌦0 )k kL1 ,

where C(⌦0 ) := hT, 'i. Replacing by , the same estimate holds with |hT, i| in the
left hand side. By Riesz representation theorem we obtain the existence of µ. ⇤

Definition 19.8 ( -convexity, uniform convexity, semiconvexity). Given 2 R, we say

that a function f : ⌦ ! R is -convex if
Z Z
@ 2'
f (x) 2 (x) dx '(x) dx
⌦ @⇠ ⌦

for every non-negative ' 2 Cc1 (⌦) and for every ⇠ 2 Rn (in short r2 f I). We say
also that
• f is uniformly convex if > 0;

• f is semiconvex if 0.

127
 Notice that, with the notation of Definition 19.8, a function f is -convex if and only
if f (x) |x|2 /2 is convex.
Analogous concepts can be given in the concave case, namely -concavity (i.e. r2 f 
I), uniform concavity, semiconcavity. An important class of semiconcave functions is
given by squared distance functions:
Example 19.9. Given a closed set E ⇢ Rn , the square of the distance from E is 2-
concave. In fact,

dist2 (x, E) |x|2 = inf (x y)2 |x|2 = inf |y|2 2hx, yi ; (19.3)
y2E y2E

since the functions x 7! |y|2 2hx, yi are affine, their infimum over y 2 E, that is (19.3),
is concave.
Particularly in the duality theory of convex functions, it is useful to extend the concept
and convexity to functions f : Rn ! R[{+1}. The concept of subdi↵erential at points x
where f (x) < 1, extends immediately and, in the interior of the convex set {f < 1}, we
recover all the properties stated before (mean value theorem, local Lipschitz continuity).
Conversely, given f : ⌦ ! R convex with ⌦ convex, a canonical extension f˜ of f to the
whole of Rn is n o
f˜(x) := inf lim inf f (xh ) : xh 2 ⌦, xh ! x .
h!1

It provides a convex and lower semicontinuous extension of f , equal to +1 on Rn \⌦. For

these reasons, in the sequel we will consider convex and lower semicontinuous functions
f : Rn ! R [ {+1}. Notice that also the notion of -convexity extends, just requiring
that f (x) |x|2 /2 is convex.
Proposition 19.10. Given a convex lower semicontinuous function f : Rn ! R [ {+1},
its subdi↵erential @f satisfies for all x, y 2 {f < 1} the monotonicity property:

hp q, x yi 0 8p 2 @f (x), 8q 2 @f (y).
Proof. It is sufficient to sum the inequalities satisfied, respectively, by p and q, i.e.

f (y) f (x) hp, y xi

f (x) f (y) hq, x yi.

Remark 19.11 (Inverse of the subdi↵erential). (i) If f : Rn ! R[{+1} is -convex,

Proposition 19.10 proves that for every p 2 @f (x) and every q 2 @f (y), we have

hp q, x yi |x y|2 . (19.4)

128
 (ii) If > 0, for every p 2 Rn no more than one x 2 {f < 1} can satisfy p 2 @f (x),
because, through (19.4), we get

p 2 @f (x) \ @f (y) =) 0 = hp p, x yi |x y|2 =) x = y .

In particular, setting [
L := @f (x) ,
f (x)<1

there exists a single-valued and onto map (@f ) 1 : L ! {x : @f (x) 6= ;} such that
p 2 @f ((@f ) 1 (p)). In addition, L = Rn : given p, to find x such that p 2 @f (x) it
suffices to minimize y 7! f (y) hp, yi and to take x as the (unique) minimum point.
1 1
(iii) Moreover, (@f ) is a Lipschitz map: rewriting (19.4) for (@f ) we get

|(@f ) 1 (p) (@f ) 1 (q)|2  hp q, (@f ) 1 (p) (@f ) 1 (q)i

 |p q||(@f ) 1 (p) (@f ) 1 (q)| ,

thus Lip((@f ) 1 )  1/ .
The conjugate of a function f : Rn ! R [ {+1}, not identically equal to +1, is
defined as
f ⇤ (x⇤ ) := sup hx⇤ , xi f (x) ;
x2Rn

we immediately point out that f ⇤ is convex and lower semicontinuous, because it is the
supremum of a family of affine functions. The assumption that f (x) < 1 for at least one
x ensures that f ⇤ : Rn ! R [ {+1}. Equivalently, f ⇤ is the smallest function satisfying

hx, yi  f (x) + f ⇤ (y) 8x, y 2 Rn . (19.5)

A similar “variational” characterization of the subdi↵erential is that x⇤ 2 @f (x) if and

only if z 7! hx⇤ , zi f (z) attains its maximum at z = x, so that:

x⇤ 2 @f (x) () f ⇤ (x⇤ ) = hx⇤ , xi f (x) . (19.6)

Theorem 19.12. Any convex lower semicontinuous function f : Rn ! R [ {+1} not

identically equal to +1 is representable as g ⇤ for some g : Rn ! R[{+1} not identically
equal to +1.
Proof. If f (x0 ) < 1 we can use Hahn-Banach theorem in Rn+1 (with a small open ball
centered at {(x0 , f (x0 ) 1)} and the hypograph of f , which is a convex set) to find an
affine function `(x) = hp, xi + c such that `  f . This yields immediately f ⇤ (p) < 1, so
that (f ⇤ )⇤ makes sense. Now, the variational characterization of the conjugate function
based on (19.5) gives that (f ⇤ )⇤  f . On the other hand, the operator g 7! (g ⇤ )⇤ is

129
order-preserving and coincides, as it is easily seen, with the identity on affine functions
`(x) = hp, xi+c (notice that `⇤ is finite only at x⇤ = p and `⇤ (p) = c). Since convex lower
semicontinuous functions are supremum of affine functions (again as an application of the
Hahn-Banach theorem), these two facts yield (f ⇤ )⇤ f on convex lower semicontinuous
functions, completing the proof. ⇤
A byproduct of the previous proof is that (f ⇤ )⇤ = f in the class of convex and lower
semicontinuous functions f : Rn ! R [ {+1}, not identically equal to +1. This way
(19.5) becomes completely symmetric and it is easily seen that (19.6) gives

x 2 @f ⇤ (x⇤ ) () x⇤ 2 @f (x) . (19.7)

In particular, in the case when f is -convex for some > 0, from the quadratic
growth of f we obtain that f ⇤ is finite and that @f ⇤ = (@f ) 1 is single-valued and
Lipschitz, therefore f ⇤ 2 C 1,1 (Rn ).

19.2 Convex functions and Measure Theory

Now we recall some classical results in Measure Theory, in order to have the necessary
tools to prove Alexandrov theorem 19.16 on di↵erentiability of convex functions.
Thanks to the next classical result we can, with a slight abuse of notation, keep the
same notation rf for the pointwise gradient and the weak derivative, at least for locally
Lipschitz functions.

Theorem 19.13 (Rademacher). Any Lipschitz function f : Rn ! R is di↵erentiable

at L n -almost every point and the pointwise gradient rf coincides L n -a.e. with the
distributional derivative rf .
Proof. Fix a point x0 which is a Lebesgue point of rf , i.e.
Z
r!0
|rf (y) rf (x0 )| dy ! 0 . (19.8)
Br (x0 )

Defining
1
fr (y) :=(f (x0 + ry) f (x0 ))
r
and noticing that rfr (y) = rf (x0 + ry) (still in the distributional sense), we are able to
rewrite (19.8) as Z
r!0
|rfr (y) rf (x0 )| dy ! 0,
B1 (0)

where (fr ) is a sequence of functions with equibounded Lipschitz constant and fr (0) = 0
for every r > 0. Thanks to the Ascoli-Arzelà theorem, as r # 0, this family of functions

130
has limit points in the uniform topology. Any limit point g obviously satisfies g(0) = 0,
and since rg is a limit point of rfr in the weak⇤ topology, the strong convergence of rfr
to rf (x0 ) gives rg ⌘ rf (x0 ), still in the weak sense. We conclude that g(x) = rf (x0 )x,
so that g is uniquely determined and
1 r!0
fr (y) = (f (x0 + ry) f (x0 )) ! rf (x0 )y
r
uniformly in B 1 (0). This convergence property is immediately seen to be equivalent to
the classical di↵erentiability of f at x0 , with gradient equal to rf (x0 ). ⇤
The proof of the following classical result can be found, for instance, in [11] and [12].

Theorem 19.14 (Area formula). Consider a locally Lipschitz function f : Rn ! Rn and

a Borel set A ⇢ Rn . Then the function
1
N (y, A) := card f (y) \ A

is L n -measurable6 and
Z Z
| det rf (x)| dx = N (y, A) dy L n (f (A)) .
A Rn

Definition 19.15 (Pointwise second order di↵erentiability). Let ⌦ ⇢ Rn be open and

x 2 ⌦. A function f : ⌦ ! R is pointwise second order di↵erentiable at x if there exist
p 2 Rn and S 2 Symn⇥n such that
1
f (y) = f (x) + hp, y xi + hS(y x), y xi + o(|y x|2 ) .
2
Notice that pointwise second order di↵erentiability implies first-order di↵erentiability,
and that p = rf (x) (here understood in the pointwise sense). Also, the symmetry
assumption on S is not restrictive, since in the formula S can also be replaced by its
symmetric part.
We are now ready to prove the main result of this section, Alexandrov theorem.

Theorem 19.16 (Alexandrov). Any convex function f : Rn ! R [ {+1} is L n -a.e.

pointwise second order di↵erentiable in the interior of {f < 1}.
Proof. The proof is based on the inverse function = (@f ) 1 , introduced in Re-
mark 19.11. Obviously, there is no loss of generality supposing that f is -convex for
some > 0.
6
In particular, notice that f (A) = {N > 0}.

131
 We briefly recall, from Remark 19.11, that @f associates to each x 2 Rn the subdif-
ferential set, on the contrary is a single-valued map which associates to each p 2 Rn
the point x such that p 2 @f (x). Let us define the set of “bad” points
⌃ := {p | @r (p) or 9r (p) and det r (p) = 0} .
Since is Lipschitz, Rademacher Theorem 19.13 and the area formula 19.14 give
Z
n
L ( (⌃))  | det r | dp = 0 .
⌃

We shall prove that the stated di↵erentiability property holds at all points x 2 / (⌃).
Let us write x = (p) with p 2 / ⌃, so that rf (x) = p, there exists the derivative r (p)
and, since it is invertible, we can name
1
S(x) := (r (p)) .
If y = (q), we get
1
S(x) (q p S(x)(y x)) = (y x r (p)(q p))
= ( (q) (p) r (p)(q p))
= o(|p q|) = o(|x y|) .
Therefore
|q rf (x) S(x)(y x)|
lim =0. (19.9)
y!x
q2@f (y)
|y x|

The result got in (19.9), together with the nonsmooth mean value Theorem 19.4, give
us the second order expansion. In fact, let
1
f˜(y) := f (y) f (x) hrf (x), (y x)i hS(x)(y x), (y x)i .
2
Since
@ f˜(y) = @f (y) rf (x) S(x)(y x)
we can read (19.9) as lim |q|/|y x| = 0. Now, choose ✓ 2 [0, 1] and a vector
q2@ f˜(x), y!x
q 2 @ f˜((1 ✓)y + ✓x) such that f˜(y) = hq, y xi (since f˜(x) = 0) to find
f˜(y) = hq, y xi = o(|y x|2 ) .
By the very definition of f˜, the statement follows. ⇤

Remark 19.17 (Characterization of S). A blow-up analysis, analogous to the one per-
formed in the proof of Rademacher’s theorem, shows that the matrix S(x) in Alexandrov’s
theorem is the density of the measure r2 f with respect to L n , see [2] for details.

132
20 Viscosity solutions
20.1 Basic definitions
In this section we want to give the notion of viscosity solution for general equations having
the form
E(x, u(x), ru(x), r2 u(x)) = 0 (20.1)
where u is defined on some locally compact domain A ⇢ Rn . This topological assumptions
is actually very useful, because we can deal at the same time with open and closed domains,
and also domain of the form Rn 1 ⇥ [0, 1), which typically occur in parabolic problems.
We first need to recall two classical ways to regularize a function.

Definition 20.1 (u.s.c. and l.s.c. regularizations). Let A0 ⇢ A be a dense subset and
u : A0 ! R. We define its upper regularization u⇤ on A by one of the following equivalent
formulas:
⇢
u (x) := sup lim sup u(xh ) | (xh ) ⇢ A0 , xh ! x
⇤
h
= inf sup u
r>0 B (x)\A0
r

= min {v | v is u.s.c. and v u} .

Similarly we can define the lower regularization u⇤

n o
0
u⇤ (x) := inf lim inf u(xh ) | (xh ) ⇢ A , xh ! x
h
= sup inf u
r>0 Br (x)\A0
= max {v | v is l.s.c. and v  u }

which is also characterized by the identity u⇤ = ( u)⇤ .

Remark 20.2. It is clear that pointwise u⇤  u  u⇤ . In fact, u is continuous at a point

x 2 A (or, more precisely, it has a continuous extension in case x 2 A \ A0 ) if and only if
u⇤ (x) = u⇤ (x).

We now assume that E : L ⇢ A ⇥ R ⇥ Rn ⇥ Symn⇥n ! R, with L dense. Here and in

the sequel we denote by Symn⇥n the space of symmetric n ⇥ n matrices.

Definition 20.3 (Subsolution). A function u : A ! R is a subsolution for the equation

(20.1) (and we write E  0) if the two following conditions hold:

(i) u⇤ is a real-valued function;

133
 (ii) for any x 2 A, if ' is C 1 in a neighbourhood of x and u⇤ ' has a local maximum
at x, then
E⇤ (x, u⇤ (x), r'(x), r2 '(x))  0 . (20.2)

It is obvious from the definition that the property of being a subsolution is invariant
under u.s.c. regularization, i.e. u is a subsolution if and only if u⇤ is a subsolution.
The geometric idea in this definition is to use a local comparison principle, since
assuming that u⇤ ' has a maximum at x implies, if u is smooth, that ru⇤ (x) = r'(x)
and r2 u⇤ (x)  r2 '(x). So, while in the classical theory of PDEs an integration by parts
formula allows to transfer derivatives from u to the test function ', here the comparison
principle allows to transfer (to some extent, since only an inequality holds for second order
derivatives) the derivatives from u to the test function '.
Similarly, we give the following:
Definition 20.4 (Supersolution). A function u : A ! R is a supersolution for the
equation (20.1) (and we write E 0) if the two following conditions hold:
(i) u⇤ is a real-valued function;

(ii) for any x 2 A, if ' is C 1 in a neighbourhood of x and u⇤ ' has a local minimum
at x, then
E ⇤ (x, u⇤ (x), r'(x), r2 '(x)) 0. (20.3)

We finally say that u is a solution of our problem if it is both a subsolution and a

supersolution.
Remark 20.5. Without loss of generality, we can always assume in the definition of
subsolution that the value of the local maximum is zero, that is u⇤ (x) '(x) = 0. This is
true because the test function ' is arbitrary and the value of ' at x does not appear in
(20.2). Also, possibly subtracting |y x|4 to ' (so that first and second derivatives of ' at
x remain unchanged), we can assume with no loss of generality that the local maximum
is strict. Analogous remarks hold for supersolutions.
Remark 20.6. A trivial example of viscosity solution is given by the Dirichlet function
Q on R, which is easily seen to be a solution to the equation u = 0 in the sense above.
0

This example shows that some continuity assumption is needed, in order to hope for
reasonable existence and uniqueness results.
Remark 20.7. Rather surprisingly, a solution of E = 0 in the viscosity sense does not
necessarily solve E = 0 in the viscosity sense. To show this, consider the equations
|f 0 | 1 = 0 and 1 |f 0 | = 0 and the function f (t) = min {1 t, 1 + t} . In this case, it
is immediate to see that f is a subsolution of the first problem (and actually a solution,
as we will see), but it is not a subsolution of the second problem, since we can choose

134
identically ' = 1 to find that the condition 1 |'0 (0)|  0, corresponding to (20.2), is
violated.
We have instead the following parity properties:
(a) Let E be odd in (u, p, S). If u verifies E  0, then u verifies E 0.
(b) Let E be even in (u, p, S). If u verifies E  0, then u verifies E 0.
We now spend some words on the ways of simplifying the conditions that have to be
checked in order prove the subsolution or supersolution property. We just examine the
case of subsolutions, the case of supersolutions being the same (with obvious variants).
We have already seen in Remark 20.5 that we can assume without loss of generality
that u⇤ ' has a strict local maximum, equal to 0, at x. We can also work equivalently with
the larger class of C 2 functions ', in a neighbourhood of x. One implication is trivial, let us
see the converse one. Let ' 2 C 2 and assume u⇤ (y) '(y)  0 for y 2 B r (x), with equality
only when y = x. By appropriate mollifiers, we can build a sequence ('k ) ⇢ C 1 (B r (x))
with 'k ! ' in C 2 (B r (x)). Let then xk be a maximum in B r (x) of the function u⇤ 'k .
Since 'k ! ' uniformly, it is easy to see that any limit point of (xk ) has to be a maximum
for u⇤ ', hence it must be x; in addition the convergence of the maximal values yields
u⇤ (xk ) ! u⇤ (x). The subsolution property, applied with 'k at xk , gives
E⇤ (xk , u⇤ (xk ), r'k (xk ), r2 'k (xk ))  0
and we can now let k ! 1 and use the lower semicontinuity of E⇤ to get the thesis.
Actually, it is rather easy now to see that the subsolution property is even equivalent
to
E⇤ (x, u⇤ (x), p, S)  0 8 (p, S) 2 J2+ u⇤ (x)
where J2+ u⇤ is the second-order super jet of u, namely
J2+ u⇤ (x) := (p, S) u⇤ (y)  u⇤ (x) + hp, y xi + 12 hS(y x), y xi + o(|y x|2 ) .
Indeed, let P (y) := u⇤ (x)+hp, y xi+ 12 hS(y x), y xi, so that u⇤ (y)  P (y)+o(|y x|2 ),
with equality when y = x. Hence, for any " > 0 we have u⇤ (y)  P (y) + "|y x|2 on a
sufficiently small neighbourhood of x with equality at y = x and we can apply (20.2) to
this smooth function to get
E⇤ (x, u⇤ (x), p, S + 2"I) = E⇤ (x, u⇤ (x), rP (x), r2 P (x) + 2"I)  0
and by lower semicontinuity we can let " ! 0 and prove the claim. Of course, if we are
dealing with first order equations, only the first order super jet is needed.
Remark 20.8. After these preliminary facts, it should be clear that this theory, despite
its elegance, has two main restrictions: on the one hand it is only suited to first or second
order equations (since no information on third derivatives comes from local comparison),
on the other hand it cannot be generalized to vector-valued functions.

135
20.2 Viscosity versus classical solutions
We first observe that a classical solution is not always a viscosity solution. To see this,
consider on R the problem u00 2 = 0. The function f (t) = t2 is clearly a classical solution,
but it is not a viscosity solution, because it is not a viscosity supersolution (take ' ⌘ 0
and study the situation at the origin).
Since we can always take u = ' if u is at least C 2 , the following theorem is trivial:

Theorem 20.9 (C 2 viscosity solutions are classical solutions). Let ⌦ ⇢ Rn be open,

u 2 C 2 (⌦) and E continuous. If u is a viscosity solution of (20.1) on ⌦, then it is also a
classical solution of the same problem.

The converse holds if S 7! E⇤ (x, u, p, S) and S 7! E ⇤ (x, u, p, S) are non-increasing in

Symn⇥n :

Theorem 20.10 (Classical solutions are viscosity solutions). If u is a classical subsolution

(resp. supersolution) of (20.1), then it is also a viscosity subsolution (resp. supersolu-
tion) of the same problem whenever E⇤ (x, u, p, ·) (resp. E ⇤ (x, u, p, ·)) is non-increasing in
Symn⇥n .
Proof. We just study the case of subsolutions. For a test function ', if u ' has a
local maximum at a point x then we know by elementary calculus that ru(x) = r'(x)
and r2 u(x)  r2 '(x) and by definition E⇤ (x, u(x), ru(x), r2 u(x))  0. Consequently,
exploiting our monotonicity assumption we obtain E⇤ (x, u(x), r'(x), r2 '(x))  0 and
the conclusion follows. ⇤
Before going further, we need to spend some words on conventions. First of all, it
should be clear that this theory also applies to parabolic equations such as (@t )u g = 0
if we let x := (y, t) 2 Rn ⇥ (0, 1) with A = Rn ⇥ (0, 1) . Secondly, it is worth remarking
that some authors adopt a di↵erent convention, which we might call elliptic convention,
which is “opposite” to the one we gave before. Indeed, according to this convention, if (for
instance) we deal with a problem of the form F (r2 u) = 0, we require for a subsolution that
u⇤ ' has a maximum at x implies F (r2 '(x)) 0 (i.e. a subsolution of F (r2 u) = 0
in our terminology). As a consequence, in the previous theorem, we should replace “non-
increasing” with “non-decreasing.”
Now, we are ready to introduce the first important tool for the following theorems.

Theorem 20.11. Let F be a family of subsolutions of (20.1) in A and let u : A ! R be

defined by
u(x) := sup {v(x) | v 2 F } .
Then u is a subsolution of the same problem on the domain A\{u⇤ < 1} (since {u⇤ < 1}
is open, the domain is still locally compact).

136
Proof. Assume as usual that u⇤ ' has a strict local maximum at x, equal to 0, and
denote by K the compact set B r (x) \ A for some r to be chosen sufficiently small, so that
x is the unique maximum of u ⇤ ' on K.
By a diagonal argument can find a sequence (xh ) inside K, convergent to x, and a
sequence of functions (vh ) ⇢ F such that u⇤ (x) = limh u(xh ) = limh vh (xh ). Hence, if we
call yh the maximum of vh⇤ ' on K, then
u⇤ (yh ) '(yh ) vh⇤ (yh ) '(yh ) vh⇤ (xh ) '(xh ) vh (xh ) '(xh ).
Since by our construction we have vh (xh ) '(xh ) ! 0 for h ! 1, we get that every limit
point y of (yh ) satisfies
u⇤ (y) '(y) 0.
Hence y is a maximum in K of u⇤ ', u⇤ (y) '(y) = 0 and y must coincide with x.
Consequently yh ! x, lim suph (u⇤ (yh ) '(yh )  u⇤ (x) '(x) and, by comparison, the
same is true for the intermediate terms, so that vh⇤ (yh ) ! u⇤ (x). In order to conclude, we
just need to consider the viscosity condition at the points yh , which reads
E⇤ (yh , vh⇤ (yh ), r'(yh ), r2 '(yh ))  0 ,
and let h ! 1 to get
E⇤ (x, u⇤ (x), r'(x), r2 '(x))  0.
⇤
We can now state a first existence result.
Theorem 20.12 (Perron). Let f and g be respectively a subsolution and a supersolution
of (20.1), such that f⇤ > 1 and g ⇤ < +1 on A. If f  g on A and the functions
E⇤ (x, u, p, ·) and E ⇤ (x, u, p, ·) are non-increasing, then there exists a solution u of (20.1)
satisfying f  u  g.
Proof. Call
F := {v | v is a subsolution of (20.1) and v  g } .
We know that f 2 F, so that this set is not empty. Hence, we can define u :=
sup {v| v 2 F} . By our definition of F, we have that u  g and therefore u⇤  g ⇤ < +1.
Since u⇤ u⇤ f⇤ > 1, in A, by Theorem 20.11 u is a subsolution on A. Consequently,
we just need to prove that it is also a supersolution on the same domain.
Pick a test function ' such that u⇤ ' has a relative minimum, equal to 0, at x0 .
Without loss of generality, we can assume that
u⇤ (x) '(x) |x x0 | 4 on A \ B r (x) (20.4)
for some sufficiently small r > 0. Assume by contradiction that
E ⇤ (x0 , u⇤ (x0 ), r'(x0 ), r2 '(x0 )) < 0 (20.5)
4
and define a function w := max{' + , u} for some parameter > 0. We claim that:

137
 (a) w is a subsolution of (20.1);

(b) {w > u} =
6 ;;

(c) w  g (and hence w 2 F),

provided we choose sufficiently small in (a) and (c).
It is easily proved, again by contradiction and exploiting the fact that E ⇤ is upper
semicontinuous, that for > 0 sufficiently small we have

E ⇤ (x, '(x) + 4
, r'(x), r2 '(x))  0 on B 2 (x0 ) \ A .

This means that '+ 4 is a classical subsolution of (20.1) on this domain and hence, by our
monotonicity hypothesis, it has to be also a viscosity subsolution. Consequently, by a very
special case of the previous theorem, we get that the function w is a viscosity subsolution
of (20.1) on B 2 (x0 )\A. Moreover, by (20.4), we know that w = u on (A \ Br (x))\B (x0 ).
Since the notions of viscosity subsolution and supersolution are clearly local, w is a global
subsolution on A.7
To prove that {w > u} = 6 ; we just need to observe that, for any > 0, u⇤ (x0 ) =
4
'(x0 ) < '(x0 ) + , and on any sequence (xh ) such that u(xh ) ! u⇤ (x0 ), we must have
for h sufficiently large the inequality u(xh ) < '(xh ) + 4 .
Finally, we have to show that w  g: this completes the proof of the claim and gives the
desired contradiction. To this aim, it is enough to prove that there exists > 0 such that
' + 4  g on A \ B (x0 ). But this readily follows, by an elementary argument, showing
that '(x0 ) = u⇤ (x0 ) < g⇤ (x0 ). Again, assume by contradiction that u⇤ (x0 ) = g⇤ (x0 ) : if
this were the case, the function g⇤ ' would have a local minimum at x0 and so, since
g⇤ is a viscosity supersolution, we would get

E ⇤ (x0 , g⇤ (x0 ), r'(x0 ), r2 '(x0 )) 0,

which is in contrast with (20.5). ⇤

20.3 The distance function

Our next goal is now to study the uniqueness problem, which is actually very delicate as
the previous examples show. We begin here with a special case.
Let C ⇢ Rn be a closed set, C 6= ; and let u(x) := dist(x, C). We claim that the
distance function is a viscosity solution of the equation |p|2 1 = 0 on A := Rn \ C.
First of all, it is clearly a viscosity supersolution in A. This follows by Theorem 20.11
(in the obvious version for supersolutions), once we observe that u(x) = inf y2C |x y| and
7
We mean that, if A = A1 [ A2 and we know that u is a subsolution both on A1 and A2 , relatively
open in A, then it is also a subsolution on A.

138
that, for any y 2 C, the function x 7! |x y| is a classical supersolution in A (because
y2/ A) and hence a viscosity supersolution of our problem.
The fact that u is also a subsolution follows by the general implication:

Lip(f )  1 ) |rf |2 10 in the sense of viscosity solutions.

Indeed, let x be a local maximum for f ', so that f (y) '(y)  f (x) '(x)
for any y 2 Br (x) (and r small enough). This is equivalent, on the same domain, to
'(y) '(x) f (y) f (x) |y x| and, by the Taylor expansion, we finally get

hr'(x), y xi + o(|y x|) |y x| .

This readily implies the claim.

The converse implication is less trivial, but still true! Namely

|rf |2 1  0 in the sense of viscosity solutions ) Lip(f )  1

for f continuous (or at least upper semicontinuous), which is proved by means of the
regularizations f " (x) := supy f (y) |x y|2 /" that we will study more in detail later
on. We just sketch here the structure of the argument:

(1) still |rf " |2 1  0 in the sense of viscosity solutions;

(2) |rf " |2 1  0 pointwise L n -a.e., because f " is semiconcave, hence locally Lipschitz,
and therefore the inequality holds at any di↵erentiaiblity point by the super-jet
characterization of viscosity subsolutions;

(3) by Proposition 1.4 one obtains Lip(f " )  1;

(4) f " # f and hence Lip(f )  1.

We now come to our uniqueness result.

Theorem 20.13. Let C ⇢ Rn be a closed set as above, A = Rn \ C and let u 2 C(A) be

a non-negative viscosity solution of |p|2 1 = 0 on A with u = 0 on @A. Then C 6= ; and
u(x) = dist(x, C).
Proof. By our assumptions we can clearly extend u continuously to Rn , so that u = 0
identically on C. It is immediate to verify that |ru|2 1  0 in the sense of viscosity
solutions on Rn . Consequently, thanks to the previous regularization argument, Lip(u) 
1 and hence, for any y 2 C, we have that u(x)  |x y|, which means u(x)  dist(x, C).
In the sequel, in order to simplify the notation, we will write w(x) for the distance function
dist(x, C),

139
 It remains to show that w  u. Assume first that A is bounded: we will show later on
that this is not restrictive. By contradiction, assume that w(x0 ) > u(x0 ) for some x0 ; in
this case there exist 0 > 0 and 0 > 0 such that
⇢
1
sup w(x) (1 + )u(y) |x y|2 0
x,y 2"
for all " > 0 and 2 (0, 0 ). Indeed, it suffices to bound from below the supremum with
w(x0 ) (1 + )u(x0 ), which is larger than 0 := (w(x0 ) u(x0 ))/2 for > 0 small enough.
Moreover, for " > 0 and 2 (0, 0 ), the supremum is actually a maximum because it
is clear that we can localize x in A (otherwise the whole sum is non-positive) and y in a
bounded set of Rn (because w is bounded on A, and again for |y x| large the whole sum
is non-positive). So, call (x, y) a maximizing couple, omitting for notational simplicity the
dependence on the parameters ", . The function x 7! w(x) 2" 1
|x y|2 has a maximum
at x = x and so we can exploit the fact that w(·) is a viscosity solution of our equation
(with respect to the test function '(x) = |x y|2 /(2")) to derive |r'|2 (x)  1, that is
|x y|
1.
"
We also claim that necessarily y 2 A, if " is sufficiently small, precisely " < 0 . Indeed,
assume by contradiction that y 2/ A, so that w(y) = 0, then by the triangle inequality
1 1
0  w(x) |x y|2  |x y| |x y|2  |x y| .
2" 2"
As a consequence, we get 0  |x y|  ", which gives a contradiction.
Now, choosing " > 0 so that y 2 A, the function y 7! (1 + )u(y) + 1
2"
|x y|2 has a
minimum at y = y and arguing as above we obtain
x y
(1 + ) ,
"
which is not compatible with |x y|  ". Hence, at least when A is bounded, we have
proved that w = u.
In the general case, fix a constant R > 0 and define uR (x) := u(x) ^ dist(x, Rn \ B R ) :
this is a supersolution of our problem on A \ BR , since u(x) is a supersolution on A
and dist(x, Rn \ B R ) is a supersolution on BR (by the infimum property). Moreover,
Lip(uR )  1 implies that uR is a global subsolution and we can apply the previous result
(special case) to the function uR to get
uR (x) = d(x, Rn \ (A \ BR )).
Letting R ! 1 we first exclude C = ; since in that case uR " 1 which is not admissible
since uR  u and then (by C 6= ;) we obtain u(x) = dist(x, C). ⇤

140
Remark 20.14. We can also give a di↵erent interpretation of the result above. In the
spirit of the classical Liouville’s theorems we can say that “the equation |ru|2 1 = 0 does
not have entire viscosity solutions on Rn that are bounded from below”. Nevertheless,
there exist trivial examples of functions that solve this equation in the viscosity sense and
are unbounded from below (e.g. take u(x) = xi for some i 2 {1, . . . , n}).

20.4 Maximum principle for semiconvex functions

We now turn to the case of second order problems having the form F (ru, r2 u) = 0 on an
open domain A ⇢ Rn . We will always assume that F (p, S) is non-increasing in its second
variable S, so that classical solutions are viscosity solutions.
Let us begin with some heuristics. Let f, g 2 C 2 (A) \ C(A), with A bounded, and
assume that f is a subsolution on A, g is a supersolution on A, f  g on @A and that
one of the inequalities F (rf, r2 f )  0, F (rg, r2 g) 0 is always strict. Then f  g in
A. Indeed, assume by contradiction supA (f g) > 0, then there exists a x0 2 A which
is a maximum for f g. Consequently rf (x0 ) = rg(x0 ) and also r2 f (x0 )  r2 g(x0 ).
These two facts imply, by the monotonicity of F, that

F (rf (x0 ), r2 f (x0 )) F (rg(x0 ), r2 g(x0 )) . (20.6)

On the other hand, f (resp. g) is also a regular subsolution (resp. supersolution) so that

F (rf (x0 ), r2 f (x0 ))  0, F (rg(x0 ), r2 g(x0 )) 0. (20.7)

Hence, if we compare (20.6) with (20.7), we find a contradiction as soon as one of the two
inequalities in (20.7) is strict.
In order to hope for a comparison principle, this argument shows the necessity to
approximate subsolutions (or supersolutions) with strict subsolutions, and this is always
linked to some form of strict monotonicity of the equation, variable from case to case (of
course in the trivial case F ⌘ 0 no comparison principle is possible). To clarify this point,
let us consider the following example. Consider the space-time coordinates x = (y, t) and
a parabolic problem
F (ry,t u, r2y,t u) = @t u G(r2y u)
with G non-decreasing, in the appropriate sense. In this case, we can reduce ourselves to
strict inequalities by performing the transformation u e t u.
In order to get a general uniqueness result for viscosity solution, we cannot just argue
as in the case of the distance function and we need to follow a strategy introduced by
Jensen. The first step is to obtain a refined versions of the maximum principle. We start
with an elementary observation.

141
Remark 20.15. If (p, S) 2 J2 u(x) and u has a relative maximum at x, then necessarily
p = 0 and S  0. To see this, it is enough to apply the definitions: by our two hypotheses
1
0 u(y) u(x) hp, y xi + hS(y x), y xi + o(|y x|2 )
2
and hence
y x
hp, i  o(|y x|) ) p = 0 ,
|y x|
hS(y x), y xi
 o(1) ) S  0 .
|y x|2
We are now ready to state and prove Jensen’s maximum principle for semiconvex
functions.
Theorem 20.16 (Jensen’s maximum principle). Let u : ⌦ ! R be semiconvex and let
x0 2 ⌦ a local maximum for u. Then, there exist a sequence (xk ) converging to x0 and
"k # 0 such that u is pointwise second order di↵erentiable at xk and

ru(xk ) ! 0 r2 u(xk )  "k I .

The proof is based on the following lemma. In the sequel we shall denote by sc(u, ⌦)
the least nonnegative constant C such that u is ( C)-convex, i.e. u + C|x|2 /2 is convex
(recall Definition 19.8).
Theorem 20.17. Let B ⇢ Rn be a ball of radius R centered at the origin and u 2 C(B)
semiconvex, with8
max u > max u .
B @B

Then, if we let

G = x 2 B 9 p 2 B s.t. u(y)  u(x) + hp, x yi, 8y 2 B

it must be
!n n
L n (G ) (20.8)
[sc(u, B)]n
for 0 < < (maxB u minB u) /(2R).
Proof. We assume first that u is also in C 1 (B). Pick a > 0, so small that 2R <
maxB u max@B u, and consider a perturbation u(y) + hp, yi with |p|  . We claim that
such function necessarily attains its maximum in B. Indeed, this immediately comes from
the two inequalities
max (u + hp, yi)  max u + R
@B @B

8
Notice that this implies sc(u, B) > 0, since maxB = max@B for convex functions.

142
and
max(u + hp, yi) max u R.
B B

Consequently, there exists x 2 B such that ru(x) = p. This shows that ru(G ) = B .
To go further, we need the area formula. In this case, it gives
Z Z
2
| det r u| dx = card ({x | ru(x) = p }) dp !n n
G B

by the previous statement. On the other hand

Z
det r2 u dx  [sc(u, B)]n L n (G ) ,
G

because the points in G are maxima for the function u(y)+hp, yi : this implies r2 u(x)  0
for any x 2 G and, by semiconvexity, r2 u(x) sc(u, B)I. If we combine these two
inequalities, we get (20.8).
In the general case we argue by approximation, finding radii rh " R and smooth functions
uh in B rh such that uh ! u locally uniformly in B and lim suph sc(uh , Brh )  sc(u, B); to
conclude, it suffices to notice that any limit of points in G (uh ) \ Brh belongs to G (u),
hence L n (G (u)) lim suph L n (G (uh ) \ Brh ). ⇤
We can now prove Jensen’s maximum principle. As a preliminary remark, observe
that, in Definition 19.8 one has (for our u) = 0 then the claim is trivial, so that we can
without loss of generality assume that < 0 and Theorem 20.17 applies.
Proof. Let x0 be a local maximum of u. We can choose R > 0 sufficiently small so
that u  u(x0 ) in B R (x0 ) and, without loss of generality, we can assume u(x0 ) = 0.
This becomes a strict local maximum for the function u e(x) = u(x) |x x0 |4 . It is also
e is semiconvex in B R (x0 ). We now apply Theorem 20.17 to u
easy to verify that u e: for any
= 1/k with k large enough we obtain that L n (G1/k ) > 0 and (thanks to the Alexandrov
theorem) this means that there exists a sequence of points (xk ) such that u e is pointwise
second order di↵erentiable at xk and, for appropriate vectors pk with |pk |  1/k, the
function u e(y) hpk , yi has a local maximum at xk . Since |pk | ! 0, any limit point of
(xk ) for k ! 1 has to be a local maximum for u e, but in B R (x0 ) this necessarily implies
2
xk ! x0 . Moreover pk = re u(xk ) ! 0 and r u e(xk )  0. As a consequence
ru(xk ) = re
u(xk ) + 4|xk x0 |2 (xk x0 ) ! 0
and the identity
r2 |z|4 = 4|z|2 I + 8z ⌦ z (20.9)
gives
r2 u(xk ) = r2 u
e(xk ) + 8(xk x0 ) ⌦ (xk x0 ) + 4|xk x0 | 2 I
2
 ru e(xk ) + 12|xk x0 |2 I .
Setting "k = 12|xk x0 |2 we get the thesis. ⇤

143
 We now introduce another important tool in the theory of viscosity solutions.

Definition 20.18 (Inf and sup-convolutions). Given u : A ! R and a parameter " > 0,
we can build the regularized functions
⇢
" 1
u (x) := sup u(y) |x y|2 (20.10)
y2A "

which are called sup-convolutions of u and satisfy u" u, and

⇢
1
u" (x) := inf u(y) + |x y|2 . (20.11)
y2A "

which are called inf-convolutions of u and satisfy u"  u.

In the next proposition we summarize the main properties of sup-convolutions; anal-

ogous properties hold for inf-convolutions.

Proposition 20.19 (Properties of sup-convolutions). Assume that u is u.s.c. on A and

that u(x)  K(1 + |x|) for some constant K 0, then

(i) u" is semiconvex and sc(u" , Rn )  2/";

(ii) u" u and u" # u pointwise in A. If u is continuous, then u" # u locally uniformly;

(iii) if F (ru, r2 u)  0 in the sense of viscosity solutions on A, then F (ru" , r2 u" )  0

on A" , where

A" := {x 2 Rn | the supremum in (20.10) is attained } .

Proof. (i) First of all, notice that, by the linear growth assumption, the function u" is
real-valued for any " > 0. Moreover, by its very definition
✓ ◆
" 1 2 1 2 2
u (x) + |x| = sup u(y) |y| + hx, yi
" y2A " "

and the functions in the right hand side are affine with respect to x. It follows that the
left hand side is convex, which means sc(u" , Rn )  2/".
(ii) The inequality u" u and the monotonicity in " are trivial. In addition, we can take
quasi-maxima (y" ) satisfying
2 2 2
" " "
u" (x)  u(y" ) + "  K(1 + |y" |) + "  K(1 + |x| + | " |) +" .
" " "

144
with " = |y" x|. Via these two inequalities, one first sees that y" ! x so that, exploiting
the upper semicontinuity of u and neglecting the quadratic term in the first inequality we
get
u(x) lim sup u(y" ) lim sup u" (x) .
"!0 "!0

If u is continuous, the claim comes from Dini’s monotone convergence theorem and the
local compactness of A.
(iii) Let x0 2 A" and let y0 2 A be a corresponding maximum, so that u" (x0 ) = u(y0 )
|x0 y0 |2 /". Let then ' be a smooth function such that u" ' has a local maximum in
x0 and, without loss of generality, we can take u" (x0 ) = '(x0 ). Let us call r the radius
such that u"  ' on Br (x0 ).
Define (x) := '(x y0 + x0 ) : we claim that u has a local maximum at y0 with value
|x0 y0 |2 /". If we prove this claim, then it must be

F (r (y0 ), r2 (y0 ))  0

and, by the definition of , this is equivalent to

F (r'(x0 ), r2 '(x0 ))  0.

This is enough to prove the claim. On the one hand

1
u(y0 ) (y0 ) = u(y0 ) '(x0 ) = u(y0 ) u" (x0 ) = |x0 y0 | 2 ,
"
while on the other hand u" (x)  '(x) in Br (x0 ) gives
1
u(y) |x y|2  '(x) 8x 2 Br (x0 ), 8y 2 A
"
and, letting y = x x0 + y0 2 A with x 2 Br (x0 ), this implies
1
u(y) (y)  |x0 y0 | 2 8y 2 A \ Br (y0 ) .
"
⇤

Remark 20.20. We will also need an x-dependent version of the previous result, that
reads as follows: if F (x, ru, r2 u)  0 in the sense of viscosity solutions on A, then for
all > 0 there holds F (x, ru" , r2 u" )  0 on A" , where

A", := {x 2 Rn | the supremum in (20.10) is attained at some y 2 B (x) \ A} ,

F (x, p, S) := inf {F (y, p, s) : y 2 B (x) \ A} . (20.12)

An analogous result holds for supersolutions

145
20.5 Existence and uniqueness results
In this section we will collect some existence and uniqueness results for second order
equations. The main tool is the comparison principle, stated below. Throughout the
section we shall always assume that A is a bounded open set in Rn .
Proposition 20.21 (Comparison principle). Let F : A ⇥ Symn⇥n ! R be continuous and
satisfying, for some > 0, the strict monotonicity condition
F (x, S + tI) F (x, S) + t 8t 0
and the uniform continuity assumption
F (·, S), S 2 Symn⇥n , are equi-continuous in A.
Let u, u : A ! R be respectively a bounded u.s.c. subsolution and a bounded l.s.c. super-
solution to F (x, r2 u) = 0 in A, with (u)⇤  (u)⇤ on @A. Then u  u on A.
Notice that the uniform continuity assumption, though restrictive, covers equations of
the form G(r2 u) + f (x) with f continuous in A.
A direct consequence of the comparison principle (take u = u = u) is the following
uniqueness result:
Theorem 20.22 (Uniqueness of continuous solutions). Let F be as in Proposition 20.21
and h 2 C(@A). Then the problem
8
< F (x, r2 u(x)) = 0 in A;
(20.13)
:
u=h on @A

admits at most one viscosity solution u 2 C(A).

At the level of existence, we can exploit Theorem 20.12 to obtain the following result.
Theorem 20.23 (Existence of continuous solutions). Let F be as in Proposition 20.21
and let f, g : A ! R be respectively a subsolution and a supersolution of F (x, D2 u) = 0
in A, such that f⇤ > 1, g ⇤ < +1 and f  g on A. If g ⇤  f⇤ on @A, then there exists
a solution to (20.13) with h = g ⇤ = f⇤ .
In order to prove this last result, it suffices to take any solution u given by Perron’s
method (see Theorem 20.12), so that f  u  g in A. It follows that u⇤  g ⇤  f⇤  u⇤
on @A and the comparison principle (with u = u⇤ , u = u⇤ ) gives u⇤  u⇤ on A, i.e. u is
continuous.
The rest of the section will be devoted to the proof of the comparison principle, which
uses besides doubling of variables, inf and sup-convolutions (see Definition 20.18) and
Jensen’s maximum principle (see Theorem 20.16).

146
Lemma 20.24. Let F, u and u be as in Proposition 20.21 and set

F (x, S) := F (x, S I)  F (x, S) ,

with > 0. For any > 0, consider the function

v , := u + |x|2 .
2
Hence:

(i) v , solves F (x, r2 v , )  0 in the viscosity sense;

(ii) if ( , A) is large enough, then v ,  u on @A and ( , A) ! 0 as # 0.

(iii) if the comparison principle holds for v , for any > ( , A), that is

v, u on A, 8 > ( , A) , (20.14)

then u  u on A.
Proof. Statements (i) follows by the translation invariance w.r.t. u of the equation, and
by r2 v = r2 u + I. Statement (ii) follows by the fact that u < u on @A.
If (20.14) holds, then
u v, u on A ,
and the comparison principle for u follows letting # 0, which allows to choose arbitrarily
small in view of (ii). ⇤

Proof. (of Proposition 20.21) Thanks to Lemma 20.24, without loss of generality we can
assume that u satisfies the stronger property

F (x, r2 u)  0

in the viscosity sense, for some > 0.

Assume by contradiction that d0 := u(x0 ) u(x0 ) > 0 for some x0 2 A, and let us
consider the sup convolution
✓ ◆ ✓ ◆
" 0 1 0 2 ⇤ 0 1 0 2
u (x) := sup u(x ) |x x | = max (u) (x ) |x x | , (20.15)
x0 2A " x0 2A "

of u and the inf convolution

✓ ◆ ✓ ◆
1 1
u" (y) := inf u(y 0 ) + |y 0 2
y| 0
= min (u)⇤ (y ) + |y 0 2
y| , (20.16)
0
y 2A " y 0 2A "

147
of u; since u" u and u"  u we have
✓ ◆
" 1 4
max u (x) u" (y) |x y| u" (x0 ) u" (x0 ) u(x0 ) u(x0 ) = d0
A⇥A 4"
and we shall denote by (x" , y" ) 2 A ⇥ A a maximizing pair, so that
1
d0 + |x" y" |4  u" (x" ) u" (y" )  sup u inf u . (20.17)
4"
Also, we denote by x0" 2 A and y"0 2 A maximizers and minimizers respectively in
(20.15) and (20.16).
Now we claim that:
(a) lim inf dist(x" , @A) > 0 and lim inf dist(y" , @A) > 0;
"#0 "#0

(b) setting M = max{osc(i), osc(u)}, for " small enough, the supremum in (20.15) with
any x 2 A satisfying |x x" | < " is attained at a point x0 2 A with |x0 x|2  M "
and the infimum in (20.16) with any y 2 A satisfying |y y" | < " is attained at a
point y 0 2 A with |y 0 y|2  M ".
To prove (a), notice that, if (x̄, ȳ) is any limit point of (x" , y" ) as " # 0, then (20.17)
gives x̄ = ȳ and
✓ ◆
⇤ 0 0 |x" x0" |2 + |y" y"0 |2
d0  lim sup (u) (x" ) (u)⇤ (y" ) .
"#0 "
Since the supremum of (u)⇤ (u)⇤ is finite, this implies that |x" x0" | ! 0, |y" y"0 | ! 0,
hence (x0" , y"0 ) ! (x̄, x̄) as well and semicontinuity gives d0  (u)⇤ (x̄) (u)⇤ (x̄). By
assumption (u)⇤  (u)⇤ on @A, therefore x̄ 2 A and this proves (a).
To prove (b), it suffices to choose, thanks to (a), "0 > 0 and 0 > 0 small enough, so
that dist(x" , @A) 0 for " 2 (0, "0 ). In general, for x 2 A we have

1 0
u(x0 )|x x|2  u(x)  u" (x)
"
which implies that the supremum in the definition
p of u" (x) is unchanged if we maximize
in the ball B x centered at x with radius M ". If |x x" | < ✏ and " < "0 , since
dist(x" , @A) 0 , this implies that the ball B x is contained in A for " small enough,
hence the supremum is attained. The argument for y" is similar.
Let us fix " small enough so that (b) holds and both x0" and y"0 belong to A, and let
us apply Jensen’s maximum principle to the (locally) semiconvex9 function
1
w(x, y) := u" (x) u" (y) |x y|4
4"
9
The local semiconvexity of w follows from Proposition 20.19.

148
to find zn := (x",n , y",n ) ! (x" , y" ) and n # 0 such that w is pointwise second order
di↵erentiable at zn , rw(zn ) ! 0 and r2 w(zn )  n I. By statement (b) and Remark 20.20,
for n large enough we have

sup F (x, r2 u" (x",n ))  0, inf F (y",n , r2 u" (y",n )) 0. (20.18)

|x x",n |2 M " |y y",n |2 M "

On the other hand, the upper bound on r2 w(zn ) together with (20.9) give
⇢
r2 u" (x",n ) 2" (x",n y",n ) ⌦ (x",n y",n ) 1" |x",n y",n |2 I  n I ,
(20.19)
r2 u" (y",n ) 2" (x",n y",n ) ⌦ (x",n y",n ) 1" |x",n y",n |2 I  n I .

By (20.19) we obtain that r2 u" (x",n ) are uniformly bounded above, and they are also uni-
formly bounded below, since u" is semiconvex. Since similar remarks apply to r2 u" (y",n ),
we can assume with no loss of generality that r2 u" (x",n ) ! X" and r2 u" (y",n ) ! Y" . If
we now di↵erentiate w along a direction (⇠, ⇠) with ⇠ 2 Rn , we may use the fact that
along these directions the fourth order term is constant to get

hr2 u" (x",n )⇠, ⇠i hr2 u" (y",n )⇠, ⇠i  2 n |⇠|2 .

Taking limits, this proves that X"  Y" . On the other hand, from (20.18) we get

sup F (x, X" )  0 and inf F (y, Y" ) 0.

x2B pM " (x" ) y2B pM " (y" )

Now, the strict monotonicity of F (x, ·) yields

sup F (x, Y" ) sup F (x, Y" ) + sup F (x, X" ) + .

x2B pM " (x" ) x2B pM " (x" ) x2B pM " (x" )

Hence
sup F (x, Y" ) inf F (y, Y" ) .
x2B pM " (x" ) y2B pM " (y" )

Since and are fixed positive constants independent of ", and since |x" y" | ! 0, this
contradicts the uniform continuity of F (·, S) for " sufficiently small. ⇤

20.6 Hölder regularity

Consider a paraboloid P , i.e. a second-order polynomial of the form
1
P (x) = c + hp, xi + hSx, xi
2

149
for some c 2 R, p 2 Rn and S 2 Symn⇥n . We say that P is a paraboloid with opening
M 2 R if S = M I, namely
M 2
P (x) = c + hp, xi +
|x| .
2
It will be occasionally convenient to center a paraboloid P with opening M at some point
x0 , writing P (x) = P (x0 ) + hrP (x0 ), x x0 i + M2 |x x0 |2 .
Definition 20.25 (Tangent paraboloids). Given a function u : ⌦ ! R and a subset
A ⇢ ⌦ ⇢ Rn , we denote
✓(x0 , A, u) := inf {M |there exists P with opening M , u(x0 ) = P (x0 ) and u  P on A } .
Moreover, we set
✓(x0 , A, u) := sup {M |there exists P with opening M , u(x0 ) = P (x0 ) and u P on A } ,
so that ✓(x0 , A, u) = ✓(x0 , A, u). Finally, denoting by ± the positive and negative
parts, we set n o
+
✓(x0 , A, u) := max ✓ (x0 , A, u), ✓ (x0 , A, u) 0.
Given a function u : ⌦ ! R and h > 0, let us consider the symmetric di↵erence
quotient in the direction ⇠ 2 Rn
2 u(x0 + h⇠) + u(x0 h⇠) 2u(x0 ) @ 2u
h,⇠ u(x0 ) := h,⇠ ( h,⇠ u)(x0 ) = ⇠ (x0 ) ,
h2 @⇠ 2
well defined if h|⇠| < dist(x0 , @⌦) and identically equal to M on paraboloids with open-
ing M . Notice that the symmetric di↵erence quotient satisfies, by applying twice the
integration by parts formula for h,⇠ ,
Z Z
2 2
u h,⇠ dx = h,⇠ u dx (20.20)
⌦ ⌦

whenever u 2 L1loc (⌦), 2 L1 (⌦) has compact support, |⇠| = 1 and the h-neighbourhood
of supp is contained in ⌦.
Remark 20.26 (Maximum principle for 2⇠ ). If a paraboloid P with opening M “touches”
u from above (i.e. P (x0 ) = u(x0 ) and P (x) u(x) in some ball Br (x0 )), then
2 2
h,⇠ u(x0 )  h,⇠ P (x0 ) =M whenever |⇠| = 1 and |h|  r ,
and a similar property holds for paraboloids touching from below. Thus, passing to the
infimum from above and the supremum from below, we deduce the inequalities
2
✓(x0 , Br (x0 ), u)  h,⇠ u(x0 )  ✓(x0 , Br (x0 ), u) whenever |⇠| = 1 and |h|  r , (20.21)
and
2
| h,⇠ u(x0 )|  ✓(x0 , Br (x0 ), u) whenever |⇠| = 1 and |h|  r . (20.22)

150
Proposition 20.27. If u : ⌦ ! R satisfies

✓" := ✓( · , B" (·) \ ⌦, u) 2 Lp (⌦)

for some " > 0 and 1 < p  1, then u belongs to W 2,p (⌦) and, more precisely,

kr2⇠⇠ ukLp (⌦)  k✓" kLp (⌦) 8⇠ 2 Sn 1 . (20.23)

Remark 20.28. By bilinearity it is possible to obtain, from (20.23), an estimate on mixed

second derivatives:

kr2⇠⌘ ukLp (⌦)  |⇠||⌘|k✓" kLp (⌦) 8⇠, ⌘ 2 Rn , ⇠ ? ⌘ .

Proof. For any ' 2 Cc1 (⌦) one has
Z Z
@ 2'
u(x) 2 (x) dx = lim u(x) 2h,⇠ '(x) dx
⌦ @⇠ h!0 ⌦
Z
= lim ( 2h,⇠ u(x))'(x) dx  k✓" kLp (⌦) k'kLp0 (⌦) ,
h!0 ⌦

where we pass from the first to the second line with (20.20) and the inequality follows
from (20.22). Thanks to Riesz representation theorem, we know that the map ' 7!
R 2

⌦
u(x) @@⇠'2 (x) dx admits a representation with an element of Lp (⌦), which represents the
derivative r2⇠⇠ u in the sense of distributions and which satisfies (20.23). ⇤
In the space of n ⇥ n matrices we will consider the operator norm | · |L and, in the
subspace of symmetric matrices, the norm k · k provided by the largest modulus of the
eigenvalues in the spectrum (M ). Obviously these two norms coincide on Symn⇥n . From
(20.21) we get
kr2 u(x0 )k  ✓(x0 , B" (x0 ), u) for all " > 0 (20.24)
at any point x0 where u has a second order Taylor expansion.
Corollary 20.29. If ⌦ ⇢ Rn is convex and ✓" 2 L1 (⌦) for some " > 0, then

Lip(ru, ⌦)  k✓" kL1 (⌦) .

Proof. The previous proposition shows that u 2 W 2,1 (⌦) and (20.24) provides a
pointwise control on r2 u (recall that semiconvex/semiconcave functions have a second
order Taylor expansion a.e.). We recall that since ⌦ is convex and v is scalar we have
krvkL1 (⌦) = Lip(v, ⌦) (while, in general, krvkL1 (⌦)  Lip(v, ⌦)). If v takes values in
Rn (in our case v = ru : ⌦ ! Rn ), then, by the same smoothing argument used in the
scalar case, we can always show that

k|rv|L kL1 (⌦) = Lip(v, ⌦) (20.25)

151
because, when v is continuously di↵erentiable, there holds
Z 1 Z 1
v(x) v(y) = Dv((1 t)x + ty)(x y) dt  |x y| |rv|L ((1 t)x + ty) dt .
0 0

Therefore from (20.24) and (20.25) we conclude. ⇤

At this point our aim is the study of a nonlinear PDE as
F (r2 u(x)) + f (x) = 0 (20.26)
with F non-decreasing on Symn⇥n (the trace, corresponding to the Laplacian, for exam-
ple).
Definition 20.30 (Ellipticity). In the problem (20.26) we have ellipticity with constants
⇤ > 0 if
kN k  F (M + N ) F (M )  ⇤kN k 8N 0 . (20.27)
Remark 20.31. Every symmetric matrix N admits a unique decomposition as a sum
N = N+ N ,
+
with
Pn N , N 0 and N + N = 0. It can P be obtained simply diagonalizing
P N =
+
i=1 ⇢i ei ⌦ ei and then choosing N := ⇢i >0 ⇢i ei ⌦ ei and N = ⇢i 0 ⇢i ei ⌦ ei .
Observing this, we are able to write the definition of elliptic problem replacing (20.27)
with
F (M + N ) F (M )  ⇤kN + k kN k 8 N 2 Symn⇥n . (20.28)
Indeed, it suffices to write
F (M + N ) F (M ) = F (M N + N +) F (M N ) + F (M N ) F (M )
and to apply to the first term the estimate from above and to the second one the estimate
from below.
Example 20.32. Consider the case
F (M ) = tr(BM )
where B = (bij )i,j=1,...,n belongs to the set
A ,⇤ := B 2 Symn⇥n | I  B  ⇤I .
Fix the symmetric matrix N 0. To verify (20.27), we choose the coordinate system in
which N = diag(⇢1 , . . . , ⇢n ), thus (since bii and ⇢i 0 for all i = 1, . . . , n)
n
X n
X
F (M + N ) F (M ) = tr(BN ) = bii ⇢i ⇢i ⇢max .
i=1 i=1

152
Analogously, since bii  ⇤ one has
n
X n
X
F (M + N ) F (M ) = tr(BN ) = bii ⇢i  ⇤ ⇢i  n⇤⇢max .
i=1 i=1

After this introductory part about definitions and notation, we enter in the core of
the matter of the Hölder regularity for viscosity solutions: as in De Giorgi’s work on the
XIX Hilbert problem, the regularity will be deduced only from inequalities derived from
ellipticity, without a specific attention to the original equation.

Definition 20.33 (Pucci’s extremal operators). Given ellipticity constants ⇤ >0

±
and a symmetric matrix M , Pucci’s extremal operators are defined by setting M ,⇤ (0) = 0
and
X X
M ,⇤ (M ) := ⇢+⇤ ⇢,
⇢2 (M )\(0,1) ⇢2 (M )\( 1,0)
X X
M+,⇤ (M ) := ⇤ ⇢+ ⇢.
⇢2 (M )\(0,1) ⇢2 (M )\( 1,0)

We will omit the dependence on and ⇤, when clear from the context.

Remark 20.34. Resuming Example 20.32, we can show that

M ,⇤ (M ) = inf tr(BM ) (20.29)

B2A ,⇤

M+,⇤ (M ) = sup tr(BM ) . (20.30)

B2A ,⇤

As a matter of fact, denoting with (bij ) the coefficients of the matrix B 2 A ,⇤ in the
system of coordinates where M is diagonal, with M = diag(⇢1 , . . . , ⇢n ) we get
n
X X X
tr(BM ) = bii ⇢i ⇢i + ⇤ ⇢i (20.31)
i=1 ⇢i >0 ⇢i <0

and the equality in (20.31) holds if

X X
B= ei ⌦ ei + ⇤ei ⌦ ei .
⇢i >0 ⇢i <0

Remark 20.35. Pucci’s extremal operators satisfy the following properties:

(a) trivially M  M+ and M ( M ) = M+ (M ) for every symmetric matrix M ,

moreover M± are positively 1-homogeneous;

153
 (b) for every M, N it is simple to obtain from (20.29) and (20.30) that

M+ (M ) + M (N )  M+ (M + N )  M+ (M ) + M+ (N )

and, similarly,

M (M ) + M (N )  M (M + N )  M (M ) + M+ (N ) ;

(c) M± are elliptic (i.e., they satisfy (20.27)) with constants , n⇤, because of Exam-
ple 20.32 and (20.29), (20.30) which represent M± as an envelope of a family of
functionals with ellipticity constants , n⇤.

(d) thanks to (20.28), one has

M ,n⇤ (M )  F (M )  M+/n,⇤ (M ) 8M 2 Symn⇥n (20.32)

whenever F is elliptic with constants , ⇤ and F (0) = 0.

Definition 20.36. With the previous notations, we will denote

Sub ,⇤ (f ) := u:⌦!R M+,⇤ (r2 u) + f  0 in ⌦

Sup ,⇤ (f ) := u:⌦!R M ,⇤ (r
2
u) + f 0 in ⌦ .

We also set
Sol ,⇤ (f ) := Sub /n,⇤ ( |f |) \ Sup ,n⇤ (|f |) . (20.33)
Remark 20.37. Roughly speaking, the classes defined above correspond to De Giorgi’s
classes DG± (⌦), since u being a solution to (20.26) with F having ellipticity constants
and ⇤ implies u 2 Sol ,⇤ (f ); thus, if we are able to infer regularity of functions in
Sol ,⇤ (f ) then we can “forget” thanks to Remark 20.35(d) the specific equation.

21 Regularity theory for viscosity solutions

21.1 The Alexandrov-Bakelman-Pucci estimate
Let us recall the notation from the previous section:

Sub(f ) := u:⌦!R M+ (r2 u) + f  0 in ⌦

Sup(f ) := u:⌦!R M (r2 u) + f 0 in ⌦ ,

where M± are Pucci’s extremal operators, and we shall not emphasize from now on the
dependence on the ellipticity coefficients and ⇤. Notice that, since M+ M , the
intersection of the two sets can be nonempty.

154
 The estimate we want to prove is named after Alexandrov, Bakelman and Pucci and is
therefore called ABP weak maximum principle. It plays the role in this regularity theory
played by the Caccioppoli inequality in the standard linear elliptic theory.
In the sequel we call “universal” a constant which depends only on the space dimension
n and on the ellipticity constants , ⇤.

Theorem 21.1 (Alexandrov-Bakelman-Pucci weak maximum principle). Let u be in

Sup(f ) \ C(B r ) with u 0 on @Br and f 2 C(B r ). Then
✓Z ◆1/n
+ n
max u  Cr f dx ,
Br {u= u}

where C is universal and u is defined below.

Since f + measures, in some sense, how far u is from being concave, the estimate above
can be seen as a quantitative formulation of the fact that a concave function in a ball
attains its minimum on the boundary of the ball.

Definition 21.2 (Definition of u ). Assume the function u is extended to all B 2r \ B r as

the null function (this extension is continuous, since u is null on @Br ). We then define

u (x) = sup L(x) L affine, L  u on B 2r .

In order to prove the ABP estimate we set M := maxB r u and assume with no loss
of generality that M > 0.
The following facts are either trivial consequences of the definitions or easy applications
of the tools introduced in the convex analysis part: firstly M  u  0, as a consequence
1,1
u 2 Wloc (B2r ) and finally since u is di↵erentiable a.e. by Rademacher’s theorem and
the graph of the subdi↵erential is closed, we get @ u (x) 6= ; for all x 2 B2r . We will use
this last property to provide a supporting hyperplane to u at any point in B r .
We need some preliminary results, here is the first one.

Theorem 21.3. Assume u 2 C(B r ), u 0 on @Br and u 2 C 1,1 (Br ). Then

✓Z ◆1/n
2
max u  cr det r u dx ,
Br Br

with c = c(n).
Proof. Let x1 2 Br be such that u (x1 ) = M . Fix ⇠ with |⇠| < M/(3r) and denote by L↵
the affine function L↵ (x) = ↵+hx, ⇠i. It is obvious that if ↵ 1, then the corresponding
hyperplane lies below the graph of u and there is a minimum value of ↵ such that this
happens, that is u L↵ on B 2r . The graph of u will then meet the corresponding

155
hyperplane at some point, say x0 2 B 2r . If it were |x0 | > r, then L↵ (x0 ) = 0, but on the
other hand |L↵ (x1 )| M and, since |x0 x1 |  3r, L↵ would have slope |⇠| M/3r,
which is a contradiction. Hence any contact point x1 must lie inside the ball Br ; from
u u L↵ we get r u (x1 ) = ⇠ and therefore BM/(3r) ⇢ r u (Br ). If we measure
the corresponding volumes and use the area formula, we get
✓ ◆n Z
M
!n  det r2 u dx
3r Br

or, equivalently,
✓Z ◆1/n
1/n 2
M  3!n r det r u dx .
Br
1/n
This proves the claim with c = 3!n . ⇤

Remark 21.4. The previous theorem implies the ABP estimate, provided we show that
• u 2 C 1,1 (Br ), as a consequence of u 2 Sup(f );

• L n -a.e. on {u > u} (the so-called non-contact region) one has det r2 u = 0;

• L n -a.e. on {u = u} (the so-called contact region) one has det r2 u  C(f + )n ,

with C universal.
Let us now come to the next steps. The next theorem shows that regularity, measured
in terms of opening of paraboloids touching u from above, propagates from the contact
set to the non-contact set. It turns out that the regularity in the contact set is a direct
consequence of the supersolution property.
Theorem 21.5 (Propagation of regularity). Let u 2 C(B r ) and suppose there exist " 2
(0, r] and M 0 such that, for all x0 2 B r \ {u = u }, there exists a paraboloid with
opening less than M which has a contact point from above with the graph of u in B" (x0 ).
Then u 2 C 1,1 (B r ) and det r2 u = 0 a.e. on {u > u } .
With the notation introduced before, the assumption of Theorem 21.5 means

✓(x0 , B" (x0 ), u) M 8x0 2 B r \ {u = u} .

Since u is convex, the corresponding quantity ✓ is null. Recall also that we have already
proved that ✓, ✓ 2 L1 implies u 2 C 1,1 in Corollary 20.29.
Theorem 21.6 (Regularity at contact points). Consider v 2 Sup(f ) in B , ' convex in
B with 0  '  v and v(0) = '(0) = 0. Then '(x)  C supB f + |x|2 in B⌫ , where ⌫
and C are universal constants.

156
 We can get a naive interpretation of this lemma (or, better, of its infinitesimal version
as # 0) by this formal argument: v ' having a local minimum at 0 implies, by the
assumption v 2 Sup(f ) M (r2 '(0))  f (0). Formally, M (r2 '(0))  M (r2 v(0)) 
f (0).
Now it is possible to see how these tools allow to prove the ABP estimate.
Proof. [of Theorem 21.1] Pick a point x0 2 Br \ {u = u } and let L be a supporting
hyperplane for u at x0 , so that u L and u (x0 ) = L (x0 ) . Recalling Theorem 21.6,
define ' := u L, v := u L (and notice that v is a supersolution because v 2
Sup(f Br )). Now, '(x0 ) = v(x0 ) implies, by means of Theorem 21.6,

✓(x0 , B⌫ (x0 ), ')  C sup f + 8x0 2 B r (21.1)

B (x0 )

with ⌫ and C universal, for all 2 (0, r). Hence

✓(x0 , B⌫ (x0 ), u)  C sup f + . (21.2)

B (x0 )

By Theorem 21.5 we get u 2 C 1,1 and det r2 u = 0 a.e. in the non-contact region.
Finally, in order to get the desired estimate, we have to show that a.e. in the contact
region one has det r2 u  c(f + )n . But this comes at once by passing to the limit as ! 0
in (21.2) at any di↵erentiability point x0 of u . In fact, all the eigenvalues of r2 u (x0 )
do not exceed Cf + (x0 ) and the conclusion follows. ⇤
Now we prove Theorem 21.6.
Proof. Let r 2 (0, /4) and call c := supB r ' /r2 . Let then x̄ 2 @Br be a maximum
point of ' on B r (by convexity the maximum is attained at the boundary). By means of
a rotation, we can write x = (x0 , xn ), x0 2 Rn 1 , xn 2 R, and assume x̄ = (0, r). Consider
the intersection A of the closed strip defined by the hyperplanes xn = r and xn = r
with the ball B /2 . We clearly have that @A = A1 [ A2 [ A3 , where A1 = B /2 \ {xn = r},
A2 = B /2 \ {xn = r} and A3 = @B /2 \ {|xn | < r}.
We claim that ' '(x̄) on A1 . To this aim, we first prove that '(y)  '(x̄)+o(|y x̄|)
for y ! x̄, y 2 H := {xn = r}. In fact, this comes from '(ry/|y|)  '(x̄) and observing
that '(y) '(ry/|y|) = o(|y x̄|), because ' is Lipschitz continuous. On the other hand,
we have that ⇠ 2 @'|H (x̄) implies '(y) '(x̄) + h⇠, y x̄i for all y 2 H. Hence, by
comparison, it must be ⇠ = 0 and so '(y) '(x̄) on A1 (this can be seen as a nonsmooth
version of the Lagrange multipliers theorem).
As a second step, set
c c
p(x) := (xn + r)2 4 2
r2 |x0 |2
8
and notice that the following properties hold:

157
 (a) on A1 , p(x)  c/(2r2 ) = '(x̄)/2  '(x)/2;
(b) on A2 , p(x)  0  '(x) (and in particular p(x)  v(x));
(c) on A3 , 2 /4 = |x0 |2 + x2n  |x0 |2 + r2  |x0 |2 + 2 /16, which implies |x0 |2 (3/16) 2 .
By means of the last estimate we get p(x)  (c/2)r2 (3/4)cr2  0  '.
Combining (a), (b), (c) above we get p  v on @A. Since p(0) = cr2 /8 > 0 = '(0)
we can rigidly move down this paraboloid until we get a limit paraboloid p0 = p ↵ (for
some translation parameter ↵ > 0) lying below the graph of v and touching it at some
point, say y. Since p  v on @A, the point y is internal to A.
By the supersolution property M (r2 p)  f (y)  supB f we get (since we have an
explicit expression for p)
c r2
8(n 1)⇤c 2  sup f.
4 B
2 2
pwe can fix r such that 8(n 1)⇤cr / 8 c/8 (it is done by taking r so that
But now
8r  /((n 1)⇤)): we have therefore c  supB f . The statement then follows
p
with C = 8/ and ⌫ := 18 /((n 1)⇤)). ⇤
It remains to prove Theorem 21.5.
Proof. Recall first that we are assuming the existence of a uniform estimate
✓(x, B" (x), u) M 8x 2 B r \ {u = u }.

Thanks to Proposition 20.27, we are able to obtain C 1,1 regularity of u as soon we are
able to propagate this estimate also to non-contact points.
Consider now any point x0 2 B r \ {u > u } and call L a supporting hyperplane for
u at x0 . Notice that x0 2 { u = L} ⇢ {u = u }. We claim that:

(a) There exist n + 1 points x1 , . . . , xn+1 such that x0 2 S := co(x1 , . . . , xn+1 ) (here
and in the sequel co stands for convex hull) and, moreover, all such points belong
to B r \ { u = L} with at most one exception lying on @B2r . In addition u ⌘ L
on S;
P
(b) x0 = n+1 i=1 ti xi with at least one index i verifying both xi 2 B r \ { u = L} and
ti 1/(3n).
To show the utility of this claim, just consider how these two facts imply the thesis: on
the one hand, if r u is di↵erentiable at x0 , we get det r2 u (x0 ) = 0 because u = L on
S and dim(S) 1. On the other hand we may assume, without loss of generality that
x1 2 {u = u } \ Br and t1 (1/3n) so that, since
✓ ◆
h
x0 + h = t 1 x1 + + t2 x2 + · · · + tn+1 xn+1 ,
t1

158
one has

u (x0 + h)  t1 (x1 + h/t1 ) + t2 u (x2 ) + · · · + tn+1 u (xn+1 )

"u #
2
h
 t1 L(x1 ) + M + t2 L(x2 ) + · · · + tn+1 L(xn+1 )
t1
= L(x0 ) + M |h|2 /t1  u (x0 ) + 3nM |h|2

and this estimate is clearly uniform since we only require |h/t1 |  ", which is implied by
|h|  "/(3n).
Hence, the problem is reduced to prove the two claims above. This is primarily based
on a standard result in convex analysis (first proved by Carathéodory for closed sets),
which is recalled here for completeness.

Theorem 21.7 (Carathéodory). Let V be a n-dimensional real vector space. If C ⇢ V ,

then for every x 2 co(C) (the convex hull of C) there exist x1 , . . . , xn+1 2 C, t1 , . . . , tn+1 2
[0, 1] such that
n+1
X n+1
X
x= t i xi and ti = 1 .
i=1 i=1

Set then C 0 := x 2 B 2r | L(x) = u (x) and C = co(C 0 ). We immediately notice

that C 0 6= ;. We claim that x0 2 C: in fact, if this were not the case, there would
exist ⌘ > 0 and a hyperplane L0 such that L0 (x0 ) > 0 and L0 (y) < 0 if y 2 C⌘ :=
y 2 B 2r | dist(y, C) < ⌘ , therefore L + L0  u on C⌘ for all > 0. Let us notice
that, on B 2r \ C⌘ ⇢ B 2r \ C 0 , the function u L is strictly positive and, thanks to the
compactness of B 2r \ C⌘ , there exists > 0 such that

L(x) + L0 (x)  u (x), 8 x 2 B 2r \ C⌘ .

Hence, we would have (L + L0 )(x0 ) > L(x0 ) and, at the same time,

L + L0  u on B 2r ,

which contradicts the maximality of L. P

Thanks to Carathéodory’s theorem, we can write x0 = n+1 i=1 ti xi with xi 2 { u = L}.
In case there were distinct points xi , xj with |xi | > r and |xj | > r (and so L(xi ) = 0,
L(xj ) = 0) then (considering a point z on the open segment between xi and xj ) the
function u would achieve its maximum, equal to 0, in the interior of B2r and so, by
the convexity of u , it would be u ⌘ 0 on B2r , in contrast with the assumption M =
max u > 0. The same argument also proves that exceptional points out of B r , if any,
must lie on @B2r .

159
Let us now prove that u (x) = L(x) on S := co(x1 , . . . , xn+1 ). The implication is
trivial, the converse one is clear for each x = xi , since L  u  u , and it is obtained
by means of the convexity of u at all points in S.
Now we prove part (b) of the claim. If all points xj verify |xj |  r, then max ti
1 1
n+1
> 3n . Otherwise, if one point, say xn+1 , satisfies |xn+1 | = 2r, then ti < 1/(3n) for all
i = 1, . . . , n implies tn+1 > 2/3 and therefore
n
X 4 n
r |x0 | 2tn+1 r ti |xi | > r r=r.
i=1
3 3n

21.2 The Harnack inequality

In this section we shall prove the Harnack inequality for functions in the class Sol(f ) :=
Sub( |f |)\Sup(|f |) where, according to Definition 20.36, the sets Sup(|f |) and Sub( |f |)
are defined through Pucci’s extremal operators (with fixed ellipticity constants 0 < 
⇤):10 , in the sense of viscosity solutions,
u 2 Sub( |f |) () M+ (r2 u) |f |  0 ; (21.3)
u 2 Sup(|f |) () M (r2 u) + |f | 0 . (21.4)
We shall use the standard notation Qr (x) for the closed n-cube in Rn with side length
r, Qr = Qr (0) and always assume that f is continuous. In the proof of Lemma 21.13
below, however, we shall apply the ABP estimate to a function w 2 Sup(g) with g upper
semicontinuous. Since there exists gn continuous with gn # g and w 2 Sup(gn ), the ABP
estimate holds, by approximation, even in this case.
Theorem 21.8. Consider a function u : Q1 ! R with u 0 and u 2 Sol(f ) \ C(Q1 ).
There exists a universal constant CH such that
✓ ◆
sup u  CH inf u + kf kLn (Q1 ) . (21.5)
Q1/2 Q1/2

Let us show how (21.5) leads to the Hölder regularity result for viscosity solutions of
the fully nonlinear elliptic PDE
F (r2 u(x)) + f (x) = 0 . (21.6)
Step 1. As usual, we need to control the oscillation (now on cubes), defined by
!r := Mr mr
10
Notice that Sup(f ) ⇢ Sup(|f |) and Sub(f ) ⇢ Sub( |f |).

160
with Mr := supQr u and mr := inf Qr u.
With the same notation of Theorem 21.8, there exists a universal constant µ 2 (0, 1) such
that
!1/2  µ !1 + 2kf kLn (Q1 ) . (21.7)
Indeed, we apply the Harnack inequality (21.5)
• to the function u m1 , so that
M1/2 m1  CH m1/2 m1 + kf kLn (Q1 ) ; (21.8)

• to the function M1 u, so that

M1 m1/2  CH M1 M1/2 + kf kLn (Q1 ) . (21.9)

Adding (21.8) and (21.9) we get

!1 + !1/2  CH !1 !1/2 + 2kf kLn (Q1 ) ,
which proves (21.7) because
CH 1 CH CH 1
!1/2  !1 + 2 kf kLn (Q1 ) < !1 + 2 kf kLn (Q1 ) .
CH + 1 CH + 1 CH + 1
We spend a line to remark that µ = (CH 1)/(CH + 1), CH being the universal constant
in (21.5). It is crucial for the decay of the oscillation that µ < 1.
Step 2. Thanks to a rescaling argument (which we will be hugely used also in the proof
of the Harnack inequality), we can generalize (21.7). Fix a radius 0 < r  1 and put
u(ry)
ur (y) := , fr (y) = f (ry) with y 2 Q1 .
r2
Notice that (21.7) holds also for ur (with the corresponding source fr ) because Pucci’s
operators are positively 1-homogeneous. Moreover, passing to a smaller scale, the Ln -
norm improves.
For simplicity we keep the notation !r for the oscillation of the function u, we use osc(·, Qr )
otherwise. We can estimate
!r/2 = r2 osc(ur , Q1/2 )  µr2 osc(ur , Q1 ) + 2r2 kfr kLn (Q1 )
= µ!r + 2rkf kLn (Qr )  µ!r + 2rkf kLn (Q1 ) .
Step 3. By the iteration lemmas we used so frequently in the elliptic regularity chapters11 ,
we are immediately able to conclude that
✓ ◆↵
min{1,↵} 1
!r  C!1 r 8r 2 (0, 1] with =µ,
2
11
See, for instance, Lemma 9.2.

161
and with C depending only on µ and kf kLn (Q1 ) , thus we have Hölder regularity.
In order to prove the Harnack inequality, we will pass through the following reformu-
lation of Theorem 21.8.
Theorem 21.9. There exist universal positive constants "0 , C such that if u : Q4pn !
[0, 1) belongs to Sol(f ) \ C(Q4pn ) on Q4pn , then

inf u  1 =) sup u  C (21.10)

Q1/4 Q1/4

provided
kf kLn (Q4pn )  "0 .
Remark 21.10. Theorem 21.8 and Theorem 21.9 are easily seen to be equivalent: since
we will prove the second one, it is more important for us to check that Theorem 21.8
follows from Theorem 21.9.
For some positive > 0 (needed to avoid a potential division by 0) consider the function
u
v := .
+ inf Q1/4 u + kf kLn (Q4pn ) /"0

Denoting by fv the source term associated with v, the homogenity of Pucci’s operators
gives kfv kLn (Q4pn )  "0 . Since inf Q1/4 v  1 we have supQ1/4 v  C, hence
✓ ◆
sup u  C inf u + + kf kLn (Q4pn ) /"0 .
Q1/4 Q1/4

We let ! 0 and we obtain Harnack inequality with the cubes Q1/4 , Q4pn ; by the same
scaling argument we already used, this means
✓ ◆
sup u  C inf u + rkf kLn (Q16rpn (x0 ) ) . (21.11)
Qr (x0 ) Qr (x0 )

Now, we pass to the cubes Q1/2 , Q1 with a simple covering argument: there exists an
integer N = N (n) such that for all x 2 Q1/2 , y 2 Q1 we can find points xi , 1  i  N ,
with xi = x, xN = y and xi+1 2 Qr (xi ) for 1  i < N , with r = r(n) so small that all
cubes Q16rpn (xi ) are contained in Q1 . Applying repeatedly (21.11) we get (21.5) with
CH ⇠ C N .
We describe the strategy of the proof of Theorem 21.9, even if the full proof will be
completed at the end of this section.
We will study the map
t 7! L n ({u > t} \ Q1 )
in order to prove:

162
 • a decay estimate of the form L n ({u > t} \ Q1 )  dt " , thanks to the fact that
u 2 Sup(|f |) (see Lemma 21.13),
• the full thesis of Theorem 21.9 using the fact that u 2 Sol(f ) ⇢ Sub( |f |), too.
The first goal will be achieved using the Alexandrov-Bakelman-Pucci inequality of the
previous section. The structure of the proof remembers that of De Giorgi’s regularity
theorem, as we said, and we will complete it through the following lemmas and remarks.
The first lemma is a particular case of Calderón-Zygmund decomposition.
Lemma 21.11 (Dyadic Lemma). Consider Borel sets A ⇢ B ⇢ Q1 with L n (A)  < 1.
If the implication
L n (A \ Q) > L n (Q) =) Q̃ ⇢ B , (21.12)
holds for any dyadic cube Q ⇢ Q1 , with Q̃ being the predecessor of Q, then

L n (A)  L n (B) .
Proof. We apply the construction of Calderón-Zygmund (seen in the proof of The-
orem 14.1) to f = A with ↵ = : there exists a countable family of cubes {Qi }i2I ,
pairwise disjoint, such that
[
A  L n -a.e. on Q1 \ Qi (21.13)
i2I

and L n (A \ Qi ) > L n (QSi ) for all i 2 I. Since < 1 and A is a characteristic function,
(21.13) means that A ⇢ i2I Qi up to Lebesgue negligible sets. Moreover, if Q̃i are the
predecessors of Qi , from (21.12) we get Q̃i ⇢ B for all i and

L n (A \ Q̃i )  L n (Q̃i ) 8i 2 I . (21.14)

This is due to the fact that a cube Q, in the Calderón-Zygmund construction, is divided
in subcubes as long as L n (A \ Q)  L n (Q). Thus (note that we sum on Q̃i rather than
on i, because di↵erent cubes might have the same predecessor)
X X
L n (A)  L n (A \ Q̃i )  L n (Q̃i )  L n (B) .
Q̃i Q̃i

⇤
It is bothering, but necessary to go on with the proof, to deal at the same time with
balls and cubes: balls emerge from the radial construction in the next lemma and cubes
are needed in Calderón-Zygmund Theorem.
Lemma 21.12 (Truncation Lemma). There exists a universal function ' 2 C 1 (Rn ) such
that

163
 (i) ' 0 on Rn \ B2pn ;

(ii) '  2 on the cube Q3 ;

(iii) M+ (r2 ')  C' Q1 on Rn .

Proof. We recall some useful inclusions:

B1/2 ⇢ Q1 ⇢ Q3 ⇢ B3pn/2 ⇢ B2pn .

For M1 , M2 > 0 and ↵ > 0 we define

↵
'(x) = M1 M2 |x| when |x| 1/2 .

Since ' is an increasing function of |x|, we can find M1 = M1 (↵) > 0 and M2 =
M2 (↵) > 0 such that

(i) ' ⌘ 0, so that ' 0 on Rn \ B2pn ;

@B2pn

(ii) ' ⌘ 2, so that '  2 on Q3 \ B1/2 .

@B3pn/2

After choosing a smooth extension for ' on B1/2 , still less than 2, we conclude check-
ing that there exists an exponent ↵ that is suitable to verify the third property of the
statement, that needs to be checked only on . We compute
↵ x⌦x
r2 |x| ↵
= ↵+2
I + ↵(↵ + 2) ↵+4 ,
|x| |x|

thus the eigenvalues of r2 ' when |x| 1/2 are M2 ↵|x| (↵+2) with multiplicity n 1
and M2 ↵(↵ + 1)|x| (↵+2) with multiplicity 1 (this is the eigenvalue due to the radial
direction). Hence, when |x| 1/4 we have

M2
M+ (r2 ') = (⇤(n 1)↵ ↵(↵ + 1))
|x|↵+2

so that M+ (r2 ')  0 on Rn \ B1/2 if we choose ↵ = ↵(n, , ⇤) 1. Since B1/2 ⇢ Q1

and ' is smooth, we conclude that (iii) holds for a suitable constant C. ⇤

Lemma 21.13 (Decay Lemma). There exist universal constants "0 > 0, M > 1 and
µ 2 (0, 1) such that if u 2 Sup(|f |), u 0 on Q4pn , inf Q3 u  1 and kf kLn (Q4pn )  "0 ,
then for every integer k 1

L n {u > M k } \ Q1  (1 µ)k . (21.15)

164
Proof. We prove the first step, that is
L n ({u > M } \ Q1 )  (1 µ) , (21.16)
with M := max ' , ' given by Lemma 21.12, and µ and "0 are respectively given by
n 1
µ := (2CABP C' ) , "0 = , (21.17)
2CABP
where CABP is the universal constant of the Alexandrov-Bakelman-Pucci estimate of The-
orem 21.1. Since u is nonnegative, in order to obtain a meaningful result from the ABP
estimate, we apply the estimate in the ball B2pn for the function w, defined as the function
u additively perturbed with the truncation function '. If w := u + ', then
(i)
w 0 on @B2pn (21.18)
because u 0 on Q4pn B2pn and ' 0 on Rn \ B2pn ;
(ii)
inf w  inf w  1 (21.19)
B2pn Q3

because Q3 ⇢ B2pn and '  2 on B2pn , and, at the same time, we are assuming
that inf Q3 u  1;
(iii) directly from the definition of Sup(|f |) we get M (r2 u) + |f | 0, moreover
M+ (r2 ')  C' Q1 . Since in general M (A + B)  M (A) + M+ (B) (see Re-
mark 20.35), then
M (r2 w) + (|f | + C' Q1 ) ( M (r2 u) + |f |)+( M+ (r2 ') + C' Q1 ) 0.
(21.20)
The inequality (21.20) means that w 2 Sup(|f | + C' Q1 ).

Thanks to the ABP estimate (which we can apply to w thanks to (21.18) and (21.20))
we get
✓Z ⇣ ⌘n ◆1/n
max w (x)  CABP |f (y)| + C' Q1 (y) dy . (21.21)
x2B 2pn {w= w}

Now, remembering that (21.19) holds and that, by definition, {w = w } ⇢ {w  0}, we

can expand (21.21) with
✓Z ⇣ ⌘n ◆1/n
1  max w (x)  CABP |f | + C' Q1 dy (21.22)
x2B 2pn {w0}

 CABP kf kLn (Q4pn ) + CABP C' L n (Q1 \ {w  0})1/n (21.23)

 CABP kf kLn (Q4pn ) + CABP C' L n (Q1 \ {u  M })1/n , (21.24)

165
where we pass from line (21.22) to line (21.23) by Minkowski inequality and from line
(21.23) to line (21.24) because, if w(x)  0, then u(x)  '(x) and then u(x)  M .
Using our choice of "0 we obtain from (21.24) the lower bound
1
L n (Q1 \ {u  M })1/n . (21.25)
2CABP C'

Thus, if µ is given by (21.17), we obtain (21.16).

We prove the inductive step: suppose that (21.15) holds for every j  k 1. We exploit
the dyadic Lemma 21.11 with A = {u > M k } \ Q1 , B = {u > M k 1 } \ Q1 and = 1 µ.
Naturally A ⇢ B ⇢ Q1 and L n (A)  ; if we are able to check that (21.12) holds, then

L n Q1 \ {u > M k }  (1 µ)L n Q1 \ {u > M k 1 }  (1 µ)k .

Concerning (21.12), suppose by contradiction that for some dyadic cube Q ⇢ Q1 we have
that
L n (A \ Q) > L n (Q) (21.26)
but Q̃ 6⇢ B, Q̃ being the predecessor of Q, as usual: there exists z 2 Q̃ such that
u(z)  M k 1 . Let us rescale and translate the problem, putting ũ(y) := u(x)M (k 1)
with x = x0 + 2 i y if Q has edge length 2 i and centre x0 (so that, in this transformation
Q becomes the unit cube and Q̃ is contained in Q3 ). Because of the rescaling technique,
we need to adapt f , that is define a new datum

f (x)
f˜(y) := .
22i M k 1
The intention of this definition of f˜ is to ensure that ũ 2 Sup(|f˜|), in fact
1
M (r2 ũ) + |f˜| = M (r2 u) + |f | 0.
22i M k 1
Since the point corresponding to z belongs to Q3 , we ge

u(z)
inf ũ(y)  1.
y2Q3 Mk 1

If kf˜kLn (Q4pn )  "0 , then, applying what we already saw in (21.25) to ũ instead of u,

µ  L n ({ũ  M } \ Q1 ) = 2ni L n {u  M k } \ Q ,

this means that µL n (Q)  L n {u  M k } \ Q and, passing to the complement,

L n {u > M k } \ Q  (1 µ)L n (Q) ,

166
which contradicts (21.26).
In order to complete our proof, we show that e↵ectively kf˜kLn (Q4pn )  "0 . In general,
let us remark that the rescaling technique does not cause any problem at the level of the
source term f . Indeed
1
kf˜kLn (Q4pn ) = kf kLn (Q4pn/2i (x0 ))  kf kLn (Q4pn ) "0 .
M k 1 2i
⇤

Corollary 21.14. There exist universal constants " > 0 and d 0 such that if u 2
Sup(|f |), u 0 on Q4pn , inf Q3 u  1 and kf kLn (Q4pn )  "0 , then

L n ({u > t} \ Q1 )  dt "

8t > 0 . (21.27)
Proof. This corollary is obtained by Lemma 21.13 choosing " such that (1 µ) = M "
and d0 = M " = (1 µ) 1 : interpolating, for every t M there exists k 2 N such that
M k 1  t < M k , so

L n ({u > t} \ Q1 )  L n {u > M k 1 } \ Q1  M "(k 1)

 d0 (M k ) "
 d0 t "
.

Choosing d d0 such that 1  dt "

for all t 2 (0, M ), we conclude. ⇤
In the next lemma we use both the subsolution and the supersolution property to
improve the decay estimate on L n ({u > t}). The statement is a bit technical and the
reader might wonder about the choice of the scale lj as given in the statement of the lemma;
it turns out, see (21.31), that this is (somehow) the smallest scale r on which we are able to
say that L n ({u ⌫ j } \ Qr ) ⌧ rn , knowing that the global volume L n ({u ⌫ j } \ Q1 )
is bounded by d(⌫ j ) " .

Lemma 21.15. Suppose that u 2 Sub( |f |) is nonnegative on Q4pn and kf kLn (Q4pn ) 
"0 , with "0 given by the decay Lemma 21.13. Assume that (21.27) holds. Then there exist
universal constants M0 > 1 and > 0 such that if

x0 2 Q1/2 and u(x0 ) M0 ⌫ j 1

for some j 1,

then
9 x1 2 Qlj (x0 ) such that u(x1 ) M0 ⌫ j ,
"/n "j/n
where ⌫ := 2M0 /(2M0 1) > 1 and lj := M0 ⌫ .
Proof. First of all, we fix a large universal constant > 0 such that
1 n
p > d2" (21.28)
2 4 n

167
and then we choose another universal constant M0 so large that
1
dM0 " < (21.29)
2
and p
"/n
M0 <2 n. (21.30)
We first estimate the superlevels

Ln u ⌫ j M0 /2 \ Qlj /(4pn) (x0 )  L n {u ⌫ j M0 /2} \ Q1

✓ ◆n
j " 1 j" "
 d ⌫ M0 /2 < p ⌫ M0
2 4 n
✓ ◆n
1 lj
= p , (21.31)
2 4 n

where we used condition (21.28) on and the definition of lj , as given in the statement
of the lemma.
By contradiction, assume that for some j 1 we have

max u < M0 ⌫ j . (21.32)

Qlj (x0 )

Under this assumption, we claim that the superlevel can be estimated as follows:
1
L n {u < ⌫ j M0 /2} \ Qlj /(4pn) (x0 ) < L n Qlj /(4pn) . (21.33)
2
Obviously the validity of (21.31) and (21.33) is the contradiction that will conclude the
proof, so we need only to show (21.33).
Define the auxiliary function

⌫M0 u(x)⌫ (j 1)
u(x)
v(y) := = 2 M0 ,
(⌫ 1)M0 ⌫j
l
where x = x0 + 4pj n y and the second equality is a consequence of the relation M0 =
⌫/[2(⌫ 1)]. Since y 2 Q4pn () x 2 Qlj (x0 ), by (21.32) the function v is defined
and positive on Q4pn . In addition, using the first equality in the definition of v, we
immediately see that u(x0 ) M0 ⌫ j 1 implies inf Q4pn v  1.
Using the second equality we see that (modulo the change of variables)

{v > M0 } = {u < ⌫ j M0 /2} .

168
Moreover,p if we compute the datum fv which corresponds to v, since the rescaling radius
is lj /(4 n), we get
2lj2
fv (y) = j f (x)
⌫
so that
2lj
kfv kLn (Q4pn ) = p j kf kLn (Qlj (x0 ))  "0 (21.34)
4 n⌫
because
"/n
2l M0
pj j = p ⌫ "j/n j < 1
4 n⌫ 2 n
thanks to (21.30). The estimate in (21.34) allows us to use Corollary 21.14 for v, that is
L n ({v > M0 } \ Q1 )  dM0 " ,
and we can use this, together with (21.29), to obtain that (21.33) holds:
1
L n {u < ⌫ j M0 /2} \ Qlj /(4pn) (x0 )  dM0 " L n Qlj /(4pn) < L n Qlj /(4pn) .
2
⇤
We can now complete the proof of Theorem 21.9, using Lemma 21.15. Notice that
in Theorem 21.9 we made all assumptions needed to apply Lemma 21.15, taking also
Corollary 21.14 into account, which ensures the validity of (21.27).
Roughly speaking, if we assume, by (a sort of) contradiction, that u is not bounded from
above by M ⌫ k0 on Q1/4 for k0 sufficiently large, then, thanks to Lemma 21.15, we should
be able to find recursively a sequence (xj ) with the property that
u(xj ) M0 ⌫ j and xj+1 2 Qlj (xj );
P
since j lj < 1, the sequence (xj ) admits a converging subsequence, and in the limit
point we find a contradiction. However, in order to iterate Lemma 21.15 we have to
confine the sequence in the cube Q1/2 (for this purpose it is convenient to use the distance
induced by the L1 norm in Rn , whose balls are cubes). P
To achieve this, we fix a universal positive integer j0 such that j j0 lj < 1/4 and we
assume, by contradiction, that there exists a point x0 2 Q1/4 with u(x0 ) M0 ⌫ j0 1 . This
time, the sequence (xk ) we generate iterating Lemma 21.15 is contained in Q1/2 and
u(xk ) M0 ⌫ j0 +k 1
. (21.35)
When k ! 1 in (21.35) we obtain the contradiction. This way, we obtained also an
“explicit” expression of the universal constant in (21.10), in fact we proved that
sup u(x)  M0 ⌫ j0 1
.
x2Q1/4
References
[1] R.Adams: Sobolev spaces. Academic Press, 1975.

[2] G.Alberti, L.Ambrosio: A geometric approach to monotone functions in Rn .

Math. Z, 230 (1999), 259–316.

[3] L.Ambrosio, N.Fusco, D.Pallara: Functions of bounded variation and free dis-
continuity problems. Oxford University Press, 2000.

[4] H.Brezis: Analyse Fonctionelle: Théorie et applications. Masson, Paris, 1983.

[5] L.A.Caffarelli, X.Cabré: Fully nonlinear elliptic equations. Colloquium Publi-

cations, 43 (1995), American Mathematical Society.

[6] E.De Giorgi: Complementi alla teoria della misura (n 1)-dimensionale in uno
spazio n-dimensionale. Seminario di Matematica della Scuola Normale Superiore di
Pisa, (1960-61), Editrice Tecnico Scientifica, Pisa.

[7] E.De Giorgi: Frontiere orientate di misura minima. Seminario di Matematica della
Scuola Normale Superiore di Pisa, (1960-61), Editrice Tecnico Scientifica, Pisa.

[8] E.De Giorgi: Un esempio di estremali discontinue per un problema variazionale di

tipo ellittico. Boll. Un. Mat. Ital. (4), 1 (1968), 135–137.

[9] E.De Giorgi: Sulla di↵erenziabilità e l’analicità degli estremali degli integrali mul-
tipli regolari. Mem. Acc. Sc. Torino, 3 (1957), 25–43.

[10] L.C.Evans: Quasiconvexity and partial regularity in the calculus of variations. Arch.
Rational Mech. Anal. 95, 3 (1986), 227–252.

[11] L.C.Evans, R.F.Gariepy: Measure Theory and Fine Properties of Functions.

Studies in Advanced Mathematics, 1992.

[12] H.Federer: Geometric Measure Theory. Die Grundlehren der mathematischen

Wissenschaft, Band 153, Springer-Verlag New York Inc., 1969.

[13] E.Gagliardo: Caratterizzazione delle tracce sulla frontiera relative ad alcune class
di funzioni in piú variabili. Rend. Sem. Mat. Univ. Padova, 27 (1957), 284–305.

[14] D.Gilbarg, N.S. Trudinger: Elliptic Partial Di↵erential Equations of Second

Order. Springer Verlag, 1983.

[15] E.Giusti: Minimal surfaces and functions of bounded variation. Birhkhäuser,

Boston, 1994.

170
[16] M.Giaquinta, E.Giusti: On the regularity of the minima of variational integrals.
Acta Math. 148, (1982), 31–46.

[17] M.Giaquinta, E.Giusti: Quasiminima. Ann. Inst. H. Poincaré Anal. Non Linéaire
1, 2 (1984), 79–107.

[18] M.Giaquinta, E.Giusti: The singular set of the minima of certain quadratic func-
tionals. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 11, 1 (1984), 45–55.

[19] M.Giaquinta: Multiple integrals in the Calculus of Variations and Nonlinear elliptic
systems. Princeton University Press, 1983.

[20] E.Hopf: Über den funktionalen, insbesondere den analytischen Charakter der
Lösungen elliptischer Di↵erentialgleichungen zweiter Ordnung. Math. Zeitschrift,
Band 34 (1932), 194–233.

[21] F.John, L.Nirenberg: On Functions of Bounded Mean Oscillation. Comm. on

Pure and Applied Math., Vol. XIV (1961), 415–426.

[22] J.Kristensen, G.Mingione: The singular set of minima of integral functionals.

Arch. Ration. Mech. Anal., 180 (2006), 331–398.

[23] The singular set of lipschitzian minima of multiple integrals. Arch. Ration. Mech.
Anal. 184 (2007), 341–369.

[24] N.G.Meyers, J.Serrin: H = W . Proc. Nat. Acad. Sci. U.S.A., 51 (1964), 1055–
1056.

[25] S.Müller, V.Sverak: Convex integration for Lipschitz mappings and counterex-
amples to regularity. Ann. of Math., 157 (2003), 715–742.

[26] E.M.Stein, G.Weiss: Introduction to Fourier Analysis on Euclidean Spaces.

Princeton University Press, 1971.

[27] N.Trudinger: On embedding into Orlicz spaces and some applications. J. Math.
Mech., 17 (1967), 473–483.

[28] K.Yosida: Functional Analysis. Mathematical surveys and monographs, 62, Amer-
ican Mathematical Society, 1998.

171